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 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT, \
101 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
102 .name = RCU_STATE_NAME(sname), \
104 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
105 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
108 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
109 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
111 static struct rcu_state *const rcu_state_p;
112 LIST_HEAD(rcu_struct_flavors);
114 /* Dump rcu_node combining tree at boot to verify correct setup. */
115 static bool dump_tree;
116 module_param(dump_tree, bool, 0444);
117 /* Control rcu_node-tree auto-balancing at boot time. */
118 static bool rcu_fanout_exact;
119 module_param(rcu_fanout_exact, bool, 0444);
120 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
121 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
122 module_param(rcu_fanout_leaf, int, 0444);
123 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
124 /* Number of rcu_nodes at specified level. */
125 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
126 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
127 /* panic() on RCU Stall sysctl. */
128 int sysctl_panic_on_rcu_stall __read_mostly;
131 * The rcu_scheduler_active variable is initialized to the value
132 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
133 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
134 * RCU can assume that there is but one task, allowing RCU to (for example)
135 * optimize synchronize_rcu() to a simple barrier(). When this variable
136 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
137 * to detect real grace periods. This variable is also used to suppress
138 * boot-time false positives from lockdep-RCU error checking. Finally, it
139 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
140 * is fully initialized, including all of its kthreads having been spawned.
142 int rcu_scheduler_active __read_mostly;
143 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
146 * The rcu_scheduler_fully_active variable transitions from zero to one
147 * during the early_initcall() processing, which is after the scheduler
148 * is capable of creating new tasks. So RCU processing (for example,
149 * creating tasks for RCU priority boosting) must be delayed until after
150 * rcu_scheduler_fully_active transitions from zero to one. We also
151 * currently delay invocation of any RCU callbacks until after this point.
153 * It might later prove better for people registering RCU callbacks during
154 * early boot to take responsibility for these callbacks, but one step at
157 static int rcu_scheduler_fully_active __read_mostly;
159 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
160 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
161 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
162 static void invoke_rcu_core(void);
163 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
164 static void rcu_report_exp_rdp(struct rcu_state *rsp,
165 struct rcu_data *rdp, bool wake);
166 static void sync_sched_exp_online_cleanup(int cpu);
168 /* rcuc/rcub kthread realtime priority */
169 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
170 module_param(kthread_prio, int, 0644);
172 /* Delay in jiffies for grace-period initialization delays, debug only. */
174 static int gp_preinit_delay;
175 module_param(gp_preinit_delay, int, 0444);
176 static int gp_init_delay;
177 module_param(gp_init_delay, int, 0444);
178 static int gp_cleanup_delay;
179 module_param(gp_cleanup_delay, int, 0444);
182 * Number of grace periods between delays, normalized by the duration of
183 * the delay. The longer the delay, the more the grace periods between
184 * each delay. The reason for this normalization is that it means that,
185 * for non-zero delays, the overall slowdown of grace periods is constant
186 * regardless of the duration of the delay. This arrangement balances
187 * the need for long delays to increase some race probabilities with the
188 * need for fast grace periods to increase other race probabilities.
190 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
193 * Track the rcutorture test sequence number and the update version
194 * number within a given test. The rcutorture_testseq is incremented
195 * on every rcutorture module load and unload, so has an odd value
196 * when a test is running. The rcutorture_vernum is set to zero
197 * when rcutorture starts and is incremented on each rcutorture update.
198 * These variables enable correlating rcutorture output with the
199 * RCU tracing information.
201 unsigned long rcutorture_testseq;
202 unsigned long rcutorture_vernum;
205 * Compute the mask of online CPUs for the specified rcu_node structure.
206 * This will not be stable unless the rcu_node structure's ->lock is
207 * held, but the bit corresponding to the current CPU will be stable
210 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
212 return READ_ONCE(rnp->qsmaskinitnext);
216 * Return true if an RCU grace period is in progress. The READ_ONCE()s
217 * permit this function to be invoked without holding the root rcu_node
218 * structure's ->lock, but of course results can be subject to change.
220 static int rcu_gp_in_progress(struct rcu_state *rsp)
222 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
226 * Note a quiescent state. Because we do not need to know
227 * how many quiescent states passed, just if there was at least
228 * one since the start of the grace period, this just sets a flag.
229 * The caller must have disabled preemption.
231 void rcu_sched_qs(void)
233 RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
234 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
236 trace_rcu_grace_period(TPS("rcu_sched"),
237 __this_cpu_read(rcu_sched_data.gpnum),
239 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
240 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
242 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
243 rcu_report_exp_rdp(&rcu_sched_state,
244 this_cpu_ptr(&rcu_sched_data), true);
249 RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
250 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
251 trace_rcu_grace_period(TPS("rcu_bh"),
252 __this_cpu_read(rcu_bh_data.gpnum),
254 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
259 * Steal a bit from the bottom of ->dynticks for idle entry/exit
260 * control. Initially this is for TLB flushing.
262 #define RCU_DYNTICK_CTRL_MASK 0x1
263 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
264 #ifndef rcu_eqs_special_exit
265 #define rcu_eqs_special_exit() do { } while (0)
268 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
269 .dynticks_nesting = 1,
270 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
271 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
275 * Record entry into an extended quiescent state. This is only to be
276 * called when not already in an extended quiescent state.
278 static void rcu_dynticks_eqs_enter(void)
280 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
284 * CPUs seeing atomic_add_return() must see prior RCU read-side
285 * critical sections, and we also must force ordering with the
288 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
289 /* Better be in an extended quiescent state! */
290 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
291 (seq & RCU_DYNTICK_CTRL_CTR));
292 /* Better not have special action (TLB flush) pending! */
293 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
294 (seq & RCU_DYNTICK_CTRL_MASK));
298 * Record exit from an extended quiescent state. This is only to be
299 * called from an extended quiescent state.
301 static void rcu_dynticks_eqs_exit(void)
303 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
307 * CPUs seeing atomic_add_return() must see prior idle sojourns,
308 * and we also must force ordering with the next RCU read-side
311 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
312 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
313 !(seq & RCU_DYNTICK_CTRL_CTR));
314 if (seq & RCU_DYNTICK_CTRL_MASK) {
315 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
316 smp_mb__after_atomic(); /* _exit after clearing mask. */
317 /* Prefer duplicate flushes to losing a flush. */
318 rcu_eqs_special_exit();
323 * Reset the current CPU's ->dynticks counter to indicate that the
324 * newly onlined CPU is no longer in an extended quiescent state.
325 * This will either leave the counter unchanged, or increment it
326 * to the next non-quiescent value.
328 * The non-atomic test/increment sequence works because the upper bits
329 * of the ->dynticks counter are manipulated only by the corresponding CPU,
330 * or when the corresponding CPU is offline.
332 static void rcu_dynticks_eqs_online(void)
334 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
336 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
338 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
342 * Is the current CPU in an extended quiescent state?
344 * No ordering, as we are sampling CPU-local information.
346 bool rcu_dynticks_curr_cpu_in_eqs(void)
348 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
350 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
354 * Snapshot the ->dynticks counter with full ordering so as to allow
355 * stable comparison of this counter with past and future snapshots.
357 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
359 int snap = atomic_add_return(0, &rdtp->dynticks);
361 return snap & ~RCU_DYNTICK_CTRL_MASK;
365 * Return true if the snapshot returned from rcu_dynticks_snap()
366 * indicates that RCU is in an extended quiescent state.
368 static bool rcu_dynticks_in_eqs(int snap)
370 return !(snap & RCU_DYNTICK_CTRL_CTR);
374 * Return true if the CPU corresponding to the specified rcu_dynticks
375 * structure has spent some time in an extended quiescent state since
376 * rcu_dynticks_snap() returned the specified snapshot.
378 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
380 return snap != rcu_dynticks_snap(rdtp);
384 * Do a double-increment of the ->dynticks counter to emulate a
385 * momentary idle-CPU quiescent state.
387 static void rcu_dynticks_momentary_idle(void)
389 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
390 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
393 /* It is illegal to call this from idle state. */
394 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
398 * Set the special (bottom) bit of the specified CPU so that it
399 * will take special action (such as flushing its TLB) on the
400 * next exit from an extended quiescent state. Returns true if
401 * the bit was successfully set, or false if the CPU was not in
402 * an extended quiescent state.
404 bool rcu_eqs_special_set(int cpu)
408 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
411 old = atomic_read(&rdtp->dynticks);
412 if (old & RCU_DYNTICK_CTRL_CTR)
414 new = old | RCU_DYNTICK_CTRL_MASK;
415 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
420 * Let the RCU core know that this CPU has gone through the scheduler,
421 * which is a quiescent state. This is called when the need for a
422 * quiescent state is urgent, so we burn an atomic operation and full
423 * memory barriers to let the RCU core know about it, regardless of what
424 * this CPU might (or might not) do in the near future.
426 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
428 * The caller must have disabled interrupts.
430 static void rcu_momentary_dyntick_idle(void)
432 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
433 rcu_dynticks_momentary_idle();
437 * Note a context switch. This is a quiescent state for RCU-sched,
438 * and requires special handling for preemptible RCU.
439 * The caller must have disabled interrupts.
441 void rcu_note_context_switch(bool preempt)
443 barrier(); /* Avoid RCU read-side critical sections leaking down. */
444 trace_rcu_utilization(TPS("Start context switch"));
446 rcu_preempt_note_context_switch(preempt);
447 /* Load rcu_urgent_qs before other flags. */
448 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
450 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
451 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
452 rcu_momentary_dyntick_idle();
453 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
455 rcu_note_voluntary_context_switch_lite(current);
457 trace_rcu_utilization(TPS("End context switch"));
458 barrier(); /* Avoid RCU read-side critical sections leaking up. */
460 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
463 * Register a quiescent state for all RCU flavors. If there is an
464 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
465 * dyntick-idle quiescent state visible to other CPUs (but only for those
466 * RCU flavors in desperate need of a quiescent state, which will normally
467 * be none of them). Either way, do a lightweight quiescent state for
470 * The barrier() calls are redundant in the common case when this is
471 * called externally, but just in case this is called from within this
475 void rcu_all_qs(void)
479 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
482 /* Load rcu_urgent_qs before other flags. */
483 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
487 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
488 barrier(); /* Avoid RCU read-side critical sections leaking down. */
489 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
490 local_irq_save(flags);
491 rcu_momentary_dyntick_idle();
492 local_irq_restore(flags);
494 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
496 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
497 barrier(); /* Avoid RCU read-side critical sections leaking up. */
500 EXPORT_SYMBOL_GPL(rcu_all_qs);
502 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
503 static long blimit = DEFAULT_RCU_BLIMIT;
504 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
505 static long qhimark = DEFAULT_RCU_QHIMARK;
506 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
507 static long qlowmark = DEFAULT_RCU_QLOMARK;
509 module_param(blimit, long, 0444);
510 module_param(qhimark, long, 0444);
511 module_param(qlowmark, long, 0444);
513 static ulong jiffies_till_first_fqs = ULONG_MAX;
514 static ulong jiffies_till_next_fqs = ULONG_MAX;
515 static bool rcu_kick_kthreads;
517 module_param(jiffies_till_first_fqs, ulong, 0644);
518 module_param(jiffies_till_next_fqs, ulong, 0644);
519 module_param(rcu_kick_kthreads, bool, 0644);
522 * How long the grace period must be before we start recruiting
523 * quiescent-state help from rcu_note_context_switch().
525 static ulong jiffies_till_sched_qs = HZ / 10;
526 module_param(jiffies_till_sched_qs, ulong, 0444);
528 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
529 static void force_quiescent_state(struct rcu_state *rsp);
530 static int rcu_pending(void);
533 * Return the number of RCU GPs completed thus far for debug & stats.
535 unsigned long rcu_get_gp_seq(void)
537 return rcu_seq_ctr(READ_ONCE(rcu_state_p->gp_seq));
539 EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
542 * Return the number of RCU-sched GPs completed thus far for debug & stats.
544 unsigned long rcu_sched_get_gp_seq(void)
546 return rcu_seq_ctr(READ_ONCE(rcu_sched_state.gp_seq));
548 EXPORT_SYMBOL_GPL(rcu_sched_get_gp_seq);
551 * Return the number of RCU-bh GPs completed thus far for debug & stats.
553 unsigned long rcu_bh_get_gp_seq(void)
555 return rcu_seq_ctr(READ_ONCE(rcu_bh_state.gp_seq));
557 EXPORT_SYMBOL_GPL(rcu_bh_get_gp_seq);
560 * Return the number of RCU expedited batches completed thus far for
561 * debug & stats. Odd numbers mean that a batch is in progress, even
562 * numbers mean idle. The value returned will thus be roughly double
563 * the cumulative batches since boot.
565 unsigned long rcu_exp_batches_completed(void)
567 return rcu_state_p->expedited_sequence;
569 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
572 * Return the number of RCU-sched expedited batches completed thus far
573 * for debug & stats. Similar to rcu_exp_batches_completed().
575 unsigned long rcu_exp_batches_completed_sched(void)
577 return rcu_sched_state.expedited_sequence;
579 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
582 * Force a quiescent state.
584 void rcu_force_quiescent_state(void)
586 force_quiescent_state(rcu_state_p);
588 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
591 * Force a quiescent state for RCU BH.
593 void rcu_bh_force_quiescent_state(void)
595 force_quiescent_state(&rcu_bh_state);
597 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
600 * Force a quiescent state for RCU-sched.
602 void rcu_sched_force_quiescent_state(void)
604 force_quiescent_state(&rcu_sched_state);
606 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
609 * Show the state of the grace-period kthreads.
611 void show_rcu_gp_kthreads(void)
613 struct rcu_state *rsp;
615 for_each_rcu_flavor(rsp) {
616 pr_info("%s: wait state: %d ->state: %#lx\n",
617 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
618 /* sched_show_task(rsp->gp_kthread); */
621 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
624 * Record the number of times rcutorture tests have been initiated and
625 * terminated. This information allows the debugfs tracing stats to be
626 * correlated to the rcutorture messages, even when the rcutorture module
627 * is being repeatedly loaded and unloaded. In other words, we cannot
628 * store this state in rcutorture itself.
