Loading...
1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
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.
8 *
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.
13 *
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.
17 *
18 * Copyright IBM Corporation, 2008
19 *
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
23 *
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
26 *
27 * For detailed explanation of Read-Copy Update mechanism see -
28 * Documentation/RCU
29 */
30#include <linux/types.h>
31#include <linux/kernel.h>
32#include <linux/init.h>
33#include <linux/spinlock.h>
34#include <linux/smp.h>
35#include <linux/rcupdate.h>
36#include <linux/interrupt.h>
37#include <linux/sched.h>
38#include <linux/nmi.h>
39#include <linux/atomic.h>
40#include <linux/bitops.h>
41#include <linux/export.h>
42#include <linux/completion.h>
43#include <linux/moduleparam.h>
44#include <linux/percpu.h>
45#include <linux/notifier.h>
46#include <linux/cpu.h>
47#include <linux/mutex.h>
48#include <linux/time.h>
49#include <linux/kernel_stat.h>
50#include <linux/wait.h>
51#include <linux/kthread.h>
52#include <linux/prefetch.h>
53#include <linux/delay.h>
54#include <linux/stop_machine.h>
55#include <linux/random.h>
56#include <linux/trace_events.h>
57#include <linux/suspend.h>
58
59#include "tree.h"
60#include "rcu.h"
61
62#ifdef MODULE_PARAM_PREFIX
63#undef MODULE_PARAM_PREFIX
64#endif
65#define MODULE_PARAM_PREFIX "rcutree."
66
67/* Data structures. */
68
69/*
70 * In order to export the rcu_state name to the tracing tools, it
71 * needs to be added in the __tracepoint_string section.
72 * This requires defining a separate variable tp_<sname>_varname
73 * that points to the string being used, and this will allow
74 * the tracing userspace tools to be able to decipher the string
75 * address to the matching string.
76 */
77#ifdef CONFIG_TRACING
78# define DEFINE_RCU_TPS(sname) \
79static char sname##_varname[] = #sname; \
80static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname;
81# define RCU_STATE_NAME(sname) sname##_varname
82#else
83# define DEFINE_RCU_TPS(sname)
84# define RCU_STATE_NAME(sname) __stringify(sname)
85#endif
86
87#define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
88DEFINE_RCU_TPS(sname) \
89static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, sname##_data); \
90struct rcu_state sname##_state = { \
91 .level = { &sname##_state.node[0] }, \
92 .rda = &sname##_data, \
93 .call = cr, \
94 .gp_state = RCU_GP_IDLE, \
95 .gpnum = 0UL - 300UL, \
96 .completed = 0UL - 300UL, \
97 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
98 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
99 .orphan_donetail = &sname##_state.orphan_donelist, \
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), \
105}
106
107RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
108RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
109
110static struct rcu_state *const rcu_state_p;
111LIST_HEAD(rcu_struct_flavors);
112
113/* Dump rcu_node combining tree at boot to verify correct setup. */
114static bool dump_tree;
115module_param(dump_tree, bool, 0444);
116/* Control rcu_node-tree auto-balancing at boot time. */
117static bool rcu_fanout_exact;
118module_param(rcu_fanout_exact, bool, 0444);
119/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
120static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
121module_param(rcu_fanout_leaf, int, 0444);
122int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
123/* Number of rcu_nodes at specified level. */
124static int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
125int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
126/* panic() on RCU Stall sysctl. */
127int sysctl_panic_on_rcu_stall __read_mostly;
128
129/*
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.
140 */
141int rcu_scheduler_active __read_mostly;
142EXPORT_SYMBOL_GPL(rcu_scheduler_active);
143
144/*
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.
151 *
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.
155 */
156static int rcu_scheduler_fully_active __read_mostly;
157
158static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
159static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
160static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
161static void invoke_rcu_core(void);
162static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
163static void rcu_report_exp_rdp(struct rcu_state *rsp,
164 struct rcu_data *rdp, bool wake);
165static void sync_sched_exp_online_cleanup(int cpu);
166
167/* rcuc/rcub kthread realtime priority */
168#ifdef CONFIG_RCU_KTHREAD_PRIO
169static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO;
170#else /* #ifdef CONFIG_RCU_KTHREAD_PRIO */
171static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
172#endif /* #else #ifdef CONFIG_RCU_KTHREAD_PRIO */
173module_param(kthread_prio, int, 0644);
174
175/* Delay in jiffies for grace-period initialization delays, debug only. */
176
177#ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT
178static int gp_preinit_delay = CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT_DELAY;
179module_param(gp_preinit_delay, int, 0644);
180#else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
181static const int gp_preinit_delay;
182#endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_PREINIT */
183
184#ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT
185static int gp_init_delay = CONFIG_RCU_TORTURE_TEST_SLOW_INIT_DELAY;
186module_param(gp_init_delay, int, 0644);
187#else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
188static const int gp_init_delay;
189#endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_INIT */
190
191#ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP
192static int gp_cleanup_delay = CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP_DELAY;
193module_param(gp_cleanup_delay, int, 0644);
194#else /* #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
195static const int gp_cleanup_delay;
196#endif /* #else #ifdef CONFIG_RCU_TORTURE_TEST_SLOW_CLEANUP */
197
198/*
199 * Number of grace periods between delays, normalized by the duration of
200 * the delay. The longer the the delay, the more the grace periods between
201 * each delay. The reason for this normalization is that it means that,
202 * for non-zero delays, the overall slowdown of grace periods is constant
203 * regardless of the duration of the delay. This arrangement balances
204 * the need for long delays to increase some race probabilities with the
205 * need for fast grace periods to increase other race probabilities.
206 */
207#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays. */
208
209/*
210 * Track the rcutorture test sequence number and the update version
211 * number within a given test. The rcutorture_testseq is incremented
212 * on every rcutorture module load and unload, so has an odd value
213 * when a test is running. The rcutorture_vernum is set to zero
214 * when rcutorture starts and is incremented on each rcutorture update.
215 * These variables enable correlating rcutorture output with the
216 * RCU tracing information.
217 */
218unsigned long rcutorture_testseq;
219unsigned long rcutorture_vernum;
220
221/*
222 * Compute the mask of online CPUs for the specified rcu_node structure.
223 * This will not be stable unless the rcu_node structure's ->lock is
224 * held, but the bit corresponding to the current CPU will be stable
225 * in most contexts.
226 */
227unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
228{
229 return READ_ONCE(rnp->qsmaskinitnext);
230}
231
232/*
233 * Return true if an RCU grace period is in progress. The READ_ONCE()s
234 * permit this function to be invoked without holding the root rcu_node
235 * structure's ->lock, but of course results can be subject to change.
236 */
237static int rcu_gp_in_progress(struct rcu_state *rsp)
238{
239 return READ_ONCE(rsp->completed) != READ_ONCE(rsp->gpnum);
240}
241
242/*
243 * Note a quiescent state. Because we do not need to know
244 * how many quiescent states passed, just if there was at least
245 * one since the start of the grace period, this just sets a flag.
246 * The caller must have disabled preemption.
247 */
248void rcu_sched_qs(void)
249{
250 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.s))
251 return;
252 trace_rcu_grace_period(TPS("rcu_sched"),
253 __this_cpu_read(rcu_sched_data.gpnum),
254 TPS("cpuqs"));
255 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.norm, false);
256 if (!__this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))
257 return;
258 __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, false);
259 rcu_report_exp_rdp(&rcu_sched_state,
260 this_cpu_ptr(&rcu_sched_data), true);
261}
262
263void rcu_bh_qs(void)
264{
265 if (__this_cpu_read(rcu_bh_data.cpu_no_qs.s)) {
266 trace_rcu_grace_period(TPS("rcu_bh"),
267 __this_cpu_read(rcu_bh_data.gpnum),
268 TPS("cpuqs"));
269 __this_cpu_write(rcu_bh_data.cpu_no_qs.b.norm, false);
270 }
271}
272
273static DEFINE_PER_CPU(int, rcu_sched_qs_mask);
274
275static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
276 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
277 .dynticks = ATOMIC_INIT(1),
278#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
279 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
280 .dynticks_idle = ATOMIC_INIT(1),
281#endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
282};
283
284DEFINE_PER_CPU_SHARED_ALIGNED(unsigned long, rcu_qs_ctr);
285EXPORT_PER_CPU_SYMBOL_GPL(rcu_qs_ctr);
286
287/*
288 * Let the RCU core know that this CPU has gone through the scheduler,
289 * which is a quiescent state. This is called when the need for a
290 * quiescent state is urgent, so we burn an atomic operation and full
291 * memory barriers to let the RCU core know about it, regardless of what
292 * this CPU might (or might not) do in the near future.
293 *
294 * We inform the RCU core by emulating a zero-duration dyntick-idle
295 * period, which we in turn do by incrementing the ->dynticks counter
296 * by two.
297 *
298 * The caller must have disabled interrupts.
299 */
300static void rcu_momentary_dyntick_idle(void)
301{
302 struct rcu_data *rdp;
303 struct rcu_dynticks *rdtp;
304 int resched_mask;
305 struct rcu_state *rsp;
306
307 /*
308 * Yes, we can lose flag-setting operations. This is OK, because
309 * the flag will be set again after some delay.
310 */
311 resched_mask = raw_cpu_read(rcu_sched_qs_mask);
312 raw_cpu_write(rcu_sched_qs_mask, 0);
313
314 /* Find the flavor that needs a quiescent state. */
315 for_each_rcu_flavor(rsp) {
316 rdp = raw_cpu_ptr(rsp->rda);
317 if (!(resched_mask & rsp->flavor_mask))
318 continue;
319 smp_mb(); /* rcu_sched_qs_mask before cond_resched_completed. */
320 if (READ_ONCE(rdp->mynode->completed) !=
321 READ_ONCE(rdp->cond_resched_completed))
322 continue;
323
324 /*
325 * Pretend to be momentarily idle for the quiescent state.
326 * This allows the grace-period kthread to record the
327 * quiescent state, with no need for this CPU to do anything
328 * further.
329 */
330 rdtp = this_cpu_ptr(&rcu_dynticks);
331 smp_mb__before_atomic(); /* Earlier stuff before QS. */
332 atomic_add(2, &rdtp->dynticks); /* QS. */
333 smp_mb__after_atomic(); /* Later stuff after QS. */
334 break;
335 }
336}
337
338/*
339 * Note a context switch. This is a quiescent state for RCU-sched,
340 * and requires special handling for preemptible RCU.
341 * The caller must have disabled interrupts.
342 */
343void rcu_note_context_switch(void)
344{
345 barrier(); /* Avoid RCU read-side critical sections leaking down. */
346 trace_rcu_utilization(TPS("Start context switch"));
347 rcu_sched_qs();
348 rcu_preempt_note_context_switch();
349 if (unlikely(raw_cpu_read(rcu_sched_qs_mask)))
350 rcu_momentary_dyntick_idle();
351 trace_rcu_utilization(TPS("End context switch"));
352 barrier(); /* Avoid RCU read-side critical sections leaking up. */
353}
354EXPORT_SYMBOL_GPL(rcu_note_context_switch);
355
356/*
357 * Register a quiescent state for all RCU flavors. If there is an
358 * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
359 * dyntick-idle quiescent state visible to other CPUs (but only for those
360 * RCU flavors in desperate need of a quiescent state, which will normally
361 * be none of them). Either way, do a lightweight quiescent state for
362 * all RCU flavors.
363 *
364 * The barrier() calls are redundant in the common case when this is
365 * called externally, but just in case this is called from within this
366 * file.
367 *
368 */
369void rcu_all_qs(void)
370{
371 unsigned long flags;
372
373 barrier(); /* Avoid RCU read-side critical sections leaking down. */
374 if (unlikely(raw_cpu_read(rcu_sched_qs_mask))) {
375 local_irq_save(flags);
376 rcu_momentary_dyntick_idle();
377 local_irq_restore(flags);
378 }
379 if (unlikely(raw_cpu_read(rcu_sched_data.cpu_no_qs.b.exp))) {
380 /*
381 * Yes, we just checked a per-CPU variable with preemption
382 * enabled, so we might be migrated to some other CPU at
383 * this point. That is OK because in that case, the
384 * migration will supply the needed quiescent state.
385 * We might end up needlessly disabling preemption and
386 * invoking rcu_sched_qs() on the destination CPU, but
387 * the probability and cost are both quite low, so this
388 * should not be a problem in practice.
389 */
390 preempt_disable();
391 rcu_sched_qs();
392 preempt_enable();
393 }
394 this_cpu_inc(rcu_qs_ctr);
395 barrier(); /* Avoid RCU read-side critical sections leaking up. */
396}
397EXPORT_SYMBOL_GPL(rcu_all_qs);
398
399static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
400static long qhimark = 10000; /* If this many pending, ignore blimit. */
401static long qlowmark = 100; /* Once only this many pending, use blimit. */
402
403module_param(blimit, long, 0444);
404module_param(qhimark, long, 0444);
405module_param(qlowmark, long, 0444);
406
407static ulong jiffies_till_first_fqs = ULONG_MAX;
408static ulong jiffies_till_next_fqs = ULONG_MAX;
409static bool rcu_kick_kthreads;
410
411module_param(jiffies_till_first_fqs, ulong, 0644);
412module_param(jiffies_till_next_fqs, ulong, 0644);
413module_param(rcu_kick_kthreads, bool, 0644);
414
415/*
416 * How long the grace period must be before we start recruiting
417 * quiescent-state help from rcu_note_context_switch().
418 */
419static ulong jiffies_till_sched_qs = HZ / 20;
420module_param(jiffies_till_sched_qs, ulong, 0644);
421
422static bool rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
423 struct rcu_data *rdp);
424static void force_qs_rnp(struct rcu_state *rsp,
425 int (*f)(struct rcu_data *rsp, bool *isidle,
426 unsigned long *maxj),
427 bool *isidle, unsigned long *maxj);
428static void force_quiescent_state(struct rcu_state *rsp);
429static int rcu_pending(void);
430
431/*
432 * Return the number of RCU batches started thus far for debug & stats.
433 */
434unsigned long rcu_batches_started(void)
435{
436 return rcu_state_p->gpnum;
437}
438EXPORT_SYMBOL_GPL(rcu_batches_started);
439
440/*
441 * Return the number of RCU-sched batches started thus far for debug & stats.
442 */
443unsigned long rcu_batches_started_sched(void)
444{
445 return rcu_sched_state.gpnum;
446}
447EXPORT_SYMBOL_GPL(rcu_batches_started_sched);
448
449/*
450 * Return the number of RCU BH batches started thus far for debug & stats.
451 */
452unsigned long rcu_batches_started_bh(void)
453{
454 return rcu_bh_state.gpnum;
455}
456EXPORT_SYMBOL_GPL(rcu_batches_started_bh);
457
458/*
459 * Return the number of RCU batches completed thus far for debug & stats.
460 */
461unsigned long rcu_batches_completed(void)
462{
463 return rcu_state_p->completed;
464}
465EXPORT_SYMBOL_GPL(rcu_batches_completed);
466
467/*
468 * Return the number of RCU-sched batches completed thus far for debug & stats.
469 */
470unsigned long rcu_batches_completed_sched(void)
471{
472 return rcu_sched_state.completed;
473}
474EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
475
476/*
477 * Return the number of RCU BH batches completed thus far for debug & stats.
478 */
479unsigned long rcu_batches_completed_bh(void)
480{
481 return rcu_bh_state.completed;
482}
483EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
484
485/*
486 * Return the number of RCU expedited batches completed thus far for
487 * debug & stats. Odd numbers mean that a batch is in progress, even
488 * numbers mean idle. The value returned will thus be roughly double
489 * the cumulative batches since boot.
490 */
491unsigned long rcu_exp_batches_completed(void)
492{
493 return rcu_state_p->expedited_sequence;
494}
495EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
496
497/*
498 * Return the number of RCU-sched expedited batches completed thus far
499 * for debug & stats. Similar to rcu_exp_batches_completed().
500 */
501unsigned long rcu_exp_batches_completed_sched(void)
502{
503 return rcu_sched_state.expedited_sequence;
504}
505EXPORT_SYMBOL_GPL(rcu_exp_batches_completed_sched);
506
507/*
508 * Force a quiescent state.
509 */
510void rcu_force_quiescent_state(void)
511{
512 force_quiescent_state(rcu_state_p);
513}
514EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
515
516/*
517 * Force a quiescent state for RCU BH.
518 */
519void rcu_bh_force_quiescent_state(void)
520{
521 force_quiescent_state(&rcu_bh_state);
522}
523EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
524
525/*
526 * Force a quiescent state for RCU-sched.
527 */
528void rcu_sched_force_quiescent_state(void)
529{
530 force_quiescent_state(&rcu_sched_state);
531}
532EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
533
534/*
535 * Show the state of the grace-period kthreads.
536 */
537void show_rcu_gp_kthreads(void)
538{
539 struct rcu_state *rsp;
540
541 for_each_rcu_flavor(rsp) {
542 pr_info("%s: wait state: %d ->state: %#lx\n",
543 rsp->name, rsp->gp_state, rsp->gp_kthread->state);
544 /* sched_show_task(rsp->gp_kthread); */
545 }
546}
547EXPORT_SYMBOL_GPL(show_rcu_gp_kthreads);
548
549/*
550 * Record the number of times rcutorture tests have been initiated and
551 * terminated. This information allows the debugfs tracing stats to be
552 * correlated to the rcutorture messages, even when the rcutorture module
553 * is being repeatedly loaded and unloaded. In other words, we cannot
554 * store this state in rcutorture itself.
555 */
556void rcutorture_record_test_transition(void)
557{
558 rcutorture_testseq++;
559 rcutorture_vernum = 0;
560}
561EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
562
563/*
564 * Send along grace-period-related data for rcutorture diagnostics.
565 */
566void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
567 unsigned long *gpnum, unsigned long *completed)
568{
569 struct rcu_state *rsp = NULL;
570
571 switch (test_type) {
572 case RCU_FLAVOR:
573 rsp = rcu_state_p;
574 break;
575 case RCU_BH_FLAVOR:
576 rsp = &rcu_bh_state;
577 break;
578 case RCU_SCHED_FLAVOR:
579 rsp = &rcu_sched_state;
580 break;
581 default:
582 break;
583 }
584 if (rsp != NULL) {
585 *flags = READ_ONCE(rsp->gp_flags);
586 *gpnum = READ_ONCE(rsp->gpnum);
587 *completed = READ_ONCE(rsp->completed);
588 return;
589 }
590 *flags = 0;
591 *gpnum = 0;
592 *completed = 0;
593}
594EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
595
596/*
597 * Record the number of writer passes through the current rcutorture test.
598 * This is also used to correlate debugfs tracing stats with the rcutorture
599 * messages.
600 */
601void rcutorture_record_progress(unsigned long vernum)
602{
603 rcutorture_vernum++;
604}
605EXPORT_SYMBOL_GPL(rcutorture_record_progress);
606
607/*
608 * Does the CPU have callbacks ready to be invoked?
609 */
610static int
611cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
612{
613 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
614 rdp->nxttail[RCU_DONE_TAIL] != NULL;
615}
616
617/*
618 * Return the root node of the specified rcu_state structure.
619 */
620static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
621{
622 return &rsp->node[0];
623}
624
625/*
626 * Is there any need for future grace periods?
627 * Interrupts must be disabled. If the caller does not hold the root
628 * rnp_node structure's ->lock, the results are advisory only.
629 */
630static int rcu_future_needs_gp(struct rcu_state *rsp)
631{
632 struct rcu_node *rnp = rcu_get_root(rsp);
633 int idx = (READ_ONCE(rnp->completed) + 1) & 0x1;
634 int *fp = &rnp->need_future_gp[idx];
635
636 return READ_ONCE(*fp);
637}
638
639/*
640 * Does the current CPU require a not-yet-started grace period?
641 * The caller must have disabled interrupts to prevent races with
642 * normal callback registry.
643 */
644static bool
645cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
646{
647 int i;
648
649 if (rcu_gp_in_progress(rsp))
650 return false; /* No, a grace period is already in progress. */
651 if (rcu_future_needs_gp(rsp))
652 return true; /* Yes, a no-CBs CPU needs one. */
653 if (!rdp->nxttail[RCU_NEXT_TAIL])
654 return false; /* No, this is a no-CBs (or offline) CPU. */
655 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
656 return true; /* Yes, CPU has newly registered callbacks. */
657 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
658 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
659 ULONG_CMP_LT(READ_ONCE(rsp->completed),
660 rdp->nxtcompleted[i]))
661 return true; /* Yes, CBs for future grace period. */
662 return false; /* No grace period needed. */
663}
664
665/*
666 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
667 *
668 * If the new value of the ->dynticks_nesting counter now is zero,
669 * we really have entered idle, and must do the appropriate accounting.
670 * The caller must have disabled interrupts.
671 */
672static void rcu_eqs_enter_common(long long oldval, bool user)
673{
674 struct rcu_state *rsp;
675 struct rcu_data *rdp;
676 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
677
678 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
679 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
680 !user && !is_idle_task(current)) {
681 struct task_struct *idle __maybe_unused =
682 idle_task(smp_processor_id());
683
684 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
685 rcu_ftrace_dump(DUMP_ORIG);
686 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
687 current->pid, current->comm,
688 idle->pid, idle->comm); /* must be idle task! */
689 }
690 for_each_rcu_flavor(rsp) {
691 rdp = this_cpu_ptr(rsp->rda);
692 do_nocb_deferred_wakeup(rdp);
693 }
694 rcu_prepare_for_idle();
695 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
696 smp_mb__before_atomic(); /* See above. */
697 atomic_inc(&rdtp->dynticks);
698 smp_mb__after_atomic(); /* Force ordering with next sojourn. */
699 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
700 atomic_read(&rdtp->dynticks) & 0x1);
701 rcu_dynticks_task_enter();
702
703 /*
704 * It is illegal to enter an extended quiescent state while
705 * in an RCU read-side critical section.
706 */
707 RCU_LOCKDEP_WARN(lock_is_held(&rcu_lock_map),
708 "Illegal idle entry in RCU read-side critical section.");
709 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map),
710 "Illegal idle entry in RCU-bh read-side critical section.");
711 RCU_LOCKDEP_WARN(lock_is_held(&rcu_sched_lock_map),
712 "Illegal idle entry in RCU-sched read-side critical section.");
713}
714
715/*
716 * Enter an RCU extended quiescent state, which can be either the
717 * idle loop or adaptive-tickless usermode execution.
718 */
719static void rcu_eqs_enter(bool user)
720{
721 long long oldval;
722 struct rcu_dynticks *rdtp;
723
724 rdtp = this_cpu_ptr(&rcu_dynticks);
725 oldval = rdtp->dynticks_nesting;
726 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
727 (oldval & DYNTICK_TASK_NEST_MASK) == 0);
728 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
729 rdtp->dynticks_nesting = 0;
730 rcu_eqs_enter_common(oldval, user);
731 } else {
732 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
733 }
734}
735
736/**
737 * rcu_idle_enter - inform RCU that current CPU is entering idle
738 *
739 * Enter idle mode, in other words, -leave- the mode in which RCU
740 * read-side critical sections can occur. (Though RCU read-side
741 * critical sections can occur in irq handlers in idle, a possibility
742 * handled by irq_enter() and irq_exit().)
743 *
744 * We crowbar the ->dynticks_nesting field to zero to allow for
745 * the possibility of usermode upcalls having messed up our count
746 * of interrupt nesting level during the prior busy period.
747 */
748void rcu_idle_enter(void)
749{
750 unsigned long flags;
751
752 local_irq_save(flags);
753 rcu_eqs_enter(false);
754 rcu_sysidle_enter(0);
755 local_irq_restore(flags);
756}
757EXPORT_SYMBOL_GPL(rcu_idle_enter);
758
759#ifdef CONFIG_NO_HZ_FULL
760/**
761 * rcu_user_enter - inform RCU that we are resuming userspace.
762 *
763 * Enter RCU idle mode right before resuming userspace. No use of RCU
764 * is permitted between this call and rcu_user_exit(). This way the
765 * CPU doesn't need to maintain the tick for RCU maintenance purposes
766 * when the CPU runs in userspace.
767 */
768void rcu_user_enter(void)
769{
770 rcu_eqs_enter(1);
771}
772#endif /* CONFIG_NO_HZ_FULL */
773
774/**
775 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
776 *
777 * Exit from an interrupt handler, which might possibly result in entering
778 * idle mode, in other words, leaving the mode in which read-side critical
779 * sections can occur. The caller must have disabled interrupts.
780 *
781 * This code assumes that the idle loop never does anything that might
782 * result in unbalanced calls to irq_enter() and irq_exit(). If your
783 * architecture violates this assumption, RCU will give you what you
784 * deserve, good and hard. But very infrequently and irreproducibly.
785 *
786 * Use things like work queues to work around this limitation.
787 *
788 * You have been warned.
789 */
790void rcu_irq_exit(void)
791{
792 long long oldval;
793 struct rcu_dynticks *rdtp;
794
795 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_exit() invoked with irqs enabled!!!");
796 rdtp = this_cpu_ptr(&rcu_dynticks);
797 oldval = rdtp->dynticks_nesting;
798 rdtp->dynticks_nesting--;
799 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
800 rdtp->dynticks_nesting < 0);
801 if (rdtp->dynticks_nesting)
802 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
803 else
804 rcu_eqs_enter_common(oldval, true);
805 rcu_sysidle_enter(1);
806}
807
808/*
809 * Wrapper for rcu_irq_exit() where interrupts are enabled.
810 */
811void rcu_irq_exit_irqson(void)
812{
813 unsigned long flags;
814
815 local_irq_save(flags);
816 rcu_irq_exit();
817 local_irq_restore(flags);
818}
819
820/*
821 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
822 *
823 * If the new value of the ->dynticks_nesting counter was previously zero,
824 * we really have exited idle, and must do the appropriate accounting.
825 * The caller must have disabled interrupts.
826 */
827static void rcu_eqs_exit_common(long long oldval, int user)
828{
829 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
830
831 rcu_dynticks_task_exit();
832 smp_mb__before_atomic(); /* Force ordering w/previous sojourn. */
833 atomic_inc(&rdtp->dynticks);
834 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
835 smp_mb__after_atomic(); /* See above. */
836 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
837 !(atomic_read(&rdtp->dynticks) & 0x1));
838 rcu_cleanup_after_idle();
839 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
840 if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
841 !user && !is_idle_task(current)) {
842 struct task_struct *idle __maybe_unused =
843 idle_task(smp_processor_id());
844
845 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
846 oldval, rdtp->dynticks_nesting);
847 rcu_ftrace_dump(DUMP_ORIG);
848 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
849 current->pid, current->comm,
850 idle->pid, idle->comm); /* must be idle task! */
851 }
852}
853
854/*
855 * Exit an RCU extended quiescent state, which can be either the
856 * idle loop or adaptive-tickless usermode execution.
857 */
858static void rcu_eqs_exit(bool user)
859{
860 struct rcu_dynticks *rdtp;
861 long long oldval;
862
863 rdtp = this_cpu_ptr(&rcu_dynticks);
864 oldval = rdtp->dynticks_nesting;
865 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
866 if (oldval & DYNTICK_TASK_NEST_MASK) {
867 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
868 } else {
869 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
870 rcu_eqs_exit_common(oldval, user);
871 }
872}
873
874/**
875 * rcu_idle_exit - inform RCU that current CPU is leaving idle
876 *
877 * Exit idle mode, in other words, -enter- the mode in which RCU
878 * read-side critical sections can occur.
879 *
880 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
881 * allow for the possibility of usermode upcalls messing up our count
882 * of interrupt nesting level during the busy period that is just
883 * now starting.
884 */
885void rcu_idle_exit(void)
886{
887 unsigned long flags;
888
889 local_irq_save(flags);
890 rcu_eqs_exit(false);
891 rcu_sysidle_exit(0);
892 local_irq_restore(flags);
893}
894EXPORT_SYMBOL_GPL(rcu_idle_exit);
895
896#ifdef CONFIG_NO_HZ_FULL
897/**
898 * rcu_user_exit - inform RCU that we are exiting userspace.
899 *
900 * Exit RCU idle mode while entering the kernel because it can
901 * run a RCU read side critical section anytime.
902 */
903void rcu_user_exit(void)
904{
905 rcu_eqs_exit(1);
906}
907#endif /* CONFIG_NO_HZ_FULL */
908
909/**
910 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
911 *
912 * Enter an interrupt handler, which might possibly result in exiting
913 * idle mode, in other words, entering the mode in which read-side critical
914 * sections can occur. The caller must have disabled interrupts.
915 *
916 * Note that the Linux kernel is fully capable of entering an interrupt
917 * handler that it never exits, for example when doing upcalls to
918 * user mode! This code assumes that the idle loop never does upcalls to
919 * user mode. If your architecture does do upcalls from the idle loop (or
920 * does anything else that results in unbalanced calls to the irq_enter()
921 * and irq_exit() functions), RCU will give you what you deserve, good
922 * and hard. But very infrequently and irreproducibly.
923 *
924 * Use things like work queues to work around this limitation.
925 *
926 * You have been warned.
927 */
928void rcu_irq_enter(void)
929{
930 struct rcu_dynticks *rdtp;
931 long long oldval;
932
933 RCU_LOCKDEP_WARN(!irqs_disabled(), "rcu_irq_enter() invoked with irqs enabled!!!");
934 rdtp = this_cpu_ptr(&rcu_dynticks);
935 oldval = rdtp->dynticks_nesting;
936 rdtp->dynticks_nesting++;
937 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
938 rdtp->dynticks_nesting == 0);
939 if (oldval)
940 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
941 else
942 rcu_eqs_exit_common(oldval, true);
943 rcu_sysidle_exit(1);
944}
945
946/*
947 * Wrapper for rcu_irq_enter() where interrupts are enabled.
948 */
949void rcu_irq_enter_irqson(void)
950{
951 unsigned long flags;
952
953 local_irq_save(flags);
954 rcu_irq_enter();
955 local_irq_restore(flags);
956}
957
958/**
959 * rcu_nmi_enter - inform RCU of entry to NMI context
960 *
961 * If the CPU was idle from RCU's viewpoint, update rdtp->dynticks and
962 * rdtp->dynticks_nmi_nesting to let the RCU grace-period handling know
963 * that the CPU is active. This implementation permits nested NMIs, as
964 * long as the nesting level does not overflow an int. (You will probably
965 * run out of stack space first.)
966 */
967void rcu_nmi_enter(void)
968{
969 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
970 int incby = 2;
971
972 /* Complain about underflow. */
973 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting < 0);
974
975 /*
976 * If idle from RCU viewpoint, atomically increment ->dynticks
977 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
978 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
979 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
980 * to be in the outermost NMI handler that interrupted an RCU-idle
981 * period (observation due to Andy Lutomirski).
982 */
983 if (!(atomic_read(&rdtp->dynticks) & 0x1)) {
984 smp_mb__before_atomic(); /* Force delay from prior write. */
985 atomic_inc(&rdtp->dynticks);
986 /* atomic_inc() before later RCU read-side crit sects */
987 smp_mb__after_atomic(); /* See above. */
988 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
989 incby = 1;
990 }
991 rdtp->dynticks_nmi_nesting += incby;
992 barrier();
993}
994
995/**
996 * rcu_nmi_exit - inform RCU of exit from NMI context
997 *
998 * If we are returning from the outermost NMI handler that interrupted an
999 * RCU-idle period, update rdtp->dynticks and rdtp->dynticks_nmi_nesting
1000 * to let the RCU grace-period handling know that the CPU is back to
1001 * being RCU-idle.
1002 */
1003void rcu_nmi_exit(void)
1004{
1005 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1006
1007 /*
1008 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
1009 * (We are exiting an NMI handler, so RCU better be paying attention
1010 * to us!)
1011 */
1012 WARN_ON_ONCE(rdtp->dynticks_nmi_nesting <= 0);
1013 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
1014
1015 /*
1016 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
1017 * leave it in non-RCU-idle state.
1018 */
1019 if (rdtp->dynticks_nmi_nesting != 1) {
1020 rdtp->dynticks_nmi_nesting -= 2;
1021 return;
1022 }
1023
1024 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
1025 rdtp->dynticks_nmi_nesting = 0;
1026 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
1027 smp_mb__before_atomic(); /* See above. */
1028 atomic_inc(&rdtp->dynticks);
1029 smp_mb__after_atomic(); /* Force delay to next write. */
1030 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
1031}
1032
1033/**
1034 * __rcu_is_watching - are RCU read-side critical sections safe?
