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), \
106 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
107 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
109 static struct rcu_state *const rcu_state_p;
110 LIST_HEAD(rcu_struct_flavors);
112 /* Dump rcu_node combining tree at boot to verify correct setup. */
113 static bool dump_tree;
114 module_param(dump_tree, bool, 0444);
115 /* Control rcu_node-tree auto-balancing at boot time. */
116 static bool rcu_fanout_exact;
117 module_param(rcu_fanout_exact, bool, 0444);
118 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
119 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
120 module_param(rcu_fanout_leaf, int, 0444);
121 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
122 /* Number of rcu_nodes at specified level. */
123 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
124 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
125 /* panic() on RCU Stall sysctl. */
126 int sysctl_panic_on_rcu_stall __read_mostly;
129 * The rcu_scheduler_active variable is initialized to the value
130 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
131 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
132 * RCU can assume that there is but one task, allowing RCU to (for example)
133 * optimize synchronize_rcu() to a simple barrier(). When this variable
134 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
135 * to detect real grace periods. This variable is also used to suppress
136 * boot-time false positives from lockdep-RCU error checking. Finally, it
137 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
138 * is fully initialized, including all of its kthreads having been spawned.
140 int rcu_scheduler_active __read_mostly;
141 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
144 * The rcu_scheduler_fully_active variable transitions from zero to one
145 * during the early_initcall() processing, which is after the scheduler
146 * is capable of creating new tasks. So RCU processing (for example,
147 * creating tasks for RCU priority boosting) must be delayed until after
148 * rcu_scheduler_fully_active transitions from zero to one. We also
149 * currently delay invocation of any RCU callbacks until after this point.
151 * It might later prove better for people registering RCU callbacks during
152 * early boot to take responsibility for these callbacks, but one step at
155 static int rcu_scheduler_fully_active __read_mostly;
157 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
158 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
159 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
160 static void invoke_rcu_core(void);
161 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
162 static void rcu_report_exp_rdp(struct rcu_state *rsp,
163 struct rcu_data *rdp, bool wake);
164 static void sync_sched_exp_online_cleanup(int cpu);
166 /* rcuc/rcub kthread realtime priority */
167 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
168 module_param(kthread_prio, int, 0644);
170 /* Delay in jiffies for grace-period initialization delays, debug only. */
172 static int gp_preinit_delay;
173 module_param(gp_preinit_delay, int, 0444);
174 static int gp_init_delay;
175 module_param(gp_init_delay, int, 0444);
176 static int gp_cleanup_delay;
177 module_param(gp_cleanup_delay, int, 0444);
180 * Number of grace periods between delays, normalized by the duration of
181 * the delay. The longer the delay, the more the grace periods between
182 * each delay. The reason for this normalization is that it means that,
183 * for non-zero delays, the overall slowdown of grace periods is constant
184 * regardless of the duration of the delay. This arrangement balances
185 * the need for long delays to increase some race probabilities with the
186 * need for fast grace periods to increase other race probabilities.
188 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
191 * Track the rcutorture test sequence number and the update version
192 * number within a given test. The rcutorture_testseq is incremented
193 * on every rcutorture module load and unload, so has an odd value
194 * when a test is running. The rcutorture_vernum is set to zero
195 * when rcutorture starts and is incremented on each rcutorture update.
196 * These variables enable correlating rcutorture output with the
197 * RCU tracing information.
199 unsigned long rcutorture_testseq;
200 unsigned long rcutorture_vernum;
203 * Compute the mask of online CPUs for the specified rcu_node structure.
204 * This will not be stable unless the rcu_node structure's ->lock is
205 * held, but the bit corresponding to the current CPU will be stable
208 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
210 return READ_ONCE(rnp->qsmaskinitnext);
214 * Return true if an RCU grace period is in progress. The READ_ONCE()s
215 * permit this function to be invoked without holding the root rcu_node
216 * structure's ->lock, but of course results can be subject to change.
218 static int rcu_gp_in_progress(struct rcu_state *rsp)
220 return rcu_seq_state(rcu_seq_current(&rsp->gp_seq));
224 * Note a quiescent state. Because we do not need to know
225 * how many quiescent states passed, just if there was at least
226 * one since the start of the grace period, this just sets a flag.
227 * The caller must have disabled preemption.
229 void rcu_sched_qs(void)
231 RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
232 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
234 trace_rcu_grace_period(TPS("rcu_sched"),
235 __this_cpu_read(rcu_sched_data.gp_seq),
237 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
238 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
240 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
241 rcu_report_exp_rdp(&rcu_sched_state,
242 this_cpu_ptr(&rcu_sched_data), true);
247 RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
248 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
249 trace_rcu_grace_period(TPS("rcu_bh"),
250 __this_cpu_read(rcu_bh_data.gp_seq),
252 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
257 * Steal a bit from the bottom of ->dynticks for idle entry/exit
258 * control. Initially this is for TLB flushing.
260 #define RCU_DYNTICK_CTRL_MASK 0x1
261 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
262 #ifndef rcu_eqs_special_exit
263 #define rcu_eqs_special_exit() do { } while (0)
266 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
267 .dynticks_nesting = 1,
268 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
269 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
273 * Record entry into an extended quiescent state. This is only to be
274 * called when not already in an extended quiescent state.
276 static void rcu_dynticks_eqs_enter(void)
278 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
282 * CPUs seeing atomic_add_return() must see prior RCU read-side
283 * critical sections, and we also must force ordering with the
286 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
287 /* Better be in an extended quiescent state! */
288 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
289 (seq & RCU_DYNTICK_CTRL_CTR));
290 /* Better not have special action (TLB flush) pending! */
291 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
292 (seq & RCU_DYNTICK_CTRL_MASK));
296 * Record exit from an extended quiescent state. This is only to be
297 * called from an extended quiescent state.
299 static void rcu_dynticks_eqs_exit(void)
301 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
305 * CPUs seeing atomic_add_return() must see prior idle sojourns,
306 * and we also must force ordering with the next RCU read-side
309 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
310 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
311 !(seq & RCU_DYNTICK_CTRL_CTR));
312 if (seq & RCU_DYNTICK_CTRL_MASK) {
313 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
314 smp_mb__after_atomic(); /* _exit after clearing mask. */
315 /* Prefer duplicate flushes to losing a flush. */
316 rcu_eqs_special_exit();
321 * Reset the current CPU's ->dynticks counter to indicate that the
322 * newly onlined CPU is no longer in an extended quiescent state.
323 * This will either leave the counter unchanged, or increment it
324 * to the next non-quiescent value.
326 * The non-atomic test/increment sequence works because the upper bits
327 * of the ->dynticks counter are manipulated only by the corresponding CPU,
328 * or when the corresponding CPU is offline.
330 static void rcu_dynticks_eqs_online(void)
332 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
334 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
336 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
340 * Is the current CPU in an extended quiescent state?
342 * No ordering, as we are sampling CPU-local information.
344 bool rcu_dynticks_curr_cpu_in_eqs(void)
346 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
348 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
352 * Snapshot the ->dynticks counter with full ordering so as to allow
353 * stable comparison of this counter with past and future snapshots.
355 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
357 int snap = atomic_add_return(0, &rdtp->dynticks);
359 return snap & ~RCU_DYNTICK_CTRL_MASK;
363 * Return true if the snapshot returned from rcu_dynticks_snap()
364 * indicates that RCU is in an extended quiescent state.
366 static bool rcu_dynticks_in_eqs(int snap)
368 return !(snap & RCU_DYNTICK_CTRL_CTR);
372 * Return true if the CPU corresponding to the specified rcu_dynticks
373 * structure has spent some time in an extended quiescent state since
374 * rcu_dynticks_snap() returned the specified snapshot.
376 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
378 return snap != rcu_dynticks_snap(rdtp);
382 * Do a double-increment of the ->dynticks counter to emulate a
383 * momentary idle-CPU quiescent state.
385 static void rcu_dynticks_momentary_idle(void)
387 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
388 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
391 /* It is illegal to call this from idle state. */
392 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
396 * Set the special (bottom) bit of the specified CPU so that it
397 * will take special action (such as flushing its TLB) on the
398 * next exit from an extended quiescent state. Returns true if
399 * the bit was successfully set, or false if the CPU was not in
400 * an extended quiescent state.
402 bool rcu_eqs_special_set(int cpu)
406 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
409 old = atomic_read(&rdtp->dynticks);
410 if (old & RCU_DYNTICK_CTRL_CTR)
412 new = old | RCU_DYNTICK_CTRL_MASK;
413 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
418 * Let the RCU core know that this CPU has gone through the scheduler,
419 * which is a quiescent state. This is called when the need for a
420 * quiescent state is urgent, so we burn an atomic operation and full
421 * memory barriers to let the RCU core know about it, regardless of what
422 * this CPU might (or might not) do in the near future.
424 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
426 * The caller must have disabled interrupts.
428 static void rcu_momentary_dyntick_idle(void)
430 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
431 rcu_dynticks_momentary_idle();
435 * Note a context switch. This is a quiescent state for RCU-sched,
436 * and requires special handling for preemptible RCU.
437 * The caller must have disabled interrupts.
439 void rcu_note_context_switch(bool preempt)
441 barrier(); /* Avoid RCU read-side critical sections leaking down. */
442 trace_rcu_utilization(TPS("Start context switch"));
444 rcu_preempt_note_context_switch(preempt);
445 /* Load rcu_urgent_qs before other flags. */
446 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
448 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
449 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
450 rcu_momentary_dyntick_idle();
451 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
453 rcu_note_voluntary_context_switch_lite(current);
455 trace_rcu_utilization(TPS("End context switch"));
456 barrier(); /* Avoid RCU read-side critical sections leaking up. */
458 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
461 * Register a quiescent state for all RCU flavors. If there is an
462 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
463 * dyntick-idle quiescent state visible to other CPUs (but only for those
464 * RCU flavors in desperate need of a quiescent state, which will normally
465 * be none of them). Either way, do a lightweight quiescent state for
468 * The barrier() calls are redundant in the common case when this is
469 * called externally, but just in case this is called from within this
473 void rcu_all_qs(void)
477 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
480 /* Load rcu_urgent_qs before other flags. */
481 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
485 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
486 barrier(); /* Avoid RCU read-side critical sections leaking down. */
487 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
488 local_irq_save(flags);
489 rcu_momentary_dyntick_idle();
490 local_irq_restore(flags);
492 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
494 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
495 barrier(); /* Avoid RCU read-side critical sections leaking up. */
498 EXPORT_SYMBOL_GPL(rcu_all_qs);
500 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
501 static long blimit = DEFAULT_RCU_BLIMIT;
502 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
503 static long qhimark = DEFAULT_RCU_QHIMARK;
504 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
505 static long qlowmark = DEFAULT_RCU_QLOMARK;
507 module_param(blimit, long, 0444);
508 module_param(qhimark, long, 0444);
509 module_param(qlowmark, long, 0444);
511 static ulong jiffies_till_first_fqs = ULONG_MAX;
512 static ulong jiffies_till_next_fqs = ULONG_MAX;
513 static bool rcu_kick_kthreads;
515 module_param(jiffies_till_first_fqs, ulong, 0644);
516 module_param(jiffies_till_next_fqs, ulong, 0644);
517 module_param(rcu_kick_kthreads, bool, 0644);
520 * How long the grace period must be before we start recruiting
521 * quiescent-state help from rcu_note_context_switch().
523 static ulong jiffies_till_sched_qs = HZ / 10;
524 module_param(jiffies_till_sched_qs, ulong, 0444);
526 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
527 static void force_quiescent_state(struct rcu_state *rsp);
528 static int rcu_pending(void);
531 * Return the number of RCU GPs completed thus far for debug & stats.
533 unsigned long rcu_get_gp_seq(void)
535 return rcu_seq_ctr(READ_ONCE(rcu_state_p->gp_seq));
537 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
540 * Return the number of RCU-sched GPs completed thus far for debug & stats.
542 unsigned long rcu_sched_get_gp_seq(void)
544 return rcu_seq_ctr(READ_ONCE(rcu_sched_state.gp_seq));
546 EXPORT_SYMBOL_GPL(rcu_sched_get_gp_seq);
549 * Return the number of RCU-bh GPs completed thus far for debug & stats.
551 unsigned long rcu_bh_get_gp_seq(void)
553 return rcu_seq_ctr(READ_ONCE(rcu_bh_state.gp_seq));
555 EXPORT_SYMBOL_GPL(rcu_bh_get_gp_seq);
558 * Return the number of RCU expedited batches completed thus far for
559 * debug & stats. Odd numbers mean that a batch is in progress, even
560 * numbers mean idle. The value returned will thus be roughly double
561 * the cumulative batches since boot.
563 unsigned long rcu_exp_batches_completed(void)
565 return rcu_state_p->expedited_sequence;
567 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
570 * Return the number of RCU-sched expedited batches completed thus far
571 * for debug & stats. Similar to rcu_exp_batches_completed().
573 unsigned long rcu_exp_batches_completed_sched(void)
575 return rcu_sched_state.expedited_sequence;
577 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
580 * Force a quiescent state.
582 void rcu_force_quiescent_state(void)
584 force_quiescent_state(rcu_state_p);
586 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
589 * Force a quiescent state for RCU BH.
591 void rcu_bh_force_quiescent_state(void)
593 force_quiescent_state(&rcu_bh_state);
595 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
598 * Force a quiescent state for RCU-sched.
600 void rcu_sched_force_quiescent_state(void)
602 force_quiescent_state(&rcu_sched_state);
604 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
607 * Show the state of the grace-period kthreads.
