Linux 4.15.6
[linux/fpc-iii.git] / kernel / rcu / tree.c
blobf9c0ca2ccf0ca530ab8bb53ad54039c62c858a87
1 /*
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 -
28 * Documentation/RCU
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>
62 #include "tree.h"
63 #include "rcu.h"
65 #ifdef MODULE_PARAM_PREFIX
66 #undef MODULE_PARAM_PREFIX
67 #endif
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.
80 #ifdef CONFIG_TRACING
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
85 #else
86 # define DEFINE_RCU_TPS(sname)
87 # define RCU_STATE_NAME(sname) __stringify(sname)
88 #endif
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, \
96 .call = cr, \
97 .gp_state = RCU_GP_IDLE, \
98 .gpnum = 0UL - 300UL, \
99 .completed = 0UL - 300UL, \
100 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
101 .name = RCU_STATE_NAME(sname), \
102 .abbr = sabbr, \
103 .exp_mutex = __MUTEX_INITIALIZER(sname##_state.exp_mutex), \
104 .exp_wake_mutex = __MUTEX_INITIALIZER(sname##_state.exp_wake_mutex), \
107 RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
108 RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
110 static struct rcu_state *const rcu_state_p;
111 LIST_HEAD(rcu_struct_flavors);
113 /* Dump rcu_node combining tree at boot to verify correct setup. */
114 static bool dump_tree;
115 module_param(dump_tree, bool, 0444);
116 /* Control rcu_node-tree auto-balancing at boot time. */
117 static bool rcu_fanout_exact;
118 module_param(rcu_fanout_exact, bool, 0444);
119 /* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120 static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
121 module_param(rcu_fanout_leaf, int, 0444);
122 int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
123 /* Number of rcu_nodes at specified level. */
124 int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
125 int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
126 /* panic() on RCU Stall sysctl. */
127 int sysctl_panic_on_rcu_stall __read_mostly;
130 * The rcu_scheduler_active variable is initialized to the value
131 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
132 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
133 * RCU can assume that there is but one task, allowing RCU to (for example)
134 * optimize synchronize_rcu() to a simple barrier(). When this variable
135 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
136 * to detect real grace periods. This variable is also used to suppress
137 * boot-time false positives from lockdep-RCU error checking. Finally, it
138 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
139 * is fully initialized, including all of its kthreads having been spawned.
141 int rcu_scheduler_active __read_mostly;
142 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
145 * The rcu_scheduler_fully_active variable transitions from zero to one
146 * during the early_initcall() processing, which is after the scheduler
147 * is capable of creating new tasks. So RCU processing (for example,
148 * creating tasks for RCU priority boosting) must be delayed until after
149 * rcu_scheduler_fully_active transitions from zero to one. We also
150 * currently delay invocation of any RCU callbacks until after this point.
152 * It might later prove better for people registering RCU callbacks during
153 * early boot to take responsibility for these callbacks, but one step at
154 * a time.
156 static int rcu_scheduler_fully_active __read_mostly;
158 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
159 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
160 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
161 static void invoke_rcu_core(void);
162 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
163 static void rcu_report_exp_rdp(struct rcu_state *rsp,
164 struct rcu_data *rdp, bool wake);
165 static void sync_sched_exp_online_cleanup(int cpu);
167 /* rcuc/rcub kthread realtime priority */
168 static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
169 module_param(kthread_prio, int, 0644);
171 /* Delay in jiffies for grace-period initialization delays, debug only. */
173 static int gp_preinit_delay;
174 module_param(gp_preinit_delay, int, 0444);
175 static int gp_init_delay;
176 module_param(gp_init_delay, int, 0444);
177 static int gp_cleanup_delay;
178 module_param(gp_cleanup_delay, int, 0444);
181 * Number of grace periods between delays, normalized by the duration of
182 * the delay. The longer the delay, the more the grace periods between
183 * each delay. The reason for this normalization is that it means that,
184 * for non-zero delays, the overall slowdown of grace periods is constant
185 * regardless of the duration of the delay. This arrangement balances
186 * the need for long delays to increase some race probabilities with the
187 * need for fast grace periods to increase other race probabilities.
189 #define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
192 * Track the rcutorture test sequence number and the update version
193 * number within a given test. The rcutorture_testseq is incremented
194 * on every rcutorture module load and unload, so has an odd value
195 * when a test is running. The rcutorture_vernum is set to zero
196 * when rcutorture starts and is incremented on each rcutorture update.
197 * These variables enable correlating rcutorture output with the
198 * RCU tracing information.
200 unsigned long rcutorture_testseq;
201 unsigned long rcutorture_vernum;
204 * Compute the mask of online CPUs for the specified rcu_node structure.
205 * This will not be stable unless the rcu_node structure's ->lock is
206 * held, but the bit corresponding to the current CPU will be stable
207 * in most contexts.
209 unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
211 return READ_ONCE(rnp->qsmaskinitnext);
215 * Return true if an RCU grace period is in progress. The READ_ONCE()s
216 * permit this function to be invoked without holding the root rcu_node
217 * structure's ->lock, but of course results can be subject to change.
219 static int rcu_gp_in_progress(struct rcu_state *rsp)
221 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
225 * Note a quiescent state. Because we do not need to know
226 * how many quiescent states passed, just if there was at least
227 * one since the start of the grace period, this just sets a flag.
228 * The caller must have disabled preemption.
230 void rcu_sched_qs(void)
232 RCU_LOCKDEP_WARN(preemptible(), "rcu_sched_qs() invoked with preemption enabled!!!");
233 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
234 return;
235 trace_rcu_grace_period(TPS("rcu_sched"),
236 __this_cpu_read(rcu_sched_data.gpnum),
237 TPS("cpuqs"));
238 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
239 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
240 return;
241 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
242 rcu_report_exp_rdp(&rcu_sched_state,
243 this_cpu_ptr(&rcu_sched_data), true);
246 void rcu_bh_qs(void)
248 RCU_LOCKDEP_WARN(preemptible(), "rcu_bh_qs() invoked with preemption enabled!!!");
249 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
250 trace_rcu_grace_period(TPS("rcu_bh"),
251 __this_cpu_read(rcu_bh_data.gpnum),
252 TPS("cpuqs"));
253 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
258 * Steal a bit from the bottom of ->dynticks for idle entry/exit
259 * control. Initially this is for TLB flushing.
261 #define RCU_DYNTICK_CTRL_MASK 0x1
262 #define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
263 #ifndef rcu_eqs_special_exit
264 #define rcu_eqs_special_exit() do { } while (0)
265 #endif
267 static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
268 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
269 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
273 * There's a few places, currently just in the tracing infrastructure,
274 * that uses rcu_irq_enter() to make sure RCU is watching. But there's
275 * a small location where that will not even work. In those cases
276 * rcu_irq_enter_disabled() needs to be checked to make sure rcu_irq_enter()
277 * can be called.
279 static DEFINE_PER_CPU(bool, disable_rcu_irq_enter);
281 bool rcu_irq_enter_disabled(void)
283 return this_cpu_read(disable_rcu_irq_enter);
287 * Record entry into an extended quiescent state. This is only to be
288 * called when not already in an extended quiescent state.
290 static void rcu_dynticks_eqs_enter(void)
292 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
293 int seq;
296 * CPUs seeing atomic_add_return() must see prior RCU read-side
297 * critical sections, and we also must force ordering with the
298 * next idle sojourn.
300 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
301 /* Better be in an extended quiescent state! */
302 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
303 (seq & RCU_DYNTICK_CTRL_CTR));
304 /* Better not have special action (TLB flush) pending! */
305 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
306 (seq & RCU_DYNTICK_CTRL_MASK));
310 * Record exit from an extended quiescent state. This is only to be
311 * called from an extended quiescent state.
313 static void rcu_dynticks_eqs_exit(void)
315 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
316 int seq;
319 * CPUs seeing atomic_add_return() must see prior idle sojourns,
320 * and we also must force ordering with the next RCU read-side
321 * critical section.
323 seq = atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
324 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
325 !(seq & RCU_DYNTICK_CTRL_CTR));
326 if (seq & RCU_DYNTICK_CTRL_MASK) {
327 atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdtp->dynticks);
328 smp_mb__after_atomic(); /* _exit after clearing mask. */
329 /* Prefer duplicate flushes to losing a flush. */
330 rcu_eqs_special_exit();
335 * Reset the current CPU's ->dynticks counter to indicate that the
336 * newly onlined CPU is no longer in an extended quiescent state.
337 * This will either leave the counter unchanged, or increment it
338 * to the next non-quiescent value.
340 * The non-atomic test/increment sequence works because the upper bits
341 * of the ->dynticks counter are manipulated only by the corresponding CPU,
342 * or when the corresponding CPU is offline.
344 static void rcu_dynticks_eqs_online(void)
346 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
348 if (atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR)
349 return;
350 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdtp->dynticks);
354 * Is the current CPU in an extended quiescent state?
356 * No ordering, as we are sampling CPU-local information.
358 bool rcu_dynticks_curr_cpu_in_eqs(void)
360 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
362 return !(atomic_read(&rdtp->dynticks) & RCU_DYNTICK_CTRL_CTR);
366 * Snapshot the ->dynticks counter with full ordering so as to allow
367 * stable comparison of this counter with past and future snapshots.
369 int rcu_dynticks_snap(struct rcu_dynticks *rdtp)
371 int snap = atomic_add_return(0, &rdtp->dynticks);
373 return snap & ~RCU_DYNTICK_CTRL_MASK;
377 * Return true if the snapshot returned from rcu_dynticks_snap()
378 * indicates that RCU is in an extended quiescent state.
380 static bool rcu_dynticks_in_eqs(int snap)
382 return !(snap & RCU_DYNTICK_CTRL_CTR);
386 * Return true if the CPU corresponding to the specified rcu_dynticks
387 * structure has spent some time in an extended quiescent state since
388 * rcu_dynticks_snap() returned the specified snapshot.
390 static bool rcu_dynticks_in_eqs_since(struct rcu_dynticks *rdtp, int snap)
392 return snap != rcu_dynticks_snap(rdtp);
396 * Do a double-increment of the ->dynticks counter to emulate a
397 * momentary idle-CPU quiescent state.
399 static void rcu_dynticks_momentary_idle(void)
401 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
402 int special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
403 &rdtp->dynticks);
405 /* It is illegal to call this from idle state. */
406 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
410 * Set the special (bottom) bit of the specified CPU so that it
411 * will take special action (such as flushing its TLB) on the
412 * next exit from an extended quiescent state. Returns true if
413 * the bit was successfully set, or false if the CPU was not in
414 * an extended quiescent state.
416 bool rcu_eqs_special_set(int cpu)
418 int old;
419 int new;
420 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
422 do {
423 old = atomic_read(&rdtp->dynticks);
424 if (old & RCU_DYNTICK_CTRL_CTR)
425 return false;
426 new = old | RCU_DYNTICK_CTRL_MASK;
427 } while (atomic_cmpxchg(&rdtp->dynticks, old, new) != old);
428 return true;
432 * Let the RCU core know that this CPU has gone through the scheduler,
433 * which is a quiescent state. This is called when the need for a
434 * quiescent state is urgent, so we burn an atomic operation and full
435 * memory barriers to let the RCU core know about it, regardless of what
436 * this CPU might (or might not) do in the near future.
438 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
440 * The caller must have disabled interrupts.
442 static void rcu_momentary_dyntick_idle(void)
444 raw_cpu_write(rcu_dynticks.rcu_need_heavy_qs, false);
445 rcu_dynticks_momentary_idle();
449 * Note a context switch. This is a quiescent state for RCU-sched,
450 * and requires special handling for preemptible RCU.
451 * The caller must have disabled interrupts.
453 void rcu_note_context_switch(bool preempt)
455 barrier(); /* Avoid RCU read-side critical sections leaking down. */
456 trace_rcu_utilization(TPS("Start context switch"));
457 rcu_sched_qs();
458 rcu_preempt_note_context_switch(preempt);
459 /* Load rcu_urgent_qs before other flags. */
460 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs)))
461 goto out;
462 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
463 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs)))
464 rcu_momentary_dyntick_idle();
465 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
466 if (!preempt)
467 rcu_note_voluntary_context_switch_lite(current);
468 out:
469 trace_rcu_utilization(TPS("End context switch"));
470 barrier(); /* Avoid RCU read-side critical sections leaking up. */
472 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
475 * Register a quiescent state for all RCU flavors. If there is an
476 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
477 * dyntick-idle quiescent state visible to other CPUs (but only for those
478 * RCU flavors in desperate need of a quiescent state, which will normally
479 * be none of them). Either way, do a lightweight quiescent state for
480 * all RCU flavors.
482 * The barrier() calls are redundant in the common case when this is
483 * called externally, but just in case this is called from within this
484 * file.
487 void rcu_all_qs(void)
489 unsigned long flags;
491 if (!raw_cpu_read(rcu_dynticks.rcu_urgent_qs))
492 return;
493 preempt_disable();
494 /* Load rcu_urgent_qs before other flags. */
495 if (!smp_load_acquire(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs))) {
496 preempt_enable();
497 return;
499 this_cpu_write(rcu_dynticks.rcu_urgent_qs, false);
500 barrier(); /* Avoid RCU read-side critical sections leaking down. */
501 if (unlikely(raw_cpu_read(rcu_dynticks.rcu_need_heavy_qs))) {
502 local_irq_save(flags);
503 rcu_momentary_dyntick_idle();
504 local_irq_restore(flags);
506 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)))
507 rcu_sched_qs();
508 this_cpu_inc(rcu_dynticks.rcu_qs_ctr);
509 barrier(); /* Avoid RCU read-side critical sections leaking up. */
510 preempt_enable();
512 EXPORT_SYMBOL_GPL(rcu_all_qs);
514 #define DEFAULT_RCU_BLIMIT 10 /* Maximum callbacks per rcu_do_batch. */
515 static long blimit = DEFAULT_RCU_BLIMIT;
516 #define DEFAULT_RCU_QHIMARK 10000 /* If this many pending, ignore blimit. */
517 static long qhimark = DEFAULT_RCU_QHIMARK;
518 #define DEFAULT_RCU_QLOMARK 100 /* Once only this many pending, use blimit. */
519 static long qlowmark = DEFAULT_RCU_QLOMARK;
521 module_param(blimit, long, 0444);
522 module_param(qhimark, long, 0444);
523 module_param(qlowmark, long, 0444);
525 static ulong jiffies_till_first_fqs = ULONG_MAX;
526 static ulong jiffies_till_next_fqs = ULONG_MAX;
527 static bool rcu_kick_kthreads;
529 module_param(jiffies_till_first_fqs, ulong, 0644);
530 module_param(jiffies_till_next_fqs, ulong, 0644);
531 module_param(rcu_kick_kthreads, bool, 0644);
534 * How long the grace period must be before we start recruiting
535 * quiescent-state help from rcu_note_context_switch().
537 static ulong jiffies_till_sched_qs = HZ / 10;
538 module_param(jiffies_till_sched_qs, ulong, 0444);
540 static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
541 struct rcu_data *rdp);
542 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp));
543 static void force_quiescent_state(struct rcu_state *rsp);
544 static int rcu_pending(void);
547 * Return the number of RCU batches started thus far for debug & stats.
549 unsigned long rcu_batches_started(void)
551 return rcu_state_p->gpnum;
553 EXPORT_SYMBOL_GPL(rcu_batches_started);
556 * Return the number of RCU-sched batches started thus far for debug & stats.
558 unsigned long rcu_batches_started_sched(void)
560 return rcu_sched_state.gpnum;
562 EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
565 * Return the number of RCU BH batches started thus far for debug & stats.
567 unsigned long rcu_batches_started_bh(void)
569 return rcu_bh_state.gpnum;
571 EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
574 * Return the number of RCU batches completed thus far for debug & stats.
576 unsigned long rcu_batches_completed(void)
578 return rcu_state_p->completed;
580 EXPORT_SYMBOL_GPL(rcu_batches_completed);
583 * Return the number of RCU-sched batches completed thus far for debug & stats.
585 unsigned long rcu_batches_completed_sched(void)
587 return rcu_sched_state.completed;
589 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
592 * Return the number of RCU BH batches completed thus far for debug & stats.
