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// SPDX-License-Identifier: GPL-2.0+
2/*
3 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
4 *
5 * Copyright IBM Corporation, 2008
6 *
7 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
8 * Manfred Spraul <manfred@colorfullife.com>
9 * Paul E. McKenney <paulmck@linux.ibm.com>
10 *
11 * Based on the original work by Paul McKenney <paulmck@linux.ibm.com>
12 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
13 *
14 * For detailed explanation of Read-Copy Update mechanism see -
15 * Documentation/RCU
16 */
17
18#define pr_fmt(fmt) "rcu: " fmt
19
20#include <linux/types.h>
21#include <linux/kernel.h>
22#include <linux/init.h>
23#include <linux/spinlock.h>
24#include <linux/smp.h>
25#include <linux/rcupdate_wait.h>
26#include <linux/interrupt.h>
27#include <linux/sched.h>
28#include <linux/sched/debug.h>
29#include <linux/nmi.h>
30#include <linux/atomic.h>
31#include <linux/bitops.h>
32#include <linux/export.h>
33#include <linux/completion.h>
34#include <linux/moduleparam.h>
35#include <linux/panic.h>
36#include <linux/panic_notifier.h>
37#include <linux/percpu.h>
38#include <linux/notifier.h>
39#include <linux/cpu.h>
40#include <linux/mutex.h>
41#include <linux/time.h>
42#include <linux/kernel_stat.h>
43#include <linux/wait.h>
44#include <linux/kthread.h>
45#include <uapi/linux/sched/types.h>
46#include <linux/prefetch.h>
47#include <linux/delay.h>
48#include <linux/random.h>
49#include <linux/trace_events.h>
50#include <linux/suspend.h>
51#include <linux/ftrace.h>
52#include <linux/tick.h>
53#include <linux/sysrq.h>
54#include <linux/kprobes.h>
55#include <linux/gfp.h>
56#include <linux/oom.h>
57#include <linux/smpboot.h>
58#include <linux/jiffies.h>
59#include <linux/slab.h>
60#include <linux/sched/isolation.h>
61#include <linux/sched/clock.h>
62#include <linux/vmalloc.h>
63#include <linux/mm.h>
64#include <linux/kasan.h>
65#include "../time/tick-internal.h"
66
67#include "tree.h"
68#include "rcu.h"
69
70#ifdef MODULE_PARAM_PREFIX
71#undef MODULE_PARAM_PREFIX
72#endif
73#define MODULE_PARAM_PREFIX "rcutree."
74
75/* Data structures. */
76
77/*
78 * Steal a bit from the bottom of ->dynticks for idle entry/exit
79 * control. Initially this is for TLB flushing.
80 */
81#define RCU_DYNTICK_CTRL_MASK 0x1
82#define RCU_DYNTICK_CTRL_CTR (RCU_DYNTICK_CTRL_MASK + 1)
83
84static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
85 .dynticks_nesting = 1,
86 .dynticks_nmi_nesting = DYNTICK_IRQ_NONIDLE,
87 .dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
88#ifdef CONFIG_RCU_NOCB_CPU
89 .cblist.flags = SEGCBLIST_SOFTIRQ_ONLY,
90#endif
91};
92static struct rcu_state rcu_state = {
93 .level = { &rcu_state.node[0] },
94 .gp_state = RCU_GP_IDLE,
95 .gp_seq = (0UL - 300UL) << RCU_SEQ_CTR_SHIFT,
96 .barrier_mutex = __MUTEX_INITIALIZER(rcu_state.barrier_mutex),
97 .name = RCU_NAME,
98 .abbr = RCU_ABBR,
99 .exp_mutex = __MUTEX_INITIALIZER(rcu_state.exp_mutex),
100 .exp_wake_mutex = __MUTEX_INITIALIZER(rcu_state.exp_wake_mutex),
101 .ofl_lock = __RAW_SPIN_LOCK_UNLOCKED(rcu_state.ofl_lock),
102};
103
104/* Dump rcu_node combining tree at boot to verify correct setup. */
105static bool dump_tree;
106module_param(dump_tree, bool, 0444);
107/* By default, use RCU_SOFTIRQ instead of rcuc kthreads. */
108static bool use_softirq = !IS_ENABLED(CONFIG_PREEMPT_RT);
109#ifndef CONFIG_PREEMPT_RT
110module_param(use_softirq, bool, 0444);
111#endif
112/* Control rcu_node-tree auto-balancing at boot time. */
113static bool rcu_fanout_exact;
114module_param(rcu_fanout_exact, bool, 0444);
115/* Increase (but not decrease) the RCU_FANOUT_LEAF at boot time. */
116static int rcu_fanout_leaf = RCU_FANOUT_LEAF;
117module_param(rcu_fanout_leaf, int, 0444);
118int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
119/* Number of rcu_nodes at specified level. */
120int num_rcu_lvl[] = NUM_RCU_LVL_INIT;
121int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
122
123/*
124 * The rcu_scheduler_active variable is initialized to the value
125 * RCU_SCHEDULER_INACTIVE and transitions RCU_SCHEDULER_INIT just before the
126 * first task is spawned. So when this variable is RCU_SCHEDULER_INACTIVE,
127 * RCU can assume that there is but one task, allowing RCU to (for example)
128 * optimize synchronize_rcu() to a simple barrier(). When this variable
129 * is RCU_SCHEDULER_INIT, RCU must actually do all the hard work required
130 * to detect real grace periods. This variable is also used to suppress
131 * boot-time false positives from lockdep-RCU error checking. Finally, it
132 * transitions from RCU_SCHEDULER_INIT to RCU_SCHEDULER_RUNNING after RCU
133 * is fully initialized, including all of its kthreads having been spawned.
134 */
135int rcu_scheduler_active __read_mostly;
136EXPORT_SYMBOL_GPL(rcu_scheduler_active);
137
138/*
139 * The rcu_scheduler_fully_active variable transitions from zero to one
140 * during the early_initcall() processing, which is after the scheduler
141 * is capable of creating new tasks. So RCU processing (for example,
142 * creating tasks for RCU priority boosting) must be delayed until after
143 * rcu_scheduler_fully_active transitions from zero to one. We also
144 * currently delay invocation of any RCU callbacks until after this point.
145 *
146 * It might later prove better for people registering RCU callbacks during
147 * early boot to take responsibility for these callbacks, but one step at
148 * a time.
149 */
150static int rcu_scheduler_fully_active __read_mostly;
151
152static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
153 unsigned long gps, unsigned long flags);
154static void rcu_init_new_rnp(struct rcu_node *rnp_leaf);
155static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf);
156static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
157static void invoke_rcu_core(void);
158static void rcu_report_exp_rdp(struct rcu_data *rdp);
159static void sync_sched_exp_online_cleanup(int cpu);
160static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp);
161static bool rcu_rdp_is_offloaded(struct rcu_data *rdp);
162
163/* rcuc/rcub kthread realtime priority */
164static int kthread_prio = IS_ENABLED(CONFIG_RCU_BOOST) ? 1 : 0;
165module_param(kthread_prio, int, 0444);
166
167/* Delay in jiffies for grace-period initialization delays, debug only. */
168
169static int gp_preinit_delay;
170module_param(gp_preinit_delay, int, 0444);
171static int gp_init_delay;
172module_param(gp_init_delay, int, 0444);
173static int gp_cleanup_delay;
174module_param(gp_cleanup_delay, int, 0444);
175
176// Add delay to rcu_read_unlock() for strict grace periods.
177static int rcu_unlock_delay;
178#ifdef CONFIG_RCU_STRICT_GRACE_PERIOD
179module_param(rcu_unlock_delay, int, 0444);
180#endif
181
182/*
183 * This rcu parameter is runtime-read-only. It reflects
184 * a minimum allowed number of objects which can be cached
185 * per-CPU. Object size is equal to one page. This value
186 * can be changed at boot time.
187 */
188static int rcu_min_cached_objs = 5;
189module_param(rcu_min_cached_objs, int, 0444);
190
191// A page shrinker can ask for pages to be freed to make them
192// available for other parts of the system. This usually happens
193// under low memory conditions, and in that case we should also
194// defer page-cache filling for a short time period.
195//
196// The default value is 5 seconds, which is long enough to reduce
197// interference with the shrinker while it asks other systems to
198// drain their caches.
199static int rcu_delay_page_cache_fill_msec = 5000;
200module_param(rcu_delay_page_cache_fill_msec, int, 0444);
201
202/* Retrieve RCU kthreads priority for rcutorture */
203int rcu_get_gp_kthreads_prio(void)
204{
205 return kthread_prio;
206}
207EXPORT_SYMBOL_GPL(rcu_get_gp_kthreads_prio);
208
209/*
210 * Number of grace periods between delays, normalized by the duration of
211 * the delay. The longer the delay, the more the grace periods between
212 * each delay. The reason for this normalization is that it means that,
213 * for non-zero delays, the overall slowdown of grace periods is constant
214 * regardless of the duration of the delay. This arrangement balances
215 * the need for long delays to increase some race probabilities with the
216 * need for fast grace periods to increase other race probabilities.
217 */
218#define PER_RCU_NODE_PERIOD 3 /* Number of grace periods between delays for debugging. */
219
220/*
221 * Compute the mask of online CPUs for the specified rcu_node structure.
222 * This will not be stable unless the rcu_node structure's ->lock is
223 * held, but the bit corresponding to the current CPU will be stable
224 * in most contexts.
225 */
226static unsigned long rcu_rnp_online_cpus(struct rcu_node *rnp)
227{
228 return READ_ONCE(rnp->qsmaskinitnext);
229}
230
231/*
232 * Return true if an RCU grace period is in progress. The READ_ONCE()s
233 * permit this function to be invoked without holding the root rcu_node
234 * structure's ->lock, but of course results can be subject to change.
235 */
236static int rcu_gp_in_progress(void)
237{
238 return rcu_seq_state(rcu_seq_current(&rcu_state.gp_seq));
239}
240
241/*
242 * Return the number of callbacks queued on the specified CPU.
243 * Handles both the nocbs and normal cases.
244 */
245static long rcu_get_n_cbs_cpu(int cpu)
246{
247 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
248
249 if (rcu_segcblist_is_enabled(&rdp->cblist))
250 return rcu_segcblist_n_cbs(&rdp->cblist);
251 return 0;
252}
253
254void rcu_softirq_qs(void)
255{
256 rcu_qs();
257 rcu_preempt_deferred_qs(current);
258 rcu_tasks_qs(current, false);
259}
260
261/*
262 * Record entry into an extended quiescent state. This is only to be
263 * called when not already in an extended quiescent state, that is,
264 * RCU is watching prior to the call to this function and is no longer
265 * watching upon return.
266 */
267static noinstr void rcu_dynticks_eqs_enter(void)
268{
269 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
270 int seq;
271
272 /*
273 * CPUs seeing atomic_add_return() must see prior RCU read-side
274 * critical sections, and we also must force ordering with the
275 * next idle sojourn.
276 */
277 rcu_dynticks_task_trace_enter(); // Before ->dynticks update!
278 seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
279 // RCU is no longer watching. Better be in extended quiescent state!
280 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
281 (seq & RCU_DYNTICK_CTRL_CTR));
282 /* Better not have special action (TLB flush) pending! */
283 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
284 (seq & RCU_DYNTICK_CTRL_MASK));
285}
286
287/*
288 * Record exit from an extended quiescent state. This is only to be
289 * called from an extended quiescent state, that is, RCU is not watching
290 * prior to the call to this function and is watching upon return.
291 */
292static noinstr void rcu_dynticks_eqs_exit(void)
293{
294 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
295 int seq;
296
297 /*
298 * CPUs seeing atomic_add_return() must see prior idle sojourns,
299 * and we also must force ordering with the next RCU read-side
300 * critical section.
301 */
302 seq = arch_atomic_add_return(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
303 // RCU is now watching. Better not be in an extended quiescent state!
304 rcu_dynticks_task_trace_exit(); // After ->dynticks update!
305 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
306 !(seq & RCU_DYNTICK_CTRL_CTR));
307 if (seq & RCU_DYNTICK_CTRL_MASK) {
308 arch_atomic_andnot(RCU_DYNTICK_CTRL_MASK, &rdp->dynticks);
309 smp_mb__after_atomic(); /* _exit after clearing mask. */
310 }
311}
312
313/*
314 * Reset the current CPU's ->dynticks counter to indicate that the
315 * newly onlined CPU is no longer in an extended quiescent state.
316 * This will either leave the counter unchanged, or increment it
317 * to the next non-quiescent value.
318 *
319 * The non-atomic test/increment sequence works because the upper bits
320 * of the ->dynticks counter are manipulated only by the corresponding CPU,
321 * or when the corresponding CPU is offline.
322 */
323static void rcu_dynticks_eqs_online(void)
324{
325 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
326
327 if (atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR)
328 return;
329 atomic_add(RCU_DYNTICK_CTRL_CTR, &rdp->dynticks);
330}
331
332/*
333 * Is the current CPU in an extended quiescent state?
334 *
335 * No ordering, as we are sampling CPU-local information.
336 */
337static __always_inline bool rcu_dynticks_curr_cpu_in_eqs(void)
338{
339 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
340
341 return !(arch_atomic_read(&rdp->dynticks) & RCU_DYNTICK_CTRL_CTR);
342}
343
344/*
345 * Snapshot the ->dynticks counter with full ordering so as to allow
346 * stable comparison of this counter with past and future snapshots.
347 */
348static int rcu_dynticks_snap(struct rcu_data *rdp)
349{
350 int snap = atomic_add_return(0, &rdp->dynticks);
351
352 return snap & ~RCU_DYNTICK_CTRL_MASK;
353}
354
355/*
356 * Return true if the snapshot returned from rcu_dynticks_snap()
357 * indicates that RCU is in an extended quiescent state.
358 */
359static bool rcu_dynticks_in_eqs(int snap)
360{
361 return !(snap & RCU_DYNTICK_CTRL_CTR);
362}
363
364/* Return true if the specified CPU is currently idle from an RCU viewpoint. */
365bool rcu_is_idle_cpu(int cpu)
366{
367 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
368
369 return rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp));
370}
371
372/*
373 * Return true if the CPU corresponding to the specified rcu_data
374 * structure has spent some time in an extended quiescent state since
375 * rcu_dynticks_snap() returned the specified snapshot.
376 */
377static bool rcu_dynticks_in_eqs_since(struct rcu_data *rdp, int snap)
378{
379 return snap != rcu_dynticks_snap(rdp);
380}
381
382/*
383 * Return true if the referenced integer is zero while the specified
384 * CPU remains within a single extended quiescent state.
385 */
386bool rcu_dynticks_zero_in_eqs(int cpu, int *vp)
387{
388 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
389 int snap;
390
391 // If not quiescent, force back to earlier extended quiescent state.
392 snap = atomic_read(&rdp->dynticks) & ~(RCU_DYNTICK_CTRL_MASK |
393 RCU_DYNTICK_CTRL_CTR);
394
395 smp_rmb(); // Order ->dynticks and *vp reads.
396 if (READ_ONCE(*vp))
397 return false; // Non-zero, so report failure;
398 smp_rmb(); // Order *vp read and ->dynticks re-read.
399
400 // If still in the same extended quiescent state, we are good!
401 return snap == (atomic_read(&rdp->dynticks) & ~RCU_DYNTICK_CTRL_MASK);
402}
403
404/*
405 * Set the special (bottom) bit of the specified CPU so that it
406 * will take special action (such as flushing its TLB) on the
407 * next exit from an extended quiescent state. Returns true if
408 * the bit was successfully set, or false if the CPU was not in
409 * an extended quiescent state.
410 */
411bool rcu_eqs_special_set(int cpu)
412{
413 int old;
414 int new;
415 int new_old;
416 struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
417
418 new_old = atomic_read(&rdp->dynticks);
419 do {
420 old = new_old;
421 if (old & RCU_DYNTICK_CTRL_CTR)
422 return false;
423 new = old | RCU_DYNTICK_CTRL_MASK;
424 new_old = atomic_cmpxchg(&rdp->dynticks, old, new);
425 } while (new_old != old);
426 return true;
427}
428
429/*
430 * Let the RCU core know that this CPU has gone through the scheduler,
431 * which is a quiescent state. This is called when the need for a
432 * quiescent state is urgent, so we burn an atomic operation and full
433 * memory barriers to let the RCU core know about it, regardless of what
434 * this CPU might (or might not) do in the near future.
435 *
436 * We inform the RCU core by emulating a zero-duration dyntick-idle period.
437 *
438 * The caller must have disabled interrupts and must not be idle.
439 */
440notrace void rcu_momentary_dyntick_idle(void)
441{
442 int special;
443
444 raw_cpu_write(rcu_data.rcu_need_heavy_qs, false);
445 special = atomic_add_return(2 * RCU_DYNTICK_CTRL_CTR,
446 &this_cpu_ptr(&rcu_data)->dynticks);
447 /* It is illegal to call this from idle state. */
448 WARN_ON_ONCE(!(special & RCU_DYNTICK_CTRL_CTR));
449 rcu_preempt_deferred_qs(current);
450}
451EXPORT_SYMBOL_GPL(rcu_momentary_dyntick_idle);
452
453/**
454 * rcu_is_cpu_rrupt_from_idle - see if 'interrupted' from idle
455 *
456 * If the current CPU is idle and running at a first-level (not nested)
457 * interrupt, or directly, from idle, return true.
458 *
459 * The caller must have at least disabled IRQs.
460 */
461static int rcu_is_cpu_rrupt_from_idle(void)
462{
463 long nesting;
464
465 /*
466 * Usually called from the tick; but also used from smp_function_call()
467 * for expedited grace periods. This latter can result in running from
468 * the idle task, instead of an actual IPI.
469 */
470 lockdep_assert_irqs_disabled();
471
472 /* Check for counter underflows */
473 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
474 "RCU dynticks_nesting counter underflow!");
475 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
476 "RCU dynticks_nmi_nesting counter underflow/zero!");
477
478 /* Are we at first interrupt nesting level? */
479 nesting = __this_cpu_read(rcu_data.dynticks_nmi_nesting);
480 if (nesting > 1)
481 return false;
482
483 /*
484 * If we're not in an interrupt, we must be in the idle task!
485 */
486 WARN_ON_ONCE(!nesting && !is_idle_task(current));
487
488 /* Does CPU appear to be idle from an RCU standpoint? */
489 return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
490}
491
492#define DEFAULT_RCU_BLIMIT (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 1000 : 10)
493 // Maximum callbacks per rcu_do_batch ...
494#define DEFAULT_MAX_RCU_BLIMIT 10000 // ... even during callback flood.
495static long blimit = DEFAULT_RCU_BLIMIT;
496#define DEFAULT_RCU_QHIMARK 10000 // If this many pending, ignore blimit.
497static long qhimark = DEFAULT_RCU_QHIMARK;
498#define DEFAULT_RCU_QLOMARK 100 // Once only this many pending, use blimit.