609 void show_rcu_gp_kthreads(void)
611 struct rcu_state *rsp;
613 for_each_rcu_flavor(rsp) {
614 pr_info("%s: wait state: %d ->state: %#lx\n",
615 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
616 /* sched_show_task(rsp->gp_kthread); */
619 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
622 * Record the number of times rcutorture tests have been initiated and
623 * terminated. This information allows the debugfs tracing stats to be
624 * correlated to the rcutorture messages, even when the rcutorture module
625 * is being repeatedly loaded and unloaded. In other words, we cannot
626 * store this state in rcutorture itself.
628 void rcutorture_record_test_transition(void)
630 rcutorture_testseq++;
631 rcutorture_vernum = 0;
633 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
636 * Send along grace-period-related data for rcutorture diagnostics.
638 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
639 unsigned long *gp_seq)
641 struct rcu_state *rsp = NULL;
650 case RCU_SCHED_FLAVOR:
651 rsp = &rcu_sched_state;
658 *flags = READ_ONCE(rsp->gp_flags);
659 *gp_seq = rcu_seq_current(&rsp->gp_seq);
661 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
664 * Record the number of writer passes through the current rcutorture test.
665 * This is also used to correlate debugfs tracing stats with the rcutorture
668 void rcutorture_record_progress(unsigned long vernum)
672 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
675 * Return the root node of the specified rcu_state structure.
677 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
679 return &rsp->node[0];
683 * Enter an RCU extended quiescent state, which can be either the
684 * idle loop or adaptive-tickless usermode execution.
686 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
687 * the possibility of usermode upcalls having messed up our count
688 * of interrupt nesting level during the prior busy period.
690 static void rcu_eqs_enter(bool user)
692 struct rcu_state *rsp;
693 struct rcu_data *rdp;
694 struct rcu_dynticks *rdtp;
696 rdtp = this_cpu_ptr(&rcu_dynticks);
697 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0);
698 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
699 rdtp->dynticks_nesting == 0);
700 if (rdtp->dynticks_nesting != 1) {
701 rdtp->dynticks_nesting--;
705 lockdep_assert_irqs_disabled();
706 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0, rdtp->dynticks);
707 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
708 for_each_rcu_flavor(rsp) {
709 rdp = this_cpu_ptr(rsp->rda);
710 do_nocb_deferred_wakeup(rdp);
712 rcu_prepare_for_idle();
713 WRITE_ONCE(rdtp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
714 rcu_dynticks_eqs_enter();
715 rcu_dynticks_task_enter();
719 * rcu_idle_enter - inform RCU that current CPU is entering idle
721 * Enter idle mode, in other words, -leave- the mode in which RCU
722 * read-side critical sections can occur. (Though RCU read-side
723 * critical sections can occur in irq handlers in idle, a possibility
724 * handled by irq_enter() and irq_exit().)
726 * If you add or remove a call to rcu_idle_enter(), be sure to test with
727 * CONFIG_RCU_EQS_DEBUG=y.
729 void rcu_idle_enter(void)
731 lockdep_assert_irqs_disabled();
732 rcu_eqs_enter(false);
735 #ifdef CONFIG_NO_HZ_FULL
737 * rcu_user_enter - inform RCU that we are resuming userspace.
739 * Enter RCU idle mode right before resuming userspace. No use of RCU
740 * is permitted between this call and rcu_user_exit(). This way the
741 * CPU doesn't need to maintain the tick for RCU maintenance purposes
742 * when the CPU runs in userspace.
744 * If you add or remove a call to rcu_user_enter(), be sure to test with
745 * CONFIG_RCU_EQS_DEBUG=y.
747 void rcu_user_enter(void)
749 lockdep_assert_irqs_disabled();
752 #endif /* CONFIG_NO_HZ_FULL */
755 * rcu_nmi_exit - inform RCU of exit from NMI context
757 * If we are returning from the outermost NMI handler that interrupted an
758 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
759 * to let the RCU grace-period handling know that the CPU is back to
762 * If you add or remove a call to rcu_nmi_exit(), be sure to test
763 * with CONFIG_RCU_EQS_DEBUG=y.
765 void rcu_nmi_exit(void)
767 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
770 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
771 * (We are exiting an NMI handler, so RCU better be paying attention
774 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
775 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
778 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
779 * leave it in non-RCU-idle state.
781 if (rdtp->dynticks_nmi_nesting != 1) {
782 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nmi_nesting, rdtp->dynticks_nmi_nesting - 2, rdtp->dynticks);
783 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* No store tearing. */
784 rdtp->dynticks_nmi_nesting - 2);
788 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
789 trace_rcu_dyntick(TPS("Startirq"), rdtp->dynticks_nmi_nesting, 0, rdtp->dynticks);
790 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
791 rcu_dynticks_eqs_enter();
795 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
797 * Exit from an interrupt handler, which might possibly result in entering
798 * idle mode, in other words, leaving the mode in which read-side critical
799 * sections can occur. The caller must have disabled interrupts.
801 * This code assumes that the idle loop never does anything that might
802 * result in unbalanced calls to irq_enter() and irq_exit(). If your
803 * architecture's idle loop violates this assumption, RCU will give you what
804 * you deserve, good and hard. But very infrequently and irreproducibly.
806 * Use things like work queues to work around this limitation.
808 * You have been warned.
810 * If you add or remove a call to rcu_irq_exit(), be sure to test with
811 * CONFIG_RCU_EQS_DEBUG=y.
813 void rcu_irq_exit(void)
815 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
817 lockdep_assert_irqs_disabled();
818 if (rdtp->dynticks_nmi_nesting == 1)
819 rcu_prepare_for_idle();
821 if (rdtp->dynticks_nmi_nesting == 0)
822 rcu_dynticks_task_enter();
826 * Wrapper for rcu_irq_exit() where interrupts are enabled.
828 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
829 * with CONFIG_RCU_EQS_DEBUG=y.
831 void rcu_irq_exit_irqson(void)
835 local_irq_save(flags);
837 local_irq_restore(flags);
841 * Exit an RCU extended quiescent state, which can be either the
842 * idle loop or adaptive-tickless usermode execution.
844 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
845 * allow for the possibility of usermode upcalls messing up our count of
846 * interrupt nesting level during the busy period that is just now starting.
848 static void rcu_eqs_exit(bool user)
850 struct rcu_dynticks *rdtp;
853 lockdep_assert_irqs_disabled();
854 rdtp = this_cpu_ptr(&rcu_dynticks);
855 oldval = rdtp->dynticks_nesting;
856 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
858 rdtp->dynticks_nesting++;
861 rcu_dynticks_task_exit();
862 rcu_dynticks_eqs_exit();
863 rcu_cleanup_after_idle();
864 trace_rcu_dyntick(TPS("End"), rdtp->dynticks_nesting, 1, rdtp->dynticks);
865 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
866 WRITE_ONCE(rdtp->dynticks_nesting, 1);
867 WRITE_ONCE(rdtp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
871 * rcu_idle_exit - inform RCU that current CPU is leaving idle
873 * Exit idle mode, in other words, -enter- the mode in which RCU
874 * read-side critical sections can occur.
876 * If you add or remove a call to rcu_idle_exit(), be sure to test with
877 * CONFIG_RCU_EQS_DEBUG=y.
879 void rcu_idle_exit(void)
883 local_irq_save(flags);
885 local_irq_restore(flags);
888 #ifdef CONFIG_NO_HZ_FULL
890 * rcu_user_exit - inform RCU that we are exiting userspace.
892 * Exit RCU idle mode while entering the kernel because it can
893 * run a RCU read side critical section anytime.
895 * If you add or remove a call to rcu_user_exit(), be sure to test with
896 * CONFIG_RCU_EQS_DEBUG=y.
898 void rcu_user_exit(void)
902 #endif /* CONFIG_NO_HZ_FULL */
905 * rcu_nmi_enter - inform RCU of entry to NMI context
907 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
908 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
909 * that the CPU is active. This implementation permits nested NMIs, as
910 * long as the nesting level does not overflow an int. (You will probably
911 * run out of stack space first.)
913 * If you add or remove a call to rcu_nmi_enter(), be sure to test
914 * with CONFIG_RCU_EQS_DEBUG=y.
916 void rcu_nmi_enter(void)
918 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
921 /* Complain about underflow. */
922 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
925 * If idle from RCU viewpoint, atomically increment ->dynticks
926 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
927 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
928 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
929 * to be in the outermost NMI handler that interrupted an RCU-idle
930 * period (observation due to Andy Lutomirski).
932 if (rcu_dynticks_curr_cpu_in_eqs()) {
933 rcu_dynticks_eqs_exit();
936 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
937 rdtp->dynticks_nmi_nesting,
938 rdtp->dynticks_nmi_nesting + incby, rdtp->dynticks);
939 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* Prevent store tearing. */
940 rdtp->dynticks_nmi_nesting + incby);
945 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
947 * Enter an interrupt handler, which might possibly result in exiting
948 * idle mode, in other words, entering the mode in which read-side critical
949 * sections can occur. The caller must have disabled interrupts.
951 * Note that the Linux kernel is fully capable of entering an interrupt
952 * handler that it never exits, for example when doing upcalls to user mode!
953 * This code assumes that the idle loop never does upcalls to user mode.
954 * If your architecture's idle loop does do upcalls to user mode (or does
955 * anything else that results in unbalanced calls to the irq_enter() and
956 * irq_exit() functions), RCU will give you what you deserve, good and hard.
957 * But very infrequently and irreproducibly.
959 * Use things like work queues to work around this limitation.
961 * You have been warned.
963 * If you add or remove a call to rcu_irq_enter(), be sure to test with
964 * CONFIG_RCU_EQS_DEBUG=y.
966 void rcu_irq_enter(void)
968 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
970 lockdep_assert_irqs_disabled();
971 if (rdtp->dynticks_nmi_nesting == 0)
972 rcu_dynticks_task_exit();
974 if (rdtp->dynticks_nmi_nesting == 1)
975 rcu_cleanup_after_idle();
979 * Wrapper for rcu_irq_enter() where interrupts are enabled.
981 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
982 * with CONFIG_RCU_EQS_DEBUG=y.
984 void rcu_irq_enter_irqson(void)
988 local_irq_save(flags);
990 local_irq_restore(flags);
994 * rcu_is_watching - see if RCU thinks that the current CPU is idle
996 * Return true if RCU is watching the running CPU, which means that this
997 * CPU can safely enter RCU read-side critical sections. In other words,
998 * if the current CPU is in its idle loop and is neither in an interrupt
999 * or NMI handler, return true.
1001 bool notrace rcu_is_watching(void)
1005 preempt_disable_notrace();
1006 ret = !rcu_dynticks_curr_cpu_in_eqs();
1007 preempt_enable_notrace();
1010 EXPORT_SYMBOL_GPL(rcu_is_watching);
1013 * If a holdout task is actually running, request an urgent quiescent
1014 * state from its CPU. This is unsynchronized, so migrations can cause
1015 * the request to go to the wrong CPU. Which is OK, all that will happen
1016 * is that the CPU's next context switch will be a bit slower and next
1017 * time around this task will generate another request.
1019 void rcu_request_urgent_qs_task(struct task_struct *t)
1026 return; /* This task is not running on that CPU. */
1027 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1030 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1033 * Is the current CPU online? Disable preemption to avoid false positives
1034 * that could otherwise happen due to the current CPU number being sampled,
1035 * this task being preempted, its old CPU being taken offline, resuming
1036 * on some other CPU, then determining that its old CPU is now offline.
1037 * It is OK to use RCU on an offline processor during initial boot, hence
1038 * the check for rcu_scheduler_fully_active. Note also that it is OK
1039 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1040 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1041 * offline to continue to use RCU for one jiffy after marking itself
1042 * offline in the cpu_online_mask. This leniency is necessary given the
1043 * non-atomic nature of the online and offline processing, for example,
1044 * the fact that a CPU enters the scheduler after completing the teardown
1047 * This is also why RCU internally marks CPUs online during in the
1048 * preparation phase and offline after the CPU has been taken down.
1050 * Disable checking if in an NMI handler because we cannot safely report
1051 * errors from NMI handlers anyway.
1053 bool rcu_lockdep_current_cpu_online(void)
1055 struct rcu_data *rdp;
1056 struct rcu_node *rnp;
1062 rdp = this_cpu_ptr(&rcu_sched_data);
1064 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1065 !rcu_scheduler_fully_active;
1069 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1071 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1074 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1076 * If the current CPU is idle or running at a first-level (not nested)
1077 * interrupt from idle, return true. The caller must have at least
1078 * disabled preemption.
1080 static int rcu_is_cpu_rrupt_from_idle(void)
1082 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 0 &&
1083 __this_cpu_read(rcu_dynticks.dynticks_nmi_nesting) <= 1;
1087 * We are reporting a quiescent state on behalf of some other CPU, so
1088 * it is our responsibility to check for and handle potential overflow
1089 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1090 * After all, the CPU might be in deep idle state, and thus executing no
1093 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1095 raw_lockdep_assert_held_rcu_node(rnp);
1096 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1098 WRITE_ONCE(rdp->gpwrap, true);
1099 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1100 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1104 * Snapshot the specified CPU's dynticks counter so that we can later
1105 * credit them with an implicit quiescent state. Return 1 if this CPU
1106 * is in dynticks idle mode, which is an extended quiescent state.
1108 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1110 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1111 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1112 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1113 rcu_gpnum_ovf(rdp->mynode, rdp);
1120 * Handler for the irq_work request posted when a grace period has
1121 * gone on for too long, but not yet long enough for an RCU CPU
1122 * stall warning. Set state appropriately, but just complain if
1123 * there is unexpected state on entry.