630 void rcutorture_record_test_transition(void)
632 rcutorture_testseq++;
633 rcutorture_vernum = 0;
635 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
638 * Send along grace-period-related data for rcutorture diagnostics.
640 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
641 unsigned long *gpnum, unsigned long *completed)
643 struct rcu_state *rsp = NULL;
652 case RCU_SCHED_FLAVOR:
653 rsp = &rcu_sched_state;
660 *flags = READ_ONCE(rsp->gp_flags);
661 *gpnum = READ_ONCE(rsp->gpnum);
662 *completed = READ_ONCE(rsp->completed);
664 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
667 * Record the number of writer passes through the current rcutorture test.
668 * This is also used to correlate debugfs tracing stats with the rcutorture
671 void rcutorture_record_progress(unsigned long vernum)
675 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
678 * Return the root node of the specified rcu_state structure.
680 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
682 return &rsp->node[0];
686 * Enter an RCU extended quiescent state, which can be either the
687 * idle loop or adaptive-tickless usermode execution.
689 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
690 * the possibility of usermode upcalls having messed up our count
691 * of interrupt nesting level during the prior busy period.
693 static void rcu_eqs_enter(bool user)
695 struct rcu_state *rsp;
696 struct rcu_data *rdp;
697 struct rcu_dynticks *rdtp;
699 rdtp = this_cpu_ptr(&rcu_dynticks);
700 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0);
701 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
702 rdtp->dynticks_nesting == 0);
703 if (rdtp->dynticks_nesting != 1) {
704 rdtp->dynticks_nesting--;
708 lockdep_assert_irqs_disabled();
709 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0, rdtp->dynticks);
710 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
711 for_each_rcu_flavor(rsp) {
712 rdp = this_cpu_ptr(rsp->rda);
713 do_nocb_deferred_wakeup(rdp);
715 rcu_prepare_for_idle();
716 WRITE_ONCE(rdtp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
717 rcu_dynticks_eqs_enter();
718 rcu_dynticks_task_enter();
722 * rcu_idle_enter - inform RCU that current CPU is entering idle
724 * Enter idle mode, in other words, -leave- the mode in which RCU
725 * read-side critical sections can occur. (Though RCU read-side
726 * critical sections can occur in irq handlers in idle, a possibility
727 * handled by irq_enter() and irq_exit().)
729 * If you add or remove a call to rcu_idle_enter(), be sure to test with
730 * CONFIG_RCU_EQS_DEBUG=y.
732 void rcu_idle_enter(void)
734 lockdep_assert_irqs_disabled();
735 rcu_eqs_enter(false);
738 #ifdef CONFIG_NO_HZ_FULL
740 * rcu_user_enter - inform RCU that we are resuming userspace.
742 * Enter RCU idle mode right before resuming userspace. No use of RCU
743 * is permitted between this call and rcu_user_exit(). This way the
744 * CPU doesn't need to maintain the tick for RCU maintenance purposes
745 * when the CPU runs in userspace.
747 * If you add or remove a call to rcu_user_enter(), be sure to test with
748 * CONFIG_RCU_EQS_DEBUG=y.
750 void rcu_user_enter(void)
752 lockdep_assert_irqs_disabled();
755 #endif /* CONFIG_NO_HZ_FULL */
758 * rcu_nmi_exit - inform RCU of exit from NMI context
760 * If we are returning from the outermost NMI handler that interrupted an
761 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
762 * to let the RCU grace-period handling know that the CPU is back to
765 * If you add or remove a call to rcu_nmi_exit(), be sure to test
766 * with CONFIG_RCU_EQS_DEBUG=y.
768 void rcu_nmi_exit(void)
770 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
773 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
774 * (We are exiting an NMI handler, so RCU better be paying attention
777 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
778 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
781 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
782 * leave it in non-RCU-idle state.
784 if (rdtp->dynticks_nmi_nesting != 1) {
785 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nmi_nesting, rdtp->dynticks_nmi_nesting - 2, rdtp->dynticks);
786 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* No store tearing. */
787 rdtp->dynticks_nmi_nesting - 2);
791 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
792 trace_rcu_dyntick(TPS("Startirq"), rdtp->dynticks_nmi_nesting, 0, rdtp->dynticks);
793 WRITE_ONCE(rdtp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
794 rcu_dynticks_eqs_enter();
798 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
800 * Exit from an interrupt handler, which might possibly result in entering
801 * idle mode, in other words, leaving the mode in which read-side critical
802 * sections can occur. The caller must have disabled interrupts.
804 * This code assumes that the idle loop never does anything that might
805 * result in unbalanced calls to irq_enter() and irq_exit(). If your
806 * architecture's idle loop violates this assumption, RCU will give you what
807 * you deserve, good and hard. But very infrequently and irreproducibly.
809 * Use things like work queues to work around this limitation.
811 * You have been warned.
813 * If you add or remove a call to rcu_irq_exit(), be sure to test with
814 * CONFIG_RCU_EQS_DEBUG=y.
816 void rcu_irq_exit(void)
818 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
820 lockdep_assert_irqs_disabled();
821 if (rdtp->dynticks_nmi_nesting == 1)
822 rcu_prepare_for_idle();
824 if (rdtp->dynticks_nmi_nesting == 0)
825 rcu_dynticks_task_enter();
829 * Wrapper for rcu_irq_exit() where interrupts are enabled.
831 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
832 * with CONFIG_RCU_EQS_DEBUG=y.
834 void rcu_irq_exit_irqson(void)
838 local_irq_save(flags);
840 local_irq_restore(flags);
844 * Exit an RCU extended quiescent state, which can be either the
845 * idle loop or adaptive-tickless usermode execution.
847 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
848 * allow for the possibility of usermode upcalls messing up our count of
849 * interrupt nesting level during the busy period that is just now starting.
851 static void rcu_eqs_exit(bool user)
853 struct rcu_dynticks *rdtp;
856 lockdep_assert_irqs_disabled();
857 rdtp = this_cpu_ptr(&rcu_dynticks);
858 oldval = rdtp->dynticks_nesting;
859 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
861 rdtp->dynticks_nesting++;
864 rcu_dynticks_task_exit();
865 rcu_dynticks_eqs_exit();
866 rcu_cleanup_after_idle();
867 trace_rcu_dyntick(TPS("End"), rdtp->dynticks_nesting, 1, rdtp->dynticks);
868 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
869 WRITE_ONCE(rdtp->dynticks_nesting, 1);
870 WRITE_ONCE(rdtp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
874 * rcu_idle_exit - inform RCU that current CPU is leaving idle
876 * Exit idle mode, in other words, -enter- the mode in which RCU
877 * read-side critical sections can occur.
879 * If you add or remove a call to rcu_idle_exit(), be sure to test with
880 * CONFIG_RCU_EQS_DEBUG=y.
882 void rcu_idle_exit(void)
886 local_irq_save(flags);
888 local_irq_restore(flags);
891 #ifdef CONFIG_NO_HZ_FULL
893 * rcu_user_exit - inform RCU that we are exiting userspace.
895 * Exit RCU idle mode while entering the kernel because it can
896 * run a RCU read side critical section anytime.
898 * If you add or remove a call to rcu_user_exit(), be sure to test with
899 * CONFIG_RCU_EQS_DEBUG=y.
901 void rcu_user_exit(void)
905 #endif /* CONFIG_NO_HZ_FULL */
908 * rcu_nmi_enter - inform RCU of entry to NMI context
910 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
911 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
912 * that the CPU is active. This implementation permits nested NMIs, as
913 * long as the nesting level does not overflow an int. (You will probably
914 * run out of stack space first.)
916 * If you add or remove a call to rcu_nmi_enter(), be sure to test
917 * with CONFIG_RCU_EQS_DEBUG=y.
919 void rcu_nmi_enter(void)
921 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
924 /* Complain about underflow. */
925 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
928 * If idle from RCU viewpoint, atomically increment ->dynticks
929 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
930 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
931 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
932 * to be in the outermost NMI handler that interrupted an RCU-idle
933 * period (observation due to Andy Lutomirski).
935 if (rcu_dynticks_curr_cpu_in_eqs()) {
936 rcu_dynticks_eqs_exit();
939 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
940 rdtp->dynticks_nmi_nesting,
941 rdtp->dynticks_nmi_nesting + incby, rdtp->dynticks);
942 WRITE_ONCE(rdtp->dynticks_nmi_nesting, /* Prevent store tearing. */
943 rdtp->dynticks_nmi_nesting + incby);
948 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
950 * Enter an interrupt handler, which might possibly result in exiting
951 * idle mode, in other words, entering the mode in which read-side critical
952 * sections can occur. The caller must have disabled interrupts.
954 * Note that the Linux kernel is fully capable of entering an interrupt
955 * handler that it never exits, for example when doing upcalls to user mode!
956 * This code assumes that the idle loop never does upcalls to user mode.
957 * If your architecture's idle loop does do upcalls to user mode (or does
958 * anything else that results in unbalanced calls to the irq_enter() and
959 * irq_exit() functions), RCU will give you what you deserve, good and hard.
960 * But very infrequently and irreproducibly.
962 * Use things like work queues to work around this limitation.
964 * You have been warned.
966 * If you add or remove a call to rcu_irq_enter(), be sure to test with
967 * CONFIG_RCU_EQS_DEBUG=y.
969 void rcu_irq_enter(void)
971 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
973 lockdep_assert_irqs_disabled();
974 if (rdtp->dynticks_nmi_nesting == 0)
975 rcu_dynticks_task_exit();
977 if (rdtp->dynticks_nmi_nesting == 1)
978 rcu_cleanup_after_idle();
982 * Wrapper for rcu_irq_enter() where interrupts are enabled.
984 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
985 * with CONFIG_RCU_EQS_DEBUG=y.
987 void rcu_irq_enter_irqson(void)
991 local_irq_save(flags);
993 local_irq_restore(flags);
997 * rcu_is_watching - see if RCU thinks that the current CPU is idle
999 * Return true if RCU is watching the running CPU, which means that this
1000 * CPU can safely enter RCU read-side critical sections. In other words,
1001 * if the current CPU is in its idle loop and is neither in an interrupt
1002 * or NMI handler, return true.
1004 bool notrace rcu_is_watching(void)
1008 preempt_disable_notrace();
1009 ret = !rcu_dynticks_curr_cpu_in_eqs();
1010 preempt_enable_notrace();
1013 EXPORT_SYMBOL_GPL(rcu_is_watching);
1016 * If a holdout task is actually running, request an urgent quiescent
1017 * state from its CPU. This is unsynchronized, so migrations can cause
1018 * the request to go to the wrong CPU. Which is OK, all that will happen
1019 * is that the CPU's next context switch will be a bit slower and next
1020 * time around this task will generate another request.
1022 void rcu_request_urgent_qs_task(struct task_struct *t)
1029 return; /* This task is not running on that CPU. */
1030 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1033 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1036 * Is the current CPU online? Disable preemption to avoid false positives
1037 * that could otherwise happen due to the current CPU number being sampled,
1038 * this task being preempted, its old CPU being taken offline, resuming
1039 * on some other CPU, then determining that its old CPU is now offline.
1040 * It is OK to use RCU on an offline processor during initial boot, hence
1041 * the check for rcu_scheduler_fully_active. Note also that it is OK
1042 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1043 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1044 * offline to continue to use RCU for one jiffy after marking itself
1045 * offline in the cpu_online_mask. This leniency is necessary given the
1046 * non-atomic nature of the online and offline processing, for example,
1047 * the fact that a CPU enters the scheduler after completing the teardown
1050 * This is also why RCU internally marks CPUs online during in the
1051 * preparation phase and offline after the CPU has been taken down.
1053 * Disable checking if in an NMI handler because we cannot safely report
1054 * errors from NMI handlers anyway.
1056 bool rcu_lockdep_current_cpu_online(void)
1058 struct rcu_data *rdp;
1059 struct rcu_node *rnp;
1065 rdp = this_cpu_ptr(&rcu_sched_data);
1067 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1068 !rcu_scheduler_fully_active;
1072 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1074 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1077 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1079 * If the current CPU is idle or running at a first-level (not nested)
1080 * interrupt from idle, return true. The caller must have at least
1081 * disabled preemption.
1083 static int rcu_is_cpu_rrupt_from_idle(void)
1085 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 0 &&
1086 __this_cpu_read(rcu_dynticks.dynticks_nmi_nesting) <= 1;
1090 * We are reporting a quiescent state on behalf of some other CPU, so
1091 * it is our responsibility to check for and handle potential overflow
1092 * of the rcu_node ->gpnum counter with respect to the rcu_data counters.
1093 * After all, the CPU might be in deep idle state, and thus executing no
1096 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1098 raw_lockdep_assert_held_rcu_node(rnp);
1099 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, rnp->gpnum))
1100 WRITE_ONCE(rdp->gpwrap, true);
1101 if (ULONG_CMP_LT(rdp->rcu_iw_gpnum + ULONG_MAX / 4, rnp->gpnum))
1102 rdp->rcu_iw_gpnum = rnp->gpnum + ULONG_MAX / 4;
1106 * Snapshot the specified CPU's dynticks counter so that we can later
1107 * credit them with an implicit quiescent state. Return 1 if this CPU
1108 * is in dynticks idle mode, which is an extended quiescent state.
1110 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1112 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1113 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1114 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1115 rcu_gpnum_ovf(rdp->mynode, rdp);
1122 * Handler for the irq_work request posted when a grace period has
1123 * gone on for too long, but not yet long enough for an RCU CPU
1124 * stall warning. Set state appropriately, but just complain if
1125 * there is unexpected state on entry.