499static long qlowmark = DEFAULT_RCU_QLOMARK;
500#define DEFAULT_RCU_QOVLD_MULT 2
501#define DEFAULT_RCU_QOVLD (DEFAULT_RCU_QOVLD_MULT * DEFAULT_RCU_QHIMARK)
502static long qovld = DEFAULT_RCU_QOVLD; // If this many pending, hammer QS.
503static long qovld_calc = -1; // No pre-initialization lock acquisitions!
504
505module_param(blimit, long, 0444);
506module_param(qhimark, long, 0444);
507module_param(qlowmark, long, 0444);
508module_param(qovld, long, 0444);
509
510static ulong jiffies_till_first_fqs = IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD) ? 0 : ULONG_MAX;
511static ulong jiffies_till_next_fqs = ULONG_MAX;
512static bool rcu_kick_kthreads;
513static int rcu_divisor = 7;
514module_param(rcu_divisor, int, 0644);
515
516/* Force an exit from rcu_do_batch() after 3 milliseconds. */
517static long rcu_resched_ns = 3 * NSEC_PER_MSEC;
518module_param(rcu_resched_ns, long, 0644);
519
520/*
521 * How long the grace period must be before we start recruiting
522 * quiescent-state help from rcu_note_context_switch().
523 */
524static ulong jiffies_till_sched_qs = ULONG_MAX;
525module_param(jiffies_till_sched_qs, ulong, 0444);
526static ulong jiffies_to_sched_qs; /* See adjust_jiffies_till_sched_qs(). */
527module_param(jiffies_to_sched_qs, ulong, 0444); /* Display only! */
528
529/*
530 * Make sure that we give the grace-period kthread time to detect any
531 * idle CPUs before taking active measures to force quiescent states.
532 * However, don't go below 100 milliseconds, adjusted upwards for really
533 * large systems.
534 */
535static void adjust_jiffies_till_sched_qs(void)
536{
537 unsigned long j;
538
539 /* If jiffies_till_sched_qs was specified, respect the request. */
540 if (jiffies_till_sched_qs != ULONG_MAX) {
541 WRITE_ONCE(jiffies_to_sched_qs, jiffies_till_sched_qs);
542 return;
543 }
544 /* Otherwise, set to third fqs scan, but bound below on large system. */
545 j = READ_ONCE(jiffies_till_first_fqs) +
546 2 * READ_ONCE(jiffies_till_next_fqs);
547 if (j < HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV)
548 j = HZ / 10 + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
549 pr_info("RCU calculated value of scheduler-enlistment delay is %ld jiffies.\n", j);
550 WRITE_ONCE(jiffies_to_sched_qs, j);
551}
552
553static int param_set_first_fqs_jiffies(const char *val, const struct kernel_param *kp)
554{
555 ulong j;
556 int ret = kstrtoul(val, 0, &j);
557
558 if (!ret) {
559 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : j);
560 adjust_jiffies_till_sched_qs();
561 }
562 return ret;
563}
564
565static int param_set_next_fqs_jiffies(const char *val, const struct kernel_param *kp)
566{
567 ulong j;
568 int ret = kstrtoul(val, 0, &j);
569
570 if (!ret) {
571 WRITE_ONCE(*(ulong *)kp->arg, (j > HZ) ? HZ : (j ?: 1));
572 adjust_jiffies_till_sched_qs();
573 }
574 return ret;
575}
576
577static const struct kernel_param_ops first_fqs_jiffies_ops = {
578 .set = param_set_first_fqs_jiffies,
579 .get = param_get_ulong,
580};
581
582static const struct kernel_param_ops next_fqs_jiffies_ops = {
583 .set = param_set_next_fqs_jiffies,
584 .get = param_get_ulong,
585};
586
587module_param_cb(jiffies_till_first_fqs, &first_fqs_jiffies_ops, &jiffies_till_first_fqs, 0644);
588module_param_cb(jiffies_till_next_fqs, &next_fqs_jiffies_ops, &jiffies_till_next_fqs, 0644);
589module_param(rcu_kick_kthreads, bool, 0644);
590
591static void force_qs_rnp(int (*f)(struct rcu_data *rdp));
592static int rcu_pending(int user);
593
594/*
595 * Return the number of RCU GPs completed thus far for debug & stats.
596 */
597unsigned long rcu_get_gp_seq(void)
598{
599 return READ_ONCE(rcu_state.gp_seq);
600}
601EXPORT_SYMBOL_GPL(rcu_get_gp_seq);
602
603/*
604 * Return the number of RCU expedited batches completed thus far for
605 * debug & stats. Odd numbers mean that a batch is in progress, even
606 * numbers mean idle. The value returned will thus be roughly double
607 * the cumulative batches since boot.
608 */
609unsigned long rcu_exp_batches_completed(void)
610{
611 return rcu_state.expedited_sequence;
612}
613EXPORT_SYMBOL_GPL(rcu_exp_batches_completed);
614
615/*
616 * Return the root node of the rcu_state structure.
617 */
618static struct rcu_node *rcu_get_root(void)
619{
620 return &rcu_state.node[0];
621}
622
623/*
624 * Send along grace-period-related data for rcutorture diagnostics.
625 */
626void rcutorture_get_gp_data(enum rcutorture_type test_type, int *flags,
627 unsigned long *gp_seq)
628{
629 switch (test_type) {
630 case RCU_FLAVOR:
631 *flags = READ_ONCE(rcu_state.gp_flags);
632 *gp_seq = rcu_seq_current(&rcu_state.gp_seq);
633 break;
634 default:
635 break;
636 }
637}
638EXPORT_SYMBOL_GPL(rcutorture_get_gp_data);
639
640/*
641 * Enter an RCU extended quiescent state, which can be either the
642 * idle loop or adaptive-tickless usermode execution.
643 *
644 * We crowbar the ->dynticks_nmi_nesting field to zero to allow for
645 * the possibility of usermode upcalls having messed up our count
646 * of interrupt nesting level during the prior busy period.
647 */
648static noinstr void rcu_eqs_enter(bool user)
649{
650 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
651
652 WARN_ON_ONCE(rdp->dynticks_nmi_nesting != DYNTICK_IRQ_NONIDLE);
653 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
654 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) &&
655 rdp->dynticks_nesting == 0);
656 if (rdp->dynticks_nesting != 1) {
657 // RCU will still be watching, so just do accounting and leave.
658 rdp->dynticks_nesting--;
659 return;
660 }
661
662 lockdep_assert_irqs_disabled();
663 instrumentation_begin();
664 trace_rcu_dyntick(TPS("Start"), rdp->dynticks_nesting, 0, atomic_read(&rdp->dynticks));
665 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
666 rcu_prepare_for_idle();
667 rcu_preempt_deferred_qs(current);
668
669 // instrumentation for the noinstr rcu_dynticks_eqs_enter()
670 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
671
672 instrumentation_end();
673 WRITE_ONCE(rdp->dynticks_nesting, 0); /* Avoid irq-access tearing. */
674 // RCU is watching here ...
675 rcu_dynticks_eqs_enter();
676 // ... but is no longer watching here.
677 rcu_dynticks_task_enter();
678}
679
680/**
681 * rcu_idle_enter - inform RCU that current CPU is entering idle
682 *
683 * Enter idle mode, in other words, -leave- the mode in which RCU
684 * read-side critical sections can occur. (Though RCU read-side
685 * critical sections can occur in irq handlers in idle, a possibility
686 * handled by irq_enter() and irq_exit().)
687 *
688 * If you add or remove a call to rcu_idle_enter(), be sure to test with
689 * CONFIG_RCU_EQS_DEBUG=y.
690 */
691void rcu_idle_enter(void)
692{
693 lockdep_assert_irqs_disabled();
694 rcu_eqs_enter(false);
695}
696EXPORT_SYMBOL_GPL(rcu_idle_enter);
697
698#ifdef CONFIG_NO_HZ_FULL
699
700#if !defined(CONFIG_GENERIC_ENTRY) || !defined(CONFIG_KVM_XFER_TO_GUEST_WORK)
701/*
702 * An empty function that will trigger a reschedule on
703 * IRQ tail once IRQs get re-enabled on userspace/guest resume.
704 */
705static void late_wakeup_func(struct irq_work *work)
706{
707}
708
709static DEFINE_PER_CPU(struct irq_work, late_wakeup_work) =
710 IRQ_WORK_INIT(late_wakeup_func);
711
712/*
713 * If either:
714 *
715 * 1) the task is about to enter in guest mode and $ARCH doesn't support KVM generic work
716 * 2) the task is about to enter in user mode and $ARCH doesn't support generic entry.
717 *
718 * In these cases the late RCU wake ups aren't supported in the resched loops and our
719 * last resort is to fire a local irq_work that will trigger a reschedule once IRQs
720 * get re-enabled again.
721 */
722noinstr static void rcu_irq_work_resched(void)
723{
724 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
725
726 if (IS_ENABLED(CONFIG_GENERIC_ENTRY) && !(current->flags & PF_VCPU))
727 return;
728
729 if (IS_ENABLED(CONFIG_KVM_XFER_TO_GUEST_WORK) && (current->flags & PF_VCPU))
730 return;
731
732 instrumentation_begin();
733 if (do_nocb_deferred_wakeup(rdp) && need_resched()) {
734 irq_work_queue(this_cpu_ptr(&late_wakeup_work));
735 }
736 instrumentation_end();
737}
738
739#else
740static inline void rcu_irq_work_resched(void) { }
741#endif
742
743/**
744 * rcu_user_enter - inform RCU that we are resuming userspace.
745 *
746 * Enter RCU idle mode right before resuming userspace. No use of RCU
747 * is permitted between this call and rcu_user_exit(). This way the
748 * CPU doesn't need to maintain the tick for RCU maintenance purposes
749 * when the CPU runs in userspace.
750 *
751 * If you add or remove a call to rcu_user_enter(), be sure to test with
752 * CONFIG_RCU_EQS_DEBUG=y.
753 */
754noinstr void rcu_user_enter(void)
755{
756 lockdep_assert_irqs_disabled();
757
758 /*
759 * Other than generic entry implementation, we may be past the last
760 * rescheduling opportunity in the entry code. Trigger a self IPI
761 * that will fire and reschedule once we resume in user/guest mode.
762 */
763 rcu_irq_work_resched();
764 rcu_eqs_enter(true);
765}
766
767#endif /* CONFIG_NO_HZ_FULL */
768
769/**
770 * rcu_nmi_exit - inform RCU of exit from NMI context
771 *
772 * If we are returning from the outermost NMI handler that interrupted an
773 * RCU-idle period, update rdp->dynticks and rdp->dynticks_nmi_nesting
774 * to let the RCU grace-period handling know that the CPU is back to
775 * being RCU-idle.
776 *
777 * If you add or remove a call to rcu_nmi_exit(), be sure to test
778 * with CONFIG_RCU_EQS_DEBUG=y.
779 */
780noinstr void rcu_nmi_exit(void)
781{
782 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
783
784 instrumentation_begin();
785 /*
786 * Check for ->dynticks_nmi_nesting underflow and bad ->dynticks.
787 * (We are exiting an NMI handler, so RCU better be paying attention
788 * to us!)
789 */
790 WARN_ON_ONCE(rdp->dynticks_nmi_nesting <= 0);
791 WARN_ON_ONCE(rcu_dynticks_curr_cpu_in_eqs());
792
793 /*
794 * If the nesting level is not 1, the CPU wasn't RCU-idle, so
795 * leave it in non-RCU-idle state.
796 */
797 if (rdp->dynticks_nmi_nesting != 1) {
798 trace_rcu_dyntick(TPS("--="), rdp->dynticks_nmi_nesting, rdp->dynticks_nmi_nesting - 2,
799 atomic_read(&rdp->dynticks));
800 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* No store tearing. */
801 rdp->dynticks_nmi_nesting - 2);
802 instrumentation_end();
803 return;
804 }
805
806 /* This NMI interrupted an RCU-idle CPU, restore RCU-idleness. */
807 trace_rcu_dyntick(TPS("Startirq"), rdp->dynticks_nmi_nesting, 0, atomic_read(&rdp->dynticks));
808 WRITE_ONCE(rdp->dynticks_nmi_nesting, 0); /* Avoid store tearing. */
809
810 if (!in_nmi())
811 rcu_prepare_for_idle();
812
813 // instrumentation for the noinstr rcu_dynticks_eqs_enter()
814 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
815 instrumentation_end();
816
817 // RCU is watching here ...
818 rcu_dynticks_eqs_enter();
819 // ... but is no longer watching here.
820
821 if (!in_nmi())
822 rcu_dynticks_task_enter();
823}
824
825/**
826 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
827 *
828 * Exit from an interrupt handler, which might possibly result in entering
829 * idle mode, in other words, leaving the mode in which read-side critical
830 * sections can occur. The caller must have disabled interrupts.
831 *
832 * This code assumes that the idle loop never does anything that might
833 * result in unbalanced calls to irq_enter() and irq_exit(). If your
834 * architecture's idle loop violates this assumption, RCU will give you what
835 * you deserve, good and hard. But very infrequently and irreproducibly.
836 *
837 * Use things like work queues to work around this limitation.
838 *
839 * You have been warned.
840 *
841 * If you add or remove a call to rcu_irq_exit(), be sure to test with
842 * CONFIG_RCU_EQS_DEBUG=y.
843 */
844void noinstr rcu_irq_exit(void)
845{
846 lockdep_assert_irqs_disabled();
847 rcu_nmi_exit();
848}
849
850#ifdef CONFIG_PROVE_RCU
851/**
852 * rcu_irq_exit_check_preempt - Validate that scheduling is possible
853 */
854void rcu_irq_exit_check_preempt(void)
855{
856 lockdep_assert_irqs_disabled();
857
858 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) <= 0,
859 "RCU dynticks_nesting counter underflow/zero!");
860 RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) !=
861 DYNTICK_IRQ_NONIDLE,
862 "Bad RCU dynticks_nmi_nesting counter\n");
863 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
864 "RCU in extended quiescent state!");
865}
866#endif /* #ifdef CONFIG_PROVE_RCU */
867
868/*
869 * Wrapper for rcu_irq_exit() where interrupts are enabled.
870 *
871 * If you add or remove a call to rcu_irq_exit_irqson(), be sure to test
872 * with CONFIG_RCU_EQS_DEBUG=y.
873 */
874void rcu_irq_exit_irqson(void)
875{
876 unsigned long flags;
877
878 local_irq_save(flags);
879 rcu_irq_exit();
880 local_irq_restore(flags);
881}
882
883/*
884 * Exit an RCU extended quiescent state, which can be either the
885 * idle loop or adaptive-tickless usermode execution.
886 *
887 * We crowbar the ->dynticks_nmi_nesting field to DYNTICK_IRQ_NONIDLE to
888 * allow for the possibility of usermode upcalls messing up our count of
889 * interrupt nesting level during the busy period that is just now starting.
890 */
891static void noinstr rcu_eqs_exit(bool user)
892{
893 struct rcu_data *rdp;
894 long oldval;
895
896 lockdep_assert_irqs_disabled();
897 rdp = this_cpu_ptr(&rcu_data);
898 oldval = rdp->dynticks_nesting;
899 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && oldval < 0);
900 if (oldval) {
901 // RCU was already watching, so just do accounting and leave.
902 rdp->dynticks_nesting++;
903 return;
904 }
905 rcu_dynticks_task_exit();
906 // RCU is not watching here ...
907 rcu_dynticks_eqs_exit();
908 // ... but is watching here.
909 instrumentation_begin();
910
911 // instrumentation for the noinstr rcu_dynticks_eqs_exit()
912 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
913
914 rcu_cleanup_after_idle();
915 trace_rcu_dyntick(TPS("End"), rdp->dynticks_nesting, 1, atomic_read(&rdp->dynticks));
916 WARN_ON_ONCE(IS_ENABLED(CONFIG_RCU_EQS_DEBUG) && !user && !is_idle_task(current));
917 WRITE_ONCE(rdp->dynticks_nesting, 1);
918 WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
919 WRITE_ONCE(rdp->dynticks_nmi_nesting, DYNTICK_IRQ_NONIDLE);
920 instrumentation_end();
921}
922
923/**
924 * rcu_idle_exit - inform RCU that current CPU is leaving idle
925 *
926 * Exit idle mode, in other words, -enter- the mode in which RCU
927 * read-side critical sections can occur.
928 *
929 * If you add or remove a call to rcu_idle_exit(), be sure to test with
930 * CONFIG_RCU_EQS_DEBUG=y.
931 */
932void rcu_idle_exit(void)
933{
934 unsigned long flags;
935
936 local_irq_save(flags);
937 rcu_eqs_exit(false);
938 local_irq_restore(flags);
939}
940EXPORT_SYMBOL_GPL(rcu_idle_exit);
941
942#ifdef CONFIG_NO_HZ_FULL
943/**
944 * rcu_user_exit - inform RCU that we are exiting userspace.
945 *
946 * Exit RCU idle mode while entering the kernel because it can
947 * run a RCU read side critical section anytime.
948 *
949 * If you add or remove a call to rcu_user_exit(), be sure to test with
950 * CONFIG_RCU_EQS_DEBUG=y.
951 */
952void noinstr rcu_user_exit(void)
953{
954 rcu_eqs_exit(true);
955}
956
957/**
958 * __rcu_irq_enter_check_tick - Enable scheduler tick on CPU if RCU needs it.
959 *
960 * The scheduler tick is not normally enabled when CPUs enter the kernel
961 * from nohz_full userspace execution. After all, nohz_full userspace
962 * execution is an RCU quiescent state and the time executing in the kernel
963 * is quite short. Except of course when it isn't. And it is not hard to
964 * cause a large system to spend tens of seconds or even minutes looping
965 * in the kernel, which can cause a number of problems, include RCU CPU
966 * stall warnings.
967 *
968 * Therefore, if a nohz_full CPU fails to report a quiescent state
969 * in a timely manner, the RCU grace-period kthread sets that CPU's
970 * ->rcu_urgent_qs flag with the expectation that the next interrupt or
971 * exception will invoke this function, which will turn on the scheduler
972 * tick, which will enable RCU to detect that CPU's quiescent states,
973 * for example, due to cond_resched() calls in CONFIG_PREEMPT=n kernels.
974 * The tick will be disabled once a quiescent state is reported for
975 * this CPU.
976 *
977 * Of course, in carefully tuned systems, there might never be an
978 * interrupt or exception. In that case, the RCU grace-period kthread
979 * will eventually cause one to happen. However, in less carefully
980 * controlled environments, this function allows RCU to get what it
981 * needs without creating otherwise useless interruptions.
982 */
983void __rcu_irq_enter_check_tick(void)
984{
985 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
986
987 // If we're here from NMI there's nothing to do.
988 if (in_nmi())
989 return;
990
991 RCU_LOCKDEP_WARN(rcu_dynticks_curr_cpu_in_eqs(),
992 "Illegal rcu_irq_enter_check_tick() from extended quiescent state");
993
994 if (!tick_nohz_full_cpu(rdp->cpu) ||
995 !READ_ONCE(rdp->rcu_urgent_qs) ||
996 READ_ONCE(rdp->rcu_forced_tick)) {
997 // RCU doesn't need nohz_full help from this CPU, or it is
998 // already getting that help.