1125 static void rcu_iw_handler(struct irq_work *iwp)
1127 struct rcu_data *rdp;
1128 struct rcu_node *rnp;
1130 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1132 raw_spin_lock_rcu_node(rnp);
1133 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1134 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1135 rdp->rcu_iw_pending = false;
1137 raw_spin_unlock_rcu_node(rnp);
1141 * Return true if the specified CPU has passed through a quiescent
1142 * state by virtue of being in or having passed through an dynticks
1143 * idle state since the last call to dyntick_save_progress_counter()
1144 * for this same CPU, or by virtue of having been offline.
1146 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1151 struct rcu_node *rnp = rdp->mynode;
1154 * If the CPU passed through or entered a dynticks idle phase with
1155 * no active irq/NMI handlers, then we can safely pretend that the CPU
1156 * already acknowledged the request to pass through a quiescent
1157 * state. Either way, that CPU cannot possibly be in an RCU
1158 * read-side critical section that started before the beginning
1159 * of the current RCU grace period.
1161 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1162 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1163 rdp->dynticks_fqs++;
1164 rcu_gpnum_ovf(rnp, rdp);
1169 * Has this CPU encountered a cond_resched() since the beginning
1170 * of the grace period? For this to be the case, the CPU has to
1171 * have noticed the current grace period. This might not be the
1172 * case for nohz_full CPUs looping in the kernel.
1174 jtsq = jiffies_till_sched_qs;
1175 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1176 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1177 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1178 rcu_seq_current(&rdp->gp_seq) == rnp->gp_seq && !rdp->gpwrap) {
1179 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("rqc"));
1180 rcu_gpnum_ovf(rnp, rdp);
1182 } else if (time_after(jiffies, rdp->rsp->gp_start + jtsq)) {
1183 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1184 smp_store_release(ruqp, true);
1187 /* Check for the CPU being offline. */
1188 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1189 trace_rcu_fqs(rdp->rsp->name, rdp->gp_seq, rdp->cpu, TPS("ofl"));
1191 rcu_gpnum_ovf(rnp, rdp);
1196 * A CPU running for an extended time within the kernel can
1197 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1198 * even context-switching back and forth between a pair of
1199 * in-kernel CPU-bound tasks cannot advance grace periods.
1200 * So if the grace period is old enough, make the CPU pay attention.
1201 * Note that the unsynchronized assignments to the per-CPU
1202 * rcu_need_heavy_qs variable are safe. Yes, setting of
1203 * bits can be lost, but they will be set again on the next
1204 * force-quiescent-state pass. So lost bit sets do not result
1205 * in incorrect behavior, merely in a grace period lasting
1206 * a few jiffies longer than it might otherwise. Because
1207 * there are at most four threads involved, and because the
1208 * updates are only once every few jiffies, the probability of
1209 * lossage (and thus of slight grace-period extension) is
1212 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1213 if (!READ_ONCE(*rnhqp) &&
1214 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1215 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1216 WRITE_ONCE(*rnhqp, true);
1217 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1218 smp_store_release(ruqp, true);
1219 rdp->rsp->jiffies_resched += jtsq; /* Re-enable beating. */
1223 * If more than halfway to RCU CPU stall-warning time, do a
1224 * resched_cpu() to try to loosen things up a bit. Also check to
1225 * see if the CPU is getting hammered with interrupts, but only
1226 * once per grace period, just to keep the IPIs down to a dull roar.
1228 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) {
1229 resched_cpu(rdp->cpu);
1230 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1231 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1232 (rnp->ffmask & rdp->grpmask)) {
1233 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1234 rdp->rcu_iw_pending = true;
1235 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1236 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1243 static void record_gp_stall_check_time(struct rcu_state *rsp)
1245 unsigned long j = jiffies;
1249 smp_wmb(); /* Record start time before stall time. */
1250 j1 = rcu_jiffies_till_stall_check();
1251 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1252 rsp->jiffies_resched = j + j1 / 2;
1253 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1257 * Convert a ->gp_state value to a character string.
1259 static const char *gp_state_getname(short gs)
1261 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1263 return gp_state_names[gs];
1267 * Complain about starvation of grace-period kthread.
1269 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1275 gpa = READ_ONCE(rsp->gp_activity);
1276 if (j - gpa > 2 * HZ) {
1277 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1279 (long)rcu_seq_current(&rsp->gp_seq),
1281 gp_state_getname(rsp->gp_state), rsp->gp_state,
1282 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1283 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1284 if (rsp->gp_kthread) {
1285 pr_err("RCU grace-period kthread stack dump:\n");
1286 sched_show_task(rsp->gp_kthread);
1287 wake_up_process(rsp->gp_kthread);
1293 * Dump stacks of all tasks running on stalled CPUs. First try using
1294 * NMIs, but fall back to manual remote stack tracing on architectures
1295 * that don't support NMI-based stack dumps. The NMI-triggered stack
1296 * traces are more accurate because they are printed by the target CPU.
1298 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1301 unsigned long flags;
1302 struct rcu_node *rnp;
1304 rcu_for_each_leaf_node(rsp, rnp) {
1305 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1306 for_each_leaf_node_possible_cpu(rnp, cpu)
1307 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1308 if (!trigger_single_cpu_backtrace(cpu))
1310 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1315 * If too much time has passed in the current grace period, and if
1316 * so configured, go kick the relevant kthreads.
1318 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1322 if (!rcu_kick_kthreads)
1324 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1325 if (time_after(jiffies, j) && rsp->gp_kthread &&
1326 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1327 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1328 rcu_ftrace_dump(DUMP_ALL);
1329 wake_up_process(rsp->gp_kthread);
1330 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1334 static inline void panic_on_rcu_stall(void)
1336 if (sysctl_panic_on_rcu_stall)
1337 panic("RCU Stall\n");
1340 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gp_seq)
1343 unsigned long flags;
1347 struct rcu_node *rnp = rcu_get_root(rsp);
1350 /* Kick and suppress, if so configured. */
1351 rcu_stall_kick_kthreads(rsp);
1352 if (rcu_cpu_stall_suppress)
1356 * OK, time to rat on our buddy...
1357 * See Documentation/RCU/stallwarn.txt for info on how to debug
1358 * RCU CPU stall warnings.
1360 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1362 print_cpu_stall_info_begin();
1363 rcu_for_each_leaf_node(rsp, rnp) {
1364 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1365 ndetected += rcu_print_task_stall(rnp);
1366 if (rnp->qsmask != 0) {
1367 for_each_leaf_node_possible_cpu(rnp, cpu)
1368 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1369 print_cpu_stall_info(rsp, cpu);
1373 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1376 print_cpu_stall_info_end();
1377 for_each_possible_cpu(cpu)
1378 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1380 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, q=%lu)\n",
1381 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1382 (long)rcu_seq_current(&rsp->gp_seq), totqlen);
1384 rcu_dump_cpu_stacks(rsp);
1386 /* Complain about tasks blocking the grace period. */
1387 rcu_print_detail_task_stall(rsp);
1389 if (rcu_seq_current(&rsp->gp_seq) != gp_seq) {
1390 pr_err("INFO: Stall ended before state dump start\n");
1393 gpa = READ_ONCE(rsp->gp_activity);
1394 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1395 rsp->name, j - gpa, j, gpa,
1396 jiffies_till_next_fqs,
1397 rcu_get_root(rsp)->qsmask);
1398 /* In this case, the current CPU might be at fault. */
1399 sched_show_task(current);
1402 /* Rewrite if needed in case of slow consoles. */
1403 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1404 WRITE_ONCE(rsp->jiffies_stall,
1405 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1407 rcu_check_gp_kthread_starvation(rsp);
1409 panic_on_rcu_stall();
1411 force_quiescent_state(rsp); /* Kick them all. */
1414 static void print_cpu_stall(struct rcu_state *rsp)
1417 unsigned long flags;
1418 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1419 struct rcu_node *rnp = rcu_get_root(rsp);
1422 /* Kick and suppress, if so configured. */
1423 rcu_stall_kick_kthreads(rsp);
1424 if (rcu_cpu_stall_suppress)
1428 * OK, time to rat on ourselves...
1429 * See Documentation/RCU/stallwarn.txt for info on how to debug
1430 * RCU CPU stall warnings.
1432 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1433 print_cpu_stall_info_begin();
1434 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1435 print_cpu_stall_info(rsp, smp_processor_id());
1436 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1437 print_cpu_stall_info_end();
1438 for_each_possible_cpu(cpu)
1439 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1441 pr_cont(" (t=%lu jiffies g=%ld q=%lu)\n",
1442 jiffies - rsp->gp_start,
1443 (long)rcu_seq_current(&rsp->gp_seq), totqlen);
1445 rcu_check_gp_kthread_starvation(rsp);
1447 rcu_dump_cpu_stacks(rsp);
1449 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1450 /* Rewrite if needed in case of slow consoles. */
1451 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1452 WRITE_ONCE(rsp->jiffies_stall,
1453 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1454 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1456 panic_on_rcu_stall();
1459 * Attempt to revive the RCU machinery by forcing a context switch.
1461 * A context switch would normally allow the RCU state machine to make
1462 * progress and it could be we're stuck in kernel space without context
1463 * switches for an entirely unreasonable amount of time.
1465 resched_cpu(smp_processor_id());
1468 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1476 struct rcu_node *rnp;
1478 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1479 !rcu_gp_in_progress(rsp))
1481 rcu_stall_kick_kthreads(rsp);
1485 * Lots of memory barriers to reject false positives.
1487 * The idea is to pick up rsp->gp_seq, then rsp->jiffies_stall,
1488 * then rsp->gp_start, and finally another copy of rsp->gp_seq.
1489 * These values are updated in the opposite order with memory
1490 * barriers (or equivalent) during grace-period initialization
1491 * and cleanup. Now, a false positive can occur if we get an new
1492 * value of rsp->gp_start and a old value of rsp->jiffies_stall.
1493 * But given the memory barriers, the only way that this can happen
1494 * is if one grace period ends and another starts between these
1495 * two fetches. This is detected by comparing the second fetch
1496 * of rsp->gp_seq with the previous fetch from rsp->gp_seq.
1498 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1499 * and rsp->gp_start suffice to forestall false positives.
1501 gs1 = READ_ONCE(rsp->gp_seq);
1502 smp_rmb(); /* Pick up ->gp_seq first... */
1503 js = READ_ONCE(rsp->jiffies_stall);
1504 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1505 gps = READ_ONCE(rsp->gp_start);
1506 smp_rmb(); /* ...and finally ->gp_start before ->gp_seq again. */
1507 gs2 = READ_ONCE(rsp->gp_seq);
1509 ULONG_CMP_LT(j, js) ||
1510 ULONG_CMP_GE(gps, js))
1511 return; /* No stall or GP completed since entering function. */
1513 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
1514 if (rcu_gp_in_progress(rsp) &&
1515 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
1516 cmpxchg(&rsp->jiffies_stall, js, jn) == js) {
1518 /* We haven't checked in, so go dump stack. */
1519 print_cpu_stall(rsp);
1521 } else if (rcu_gp_in_progress(rsp) &&
1522 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
1523 cmpxchg(&rsp->jiffies_stall, js, jn) == js) {
1525 /* They had a few time units to dump stack, so complain. */
1526 print_other_cpu_stall(rsp, gs2);
1531 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1533 * Set the stall-warning timeout way off into the future, thus preventing
1534 * any RCU CPU stall-warning messages from appearing in the current set of
1535 * RCU grace periods.
1537 * The caller must disable hard irqs.
1539 void rcu_cpu_stall_reset(void)
1541 struct rcu_state *rsp;
1543 for_each_rcu_flavor(rsp)
1544 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1547 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1548 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1549 unsigned long c, const char *s)
1551 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gp_seq, c,
1552 rnp->level, rnp->grplo, rnp->grphi, s);
1556 * Start the specified grace period, as needed to handle newly arrived
1557 * callbacks. The required future grace periods are recorded in each
1558 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1559 * is reason to awaken the grace-period kthread.
1561 * The caller must hold the specified rcu_node structure's ->lock, which
1562 * is why the caller is responsible for waking the grace-period kthread.
1564 static bool rcu_start_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1568 struct rcu_state *rsp = rdp->rsp;
1569 struct rcu_node *rnp_root;
1572 * Use funnel locking to either acquire the root rcu_node
1573 * structure's lock or bail out if the need for this grace period
1574 * has already been recorded -- or has already started. If there
1575 * is already a grace period in progress in a non-leaf node, no
1576 * recording is needed because the end of the grace period will
1577 * scan the leaf rcu_node structures. Note that rnp->lock must
1580 raw_lockdep_assert_held_rcu_node(rnp);
1581 trace_rcu_this_gp(rnp, rdp, c, TPS("Startleaf"));
1582 for (rnp_root = rnp; 1; rnp_root = rnp_root->parent) {
1583 if (rnp_root != rnp)
1584 raw_spin_lock_rcu_node(rnp_root);
1585 if (ULONG_CMP_GE(rnp_root->gp_seq_needed, c) ||
1586 rcu_seq_done(&rnp_root->gp_seq, c) ||
1588 rcu_seq_state(rcu_seq_current(&rnp_root->gp_seq)))) {
1589 trace_rcu_this_gp(rnp_root, rdp, c, TPS("Prestarted"));
1592 rnp_root->gp_seq_needed = c;
1593 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1595 * We just marked the leaf, and a grace period
1596 * is in progress, which means that rcu_gp_cleanup()
1597 * will see the marking. Bail to reduce contention.