1035 *
1036 * Return true if RCU is watching the running CPU, which means that
1037 * this CPU can safely enter RCU read-side critical sections. Unlike
1038 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
1039 * least disabled preemption.
1040 */
1041bool notrace __rcu_is_watching(void)
1042{
1043 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
1044}
1045
1046/**
1047 * rcu_is_watching - see if RCU thinks that the current CPU is idle
1048 *
1049 * If the current CPU is in its idle loop and is neither in an interrupt
1050 * or NMI handler, return true.
1051 */
1052bool notrace rcu_is_watching(void)
1053{
1054 bool ret;
1055
1056 preempt_disable_notrace();
1057 ret = __rcu_is_watching();
1058 preempt_enable_notrace();
1059 return ret;
1060}
1061EXPORT_SYMBOL_GPL(rcu_is_watching);
1062
1063#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1064
1065/*
1066 * Is the current CPU online? Disable preemption to avoid false positives
1067 * that could otherwise happen due to the current CPU number being sampled,
1068 * this task being preempted, its old CPU being taken offline, resuming
1069 * on some other CPU, then determining that its old CPU is now offline.
1070 * It is OK to use RCU on an offline processor during initial boot, hence
1071 * the check for rcu_scheduler_fully_active. Note also that it is OK
1072 * for a CPU coming online to use RCU for one jiffy prior to marking itself
1073 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
1074 * offline to continue to use RCU for one jiffy after marking itself
1075 * offline in the cpu_online_mask. This leniency is necessary given the
1076 * non-atomic nature of the online and offline processing, for example,
1077 * the fact that a CPU enters the scheduler after completing the teardown
1078 * of the CPU.
1079 *
1080 * This is also why RCU internally marks CPUs online during in the
1081 * preparation phase and offline after the CPU has been taken down.
1082 *
1083 * Disable checking if in an NMI handler because we cannot safely report
1084 * errors from NMI handlers anyway.
1085 */
1086bool rcu_lockdep_current_cpu_online(void)
1087{
1088 struct rcu_data *rdp;
1089 struct rcu_node *rnp;
1090 bool ret;
1091
1092 if (in_nmi())
1093 return true;
1094 preempt_disable();
1095 rdp = this_cpu_ptr(&rcu_sched_data);
1096 rnp = rdp->mynode;
1097 ret = (rdp->grpmask & rcu_rnp_online_cpus(rnp)) ||
1098 !rcu_scheduler_fully_active;
1099 preempt_enable();
1100 return ret;
1101}
1102EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1103
1104#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1105
1106/**
1107 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
1108 *
1109 * If the current CPU is idle or running at a first-level (not nested)
1110 * interrupt from idle, return true. The caller must have at least
1111 * disabled preemption.
1112 */
1113static int rcu_is_cpu_rrupt_from_idle(void)
1114{
1115 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
1116}
1117
1118/*
1119 * Snapshot the specified CPU's dynticks counter so that we can later
1120 * credit them with an implicit quiescent state. Return 1 if this CPU
1121 * is in dynticks idle mode, which is an extended quiescent state.
1122 */
1123static int dyntick_save_progress_counter(struct rcu_data *rdp,
1124 bool *isidle, unsigned long *maxj)
1125{
1126 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
1127 rcu_sysidle_check_cpu(rdp, isidle, maxj);
1128 if ((rdp->dynticks_snap & 0x1) == 0) {
1129 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1130 if (ULONG_CMP_LT(READ_ONCE(rdp->gpnum) + ULONG_MAX / 4,
1131 rdp->mynode->gpnum))
1132 WRITE_ONCE(rdp->gpwrap, true);
1133 return 1;
1134 }
1135 return 0;
1136}
1137
1138/*
1139 * Return true if the specified CPU has passed through a quiescent
1140 * state by virtue of being in or having passed through an dynticks
1141 * idle state since the last call to dyntick_save_progress_counter()
1142 * for this same CPU, or by virtue of having been offline.
1143 */
1144static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
1145 bool *isidle, unsigned long *maxj)
1146{
1147 unsigned int curr;
1148 int *rcrmp;
1149 unsigned int snap;
1150
1151 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
1152 snap = (unsigned int)rdp->dynticks_snap;
1153
1154 /*
1155 * If the CPU passed through or entered a dynticks idle phase with
1156 * no active irq/NMI handlers, then we can safely pretend that the CPU
1157 * already acknowledged the request to pass through a quiescent
1158 * state. Either way, that CPU cannot possibly be in an RCU
1159 * read-side critical section that started before the beginning
1160 * of the current RCU grace period.
1161 */
1162 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
1163 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
1164 rdp->dynticks_fqs++;
1165 return 1;
1166 }
1167
1168 /*
1169 * Check for the CPU being offline, but only if the grace period
1170 * is old enough. We don't need to worry about the CPU changing
1171 * state: If we see it offline even once, it has been through a
1172 * quiescent state.
1173 *
1174 * The reason for insisting that the grace period be at least
1175 * one jiffy old is that CPUs that are not quite online and that
1176 * have just gone offline can still execute RCU read-side critical
1177 * sections.
1178 */
1179 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
1180 return 0; /* Grace period is not old enough. */
1181 barrier();
1182 if (cpu_is_offline(rdp->cpu)) {
1183 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
1184 rdp->offline_fqs++;
1185 return 1;
1186 }
1187
1188 /*
1189 * A CPU running for an extended time within the kernel can
1190 * delay RCU grace periods. When the CPU is in NO_HZ_FULL mode,
1191 * even context-switching back and forth between a pair of
1192 * in-kernel CPU-bound tasks cannot advance grace periods.
1193 * So if the grace period is old enough, make the CPU pay attention.
1194 * Note that the unsynchronized assignments to the per-CPU
1195 * rcu_sched_qs_mask variable are safe. Yes, setting of
1196 * bits can be lost, but they will be set again on the next
1197 * force-quiescent-state pass. So lost bit sets do not result
1198 * in incorrect behavior, merely in a grace period lasting
1199 * a few jiffies longer than it might otherwise. Because
1200 * there are at most four threads involved, and because the
1201 * updates are only once every few jiffies, the probability of
1202 * lossage (and thus of slight grace-period extension) is
1203 * quite low.
1204 *
1205 * Note that if the jiffies_till_sched_qs boot/sysfs parameter
1206 * is set too high, we override with half of the RCU CPU stall
1207 * warning delay.
1208 */
1209 rcrmp = &per_cpu(rcu_sched_qs_mask, rdp->cpu);
1210 if (ULONG_CMP_GE(jiffies,
1211 rdp->rsp->gp_start + jiffies_till_sched_qs) ||
1212 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
1213 if (!(READ_ONCE(*rcrmp) & rdp->rsp->flavor_mask)) {
1214 WRITE_ONCE(rdp->cond_resched_completed,
1215 READ_ONCE(rdp->mynode->completed));
1216 smp_mb(); /* ->cond_resched_completed before *rcrmp. */
1217 WRITE_ONCE(*rcrmp,
1218 READ_ONCE(*rcrmp) + rdp->rsp->flavor_mask);
1219 }
1220 rdp->rsp->jiffies_resched += 5; /* Re-enable beating. */
1221 }
1222
1223 /* And if it has been a really long time, kick the CPU as well. */
1224 if (ULONG_CMP_GE(jiffies,
1225 rdp->rsp->gp_start + 2 * jiffies_till_sched_qs) ||
1226 ULONG_CMP_GE(jiffies, rdp->rsp->gp_start + jiffies_till_sched_qs))
1227 resched_cpu(rdp->cpu); /* Force CPU into scheduler. */
1228
1229 return 0;
1230}
1231
1232static void record_gp_stall_check_time(struct rcu_state *rsp)
1233{
1234 unsigned long j = jiffies;
1235 unsigned long j1;
1236
1237 rsp->gp_start = j;
1238 smp_wmb(); /* Record start time before stall time. */
1239 j1 = rcu_jiffies_till_stall_check();
1240 WRITE_ONCE(rsp->jiffies_stall, j + j1);
1241 rsp->jiffies_resched = j + j1 / 2;
1242 rsp->n_force_qs_gpstart = READ_ONCE(rsp->n_force_qs);
1243}
1244
1245/*
1246 * Convert a ->gp_state value to a character string.
1247 */
1248static const char *gp_state_getname(short gs)
1249{
1250 if (gs < 0 || gs >= ARRAY_SIZE(gp_state_names))
1251 return "???";
1252 return gp_state_names[gs];
1253}
1254
1255/*
1256 * Complain about starvation of grace-period kthread.
1257 */
1258static void rcu_check_gp_kthread_starvation(struct rcu_state *rsp)
1259{
1260 unsigned long gpa;
1261 unsigned long j;
1262
1263 j = jiffies;
1264 gpa = READ_ONCE(rsp->gp_activity);
1265 if (j - gpa > 2 * HZ) {
1266 pr_err("%s kthread starved for %ld jiffies! g%lu c%lu f%#x %s(%d) ->state=%#lx\n",
1267 rsp->name, j - gpa,
1268 rsp->gpnum, rsp->completed,
1269 rsp->gp_flags,
1270 gp_state_getname(rsp->gp_state), rsp->gp_state,
1271 rsp->gp_kthread ? rsp->gp_kthread->state : ~0);
1272 if (rsp->gp_kthread) {
1273 sched_show_task(rsp->gp_kthread);
1274 wake_up_process(rsp->gp_kthread);
1275 }
1276 }
1277}
1278
1279/*
1280 * Dump stacks of all tasks running on stalled CPUs.
1281 */
1282static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
1283{
1284 int cpu;
1285 unsigned long flags;
1286 struct rcu_node *rnp;
1287
1288 rcu_for_each_leaf_node(rsp, rnp) {
1289 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1290 if (rnp->qsmask != 0) {
1291 for_each_leaf_node_possible_cpu(rnp, cpu)
1292 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu))
1293 dump_cpu_task(cpu);
1294 }
1295 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1296 }
1297}
1298
1299/*
1300 * If too much time has passed in the current grace period, and if
1301 * so configured, go kick the relevant kthreads.
1302 */
1303static void rcu_stall_kick_kthreads(struct rcu_state *rsp)
1304{
1305 unsigned long j;
1306
1307 if (!rcu_kick_kthreads)
1308 return;
1309 j = READ_ONCE(rsp->jiffies_kick_kthreads);
1310 if (time_after(jiffies, j) && rsp->gp_kthread &&
1311 (rcu_gp_in_progress(rsp) || READ_ONCE(rsp->gp_flags))) {
1312 WARN_ONCE(1, "Kicking %s grace-period kthread\n", rsp->name);
1313 rcu_ftrace_dump(DUMP_ALL);
1314 wake_up_process(rsp->gp_kthread);
1315 WRITE_ONCE(rsp->jiffies_kick_kthreads, j + HZ);
1316 }
1317}
1318
1319static inline void panic_on_rcu_stall(void)
1320{
1321 if (sysctl_panic_on_rcu_stall)
1322 panic("RCU Stall\n");
1323}
1324
1325static void print_other_cpu_stall(struct rcu_state *rsp, unsigned long gpnum)
1326{
1327 int cpu;
1328 long delta;
1329 unsigned long flags;
1330 unsigned long gpa;
1331 unsigned long j;
1332 int ndetected = 0;
1333 struct rcu_node *rnp = rcu_get_root(rsp);
1334 long totqlen = 0;
1335
1336 /* Kick and suppress, if so configured. */
1337 rcu_stall_kick_kthreads(rsp);
1338 if (rcu_cpu_stall_suppress)
1339 return;
1340
1341 /* Only let one CPU complain about others per time interval. */
1342
1343 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1344 delta = jiffies - READ_ONCE(rsp->jiffies_stall);
1345 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
1346 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1347 return;
1348 }
1349 WRITE_ONCE(rsp->jiffies_stall,
1350 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1351 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1352
1353 /*
1354 * OK, time to rat on our buddy...
1355 * See Documentation/RCU/stallwarn.txt for info on how to debug
1356 * RCU CPU stall warnings.
1357 */
1358 pr_err("INFO: %s detected stalls on CPUs/tasks:",
1359 rsp->name);
1360 print_cpu_stall_info_begin();
1361 rcu_for_each_leaf_node(rsp, rnp) {
1362 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1363 ndetected += rcu_print_task_stall(rnp);
1364 if (rnp->qsmask != 0) {
1365 for_each_leaf_node_possible_cpu(rnp, cpu)
1366 if (rnp->qsmask & leaf_node_cpu_bit(rnp, cpu)) {
1367 print_cpu_stall_info(rsp, cpu);
1368 ndetected++;
1369 }
1370 }
1371 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1372 }
1373
1374 print_cpu_stall_info_end();
1375 for_each_possible_cpu(cpu)
1376 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1377 pr_cont("(detected by %d, t=%ld jiffies, g=%ld, c=%ld, q=%lu)\n",
1378 smp_processor_id(), (long)(jiffies - rsp->gp_start),
1379 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1380 if (ndetected) {
1381 rcu_dump_cpu_stacks(rsp);
1382 } else {
1383 if (READ_ONCE(rsp->gpnum) != gpnum ||
1384 READ_ONCE(rsp->completed) == gpnum) {
1385 pr_err("INFO: Stall ended before state dump start\n");
1386 } else {
1387 j = jiffies;
1388 gpa = READ_ONCE(rsp->gp_activity);
1389 pr_err("All QSes seen, last %s kthread activity %ld (%ld-%ld), jiffies_till_next_fqs=%ld, root ->qsmask %#lx\n",
1390 rsp->name, j - gpa, j, gpa,
1391 jiffies_till_next_fqs,
1392 rcu_get_root(rsp)->qsmask);
1393 /* In this case, the current CPU might be at fault. */
1394 sched_show_task(current);
1395 }
1396 }
1397
1398 /* Complain about tasks blocking the grace period. */
1399 rcu_print_detail_task_stall(rsp);
1400
1401 rcu_check_gp_kthread_starvation(rsp);
1402
1403 panic_on_rcu_stall();
1404
1405 force_quiescent_state(rsp); /* Kick them all. */
1406}
1407
1408static void print_cpu_stall(struct rcu_state *rsp)
1409{
1410 int cpu;
1411 unsigned long flags;
1412 struct rcu_node *rnp = rcu_get_root(rsp);
1413 long totqlen = 0;
1414
1415 /* Kick and suppress, if so configured. */
1416 rcu_stall_kick_kthreads(rsp);
1417 if (rcu_cpu_stall_suppress)
1418 return;
1419
1420 /*
1421 * OK, time to rat on ourselves...
1422 * See Documentation/RCU/stallwarn.txt for info on how to debug
1423 * RCU CPU stall warnings.
1424 */
1425 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
1426 print_cpu_stall_info_begin();
1427 print_cpu_stall_info(rsp, smp_processor_id());
1428 print_cpu_stall_info_end();
1429 for_each_possible_cpu(cpu)
1430 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
1431 pr_cont(" (t=%lu jiffies g=%ld c=%ld q=%lu)\n",
1432 jiffies - rsp->gp_start,
1433 (long)rsp->gpnum, (long)rsp->completed, totqlen);
1434
1435 rcu_check_gp_kthread_starvation(rsp);
1436
1437 rcu_dump_cpu_stacks(rsp);
1438
1439 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1440 if (ULONG_CMP_GE(jiffies, READ_ONCE(rsp->jiffies_stall)))
1441 WRITE_ONCE(rsp->jiffies_stall,
1442 jiffies + 3 * rcu_jiffies_till_stall_check() + 3);
1443 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1444
1445 panic_on_rcu_stall();
1446
1447 /*
1448 * Attempt to revive the RCU machinery by forcing a context switch.
1449 *
1450 * A context switch would normally allow the RCU state machine to make
1451 * progress and it could be we're stuck in kernel space without context
1452 * switches for an entirely unreasonable amount of time.
1453 */
1454 resched_cpu(smp_processor_id());
1455}
1456
1457static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
1458{
1459 unsigned long completed;
1460 unsigned long gpnum;
1461 unsigned long gps;
1462 unsigned long j;
1463 unsigned long js;
1464 struct rcu_node *rnp;
1465
1466 if ((rcu_cpu_stall_suppress && !rcu_kick_kthreads) ||
1467 !rcu_gp_in_progress(rsp))
1468 return;
1469 rcu_stall_kick_kthreads(rsp);
1470 j = jiffies;
1471
1472 /*
1473 * Lots of memory barriers to reject false positives.
1474 *
1475 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1476 * then rsp->gp_start, and finally rsp->completed. These values
1477 * are updated in the opposite order with memory barriers (or
1478 * equivalent) during grace-period initialization and cleanup.
1479 * Now, a false positive can occur if we get an new value of
1480 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1481 * the memory barriers, the only way that this can happen is if one
1482 * grace period ends and another starts between these two fetches.
1483 * Detect this by comparing rsp->completed with the previous fetch
1484 * from rsp->gpnum.
1485 *
1486 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1487 * and rsp->gp_start suffice to forestall false positives.
1488 */
1489 gpnum = READ_ONCE(rsp->gpnum);
1490 smp_rmb(); /* Pick up ->gpnum first... */
1491 js = READ_ONCE(rsp->jiffies_stall);
1492 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1493 gps = READ_ONCE(rsp->gp_start);
1494 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1495 completed = READ_ONCE(rsp->completed);
1496 if (ULONG_CMP_GE(completed, gpnum) ||
1497 ULONG_CMP_LT(j, js) ||
1498 ULONG_CMP_GE(gps, js))
1499 return; /* No stall or GP completed since entering function. */
1500 rnp = rdp->mynode;
1501 if (rcu_gp_in_progress(rsp) &&
1502 (READ_ONCE(rnp->qsmask) & rdp->grpmask)) {
1503
1504 /* We haven't checked in, so go dump stack. */
1505 print_cpu_stall(rsp);
1506
1507 } else if (rcu_gp_in_progress(rsp) &&
1508 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1509
1510 /* They had a few time units to dump stack, so complain. */
1511 print_other_cpu_stall(rsp, gpnum);
1512 }
1513}
1514
1515/**
1516 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1517 *
1518 * Set the stall-warning timeout way off into the future, thus preventing
1519 * any RCU CPU stall-warning messages from appearing in the current set of
1520 * RCU grace periods.
1521 *
1522 * The caller must disable hard irqs.
1523 */
1524void rcu_cpu_stall_reset(void)
1525{
1526 struct rcu_state *rsp;
1527
1528 for_each_rcu_flavor(rsp)
1529 WRITE_ONCE(rsp->jiffies_stall, jiffies + ULONG_MAX / 2);
1530}
1531
1532/*
1533 * Initialize the specified rcu_data structure's default callback list
1534 * to empty. The default callback list is the one that is not used by
1535 * no-callbacks CPUs.
1536 */
1537static void init_default_callback_list(struct rcu_data *rdp)
1538{
1539 int i;
1540
1541 rdp->nxtlist = NULL;
1542 for (i = 0; i < RCU_NEXT_SIZE; i++)
1543 rdp->nxttail[i] = &rdp->nxtlist;
1544}
1545
1546/*
1547 * Initialize the specified rcu_data structure's callback list to empty.
1548 */
1549static void init_callback_list(struct rcu_data *rdp)
1550{
1551 if (init_nocb_callback_list(rdp))
1552 return;
1553 init_default_callback_list(rdp);
1554}
1555
1556/*
1557 * Determine the value that ->completed will have at the end of the
1558 * next subsequent grace period. This is used to tag callbacks so that
1559 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1560 * been dyntick-idle for an extended period with callbacks under the
1561 * influence of RCU_FAST_NO_HZ.
1562 *
1563 * The caller must hold rnp->lock with interrupts disabled.
1564 */
1565static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1566 struct rcu_node *rnp)
1567{
1568 /*
1569 * If RCU is idle, we just wait for the next grace period.
1570 * But we can only be sure that RCU is idle if we are looking
1571 * at the root rcu_node structure -- otherwise, a new grace
1572 * period might have started, but just not yet gotten around
1573 * to initializing the current non-root rcu_node structure.
1574 */
1575 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1576 return rnp->completed + 1;
1577
1578 /*
1579 * Otherwise, wait for a possible partial grace period and
1580 * then the subsequent full grace period.
1581 */
1582 return rnp->completed + 2;
1583}
1584
1585/*
1586 * Trace-event helper function for rcu_start_future_gp() and
1587 * rcu_nocb_wait_gp().
1588 */
1589static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1590 unsigned long c, const char *s)
1591{
1592 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1593 rnp->completed, c, rnp->level,
1594 rnp->grplo, rnp->grphi, s);
1595}
1596
1597/*
1598 * Start some future grace period, as needed to handle newly arrived
1599 * callbacks. The required future grace periods are recorded in each
1600 * rcu_node structure's ->need_future_gp field. Returns true if there
1601 * is reason to awaken the grace-period kthread.
1602 *
1603 * The caller must hold the specified rcu_node structure's ->lock.
1604 */
1605static bool __maybe_unused
1606rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1607 unsigned long *c_out)
1608{
1609 unsigned long c;
1610 int i;
1611 bool ret = false;
1612 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1613
1614 /*
1615 * Pick up grace-period number for new callbacks. If this
1616 * grace period is already marked as needed, return to the caller.
1617 */
1618 c = rcu_cbs_completed(rdp->rsp, rnp);
1619 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1620 if (rnp->need_future_gp[c & 0x1]) {
1621 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1622 goto out;
1623 }
1624
1625 /*
1626 * If either this rcu_node structure or the root rcu_node structure
1627 * believe that a grace period is in progress, then we must wait
1628 * for the one following, which is in "c". Because our request
1629 * will be noticed at the end of the current grace period, we don't
1630 * need to explicitly start one. We only do the lockless check
1631 * of rnp_root's fields if the current rcu_node structure thinks
1632 * there is no grace period in flight, and because we hold rnp->lock,
1633 * the only possible change is when rnp_root's two fields are
1634 * equal, in which case rnp_root->gpnum might be concurrently
1635 * incremented. But that is OK, as it will just result in our
1636 * doing some extra useless work.
1637 */
1638 if (rnp->gpnum != rnp->completed ||
1639 READ_ONCE(rnp_root->gpnum) != READ_ONCE(rnp_root->completed)) {
1640 rnp->need_future_gp[c & 0x1]++;
1641 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1642 goto out;
1643 }
1644
1645 /*
1646 * There might be no grace period in progress. If we don't already
1647 * hold it, acquire the root rcu_node structure's lock in order to
1648 * start one (if needed).
1649 */
1650 if (rnp != rnp_root)
1651 raw_spin_lock_rcu_node(rnp_root);
1652
1653 /*
1654 * Get a new grace-period number. If there really is no grace
1655 * period in progress, it will be smaller than the one we obtained
1656 * earlier. Adjust callbacks as needed. Note that even no-CBs
1657 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1658 */
1659 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1660 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1661 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1662 rdp->nxtcompleted[i] = c;
1663
1664 /*
1665 * If the needed for the required grace period is already
1666 * recorded, trace and leave.
1667 */
1668 if (rnp_root->need_future_gp[c & 0x1]) {
1669 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1670 goto unlock_out;
1671 }
1672
1673 /* Record the need for the future grace period. */
1674 rnp_root->need_future_gp[c & 0x1]++;
1675
1676 /* If a grace period is not already in progress, start one. */
1677 if (rnp_root->gpnum != rnp_root->completed) {
1678 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1679 } else {
1680 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1681 ret = rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1682 }
1683unlock_out:
1684 if (rnp != rnp_root)
1685 raw_spin_unlock_rcu_node(rnp_root);
1686out:
1687 if (c_out != NULL)
1688 *c_out = c;
1689 return ret;
1690}
1691
1692/*
1693 * Clean up any old requests for the just-ended grace period. Also return
1694 * whether any additional grace periods have been requested. Also invoke
1695 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1696 * waiting for this grace period to complete.
1697 */
1698static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1699{
1700 int c = rnp->completed;
1701 int needmore;
1702 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1703
1704 rnp->need_future_gp[c & 0x1] = 0;
1705 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1706 trace_rcu_future_gp(rnp, rdp, c,
1707 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1708 return needmore;
1709}
1710
1711/*
1712 * Awaken the grace-period kthread for the specified flavor of RCU.
1713 * Don't do a self-awaken, and don't bother awakening when there is
1714 * nothing for the grace-period kthread to do (as in several CPUs
1715 * raced to awaken, and we lost), and finally don't try to awaken
1716 * a kthread that has not yet been created.
1717 */
1718static void rcu_gp_kthread_wake(struct rcu_state *rsp)
1719{
1720 if (current == rsp->gp_kthread ||
1721 !READ_ONCE(rsp->gp_flags) ||
1722 !rsp->gp_kthread)
1723 return;
1724 swake_up(&rsp->gp_wq);
1725}
1726
1727/*
1728 * If there is room, assign a ->completed number to any callbacks on
1729 * this CPU that have not already been assigned. Also accelerate any
1730 * callbacks that were previously assigned a ->completed number that has
1731 * since proven to be too conservative, which can happen if callbacks get
1732 * assigned a ->completed number while RCU is idle, but with reference to
1733 * a non-root rcu_node structure. This function is idempotent, so it does
1734 * not hurt to call it repeatedly. Returns an flag saying that we should
1735 * awaken the RCU grace-period kthread.
1736 *
1737 * The caller must hold rnp->lock with interrupts disabled.
1738 */
1739static bool rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1740 struct rcu_data *rdp)
1741{
1742 unsigned long c;
1743 int i;
1744 bool ret;
1745
1746 /* If the CPU has no callbacks, nothing to do. */
1747 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1748 return false;
1749
1750 /*
1751 * Starting from the sublist containing the callbacks most
1752 * recently assigned a ->completed number and working down, find the
1753 * first sublist that is not assignable to an upcoming grace period.
1754 * Such a sublist has something in it (first two tests) and has
1755 * a ->completed number assigned that will complete sooner than
1756 * the ->completed number for newly arrived callbacks (last test).
1757 *
1758 * The key point is that any later sublist can be assigned the
1759 * same ->completed number as the newly arrived callbacks, which
1760 * means that the callbacks in any of these later sublist can be
1761 * grouped into a single sublist, whether or not they have already
1762 * been assigned a ->completed number.
1763 */
1764 c = rcu_cbs_completed(rsp, rnp);
1765 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1766 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1767 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1768 break;
1769
1770 /*
1771 * If there are no sublist for unassigned callbacks, leave.
1772 * At the same time, advance "i" one sublist, so that "i" will
1773 * index into the sublist where all the remaining callbacks should
1774 * be grouped into.
1775 */
1776 if (++i >= RCU_NEXT_TAIL)
1777 return false;
1778
1779 /*
1780 * Assign all subsequent callbacks' ->completed number to the next
1781 * full grace period and group them all in the sublist initially
1782 * indexed by "i".
1783 */
1784 for (; i <= RCU_NEXT_TAIL; i++) {
1785 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1786 rdp->nxtcompleted[i] = c;
1787 }
1788 /* Record any needed additional grace periods. */
1789 ret = rcu_start_future_gp(rnp, rdp, NULL);
1790
1791 /* Trace depending on how much we were able to accelerate. */
1792 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1793 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1794 else
1795 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1796 return ret;
1797}
1798
1799/*
1800 * Move any callbacks whose grace period has completed to the
1801 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1802 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1803 * sublist. This function is idempotent, so it does not hurt to
1804 * invoke it repeatedly. As long as it is not invoked -too- often...
1805 * Returns true if the RCU grace-period kthread needs to be awakened.
1806 *
1807 * The caller must hold rnp->lock with interrupts disabled.
1808 */
1809static bool rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1810 struct rcu_data *rdp)
1811{
1812 int i, j;
1813
1814 /* If the CPU has no callbacks, nothing to do. */
1815 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1816 return false;
1817
1818 /*
1819 * Find all callbacks whose ->completed numbers indicate that they
1820 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1821 */
1822 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1823 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1824 break;
1825 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1826 }
1827 /* Clean up any sublist tail pointers that were misordered above. */
1828 for (j = RCU_WAIT_TAIL; j < i; j++)
1829 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1830
1831 /* Copy down callbacks to fill in empty sublists. */
1832 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1833 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1834 break;
1835 rdp->nxttail[j] = rdp->nxttail[i];
1836 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1837 }
1838
1839 /* Classify any remaining callbacks. */
1840 return rcu_accelerate_cbs(rsp, rnp, rdp);
1841}
1842
1843/*
1844 * Update CPU-local rcu_data state to record the beginnings and ends of
1845 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1846 * structure corresponding to the current CPU, and must have irqs disabled.
1847 * Returns true if the grace-period kthread needs to be awakened.
1848 */
1849static bool __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp,
1850 struct rcu_data *rdp)
1851{
1852 bool ret;
1853 bool need_gp;
1854
1855 /* Handle the ends of any preceding grace periods first. */
1856 if (rdp->completed == rnp->completed &&
1857 !unlikely(READ_ONCE(rdp->gpwrap))) {
1858
1859 /* No grace period end, so just accelerate recent callbacks. */
1860 ret = rcu_accelerate_cbs(rsp, rnp, rdp);
1861
1862 } else {
1863
1864 /* Advance callbacks. */
1865 ret = rcu_advance_cbs(rsp, rnp, rdp);
1866
1867 /* Remember that we saw this grace-period completion. */
1868 rdp->completed = rnp->completed;
1869 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1870 }
1871
1872 if (rdp->gpnum != rnp->gpnum || unlikely(READ_ONCE(rdp->gpwrap))) {
1873 /*
1874 * If the current grace period is waiting for this CPU,
1875 * set up to detect a quiescent state, otherwise don't
1876 * go looking for one.
1877 */
1878 rdp->gpnum = rnp->gpnum;
1879 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1880 need_gp = !!(rnp->qsmask & rdp->grpmask);
1881 rdp->cpu_no_qs.b.norm = need_gp;
1882 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
1883 rdp->core_needs_qs = need_gp;
1884 zero_cpu_stall_ticks(rdp);
1885 WRITE_ONCE(rdp->gpwrap, false);
1886 }
1887 return ret;
1888}
1889
1890static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1891{
1892 unsigned long flags;
1893 bool needwake;
1894 struct rcu_node *rnp;
1895
1896 local_irq_save(flags);
1897 rnp = rdp->mynode;
1898 if ((rdp->gpnum == READ_ONCE(rnp->gpnum) &&
1899 rdp->completed == READ_ONCE(rnp->completed) &&
1900 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1901 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1902 local_irq_restore(flags);
1903 return;
1904 }
1905 needwake = __note_gp_changes(rsp, rnp, rdp);
1906 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1907 if (needwake)
1908 rcu_gp_kthread_wake(rsp);
1909}
1910
1911static void rcu_gp_slow(struct rcu_state *rsp, int delay)
1912{
1913 if (delay > 0 &&
1914 !(rsp->gpnum % (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1915 schedule_timeout_uninterruptible(delay);
1916}
1917
1918/*
1919 * Initialize a new grace period. Return false if no grace period required.
1920 */
1921static bool rcu_gp_init(struct rcu_state *rsp)
1922{
1923 unsigned long oldmask;
1924 struct rcu_data *rdp;
1925 struct rcu_node *rnp = rcu_get_root(rsp);
1926
1927 WRITE_ONCE(rsp->gp_activity, jiffies);
1928 raw_spin_lock_irq_rcu_node(rnp);
1929 if (!READ_ONCE(rsp->gp_flags)) {
1930 /* Spurious wakeup, tell caller to go back to sleep. */
1931 raw_spin_unlock_irq_rcu_node(rnp);
1932 return false;
1933 }
1934 WRITE_ONCE(rsp->gp_flags, 0); /* Clear all flags: New grace period. */
1935
1936 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1937 /*
1938 * Grace period already in progress, don't start another.
1939 * Not supposed to be able to happen.