1127 static void rcu_iw_handler(struct irq_work *iwp)
1129 struct rcu_data *rdp;
1130 struct rcu_node *rnp;
1132 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1134 raw_spin_lock_rcu_node(rnp);
1135 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1136 rdp->rcu_iw_gpnum = rnp->gpnum;
1137 rdp->rcu_iw_pending = false;
1139 raw_spin_unlock_rcu_node(rnp);
1143 * Return true if the specified CPU has passed through a quiescent
1144 * state by virtue of being in or having passed through an dynticks
1145 * idle state since the last call to dyntick_save_progress_counter()
1146 * for this same CPU, or by virtue of having been offline.
1148 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1153 struct rcu_node *rnp = rdp->mynode;
1156 * If the CPU passed through or entered a dynticks idle phase with
1157 * no active irq/NMI handlers, then we can safely pretend that the CPU
1158 * already acknowledged the request to pass through a quiescent
1159 * state. Either way, that CPU cannot possibly be in an RCU
1160 * read-side critical section that started before the beginning
1161 * of the current RCU grace period.
1163 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1164 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1165 rdp->dynticks_fqs++;
1166 rcu_gpnum_ovf(rnp, rdp);
1171 * Has this CPU encountered a cond_resched() since the beginning
1172 * of the grace period? For this to be the case, the CPU has to
1173 * have noticed the current grace period. This might not be the
1174 * case for nohz_full CPUs looping in the kernel.
1176 jtsq = jiffies_till_sched_qs;
1177 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1178 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1179 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1180 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
1181 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
1182 rcu_gpnum_ovf(rnp, rdp);
1184 } else if (time_after(jiffies, rdp->rsp->gp_start + jtsq)) {
1185 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1186 smp_store_release(ruqp, true);
1189 /* Check for the CPU being offline. */
1190 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1191 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1193 rcu_gpnum_ovf(rnp, rdp);
1198 * A CPU running for an extended time within the kernel can
1199 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1200 * even context-switching back and forth between a pair of
1201 * in-kernel CPU-bound tasks cannot advance grace periods.
1202 * So if the grace period is old enough, make the CPU pay attention.
1203 * Note that the unsynchronized assignments to the per-CPU
1204 * rcu_need_heavy_qs variable are safe. Yes, setting of
1205 * bits can be lost, but they will be set again on the next
1206 * force-quiescent-state pass. So lost bit sets do not result
1207 * in incorrect behavior, merely in a grace period lasting
1208 * a few jiffies longer than it might otherwise. Because
1209 * there are at most four threads involved, and because the
1210 * updates are only once every few jiffies, the probability of
1211 * lossage (and thus of slight grace-period extension) is
1214 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1215 if (!READ_ONCE(*rnhqp) &&
1216 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1217 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1218 WRITE_ONCE(*rnhqp, true);
1219 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1220 smp_store_release(ruqp, true);
1221 rdp->rsp->jiffies_resched += jtsq; /* Re-enable beating. */
1225 * If more than halfway to RCU CPU stall-warning time, do a
1226 * resched_cpu() to try to loosen things up a bit. Also check to
1227 * see if the CPU is getting hammered with interrupts, but only
1228 * once per grace period, just to keep the IPIs down to a dull roar.
1230 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) {
1231 resched_cpu(rdp->cpu);
1232 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1233 !rdp->rcu_iw_pending && rdp->rcu_iw_gpnum != rnp->gpnum &&
1234 (rnp->ffmask & rdp->grpmask)) {
1235 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1236 rdp->rcu_iw_pending = true;
1237 rdp->rcu_iw_gpnum = rnp->gpnum;
1238 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1245 static void record_gp_stall_check_time(struct rcu_state *rsp)
1247 unsigned long j = jiffies;
1251 smp_wmb(); /* Record start time before stall time. */
1252 j1 = rcu_jiffies_till_stall_check();
1253 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1254 rsp->jiffies_resched = j + j1 / 2;
1255 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1259 * Convert a ->gp_state value to a character string.
1261 static const char *gp_state_getname(short gs)
1263 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1265 return gp_state_names[gs];
1269 * Complain about starvation of grace-period kthread.
1271 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1277 gpa = READ_ONCE(rsp->gp_activity);
1278 if (j - gpa > 2 * HZ) {
1279 pr_err("%s kthread starved for %ld jiffies! g%ld f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1281 (long)rcu_seq_current(&rsp->gp_seq),
1283 gp_state_getname(rsp->gp_state), rsp->gp_state,
1284 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1285 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1286 if (rsp->gp_kthread) {
1287 pr_err("RCU grace-period kthread stack dump:\n");
1288 sched_show_task(rsp->gp_kthread);
1289 wake_up_process(rsp->gp_kthread);
1295 * Dump stacks of all tasks running on stalled CPUs. First try using
1296 * NMIs, but fall back to manual remote stack tracing on architectures
1297 * that don't support NMI-based stack dumps. The NMI-triggered stack
1298 * traces are more accurate because they are printed by the target CPU.
1300 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1303 unsigned long flags;
1304 struct rcu_node *rnp;
1306 rcu_for_each_leaf_node(rsp, rnp) {
1307 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1308 for_each_leaf_node_possible_cpu(rnp, cpu)
1309 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1310 if (!trigger_single_cpu_backtrace(cpu))
1312 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1317 * If too much time has passed in the current grace period, and if
1318 * so configured, go kick the relevant kthreads.
1320 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1324 if (!rcu_kick_kthreads)
1326 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1327 if (time_after(jiffies, j) && rsp->gp_kthread &&
1328 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1329 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1330 rcu_ftrace_dump(DUMP_ALL);
1331 wake_up_process(rsp->gp_kthread);
1332 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1336 static inline void panic_on_rcu_stall(void)
1338 if (sysctl_panic_on_rcu_stall)
1339 panic("RCU Stall\n");
1342 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1345 unsigned long flags;
1349 struct rcu_node *rnp = rcu_get_root(rsp);
1352 /* Kick and suppress, if so configured. */
1353 rcu_stall_kick_kthreads(rsp);
1354 if (rcu_cpu_stall_suppress)
1358 * OK, time to rat on our buddy...
1359 * See Documentation/RCU/stallwarn.txt for info on how to debug
1360 * RCU CPU stall warnings.
1362 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1364 print_cpu_stall_info_begin();
1365 rcu_for_each_leaf_node(rsp, rnp) {
1366 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1367 ndetected += rcu_print_task_stall(rnp);
1368 if (rnp->qsmask != 0) {
1369 for_each_leaf_node_possible_cpu(rnp, cpu)
1370 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1371 print_cpu_stall_info(rsp, cpu);
1375 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1378 print_cpu_stall_info_end();
1379 for_each_possible_cpu(cpu)
1380 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1382 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1383 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1384 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1386 rcu_dump_cpu_stacks(rsp);
1388 /* Complain about tasks blocking the grace period. */
1389 rcu_print_detail_task_stall(rsp);
1391 if (READ_ONCE(rsp->gpnum) != gpnum ||
1392 READ_ONCE(rsp->completed) == gpnum) {
1393 pr_err("INFO: Stall ended before state dump start\n");
1396 gpa = READ_ONCE(rsp->gp_activity);
1397 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1398 rsp->name, j - gpa, j, gpa,
1399 jiffies_till_next_fqs,
1400 rcu_get_root(rsp)->qsmask);
1401 /* In this case, the current CPU might be at fault. */
1402 sched_show_task(current);
1405 /* Rewrite if needed in case of slow consoles. */
1406 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1407 WRITE_ONCE(rsp->jiffies_stall,
1408 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1410 rcu_check_gp_kthread_starvation(rsp);
1412 panic_on_rcu_stall();
1414 force_quiescent_state(rsp); /* Kick them all. */
1417 static void print_cpu_stall(struct rcu_state *rsp)
1420 unsigned long flags;
1421 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1422 struct rcu_node *rnp = rcu_get_root(rsp);
1425 /* Kick and suppress, if so configured. */
1426 rcu_stall_kick_kthreads(rsp);
1427 if (rcu_cpu_stall_suppress)
1431 * OK, time to rat on ourselves...
1432 * See Documentation/RCU/stallwarn.txt for info on how to debug
1433 * RCU CPU stall warnings.
1435 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1436 print_cpu_stall_info_begin();
1437 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1438 print_cpu_stall_info(rsp, smp_processor_id());
1439 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1440 print_cpu_stall_info_end();
1441 for_each_possible_cpu(cpu)
1442 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1444 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1445 jiffies - rsp->gp_start,
1446 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1448 rcu_check_gp_kthread_starvation(rsp);
1450 rcu_dump_cpu_stacks(rsp);
1452 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1453 /* Rewrite if needed in case of slow consoles. */
1454 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1455 WRITE_ONCE(rsp->jiffies_stall,
1456 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1457 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1459 panic_on_rcu_stall();
1462 * Attempt to revive the RCU machinery by forcing a context switch.
1464 * A context switch would normally allow the RCU state machine to make
1465 * progress and it could be we're stuck in kernel space without context
1466 * switches for an entirely unreasonable amount of time.
1468 resched_cpu(smp_processor_id());
1471 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1473 unsigned long completed;
1474 unsigned long gpnum;
1479 struct rcu_node *rnp;
1481 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1482 !rcu_gp_in_progress(rsp))
1484 rcu_stall_kick_kthreads(rsp);
1488 * Lots of memory barriers to reject false positives.
1490 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1491 * then rsp->gp_start, and finally rsp->completed. These values
1492 * are updated in the opposite order with memory barriers (or
1493 * equivalent) during grace-period initialization and cleanup.
1494 * Now, a false positive can occur if we get an new value of
1495 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1496 * the memory barriers, the only way that this can happen is if one
1497 * grace period ends and another starts between these two fetches.
1498 * Detect this by comparing rsp->completed with the previous fetch
1501 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1502 * and rsp->gp_start suffice to forestall false positives.
1504 gpnum = READ_ONCE(rsp->gpnum);
1505 smp_rmb(); /* Pick up ->gpnum first... */
1506 js = READ_ONCE(rsp->jiffies_stall);
1507 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1508 gps = READ_ONCE(rsp->gp_start);
1509 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1510 completed = READ_ONCE(rsp->completed);
1511 if (ULONG_CMP_GE(completed, gpnum) ||
1512 ULONG_CMP_LT(j, js) ||
1513 ULONG_CMP_GE(gps, js))
1514 return; /* No stall or GP completed since entering function. */
1516 jn = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
1517 if (rcu_gp_in_progress(rsp) &&
1518 (READ_ONCE(rnp->qsmask) & rdp->grpmask) &&
1519 cmpxchg(&rsp->jiffies_stall, js, jn) == js) {
1521 /* We haven't checked in, so go dump stack. */
1522 print_cpu_stall(rsp);
1524 } else if (rcu_gp_in_progress(rsp) &&
1525 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY) &&
1526 cmpxchg(&rsp->jiffies_stall, js, jn) == js) {
1528 /* They had a few time units to dump stack, so complain. */
1529 print_other_cpu_stall(rsp, gpnum);
1534 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1536 * Set the stall-warning timeout way off into the future, thus preventing
1537 * any RCU CPU stall-warning messages from appearing in the current set of
1538 * RCU grace periods.
1540 * The caller must disable hard irqs.
1542 void rcu_cpu_stall_reset(void)
1544 struct rcu_state *rsp;
1546 for_each_rcu_flavor(rsp)
1547 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1551 * Determine the value that ->completed will have at the end of the
1552 * next subsequent grace period. This is used to tag callbacks so that
1553 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1554 * been dyntick-idle for an extended period with callbacks under the
1555 * influence of RCU_FAST_NO_HZ.
1557 * The caller must hold rnp->lock with interrupts disabled.
1559 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1560 struct rcu_node *rnp)
1562 raw_lockdep_assert_held_rcu_node(rnp);
1565 * If RCU is idle, we just wait for the next grace period.
1566 * But we can only be sure that RCU is idle if we are looking
1567 * at the root rcu_node structure -- otherwise, a new grace
1568 * period might have started, but just not yet gotten around
1569 * to initializing the current non-root rcu_node structure.
1571 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1572 return rnp->completed + 1;
1575 * If the current rcu_node structure believes that RCU is
1576 * idle, and if the rcu_state structure does not yet reflect
1577 * the start of a new grace period, then the next grace period
1578 * will suffice. The memory barrier is needed to accurately
1579 * sample the rsp->gpnum, and pairs with the second lock
1580 * acquisition in rcu_gp_init(), which is augmented with
1581 * smp_mb__after_unlock_lock() for this purpose.
1583 if (rnp->gpnum == rnp->completed) {
1584 smp_mb(); /* See above block comment. */
1585 if (READ_ONCE(rsp->gpnum) == rnp->completed)
1586 return rnp->completed + 1;
1590 * Otherwise, wait for a possible partial grace period and
1591 * then the subsequent full grace period.
1593 return rnp->completed + 2;
1596 /* Trace-event wrapper function for trace_rcu_future_grace_period. */
1597 static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1598 unsigned long c, const char *s)
1600 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1601 rnp->completed, c, rnp->level,
1602 rnp->grplo, rnp->grphi, s);
1606 * Start the specified grace period, as needed to handle newly arrived
1607 * callbacks. The required future grace periods are recorded in each
1608 * rcu_node structure's ->need_future_gp[] field. Returns true if there
1609 * is reason to awaken the grace-period kthread.
1611 * The caller must hold the specified rcu_node structure's ->lock, which
1612 * is why the caller is responsible for waking the grace-period kthread.