594 unsigned long rcu_batches_completed_bh(void)
596 return rcu_bh_state.completed;
598 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
601 * Return the number of RCU expedited batches completed thus far for
602 * debug & stats. Odd numbers mean that a batch is in progress, even
603 * numbers mean idle. The value returned will thus be roughly double
604 * the cumulative batches since boot.
606 unsigned long rcu_exp_batches_completed(void)
608 return rcu_state_p->expedited_sequence;
610 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
613 * Return the number of RCU-sched expedited batches completed thus far
614 * for debug & stats. Similar to rcu_exp_batches_completed().
616 unsigned long rcu_exp_batches_completed_sched(void)
618 return rcu_sched_state.expedited_sequence;
620 EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
623 * Force a quiescent state.
625 void rcu_force_quiescent_state(void)
627 force_quiescent_state(rcu_state_p);
629 EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
632 * Force a quiescent state for RCU BH.
634 void rcu_bh_force_quiescent_state(void)
636 force_quiescent_state(&rcu_bh_state);
638 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
641 * Force a quiescent state for RCU-sched.
643 void rcu_sched_force_quiescent_state(void)
645 force_quiescent_state(&rcu_sched_state);
647 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
650 * Show the state of the grace-period kthreads.
652 void show_rcu_gp_kthreads(void)
654 struct rcu_state *rsp;
656 for_each_rcu_flavor(rsp) {
657 pr_info("%s: wait state: %d ->state: %#lx\n",
658 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
659 /* sched_show_task(rsp->gp_kthread); */
662 EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
665 * Record the number of times rcutorture tests have been initiated and
666 * terminated. This information allows the debugfs tracing stats to be
667 * correlated to the rcutorture messages, even when the rcutorture module
668 * is being repeatedly loaded and unloaded. In other words, we cannot
669 * store this state in rcutorture itself.
671 void rcutorture_record_test_transition(void)
673 rcutorture_testseq++;
674 rcutorture_vernum = 0;
676 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
679 * Send along grace-period-related data for rcutorture diagnostics.
681 void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
682 unsigned long *gpnum, unsigned long *completed)
684 struct rcu_state *rsp = NULL;
686 switch (test_type) {
687 case RCU_FLAVOR:
688 rsp = rcu_state_p;
689 break;
690 case RCU_BH_FLAVOR:
691 rsp = &rcu_bh_state;
692 break;
693 case RCU_SCHED_FLAVOR:
694 rsp = &rcu_sched_state;
695 break;
696 default:
697 break;
699 if (rsp == NULL)
700 return;
701 *flags = READ_ONCE(rsp->gp_flags);
702 *gpnum = READ_ONCE(rsp->gpnum);
703 *completed = READ_ONCE(rsp->completed);
705 EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
708 * Record the number of writer passes through the current rcutorture test.
709 * This is also used to correlate debugfs tracing stats with the rcutorture
710 * messages.
712 void rcutorture_record_progress(unsigned long vernum)
714 rcutorture_vernum++;
716 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
719 * Return the root node of the specified rcu_state structure.
721 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
723 return &rsp->node[0];
727 * Is there any need for future grace periods?
728 * Interrupts must be disabled. If the caller does not hold the root
729 * rnp_node structure's ->lock, the results are advisory only.
731 static int rcu_future_needs_gp(struct rcu_state *rsp)
733 struct rcu_node *rnp = rcu_get_root(rsp);
734 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
735 int *fp = &rnp->need_future_gp[idx];
737 lockdep_assert_irqs_disabled();
738 return READ_ONCE(*fp);
742 * Does the current CPU require a not-yet-started grace period?
743 * The caller must have disabled interrupts to prevent races with
744 * normal callback registry.
746 static bool
747 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
749 lockdep_assert_irqs_disabled();
750 if (rcu_gp_in_progress(rsp))
751 return false; /* No, a grace period is already in progress. */
752 if (rcu_future_needs_gp(rsp))
753 return true; /* Yes, a no-CBs CPU needs one. */
754 if (!rcu_segcblist_is_enabled(&rdp->cblist))
755 return false; /* No, this is a no-CBs (or offline) CPU. */
756 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
757 return true; /* Yes, CPU has newly registered callbacks. */
758 if (rcu_segcblist_future_gp_needed(&rdp->cblist,
759 READ_ONCE(rsp->completed)))
760 return true; /* Yes, CBs for future grace period. */
761 return false; /* No grace period needed. */
765 * rcu_eqs_enter_common - current CPU is entering an extended quiescent state
767 * Enter idle, doing appropriate accounting. The caller must have
768 * disabled interrupts.
770 static void rcu_eqs_enter_common(bool user)
772 struct rcu_state *rsp;
773 struct rcu_data *rdp;
774 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
776 lockdep_assert_irqs_disabled();
777 trace_rcu_dyntick(TPS("Start"), rdtp->dynticks_nesting, 0);
778 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
779 !user && !is_idle_task(current)) {
780 struct task_struct *idle __maybe_unused =
781 idle_task(smp_processor_id());
783 trace_rcu_dyntick(TPS("Error on entry: not idle task"), rdtp->dynticks_nesting, 0);
784 rcu_ftrace_dump(DUMP_ORIG);
785 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
786 current->pid, current->comm,
787 idle->pid, idle->comm); /* must be idle task! */
789 for_each_rcu_flavor(rsp) {
790 rdp = this_cpu_ptr(rsp->rda);
791 do_nocb_deferred_wakeup(rdp);
793 rcu_prepare_for_idle();
794 __this_cpu_inc(disable_rcu_irq_enter);
795 rdtp->dynticks_nesting = 0; /* Breaks tracing momentarily. */
796 rcu_dynticks_eqs_enter(); /* After this, tracing works again. */
797 __this_cpu_dec(disable_rcu_irq_enter);
798 rcu_dynticks_task_enter();
801 * It is illegal to enter an extended quiescent state while
802 * in an RCU read-side critical section.
804 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
805 "Illegal idle entry in RCU read-side critical section.");
806 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
807 "Illegal idle entry in RCU-bh read-side critical section.");
808 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
809 "Illegal idle entry in RCU-sched read-side critical section.");
813 * Enter an RCU extended quiescent state, which can be either the
814 * idle loop or adaptive-tickless usermode execution.
816 static void rcu_eqs_enter(bool user)
818 struct rcu_dynticks *rdtp;
820 rdtp = this_cpu_ptr(&rcu_dynticks);
821 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
822 (rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == 0);
823 if ((rdtp->dynticks_nesting & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
824 rcu_eqs_enter_common(user);
825 else
826 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
830 * rcu_idle_enter - inform RCU that current CPU is entering idle
832 * Enter idle mode, in other words, -leave- the mode in which RCU
833 * read-side critical sections can occur. (Though RCU read-side
834 * critical sections can occur in irq handlers in idle, a possibility
835 * handled by irq_enter() and irq_exit().)
837 * We crowbar the ->dynticks_nesting field to zero to allow for
838 * the possibility of usermode upcalls having messed up our count
839 * of interrupt nesting level during the prior busy period.
841 * If you add or remove a call to rcu_idle_enter(), be sure to test with
842 * CONFIG_RCU_EQS_DEBUG=y.
844 void rcu_idle_enter(void)
846 lockdep_assert_irqs_disabled();
847 rcu_eqs_enter(false);
850 #ifdef CONFIG_NO_HZ_FULL
852 * rcu_user_enter - inform RCU that we are resuming userspace.
854 * Enter RCU idle mode right before resuming userspace. No use of RCU
855 * is permitted between this call and rcu_user_exit(). This way the
856 * CPU doesn't need to maintain the tick for RCU maintenance purposes
857 * when the CPU runs in userspace.
859 * If you add or remove a call to rcu_user_enter(), be sure to test with
860 * CONFIG_RCU_EQS_DEBUG=y.
862 void rcu_user_enter(void)
864 lockdep_assert_irqs_disabled();
865 rcu_eqs_enter(true);
867 #endif /* CONFIG_NO_HZ_FULL */
870 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
872 * Exit from an interrupt handler, which might possibly result in entering
873 * idle mode, in other words, leaving the mode in which read-side critical
874 * sections can occur. The caller must have disabled interrupts.
876 * This code assumes that the idle loop never does anything that might
877 * result in unbalanced calls to irq_enter() and irq_exit(). If your
878 * architecture violates this assumption, RCU will give you what you
879 * deserve, good and hard. But very infrequently and irreproducibly.
881 * Use things like work queues to work around this limitation.
883 * You have been warned.
885 * If you add or remove a call to rcu_irq_exit(), be sure to test with
886 * CONFIG_RCU_EQS_DEBUG=y.
888 void rcu_irq_exit(void)
890 struct rcu_dynticks *rdtp;
892 lockdep_assert_irqs_disabled();
893 rdtp = this_cpu_ptr(&rcu_dynticks);
895 /* Page faults can happen in NMI handlers, so check... */
896 if (rdtp->dynticks_nmi_nesting)
897 return;
899 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
900 rdtp->dynticks_nesting < 1);
901 if (rdtp->dynticks_nesting <= 1) {
902 rcu_eqs_enter_common(true);
903 } else {
904 trace_rcu_dyntick(TPS("--="), rdtp->dynticks_nesting, rdtp->dynticks_nesting - 1);
905 rdtp->dynticks_nesting--;
910 * Wrapper for rcu_irq_exit() where interrupts are enabled.
912 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
913 * with CONFIG_RCU_EQS_DEBUG=y.
915 void rcu_irq_exit_irqson(void)
917 unsigned long flags;
919 local_irq_save(flags);
920 rcu_irq_exit();
921 local_irq_restore(flags);
925 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
927 * If the new value of the ->dynticks_nesting counter was previously zero,
928 * we really have exited idle, and must do the appropriate accounting.
929 * The caller must have disabled interrupts.
931 static void rcu_eqs_exit_common(long long oldval, int user)
933 RCU_TRACE(struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);)
935 rcu_dynticks_task_exit();
936 rcu_dynticks_eqs_exit();
937 rcu_cleanup_after_idle();
938 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
939 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
940 !user && !is_idle_task(current)) {
941 struct task_struct *idle __maybe_unused =
942 idle_task(smp_processor_id());
944 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
945 oldval, rdtp->dynticks_nesting);
946 rcu_ftrace_dump(DUMP_ORIG);
947 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
948 current->pid, current->comm,
949 idle->pid, idle->comm); /* must be idle task! */
954 * Exit an RCU extended quiescent state, which can be either the
955 * idle loop or adaptive-tickless usermode execution.
957 static void rcu_eqs_exit(bool user)
959 struct rcu_dynticks *rdtp;
960 long long oldval;
962 lockdep_assert_irqs_disabled();
963 rdtp = this_cpu_ptr(&rcu_dynticks);
964 oldval = rdtp->dynticks_nesting;
965 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
966 if (oldval & DYNTICK_TASK_NEST_MASK) {
967 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
968 } else {
969 __this_cpu_inc(disable_rcu_irq_enter);
970 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
971 rcu_eqs_exit_common(oldval, user);
972 __this_cpu_dec(disable_rcu_irq_enter);
977 * rcu_idle_exit - inform RCU that current CPU is leaving idle
979 * Exit idle mode, in other words, -enter- the mode in which RCU
980 * read-side critical sections can occur.
982 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
983 * allow for the possibility of usermode upcalls messing up our count
984 * of interrupt nesting level during the busy period that is just
985 * now starting.
987 * If you add or remove a call to rcu_idle_exit(), be sure to test with
988 * CONFIG_RCU_EQS_DEBUG=y.
990 void rcu_idle_exit(void)
992 unsigned long flags;
994 local_irq_save(flags);
995 rcu_eqs_exit(false);
996 local_irq_restore(flags);
999 #ifdef CONFIG_NO_HZ_FULL
1001 * rcu_user_exit - inform RCU that we are exiting userspace.
1003 * Exit RCU idle mode while entering the kernel because it can
1004 * run a RCU read side critical section anytime.
1006 * If you add or remove a call to rcu_user_exit(), be sure to test with
1007 * CONFIG_RCU_EQS_DEBUG=y.
1009 void rcu_user_exit(void)
1011 rcu_eqs_exit(1);
1013 #endif /* CONFIG_NO_HZ_FULL */
1016 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1018 * Enter an interrupt handler, which might possibly result in exiting
1019 * idle mode, in other words, entering the mode in which read-side critical
1020 * sections can occur. The caller must have disabled interrupts.
1022 * Note that the Linux kernel is fully capable of entering an interrupt
1023 * handler that it never exits, for example when doing upcalls to
1024 * user mode! This code assumes that the idle loop never does upcalls to
1025 * user mode. If your architecture does do upcalls from the idle loop (or
1026 * does anything else that results in unbalanced calls to the irq_enter()
1027 * and irq_exit() functions), RCU will give you what you deserve, good
1028 * and hard. But very infrequently and irreproducibly.
1030 * Use things like work queues to work around this limitation.
1032 * You have been warned.
1034 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1035 * CONFIG_RCU_EQS_DEBUG=y.
1037 void rcu_irq_enter(void)
1039 struct rcu_dynticks *rdtp;
1040 long long oldval;
1042 lockdep_assert_irqs_disabled();
1043 rdtp = this_cpu_ptr(&rcu_dynticks);
1045 /* Page faults can happen in NMI handlers, so check... */
1046 if (rdtp->dynticks_nmi_nesting)
1047 return;
1049 oldval = rdtp->dynticks_nesting;
1050 rdtp->dynticks_nesting++;
1051 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
1052 rdtp->dynticks_nesting == 0);
1053 if (oldval)
1054 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
1055 else
1056 rcu_eqs_exit_common(oldval, true);
1060 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1062 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1063 * with CONFIG_RCU_EQS_DEBUG=y.
1065 void rcu_irq_enter_irqson(void)
1067 unsigned long flags;
1069 local_irq_save(flags);
1070 rcu_irq_enter();
1071 local_irq_restore(flags);
1075 * rcu_nmi_enter - inform RCU of entry to NMI context
1077 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
1078 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
1079 * that the CPU is active. This implementation permits nested NMIs, as
1080 * long as the nesting level does not overflow an int. (You will probably
1081 * run out of stack space first.)
1083 * If you add or remove a call to rcu_nmi_enter(), be sure to test
1084 * with CONFIG_RCU_EQS_DEBUG=y.
1086 void rcu_nmi_enter(void)
1088 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1089 int incby = 2;
1091 /* Complain about underflow. */
1092 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
1095 * If idle from RCU viewpoint, atomically increment ->dynticks
1096 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1097 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1098 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1099 * to be in the outermost NMI handler that interrupted an RCU-idle
1100 * period (observation due to Andy Lutomirski).
1102 if (rcu_dynticks_curr_cpu_in_eqs()) {
1103 rcu_dynticks_eqs_exit();
1104 incby = 1;
1106 rdtp->dynticks_nmi_nesting += incby;
1107 barrier();
1111 * rcu_nmi_exit - inform RCU of exit from NMI context
1113 * If we are returning from the outermost NMI handler that interrupted an
1114 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
1115 * to let the RCU grace-period handling know that the CPU is back to
1116 * being RCU-idle.
1118 * If you add or remove a call to rcu_nmi_exit(), be sure to test
1119 * with CONFIG_RCU_EQS_DEBUG=y.
1121 void rcu_nmi_exit(void)
1123 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1126 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
1127 * (We are exiting an NMI handler, so RCU better be paying attention
1128 * to us!)
1130 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
1131 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
1134 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
1135 * leave it in non-RCU-idle state.
1137 if (rdtp->dynticks_nmi_nesting != 1) {
1138 rdtp->dynticks_nmi_nesting -= 2;
1139 return;
1142 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
1143 rdtp->dynticks_nmi_nesting = 0;
1144 rcu_dynticks_eqs_enter();
1148 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1150 * Return true if RCU is watching the running CPU, which means that this
1151 * CPU can safely enter RCU read-side critical sections. In other words,
1152 * if the current CPU is in its idle loop and is neither in an interrupt
1153 * or NMI handler, return true.