1940 */
1941 raw_spin_unlock_irq_rcu_node(rnp);
1942 return false;
1943 }
1944
1945 /* Advance to a new grace period and initialize state. */
1946 record_gp_stall_check_time(rsp);
1947 /* Record GP times before starting GP, hence smp_store_release(). */
1948 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1949 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1950 raw_spin_unlock_irq_rcu_node(rnp);
1951
1952 /*
1953 * Apply per-leaf buffered online and offline operations to the
1954 * rcu_node tree. Note that this new grace period need not wait
1955 * for subsequent online CPUs, and that quiescent-state forcing
1956 * will handle subsequent offline CPUs.
1957 */
1958 rcu_for_each_leaf_node(rsp, rnp) {
1959 rcu_gp_slow(rsp, gp_preinit_delay);
1960 raw_spin_lock_irq_rcu_node(rnp);
1961 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1962 !rnp->wait_blkd_tasks) {
1963 /* Nothing to do on this leaf rcu_node structure. */
1964 raw_spin_unlock_irq_rcu_node(rnp);
1965 continue;
1966 }
1967
1968 /* Record old state, apply changes to ->qsmaskinit field. */
1969 oldmask = rnp->qsmaskinit;
1970 rnp->qsmaskinit = rnp->qsmaskinitnext;
1971
1972 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1973 if (!oldmask != !rnp->qsmaskinit) {
1974 if (!oldmask) /* First online CPU for this rcu_node. */
1975 rcu_init_new_rnp(rnp);
1976 else if (rcu_preempt_has_tasks(rnp)) /* blocked tasks */
1977 rnp->wait_blkd_tasks = true;
1978 else /* Last offline CPU and can propagate. */
1979 rcu_cleanup_dead_rnp(rnp);
1980 }
1981
1982 /*
1983 * If all waited-on tasks from prior grace period are
1984 * done, and if all this rcu_node structure's CPUs are
1985 * still offline, propagate up the rcu_node tree and
1986 * clear ->wait_blkd_tasks. Otherwise, if one of this
1987 * rcu_node structure's CPUs has since come back online,
1988 * simply clear ->wait_blkd_tasks (but rcu_cleanup_dead_rnp()
1989 * checks for this, so just call it unconditionally).
1990 */
1991 if (rnp->wait_blkd_tasks &&
1992 (!rcu_preempt_has_tasks(rnp) ||
1993 rnp->qsmaskinit)) {
1994 rnp->wait_blkd_tasks = false;
1995 rcu_cleanup_dead_rnp(rnp);
1996 }
1997
1998 raw_spin_unlock_irq_rcu_node(rnp);
1999 }
2000
2001 /*
2002 * Set the quiescent-state-needed bits in all the rcu_node
2003 * structures for all currently online CPUs in breadth-first order,
2004 * starting from the root rcu_node structure, relying on the layout
2005 * of the tree within the rsp->node[] array. Note that other CPUs
2006 * will access only the leaves of the hierarchy, thus seeing that no
2007 * grace period is in progress, at least until the corresponding
2008 * leaf node has been initialized.
2009 *
2010 * The grace period cannot complete until the initialization
2011 * process finishes, because this kthread handles both.
2012 */
2013 rcu_for_each_node_breadth_first(rsp, rnp) {
2014 rcu_gp_slow(rsp, gp_init_delay);
2015 raw_spin_lock_irq_rcu_node(rnp);
2016 rdp = this_cpu_ptr(rsp->rda);
2017 rcu_preempt_check_blocked_tasks(rnp);
2018 rnp->qsmask = rnp->qsmaskinit;
2019 WRITE_ONCE(rnp->gpnum, rsp->gpnum);
2020 if (WARN_ON_ONCE(rnp->completed != rsp->completed))
2021 WRITE_ONCE(rnp->completed, rsp->completed);
2022 if (rnp == rdp->mynode)
2023 (void)__note_gp_changes(rsp, rnp, rdp);
2024 rcu_preempt_boost_start_gp(rnp);
2025 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
2026 rnp->level, rnp->grplo,
2027 rnp->grphi, rnp->qsmask);
2028 raw_spin_unlock_irq_rcu_node(rnp);
2029 cond_resched_rcu_qs();
2030 WRITE_ONCE(rsp->gp_activity, jiffies);
2031 }
2032
2033 return true;
2034}
2035
2036/*
2037 * Helper function for wait_event_interruptible_timeout() wakeup
2038 * at force-quiescent-state time.
2039 */
2040static bool rcu_gp_fqs_check_wake(struct rcu_state *rsp, int *gfp)
2041{
2042 struct rcu_node *rnp = rcu_get_root(rsp);
2043
2044 /* Someone like call_rcu() requested a force-quiescent-state scan. */
2045 *gfp = READ_ONCE(rsp->gp_flags);
2046 if (*gfp & RCU_GP_FLAG_FQS)
2047 return true;
2048
2049 /* The current grace period has completed. */
2050 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
2051 return true;
2052
2053 return false;
2054}
2055
2056/*
2057 * Do one round of quiescent-state forcing.
2058 */
2059static void rcu_gp_fqs(struct rcu_state *rsp, bool first_time)
2060{
2061 bool isidle = false;
2062 unsigned long maxj;
2063 struct rcu_node *rnp = rcu_get_root(rsp);
2064
2065 WRITE_ONCE(rsp->gp_activity, jiffies);
2066 rsp->n_force_qs++;
2067 if (first_time) {
2068 /* Collect dyntick-idle snapshots. */
2069 if (is_sysidle_rcu_state(rsp)) {
2070 isidle = true;
2071 maxj = jiffies - ULONG_MAX / 4;
2072 }
2073 force_qs_rnp(rsp, dyntick_save_progress_counter,
2074 &isidle, &maxj);
2075 rcu_sysidle_report_gp(rsp, isidle, maxj);
2076 } else {
2077 /* Handle dyntick-idle and offline CPUs. */
2078 isidle = true;
2079 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
2080 }
2081 /* Clear flag to prevent immediate re-entry. */
2082 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2083 raw_spin_lock_irq_rcu_node(rnp);
2084 WRITE_ONCE(rsp->gp_flags,
2085 READ_ONCE(rsp->gp_flags) & ~RCU_GP_FLAG_FQS);
2086 raw_spin_unlock_irq_rcu_node(rnp);
2087 }
2088}
2089
2090/*
2091 * Clean up after the old grace period.
2092 */
2093static void rcu_gp_cleanup(struct rcu_state *rsp)
2094{
2095 unsigned long gp_duration;
2096 bool needgp = false;
2097 int nocb = 0;
2098 struct rcu_data *rdp;
2099 struct rcu_node *rnp = rcu_get_root(rsp);
2100 struct swait_queue_head *sq;
2101
2102 WRITE_ONCE(rsp->gp_activity, jiffies);
2103 raw_spin_lock_irq_rcu_node(rnp);
2104 gp_duration = jiffies - rsp->gp_start;
2105 if (gp_duration > rsp->gp_max)
2106 rsp->gp_max = gp_duration;
2107
2108 /*
2109 * We know the grace period is complete, but to everyone else
2110 * it appears to still be ongoing. But it is also the case
2111 * that to everyone else it looks like there is nothing that
2112 * they can do to advance the grace period. It is therefore
2113 * safe for us to drop the lock in order to mark the grace
2114 * period as completed in all of the rcu_node structures.
2115 */
2116 raw_spin_unlock_irq_rcu_node(rnp);
2117
2118 /*
2119 * Propagate new ->completed value to rcu_node structures so
2120 * that other CPUs don't have to wait until the start of the next
2121 * grace period to process their callbacks. This also avoids
2122 * some nasty RCU grace-period initialization races by forcing
2123 * the end of the current grace period to be completely recorded in
2124 * all of the rcu_node structures before the beginning of the next
2125 * grace period is recorded in any of the rcu_node structures.
2126 */
2127 rcu_for_each_node_breadth_first(rsp, rnp) {
2128 raw_spin_lock_irq_rcu_node(rnp);
2129 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
2130 WARN_ON_ONCE(rnp->qsmask);
2131 WRITE_ONCE(rnp->completed, rsp->gpnum);
2132 rdp = this_cpu_ptr(rsp->rda);
2133 if (rnp == rdp->mynode)
2134 needgp = __note_gp_changes(rsp, rnp, rdp) || needgp;
2135 /* smp_mb() provided by prior unlock-lock pair. */
2136 nocb += rcu_future_gp_cleanup(rsp, rnp);
2137 sq = rcu_nocb_gp_get(rnp);
2138 raw_spin_unlock_irq_rcu_node(rnp);
2139 rcu_nocb_gp_cleanup(sq);
2140 cond_resched_rcu_qs();
2141 WRITE_ONCE(rsp->gp_activity, jiffies);
2142 rcu_gp_slow(rsp, gp_cleanup_delay);
2143 }
2144 rnp = rcu_get_root(rsp);
2145 raw_spin_lock_irq_rcu_node(rnp); /* Order GP before ->completed update. */
2146 rcu_nocb_gp_set(rnp, nocb);
2147
2148 /* Declare grace period done. */
2149 WRITE_ONCE(rsp->completed, rsp->gpnum);
2150 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
2151 rsp->gp_state = RCU_GP_IDLE;
2152 rdp = this_cpu_ptr(rsp->rda);
2153 /* Advance CBs to reduce false positives below. */
2154 needgp = rcu_advance_cbs(rsp, rnp, rdp) || needgp;
2155 if (needgp || cpu_needs_another_gp(rsp, rdp)) {
2156 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2157 trace_rcu_grace_period(rsp->name,
2158 READ_ONCE(rsp->gpnum),
2159 TPS("newreq"));
2160 }
2161 raw_spin_unlock_irq_rcu_node(rnp);
2162}
2163
2164/*
2165 * Body of kthread that handles grace periods.
2166 */
2167static int __noreturn rcu_gp_kthread(void *arg)
2168{
2169 bool first_gp_fqs;
2170 int gf;
2171 unsigned long j;
2172 int ret;
2173 struct rcu_state *rsp = arg;
2174 struct rcu_node *rnp = rcu_get_root(rsp);
2175
2176 rcu_bind_gp_kthread();
2177 for (;;) {
2178
2179 /* Handle grace-period start. */
2180 for (;;) {
2181 trace_rcu_grace_period(rsp->name,
2182 READ_ONCE(rsp->gpnum),
2183 TPS("reqwait"));
2184 rsp->gp_state = RCU_GP_WAIT_GPS;
2185 swait_event_interruptible(rsp->gp_wq,
2186 READ_ONCE(rsp->gp_flags) &
2187 RCU_GP_FLAG_INIT);
2188 rsp->gp_state = RCU_GP_DONE_GPS;
2189 /* Locking provides needed memory barrier. */
2190 if (rcu_gp_init(rsp))
2191 break;
2192 cond_resched_rcu_qs();
2193 WRITE_ONCE(rsp->gp_activity, jiffies);
2194 WARN_ON(signal_pending(current));
2195 trace_rcu_grace_period(rsp->name,
2196 READ_ONCE(rsp->gpnum),
2197 TPS("reqwaitsig"));
2198 }
2199
2200 /* Handle quiescent-state forcing. */
2201 first_gp_fqs = true;
2202 j = jiffies_till_first_fqs;
2203 if (j > HZ) {
2204 j = HZ;
2205 jiffies_till_first_fqs = HZ;
2206 }
2207 ret = 0;
2208 for (;;) {
2209 if (!ret) {
2210 rsp->jiffies_force_qs = jiffies + j;
2211 WRITE_ONCE(rsp->jiffies_kick_kthreads,
2212 jiffies + 3 * j);
2213 }
2214 trace_rcu_grace_period(rsp->name,
2215 READ_ONCE(rsp->gpnum),
2216 TPS("fqswait"));
2217 rsp->gp_state = RCU_GP_WAIT_FQS;
2218 ret = swait_event_interruptible_timeout(rsp->gp_wq,
2219 rcu_gp_fqs_check_wake(rsp, &gf), j);
2220 rsp->gp_state = RCU_GP_DOING_FQS;
2221 /* Locking provides needed memory barriers. */
2222 /* If grace period done, leave loop. */
2223 if (!READ_ONCE(rnp->qsmask) &&
2224 !rcu_preempt_blocked_readers_cgp(rnp))
2225 break;
2226 /* If time for quiescent-state forcing, do it. */
2227 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
2228 (gf & RCU_GP_FLAG_FQS)) {
2229 trace_rcu_grace_period(rsp->name,
2230 READ_ONCE(rsp->gpnum),
2231 TPS("fqsstart"));
2232 rcu_gp_fqs(rsp, first_gp_fqs);
2233 first_gp_fqs = false;
2234 trace_rcu_grace_period(rsp->name,
2235 READ_ONCE(rsp->gpnum),
2236 TPS("fqsend"));
2237 cond_resched_rcu_qs();
2238 WRITE_ONCE(rsp->gp_activity, jiffies);
2239 ret = 0; /* Force full wait till next FQS. */
2240 j = jiffies_till_next_fqs;
2241 if (j > HZ) {
2242 j = HZ;
2243 jiffies_till_next_fqs = HZ;
2244 } else if (j < 1) {
2245 j = 1;
2246 jiffies_till_next_fqs = 1;
2247 }
2248 } else {
2249 /* Deal with stray signal. */
2250 cond_resched_rcu_qs();
2251 WRITE_ONCE(rsp->gp_activity, jiffies);
2252 WARN_ON(signal_pending(current));
2253 trace_rcu_grace_period(rsp->name,
2254 READ_ONCE(rsp->gpnum),
2255 TPS("fqswaitsig"));
2256 ret = 1; /* Keep old FQS timing. */
2257 j = jiffies;
2258 if (time_after(jiffies, rsp->jiffies_force_qs))
2259 j = 1;
2260 else
2261 j = rsp->jiffies_force_qs - j;
2262 }
2263 }
2264
2265 /* Handle grace-period end. */
2266 rsp->gp_state = RCU_GP_CLEANUP;
2267 rcu_gp_cleanup(rsp);
2268 rsp->gp_state = RCU_GP_CLEANED;
2269 }
2270}
2271
2272/*
2273 * Start a new RCU grace period if warranted, re-initializing the hierarchy
2274 * in preparation for detecting the next grace period. The caller must hold
2275 * the root node's ->lock and hard irqs must be disabled.
2276 *
2277 * Note that it is legal for a dying CPU (which is marked as offline) to
2278 * invoke this function. This can happen when the dying CPU reports its
2279 * quiescent state.
2280 *
2281 * Returns true if the grace-period kthread must be awakened.
2282 */
2283static bool
2284rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
2285 struct rcu_data *rdp)
2286{
2287 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
2288 /*
2289 * Either we have not yet spawned the grace-period
2290 * task, this CPU does not need another grace period,
2291 * or a grace period is already in progress.
2292 * Either way, don't start a new grace period.
2293 */
2294 return false;
2295 }
2296 WRITE_ONCE(rsp->gp_flags, RCU_GP_FLAG_INIT);
2297 trace_rcu_grace_period(rsp->name, READ_ONCE(rsp->gpnum),
2298 TPS("newreq"));
2299
2300 /*
2301 * We can't do wakeups while holding the rnp->lock, as that
2302 * could cause possible deadlocks with the rq->lock. Defer
2303 * the wakeup to our caller.
2304 */
2305 return true;
2306}
2307
2308/*
2309 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
2310 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
2311 * is invoked indirectly from rcu_advance_cbs(), which would result in
2312 * endless recursion -- or would do so if it wasn't for the self-deadlock
2313 * that is encountered beforehand.
2314 *
2315 * Returns true if the grace-period kthread needs to be awakened.
2316 */
2317static bool rcu_start_gp(struct rcu_state *rsp)
2318{
2319 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
2320 struct rcu_node *rnp = rcu_get_root(rsp);
2321 bool ret = false;
2322
2323 /*
2324 * If there is no grace period in progress right now, any
2325 * callbacks we have up to this point will be satisfied by the
2326 * next grace period. Also, advancing the callbacks reduces the
2327 * probability of false positives from cpu_needs_another_gp()
2328 * resulting in pointless grace periods. So, advance callbacks
2329 * then start the grace period!
2330 */
2331 ret = rcu_advance_cbs(rsp, rnp, rdp) || ret;
2332 ret = rcu_start_gp_advanced(rsp, rnp, rdp) || ret;
2333 return ret;
2334}
2335
2336/*
2337 * Report a full set of quiescent states to the specified rcu_state data
2338 * structure. Invoke rcu_gp_kthread_wake() to awaken the grace-period
2339 * kthread if another grace period is required. Whether we wake
2340 * the grace-period kthread or it awakens itself for the next round
2341 * of quiescent-state forcing, that kthread will clean up after the
2342 * just-completed grace period. Note that the caller must hold rnp->lock,
2343 * which is released before return.
2344 */
2345static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
2346 __releases(rcu_get_root(rsp)->lock)
2347{
2348 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
2349 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2350 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
2351 rcu_gp_kthread_wake(rsp);
2352}
2353
2354/*
2355 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2356 * Allows quiescent states for a group of CPUs to be reported at one go
2357 * to the specified rcu_node structure, though all the CPUs in the group
2358 * must be represented by the same rcu_node structure (which need not be a
2359 * leaf rcu_node structure, though it often will be). The gps parameter
2360 * is the grace-period snapshot, which means that the quiescent states
2361 * are valid only if rnp->gpnum is equal to gps. That structure's lock
2362 * must be held upon entry, and it is released before return.
2363 */
2364static void
2365rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
2366 struct rcu_node *rnp, unsigned long gps, unsigned long flags)
2367 __releases(rnp->lock)
2368{
2369 unsigned long oldmask = 0;
2370 struct rcu_node *rnp_c;
2371
2372 /* Walk up the rcu_node hierarchy. */
2373 for (;;) {
2374 if (!(rnp->qsmask & mask) || rnp->gpnum != gps) {
2375
2376 /*
2377 * Our bit has already been cleared, or the
2378 * relevant grace period is already over, so done.
2379 */
2380 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2381 return;
2382 }
2383 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2384 rnp->qsmask &= ~mask;
2385 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
2386 mask, rnp->qsmask, rnp->level,
2387 rnp->grplo, rnp->grphi,
2388 !!rnp->gp_tasks);
2389 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2390
2391 /* Other bits still set at this level, so done. */
2392 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2393 return;
2394 }
2395 mask = rnp->grpmask;
2396 if (rnp->parent == NULL) {
2397
2398 /* No more levels. Exit loop holding root lock. */
2399
2400 break;
2401 }
2402 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2403 rnp_c = rnp;
2404 rnp = rnp->parent;
2405 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2406 oldmask = rnp_c->qsmask;
2407 }
2408
2409 /*
2410 * Get here if we are the last CPU to pass through a quiescent
2411 * state for this grace period. Invoke rcu_report_qs_rsp()
2412 * to clean up and start the next grace period if one is needed.
2413 */
2414 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
2415}
2416
2417/*
2418 * Record a quiescent state for all tasks that were previously queued
2419 * on the specified rcu_node structure and that were blocking the current
2420 * RCU grace period. The caller must hold the specified rnp->lock with
2421 * irqs disabled, and this lock is released upon return, but irqs remain
2422 * disabled.
2423 */
2424static void rcu_report_unblock_qs_rnp(struct rcu_state *rsp,
2425 struct rcu_node *rnp, unsigned long flags)
2426 __releases(rnp->lock)
2427{
2428 unsigned long gps;
2429 unsigned long mask;
2430 struct rcu_node *rnp_p;
2431
2432 if (rcu_state_p == &rcu_sched_state || rsp != rcu_state_p ||
2433 rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2434 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2435 return; /* Still need more quiescent states! */
2436 }
2437
2438 rnp_p = rnp->parent;
2439 if (rnp_p == NULL) {
2440 /*
2441 * Only one rcu_node structure in the tree, so don't
2442 * try to report up to its nonexistent parent!
2443 */
2444 rcu_report_qs_rsp(rsp, flags);
2445 return;
2446 }
2447
2448 /* Report up the rest of the hierarchy, tracking current ->gpnum. */
2449 gps = rnp->gpnum;
2450 mask = rnp->grpmask;
2451 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2452 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2453 rcu_report_qs_rnp(mask, rsp, rnp_p, gps, flags);
2454}
2455
2456/*
2457 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2458 * structure. This must be called from the specified CPU.
2459 */
2460static void
2461rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
2462{
2463 unsigned long flags;
2464 unsigned long mask;
2465 bool needwake;
2466 struct rcu_node *rnp;
2467
2468 rnp = rdp->mynode;
2469 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2470 if ((rdp->cpu_no_qs.b.norm &&
2471 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) ||
2472 rdp->gpnum != rnp->gpnum || rnp->completed == rnp->gpnum ||
2473 rdp->gpwrap) {
2474
2475 /*
2476 * The grace period in which this quiescent state was
2477 * recorded has ended, so don't report it upwards.
2478 * We will instead need a new quiescent state that lies
2479 * within the current grace period.
2480 */
2481 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2482 rdp->rcu_qs_ctr_snap = __this_cpu_read(rcu_qs_ctr);
2483 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2484 return;
2485 }
2486 mask = rdp->grpmask;
2487 if ((rnp->qsmask & mask) == 0) {
2488 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2489 } else {
2490 rdp->core_needs_qs = false;
2491
2492 /*
2493 * This GP can't end until cpu checks in, so all of our
2494 * callbacks can be processed during the next GP.
2495 */
2496 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
2497
2498 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2499 /* ^^^ Released rnp->lock */
2500 if (needwake)
2501 rcu_gp_kthread_wake(rsp);
2502 }
2503}
2504
2505/*
2506 * Check to see if there is a new grace period of which this CPU
2507 * is not yet aware, and if so, set up local rcu_data state for it.
2508 * Otherwise, see if this CPU has just passed through its first
2509 * quiescent state for this grace period, and record that fact if so.
2510 */
2511static void
2512rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
2513{
2514 /* Check for grace-period ends and beginnings. */
2515 note_gp_changes(rsp, rdp);
2516
2517 /*
2518 * Does this CPU still need to do its part for current grace period?
2519 * If no, return and let the other CPUs do their part as well.
2520 */
2521 if (!rdp->core_needs_qs)
2522 return;
2523
2524 /*
2525 * Was there a quiescent state since the beginning of the grace
2526 * period? If no, then exit and wait for the next call.
2527 */
2528 if (rdp->cpu_no_qs.b.norm &&
2529 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr))
2530 return;
2531
2532 /*
2533 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2534 * judge of that).
2535 */
2536 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
2537}
2538
2539/*
2540 * Send the specified CPU's RCU callbacks to the orphanage. The
2541 * specified CPU must be offline, and the caller must hold the
2542 * ->orphan_lock.
2543 */
2544static void
2545rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
2546 struct rcu_node *rnp, struct rcu_data *rdp)
2547{
2548 /* No-CBs CPUs do not have orphanable callbacks. */
2549 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) || rcu_is_nocb_cpu(rdp->cpu))
2550 return;
2551
2552 /*
2553 * Orphan the callbacks. First adjust the counts. This is safe
2554 * because _rcu_barrier() excludes CPU-hotplug operations, so it
2555 * cannot be running now. Thus no memory barrier is required.
2556 */
2557 if (rdp->nxtlist != NULL) {
2558 rsp->qlen_lazy += rdp->qlen_lazy;
2559 rsp->qlen += rdp->qlen;
2560 rdp->n_cbs_orphaned += rdp->qlen;
2561 rdp->qlen_lazy = 0;
2562 WRITE_ONCE(rdp->qlen, 0);
2563 }
2564
2565 /*
2566 * Next, move those callbacks still needing a grace period to
2567 * the orphanage, where some other CPU will pick them up.
2568 * Some of the callbacks might have gone partway through a grace
2569 * period, but that is too bad. They get to start over because we
2570 * cannot assume that grace periods are synchronized across CPUs.
2571 * We don't bother updating the ->nxttail[] array yet, instead
2572 * we just reset the whole thing later on.
2573 */
2574 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
2575 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
2576 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
2577 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2578 }
2579
2580 /*
2581 * Then move the ready-to-invoke callbacks to the orphanage,
2582 * where some other CPU will pick them up. These will not be
2583 * required to pass though another grace period: They are done.
2584 */
2585 if (rdp->nxtlist != NULL) {
2586 *rsp->orphan_donetail = rdp->nxtlist;
2587 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
2588 }
2589
2590 /*
2591 * Finally, initialize the rcu_data structure's list to empty and
2592 * disallow further callbacks on this CPU.
2593 */
2594 init_callback_list(rdp);
2595 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2596}
2597
2598/*
2599 * Adopt the RCU callbacks from the specified rcu_state structure's
2600 * orphanage. The caller must hold the ->orphan_lock.
2601 */
2602static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
2603{
2604 int i;
2605 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2606
2607 /* No-CBs CPUs are handled specially. */
2608 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2609 rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
2610 return;
2611
2612 /* Do the accounting first. */
2613 rdp->qlen_lazy += rsp->qlen_lazy;
2614 rdp->qlen += rsp->qlen;
2615 rdp->n_cbs_adopted += rsp->qlen;
2616 if (rsp->qlen_lazy != rsp->qlen)
2617 rcu_idle_count_callbacks_posted();
2618 rsp->qlen_lazy = 0;
2619 rsp->qlen = 0;
2620
2621 /*
2622 * We do not need a memory barrier here because the only way we
2623 * can get here if there is an rcu_barrier() in flight is if
2624 * we are the task doing the rcu_barrier().
2625 */
2626
2627 /* First adopt the ready-to-invoke callbacks. */
2628 if (rsp->orphan_donelist != NULL) {
2629 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
2630 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
2631 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
2632 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2633 rdp->nxttail[i] = rsp->orphan_donetail;
2634 rsp->orphan_donelist = NULL;
2635 rsp->orphan_donetail = &rsp->orphan_donelist;
2636 }
2637
2638 /* And then adopt the callbacks that still need a grace period. */
2639 if (rsp->orphan_nxtlist != NULL) {
2640 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
2641 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
2642 rsp->orphan_nxtlist = NULL;
2643 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2644 }
2645}
2646
2647/*
2648 * Trace the fact that this CPU is going offline.
2649 */
2650static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2651{
2652 RCU_TRACE(unsigned long mask);
2653 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2654 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2655
2656 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2657 return;
2658
2659 RCU_TRACE(mask = rdp->grpmask);
2660 trace_rcu_grace_period(rsp->name,
2661 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2662 TPS("cpuofl"));
2663}
2664
2665/*
2666 * All CPUs for the specified rcu_node structure have gone offline,
2667 * and all tasks that were preempted within an RCU read-side critical
2668 * section while running on one of those CPUs have since exited their RCU
2669 * read-side critical section. Some other CPU is reporting this fact with
2670 * the specified rcu_node structure's ->lock held and interrupts disabled.
2671 * This function therefore goes up the tree of rcu_node structures,
2672 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2673 * the leaf rcu_node structure's ->qsmaskinit field has already been
2674 * updated
2675 *
2676 * This function does check that the specified rcu_node structure has
2677 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2678 * prematurely. That said, invoking it after the fact will cost you
2679 * a needless lock acquisition. So once it has done its work, don't
2680 * invoke it again.
2681 */
2682static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2683{
2684 long mask;
2685 struct rcu_node *rnp = rnp_leaf;
2686
2687 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2688 rnp->qsmaskinit || rcu_preempt_has_tasks(rnp))
2689 return;
2690 for (;;) {
2691 mask = rnp->grpmask;
2692 rnp = rnp->parent;
2693 if (!rnp)
2694 break;
2695 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2696 rnp->qsmaskinit &= ~mask;
2697 rnp->qsmask &= ~mask;
2698 if (rnp->qsmaskinit) {
2699 raw_spin_unlock_rcu_node(rnp);
2700 /* irqs remain disabled. */
2701 return;
2702 }
2703 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2704 }
2705}
2706
2707/*
2708 * The CPU has been completely removed, and some other CPU is reporting
2709 * this fact from process context. Do the remainder of the cleanup,
2710 * including orphaning the outgoing CPU's RCU callbacks, and also
2711 * adopting them. There can only be one CPU hotplug operation at a time,
2712 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2713 */
2714static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2715{
2716 unsigned long flags;
2717 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2718 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2719
2720 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2721 return;
2722
2723 /* Adjust any no-longer-needed kthreads. */
2724 rcu_boost_kthread_setaffinity(rnp, -1);
2725
2726 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2727 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2728 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2729 rcu_adopt_orphan_cbs(rsp, flags);
2730 raw_spin_unlock_irqrestore(&rsp->orphan_lock, flags);
2731
2732 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2733 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2734 cpu, rdp->qlen, rdp->nxtlist);
2735}
2736
2737/*
2738 * Invoke any RCU callbacks that have made it to the end of their grace
2739 * period. Thottle as specified by rdp->blimit.
2740 */
2741static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2742{
2743 unsigned long flags;
2744 struct rcu_head *next, *list, **tail;
2745 long bl, count, count_lazy;
2746 int i;
2747
2748 /* If no callbacks are ready, just return. */
2749 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2750 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2751 trace_rcu_batch_end(rsp->name, 0, !!READ_ONCE(rdp->nxtlist),
2752 need_resched(), is_idle_task(current),
2753 rcu_is_callbacks_kthread());
2754 return;
2755 }
2756
2757 /*
2758 * Extract the list of ready callbacks, disabling to prevent
2759 * races with call_rcu() from interrupt handlers.
2760 */
2761 local_irq_save(flags);
2762 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2763 bl = rdp->blimit;
2764 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2765 list = rdp->nxtlist;
2766 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2767 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2768 tail = rdp->nxttail[RCU_DONE_TAIL];
2769 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2770 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2771 rdp->nxttail[i] = &rdp->nxtlist;
2772 local_irq_restore(flags);
2773
2774 /* Invoke callbacks. */
2775 count = count_lazy = 0;
2776 while (list) {
2777 next = list->next;
2778 prefetch(next);
2779 debug_rcu_head_unqueue(list);
2780 if (__rcu_reclaim(rsp->name, list))
2781 count_lazy++;
2782 list = next;
2783 /* Stop only if limit reached and CPU has something to do. */
2784 if (++count >= bl &&
2785 (need_resched() ||
2786 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2787 break;
2788 }
2789
2790 local_irq_save(flags);
2791 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2792 is_idle_task(current),
2793 rcu_is_callbacks_kthread());
2794
2795 /* Update count, and requeue any remaining callbacks. */
2796 if (list != NULL) {
2797 *tail = rdp->nxtlist;
2798 rdp->nxtlist = list;
2799 for (i = 0; i < RCU_NEXT_SIZE; i++)
2800 if (&rdp->nxtlist == rdp->nxttail[i])
2801 rdp->nxttail[i] = tail;
2802 else
2803 break;
2804 }
2805 smp_mb(); /* List handling before counting for rcu_barrier(). */
2806 rdp->qlen_lazy -= count_lazy;
2807 WRITE_ONCE(rdp->qlen, rdp->qlen - count);
2808 rdp->n_cbs_invoked += count;
2809
2810 /* Reinstate batch limit if we have worked down the excess. */
2811 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2812 rdp->blimit = blimit;
2813
2814 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2815 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2816 rdp->qlen_last_fqs_check = 0;
2817 rdp->n_force_qs_snap = rsp->n_force_qs;
2818 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2819 rdp->qlen_last_fqs_check = rdp->qlen;
2820 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2821
2822 local_irq_restore(flags);
2823
2824 /* Re-invoke RCU core processing if there are callbacks remaining. */
2825 if (cpu_has_callbacks_ready_to_invoke(rdp))
2826 invoke_rcu_core();
2827}
2828
2829/*
2830 * Check to see if this CPU is in a non-context-switch quiescent state
2831 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2832 * Also schedule RCU core processing.
2833 *
2834 * This function must be called from hardirq context. It is normally
2835 * invoked from the scheduling-clock interrupt.
2836 */
2837void rcu_check_callbacks(int user)
2838{
2839 trace_rcu_utilization(TPS("Start scheduler-tick"));
2840 increment_cpu_stall_ticks();
2841 if (user || rcu_is_cpu_rrupt_from_idle()) {
2842
2843 /*
2844 * Get here if this CPU took its interrupt from user
2845 * mode or from the idle loop, and if this is not a
2846 * nested interrupt. In this case, the CPU is in
2847 * a quiescent state, so note it.
2848 *
2849 * No memory barrier is required here because both
2850 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2851 * variables that other CPUs neither access nor modify,
2852 * at least not while the corresponding CPU is online.