1614 static bool rcu_start_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1618 struct rcu_state *rsp = rdp->rsp;
1619 struct rcu_node *rnp_root;
1622 * Use funnel locking to either acquire the root rcu_node
1623 * structure's lock or bail out if the need for this grace period
1624 * has already been recorded -- or has already started. If there
1625 * is already a grace period in progress in a non-leaf node, no
1626 * recording is needed because the end of the grace period will
1627 * scan the leaf rcu_node structures. Note that rnp->lock must
1630 raw_lockdep_assert_held_rcu_node(rnp);
1631 trace_rcu_this_gp(rnp, rdp, c, TPS("Startleaf"));
1632 for (rnp_root = rnp; 1; rnp_root = rnp_root->parent) {
1633 if (rnp_root != rnp)
1634 raw_spin_lock_rcu_node(rnp_root);
1635 WARN_ON_ONCE(ULONG_CMP_LT(rnp_root->gpnum +
1636 need_future_gp_mask(), c));
1637 if (need_future_gp_element(rnp_root, c) ||
1638 ULONG_CMP_GE(rnp_root->gpnum, c) ||
1640 rnp_root->gpnum != rnp_root->completed)) {
1641 trace_rcu_this_gp(rnp_root, rdp, c, TPS("Prestarted"));
1644 need_future_gp_element(rnp_root, c) = true;
1645 if (rnp_root != rnp && rnp_root->parent != NULL)
1646 raw_spin_unlock_rcu_node(rnp_root);
1647 if (!rnp_root->parent)
1648 break; /* At root, and perhaps also leaf. */
1651 /* If GP already in progress, just leave, otherwise start one. */
1652 if (rnp_root->gpnum != rnp_root->completed) {
1653 trace_rcu_this_gp(rnp_root, rdp, c, TPS("Startedleafroot"));
1656 trace_rcu_this_gp(rnp_root, rdp, c, TPS("Startedroot"));
1657 WRITE_ONCE(rsp->gp_flags, rsp->gp_flags | RCU_GP_FLAG_INIT);
1658 rsp->gp_req_activity = jiffies;
1659 if (!rsp->gp_kthread) {
1660 trace_rcu_this_gp(rnp_root, rdp, c, TPS("NoGPkthread"));
1663 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum), TPS("newreq"));
1664 ret = true; /* Caller must wake GP kthread. */
1666 if (rnp != rnp_root)
1667 raw_spin_unlock_rcu_node(rnp_root);
1672 * Clean up any old requests for the just-ended grace period. Also return
1673 * whether any additional grace periods have been requested.
1675 static bool rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1677 unsigned long c = rnp->completed;
1679 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1681 need_future_gp_element(rnp, c) = false;
1682 needmore = need_any_future_gp(rnp);
1683 trace_rcu_this_gp(rnp, rdp, c,
1684 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1689 * Awaken the grace-period kthread for the specified flavor of RCU.
1690 * Don't do a self-awaken, and don't bother awakening when there is
1691 * nothing for the grace-period kthread to do (as in several CPUs
1692 * raced to awaken, and we lost), and finally don't try to awaken
1693 * a kthread that has not yet been created.
1695 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1697 if (current == rsp->gp_kthread ||
1698 !READ_ONCE(rsp->gp_flags) ||
1701 swake_up(&rsp->gp_wq);
1705 * If there is room, assign a ->completed number to any callbacks on
1706 * this CPU that have not already been assigned. Also accelerate any
1707 * callbacks that were previously assigned a ->completed number that has
1708 * since proven to be too conservative, which can happen if callbacks get
1709 * assigned a ->completed number while RCU is idle, but with reference to
1710 * a non-root rcu_node structure. This function is idempotent, so it does
1711 * not hurt to call it repeatedly. Returns an flag saying that we should
1712 * awaken the RCU grace-period kthread.
1714 * The caller must hold rnp->lock with interrupts disabled.
1716 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1717 struct rcu_data *rdp)
1722 raw_lockdep_assert_held_rcu_node(rnp);
1724 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1725 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1729 * Callbacks are often registered with incomplete grace-period
1730 * information. Something about the fact that getting exact
1731 * information requires acquiring a global lock... RCU therefore
1732 * makes a conservative estimate of the grace period number at which
1733 * a given callback will become ready to invoke. The following
1734 * code checks this estimate and improves it when possible, thus
1735 * accelerating callback invocation to an earlier grace-period
1738 c = rcu_cbs_completed(rsp, rnp);
1739 if (rcu_segcblist_accelerate(&rdp->cblist, c))
1740 ret = rcu_start_this_gp(rnp, rdp, c);
1742 /* Trace depending on how much we were able to accelerate. */
1743 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1744 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1746 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1751 * Move any callbacks whose grace period has completed to the
1752 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1753 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1754 * sublist. This function is idempotent, so it does not hurt to
1755 * invoke it repeatedly. As long as it is not invoked -too- often...
1756 * Returns true if the RCU grace-period kthread needs to be awakened.
1758 * The caller must hold rnp->lock with interrupts disabled.
1760 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1761 struct rcu_data *rdp)
1763 raw_lockdep_assert_held_rcu_node(rnp);
1765 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1766 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1770 * Find all callbacks whose ->completed numbers indicate that they
1771 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1773 rcu_segcblist_advance(&rdp->cblist, rnp->completed);
1775 /* Classify any remaining callbacks. */
1776 return rcu_accelerate_cbs(rsp, rnp, rdp);
1780 * Update CPU-local rcu_data state to record the beginnings and ends of
1781 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1782 * structure corresponding to the current CPU, and must have irqs disabled.
1783 * Returns true if the grace-period kthread needs to be awakened.
1785 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1786 struct rcu_data *rdp)
1791 raw_lockdep_assert_held_rcu_node(rnp);
1793 /* Handle the ends of any preceding grace periods first. */
1794 if (rdp->completed == rnp->completed &&
1795 !unlikely(READ_ONCE(rdp->gpwrap))) {
1797 /* No grace period end, so just accelerate recent callbacks. */
1798 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1802 /* Advance callbacks. */
1803 ret = rcu_advance_cbs(rsp, rnp, rdp);
1805 /* Remember that we saw this grace-period completion. */
1806 rdp->completed = rnp->completed;
1807 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1810 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1812 * If the current grace period is waiting for this CPU,
1813 * set up to detect a quiescent state, otherwise don't
1814 * go looking for one.
1816 rdp->gpnum = rnp->gpnum;
1817 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1818 need_gp = !!(rnp->qsmask & rdp->grpmask);
1819 rdp->cpu_no_qs.b.norm = need_gp;
1820 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1821 rdp->core_needs_qs = need_gp;
1822 zero_cpu_stall_ticks(rdp);
1823 WRITE_ONCE(rdp->gpwrap, false);
1824 rcu_gpnum_ovf(rnp, rdp);
1826 if (rdp->gp_seq != rnp->gp_seq)
1827 rdp->gp_seq = rnp->gp_seq;
1831 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1833 unsigned long flags;
1835 struct rcu_node *rnp;
1837 local_irq_save(flags);
1839 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1840 rdp->completed == READ_ONCE(rnp->completed) &&
1841 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1842 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1843 local_irq_restore(flags);
1846 needwake = __note_gp_changes(rsp, rnp, rdp);
1847 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1849 rcu_gp_kthread_wake(rsp);
1852 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1855 !(rcu_seq_ctr(rsp->gp_seq) %
1856 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1857 schedule_timeout_uninterruptible(delay);
1861 * Initialize a new grace period. Return false if no grace period required.
1863 static bool rcu_gp_init(struct rcu_state *rsp)
1865 unsigned long oldmask;
1866 struct rcu_data *rdp;
1867 struct rcu_node *rnp = rcu_get_root(rsp);
1869 WRITE_ONCE(rsp->gp_activity, jiffies);
1870 raw_spin_lock_irq_rcu_node(rnp);
1871 if (!READ_ONCE(rsp->gp_flags)) {
1872 /* Spurious wakeup, tell caller to go back to sleep. */
1873 raw_spin_unlock_irq_rcu_node(rnp);
1876 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1878 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1880 * Grace period already in progress, don't start another.
1881 * Not supposed to be able to happen.
1883 raw_spin_unlock_irq_rcu_node(rnp);
1887 /* Advance to a new grace period and initialize state. */
1888 record_gp_stall_check_time(rsp);
1889 /* Record GP times before starting GP, hence smp_store_release(). */
1890 WARN_ON_ONCE(rsp->gpnum << RCU_SEQ_CTR_SHIFT != rsp->gp_seq);
1891 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1892 smp_mb(); /* Pairs with barriers in stall-warning code. */
1893 rcu_seq_start(&rsp->gp_seq);
1894 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1895 raw_spin_unlock_irq_rcu_node(rnp);
1898 * Apply per-leaf buffered online and offline operations to the
1899 * rcu_node tree. Note that this new grace period need not wait
1900 * for subsequent online CPUs, and that quiescent-state forcing
1901 * will handle subsequent offline CPUs.
1903 rcu_for_each_leaf_node(rsp, rnp) {
1904 rcu_gp_slow(rsp, gp_preinit_delay);
1905 raw_spin_lock_irq_rcu_node(rnp);
1906 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1907 !rnp->wait_blkd_tasks) {
1908 /* Nothing to do on this leaf rcu_node structure. */
1909 raw_spin_unlock_irq_rcu_node(rnp);
1913 /* Record old state, apply changes to ->qsmaskinit field. */
1914 oldmask = rnp->qsmaskinit;
1915 rnp->qsmaskinit = rnp->qsmaskinitnext;
1917 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1918 if (!oldmask != !rnp->qsmaskinit) {
1919 if (!oldmask) /* First online CPU for this rcu_node. */
1920 rcu_init_new_rnp(rnp);
1921 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1922 rnp->wait_blkd_tasks = true;
1923 else /* Last offline CPU and can propagate. */
1924 rcu_cleanup_dead_rnp(rnp);
1928 * If all waited-on tasks from prior grace period are
1929 * done, and if all this rcu_node structure's CPUs are
1930 * still offline, propagate up the rcu_node tree and
1931 * clear ->wait_blkd_tasks. Otherwise, if one of this
1932 * rcu_node structure's CPUs has since come back online,
1933 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1934 * checks for this, so just call it unconditionally).
1936 if (rnp->wait_blkd_tasks &&
1937 (!rcu_preempt_has_tasks(rnp) ||
1939 rnp->wait_blkd_tasks = false;
1940 rcu_cleanup_dead_rnp(rnp);
1943 raw_spin_unlock_irq_rcu_node(rnp);
1947 * Set the quiescent-state-needed bits in all the rcu_node
1948 * structures for all currently online CPUs in breadth-first order,
1949 * starting from the root rcu_node structure, relying on the layout
1950 * of the tree within the rsp->node[] array. Note that other CPUs
1951 * will access only the leaves of the hierarchy, thus seeing that no
1952 * grace period is in progress, at least until the corresponding
1953 * leaf node has been initialized.
1955 * The grace period cannot complete until the initialization
1956 * process finishes, because this kthread handles both.
1958 rcu_for_each_node_breadth_first(rsp, rnp) {
1959 rcu_gp_slow(rsp, gp_init_delay);
1960 raw_spin_lock_irq_rcu_node(rnp);
1961 rdp = this_cpu_ptr(rsp->rda);
1962 rcu_preempt_check_blocked_tasks(rnp);
1963 rnp->qsmask = rnp->qsmaskinit;
1964 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
1965 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
1966 WRITE_ONCE(rnp->completed, rsp->completed);
1967 WRITE_ONCE(rnp->gp_seq, rsp->gp_seq);
1968 if (rnp == rdp->mynode)
1969 (void)__note_gp_changes(rsp, rnp, rdp);
1970 rcu_preempt_boost_start_gp(rnp);
1971 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1972 rnp->level, rnp->grplo,
1973 rnp->grphi, rnp->qsmask);
1974 raw_spin_unlock_irq_rcu_node(rnp);
1975 cond_resched_tasks_rcu_qs();
1976 WRITE_ONCE(rsp->gp_activity, jiffies);
1983 * Helper function for swait_event_idle() wakeup at force-quiescent-state
1986 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
1988 struct rcu_node *rnp = rcu_get_root(rsp);
1990 /* Someone like call_rcu() requested a force-quiescent-state scan. */
1991 *gfp = READ_ONCE(rsp->gp_flags);
1992 if (*gfp & RCU_GP_FLAG_FQS)
1995 /* The current grace period has completed. */
1996 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2003 * Do one round of quiescent-state forcing.
2005 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2007 struct rcu_node *rnp = rcu_get_root(rsp);
2009 WRITE_ONCE(rsp->gp_activity, jiffies);
2012 /* Collect dyntick-idle snapshots. */
2013 force_qs_rnp(rsp, dyntick_save_progress_counter);
2015 /* Handle dyntick-idle and offline CPUs. */
2016 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
2018 /* Clear flag to prevent immediate re-entry. */
2019 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2020 raw_spin_lock_irq_rcu_node(rnp);
2021 WRITE_ONCE(rsp->gp_flags,
2022 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2023 raw_spin_unlock_irq_rcu_node(rnp);
2028 * Clean up after the old grace period.
2030 static void rcu_gp_cleanup(struct rcu_state *rsp)
2032 unsigned long gp_duration;
2033 bool needgp = false;
2034 unsigned long new_gp_seq;
2035 struct rcu_data *rdp;
2036 struct rcu_node *rnp = rcu_get_root(rsp);
2037 struct swait_queue_head *sq;
2039 WRITE_ONCE(rsp->gp_activity, jiffies);
2040 raw_spin_lock_irq_rcu_node(rnp);
2041 gp_duration = jiffies - rsp->gp_start;
2042 if (gp_duration > rsp->gp_max)
2043 rsp->gp_max = gp_duration;
2046 * We know the grace period is complete, but to everyone else
2047 * it appears to still be ongoing. But it is also the case
2048 * that to everyone else it looks like there is nothing that
2049 * they can do to advance the grace period. It is therefore
2050 * safe for us to drop the lock in order to mark the grace
2051 * period as completed in all of the rcu_node structures.
2053 raw_spin_unlock_irq_rcu_node(rnp);
2056 * Propagate new ->completed value to rcu_node structures so
2057 * that other CPUs don't have to wait until the start of the next
2058 * grace period to process their callbacks. This also avoids
2059 * some nasty RCU grace-period initialization races by forcing
2060 * the end of the current grace period to be completely recorded in
2061 * all of the rcu_node structures before the beginning of the next
2062 * grace period is recorded in any of the rcu_node structures.