1155 bool notrace rcu_is_watching(void)
1157 bool ret;
1159 preempt_disable_notrace();
1160 ret = !rcu_dynticks_curr_cpu_in_eqs();
1161 preempt_enable_notrace();
1162 return ret;
1164 EXPORT_SYMBOL_GPL(rcu_is_watching);
1167 * If a holdout task is actually running, request an urgent quiescent
1168 * state from its CPU. This is unsynchronized, so migrations can cause
1169 * the request to go to the wrong CPU. Which is OK, all that will happen
1170 * is that the CPU's next context switch will be a bit slower and next
1171 * time around this task will generate another request.
1173 void rcu_request_urgent_qs_task(struct task_struct *t)
1175 int cpu;
1177 barrier();
1178 cpu = task_cpu(t);
1179 if (!task_curr(t))
1180 return; /* This task is not running on that CPU. */
1181 smp_store_release(per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, cpu), true);
1184 #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1187 * Is the current CPU online? Disable preemption to avoid false positives
1188 * that could otherwise happen due to the current CPU number being sampled,
1189 * this task being preempted, its old CPU being taken offline, resuming
1190 * on some other CPU, then determining that its old CPU is now offline.
1191 * It is OK to use RCU on an offline processor during initial boot, hence
1192 * the check for rcu_scheduler_fully_active. Note also that it is OK
1193 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1194 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1195 * offline to continue to use RCU for one jiffy after marking itself
1196 * offline in the cpu_online_mask. This leniency is necessary given the
1197 * non-atomic nature of the online and offline processing, for example,
1198 * the fact that a CPU enters the scheduler after completing the teardown
1199 * of the CPU.
1201 * This is also why RCU internally marks CPUs online during in the
1202 * preparation phase and offline after the CPU has been taken down.
1204 * Disable checking if in an NMI handler because we cannot safely report
1205 * errors from NMI handlers anyway.
1207 bool rcu_lockdep_current_cpu_online(void)
1209 struct rcu_data *rdp;
1210 struct rcu_node *rnp;
1211 bool ret;
1213 if (in_nmi())
1214 return true;
1215 preempt_disable();
1216 rdp = this_cpu_ptr(&rcu_sched_data);
1217 rnp = rdp->mynode;
1218 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1219 !rcu_scheduler_fully_active;
1220 preempt_enable();
1221 return ret;
1223 EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1225 #endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1228 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1230 * If the current CPU is idle or running at a first-level (not nested)
1231 * interrupt from idle, return true. The caller must have at least
1232 * disabled preemption.
1234 static int rcu_is_cpu_rrupt_from_idle(void)
1236 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1240 * We are reporting a quiescent state on behalf of some other CPU, so
1241 * it is our responsibility to check for and handle potential overflow
1242 * of the rcu_node ->gpnum counter with respect to the rcu_data counters.
1243 * After all, the CPU might be in deep idle state, and thus executing no
1244 * code whatsoever.
1246 static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1248 lockdep_assert_held(&rnp->lock);
1249 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4, rnp->gpnum))
1250 WRITE_ONCE(rdp->gpwrap, true);
1251 if (ULONG_CMP_LT(rdp->rcu_iw_gpnum + ULONG_MAX / 4, rnp->gpnum))
1252 rdp->rcu_iw_gpnum = rnp->gpnum + ULONG_MAX / 4;
1256 * Snapshot the specified CPU's dynticks counter so that we can later
1257 * credit them with an implicit quiescent state. Return 1 if this CPU
1258 * is in dynticks idle mode, which is an extended quiescent state.
1260 static int dyntick_save_progress_counter(struct rcu_data *rdp)
1262 rdp->dynticks_snap = rcu_dynticks_snap(rdp->dynticks);
1263 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1264 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1265 rcu_gpnum_ovf(rdp->mynode, rdp);
1266 return 1;
1268 return 0;
1272 * Handler for the irq_work request posted when a grace period has
1273 * gone on for too long, but not yet long enough for an RCU CPU
1274 * stall warning. Set state appropriately, but just complain if
1275 * there is unexpected state on entry.
1277 static void rcu_iw_handler(struct irq_work *iwp)
1279 struct rcu_data *rdp;
1280 struct rcu_node *rnp;
1282 rdp = container_of(iwp, struct rcu_data, rcu_iw);
1283 rnp = rdp->mynode;
1284 raw_spin_lock_rcu_node(rnp);
1285 if (!WARN_ON_ONCE(!rdp->rcu_iw_pending)) {
1286 rdp->rcu_iw_gpnum = rnp->gpnum;
1287 rdp->rcu_iw_pending = false;
1289 raw_spin_unlock_rcu_node(rnp);
1293 * Return true if the specified CPU has passed through a quiescent
1294 * state by virtue of being in or having passed through an dynticks
1295 * idle state since the last call to dyntick_save_progress_counter()
1296 * for this same CPU, or by virtue of having been offline.
1298 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1300 unsigned long jtsq;
1301 bool *rnhqp;
1302 bool *ruqp;
1303 struct rcu_node *rnp = rdp->mynode;
1306 * If the CPU passed through or entered a dynticks idle phase with
1307 * no active irq/NMI handlers, then we can safely pretend that the CPU
1308 * already acknowledged the request to pass through a quiescent
1309 * state. Either way, that CPU cannot possibly be in an RCU
1310 * read-side critical section that started before the beginning
1311 * of the current RCU grace period.
1313 if (rcu_dynticks_in_eqs_since(rdp->dynticks, rdp->dynticks_snap)) {
1314 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1315 rdp->dynticks_fqs++;
1316 rcu_gpnum_ovf(rnp, rdp);
1317 return 1;
1321 * Has this CPU encountered a cond_resched_rcu_qs() since the
1322 * beginning of the grace period? For this to be the case,
1323 * the CPU has to have noticed the current grace period. This
1324 * might not be the case for nohz_full CPUs looping in the kernel.
1326 jtsq = jiffies_till_sched_qs;
1327 ruqp = per_cpu_ptr(&rcu_dynticks.rcu_urgent_qs, rdp->cpu);
1328 if (time_after(jiffies, rdp->rsp->gp_start + jtsq) &&
1329 READ_ONCE(rdp->rcu_qs_ctr_snap) != per_cpu(rcu_dynticks.rcu_qs_ctr, rdp->cpu) &&
1330 READ_ONCE(rdp->gpnum) == rnp->gpnum && !rdp->gpwrap) {
1331 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("rqc"));
1332 rcu_gpnum_ovf(rnp, rdp);
1333 return 1;
1334 } else if (time_after(jiffies, rdp->rsp->gp_start + jtsq)) {
1335 /* Load rcu_qs_ctr before store to rcu_urgent_qs. */
1336 smp_store_release(ruqp, true);
1339 /* Check for the CPU being offline. */
1340 if (!(rdp->grpmask & rcu_rnp_online_cpus(rnp))) {
1341 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1342 rdp->offline_fqs++;
1343 rcu_gpnum_ovf(rnp, rdp);
1344 return 1;
1348 * A CPU running for an extended time within the kernel can
1349 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1350 * even context-switching back and forth between a pair of
1351 * in-kernel CPU-bound tasks cannot advance grace periods.
1352 * So if the grace period is old enough, make the CPU pay attention.
1353 * Note that the unsynchronized assignments to the per-CPU
1354 * rcu_need_heavy_qs variable are safe. Yes, setting of
1355 * bits can be lost, but they will be set again on the next
1356 * force-quiescent-state pass. So lost bit sets do not result
1357 * in incorrect behavior, merely in a grace period lasting
1358 * a few jiffies longer than it might otherwise. Because
1359 * there are at most four threads involved, and because the
1360 * updates are only once every few jiffies, the probability of
1361 * lossage (and thus of slight grace-period extension) is
1362 * quite low.
1364 rnhqp = &per_cpu(rcu_dynticks.rcu_need_heavy_qs, rdp->cpu);
1365 if (!READ_ONCE(*rnhqp) &&
1366 (time_after(jiffies, rdp->rsp->gp_start + jtsq) ||
1367 time_after(jiffies, rdp->rsp->jiffies_resched))) {
1368 WRITE_ONCE(*rnhqp, true);
1369 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1370 smp_store_release(ruqp, true);
1371 rdp->rsp->jiffies_resched += jtsq; /* Re-enable beating. */
1375 * If more than halfway to RCU CPU stall-warning time, do a
1376 * resched_cpu() to try to loosen things up a bit. Also check to
1377 * see if the CPU is getting hammered with interrupts, but only
1378 * once per grace period, just to keep the IPIs down to a dull roar.
1380 if (jiffies - rdp->rsp->gp_start > rcu_jiffies_till_stall_check() / 2) {
1381 resched_cpu(rdp->cpu);
1382 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1383 !rdp->rcu_iw_pending && rdp->rcu_iw_gpnum != rnp->gpnum &&
1384 (rnp->ffmask & rdp->grpmask)) {
1385 init_irq_work(&rdp->rcu_iw, rcu_iw_handler);
1386 rdp->rcu_iw_pending = true;
1387 rdp->rcu_iw_gpnum = rnp->gpnum;
1388 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1392 return 0;
1395 static void record_gp_stall_check_time(struct rcu_state *rsp)
1397 unsigned long j = jiffies;
1398 unsigned long j1;
1400 rsp->gp_start = j;
1401 smp_wmb(); /* Record start time before stall time. */
1402 j1 = rcu_jiffies_till_stall_check();
1403 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1404 rsp->jiffies_resched = j + j1 / 2;
1405 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1409 * Convert a ->gp_state value to a character string.
1411 static const char *gp_state_getname(short gs)
1413 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1414 return "???";
1415 return gp_state_names[gs];
1419 * Complain about starvation of grace-period kthread.
1421 static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1423 unsigned long gpa;
1424 unsigned long j;
1426 j = jiffies;
1427 gpa = READ_ONCE(rsp->gp_activity);
1428 if (j - gpa > 2 * HZ) {
1429 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx ->cpu=%d\n",
1430 rsp->name, j - gpa,
1431 rsp->gpnum, rsp->completed,
1432 rsp->gp_flags,
1433 gp_state_getname(rsp->gp_state), rsp->gp_state,
1434 rsp->gp_kthread ? rsp->gp_kthread->state : ~0,
1435 rsp->gp_kthread ? task_cpu(rsp->gp_kthread) : -1);
1436 if (rsp->gp_kthread) {
1437 sched_show_task(rsp->gp_kthread);
1438 wake_up_process(rsp->gp_kthread);
1444 * Dump stacks of all tasks running on stalled CPUs. First try using
1445 * NMIs, but fall back to manual remote stack tracing on architectures
1446 * that don't support NMI-based stack dumps. The NMI-triggered stack
1447 * traces are more accurate because they are printed by the target CPU.
1449 static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1451 int cpu;
1452 unsigned long flags;
1453 struct rcu_node *rnp;
1455 rcu_for_each_leaf_node(rsp, rnp) {
1456 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1457 for_each_leaf_node_possible_cpu(rnp, cpu)
1458 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1459 if (!trigger_single_cpu_backtrace(cpu))
1460 dump_cpu_task(cpu);
1461 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1466 * If too much time has passed in the current grace period, and if
1467 * so configured, go kick the relevant kthreads.
1469 static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1471 unsigned long j;
1473 if (!rcu_kick_kthreads)
1474 return;
1475 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1476 if (time_after(jiffies, j) && rsp->gp_kthread &&
1477 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1478 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1479 rcu_ftrace_dump(DUMP_ALL);
1480 wake_up_process(rsp->gp_kthread);
1481 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1485 static inline void panic_on_rcu_stall(void)
1487 if (sysctl_panic_on_rcu_stall)
1488 panic("RCU Stall\n");
1491 static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1493 int cpu;
1494 long delta;
1495 unsigned long flags;
1496 unsigned long gpa;
1497 unsigned long j;
1498 int ndetected = 0;
1499 struct rcu_node *rnp = rcu_get_root(rsp);
1500 long totqlen = 0;
1502 /* Kick and suppress, if so configured. */
1503 rcu_stall_kick_kthreads(rsp);
1504 if (rcu_cpu_stall_suppress)
1505 return;
1507 /* Only let one CPU complain about others per time interval. */
1509 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1510 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1511 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1512 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1513 return;
1515 WRITE_ONCE(rsp->jiffies_stall,
1516 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1517 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1520 * OK, time to rat on our buddy...
1521 * See Documentation/RCU/stallwarn.txt for info on how to debug
1522 * RCU CPU stall warnings.
1524 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1525 rsp->name);
1526 print_cpu_stall_info_begin();
1527 rcu_for_each_leaf_node(rsp, rnp) {
1528 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1529 ndetected += rcu_print_task_stall(rnp);
1530 if (rnp->qsmask != 0) {
1531 for_each_leaf_node_possible_cpu(rnp, cpu)
1532 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1533 print_cpu_stall_info(rsp, cpu);
1534 ndetected++;
1537 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1540 print_cpu_stall_info_end();
1541 for_each_possible_cpu(cpu)
1542 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1543 cpu)->cblist);
1544 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1545 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1546 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1547 if (ndetected) {
1548 rcu_dump_cpu_stacks(rsp);
1550 /* Complain about tasks blocking the grace period. */
1551 rcu_print_detail_task_stall(rsp);
1552 } else {
1553 if (READ_ONCE(rsp->gpnum) != gpnum ||
1554 READ_ONCE(rsp->completed) == gpnum) {
1555 pr_err("INFO: Stall ended before state dump start\n");
1556 } else {
1557 j = jiffies;
1558 gpa = READ_ONCE(rsp->gp_activity);
1559 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1560 rsp->name, j - gpa, j, gpa,
1561 jiffies_till_next_fqs,
1562 rcu_get_root(rsp)->qsmask);
1563 /* In this case, the current CPU might be at fault. */
1564 sched_show_task(current);
1568 rcu_check_gp_kthread_starvation(rsp);
1570 panic_on_rcu_stall();
1572 force_quiescent_state(rsp); /* Kick them all. */
1575 static void print_cpu_stall(struct rcu_state *rsp)
1577 int cpu;
1578 unsigned long flags;
1579 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1580 struct rcu_node *rnp = rcu_get_root(rsp);
1581 long totqlen = 0;
1583 /* Kick and suppress, if so configured. */
1584 rcu_stall_kick_kthreads(rsp);
1585 if (rcu_cpu_stall_suppress)
1586 return;
1589 * OK, time to rat on ourselves...
1590 * See Documentation/RCU/stallwarn.txt for info on how to debug
1591 * RCU CPU stall warnings.
1593 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1594 print_cpu_stall_info_begin();
1595 raw_spin_lock_irqsave_rcu_node(rdp->mynode, flags);
1596 print_cpu_stall_info(rsp, smp_processor_id());
1597 raw_spin_unlock_irqrestore_rcu_node(rdp->mynode, flags);
1598 print_cpu_stall_info_end();
1599 for_each_possible_cpu(cpu)
1600 totqlen += rcu_segcblist_n_cbs(&per_cpu_ptr(rsp->rda,
1601 cpu)->cblist);
1602 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1603 jiffies - rsp->gp_start,
1604 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1606 rcu_check_gp_kthread_starvation(rsp);
1608 rcu_dump_cpu_stacks(rsp);
1610 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1611 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1612 WRITE_ONCE(rsp->jiffies_stall,
1613 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1614 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1616 panic_on_rcu_stall();
1619 * Attempt to revive the RCU machinery by forcing a context switch.
1621 * A context switch would normally allow the RCU state machine to make
1622 * progress and it could be we're stuck in kernel space without context
1623 * switches for an entirely unreasonable amount of time.
1625 resched_cpu(smp_processor_id());
1628 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1630 unsigned long completed;
1631 unsigned long gpnum;
1632 unsigned long gps;
1633 unsigned long j;
1634 unsigned long js;
1635 struct rcu_node *rnp;
1637 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1638 !rcu_gp_in_progress(rsp))
1639 return;
1640 rcu_stall_kick_kthreads(rsp);
1641 j = jiffies;
1644 * Lots of memory barriers to reject false positives.
1646 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1647 * then rsp->gp_start, and finally rsp->completed. These values
1648 * are updated in the opposite order with memory barriers (or
1649 * equivalent) during grace-period initialization and cleanup.