2853 */
2854
2855 rcu_sched_qs();
2856 rcu_bh_qs();
2857
2858 } else if (!in_softirq()) {
2859
2860 /*
2861 * Get here if this CPU did not take its interrupt from
2862 * softirq, in other words, if it is not interrupting
2863 * a rcu_bh read-side critical section. This is an _bh
2864 * critical section, so note it.
2865 */
2866
2867 rcu_bh_qs();
2868 }
2869 rcu_preempt_check_callbacks();
2870 if (rcu_pending())
2871 invoke_rcu_core();
2872 if (user)
2873 rcu_note_voluntary_context_switch(current);
2874 trace_rcu_utilization(TPS("End scheduler-tick"));
2875}
2876
2877/*
2878 * Scan the leaf rcu_node structures, processing dyntick state for any that
2879 * have not yet encountered a quiescent state, using the function specified.
2880 * Also initiate boosting for any threads blocked on the root rcu_node.
2881 *
2882 * The caller must have suppressed start of new grace periods.
2883 */
2884static void force_qs_rnp(struct rcu_state *rsp,
2885 int (*f)(struct rcu_data *rsp, bool *isidle,
2886 unsigned long *maxj),
2887 bool *isidle, unsigned long *maxj)
2888{
2889 int cpu;
2890 unsigned long flags;
2891 unsigned long mask;
2892 struct rcu_node *rnp;
2893
2894 rcu_for_each_leaf_node(rsp, rnp) {
2895 cond_resched_rcu_qs();
2896 mask = 0;
2897 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2898 if (rnp->qsmask == 0) {
2899 if (rcu_state_p == &rcu_sched_state ||
2900 rsp != rcu_state_p ||
2901 rcu_preempt_blocked_readers_cgp(rnp)) {
2902 /*
2903 * No point in scanning bits because they
2904 * are all zero. But we might need to
2905 * priority-boost blocked readers.
2906 */
2907 rcu_initiate_boost(rnp, flags);
2908 /* rcu_initiate_boost() releases rnp->lock */
2909 continue;
2910 }
2911 if (rnp->parent &&
2912 (rnp->parent->qsmask & rnp->grpmask)) {
2913 /*
2914 * Race between grace-period
2915 * initialization and task exiting RCU
2916 * read-side critical section: Report.
2917 */
2918 rcu_report_unblock_qs_rnp(rsp, rnp, flags);
2919 /* rcu_report_unblock_qs_rnp() rlses ->lock */
2920 continue;
2921 }
2922 }
2923 for_each_leaf_node_possible_cpu(rnp, cpu) {
2924 unsigned long bit = leaf_node_cpu_bit(rnp, cpu);
2925 if ((rnp->qsmask & bit) != 0) {
2926 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2927 mask |= bit;
2928 }
2929 }
2930 if (mask != 0) {
2931 /* Idle/offline CPUs, report (releases rnp->lock. */
2932 rcu_report_qs_rnp(mask, rsp, rnp, rnp->gpnum, flags);
2933 } else {
2934 /* Nothing to do here, so just drop the lock. */
2935 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2936 }
2937 }
2938}
2939
2940/*
2941 * Force quiescent states on reluctant CPUs, and also detect which
2942 * CPUs are in dyntick-idle mode.
2943 */
2944static void force_quiescent_state(struct rcu_state *rsp)
2945{
2946 unsigned long flags;
2947 bool ret;
2948 struct rcu_node *rnp;
2949 struct rcu_node *rnp_old = NULL;
2950
2951 /* Funnel through hierarchy to reduce memory contention. */
2952 rnp = __this_cpu_read(rsp->rda->mynode);
2953 for (; rnp != NULL; rnp = rnp->parent) {
2954 ret = (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2955 !raw_spin_trylock(&rnp->fqslock);
2956 if (rnp_old != NULL)
2957 raw_spin_unlock(&rnp_old->fqslock);
2958 if (ret) {
2959 rsp->n_force_qs_lh++;
2960 return;
2961 }
2962 rnp_old = rnp;
2963 }
2964 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2965
2966 /* Reached the root of the rcu_node tree, acquire lock. */
2967 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2968 raw_spin_unlock(&rnp_old->fqslock);
2969 if (READ_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2970 rsp->n_force_qs_lh++;
2971 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2972 return; /* Someone beat us to it. */
2973 }
2974 WRITE_ONCE(rsp->gp_flags, READ_ONCE(rsp->gp_flags) | RCU_GP_FLAG_FQS);
2975 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2976 rcu_gp_kthread_wake(rsp);
2977}
2978
2979/*
2980 * This does the RCU core processing work for the specified rcu_state
2981 * and rcu_data structures. This may be called only from the CPU to
2982 * whom the rdp belongs.
2983 */
2984static void
2985__rcu_process_callbacks(struct rcu_state *rsp)
2986{
2987 unsigned long flags;
2988 bool needwake;
2989 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
2990
2991 WARN_ON_ONCE(rdp->beenonline == 0);
2992
2993 /* Update RCU state based on any recent quiescent states. */
2994 rcu_check_quiescent_state(rsp, rdp);
2995
2996 /* Does this CPU require a not-yet-started grace period? */
2997 local_irq_save(flags);
2998 if (cpu_needs_another_gp(rsp, rdp)) {
2999 raw_spin_lock_rcu_node(rcu_get_root(rsp)); /* irqs disabled. */
3000 needwake = rcu_start_gp(rsp);
3001 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(rsp), flags);
3002 if (needwake)
3003 rcu_gp_kthread_wake(rsp);
3004 } else {
3005 local_irq_restore(flags);
3006 }
3007
3008 /* If there are callbacks ready, invoke them. */
3009 if (cpu_has_callbacks_ready_to_invoke(rdp))
3010 invoke_rcu_callbacks(rsp, rdp);
3011
3012 /* Do any needed deferred wakeups of rcuo kthreads. */
3013 do_nocb_deferred_wakeup(rdp);
3014}
3015
3016/*
3017 * Do RCU core processing for the current CPU.
3018 */
3019static __latent_entropy void rcu_process_callbacks(struct softirq_action *unused)
3020{
3021 struct rcu_state *rsp;
3022
3023 if (cpu_is_offline(smp_processor_id()))
3024 return;
3025 trace_rcu_utilization(TPS("Start RCU core"));
3026 for_each_rcu_flavor(rsp)
3027 __rcu_process_callbacks(rsp);
3028 trace_rcu_utilization(TPS("End RCU core"));
3029}
3030
3031/*
3032 * Schedule RCU callback invocation. If the specified type of RCU
3033 * does not support RCU priority boosting, just do a direct call,
3034 * otherwise wake up the per-CPU kernel kthread. Note that because we
3035 * are running on the current CPU with softirqs disabled, the
3036 * rcu_cpu_kthread_task cannot disappear out from under us.
3037 */
3038static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
3039{
3040 if (unlikely(!READ_ONCE(rcu_scheduler_fully_active)))
3041 return;
3042 if (likely(!rsp->boost)) {
3043 rcu_do_batch(rsp, rdp);
3044 return;
3045 }
3046 invoke_rcu_callbacks_kthread();
3047}
3048
3049static void invoke_rcu_core(void)
3050{
3051 if (cpu_online(smp_processor_id()))
3052 raise_softirq(RCU_SOFTIRQ);
3053}
3054
3055/*
3056 * Handle any core-RCU processing required by a call_rcu() invocation.
3057 */
3058static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
3059 struct rcu_head *head, unsigned long flags)
3060{
3061 bool needwake;
3062
3063 /*
3064 * If called from an extended quiescent state, invoke the RCU
3065 * core in order to force a re-evaluation of RCU's idleness.
3066 */
3067 if (!rcu_is_watching())
3068 invoke_rcu_core();
3069
3070 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
3071 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
3072 return;
3073
3074 /*
3075 * Force the grace period if too many callbacks or too long waiting.
3076 * Enforce hysteresis, and don't invoke force_quiescent_state()
3077 * if some other CPU has recently done so. Also, don't bother
3078 * invoking force_quiescent_state() if the newly enqueued callback
3079 * is the only one waiting for a grace period to complete.
3080 */
3081 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
3082
3083 /* Are we ignoring a completed grace period? */
3084 note_gp_changes(rsp, rdp);
3085
3086 /* Start a new grace period if one not already started. */
3087 if (!rcu_gp_in_progress(rsp)) {
3088 struct rcu_node *rnp_root = rcu_get_root(rsp);
3089
3090 raw_spin_lock_rcu_node(rnp_root);
3091 needwake = rcu_start_gp(rsp);
3092 raw_spin_unlock_rcu_node(rnp_root);
3093 if (needwake)
3094 rcu_gp_kthread_wake(rsp);
3095 } else {
3096 /* Give the grace period a kick. */
3097 rdp->blimit = LONG_MAX;
3098 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
3099 *rdp->nxttail[RCU_DONE_TAIL] != head)
3100 force_quiescent_state(rsp);
3101 rdp->n_force_qs_snap = rsp->n_force_qs;
3102 rdp->qlen_last_fqs_check = rdp->qlen;
3103 }
3104 }
3105}
3106
3107/*
3108 * RCU callback function to leak a callback.
3109 */
3110static void rcu_leak_callback(struct rcu_head *rhp)
3111{
3112}
3113
3114/*
3115 * Helper function for call_rcu() and friends. The cpu argument will
3116 * normally be -1, indicating "currently running CPU". It may specify
3117 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
3118 * is expected to specify a CPU.
3119 */
3120static void
3121__call_rcu(struct rcu_head *head, rcu_callback_t func,
3122 struct rcu_state *rsp, int cpu, bool lazy)
3123{
3124 unsigned long flags;
3125 struct rcu_data *rdp;
3126
3127 /* Misaligned rcu_head! */
3128 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3129
3130 if (debug_rcu_head_queue(head)) {
3131 /* Probable double call_rcu(), so leak the callback. */
3132 WRITE_ONCE(head->func, rcu_leak_callback);
3133 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
3134 return;
3135 }
3136 head->func = func;
3137 head->next = NULL;
3138 local_irq_save(flags);
3139 rdp = this_cpu_ptr(rsp->rda);
3140
3141 /* Add the callback to our list. */
3142 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
3143 int offline;
3144
3145 if (cpu != -1)
3146 rdp = per_cpu_ptr(rsp->rda, cpu);
3147 if (likely(rdp->mynode)) {
3148 /* Post-boot, so this should be for a no-CBs CPU. */
3149 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
3150 WARN_ON_ONCE(offline);
3151 /* Offline CPU, _call_rcu() illegal, leak callback. */
3152 local_irq_restore(flags);
3153 return;
3154 }
3155 /*
3156 * Very early boot, before rcu_init(). Initialize if needed
3157 * and then drop through to queue the callback.
3158 */
3159 BUG_ON(cpu != -1);
3160 WARN_ON_ONCE(!rcu_is_watching());
3161 if (!likely(rdp->nxtlist))
3162 init_default_callback_list(rdp);
3163 }
3164 WRITE_ONCE(rdp->qlen, rdp->qlen + 1);
3165 if (lazy)
3166 rdp->qlen_lazy++;
3167 else
3168 rcu_idle_count_callbacks_posted();
3169 smp_mb(); /* Count before adding callback for rcu_barrier(). */
3170 *rdp->nxttail[RCU_NEXT_TAIL] = head;
3171 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
3172
3173 if (__is_kfree_rcu_offset((unsigned long)func))
3174 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
3175 rdp->qlen_lazy, rdp->qlen);
3176 else
3177 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
3178
3179 /* Go handle any RCU core processing required. */
3180 __call_rcu_core(rsp, rdp, head, flags);
3181 local_irq_restore(flags);
3182}
3183
3184/*
3185 * Queue an RCU-sched callback for invocation after a grace period.
3186 */
3187void call_rcu_sched(struct rcu_head *head, rcu_callback_t func)
3188{
3189 __call_rcu(head, func, &rcu_sched_state, -1, 0);
3190}
3191EXPORT_SYMBOL_GPL(call_rcu_sched);
3192
3193/*
3194 * Queue an RCU callback for invocation after a quicker grace period.
3195 */
3196void call_rcu_bh(struct rcu_head *head, rcu_callback_t func)
3197{
3198 __call_rcu(head, func, &rcu_bh_state, -1, 0);
3199}
3200EXPORT_SYMBOL_GPL(call_rcu_bh);
3201
3202/*
3203 * Queue an RCU callback for lazy invocation after a grace period.
3204 * This will likely be later named something like "call_rcu_lazy()",
3205 * but this change will require some way of tagging the lazy RCU
3206 * callbacks in the list of pending callbacks. Until then, this
3207 * function may only be called from __kfree_rcu().
3208 */
3209void kfree_call_rcu(struct rcu_head *head,
3210 rcu_callback_t func)
3211{
3212 __call_rcu(head, func, rcu_state_p, -1, 1);
3213}
3214EXPORT_SYMBOL_GPL(kfree_call_rcu);
3215
3216/*
3217 * Because a context switch is a grace period for RCU-sched and RCU-bh,
3218 * any blocking grace-period wait automatically implies a grace period
3219 * if there is only one CPU online at any point time during execution
3220 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
3221 * occasionally incorrectly indicate that there are multiple CPUs online
3222 * when there was in fact only one the whole time, as this just adds
3223 * some overhead: RCU still operates correctly.
3224 */
3225static inline int rcu_blocking_is_gp(void)
3226{
3227 int ret;
3228
3229 might_sleep(); /* Check for RCU read-side critical section. */
3230 preempt_disable();
3231 ret = num_online_cpus() <= 1;
3232 preempt_enable();
3233 return ret;
3234}
3235
3236/**
3237 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
3238 *
3239 * Control will return to the caller some time after a full rcu-sched
3240 * grace period has elapsed, in other words after all currently executing
3241 * rcu-sched read-side critical sections have completed. These read-side
3242 * critical sections are delimited by rcu_read_lock_sched() and
3243 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
3244 * local_irq_disable(), and so on may be used in place of
3245 * rcu_read_lock_sched().
3246 *
3247 * This means that all preempt_disable code sequences, including NMI and
3248 * non-threaded hardware-interrupt handlers, in progress on entry will
3249 * have completed before this primitive returns. However, this does not
3250 * guarantee that softirq handlers will have completed, since in some
3251 * kernels, these handlers can run in process context, and can block.
3252 *
3253 * Note that this guarantee implies further memory-ordering guarantees.
3254 * On systems with more than one CPU, when synchronize_sched() returns,
3255 * each CPU is guaranteed to have executed a full memory barrier since the
3256 * end of its last RCU-sched read-side critical section whose beginning
3257 * preceded the call to synchronize_sched(). In addition, each CPU having
3258 * an RCU read-side critical section that extends beyond the return from
3259 * synchronize_sched() is guaranteed to have executed a full memory barrier
3260 * after the beginning of synchronize_sched() and before the beginning of
3261 * that RCU read-side critical section. Note that these guarantees include
3262 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3263 * that are executing in the kernel.
3264 *
3265 * Furthermore, if CPU A invoked synchronize_sched(), which returned
3266 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3267 * to have executed a full memory barrier during the execution of
3268 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
3269 * again only if the system has more than one CPU).
3270 *
3271 * This primitive provides the guarantees made by the (now removed)
3272 * synchronize_kernel() API. In contrast, synchronize_rcu() only
3273 * guarantees that rcu_read_lock() sections will have completed.
3274 * In "classic RCU", these two guarantees happen to be one and
3275 * the same, but can differ in realtime RCU implementations.
3276 */
3277void synchronize_sched(void)
3278{
3279 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3280 lock_is_held(&rcu_lock_map) ||
3281 lock_is_held(&rcu_sched_lock_map),
3282 "Illegal synchronize_sched() in RCU-sched read-side critical section");
3283 if (rcu_blocking_is_gp())
3284 return;
3285 if (rcu_gp_is_expedited())
3286 synchronize_sched_expedited();
3287 else
3288 wait_rcu_gp(call_rcu_sched);
3289}
3290EXPORT_SYMBOL_GPL(synchronize_sched);
3291
3292/**
3293 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
3294 *
3295 * Control will return to the caller some time after a full rcu_bh grace
3296 * period has elapsed, in other words after all currently executing rcu_bh
3297 * read-side critical sections have completed. RCU read-side critical
3298 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
3299 * and may be nested.
3300 *
3301 * See the description of synchronize_sched() for more detailed information
3302 * on memory ordering guarantees.
3303 */
3304void synchronize_rcu_bh(void)
3305{
3306 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3307 lock_is_held(&rcu_lock_map) ||
3308 lock_is_held(&rcu_sched_lock_map),
3309 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
3310 if (rcu_blocking_is_gp())
3311 return;
3312 if (rcu_gp_is_expedited())
3313 synchronize_rcu_bh_expedited();
3314 else
3315 wait_rcu_gp(call_rcu_bh);
3316}
3317EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
3318
3319/**
3320 * get_state_synchronize_rcu - Snapshot current RCU state
3321 *
3322 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3323 * to determine whether or not a full grace period has elapsed in the
3324 * meantime.
3325 */
3326unsigned long get_state_synchronize_rcu(void)
3327{
3328 /*
3329 * Any prior manipulation of RCU-protected data must happen
3330 * before the load from ->gpnum.
3331 */
3332 smp_mb(); /* ^^^ */
3333
3334 /*
3335 * Make sure this load happens before the purportedly
3336 * time-consuming work between get_state_synchronize_rcu()
3337 * and cond_synchronize_rcu().
3338 */
3339 return smp_load_acquire(&rcu_state_p->gpnum);
3340}
3341EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3342
3343/**
3344 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3345 *
3346 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
3347 *
3348 * If a full RCU grace period has elapsed since the earlier call to
3349 * get_state_synchronize_rcu(), just return. Otherwise, invoke
3350 * synchronize_rcu() to wait for a full grace period.
3351 *
3352 * Yes, this function does not take counter wrap into account. But
3353 * counter wrap is harmless. If the counter wraps, we have waited for
3354 * more than 2 billion grace periods (and way more on a 64-bit system!),
3355 * so waiting for one additional grace period should be just fine.
3356 */
3357void cond_synchronize_rcu(unsigned long oldstate)
3358{
3359 unsigned long newstate;
3360
3361 /*
3362 * Ensure that this load happens before any RCU-destructive
3363 * actions the caller might carry out after we return.
3364 */
3365 newstate = smp_load_acquire(&rcu_state_p->completed);
3366 if (ULONG_CMP_GE(oldstate, newstate))
3367 synchronize_rcu();
3368}
3369EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3370
3371/**
3372 * get_state_synchronize_sched - Snapshot current RCU-sched state
3373 *
3374 * Returns a cookie that is used by a later call to cond_synchronize_sched()
3375 * to determine whether or not a full grace period has elapsed in the
3376 * meantime.
3377 */
3378unsigned long get_state_synchronize_sched(void)
3379{
3380 /*
3381 * Any prior manipulation of RCU-protected data must happen
3382 * before the load from ->gpnum.
3383 */
3384 smp_mb(); /* ^^^ */
3385
3386 /*
3387 * Make sure this load happens before the purportedly
3388 * time-consuming work between get_state_synchronize_sched()
3389 * and cond_synchronize_sched().
3390 */
3391 return smp_load_acquire(&rcu_sched_state.gpnum);
3392}
3393EXPORT_SYMBOL_GPL(get_state_synchronize_sched);
3394
3395/**
3396 * cond_synchronize_sched - Conditionally wait for an RCU-sched grace period
3397 *
3398 * @oldstate: return value from earlier call to get_state_synchronize_sched()
3399 *
3400 * If a full RCU-sched grace period has elapsed since the earlier call to
3401 * get_state_synchronize_sched(), just return. Otherwise, invoke
3402 * synchronize_sched() to wait for a full grace period.
3403 *
3404 * Yes, this function does not take counter wrap into account. But
3405 * counter wrap is harmless. If the counter wraps, we have waited for
3406 * more than 2 billion grace periods (and way more on a 64-bit system!),
3407 * so waiting for one additional grace period should be just fine.
3408 */
3409void cond_synchronize_sched(unsigned long oldstate)
3410{
3411 unsigned long newstate;
3412
3413 /*
3414 * Ensure that this load happens before any RCU-destructive
3415 * actions the caller might carry out after we return.
3416 */
3417 newstate = smp_load_acquire(&rcu_sched_state.completed);
3418 if (ULONG_CMP_GE(oldstate, newstate))
3419 synchronize_sched();
3420}
3421EXPORT_SYMBOL_GPL(cond_synchronize_sched);
3422
3423/* Adjust sequence number for start of update-side operation. */
3424static void rcu_seq_start(unsigned long *sp)
3425{
3426 WRITE_ONCE(*sp, *sp + 1);
3427 smp_mb(); /* Ensure update-side operation after counter increment. */
3428 WARN_ON_ONCE(!(*sp & 0x1));
3429}
3430
3431/* Adjust sequence number for end of update-side operation. */
3432static void rcu_seq_end(unsigned long *sp)
3433{
3434 smp_mb(); /* Ensure update-side operation before counter increment. */
3435 WRITE_ONCE(*sp, *sp + 1);
3436 WARN_ON_ONCE(*sp & 0x1);
3437}
3438
3439/* Take a snapshot of the update side's sequence number. */
3440static unsigned long rcu_seq_snap(unsigned long *sp)
3441{
3442 unsigned long s;
3443
3444 s = (READ_ONCE(*sp) + 3) & ~0x1;
3445 smp_mb(); /* Above access must not bleed into critical section. */
3446 return s;
3447}
3448
3449/*
3450 * Given a snapshot from rcu_seq_snap(), determine whether or not a
3451 * full update-side operation has occurred.
3452 */
3453static bool rcu_seq_done(unsigned long *sp, unsigned long s)
3454{
3455 return ULONG_CMP_GE(READ_ONCE(*sp), s);
3456}
3457
3458/*
3459 * Check to see if there is any immediate RCU-related work to be done
3460 * by the current CPU, for the specified type of RCU, returning 1 if so.
3461 * The checks are in order of increasing expense: checks that can be
3462 * carried out against CPU-local state are performed first. However,
3463 * we must check for CPU stalls first, else we might not get a chance.
3464 */
3465static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
3466{
3467 struct rcu_node *rnp = rdp->mynode;
3468
3469 rdp->n_rcu_pending++;
3470
3471 /* Check for CPU stalls, if enabled. */
3472 check_cpu_stall(rsp, rdp);
3473
3474 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
3475 if (rcu_nohz_full_cpu(rsp))
3476 return 0;
3477
3478 /* Is the RCU core waiting for a quiescent state from this CPU? */
3479 if (rcu_scheduler_fully_active &&
3480 rdp->core_needs_qs && rdp->cpu_no_qs.b.norm &&
3481 rdp->rcu_qs_ctr_snap == __this_cpu_read(rcu_qs_ctr)) {
3482 rdp->n_rp_core_needs_qs++;
3483 } else if (rdp->core_needs_qs &&
3484 (!rdp->cpu_no_qs.b.norm ||
3485 rdp->rcu_qs_ctr_snap != __this_cpu_read(rcu_qs_ctr))) {
3486 rdp->n_rp_report_qs++;
3487 return 1;
3488 }
3489
3490 /* Does this CPU have callbacks ready to invoke? */
3491 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
3492 rdp->n_rp_cb_ready++;
3493 return 1;
3494 }
3495
3496 /* Has RCU gone idle with this CPU needing another grace period? */
3497 if (cpu_needs_another_gp(rsp, rdp)) {
3498 rdp->n_rp_cpu_needs_gp++;
3499 return 1;
3500 }
3501
3502 /* Has another RCU grace period completed? */
3503 if (READ_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
3504 rdp->n_rp_gp_completed++;
3505 return 1;
3506 }
3507
3508 /* Has a new RCU grace period started? */
3509 if (READ_ONCE(rnp->gpnum) != rdp->gpnum ||
3510 unlikely(READ_ONCE(rdp->gpwrap))) { /* outside lock */
3511 rdp->n_rp_gp_started++;
3512 return 1;
3513 }
3514
3515 /* Does this CPU need a deferred NOCB wakeup? */
3516 if (rcu_nocb_need_deferred_wakeup(rdp)) {
3517 rdp->n_rp_nocb_defer_wakeup++;
3518 return 1;
3519 }
3520
3521 /* nothing to do */
3522 rdp->n_rp_need_nothing++;
3523 return 0;
3524}
3525
3526/*
3527 * Check to see if there is any immediate RCU-related work to be done
3528 * by the current CPU, returning 1 if so. This function is part of the
3529 * RCU implementation; it is -not- an exported member of the RCU API.
3530 */
3531static int rcu_pending(void)
3532{
3533 struct rcu_state *rsp;
3534
3535 for_each_rcu_flavor(rsp)
3536 if (__rcu_pending(rsp, this_cpu_ptr(rsp->rda)))
3537 return 1;
3538 return 0;
3539}
3540
3541/*
3542 * Return true if the specified CPU has any callback. If all_lazy is
3543 * non-NULL, store an indication of whether all callbacks are lazy.
3544 * (If there are no callbacks, all of them are deemed to be lazy.)
3545 */
3546static bool __maybe_unused rcu_cpu_has_callbacks(bool *all_lazy)
3547{
3548 bool al = true;
3549 bool hc = false;
3550 struct rcu_data *rdp;
3551 struct rcu_state *rsp;
3552
3553 for_each_rcu_flavor(rsp) {
3554 rdp = this_cpu_ptr(rsp->rda);
3555 if (!rdp->nxtlist)
3556 continue;
3557 hc = true;
3558 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
3559 al = false;
3560 break;
3561 }
3562 }
3563 if (all_lazy)
3564 *all_lazy = al;
3565 return hc;
3566}
3567
3568/*
3569 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
3570 * the compiler is expected to optimize this away.
3571 */
3572static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
3573 int cpu, unsigned long done)
3574{
3575 trace_rcu_barrier(rsp->name, s, cpu,
3576 atomic_read(&rsp->barrier_cpu_count), done);
3577}
3578
3579/*
3580 * RCU callback function for _rcu_barrier(). If we are last, wake
3581 * up the task executing _rcu_barrier().
3582 */
3583static void rcu_barrier_callback(struct rcu_head *rhp)
3584{
3585 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
3586 struct rcu_state *rsp = rdp->rsp;
3587
3588 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
3589 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->barrier_sequence);
3590 complete(&rsp->barrier_completion);
3591 } else {
3592 _rcu_barrier_trace(rsp, "CB", -1, rsp->barrier_sequence);
3593 }
3594}
3595
3596/*
3597 * Called with preemption disabled, and from cross-cpu IRQ context.
3598 */
3599static void rcu_barrier_func(void *type)
3600{
3601 struct rcu_state *rsp = type;
3602 struct rcu_data *rdp = raw_cpu_ptr(rsp->rda);
3603
3604 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->barrier_sequence);
3605 atomic_inc(&rsp->barrier_cpu_count);
3606 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
3607}
3608
3609/*
3610 * Orchestrate the specified type of RCU barrier, waiting for all
3611 * RCU callbacks of the specified type to complete.
3612 */
3613static void _rcu_barrier(struct rcu_state *rsp)
3614{
3615 int cpu;
3616 struct rcu_data *rdp;
3617 unsigned long s = rcu_seq_snap(&rsp->barrier_sequence);
3618
3619 _rcu_barrier_trace(rsp, "Begin", -1, s);
3620
3621 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3622 mutex_lock(&rsp->barrier_mutex);
3623
3624 /* Did someone else do our work for us? */
3625 if (rcu_seq_done(&rsp->barrier_sequence, s)) {
3626 _rcu_barrier_trace(rsp, "EarlyExit", -1, rsp->barrier_sequence);
3627 smp_mb(); /* caller's subsequent code after above check. */
3628 mutex_unlock(&rsp->barrier_mutex);
3629 return;
3630 }
3631
3632 /* Mark the start of the barrier operation. */
3633 rcu_seq_start(&rsp->barrier_sequence);
3634 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->barrier_sequence);
3635
3636 /*
3637 * Initialize the count to one rather than to zero in order to
3638 * avoid a too-soon return to zero in case of a short grace period
3639 * (or preemption of this task). Exclude CPU-hotplug operations
3640 * to ensure that no offline CPU has callbacks queued.
3641 */
3642 init_completion(&rsp->barrier_completion);
3643 atomic_set(&rsp->barrier_cpu_count, 1);
3644 get_online_cpus();
3645
3646 /*
3647 * Force each CPU with callbacks to register a new callback.
3648 * When that callback is invoked, we will know that all of the
3649 * corresponding CPU's preceding callbacks have been invoked.
3650 */
3651 for_each_possible_cpu(cpu) {
3652 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3653 continue;
3654 rdp = per_cpu_ptr(rsp->rda, cpu);
3655 if (rcu_is_nocb_cpu(cpu)) {
3656 if (!rcu_nocb_cpu_needs_barrier(rsp, cpu)) {
3657 _rcu_barrier_trace(rsp, "OfflineNoCB", cpu,
3658 rsp->barrier_sequence);
3659 } else {
3660 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3661 rsp->barrier_sequence);
3662 smp_mb__before_atomic();
3663 atomic_inc(&rsp->barrier_cpu_count);
3664 __call_rcu(&rdp->barrier_head,
3665 rcu_barrier_callback, rsp, cpu, 0);
3666 }
3667 } else if (READ_ONCE(rdp->qlen)) {
3668 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3669 rsp->barrier_sequence);
3670 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3671 } else {
3672 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3673 rsp->barrier_sequence);
3674 }
3675 }
3676 put_online_cpus();
3677
3678 /*
3679 * Now that we have an rcu_barrier_callback() callback on each
3680 * CPU, and thus each counted, remove the initial count.
3681 */
3682 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3683 complete(&rsp->barrier_completion);
3684
3685 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3686 wait_for_completion(&rsp->barrier_completion);
3687
3688 /* Mark the end of the barrier operation. */
3689 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->barrier_sequence);
3690 rcu_seq_end(&rsp->barrier_sequence);
3691
3692 /* Other rcu_barrier() invocations can now safely proceed. */
3693 mutex_unlock(&rsp->barrier_mutex);
3694}
3695
3696/**
3697 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3698 */
3699void rcu_barrier_bh(void)
3700{
3701 _rcu_barrier(&rcu_bh_state);
3702}
3703EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3704
3705/**
3706 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3707 */
3708void rcu_barrier_sched(void)
3709{
3710 _rcu_barrier(&rcu_sched_state);
3711}
3712EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3713
3714/*
3715 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
3716 * first CPU in a given leaf rcu_node structure coming online. The caller
3717 * must hold the corresponding leaf rcu_node ->lock with interrrupts
3718 * disabled.
3719 */
3720static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
3721{
3722 long mask;
3723 struct rcu_node *rnp = rnp_leaf;
3724
3725 for (;;) {
3726 mask = rnp->grpmask;
3727 rnp = rnp->parent;
3728 if (rnp == NULL)
3729 return;
3730 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
3731 rnp->qsmaskinit |= mask;
3732 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
3733 }
3734}
3735
3736/*
3737 * Do boot-time initialization of a CPU's per-CPU RCU data.
3738 */
3739static void __init
3740rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3741{
3742 unsigned long flags;
3743 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3744 struct rcu_node *rnp = rcu_get_root(rsp);
3745
3746 /* Set up local state, ensuring consistent view of global state. */
3747 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3748 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
3749 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3750 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3751 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3752 rdp->cpu = cpu;
3753 rdp->rsp = rsp;
3754 rcu_boot_init_nocb_percpu_data(rdp);
3755 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3756}
3757
3758/*
3759 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3760 * offline event can be happening at a given time. Note also that we
3761 * can accept some slop in the rsp->completed access due to the fact
3762 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3763 */
3764static void
3765rcu_init_percpu_data(int cpu, struct rcu_state *rsp)
3766{
3767 unsigned long flags;
3768 unsigned long mask;
3769 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3770 struct rcu_node *rnp = rcu_get_root(rsp);
3771
3772 /* Set up local state, ensuring consistent view of global state. */
3773 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3774 rdp->qlen_last_fqs_check = 0;
3775 rdp->n_force_qs_snap = rsp->n_force_qs;
3776 rdp->blimit = blimit;
3777 if (!rdp->nxtlist)
3778 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3779 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3780 rcu_sysidle_init_percpu_data(rdp->dynticks);
3781 atomic_set(&rdp->dynticks->dynticks,
3782 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3783 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
3784
3785 /*
3786 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
3787 * propagation up the rcu_node tree will happen at the beginning
3788 * of the next grace period.