2064 new_gp_seq = rsp->gp_seq;
2065 rcu_seq_end(&new_gp_seq);
2066 rcu_for_each_node_breadth_first(rsp, rnp) {
2067 raw_spin_lock_irq_rcu_node(rnp);
2068 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2069 dump_blkd_tasks(rnp, 10);
2070 WARN_ON_ONCE(rnp->qsmask);
2071 WRITE_ONCE(rnp->completed, rsp->gpnum);
2072 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2073 rdp = this_cpu_ptr(rsp->rda);
2074 if (rnp == rdp->mynode)
2075 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2076 /* smp_mb() provided by prior unlock-lock pair. */
2077 needgp = rcu_future_gp_cleanup(rsp, rnp) || needgp;
2078 sq = rcu_nocb_gp_get(rnp);
2079 raw_spin_unlock_irq_rcu_node(rnp);
2080 rcu_nocb_gp_cleanup(sq);
2081 cond_resched_tasks_rcu_qs();
2082 WRITE_ONCE(rsp->gp_activity, jiffies);
2083 rcu_gp_slow(rsp, gp_cleanup_delay);
2085 rnp = rcu_get_root(rsp);
2086 raw_spin_lock_irq_rcu_node(rnp); /* GP before rsp->gp_seq update. */
2088 /* Declare grace period done. */
2089 WRITE_ONCE(rsp->completed, rsp->gpnum);
2090 rcu_seq_end(&rsp->gp_seq);
2091 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2092 rsp->gp_state = RCU_GP_IDLE;
2093 /* Check for GP requests since above loop. */
2094 rdp = this_cpu_ptr(rsp->rda);
2095 if (need_any_future_gp(rnp)) {
2096 trace_rcu_this_gp(rnp, rdp, rsp->completed - 1,
2097 TPS("CleanupMore"));
2100 /* Advance CBs to reduce false positives below. */
2101 if (!rcu_accelerate_cbs(rsp, rnp, rdp) && needgp) {
2102 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2103 rsp->gp_req_activity = jiffies;
2104 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2107 WRITE_ONCE(rsp->gp_flags, rsp->gp_flags & RCU_GP_FLAG_INIT);
2109 raw_spin_unlock_irq_rcu_node(rnp);
2113 * Body of kthread that handles grace periods.
2115 static int __noreturn rcu_gp_kthread(void *arg)
2121 struct rcu_state *rsp = arg;
2122 struct rcu_node *rnp = rcu_get_root(rsp);
2124 rcu_bind_gp_kthread();
2127 /* Handle grace-period start. */
2129 trace_rcu_grace_period(rsp->name,
2130 READ_ONCE(rsp->gpnum),
2132 rsp->gp_state = RCU_GP_WAIT_GPS;
2133 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2135 rsp->gp_state = RCU_GP_DONE_GPS;
2136 /* Locking provides needed memory barrier. */
2137 if (rcu_gp_init(rsp))
2139 cond_resched_tasks_rcu_qs();
2140 WRITE_ONCE(rsp->gp_activity, jiffies);
2141 WARN_ON(signal_pending(current));
2142 trace_rcu_grace_period(rsp->name,
2143 READ_ONCE(rsp->gpnum),
2147 /* Handle quiescent-state forcing. */
2148 first_gp_fqs = true;
2149 j = jiffies_till_first_fqs;
2152 jiffies_till_first_fqs = HZ;
2157 rsp->jiffies_force_qs = jiffies + j;
2158 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2161 trace_rcu_grace_period(rsp->name,
2162 READ_ONCE(rsp->gpnum),
2164 rsp->gp_state = RCU_GP_WAIT_FQS;
2165 ret = swait_event_idle_timeout(rsp->gp_wq,
2166 rcu_gp_fqs_check_wake(rsp, &gf), j);
2167 rsp->gp_state = RCU_GP_DOING_FQS;
2168 /* Locking provides needed memory barriers. */
2169 /* If grace period done, leave loop. */
2170 if (!READ_ONCE(rnp->qsmask) &&
2171 !rcu_preempt_blocked_readers_cgp(rnp))
2173 /* If time for quiescent-state forcing, do it. */
2174 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2175 (gf & RCU_GP_FLAG_FQS)) {
2176 trace_rcu_grace_period(rsp->name,
2177 READ_ONCE(rsp->gpnum),
2179 rcu_gp_fqs(rsp, first_gp_fqs);
2180 first_gp_fqs = false;
2181 trace_rcu_grace_period(rsp->name,
2182 READ_ONCE(rsp->gpnum),
2184 cond_resched_tasks_rcu_qs();
2185 WRITE_ONCE(rsp->gp_activity, jiffies);
2186 ret = 0; /* Force full wait till next FQS. */
2187 j = jiffies_till_next_fqs;
2190 jiffies_till_next_fqs = HZ;
2193 jiffies_till_next_fqs = 1;
2196 /* Deal with stray signal. */
2197 cond_resched_tasks_rcu_qs();
2198 WRITE_ONCE(rsp->gp_activity, jiffies);
2199 WARN_ON(signal_pending(current));
2200 trace_rcu_grace_period(rsp->name,
2201 READ_ONCE(rsp->gpnum),
2203 ret = 1; /* Keep old FQS timing. */
2205 if (time_after(jiffies, rsp->jiffies_force_qs))
2208 j = rsp->jiffies_force_qs - j;
2212 /* Handle grace-period end. */
2213 rsp->gp_state = RCU_GP_CLEANUP;
2214 rcu_gp_cleanup(rsp);
2215 rsp->gp_state = RCU_GP_CLEANED;
2220 * Report a full set of quiescent states to the specified rcu_state data
2221 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2222 * kthread if another grace period is required. Whether we wake
2223 * the grace-period kthread or it awakens itself for the next round
2224 * of quiescent-state forcing, that kthread will clean up after the
2225 * just-completed grace period. Note that the caller must hold rnp->lock,
2226 * which is released before return.
2228 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2229 __releases(rcu_get_root(rsp)->lock)
2231 raw_lockdep_assert_held_rcu_node(rcu_get_root(rsp));
2232 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2233 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2234 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2235 rcu_gp_kthread_wake(rsp);
2239 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2240 * Allows quiescent states for a group of CPUs to be reported at one go
2241 * to the specified rcu_node structure, though all the CPUs in the group
2242 * must be represented by the same rcu_node structure (which need not be a
2243 * leaf rcu_node structure, though it often will be). The gps parameter
2244 * is the grace-period snapshot, which means that the quiescent states
2245 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2246 * must be held upon entry, and it is released before return.
2249 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2250 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2251 __releases(rnp->lock)
2253 unsigned long oldmask = 0;
2254 struct rcu_node *rnp_c;
2256 raw_lockdep_assert_held_rcu_node(rnp);
2258 /* Walk up the rcu_node hierarchy. */
2260 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2263 * Our bit has already been cleared, or the
2264 * relevant grace period is already over, so done.
2266 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2269 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2270 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2271 rcu_preempt_blocked_readers_cgp(rnp));
2272 rnp->qsmask &= ~mask;
2273 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2274 mask, rnp->qsmask, rnp->level,
2275 rnp->grplo, rnp->grphi,
2277 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2279 /* Other bits still set at this level, so done. */
2280 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2283 rnp->completedqs = rnp->gpnum;
2284 mask = rnp->grpmask;
2285 if (rnp->parent == NULL) {
2287 /* No more levels. Exit loop holding root lock. */
2291 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2294 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2295 oldmask = rnp_c->qsmask;
2299 * Get here if we are the last CPU to pass through a quiescent
2300 * state for this grace period. Invoke rcu_report_qs_rsp()
2301 * to clean up and start the next grace period if one is needed.
2303 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2307 * Record a quiescent state for all tasks that were previously queued
2308 * on the specified rcu_node structure and that were blocking the current
2309 * RCU grace period. The caller must hold the specified rnp->lock with
2310 * irqs disabled, and this lock is released upon return, but irqs remain
2313 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2314 struct rcu_node *rnp, unsigned long flags)
2315 __releases(rnp->lock)
2319 struct rcu_node *rnp_p;
2321 raw_lockdep_assert_held_rcu_node(rnp);
2322 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2323 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2324 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2325 return; /* Still need more quiescent states! */
2328 rnp_p = rnp->parent;
2329 if (rnp_p == NULL) {
2331 * Only one rcu_node structure in the tree, so don't
2332 * try to report up to its nonexistent parent!
2334 rcu_report_qs_rsp(rsp, flags);
2338 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2340 mask = rnp->grpmask;
2341 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2342 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2343 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2347 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2348 * structure. This must be called from the specified CPU.
2351 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2353 unsigned long flags;
2356 struct rcu_node *rnp;
2359 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2360 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2361 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2364 * The grace period in which this quiescent state was
2365 * recorded has ended, so don't report it upwards.
2366 * We will instead need a new quiescent state that lies
2367 * within the current grace period.
2369 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2370 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2371 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2374 mask = rdp->grpmask;
2375 if ((rnp->qsmask & mask) == 0) {
2376 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2378 rdp->core_needs_qs = false;
2381 * This GP can't end until cpu checks in, so all of our
2382 * callbacks can be processed during the next GP.
2384 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2386 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2387 /* ^^^ Released rnp->lock */
2389 rcu_gp_kthread_wake(rsp);
2394 * Check to see if there is a new grace period of which this CPU
2395 * is not yet aware, and if so, set up local rcu_data state for it.
2396 * Otherwise, see if this CPU has just passed through its first
2397 * quiescent state for this grace period, and record that fact if so.
2400 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2402 /* Check for grace-period ends and beginnings. */
2403 note_gp_changes(rsp, rdp);
2406 * Does this CPU still need to do its part for current grace period?
2407 * If no, return and let the other CPUs do their part as well.
2409 if (!rdp->core_needs_qs)
2413 * Was there a quiescent state since the beginning of the grace
2414 * period? If no, then exit and wait for the next call.
2416 if (rdp->cpu_no_qs.b.norm)
2420 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2423 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2427 * Trace the fact that this CPU is going offline.
2429 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2431 RCU_TRACE(unsigned long mask;)
2432 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2433 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2435 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2438 RCU_TRACE(mask = rdp->grpmask;)
2439 trace_rcu_grace_period(rsp->name,
2440 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2445 * All CPUs for the specified rcu_node structure have gone offline,
2446 * and all tasks that were preempted within an RCU read-side critical
2447 * section while running on one of those CPUs have since exited their RCU
2448 * read-side critical section. Some other CPU is reporting this fact with
2449 * the specified rcu_node structure's ->lock held and interrupts disabled.
2450 * This function therefore goes up the tree of rcu_node structures,
2451 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2452 * the leaf rcu_node structure's ->qsmaskinit field has already been
2455 * This function does check that the specified rcu_node structure has
2456 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2457 * prematurely. That said, invoking it after the fact will cost you
2458 * a needless lock acquisition. So once it has done its work, don't
2461 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2464 struct rcu_node *rnp = rnp_leaf;
2466 raw_lockdep_assert_held_rcu_node(rnp);
2467 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2468 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2471 mask = rnp->grpmask;
2475 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2476 rnp->qsmaskinit &= ~mask;
2477 rnp->qsmask &= ~mask;
2478 if (rnp->qsmaskinit) {
2479 raw_spin_unlock_rcu_node(rnp);
2480 /* irqs remain disabled. */
2483 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2488 * The CPU has been completely removed, and some other CPU is reporting
2489 * this fact from process context. Do the remainder of the cleanup.
2490 * There can only be one CPU hotplug operation at a time, so no need for
2493 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2495 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2496 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2498 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2501 /* Adjust any no-longer-needed kthreads. */
2502 rcu_boost_kthread_setaffinity(rnp, -1);
2506 * Invoke any RCU callbacks that have made it to the end of their grace
2507 * period. Thottle as specified by rdp->blimit.
2509 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2511 unsigned long flags;
2512 struct rcu_head *rhp;
2513 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2516 /* If no callbacks are ready, just return. */
2517 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2518 trace_rcu_batch_start(rsp->name,
2519 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2520 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2521 trace_rcu_batch_end(rsp->name, 0,
2522 !rcu_segcblist_empty(&rdp->cblist),
2523 need_resched(), is_idle_task(current),
2524 rcu_is_callbacks_kthread());
2529 * Extract the list of ready callbacks, disabling to prevent
2530 * races with call_rcu() from interrupt handlers. Leave the
2531 * callback counts, as rcu_barrier() needs to be conservative.
2533 local_irq_save(flags);
2534 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2536 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2537 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2538 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2539 local_irq_restore(flags);
2541 /* Invoke callbacks. */
2542 rhp = rcu_cblist_dequeue(&rcl);
2543 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2544 debug_rcu_head_unqueue(rhp);
2545 if (__rcu_reclaim(rsp->name, rhp))
2546 rcu_cblist_dequeued_lazy(&rcl);
2548 * Stop only if limit reached and CPU has something to do.
2549 * Note: The rcl structure counts down from zero.
2551 if (-rcl.len >= bl &&
2553 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2557 local_irq_save(flags);
2559 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2560 is_idle_task(current), rcu_is_callbacks_kthread());
2562 /* Update counts and requeue any remaining callbacks. */
2563 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2564 smp_mb(); /* List handling before counting for rcu_barrier(). */
2565 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2567 /* Reinstate batch limit if we have worked down the excess. */
2568 count = rcu_segcblist_n_cbs(&rdp->cblist);
2569 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2570 rdp->blimit = blimit;
2572 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2573 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2574 rdp->qlen_last_fqs_check = 0;
2575 rdp->n_force_qs_snap = rsp->n_force_qs;
2576 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2577 rdp->qlen_last_fqs_check = count;
2580 * The following usually indicates a double call_rcu(). To track
2581 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2583 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2585 local_irq_restore(flags);
2587 /* Re-invoke RCU core processing if there are callbacks remaining. */
2588 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2593 * Check to see if this CPU is in a non-context-switch quiescent state
2594 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2595 * Also schedule RCU core processing.
2597 * This function must be called from hardirq context. It is normally
2598 * invoked from the scheduling-clock interrupt.