1650 * Now, a false positive can occur if we get an new value of
1651 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1652 * the memory barriers, the only way that this can happen is if one
1653 * grace period ends and another starts between these two fetches.
1654 * Detect this by comparing rsp->completed with the previous fetch
1655 * from rsp->gpnum.
1657 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1658 * and rsp->gp_start suffice to forestall false positives.
1660 gpnum = READ_ONCE(rsp->gpnum);
1661 smp_rmb(); /* Pick up ->gpnum first... */
1662 js = READ_ONCE(rsp->jiffies_stall);
1663 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1664 gps = READ_ONCE(rsp->gp_start);
1665 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1666 completed = READ_ONCE(rsp->completed);
1667 if (ULONG_CMP_GE(completed, gpnum) ||
1668 ULONG_CMP_LT(j, js) ||
1669 ULONG_CMP_GE(gps, js))
1670 return; /* No stall or GP completed since entering function. */
1671 rnp = rdp->mynode;
1672 if (rcu_gp_in_progress(rsp) &&
1673 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1675 /* We haven't checked in, so go dump stack. */
1676 print_cpu_stall(rsp);
1678 } else if (rcu_gp_in_progress(rsp) &&
1679 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1681 /* They had a few time units to dump stack, so complain. */
1682 print_other_cpu_stall(rsp, gpnum);
1687 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1689 * Set the stall-warning timeout way off into the future, thus preventing
1690 * any RCU CPU stall-warning messages from appearing in the current set of
1691 * RCU grace periods.
1693 * The caller must disable hard irqs.
1695 void rcu_cpu_stall_reset(void)
1697 struct rcu_state *rsp;
1699 for_each_rcu_flavor(rsp)
1700 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1704 * Determine the value that ->completed will have at the end of the
1705 * next subsequent grace period. This is used to tag callbacks so that
1706 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1707 * been dyntick-idle for an extended period with callbacks under the
1708 * influence of RCU_FAST_NO_HZ.
1710 * The caller must hold rnp->lock with interrupts disabled.
1712 static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1713 struct rcu_node *rnp)
1715 lockdep_assert_held(&rnp->lock);
1718 * If RCU is idle, we just wait for the next grace period.
1719 * But we can only be sure that RCU is idle if we are looking
1720 * at the root rcu_node structure -- otherwise, a new grace
1721 * period might have started, but just not yet gotten around
1722 * to initializing the current non-root rcu_node structure.
1724 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1725 return rnp->completed + 1;
1728 * Otherwise, wait for a possible partial grace period and
1729 * then the subsequent full grace period.
1731 return rnp->completed + 2;
1735 * Trace-event helper function for rcu_start_future_gp() and
1736 * rcu_nocb_wait_gp().
1738 static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1739 unsigned long c, const char *s)
1741 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1742 rnp->completed, c, rnp->level,
1743 rnp->grplo, rnp->grphi, s);
1747 * Start some future grace period, as needed to handle newly arrived
1748 * callbacks. The required future grace periods are recorded in each
1749 * rcu_node structure's ->need_future_gp field. Returns true if there
1750 * is reason to awaken the grace-period kthread.
1752 * The caller must hold the specified rcu_node structure's ->lock.
1754 static bool __maybe_unused
1755 rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1756 unsigned long *c_out)
1758 unsigned long c;
1759 bool ret = false;
1760 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1762 lockdep_assert_held(&rnp->lock);
1765 * Pick up grace-period number for new callbacks. If this
1766 * grace period is already marked as needed, return to the caller.
1768 c = rcu_cbs_completed(rdp->rsp, rnp);
1769 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1770 if (rnp->need_future_gp[c & 0x1]) {
1771 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1772 goto out;
1776 * If either this rcu_node structure or the root rcu_node structure
1777 * believe that a grace period is in progress, then we must wait
1778 * for the one following, which is in "c". Because our request
1779 * will be noticed at the end of the current grace period, we don't
1780 * need to explicitly start one. We only do the lockless check
1781 * of rnp_root's fields if the current rcu_node structure thinks
1782 * there is no grace period in flight, and because we hold rnp->lock,
1783 * the only possible change is when rnp_root's two fields are
1784 * equal, in which case rnp_root->gpnum might be concurrently
1785 * incremented. But that is OK, as it will just result in our
1786 * doing some extra useless work.
1788 if (rnp->gpnum != rnp->completed ||
1789 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1790 rnp->need_future_gp[c & 0x1]++;
1791 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1792 goto out;
1796 * There might be no grace period in progress. If we don't already
1797 * hold it, acquire the root rcu_node structure's lock in order to
1798 * start one (if needed).
1800 if (rnp != rnp_root)
1801 raw_spin_lock_rcu_node(rnp_root);
1804 * Get a new grace-period number. If there really is no grace
1805 * period in progress, it will be smaller than the one we obtained
1806 * earlier. Adjust callbacks as needed.
1808 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1809 if (!rcu_is_nocb_cpu(rdp->cpu))
1810 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1813 * If the needed for the required grace period is already
1814 * recorded, trace and leave.
1816 if (rnp_root->need_future_gp[c & 0x1]) {
1817 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1818 goto unlock_out;
1821 /* Record the need for the future grace period. */
1822 rnp_root->need_future_gp[c & 0x1]++;
1824 /* If a grace period is not already in progress, start one. */
1825 if (rnp_root->gpnum != rnp_root->completed) {
1826 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1827 } else {
1828 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1829 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1831 unlock_out:
1832 if (rnp != rnp_root)
1833 raw_spin_unlock_rcu_node(rnp_root);
1834 out:
1835 if (c_out != NULL)
1836 *c_out = c;
1837 return ret;
1841 * Clean up any old requests for the just-ended grace period. Also return
1842 * whether any additional grace periods have been requested.
1844 static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1846 int c = rnp->completed;
1847 int needmore;
1848 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1850 rnp->need_future_gp[c & 0x1] = 0;
1851 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1852 trace_rcu_future_gp(rnp, rdp, c,
1853 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1854 return needmore;
1858 * Awaken the grace-period kthread for the specified flavor of RCU.
1859 * Don't do a self-awaken, and don't bother awakening when there is
1860 * nothing for the grace-period kthread to do (as in several CPUs
1861 * raced to awaken, and we lost), and finally don't try to awaken
1862 * a kthread that has not yet been created.
1864 static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1866 if (current == rsp->gp_kthread ||
1867 !READ_ONCE(rsp->gp_flags) ||
1868 !rsp->gp_kthread)
1869 return;
1870 swake_up(&rsp->gp_wq);
1874 * If there is room, assign a ->completed number to any callbacks on
1875 * this CPU that have not already been assigned. Also accelerate any
1876 * callbacks that were previously assigned a ->completed number that has
1877 * since proven to be too conservative, which can happen if callbacks get
1878 * assigned a ->completed number while RCU is idle, but with reference to
1879 * a non-root rcu_node structure. This function is idempotent, so it does
1880 * not hurt to call it repeatedly. Returns an flag saying that we should
1881 * awaken the RCU grace-period kthread.
1883 * The caller must hold rnp->lock with interrupts disabled.
1885 static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1886 struct rcu_data *rdp)
1888 bool ret = false;
1890 lockdep_assert_held(&rnp->lock);
1892 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1893 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1894 return false;
1897 * Callbacks are often registered with incomplete grace-period
1898 * information. Something about the fact that getting exact
1899 * information requires acquiring a global lock... RCU therefore
1900 * makes a conservative estimate of the grace period number at which
1901 * a given callback will become ready to invoke. The following
1902 * code checks this estimate and improves it when possible, thus
1903 * accelerating callback invocation to an earlier grace-period
1904 * number.
1906 if (rcu_segcblist_accelerate(&rdp->cblist, rcu_cbs_completed(rsp, rnp)))
1907 ret = rcu_start_future_gp(rnp, rdp, NULL);
1909 /* Trace depending on how much we were able to accelerate. */
1910 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1911 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1912 else
1913 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1914 return ret;
1918 * Move any callbacks whose grace period has completed to the
1919 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1920 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1921 * sublist. This function is idempotent, so it does not hurt to
1922 * invoke it repeatedly. As long as it is not invoked -too- often...
1923 * Returns true if the RCU grace-period kthread needs to be awakened.
1925 * The caller must hold rnp->lock with interrupts disabled.
1927 static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1928 struct rcu_data *rdp)
1930 lockdep_assert_held(&rnp->lock);
1932 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1933 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1934 return false;
1937 * Find all callbacks whose ->completed numbers indicate that they
1938 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1940 rcu_segcblist_advance(&rdp->cblist, rnp->completed);
1942 /* Classify any remaining callbacks. */
1943 return rcu_accelerate_cbs(rsp, rnp, rdp);
1947 * Update CPU-local rcu_data state to record the beginnings and ends of
1948 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1949 * structure corresponding to the current CPU, and must have irqs disabled.
1950 * Returns true if the grace-period kthread needs to be awakened.
1952 static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1953 struct rcu_data *rdp)
1955 bool ret;
1956 bool need_gp;
1958 lockdep_assert_held(&rnp->lock);
1960 /* Handle the ends of any preceding grace periods first. */
1961 if (rdp->completed == rnp->completed &&
1962 !unlikely(READ_ONCE(rdp->gpwrap))) {
1964 /* No grace period end, so just accelerate recent callbacks. */
1965 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1967 } else {
1969 /* Advance callbacks. */
1970 ret = rcu_advance_cbs(rsp, rnp, rdp);
1972 /* Remember that we saw this grace-period completion. */
1973 rdp->completed = rnp->completed;
1974 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1977 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1979 * If the current grace period is waiting for this CPU,
1980 * set up to detect a quiescent state, otherwise don't
1981 * go looking for one.
1983 rdp->gpnum = rnp->gpnum;
1984 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1985 need_gp = !!(rnp->qsmask & rdp->grpmask);
1986 rdp->cpu_no_qs.b.norm = need_gp;
1987 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
1988 rdp->core_needs_qs = need_gp;
1989 zero_cpu_stall_ticks(rdp);
1990 WRITE_ONCE(rdp->gpwrap, false);
1991 rcu_gpnum_ovf(rnp, rdp);
1993 return ret;
1996 static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1998 unsigned long flags;
1999 bool needwake;
2000 struct rcu_node *rnp;
2002 local_irq_save(flags);
2003 rnp = rdp->mynode;
2004 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
2005 rdp->completed == READ_ONCE(rnp->completed) &&
2006 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
2007 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
2008 local_irq_restore(flags);
2009 return;
2011 needwake = __note_gp_changes(rsp, rnp, rdp);
2012 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2013 if (needwake)
2014 rcu_gp_kthread_wake(rsp);
2017 static void rcu_gp_slow(struct rcu_state *rsp, int delay)
2019 if (delay > 0 &&
2020 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
2021 schedule_timeout_uninterruptible(delay);
2025 * Initialize a new grace period. Return false if no grace period required.
2027 static bool rcu_gp_init(struct rcu_state *rsp)
2029 unsigned long oldmask;
2030 struct rcu_data *rdp;
2031 struct rcu_node *rnp = rcu_get_root(rsp);
2033 WRITE_ONCE(rsp->gp_activity, jiffies);
2034 raw_spin_lock_irq_rcu_node(rnp);
2035 if (!READ_ONCE(rsp->gp_flags)) {
2036 /* Spurious wakeup, tell caller to go back to sleep. */
2037 raw_spin_unlock_irq_rcu_node(rnp);
2038 return false;
2040 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
2042 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
2044 * Grace period already in progress, don't start another.
2045 * Not supposed to be able to happen.
2047 raw_spin_unlock_irq_rcu_node(rnp);
2048 return false;
2051 /* Advance to a new grace period and initialize state. */
2052 record_gp_stall_check_time(rsp);
2053 /* Record GP times before starting GP, hence smp_store_release(). */
2054 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
2055 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
2056 raw_spin_unlock_irq_rcu_node(rnp);
2059 * Apply per-leaf buffered online and offline operations to the
2060 * rcu_node tree. Note that this new grace period need not wait
2061 * for subsequent online CPUs, and that quiescent-state forcing
2062 * will handle subsequent offline CPUs.
2064 rcu_for_each_leaf_node(rsp, rnp) {
2065 rcu_gp_slow(rsp, gp_preinit_delay);
2066 raw_spin_lock_irq_rcu_node(rnp);
2067 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
2068 !rnp->wait_blkd_tasks) {
2069 /* Nothing to do on this leaf rcu_node structure. */
2070 raw_spin_unlock_irq_rcu_node(rnp);
2071 continue;
2074 /* Record old state, apply changes to ->qsmaskinit field. */
2075 oldmask = rnp->qsmaskinit;
2076 rnp->qsmaskinit = rnp->qsmaskinitnext;
2078 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
2079 if (!oldmask != !rnp->qsmaskinit) {
2080 if (!oldmask) /* First online CPU for this rcu_node. */
2081 rcu_init_new_rnp(rnp);
2082 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
2083 rnp->wait_blkd_tasks = true;
2084 else /* Last offline CPU and can propagate. */
2085 rcu_cleanup_dead_rnp(rnp);
2089 * If all waited-on tasks from prior grace period are
2090 * done, and if all this rcu_node structure's CPUs are
2091 * still offline, propagate up the rcu_node tree and
2092 * clear ->wait_blkd_tasks. Otherwise, if one of this
2093 * rcu_node structure's CPUs has since come back online,
2094 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
2095 * checks for this, so just call it unconditionally).
2097 if (rnp->wait_blkd_tasks &&
2098 (!rcu_preempt_has_tasks(rnp) ||
2099 rnp->qsmaskinit)) {
2100 rnp->wait_blkd_tasks = false;
2101 rcu_cleanup_dead_rnp(rnp);
2104 raw_spin_unlock_irq_rcu_node(rnp);
2108 * Set the quiescent-state-needed bits in all the rcu_node
2109 * structures for all currently online CPUs in breadth-first order,
2110 * starting from the root rcu_node structure, relying on the layout
2111 * of the tree within the rsp->node[] array. Note that other CPUs
2112 * will access only the leaves of the hierarchy, thus seeing that no
2113 * grace period is in progress, at least until the corresponding
2114 * leaf node has been initialized.
2116 * The grace period cannot complete until the initialization
2117 * process finishes, because this kthread handles both.
2119 rcu_for_each_node_breadth_first(rsp, rnp) {
2120 rcu_gp_slow(rsp, gp_init_delay);
2121 raw_spin_lock_irq_rcu_node(rnp);
2122 rdp = this_cpu_ptr(rsp->rda);
2123 rcu_preempt_check_blocked_tasks(rnp);
2124 rnp->qsmask = rnp->qsmaskinit;
2125 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2126 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2127 WRITE_ONCE(rnp->completed, rsp->completed);
2128 if (rnp == rdp->mynode)
2129 (void)__note_gp_changes(rsp, rnp, rdp);
2130 rcu_preempt_boost_start_gp(rnp);
2131 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2132 rnp->level, rnp->grplo,
2133 rnp->grphi, rnp->qsmask);
2134 raw_spin_unlock_irq_rcu_node(rnp);
2135 cond_resched_rcu_qs();
2136 WRITE_ONCE(rsp->gp_activity, jiffies);
2139 return true;
2143 * Helper function for swait_event_idle() wakeup at force-quiescent-state
2144 * time.
2146 static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2148 struct rcu_node *rnp = rcu_get_root(rsp);
2150 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2151 *gfp = READ_ONCE(rsp->gp_flags);
2152 if (*gfp & RCU_GP_FLAG_FQS)
2153 return true;
2155 /* The current grace period has completed. */
2156 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2157 return true;
2159 return false;
2163 * Do one round of quiescent-state forcing.
2165 static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2167 struct rcu_node *rnp = rcu_get_root(rsp);
2169 WRITE_ONCE(rsp->gp_activity, jiffies);
2170 rsp->n_force_qs++;
2171 if (first_time) {
2172 /* Collect dyntick-idle snapshots. */
2173 force_qs_rnp(rsp, dyntick_save_progress_counter);
2174 } else {
2175 /* Handle dyntick-idle and offline CPUs. */
2176 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
2178 /* Clear flag to prevent immediate re-entry. */
2179 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2180 raw_spin_lock_irq_rcu_node(rnp);
2181 WRITE_ONCE(rsp->gp_flags,
2182 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2183 raw_spin_unlock_irq_rcu_node(rnp);
2188 * Clean up after the old grace period.