999 return;
1000 }
1001
1002 // We get here only when not in an extended quiescent state and
1003 // from interrupts (as opposed to NMIs). Therefore, (1) RCU is
1004 // already watching and (2) The fact that we are in an interrupt
1005 // handler and that the rcu_node lock is an irq-disabled lock
1006 // prevents self-deadlock. So we can safely recheck under the lock.
1007 // Note that the nohz_full state currently cannot change.
1008 raw_spin_lock_rcu_node(rdp->mynode);
1009 if (rdp->rcu_urgent_qs && !rdp->rcu_forced_tick) {
1010 // A nohz_full CPU is in the kernel and RCU needs a
1011 // quiescent state. Turn on the tick!
1012 WRITE_ONCE(rdp->rcu_forced_tick, true);
1013 tick_dep_set_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
1014 }
1015 raw_spin_unlock_rcu_node(rdp->mynode);
1016}
1017#endif /* CONFIG_NO_HZ_FULL */
1018
1019/**
1020 * rcu_nmi_enter - inform RCU of entry to NMI context
1021 *
1022 * If the CPU was idle from RCU's viewpoint, update rdp->dynticks and
1023 * rdp->dynticks_nmi_nesting to let the RCU grace-period handling know
1024 * that the CPU is active. This implementation permits nested NMIs, as
1025 * long as the nesting level does not overflow an int. (You will probably
1026 * run out of stack space first.)
1027 *
1028 * If you add or remove a call to rcu_nmi_enter(), be sure to test
1029 * with CONFIG_RCU_EQS_DEBUG=y.
1030 */
1031noinstr void rcu_nmi_enter(void)
1032{
1033 long incby = 2;
1034 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1035
1036 /* Complain about underflow. */
1037 WARN_ON_ONCE(rdp->dynticks_nmi_nesting < 0);
1038
1039 /*
1040 * If idle from RCU viewpoint, atomically increment ->dynticks
1041 * to mark non-idle and increment ->dynticks_nmi_nesting by one.
1042 * Otherwise, increment ->dynticks_nmi_nesting by two. This means
1043 * if ->dynticks_nmi_nesting is equal to one, we are guaranteed
1044 * to be in the outermost NMI handler that interrupted an RCU-idle
1045 * period (observation due to Andy Lutomirski).
1046 */
1047 if (rcu_dynticks_curr_cpu_in_eqs()) {
1048
1049 if (!in_nmi())
1050 rcu_dynticks_task_exit();
1051
1052 // RCU is not watching here ...
1053 rcu_dynticks_eqs_exit();
1054 // ... but is watching here.
1055
1056 if (!in_nmi()) {
1057 instrumentation_begin();
1058 rcu_cleanup_after_idle();
1059 instrumentation_end();
1060 }
1061
1062 instrumentation_begin();
1063 // instrumentation for the noinstr rcu_dynticks_curr_cpu_in_eqs()
1064 instrument_atomic_read(&rdp->dynticks, sizeof(rdp->dynticks));
1065 // instrumentation for the noinstr rcu_dynticks_eqs_exit()
1066 instrument_atomic_write(&rdp->dynticks, sizeof(rdp->dynticks));
1067
1068 incby = 1;
1069 } else if (!in_nmi()) {
1070 instrumentation_begin();
1071 rcu_irq_enter_check_tick();
1072 } else {
1073 instrumentation_begin();
1074 }
1075
1076 trace_rcu_dyntick(incby == 1 ? TPS("Endirq") : TPS("++="),
1077 rdp->dynticks_nmi_nesting,
1078 rdp->dynticks_nmi_nesting + incby, atomic_read(&rdp->dynticks));
1079 instrumentation_end();
1080 WRITE_ONCE(rdp->dynticks_nmi_nesting, /* Prevent store tearing. */
1081 rdp->dynticks_nmi_nesting + incby);
1082 barrier();
1083}
1084
1085/**
1086 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
1087 *
1088 * Enter an interrupt handler, which might possibly result in exiting
1089 * idle mode, in other words, entering the mode in which read-side critical
1090 * sections can occur. The caller must have disabled interrupts.
1091 *
1092 * Note that the Linux kernel is fully capable of entering an interrupt
1093 * handler that it never exits, for example when doing upcalls to user mode!
1094 * This code assumes that the idle loop never does upcalls to user mode.
1095 * If your architecture's idle loop does do upcalls to user mode (or does
1096 * anything else that results in unbalanced calls to the irq_enter() and
1097 * irq_exit() functions), RCU will give you what you deserve, good and hard.
1098 * But very infrequently and irreproducibly.
1099 *
1100 * Use things like work queues to work around this limitation.
1101 *
1102 * You have been warned.
1103 *
1104 * If you add or remove a call to rcu_irq_enter(), be sure to test with
1105 * CONFIG_RCU_EQS_DEBUG=y.
1106 */
1107noinstr void rcu_irq_enter(void)
1108{
1109 lockdep_assert_irqs_disabled();
1110 rcu_nmi_enter();
1111}
1112
1113/*
1114 * Wrapper for rcu_irq_enter() where interrupts are enabled.
1115 *
1116 * If you add or remove a call to rcu_irq_enter_irqson(), be sure to test
1117 * with CONFIG_RCU_EQS_DEBUG=y.
1118 */
1119void rcu_irq_enter_irqson(void)
1120{
1121 unsigned long flags;
1122
1123 local_irq_save(flags);
1124 rcu_irq_enter();
1125 local_irq_restore(flags);
1126}
1127
1128/*
1129 * If any sort of urgency was applied to the current CPU (for example,
1130 * the scheduler-clock interrupt was enabled on a nohz_full CPU) in order
1131 * to get to a quiescent state, disable it.
1132 */
1133static void rcu_disable_urgency_upon_qs(struct rcu_data *rdp)
1134{
1135 raw_lockdep_assert_held_rcu_node(rdp->mynode);
1136 WRITE_ONCE(rdp->rcu_urgent_qs, false);
1137 WRITE_ONCE(rdp->rcu_need_heavy_qs, false);
1138 if (tick_nohz_full_cpu(rdp->cpu) && rdp->rcu_forced_tick) {
1139 tick_dep_clear_cpu(rdp->cpu, TICK_DEP_BIT_RCU);
1140 WRITE_ONCE(rdp->rcu_forced_tick, false);
1141 }
1142}
1143
1144/**
1145 * rcu_is_watching - see if RCU thinks that the current CPU is not idle
1146 *
1147 * Return true if RCU is watching the running CPU, which means that this
1148 * CPU can safely enter RCU read-side critical sections. In other words,
1149 * if the current CPU is not in its idle loop or is in an interrupt or
1150 * NMI handler, return true.
1151 *
1152 * Make notrace because it can be called by the internal functions of
1153 * ftrace, and making this notrace removes unnecessary recursion calls.
1154 */
1155notrace bool rcu_is_watching(void)
1156{
1157 bool ret;
1158
1159 preempt_disable_notrace();
1160 ret = !rcu_dynticks_curr_cpu_in_eqs();
1161 preempt_enable_notrace();
1162 return ret;
1163}
1164EXPORT_SYMBOL_GPL(rcu_is_watching);
1165
1166/*
1167 * If a holdout task is actually running, request an urgent quiescent
1168 * state from its CPU. This is unsynchronized, so migrations can cause
1169 * the request to go to the wrong CPU. Which is OK, all that will happen
1170 * is that the CPU's next context switch will be a bit slower and next
1171 * time around this task will generate another request.
1172 */
1173void rcu_request_urgent_qs_task(struct task_struct *t)
1174{
1175 int cpu;
1176
1177 barrier();
1178 cpu = task_cpu(t);
1179 if (!task_curr(t))
1180 return; /* This task is not running on that CPU. */
1181 smp_store_release(per_cpu_ptr(&rcu_data.rcu_urgent_qs, cpu), true);
1182}
1183
1184#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
1185
1186/*
1187 * Is the current CPU online as far as RCU is concerned?
1188 *
1189 * Disable preemption to avoid false positives that could otherwise
1190 * happen due to the current CPU number being sampled, this task being
1191 * preempted, its old CPU being taken offline, resuming on some other CPU,
1192 * then determining that its old CPU is now offline.
1193 *
1194 * Disable checking if in an NMI handler because we cannot safely
1195 * report errors from NMI handlers anyway. In addition, it is OK to use
1196 * RCU on an offline processor during initial boot, hence the check for
1197 * rcu_scheduler_fully_active.
1198 */
1199bool rcu_lockdep_current_cpu_online(void)
1200{
1201 struct rcu_data *rdp;
1202 struct rcu_node *rnp;
1203 bool ret = false;
1204
1205 if (in_nmi() || !rcu_scheduler_fully_active)
1206 return true;
1207 preempt_disable_notrace();
1208 rdp = this_cpu_ptr(&rcu_data);
1209 rnp = rdp->mynode;
1210 if (rdp->grpmask & rcu_rnp_online_cpus(rnp) || READ_ONCE(rnp->ofl_seq) & 0x1)
1211 ret = true;
1212 preempt_enable_notrace();
1213 return ret;
1214}
1215EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
1216
1217#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
1218
1219/*
1220 * When trying to report a quiescent state on behalf of some other CPU,
1221 * it is our responsibility to check for and handle potential overflow
1222 * of the rcu_node ->gp_seq counter with respect to the rcu_data counters.
1223 * After all, the CPU might be in deep idle state, and thus executing no
1224 * code whatsoever.
1225 */
1226static void rcu_gpnum_ovf(struct rcu_node *rnp, struct rcu_data *rdp)
1227{
1228 raw_lockdep_assert_held_rcu_node(rnp);
1229 if (ULONG_CMP_LT(rcu_seq_current(&rdp->gp_seq) + ULONG_MAX / 4,
1230 rnp->gp_seq))
1231 WRITE_ONCE(rdp->gpwrap, true);
1232 if (ULONG_CMP_LT(rdp->rcu_iw_gp_seq + ULONG_MAX / 4, rnp->gp_seq))
1233 rdp->rcu_iw_gp_seq = rnp->gp_seq + ULONG_MAX / 4;
1234}
1235
1236/*
1237 * Snapshot the specified CPU's dynticks counter so that we can later
1238 * credit them with an implicit quiescent state. Return 1 if this CPU
1239 * is in dynticks idle mode, which is an extended quiescent state.
1240 */
1241static int dyntick_save_progress_counter(struct rcu_data *rdp)
1242{
1243 rdp->dynticks_snap = rcu_dynticks_snap(rdp);
1244 if (rcu_dynticks_in_eqs(rdp->dynticks_snap)) {
1245 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1246 rcu_gpnum_ovf(rdp->mynode, rdp);
1247 return 1;
1248 }
1249 return 0;
1250}
1251
1252/*
1253 * Return true if the specified CPU has passed through a quiescent
1254 * state by virtue of being in or having passed through an dynticks
1255 * idle state since the last call to dyntick_save_progress_counter()
1256 * for this same CPU, or by virtue of having been offline.
1257 */
1258static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
1259{
1260 unsigned long jtsq;
1261 bool *rnhqp;
1262 bool *ruqp;
1263 struct rcu_node *rnp = rdp->mynode;
1264
1265 /*
1266 * If the CPU passed through or entered a dynticks idle phase with
1267 * no active irq/NMI handlers, then we can safely pretend that the CPU
1268 * already acknowledged the request to pass through a quiescent
1269 * state. Either way, that CPU cannot possibly be in an RCU
1270 * read-side critical section that started before the beginning
1271 * of the current RCU grace period.
1272 */
1273 if (rcu_dynticks_in_eqs_since(rdp, rdp->dynticks_snap)) {
1274 trace_rcu_fqs(rcu_state.name, rdp->gp_seq, rdp->cpu, TPS("dti"));
1275 rcu_gpnum_ovf(rnp, rdp);
1276 return 1;
1277 }
1278
1279 /*
1280 * Complain if a CPU that is considered to be offline from RCU's
1281 * perspective has not yet reported a quiescent state. After all,
1282 * the offline CPU should have reported a quiescent state during
1283 * the CPU-offline process, or, failing that, by rcu_gp_init()
1284 * if it ran concurrently with either the CPU going offline or the
1285 * last task on a leaf rcu_node structure exiting its RCU read-side
1286 * critical section while all CPUs corresponding to that structure
1287 * are offline. This added warning detects bugs in any of these
1288 * code paths.
1289 *
1290 * The rcu_node structure's ->lock is held here, which excludes
1291 * the relevant portions the CPU-hotplug code, the grace-period
1292 * initialization code, and the rcu_read_unlock() code paths.
1293 *
1294 * For more detail, please refer to the "Hotplug CPU" section
1295 * of RCU's Requirements documentation.
1296 */
1297 if (WARN_ON_ONCE(!(rdp->grpmask & rcu_rnp_online_cpus(rnp)))) {
1298 bool onl;
1299 struct rcu_node *rnp1;
1300
1301 pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
1302 __func__, rnp->grplo, rnp->grphi, rnp->level,
1303 (long)rnp->gp_seq, (long)rnp->completedqs);
1304 for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
1305 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx ->rcu_gp_init_mask %#lx\n",
1306 __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext, rnp1->rcu_gp_init_mask);
1307 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
1308 pr_info("%s %d: %c online: %ld(%d) offline: %ld(%d)\n",
1309 __func__, rdp->cpu, ".o"[onl],
1310 (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
1311 (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
1312 return 1; /* Break things loose after complaining. */
1313 }
1314
1315 /*
1316 * A CPU running for an extended time within the kernel can
1317 * delay RCU grace periods: (1) At age jiffies_to_sched_qs,
1318 * set .rcu_urgent_qs, (2) At age 2*jiffies_to_sched_qs, set
1319 * both .rcu_need_heavy_qs and .rcu_urgent_qs. Note that the
1320 * unsynchronized assignments to the per-CPU rcu_need_heavy_qs
1321 * variable are safe because the assignments are repeated if this
1322 * CPU failed to pass through a quiescent state. This code
1323 * also checks .jiffies_resched in case jiffies_to_sched_qs
1324 * is set way high.
1325 */
1326 jtsq = READ_ONCE(jiffies_to_sched_qs);
1327 ruqp = per_cpu_ptr(&rcu_data.rcu_urgent_qs, rdp->cpu);
1328 rnhqp = &per_cpu(rcu_data.rcu_need_heavy_qs, rdp->cpu);
1329 if (!READ_ONCE(*rnhqp) &&
1330 (time_after(jiffies, rcu_state.gp_start + jtsq * 2) ||
1331 time_after(jiffies, rcu_state.jiffies_resched) ||
1332 rcu_state.cbovld)) {
1333 WRITE_ONCE(*rnhqp, true);
1334 /* Store rcu_need_heavy_qs before rcu_urgent_qs. */
1335 smp_store_release(ruqp, true);
1336 } else if (time_after(jiffies, rcu_state.gp_start + jtsq)) {
1337 WRITE_ONCE(*ruqp, true);
1338 }
1339
1340 /*
1341 * NO_HZ_FULL CPUs can run in-kernel without rcu_sched_clock_irq!
1342 * The above code handles this, but only for straight cond_resched().
1343 * And some in-kernel loops check need_resched() before calling
1344 * cond_resched(), which defeats the above code for CPUs that are
1345 * running in-kernel with scheduling-clock interrupts disabled.
1346 * So hit them over the head with the resched_cpu() hammer!
1347 */
1348 if (tick_nohz_full_cpu(rdp->cpu) &&
1349 (time_after(jiffies, READ_ONCE(rdp->last_fqs_resched) + jtsq * 3) ||
1350 rcu_state.cbovld)) {
1351 WRITE_ONCE(*ruqp, true);
1352 resched_cpu(rdp->cpu);
1353 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1354 }
1355
1356 /*
1357 * If more than halfway to RCU CPU stall-warning time, invoke
1358 * resched_cpu() more frequently to try to loosen things up a bit.
1359 * Also check to see if the CPU is getting hammered with interrupts,
1360 * but only once per grace period, just to keep the IPIs down to
1361 * a dull roar.
1362 */
1363 if (time_after(jiffies, rcu_state.jiffies_resched)) {
1364 if (time_after(jiffies,
1365 READ_ONCE(rdp->last_fqs_resched) + jtsq)) {
1366 resched_cpu(rdp->cpu);
1367 WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1368 }
1369 if (IS_ENABLED(CONFIG_IRQ_WORK) &&
1370 !rdp->rcu_iw_pending && rdp->rcu_iw_gp_seq != rnp->gp_seq &&
1371 (rnp->ffmask & rdp->grpmask)) {
1372 rdp->rcu_iw_pending = true;
1373 rdp->rcu_iw_gp_seq = rnp->gp_seq;
1374 irq_work_queue_on(&rdp->rcu_iw, rdp->cpu);
1375 }
1376 }
1377
1378 return 0;
1379}
1380
1381/* Trace-event wrapper function for trace_rcu_future_grace_period. */
1382static void trace_rcu_this_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1383 unsigned long gp_seq_req, const char *s)
1384{
1385 trace_rcu_future_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
1386 gp_seq_req, rnp->level,
1387 rnp->grplo, rnp->grphi, s);
1388}
1389
1390/*
1391 * rcu_start_this_gp - Request the start of a particular grace period
1392 * @rnp_start: The leaf node of the CPU from which to start.
1393 * @rdp: The rcu_data corresponding to the CPU from which to start.
1394 * @gp_seq_req: The gp_seq of the grace period to start.
1395 *
1396 * Start the specified grace period, as needed to handle newly arrived
1397 * callbacks. The required future grace periods are recorded in each
1398 * rcu_node structure's ->gp_seq_needed field. Returns true if there
1399 * is reason to awaken the grace-period kthread.
1400 *
1401 * The caller must hold the specified rcu_node structure's ->lock, which
1402 * is why the caller is responsible for waking the grace-period kthread.
1403 *
1404 * Returns true if the GP thread needs to be awakened else false.
1405 */
1406static bool rcu_start_this_gp(struct rcu_node *rnp_start, struct rcu_data *rdp,
1407 unsigned long gp_seq_req)
1408{
1409 bool ret = false;
1410 struct rcu_node *rnp;
1411
1412 /*
1413 * Use funnel locking to either acquire the root rcu_node
1414 * structure's lock or bail out if the need for this grace period
1415 * has already been recorded -- or if that grace period has in
1416 * fact already started. If there is already a grace period in
1417 * progress in a non-leaf node, no recording is needed because the
1418 * end of the grace period will scan the leaf rcu_node structures.
1419 * Note that rnp_start->lock must not be released.
1420 */
1421 raw_lockdep_assert_held_rcu_node(rnp_start);
1422 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req, TPS("Startleaf"));
1423 for (rnp = rnp_start; 1; rnp = rnp->parent) {
1424 if (rnp != rnp_start)
1425 raw_spin_lock_rcu_node(rnp);
1426 if (ULONG_CMP_GE(rnp->gp_seq_needed, gp_seq_req) ||
1427 rcu_seq_started(&rnp->gp_seq, gp_seq_req) ||
1428 (rnp != rnp_start &&
1429 rcu_seq_state(rcu_seq_current(&rnp->gp_seq)))) {
1430 trace_rcu_this_gp(rnp, rdp, gp_seq_req,
1431 TPS("Prestarted"));
1432 goto unlock_out;
1433 }
1434 WRITE_ONCE(rnp->gp_seq_needed, gp_seq_req);
1435 if (rcu_seq_state(rcu_seq_current(&rnp->gp_seq))) {
1436 /*
1437 * We just marked the leaf or internal node, and a
1438 * grace period is in progress, which means that
1439 * rcu_gp_cleanup() will see the marking. Bail to
1440 * reduce contention.