3789 */
3790 rnp = rdp->mynode;
3791 mask = rdp->grpmask;
3792 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
3793 if (!rdp->beenonline)
3794 WRITE_ONCE(rsp->ncpus, READ_ONCE(rsp->ncpus) + 1);
3795 rdp->beenonline = true; /* We have now been online. */
3796 rdp->gpnum = rnp->completed; /* Make CPU later note any new GP. */
3797 rdp->completed = rnp->completed;
3798 rdp->cpu_no_qs.b.norm = true;
3799 rdp->rcu_qs_ctr_snap = per_cpu(rcu_qs_ctr, cpu);
3800 rdp->core_needs_qs = false;
3801 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3802 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3803}
3804
3805int rcutree_prepare_cpu(unsigned int cpu)
3806{
3807 struct rcu_state *rsp;
3808
3809 for_each_rcu_flavor(rsp)
3810 rcu_init_percpu_data(cpu, rsp);
3811
3812 rcu_prepare_kthreads(cpu);
3813 rcu_spawn_all_nocb_kthreads(cpu);
3814
3815 return 0;
3816}
3817
3818static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
3819{
3820 struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
3821
3822 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
3823}
3824
3825int rcutree_online_cpu(unsigned int cpu)
3826{
3827 sync_sched_exp_online_cleanup(cpu);
3828 rcutree_affinity_setting(cpu, -1);
3829 return 0;
3830}
3831
3832int rcutree_offline_cpu(unsigned int cpu)
3833{
3834 rcutree_affinity_setting(cpu, cpu);
3835 return 0;
3836}
3837
3838
3839int rcutree_dying_cpu(unsigned int cpu)
3840{
3841 struct rcu_state *rsp;
3842
3843 for_each_rcu_flavor(rsp)
3844 rcu_cleanup_dying_cpu(rsp);
3845 return 0;
3846}
3847
3848int rcutree_dead_cpu(unsigned int cpu)
3849{
3850 struct rcu_state *rsp;
3851
3852 for_each_rcu_flavor(rsp) {
3853 rcu_cleanup_dead_cpu(cpu, rsp);
3854 do_nocb_deferred_wakeup(per_cpu_ptr(rsp->rda, cpu));
3855 }
3856 return 0;
3857}
3858
3859/*
3860 * Mark the specified CPU as being online so that subsequent grace periods
3861 * (both expedited and normal) will wait on it. Note that this means that
3862 * incoming CPUs are not allowed to use RCU read-side critical sections
3863 * until this function is called. Failing to observe this restriction
3864 * will result in lockdep splats.
3865 */
3866void rcu_cpu_starting(unsigned int cpu)
3867{
3868 unsigned long flags;
3869 unsigned long mask;
3870 struct rcu_data *rdp;
3871 struct rcu_node *rnp;
3872 struct rcu_state *rsp;
3873
3874 for_each_rcu_flavor(rsp) {
3875 rdp = this_cpu_ptr(rsp->rda);
3876 rnp = rdp->mynode;
3877 mask = rdp->grpmask;
3878 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3879 rnp->qsmaskinitnext |= mask;
3880 rnp->expmaskinitnext |= mask;
3881 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3882 }
3883}
3884
3885#ifdef CONFIG_HOTPLUG_CPU
3886/*
3887 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3888 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3889 * bit masks.
3890 * The CPU is exiting the idle loop into the arch_cpu_idle_dead()
3891 * function. We now remove it from the rcu_node tree's ->qsmaskinit
3892 * bit masks.
3893 */
3894static void rcu_cleanup_dying_idle_cpu(int cpu, struct rcu_state *rsp)
3895{
3896 unsigned long flags;
3897 unsigned long mask;
3898 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3899 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
3900
3901 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
3902 mask = rdp->grpmask;
3903 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
3904 rnp->qsmaskinitnext &= ~mask;
3905 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3906}
3907
3908void rcu_report_dead(unsigned int cpu)
3909{
3910 struct rcu_state *rsp;
3911
3912 /* QS for any half-done expedited RCU-sched GP. */
3913 preempt_disable();
3914 rcu_report_exp_rdp(&rcu_sched_state,
3915 this_cpu_ptr(rcu_sched_state.rda), true);
3916 preempt_enable();
3917 for_each_rcu_flavor(rsp)
3918 rcu_cleanup_dying_idle_cpu(cpu, rsp);
3919}
3920#endif
3921
3922static int rcu_pm_notify(struct notifier_block *self,
3923 unsigned long action, void *hcpu)
3924{
3925 switch (action) {
3926 case PM_HIBERNATION_PREPARE:
3927 case PM_SUSPEND_PREPARE:
3928 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3929 rcu_expedite_gp();
3930 break;
3931 case PM_POST_HIBERNATION:
3932 case PM_POST_SUSPEND:
3933 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3934 rcu_unexpedite_gp();
3935 break;
3936 default:
3937 break;
3938 }
3939 return NOTIFY_OK;
3940}
3941
3942/*
3943 * Spawn the kthreads that handle each RCU flavor's grace periods.
3944 */
3945static int __init rcu_spawn_gp_kthread(void)
3946{
3947 unsigned long flags;
3948 int kthread_prio_in = kthread_prio;
3949 struct rcu_node *rnp;
3950 struct rcu_state *rsp;
3951 struct sched_param sp;
3952 struct task_struct *t;
3953
3954 /* Force priority into range. */
3955 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
3956 kthread_prio = 1;
3957 else if (kthread_prio < 0)
3958 kthread_prio = 0;
3959 else if (kthread_prio > 99)
3960 kthread_prio = 99;
3961 if (kthread_prio != kthread_prio_in)
3962 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
3963 kthread_prio, kthread_prio_in);
3964
3965 rcu_scheduler_fully_active = 1;
3966 for_each_rcu_flavor(rsp) {
3967 t = kthread_create(rcu_gp_kthread, rsp, "%s", rsp->name);
3968 BUG_ON(IS_ERR(t));
3969 rnp = rcu_get_root(rsp);
3970 raw_spin_lock_irqsave_rcu_node(rnp, flags);
3971 rsp->gp_kthread = t;
3972 if (kthread_prio) {
3973 sp.sched_priority = kthread_prio;
3974 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
3975 }
3976 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3977 wake_up_process(t);
3978 }
3979 rcu_spawn_nocb_kthreads();
3980 rcu_spawn_boost_kthreads();
3981 return 0;
3982}
3983early_initcall(rcu_spawn_gp_kthread);
3984
3985/*
3986 * This function is invoked towards the end of the scheduler's
3987 * initialization process. Before this is called, the idle task might
3988 * contain synchronous grace-period primitives (during which time, this idle
3989 * task is booting the system, and such primitives are no-ops). After this
3990 * function is called, any synchronous grace-period primitives are run as
3991 * expedited, with the requesting task driving the grace period forward.
3992 * A later core_initcall() rcu_exp_runtime_mode() will switch to full
3993 * runtime RCU functionality.
3994 */
3995void rcu_scheduler_starting(void)
3996{
3997 WARN_ON(num_online_cpus() != 1);
3998 WARN_ON(nr_context_switches() > 0);
3999 rcu_test_sync_prims();
4000 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4001 rcu_test_sync_prims();
4002}
4003
4004/*
4005 * Compute the per-level fanout, either using the exact fanout specified
4006 * or balancing the tree, depending on the rcu_fanout_exact boot parameter.
4007 */
4008static void __init rcu_init_levelspread(int *levelspread, const int *levelcnt)
4009{
4010 int i;
4011
4012 if (rcu_fanout_exact) {
4013 levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
4014 for (i = rcu_num_lvls - 2; i >= 0; i--)
4015 levelspread[i] = RCU_FANOUT;
4016 } else {
4017 int ccur;
4018 int cprv;
4019
4020 cprv = nr_cpu_ids;
4021 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4022 ccur = levelcnt[i];
4023 levelspread[i] = (cprv + ccur - 1) / ccur;
4024 cprv = ccur;
4025 }
4026 }
4027}
4028
4029/*
4030 * Helper function for rcu_init() that initializes one rcu_state structure.
4031 */
4032static void __init rcu_init_one(struct rcu_state *rsp)
4033{
4034 static const char * const buf[] = RCU_NODE_NAME_INIT;
4035 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4036 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4037 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4038 static u8 fl_mask = 0x1;
4039
4040 int levelcnt[RCU_NUM_LVLS]; /* # nodes in each level. */
4041 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4042 int cpustride = 1;
4043 int i;
4044 int j;
4045 struct rcu_node *rnp;
4046
4047 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4048
4049 /* Silence gcc 4.8 false positive about array index out of range. */
4050 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4051 panic("rcu_init_one: rcu_num_lvls out of range");
4052
4053 /* Initialize the level-tracking arrays. */
4054
4055 for (i = 0; i < rcu_num_lvls; i++)
4056 levelcnt[i] = num_rcu_lvl[i];
4057 for (i = 1; i < rcu_num_lvls; i++)
4058 rsp->level[i] = rsp->level[i - 1] + levelcnt[i - 1];
4059 rcu_init_levelspread(levelspread, levelcnt);
4060 rsp->flavor_mask = fl_mask;
4061 fl_mask <<= 1;
4062
4063 /* Initialize the elements themselves, starting from the leaves. */
4064
4065 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4066 cpustride *= levelspread[i];
4067 rnp = rsp->level[i];
4068 for (j = 0; j < levelcnt[i]; j++, rnp++) {
4069 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4070 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4071 &rcu_node_class[i], buf[i]);
4072 raw_spin_lock_init(&rnp->fqslock);
4073 lockdep_set_class_and_name(&rnp->fqslock,
4074 &rcu_fqs_class[i], fqs[i]);
4075 rnp->gpnum = rsp->gpnum;
4076 rnp->completed = rsp->completed;
4077 rnp->qsmask = 0;
4078 rnp->qsmaskinit = 0;
4079 rnp->grplo = j * cpustride;
4080 rnp->grphi = (j + 1) * cpustride - 1;
4081 if (rnp->grphi >= nr_cpu_ids)
4082 rnp->grphi = nr_cpu_ids - 1;
4083 if (i == 0) {
4084 rnp->grpnum = 0;
4085 rnp->grpmask = 0;
4086 rnp->parent = NULL;
4087 } else {
4088 rnp->grpnum = j % levelspread[i - 1];
4089 rnp->grpmask = 1UL << rnp->grpnum;
4090 rnp->parent = rsp->level[i - 1] +
4091 j / levelspread[i - 1];
4092 }
4093 rnp->level = i;
4094 INIT_LIST_HEAD(&rnp->blkd_tasks);
4095 rcu_init_one_nocb(rnp);
4096 init_waitqueue_head(&rnp->exp_wq[0]);
4097 init_waitqueue_head(&rnp->exp_wq[1]);
4098 init_waitqueue_head(&rnp->exp_wq[2]);
4099 init_waitqueue_head(&rnp->exp_wq[3]);
4100 spin_lock_init(&rnp->exp_lock);
4101 }
4102 }
4103
4104 init_swait_queue_head(&rsp->gp_wq);
4105 init_swait_queue_head(&rsp->expedited_wq);
4106 rnp = rsp->level[rcu_num_lvls - 1];
4107 for_each_possible_cpu(i) {
4108 while (i > rnp->grphi)
4109 rnp++;
4110 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
4111 rcu_boot_init_percpu_data(i, rsp);
4112 }
4113 list_add(&rsp->flavors, &rcu_struct_flavors);
4114}
4115
4116/*
4117 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4118 * replace the definitions in tree.h because those are needed to size
4119 * the ->node array in the rcu_state structure.
4120 */
4121static void __init rcu_init_geometry(void)
4122{
4123 ulong d;
4124 int i;
4125 int rcu_capacity[RCU_NUM_LVLS];
4126
4127 /*
4128 * Initialize any unspecified boot parameters.
4129 * The default values of jiffies_till_first_fqs and
4130 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4131 * value, which is a function of HZ, then adding one for each
4132 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4133 */
4134 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4135 if (jiffies_till_first_fqs == ULONG_MAX)
4136 jiffies_till_first_fqs = d;
4137 if (jiffies_till_next_fqs == ULONG_MAX)
4138 jiffies_till_next_fqs = d;
4139
4140 /* If the compile-time values are accurate, just leave. */
4141 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4142 nr_cpu_ids == NR_CPUS)
4143 return;
4144 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
4145 rcu_fanout_leaf, nr_cpu_ids);
4146
4147 /*
4148 * The boot-time rcu_fanout_leaf parameter must be at least two
4149 * and cannot exceed the number of bits in the rcu_node masks.
4150 * Complain and fall back to the compile-time values if this
4151 * limit is exceeded.
4152 */
4153 if (rcu_fanout_leaf < 2 ||
4154 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4155 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4156 WARN_ON(1);
4157 return;
4158 }
4159
4160 /*
4161 * Compute number of nodes that can be handled an rcu_node tree
4162 * with the given number of levels.
4163 */
4164 rcu_capacity[0] = rcu_fanout_leaf;
4165 for (i = 1; i < RCU_NUM_LVLS; i++)
4166 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4167
4168 /*
4169 * The tree must be able to accommodate the configured number of CPUs.
4170 * If this limit is exceeded, fall back to the compile-time values.
4171 */
4172 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4173 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4174 WARN_ON(1);
4175 return;
4176 }
4177
4178 /* Calculate the number of levels in the tree. */
4179 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4180 }
4181 rcu_num_lvls = i + 1;
4182
4183 /* Calculate the number of rcu_nodes at each level of the tree. */
4184 for (i = 0; i < rcu_num_lvls; i++) {
4185 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4186 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4187 }
4188
4189 /* Calculate the total number of rcu_node structures. */
4190 rcu_num_nodes = 0;
4191 for (i = 0; i < rcu_num_lvls; i++)
4192 rcu_num_nodes += num_rcu_lvl[i];
4193}
4194
4195/*
4196 * Dump out the structure of the rcu_node combining tree associated
4197 * with the rcu_state structure referenced by rsp.
4198 */
4199static void __init rcu_dump_rcu_node_tree(struct rcu_state *rsp)
4200{
4201 int level = 0;
4202 struct rcu_node *rnp;
4203
4204 pr_info("rcu_node tree layout dump\n");
4205 pr_info(" ");
4206 rcu_for_each_node_breadth_first(rsp, rnp) {
4207 if (rnp->level != level) {
4208 pr_cont("\n");
4209 pr_info(" ");
4210 level = rnp->level;
4211 }
4212 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4213 }
4214 pr_cont("\n");
4215}
4216
4217void __init rcu_init(void)
4218{
4219 int cpu;
4220
4221 rcu_early_boot_tests();
4222
4223 rcu_bootup_announce();
4224 rcu_init_geometry();
4225 rcu_init_one(&rcu_bh_state);
4226 rcu_init_one(&rcu_sched_state);
4227 if (dump_tree)
4228 rcu_dump_rcu_node_tree(&rcu_sched_state);
4229 __rcu_init_preempt();
4230 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
4231
4232 /*
4233 * We don't need protection against CPU-hotplug here because
4234 * this is called early in boot, before either interrupts
4235 * or the scheduler are operational.
4236 */
4237 pm_notifier(rcu_pm_notify, 0);
4238 for_each_online_cpu(cpu) {
4239 rcutree_prepare_cpu(cpu);
4240 rcu_cpu_starting(cpu);
4241 }
4242}
4243
4244#include "tree_exp.h"
4245#include "tree_plugin.h"
1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
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.
8 *
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.
13 *
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.
17 *
18 * Copyright IBM Corporation, 2008
19 *
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
23 *
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
26 *
27 * For detailed explanation of Read-Copy Update mechanism see -
28 * Documentation/RCU
29 */
30#include <linux/types.h>
31#include <linux/kernel.h>
32#include <linux/init.h>
33#include <linux/spinlock.h>
34#include <linux/smp.h>
35#include <linux/rcupdate.h>
36#include <linux/interrupt.h>
37#include <linux/sched.h>
38#include <linux/nmi.h>
39#include <linux/atomic.h>
40#include <linux/bitops.h>
41#include <linux/export.h>
42#include <linux/completion.h>
43#include <linux/moduleparam.h>
44#include <linux/module.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 <linux/prefetch.h>
54#include <linux/delay.h>
55#include <linux/stop_machine.h>
56#include <linux/random.h>
57#include <linux/ftrace_event.h>
58#include <linux/suspend.h>
59
60#include "tree.h"
61#include "rcu.h"
62
63MODULE_ALIAS("rcutree");
64#ifdef MODULE_PARAM_PREFIX
65#undef MODULE_PARAM_PREFIX
66#endif
67#define MODULE_PARAM_PREFIX "rcutree."
68
69/* Data structures. */
70
71static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
72static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
73
74/*
75 * In order to export the rcu_state name to the tracing tools, it
76 * needs to be added in the __tracepoint_string section.
77 * This requires defining a separate variable tp_<sname>_varname
78 * that points to the string being used, and this will allow
79 * the tracing userspace tools to be able to decipher the string
80 * address to the matching string.
81 */
82#define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
83static char sname##_varname[] = #sname; \
84static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
85struct rcu_state sname##_state = { \
86 .level = { &sname##_state.node[0] }, \
87 .call = cr, \
88 .fqs_state = RCU_GP_IDLE, \
89 .gpnum = 0UL - 300UL, \
90 .completed = 0UL - 300UL, \
91 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
92 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
93 .orphan_donetail = &sname##_state.orphan_donelist, \
94 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
95 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
96 .name = sname##_varname, \
97 .abbr = sabbr, \
98}; \
99DEFINE_PER_CPU(struct rcu_data, sname##_data)
100
101RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
102RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
103
104static struct rcu_state *rcu_state;
105LIST_HEAD(rcu_struct_flavors);
106
107/* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
108static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
109module_param(rcu_fanout_leaf, int, 0444);
110int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
111static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
112 NUM_RCU_LVL_0,
113 NUM_RCU_LVL_1,
114 NUM_RCU_LVL_2,
115 NUM_RCU_LVL_3,
116 NUM_RCU_LVL_4,
117};
118int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
119
120/*
121 * The rcu_scheduler_active variable transitions from zero to one just
122 * before the first task is spawned. So when this variable is zero, RCU
123 * can assume that there is but one task, allowing RCU to (for example)
124 * optimize synchronize_sched() to a simple barrier(). When this variable
125 * is one, RCU must actually do all the hard work required to detect real
126 * grace periods. This variable is also used to suppress boot-time false
127 * positives from lockdep-RCU error checking.
128 */
129int rcu_scheduler_active __read_mostly;
130EXPORT_SYMBOL_GPL(rcu_scheduler_active);
131
132/*
133 * The rcu_scheduler_fully_active variable transitions from zero to one
134 * during the early_initcall() processing, which is after the scheduler
135 * is capable of creating new tasks. So RCU processing (for example,
136 * creating tasks for RCU priority boosting) must be delayed until after
137 * rcu_scheduler_fully_active transitions from zero to one. We also
138 * currently delay invocation of any RCU callbacks until after this point.
139 *
140 * It might later prove better for people registering RCU callbacks during
141 * early boot to take responsibility for these callbacks, but one step at
142 * a time.
143 */
144static int rcu_scheduler_fully_active __read_mostly;
145
146#ifdef CONFIG_RCU_BOOST
147
148/*
149 * Control variables for per-CPU and per-rcu_node kthreads. These
150 * handle all flavors of RCU.
151 */
152static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
153DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
154DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
155DEFINE_PER_CPU(char, rcu_cpu_has_work);
156
157#endif /* #ifdef CONFIG_RCU_BOOST */
158
159static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
160static void invoke_rcu_core(void);
161static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
162
163/*
164 * Track the rcutorture test sequence number and the update version
165 * number within a given test. The rcutorture_testseq is incremented
166 * on every rcutorture module load and unload, so has an odd value
167 * when a test is running. The rcutorture_vernum is set to zero
168 * when rcutorture starts and is incremented on each rcutorture update.
169 * These variables enable correlating rcutorture output with the
170 * RCU tracing information.
171 */
172unsigned long rcutorture_testseq;
173unsigned long rcutorture_vernum;
174
175/*
176 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
177 * permit this function to be invoked without holding the root rcu_node
178 * structure's ->lock, but of course results can be subject to change.
179 */
180static int rcu_gp_in_progress(struct rcu_state *rsp)
181{
182 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
183}
184
185/*
186 * Note a quiescent state. Because we do not need to know
187 * how many quiescent states passed, just if there was at least
188 * one since the start of the grace period, this just sets a flag.
189 * The caller must have disabled preemption.
190 */
191void rcu_sched_qs(int cpu)
192{
193 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
194
195 if (rdp->passed_quiesce == 0)
196 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
197 rdp->passed_quiesce = 1;
198}
199
200void rcu_bh_qs(int cpu)
201{
202 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
203
204 if (rdp->passed_quiesce == 0)
205 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
206 rdp->passed_quiesce = 1;
207}
208
209/*
210 * Note a context switch. This is a quiescent state for RCU-sched,
211 * and requires special handling for preemptible RCU.
212 * The caller must have disabled preemption.
213 */
214void rcu_note_context_switch(int cpu)
215{
216 trace_rcu_utilization(TPS("Start context switch"));
217 rcu_sched_qs(cpu);
218 rcu_preempt_note_context_switch(cpu);
219 trace_rcu_utilization(TPS("End context switch"));
220}
221EXPORT_SYMBOL_GPL(rcu_note_context_switch);
222
223static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
224 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
225 .dynticks = ATOMIC_INIT(1),
226#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
227 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
228 .dynticks_idle = ATOMIC_INIT(1),
229#endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
230};
231
232static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
233static long qhimark = 10000; /* If this many pending, ignore blimit. */
234static long qlowmark = 100; /* Once only this many pending, use blimit. */
235
236module_param(blimit, long, 0444);
237module_param(qhimark, long, 0444);
238module_param(qlowmark, long, 0444);
239
240static ulong jiffies_till_first_fqs = ULONG_MAX;
241static ulong jiffies_till_next_fqs = ULONG_MAX;
242
243module_param(jiffies_till_first_fqs, ulong, 0644);
244module_param(jiffies_till_next_fqs, ulong, 0644);
245
246static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
247 struct rcu_data *rdp);
248static void force_qs_rnp(struct rcu_state *rsp,
249 int (*f)(struct rcu_data *rsp, bool *isidle,
250 unsigned long *maxj),
251 bool *isidle, unsigned long *maxj);
252static void force_quiescent_state(struct rcu_state *rsp);
253static int rcu_pending(int cpu);
254
255/*
256 * Return the number of RCU-sched batches processed thus far for debug & stats.
257 */
258long rcu_batches_completed_sched(void)
259{
260 return rcu_sched_state.completed;
261}
262EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
263
264/*
265 * Return the number of RCU BH batches processed thus far for debug & stats.
266 */
267long rcu_batches_completed_bh(void)
268{
269 return rcu_bh_state.completed;
270}
271EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
272
273/*
274 * Force a quiescent state for RCU BH.
275 */
276void rcu_bh_force_quiescent_state(void)
277{
278 force_quiescent_state(&rcu_bh_state);
279}
280EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
281
282/*
283 * Record the number of times rcutorture tests have been initiated and
284 * terminated. This information allows the debugfs tracing stats to be
285 * correlated to the rcutorture messages, even when the rcutorture module
286 * is being repeatedly loaded and unloaded. In other words, we cannot
287 * store this state in rcutorture itself.
288 */
289void rcutorture_record_test_transition(void)
290{
291 rcutorture_testseq++;
292 rcutorture_vernum = 0;
293}
294EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
295
296/*
297 * Record the number of writer passes through the current rcutorture test.
298 * This is also used to correlate debugfs tracing stats with the rcutorture
299 * messages.
300 */
301void rcutorture_record_progress(unsigned long vernum)
302{
303 rcutorture_vernum++;
304}
305EXPORT_SYMBOL_GPL(rcutorture_record_progress);
306
307/*
308 * Force a quiescent state for RCU-sched.
309 */
310void rcu_sched_force_quiescent_state(void)
311{
312 force_quiescent_state(&rcu_sched_state);
313}
314EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
315
316/*
317 * Does the CPU have callbacks ready to be invoked?
318 */
319static int
320cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
321{
322 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
323 rdp->nxttail[RCU_DONE_TAIL] != NULL;
324}
325
326/*
327 * Does the current CPU require a not-yet-started grace period?
328 * The caller must have disabled interrupts to prevent races with
329 * normal callback registry.
330 */
331static int
332cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
333{
334 int i;
335
336 if (rcu_gp_in_progress(rsp))
337 return 0; /* No, a grace period is already in progress. */
338 if (rcu_nocb_needs_gp(rsp))
339 return 1; /* Yes, a no-CBs CPU needs one. */
340 if (!rdp->nxttail[RCU_NEXT_TAIL])
341 return 0; /* No, this is a no-CBs (or offline) CPU. */
342 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
343 return 1; /* Yes, this CPU has newly registered callbacks. */
344 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
345 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
346 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
347 rdp->nxtcompleted[i]))
348 return 1; /* Yes, CBs for future grace period. */
349 return 0; /* No grace period needed. */
350}
351
352/*
353 * Return the root node of the specified rcu_state structure.
354 */
355static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
356{
357 return &rsp->node[0];
358}
359
360/*
361 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
362 *
363 * If the new value of the ->dynticks_nesting counter now is zero,
364 * we really have entered idle, and must do the appropriate accounting.
365 * The caller must have disabled interrupts.
366 */
367static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
368 bool user)
369{
370 struct rcu_state *rsp;
371 struct rcu_data *rdp;
372
373 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
374 if (!user && !is_idle_task(current)) {
375 struct task_struct *idle __maybe_unused =
376 idle_task(smp_processor_id());
377
378 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
379 ftrace_dump(DUMP_ORIG);
380 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
381 current->pid, current->comm,
382 idle->pid, idle->comm); /* must be idle task! */
383 }
384 for_each_rcu_flavor(rsp) {
385 rdp = this_cpu_ptr(rsp->rda);
386 do_nocb_deferred_wakeup(rdp);
387 }
388 rcu_prepare_for_idle(smp_processor_id());
389 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
390 smp_mb__before_atomic_inc(); /* See above. */
391 atomic_inc(&rdtp->dynticks);
392 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
393 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
394
395 /*
396 * It is illegal to enter an extended quiescent state while
397 * in an RCU read-side critical section.
398 */
399 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
400 "Illegal idle entry in RCU read-side critical section.");
401 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
402 "Illegal idle entry in RCU-bh read-side critical section.");
403 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
404 "Illegal idle entry in RCU-sched read-side critical section.");
405}
406
407/*
408 * Enter an RCU extended quiescent state, which can be either the
409 * idle loop or adaptive-tickless usermode execution.
410 */
411static void rcu_eqs_enter(bool user)
412{
413 long long oldval;
414 struct rcu_dynticks *rdtp;
415
416 rdtp = this_cpu_ptr(&rcu_dynticks);
417 oldval = rdtp->dynticks_nesting;
418 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
419 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE) {
420 rdtp->dynticks_nesting = 0;
421 rcu_eqs_enter_common(rdtp, oldval, user);
422 } else {
423 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
424 }
425}
426
427/**
428 * rcu_idle_enter - inform RCU that current CPU is entering idle
429 *
430 * Enter idle mode, in other words, -leave- the mode in which RCU
431 * read-side critical sections can occur. (Though RCU read-side
432 * critical sections can occur in irq handlers in idle, a possibility
433 * handled by irq_enter() and irq_exit().)
434 *
435 * We crowbar the ->dynticks_nesting field to zero to allow for
436 * the possibility of usermode upcalls having messed up our count
437 * of interrupt nesting level during the prior busy period.
438 */
439void rcu_idle_enter(void)
440{
441 unsigned long flags;
442
443 local_irq_save(flags);
444 rcu_eqs_enter(false);
445 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
446 local_irq_restore(flags);
447}
448EXPORT_SYMBOL_GPL(rcu_idle_enter);
449
450#ifdef CONFIG_RCU_USER_QS
451/**
452 * rcu_user_enter - inform RCU that we are resuming userspace.
453 *
454 * Enter RCU idle mode right before resuming userspace. No use of RCU
455 * is permitted between this call and rcu_user_exit(). This way the
456 * CPU doesn't need to maintain the tick for RCU maintenance purposes
457 * when the CPU runs in userspace.
458 */
459void rcu_user_enter(void)
460{
461 rcu_eqs_enter(1);
462}
463#endif /* CONFIG_RCU_USER_QS */
464
465/**
466 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
467 *
468 * Exit from an interrupt handler, which might possibly result in entering
469 * idle mode, in other words, leaving the mode in which read-side critical
470 * sections can occur.
471 *
472 * This code assumes that the idle loop never does anything that might
473 * result in unbalanced calls to irq_enter() and irq_exit(). If your
474 * architecture violates this assumption, RCU will give you what you
475 * deserve, good and hard. But very infrequently and irreproducibly.
476 *
477 * Use things like work queues to work around this limitation.
478 *
479 * You have been warned.
480 */
481void rcu_irq_exit(void)
482{
483 unsigned long flags;
484 long long oldval;
485 struct rcu_dynticks *rdtp;
486
487 local_irq_save(flags);
488 rdtp = this_cpu_ptr(&rcu_dynticks);
489 oldval = rdtp->dynticks_nesting;
490 rdtp->dynticks_nesting--;
491 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
492 if (rdtp->dynticks_nesting)
493 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
494 else
495 rcu_eqs_enter_common(rdtp, oldval, true);
496 rcu_sysidle_enter(rdtp, 1);
497 local_irq_restore(flags);
498}
499
500/*
501 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
502 *
503 * If the new value of the ->dynticks_nesting counter was previously zero,
504 * we really have exited idle, and must do the appropriate accounting.
505 * The caller must have disabled interrupts.
506 */
507static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
508 int user)
509{
510 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
511 atomic_inc(&rdtp->dynticks);
512 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
513 smp_mb__after_atomic_inc(); /* See above. */
514 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
515 rcu_cleanup_after_idle(smp_processor_id());
516 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
517 if (!user && !is_idle_task(current)) {
518 struct task_struct *idle __maybe_unused =
519 idle_task(smp_processor_id());
520
521 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
522 oldval, rdtp->dynticks_nesting);
523 ftrace_dump(DUMP_ORIG);
524 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
525 current->pid, current->comm,
526 idle->pid, idle->comm); /* must be idle task! */
527 }
528}
529
530/*
531 * Exit an RCU extended quiescent state, which can be either the
532 * idle loop or adaptive-tickless usermode execution.
533 */
534static void rcu_eqs_exit(bool user)
535{
536 struct rcu_dynticks *rdtp;
537 long long oldval;
538
539 rdtp = this_cpu_ptr(&rcu_dynticks);
540 oldval = rdtp->dynticks_nesting;
541 WARN_ON_ONCE(oldval < 0);
542 if (oldval & DYNTICK_TASK_NEST_MASK) {
543 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
544 } else {
545 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
546 rcu_eqs_exit_common(rdtp, oldval, user);
547 }
548}
549
550/**
551 * rcu_idle_exit - inform RCU that current CPU is leaving idle
552 *
553 * Exit idle mode, in other words, -enter- the mode in which RCU
554 * read-side critical sections can occur.
555 *
556 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
557 * allow for the possibility of usermode upcalls messing up our count
558 * of interrupt nesting level during the busy period that is just
559 * now starting.
560 */
561void rcu_idle_exit(void)
562{
563 unsigned long flags;
564
565 local_irq_save(flags);
566 rcu_eqs_exit(false);
567 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
568 local_irq_restore(flags);
569}
570EXPORT_SYMBOL_GPL(rcu_idle_exit);
571
572#ifdef CONFIG_RCU_USER_QS
573/**
574 * rcu_user_exit - inform RCU that we are exiting userspace.
575 *
576 * Exit RCU idle mode while entering the kernel because it can
577 * run a RCU read side critical section anytime.