1599 trace_rcu_this_gp(rnp, rdp, c, TPS("Startedleaf"));
1602 if (rnp_root != rnp && rnp_root->parent != NULL)
1603 raw_spin_unlock_rcu_node(rnp_root);
1604 if (!rnp_root->parent)
1605 break; /* At root, and perhaps also leaf. */
1608 /* If GP already in progress, just leave, otherwise start one. */
1609 if (rcu_gp_in_progress(rsp)) {
1610 trace_rcu_this_gp(rnp_root, rdp, c, TPS("Startedleafroot"));
1613 trace_rcu_this_gp(rnp_root, rdp, c, TPS("Startedroot"));
1614 WRITE_ONCE(rsp->gp_flags, rsp->gp_flags | RCU_GP_FLAG_INIT);
1615 rsp->gp_req_activity = jiffies;
1616 if (!rsp->gp_kthread) {
1617 trace_rcu_this_gp(rnp_root, rdp, c, TPS("NoGPkthread"));
1620 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gp_seq), TPS("newreq"));
1621 ret = true; /* Caller must wake GP kthread. */
1623 /* Push furthest requested GP to leaf node and rcu_data structure. */
1624 if (ULONG_CMP_LT(c, rnp_root->gp_seq_needed)) {
1625 rnp->gp_seq_needed = rnp_root->gp_seq_needed;
1626 rdp->gp_seq_needed = rnp_root->gp_seq_needed;
1628 if (rnp != rnp_root)
1629 raw_spin_unlock_rcu_node(rnp_root);
1634 * Clean up any old requests for the just-ended grace period. Also return
1635 * whether any additional grace periods have been requested.
1637 static bool rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1639 unsigned long c = rnp->gp_seq;
1641 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1643 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1645 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1646 trace_rcu_this_gp(rnp, rdp, c,
1647 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1652 * Awaken the grace-period kthread for the specified flavor of RCU.
1653 * Don't do a self-awaken, and don't bother awakening when there is
1654 * nothing for the grace-period kthread to do (as in several CPUs
1655 * raced to awaken, and we lost), and finally don't try to awaken
1656 * a kthread that has not yet been created.
1658 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1660 if (current == rsp->gp_kthread ||
1661 !READ_ONCE(rsp->gp_flags) ||
1664 swake_up(&rsp->gp_wq);
1668 * If there is room, assign a ->gp_seq number to any callbacks on this
1669 * CPU that have not already been assigned. Also accelerate any callbacks
1670 * that were previously assigned a ->gp_seq number that has since proven
1671 * to be too conservative, which can happen if callbacks get assigned a
1672 * ->gp_seq number while RCU is idle, but with reference to a non-root
1673 * rcu_node structure. This function is idempotent, so it does not hurt
1674 * to call it repeatedly. Returns an flag saying that we should awaken
1675 * the RCU grace-period kthread.
1677 * The caller must hold rnp->lock with interrupts disabled.
1679 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1680 struct rcu_data *rdp)
1685 raw_lockdep_assert_held_rcu_node(rnp);
1687 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1688 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1692 * Callbacks are often registered with incomplete grace-period
1693 * information. Something about the fact that getting exact
1694 * information requires acquiring a global lock... RCU therefore
1695 * makes a conservative estimate of the grace period number at which
1696 * a given callback will become ready to invoke. The following
1697 * code checks this estimate and improves it when possible, thus
1698 * accelerating callback invocation to an earlier grace-period
1701 c = rcu_seq_snap(&rsp->gp_seq);
1702 if (rcu_segcblist_accelerate(&rdp->cblist, c))
1703 ret = rcu_start_this_gp(rnp, rdp, c);
1705 /* Trace depending on how much we were able to accelerate. */
1706 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1707 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("AccWaitCB"));
1709 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("AccReadyCB"));
1714 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1715 * rcu_node structure's ->lock be held. It consults the cached value
1716 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1717 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1718 * while holding the leaf rcu_node structure's ->lock.
1720 static void rcu_accelerate_cbs_unlocked(struct rcu_state *rsp,
1721 struct rcu_node *rnp,
1722 struct rcu_data *rdp)
1727 lockdep_assert_irqs_disabled();
1728 c = rcu_seq_snap(&rsp->gp_seq);
1729 if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1730 /* Old request still live, so mark recent callbacks. */
1731 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1734 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1735 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1736 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1738 rcu_gp_kthread_wake(rsp);
1742 * Move any callbacks whose grace period has completed to the
1743 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1744 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1745 * sublist. This function is idempotent, so it does not hurt to
1746 * invoke it repeatedly. As long as it is not invoked -too- often...
1747 * Returns true if the RCU grace-period kthread needs to be awakened.
1749 * The caller must hold rnp->lock with interrupts disabled.
1751 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1752 struct rcu_data *rdp)
1754 raw_lockdep_assert_held_rcu_node(rnp);
1756 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1757 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1761 * Find all callbacks whose ->gp_seq numbers indicate that they
1762 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1764 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1766 /* Classify any remaining callbacks. */
1767 return rcu_accelerate_cbs(rsp, rnp, rdp);
1771 * Update CPU-local rcu_data state to record the beginnings and ends of
1772 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1773 * structure corresponding to the current CPU, and must have irqs disabled.
1774 * Returns true if the grace-period kthread needs to be awakened.
1776 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1777 struct rcu_data *rdp)
1782 raw_lockdep_assert_held_rcu_node(rnp);
1784 if (rdp->gp_seq == rnp->gp_seq)
1785 return false; /* Nothing to do. */
1787 /* Handle the ends of any preceding grace periods first. */
1788 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1789 unlikely(READ_ONCE(rdp->gpwrap))) {
1790 ret = rcu_advance_cbs(rsp, rnp, rdp); /* Advance callbacks. */
1791 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("cpuend"));
1793 ret = rcu_accelerate_cbs(rsp, rnp, rdp); /* Recent callbacks. */
1796 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1797 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1798 unlikely(READ_ONCE(rdp->gpwrap))) {
1800 * If the current grace period is waiting for this CPU,
1801 * set up to detect a quiescent state, otherwise don't
1802 * go looking for one.
1804 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("cpustart"));
1805 need_gp = !!(rnp->qsmask & rdp->grpmask);
1806 rdp->cpu_no_qs.b.norm = need_gp;
1807 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1808 rdp->core_needs_qs = need_gp;
1809 zero_cpu_stall_ticks(rdp);
1810 WRITE_ONCE(rdp->gpwrap, false);
1811 rcu_gpnum_ovf(rnp, rdp);
1813 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1817 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1819 unsigned long flags;
1821 struct rcu_node *rnp;
1823 local_irq_save(flags);
1825 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1826 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1827 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1828 local_irq_restore(flags);
1831 needwake = __note_gp_changes(rsp, rnp, rdp);
1832 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1834 rcu_gp_kthread_wake(rsp);
1837 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1840 !(rcu_seq_ctr(rsp->gp_seq) %
1841 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1842 schedule_timeout_uninterruptible(delay);
1846 * Initialize a new grace period. Return false if no grace period required.
1848 static bool rcu_gp_init(struct rcu_state *rsp)
1850 unsigned long oldmask;
1851 struct rcu_data *rdp;
1852 struct rcu_node *rnp = rcu_get_root(rsp);
1854 WRITE_ONCE(rsp->gp_activity, jiffies);
1855 raw_spin_lock_irq_rcu_node(rnp);
1856 if (!READ_ONCE(rsp->gp_flags)) {
1857 /* Spurious wakeup, tell caller to go back to sleep. */
1858 raw_spin_unlock_irq_rcu_node(rnp);
1861 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1863 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1865 * Grace period already in progress, don't start another.
1866 * Not supposed to be able to happen.
1868 raw_spin_unlock_irq_rcu_node(rnp);
1872 /* Advance to a new grace period and initialize state. */
1873 record_gp_stall_check_time(rsp);
1874 /* Record GP times before starting GP, hence rcu_seq_start(). */
1875 rcu_seq_start(&rsp->gp_seq);
1876 trace_rcu_grace_period(rsp->name, rsp->gp_seq, TPS("start"));
1877 raw_spin_unlock_irq_rcu_node(rnp);
1880 * Apply per-leaf buffered online and offline operations to the
1881 * rcu_node tree. Note that this new grace period need not wait
1882 * for subsequent online CPUs, and that quiescent-state forcing
1883 * will handle subsequent offline CPUs.
1885 rcu_for_each_leaf_node(rsp, rnp) {
1886 rcu_gp_slow(rsp, gp_preinit_delay);
1887 raw_spin_lock_irq_rcu_node(rnp);
1888 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1889 !rnp->wait_blkd_tasks) {
1890 /* Nothing to do on this leaf rcu_node structure. */
1891 raw_spin_unlock_irq_rcu_node(rnp);
1895 /* Record old state, apply changes to ->qsmaskinit field. */
1896 oldmask = rnp->qsmaskinit;
1897 rnp->qsmaskinit = rnp->qsmaskinitnext;
1899 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1900 if (!oldmask != !rnp->qsmaskinit) {
1901 if (!oldmask) /* First online CPU for this rcu_node. */
1902 rcu_init_new_rnp(rnp);
1903 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1904 rnp->wait_blkd_tasks = true;
1905 else /* Last offline CPU and can propagate. */
1906 rcu_cleanup_dead_rnp(rnp);
1910 * If all waited-on tasks from prior grace period are
1911 * done, and if all this rcu_node structure's CPUs are
1912 * still offline, propagate up the rcu_node tree and
1913 * clear ->wait_blkd_tasks. Otherwise, if one of this
1914 * rcu_node structure's CPUs has since come back online,
1915 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1916 * checks for this, so just call it unconditionally).
1918 if (rnp->wait_blkd_tasks &&
1919 (!rcu_preempt_has_tasks(rnp) ||
1921 rnp->wait_blkd_tasks = false;
1922 rcu_cleanup_dead_rnp(rnp);
1925 raw_spin_unlock_irq_rcu_node(rnp);
1929 * Set the quiescent-state-needed bits in all the rcu_node
1930 * structures for all currently online CPUs in breadth-first order,
1931 * starting from the root rcu_node structure, relying on the layout
1932 * of the tree within the rsp->node[] array. Note that other CPUs
1933 * will access only the leaves of the hierarchy, thus seeing that no
1934 * grace period is in progress, at least until the corresponding
1935 * leaf node has been initialized.
1937 * The grace period cannot complete until the initialization
1938 * process finishes, because this kthread handles both.
1940 rcu_for_each_node_breadth_first(rsp, rnp) {
1941 rcu_gp_slow(rsp, gp_init_delay);
1942 raw_spin_lock_irq_rcu_node(rnp);
1943 rdp = this_cpu_ptr(rsp->rda);
1944 rcu_preempt_check_blocked_tasks(rnp);
1945 rnp->qsmask = rnp->qsmaskinit;
1946 WRITE_ONCE(rnp->gp_seq, rsp->gp_seq);
1947 if (rnp == rdp->mynode)
1948 (void)__note_gp_changes(rsp, rnp, rdp);
1949 rcu_preempt_boost_start_gp(rnp);
1950 trace_rcu_grace_period_init(rsp->name, rnp->gp_seq,
1951 rnp->level, rnp->grplo,
1952 rnp->grphi, rnp->qsmask);
1953 raw_spin_unlock_irq_rcu_node(rnp);
1954 cond_resched_tasks_rcu_qs();
1955 WRITE_ONCE(rsp->gp_activity, jiffies);
1962 * Helper function for swait_event_idle() wakeup at force-quiescent-state
1965 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
1967 struct rcu_node *rnp = rcu_get_root(rsp);
1969 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1970 *gfp = READ_ONCE(rsp->gp_flags);
1971 if (*gfp & RCU_GP_FLAG_FQS)
1974 /* The current grace period has completed. */
1975 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1982 * Do one round of quiescent-state forcing.
1984 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
1986 struct rcu_node *rnp = rcu_get_root(rsp);
1988 WRITE_ONCE(rsp->gp_activity, jiffies);
1991 /* Collect dyntick-idle snapshots. */
1992 force_qs_rnp(rsp, dyntick_save_progress_counter);
1994 /* Handle dyntick-idle and offline CPUs. */
1995 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1997 /* Clear flag to prevent immediate re-entry. */
1998 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1999 raw_spin_lock_irq_rcu_node(rnp);
2000 WRITE_ONCE(rsp->gp_flags,
2001 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2002 raw_spin_unlock_irq_rcu_node(rnp);
2007 * Clean up after the old grace period.
2009 static void rcu_gp_cleanup(struct rcu_state *rsp)
2011 unsigned long gp_duration;
2012 bool needgp = false;
2013 unsigned long new_gp_seq;
2014 struct rcu_data *rdp;
2015 struct rcu_node *rnp = rcu_get_root(rsp);
2016 struct swait_queue_head *sq;
2018 WRITE_ONCE(rsp->gp_activity, jiffies);
2019 raw_spin_lock_irq_rcu_node(rnp);
2020 gp_duration = jiffies - rsp->gp_start;
2021 if (gp_duration > rsp->gp_max)
2022 rsp->gp_max = gp_duration;
2025 * We know the grace period is complete, but to everyone else
2026 * it appears to still be ongoing. But it is also the case
2027 * that to everyone else it looks like there is nothing that
2028 * they can do to advance the grace period. It is therefore
2029 * safe for us to drop the lock in order to mark the grace
2030 * period as completed in all of the rcu_node structures.
2032 raw_spin_unlock_irq_rcu_node(rnp);
2035 * Propagate new ->gp_seq value to rcu_node structures so that
2036 * other CPUs don't have to wait until the start of the next grace
2037 * period to process their callbacks. This also avoids some nasty
2038 * RCU grace-period initialization races by forcing the end of
2039 * the current grace period to be completely recorded in all of
2040 * the rcu_node structures before the beginning of the next grace
2041 * period is recorded in any of the rcu_node structures.