1441 */
1442 trace_rcu_this_gp(rnp_start, rdp, gp_seq_req,
1443 TPS("Startedleaf"));
1444 goto unlock_out;
1445 }
1446 if (rnp != rnp_start && rnp->parent != NULL)
1447 raw_spin_unlock_rcu_node(rnp);
1448 if (!rnp->parent)
1449 break; /* At root, and perhaps also leaf. */
1450 }
1451
1452 /* If GP already in progress, just leave, otherwise start one. */
1453 if (rcu_gp_in_progress()) {
1454 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedleafroot"));
1455 goto unlock_out;
1456 }
1457 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("Startedroot"));
1458 WRITE_ONCE(rcu_state.gp_flags, rcu_state.gp_flags | RCU_GP_FLAG_INIT);
1459 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
1460 if (!READ_ONCE(rcu_state.gp_kthread)) {
1461 trace_rcu_this_gp(rnp, rdp, gp_seq_req, TPS("NoGPkthread"));
1462 goto unlock_out;
1463 }
1464 trace_rcu_grace_period(rcu_state.name, data_race(rcu_state.gp_seq), TPS("newreq"));
1465 ret = true; /* Caller must wake GP kthread. */
1466unlock_out:
1467 /* Push furthest requested GP to leaf node and rcu_data structure. */
1468 if (ULONG_CMP_LT(gp_seq_req, rnp->gp_seq_needed)) {
1469 WRITE_ONCE(rnp_start->gp_seq_needed, rnp->gp_seq_needed);
1470 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1471 }
1472 if (rnp != rnp_start)
1473 raw_spin_unlock_rcu_node(rnp);
1474 return ret;
1475}
1476
1477/*
1478 * Clean up any old requests for the just-ended grace period. Also return
1479 * whether any additional grace periods have been requested.
1480 */
1481static bool rcu_future_gp_cleanup(struct rcu_node *rnp)
1482{
1483 bool needmore;
1484 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1485
1486 needmore = ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed);
1487 if (!needmore)
1488 rnp->gp_seq_needed = rnp->gp_seq; /* Avoid counter wrap. */
1489 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq,
1490 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1491 return needmore;
1492}
1493
1494/*
1495 * Awaken the grace-period kthread. Don't do a self-awaken (unless in an
1496 * interrupt or softirq handler, in which case we just might immediately
1497 * sleep upon return, resulting in a grace-period hang), and don't bother
1498 * awakening when there is nothing for the grace-period kthread to do
1499 * (as in several CPUs raced to awaken, we lost), and finally don't try
1500 * to awaken a kthread that has not yet been created. If all those checks
1501 * are passed, track some debug information and awaken.
1502 *
1503 * So why do the self-wakeup when in an interrupt or softirq handler
1504 * in the grace-period kthread's context? Because the kthread might have
1505 * been interrupted just as it was going to sleep, and just after the final
1506 * pre-sleep check of the awaken condition. In this case, a wakeup really
1507 * is required, and is therefore supplied.
1508 */
1509static void rcu_gp_kthread_wake(void)
1510{
1511 struct task_struct *t = READ_ONCE(rcu_state.gp_kthread);
1512
1513 if ((current == t && !in_irq() && !in_serving_softirq()) ||
1514 !READ_ONCE(rcu_state.gp_flags) || !t)
1515 return;
1516 WRITE_ONCE(rcu_state.gp_wake_time, jiffies);
1517 WRITE_ONCE(rcu_state.gp_wake_seq, READ_ONCE(rcu_state.gp_seq));
1518 swake_up_one(&rcu_state.gp_wq);
1519}
1520
1521/*
1522 * If there is room, assign a ->gp_seq number to any callbacks on this
1523 * CPU that have not already been assigned. Also accelerate any callbacks
1524 * that were previously assigned a ->gp_seq number that has since proven
1525 * to be too conservative, which can happen if callbacks get assigned a
1526 * ->gp_seq number while RCU is idle, but with reference to a non-root
1527 * rcu_node structure. This function is idempotent, so it does not hurt
1528 * to call it repeatedly. Returns an flag saying that we should awaken
1529 * the RCU grace-period kthread.
1530 *
1531 * The caller must hold rnp->lock with interrupts disabled.
1532 */
1533static bool rcu_accelerate_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1534{
1535 unsigned long gp_seq_req;
1536 bool ret = false;
1537
1538 rcu_lockdep_assert_cblist_protected(rdp);
1539 raw_lockdep_assert_held_rcu_node(rnp);
1540
1541 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1542 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1543 return false;
1544
1545 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPreAcc"));
1546
1547 /*
1548 * Callbacks are often registered with incomplete grace-period
1549 * information. Something about the fact that getting exact
1550 * information requires acquiring a global lock... RCU therefore
1551 * makes a conservative estimate of the grace period number at which
1552 * a given callback will become ready to invoke. The following
1553 * code checks this estimate and improves it when possible, thus
1554 * accelerating callback invocation to an earlier grace-period
1555 * number.
1556 */
1557 gp_seq_req = rcu_seq_snap(&rcu_state.gp_seq);
1558 if (rcu_segcblist_accelerate(&rdp->cblist, gp_seq_req))
1559 ret = rcu_start_this_gp(rnp, rdp, gp_seq_req);
1560
1561 /* Trace depending on how much we were able to accelerate. */
1562 if (rcu_segcblist_restempty(&rdp->cblist, RCU_WAIT_TAIL))
1563 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccWaitCB"));
1564 else
1565 trace_rcu_grace_period(rcu_state.name, gp_seq_req, TPS("AccReadyCB"));
1566
1567 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbPostAcc"));
1568
1569 return ret;
1570}
1571
1572/*
1573 * Similar to rcu_accelerate_cbs(), but does not require that the leaf
1574 * rcu_node structure's ->lock be held. It consults the cached value
1575 * of ->gp_seq_needed in the rcu_data structure, and if that indicates
1576 * that a new grace-period request be made, invokes rcu_accelerate_cbs()
1577 * while holding the leaf rcu_node structure's ->lock.
1578 */
1579static void rcu_accelerate_cbs_unlocked(struct rcu_node *rnp,
1580 struct rcu_data *rdp)
1581{
1582 unsigned long c;
1583 bool needwake;
1584
1585 rcu_lockdep_assert_cblist_protected(rdp);
1586 c = rcu_seq_snap(&rcu_state.gp_seq);
1587 if (!READ_ONCE(rdp->gpwrap) && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
1588 /* Old request still live, so mark recent callbacks. */
1589 (void)rcu_segcblist_accelerate(&rdp->cblist, c);
1590 return;
1591 }
1592 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1593 needwake = rcu_accelerate_cbs(rnp, rdp);
1594 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1595 if (needwake)
1596 rcu_gp_kthread_wake();
1597}
1598
1599/*
1600 * Move any callbacks whose grace period has completed to the
1601 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1602 * assign ->gp_seq numbers to any callbacks in the RCU_NEXT_TAIL
1603 * sublist. This function is idempotent, so it does not hurt to
1604 * invoke it repeatedly. As long as it is not invoked -too- often...
1605 * Returns true if the RCU grace-period kthread needs to be awakened.
1606 *
1607 * The caller must hold rnp->lock with interrupts disabled.
1608 */
1609static bool rcu_advance_cbs(struct rcu_node *rnp, struct rcu_data *rdp)
1610{
1611 rcu_lockdep_assert_cblist_protected(rdp);
1612 raw_lockdep_assert_held_rcu_node(rnp);
1613
1614 /* If no pending (not yet ready to invoke) callbacks, nothing to do. */
1615 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1616 return false;
1617
1618 /*
1619 * Find all callbacks whose ->gp_seq numbers indicate that they
1620 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1621 */
1622 rcu_segcblist_advance(&rdp->cblist, rnp->gp_seq);
1623
1624 /* Classify any remaining callbacks. */
1625 return rcu_accelerate_cbs(rnp, rdp);
1626}
1627
1628/*
1629 * Move and classify callbacks, but only if doing so won't require
1630 * that the RCU grace-period kthread be awakened.
1631 */
1632static void __maybe_unused rcu_advance_cbs_nowake(struct rcu_node *rnp,
1633 struct rcu_data *rdp)
1634{
1635 rcu_lockdep_assert_cblist_protected(rdp);
1636 if (!rcu_seq_state(rcu_seq_current(&rnp->gp_seq)) ||
1637 !raw_spin_trylock_rcu_node(rnp))
1638 return;
1639 WARN_ON_ONCE(rcu_advance_cbs(rnp, rdp));
1640 raw_spin_unlock_rcu_node(rnp);
1641}
1642
1643/*
1644 * In CONFIG_RCU_STRICT_GRACE_PERIOD=y kernels, attempt to generate a
1645 * quiescent state. This is intended to be invoked when the CPU notices
1646 * a new grace period.
1647 */
1648static void rcu_strict_gp_check_qs(void)
1649{
1650 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD)) {
1651 rcu_read_lock();
1652 rcu_read_unlock();
1653 }
1654}
1655
1656/*
1657 * Update CPU-local rcu_data state to record the beginnings and ends of
1658 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1659 * structure corresponding to the current CPU, and must have irqs disabled.
1660 * Returns true if the grace-period kthread needs to be awakened.
1661 */
1662static bool __note_gp_changes(struct rcu_node *rnp, struct rcu_data *rdp)
1663{
1664 bool ret = false;
1665 bool need_qs;
1666 const bool offloaded = rcu_rdp_is_offloaded(rdp);
1667
1668 raw_lockdep_assert_held_rcu_node(rnp);
1669
1670 if (rdp->gp_seq == rnp->gp_seq)
1671 return false; /* Nothing to do. */
1672
1673 /* Handle the ends of any preceding grace periods first. */
1674 if (rcu_seq_completed_gp(rdp->gp_seq, rnp->gp_seq) ||
1675 unlikely(READ_ONCE(rdp->gpwrap))) {
1676 if (!offloaded)
1677 ret = rcu_advance_cbs(rnp, rdp); /* Advance CBs. */
1678 rdp->core_needs_qs = false;
1679 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuend"));
1680 } else {
1681 if (!offloaded)
1682 ret = rcu_accelerate_cbs(rnp, rdp); /* Recent CBs. */
1683 if (rdp->core_needs_qs)
1684 rdp->core_needs_qs = !!(rnp->qsmask & rdp->grpmask);
1685 }
1686
1687 /* Now handle the beginnings of any new-to-this-CPU grace periods. */
1688 if (rcu_seq_new_gp(rdp->gp_seq, rnp->gp_seq) ||
1689 unlikely(READ_ONCE(rdp->gpwrap))) {
1690 /*
1691 * If the current grace period is waiting for this CPU,
1692 * set up to detect a quiescent state, otherwise don't
1693 * go looking for one.
1694 */
1695 trace_rcu_grace_period(rcu_state.name, rnp->gp_seq, TPS("cpustart"));
1696 need_qs = !!(rnp->qsmask & rdp->grpmask);
1697 rdp->cpu_no_qs.b.norm = need_qs;
1698 rdp->core_needs_qs = need_qs;
1699 zero_cpu_stall_ticks(rdp);
1700 }
1701 rdp->gp_seq = rnp->gp_seq; /* Remember new grace-period state. */
1702 if (ULONG_CMP_LT(rdp->gp_seq_needed, rnp->gp_seq_needed) || rdp->gpwrap)
1703 WRITE_ONCE(rdp->gp_seq_needed, rnp->gp_seq_needed);
1704 WRITE_ONCE(rdp->gpwrap, false);
1705 rcu_gpnum_ovf(rnp, rdp);
1706 return ret;
1707}
1708
1709static void note_gp_changes(struct rcu_data *rdp)
1710{
1711 unsigned long flags;
1712 bool needwake;
1713 struct rcu_node *rnp;
1714
1715 local_irq_save(flags);
1716 rnp = rdp->mynode;
1717 if ((rdp->gp_seq == rcu_seq_current(&rnp->gp_seq) &&
1718 !unlikely(READ_ONCE(rdp->gpwrap))) || /* w/out lock. */
1719 !raw_spin_trylock_rcu_node(rnp)) { /* irqs already off, so later. */
1720 local_irq_restore(flags);
1721 return;
1722 }
1723 needwake = __note_gp_changes(rnp, rdp);
1724 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1725 rcu_strict_gp_check_qs();
1726 if (needwake)
1727 rcu_gp_kthread_wake();
1728}
1729
1730static void rcu_gp_slow(int delay)
1731{
1732 if (delay > 0 &&
1733 !(rcu_seq_ctr(rcu_state.gp_seq) %
1734 (rcu_num_nodes * PER_RCU_NODE_PERIOD * delay)))
1735 schedule_timeout_idle(delay);
1736}
1737
1738static unsigned long sleep_duration;
1739
1740/* Allow rcutorture to stall the grace-period kthread. */
1741void rcu_gp_set_torture_wait(int duration)
1742{
1743 if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST) && duration > 0)
1744 WRITE_ONCE(sleep_duration, duration);
1745}
1746EXPORT_SYMBOL_GPL(rcu_gp_set_torture_wait);
1747
1748/* Actually implement the aforementioned wait. */
1749static void rcu_gp_torture_wait(void)
1750{
1751 unsigned long duration;
1752
1753 if (!IS_ENABLED(CONFIG_RCU_TORTURE_TEST))
1754 return;
1755 duration = xchg(&sleep_duration, 0UL);
1756 if (duration > 0) {
1757 pr_alert("%s: Waiting %lu jiffies\n", __func__, duration);
1758 schedule_timeout_idle(duration);
1759 pr_alert("%s: Wait complete\n", __func__);
1760 }
1761}
1762
1763/*
1764 * Handler for on_each_cpu() to invoke the target CPU's RCU core
1765 * processing.
1766 */
1767static void rcu_strict_gp_boundary(void *unused)
1768{
1769 invoke_rcu_core();
1770}
1771
1772/*
1773 * Initialize a new grace period. Return false if no grace period required.
1774 */
1775static bool rcu_gp_init(void)
1776{
1777 unsigned long firstseq;
1778 unsigned long flags;
1779 unsigned long oldmask;
1780 unsigned long mask;
1781 struct rcu_data *rdp;
1782 struct rcu_node *rnp = rcu_get_root();
1783
1784 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1785 raw_spin_lock_irq_rcu_node(rnp);
1786 if (!READ_ONCE(rcu_state.gp_flags)) {
1787 /* Spurious wakeup, tell caller to go back to sleep. */
1788 raw_spin_unlock_irq_rcu_node(rnp);
1789 return false;
1790 }
1791 WRITE_ONCE(rcu_state.gp_flags, 0); /* Clear all flags: New GP. */
1792
1793 if (WARN_ON_ONCE(rcu_gp_in_progress())) {
1794 /*
1795 * Grace period already in progress, don't start another.
1796 * Not supposed to be able to happen.
1797 */
1798 raw_spin_unlock_irq_rcu_node(rnp);
1799 return false;
1800 }
1801
1802 /* Advance to a new grace period and initialize state. */
1803 record_gp_stall_check_time();
1804 /* Record GP times before starting GP, hence rcu_seq_start(). */
1805 rcu_seq_start(&rcu_state.gp_seq);
1806 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
1807 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("start"));
1808 raw_spin_unlock_irq_rcu_node(rnp);
1809
1810 /*
1811 * Apply per-leaf buffered online and offline operations to
1812 * the rcu_node tree. Note that this new grace period need not
1813 * wait for subsequent online CPUs, and that RCU hooks in the CPU
1814 * offlining path, when combined with checks in this function,
1815 * will handle CPUs that are currently going offline or that will
1816 * go offline later. Please also refer to "Hotplug CPU" section
1817 * of RCU's Requirements documentation.
1818 */
1819 WRITE_ONCE(rcu_state.gp_state, RCU_GP_ONOFF);
1820 rcu_for_each_leaf_node(rnp) {
1821 smp_mb(); // Pair with barriers used when updating ->ofl_seq to odd values.
1822 firstseq = READ_ONCE(rnp->ofl_seq);
1823 if (firstseq & 0x1)
1824 while (firstseq == READ_ONCE(rnp->ofl_seq))
1825 schedule_timeout_idle(1); // Can't wake unless RCU is watching.
1826 smp_mb(); // Pair with barriers used when updating ->ofl_seq to even values.
1827 raw_spin_lock(&rcu_state.ofl_lock);
1828 raw_spin_lock_irq_rcu_node(rnp);
1829 if (rnp->qsmaskinit == rnp->qsmaskinitnext &&
1830 !rnp->wait_blkd_tasks) {
1831 /* Nothing to do on this leaf rcu_node structure. */
1832 raw_spin_unlock_irq_rcu_node(rnp);
1833 raw_spin_unlock(&rcu_state.ofl_lock);
1834 continue;
1835 }
1836
1837 /* Record old state, apply changes to ->qsmaskinit field. */
1838 oldmask = rnp->qsmaskinit;
1839 rnp->qsmaskinit = rnp->qsmaskinitnext;
1840
1841 /* If zero-ness of ->qsmaskinit changed, propagate up tree. */
1842 if (!oldmask != !rnp->qsmaskinit) {
1843 if (!oldmask) { /* First online CPU for rcu_node. */
1844 if (!rnp->wait_blkd_tasks) /* Ever offline? */
1845 rcu_init_new_rnp(rnp);
1846 } else if (rcu_preempt_has_tasks(rnp)) {
1847 rnp->wait_blkd_tasks = true; /* blocked tasks */
1848 } else { /* Last offline CPU and can propagate. */
1849 rcu_cleanup_dead_rnp(rnp);
1850 }
1851 }
1852
1853 /*
1854 * If all waited-on tasks from prior grace period are
1855 * done, and if all this rcu_node structure's CPUs are
1856 * still offline, propagate up the rcu_node tree and
1857 * clear ->wait_blkd_tasks. Otherwise, if one of this
1858 * rcu_node structure's CPUs has since come back online,
1859 * simply clear ->wait_blkd_tasks.
1860 */
1861 if (rnp->wait_blkd_tasks &&
1862 (!rcu_preempt_has_tasks(rnp) || rnp->qsmaskinit)) {
1863 rnp->wait_blkd_tasks = false;
1864 if (!rnp->qsmaskinit)
1865 rcu_cleanup_dead_rnp(rnp);
1866 }
1867
1868 raw_spin_unlock_irq_rcu_node(rnp);
1869 raw_spin_unlock(&rcu_state.ofl_lock);
1870 }
1871 rcu_gp_slow(gp_preinit_delay); /* Races with CPU hotplug. */
1872
1873 /*
1874 * Set the quiescent-state-needed bits in all the rcu_node
1875 * structures for all currently online CPUs in breadth-first
1876 * order, starting from the root rcu_node structure, relying on the
1877 * layout of the tree within the rcu_state.node[] array. Note that
1878 * other CPUs will access only the leaves of the hierarchy, thus
1879 * seeing that no grace period is in progress, at least until the
1880 * corresponding leaf node has been initialized.