578 */
579void rcu_user_exit(void)
580{
581 rcu_eqs_exit(1);
582}
583#endif /* CONFIG_RCU_USER_QS */
584
585/**
586 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
587 *
588 * Enter an interrupt handler, which might possibly result in exiting
589 * idle mode, in other words, entering the mode in which read-side critical
590 * sections can occur.
591 *
592 * Note that the Linux kernel is fully capable of entering an interrupt
593 * handler that it never exits, for example when doing upcalls to
594 * user mode! This code assumes that the idle loop never does upcalls to
595 * user mode. If your architecture does do upcalls from the idle loop (or
596 * does anything else that results in unbalanced calls to the irq_enter()
597 * and irq_exit() functions), RCU will give you what you deserve, good
598 * and hard. But very infrequently and irreproducibly.
599 *
600 * Use things like work queues to work around this limitation.
601 *
602 * You have been warned.
603 */
604void rcu_irq_enter(void)
605{
606 unsigned long flags;
607 struct rcu_dynticks *rdtp;
608 long long oldval;
609
610 local_irq_save(flags);
611 rdtp = this_cpu_ptr(&rcu_dynticks);
612 oldval = rdtp->dynticks_nesting;
613 rdtp->dynticks_nesting++;
614 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
615 if (oldval)
616 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
617 else
618 rcu_eqs_exit_common(rdtp, oldval, true);
619 rcu_sysidle_exit(rdtp, 1);
620 local_irq_restore(flags);
621}
622
623/**
624 * rcu_nmi_enter - inform RCU of entry to NMI context
625 *
626 * If the CPU was idle with dynamic ticks active, and there is no
627 * irq handler running, this updates rdtp->dynticks_nmi to let the
628 * RCU grace-period handling know that the CPU is active.
629 */
630void rcu_nmi_enter(void)
631{
632 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
633
634 if (rdtp->dynticks_nmi_nesting == 0 &&
635 (atomic_read(&rdtp->dynticks) & 0x1))
636 return;
637 rdtp->dynticks_nmi_nesting++;
638 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
639 atomic_inc(&rdtp->dynticks);
640 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
641 smp_mb__after_atomic_inc(); /* See above. */
642 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
643}
644
645/**
646 * rcu_nmi_exit - inform RCU of exit from NMI context
647 *
648 * If the CPU was idle with dynamic ticks active, and there is no
649 * irq handler running, this updates rdtp->dynticks_nmi to let the
650 * RCU grace-period handling know that the CPU is no longer active.
651 */
652void rcu_nmi_exit(void)
653{
654 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
655
656 if (rdtp->dynticks_nmi_nesting == 0 ||
657 --rdtp->dynticks_nmi_nesting != 0)
658 return;
659 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
660 smp_mb__before_atomic_inc(); /* See above. */
661 atomic_inc(&rdtp->dynticks);
662 smp_mb__after_atomic_inc(); /* Force delay to next write. */
663 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
664}
665
666/**
667 * __rcu_is_watching - are RCU read-side critical sections safe?
668 *
669 * Return true if RCU is watching the running CPU, which means that
670 * this CPU can safely enter RCU read-side critical sections. Unlike
671 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
672 * least disabled preemption.
673 */
674bool notrace __rcu_is_watching(void)
675{
676 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
677}
678
679/**
680 * rcu_is_watching - see if RCU thinks that the current CPU is idle
681 *
682 * If the current CPU is in its idle loop and is neither in an interrupt
683 * or NMI handler, return true.
684 */
685bool notrace rcu_is_watching(void)
686{
687 int ret;
688
689 preempt_disable();
690 ret = __rcu_is_watching();
691 preempt_enable();
692 return ret;
693}
694EXPORT_SYMBOL_GPL(rcu_is_watching);
695
696#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
697
698/*
699 * Is the current CPU online? Disable preemption to avoid false positives
700 * that could otherwise happen due to the current CPU number being sampled,
701 * this task being preempted, its old CPU being taken offline, resuming
702 * on some other CPU, then determining that its old CPU is now offline.
703 * It is OK to use RCU on an offline processor during initial boot, hence
704 * the check for rcu_scheduler_fully_active. Note also that it is OK
705 * for a CPU coming online to use RCU for one jiffy prior to marking itself
706 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
707 * offline to continue to use RCU for one jiffy after marking itself
708 * offline in the cpu_online_mask. This leniency is necessary given the
709 * non-atomic nature of the online and offline processing, for example,
710 * the fact that a CPU enters the scheduler after completing the CPU_DYING
711 * notifiers.
712 *
713 * This is also why RCU internally marks CPUs online during the
714 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
715 *
716 * Disable checking if in an NMI handler because we cannot safely report
717 * errors from NMI handlers anyway.
718 */
719bool rcu_lockdep_current_cpu_online(void)
720{
721 struct rcu_data *rdp;
722 struct rcu_node *rnp;
723 bool ret;
724
725 if (in_nmi())
726 return true;
727 preempt_disable();
728 rdp = this_cpu_ptr(&rcu_sched_data);
729 rnp = rdp->mynode;
730 ret = (rdp->grpmask & rnp->qsmaskinit) ||
731 !rcu_scheduler_fully_active;
732 preempt_enable();
733 return ret;
734}
735EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
736
737#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
738
739/**
740 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
741 *
742 * If the current CPU is idle or running at a first-level (not nested)
743 * interrupt from idle, return true. The caller must have at least
744 * disabled preemption.
745 */
746static int rcu_is_cpu_rrupt_from_idle(void)
747{
748 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
749}
750
751/*
752 * Snapshot the specified CPU's dynticks counter so that we can later
753 * credit them with an implicit quiescent state. Return 1 if this CPU
754 * is in dynticks idle mode, which is an extended quiescent state.
755 */
756static int dyntick_save_progress_counter(struct rcu_data *rdp,
757 bool *isidle, unsigned long *maxj)
758{
759 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
760 rcu_sysidle_check_cpu(rdp, isidle, maxj);
761 return (rdp->dynticks_snap & 0x1) == 0;
762}
763
764/*
765 * This function really isn't for public consumption, but RCU is special in
766 * that context switches can allow the state machine to make progress.
767 */
768extern void resched_cpu(int cpu);
769
770/*
771 * Return true if the specified CPU has passed through a quiescent
772 * state by virtue of being in or having passed through an dynticks
773 * idle state since the last call to dyntick_save_progress_counter()
774 * for this same CPU, or by virtue of having been offline.
775 */
776static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
777 bool *isidle, unsigned long *maxj)
778{
779 unsigned int curr;
780 unsigned int snap;
781
782 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
783 snap = (unsigned int)rdp->dynticks_snap;
784
785 /*
786 * If the CPU passed through or entered a dynticks idle phase with
787 * no active irq/NMI handlers, then we can safely pretend that the CPU
788 * already acknowledged the request to pass through a quiescent
789 * state. Either way, that CPU cannot possibly be in an RCU
790 * read-side critical section that started before the beginning
791 * of the current RCU grace period.
792 */
793 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
794 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
795 rdp->dynticks_fqs++;
796 return 1;
797 }
798
799 /*
800 * Check for the CPU being offline, but only if the grace period
801 * is old enough. We don't need to worry about the CPU changing
802 * state: If we see it offline even once, it has been through a
803 * quiescent state.
804 *
805 * The reason for insisting that the grace period be at least
806 * one jiffy old is that CPUs that are not quite online and that
807 * have just gone offline can still execute RCU read-side critical
808 * sections.
809 */
810 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
811 return 0; /* Grace period is not old enough. */
812 barrier();
813 if (cpu_is_offline(rdp->cpu)) {
814 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
815 rdp->offline_fqs++;
816 return 1;
817 }
818
819 /*
820 * There is a possibility that a CPU in adaptive-ticks state
821 * might run in the kernel with the scheduling-clock tick disabled
822 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
823 * force the CPU to restart the scheduling-clock tick in this
824 * CPU is in this state.
825 */
826 rcu_kick_nohz_cpu(rdp->cpu);
827
828 /*
829 * Alternatively, the CPU might be running in the kernel
830 * for an extended period of time without a quiescent state.
831 * Attempt to force the CPU through the scheduler to gain the
832 * needed quiescent state, but only if the grace period has gone
833 * on for an uncommonly long time. If there are many stuck CPUs,
834 * we will beat on the first one until it gets unstuck, then move
835 * to the next. Only do this for the primary flavor of RCU.
836 */
837 if (rdp->rsp == rcu_state &&
838 ULONG_CMP_GE(jiffies, rdp->rsp->jiffies_resched)) {
839 rdp->rsp->jiffies_resched += 5;
840 resched_cpu(rdp->cpu);
841 }
842
843 return 0;
844}
845
846static void record_gp_stall_check_time(struct rcu_state *rsp)
847{
848 unsigned long j = jiffies;
849 unsigned long j1;
850
851 rsp->gp_start = j;
852 smp_wmb(); /* Record start time before stall time. */
853 j1 = rcu_jiffies_till_stall_check();
854 rsp->jiffies_stall = j + j1;
855 rsp->jiffies_resched = j + j1 / 2;
856}
857
858/*
859 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
860 * for architectures that do not implement trigger_all_cpu_backtrace().
861 * The NMI-triggered stack traces are more accurate because they are
862 * printed by the target CPU.
863 */
864static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
865{
866 int cpu;
867 unsigned long flags;
868 struct rcu_node *rnp;
869
870 rcu_for_each_leaf_node(rsp, rnp) {
871 raw_spin_lock_irqsave(&rnp->lock, flags);
872 if (rnp->qsmask != 0) {
873 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
874 if (rnp->qsmask & (1UL << cpu))
875 dump_cpu_task(rnp->grplo + cpu);
876 }
877 raw_spin_unlock_irqrestore(&rnp->lock, flags);
878 }
879}
880
881static void print_other_cpu_stall(struct rcu_state *rsp)
882{
883 int cpu;
884 long delta;
885 unsigned long flags;
886 int ndetected = 0;
887 struct rcu_node *rnp = rcu_get_root(rsp);
888 long totqlen = 0;
889
890 /* Only let one CPU complain about others per time interval. */
891
892 raw_spin_lock_irqsave(&rnp->lock, flags);
893 delta = jiffies - rsp->jiffies_stall;
894 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
895 raw_spin_unlock_irqrestore(&rnp->lock, flags);
896 return;
897 }
898 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
899 raw_spin_unlock_irqrestore(&rnp->lock, flags);
900
901 /*
902 * OK, time to rat on our buddy...
903 * See Documentation/RCU/stallwarn.txt for info on how to debug
904 * RCU CPU stall warnings.
905 */
906 pr_err("INFO: %s detected stalls on CPUs/tasks:",
907 rsp->name);
908 print_cpu_stall_info_begin();
909 rcu_for_each_leaf_node(rsp, rnp) {
910 raw_spin_lock_irqsave(&rnp->lock, flags);
911 ndetected += rcu_print_task_stall(rnp);
912 if (rnp->qsmask != 0) {
913 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
914 if (rnp->qsmask & (1UL << cpu)) {
915 print_cpu_stall_info(rsp,
916 rnp->grplo + cpu);
917 ndetected++;
918 }
919 }
920 raw_spin_unlock_irqrestore(&rnp->lock, flags);
921 }
922
923 /*
924 * Now rat on any tasks that got kicked up to the root rcu_node
925 * due to CPU offlining.
926 */
927 rnp = rcu_get_root(rsp);
928 raw_spin_lock_irqsave(&rnp->lock, flags);
929 ndetected += rcu_print_task_stall(rnp);
930 raw_spin_unlock_irqrestore(&rnp->lock, flags);
931
932 print_cpu_stall_info_end();
933 for_each_possible_cpu(cpu)
934 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
935 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
936 smp_processor_id(), (long)(jiffies - rsp->gp_start),
937 rsp->gpnum, rsp->completed, totqlen);
938 if (ndetected == 0)
939 pr_err("INFO: Stall ended before state dump start\n");
940 else if (!trigger_all_cpu_backtrace())
941 rcu_dump_cpu_stacks(rsp);
942
943 /* Complain about tasks blocking the grace period. */
944
945 rcu_print_detail_task_stall(rsp);
946
947 force_quiescent_state(rsp); /* Kick them all. */
948}
949
950/*
951 * This function really isn't for public consumption, but RCU is special in
952 * that context switches can allow the state machine to make progress.
953 */
954extern void resched_cpu(int cpu);
955
956static void print_cpu_stall(struct rcu_state *rsp)
957{
958 int cpu;
959 unsigned long flags;
960 struct rcu_node *rnp = rcu_get_root(rsp);
961 long totqlen = 0;
962
963 /*
964 * OK, time to rat on ourselves...
965 * See Documentation/RCU/stallwarn.txt for info on how to debug
966 * RCU CPU stall warnings.
967 */
968 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
969 print_cpu_stall_info_begin();
970 print_cpu_stall_info(rsp, smp_processor_id());
971 print_cpu_stall_info_end();
972 for_each_possible_cpu(cpu)
973 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
974 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
975 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
976 if (!trigger_all_cpu_backtrace())
977 dump_stack();
978
979 raw_spin_lock_irqsave(&rnp->lock, flags);
980 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
981 rsp->jiffies_stall = jiffies +
982 3 * rcu_jiffies_till_stall_check() + 3;
983 raw_spin_unlock_irqrestore(&rnp->lock, flags);
984
985 /*
986 * Attempt to revive the RCU machinery by forcing a context switch.
987 *
988 * A context switch would normally allow the RCU state machine to make
989 * progress and it could be we're stuck in kernel space without context
990 * switches for an entirely unreasonable amount of time.
991 */
992 resched_cpu(smp_processor_id());
993}
994
995static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
996{
997 unsigned long completed;
998 unsigned long gpnum;
999 unsigned long gps;
1000 unsigned long j;
1001 unsigned long js;
1002 struct rcu_node *rnp;
1003
1004 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
1005 return;
1006 j = jiffies;
1007
1008 /*
1009 * Lots of memory barriers to reject false positives.
1010 *
1011 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
1012 * then rsp->gp_start, and finally rsp->completed. These values
1013 * are updated in the opposite order with memory barriers (or
1014 * equivalent) during grace-period initialization and cleanup.
1015 * Now, a false positive can occur if we get an new value of
1016 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
1017 * the memory barriers, the only way that this can happen is if one
1018 * grace period ends and another starts between these two fetches.
1019 * Detect this by comparing rsp->completed with the previous fetch
1020 * from rsp->gpnum.
1021 *
1022 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
1023 * and rsp->gp_start suffice to forestall false positives.
1024 */
1025 gpnum = ACCESS_ONCE(rsp->gpnum);
1026 smp_rmb(); /* Pick up ->gpnum first... */
1027 js = ACCESS_ONCE(rsp->jiffies_stall);
1028 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
1029 gps = ACCESS_ONCE(rsp->gp_start);
1030 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1031 completed = ACCESS_ONCE(rsp->completed);
1032 if (ULONG_CMP_GE(completed, gpnum) ||
1033 ULONG_CMP_LT(j, js) ||
1034 ULONG_CMP_GE(gps, js))
1035 return; /* No stall or GP completed since entering function. */
1036 rnp = rdp->mynode;
1037 if (rcu_gp_in_progress(rsp) &&
1038 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1039
1040 /* We haven't checked in, so go dump stack. */
1041 print_cpu_stall(rsp);
1042
1043 } else if (rcu_gp_in_progress(rsp) &&
1044 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1045
1046 /* They had a few time units to dump stack, so complain. */
1047 print_other_cpu_stall(rsp);
1048 }
1049}
1050
1051/**
1052 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1053 *
1054 * Set the stall-warning timeout way off into the future, thus preventing
1055 * any RCU CPU stall-warning messages from appearing in the current set of
1056 * RCU grace periods.
1057 *
1058 * The caller must disable hard irqs.
1059 */
1060void rcu_cpu_stall_reset(void)
1061{
1062 struct rcu_state *rsp;
1063
1064 for_each_rcu_flavor(rsp)
1065 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1066}
1067
1068/*
1069 * Initialize the specified rcu_data structure's callback list to empty.
1070 */
1071static void init_callback_list(struct rcu_data *rdp)
1072{
1073 int i;
1074
1075 if (init_nocb_callback_list(rdp))
1076 return;
1077 rdp->nxtlist = NULL;
1078 for (i = 0; i < RCU_NEXT_SIZE; i++)
1079 rdp->nxttail[i] = &rdp->nxtlist;
1080}
1081
1082/*
1083 * Determine the value that ->completed will have at the end of the
1084 * next subsequent grace period. This is used to tag callbacks so that
1085 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1086 * been dyntick-idle for an extended period with callbacks under the
1087 * influence of RCU_FAST_NO_HZ.
1088 *
1089 * The caller must hold rnp->lock with interrupts disabled.
1090 */
1091static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1092 struct rcu_node *rnp)
1093{
1094 /*
1095 * If RCU is idle, we just wait for the next grace period.
1096 * But we can only be sure that RCU is idle if we are looking
1097 * at the root rcu_node structure -- otherwise, a new grace
1098 * period might have started, but just not yet gotten around
1099 * to initializing the current non-root rcu_node structure.
1100 */
1101 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1102 return rnp->completed + 1;
1103
1104 /*
1105 * Otherwise, wait for a possible partial grace period and
1106 * then the subsequent full grace period.
1107 */
1108 return rnp->completed + 2;
1109}
1110
1111/*
1112 * Trace-event helper function for rcu_start_future_gp() and
1113 * rcu_nocb_wait_gp().
1114 */
1115static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1116 unsigned long c, const char *s)
1117{
1118 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1119 rnp->completed, c, rnp->level,
1120 rnp->grplo, rnp->grphi, s);
1121}
1122
1123/*
1124 * Start some future grace period, as needed to handle newly arrived
1125 * callbacks. The required future grace periods are recorded in each
1126 * rcu_node structure's ->need_future_gp field.
1127 *
1128 * The caller must hold the specified rcu_node structure's ->lock.
1129 */
1130static unsigned long __maybe_unused
1131rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1132{
1133 unsigned long c;
1134 int i;
1135 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1136
1137 /*
1138 * Pick up grace-period number for new callbacks. If this
1139 * grace period is already marked as needed, return to the caller.
1140 */
1141 c = rcu_cbs_completed(rdp->rsp, rnp);
1142 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1143 if (rnp->need_future_gp[c & 0x1]) {
1144 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1145 return c;
1146 }
1147
1148 /*
1149 * If either this rcu_node structure or the root rcu_node structure
1150 * believe that a grace period is in progress, then we must wait
1151 * for the one following, which is in "c". Because our request
1152 * will be noticed at the end of the current grace period, we don't
1153 * need to explicitly start one.
1154 */
1155 if (rnp->gpnum != rnp->completed ||
1156 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1157 rnp->need_future_gp[c & 0x1]++;
1158 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1159 return c;
1160 }
1161
1162 /*
1163 * There might be no grace period in progress. If we don't already
1164 * hold it, acquire the root rcu_node structure's lock in order to
1165 * start one (if needed).
1166 */
1167 if (rnp != rnp_root) {
1168 raw_spin_lock(&rnp_root->lock);
1169 smp_mb__after_unlock_lock();
1170 }
1171
1172 /*
1173 * Get a new grace-period number. If there really is no grace
1174 * period in progress, it will be smaller than the one we obtained
1175 * earlier. Adjust callbacks as needed. Note that even no-CBs
1176 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1177 */
1178 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1179 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1180 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1181 rdp->nxtcompleted[i] = c;
1182
1183 /*
1184 * If the needed for the required grace period is already
1185 * recorded, trace and leave.
1186 */
1187 if (rnp_root->need_future_gp[c & 0x1]) {
1188 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1189 goto unlock_out;
1190 }
1191
1192 /* Record the need for the future grace period. */
1193 rnp_root->need_future_gp[c & 0x1]++;
1194
1195 /* If a grace period is not already in progress, start one. */
1196 if (rnp_root->gpnum != rnp_root->completed) {
1197 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1198 } else {
1199 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1200 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1201 }
1202unlock_out:
1203 if (rnp != rnp_root)
1204 raw_spin_unlock(&rnp_root->lock);
1205 return c;
1206}
1207
1208/*
1209 * Clean up any old requests for the just-ended grace period. Also return
1210 * whether any additional grace periods have been requested. Also invoke
1211 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1212 * waiting for this grace period to complete.
1213 */
1214static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1215{
1216 int c = rnp->completed;
1217 int needmore;
1218 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1219
1220 rcu_nocb_gp_cleanup(rsp, rnp);
1221 rnp->need_future_gp[c & 0x1] = 0;
1222 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1223 trace_rcu_future_gp(rnp, rdp, c,
1224 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1225 return needmore;
1226}
1227
1228/*
1229 * If there is room, assign a ->completed number to any callbacks on
1230 * this CPU that have not already been assigned. Also accelerate any
1231 * callbacks that were previously assigned a ->completed number that has
1232 * since proven to be too conservative, which can happen if callbacks get
1233 * assigned a ->completed number while RCU is idle, but with reference to
1234 * a non-root rcu_node structure. This function is idempotent, so it does
1235 * not hurt to call it repeatedly.
1236 *
1237 * The caller must hold rnp->lock with interrupts disabled.
1238 */
1239static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1240 struct rcu_data *rdp)
1241{
1242 unsigned long c;
1243 int i;
1244
1245 /* If the CPU has no callbacks, nothing to do. */
1246 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1247 return;
1248
1249 /*
1250 * Starting from the sublist containing the callbacks most
1251 * recently assigned a ->completed number and working down, find the
1252 * first sublist that is not assignable to an upcoming grace period.
1253 * Such a sublist has something in it (first two tests) and has
1254 * a ->completed number assigned that will complete sooner than
1255 * the ->completed number for newly arrived callbacks (last test).
1256 *
1257 * The key point is that any later sublist can be assigned the
1258 * same ->completed number as the newly arrived callbacks, which
1259 * means that the callbacks in any of these later sublist can be
1260 * grouped into a single sublist, whether or not they have already
1261 * been assigned a ->completed number.
1262 */
1263 c = rcu_cbs_completed(rsp, rnp);
1264 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1265 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1266 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1267 break;
1268
1269 /*
1270 * If there are no sublist for unassigned callbacks, leave.
1271 * At the same time, advance "i" one sublist, so that "i" will
1272 * index into the sublist where all the remaining callbacks should
1273 * be grouped into.
1274 */
1275 if (++i >= RCU_NEXT_TAIL)
1276 return;
1277
1278 /*
1279 * Assign all subsequent callbacks' ->completed number to the next
1280 * full grace period and group them all in the sublist initially
1281 * indexed by "i".
1282 */
1283 for (; i <= RCU_NEXT_TAIL; i++) {
1284 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1285 rdp->nxtcompleted[i] = c;
1286 }
1287 /* Record any needed additional grace periods. */
1288 rcu_start_future_gp(rnp, rdp);
1289
1290 /* Trace depending on how much we were able to accelerate. */
1291 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1292 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1293 else
1294 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1295}
1296
1297/*
1298 * Move any callbacks whose grace period has completed to the
1299 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1300 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1301 * sublist. This function is idempotent, so it does not hurt to
1302 * invoke it repeatedly. As long as it is not invoked -too- often...
1303 *
1304 * The caller must hold rnp->lock with interrupts disabled.
1305 */
1306static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1307 struct rcu_data *rdp)
1308{
1309 int i, j;
1310
1311 /* If the CPU has no callbacks, nothing to do. */
1312 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1313 return;
1314
1315 /*
1316 * Find all callbacks whose ->completed numbers indicate that they
1317 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1318 */
1319 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1320 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1321 break;
1322 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1323 }
1324 /* Clean up any sublist tail pointers that were misordered above. */
1325 for (j = RCU_WAIT_TAIL; j < i; j++)
1326 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1327
1328 /* Copy down callbacks to fill in empty sublists. */
1329 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1330 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1331 break;
1332 rdp->nxttail[j] = rdp->nxttail[i];
1333 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1334 }
1335
1336 /* Classify any remaining callbacks. */
1337 rcu_accelerate_cbs(rsp, rnp, rdp);
1338}
1339
1340/*
1341 * Update CPU-local rcu_data state to record the beginnings and ends of
1342 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1343 * structure corresponding to the current CPU, and must have irqs disabled.
1344 */
1345static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1346{
1347 /* Handle the ends of any preceding grace periods first. */
1348 if (rdp->completed == rnp->completed) {
1349
1350 /* No grace period end, so just accelerate recent callbacks. */
1351 rcu_accelerate_cbs(rsp, rnp, rdp);
1352
1353 } else {
1354
1355 /* Advance callbacks. */
1356 rcu_advance_cbs(rsp, rnp, rdp);
1357
1358 /* Remember that we saw this grace-period completion. */
1359 rdp->completed = rnp->completed;
1360 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1361 }
1362
1363 if (rdp->gpnum != rnp->gpnum) {
1364 /*
1365 * If the current grace period is waiting for this CPU,
1366 * set up to detect a quiescent state, otherwise don't
1367 * go looking for one.
1368 */
1369 rdp->gpnum = rnp->gpnum;
1370 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1371 rdp->passed_quiesce = 0;
1372 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1373 zero_cpu_stall_ticks(rdp);
1374 }
1375}
1376
1377static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1378{
1379 unsigned long flags;
1380 struct rcu_node *rnp;
1381
1382 local_irq_save(flags);
1383 rnp = rdp->mynode;
1384 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1385 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1386 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1387 local_irq_restore(flags);
1388 return;
1389 }
1390 smp_mb__after_unlock_lock();
1391 __note_gp_changes(rsp, rnp, rdp);
1392 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1393}
1394
1395/*
1396 * Initialize a new grace period. Return 0 if no grace period required.
1397 */
1398static int rcu_gp_init(struct rcu_state *rsp)
1399{
1400 struct rcu_data *rdp;
1401 struct rcu_node *rnp = rcu_get_root(rsp);
1402
1403 rcu_bind_gp_kthread();
1404 raw_spin_lock_irq(&rnp->lock);
1405 smp_mb__after_unlock_lock();
1406 if (rsp->gp_flags == 0) {
1407 /* Spurious wakeup, tell caller to go back to sleep. */
1408 raw_spin_unlock_irq(&rnp->lock);
1409 return 0;
1410 }
1411 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1412
1413 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1414 /*
1415 * Grace period already in progress, don't start another.
1416 * Not supposed to be able to happen.
1417 */
1418 raw_spin_unlock_irq(&rnp->lock);
1419 return 0;
1420 }
1421
1422 /* Advance to a new grace period and initialize state. */
1423 record_gp_stall_check_time(rsp);
1424 /* Record GP times before starting GP, hence smp_store_release(). */
1425 smp_store_release(&rsp->gpnum, rsp->gpnum + 1);
1426 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1427 raw_spin_unlock_irq(&rnp->lock);
1428
1429 /* Exclude any concurrent CPU-hotplug operations. */
1430 mutex_lock(&rsp->onoff_mutex);
1431 smp_mb__after_unlock_lock(); /* ->gpnum increment before GP! */
1432
1433 /*
1434 * Set the quiescent-state-needed bits in all the rcu_node
1435 * structures for all currently online CPUs in breadth-first order,
1436 * starting from the root rcu_node structure, relying on the layout
1437 * of the tree within the rsp->node[] array. Note that other CPUs
1438 * will access only the leaves of the hierarchy, thus seeing that no
1439 * grace period is in progress, at least until the corresponding
1440 * leaf node has been initialized. In addition, we have excluded
1441 * CPU-hotplug operations.
1442 *
1443 * The grace period cannot complete until the initialization
1444 * process finishes, because this kthread handles both.
1445 */
1446 rcu_for_each_node_breadth_first(rsp, rnp) {
1447 raw_spin_lock_irq(&rnp->lock);
1448 smp_mb__after_unlock_lock();
1449 rdp = this_cpu_ptr(rsp->rda);
1450 rcu_preempt_check_blocked_tasks(rnp);
1451 rnp->qsmask = rnp->qsmaskinit;
1452 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1453 WARN_ON_ONCE(rnp->completed != rsp->completed);
1454 ACCESS_ONCE(rnp->completed) = rsp->completed;
1455 if (rnp == rdp->mynode)
1456 __note_gp_changes(rsp, rnp, rdp);
1457 rcu_preempt_boost_start_gp(rnp);
1458 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1459 rnp->level, rnp->grplo,
1460 rnp->grphi, rnp->qsmask);
1461 raw_spin_unlock_irq(&rnp->lock);
1462#ifdef CONFIG_PROVE_RCU_DELAY
1463 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1464 system_state == SYSTEM_RUNNING)
1465 udelay(200);
1466#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1467 cond_resched();
1468 }
1469
1470 mutex_unlock(&rsp->onoff_mutex);
1471 return 1;
1472}
1473
1474/*
1475 * Do one round of quiescent-state forcing.
1476 */
1477static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1478{
1479 int fqs_state = fqs_state_in;
1480 bool isidle = false;
1481 unsigned long maxj;
1482 struct rcu_node *rnp = rcu_get_root(rsp);
1483
1484 rsp->n_force_qs++;
1485 if (fqs_state == RCU_SAVE_DYNTICK) {
1486 /* Collect dyntick-idle snapshots. */
1487 if (is_sysidle_rcu_state(rsp)) {
1488 isidle = 1;
1489 maxj = jiffies - ULONG_MAX / 4;
1490 }
1491 force_qs_rnp(rsp, dyntick_save_progress_counter,
1492 &isidle, &maxj);
1493 rcu_sysidle_report_gp(rsp, isidle, maxj);
1494 fqs_state = RCU_FORCE_QS;
1495 } else {
1496 /* Handle dyntick-idle and offline CPUs. */
1497 isidle = 0;
1498 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1499 }
1500 /* Clear flag to prevent immediate re-entry. */
1501 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1502 raw_spin_lock_irq(&rnp->lock);
1503 smp_mb__after_unlock_lock();
1504 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1505 raw_spin_unlock_irq(&rnp->lock);
1506 }
1507 return fqs_state;
1508}
1509
1510/*
1511 * Clean up after the old grace period.
1512 */
1513static void rcu_gp_cleanup(struct rcu_state *rsp)
1514{
1515 unsigned long gp_duration;
1516 int nocb = 0;
1517 struct rcu_data *rdp;
1518 struct rcu_node *rnp = rcu_get_root(rsp);
1519
1520 raw_spin_lock_irq(&rnp->lock);
1521 smp_mb__after_unlock_lock();
1522 gp_duration = jiffies - rsp->gp_start;
1523 if (gp_duration > rsp->gp_max)
1524 rsp->gp_max = gp_duration;
1525
1526 /*
1527 * We know the grace period is complete, but to everyone else
1528 * it appears to still be ongoing. But it is also the case
1529 * that to everyone else it looks like there is nothing that
1530 * they can do to advance the grace period. It is therefore
1531 * safe for us to drop the lock in order to mark the grace
1532 * period as completed in all of the rcu_node structures.
1533 */
1534 raw_spin_unlock_irq(&rnp->lock);
1535
1536 /*
1537 * Propagate new ->completed value to rcu_node structures so
1538 * that other CPUs don't have to wait until the start of the next
1539 * grace period to process their callbacks. This also avoids
1540 * some nasty RCU grace-period initialization races by forcing
1541 * the end of the current grace period to be completely recorded in
1542 * all of the rcu_node structures before the beginning of the next
1543 * grace period is recorded in any of the rcu_node structures.
1544 */
1545 rcu_for_each_node_breadth_first(rsp, rnp) {
1546 raw_spin_lock_irq(&rnp->lock);
1547 smp_mb__after_unlock_lock();
1548 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1549 rdp = this_cpu_ptr(rsp->rda);
1550 if (rnp == rdp->mynode)
1551 __note_gp_changes(rsp, rnp, rdp);
1552 /* smp_mb() provided by prior unlock-lock pair. */
1553 nocb += rcu_future_gp_cleanup(rsp, rnp);
1554 raw_spin_unlock_irq(&rnp->lock);
1555 cond_resched();
1556 }
1557 rnp = rcu_get_root(rsp);
1558 raw_spin_lock_irq(&rnp->lock);
1559 smp_mb__after_unlock_lock(); /* Order GP before ->completed update. */
1560 rcu_nocb_gp_set(rnp, nocb);
1561
1562 /* Declare grace period done. */
1563 ACCESS_ONCE(rsp->completed) = rsp->gpnum;
1564 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1565 rsp->fqs_state = RCU_GP_IDLE;
1566 rdp = this_cpu_ptr(rsp->rda);
1567 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1568 if (cpu_needs_another_gp(rsp, rdp)) {
1569 rsp->gp_flags = RCU_GP_FLAG_INIT;
1570 trace_rcu_grace_period(rsp->name,
1571 ACCESS_ONCE(rsp->gpnum),
1572 TPS("newreq"));
1573 }
1574 raw_spin_unlock_irq(&rnp->lock);
1575}
1576
1577/*
1578 * Body of kthread that handles grace periods.