2043 new_gp_seq = rsp->gp_seq;
2044 rcu_seq_end(&new_gp_seq);
2045 rcu_for_each_node_breadth_first(rsp, rnp) {
2046 raw_spin_lock_irq_rcu_node(rnp);
2047 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2048 dump_blkd_tasks(rnp, 10);
2049 WARN_ON_ONCE(rnp->qsmask);
2050 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2051 rdp = this_cpu_ptr(rsp->rda);
2052 if (rnp == rdp->mynode)
2053 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2054 /* smp_mb() provided by prior unlock-lock pair. */
2055 needgp = rcu_future_gp_cleanup(rsp, rnp) || needgp;
2056 sq = rcu_nocb_gp_get(rnp);
2057 raw_spin_unlock_irq_rcu_node(rnp);
2058 rcu_nocb_gp_cleanup(sq);
2059 cond_resched_tasks_rcu_qs();
2060 WRITE_ONCE(rsp->gp_activity, jiffies);
2061 rcu_gp_slow(rsp, gp_cleanup_delay);
2063 rnp = rcu_get_root(rsp);
2064 raw_spin_lock_irq_rcu_node(rnp); /* GP before rsp->gp_seq update. */
2066 /* Declare grace period done. */
2067 rcu_seq_end(&rsp->gp_seq);
2068 trace_rcu_grace_period(rsp->name, rsp->gp_seq, TPS("end"));
2069 rsp->gp_state = RCU_GP_IDLE;
2070 /* Check for GP requests since above loop. */
2071 rdp = this_cpu_ptr(rsp->rda);
2072 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2073 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2074 TPS("CleanupMore"));
2077 /* Advance CBs to reduce false positives below. */
2078 if (!rcu_accelerate_cbs(rsp, rnp, rdp) && needgp) {
2079 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2080 rsp->gp_req_activity = jiffies;
2081 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gp_seq),
2084 WRITE_ONCE(rsp->gp_flags, rsp->gp_flags & RCU_GP_FLAG_INIT);
2086 raw_spin_unlock_irq_rcu_node(rnp);
2090 * Body of kthread that handles grace periods.
2092 static int __noreturn rcu_gp_kthread(void *arg)
2098 struct rcu_state *rsp = arg;
2099 struct rcu_node *rnp = rcu_get_root(rsp);
2101 rcu_bind_gp_kthread();
2104 /* Handle grace-period start. */
2106 trace_rcu_grace_period(rsp->name,
2107 READ_ONCE(rsp->gp_seq),
2109 rsp->gp_state = RCU_GP_WAIT_GPS;
2110 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2112 rsp->gp_state = RCU_GP_DONE_GPS;
2113 /* Locking provides needed memory barrier. */
2114 if (rcu_gp_init(rsp))
2116 cond_resched_tasks_rcu_qs();
2117 WRITE_ONCE(rsp->gp_activity, jiffies);
2118 WARN_ON(signal_pending(current));
2119 trace_rcu_grace_period(rsp->name,
2120 READ_ONCE(rsp->gp_seq),
2124 /* Handle quiescent-state forcing. */
2125 first_gp_fqs = true;
2126 j = jiffies_till_first_fqs;
2129 jiffies_till_first_fqs = HZ;
2134 rsp->jiffies_force_qs = jiffies + j;
2135 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2138 trace_rcu_grace_period(rsp->name,
2139 READ_ONCE(rsp->gp_seq),
2141 rsp->gp_state = RCU_GP_WAIT_FQS;
2142 ret = swait_event_idle_timeout(rsp->gp_wq,
2143 rcu_gp_fqs_check_wake(rsp, &gf), j);
2144 rsp->gp_state = RCU_GP_DOING_FQS;
2145 /* Locking provides needed memory barriers. */
2146 /* If grace period done, leave loop. */
2147 if (!READ_ONCE(rnp->qsmask) &&
2148 !rcu_preempt_blocked_readers_cgp(rnp))
2150 /* If time for quiescent-state forcing, do it. */
2151 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2152 (gf & RCU_GP_FLAG_FQS)) {
2153 trace_rcu_grace_period(rsp->name,
2154 READ_ONCE(rsp->gp_seq),
2156 rcu_gp_fqs(rsp, first_gp_fqs);
2157 first_gp_fqs = false;
2158 trace_rcu_grace_period(rsp->name,
2159 READ_ONCE(rsp->gp_seq),
2161 cond_resched_tasks_rcu_qs();
2162 WRITE_ONCE(rsp->gp_activity, jiffies);
2163 ret = 0; /* Force full wait till next FQS. */
2164 j = jiffies_till_next_fqs;
2167 jiffies_till_next_fqs = HZ;
2170 jiffies_till_next_fqs = 1;
2173 /* Deal with stray signal. */
2174 cond_resched_tasks_rcu_qs();
2175 WRITE_ONCE(rsp->gp_activity, jiffies);
2176 WARN_ON(signal_pending(current));
2177 trace_rcu_grace_period(rsp->name,
2178 READ_ONCE(rsp->gp_seq),
2180 ret = 1; /* Keep old FQS timing. */
2182 if (time_after(jiffies, rsp->jiffies_force_qs))
2185 j = rsp->jiffies_force_qs - j;
2189 /* Handle grace-period end. */
2190 rsp->gp_state = RCU_GP_CLEANUP;
2191 rcu_gp_cleanup(rsp);
2192 rsp->gp_state = RCU_GP_CLEANED;
2197 * Report a full set of quiescent states to the specified rcu_state data
2198 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2199 * kthread if another grace period is required. Whether we wake
2200 * the grace-period kthread or it awakens itself for the next round
2201 * of quiescent-state forcing, that kthread will clean up after the
2202 * just-completed grace period. Note that the caller must hold rnp->lock,
2203 * which is released before return.
2205 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2206 __releases(rcu_get_root(rsp)->lock)
2208 raw_lockdep_assert_held_rcu_node(rcu_get_root(rsp));
2209 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2210 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2211 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2212 rcu_gp_kthread_wake(rsp);
2216 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2217 * Allows quiescent states for a group of CPUs to be reported at one go
2218 * to the specified rcu_node structure, though all the CPUs in the group
2219 * must be represented by the same rcu_node structure (which need not be a
2220 * leaf rcu_node structure, though it often will be). The gps parameter
2221 * is the grace-period snapshot, which means that the quiescent states
2222 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2223 * must be held upon entry, and it is released before return.
2226 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2227 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2228 __releases(rnp->lock)
2230 unsigned long oldmask = 0;
2231 struct rcu_node *rnp_c;
2233 raw_lockdep_assert_held_rcu_node(rnp);
2235 /* Walk up the rcu_node hierarchy. */
2237 if (!(rnp->qsmask & mask) || rnp->gp_seq != gps) {
2240 * Our bit has already been cleared, or the
2241 * relevant grace period is already over, so done.
2243 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2246 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2247 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2248 rcu_preempt_blocked_readers_cgp(rnp));
2249 rnp->qsmask &= ~mask;
2250 trace_rcu_quiescent_state_report(rsp->name, rnp->gp_seq,
2251 mask, rnp->qsmask, rnp->level,
2252 rnp->grplo, rnp->grphi,
2254 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2256 /* Other bits still set at this level, so done. */
2257 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2260 rnp->completedqs = rnp->gp_seq;
2261 mask = rnp->grpmask;
2262 if (rnp->parent == NULL) {
2264 /* No more levels. Exit loop holding root lock. */
2268 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2271 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2272 oldmask = rnp_c->qsmask;
2276 * Get here if we are the last CPU to pass through a quiescent
2277 * state for this grace period. Invoke rcu_report_qs_rsp()
2278 * to clean up and start the next grace period if one is needed.
2280 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2284 * Record a quiescent state for all tasks that were previously queued
2285 * on the specified rcu_node structure and that were blocking the current
2286 * RCU grace period. The caller must hold the specified rnp->lock with
2287 * irqs disabled, and this lock is released upon return, but irqs remain
2290 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2291 struct rcu_node *rnp, unsigned long flags)
2292 __releases(rnp->lock)
2296 struct rcu_node *rnp_p;
2298 raw_lockdep_assert_held_rcu_node(rnp);
2299 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2300 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2301 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2302 return; /* Still need more quiescent states! */
2305 rnp_p = rnp->parent;
2306 if (rnp_p == NULL) {
2308 * Only one rcu_node structure in the tree, so don't
2309 * try to report up to its nonexistent parent!
2311 rcu_report_qs_rsp(rsp, flags);
2315 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2317 mask = rnp->grpmask;
2318 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2319 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2320 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2324 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2325 * structure. This must be called from the specified CPU.
2328 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2330 unsigned long flags;
2333 struct rcu_node *rnp;
2336 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2337 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2341 * The grace period in which this quiescent state was
2342 * recorded has ended, so don't report it upwards.
2343 * We will instead need a new quiescent state that lies
2344 * within the current grace period.
2346 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2347 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2348 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2351 mask = rdp->grpmask;
2352 if ((rnp->qsmask & mask) == 0) {
2353 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2355 rdp->core_needs_qs = false;
2358 * This GP can't end until cpu checks in, so all of our
2359 * callbacks can be processed during the next GP.
2361 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2363 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
2364 /* ^^^ Released rnp->lock */
2366 rcu_gp_kthread_wake(rsp);
2371 * Check to see if there is a new grace period of which this CPU
2372 * is not yet aware, and if so, set up local rcu_data state for it.
2373 * Otherwise, see if this CPU has just passed through its first
2374 * quiescent state for this grace period, and record that fact if so.
2377 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2379 /* Check for grace-period ends and beginnings. */
2380 note_gp_changes(rsp, rdp);
2383 * Does this CPU still need to do its part for current grace period?
2384 * If no, return and let the other CPUs do their part as well.
2386 if (!rdp->core_needs_qs)
2390 * Was there a quiescent state since the beginning of the grace
2391 * period? If no, then exit and wait for the next call.
2393 if (rdp->cpu_no_qs.b.norm)
2397 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2400 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2404 * Trace the fact that this CPU is going offline.
2406 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2408 RCU_TRACE(bool blkd;)
2409 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2410 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2412 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2415 RCU_TRACE(blkd = !!(rnp->qsmask & rdp->grpmask);)
2416 trace_rcu_grace_period(rsp->name, rnp->gp_seq,
2417 blkd ? TPS("cpuofl") : TPS("cpuofl-bgp"));
2421 * All CPUs for the specified rcu_node structure have gone offline,
2422 * and all tasks that were preempted within an RCU read-side critical
2423 * section while running on one of those CPUs have since exited their RCU
2424 * read-side critical section. Some other CPU is reporting this fact with
2425 * the specified rcu_node structure's ->lock held and interrupts disabled.
2426 * This function therefore goes up the tree of rcu_node structures,
2427 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2428 * the leaf rcu_node structure's ->qsmaskinit field has already been
2431 * This function does check that the specified rcu_node structure has
2432 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2433 * prematurely. That said, invoking it after the fact will cost you
2434 * a needless lock acquisition. So once it has done its work, don't
2437 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2440 struct rcu_node *rnp = rnp_leaf;
2442 raw_lockdep_assert_held_rcu_node(rnp);
2443 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2444 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2447 mask = rnp->grpmask;
2451 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2452 rnp->qsmaskinit &= ~mask;
2453 rnp->qsmask &= ~mask;
2454 if (rnp->qsmaskinit) {
2455 raw_spin_unlock_rcu_node(rnp);
2456 /* irqs remain disabled. */
2459 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2464 * The CPU has been completely removed, and some other CPU is reporting
2465 * this fact from process context. Do the remainder of the cleanup.
2466 * There can only be one CPU hotplug operation at a time, so no need for
2469 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2471 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2472 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2474 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2477 /* Adjust any no-longer-needed kthreads. */
2478 rcu_boost_kthread_setaffinity(rnp, -1);
2482 * Invoke any RCU callbacks that have made it to the end of their grace
2483 * period. Thottle as specified by rdp->blimit.
2485 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2487 unsigned long flags;
2488 struct rcu_head *rhp;
2489 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2492 /* If no callbacks are ready, just return. */
2493 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2494 trace_rcu_batch_start(rsp->name,
2495 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2496 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2497 trace_rcu_batch_end(rsp->name, 0,
2498 !rcu_segcblist_empty(&rdp->cblist),
2499 need_resched(), is_idle_task(current),
2500 rcu_is_callbacks_kthread());
2505 * Extract the list of ready callbacks, disabling to prevent
2506 * races with call_rcu() from interrupt handlers. Leave the
2507 * callback counts, as rcu_barrier() needs to be conservative.
2509 local_irq_save(flags);
2510 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2512 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2513 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2514 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2515 local_irq_restore(flags);
2517 /* Invoke callbacks. */
2518 rhp = rcu_cblist_dequeue(&rcl);
2519 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2520 debug_rcu_head_unqueue(rhp);
2521 if (__rcu_reclaim(rsp->name, rhp))
2522 rcu_cblist_dequeued_lazy(&rcl);
2524 * Stop only if limit reached and CPU has something to do.
2525 * Note: The rcl structure counts down from zero.
2527 if (-rcl.len >= bl &&
2529 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2533 local_irq_save(flags);
2535 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2536 is_idle_task(current), rcu_is_callbacks_kthread());
2538 /* Update counts and requeue any remaining callbacks. */
2539 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2540 smp_mb(); /* List handling before counting for rcu_barrier(). */
2541 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2543 /* Reinstate batch limit if we have worked down the excess. */
2544 count = rcu_segcblist_n_cbs(&rdp->cblist);
2545 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2546 rdp->blimit = blimit;
2548 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2549 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2550 rdp->qlen_last_fqs_check = 0;
2551 rdp->n_force_qs_snap = rsp->n_force_qs;
2552 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2553 rdp->qlen_last_fqs_check = count;
2556 * The following usually indicates a double call_rcu(). To track
2557 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2559 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2561 local_irq_restore(flags);
2563 /* Re-invoke RCU core processing if there are callbacks remaining. */
2564 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2569 * Check to see if this CPU is in a non-context-switch quiescent state
2570 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2571 * Also schedule RCU core processing.