1881 *
1882 * The grace period cannot complete until the initialization
1883 * process finishes, because this kthread handles both.
1884 */
1885 WRITE_ONCE(rcu_state.gp_state, RCU_GP_INIT);
1886 rcu_for_each_node_breadth_first(rnp) {
1887 rcu_gp_slow(gp_init_delay);
1888 raw_spin_lock_irqsave_rcu_node(rnp, flags);
1889 rdp = this_cpu_ptr(&rcu_data);
1890 rcu_preempt_check_blocked_tasks(rnp);
1891 rnp->qsmask = rnp->qsmaskinit;
1892 WRITE_ONCE(rnp->gp_seq, rcu_state.gp_seq);
1893 if (rnp == rdp->mynode)
1894 (void)__note_gp_changes(rnp, rdp);
1895 rcu_preempt_boost_start_gp(rnp);
1896 trace_rcu_grace_period_init(rcu_state.name, rnp->gp_seq,
1897 rnp->level, rnp->grplo,
1898 rnp->grphi, rnp->qsmask);
1899 /* Quiescent states for tasks on any now-offline CPUs. */
1900 mask = rnp->qsmask & ~rnp->qsmaskinitnext;
1901 rnp->rcu_gp_init_mask = mask;
1902 if ((mask || rnp->wait_blkd_tasks) && rcu_is_leaf_node(rnp))
1903 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
1904 else
1905 raw_spin_unlock_irq_rcu_node(rnp);
1906 cond_resched_tasks_rcu_qs();
1907 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1908 }
1909
1910 // If strict, make all CPUs aware of new grace period.
1911 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
1912 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
1913
1914 return true;
1915}
1916
1917/*
1918 * Helper function for swait_event_idle_exclusive() wakeup at force-quiescent-state
1919 * time.
1920 */
1921static bool rcu_gp_fqs_check_wake(int *gfp)
1922{
1923 struct rcu_node *rnp = rcu_get_root();
1924
1925 // If under overload conditions, force an immediate FQS scan.
1926 if (*gfp & RCU_GP_FLAG_OVLD)
1927 return true;
1928
1929 // Someone like call_rcu() requested a force-quiescent-state scan.
1930 *gfp = READ_ONCE(rcu_state.gp_flags);
1931 if (*gfp & RCU_GP_FLAG_FQS)
1932 return true;
1933
1934 // The current grace period has completed.
1935 if (!READ_ONCE(rnp->qsmask) && !rcu_preempt_blocked_readers_cgp(rnp))
1936 return true;
1937
1938 return false;
1939}
1940
1941/*
1942 * Do one round of quiescent-state forcing.
1943 */
1944static void rcu_gp_fqs(bool first_time)
1945{
1946 struct rcu_node *rnp = rcu_get_root();
1947
1948 WRITE_ONCE(rcu_state.gp_activity, jiffies);
1949 rcu_state.n_force_qs++;
1950 if (first_time) {
1951 /* Collect dyntick-idle snapshots. */
1952 force_qs_rnp(dyntick_save_progress_counter);
1953 } else {
1954 /* Handle dyntick-idle and offline CPUs. */
1955 force_qs_rnp(rcu_implicit_dynticks_qs);
1956 }
1957 /* Clear flag to prevent immediate re-entry. */
1958 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
1959 raw_spin_lock_irq_rcu_node(rnp);
1960 WRITE_ONCE(rcu_state.gp_flags,
1961 READ_ONCE(rcu_state.gp_flags) & ~RCU_GP_FLAG_FQS);
1962 raw_spin_unlock_irq_rcu_node(rnp);
1963 }
1964}
1965
1966/*
1967 * Loop doing repeated quiescent-state forcing until the grace period ends.
1968 */
1969static void rcu_gp_fqs_loop(void)
1970{
1971 bool first_gp_fqs;
1972 int gf = 0;
1973 unsigned long j;
1974 int ret;
1975 struct rcu_node *rnp = rcu_get_root();
1976
1977 first_gp_fqs = true;
1978 j = READ_ONCE(jiffies_till_first_fqs);
1979 if (rcu_state.cbovld)
1980 gf = RCU_GP_FLAG_OVLD;
1981 ret = 0;
1982 for (;;) {
1983 if (!ret) {
1984 WRITE_ONCE(rcu_state.jiffies_force_qs, jiffies + j);
1985 /*
1986 * jiffies_force_qs before RCU_GP_WAIT_FQS state
1987 * update; required for stall checks.
1988 */
1989 smp_wmb();
1990 WRITE_ONCE(rcu_state.jiffies_kick_kthreads,
1991 jiffies + (j ? 3 * j : 2));
1992 }
1993 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
1994 TPS("fqswait"));
1995 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_FQS);
1996 ret = swait_event_idle_timeout_exclusive(
1997 rcu_state.gp_wq, rcu_gp_fqs_check_wake(&gf), j);
1998 rcu_gp_torture_wait();
1999 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DOING_FQS);
2000 /* Locking provides needed memory barriers. */
2001 /* If grace period done, leave loop. */
2002 if (!READ_ONCE(rnp->qsmask) &&
2003 !rcu_preempt_blocked_readers_cgp(rnp))
2004 break;
2005 /* If time for quiescent-state forcing, do it. */
2006 if (!time_after(rcu_state.jiffies_force_qs, jiffies) ||
2007 (gf & (RCU_GP_FLAG_FQS | RCU_GP_FLAG_OVLD))) {
2008 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2009 TPS("fqsstart"));
2010 rcu_gp_fqs(first_gp_fqs);
2011 gf = 0;
2012 if (first_gp_fqs) {
2013 first_gp_fqs = false;
2014 gf = rcu_state.cbovld ? RCU_GP_FLAG_OVLD : 0;
2015 }
2016 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2017 TPS("fqsend"));
2018 cond_resched_tasks_rcu_qs();
2019 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2020 ret = 0; /* Force full wait till next FQS. */
2021 j = READ_ONCE(jiffies_till_next_fqs);
2022 } else {
2023 /* Deal with stray signal. */
2024 cond_resched_tasks_rcu_qs();
2025 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2026 WARN_ON(signal_pending(current));
2027 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2028 TPS("fqswaitsig"));
2029 ret = 1; /* Keep old FQS timing. */
2030 j = jiffies;
2031 if (time_after(jiffies, rcu_state.jiffies_force_qs))
2032 j = 1;
2033 else
2034 j = rcu_state.jiffies_force_qs - j;
2035 gf = 0;
2036 }
2037 }
2038}
2039
2040/*
2041 * Clean up after the old grace period.
2042 */
2043static noinline void rcu_gp_cleanup(void)
2044{
2045 int cpu;
2046 bool needgp = false;
2047 unsigned long gp_duration;
2048 unsigned long new_gp_seq;
2049 bool offloaded;
2050 struct rcu_data *rdp;
2051 struct rcu_node *rnp = rcu_get_root();
2052 struct swait_queue_head *sq;
2053
2054 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2055 raw_spin_lock_irq_rcu_node(rnp);
2056 rcu_state.gp_end = jiffies;
2057 gp_duration = rcu_state.gp_end - rcu_state.gp_start;
2058 if (gp_duration > rcu_state.gp_max)
2059 rcu_state.gp_max = gp_duration;
2060
2061 /*
2062 * We know the grace period is complete, but to everyone else
2063 * it appears to still be ongoing. But it is also the case
2064 * that to everyone else it looks like there is nothing that
2065 * they can do to advance the grace period. It is therefore
2066 * safe for us to drop the lock in order to mark the grace
2067 * period as completed in all of the rcu_node structures.
2068 */
2069 raw_spin_unlock_irq_rcu_node(rnp);
2070
2071 /*
2072 * Propagate new ->gp_seq value to rcu_node structures so that
2073 * other CPUs don't have to wait until the start of the next grace
2074 * period to process their callbacks. This also avoids some nasty
2075 * RCU grace-period initialization races by forcing the end of
2076 * the current grace period to be completely recorded in all of
2077 * the rcu_node structures before the beginning of the next grace
2078 * period is recorded in any of the rcu_node structures.
2079 */
2080 new_gp_seq = rcu_state.gp_seq;
2081 rcu_seq_end(&new_gp_seq);
2082 rcu_for_each_node_breadth_first(rnp) {
2083 raw_spin_lock_irq_rcu_node(rnp);
2084 if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
2085 dump_blkd_tasks(rnp, 10);
2086 WARN_ON_ONCE(rnp->qsmask);
2087 WRITE_ONCE(rnp->gp_seq, new_gp_seq);
2088 rdp = this_cpu_ptr(&rcu_data);
2089 if (rnp == rdp->mynode)
2090 needgp = __note_gp_changes(rnp, rdp) || needgp;
2091 /* smp_mb() provided by prior unlock-lock pair. */
2092 needgp = rcu_future_gp_cleanup(rnp) || needgp;
2093 // Reset overload indication for CPUs no longer overloaded
2094 if (rcu_is_leaf_node(rnp))
2095 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->cbovldmask) {
2096 rdp = per_cpu_ptr(&rcu_data, cpu);
2097 check_cb_ovld_locked(rdp, rnp);
2098 }
2099 sq = rcu_nocb_gp_get(rnp);
2100 raw_spin_unlock_irq_rcu_node(rnp);
2101 rcu_nocb_gp_cleanup(sq);
2102 cond_resched_tasks_rcu_qs();
2103 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2104 rcu_gp_slow(gp_cleanup_delay);
2105 }
2106 rnp = rcu_get_root();
2107 raw_spin_lock_irq_rcu_node(rnp); /* GP before ->gp_seq update. */
2108
2109 /* Declare grace period done, trace first to use old GP number. */
2110 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq, TPS("end"));
2111 rcu_seq_end(&rcu_state.gp_seq);
2112 ASSERT_EXCLUSIVE_WRITER(rcu_state.gp_seq);
2113 WRITE_ONCE(rcu_state.gp_state, RCU_GP_IDLE);
2114 /* Check for GP requests since above loop. */
2115 rdp = this_cpu_ptr(&rcu_data);
2116 if (!needgp && ULONG_CMP_LT(rnp->gp_seq, rnp->gp_seq_needed)) {
2117 trace_rcu_this_gp(rnp, rdp, rnp->gp_seq_needed,
2118 TPS("CleanupMore"));
2119 needgp = true;
2120 }
2121 /* Advance CBs to reduce false positives below. */
2122 offloaded = rcu_rdp_is_offloaded(rdp);
2123 if ((offloaded || !rcu_accelerate_cbs(rnp, rdp)) && needgp) {
2124 WRITE_ONCE(rcu_state.gp_flags, RCU_GP_FLAG_INIT);
2125 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
2126 trace_rcu_grace_period(rcu_state.name,
2127 rcu_state.gp_seq,
2128 TPS("newreq"));
2129 } else {
2130 WRITE_ONCE(rcu_state.gp_flags,
2131 rcu_state.gp_flags & RCU_GP_FLAG_INIT);
2132 }
2133 raw_spin_unlock_irq_rcu_node(rnp);
2134
2135 // If strict, make all CPUs aware of the end of the old grace period.
2136 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2137 on_each_cpu(rcu_strict_gp_boundary, NULL, 0);
2138}
2139
2140/*
2141 * Body of kthread that handles grace periods.
2142 */
2143static int __noreturn rcu_gp_kthread(void *unused)
2144{
2145 rcu_bind_gp_kthread();
2146 for (;;) {
2147
2148 /* Handle grace-period start. */
2149 for (;;) {
2150 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2151 TPS("reqwait"));
2152 WRITE_ONCE(rcu_state.gp_state, RCU_GP_WAIT_GPS);
2153 swait_event_idle_exclusive(rcu_state.gp_wq,
2154 READ_ONCE(rcu_state.gp_flags) &
2155 RCU_GP_FLAG_INIT);
2156 rcu_gp_torture_wait();
2157 WRITE_ONCE(rcu_state.gp_state, RCU_GP_DONE_GPS);
2158 /* Locking provides needed memory barrier. */
2159 if (rcu_gp_init())
2160 break;
2161 cond_resched_tasks_rcu_qs();
2162 WRITE_ONCE(rcu_state.gp_activity, jiffies);
2163 WARN_ON(signal_pending(current));
2164 trace_rcu_grace_period(rcu_state.name, rcu_state.gp_seq,
2165 TPS("reqwaitsig"));
2166 }
2167
2168 /* Handle quiescent-state forcing. */
2169 rcu_gp_fqs_loop();
2170
2171 /* Handle grace-period end. */
2172 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANUP);
2173 rcu_gp_cleanup();
2174 WRITE_ONCE(rcu_state.gp_state, RCU_GP_CLEANED);
2175 }
2176}
2177
2178/*
2179 * Report a full set of quiescent states to the rcu_state data structure.
2180 * Invoke rcu_gp_kthread_wake() to awaken the grace-period kthread if
2181 * another grace period is required. Whether we wake the grace-period
2182 * kthread or it awakens itself for the next round of quiescent-state
2183 * forcing, that kthread will clean up after the just-completed grace
2184 * period. Note that the caller must hold rnp->lock, which is released
2185 * before return.
2186 */
2187static void rcu_report_qs_rsp(unsigned long flags)
2188 __releases(rcu_get_root()->lock)
2189{
2190 raw_lockdep_assert_held_rcu_node(rcu_get_root());
2191 WARN_ON_ONCE(!rcu_gp_in_progress());
2192 WRITE_ONCE(rcu_state.gp_flags,
2193 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2194 raw_spin_unlock_irqrestore_rcu_node(rcu_get_root(), flags);
2195 rcu_gp_kthread_wake();
2196}
2197
2198/*
2199 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
2200 * Allows quiescent states for a group of CPUs to be reported at one go
2201 * to the specified rcu_node structure, though all the CPUs in the group
2202 * must be represented by the same rcu_node structure (which need not be a
2203 * leaf rcu_node structure, though it often will be). The gps parameter
2204 * is the grace-period snapshot, which means that the quiescent states
2205 * are valid only if rnp->gp_seq is equal to gps. That structure's lock
2206 * must be held upon entry, and it is released before return.
2207 *
2208 * As a special case, if mask is zero, the bit-already-cleared check is
2209 * disabled. This allows propagating quiescent state due to resumed tasks
2210 * during grace-period initialization.
2211 */
2212static void rcu_report_qs_rnp(unsigned long mask, struct rcu_node *rnp,
2213 unsigned long gps, unsigned long flags)
2214 __releases(rnp->lock)
2215{
2216 unsigned long oldmask = 0;
2217 struct rcu_node *rnp_c;
2218
2219 raw_lockdep_assert_held_rcu_node(rnp);
2220
2221 /* Walk up the rcu_node hierarchy. */
2222 for (;;) {
2223 if ((!(rnp->qsmask & mask) && mask) || rnp->gp_seq != gps) {
2224
2225 /*
2226 * Our bit has already been cleared, or the
2227 * relevant grace period is already over, so done.
2228 */
2229 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2230 return;
2231 }
2232 WARN_ON_ONCE(oldmask); /* Any child must be all zeroed! */
2233 WARN_ON_ONCE(!rcu_is_leaf_node(rnp) &&
2234 rcu_preempt_blocked_readers_cgp(rnp));
2235 WRITE_ONCE(rnp->qsmask, rnp->qsmask & ~mask);
2236 trace_rcu_quiescent_state_report(rcu_state.name, rnp->gp_seq,
2237 mask, rnp->qsmask, rnp->level,
2238 rnp->grplo, rnp->grphi,
2239 !!rnp->gp_tasks);
2240 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
2241
2242 /* Other bits still set at this level, so done. */
2243 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2244 return;
2245 }
2246 rnp->completedqs = rnp->gp_seq;
2247 mask = rnp->grpmask;
2248 if (rnp->parent == NULL) {
2249
2250 /* No more levels. Exit loop holding root lock. */
2251
2252 break;
2253 }
2254 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2255 rnp_c = rnp;
2256 rnp = rnp->parent;
2257 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2258 oldmask = READ_ONCE(rnp_c->qsmask);
2259 }
2260
2261 /*
2262 * Get here if we are the last CPU to pass through a quiescent
2263 * state for this grace period. Invoke rcu_report_qs_rsp()
2264 * to clean up and start the next grace period if one is needed.
2265 */
2266 rcu_report_qs_rsp(flags); /* releases rnp->lock. */
2267}
2268
2269/*
2270 * Record a quiescent state for all tasks that were previously queued
2271 * on the specified rcu_node structure and that were blocking the current
2272 * RCU grace period. The caller must hold the corresponding rnp->lock with
2273 * irqs disabled, and this lock is released upon return, but irqs remain
2274 * disabled.
2275 */
2276static void __maybe_unused
2277rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
2278 __releases(rnp->lock)
2279{
2280 unsigned long gps;
2281 unsigned long mask;
2282 struct rcu_node *rnp_p;
2283
2284 raw_lockdep_assert_held_rcu_node(rnp);
2285 if (WARN_ON_ONCE(!IS_ENABLED(CONFIG_PREEMPT_RCU)) ||
2286 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)) ||
2287 rnp->qsmask != 0) {
2288 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2289 return; /* Still need more quiescent states! */
2290 }
2291
2292 rnp->completedqs = rnp->gp_seq;
2293 rnp_p = rnp->parent;
2294 if (rnp_p == NULL) {
2295 /*
2296 * Only one rcu_node structure in the tree, so don't
2297 * try to report up to its nonexistent parent!
2298 */
2299 rcu_report_qs_rsp(flags);
2300 return;
2301 }
2302
2303 /* Report up the rest of the hierarchy, tracking current ->gp_seq. */
2304 gps = rnp->gp_seq;
2305 mask = rnp->grpmask;
2306 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2307 raw_spin_lock_rcu_node(rnp_p); /* irqs already disabled. */
2308 rcu_report_qs_rnp(mask, rnp_p, gps, flags);
2309}
2310
2311/*
2312 * Record a quiescent state for the specified CPU to that CPU's rcu_data
2313 * structure. This must be called from the specified CPU.
2314 */
2315static void
2316rcu_report_qs_rdp(struct rcu_data *rdp)
2317{
2318 unsigned long flags;
2319 unsigned long mask;
2320 bool needwake = false;
2321 const bool offloaded = rcu_rdp_is_offloaded(rdp);
2322 struct rcu_node *rnp;
2323
2324 WARN_ON_ONCE(rdp->cpu != smp_processor_id());
2325 rnp = rdp->mynode;
2326 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2327 if (rdp->cpu_no_qs.b.norm || rdp->gp_seq != rnp->gp_seq ||
2328 rdp->gpwrap) {
2329
2330 /*
2331 * The grace period in which this quiescent state was
2332 * recorded has ended, so don't report it upwards.
2333 * We will instead need a new quiescent state that lies
2334 * within the current grace period.