1579 */
1580static int __noreturn rcu_gp_kthread(void *arg)
1581{
1582 int fqs_state;
1583 int gf;
1584 unsigned long j;
1585 int ret;
1586 struct rcu_state *rsp = arg;
1587 struct rcu_node *rnp = rcu_get_root(rsp);
1588
1589 for (;;) {
1590
1591 /* Handle grace-period start. */
1592 for (;;) {
1593 trace_rcu_grace_period(rsp->name,
1594 ACCESS_ONCE(rsp->gpnum),
1595 TPS("reqwait"));
1596 wait_event_interruptible(rsp->gp_wq,
1597 ACCESS_ONCE(rsp->gp_flags) &
1598 RCU_GP_FLAG_INIT);
1599 /* Locking provides needed memory barrier. */
1600 if (rcu_gp_init(rsp))
1601 break;
1602 cond_resched();
1603 flush_signals(current);
1604 trace_rcu_grace_period(rsp->name,
1605 ACCESS_ONCE(rsp->gpnum),
1606 TPS("reqwaitsig"));
1607 }
1608
1609 /* Handle quiescent-state forcing. */
1610 fqs_state = RCU_SAVE_DYNTICK;
1611 j = jiffies_till_first_fqs;
1612 if (j > HZ) {
1613 j = HZ;
1614 jiffies_till_first_fqs = HZ;
1615 }
1616 ret = 0;
1617 for (;;) {
1618 if (!ret)
1619 rsp->jiffies_force_qs = jiffies + j;
1620 trace_rcu_grace_period(rsp->name,
1621 ACCESS_ONCE(rsp->gpnum),
1622 TPS("fqswait"));
1623 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1624 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1625 RCU_GP_FLAG_FQS) ||
1626 (!ACCESS_ONCE(rnp->qsmask) &&
1627 !rcu_preempt_blocked_readers_cgp(rnp)),
1628 j);
1629 /* Locking provides needed memory barriers. */
1630 /* If grace period done, leave loop. */
1631 if (!ACCESS_ONCE(rnp->qsmask) &&
1632 !rcu_preempt_blocked_readers_cgp(rnp))
1633 break;
1634 /* If time for quiescent-state forcing, do it. */
1635 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1636 (gf & RCU_GP_FLAG_FQS)) {
1637 trace_rcu_grace_period(rsp->name,
1638 ACCESS_ONCE(rsp->gpnum),
1639 TPS("fqsstart"));
1640 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1641 trace_rcu_grace_period(rsp->name,
1642 ACCESS_ONCE(rsp->gpnum),
1643 TPS("fqsend"));
1644 cond_resched();
1645 } else {
1646 /* Deal with stray signal. */
1647 cond_resched();
1648 flush_signals(current);
1649 trace_rcu_grace_period(rsp->name,
1650 ACCESS_ONCE(rsp->gpnum),
1651 TPS("fqswaitsig"));
1652 }
1653 j = jiffies_till_next_fqs;
1654 if (j > HZ) {
1655 j = HZ;
1656 jiffies_till_next_fqs = HZ;
1657 } else if (j < 1) {
1658 j = 1;
1659 jiffies_till_next_fqs = 1;
1660 }
1661 }
1662
1663 /* Handle grace-period end. */
1664 rcu_gp_cleanup(rsp);
1665 }
1666}
1667
1668static void rsp_wakeup(struct irq_work *work)
1669{
1670 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1671
1672 /* Wake up rcu_gp_kthread() to start the grace period. */
1673 wake_up(&rsp->gp_wq);
1674}
1675
1676/*
1677 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1678 * in preparation for detecting the next grace period. The caller must hold
1679 * the root node's ->lock and hard irqs must be disabled.
1680 *
1681 * Note that it is legal for a dying CPU (which is marked as offline) to
1682 * invoke this function. This can happen when the dying CPU reports its
1683 * quiescent state.
1684 */
1685static void
1686rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1687 struct rcu_data *rdp)
1688{
1689 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1690 /*
1691 * Either we have not yet spawned the grace-period
1692 * task, this CPU does not need another grace period,
1693 * or a grace period is already in progress.
1694 * Either way, don't start a new grace period.
1695 */
1696 return;
1697 }
1698 rsp->gp_flags = RCU_GP_FLAG_INIT;
1699 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1700 TPS("newreq"));
1701
1702 /*
1703 * We can't do wakeups while holding the rnp->lock, as that
1704 * could cause possible deadlocks with the rq->lock. Defer
1705 * the wakeup to interrupt context. And don't bother waking
1706 * up the running kthread.
1707 */
1708 if (current != rsp->gp_kthread)
1709 irq_work_queue(&rsp->wakeup_work);
1710}
1711
1712/*
1713 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1714 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1715 * is invoked indirectly from rcu_advance_cbs(), which would result in
1716 * endless recursion -- or would do so if it wasn't for the self-deadlock
1717 * that is encountered beforehand.
1718 */
1719static void
1720rcu_start_gp(struct rcu_state *rsp)
1721{
1722 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1723 struct rcu_node *rnp = rcu_get_root(rsp);
1724
1725 /*
1726 * If there is no grace period in progress right now, any
1727 * callbacks we have up to this point will be satisfied by the
1728 * next grace period. Also, advancing the callbacks reduces the
1729 * probability of false positives from cpu_needs_another_gp()
1730 * resulting in pointless grace periods. So, advance callbacks
1731 * then start the grace period!
1732 */
1733 rcu_advance_cbs(rsp, rnp, rdp);
1734 rcu_start_gp_advanced(rsp, rnp, rdp);
1735}
1736
1737/*
1738 * Report a full set of quiescent states to the specified rcu_state
1739 * data structure. This involves cleaning up after the prior grace
1740 * period and letting rcu_start_gp() start up the next grace period
1741 * if one is needed. Note that the caller must hold rnp->lock, which
1742 * is released before return.
1743 */
1744static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1745 __releases(rcu_get_root(rsp)->lock)
1746{
1747 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1748 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1749 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1750}
1751
1752/*
1753 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1754 * Allows quiescent states for a group of CPUs to be reported at one go
1755 * to the specified rcu_node structure, though all the CPUs in the group
1756 * must be represented by the same rcu_node structure (which need not be
1757 * a leaf rcu_node structure, though it often will be). That structure's
1758 * lock must be held upon entry, and it is released before return.
1759 */
1760static void
1761rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1762 struct rcu_node *rnp, unsigned long flags)
1763 __releases(rnp->lock)
1764{
1765 struct rcu_node *rnp_c;
1766
1767 /* Walk up the rcu_node hierarchy. */
1768 for (;;) {
1769 if (!(rnp->qsmask & mask)) {
1770
1771 /* Our bit has already been cleared, so done. */
1772 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1773 return;
1774 }
1775 rnp->qsmask &= ~mask;
1776 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1777 mask, rnp->qsmask, rnp->level,
1778 rnp->grplo, rnp->grphi,
1779 !!rnp->gp_tasks);
1780 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1781
1782 /* Other bits still set at this level, so done. */
1783 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1784 return;
1785 }
1786 mask = rnp->grpmask;
1787 if (rnp->parent == NULL) {
1788
1789 /* No more levels. Exit loop holding root lock. */
1790
1791 break;
1792 }
1793 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1794 rnp_c = rnp;
1795 rnp = rnp->parent;
1796 raw_spin_lock_irqsave(&rnp->lock, flags);
1797 smp_mb__after_unlock_lock();
1798 WARN_ON_ONCE(rnp_c->qsmask);
1799 }
1800
1801 /*
1802 * Get here if we are the last CPU to pass through a quiescent
1803 * state for this grace period. Invoke rcu_report_qs_rsp()
1804 * to clean up and start the next grace period if one is needed.
1805 */
1806 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1807}
1808
1809/*
1810 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1811 * structure. This must be either called from the specified CPU, or
1812 * called when the specified CPU is known to be offline (and when it is
1813 * also known that no other CPU is concurrently trying to help the offline
1814 * CPU). The lastcomp argument is used to make sure we are still in the
1815 * grace period of interest. We don't want to end the current grace period
1816 * based on quiescent states detected in an earlier grace period!
1817 */
1818static void
1819rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1820{
1821 unsigned long flags;
1822 unsigned long mask;
1823 struct rcu_node *rnp;
1824
1825 rnp = rdp->mynode;
1826 raw_spin_lock_irqsave(&rnp->lock, flags);
1827 smp_mb__after_unlock_lock();
1828 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1829 rnp->completed == rnp->gpnum) {
1830
1831 /*
1832 * The grace period in which this quiescent state was
1833 * recorded has ended, so don't report it upwards.
1834 * We will instead need a new quiescent state that lies
1835 * within the current grace period.
1836 */
1837 rdp->passed_quiesce = 0; /* need qs for new gp. */
1838 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1839 return;
1840 }
1841 mask = rdp->grpmask;
1842 if ((rnp->qsmask & mask) == 0) {
1843 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1844 } else {
1845 rdp->qs_pending = 0;
1846
1847 /*
1848 * This GP can't end until cpu checks in, so all of our
1849 * callbacks can be processed during the next GP.
1850 */
1851 rcu_accelerate_cbs(rsp, rnp, rdp);
1852
1853 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1854 }
1855}
1856
1857/*
1858 * Check to see if there is a new grace period of which this CPU
1859 * is not yet aware, and if so, set up local rcu_data state for it.
1860 * Otherwise, see if this CPU has just passed through its first
1861 * quiescent state for this grace period, and record that fact if so.
1862 */
1863static void
1864rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1865{
1866 /* Check for grace-period ends and beginnings. */
1867 note_gp_changes(rsp, rdp);
1868
1869 /*
1870 * Does this CPU still need to do its part for current grace period?
1871 * If no, return and let the other CPUs do their part as well.
1872 */
1873 if (!rdp->qs_pending)
1874 return;
1875
1876 /*
1877 * Was there a quiescent state since the beginning of the grace
1878 * period? If no, then exit and wait for the next call.
1879 */
1880 if (!rdp->passed_quiesce)
1881 return;
1882
1883 /*
1884 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1885 * judge of that).
1886 */
1887 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1888}
1889
1890#ifdef CONFIG_HOTPLUG_CPU
1891
1892/*
1893 * Send the specified CPU's RCU callbacks to the orphanage. The
1894 * specified CPU must be offline, and the caller must hold the
1895 * ->orphan_lock.
1896 */
1897static void
1898rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1899 struct rcu_node *rnp, struct rcu_data *rdp)
1900{
1901 /* No-CBs CPUs do not have orphanable callbacks. */
1902 if (rcu_is_nocb_cpu(rdp->cpu))
1903 return;
1904
1905 /*
1906 * Orphan the callbacks. First adjust the counts. This is safe
1907 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1908 * cannot be running now. Thus no memory barrier is required.
1909 */
1910 if (rdp->nxtlist != NULL) {
1911 rsp->qlen_lazy += rdp->qlen_lazy;
1912 rsp->qlen += rdp->qlen;
1913 rdp->n_cbs_orphaned += rdp->qlen;
1914 rdp->qlen_lazy = 0;
1915 ACCESS_ONCE(rdp->qlen) = 0;
1916 }
1917
1918 /*
1919 * Next, move those callbacks still needing a grace period to
1920 * the orphanage, where some other CPU will pick them up.
1921 * Some of the callbacks might have gone partway through a grace
1922 * period, but that is too bad. They get to start over because we
1923 * cannot assume that grace periods are synchronized across CPUs.
1924 * We don't bother updating the ->nxttail[] array yet, instead
1925 * we just reset the whole thing later on.
1926 */
1927 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1928 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1929 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1930 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1931 }
1932
1933 /*
1934 * Then move the ready-to-invoke callbacks to the orphanage,
1935 * where some other CPU will pick them up. These will not be
1936 * required to pass though another grace period: They are done.
1937 */
1938 if (rdp->nxtlist != NULL) {
1939 *rsp->orphan_donetail = rdp->nxtlist;
1940 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1941 }
1942
1943 /* Finally, initialize the rcu_data structure's list to empty. */
1944 init_callback_list(rdp);
1945}
1946
1947/*
1948 * Adopt the RCU callbacks from the specified rcu_state structure's
1949 * orphanage. The caller must hold the ->orphan_lock.
1950 */
1951static void rcu_adopt_orphan_cbs(struct rcu_state *rsp, unsigned long flags)
1952{
1953 int i;
1954 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1955
1956 /* No-CBs CPUs are handled specially. */
1957 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp, flags))
1958 return;
1959
1960 /* Do the accounting first. */
1961 rdp->qlen_lazy += rsp->qlen_lazy;
1962 rdp->qlen += rsp->qlen;
1963 rdp->n_cbs_adopted += rsp->qlen;
1964 if (rsp->qlen_lazy != rsp->qlen)
1965 rcu_idle_count_callbacks_posted();
1966 rsp->qlen_lazy = 0;
1967 rsp->qlen = 0;
1968
1969 /*
1970 * We do not need a memory barrier here because the only way we
1971 * can get here if there is an rcu_barrier() in flight is if
1972 * we are the task doing the rcu_barrier().
1973 */
1974
1975 /* First adopt the ready-to-invoke callbacks. */
1976 if (rsp->orphan_donelist != NULL) {
1977 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1978 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1979 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1980 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1981 rdp->nxttail[i] = rsp->orphan_donetail;
1982 rsp->orphan_donelist = NULL;
1983 rsp->orphan_donetail = &rsp->orphan_donelist;
1984 }
1985
1986 /* And then adopt the callbacks that still need a grace period. */
1987 if (rsp->orphan_nxtlist != NULL) {
1988 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1989 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1990 rsp->orphan_nxtlist = NULL;
1991 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1992 }
1993}
1994
1995/*
1996 * Trace the fact that this CPU is going offline.
1997 */
1998static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1999{
2000 RCU_TRACE(unsigned long mask);
2001 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
2002 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
2003
2004 RCU_TRACE(mask = rdp->grpmask);
2005 trace_rcu_grace_period(rsp->name,
2006 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
2007 TPS("cpuofl"));
2008}
2009
2010/*
2011 * The CPU has been completely removed, and some other CPU is reporting
2012 * this fact from process context. Do the remainder of the cleanup,
2013 * including orphaning the outgoing CPU's RCU callbacks, and also
2014 * adopting them. There can only be one CPU hotplug operation at a time,
2015 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
2016 */
2017static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2018{
2019 unsigned long flags;
2020 unsigned long mask;
2021 int need_report = 0;
2022 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
2023 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2024
2025 /* Adjust any no-longer-needed kthreads. */
2026 rcu_boost_kthread_setaffinity(rnp, -1);
2027
2028 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
2029
2030 /* Exclude any attempts to start a new grace period. */
2031 mutex_lock(&rsp->onoff_mutex);
2032 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
2033
2034 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
2035 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
2036 rcu_adopt_orphan_cbs(rsp, flags);
2037
2038 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
2039 mask = rdp->grpmask; /* rnp->grplo is constant. */
2040 do {
2041 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
2042 smp_mb__after_unlock_lock();
2043 rnp->qsmaskinit &= ~mask;
2044 if (rnp->qsmaskinit != 0) {
2045 if (rnp != rdp->mynode)
2046 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2047 break;
2048 }
2049 if (rnp == rdp->mynode)
2050 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2051 else
2052 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2053 mask = rnp->grpmask;
2054 rnp = rnp->parent;
2055 } while (rnp != NULL);
2056
2057 /*
2058 * We still hold the leaf rcu_node structure lock here, and
2059 * irqs are still disabled. The reason for this subterfuge is
2060 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2061 * held leads to deadlock.
2062 */
2063 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2064 rnp = rdp->mynode;
2065 if (need_report & RCU_OFL_TASKS_NORM_GP)
2066 rcu_report_unblock_qs_rnp(rnp, flags);
2067 else
2068 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2069 if (need_report & RCU_OFL_TASKS_EXP_GP)
2070 rcu_report_exp_rnp(rsp, rnp, true);
2071 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2072 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2073 cpu, rdp->qlen, rdp->nxtlist);
2074 init_callback_list(rdp);
2075 /* Disallow further callbacks on this CPU. */
2076 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2077 mutex_unlock(&rsp->onoff_mutex);
2078}
2079
2080#else /* #ifdef CONFIG_HOTPLUG_CPU */
2081
2082static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2083{
2084}
2085
2086static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2087{
2088}
2089
2090#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2091
2092/*
2093 * Invoke any RCU callbacks that have made it to the end of their grace
2094 * period. Thottle as specified by rdp->blimit.
2095 */
2096static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2097{
2098 unsigned long flags;
2099 struct rcu_head *next, *list, **tail;
2100 long bl, count, count_lazy;
2101 int i;
2102
2103 /* If no callbacks are ready, just return. */
2104 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2105 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2106 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2107 need_resched(), is_idle_task(current),
2108 rcu_is_callbacks_kthread());
2109 return;
2110 }
2111
2112 /*
2113 * Extract the list of ready callbacks, disabling to prevent
2114 * races with call_rcu() from interrupt handlers.
2115 */
2116 local_irq_save(flags);
2117 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2118 bl = rdp->blimit;
2119 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2120 list = rdp->nxtlist;
2121 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2122 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2123 tail = rdp->nxttail[RCU_DONE_TAIL];
2124 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2125 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2126 rdp->nxttail[i] = &rdp->nxtlist;
2127 local_irq_restore(flags);
2128
2129 /* Invoke callbacks. */
2130 count = count_lazy = 0;
2131 while (list) {
2132 next = list->next;
2133 prefetch(next);
2134 debug_rcu_head_unqueue(list);
2135 if (__rcu_reclaim(rsp->name, list))
2136 count_lazy++;
2137 list = next;
2138 /* Stop only if limit reached and CPU has something to do. */
2139 if (++count >= bl &&
2140 (need_resched() ||
2141 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2142 break;
2143 }
2144
2145 local_irq_save(flags);
2146 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2147 is_idle_task(current),
2148 rcu_is_callbacks_kthread());
2149
2150 /* Update count, and requeue any remaining callbacks. */
2151 if (list != NULL) {
2152 *tail = rdp->nxtlist;
2153 rdp->nxtlist = list;
2154 for (i = 0; i < RCU_NEXT_SIZE; i++)
2155 if (&rdp->nxtlist == rdp->nxttail[i])
2156 rdp->nxttail[i] = tail;
2157 else
2158 break;
2159 }
2160 smp_mb(); /* List handling before counting for rcu_barrier(). */
2161 rdp->qlen_lazy -= count_lazy;
2162 ACCESS_ONCE(rdp->qlen) -= count;
2163 rdp->n_cbs_invoked += count;
2164
2165 /* Reinstate batch limit if we have worked down the excess. */
2166 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2167 rdp->blimit = blimit;
2168
2169 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2170 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2171 rdp->qlen_last_fqs_check = 0;
2172 rdp->n_force_qs_snap = rsp->n_force_qs;
2173 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2174 rdp->qlen_last_fqs_check = rdp->qlen;
2175 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2176
2177 local_irq_restore(flags);
2178
2179 /* Re-invoke RCU core processing if there are callbacks remaining. */
2180 if (cpu_has_callbacks_ready_to_invoke(rdp))
2181 invoke_rcu_core();
2182}
2183
2184/*
2185 * Check to see if this CPU is in a non-context-switch quiescent state
2186 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2187 * Also schedule RCU core processing.
2188 *
2189 * This function must be called from hardirq context. It is normally
2190 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2191 * false, there is no point in invoking rcu_check_callbacks().
2192 */
2193void rcu_check_callbacks(int cpu, int user)
2194{
2195 trace_rcu_utilization(TPS("Start scheduler-tick"));
2196 increment_cpu_stall_ticks();
2197 if (user || rcu_is_cpu_rrupt_from_idle()) {
2198
2199 /*
2200 * Get here if this CPU took its interrupt from user
2201 * mode or from the idle loop, and if this is not a
2202 * nested interrupt. In this case, the CPU is in
2203 * a quiescent state, so note it.
2204 *
2205 * No memory barrier is required here because both
2206 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2207 * variables that other CPUs neither access nor modify,
2208 * at least not while the corresponding CPU is online.
2209 */
2210
2211 rcu_sched_qs(cpu);
2212 rcu_bh_qs(cpu);
2213
2214 } else if (!in_softirq()) {
2215
2216 /*
2217 * Get here if this CPU did not take its interrupt from
2218 * softirq, in other words, if it is not interrupting
2219 * a rcu_bh read-side critical section. This is an _bh
2220 * critical section, so note it.
2221 */
2222
2223 rcu_bh_qs(cpu);
2224 }
2225 rcu_preempt_check_callbacks(cpu);
2226 if (rcu_pending(cpu))
2227 invoke_rcu_core();
2228 trace_rcu_utilization(TPS("End scheduler-tick"));
2229}
2230
2231/*
2232 * Scan the leaf rcu_node structures, processing dyntick state for any that
2233 * have not yet encountered a quiescent state, using the function specified.
2234 * Also initiate boosting for any threads blocked on the root rcu_node.
2235 *
2236 * The caller must have suppressed start of new grace periods.
2237 */
2238static void force_qs_rnp(struct rcu_state *rsp,
2239 int (*f)(struct rcu_data *rsp, bool *isidle,
2240 unsigned long *maxj),
2241 bool *isidle, unsigned long *maxj)
2242{
2243 unsigned long bit;
2244 int cpu;
2245 unsigned long flags;
2246 unsigned long mask;
2247 struct rcu_node *rnp;
2248
2249 rcu_for_each_leaf_node(rsp, rnp) {
2250 cond_resched();
2251 mask = 0;
2252 raw_spin_lock_irqsave(&rnp->lock, flags);
2253 smp_mb__after_unlock_lock();
2254 if (!rcu_gp_in_progress(rsp)) {
2255 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2256 return;
2257 }
2258 if (rnp->qsmask == 0) {
2259 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2260 continue;
2261 }
2262 cpu = rnp->grplo;
2263 bit = 1;
2264 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2265 if ((rnp->qsmask & bit) != 0) {
2266 if ((rnp->qsmaskinit & bit) != 0)
2267 *isidle = 0;
2268 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2269 mask |= bit;
2270 }
2271 }
2272 if (mask != 0) {
2273
2274 /* rcu_report_qs_rnp() releases rnp->lock. */
2275 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2276 continue;
2277 }
2278 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2279 }
2280 rnp = rcu_get_root(rsp);
2281 if (rnp->qsmask == 0) {
2282 raw_spin_lock_irqsave(&rnp->lock, flags);
2283 smp_mb__after_unlock_lock();
2284 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2285 }
2286}
2287
2288/*
2289 * Force quiescent states on reluctant CPUs, and also detect which
2290 * CPUs are in dyntick-idle mode.
2291 */
2292static void force_quiescent_state(struct rcu_state *rsp)
2293{
2294 unsigned long flags;
2295 bool ret;
2296 struct rcu_node *rnp;
2297 struct rcu_node *rnp_old = NULL;
2298
2299 /* Funnel through hierarchy to reduce memory contention. */
2300 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2301 for (; rnp != NULL; rnp = rnp->parent) {
2302 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2303 !raw_spin_trylock(&rnp->fqslock);
2304 if (rnp_old != NULL)
2305 raw_spin_unlock(&rnp_old->fqslock);
2306 if (ret) {
2307 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2308 return;
2309 }
2310 rnp_old = rnp;
2311 }
2312 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2313
2314 /* Reached the root of the rcu_node tree, acquire lock. */
2315 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2316 smp_mb__after_unlock_lock();
2317 raw_spin_unlock(&rnp_old->fqslock);
2318 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2319 ACCESS_ONCE(rsp->n_force_qs_lh)++;
2320 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2321 return; /* Someone beat us to it. */
2322 }
2323 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2324 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2325 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2326}
2327
2328/*
2329 * This does the RCU core processing work for the specified rcu_state
2330 * and rcu_data structures. This may be called only from the CPU to
2331 * whom the rdp belongs.
2332 */
2333static void
2334__rcu_process_callbacks(struct rcu_state *rsp)
2335{
2336 unsigned long flags;
2337 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2338
2339 WARN_ON_ONCE(rdp->beenonline == 0);
2340
2341 /* Update RCU state based on any recent quiescent states. */
2342 rcu_check_quiescent_state(rsp, rdp);
2343
2344 /* Does this CPU require a not-yet-started grace period? */
2345 local_irq_save(flags);
2346 if (cpu_needs_another_gp(rsp, rdp)) {
2347 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2348 rcu_start_gp(rsp);
2349 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2350 } else {
2351 local_irq_restore(flags);
2352 }
2353
2354 /* If there are callbacks ready, invoke them. */
2355 if (cpu_has_callbacks_ready_to_invoke(rdp))
2356 invoke_rcu_callbacks(rsp, rdp);
2357
2358 /* Do any needed deferred wakeups of rcuo kthreads. */
2359 do_nocb_deferred_wakeup(rdp);
2360}
2361
2362/*
2363 * Do RCU core processing for the current CPU.
2364 */
2365static void rcu_process_callbacks(struct softirq_action *unused)
2366{
2367 struct rcu_state *rsp;
2368
2369 if (cpu_is_offline(smp_processor_id()))
2370 return;
2371 trace_rcu_utilization(TPS("Start RCU core"));
2372 for_each_rcu_flavor(rsp)
2373 __rcu_process_callbacks(rsp);
2374 trace_rcu_utilization(TPS("End RCU core"));
2375}
2376
2377/*
2378 * Schedule RCU callback invocation. If the specified type of RCU
2379 * does not support RCU priority boosting, just do a direct call,
2380 * otherwise wake up the per-CPU kernel kthread. Note that because we
2381 * are running on the current CPU with interrupts disabled, the
2382 * rcu_cpu_kthread_task cannot disappear out from under us.
2383 */
2384static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2385{
2386 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2387 return;
2388 if (likely(!rsp->boost)) {
2389 rcu_do_batch(rsp, rdp);
2390 return;
2391 }
2392 invoke_rcu_callbacks_kthread();
2393}
2394
2395static void invoke_rcu_core(void)
2396{
2397 if (cpu_online(smp_processor_id()))
2398 raise_softirq(RCU_SOFTIRQ);
2399}
2400
2401/*
2402 * Handle any core-RCU processing required by a call_rcu() invocation.
2403 */
2404static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2405 struct rcu_head *head, unsigned long flags)
2406{
2407 /*
2408 * If called from an extended quiescent state, invoke the RCU
2409 * core in order to force a re-evaluation of RCU's idleness.
2410 */
2411 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2412 invoke_rcu_core();
2413
2414 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2415 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2416 return;
2417
2418 /*
2419 * Force the grace period if too many callbacks or too long waiting.
2420 * Enforce hysteresis, and don't invoke force_quiescent_state()
2421 * if some other CPU has recently done so. Also, don't bother
2422 * invoking force_quiescent_state() if the newly enqueued callback
2423 * is the only one waiting for a grace period to complete.
2424 */
2425 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2426
2427 /* Are we ignoring a completed grace period? */
2428 note_gp_changes(rsp, rdp);
2429
2430 /* Start a new grace period if one not already started. */
2431 if (!rcu_gp_in_progress(rsp)) {
2432 struct rcu_node *rnp_root = rcu_get_root(rsp);
2433
2434 raw_spin_lock(&rnp_root->lock);
2435 smp_mb__after_unlock_lock();
2436 rcu_start_gp(rsp);
2437 raw_spin_unlock(&rnp_root->lock);
2438 } else {
2439 /* Give the grace period a kick. */
2440 rdp->blimit = LONG_MAX;
2441 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2442 *rdp->nxttail[RCU_DONE_TAIL] != head)
2443 force_quiescent_state(rsp);
2444 rdp->n_force_qs_snap = rsp->n_force_qs;
2445 rdp->qlen_last_fqs_check = rdp->qlen;
2446 }
2447 }
2448}
2449
2450/*
2451 * RCU callback function to leak a callback.
2452 */
2453static void rcu_leak_callback(struct rcu_head *rhp)
2454{
2455}
2456
2457/*
2458 * Helper function for call_rcu() and friends. The cpu argument will
2459 * normally be -1, indicating "currently running CPU". It may specify
2460 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2461 * is expected to specify a CPU.
2462 */
2463static void
2464__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2465 struct rcu_state *rsp, int cpu, bool lazy)
2466{
2467 unsigned long flags;
2468 struct rcu_data *rdp;
2469
2470 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2471 if (debug_rcu_head_queue(head)) {
2472 /* Probable double call_rcu(), so leak the callback. */
2473 ACCESS_ONCE(head->func) = rcu_leak_callback;
2474 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2475 return;
2476 }
2477 head->func = func;
2478 head->next = NULL;
2479
2480 /*
2481 * Opportunistically note grace-period endings and beginnings.
2482 * Note that we might see a beginning right after we see an
2483 * end, but never vice versa, since this CPU has to pass through
2484 * a quiescent state betweentimes.
2485 */
2486 local_irq_save(flags);
2487 rdp = this_cpu_ptr(rsp->rda);
2488
2489 /* Add the callback to our list. */
2490 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2491 int offline;
2492
2493 if (cpu != -1)
2494 rdp = per_cpu_ptr(rsp->rda, cpu);
2495 offline = !__call_rcu_nocb(rdp, head, lazy, flags);
2496 WARN_ON_ONCE(offline);
2497 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2498 local_irq_restore(flags);
2499 return;
2500 }
2501 ACCESS_ONCE(rdp->qlen)++;
2502 if (lazy)
2503 rdp->qlen_lazy++;
2504 else
2505 rcu_idle_count_callbacks_posted();
2506 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2507 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2508 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2509
2510 if (__is_kfree_rcu_offset((unsigned long)func))
2511 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2512 rdp->qlen_lazy, rdp->qlen);
2513 else
2514 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2515
2516 /* Go handle any RCU core processing required. */
2517 __call_rcu_core(rsp, rdp, head, flags);
2518 local_irq_restore(flags);
2519}
2520
2521/*
2522 * Queue an RCU-sched callback for invocation after a grace period.
2523 */
2524void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2525{
2526 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2527}
2528EXPORT_SYMBOL_GPL(call_rcu_sched);
2529
2530/*
2531 * Queue an RCU callback for invocation after a quicker grace period.
2532 */
2533void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2534{
2535 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2536}
2537EXPORT_SYMBOL_GPL(call_rcu_bh);
2538
2539/*
2540 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2541 * any blocking grace-period wait automatically implies a grace period
2542 * if there is only one CPU online at any point time during execution
2543 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2544 * occasionally incorrectly indicate that there are multiple CPUs online
2545 * when there was in fact only one the whole time, as this just adds
2546 * some overhead: RCU still operates correctly.
2547 */
2548static inline int rcu_blocking_is_gp(void)
2549{
2550 int ret;
2551
2552 might_sleep(); /* Check for RCU read-side critical section. */
2553 preempt_disable();
2554 ret = num_online_cpus() <= 1;
2555 preempt_enable();
2556 return ret;
2557}
2558
2559/**
2560 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2561 *
2562 * Control will return to the caller some time after a full rcu-sched
2563 * grace period has elapsed, in other words after all currently executing
2564 * rcu-sched read-side critical sections have completed. These read-side
2565 * critical sections are delimited by rcu_read_lock_sched() and
2566 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2567 * local_irq_disable(), and so on may be used in place of
2568 * rcu_read_lock_sched().
2569 *
2570 * This means that all preempt_disable code sequences, including NMI and
2571 * non-threaded hardware-interrupt handlers, in progress on entry will
2572 * have completed before this primitive returns. However, this does not
2573 * guarantee that softirq handlers will have completed, since in some
2574 * kernels, these handlers can run in process context, and can block.
2575 *
2576 * Note that this guarantee implies further memory-ordering guarantees.