2573 * This function must be called from hardirq context. It is normally
2574 * invoked from the scheduling-clock interrupt.
2576 void rcu_check_callbacks(int user)
2578 trace_rcu_utilization(TPS("Start scheduler-tick"));
2579 increment_cpu_stall_ticks();
2580 if (user || rcu_is_cpu_rrupt_from_idle()) {
2583 * Get here if this CPU took its interrupt from user
2584 * mode or from the idle loop, and if this is not a
2585 * nested interrupt. In this case, the CPU is in
2586 * a quiescent state, so note it.
2588 * No memory barrier is required here because both
2589 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2590 * variables that other CPUs neither access nor modify,
2591 * at least not while the corresponding CPU is online.
2597 } else if (!in_softirq()) {
2600 * Get here if this CPU did not take its interrupt from
2601 * softirq, in other words, if it is not interrupting
2602 * a rcu_bh read-side critical section. This is an _bh
2603 * critical section, so note it.
2608 rcu_preempt_check_callbacks();
2612 rcu_note_voluntary_context_switch(current);
2613 trace_rcu_utilization(TPS("End scheduler-tick"));
2617 * Scan the leaf rcu_node structures, processing dyntick state for any that
2618 * have not yet encountered a quiescent state, using the function specified.
2619 * Also initiate boosting for any threads blocked on the root rcu_node.
2621 * The caller must have suppressed start of new grace periods.
2623 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2626 unsigned long flags;
2628 struct rcu_node *rnp;
2630 rcu_for_each_leaf_node(rsp, rnp) {
2631 cond_resched_tasks_rcu_qs();
2633 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2634 if (rnp->qsmask == 0) {
2635 if (rcu_state_p == &rcu_sched_state ||
2636 rsp != rcu_state_p ||
2637 rcu_preempt_blocked_readers_cgp(rnp)) {
2639 * No point in scanning bits because they
2640 * are all zero. But we might need to
2641 * priority-boost blocked readers.
2643 rcu_initiate_boost(rnp, flags);
2644 /* rcu_initiate_boost() releases rnp->lock */
2648 (rnp->parent->qsmask & rnp->grpmask)) {
2650 * Race between grace-period
2651 * initialization and task exiting RCU
2652 * read-side critical section: Report.
2654 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2655 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2659 for_each_leaf_node_possible_cpu(rnp, cpu) {
2660 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2661 if ((rnp->qsmask & bit) != 0) {
2662 if (f(per_cpu_ptr(rsp->rda, cpu)))
2667 /* Idle/offline CPUs, report (releases rnp->lock). */
2668 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gp_seq, flags);
2670 /* Nothing to do here, so just drop the lock. */
2671 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2677 * Force quiescent states on reluctant CPUs, and also detect which
2678 * CPUs are in dyntick-idle mode.
2680 static void force_quiescent_state(struct rcu_state *rsp)
2682 unsigned long flags;
2684 struct rcu_node *rnp;
2685 struct rcu_node *rnp_old = NULL;
2687 /* Funnel through hierarchy to reduce memory contention. */
2688 rnp = __this_cpu_read(rsp->rda->mynode);
2689 for (; rnp != NULL; rnp = rnp->parent) {
2690 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2691 !raw_spin_trylock(&rnp->fqslock);
2692 if (rnp_old != NULL)
2693 raw_spin_unlock(&rnp_old->fqslock);
2698 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2700 /* Reached the root of the rcu_node tree, acquire lock. */
2701 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2702 raw_spin_unlock(&rnp_old->fqslock);
2703 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2704 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2705 return; /* Someone beat us to it. */
2707 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2708 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2709 rcu_gp_kthread_wake(rsp);
2713 * This function checks for grace-period requests that fail to motivate
2714 * RCU to come out of its idle mode.
2717 rcu_check_gp_start_stall(struct rcu_state *rsp, struct rcu_node *rnp,
2718 struct rcu_data *rdp)
2720 unsigned long flags;
2722 struct rcu_node *rnp_root = rcu_get_root(rsp);
2723 static atomic_t warned = ATOMIC_INIT(0);
2725 if (!IS_ENABLED(CONFIG_PROVE_RCU) || rcu_gp_in_progress(rsp) ||
2726 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed))
2728 j = jiffies; /* Expensive access, and in common case don't get here. */
2729 if (time_before(j, READ_ONCE(rsp->gp_req_activity) + HZ) ||
2730 time_before(j, READ_ONCE(rsp->gp_activity) + HZ) ||
2731 atomic_read(&warned))
2734 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2736 if (rcu_gp_in_progress(rsp) ||
2737 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
2738 time_before(j, READ_ONCE(rsp->gp_req_activity) + HZ) ||
2739 time_before(j, READ_ONCE(rsp->gp_activity) + HZ) ||
2740 atomic_read(&warned)) {
2741 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2744 /* Hold onto the leaf lock to make others see warned==1. */
2746 if (rnp_root != rnp)
2747 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
2749 if (rcu_gp_in_progress(rsp) ||
2750 ULONG_CMP_GE(rnp_root->gp_seq, rnp_root->gp_seq_needed) ||
2751 time_before(j, rsp->gp_req_activity + HZ) ||
2752 time_before(j, rsp->gp_activity + HZ) ||
2753 atomic_xchg(&warned, 1)) {
2754 raw_spin_unlock_rcu_node(rnp_root); /* irqs remain disabled. */
2755 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2758 pr_alert("%s: g%ld->%ld gar:%lu ga:%lu f%#x %s->state:%#lx\n",
2759 __func__, (long)READ_ONCE(rsp->gp_seq),
2760 (long)READ_ONCE(rnp_root->gp_seq_needed),
2761 j - rsp->gp_req_activity, j - rsp->gp_activity,
2762 rsp->gp_flags, rsp->name,
2763 rsp->gp_kthread ? rsp->gp_kthread->state : 0x1ffffL);
2765 if (rnp_root != rnp)
2766 raw_spin_unlock_rcu_node(rnp_root);
2767 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2771 * This does the RCU core processing work for the specified rcu_state
2772 * and rcu_data structures. This may be called only from the CPU to
2773 * whom the rdp belongs.
2776 __rcu_process_callbacks(struct rcu_state *rsp)
2778 unsigned long flags;
2779 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2780 struct rcu_node *rnp = rdp->mynode;
2782 WARN_ON_ONCE(!rdp->beenonline);
2784 /* Update RCU state based on any recent quiescent states. */
2785 rcu_check_quiescent_state(rsp, rdp);
2787 /* No grace period and unregistered callbacks? */
2788 if (!rcu_gp_in_progress(rsp) &&
2789 rcu_segcblist_is_enabled(&rdp->cblist)) {
2790 local_irq_save(flags);
2791 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2792 rcu_accelerate_cbs_unlocked(rsp, rnp, rdp);
2793 local_irq_restore(flags);
2796 rcu_check_gp_start_stall(rsp, rnp, rdp);
2798 /* If there are callbacks ready, invoke them. */
2799 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2800 invoke_rcu_callbacks(rsp, rdp);
2802 /* Do any needed deferred wakeups of rcuo kthreads. */
2803 do_nocb_deferred_wakeup(rdp);
2807 * Do RCU core processing for the current CPU.
2809 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2811 struct rcu_state *rsp;
2813 if (cpu_is_offline(smp_processor_id()))
2815 trace_rcu_utilization(TPS("Start RCU core"));
2816 for_each_rcu_flavor(rsp)
2817 __rcu_process_callbacks(rsp);
2818 trace_rcu_utilization(TPS("End RCU core"));
2822 * Schedule RCU callback invocation. If the specified type of RCU
2823 * does not support RCU priority boosting, just do a direct call,
2824 * otherwise wake up the per-CPU kernel kthread. Note that because we
2825 * are running on the current CPU with softirqs disabled, the
2826 * rcu_cpu_kthread_task cannot disappear out from under us.
2828 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2830 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2832 if (likely(!rsp->boost)) {
2833 rcu_do_batch(rsp, rdp);
2836 invoke_rcu_callbacks_kthread();
2839 static void invoke_rcu_core(void)
2841 if (cpu_online(smp_processor_id()))
2842 raise_softirq(RCU_SOFTIRQ);
2846 * Handle any core-RCU processing required by a call_rcu() invocation.
2848 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2849 struct rcu_head *head, unsigned long flags)
2852 * If called from an extended quiescent state, invoke the RCU
2853 * core in order to force a re-evaluation of RCU's idleness.
2855 if (!rcu_is_watching())
2858 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2859 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2863 * Force the grace period if too many callbacks or too long waiting.
2864 * Enforce hysteresis, and don't invoke force_quiescent_state()
2865 * if some other CPU has recently done so. Also, don't bother
2866 * invoking force_quiescent_state() if the newly enqueued callback
2867 * is the only one waiting for a grace period to complete.
2869 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2870 rdp->qlen_last_fqs_check + qhimark)) {
2872 /* Are we ignoring a completed grace period? */
2873 note_gp_changes(rsp, rdp);
2875 /* Start a new grace period if one not already started. */
2876 if (!rcu_gp_in_progress(rsp)) {
2877 rcu_accelerate_cbs_unlocked(rsp, rdp->mynode, rdp);
2879 /* Give the grace period a kick. */
2880 rdp->blimit = LONG_MAX;
2881 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2882 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2883 force_quiescent_state(rsp);
2884 rdp->n_force_qs_snap = rsp->n_force_qs;
2885 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2891 * RCU callback function to leak a callback.
2893 static void rcu_leak_callback(struct rcu_head *rhp)
2898 * Helper function for call_rcu() and friends. The cpu argument will
2899 * normally be -1, indicating "currently running CPU". It may specify
2900 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2901 * is expected to specify a CPU.
2904 __call_rcu(struct rcu_head *head, rcu_callback_t func,
2905 struct rcu_state *rsp, int cpu, bool lazy)
2907 unsigned long flags;
2908 struct rcu_data *rdp;
2910 /* Misaligned rcu_head! */
2911 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2913 if (debug_rcu_head_queue(head)) {
2915 * Probable double call_rcu(), so leak the callback.
2916 * Use rcu:rcu_callback trace event to find the previous
2917 * time callback was passed to __call_rcu().
2919 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
2921 WRITE_ONCE(head->func, rcu_leak_callback);
2926 local_irq_save(flags);
2927 rdp = this_cpu_ptr(rsp->rda);
2929 /* Add the callback to our list. */
2930 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
2934 rdp = per_cpu_ptr(rsp->rda, cpu);
2935 if (likely(rdp->mynode)) {
2936 /* Post-boot, so this should be for a no-CBs CPU. */
2937 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2938 WARN_ON_ONCE(offline);
2939 /* Offline CPU, _call_rcu() illegal, leak callback. */
2940 local_irq_restore(flags);
2944 * Very early boot, before rcu_init(). Initialize if needed
2945 * and then drop through to queue the callback.
2948 WARN_ON_ONCE(!rcu_is_watching());
2949 if (rcu_segcblist_empty(&rdp->cblist))
2950 rcu_segcblist_init(&rdp->cblist);
2952 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2954 rcu_idle_count_callbacks_posted();
2956 if (__is_kfree_rcu_offset((unsigned long)func))
2957 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2958 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2959 rcu_segcblist_n_cbs(&rdp->cblist));
2961 trace_rcu_callback(rsp->name, head,
2962 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2963 rcu_segcblist_n_cbs(&rdp->cblist));
2965 /* Go handle any RCU core processing required. */
2966 __call_rcu_core(rsp, rdp, head, flags);
2967 local_irq_restore(flags);
2971 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
2972 * @head: structure to be used for queueing the RCU updates.
2973 * @func: actual callback function to be invoked after the grace period
2975 * The callback function will be invoked some time after a full grace
2976 * period elapses, in other words after all currently executing RCU
2977 * read-side critical sections have completed. call_rcu_sched() assumes
2978 * that the read-side critical sections end on enabling of preemption
2979 * or on voluntary preemption.
2980 * RCU read-side critical sections are delimited by:
2982 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
2983 * - anything that disables preemption.
2985 * These may be nested.
2987 * See the description of call_rcu() for more detailed information on
2988 * memory ordering guarantees.
2990 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
2992 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2994 EXPORT_SYMBOL_GPL(call_rcu_sched);
2997 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
2998 * @head: structure to be used for queueing the RCU updates.
2999 * @func: actual callback function to be invoked after the grace period
3001 * The callback function will be invoked some time after a full grace
3002 * period elapses, in other words after all currently executing RCU
3003 * read-side critical sections have completed. call_rcu_bh() assumes
3004 * that the read-side critical sections end on completion of a softirq
3005 * handler. This means that read-side critical sections in process
3006 * context must not be interrupted by softirqs. This interface is to be
3007 * used when most of the read-side critical sections are in softirq context.
3008 * RCU read-side critical sections are delimited by:
3010 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3011 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3013 * These may be nested.
3015 * See the description of call_rcu() for more detailed information on
3016 * memory ordering guarantees.
3018 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3020 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3022 EXPORT_SYMBOL_GPL(call_rcu_bh);
3025 * Queue an RCU callback for lazy invocation after a grace period.
3026 * This will likely be later named something like "call_rcu_lazy()",
3027 * but this change will require some way of tagging the lazy RCU
3028 * callbacks in the list of pending callbacks. Until then, this
3029 * function may only be called from __kfree_rcu().