2600 void rcu_check_callbacks(int user)
2602 trace_rcu_utilization(TPS("Start scheduler-tick"));
2603 increment_cpu_stall_ticks();
2604 if (user || rcu_is_cpu_rrupt_from_idle()) {
2607 * Get here if this CPU took its interrupt from user
2608 * mode or from the idle loop, and if this is not a
2609 * nested interrupt. In this case, the CPU is in
2610 * a quiescent state, so note it.
2612 * No memory barrier is required here because both
2613 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2614 * variables that other CPUs neither access nor modify,
2615 * at least not while the corresponding CPU is online.
2621 } else if (!in_softirq()) {
2624 * Get here if this CPU did not take its interrupt from
2625 * softirq, in other words, if it is not interrupting
2626 * a rcu_bh read-side critical section. This is an _bh
2627 * critical section, so note it.
2632 rcu_preempt_check_callbacks();
2636 rcu_note_voluntary_context_switch(current);
2637 trace_rcu_utilization(TPS("End scheduler-tick"));
2641 * Scan the leaf rcu_node structures, processing dyntick state for any that
2642 * have not yet encountered a quiescent state, using the function specified.
2643 * Also initiate boosting for any threads blocked on the root rcu_node.
2645 * The caller must have suppressed start of new grace periods.
2647 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2650 unsigned long flags;
2652 struct rcu_node *rnp;
2654 rcu_for_each_leaf_node(rsp, rnp) {
2655 cond_resched_tasks_rcu_qs();
2657 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2658 if (rnp->qsmask == 0) {
2659 if (rcu_state_p == &rcu_sched_state ||
2660 rsp != rcu_state_p ||
2661 rcu_preempt_blocked_readers_cgp(rnp)) {
2663 * No point in scanning bits because they
2664 * are all zero. But we might need to
2665 * priority-boost blocked readers.
2667 rcu_initiate_boost(rnp, flags);
2668 /* rcu_initiate_boost() releases rnp->lock */
2672 (rnp->parent->qsmask & rnp->grpmask)) {
2674 * Race between grace-period
2675 * initialization and task exiting RCU
2676 * read-side critical section: Report.
2678 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2679 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2683 for_each_leaf_node_possible_cpu(rnp, cpu) {
2684 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2685 if ((rnp->qsmask & bit) != 0) {
2686 if (f(per_cpu_ptr(rsp->rda, cpu)))
2691 /* Idle/offline CPUs, report (releases rnp->lock. */
2692 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2694 /* Nothing to do here, so just drop the lock. */
2695 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2701 * Force quiescent states on reluctant CPUs, and also detect which
2702 * CPUs are in dyntick-idle mode.
2704 static void force_quiescent_state(struct rcu_state *rsp)
2706 unsigned long flags;
2708 struct rcu_node *rnp;
2709 struct rcu_node *rnp_old = NULL;
2711 /* Funnel through hierarchy to reduce memory contention. */
2712 rnp = __this_cpu_read(rsp->rda->mynode);
2713 for (; rnp != NULL; rnp = rnp->parent) {
2714 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2715 !raw_spin_trylock(&rnp->fqslock);
2716 if (rnp_old != NULL)
2717 raw_spin_unlock(&rnp_old->fqslock);
2722 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2724 /* Reached the root of the rcu_node tree, acquire lock. */
2725 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2726 raw_spin_unlock(&rnp_old->fqslock);
2727 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2728 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2729 return; /* Someone beat us to it. */
2731 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2732 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2733 rcu_gp_kthread_wake(rsp);
2737 * This function checks for grace-period requests that fail to motivate
2738 * RCU to come out of its idle mode.
2741 rcu_check_gp_start_stall(struct rcu_state *rsp, struct rcu_node *rnp,
2742 struct rcu_data *rdp)
2744 unsigned long flags;
2746 struct rcu_node *rnp_root = rcu_get_root(rsp);
2747 static atomic_t warned = ATOMIC_INIT(0);
2749 if (!IS_ENABLED(CONFIG_PROVE_RCU) ||
2750 rcu_gp_in_progress(rsp) || !need_any_future_gp(rcu_get_root(rsp)))
2752 j = jiffies; /* Expensive access, and in common case don't get here. */
2753 if (time_before(j, READ_ONCE(rsp->gp_req_activity) + HZ) ||
2754 time_before(j, READ_ONCE(rsp->gp_activity) + HZ) ||
2755 atomic_read(&warned))
2758 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2760 if (rcu_gp_in_progress(rsp) || !need_any_future_gp(rcu_get_root(rsp)) ||
2761 time_before(j, READ_ONCE(rsp->gp_req_activity) + HZ) ||
2762 time_before(j, READ_ONCE(rsp->gp_activity) + HZ) ||
2763 atomic_read(&warned)) {
2764 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2767 /* Hold onto the leaf lock to make others see warned==1. */
2769 if (rnp_root != rnp)
2770 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
2772 if (rcu_gp_in_progress(rsp) || !need_any_future_gp(rcu_get_root(rsp)) ||
2773 time_before(j, rsp->gp_req_activity + HZ) ||
2774 time_before(j, rsp->gp_activity + HZ) ||
2775 atomic_xchg(&warned, 1)) {
2776 raw_spin_unlock_rcu_node(rnp_root); /* irqs remain disabled. */
2777 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2780 pr_alert("%s: g%lu %d%d%d%d gar:%lu ga:%lu f%#x %s->state:%#lx\n",
2781 __func__, READ_ONCE(rsp->gpnum),
2782 need_future_gp_element(rcu_get_root(rsp), 0),
2783 need_future_gp_element(rcu_get_root(rsp), 1),
2784 need_future_gp_element(rcu_get_root(rsp), 2),
2785 need_future_gp_element(rcu_get_root(rsp), 3),
2786 j - rsp->gp_req_activity, j - rsp->gp_activity,
2787 rsp->gp_flags, rsp->name,
2788 rsp->gp_kthread ? rsp->gp_kthread->state : 0x1ffffL);
2790 if (rnp_root != rnp)
2791 raw_spin_unlock_rcu_node(rnp_root);
2792 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2796 * This does the RCU core processing work for the specified rcu_state
2797 * and rcu_data structures. This may be called only from the CPU to
2798 * whom the rdp belongs.
2801 __rcu_process_callbacks(struct rcu_state *rsp)
2803 unsigned long flags;
2805 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2806 struct rcu_node *rnp = rdp->mynode;
2808 WARN_ON_ONCE(!rdp->beenonline);
2810 /* Update RCU state based on any recent quiescent states. */
2811 rcu_check_quiescent_state(rsp, rdp);
2813 /* No grace period and unregistered callbacks? */
2814 if (!rcu_gp_in_progress(rsp) &&
2815 rcu_segcblist_is_enabled(&rdp->cblist)) {
2816 local_irq_save(flags);
2817 if (rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL)) {
2818 local_irq_restore(flags);
2820 raw_spin_lock_rcu_node(rnp); /* irqs disabled. */
2821 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2822 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2824 rcu_gp_kthread_wake(rsp);
2828 rcu_check_gp_start_stall(rsp, rnp, rdp);
2830 /* If there are callbacks ready, invoke them. */
2831 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2832 invoke_rcu_callbacks(rsp, rdp);
2834 /* Do any needed deferred wakeups of rcuo kthreads. */
2835 do_nocb_deferred_wakeup(rdp);
2839 * Do RCU core processing for the current CPU.
2841 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2843 struct rcu_state *rsp;
2845 if (cpu_is_offline(smp_processor_id()))
2847 trace_rcu_utilization(TPS("Start RCU core"));
2848 for_each_rcu_flavor(rsp)
2849 __rcu_process_callbacks(rsp);
2850 trace_rcu_utilization(TPS("End RCU core"));
2854 * Schedule RCU callback invocation. If the specified type of RCU
2855 * does not support RCU priority boosting, just do a direct call,
2856 * otherwise wake up the per-CPU kernel kthread. Note that because we
2857 * are running on the current CPU with softirqs disabled, the
2858 * rcu_cpu_kthread_task cannot disappear out from under us.
2860 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2862 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
2864 if (likely(!rsp->boost)) {
2865 rcu_do_batch(rsp, rdp);
2868 invoke_rcu_callbacks_kthread();
2871 static void invoke_rcu_core(void)
2873 if (cpu_online(smp_processor_id()))
2874 raise_softirq(RCU_SOFTIRQ);
2878 * Handle any core-RCU processing required by a call_rcu() invocation.
2880 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2881 struct rcu_head *head, unsigned long flags)
2886 * If called from an extended quiescent state, invoke the RCU
2887 * core in order to force a re-evaluation of RCU's idleness.
2889 if (!rcu_is_watching())
2892 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2893 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2897 * Force the grace period if too many callbacks or too long waiting.
2898 * Enforce hysteresis, and don't invoke force_quiescent_state()
2899 * if some other CPU has recently done so. Also, don't bother
2900 * invoking force_quiescent_state() if the newly enqueued callback
2901 * is the only one waiting for a grace period to complete.
2903 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2904 rdp->qlen_last_fqs_check + qhimark)) {
2906 /* Are we ignoring a completed grace period? */
2907 note_gp_changes(rsp, rdp);
2909 /* Start a new grace period if one not already started. */
2910 if (!rcu_gp_in_progress(rsp)) {
2911 struct rcu_node *rnp = rdp->mynode;
2913 raw_spin_lock_rcu_node(rnp);
2914 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2915 raw_spin_unlock_rcu_node(rnp);
2917 rcu_gp_kthread_wake(rsp);
2919 /* Give the grace period a kick. */
2920 rdp->blimit = LONG_MAX;
2921 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2922 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2923 force_quiescent_state(rsp);
2924 rdp->n_force_qs_snap = rsp->n_force_qs;
2925 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2931 * RCU callback function to leak a callback.
2933 static void rcu_leak_callback(struct rcu_head *rhp)
2938 * Helper function for call_rcu() and friends. The cpu argument will
2939 * normally be -1, indicating "currently running CPU". It may specify
2940 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2941 * is expected to specify a CPU.
2944 __call_rcu(struct rcu_head *head, rcu_callback_t func,
2945 struct rcu_state *rsp, int cpu, bool lazy)
2947 unsigned long flags;
2948 struct rcu_data *rdp;
2950 /* Misaligned rcu_head! */
2951 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
2953 if (debug_rcu_head_queue(head)) {
2955 * Probable double call_rcu(), so leak the callback.
2956 * Use rcu:rcu_callback trace event to find the previous
2957 * time callback was passed to __call_rcu().
2959 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
2961 WRITE_ONCE(head->func, rcu_leak_callback);
2966 local_irq_save(flags);
2967 rdp = this_cpu_ptr(rsp->rda);
2969 /* Add the callback to our list. */
2970 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
2974 rdp = per_cpu_ptr(rsp->rda, cpu);
2975 if (likely(rdp->mynode)) {
2976 /* Post-boot, so this should be for a no-CBs CPU. */
2977 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2978 WARN_ON_ONCE(offline);
2979 /* Offline CPU, _call_rcu() illegal, leak callback. */
2980 local_irq_restore(flags);
2984 * Very early boot, before rcu_init(). Initialize if needed
2985 * and then drop through to queue the callback.
2988 WARN_ON_ONCE(!rcu_is_watching());
2989 if (rcu_segcblist_empty(&rdp->cblist))
2990 rcu_segcblist_init(&rdp->cblist);
2992 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
2994 rcu_idle_count_callbacks_posted();
2996 if (__is_kfree_rcu_offset((unsigned long)func))
2997 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2998 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2999 rcu_segcblist_n_cbs(&rdp->cblist));
3001 trace_rcu_callback(rsp->name, head,
3002 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3003 rcu_segcblist_n_cbs(&rdp->cblist));
3005 /* Go handle any RCU core processing required. */
3006 __call_rcu_core(rsp, rdp, head, flags);
3007 local_irq_restore(flags);
3011 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
3012 * @head: structure to be used for queueing the RCU updates.
3013 * @func: actual callback function to be invoked after the grace period
3015 * The callback function will be invoked some time after a full grace
3016 * period elapses, in other words after all currently executing RCU
3017 * read-side critical sections have completed. call_rcu_sched() assumes
3018 * that the read-side critical sections end on enabling of preemption
3019 * or on voluntary preemption.
3020 * RCU read-side critical sections are delimited by:
3022 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
3023 * - anything that disables preemption.
3025 * These may be nested.
3027 * See the description of call_rcu() for more detailed information on
3028 * memory ordering guarantees.
3030 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3032 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3034 EXPORT_SYMBOL_GPL(call_rcu_sched);
3037 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
3038 * @head: structure to be used for queueing the RCU updates.
3039 * @func: actual callback function to be invoked after the grace period
3041 * The callback function will be invoked some time after a full grace
3042 * period elapses, in other words after all currently executing RCU
3043 * read-side critical sections have completed. call_rcu_bh() assumes
3044 * that the read-side critical sections end on completion of a softirq
3045 * handler. This means that read-side critical sections in process
3046 * context must not be interrupted by softirqs. This interface is to be
3047 * used when most of the read-side critical sections are in softirq context.
3048 * RCU read-side critical sections are delimited by:
3050 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3051 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3053 * These may be nested.
3055 * See the description of call_rcu() for more detailed information on
3056 * memory ordering guarantees.
3058 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3060 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3062 EXPORT_SYMBOL_GPL(call_rcu_bh);
3065 * Queue an RCU callback for lazy invocation after a grace period.
3066 * This will likely be later named something like "call_rcu_lazy()",
3067 * but this change will require some way of tagging the lazy RCU
3068 * callbacks in the list of pending callbacks. Until then, this
3069 * function may only be called from __kfree_rcu().
3071 void kfree_call_rcu(struct rcu_head *head,
3072 rcu_callback_t func)
3074 __call_rcu(head, func, rcu_state_p, -1, 1);
3076 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3079 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3080 * any blocking grace-period wait automatically implies a grace period
3081 * if there is only one CPU online at any point time during execution
3082 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3083 * occasionally incorrectly indicate that there are multiple CPUs online
3084 * when there was in fact only one the whole time, as this just adds
3085 * some overhead: RCU still operates correctly.