2577 * On systems with more than one CPU, when synchronize_sched() returns,
2578 * each CPU is guaranteed to have executed a full memory barrier since the
2579 * end of its last RCU-sched read-side critical section whose beginning
2580 * preceded the call to synchronize_sched(). In addition, each CPU having
2581 * an RCU read-side critical section that extends beyond the return from
2582 * synchronize_sched() is guaranteed to have executed a full memory barrier
2583 * after the beginning of synchronize_sched() and before the beginning of
2584 * that RCU read-side critical section. Note that these guarantees include
2585 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2586 * that are executing in the kernel.
2587 *
2588 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2589 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2590 * to have executed a full memory barrier during the execution of
2591 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2592 * again only if the system has more than one CPU).
2593 *
2594 * This primitive provides the guarantees made by the (now removed)
2595 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2596 * guarantees that rcu_read_lock() sections will have completed.
2597 * In "classic RCU", these two guarantees happen to be one and
2598 * the same, but can differ in realtime RCU implementations.
2599 */
2600void synchronize_sched(void)
2601{
2602 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2603 !lock_is_held(&rcu_lock_map) &&
2604 !lock_is_held(&rcu_sched_lock_map),
2605 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2606 if (rcu_blocking_is_gp())
2607 return;
2608 if (rcu_expedited)
2609 synchronize_sched_expedited();
2610 else
2611 wait_rcu_gp(call_rcu_sched);
2612}
2613EXPORT_SYMBOL_GPL(synchronize_sched);
2614
2615/**
2616 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2617 *
2618 * Control will return to the caller some time after a full rcu_bh grace
2619 * period has elapsed, in other words after all currently executing rcu_bh
2620 * read-side critical sections have completed. RCU read-side critical
2621 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2622 * and may be nested.
2623 *
2624 * See the description of synchronize_sched() for more detailed information
2625 * on memory ordering guarantees.
2626 */
2627void synchronize_rcu_bh(void)
2628{
2629 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2630 !lock_is_held(&rcu_lock_map) &&
2631 !lock_is_held(&rcu_sched_lock_map),
2632 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2633 if (rcu_blocking_is_gp())
2634 return;
2635 if (rcu_expedited)
2636 synchronize_rcu_bh_expedited();
2637 else
2638 wait_rcu_gp(call_rcu_bh);
2639}
2640EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2641
2642/**
2643 * get_state_synchronize_rcu - Snapshot current RCU state
2644 *
2645 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
2646 * to determine whether or not a full grace period has elapsed in the
2647 * meantime.
2648 */
2649unsigned long get_state_synchronize_rcu(void)
2650{
2651 /*
2652 * Any prior manipulation of RCU-protected data must happen
2653 * before the load from ->gpnum.
2654 */
2655 smp_mb(); /* ^^^ */
2656
2657 /*
2658 * Make sure this load happens before the purportedly
2659 * time-consuming work between get_state_synchronize_rcu()
2660 * and cond_synchronize_rcu().
2661 */
2662 return smp_load_acquire(&rcu_state->gpnum);
2663}
2664EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
2665
2666/**
2667 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
2668 *
2669 * @oldstate: return value from earlier call to get_state_synchronize_rcu()
2670 *
2671 * If a full RCU grace period has elapsed since the earlier call to
2672 * get_state_synchronize_rcu(), just return. Otherwise, invoke
2673 * synchronize_rcu() to wait for a full grace period.
2674 *
2675 * Yes, this function does not take counter wrap into account. But
2676 * counter wrap is harmless. If the counter wraps, we have waited for
2677 * more than 2 billion grace periods (and way more on a 64-bit system!),
2678 * so waiting for one additional grace period should be just fine.
2679 */
2680void cond_synchronize_rcu(unsigned long oldstate)
2681{
2682 unsigned long newstate;
2683
2684 /*
2685 * Ensure that this load happens before any RCU-destructive
2686 * actions the caller might carry out after we return.
2687 */
2688 newstate = smp_load_acquire(&rcu_state->completed);
2689 if (ULONG_CMP_GE(oldstate, newstate))
2690 synchronize_rcu();
2691}
2692EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
2693
2694static int synchronize_sched_expedited_cpu_stop(void *data)
2695{
2696 /*
2697 * There must be a full memory barrier on each affected CPU
2698 * between the time that try_stop_cpus() is called and the
2699 * time that it returns.
2700 *
2701 * In the current initial implementation of cpu_stop, the
2702 * above condition is already met when the control reaches
2703 * this point and the following smp_mb() is not strictly
2704 * necessary. Do smp_mb() anyway for documentation and
2705 * robustness against future implementation changes.
2706 */
2707 smp_mb(); /* See above comment block. */
2708 return 0;
2709}
2710
2711/**
2712 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2713 *
2714 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2715 * approach to force the grace period to end quickly. This consumes
2716 * significant time on all CPUs and is unfriendly to real-time workloads,
2717 * so is thus not recommended for any sort of common-case code. In fact,
2718 * if you are using synchronize_sched_expedited() in a loop, please
2719 * restructure your code to batch your updates, and then use a single
2720 * synchronize_sched() instead.
2721 *
2722 * Note that it is illegal to call this function while holding any lock
2723 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2724 * to call this function from a CPU-hotplug notifier. Failing to observe
2725 * these restriction will result in deadlock.
2726 *
2727 * This implementation can be thought of as an application of ticket
2728 * locking to RCU, with sync_sched_expedited_started and
2729 * sync_sched_expedited_done taking on the roles of the halves
2730 * of the ticket-lock word. Each task atomically increments
2731 * sync_sched_expedited_started upon entry, snapshotting the old value,
2732 * then attempts to stop all the CPUs. If this succeeds, then each
2733 * CPU will have executed a context switch, resulting in an RCU-sched
2734 * grace period. We are then done, so we use atomic_cmpxchg() to
2735 * update sync_sched_expedited_done to match our snapshot -- but
2736 * only if someone else has not already advanced past our snapshot.
2737 *
2738 * On the other hand, if try_stop_cpus() fails, we check the value
2739 * of sync_sched_expedited_done. If it has advanced past our
2740 * initial snapshot, then someone else must have forced a grace period
2741 * some time after we took our snapshot. In this case, our work is
2742 * done for us, and we can simply return. Otherwise, we try again,
2743 * but keep our initial snapshot for purposes of checking for someone
2744 * doing our work for us.
2745 *
2746 * If we fail too many times in a row, we fall back to synchronize_sched().
2747 */
2748void synchronize_sched_expedited(void)
2749{
2750 long firstsnap, s, snap;
2751 int trycount = 0;
2752 struct rcu_state *rsp = &rcu_sched_state;
2753
2754 /*
2755 * If we are in danger of counter wrap, just do synchronize_sched().
2756 * By allowing sync_sched_expedited_started to advance no more than
2757 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2758 * that more than 3.5 billion CPUs would be required to force a
2759 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2760 * course be required on a 64-bit system.
2761 */
2762 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2763 (ulong)atomic_long_read(&rsp->expedited_done) +
2764 ULONG_MAX / 8)) {
2765 synchronize_sched();
2766 atomic_long_inc(&rsp->expedited_wrap);
2767 return;
2768 }
2769
2770 /*
2771 * Take a ticket. Note that atomic_inc_return() implies a
2772 * full memory barrier.
2773 */
2774 snap = atomic_long_inc_return(&rsp->expedited_start);
2775 firstsnap = snap;
2776 get_online_cpus();
2777 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2778
2779 /*
2780 * Each pass through the following loop attempts to force a
2781 * context switch on each CPU.
2782 */
2783 while (try_stop_cpus(cpu_online_mask,
2784 synchronize_sched_expedited_cpu_stop,
2785 NULL) == -EAGAIN) {
2786 put_online_cpus();
2787 atomic_long_inc(&rsp->expedited_tryfail);
2788
2789 /* Check to see if someone else did our work for us. */
2790 s = atomic_long_read(&rsp->expedited_done);
2791 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2792 /* ensure test happens before caller kfree */
2793 smp_mb__before_atomic_inc(); /* ^^^ */
2794 atomic_long_inc(&rsp->expedited_workdone1);
2795 return;
2796 }
2797
2798 /* No joy, try again later. Or just synchronize_sched(). */
2799 if (trycount++ < 10) {
2800 udelay(trycount * num_online_cpus());
2801 } else {
2802 wait_rcu_gp(call_rcu_sched);
2803 atomic_long_inc(&rsp->expedited_normal);
2804 return;
2805 }
2806
2807 /* Recheck to see if someone else did our work for us. */
2808 s = atomic_long_read(&rsp->expedited_done);
2809 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2810 /* ensure test happens before caller kfree */
2811 smp_mb__before_atomic_inc(); /* ^^^ */
2812 atomic_long_inc(&rsp->expedited_workdone2);
2813 return;
2814 }
2815
2816 /*
2817 * Refetching sync_sched_expedited_started allows later
2818 * callers to piggyback on our grace period. We retry
2819 * after they started, so our grace period works for them,
2820 * and they started after our first try, so their grace
2821 * period works for us.
2822 */
2823 get_online_cpus();
2824 snap = atomic_long_read(&rsp->expedited_start);
2825 smp_mb(); /* ensure read is before try_stop_cpus(). */
2826 }
2827 atomic_long_inc(&rsp->expedited_stoppedcpus);
2828
2829 /*
2830 * Everyone up to our most recent fetch is covered by our grace
2831 * period. Update the counter, but only if our work is still
2832 * relevant -- which it won't be if someone who started later
2833 * than we did already did their update.
2834 */
2835 do {
2836 atomic_long_inc(&rsp->expedited_done_tries);
2837 s = atomic_long_read(&rsp->expedited_done);
2838 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2839 /* ensure test happens before caller kfree */
2840 smp_mb__before_atomic_inc(); /* ^^^ */
2841 atomic_long_inc(&rsp->expedited_done_lost);
2842 break;
2843 }
2844 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2845 atomic_long_inc(&rsp->expedited_done_exit);
2846
2847 put_online_cpus();
2848}
2849EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2850
2851/*
2852 * Check to see if there is any immediate RCU-related work to be done
2853 * by the current CPU, for the specified type of RCU, returning 1 if so.
2854 * The checks are in order of increasing expense: checks that can be
2855 * carried out against CPU-local state are performed first. However,
2856 * we must check for CPU stalls first, else we might not get a chance.
2857 */
2858static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2859{
2860 struct rcu_node *rnp = rdp->mynode;
2861
2862 rdp->n_rcu_pending++;
2863
2864 /* Check for CPU stalls, if enabled. */
2865 check_cpu_stall(rsp, rdp);
2866
2867 /* Is this CPU a NO_HZ_FULL CPU that should ignore RCU? */
2868 if (rcu_nohz_full_cpu(rsp))
2869 return 0;
2870
2871 /* Is the RCU core waiting for a quiescent state from this CPU? */
2872 if (rcu_scheduler_fully_active &&
2873 rdp->qs_pending && !rdp->passed_quiesce) {
2874 rdp->n_rp_qs_pending++;
2875 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2876 rdp->n_rp_report_qs++;
2877 return 1;
2878 }
2879
2880 /* Does this CPU have callbacks ready to invoke? */
2881 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2882 rdp->n_rp_cb_ready++;
2883 return 1;
2884 }
2885
2886 /* Has RCU gone idle with this CPU needing another grace period? */
2887 if (cpu_needs_another_gp(rsp, rdp)) {
2888 rdp->n_rp_cpu_needs_gp++;
2889 return 1;
2890 }
2891
2892 /* Has another RCU grace period completed? */
2893 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2894 rdp->n_rp_gp_completed++;
2895 return 1;
2896 }
2897
2898 /* Has a new RCU grace period started? */
2899 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2900 rdp->n_rp_gp_started++;
2901 return 1;
2902 }
2903
2904 /* Does this CPU need a deferred NOCB wakeup? */
2905 if (rcu_nocb_need_deferred_wakeup(rdp)) {
2906 rdp->n_rp_nocb_defer_wakeup++;
2907 return 1;
2908 }
2909
2910 /* nothing to do */
2911 rdp->n_rp_need_nothing++;
2912 return 0;
2913}
2914
2915/*
2916 * Check to see if there is any immediate RCU-related work to be done
2917 * by the current CPU, returning 1 if so. This function is part of the
2918 * RCU implementation; it is -not- an exported member of the RCU API.
2919 */
2920static int rcu_pending(int cpu)
2921{
2922 struct rcu_state *rsp;
2923
2924 for_each_rcu_flavor(rsp)
2925 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2926 return 1;
2927 return 0;
2928}
2929
2930/*
2931 * Return true if the specified CPU has any callback. If all_lazy is
2932 * non-NULL, store an indication of whether all callbacks are lazy.
2933 * (If there are no callbacks, all of them are deemed to be lazy.)
2934 */
2935static int __maybe_unused rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2936{
2937 bool al = true;
2938 bool hc = false;
2939 struct rcu_data *rdp;
2940 struct rcu_state *rsp;
2941
2942 for_each_rcu_flavor(rsp) {
2943 rdp = per_cpu_ptr(rsp->rda, cpu);
2944 if (!rdp->nxtlist)
2945 continue;
2946 hc = true;
2947 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2948 al = false;
2949 break;
2950 }
2951 }
2952 if (all_lazy)
2953 *all_lazy = al;
2954 return hc;
2955}
2956
2957/*
2958 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2959 * the compiler is expected to optimize this away.
2960 */
2961static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2962 int cpu, unsigned long done)
2963{
2964 trace_rcu_barrier(rsp->name, s, cpu,
2965 atomic_read(&rsp->barrier_cpu_count), done);
2966}
2967
2968/*
2969 * RCU callback function for _rcu_barrier(). If we are last, wake
2970 * up the task executing _rcu_barrier().
2971 */
2972static void rcu_barrier_callback(struct rcu_head *rhp)
2973{
2974 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2975 struct rcu_state *rsp = rdp->rsp;
2976
2977 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2978 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2979 complete(&rsp->barrier_completion);
2980 } else {
2981 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2982 }
2983}
2984
2985/*
2986 * Called with preemption disabled, and from cross-cpu IRQ context.
2987 */
2988static void rcu_barrier_func(void *type)
2989{
2990 struct rcu_state *rsp = type;
2991 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2992
2993 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2994 atomic_inc(&rsp->barrier_cpu_count);
2995 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2996}
2997
2998/*
2999 * Orchestrate the specified type of RCU barrier, waiting for all
3000 * RCU callbacks of the specified type to complete.
3001 */
3002static void _rcu_barrier(struct rcu_state *rsp)
3003{
3004 int cpu;
3005 struct rcu_data *rdp;
3006 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
3007 unsigned long snap_done;
3008
3009 _rcu_barrier_trace(rsp, "Begin", -1, snap);
3010
3011 /* Take mutex to serialize concurrent rcu_barrier() requests. */
3012 mutex_lock(&rsp->barrier_mutex);
3013
3014 /*
3015 * Ensure that all prior references, including to ->n_barrier_done,
3016 * are ordered before the _rcu_barrier() machinery.
3017 */
3018 smp_mb(); /* See above block comment. */
3019
3020 /*
3021 * Recheck ->n_barrier_done to see if others did our work for us.
3022 * This means checking ->n_barrier_done for an even-to-odd-to-even
3023 * transition. The "if" expression below therefore rounds the old
3024 * value up to the next even number and adds two before comparing.
3025 */
3026 snap_done = rsp->n_barrier_done;
3027 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
3028
3029 /*
3030 * If the value in snap is odd, we needed to wait for the current
3031 * rcu_barrier() to complete, then wait for the next one, in other
3032 * words, we need the value of snap_done to be three larger than
3033 * the value of snap. On the other hand, if the value in snap is
3034 * even, we only had to wait for the next rcu_barrier() to complete,
3035 * in other words, we need the value of snap_done to be only two
3036 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
3037 * this for us (thank you, Linus!).
3038 */
3039 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
3040 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
3041 smp_mb(); /* caller's subsequent code after above check. */
3042 mutex_unlock(&rsp->barrier_mutex);
3043 return;
3044 }
3045
3046 /*
3047 * Increment ->n_barrier_done to avoid duplicate work. Use
3048 * ACCESS_ONCE() to prevent the compiler from speculating
3049 * the increment to precede the early-exit check.
3050 */
3051 ACCESS_ONCE(rsp->n_barrier_done)++;
3052 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
3053 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
3054 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
3055
3056 /*
3057 * Initialize the count to one rather than to zero in order to
3058 * avoid a too-soon return to zero in case of a short grace period
3059 * (or preemption of this task). Exclude CPU-hotplug operations
3060 * to ensure that no offline CPU has callbacks queued.
3061 */
3062 init_completion(&rsp->barrier_completion);
3063 atomic_set(&rsp->barrier_cpu_count, 1);
3064 get_online_cpus();
3065
3066 /*
3067 * Force each CPU with callbacks to register a new callback.
3068 * When that callback is invoked, we will know that all of the
3069 * corresponding CPU's preceding callbacks have been invoked.
3070 */
3071 for_each_possible_cpu(cpu) {
3072 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
3073 continue;
3074 rdp = per_cpu_ptr(rsp->rda, cpu);
3075 if (rcu_is_nocb_cpu(cpu)) {
3076 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
3077 rsp->n_barrier_done);
3078 atomic_inc(&rsp->barrier_cpu_count);
3079 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
3080 rsp, cpu, 0);
3081 } else if (ACCESS_ONCE(rdp->qlen)) {
3082 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
3083 rsp->n_barrier_done);
3084 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
3085 } else {
3086 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
3087 rsp->n_barrier_done);
3088 }
3089 }
3090 put_online_cpus();
3091
3092 /*
3093 * Now that we have an rcu_barrier_callback() callback on each
3094 * CPU, and thus each counted, remove the initial count.
3095 */
3096 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
3097 complete(&rsp->barrier_completion);
3098
3099 /* Increment ->n_barrier_done to prevent duplicate work. */
3100 smp_mb(); /* Keep increment after above mechanism. */
3101 ACCESS_ONCE(rsp->n_barrier_done)++;
3102 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
3103 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
3104 smp_mb(); /* Keep increment before caller's subsequent code. */
3105
3106 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
3107 wait_for_completion(&rsp->barrier_completion);
3108
3109 /* Other rcu_barrier() invocations can now safely proceed. */
3110 mutex_unlock(&rsp->barrier_mutex);
3111}
3112
3113/**
3114 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
3115 */
3116void rcu_barrier_bh(void)
3117{
3118 _rcu_barrier(&rcu_bh_state);
3119}
3120EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3121
3122/**
3123 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3124 */
3125void rcu_barrier_sched(void)
3126{
3127 _rcu_barrier(&rcu_sched_state);
3128}
3129EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3130
3131/*
3132 * Do boot-time initialization of a CPU's per-CPU RCU data.
3133 */
3134static void __init
3135rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3136{
3137 unsigned long flags;
3138 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3139 struct rcu_node *rnp = rcu_get_root(rsp);
3140
3141 /* Set up local state, ensuring consistent view of global state. */
3142 raw_spin_lock_irqsave(&rnp->lock, flags);
3143 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3144 init_callback_list(rdp);
3145 rdp->qlen_lazy = 0;
3146 ACCESS_ONCE(rdp->qlen) = 0;
3147 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3148 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3149 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3150 rdp->cpu = cpu;
3151 rdp->rsp = rsp;
3152 rcu_boot_init_nocb_percpu_data(rdp);
3153 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3154}
3155
3156/*
3157 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3158 * offline event can be happening at a given time. Note also that we
3159 * can accept some slop in the rsp->completed access due to the fact
3160 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3161 */
3162static void
3163rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3164{
3165 unsigned long flags;
3166 unsigned long mask;
3167 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3168 struct rcu_node *rnp = rcu_get_root(rsp);
3169
3170 /* Exclude new grace periods. */
3171 mutex_lock(&rsp->onoff_mutex);
3172
3173 /* Set up local state, ensuring consistent view of global state. */
3174 raw_spin_lock_irqsave(&rnp->lock, flags);
3175 rdp->beenonline = 1; /* We have now been online. */
3176 rdp->preemptible = preemptible;
3177 rdp->qlen_last_fqs_check = 0;
3178 rdp->n_force_qs_snap = rsp->n_force_qs;
3179 rdp->blimit = blimit;
3180 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3181 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3182 rcu_sysidle_init_percpu_data(rdp->dynticks);
3183 atomic_set(&rdp->dynticks->dynticks,
3184 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3185 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3186
3187 /* Add CPU to rcu_node bitmasks. */
3188 rnp = rdp->mynode;
3189 mask = rdp->grpmask;
3190 do {
3191 /* Exclude any attempts to start a new GP on small systems. */
3192 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3193 rnp->qsmaskinit |= mask;
3194 mask = rnp->grpmask;
3195 if (rnp == rdp->mynode) {
3196 /*
3197 * If there is a grace period in progress, we will
3198 * set up to wait for it next time we run the
3199 * RCU core code.
3200 */
3201 rdp->gpnum = rnp->completed;
3202 rdp->completed = rnp->completed;
3203 rdp->passed_quiesce = 0;
3204 rdp->qs_pending = 0;
3205 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3206 }
3207 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3208 rnp = rnp->parent;
3209 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3210 local_irq_restore(flags);
3211
3212 mutex_unlock(&rsp->onoff_mutex);
3213}
3214
3215static void rcu_prepare_cpu(int cpu)
3216{
3217 struct rcu_state *rsp;
3218
3219 for_each_rcu_flavor(rsp)
3220 rcu_init_percpu_data(cpu, rsp,
3221 strcmp(rsp->name, "rcu_preempt") == 0);
3222}
3223
3224/*
3225 * Handle CPU online/offline notification events.
3226 */
3227static int rcu_cpu_notify(struct notifier_block *self,
3228 unsigned long action, void *hcpu)
3229{
3230 long cpu = (long)hcpu;
3231 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3232 struct rcu_node *rnp = rdp->mynode;
3233 struct rcu_state *rsp;
3234
3235 trace_rcu_utilization(TPS("Start CPU hotplug"));
3236 switch (action) {
3237 case CPU_UP_PREPARE:
3238 case CPU_UP_PREPARE_FROZEN:
3239 rcu_prepare_cpu(cpu);
3240 rcu_prepare_kthreads(cpu);
3241 break;
3242 case CPU_ONLINE:
3243 case CPU_DOWN_FAILED:
3244 rcu_boost_kthread_setaffinity(rnp, -1);
3245 break;
3246 case CPU_DOWN_PREPARE:
3247 rcu_boost_kthread_setaffinity(rnp, cpu);
3248 break;
3249 case CPU_DYING:
3250 case CPU_DYING_FROZEN:
3251 for_each_rcu_flavor(rsp)
3252 rcu_cleanup_dying_cpu(rsp);
3253 break;
3254 case CPU_DEAD:
3255 case CPU_DEAD_FROZEN:
3256 case CPU_UP_CANCELED:
3257 case CPU_UP_CANCELED_FROZEN:
3258 for_each_rcu_flavor(rsp)
3259 rcu_cleanup_dead_cpu(cpu, rsp);
3260 break;
3261 default:
3262 break;
3263 }
3264 trace_rcu_utilization(TPS("End CPU hotplug"));
3265 return NOTIFY_OK;
3266}
3267
3268static int rcu_pm_notify(struct notifier_block *self,
3269 unsigned long action, void *hcpu)
3270{
3271 switch (action) {
3272 case PM_HIBERNATION_PREPARE:
3273 case PM_SUSPEND_PREPARE:
3274 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3275 rcu_expedited = 1;
3276 break;
3277 case PM_POST_HIBERNATION:
3278 case PM_POST_SUSPEND:
3279 rcu_expedited = 0;
3280 break;
3281 default:
3282 break;
3283 }
3284 return NOTIFY_OK;
3285}
3286
3287/*
3288 * Spawn the kthread that handles this RCU flavor's grace periods.
3289 */
3290static int __init rcu_spawn_gp_kthread(void)
3291{
3292 unsigned long flags;
3293 struct rcu_node *rnp;
3294 struct rcu_state *rsp;
3295 struct task_struct *t;
3296
3297 for_each_rcu_flavor(rsp) {
3298 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3299 BUG_ON(IS_ERR(t));
3300 rnp = rcu_get_root(rsp);
3301 raw_spin_lock_irqsave(&rnp->lock, flags);
3302 rsp->gp_kthread = t;
3303 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3304 rcu_spawn_nocb_kthreads(rsp);
3305 }
3306 return 0;
3307}
3308early_initcall(rcu_spawn_gp_kthread);
3309
3310/*
3311 * This function is invoked towards the end of the scheduler's initialization
3312 * process. Before this is called, the idle task might contain
3313 * RCU read-side critical sections (during which time, this idle
3314 * task is booting the system). After this function is called, the
3315 * idle tasks are prohibited from containing RCU read-side critical
3316 * sections. This function also enables RCU lockdep checking.
3317 */
3318void rcu_scheduler_starting(void)
3319{
3320 WARN_ON(num_online_cpus() != 1);
3321 WARN_ON(nr_context_switches() > 0);
3322 rcu_scheduler_active = 1;
3323}
3324
3325/*
3326 * Compute the per-level fanout, either using the exact fanout specified
3327 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3328 */
3329#ifdef CONFIG_RCU_FANOUT_EXACT
3330static void __init rcu_init_levelspread(struct rcu_state *rsp)
3331{
3332 int i;
3333
3334 rsp->levelspread[rcu_num_lvls - 1] = rcu_fanout_leaf;
3335 for (i = rcu_num_lvls - 2; i >= 0; i--)
3336 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3337}
3338#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3339static void __init rcu_init_levelspread(struct rcu_state *rsp)
3340{
3341 int ccur;
3342 int cprv;
3343 int i;
3344
3345 cprv = nr_cpu_ids;
3346 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3347 ccur = rsp->levelcnt[i];
3348 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3349 cprv = ccur;
3350 }
3351}
3352#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3353
3354/*
3355 * Helper function for rcu_init() that initializes one rcu_state structure.
3356 */
3357static void __init rcu_init_one(struct rcu_state *rsp,
3358 struct rcu_data __percpu *rda)
3359{
3360 static char *buf[] = { "rcu_node_0",
3361 "rcu_node_1",
3362 "rcu_node_2",
3363 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3364 static char *fqs[] = { "rcu_node_fqs_0",
3365 "rcu_node_fqs_1",
3366 "rcu_node_fqs_2",
3367 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3368 int cpustride = 1;
3369 int i;
3370 int j;
3371 struct rcu_node *rnp;
3372
3373 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3374
3375 /* Silence gcc 4.8 warning about array index out of range. */
3376 if (rcu_num_lvls > RCU_NUM_LVLS)
3377 panic("rcu_init_one: rcu_num_lvls overflow");
3378
3379 /* Initialize the level-tracking arrays. */
3380
3381 for (i = 0; i < rcu_num_lvls; i++)
3382 rsp->levelcnt[i] = num_rcu_lvl[i];
3383 for (i = 1; i < rcu_num_lvls; i++)
3384 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3385 rcu_init_levelspread(rsp);
3386
3387 /* Initialize the elements themselves, starting from the leaves. */
3388
3389 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3390 cpustride *= rsp->levelspread[i];
3391 rnp = rsp->level[i];
3392 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3393 raw_spin_lock_init(&rnp->lock);
3394 lockdep_set_class_and_name(&rnp->lock,
3395 &rcu_node_class[i], buf[i]);
3396 raw_spin_lock_init(&rnp->fqslock);
3397 lockdep_set_class_and_name(&rnp->fqslock,
3398 &rcu_fqs_class[i], fqs[i]);
3399 rnp->gpnum = rsp->gpnum;
3400 rnp->completed = rsp->completed;
3401 rnp->qsmask = 0;
3402 rnp->qsmaskinit = 0;
3403 rnp->grplo = j * cpustride;
3404 rnp->grphi = (j + 1) * cpustride - 1;
3405 if (rnp->grphi >= NR_CPUS)
3406 rnp->grphi = NR_CPUS - 1;
3407 if (i == 0) {
3408 rnp->grpnum = 0;
3409 rnp->grpmask = 0;
3410 rnp->parent = NULL;
3411 } else {
3412 rnp->grpnum = j % rsp->levelspread[i - 1];
3413 rnp->grpmask = 1UL << rnp->grpnum;
3414 rnp->parent = rsp->level[i - 1] +
3415 j / rsp->levelspread[i - 1];
3416 }
3417 rnp->level = i;
3418 INIT_LIST_HEAD(&rnp->blkd_tasks);
3419 rcu_init_one_nocb(rnp);
3420 }
3421 }
3422
3423 rsp->rda = rda;
3424 init_waitqueue_head(&rsp->gp_wq);
3425 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3426 rnp = rsp->level[rcu_num_lvls - 1];
3427 for_each_possible_cpu(i) {
3428 while (i > rnp->grphi)
3429 rnp++;
3430 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3431 rcu_boot_init_percpu_data(i, rsp);
3432 }
3433 list_add(&rsp->flavors, &rcu_struct_flavors);
3434}
3435
3436/*
3437 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3438 * replace the definitions in tree.h because those are needed to size
3439 * the ->node array in the rcu_state structure.
3440 */
3441static void __init rcu_init_geometry(void)
3442{
3443 ulong d;
3444 int i;
3445 int j;
3446 int n = nr_cpu_ids;
3447 int rcu_capacity[MAX_RCU_LVLS + 1];
3448
3449 /*
3450 * Initialize any unspecified boot parameters.
3451 * The default values of jiffies_till_first_fqs and
3452 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3453 * value, which is a function of HZ, then adding one for each
3454 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3455 */
3456 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3457 if (jiffies_till_first_fqs == ULONG_MAX)
3458 jiffies_till_first_fqs = d;
3459 if (jiffies_till_next_fqs == ULONG_MAX)
3460 jiffies_till_next_fqs = d;
3461
3462 /* If the compile-time values are accurate, just leave. */
3463 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3464 nr_cpu_ids == NR_CPUS)
3465 return;
3466 pr_info("RCU: Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%d\n",
3467 rcu_fanout_leaf, nr_cpu_ids);
3468
3469 /*
3470 * Compute number of nodes that can be handled an rcu_node tree
3471 * with the given number of levels. Setting rcu_capacity[0] makes
3472 * some of the arithmetic easier.
3473 */
3474 rcu_capacity[0] = 1;
3475 rcu_capacity[1] = rcu_fanout_leaf;
3476 for (i = 2; i <= MAX_RCU_LVLS; i++)
3477 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3478
3479 /*
3480 * The boot-time rcu_fanout_leaf parameter is only permitted
3481 * to increase the leaf-level fanout, not decrease it. Of course,
3482 * the leaf-level fanout cannot exceed the number of bits in
3483 * the rcu_node masks. Finally, the tree must be able to accommodate
3484 * the configured number of CPUs. Complain and fall back to the
3485 * compile-time values if these limits are exceeded.
3486 */
3487 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3488 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3489 n > rcu_capacity[MAX_RCU_LVLS]) {
3490 WARN_ON(1);
3491 return;
3492 }
3493
3494 /* Calculate the number of rcu_nodes at each level of the tree. */
3495 for (i = 1; i <= MAX_RCU_LVLS; i++)
3496 if (n <= rcu_capacity[i]) {
3497 for (j = 0; j <= i; j++)
3498 num_rcu_lvl[j] =
3499 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3500 rcu_num_lvls = i;
3501 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3502 num_rcu_lvl[j] = 0;
3503 break;
3504 }
3505
3506 /* Calculate the total number of rcu_node structures. */
3507 rcu_num_nodes = 0;
3508 for (i = 0; i <= MAX_RCU_LVLS; i++)
3509 rcu_num_nodes += num_rcu_lvl[i];
3510 rcu_num_nodes -= n;
3511}
3512
3513void __init rcu_init(void)
3514{
3515 int cpu;
3516
3517 rcu_bootup_announce();
3518 rcu_init_geometry();
3519 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3520 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3521 __rcu_init_preempt();
3522 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3523
3524 /*
3525 * We don't need protection against CPU-hotplug here because
3526 * this is called early in boot, before either interrupts
3527 * or the scheduler are operational.
3528 */
3529 cpu_notifier(rcu_cpu_notify, 0);
3530 pm_notifier(rcu_pm_notify, 0);
3531 for_each_online_cpu(cpu)
3532 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3533}
3534
3535#include "tree_plugin.h"