3031 void kfree_call_rcu(struct rcu_head *head,
3032 rcu_callback_t func)
3034 __call_rcu(head, func, rcu_state_p, -1, 1);
3036 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3039 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3040 * any blocking grace-period wait automatically implies a grace period
3041 * if there is only one CPU online at any point time during execution
3042 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3043 * occasionally incorrectly indicate that there are multiple CPUs online
3044 * when there was in fact only one the whole time, as this just adds
3045 * some overhead: RCU still operates correctly.
3047 static inline int rcu_blocking_is_gp(void)
3051 might_sleep(); /* Check for RCU read-side critical section. */
3053 ret = num_online_cpus() <= 1;
3059 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3061 * Control will return to the caller some time after a full rcu-sched
3062 * grace period has elapsed, in other words after all currently executing
3063 * rcu-sched read-side critical sections have completed. These read-side
3064 * critical sections are delimited by rcu_read_lock_sched() and
3065 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3066 * local_irq_disable(), and so on may be used in place of
3067 * rcu_read_lock_sched().
3069 * This means that all preempt_disable code sequences, including NMI and
3070 * non-threaded hardware-interrupt handlers, in progress on entry will
3071 * have completed before this primitive returns. However, this does not
3072 * guarantee that softirq handlers will have completed, since in some
3073 * kernels, these handlers can run in process context, and can block.
3075 * Note that this guarantee implies further memory-ordering guarantees.
3076 * On systems with more than one CPU, when synchronize_sched() returns,
3077 * each CPU is guaranteed to have executed a full memory barrier since the
3078 * end of its last RCU-sched read-side critical section whose beginning
3079 * preceded the call to synchronize_sched(). In addition, each CPU having
3080 * an RCU read-side critical section that extends beyond the return from
3081 * synchronize_sched() is guaranteed to have executed a full memory barrier
3082 * after the beginning of synchronize_sched() and before the beginning of
3083 * that RCU read-side critical section. Note that these guarantees include
3084 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3085 * that are executing in the kernel.
3087 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3088 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3089 * to have executed a full memory barrier during the execution of
3090 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3091 * again only if the system has more than one CPU).
3093 void synchronize_sched(void)
3095 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3096 lock_is_held(&rcu_lock_map) ||
3097 lock_is_held(&rcu_sched_lock_map),
3098 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3099 if (rcu_blocking_is_gp())
3101 if (rcu_gp_is_expedited())
3102 synchronize_sched_expedited();
3104 wait_rcu_gp(call_rcu_sched);
3106 EXPORT_SYMBOL_GPL(synchronize_sched);
3109 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3111 * Control will return to the caller some time after a full rcu_bh grace
3112 * period has elapsed, in other words after all currently executing rcu_bh
3113 * read-side critical sections have completed. RCU read-side critical
3114 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3115 * and may be nested.
3117 * See the description of synchronize_sched() for more detailed information
3118 * on memory ordering guarantees.
3120 void synchronize_rcu_bh(void)
3122 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3123 lock_is_held(&rcu_lock_map) ||
3124 lock_is_held(&rcu_sched_lock_map),
3125 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3126 if (rcu_blocking_is_gp())
3128 if (rcu_gp_is_expedited())
3129 synchronize_rcu_bh_expedited();
3131 wait_rcu_gp(call_rcu_bh);
3133 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3136 * get_state_synchronize_rcu - Snapshot current RCU state
3138 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3139 * to determine whether or not a full grace period has elapsed in the
3142 unsigned long get_state_synchronize_rcu(void)
3145 * Any prior manipulation of RCU-protected data must happen
3146 * before the load from ->gp_seq.
3149 return rcu_seq_snap(&rcu_state_p->gp_seq);
3151 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3154 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3156 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3158 * If a full RCU grace period has elapsed since the earlier call to
3159 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3160 * synchronize_rcu() to wait for a full grace period.
3162 * Yes, this function does not take counter wrap into account. But
3163 * counter wrap is harmless. If the counter wraps, we have waited for
3164 * more than 2 billion grace periods (and way more on a 64-bit system!),
3165 * so waiting for one additional grace period should be just fine.
3167 void cond_synchronize_rcu(unsigned long oldstate)
3169 if (!rcu_seq_done(&rcu_state_p->gp_seq, oldstate))
3172 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3174 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3177 * get_state_synchronize_sched - Snapshot current RCU-sched state
3179 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3180 * to determine whether or not a full grace period has elapsed in the
3183 unsigned long get_state_synchronize_sched(void)
3186 * Any prior manipulation of RCU-protected data must happen
3187 * before the load from ->gp_seq.
3190 return rcu_seq_snap(&rcu_sched_state.gp_seq);
3192 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3195 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3197 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3199 * If a full RCU-sched grace period has elapsed since the earlier call to
3200 * get_state_synchronize_sched(), just return. Otherwise, invoke
3201 * synchronize_sched() to wait for a full grace period.
3203 * Yes, this function does not take counter wrap into account. But
3204 * counter wrap is harmless. If the counter wraps, we have waited for
3205 * more than 2 billion grace periods (and way more on a 64-bit system!),
3206 * so waiting for one additional grace period should be just fine.
3208 void cond_synchronize_sched(unsigned long oldstate)
3210 if (!rcu_seq_done(&rcu_sched_state.gp_seq, oldstate))
3211 synchronize_sched();
3213 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3215 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3218 * Check to see if there is any immediate RCU-related work to be done
3219 * by the current CPU, for the specified type of RCU, returning 1 if so.
3220 * The checks are in order of increasing expense: checks that can be
3221 * carried out against CPU-local state are performed first. However,
3222 * we must check for CPU stalls first, else we might not get a chance.
3224 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3226 struct rcu_node *rnp = rdp->mynode;
3228 /* Check for CPU stalls, if enabled. */
3229 check_cpu_stall(rsp, rdp);
3231 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3232 if (rcu_nohz_full_cpu(rsp))
3235 /* Is the RCU core waiting for a quiescent state from this CPU? */
3236 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
3239 /* Does this CPU have callbacks ready to invoke? */
3240 if (rcu_segcblist_ready_cbs(&rdp->cblist))
3243 /* Has RCU gone idle with this CPU needing another grace period? */
3244 if (!rcu_gp_in_progress(rsp) &&
3245 rcu_segcblist_is_enabled(&rdp->cblist) &&
3246 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3249 /* Have RCU grace period completed or started? */
3250 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3251 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3254 /* Does this CPU need a deferred NOCB wakeup? */
3255 if (rcu_nocb_need_deferred_wakeup(rdp))
3263 * Check to see if there is any immediate RCU-related work to be done
3264 * by the current CPU, returning 1 if so. This function is part of the
3265 * RCU implementation; it is -not- an exported member of the RCU API.
3267 static int rcu_pending(void)
3269 struct rcu_state *rsp;
3271 for_each_rcu_flavor(rsp)
3272 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3278 * Return true if the specified CPU has any callback. If all_lazy is
3279 * non-NULL, store an indication of whether all callbacks are lazy.
3280 * (If there are no callbacks, all of them are deemed to be lazy.)
3282 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3286 struct rcu_data *rdp;
3287 struct rcu_state *rsp;
3289 for_each_rcu_flavor(rsp) {
3290 rdp = this_cpu_ptr(rsp->rda);
3291 if (rcu_segcblist_empty(&rdp->cblist))
3294 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3305 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3306 * the compiler is expected to optimize this away.
3308 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3309 int cpu, unsigned long done)
3311 trace_rcu_barrier(rsp->name, s, cpu,
3312 atomic_read(&rsp->barrier_cpu_count), done);
3316 * RCU callback function for _rcu_barrier(). If we are last, wake
3317 * up the task executing _rcu_barrier().
3319 static void rcu_barrier_callback(struct rcu_head *rhp)
3321 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3322 struct rcu_state *rsp = rdp->rsp;
3324 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3325 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3326 rsp->barrier_sequence);
3327 complete(&rsp->barrier_completion);
3329 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3334 * Called with preemption disabled, and from cross-cpu IRQ context.
3336 static void rcu_barrier_func(void *type)
3338 struct rcu_state *rsp = type;
3339 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3341 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3342 rdp->barrier_head.func = rcu_barrier_callback;
3343 debug_rcu_head_queue(&rdp->barrier_head);
3344 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3345 atomic_inc(&rsp->barrier_cpu_count);
3347 debug_rcu_head_unqueue(&rdp->barrier_head);
3348 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3349 rsp->barrier_sequence);
3354 * Orchestrate the specified type of RCU barrier, waiting for all
3355 * RCU callbacks of the specified type to complete.
3357 static void _rcu_barrier(struct rcu_state *rsp)
3360 struct rcu_data *rdp;
3361 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3363 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3365 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3366 mutex_lock(&rsp->barrier_mutex);
3368 /* Did someone else do our work for us? */
3369 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3370 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3371 rsp->barrier_sequence);
3372 smp_mb(); /* caller's subsequent code after above check. */
3373 mutex_unlock(&rsp->barrier_mutex);
3377 /* Mark the start of the barrier operation. */
3378 rcu_seq_start(&rsp->barrier_sequence);
3379 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3382 * Initialize the count to one rather than to zero in order to
3383 * avoid a too-soon return to zero in case of a short grace period
3384 * (or preemption of this task). Exclude CPU-hotplug operations
3385 * to ensure that no offline CPU has callbacks queued.
3387 init_completion(&rsp->barrier_completion);
3388 atomic_set(&rsp->barrier_cpu_count, 1);
3392 * Force each CPU with callbacks to register a new callback.
3393 * When that callback is invoked, we will know that all of the
3394 * corresponding CPU's preceding callbacks have been invoked.
3396 for_each_possible_cpu(cpu) {
3397 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3399 rdp = per_cpu_ptr(rsp->rda, cpu);
3400 if (rcu_is_nocb_cpu(cpu)) {
3401 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3402 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3403 rsp->barrier_sequence);
3405 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3406 rsp->barrier_sequence);
3407 smp_mb__before_atomic();
3408 atomic_inc(&rsp->barrier_cpu_count);
3409 __call_rcu(&rdp->barrier_head,
3410 rcu_barrier_callback, rsp, cpu, 0);
3412 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3413 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3414 rsp->barrier_sequence);
3415 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3417 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3418 rsp->barrier_sequence);
3424 * Now that we have an rcu_barrier_callback() callback on each
3425 * CPU, and thus each counted, remove the initial count.
3427 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3428 complete(&rsp->barrier_completion);
3430 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3431 wait_for_completion(&rsp->barrier_completion);
3433 /* Mark the end of the barrier operation. */
3434 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3435 rcu_seq_end(&rsp->barrier_sequence);
3437 /* Other rcu_barrier() invocations can now safely proceed. */
3438 mutex_unlock(&rsp->barrier_mutex);
3442 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3444 void rcu_barrier_bh(void)
3446 _rcu_barrier(&rcu_bh_state);
3448 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3451 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3453 void rcu_barrier_sched(void)
3455 _rcu_barrier(&rcu_sched_state);
3457 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3460 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3461 * first CPU in a given leaf rcu_node structure coming online. The caller
3462 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3465 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3468 struct rcu_node *rnp = rnp_leaf;
3470 raw_lockdep_assert_held_rcu_node(rnp);
3472 mask = rnp->grpmask;
3476 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3477 rnp->qsmaskinit |= mask;
3478 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3483 * Do boot-time initialization of a CPU's per-CPU RCU data.
3486 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3488 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3490 /* Set up local state, ensuring consistent view of global state. */
3491 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3492 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3493 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != 1);
3494 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3497 rcu_boot_init_nocb_percpu_data(rdp);
3501 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3502 * offline event can be happening at a given time. Note also that we can
3503 * accept some slop in the rsp->gp_seq access due to the fact that this
3504 * CPU cannot possibly have any RCU callbacks in flight yet.
3507 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3509 unsigned long flags;
3510 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3511 struct rcu_node *rnp = rcu_get_root(rsp);
3513 /* Set up local state, ensuring consistent view of global state. */
3514 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3515 rdp->qlen_last_fqs_check = 0;
3516 rdp->n_force_qs_snap = rsp->n_force_qs;
3517 rdp->blimit = blimit;
3518 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3519 !init_nocb_callback_list(rdp))
3520 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3521 rdp->dynticks->dynticks_nesting = 1; /* CPU not up, no tearing. */
3522 rcu_dynticks_eqs_online();
3523 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3526 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3527 * propagation up the rcu_node tree will happen at the beginning
3528 * of the next grace period.
3531 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3532 rdp->beenonline = true; /* We have now been online. */
3533 rdp->gp_seq = rnp->gp_seq;
3534 rdp->gp_seq_needed = rnp->gp_seq;
3535 rdp->cpu_no_qs.b.norm = true;
3536 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3537 rdp->core_needs_qs = false;
3538 rdp->rcu_iw_pending = false;
3539 rdp->rcu_iw_gp_seq = rnp->gp_seq - 1;
3540 trace_rcu_grace_period(rsp->name, rdp->gp_seq, TPS("cpuonl"));
3541 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3545 * Invoked early in the CPU-online process, when pretty much all
3546 * services are available. The incoming CPU is not present.
3548 int rcutree_prepare_cpu(unsigned int cpu)
3550 struct rcu_state *rsp;
3552 for_each_rcu_flavor(rsp)
3553 rcu_init_percpu_data(cpu, rsp);
3555 rcu_prepare_kthreads(cpu);
3556 rcu_spawn_all_nocb_kthreads(cpu);
3562 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3564 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3566 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3568 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3572 * Near the end of the CPU-online process. Pretty much all services
3573 * enabled, and the CPU is now very much alive.