2190 static void rcu_gp_cleanup(struct rcu_state *rsp)
2192 unsigned long gp_duration;
2193 bool needgp = false;
2194 int nocb = 0;
2195 struct rcu_data *rdp;
2196 struct rcu_node *rnp = rcu_get_root(rsp);
2197 struct swait_queue_head *sq;
2199 WRITE_ONCE(rsp->gp_activity, jiffies);
2200 raw_spin_lock_irq_rcu_node(rnp);
2201 gp_duration = jiffies - rsp->gp_start;
2202 if (gp_duration > rsp->gp_max)
2203 rsp->gp_max = gp_duration;
2206 * We know the grace period is complete, but to everyone else
2207 * it appears to still be ongoing. But it is also the case
2208 * that to everyone else it looks like there is nothing that
2209 * they can do to advance the grace period. It is therefore
2210 * safe for us to drop the lock in order to mark the grace
2211 * period as completed in all of the rcu_node structures.
2213 raw_spin_unlock_irq_rcu_node(rnp);
2216 * Propagate new ->completed value to rcu_node structures so
2217 * that other CPUs don't have to wait until the start of the next
2218 * grace period to process their callbacks. This also avoids
2219 * some nasty RCU grace-period initialization races by forcing
2220 * the end of the current grace period to be completely recorded in
2221 * all of the rcu_node structures before the beginning of the next
2222 * grace period is recorded in any of the rcu_node structures.
2224 rcu_for_each_node_breadth_first(rsp, rnp) {
2225 raw_spin_lock_irq_rcu_node(rnp);
2226 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2227 WARN_ON_ONCE(rnp->qsmask);
2228 WRITE_ONCE(rnp->completed, rsp->gpnum);
2229 rdp = this_cpu_ptr(rsp->rda);
2230 if (rnp == rdp->mynode)
2231 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2232 /* smp_mb() provided by prior unlock-lock pair. */
2233 nocb += rcu_future_gp_cleanup(rsp, rnp);
2234 sq = rcu_nocb_gp_get(rnp);
2235 raw_spin_unlock_irq_rcu_node(rnp);
2236 rcu_nocb_gp_cleanup(sq);
2237 cond_resched_rcu_qs();
2238 WRITE_ONCE(rsp->gp_activity, jiffies);
2239 rcu_gp_slow(rsp, gp_cleanup_delay);
2241 rnp = rcu_get_root(rsp);
2242 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2243 rcu_nocb_gp_set(rnp, nocb);
2245 /* Declare grace period done. */
2246 WRITE_ONCE(rsp->completed, rsp->gpnum);
2247 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2248 rsp->gp_state = RCU_GP_IDLE;
2249 rdp = this_cpu_ptr(rsp->rda);
2250 /* Advance CBs to reduce false positives below. */
2251 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2252 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2253 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2254 trace_rcu_grace_period(rsp->name,
2255 READ_ONCE(rsp->gpnum),
2256 TPS("newreq"));
2258 raw_spin_unlock_irq_rcu_node(rnp);
2262 * Body of kthread that handles grace periods.
2264 static int __noreturn rcu_gp_kthread(void *arg)
2266 bool first_gp_fqs;
2267 int gf;
2268 unsigned long j;
2269 int ret;
2270 struct rcu_state *rsp = arg;
2271 struct rcu_node *rnp = rcu_get_root(rsp);
2273 rcu_bind_gp_kthread();
2274 for (;;) {
2276 /* Handle grace-period start. */
2277 for (;;) {
2278 trace_rcu_grace_period(rsp->name,
2279 READ_ONCE(rsp->gpnum),
2280 TPS("reqwait"));
2281 rsp->gp_state = RCU_GP_WAIT_GPS;
2282 swait_event_idle(rsp->gp_wq, READ_ONCE(rsp->gp_flags) &
2283 RCU_GP_FLAG_INIT);
2284 rsp->gp_state = RCU_GP_DONE_GPS;
2285 /* Locking provides needed memory barrier. */
2286 if (rcu_gp_init(rsp))
2287 break;
2288 cond_resched_rcu_qs();
2289 WRITE_ONCE(rsp->gp_activity, jiffies);
2290 WARN_ON(signal_pending(current));
2291 trace_rcu_grace_period(rsp->name,
2292 READ_ONCE(rsp->gpnum),
2293 TPS("reqwaitsig"));
2296 /* Handle quiescent-state forcing. */
2297 first_gp_fqs = true;
2298 j = jiffies_till_first_fqs;
2299 if (j > HZ) {
2300 j = HZ;
2301 jiffies_till_first_fqs = HZ;
2303 ret = 0;
2304 for (;;) {
2305 if (!ret) {
2306 rsp->jiffies_force_qs = jiffies + j;
2307 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2308 jiffies + 3 * j);
2310 trace_rcu_grace_period(rsp->name,
2311 READ_ONCE(rsp->gpnum),
2312 TPS("fqswait"));
2313 rsp->gp_state = RCU_GP_WAIT_FQS;
2314 ret = swait_event_idle_timeout(rsp->gp_wq,
2315 rcu_gp_fqs_check_wake(rsp, &gf), j);
2316 rsp->gp_state = RCU_GP_DOING_FQS;
2317 /* Locking provides needed memory barriers. */
2318 /* If grace period done, leave loop. */
2319 if (!READ_ONCE(rnp->qsmask) &&
2320 !rcu_preempt_blocked_readers_cgp(rnp))
2321 break;
2322 /* If time for quiescent-state forcing, do it. */
2323 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2324 (gf & RCU_GP_FLAG_FQS)) {
2325 trace_rcu_grace_period(rsp->name,
2326 READ_ONCE(rsp->gpnum),
2327 TPS("fqsstart"));
2328 rcu_gp_fqs(rsp, first_gp_fqs);
2329 first_gp_fqs = false;
2330 trace_rcu_grace_period(rsp->name,
2331 READ_ONCE(rsp->gpnum),
2332 TPS("fqsend"));
2333 cond_resched_rcu_qs();
2334 WRITE_ONCE(rsp->gp_activity, jiffies);
2335 ret = 0; /* Force full wait till next FQS. */
2336 j = jiffies_till_next_fqs;
2337 if (j > HZ) {
2338 j = HZ;
2339 jiffies_till_next_fqs = HZ;
2340 } else if (j < 1) {
2341 j = 1;
2342 jiffies_till_next_fqs = 1;
2344 } else {
2345 /* Deal with stray signal. */
2346 cond_resched_rcu_qs();
2347 WRITE_ONCE(rsp->gp_activity, jiffies);
2348 WARN_ON(signal_pending(current));
2349 trace_rcu_grace_period(rsp->name,
2350 READ_ONCE(rsp->gpnum),
2351 TPS("fqswaitsig"));
2352 ret = 1; /* Keep old FQS timing. */
2353 j = jiffies;
2354 if (time_after(jiffies, rsp->jiffies_force_qs))
2355 j = 1;
2356 else
2357 j = rsp->jiffies_force_qs - j;
2361 /* Handle grace-period end. */
2362 rsp->gp_state = RCU_GP_CLEANUP;
2363 rcu_gp_cleanup(rsp);
2364 rsp->gp_state = RCU_GP_CLEANED;
2369 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2370 * in preparation for detecting the next grace period. The caller must hold
2371 * the root node's ->lock and hard irqs must be disabled.
2373 * Note that it is legal for a dying CPU (which is marked as offline) to
2374 * invoke this function. This can happen when the dying CPU reports its
2375 * quiescent state.
2377 * Returns true if the grace-period kthread must be awakened.
2379 static bool
2380 rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2381 struct rcu_data *rdp)
2383 lockdep_assert_held(&rnp->lock);
2384 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2386 * Either we have not yet spawned the grace-period
2387 * task, this CPU does not need another grace period,
2388 * or a grace period is already in progress.
2389 * Either way, don't start a new grace period.
2391 return false;
2393 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2394 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2395 TPS("newreq"));
2398 * We can't do wakeups while holding the rnp->lock, as that
2399 * could cause possible deadlocks with the rq->lock. Defer
2400 * the wakeup to our caller.
2402 return true;
2406 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2407 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2408 * is invoked indirectly from rcu_advance_cbs(), which would result in
2409 * endless recursion -- or would do so if it wasn't for the self-deadlock
2410 * that is encountered beforehand.
2412 * Returns true if the grace-period kthread needs to be awakened.
2414 static bool rcu_start_gp(struct rcu_state *rsp)
2416 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2417 struct rcu_node *rnp = rcu_get_root(rsp);
2418 bool ret = false;
2421 * If there is no grace period in progress right now, any
2422 * callbacks we have up to this point will be satisfied by the
2423 * next grace period. Also, advancing the callbacks reduces the
2424 * probability of false positives from cpu_needs_another_gp()
2425 * resulting in pointless grace periods. So, advance callbacks
2426 * then start the grace period!
2428 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2429 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2430 return ret;
2434 * Report a full set of quiescent states to the specified rcu_state data
2435 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2436 * kthread if another grace period is required. Whether we wake
2437 * the grace-period kthread or it awakens itself for the next round
2438 * of quiescent-state forcing, that kthread will clean up after the
2439 * just-completed grace period. Note that the caller must hold rnp->lock,
2440 * which is released before return.
2442 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2443 __releases(rcu_get_root(rsp)->lock)
2445 lockdep_assert_held(&rcu_get_root(rsp)->lock);
2446 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2447 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2448 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2449 rcu_gp_kthread_wake(rsp);
2453 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2454 * Allows quiescent states for a group of CPUs to be reported at one go
2455 * to the specified rcu_node structure, though all the CPUs in the group
2456 * must be represented by the same rcu_node structure (which need not be a
2457 * leaf rcu_node structure, though it often will be). The gps parameter
2458 * is the grace-period snapshot, which means that the quiescent states
2459 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2460 * must be held upon entry, and it is released before return.
2462 static void
2463 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2464 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2465 __releases(rnp->lock)
2467 unsigned long oldmask = 0;
2468 struct rcu_node *rnp_c;
2470 lockdep_assert_held(&rnp->lock);
2472 /* Walk up the rcu_node hierarchy. */
2473 for (;;) {
2474 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2477 * Our bit has already been cleared, or the
2478 * relevant grace period is already over, so done.
2480 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2481 return;
2483 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2484 WARN_ON_ONCE(rnp->level != rcu_num_lvls - 1 &&
2485 rcu_preempt_blocked_readers_cgp(rnp));
2486 rnp->qsmask &= ~mask;
2487 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2488 mask, rnp->qsmask, rnp->level,
2489 rnp->grplo, rnp->grphi,
2490 !!rnp->gp_tasks);
2491 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2493 /* Other bits still set at this level, so done. */
2494 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2495 return;
2497 mask = rnp->grpmask;
2498 if (rnp->parent == NULL) {
2500 /* No more levels. Exit loop holding root lock. */
2502 break;
2504 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2505 rnp_c = rnp;
2506 rnp = rnp->parent;
2507 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2508 oldmask = rnp_c->qsmask;
2512 * Get here if we are the last CPU to pass through a quiescent
2513 * state for this grace period. Invoke rcu_report_qs_rsp()
2514 * to clean up and start the next grace period if one is needed.
2516 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2520 * Record a quiescent state for all tasks that were previously queued
2521 * on the specified rcu_node structure and that were blocking the current
2522 * RCU grace period. The caller must hold the specified rnp->lock with
2523 * irqs disabled, and this lock is released upon return, but irqs remain
2524 * disabled.
2526 static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2527 struct rcu_node *rnp, unsigned long flags)
2528 __releases(rnp->lock)
2530 unsigned long gps;
2531 unsigned long mask;
2532 struct rcu_node *rnp_p;
2534 lockdep_assert_held(&rnp->lock);
2535 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2536 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2537 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2538 return; /* Still need more quiescent states! */
2541 rnp_p = rnp->parent;
2542 if (rnp_p == NULL) {
2544 * Only one rcu_node structure in the tree, so don't
2545 * try to report up to its nonexistent parent!
2547 rcu_report_qs_rsp(rsp, flags);
2548 return;
2551 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2552 gps = rnp->gpnum;
2553 mask = rnp->grpmask;
2554 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2555 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2556 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2560 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2561 * structure. This must be called from the specified CPU.
2563 static void
2564 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2566 unsigned long flags;
2567 unsigned long mask;
2568 bool needwake;
2569 struct rcu_node *rnp;
2571 rnp = rdp->mynode;
2572 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2573 if (rdp->cpu_no_qs.b.norm || rdp->gpnum != rnp->gpnum ||
2574 rnp->completed == rnp->gpnum || rdp->gpwrap) {
2577 * The grace period in which this quiescent state was
2578 * recorded has ended, so don't report it upwards.
2579 * We will instead need a new quiescent state that lies
2580 * within the current grace period.
2582 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2583 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_dynticks.rcu_qs_ctr);
2584 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2585 return;
2587 mask = rdp->grpmask;
2588 if ((rnp->qsmask & mask) == 0) {
2589 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2590 } else {
2591 rdp->core_needs_qs = false;
2594 * This GP can't end until cpu checks in, so all of our
2595 * callbacks can be processed during the next GP.
2597 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2599 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2600 /* ^^^ Released rnp->lock */
2601 if (needwake)
2602 rcu_gp_kthread_wake(rsp);
2607 * Check to see if there is a new grace period of which this CPU
2608 * is not yet aware, and if so, set up local rcu_data state for it.
2609 * Otherwise, see if this CPU has just passed through its first
2610 * quiescent state for this grace period, and record that fact if so.
2612 static void
2613 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2615 /* Check for grace-period ends and beginnings. */
2616 note_gp_changes(rsp, rdp);
2619 * Does this CPU still need to do its part for current grace period?
2620 * If no, return and let the other CPUs do their part as well.
2622 if (!rdp->core_needs_qs)
2623 return;
2626 * Was there a quiescent state since the beginning of the grace
2627 * period? If no, then exit and wait for the next call.
2629 if (rdp->cpu_no_qs.b.norm)
2630 return;
2633 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2634 * judge of that).
2636 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2640 * Trace the fact that this CPU is going offline.
2642 static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2644 RCU_TRACE(unsigned long mask;)
2645 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda);)
2646 RCU_TRACE(struct rcu_node *rnp = rdp->mynode;)
2648 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2649 return;
2651 RCU_TRACE(mask = rdp->grpmask;)
2652 trace_rcu_grace_period(rsp->name,
2653 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2654 TPS("cpuofl"));
2658 * All CPUs for the specified rcu_node structure have gone offline,
2659 * and all tasks that were preempted within an RCU read-side critical
2660 * section while running on one of those CPUs have since exited their RCU
2661 * read-side critical section. Some other CPU is reporting this fact with
2662 * the specified rcu_node structure's ->lock held and interrupts disabled.
2663 * This function therefore goes up the tree of rcu_node structures,
2664 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2665 * the leaf rcu_node structure's ->qsmaskinit field has already been
2666 * updated
2668 * This function does check that the specified rcu_node structure has
2669 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2670 * prematurely. That said, invoking it after the fact will cost you
2671 * a needless lock acquisition. So once it has done its work, don't
2672 * invoke it again.
2674 static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2676 long mask;
2677 struct rcu_node *rnp = rnp_leaf;
2679 lockdep_assert_held(&rnp->lock);
2680 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2681 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2682 return;
2683 for (;;) {
2684 mask = rnp->grpmask;
2685 rnp = rnp->parent;
2686 if (!rnp)
2687 break;
2688 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2689 rnp->qsmaskinit &= ~mask;
2690 rnp->qsmask &= ~mask;
2691 if (rnp->qsmaskinit) {
2692 raw_spin_unlock_rcu_node(rnp);
2693 /* irqs remain disabled. */
2694 return;
2696 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2701 * The CPU has been completely removed, and some other CPU is reporting
2702 * this fact from process context. Do the remainder of the cleanup.
2703 * There can only be one CPU hotplug operation at a time, so no need for
2704 * explicit locking.