2335 */
2336 rdp->cpu_no_qs.b.norm = true; /* need qs for new gp. */
2337 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2338 return;
2339 }
2340 mask = rdp->grpmask;
2341 rdp->core_needs_qs = false;
2342 if ((rnp->qsmask & mask) == 0) {
2343 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2344 } else {
2345 /*
2346 * This GP can't end until cpu checks in, so all of our
2347 * callbacks can be processed during the next GP.
2348 */
2349 if (!offloaded)
2350 needwake = rcu_accelerate_cbs(rnp, rdp);
2351
2352 rcu_disable_urgency_upon_qs(rdp);
2353 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2354 /* ^^^ Released rnp->lock */
2355 if (needwake)
2356 rcu_gp_kthread_wake();
2357 }
2358}
2359
2360/*
2361 * Check to see if there is a new grace period of which this CPU
2362 * is not yet aware, and if so, set up local rcu_data state for it.
2363 * Otherwise, see if this CPU has just passed through its first
2364 * quiescent state for this grace period, and record that fact if so.
2365 */
2366static void
2367rcu_check_quiescent_state(struct rcu_data *rdp)
2368{
2369 /* Check for grace-period ends and beginnings. */
2370 note_gp_changes(rdp);
2371
2372 /*
2373 * Does this CPU still need to do its part for current grace period?
2374 * If no, return and let the other CPUs do their part as well.
2375 */
2376 if (!rdp->core_needs_qs)
2377 return;
2378
2379 /*
2380 * Was there a quiescent state since the beginning of the grace
2381 * period? If no, then exit and wait for the next call.
2382 */
2383 if (rdp->cpu_no_qs.b.norm)
2384 return;
2385
2386 /*
2387 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
2388 * judge of that).
2389 */
2390 rcu_report_qs_rdp(rdp);
2391}
2392
2393/*
2394 * Near the end of the offline process. Trace the fact that this CPU
2395 * is going offline.
2396 */
2397int rcutree_dying_cpu(unsigned int cpu)
2398{
2399 bool blkd;
2400 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
2401 struct rcu_node *rnp = rdp->mynode;
2402
2403 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2404 return 0;
2405
2406 blkd = !!(rnp->qsmask & rdp->grpmask);
2407 trace_rcu_grace_period(rcu_state.name, READ_ONCE(rnp->gp_seq),
2408 blkd ? TPS("cpuofl-bgp") : TPS("cpuofl"));
2409 return 0;
2410}
2411
2412/*
2413 * All CPUs for the specified rcu_node structure have gone offline,
2414 * and all tasks that were preempted within an RCU read-side critical
2415 * section while running on one of those CPUs have since exited their RCU
2416 * read-side critical section. Some other CPU is reporting this fact with
2417 * the specified rcu_node structure's ->lock held and interrupts disabled.
2418 * This function therefore goes up the tree of rcu_node structures,
2419 * clearing the corresponding bits in the ->qsmaskinit fields. Note that
2420 * the leaf rcu_node structure's ->qsmaskinit field has already been
2421 * updated.
2422 *
2423 * This function does check that the specified rcu_node structure has
2424 * all CPUs offline and no blocked tasks, so it is OK to invoke it
2425 * prematurely. That said, invoking it after the fact will cost you
2426 * a needless lock acquisition. So once it has done its work, don't
2427 * invoke it again.
2428 */
2429static void rcu_cleanup_dead_rnp(struct rcu_node *rnp_leaf)
2430{
2431 long mask;
2432 struct rcu_node *rnp = rnp_leaf;
2433
2434 raw_lockdep_assert_held_rcu_node(rnp_leaf);
2435 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU) ||
2436 WARN_ON_ONCE(rnp_leaf->qsmaskinit) ||
2437 WARN_ON_ONCE(rcu_preempt_has_tasks(rnp_leaf)))
2438 return;
2439 for (;;) {
2440 mask = rnp->grpmask;
2441 rnp = rnp->parent;
2442 if (!rnp)
2443 break;
2444 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2445 rnp->qsmaskinit &= ~mask;
2446 /* Between grace periods, so better already be zero! */
2447 WARN_ON_ONCE(rnp->qsmask);
2448 if (rnp->qsmaskinit) {
2449 raw_spin_unlock_rcu_node(rnp);
2450 /* irqs remain disabled. */
2451 return;
2452 }
2453 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
2454 }
2455}
2456
2457/*
2458 * The CPU has been completely removed, and some other CPU is reporting
2459 * this fact from process context. Do the remainder of the cleanup.
2460 * There can only be one CPU hotplug operation at a time, so no need for
2461 * explicit locking.
2462 */
2463int rcutree_dead_cpu(unsigned int cpu)
2464{
2465 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
2466 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
2467
2468 if (!IS_ENABLED(CONFIG_HOTPLUG_CPU))
2469 return 0;
2470
2471 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus - 1);
2472 /* Adjust any no-longer-needed kthreads. */
2473 rcu_boost_kthread_setaffinity(rnp, -1);
2474 /* Do any needed no-CB deferred wakeups from this CPU. */
2475 do_nocb_deferred_wakeup(per_cpu_ptr(&rcu_data, cpu));
2476
2477 // Stop-machine done, so allow nohz_full to disable tick.
2478 tick_dep_clear(TICK_DEP_BIT_RCU);
2479 return 0;
2480}
2481
2482/*
2483 * Invoke any RCU callbacks that have made it to the end of their grace
2484 * period. Throttle as specified by rdp->blimit.
2485 */
2486static void rcu_do_batch(struct rcu_data *rdp)
2487{
2488 int div;
2489 bool __maybe_unused empty;
2490 unsigned long flags;
2491 const bool offloaded = rcu_rdp_is_offloaded(rdp);
2492 struct rcu_head *rhp;
2493 struct rcu_cblist rcl = RCU_CBLIST_INITIALIZER(rcl);
2494 long bl, count = 0;
2495 long pending, tlimit = 0;
2496
2497 /* If no callbacks are ready, just return. */
2498 if (!rcu_segcblist_ready_cbs(&rdp->cblist)) {
2499 trace_rcu_batch_start(rcu_state.name,
2500 rcu_segcblist_n_cbs(&rdp->cblist), 0);
2501 trace_rcu_batch_end(rcu_state.name, 0,
2502 !rcu_segcblist_empty(&rdp->cblist),
2503 need_resched(), is_idle_task(current),
2504 rcu_is_callbacks_kthread());
2505 return;
2506 }
2507
2508 /*
2509 * Extract the list of ready callbacks, disabling to prevent
2510 * races with call_rcu() from interrupt handlers. Leave the
2511 * callback counts, as rcu_barrier() needs to be conservative.
2512 */
2513 local_irq_save(flags);
2514 rcu_nocb_lock(rdp);
2515 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2516 pending = rcu_segcblist_n_cbs(&rdp->cblist);
2517 div = READ_ONCE(rcu_divisor);
2518 div = div < 0 ? 7 : div > sizeof(long) * 8 - 2 ? sizeof(long) * 8 - 2 : div;
2519 bl = max(rdp->blimit, pending >> div);
2520 if (unlikely(bl > 100)) {
2521 long rrn = READ_ONCE(rcu_resched_ns);
2522
2523 rrn = rrn < NSEC_PER_MSEC ? NSEC_PER_MSEC : rrn > NSEC_PER_SEC ? NSEC_PER_SEC : rrn;
2524 tlimit = local_clock() + rrn;
2525 }
2526 trace_rcu_batch_start(rcu_state.name,
2527 rcu_segcblist_n_cbs(&rdp->cblist), bl);
2528 rcu_segcblist_extract_done_cbs(&rdp->cblist, &rcl);
2529 if (offloaded)
2530 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2531
2532 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCbDequeued"));
2533 rcu_nocb_unlock_irqrestore(rdp, flags);
2534
2535 /* Invoke callbacks. */
2536 tick_dep_set_task(current, TICK_DEP_BIT_RCU);
2537 rhp = rcu_cblist_dequeue(&rcl);
2538
2539 for (; rhp; rhp = rcu_cblist_dequeue(&rcl)) {
2540 rcu_callback_t f;
2541
2542 count++;
2543 debug_rcu_head_unqueue(rhp);
2544
2545 rcu_lock_acquire(&rcu_callback_map);
2546 trace_rcu_invoke_callback(rcu_state.name, rhp);
2547
2548 f = rhp->func;
2549 WRITE_ONCE(rhp->func, (rcu_callback_t)0L);
2550 f(rhp);
2551
2552 rcu_lock_release(&rcu_callback_map);
2553
2554 /*
2555 * Stop only if limit reached and CPU has something to do.
2556 */
2557 if (count >= bl && !offloaded &&
2558 (need_resched() ||
2559 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2560 break;
2561 if (unlikely(tlimit)) {
2562 /* only call local_clock() every 32 callbacks */
2563 if (likely((count & 31) || local_clock() < tlimit))
2564 continue;
2565 /* Exceeded the time limit, so leave. */
2566 break;
2567 }
2568 if (!in_serving_softirq()) {
2569 local_bh_enable();
2570 lockdep_assert_irqs_enabled();
2571 cond_resched_tasks_rcu_qs();
2572 lockdep_assert_irqs_enabled();
2573 local_bh_disable();
2574 }
2575 }
2576
2577 local_irq_save(flags);
2578 rcu_nocb_lock(rdp);
2579 rdp->n_cbs_invoked += count;
2580 trace_rcu_batch_end(rcu_state.name, count, !!rcl.head, need_resched(),
2581 is_idle_task(current), rcu_is_callbacks_kthread());
2582
2583 /* Update counts and requeue any remaining callbacks. */
2584 rcu_segcblist_insert_done_cbs(&rdp->cblist, &rcl);
2585 rcu_segcblist_add_len(&rdp->cblist, -count);
2586
2587 /* Reinstate batch limit if we have worked down the excess. */
2588 count = rcu_segcblist_n_cbs(&rdp->cblist);
2589 if (rdp->blimit >= DEFAULT_MAX_RCU_BLIMIT && count <= qlowmark)
2590 rdp->blimit = blimit;
2591
2592 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2593 if (count == 0 && rdp->qlen_last_fqs_check != 0) {
2594 rdp->qlen_last_fqs_check = 0;
2595 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2596 } else if (count < rdp->qlen_last_fqs_check - qhimark)
2597 rdp->qlen_last_fqs_check = count;
2598
2599 /*
2600 * The following usually indicates a double call_rcu(). To track
2601 * this down, try building with CONFIG_DEBUG_OBJECTS_RCU_HEAD=y.
2602 */
2603 empty = rcu_segcblist_empty(&rdp->cblist);
2604 WARN_ON_ONCE(count == 0 && !empty);
2605 WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) &&
2606 count != 0 && empty);
2607 WARN_ON_ONCE(count == 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) != 0);
2608 WARN_ON_ONCE(!empty && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
2609
2610 rcu_nocb_unlock_irqrestore(rdp, flags);
2611
2612 /* Re-invoke RCU core processing if there are callbacks remaining. */
2613 if (!offloaded && rcu_segcblist_ready_cbs(&rdp->cblist))
2614 invoke_rcu_core();
2615 tick_dep_clear_task(current, TICK_DEP_BIT_RCU);
2616}
2617
2618/*
2619 * This function is invoked from each scheduling-clock interrupt,
2620 * and checks to see if this CPU is in a non-context-switch quiescent
2621 * state, for example, user mode or idle loop. It also schedules RCU
2622 * core processing. If the current grace period has gone on too long,
2623 * it will ask the scheduler to manufacture a context switch for the sole
2624 * purpose of providing the needed quiescent state.
2625 */
2626void rcu_sched_clock_irq(int user)
2627{
2628 trace_rcu_utilization(TPS("Start scheduler-tick"));
2629 lockdep_assert_irqs_disabled();
2630 raw_cpu_inc(rcu_data.ticks_this_gp);
2631 /* The load-acquire pairs with the store-release setting to true. */
2632 if (smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
2633 /* Idle and userspace execution already are quiescent states. */
2634 if (!rcu_is_cpu_rrupt_from_idle() && !user) {
2635 set_tsk_need_resched(current);
2636 set_preempt_need_resched();
2637 }
2638 __this_cpu_write(rcu_data.rcu_urgent_qs, false);
2639 }
2640 rcu_flavor_sched_clock_irq(user);
2641 if (rcu_pending(user))
2642 invoke_rcu_core();
2643 lockdep_assert_irqs_disabled();
2644
2645 trace_rcu_utilization(TPS("End scheduler-tick"));
2646}
2647
2648/*
2649 * Scan the leaf rcu_node structures. For each structure on which all
2650 * CPUs have reported a quiescent state and on which there are tasks
2651 * blocking the current grace period, initiate RCU priority boosting.
2652 * Otherwise, invoke the specified function to check dyntick state for
2653 * each CPU that has not yet reported a quiescent state.
2654 */
2655static void force_qs_rnp(int (*f)(struct rcu_data *rdp))
2656{
2657 int cpu;
2658 unsigned long flags;
2659 unsigned long mask;
2660 struct rcu_data *rdp;
2661 struct rcu_node *rnp;
2662
2663 rcu_state.cbovld = rcu_state.cbovldnext;
2664 rcu_state.cbovldnext = false;
2665 rcu_for_each_leaf_node(rnp) {
2666 cond_resched_tasks_rcu_qs();
2667 mask = 0;
2668 raw_spin_lock_irqsave_rcu_node(rnp, flags);
2669 rcu_state.cbovldnext |= !!rnp->cbovldmask;
2670 if (rnp->qsmask == 0) {
2671 if (rcu_preempt_blocked_readers_cgp(rnp)) {
2672 /*
2673 * No point in scanning bits because they
2674 * are all zero. But we might need to
2675 * priority-boost blocked readers.
2676 */
2677 rcu_initiate_boost(rnp, flags);
2678 /* rcu_initiate_boost() releases rnp->lock */
2679 continue;
2680 }
2681 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2682 continue;
2683 }
2684 for_each_leaf_node_cpu_mask(rnp, cpu, rnp->qsmask) {
2685 rdp = per_cpu_ptr(&rcu_data, cpu);
2686 if (f(rdp)) {
2687 mask |= rdp->grpmask;
2688 rcu_disable_urgency_upon_qs(rdp);
2689 }
2690 }
2691 if (mask != 0) {
2692 /* Idle/offline CPUs, report (releases rnp->lock). */
2693 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
2694 } else {
2695 /* Nothing to do here, so just drop the lock. */
2696 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2697 }
2698 }
2699}
2700
2701/*
2702 * Force quiescent states on reluctant CPUs, and also detect which
2703 * CPUs are in dyntick-idle mode.
2704 */
2705void rcu_force_quiescent_state(void)
2706{
2707 unsigned long flags;
2708 bool ret;
2709 struct rcu_node *rnp;
2710 struct rcu_node *rnp_old = NULL;
2711
2712 /* Funnel through hierarchy to reduce memory contention. */
2713 rnp = __this_cpu_read(rcu_data.mynode);
2714 for (; rnp != NULL; rnp = rnp->parent) {
2715 ret = (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) ||
2716 !raw_spin_trylock(&rnp->fqslock);
2717 if (rnp_old != NULL)
2718 raw_spin_unlock(&rnp_old->fqslock);
2719 if (ret)
2720 return;
2721 rnp_old = rnp;
2722 }
2723 /* rnp_old == rcu_get_root(), rnp == NULL. */
2724
2725 /* Reached the root of the rcu_node tree, acquire lock. */
2726 raw_spin_lock_irqsave_rcu_node(rnp_old, flags);
2727 raw_spin_unlock(&rnp_old->fqslock);
2728 if (READ_ONCE(rcu_state.gp_flags) & RCU_GP_FLAG_FQS) {
2729 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2730 return; /* Someone beat us to it. */
2731 }
2732 WRITE_ONCE(rcu_state.gp_flags,
2733 READ_ONCE(rcu_state.gp_flags) | RCU_GP_FLAG_FQS);
2734 raw_spin_unlock_irqrestore_rcu_node(rnp_old, flags);
2735 rcu_gp_kthread_wake();
2736}
2737EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
2738
2739// Workqueue handler for an RCU reader for kernels enforcing struct RCU
2740// grace periods.
2741static void strict_work_handler(struct work_struct *work)
2742{
2743 rcu_read_lock();
2744 rcu_read_unlock();
2745}
2746
2747/* Perform RCU core processing work for the current CPU. */
2748static __latent_entropy void rcu_core(void)
2749{
2750 unsigned long flags;
2751 struct rcu_data *rdp = raw_cpu_ptr(&rcu_data);
2752 struct rcu_node *rnp = rdp->mynode;
2753 const bool do_batch = !rcu_segcblist_completely_offloaded(&rdp->cblist);
2754
2755 if (cpu_is_offline(smp_processor_id()))
2756 return;
2757 trace_rcu_utilization(TPS("Start RCU core"));
2758 WARN_ON_ONCE(!rdp->beenonline);
2759
2760 /* Report any deferred quiescent states if preemption enabled. */
2761 if (!(preempt_count() & PREEMPT_MASK)) {
2762 rcu_preempt_deferred_qs(current);
2763 } else if (rcu_preempt_need_deferred_qs(current)) {
2764 set_tsk_need_resched(current);
2765 set_preempt_need_resched();
2766 }
2767
2768 /* Update RCU state based on any recent quiescent states. */
2769 rcu_check_quiescent_state(rdp);
2770
2771 /* No grace period and unregistered callbacks? */
2772 if (!rcu_gp_in_progress() &&
2773 rcu_segcblist_is_enabled(&rdp->cblist) && do_batch) {
2774 rcu_nocb_lock_irqsave(rdp, flags);
2775 if (!rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
2776 rcu_accelerate_cbs_unlocked(rnp, rdp);
2777 rcu_nocb_unlock_irqrestore(rdp, flags);
2778 }
2779
2780 rcu_check_gp_start_stall(rnp, rdp, rcu_jiffies_till_stall_check());
2781
2782 /* If there are callbacks ready, invoke them. */
2783 if (do_batch && rcu_segcblist_ready_cbs(&rdp->cblist) &&
2784 likely(READ_ONCE(rcu_scheduler_fully_active)))
2785 rcu_do_batch(rdp);
2786
2787 /* Do any needed deferred wakeups of rcuo kthreads. */
2788 do_nocb_deferred_wakeup(rdp);
2789 trace_rcu_utilization(TPS("End RCU core"));
2790
2791 // If strict GPs, schedule an RCU reader in a clean environment.
2792 if (IS_ENABLED(CONFIG_RCU_STRICT_GRACE_PERIOD))
2793 queue_work_on(rdp->cpu, rcu_gp_wq, &rdp->strict_work);
2794}
2795
2796static void rcu_core_si(struct softirq_action *h)
2797{
2798 rcu_core();
2799}
2800
2801static void rcu_wake_cond(struct task_struct *t, int status)
2802{
2803 /*
2804 * If the thread is yielding, only wake it when this
2805 * is invoked from idle
2806 */
2807 if (t && (status != RCU_KTHREAD_YIELDING || is_idle_task(current)))
2808 wake_up_process(t);
2809}
2810
2811static void invoke_rcu_core_kthread(void)
2812{
2813 struct task_struct *t;
2814 unsigned long flags;
2815
2816 local_irq_save(flags);
2817 __this_cpu_write(rcu_data.rcu_cpu_has_work, 1);
2818 t = __this_cpu_read(rcu_data.rcu_cpu_kthread_task);
2819 if (t != NULL && t != current)
2820 rcu_wake_cond(t, __this_cpu_read(rcu_data.rcu_cpu_kthread_status));
2821 local_irq_restore(flags);
2822}
2823
2824/*
2825 * Wake up this CPU's rcuc kthread to do RCU core processing.