3575 int rcutree_online_cpu(unsigned int cpu)
3577 unsigned long flags;
3578 struct rcu_data *rdp;
3579 struct rcu_node *rnp;
3580 struct rcu_state *rsp;
3582 for_each_rcu_flavor(rsp) {
3583 rdp = per_cpu_ptr(rsp->rda, cpu);
3585 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3586 rnp->ffmask |= rdp->grpmask;
3587 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3589 if (IS_ENABLED(CONFIG_TREE_SRCU))
3590 srcu_online_cpu(cpu);
3591 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3592 return 0; /* Too early in boot for scheduler work. */
3593 sync_sched_exp_online_cleanup(cpu);
3594 rcutree_affinity_setting(cpu, -1);
3599 * Near the beginning of the process. The CPU is still very much alive
3600 * with pretty much all services enabled.
3602 int rcutree_offline_cpu(unsigned int cpu)
3604 unsigned long flags;
3605 struct rcu_data *rdp;
3606 struct rcu_node *rnp;
3607 struct rcu_state *rsp;
3609 for_each_rcu_flavor(rsp) {
3610 rdp = per_cpu_ptr(rsp->rda, cpu);
3612 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3613 rnp->ffmask &= ~rdp->grpmask;
3614 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3617 rcutree_affinity_setting(cpu, cpu);
3618 if (IS_ENABLED(CONFIG_TREE_SRCU))
3619 srcu_offline_cpu(cpu);
3624 * Near the end of the offline process. We do only tracing here.
3626 int rcutree_dying_cpu(unsigned int cpu)
3628 struct rcu_state *rsp;
3630 for_each_rcu_flavor(rsp)
3631 rcu_cleanup_dying_cpu(rsp);
3636 * The outgoing CPU is gone and we are running elsewhere.
3638 int rcutree_dead_cpu(unsigned int cpu)
3640 struct rcu_state *rsp;
3642 for_each_rcu_flavor(rsp) {
3643 rcu_cleanup_dead_cpu(cpu, rsp);
3644 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3649 static DEFINE_PER_CPU(int, rcu_cpu_started);
3652 * Mark the specified CPU as being online so that subsequent grace periods
3653 * (both expedited and normal) will wait on it. Note that this means that
3654 * incoming CPUs are not allowed to use RCU read-side critical sections
3655 * until this function is called. Failing to observe this restriction
3656 * will result in lockdep splats.
3658 * Note that this function is special in that it is invoked directly
3659 * from the incoming CPU rather than from the cpuhp_step mechanism.
3660 * This is because this function must be invoked at a precise location.
3662 void rcu_cpu_starting(unsigned int cpu)
3664 unsigned long flags;
3667 unsigned long oldmask;
3668 struct rcu_data *rdp;
3669 struct rcu_node *rnp;
3670 struct rcu_state *rsp;
3672 if (per_cpu(rcu_cpu_started, cpu))
3675 per_cpu(rcu_cpu_started, cpu) = 1;
3677 for_each_rcu_flavor(rsp) {
3678 rdp = per_cpu_ptr(rsp->rda, cpu);
3680 mask = rdp->grpmask;
3681 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3682 rnp->qsmaskinitnext |= mask;
3683 oldmask = rnp->expmaskinitnext;
3684 rnp->expmaskinitnext |= mask;
3685 oldmask ^= rnp->expmaskinitnext;
3686 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3687 /* Allow lockless access for expedited grace periods. */
3688 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3689 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3691 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3694 #ifdef CONFIG_HOTPLUG_CPU
3696 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3697 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3700 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3702 unsigned long flags;
3704 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3705 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3707 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3708 mask = rdp->grpmask;
3709 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3710 rnp->qsmaskinitnext &= ~mask;
3711 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3715 * The outgoing function has no further need of RCU, so remove it from
3716 * the list of CPUs that RCU must track.
3718 * Note that this function is special in that it is invoked directly
3719 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3720 * This is because this function must be invoked at a precise location.
3722 void rcu_report_dead(unsigned int cpu)
3724 struct rcu_state *rsp;
3726 /* QS for any half-done expedited RCU-sched GP. */
3728 rcu_report_exp_rdp(&rcu_sched_state,
3729 this_cpu_ptr(rcu_sched_state.rda), true);
3731 for_each_rcu_flavor(rsp)
3732 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3734 per_cpu(rcu_cpu_started, cpu) = 0;
3737 /* Migrate the dead CPU's callbacks to the current CPU. */
3738 static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3740 unsigned long flags;
3741 struct rcu_data *my_rdp;
3742 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3743 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3746 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3747 return; /* No callbacks to migrate. */
3749 local_irq_save(flags);
3750 my_rdp = this_cpu_ptr(rsp->rda);
3751 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3752 local_irq_restore(flags);
3755 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3756 /* Leverage recent GPs and set GP for new callbacks. */
3757 needwake = rcu_advance_cbs(rsp, rnp_root, rdp) ||
3758 rcu_advance_cbs(rsp, rnp_root, my_rdp);
3759 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3760 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3761 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3762 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3764 rcu_gp_kthread_wake(rsp);
3765 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3766 !rcu_segcblist_empty(&rdp->cblist),
3767 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3768 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3769 rcu_segcblist_first_cb(&rdp->cblist));
3773 * The outgoing CPU has just passed through the dying-idle state,
3774 * and we are being invoked from the CPU that was IPIed to continue the
3775 * offline operation. We need to migrate the outgoing CPU's callbacks.
3777 void rcutree_migrate_callbacks(int cpu)
3779 struct rcu_state *rsp;
3781 for_each_rcu_flavor(rsp)
3782 rcu_migrate_callbacks(cpu, rsp);
3787 * On non-huge systems, use expedited RCU grace periods to make suspend
3788 * and hibernation run faster.
3790 static int rcu_pm_notify(struct notifier_block *self,
3791 unsigned long action, void *hcpu)
3794 case PM_HIBERNATION_PREPARE:
3795 case PM_SUSPEND_PREPARE:
3796 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3799 case PM_POST_HIBERNATION:
3800 case PM_POST_SUSPEND:
3801 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3802 rcu_unexpedite_gp();
3811 * Spawn the kthreads that handle each RCU flavor's grace periods.
3813 static int __init rcu_spawn_gp_kthread(void)
3815 unsigned long flags;
3816 int kthread_prio_in = kthread_prio;
3817 struct rcu_node *rnp;
3818 struct rcu_state *rsp;
3819 struct sched_param sp;
3820 struct task_struct *t;
3822 /* Force priority into range. */
3823 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3825 else if (kthread_prio < 0)
3827 else if (kthread_prio > 99)
3829 if (kthread_prio != kthread_prio_in)
3830 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3831 kthread_prio, kthread_prio_in);
3833 rcu_scheduler_fully_active = 1;
3834 for_each_rcu_flavor(rsp) {
3835 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3837 rnp = rcu_get_root(rsp);
3838 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3839 rsp->gp_kthread = t;
3841 sp.sched_priority = kthread_prio;
3842 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3844 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3847 rcu_spawn_nocb_kthreads();
3848 rcu_spawn_boost_kthreads();
3851 early_initcall(rcu_spawn_gp_kthread);
3854 * This function is invoked towards the end of the scheduler's
3855 * initialization process. Before this is called, the idle task might
3856 * contain synchronous grace-period primitives (during which time, this idle
3857 * task is booting the system, and such primitives are no-ops). After this
3858 * function is called, any synchronous grace-period primitives are run as
3859 * expedited, with the requesting task driving the grace period forward.
3860 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3861 * runtime RCU functionality.
3863 void rcu_scheduler_starting(void)
3865 WARN_ON(num_online_cpus() != 1);
3866 WARN_ON(nr_context_switches() > 0);
3867 rcu_test_sync_prims();
3868 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3869 rcu_test_sync_prims();
3873 * Helper function for rcu_init() that initializes one rcu_state structure.
3875 static void __init rcu_init_one(struct rcu_state *rsp)
3877 static const char * const buf[] = RCU_NODE_NAME_INIT;
3878 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3879 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3880 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3882 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3886 struct rcu_node *rnp;
3888 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3890 /* Silence gcc 4.8 false positive about array index out of range. */
3891 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3892 panic("rcu_init_one: rcu_num_lvls out of range");
3894 /* Initialize the level-tracking arrays. */
3896 for (i = 1; i < rcu_num_lvls; i++)
3897 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
3898 rcu_init_levelspread(levelspread, num_rcu_lvl);
3900 /* Initialize the elements themselves, starting from the leaves. */
3902 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3903 cpustride *= levelspread[i];
3904 rnp = rsp->level[i];
3905 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3906 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3907 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3908 &rcu_node_class[i], buf[i]);
3909 raw_spin_lock_init(&rnp->fqslock);
3910 lockdep_set_class_and_name(&rnp->fqslock,
3911 &rcu_fqs_class[i], fqs[i]);
3912 rnp->gp_seq = rsp->gp_seq;
3913 rnp->gp_seq_needed = rsp->gp_seq;
3914 rnp->completedqs = rsp->gp_seq;
3916 rnp->qsmaskinit = 0;
3917 rnp->grplo = j * cpustride;
3918 rnp->grphi = (j + 1) * cpustride - 1;
3919 if (rnp->grphi >= nr_cpu_ids)
3920 rnp->grphi = nr_cpu_ids - 1;
3926 rnp->grpnum = j % levelspread[i - 1];
3927 rnp->grpmask = 1UL << rnp->grpnum;
3928 rnp->parent = rsp->level[i - 1] +
3929 j / levelspread[i - 1];
3932 INIT_LIST_HEAD(&rnp->blkd_tasks);
3933 rcu_init_one_nocb(rnp);
3934 init_waitqueue_head(&rnp->exp_wq[0]);
3935 init_waitqueue_head(&rnp->exp_wq[1]);
3936 init_waitqueue_head(&rnp->exp_wq[2]);
3937 init_waitqueue_head(&rnp->exp_wq[3]);
3938 spin_lock_init(&rnp->exp_lock);
3942 init_swait_queue_head(&rsp->gp_wq);
3943 init_swait_queue_head(&rsp->expedited_wq);
3944 rnp = rcu_first_leaf_node(rsp);
3945 for_each_possible_cpu(i) {
3946 while (i > rnp->grphi)
3948 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3949 rcu_boot_init_percpu_data(i, rsp);
3951 list_add(&rsp->flavors, &rcu_struct_flavors);
3955 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3956 * replace the definitions in tree.h because those are needed to size
3957 * the ->node array in the rcu_state structure.
3959 static void __init rcu_init_geometry(void)
3963 int rcu_capacity[RCU_NUM_LVLS];
3966 * Initialize any unspecified boot parameters.
3967 * The default values of jiffies_till_first_fqs and
3968 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3969 * value, which is a function of HZ, then adding one for each
3970 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3972 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3973 if (jiffies_till_first_fqs == ULONG_MAX)
3974 jiffies_till_first_fqs = d;
3975 if (jiffies_till_next_fqs == ULONG_MAX)
3976 jiffies_till_next_fqs = d;
3978 /* If the compile-time values are accurate, just leave. */
3979 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
3980 nr_cpu_ids == NR_CPUS)
3982 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
3983 rcu_fanout_leaf, nr_cpu_ids);
3986 * The boot-time rcu_fanout_leaf parameter must be at least two
3987 * and cannot exceed the number of bits in the rcu_node masks.
3988 * Complain and fall back to the compile-time values if this
3989 * limit is exceeded.
3991 if (rcu_fanout_leaf < 2 ||
3992 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
3993 rcu_fanout_leaf = RCU_FANOUT_LEAF;
3999 * Compute number of nodes that can be handled an rcu_node tree
4000 * with the given number of levels.
4002 rcu_capacity[0] = rcu_fanout_leaf;
4003 for (i = 1; i < RCU_NUM_LVLS; i++)
4004 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4007 * The tree must be able to accommodate the configured number of CPUs.
4008 * If this limit is exceeded, fall back to the compile-time values.
4010 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4011 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4016 /* Calculate the number of levels in the tree. */
4017 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4019 rcu_num_lvls = i + 1;
4021 /* Calculate the number of rcu_nodes at each level of the tree. */
4022 for (i = 0; i < rcu_num_lvls; i++) {
4023 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4024 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4027 /* Calculate the total number of rcu_node structures. */
4029 for (i = 0; i < rcu_num_lvls; i++)
4030 rcu_num_nodes += num_rcu_lvl[i];
4034 * Dump out the structure of the rcu_node combining tree associated
4035 * with the rcu_state structure referenced by rsp.
4037 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4040 struct rcu_node *rnp;
4042 pr_info("rcu_node tree layout dump\n");
4044 rcu_for_each_node_breadth_first(rsp, rnp) {
4045 if (rnp->level != level) {
4050 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4055 struct workqueue_struct *rcu_gp_wq;
4056 struct workqueue_struct *rcu_par_gp_wq;
4058 void __init rcu_init(void)
4062 rcu_early_boot_tests();
4064 rcu_bootup_announce();
4065 rcu_init_geometry();
4066 rcu_init_one(&rcu_bh_state);
4067 rcu_init_one(&rcu_sched_state);
4069 rcu_dump_rcu_node_tree(&rcu_sched_state);
4070 __rcu_init_preempt();
4071 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4074 * We don't need protection against CPU-hotplug here because
4075 * this is called early in boot, before either interrupts
4076 * or the scheduler are operational.
4078 pm_notifier(rcu_pm_notify, 0);
4079 for_each_online_cpu(cpu) {
4080 rcutree_prepare_cpu(cpu);
4081 rcu_cpu_starting(cpu);
4082 rcutree_online_cpu(cpu);
4085 /* Create workqueue for expedited GPs and for Tree SRCU. */
4086 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4087 WARN_ON(!rcu_gp_wq);
4088 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4089 WARN_ON(!rcu_par_gp_wq);
4092 #include "tree_exp.h"
4093 #include "tree_plugin.h"