2706 static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2708 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2709 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2711 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2712 return;
2714 /* Adjust any no-longer-needed kthreads. */
2715 rcu_boost_kthread_setaffinity(rnp, -1);
2719 * Invoke any RCU callbacks that have made it to the end of their grace
2720 * period. Thottle as specified by rdp->blimit.
2722 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2724 unsigned long flags;
2725 struct rcu_head *rhp;
2726 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2727 long bl, count;
2729 /* If no callbacks are ready, just return. */
2730 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2731 trace_rcu_batch_start(rsp->name,
2732 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2733 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2734 trace_rcu_batch_end(rsp->name, 0,
2735 !rcu_segcblist_empty(&rdp->cblist),
2736 need_resched(), is_idle_task(current),
2737 rcu_is_callbacks_kthread());
2738 return;
2742 * Extract the list of ready callbacks, disabling to prevent
2743 * races with call_rcu() from interrupt handlers. Leave the
2744 * callback counts, as rcu_barrier() needs to be conservative.
2746 local_irq_save(flags);
2747 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2748 bl = rdp->blimit;
2749 trace_rcu_batch_start(rsp->name, rcu_segcblist_n_lazy_cbs(&rdp->cblist),
2750 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2751 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2752 local_irq_restore(flags);
2754 /* Invoke callbacks. */
2755 rhp = rcu_cblist_dequeue(&rcl);
2756 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2757 debug_rcu_head_unqueue(rhp);
2758 if (__rcu_reclaim(rsp->name, rhp))
2759 rcu_cblist_dequeued_lazy(&rcl);
2761 * Stop only if limit reached and CPU has something to do.
2762 * Note: The rcl structure counts down from zero.
2764 if (-rcl.len >= bl &&
2765 (need_resched() ||
2766 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2767 break;
2770 local_irq_save(flags);
2771 count = -rcl.len;
2772 trace_rcu_batch_end(rsp->name, count, !!rcl.head, need_resched(),
2773 is_idle_task(current), rcu_is_callbacks_kthread());
2775 /* Update counts and requeue any remaining callbacks. */
2776 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2777 smp_mb(); /* List handling before counting for rcu_barrier(). */
2778 rdp->n_cbs_invoked += count;
2779 rcu_segcblist_insert_count(&rdp->cblist, &rcl);
2781 /* Reinstate batch limit if we have worked down the excess. */
2782 count = rcu_segcblist_n_cbs(&rdp->cblist);
2783 if (rdp->blimit == LONG_MAX && count <= qlowmark)
2784 rdp->blimit = blimit;
2786 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2787 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2788 rdp->qlen_last_fqs_check = 0;
2789 rdp->n_force_qs_snap = rsp->n_force_qs;
2790 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2791 rdp->qlen_last_fqs_check = count;
2792 WARN_ON_ONCE(rcu_segcblist_empty(&rdp->cblist) != (count == 0));
2794 local_irq_restore(flags);
2796 /* Re-invoke RCU core processing if there are callbacks remaining. */
2797 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2798 invoke_rcu_core();
2802 * Check to see if this CPU is in a non-context-switch quiescent state
2803 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2804 * Also schedule RCU core processing.
2806 * This function must be called from hardirq context. It is normally
2807 * invoked from the scheduling-clock interrupt.
2809 void rcu_check_callbacks(int user)
2811 trace_rcu_utilization(TPS("Start scheduler-tick"));
2812 increment_cpu_stall_ticks();
2813 if (user || rcu_is_cpu_rrupt_from_idle()) {
2816 * Get here if this CPU took its interrupt from user
2817 * mode or from the idle loop, and if this is not a
2818 * nested interrupt. In this case, the CPU is in
2819 * a quiescent state, so note it.
2821 * No memory barrier is required here because both
2822 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2823 * variables that other CPUs neither access nor modify,
2824 * at least not while the corresponding CPU is online.
2827 rcu_sched_qs();
2828 rcu_bh_qs();
2830 } else if (!in_softirq()) {
2833 * Get here if this CPU did not take its interrupt from
2834 * softirq, in other words, if it is not interrupting
2835 * a rcu_bh read-side critical section. This is an _bh
2836 * critical section, so note it.
2839 rcu_bh_qs();
2841 rcu_preempt_check_callbacks();
2842 if (rcu_pending())
2843 invoke_rcu_core();
2844 if (user)
2845 rcu_note_voluntary_context_switch(current);
2846 trace_rcu_utilization(TPS("End scheduler-tick"));
2850 * Scan the leaf rcu_node structures, processing dyntick state for any that
2851 * have not yet encountered a quiescent state, using the function specified.
2852 * Also initiate boosting for any threads blocked on the root rcu_node.
2854 * The caller must have suppressed start of new grace periods.
2856 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *rsp))
2858 int cpu;
2859 unsigned long flags;
2860 unsigned long mask;
2861 struct rcu_node *rnp;
2863 rcu_for_each_leaf_node(rsp, rnp) {
2864 cond_resched_rcu_qs();
2865 mask = 0;
2866 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2867 if (rnp->qsmask == 0) {
2868 if (rcu_state_p == &rcu_sched_state ||
2869 rsp != rcu_state_p ||
2870 rcu_preempt_blocked_readers_cgp(rnp)) {
2872 * No point in scanning bits because they
2873 * are all zero. But we might need to
2874 * priority-boost blocked readers.
2876 rcu_initiate_boost(rnp, flags);
2877 /* rcu_initiate_boost() releases rnp->lock */
2878 continue;
2880 if (rnp->parent &&
2881 (rnp->parent->qsmask & rnp->grpmask)) {
2883 * Race between grace-period
2884 * initialization and task exiting RCU
2885 * read-side critical section: Report.
2887 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2888 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2889 continue;
2892 for_each_leaf_node_possible_cpu(rnp, cpu) {
2893 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2894 if ((rnp->qsmask & bit) != 0) {
2895 if (f(per_cpu_ptr(rsp->rda, cpu)))
2896 mask |= bit;
2899 if (mask != 0) {
2900 /* Idle/offline CPUs, report (releases rnp->lock. */
2901 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2902 } else {
2903 /* Nothing to do here, so just drop the lock. */
2904 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2910 * Force quiescent states on reluctant CPUs, and also detect which
2911 * CPUs are in dyntick-idle mode.
2913 static void force_quiescent_state(struct rcu_state *rsp)
2915 unsigned long flags;
2916 bool ret;
2917 struct rcu_node *rnp;
2918 struct rcu_node *rnp_old = NULL;
2920 /* Funnel through hierarchy to reduce memory contention. */
2921 rnp = __this_cpu_read(rsp->rda->mynode);
2922 for (; rnp != NULL; rnp = rnp->parent) {
2923 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2924 !raw_spin_trylock(&rnp->fqslock);
2925 if (rnp_old != NULL)
2926 raw_spin_unlock(&rnp_old->fqslock);
2927 if (ret) {
2928 rsp->n_force_qs_lh++;
2929 return;
2931 rnp_old = rnp;
2933 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2935 /* Reached the root of the rcu_node tree, acquire lock. */
2936 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2937 raw_spin_unlock(&rnp_old->fqslock);
2938 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2939 rsp->n_force_qs_lh++;
2940 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2941 return; /* Someone beat us to it. */
2943 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2944 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2945 rcu_gp_kthread_wake(rsp);
2949 * This does the RCU core processing work for the specified rcu_state
2950 * and rcu_data structures. This may be called only from the CPU to
2951 * whom the rdp belongs.
2953 static void
2954 __rcu_process_callbacks(struct rcu_state *rsp)
2956 unsigned long flags;
2957 bool needwake;
2958 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2960 WARN_ON_ONCE(!rdp->beenonline);
2962 /* Update RCU state based on any recent quiescent states. */
2963 rcu_check_quiescent_state(rsp, rdp);
2965 /* Does this CPU require a not-yet-started grace period? */
2966 local_irq_save(flags);
2967 if (cpu_needs_another_gp(rsp, rdp)) {
2968 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
2969 needwake = rcu_start_gp(rsp);
2970 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2971 if (needwake)
2972 rcu_gp_kthread_wake(rsp);
2973 } else {
2974 local_irq_restore(flags);
2977 /* If there are callbacks ready, invoke them. */
2978 if (rcu_segcblist_ready_cbs(&rdp->cblist))
2979 invoke_rcu_callbacks(rsp, rdp);
2981 /* Do any needed deferred wakeups of rcuo kthreads. */
2982 do_nocb_deferred_wakeup(rdp);
2986 * Do RCU core processing for the current CPU.
2988 static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
2990 struct rcu_state *rsp;
2992 if (cpu_is_offline(smp_processor_id()))
2993 return;
2994 trace_rcu_utilization(TPS("Start RCU core"));
2995 for_each_rcu_flavor(rsp)
2996 __rcu_process_callbacks(rsp);
2997 trace_rcu_utilization(TPS("End RCU core"));
3001 * Schedule RCU callback invocation. If the specified type of RCU
3002 * does not support RCU priority boosting, just do a direct call,
3003 * otherwise wake up the per-CPU kernel kthread. Note that because we
3004 * are running on the current CPU with softirqs disabled, the
3005 * rcu_cpu_kthread_task cannot disappear out from under us.
3007 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
3009 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
3010 return;
3011 if (likely(!rsp->boost)) {
3012 rcu_do_batch(rsp, rdp);
3013 return;
3015 invoke_rcu_callbacks_kthread();
3018 static void invoke_rcu_core(void)
3020 if (cpu_online(smp_processor_id()))
3021 raise_softirq(RCU_SOFTIRQ);
3025 * Handle any core-RCU processing required by a call_rcu() invocation.
3027 static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
3028 struct rcu_head *head, unsigned long flags)
3030 bool needwake;
3033 * If called from an extended quiescent state, invoke the RCU
3034 * core in order to force a re-evaluation of RCU's idleness.
3036 if (!rcu_is_watching())
3037 invoke_rcu_core();
3039 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3040 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
3041 return;
3044 * Force the grace period if too many callbacks or too long waiting.
3045 * Enforce hysteresis, and don't invoke force_quiescent_state()
3046 * if some other CPU has recently done so. Also, don't bother
3047 * invoking force_quiescent_state() if the newly enqueued callback
3048 * is the only one waiting for a grace period to complete.
3050 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
3051 rdp->qlen_last_fqs_check + qhimark)) {
3053 /* Are we ignoring a completed grace period? */
3054 note_gp_changes(rsp, rdp);
3056 /* Start a new grace period if one not already started. */
3057 if (!rcu_gp_in_progress(rsp)) {
3058 struct rcu_node *rnp_root = rcu_get_root(rsp);
3060 raw_spin_lock_rcu_node(rnp_root);
3061 needwake = rcu_start_gp(rsp);
3062 raw_spin_unlock_rcu_node(rnp_root);
3063 if (needwake)
3064 rcu_gp_kthread_wake(rsp);
3065 } else {
3066 /* Give the grace period a kick. */
3067 rdp->blimit = LONG_MAX;
3068 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3069 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
3070 force_quiescent_state(rsp);
3071 rdp->n_force_qs_snap = rsp->n_force_qs;
3072 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
3078 * RCU callback function to leak a callback.
3080 static void rcu_leak_callback(struct rcu_head *rhp)
3085 * Helper function for call_rcu() and friends. The cpu argument will
3086 * normally be -1, indicating "currently running CPU". It may specify
3087 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3088 * is expected to specify a CPU.
3090 static void
3091 __call_rcu(struct rcu_head *head, rcu_callback_t func,
3092 struct rcu_state *rsp, int cpu, bool lazy)
3094 unsigned long flags;
3095 struct rcu_data *rdp;
3097 /* Misaligned rcu_head! */
3098 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3100 if (debug_rcu_head_queue(head)) {
3102 * Probable double call_rcu(), so leak the callback.
3103 * Use rcu:rcu_callback trace event to find the previous
3104 * time callback was passed to __call_rcu().
3106 WARN_ONCE(1, "__call_rcu(): Double-freed CB %p->%pF()!!!\n",
3107 head, head->func);
3108 WRITE_ONCE(head->func, rcu_leak_callback);
3109 return;
3111 head->func = func;
3112 head->next = NULL;
3113 local_irq_save(flags);
3114 rdp = this_cpu_ptr(rsp->rda);
3116 /* Add the callback to our list. */
3117 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist)) || cpu != -1) {
3118 int offline;
3120 if (cpu != -1)
3121 rdp = per_cpu_ptr(rsp->rda, cpu);
3122 if (likely(rdp->mynode)) {
3123 /* Post-boot, so this should be for a no-CBs CPU. */
3124 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3125 WARN_ON_ONCE(offline);
3126 /* Offline CPU, _call_rcu() illegal, leak callback. */
3127 local_irq_restore(flags);
3128 return;
3131 * Very early boot, before rcu_init(). Initialize if needed
3132 * and then drop through to queue the callback.
3134 BUG_ON(cpu != -1);
3135 WARN_ON_ONCE(!rcu_is_watching());
3136 if (rcu_segcblist_empty(&rdp->cblist))
3137 rcu_segcblist_init(&rdp->cblist);
3139 rcu_segcblist_enqueue(&rdp->cblist, head, lazy);
3140 if (!lazy)
3141 rcu_idle_count_callbacks_posted();
3143 if (__is_kfree_rcu_offset((unsigned long)func))
3144 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3145 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3146 rcu_segcblist_n_cbs(&rdp->cblist));
3147 else
3148 trace_rcu_callback(rsp->name, head,
3149 rcu_segcblist_n_lazy_cbs(&rdp->cblist),
3150 rcu_segcblist_n_cbs(&rdp->cblist));
3152 /* Go handle any RCU core processing required. */
3153 __call_rcu_core(rsp, rdp, head, flags);
3154 local_irq_restore(flags);
3158 * call_rcu_sched() - Queue an RCU for invocation after sched grace period.
3159 * @head: structure to be used for queueing the RCU updates.
3160 * @func: actual callback function to be invoked after the grace period
3162 * The callback function will be invoked some time after a full grace
3163 * period elapses, in other words after all currently executing RCU
3164 * read-side critical sections have completed. call_rcu_sched() assumes
3165 * that the read-side critical sections end on enabling of preemption
3166 * or on voluntary preemption.
3167 * RCU read-side critical sections are delimited by:
3169 * - rcu_read_lock_sched() and rcu_read_unlock_sched(), OR
3170 * - anything that disables preemption.
3172 * These may be nested.
3174 * See the description of call_rcu() for more detailed information on
3175 * memory ordering guarantees.
3177 void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3179 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3181 EXPORT_SYMBOL_GPL(call_rcu_sched);
3184 * call_rcu_bh() - Queue an RCU for invocation after a quicker grace period.
3185 * @head: structure to be used for queueing the RCU updates.
3186 * @func: actual callback function to be invoked after the grace period
3188 * The callback function will be invoked some time after a full grace
3189 * period elapses, in other words after all currently executing RCU
3190 * read-side critical sections have completed. call_rcu_bh() assumes
3191 * that the read-side critical sections end on completion of a softirq
3192 * handler. This means that read-side critical sections in process
3193 * context must not be interrupted by softirqs. This interface is to be
3194 * used when most of the read-side critical sections are in softirq context.
3195 * RCU read-side critical sections are delimited by:
3197 * - rcu_read_lock() and rcu_read_unlock(), if in interrupt context, OR
3198 * - rcu_read_lock_bh() and rcu_read_unlock_bh(), if in process context.
3200 * These may be nested.
3202 * See the description of call_rcu() for more detailed information on
3203 * memory ordering guarantees.
3205 void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3207 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3209 EXPORT_SYMBOL_GPL(call_rcu_bh);
3212 * Queue an RCU callback for lazy invocation after a grace period.
3213 * This will likely be later named something like "call_rcu_lazy()",
3214 * but this change will require some way of tagging the lazy RCU
3215 * callbacks in the list of pending callbacks. Until then, this
3216 * function may only be called from __kfree_rcu().
3218 void kfree_call_rcu(struct rcu_head *head,
3219 rcu_callback_t func)
3221 __call_rcu(head, func, rcu_state_p, -1, 1);
3223 EXPORT_SYMBOL_GPL(kfree_call_rcu);
3226 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3227 * any blocking grace-period wait automatically implies a grace period
3228 * if there is only one CPU online at any point time during execution
3229 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3230 * occasionally incorrectly indicate that there are multiple CPUs online
3231 * when there was in fact only one the whole time, as this just adds
3232 * some overhead: RCU still operates correctly.