2826 */
2827static void invoke_rcu_core(void)
2828{
2829 if (!cpu_online(smp_processor_id()))
2830 return;
2831 if (use_softirq)
2832 raise_softirq(RCU_SOFTIRQ);
2833 else
2834 invoke_rcu_core_kthread();
2835}
2836
2837static void rcu_cpu_kthread_park(unsigned int cpu)
2838{
2839 per_cpu(rcu_data.rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
2840}
2841
2842static int rcu_cpu_kthread_should_run(unsigned int cpu)
2843{
2844 return __this_cpu_read(rcu_data.rcu_cpu_has_work);
2845}
2846
2847/*
2848 * Per-CPU kernel thread that invokes RCU callbacks. This replaces
2849 * the RCU softirq used in configurations of RCU that do not support RCU
2850 * priority boosting.
2851 */
2852static void rcu_cpu_kthread(unsigned int cpu)
2853{
2854 unsigned int *statusp = this_cpu_ptr(&rcu_data.rcu_cpu_kthread_status);
2855 char work, *workp = this_cpu_ptr(&rcu_data.rcu_cpu_has_work);
2856 int spincnt;
2857
2858 trace_rcu_utilization(TPS("Start CPU kthread@rcu_run"));
2859 for (spincnt = 0; spincnt < 10; spincnt++) {
2860 local_bh_disable();
2861 *statusp = RCU_KTHREAD_RUNNING;
2862 local_irq_disable();
2863 work = *workp;
2864 *workp = 0;
2865 local_irq_enable();
2866 if (work)
2867 rcu_core();
2868 local_bh_enable();
2869 if (*workp == 0) {
2870 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
2871 *statusp = RCU_KTHREAD_WAITING;
2872 return;
2873 }
2874 }
2875 *statusp = RCU_KTHREAD_YIELDING;
2876 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
2877 schedule_timeout_idle(2);
2878 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
2879 *statusp = RCU_KTHREAD_WAITING;
2880}
2881
2882static struct smp_hotplug_thread rcu_cpu_thread_spec = {
2883 .store = &rcu_data.rcu_cpu_kthread_task,
2884 .thread_should_run = rcu_cpu_kthread_should_run,
2885 .thread_fn = rcu_cpu_kthread,
2886 .thread_comm = "rcuc/%u",
2887 .setup = rcu_cpu_kthread_setup,
2888 .park = rcu_cpu_kthread_park,
2889};
2890
2891/*
2892 * Spawn per-CPU RCU core processing kthreads.
2893 */
2894static int __init rcu_spawn_core_kthreads(void)
2895{
2896 int cpu;
2897
2898 for_each_possible_cpu(cpu)
2899 per_cpu(rcu_data.rcu_cpu_has_work, cpu) = 0;
2900 if (!IS_ENABLED(CONFIG_RCU_BOOST) && use_softirq)
2901 return 0;
2902 WARN_ONCE(smpboot_register_percpu_thread(&rcu_cpu_thread_spec),
2903 "%s: Could not start rcuc kthread, OOM is now expected behavior\n", __func__);
2904 return 0;
2905}
2906
2907/*
2908 * Handle any core-RCU processing required by a call_rcu() invocation.
2909 */
2910static void __call_rcu_core(struct rcu_data *rdp, struct rcu_head *head,
2911 unsigned long flags)
2912{
2913 /*
2914 * If called from an extended quiescent state, invoke the RCU
2915 * core in order to force a re-evaluation of RCU's idleness.
2916 */
2917 if (!rcu_is_watching())
2918 invoke_rcu_core();
2919
2920 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2921 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2922 return;
2923
2924 /*
2925 * Force the grace period if too many callbacks or too long waiting.
2926 * Enforce hysteresis, and don't invoke rcu_force_quiescent_state()
2927 * if some other CPU has recently done so. Also, don't bother
2928 * invoking rcu_force_quiescent_state() if the newly enqueued callback
2929 * is the only one waiting for a grace period to complete.
2930 */
2931 if (unlikely(rcu_segcblist_n_cbs(&rdp->cblist) >
2932 rdp->qlen_last_fqs_check + qhimark)) {
2933
2934 /* Are we ignoring a completed grace period? */
2935 note_gp_changes(rdp);
2936
2937 /* Start a new grace period if one not already started. */
2938 if (!rcu_gp_in_progress()) {
2939 rcu_accelerate_cbs_unlocked(rdp->mynode, rdp);
2940 } else {
2941 /* Give the grace period a kick. */
2942 rdp->blimit = DEFAULT_MAX_RCU_BLIMIT;
2943 if (rcu_state.n_force_qs == rdp->n_force_qs_snap &&
2944 rcu_segcblist_first_pend_cb(&rdp->cblist) != head)
2945 rcu_force_quiescent_state();
2946 rdp->n_force_qs_snap = rcu_state.n_force_qs;
2947 rdp->qlen_last_fqs_check = rcu_segcblist_n_cbs(&rdp->cblist);
2948 }
2949 }
2950}
2951
2952/*
2953 * RCU callback function to leak a callback.
2954 */
2955static void rcu_leak_callback(struct rcu_head *rhp)
2956{
2957}
2958
2959/*
2960 * Check and if necessary update the leaf rcu_node structure's
2961 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2962 * number of queued RCU callbacks. The caller must hold the leaf rcu_node
2963 * structure's ->lock.
2964 */
2965static void check_cb_ovld_locked(struct rcu_data *rdp, struct rcu_node *rnp)
2966{
2967 raw_lockdep_assert_held_rcu_node(rnp);
2968 if (qovld_calc <= 0)
2969 return; // Early boot and wildcard value set.
2970 if (rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc)
2971 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask | rdp->grpmask);
2972 else
2973 WRITE_ONCE(rnp->cbovldmask, rnp->cbovldmask & ~rdp->grpmask);
2974}
2975
2976/*
2977 * Check and if necessary update the leaf rcu_node structure's
2978 * ->cbovldmask bit corresponding to the current CPU based on that CPU's
2979 * number of queued RCU callbacks. No locks need be held, but the
2980 * caller must have disabled interrupts.
2981 *
2982 * Note that this function ignores the possibility that there are a lot
2983 * of callbacks all of which have already seen the end of their respective
2984 * grace periods. This omission is due to the need for no-CBs CPUs to
2985 * be holding ->nocb_lock to do this check, which is too heavy for a
2986 * common-case operation.
2987 */
2988static void check_cb_ovld(struct rcu_data *rdp)
2989{
2990 struct rcu_node *const rnp = rdp->mynode;
2991
2992 if (qovld_calc <= 0 ||
2993 ((rcu_segcblist_n_cbs(&rdp->cblist) >= qovld_calc) ==
2994 !!(READ_ONCE(rnp->cbovldmask) & rdp->grpmask)))
2995 return; // Early boot wildcard value or already set correctly.
2996 raw_spin_lock_rcu_node(rnp);
2997 check_cb_ovld_locked(rdp, rnp);
2998 raw_spin_unlock_rcu_node(rnp);
2999}
3000
3001/* Helper function for call_rcu() and friends. */
3002static void
3003__call_rcu(struct rcu_head *head, rcu_callback_t func)
3004{
3005 static atomic_t doublefrees;
3006 unsigned long flags;
3007 struct rcu_data *rdp;
3008 bool was_alldone;
3009
3010 /* Misaligned rcu_head! */
3011 WARN_ON_ONCE((unsigned long)head & (sizeof(void *) - 1));
3012
3013 if (debug_rcu_head_queue(head)) {
3014 /*
3015 * Probable double call_rcu(), so leak the callback.
3016 * Use rcu:rcu_callback trace event to find the previous
3017 * time callback was passed to __call_rcu().
3018 */
3019 if (atomic_inc_return(&doublefrees) < 4) {
3020 pr_err("%s(): Double-freed CB %p->%pS()!!! ", __func__, head, head->func);
3021 mem_dump_obj(head);
3022 }
3023 WRITE_ONCE(head->func, rcu_leak_callback);
3024 return;
3025 }
3026 head->func = func;
3027 head->next = NULL;
3028 local_irq_save(flags);
3029 kasan_record_aux_stack(head);
3030 rdp = this_cpu_ptr(&rcu_data);
3031
3032 /* Add the callback to our list. */
3033 if (unlikely(!rcu_segcblist_is_enabled(&rdp->cblist))) {
3034 // This can trigger due to call_rcu() from offline CPU:
3035 WARN_ON_ONCE(rcu_scheduler_active != RCU_SCHEDULER_INACTIVE);
3036 WARN_ON_ONCE(!rcu_is_watching());
3037 // Very early boot, before rcu_init(). Initialize if needed
3038 // and then drop through to queue the callback.
3039 if (rcu_segcblist_empty(&rdp->cblist))
3040 rcu_segcblist_init(&rdp->cblist);
3041 }
3042
3043 check_cb_ovld(rdp);
3044 if (rcu_nocb_try_bypass(rdp, head, &was_alldone, flags))
3045 return; // Enqueued onto ->nocb_bypass, so just leave.
3046 // If no-CBs CPU gets here, rcu_nocb_try_bypass() acquired ->nocb_lock.
3047 rcu_segcblist_enqueue(&rdp->cblist, head);
3048 if (__is_kvfree_rcu_offset((unsigned long)func))
3049 trace_rcu_kvfree_callback(rcu_state.name, head,
3050 (unsigned long)func,
3051 rcu_segcblist_n_cbs(&rdp->cblist));
3052 else
3053 trace_rcu_callback(rcu_state.name, head,
3054 rcu_segcblist_n_cbs(&rdp->cblist));
3055
3056 trace_rcu_segcb_stats(&rdp->cblist, TPS("SegCBQueued"));
3057
3058 /* Go handle any RCU core processing required. */
3059 if (unlikely(rcu_rdp_is_offloaded(rdp))) {
3060 __call_rcu_nocb_wake(rdp, was_alldone, flags); /* unlocks */
3061 } else {
3062 __call_rcu_core(rdp, head, flags);
3063 local_irq_restore(flags);
3064 }
3065}
3066
3067/**
3068 * call_rcu() - Queue an RCU callback for invocation after a grace period.
3069 * @head: structure to be used for queueing the RCU updates.
3070 * @func: actual callback function to be invoked after the grace period
3071 *
3072 * The callback function will be invoked some time after a full grace
3073 * period elapses, in other words after all pre-existing RCU read-side
3074 * critical sections have completed. However, the callback function
3075 * might well execute concurrently with RCU read-side critical sections
3076 * that started after call_rcu() was invoked.
3077 *
3078 * RCU read-side critical sections are delimited by rcu_read_lock()
3079 * and rcu_read_unlock(), and may be nested. In addition, but only in
3080 * v5.0 and later, regions of code across which interrupts, preemption,
3081 * or softirqs have been disabled also serve as RCU read-side critical
3082 * sections. This includes hardware interrupt handlers, softirq handlers,
3083 * and NMI handlers.
3084 *
3085 * Note that all CPUs must agree that the grace period extended beyond
3086 * all pre-existing RCU read-side critical section. On systems with more
3087 * than one CPU, this means that when "func()" is invoked, each CPU is
3088 * guaranteed to have executed a full memory barrier since the end of its
3089 * last RCU read-side critical section whose beginning preceded the call
3090 * to call_rcu(). It also means that each CPU executing an RCU read-side
3091 * critical section that continues beyond the start of "func()" must have
3092 * executed a memory barrier after the call_rcu() but before the beginning
3093 * of that RCU read-side critical section. Note that these guarantees
3094 * include CPUs that are offline, idle, or executing in user mode, as
3095 * well as CPUs that are executing in the kernel.
3096 *
3097 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
3098 * resulting RCU callback function "func()", then both CPU A and CPU B are
3099 * guaranteed to execute a full memory barrier during the time interval
3100 * between the call to call_rcu() and the invocation of "func()" -- even
3101 * if CPU A and CPU B are the same CPU (but again only if the system has
3102 * more than one CPU).
3103 *
3104 * Implementation of these memory-ordering guarantees is described here:
3105 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3106 */
3107void call_rcu(struct rcu_head *head, rcu_callback_t func)
3108{
3109 __call_rcu(head, func);
3110}
3111EXPORT_SYMBOL_GPL(call_rcu);
3112
3113
3114/* Maximum number of jiffies to wait before draining a batch. */
3115#define KFREE_DRAIN_JIFFIES (HZ / 50)
3116#define KFREE_N_BATCHES 2
3117#define FREE_N_CHANNELS 2
3118
3119/**
3120 * struct kvfree_rcu_bulk_data - single block to store kvfree_rcu() pointers
3121 * @nr_records: Number of active pointers in the array
3122 * @next: Next bulk object in the block chain
3123 * @records: Array of the kvfree_rcu() pointers
3124 */
3125struct kvfree_rcu_bulk_data {
3126 unsigned long nr_records;
3127 struct kvfree_rcu_bulk_data *next;
3128 void *records[];
3129};
3130
3131/*
3132 * This macro defines how many entries the "records" array
3133 * will contain. It is based on the fact that the size of
3134 * kvfree_rcu_bulk_data structure becomes exactly one page.
3135 */
3136#define KVFREE_BULK_MAX_ENTR \
3137 ((PAGE_SIZE - sizeof(struct kvfree_rcu_bulk_data)) / sizeof(void *))
3138
3139/**
3140 * struct kfree_rcu_cpu_work - single batch of kfree_rcu() requests
3141 * @rcu_work: Let queue_rcu_work() invoke workqueue handler after grace period
3142 * @head_free: List of kfree_rcu() objects waiting for a grace period
3143 * @bkvhead_free: Bulk-List of kvfree_rcu() objects waiting for a grace period
3144 * @krcp: Pointer to @kfree_rcu_cpu structure
3145 */
3146
3147struct kfree_rcu_cpu_work {
3148 struct rcu_work rcu_work;
3149 struct rcu_head *head_free;
3150 struct kvfree_rcu_bulk_data *bkvhead_free[FREE_N_CHANNELS];
3151 struct kfree_rcu_cpu *krcp;
3152};
3153
3154/**
3155 * struct kfree_rcu_cpu - batch up kfree_rcu() requests for RCU grace period
3156 * @head: List of kfree_rcu() objects not yet waiting for a grace period
3157 * @bkvhead: Bulk-List of kvfree_rcu() objects not yet waiting for a grace period
3158 * @krw_arr: Array of batches of kfree_rcu() objects waiting for a grace period
3159 * @lock: Synchronize access to this structure
3160 * @monitor_work: Promote @head to @head_free after KFREE_DRAIN_JIFFIES
3161 * @monitor_todo: Tracks whether a @monitor_work delayed work is pending
3162 * @initialized: The @rcu_work fields have been initialized
3163 * @count: Number of objects for which GP not started
3164 * @bkvcache:
3165 * A simple cache list that contains objects for reuse purpose.
3166 * In order to save some per-cpu space the list is singular.
3167 * Even though it is lockless an access has to be protected by the
3168 * per-cpu lock.
3169 * @page_cache_work: A work to refill the cache when it is empty
3170 * @backoff_page_cache_fill: Delay cache refills
3171 * @work_in_progress: Indicates that page_cache_work is running
3172 * @hrtimer: A hrtimer for scheduling a page_cache_work
3173 * @nr_bkv_objs: number of allocated objects at @bkvcache.
3174 *
3175 * This is a per-CPU structure. The reason that it is not included in
3176 * the rcu_data structure is to permit this code to be extracted from
3177 * the RCU files. Such extraction could allow further optimization of
3178 * the interactions with the slab allocators.
3179 */
3180struct kfree_rcu_cpu {
3181 struct rcu_head *head;
3182 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS];
3183 struct kfree_rcu_cpu_work krw_arr[KFREE_N_BATCHES];
3184 raw_spinlock_t lock;
3185 struct delayed_work monitor_work;
3186 bool monitor_todo;
3187 bool initialized;
3188 int count;
3189
3190 struct delayed_work page_cache_work;
3191 atomic_t backoff_page_cache_fill;
3192 atomic_t work_in_progress;
3193 struct hrtimer hrtimer;
3194
3195 struct llist_head bkvcache;
3196 int nr_bkv_objs;
3197};
3198
3199static DEFINE_PER_CPU(struct kfree_rcu_cpu, krc) = {
3200 .lock = __RAW_SPIN_LOCK_UNLOCKED(krc.lock),
3201};
3202
3203static __always_inline void
3204debug_rcu_bhead_unqueue(struct kvfree_rcu_bulk_data *bhead)
3205{
3206#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
3207 int i;
3208
3209 for (i = 0; i < bhead->nr_records; i++)
3210 debug_rcu_head_unqueue((struct rcu_head *)(bhead->records[i]));
3211#endif
3212}
3213
3214static inline struct kfree_rcu_cpu *
3215krc_this_cpu_lock(unsigned long *flags)
3216{
3217 struct kfree_rcu_cpu *krcp;
3218
3219 local_irq_save(*flags); // For safely calling this_cpu_ptr().
3220 krcp = this_cpu_ptr(&krc);
3221 raw_spin_lock(&krcp->lock);
3222
3223 return krcp;
3224}
3225
3226static inline void
3227krc_this_cpu_unlock(struct kfree_rcu_cpu *krcp, unsigned long flags)
3228{
3229 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3230}
3231
3232static inline struct kvfree_rcu_bulk_data *
3233get_cached_bnode(struct kfree_rcu_cpu *krcp)
3234{
3235 if (!krcp->nr_bkv_objs)
3236 return NULL;
3237
3238 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs - 1);
3239 return (struct kvfree_rcu_bulk_data *)
3240 llist_del_first(&krcp->bkvcache);
3241}
3242
3243static inline bool
3244put_cached_bnode(struct kfree_rcu_cpu *krcp,
3245 struct kvfree_rcu_bulk_data *bnode)
3246{
3247 // Check the limit.
3248 if (krcp->nr_bkv_objs >= rcu_min_cached_objs)
3249 return false;
3250
3251 llist_add((struct llist_node *) bnode, &krcp->bkvcache);
3252 WRITE_ONCE(krcp->nr_bkv_objs, krcp->nr_bkv_objs + 1);
3253 return true;
3254}
3255
3256static int
3257drain_page_cache(struct kfree_rcu_cpu *krcp)
3258{
3259 unsigned long flags;
3260 struct llist_node *page_list, *pos, *n;
3261 int freed = 0;
3262
3263 raw_spin_lock_irqsave(&krcp->lock, flags);
3264 page_list = llist_del_all(&krcp->bkvcache);
3265 WRITE_ONCE(krcp->nr_bkv_objs, 0);
3266 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3267
3268 llist_for_each_safe(pos, n, page_list) {
3269 free_page((unsigned long)pos);
3270 freed++;
3271 }
3272
3273 return freed;
3274}
3275
3276/*
3277 * This function is invoked in workqueue context after a grace period.