3087 static inline int rcu_blocking_is_gp(void)
3091 might_sleep(); /* Check for RCU read-side critical section. */
3093 ret = num_online_cpus() <= 1;
3099 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3101 * Control will return to the caller some time after a full rcu-sched
3102 * grace period has elapsed, in other words after all currently executing
3103 * rcu-sched read-side critical sections have completed. These read-side
3104 * critical sections are delimited by rcu_read_lock_sched() and
3105 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3106 * local_irq_disable(), and so on may be used in place of
3107 * rcu_read_lock_sched().
3109 * This means that all preempt_disable code sequences, including NMI and
3110 * non-threaded hardware-interrupt handlers, in progress on entry will
3111 * have completed before this primitive returns. However, this does not
3112 * guarantee that softirq handlers will have completed, since in some
3113 * kernels, these handlers can run in process context, and can block.
3115 * Note that this guarantee implies further memory-ordering guarantees.
3116 * On systems with more than one CPU, when synchronize_sched() returns,
3117 * each CPU is guaranteed to have executed a full memory barrier since the
3118 * end of its last RCU-sched read-side critical section whose beginning
3119 * preceded the call to synchronize_sched(). In addition, each CPU having
3120 * an RCU read-side critical section that extends beyond the return from
3121 * synchronize_sched() is guaranteed to have executed a full memory barrier
3122 * after the beginning of synchronize_sched() and before the beginning of
3123 * that RCU read-side critical section. Note that these guarantees include
3124 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3125 * that are executing in the kernel.
3127 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3128 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3129 * to have executed a full memory barrier during the execution of
3130 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3131 * again only if the system has more than one CPU).
3133 void synchronize_sched(void)
3135 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3136 lock_is_held(&rcu_lock_map) ||
3137 lock_is_held(&rcu_sched_lock_map),
3138 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3139 if (rcu_blocking_is_gp())
3141 if (rcu_gp_is_expedited())
3142 synchronize_sched_expedited();
3144 wait_rcu_gp(call_rcu_sched);
3146 EXPORT_SYMBOL_GPL(synchronize_sched);
3149 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3151 * Control will return to the caller some time after a full rcu_bh grace
3152 * period has elapsed, in other words after all currently executing rcu_bh
3153 * read-side critical sections have completed. RCU read-side critical
3154 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3155 * and may be nested.
3157 * See the description of synchronize_sched() for more detailed information
3158 * on memory ordering guarantees.
3160 void synchronize_rcu_bh(void)
3162 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3163 lock_is_held(&rcu_lock_map) ||
3164 lock_is_held(&rcu_sched_lock_map),
3165 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3166 if (rcu_blocking_is_gp())
3168 if (rcu_gp_is_expedited())
3169 synchronize_rcu_bh_expedited();
3171 wait_rcu_gp(call_rcu_bh);
3173 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3176 * get_state_synchronize_rcu - Snapshot current RCU state
3178 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3179 * to determine whether or not a full grace period has elapsed in the
3182 unsigned long get_state_synchronize_rcu(void)
3185 * Any prior manipulation of RCU-protected data must happen
3186 * before the load from ->gpnum.
3191 * Make sure this load happens before the purportedly
3192 * time-consuming work between get_state_synchronize_rcu()
3193 * and cond_synchronize_rcu().
3195 return smp_load_acquire(&rcu_state_p->gpnum);
3197 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3200 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3202 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3204 * If a full RCU grace period has elapsed since the earlier call to
3205 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3206 * synchronize_rcu() to wait for a full grace period.
3208 * Yes, this function does not take counter wrap into account. But
3209 * counter wrap is harmless. If the counter wraps, we have waited for
3210 * more than 2 billion grace periods (and way more on a 64-bit system!),
3211 * so waiting for one additional grace period should be just fine.
3213 void cond_synchronize_rcu(unsigned long oldstate)
3215 unsigned long newstate;
3218 * Ensure that this load happens before any RCU-destructive
3219 * actions the caller might carry out after we return.
3221 newstate = smp_load_acquire(&rcu_state_p->completed);
3222 if (ULONG_CMP_GE(oldstate, newstate))
3225 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3228 * get_state_synchronize_sched - Snapshot current RCU-sched state
3230 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3231 * to determine whether or not a full grace period has elapsed in the
3234 unsigned long get_state_synchronize_sched(void)
3237 * Any prior manipulation of RCU-protected data must happen
3238 * before the load from ->gpnum.
3243 * Make sure this load happens before the purportedly
3244 * time-consuming work between get_state_synchronize_sched()
3245 * and cond_synchronize_sched().
3247 return smp_load_acquire(&rcu_sched_state.gpnum);
3249 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3252 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3254 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3256 * If a full RCU-sched grace period has elapsed since the earlier call to
3257 * get_state_synchronize_sched(), just return. Otherwise, invoke
3258 * synchronize_sched() to wait for a full grace period.
3260 * Yes, this function does not take counter wrap into account. But
3261 * counter wrap is harmless. If the counter wraps, we have waited for
3262 * more than 2 billion grace periods (and way more on a 64-bit system!),
3263 * so waiting for one additional grace period should be just fine.
3265 void cond_synchronize_sched(unsigned long oldstate)
3267 unsigned long newstate;
3270 * Ensure that this load happens before any RCU-destructive
3271 * actions the caller might carry out after we return.
3273 newstate = smp_load_acquire(&rcu_sched_state.completed);
3274 if (ULONG_CMP_GE(oldstate, newstate))
3275 synchronize_sched();
3277 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3280 * Check to see if there is any immediate RCU-related work to be done
3281 * by the current CPU, for the specified type of RCU, returning 1 if so.
3282 * The checks are in order of increasing expense: checks that can be
3283 * carried out against CPU-local state are performed first. However,
3284 * we must check for CPU stalls first, else we might not get a chance.
3286 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3288 struct rcu_node *rnp = rdp->mynode;
3290 /* Check for CPU stalls, if enabled. */
3291 check_cpu_stall(rsp, rdp);
3293 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3294 if (rcu_nohz_full_cpu(rsp))
3297 /* Is the RCU core waiting for a quiescent state from this CPU? */
3298 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm)
3301 /* Does this CPU have callbacks ready to invoke? */
3302 if (rcu_segcblist_ready_cbs(&rdp->cblist))
3305 /* Has RCU gone idle with this CPU needing another grace period? */
3306 if (!rcu_gp_in_progress(rsp) &&
3307 rcu_segcblist_is_enabled(&rdp->cblist) &&
3308 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3311 /* Has another RCU grace period completed? */
3312 if (READ_ONCE(rnp->completed) != rdp->completed) /* outside lock */
3315 /* Has a new RCU grace period started? */
3316 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3317 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3320 /* Does this CPU need a deferred NOCB wakeup? */
3321 if (rcu_nocb_need_deferred_wakeup(rdp))
3329 * Check to see if there is any immediate RCU-related work to be done
3330 * by the current CPU, returning 1 if so. This function is part of the
3331 * RCU implementation; it is -not- an exported member of the RCU API.
3333 static int rcu_pending(void)
3335 struct rcu_state *rsp;
3337 for_each_rcu_flavor(rsp)
3338 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3344 * Return true if the specified CPU has any callback. If all_lazy is
3345 * non-NULL, store an indication of whether all callbacks are lazy.
3346 * (If there are no callbacks, all of them are deemed to be lazy.)
3348 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3352 struct rcu_data *rdp;
3353 struct rcu_state *rsp;
3355 for_each_rcu_flavor(rsp) {
3356 rdp = this_cpu_ptr(rsp->rda);
3357 if (rcu_segcblist_empty(&rdp->cblist))
3360 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3371 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3372 * the compiler is expected to optimize this away.
3374 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3375 int cpu, unsigned long done)
3377 trace_rcu_barrier(rsp->name, s, cpu,
3378 atomic_read(&rsp->barrier_cpu_count), done);
3382 * RCU callback function for _rcu_barrier(). If we are last, wake
3383 * up the task executing _rcu_barrier().
3385 static void rcu_barrier_callback(struct rcu_head *rhp)
3387 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3388 struct rcu_state *rsp = rdp->rsp;
3390 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3391 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3392 rsp->barrier_sequence);
3393 complete(&rsp->barrier_completion);
3395 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3400 * Called with preemption disabled, and from cross-cpu IRQ context.
3402 static void rcu_barrier_func(void *type)
3404 struct rcu_state *rsp = type;
3405 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3407 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3408 rdp->barrier_head.func = rcu_barrier_callback;
3409 debug_rcu_head_queue(&rdp->barrier_head);
3410 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3411 atomic_inc(&rsp->barrier_cpu_count);
3413 debug_rcu_head_unqueue(&rdp->barrier_head);
3414 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3415 rsp->barrier_sequence);
3420 * Orchestrate the specified type of RCU barrier, waiting for all
3421 * RCU callbacks of the specified type to complete.
3423 static void _rcu_barrier(struct rcu_state *rsp)
3426 struct rcu_data *rdp;
3427 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3429 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3431 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3432 mutex_lock(&rsp->barrier_mutex);
3434 /* Did someone else do our work for us? */
3435 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3436 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3437 rsp->barrier_sequence);
3438 smp_mb(); /* caller's subsequent code after above check. */
3439 mutex_unlock(&rsp->barrier_mutex);
3443 /* Mark the start of the barrier operation. */
3444 rcu_seq_start(&rsp->barrier_sequence);
3445 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3448 * Initialize the count to one rather than to zero in order to
3449 * avoid a too-soon return to zero in case of a short grace period
3450 * (or preemption of this task). Exclude CPU-hotplug operations
3451 * to ensure that no offline CPU has callbacks queued.
3453 init_completion(&rsp->barrier_completion);
3454 atomic_set(&rsp->barrier_cpu_count, 1);
3458 * Force each CPU with callbacks to register a new callback.
3459 * When that callback is invoked, we will know that all of the
3460 * corresponding CPU's preceding callbacks have been invoked.
3462 for_each_possible_cpu(cpu) {
3463 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3465 rdp = per_cpu_ptr(rsp->rda, cpu);
3466 if (rcu_is_nocb_cpu(cpu)) {
3467 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3468 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3469 rsp->barrier_sequence);
3471 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3472 rsp->barrier_sequence);
3473 smp_mb__before_atomic();
3474 atomic_inc(&rsp->barrier_cpu_count);
3475 __call_rcu(&rdp->barrier_head,
3476 rcu_barrier_callback, rsp, cpu, 0);
3478 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3479 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3480 rsp->barrier_sequence);
3481 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3483 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3484 rsp->barrier_sequence);
3490 * Now that we have an rcu_barrier_callback() callback on each
3491 * CPU, and thus each counted, remove the initial count.
3493 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3494 complete(&rsp->barrier_completion);
3496 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3497 wait_for_completion(&rsp->barrier_completion);
3499 /* Mark the end of the barrier operation. */
3500 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3501 rcu_seq_end(&rsp->barrier_sequence);
3503 /* Other rcu_barrier() invocations can now safely proceed. */
3504 mutex_unlock(&rsp->barrier_mutex);
3508 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3510 void rcu_barrier_bh(void)
3512 _rcu_barrier(&rcu_bh_state);
3514 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3517 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3519 void rcu_barrier_sched(void)
3521 _rcu_barrier(&rcu_sched_state);
3523 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3526 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3527 * first CPU in a given leaf rcu_node structure coming online. The caller
3528 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3531 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3534 struct rcu_node *rnp = rnp_leaf;
3536 raw_lockdep_assert_held_rcu_node(rnp);
3538 mask = rnp->grpmask;
3542 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3543 rnp->qsmaskinit |= mask;
3544 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3549 * Do boot-time initialization of a CPU's per-CPU RCU data.
3552 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3554 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3556 /* Set up local state, ensuring consistent view of global state. */
3557 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3558 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3559 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != 1);
3560 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3563 rcu_boot_init_nocb_percpu_data(rdp);
3567 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3568 * offline event can be happening at a given time. Note also that we
3569 * can accept some slop in the rsp->completed access due to the fact
3570 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3573 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3575 unsigned long flags;
3576 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3577 struct rcu_node *rnp = rcu_get_root(rsp);
3579 /* Set up local state, ensuring consistent view of global state. */
3580 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3581 rdp->qlen_last_fqs_check = 0;
3582 rdp->n_force_qs_snap = rsp->n_force_qs;
3583 rdp->blimit = blimit;
3584 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3585 !init_nocb_callback_list(rdp))
3586 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3587 rdp->dynticks->dynticks_nesting = 1; /* CPU not up, no tearing. */
3588 rcu_dynticks_eqs_online();
3589 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3592 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3593 * propagation up the rcu_node tree will happen at the beginning
3594 * of the next grace period.
3597 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3598 rdp->beenonline = true; /* We have now been online. */
3599 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3600 rdp->completed = rnp->completed;
3601 rdp->gp_seq = rnp->gp_seq;
3602 rdp->cpu_no_qs.b.norm = true;
3603 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3604 rdp->core_needs_qs = false;
3605 rdp->rcu_iw_pending = false;
3606 rdp->rcu_iw_gpnum = rnp->gpnum - 1;
3607 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3608 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3612 * Invoked early in the CPU-online process, when pretty much all
3613 * services are available. The incoming CPU is not present.
3615 int rcutree_prepare_cpu(unsigned int cpu)
3617 struct rcu_state *rsp;
3619 for_each_rcu_flavor(rsp)
3620 rcu_init_percpu_data(cpu, rsp);
3622 rcu_prepare_kthreads(cpu);
3623 rcu_spawn_all_nocb_kthreads(cpu);
3629 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3631 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3633 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3635 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3639 * Near the end of the CPU-online process. Pretty much all services
3640 * enabled, and the CPU is now very much alive.