3234 static inline int rcu_blocking_is_gp(void)
3236 int ret;
3238 might_sleep(); /* Check for RCU read-side critical section. */
3239 preempt_disable();
3240 ret = num_online_cpus() <= 1;
3241 preempt_enable();
3242 return ret;
3246 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3248 * Control will return to the caller some time after a full rcu-sched
3249 * grace period has elapsed, in other words after all currently executing
3250 * rcu-sched read-side critical sections have completed. These read-side
3251 * critical sections are delimited by rcu_read_lock_sched() and
3252 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3253 * local_irq_disable(), and so on may be used in place of
3254 * rcu_read_lock_sched().
3256 * This means that all preempt_disable code sequences, including NMI and
3257 * non-threaded hardware-interrupt handlers, in progress on entry will
3258 * have completed before this primitive returns. However, this does not
3259 * guarantee that softirq handlers will have completed, since in some
3260 * kernels, these handlers can run in process context, and can block.
3262 * Note that this guarantee implies further memory-ordering guarantees.
3263 * On systems with more than one CPU, when synchronize_sched() returns,
3264 * each CPU is guaranteed to have executed a full memory barrier since the
3265 * end of its last RCU-sched read-side critical section whose beginning
3266 * preceded the call to synchronize_sched(). In addition, each CPU having
3267 * an RCU read-side critical section that extends beyond the return from
3268 * synchronize_sched() is guaranteed to have executed a full memory barrier
3269 * after the beginning of synchronize_sched() and before the beginning of
3270 * that RCU read-side critical section. Note that these guarantees include
3271 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3272 * that are executing in the kernel.
3274 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3275 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3276 * to have executed a full memory barrier during the execution of
3277 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3278 * again only if the system has more than one CPU).
3280 void synchronize_sched(void)
3282 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3283 lock_is_held(&rcu_lock_map) ||
3284 lock_is_held(&rcu_sched_lock_map),
3285 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3286 if (rcu_blocking_is_gp())
3287 return;
3288 if (rcu_gp_is_expedited())
3289 synchronize_sched_expedited();
3290 else
3291 wait_rcu_gp(call_rcu_sched);
3293 EXPORT_SYMBOL_GPL(synchronize_sched);
3296 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3298 * Control will return to the caller some time after a full rcu_bh grace
3299 * period has elapsed, in other words after all currently executing rcu_bh
3300 * read-side critical sections have completed. RCU read-side critical
3301 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3302 * and may be nested.
3304 * See the description of synchronize_sched() for more detailed information
3305 * on memory ordering guarantees.
3307 void synchronize_rcu_bh(void)
3309 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3310 lock_is_held(&rcu_lock_map) ||
3311 lock_is_held(&rcu_sched_lock_map),
3312 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3313 if (rcu_blocking_is_gp())
3314 return;
3315 if (rcu_gp_is_expedited())
3316 synchronize_rcu_bh_expedited();
3317 else
3318 wait_rcu_gp(call_rcu_bh);
3320 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3323 * get_state_synchronize_rcu - Snapshot current RCU state
3325 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3326 * to determine whether or not a full grace period has elapsed in the
3327 * meantime.
3329 unsigned long get_state_synchronize_rcu(void)
3332 * Any prior manipulation of RCU-protected data must happen
3333 * before the load from ->gpnum.
3335 smp_mb(); /* ^^^ */
3338 * Make sure this load happens before the purportedly
3339 * time-consuming work between get_state_synchronize_rcu()
3340 * and cond_synchronize_rcu().
3342 return smp_load_acquire(&rcu_state_p->gpnum);
3344 EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3347 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3349 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3351 * If a full RCU grace period has elapsed since the earlier call to
3352 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3353 * synchronize_rcu() to wait for a full grace period.
3355 * Yes, this function does not take counter wrap into account. But
3356 * counter wrap is harmless. If the counter wraps, we have waited for
3357 * more than 2 billion grace periods (and way more on a 64-bit system!),
3358 * so waiting for one additional grace period should be just fine.
3360 void cond_synchronize_rcu(unsigned long oldstate)
3362 unsigned long newstate;
3365 * Ensure that this load happens before any RCU-destructive
3366 * actions the caller might carry out after we return.
3368 newstate = smp_load_acquire(&rcu_state_p->completed);
3369 if (ULONG_CMP_GE(oldstate, newstate))
3370 synchronize_rcu();
3372 EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3375 * get_state_synchronize_sched - Snapshot current RCU-sched state
3377 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3378 * to determine whether or not a full grace period has elapsed in the
3379 * meantime.
3381 unsigned long get_state_synchronize_sched(void)
3384 * Any prior manipulation of RCU-protected data must happen
3385 * before the load from ->gpnum.
3387 smp_mb(); /* ^^^ */
3390 * Make sure this load happens before the purportedly
3391 * time-consuming work between get_state_synchronize_sched()
3392 * and cond_synchronize_sched().
3394 return smp_load_acquire(&rcu_sched_state.gpnum);
3396 EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3399 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3401 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3403 * If a full RCU-sched grace period has elapsed since the earlier call to
3404 * get_state_synchronize_sched(), just return. Otherwise, invoke
3405 * synchronize_sched() to wait for a full grace period.
3407 * Yes, this function does not take counter wrap into account. But
3408 * counter wrap is harmless. If the counter wraps, we have waited for
3409 * more than 2 billion grace periods (and way more on a 64-bit system!),
3410 * so waiting for one additional grace period should be just fine.
3412 void cond_synchronize_sched(unsigned long oldstate)
3414 unsigned long newstate;
3417 * Ensure that this load happens before any RCU-destructive
3418 * actions the caller might carry out after we return.
3420 newstate = smp_load_acquire(&rcu_sched_state.completed);
3421 if (ULONG_CMP_GE(oldstate, newstate))
3422 synchronize_sched();
3424 EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3427 * Check to see if there is any immediate RCU-related work to be done
3428 * by the current CPU, for the specified type of RCU, returning 1 if so.
3429 * The checks are in order of increasing expense: checks that can be
3430 * carried out against CPU-local state are performed first. However,
3431 * we must check for CPU stalls first, else we might not get a chance.
3433 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3435 struct rcu_node *rnp = rdp->mynode;
3437 rdp->n_rcu_pending++;
3439 /* Check for CPU stalls, if enabled. */
3440 check_cpu_stall(rsp, rdp);
3442 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3443 if (rcu_nohz_full_cpu(rsp))
3444 return 0;
3446 /* Is the RCU core waiting for a quiescent state from this CPU? */
3447 if (rcu_scheduler_fully_active &&
3448 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3449 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_dynticks.rcu_qs_ctr)) {
3450 rdp->n_rp_core_needs_qs++;
3451 } else if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm) {
3452 rdp->n_rp_report_qs++;
3453 return 1;
3456 /* Does this CPU have callbacks ready to invoke? */
3457 if (rcu_segcblist_ready_cbs(&rdp->cblist)) {
3458 rdp->n_rp_cb_ready++;
3459 return 1;
3462 /* Has RCU gone idle with this CPU needing another grace period? */
3463 if (cpu_needs_another_gp(rsp, rdp)) {
3464 rdp->n_rp_cpu_needs_gp++;
3465 return 1;
3468 /* Has another RCU grace period completed? */
3469 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3470 rdp->n_rp_gp_completed++;
3471 return 1;
3474 /* Has a new RCU grace period started? */
3475 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3476 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3477 rdp->n_rp_gp_started++;
3478 return 1;
3481 /* Does this CPU need a deferred NOCB wakeup? */
3482 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3483 rdp->n_rp_nocb_defer_wakeup++;
3484 return 1;
3487 /* nothing to do */
3488 rdp->n_rp_need_nothing++;
3489 return 0;
3493 * Check to see if there is any immediate RCU-related work to be done
3494 * by the current CPU, returning 1 if so. This function is part of the
3495 * RCU implementation; it is -not- an exported member of the RCU API.
3497 static int rcu_pending(void)
3499 struct rcu_state *rsp;
3501 for_each_rcu_flavor(rsp)
3502 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3503 return 1;
3504 return 0;
3508 * Return true if the specified CPU has any callback. If all_lazy is
3509 * non-NULL, store an indication of whether all callbacks are lazy.
3510 * (If there are no callbacks, all of them are deemed to be lazy.)
3512 static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3514 bool al = true;
3515 bool hc = false;
3516 struct rcu_data *rdp;
3517 struct rcu_state *rsp;
3519 for_each_rcu_flavor(rsp) {
3520 rdp = this_cpu_ptr(rsp->rda);
3521 if (rcu_segcblist_empty(&rdp->cblist))
3522 continue;
3523 hc = true;
3524 if (rcu_segcblist_n_nonlazy_cbs(&rdp->cblist) || !all_lazy) {
3525 al = false;
3526 break;
3529 if (all_lazy)
3530 *all_lazy = al;
3531 return hc;
3535 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3536 * the compiler is expected to optimize this away.
3538 static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3539 int cpu, unsigned long done)
3541 trace_rcu_barrier(rsp->name, s, cpu,
3542 atomic_read(&rsp->barrier_cpu_count), done);
3546 * RCU callback function for _rcu_barrier(). If we are last, wake
3547 * up the task executing _rcu_barrier().
3549 static void rcu_barrier_callback(struct rcu_head *rhp)
3551 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3552 struct rcu_state *rsp = rdp->rsp;
3554 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3555 _rcu_barrier_trace(rsp, TPS("LastCB"), -1,
3556 rsp->barrier_sequence);
3557 complete(&rsp->barrier_completion);
3558 } else {
3559 _rcu_barrier_trace(rsp, TPS("CB"), -1, rsp->barrier_sequence);
3564 * Called with preemption disabled, and from cross-cpu IRQ context.
3566 static void rcu_barrier_func(void *type)
3568 struct rcu_state *rsp = type;
3569 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3571 _rcu_barrier_trace(rsp, TPS("IRQ"), -1, rsp->barrier_sequence);
3572 rdp->barrier_head.func = rcu_barrier_callback;
3573 debug_rcu_head_queue(&rdp->barrier_head);
3574 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head, 0)) {
3575 atomic_inc(&rsp->barrier_cpu_count);
3576 } else {
3577 debug_rcu_head_unqueue(&rdp->barrier_head);
3578 _rcu_barrier_trace(rsp, TPS("IRQNQ"), -1,
3579 rsp->barrier_sequence);
3584 * Orchestrate the specified type of RCU barrier, waiting for all
3585 * RCU callbacks of the specified type to complete.
3587 static void _rcu_barrier(struct rcu_state *rsp)
3589 int cpu;
3590 struct rcu_data *rdp;
3591 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3593 _rcu_barrier_trace(rsp, TPS("Begin"), -1, s);
3595 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3596 mutex_lock(&rsp->barrier_mutex);
3598 /* Did someone else do our work for us? */
3599 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3600 _rcu_barrier_trace(rsp, TPS("EarlyExit"), -1,
3601 rsp->barrier_sequence);
3602 smp_mb(); /* caller's subsequent code after above check. */
3603 mutex_unlock(&rsp->barrier_mutex);
3604 return;
3607 /* Mark the start of the barrier operation. */
3608 rcu_seq_start(&rsp->barrier_sequence);
3609 _rcu_barrier_trace(rsp, TPS("Inc1"), -1, rsp->barrier_sequence);
3612 * Initialize the count to one rather than to zero in order to
3613 * avoid a too-soon return to zero in case of a short grace period
3614 * (or preemption of this task). Exclude CPU-hotplug operations
3615 * to ensure that no offline CPU has callbacks queued.
3617 init_completion(&rsp->barrier_completion);
3618 atomic_set(&rsp->barrier_cpu_count, 1);
3619 get_online_cpus();
3622 * Force each CPU with callbacks to register a new callback.
3623 * When that callback is invoked, we will know that all of the
3624 * corresponding CPU's preceding callbacks have been invoked.
3626 for_each_possible_cpu(cpu) {
3627 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3628 continue;
3629 rdp = per_cpu_ptr(rsp->rda, cpu);
3630 if (rcu_is_nocb_cpu(cpu)) {
3631 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3632 _rcu_barrier_trace(rsp, TPS("OfflineNoCB"), cpu,
3633 rsp->barrier_sequence);
3634 } else {
3635 _rcu_barrier_trace(rsp, TPS("OnlineNoCB"), cpu,
3636 rsp->barrier_sequence);
3637 smp_mb__before_atomic();
3638 atomic_inc(&rsp->barrier_cpu_count);
3639 __call_rcu(&rdp->barrier_head,
3640 rcu_barrier_callback, rsp, cpu, 0);
3642 } else if (rcu_segcblist_n_cbs(&rdp->cblist)) {
3643 _rcu_barrier_trace(rsp, TPS("OnlineQ"), cpu,
3644 rsp->barrier_sequence);
3645 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3646 } else {
3647 _rcu_barrier_trace(rsp, TPS("OnlineNQ"), cpu,
3648 rsp->barrier_sequence);
3651 put_online_cpus();
3654 * Now that we have an rcu_barrier_callback() callback on each
3655 * CPU, and thus each counted, remove the initial count.
3657 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3658 complete(&rsp->barrier_completion);
3660 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3661 wait_for_completion(&rsp->barrier_completion);
3663 /* Mark the end of the barrier operation. */
3664 _rcu_barrier_trace(rsp, TPS("Inc2"), -1, rsp->barrier_sequence);
3665 rcu_seq_end(&rsp->barrier_sequence);
3667 /* Other rcu_barrier() invocations can now safely proceed. */
3668 mutex_unlock(&rsp->barrier_mutex);
3672 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3674 void rcu_barrier_bh(void)
3676 _rcu_barrier(&rcu_bh_state);
3678 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3681 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3683 void rcu_barrier_sched(void)
3685 _rcu_barrier(&rcu_sched_state);
3687 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3690 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3691 * first CPU in a given leaf rcu_node structure coming online. The caller
3692 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3693 * disabled.
3695 static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3697 long mask;
3698 struct rcu_node *rnp = rnp_leaf;
3700 lockdep_assert_held(&rnp->lock);
3701 for (;;) {
3702 mask = rnp->grpmask;
3703 rnp = rnp->parent;
3704 if (rnp == NULL)
3705 return;
3706 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3707 rnp->qsmaskinit |= mask;
3708 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3713 * Do boot-time initialization of a CPU's per-CPU RCU data.
3715 static void __init
3716 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3718 unsigned long flags;
3719 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3720 struct rcu_node *rnp = rcu_get_root(rsp);
3722 /* Set up local state, ensuring consistent view of global state. */
3723 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3724 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3725 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3726 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3727 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp->dynticks)));
3728 rdp->cpu = cpu;
3729 rdp->rsp = rsp;
3730 rcu_boot_init_nocb_percpu_data(rdp);
3731 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3735 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3736 * offline event can be happening at a given time. Note also that we
3737 * can accept some slop in the rsp->completed access due to the fact
3738 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3740 static void
3741 rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3743 unsigned long flags;
3744 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3745 struct rcu_node *rnp = rcu_get_root(rsp);
3747 /* Set up local state, ensuring consistent view of global state. */
3748 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3749 rdp->qlen_last_fqs_check = 0;
3750 rdp->n_force_qs_snap = rsp->n_force_qs;
3751 rdp->blimit = blimit;
3752 if (rcu_segcblist_empty(&rdp->cblist) && /* No early-boot CBs? */
3753 !init_nocb_callback_list(rdp))
3754 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
3755 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3756 rcu_dynticks_eqs_online();
3757 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3760 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3761 * propagation up the rcu_node tree will happen at the beginning
3762 * of the next grace period.
3764 rnp = rdp->mynode;
3765 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3766 rdp->beenonline = true; /* We have now been online. */
3767 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3768 rdp->completed = rnp->completed;
3769 rdp->cpu_no_qs.b.norm = true;
3770 rdp->rcu_qs_ctr_snap = per_cpu(rcu_dynticks.rcu_qs_ctr, cpu);
3771 rdp->core_needs_qs = false;
3772 rdp->rcu_iw_pending = false;
3773 rdp->rcu_iw_gpnum = rnp->gpnum - 1;
3774 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3775 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3779 * Invoked early in the CPU-online process, when pretty much all
3780 * services are available. The incoming CPU is not present.
3782 int rcutree_prepare_cpu(unsigned int cpu)
3784 struct rcu_state *rsp;
3786 for_each_rcu_flavor(rsp)
3787 rcu_init_percpu_data(cpu, rsp);
3789 rcu_prepare_kthreads(cpu);
3790 rcu_spawn_all_nocb_kthreads(cpu);
3792 return 0;
3796 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
3798 static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3800 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3802 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3806 * Near the end of the CPU-online process. Pretty much all services
3807 * enabled, and the CPU is now very much alive.
3809 int rcutree_online_cpu(unsigned int cpu)
3811 unsigned long flags;
3812 struct rcu_data *rdp;
3813 struct rcu_node *rnp;
3814 struct rcu_state *rsp;
3816 for_each_rcu_flavor(rsp) {
3817 rdp = per_cpu_ptr(rsp->rda, cpu);
3818 rnp = rdp->mynode;
3819 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3820 rnp->ffmask |= rdp->grpmask;
3821 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3823 if (IS_ENABLED(CONFIG_TREE_SRCU))
3824 srcu_online_cpu(cpu);
3825 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
3826 return 0; /* Too early in boot for scheduler work. */
3827 sync_sched_exp_online_cleanup(cpu);
3828 rcutree_affinity_setting(cpu, -1);
3829 return 0;
3833 * Near the beginning of the process. The CPU is still very much alive
3834 * with pretty much all services enabled.
3836 int rcutree_offline_cpu(unsigned int cpu)
3838 unsigned long flags;
3839 struct rcu_data *rdp;
3840 struct rcu_node *rnp;
3841 struct rcu_state *rsp;
3843 for_each_rcu_flavor(rsp) {
3844 rdp = per_cpu_ptr(rsp->rda, cpu);
3845 rnp = rdp->mynode;
3846 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3847 rnp->ffmask &= ~rdp->grpmask;
3848 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3851 rcutree_affinity_setting(cpu, cpu);
3852 if (IS_ENABLED(CONFIG_TREE_SRCU))
3853 srcu_offline_cpu(cpu);
3854 return 0;
3858 * Near the end of the offline process. We do only tracing here.
3860 int rcutree_dying_cpu(unsigned int cpu)
3862 struct rcu_state *rsp;
3864 for_each_rcu_flavor(rsp)
3865 rcu_cleanup_dying_cpu(rsp);
3866 return 0;
3870 * The outgoing CPU is gone and we are running elsewhere.
3872 int rcutree_dead_cpu(unsigned int cpu)
3874 struct rcu_state *rsp;
3876 for_each_rcu_flavor(rsp) {
3877 rcu_cleanup_dead_cpu(cpu, rsp);
3878 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3880 return 0;
3884 * Mark the specified CPU as being online so that subsequent grace periods
3885 * (both expedited and normal) will wait on it. Note that this means that
3886 * incoming CPUs are not allowed to use RCU read-side critical sections
3887 * until this function is called. Failing to observe this restriction
3888 * will result in lockdep splats.
3890 * Note that this function is special in that it is invoked directly
3891 * from the incoming CPU rather than from the cpuhp_step mechanism.
3892 * This is because this function must be invoked at a precise location.
3894 void rcu_cpu_starting(unsigned int cpu)
3896 unsigned long flags;
3897 unsigned long mask;
3898 int nbits;
3899 unsigned long oldmask;
3900 struct rcu_data *rdp;
3901 struct rcu_node *rnp;
3902 struct rcu_state *rsp;
3904 for_each_rcu_flavor(rsp) {
3905 rdp = per_cpu_ptr(rsp->rda, cpu);
3906 rnp = rdp->mynode;
3907 mask = rdp->grpmask;
3908 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3909 rnp->qsmaskinitnext |= mask;
3910 oldmask = rnp->expmaskinitnext;
3911 rnp->expmaskinitnext |= mask;
3912 oldmask ^= rnp->expmaskinitnext;
3913 nbits = bitmap_weight(&oldmask, BITS_PER_LONG);
3914 /* Allow lockless access for expedited grace periods. */
3915 smp_store_release(&rsp->ncpus, rsp->ncpus + nbits); /* ^^^ */
3916 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3918 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
3921 #ifdef CONFIG_HOTPLUG_CPU
3923 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3924 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3925 * bit masks.
3927 static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3929 unsigned long flags;
3930 unsigned long mask;
3931 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3932 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3934 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3935 mask = rdp->grpmask;
3936 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3937 rnp->qsmaskinitnext &= ~mask;
3938 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3942 * The outgoing function has no further need of RCU, so remove it from
3943 * the list of CPUs that RCU must track.
3945 * Note that this function is special in that it is invoked directly
3946 * from the outgoing CPU rather than from the cpuhp_step mechanism.
3947 * This is because this function must be invoked at a precise location.
3949 void rcu_report_dead(unsigned int cpu)
3951 struct rcu_state *rsp;
3953 /* QS for any half-done expedited RCU-sched GP. */
3954 preempt_disable();
3955 rcu_report_exp_rdp(&rcu_sched_state,
3956 this_cpu_ptr(rcu_sched_state.rda), true);
3957 preempt_enable();
3958 for_each_rcu_flavor(rsp)
3959 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3962 /* Migrate the dead CPU's callbacks to the current CPU. */
3963 static void rcu_migrate_callbacks(int cpu, struct rcu_state *rsp)
3965 unsigned long flags;
3966 struct rcu_data *my_rdp;
3967 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3968 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
3970 if (rcu_is_nocb_cpu(cpu) || rcu_segcblist_empty(&rdp->cblist))
3971 return; /* No callbacks to migrate. */
3973 local_irq_save(flags);
3974 my_rdp = this_cpu_ptr(rsp->rda);
3975 if (rcu_nocb_adopt_orphan_cbs(my_rdp, rdp, flags)) {
3976 local_irq_restore(flags);
3977 return;
3979 raw_spin_lock_rcu_node(rnp_root); /* irqs already disabled. */
3980 rcu_advance_cbs(rsp, rnp_root, rdp); /* Leverage recent GPs. */
3981 rcu_advance_cbs(rsp, rnp_root, my_rdp); /* Assign GP to pending CBs. */
3982 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
3983 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
3984 !rcu_segcblist_n_cbs(&my_rdp->cblist));
3985 raw_spin_unlock_irqrestore_rcu_node(rnp_root, flags);
3986 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
3987 !rcu_segcblist_empty(&rdp->cblist),
3988 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
3989 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
3990 rcu_segcblist_first_cb(&rdp->cblist));
3994 * The outgoing CPU has just passed through the dying-idle state,
3995 * and we are being invoked from the CPU that was IPIed to continue the
3996 * offline operation. We need to migrate the outgoing CPU's callbacks.
3998 void rcutree_migrate_callbacks(int cpu)
4000 struct rcu_state *rsp;
4002 for_each_rcu_flavor(rsp)
4003 rcu_migrate_callbacks(cpu, rsp);
4005 #endif
4008 * On non-huge systems, use expedited RCU grace periods to make suspend
4009 * and hibernation run faster.
4011 static int rcu_pm_notify(struct notifier_block *self,
4012 unsigned long action, void *hcpu)
4014 switch (action) {
4015 case PM_HIBERNATION_PREPARE:
4016 case PM_SUSPEND_PREPARE:
4017 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4018 rcu_expedite_gp();
4019 break;
4020 case PM_POST_HIBERNATION:
4021 case PM_POST_SUSPEND:
4022 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
4023 rcu_unexpedite_gp();
4024 break;
4025 default:
4026 break;
4028 return NOTIFY_OK;
4032 * Spawn the kthreads that handle each RCU flavor's grace periods.
4034 static int __init rcu_spawn_gp_kthread(void)
4036 unsigned long flags;
4037 int kthread_prio_in = kthread_prio;
4038 struct rcu_node *rnp;
4039 struct rcu_state *rsp;
4040 struct sched_param sp;
4041 struct task_struct *t;
4043 /* Force priority into range. */
4044 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4045 kthread_prio = 1;
4046 else if (kthread_prio < 0)
4047 kthread_prio = 0;
4048 else if (kthread_prio > 99)
4049 kthread_prio = 99;
4050 if (kthread_prio != kthread_prio_in)
4051 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4052 kthread_prio, kthread_prio_in);
4054 rcu_scheduler_fully_active = 1;
4055 for_each_rcu_flavor(rsp) {
4056 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
4057 BUG_ON(IS_ERR(t));
4058 rnp = rcu_get_root(rsp);
4059 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4060 rsp->gp_kthread = t;
4061 if (kthread_prio) {
4062 sp.sched_priority = kthread_prio;
4063 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4065 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4066 wake_up_process(t);
4068 rcu_spawn_nocb_kthreads();
4069 rcu_spawn_boost_kthreads();
4070 return 0;
4072 early_initcall(rcu_spawn_gp_kthread);
4075 * This function is invoked towards the end of the scheduler's
4076 * initialization process. Before this is called, the idle task might
4077 * contain synchronous grace-period primitives (during which time, this idle
4078 * task is booting the system, and such primitives are no-ops). After this
4079 * function is called, any synchronous grace-period primitives are run as
4080 * expedited, with the requesting task driving the grace period forward.
4081 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4082 * runtime RCU functionality.
4084 void rcu_scheduler_starting(void)
4086 WARN_ON(num_online_cpus() != 1);
4087 WARN_ON(nr_context_switches() > 0);
4088 rcu_test_sync_prims();
4089 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4090 rcu_test_sync_prims();
4094 * Helper function for rcu_init() that initializes one rcu_state structure.
4096 static void __init rcu_init_one(struct rcu_state *rsp)
4098 static const char * const buf[] = RCU_NODE_NAME_INIT;
4099 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4100 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4101 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4103 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4104 int cpustride = 1;
4105 int i;
4106 int j;
4107 struct rcu_node *rnp;
4109 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4111 /* Silence gcc 4.8 false positive about array index out of range. */
4112 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4113 panic("rcu_init_one: rcu_num_lvls out of range");
4115 /* Initialize the level-tracking arrays. */
4117 for (i = 1; i < rcu_num_lvls; i++)
4118 rsp->level[i] = rsp->level[i - 1] + num_rcu_lvl[i - 1];
4119 rcu_init_levelspread(levelspread, num_rcu_lvl);
4121 /* Initialize the elements themselves, starting from the leaves. */
4123 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4124 cpustride *= levelspread[i];
4125 rnp = rsp->level[i];
4126 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4127 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4128 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4129 &rcu_node_class[i], buf[i]);
4130 raw_spin_lock_init(&rnp->fqslock);
4131 lockdep_set_class_and_name(&rnp->fqslock,
4132 &rcu_fqs_class[i], fqs[i]);
4133 rnp->gpnum = rsp->gpnum;
4134 rnp->completed = rsp->completed;
4135 rnp->qsmask = 0;
4136 rnp->qsmaskinit = 0;
4137 rnp->grplo = j * cpustride;
4138 rnp->grphi = (j + 1) * cpustride - 1;
4139 if (rnp->grphi >= nr_cpu_ids)
4140 rnp->grphi = nr_cpu_ids - 1;
4141 if (i == 0) {
4142 rnp->grpnum = 0;
4143 rnp->grpmask = 0;
4144 rnp->parent = NULL;
4145 } else {
4146 rnp->grpnum = j % levelspread[i - 1];
4147 rnp->grpmask = 1UL << rnp->grpnum;
4148 rnp->parent = rsp->level[i - 1] +
4149 j / levelspread[i - 1];
4151 rnp->level = i;
4152 INIT_LIST_HEAD(&rnp->blkd_tasks);
4153 rcu_init_one_nocb(rnp);
4154 init_waitqueue_head(&rnp->exp_wq[0]);
4155 init_waitqueue_head(&rnp->exp_wq[1]);
4156 init_waitqueue_head(&rnp->exp_wq[2]);
4157 init_waitqueue_head(&rnp->exp_wq[3]);
4158 spin_lock_init(&rnp->exp_lock);
4162 init_swait_queue_head(&rsp->gp_wq);
4163 init_swait_queue_head(&rsp->expedited_wq);
4164 rnp = rsp->level[rcu_num_lvls - 1];
4165 for_each_possible_cpu(i) {
4166 while (i > rnp->grphi)
4167 rnp++;
4168 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4169 rcu_boot_init_percpu_data(i, rsp);
4171 list_add(&rsp->flavors, &rcu_struct_flavors);
4175 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4176 * replace the definitions in tree.h because those are needed to size
4177 * the ->node array in the rcu_state structure.
4179 static void __init rcu_init_geometry(void)
4181 ulong d;
4182 int i;
4183 int rcu_capacity[RCU_NUM_LVLS];
4186 * Initialize any unspecified boot parameters.
4187 * The default values of jiffies_till_first_fqs and
4188 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4189 * value, which is a function of HZ, then adding one for each
4190 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4192 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4193 if (jiffies_till_first_fqs == ULONG_MAX)
4194 jiffies_till_first_fqs = d;
4195 if (jiffies_till_next_fqs == ULONG_MAX)
4196 jiffies_till_next_fqs = d;
4198 /* If the compile-time values are accurate, just leave. */
4199 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4200 nr_cpu_ids == NR_CPUS)
4201 return;
4202 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4203 rcu_fanout_leaf, nr_cpu_ids);
4206 * The boot-time rcu_fanout_leaf parameter must be at least two
4207 * and cannot exceed the number of bits in the rcu_node masks.
4208 * Complain and fall back to the compile-time values if this
4209 * limit is exceeded.
4211 if (rcu_fanout_leaf < 2 ||
4212 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4213 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4214 WARN_ON(1);
4215 return;
4219 * Compute number of nodes that can be handled an rcu_node tree
4220 * with the given number of levels.
4222 rcu_capacity[0] = rcu_fanout_leaf;
4223 for (i = 1; i < RCU_NUM_LVLS; i++)
4224 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4227 * The tree must be able to accommodate the configured number of CPUs.
4228 * If this limit is exceeded, fall back to the compile-time values.
4230 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4231 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4232 WARN_ON(1);
4233 return;
4236 /* Calculate the number of levels in the tree. */
4237 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4239 rcu_num_lvls = i + 1;
4241 /* Calculate the number of rcu_nodes at each level of the tree. */
4242 for (i = 0; i < rcu_num_lvls; i++) {
4243 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4244 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4247 /* Calculate the total number of rcu_node structures. */
4248 rcu_num_nodes = 0;
4249 for (i = 0; i < rcu_num_lvls; i++)
4250 rcu_num_nodes += num_rcu_lvl[i];
4254 * Dump out the structure of the rcu_node combining tree associated
4255 * with the rcu_state structure referenced by rsp.
4257 static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4259 int level = 0;
4260 struct rcu_node *rnp;
4262 pr_info("rcu_node tree layout dump\n");
4263 pr_info(" ");
4264 rcu_for_each_node_breadth_first(rsp, rnp) {
4265 if (rnp->level != level) {
4266 pr_cont("\n");
4267 pr_info(" ");
4268 level = rnp->level;
4270 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4272 pr_cont("\n");
4275 void __init rcu_init(void)
4277 int cpu;
4279 rcu_early_boot_tests();
4281 rcu_bootup_announce();
4282 rcu_init_geometry();
4283 rcu_init_one(&rcu_bh_state);
4284 rcu_init_one(&rcu_sched_state);
4285 if (dump_tree)
4286 rcu_dump_rcu_node_tree(&rcu_sched_state);
4287 __rcu_init_preempt();
4288 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4291 * We don't need protection against CPU-hotplug here because
4292 * this is called early in boot, before either interrupts
4293 * or the scheduler are operational.
4295 pm_notifier(rcu_pm_notify, 0);
4296 for_each_online_cpu(cpu) {
4297 rcutree_prepare_cpu(cpu);
4298 rcu_cpu_starting(cpu);
4299 rcutree_online_cpu(cpu);
4303 #include "tree_exp.h"
4304 #include "tree_plugin.h"