3642 int rcutree_online_cpu(unsigned int cpu)
3644 unsigned long flags;
3645 struct rcu_data *rdp;
3646 struct rcu_node *rnp;
3647 struct rcu_state *rsp;
3649 for_each_rcu_flavor(rsp) {
3650 rdp = per_cpu_ptr(rsp->rda, cpu);
3652 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3653 rnp->ffmask |= rdp->grpmask;
3654 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3656 if (IS_ENABLED(CONFIG_TREE_SRCU))
3657 srcu_online_cpu(cpu);
3658 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3659 return 0; /* Too early in boot for scheduler work. */
3660 sync_sched_exp_online_cleanup(cpu);
3661 rcutree_affinity_setting(cpu, -1);
3666 * Near the beginning of the process. The CPU is still very much alive
3667 * with pretty much all services enabled.
3669 int rcutree_offline_cpu(unsigned int cpu)
3671 unsigned long flags;
3672 struct rcu_data *rdp;
3673 struct rcu_node *rnp;
3674 struct rcu_state *rsp;
3676 for_each_rcu_flavor(rsp) {
3677 rdp = per_cpu_ptr(rsp->rda, cpu);
3679 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3680 rnp->ffmask &= ~rdp->grpmask;
3681 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3684 rcutree_affinity_setting(cpu, cpu);
3685 if (IS_ENABLED(CONFIG_TREE_SRCU))
3686 srcu_offline_cpu(cpu);
3691 * Near the end of the offline process. We do only tracing here.
3693 int rcutree_dying_cpu(unsigned int cpu)
3695 struct rcu_state *rsp;
3697 for_each_rcu_flavor(rsp)
3698 rcu_cleanup_dying_cpu(rsp);
3703 * The outgoing CPU is gone and we are running elsewhere.
3705 int rcutree_dead_cpu(unsigned int cpu)
3707 struct rcu_state *rsp;
3709 for_each_rcu_flavor(rsp) {
3710 rcu_cleanup_dead_cpu(cpu, rsp);
3711 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3716 static DEFINE_PER_CPU(int, rcu_cpu_started);
3719 * Mark the specified CPU as being online so that subsequent grace periods
3720 * (both expedited and normal) will wait on it. Note that this means that
3721 * incoming CPUs are not allowed to use RCU read-side critical sections
3722 * until this function is called. Failing to observe this restriction
3723 * will result in lockdep splats.
3725 * Note that this function is special in that it is invoked directly
3726 * from the incoming CPU rather than from the cpuhp_step mechanism.
3727 * This is because this function must be invoked at a precise location.
3729 void rcu_cpu_starting(unsigned int cpu)
3731 unsigned long flags;
3734 unsigned long oldmask;
3735 struct rcu_data *rdp;
3736 struct rcu_node *rnp;
3737 struct rcu_state *rsp;
3739 if (per_cpu(rcu_cpu_started, cpu))
3742 per_cpu(rcu_cpu_started, cpu) = 1;
3744 for_each_rcu_flavor(rsp) {
3745 rdp = per_cpu_ptr(rsp->rda, cpu);
3747 mask = rdp->grpmask;
3748 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3749 rnp->qsmaskinitnext |= mask;
3750 oldmask = rnp->expmaskinitnext;
3751 rnp->expmaskinitnext |= mask;
3752 oldmask ^= rnp->expmaskinitnext;
3753 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3754 /* Allow lockless access for expedited grace periods. */
3755 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3756 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3758 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3761 #ifdef CONFIG_HOTPLUG_CPU
3763 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3764 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3767 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3769 unsigned long flags;
3771 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3772 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3774 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3775 mask = rdp->grpmask;
3776 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3777 rnp->qsmaskinitnext &= ~mask;
3778 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3782 * The outgoing function has no further need of RCU, so remove it from
3783 * the list of CPUs that RCU must track.
3785 * Note that this function is special in that it is invoked directly
3786 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3787 * This is because this function must be invoked at a precise location.
3789 void rcu_report_dead(unsigned int cpu)
3791 struct rcu_state *rsp;
3793 /* QS for any half-done expedited RCU-sched GP. */
3795 rcu_report_exp_rdp(&rcu_sched_state,
3796 this_cpu_ptr(rcu_sched_state.rda), true);
3798 for_each_rcu_flavor(rsp)
3799 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3801 per_cpu(rcu_cpu_started, cpu) = 0;
3804 /* Migrate the dead CPU's callbacks to the current CPU. */
3805 static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3807 unsigned long flags;
3808 struct rcu_data *my_rdp;
3809 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3810 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3813 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3814 return; /* No callbacks to migrate. */
3816 local_irq_save(flags);
3817 my_rdp = this_cpu_ptr(rsp->rda);
3818 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3819 local_irq_restore(flags);
3822 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3823 /* Leverage recent GPs and set GP for new callbacks. */
3824 needwake = rcu_advance_cbs(rsp, rnp_root, rdp) ||
3825 rcu_advance_cbs(rsp, rnp_root, my_rdp);
3826 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3827 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3828 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3829 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3831 rcu_gp_kthread_wake(rsp);
3832 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3833 !rcu_segcblist_empty(&rdp->cblist),
3834 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3835 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3836 rcu_segcblist_first_cb(&rdp->cblist));
3840 * The outgoing CPU has just passed through the dying-idle state,
3841 * and we are being invoked from the CPU that was IPIed to continue the
3842 * offline operation. We need to migrate the outgoing CPU's callbacks.
3844 void rcutree_migrate_callbacks(int cpu)
3846 struct rcu_state *rsp;
3848 for_each_rcu_flavor(rsp)
3849 rcu_migrate_callbacks(cpu, rsp);
3854 * On non-huge systems, use expedited RCU grace periods to make suspend
3855 * and hibernation run faster.
3857 static int rcu_pm_notify(struct notifier_block *self,
3858 unsigned long action, void *hcpu)
3861 case PM_HIBERNATION_PREPARE:
3862 case PM_SUSPEND_PREPARE:
3863 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3866 case PM_POST_HIBERNATION:
3867 case PM_POST_SUSPEND:
3868 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3869 rcu_unexpedite_gp();
3878 * Spawn the kthreads that handle each RCU flavor's grace periods.
3880 static int __init rcu_spawn_gp_kthread(void)
3882 unsigned long flags;
3883 int kthread_prio_in = kthread_prio;
3884 struct rcu_node *rnp;
3885 struct rcu_state *rsp;
3886 struct sched_param sp;
3887 struct task_struct *t;
3889 /* Force priority into range. */
3890 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3892 else if (kthread_prio < 0)
3894 else if (kthread_prio > 99)
3896 if (kthread_prio != kthread_prio_in)
3897 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3898 kthread_prio, kthread_prio_in);
3900 rcu_scheduler_fully_active = 1;
3901 for_each_rcu_flavor(rsp) {
3902 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3904 rnp = rcu_get_root(rsp);
3905 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3906 rsp->gp_kthread = t;
3908 sp.sched_priority = kthread_prio;
3909 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3911 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3914 rcu_spawn_nocb_kthreads();
3915 rcu_spawn_boost_kthreads();
3918 early_initcall(rcu_spawn_gp_kthread);
3921 * This function is invoked towards the end of the scheduler's
3922 * initialization process. Before this is called, the idle task might
3923 * contain synchronous grace-period primitives (during which time, this idle
3924 * task is booting the system, and such primitives are no-ops). After this
3925 * function is called, any synchronous grace-period primitives are run as
3926 * expedited, with the requesting task driving the grace period forward.
3927 * A later core_initcall() rcu_set_runtime_mode() will switch to full
3928 * runtime RCU functionality.
3930 void rcu_scheduler_starting(void)
3932 WARN_ON(num_online_cpus() != 1);
3933 WARN_ON(nr_context_switches() > 0);
3934 rcu_test_sync_prims();
3935 rcu_scheduler_active = RCU_SCHEDULER_INIT;
3936 rcu_test_sync_prims();
3940 * Helper function for rcu_init() that initializes one rcu_state structure.
3942 static void __init rcu_init_one(struct rcu_state *rsp)
3944 static const char * const buf[] = RCU_NODE_NAME_INIT;
3945 static const char * const fqs[] = RCU_FQS_NAME_INIT;
3946 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
3947 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
3949 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
3953 struct rcu_node *rnp;
3955 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3957 /* Silence gcc 4.8 false positive about array index out of range. */
3958 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
3959 panic("rcu_init_one: rcu_num_lvls out of range");
3961 /* Initialize the level-tracking arrays. */
3963 for (i = 1; i < rcu_num_lvls; i++)
3964 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
3965 rcu_init_levelspread(levelspread, num_rcu_lvl);
3967 /* Initialize the elements themselves, starting from the leaves. */
3969 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3970 cpustride *= levelspread[i];
3971 rnp = rsp->level[i];
3972 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
3973 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
3974 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
3975 &rcu_node_class[i], buf[i]);
3976 raw_spin_lock_init(&rnp->fqslock);
3977 lockdep_set_class_and_name(&rnp->fqslock,
3978 &rcu_fqs_class[i], fqs[i]);
3979 rnp->gpnum = rsp->gpnum;
3980 rnp->completed = rsp->completed;
3981 rnp->gp_seq = rsp->gp_seq;
3982 rnp->completedqs = rsp->completed;
3984 rnp->qsmaskinit = 0;
3985 rnp->grplo = j * cpustride;
3986 rnp->grphi = (j + 1) * cpustride - 1;
3987 if (rnp->grphi >= nr_cpu_ids)
3988 rnp->grphi = nr_cpu_ids - 1;
3994 rnp->grpnum = j % levelspread[i - 1];
3995 rnp->grpmask = 1UL << rnp->grpnum;
3996 rnp->parent = rsp->level[i - 1] +
3997 j / levelspread[i - 1];
4000 INIT_LIST_HEAD(&rnp->blkd_tasks);
4001 rcu_init_one_nocb(rnp);
4002 init_waitqueue_head(&rnp->exp_wq[0]);
4003 init_waitqueue_head(&rnp->exp_wq[1]);
4004 init_waitqueue_head(&rnp->exp_wq[2]);
4005 init_waitqueue_head(&rnp->exp_wq[3]);
4006 spin_lock_init(&rnp->exp_lock);
4010 init_swait_queue_head(&rsp->gp_wq);
4011 init_swait_queue_head(&rsp->expedited_wq);
4012 rnp = rcu_first_leaf_node(rsp);
4013 for_each_possible_cpu(i) {
4014 while (i > rnp->grphi)
4016 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4017 rcu_boot_init_percpu_data(i, rsp);
4019 list_add(&rsp->flavors, &rcu_struct_flavors);
4023 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4024 * replace the definitions in tree.h because those are needed to size
4025 * the ->node array in the rcu_state structure.
4027 static void __init rcu_init_geometry(void)
4031 int rcu_capacity[RCU_NUM_LVLS];
4034 * Initialize any unspecified boot parameters.
4035 * The default values of jiffies_till_first_fqs and
4036 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4037 * value, which is a function of HZ, then adding one for each
4038 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4040 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4041 if (jiffies_till_first_fqs == ULONG_MAX)
4042 jiffies_till_first_fqs = d;
4043 if (jiffies_till_next_fqs == ULONG_MAX)
4044 jiffies_till_next_fqs = d;
4046 /* If the compile-time values are accurate, just leave. */
4047 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4048 nr_cpu_ids == NR_CPUS)
4050 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4051 rcu_fanout_leaf, nr_cpu_ids);
4054 * The boot-time rcu_fanout_leaf parameter must be at least two
4055 * and cannot exceed the number of bits in the rcu_node masks.
4056 * Complain and fall back to the compile-time values if this
4057 * limit is exceeded.
4059 if (rcu_fanout_leaf < 2 ||
4060 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4061 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4067 * Compute number of nodes that can be handled an rcu_node tree
4068 * with the given number of levels.
4070 rcu_capacity[0] = rcu_fanout_leaf;
4071 for (i = 1; i < RCU_NUM_LVLS; i++)
4072 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4075 * The tree must be able to accommodate the configured number of CPUs.
4076 * If this limit is exceeded, fall back to the compile-time values.
4078 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4079 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4084 /* Calculate the number of levels in the tree. */
4085 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4087 rcu_num_lvls = i + 1;
4089 /* Calculate the number of rcu_nodes at each level of the tree. */
4090 for (i = 0; i < rcu_num_lvls; i++) {
4091 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4092 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4095 /* Calculate the total number of rcu_node structures. */
4097 for (i = 0; i < rcu_num_lvls; i++)
4098 rcu_num_nodes += num_rcu_lvl[i];
4102 * Dump out the structure of the rcu_node combining tree associated
4103 * with the rcu_state structure referenced by rsp.
4105 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4108 struct rcu_node *rnp;
4110 pr_info("rcu_node tree layout dump\n");
4112 rcu_for_each_node_breadth_first(rsp, rnp) {
4113 if (rnp->level != level) {
4118 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4123 struct workqueue_struct *rcu_gp_wq;
4124 struct workqueue_struct *rcu_par_gp_wq;
4126 void __init rcu_init(void)
4130 rcu_early_boot_tests();
4132 rcu_bootup_announce();
4133 rcu_init_geometry();
4134 rcu_init_one(&rcu_bh_state);
4135 rcu_init_one(&rcu_sched_state);
4137 rcu_dump_rcu_node_tree(&rcu_sched_state);
4138 __rcu_init_preempt();
4139 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4142 * We don't need protection against CPU-hotplug here because
4143 * this is called early in boot, before either interrupts
4144 * or the scheduler are operational.
4146 pm_notifier(rcu_pm_notify, 0);
4147 for_each_online_cpu(cpu) {
4148 rcutree_prepare_cpu(cpu);
4149 rcu_cpu_starting(cpu);
4150 rcutree_online_cpu(cpu);
4153 /* Create workqueue for expedited GPs and for Tree SRCU. */
4154 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4155 WARN_ON(!rcu_gp_wq);
4156 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4157 WARN_ON(!rcu_par_gp_wq);
4160 #include "tree_exp.h"
4161 #include "tree_plugin.h"