3278 * It frees all the objects queued on ->bkvhead_free or ->head_free.
3279 */
3280static void kfree_rcu_work(struct work_struct *work)
3281{
3282 unsigned long flags;
3283 struct kvfree_rcu_bulk_data *bkvhead[FREE_N_CHANNELS], *bnext;
3284 struct rcu_head *head, *next;
3285 struct kfree_rcu_cpu *krcp;
3286 struct kfree_rcu_cpu_work *krwp;
3287 int i, j;
3288
3289 krwp = container_of(to_rcu_work(work),
3290 struct kfree_rcu_cpu_work, rcu_work);
3291 krcp = krwp->krcp;
3292
3293 raw_spin_lock_irqsave(&krcp->lock, flags);
3294 // Channels 1 and 2.
3295 for (i = 0; i < FREE_N_CHANNELS; i++) {
3296 bkvhead[i] = krwp->bkvhead_free[i];
3297 krwp->bkvhead_free[i] = NULL;
3298 }
3299
3300 // Channel 3.
3301 head = krwp->head_free;
3302 krwp->head_free = NULL;
3303 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3304
3305 // Handle the first two channels.
3306 for (i = 0; i < FREE_N_CHANNELS; i++) {
3307 for (; bkvhead[i]; bkvhead[i] = bnext) {
3308 bnext = bkvhead[i]->next;
3309 debug_rcu_bhead_unqueue(bkvhead[i]);
3310
3311 rcu_lock_acquire(&rcu_callback_map);
3312 if (i == 0) { // kmalloc() / kfree().
3313 trace_rcu_invoke_kfree_bulk_callback(
3314 rcu_state.name, bkvhead[i]->nr_records,
3315 bkvhead[i]->records);
3316
3317 kfree_bulk(bkvhead[i]->nr_records,
3318 bkvhead[i]->records);
3319 } else { // vmalloc() / vfree().
3320 for (j = 0; j < bkvhead[i]->nr_records; j++) {
3321 trace_rcu_invoke_kvfree_callback(
3322 rcu_state.name,
3323 bkvhead[i]->records[j], 0);
3324
3325 vfree(bkvhead[i]->records[j]);
3326 }
3327 }
3328 rcu_lock_release(&rcu_callback_map);
3329
3330 raw_spin_lock_irqsave(&krcp->lock, flags);
3331 if (put_cached_bnode(krcp, bkvhead[i]))
3332 bkvhead[i] = NULL;
3333 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3334
3335 if (bkvhead[i])
3336 free_page((unsigned long) bkvhead[i]);
3337
3338 cond_resched_tasks_rcu_qs();
3339 }
3340 }
3341
3342 /*
3343 * This is used when the "bulk" path can not be used for the
3344 * double-argument of kvfree_rcu(). This happens when the
3345 * page-cache is empty, which means that objects are instead
3346 * queued on a linked list through their rcu_head structures.
3347 * This list is named "Channel 3".
3348 */
3349 for (; head; head = next) {
3350 unsigned long offset = (unsigned long)head->func;
3351 void *ptr = (void *)head - offset;
3352
3353 next = head->next;
3354 debug_rcu_head_unqueue((struct rcu_head *)ptr);
3355 rcu_lock_acquire(&rcu_callback_map);
3356 trace_rcu_invoke_kvfree_callback(rcu_state.name, head, offset);
3357
3358 if (!WARN_ON_ONCE(!__is_kvfree_rcu_offset(offset)))
3359 kvfree(ptr);
3360
3361 rcu_lock_release(&rcu_callback_map);
3362 cond_resched_tasks_rcu_qs();
3363 }
3364}
3365
3366/*
3367 * This function is invoked after the KFREE_DRAIN_JIFFIES timeout.
3368 */
3369static void kfree_rcu_monitor(struct work_struct *work)
3370{
3371 struct kfree_rcu_cpu *krcp = container_of(work,
3372 struct kfree_rcu_cpu, monitor_work.work);
3373 unsigned long flags;
3374 int i, j;
3375
3376 raw_spin_lock_irqsave(&krcp->lock, flags);
3377
3378 // Attempt to start a new batch.
3379 for (i = 0; i < KFREE_N_BATCHES; i++) {
3380 struct kfree_rcu_cpu_work *krwp = &(krcp->krw_arr[i]);
3381
3382 // Try to detach bkvhead or head and attach it over any
3383 // available corresponding free channel. It can be that
3384 // a previous RCU batch is in progress, it means that
3385 // immediately to queue another one is not possible so
3386 // in that case the monitor work is rearmed.
3387 if ((krcp->bkvhead[0] && !krwp->bkvhead_free[0]) ||
3388 (krcp->bkvhead[1] && !krwp->bkvhead_free[1]) ||
3389 (krcp->head && !krwp->head_free)) {
3390 // Channel 1 corresponds to the SLAB-pointer bulk path.
3391 // Channel 2 corresponds to vmalloc-pointer bulk path.
3392 for (j = 0; j < FREE_N_CHANNELS; j++) {
3393 if (!krwp->bkvhead_free[j]) {
3394 krwp->bkvhead_free[j] = krcp->bkvhead[j];
3395 krcp->bkvhead[j] = NULL;
3396 }
3397 }
3398
3399 // Channel 3 corresponds to both SLAB and vmalloc
3400 // objects queued on the linked list.
3401 if (!krwp->head_free) {
3402 krwp->head_free = krcp->head;
3403 krcp->head = NULL;
3404 }
3405
3406 WRITE_ONCE(krcp->count, 0);
3407
3408 // One work is per one batch, so there are three
3409 // "free channels", the batch can handle. It can
3410 // be that the work is in the pending state when
3411 // channels have been detached following by each
3412 // other.
3413 queue_rcu_work(system_wq, &krwp->rcu_work);
3414 }
3415 }
3416
3417 // If there is nothing to detach, it means that our job is
3418 // successfully done here. In case of having at least one
3419 // of the channels that is still busy we should rearm the
3420 // work to repeat an attempt. Because previous batches are
3421 // still in progress.
3422 if (!krcp->bkvhead[0] && !krcp->bkvhead[1] && !krcp->head)
3423 krcp->monitor_todo = false;
3424 else
3425 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3426
3427 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3428}
3429
3430static enum hrtimer_restart
3431schedule_page_work_fn(struct hrtimer *t)
3432{
3433 struct kfree_rcu_cpu *krcp =
3434 container_of(t, struct kfree_rcu_cpu, hrtimer);
3435
3436 queue_delayed_work(system_highpri_wq, &krcp->page_cache_work, 0);
3437 return HRTIMER_NORESTART;
3438}
3439
3440static void fill_page_cache_func(struct work_struct *work)
3441{
3442 struct kvfree_rcu_bulk_data *bnode;
3443 struct kfree_rcu_cpu *krcp =
3444 container_of(work, struct kfree_rcu_cpu,
3445 page_cache_work.work);
3446 unsigned long flags;
3447 int nr_pages;
3448 bool pushed;
3449 int i;
3450
3451 nr_pages = atomic_read(&krcp->backoff_page_cache_fill) ?
3452 1 : rcu_min_cached_objs;
3453
3454 for (i = 0; i < nr_pages; i++) {
3455 bnode = (struct kvfree_rcu_bulk_data *)
3456 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3457
3458 if (bnode) {
3459 raw_spin_lock_irqsave(&krcp->lock, flags);
3460 pushed = put_cached_bnode(krcp, bnode);
3461 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3462
3463 if (!pushed) {
3464 free_page((unsigned long) bnode);
3465 break;
3466 }
3467 }
3468 }
3469
3470 atomic_set(&krcp->work_in_progress, 0);
3471 atomic_set(&krcp->backoff_page_cache_fill, 0);
3472}
3473
3474static void
3475run_page_cache_worker(struct kfree_rcu_cpu *krcp)
3476{
3477 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3478 !atomic_xchg(&krcp->work_in_progress, 1)) {
3479 if (atomic_read(&krcp->backoff_page_cache_fill)) {
3480 queue_delayed_work(system_wq,
3481 &krcp->page_cache_work,
3482 msecs_to_jiffies(rcu_delay_page_cache_fill_msec));
3483 } else {
3484 hrtimer_init(&krcp->hrtimer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
3485 krcp->hrtimer.function = schedule_page_work_fn;
3486 hrtimer_start(&krcp->hrtimer, 0, HRTIMER_MODE_REL);
3487 }
3488 }
3489}
3490
3491// Record ptr in a page managed by krcp, with the pre-krc_this_cpu_lock()
3492// state specified by flags. If can_alloc is true, the caller must
3493// be schedulable and not be holding any locks or mutexes that might be
3494// acquired by the memory allocator or anything that it might invoke.
3495// Returns true if ptr was successfully recorded, else the caller must
3496// use a fallback.
3497static inline bool
3498add_ptr_to_bulk_krc_lock(struct kfree_rcu_cpu **krcp,
3499 unsigned long *flags, void *ptr, bool can_alloc)
3500{
3501 struct kvfree_rcu_bulk_data *bnode;
3502 int idx;
3503
3504 *krcp = krc_this_cpu_lock(flags);
3505 if (unlikely(!(*krcp)->initialized))
3506 return false;
3507
3508 idx = !!is_vmalloc_addr(ptr);
3509
3510 /* Check if a new block is required. */
3511 if (!(*krcp)->bkvhead[idx] ||
3512 (*krcp)->bkvhead[idx]->nr_records == KVFREE_BULK_MAX_ENTR) {
3513 bnode = get_cached_bnode(*krcp);
3514 if (!bnode && can_alloc) {
3515 krc_this_cpu_unlock(*krcp, *flags);
3516
3517 // __GFP_NORETRY - allows a light-weight direct reclaim
3518 // what is OK from minimizing of fallback hitting point of
3519 // view. Apart of that it forbids any OOM invoking what is
3520 // also beneficial since we are about to release memory soon.
3521 //
3522 // __GFP_NOMEMALLOC - prevents from consuming of all the
3523 // memory reserves. Please note we have a fallback path.
3524 //
3525 // __GFP_NOWARN - it is supposed that an allocation can
3526 // be failed under low memory or high memory pressure
3527 // scenarios.
3528 bnode = (struct kvfree_rcu_bulk_data *)
3529 __get_free_page(GFP_KERNEL | __GFP_NORETRY | __GFP_NOMEMALLOC | __GFP_NOWARN);
3530 *krcp = krc_this_cpu_lock(flags);
3531 }
3532
3533 if (!bnode)
3534 return false;
3535
3536 /* Initialize the new block. */
3537 bnode->nr_records = 0;
3538 bnode->next = (*krcp)->bkvhead[idx];
3539
3540 /* Attach it to the head. */
3541 (*krcp)->bkvhead[idx] = bnode;
3542 }
3543
3544 /* Finally insert. */
3545 (*krcp)->bkvhead[idx]->records
3546 [(*krcp)->bkvhead[idx]->nr_records++] = ptr;
3547
3548 return true;
3549}
3550
3551/*
3552 * Queue a request for lazy invocation of the appropriate free routine
3553 * after a grace period. Please note that three paths are maintained,
3554 * two for the common case using arrays of pointers and a third one that
3555 * is used only when the main paths cannot be used, for example, due to
3556 * memory pressure.
3557 *
3558 * Each kvfree_call_rcu() request is added to a batch. The batch will be drained
3559 * every KFREE_DRAIN_JIFFIES number of jiffies. All the objects in the batch will
3560 * be free'd in workqueue context. This allows us to: batch requests together to
3561 * reduce the number of grace periods during heavy kfree_rcu()/kvfree_rcu() load.
3562 */
3563void kvfree_call_rcu(struct rcu_head *head, rcu_callback_t func)
3564{
3565 unsigned long flags;
3566 struct kfree_rcu_cpu *krcp;
3567 bool success;
3568 void *ptr;
3569
3570 if (head) {
3571 ptr = (void *) head - (unsigned long) func;
3572 } else {
3573 /*
3574 * Please note there is a limitation for the head-less
3575 * variant, that is why there is a clear rule for such
3576 * objects: it can be used from might_sleep() context
3577 * only. For other places please embed an rcu_head to
3578 * your data.
3579 */
3580 might_sleep();
3581 ptr = (unsigned long *) func;
3582 }
3583
3584 // Queue the object but don't yet schedule the batch.
3585 if (debug_rcu_head_queue(ptr)) {
3586 // Probable double kfree_rcu(), just leak.
3587 WARN_ONCE(1, "%s(): Double-freed call. rcu_head %p\n",
3588 __func__, head);
3589
3590 // Mark as success and leave.
3591 return;
3592 }
3593
3594 kasan_record_aux_stack(ptr);
3595 success = add_ptr_to_bulk_krc_lock(&krcp, &flags, ptr, !head);
3596 if (!success) {
3597 run_page_cache_worker(krcp);
3598
3599 if (head == NULL)
3600 // Inline if kvfree_rcu(one_arg) call.
3601 goto unlock_return;
3602
3603 head->func = func;
3604 head->next = krcp->head;
3605 krcp->head = head;
3606 success = true;
3607 }
3608
3609 WRITE_ONCE(krcp->count, krcp->count + 1);
3610
3611 // Set timer to drain after KFREE_DRAIN_JIFFIES.
3612 if (rcu_scheduler_active == RCU_SCHEDULER_RUNNING &&
3613 !krcp->monitor_todo) {
3614 krcp->monitor_todo = true;
3615 schedule_delayed_work(&krcp->monitor_work, KFREE_DRAIN_JIFFIES);
3616 }
3617
3618unlock_return:
3619 krc_this_cpu_unlock(krcp, flags);
3620
3621 /*
3622 * Inline kvfree() after synchronize_rcu(). We can do
3623 * it from might_sleep() context only, so the current
3624 * CPU can pass the QS state.
3625 */
3626 if (!success) {
3627 debug_rcu_head_unqueue((struct rcu_head *) ptr);
3628 synchronize_rcu();
3629 kvfree(ptr);
3630 }
3631}
3632EXPORT_SYMBOL_GPL(kvfree_call_rcu);
3633
3634static unsigned long
3635kfree_rcu_shrink_count(struct shrinker *shrink, struct shrink_control *sc)
3636{
3637 int cpu;
3638 unsigned long count = 0;
3639
3640 /* Snapshot count of all CPUs */
3641 for_each_possible_cpu(cpu) {
3642 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3643
3644 count += READ_ONCE(krcp->count);
3645 count += READ_ONCE(krcp->nr_bkv_objs);
3646 atomic_set(&krcp->backoff_page_cache_fill, 1);
3647 }
3648
3649 return count;
3650}
3651
3652static unsigned long
3653kfree_rcu_shrink_scan(struct shrinker *shrink, struct shrink_control *sc)
3654{
3655 int cpu, freed = 0;
3656
3657 for_each_possible_cpu(cpu) {
3658 int count;
3659 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3660
3661 count = krcp->count;
3662 count += drain_page_cache(krcp);
3663 kfree_rcu_monitor(&krcp->monitor_work.work);
3664
3665 sc->nr_to_scan -= count;
3666 freed += count;
3667
3668 if (sc->nr_to_scan <= 0)
3669 break;
3670 }
3671
3672 return freed == 0 ? SHRINK_STOP : freed;
3673}
3674
3675static struct shrinker kfree_rcu_shrinker = {
3676 .count_objects = kfree_rcu_shrink_count,
3677 .scan_objects = kfree_rcu_shrink_scan,
3678 .batch = 0,
3679 .seeks = DEFAULT_SEEKS,
3680};
3681
3682void __init kfree_rcu_scheduler_running(void)
3683{
3684 int cpu;
3685 unsigned long flags;
3686
3687 for_each_possible_cpu(cpu) {
3688 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
3689
3690 raw_spin_lock_irqsave(&krcp->lock, flags);
3691 if ((!krcp->bkvhead[0] && !krcp->bkvhead[1] && !krcp->head) ||
3692 krcp->monitor_todo) {
3693 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3694 continue;
3695 }
3696 krcp->monitor_todo = true;
3697 schedule_delayed_work_on(cpu, &krcp->monitor_work,
3698 KFREE_DRAIN_JIFFIES);
3699 raw_spin_unlock_irqrestore(&krcp->lock, flags);
3700 }
3701}
3702
3703/*
3704 * During early boot, any blocking grace-period wait automatically
3705 * implies a grace period. Later on, this is never the case for PREEMPTION.
3706 *
3707 * However, because a context switch is a grace period for !PREEMPTION, any
3708 * blocking grace-period wait automatically implies a grace period if
3709 * there is only one CPU online at any point time during execution of
3710 * either synchronize_rcu() or synchronize_rcu_expedited(). It is OK to
3711 * occasionally incorrectly indicate that there are multiple CPUs online
3712 * when there was in fact only one the whole time, as this just adds some
3713 * overhead: RCU still operates correctly.
3714 */
3715static int rcu_blocking_is_gp(void)
3716{
3717 int ret;
3718
3719 if (IS_ENABLED(CONFIG_PREEMPTION))
3720 return rcu_scheduler_active == RCU_SCHEDULER_INACTIVE;
3721 might_sleep(); /* Check for RCU read-side critical section. */
3722 preempt_disable();
3723 /*
3724 * If the rcu_state.n_online_cpus counter is equal to one,
3725 * there is only one CPU, and that CPU sees all prior accesses
3726 * made by any CPU that was online at the time of its access.
3727 * Furthermore, if this counter is equal to one, its value cannot
3728 * change until after the preempt_enable() below.
3729 *
3730 * Furthermore, if rcu_state.n_online_cpus is equal to one here,
3731 * all later CPUs (both this one and any that come online later
3732 * on) are guaranteed to see all accesses prior to this point
3733 * in the code, without the need for additional memory barriers.
3734 * Those memory barriers are provided by CPU-hotplug code.
3735 */
3736 ret = READ_ONCE(rcu_state.n_online_cpus) <= 1;
3737 preempt_enable();
3738 return ret;
3739}
3740
3741/**
3742 * synchronize_rcu - wait until a grace period has elapsed.
3743 *
3744 * Control will return to the caller some time after a full grace
3745 * period has elapsed, in other words after all currently executing RCU
3746 * read-side critical sections have completed. Note, however, that
3747 * upon return from synchronize_rcu(), the caller might well be executing
3748 * concurrently with new RCU read-side critical sections that began while
3749 * synchronize_rcu() was waiting.
3750 *
3751 * RCU read-side critical sections are delimited by rcu_read_lock()
3752 * and rcu_read_unlock(), and may be nested. In addition, but only in
3753 * v5.0 and later, regions of code across which interrupts, preemption,
3754 * or softirqs have been disabled also serve as RCU read-side critical
3755 * sections. This includes hardware interrupt handlers, softirq handlers,
3756 * and NMI handlers.
3757 *
3758 * Note that this guarantee implies further memory-ordering guarantees.
3759 * On systems with more than one CPU, when synchronize_rcu() returns,
3760 * each CPU is guaranteed to have executed a full memory barrier since
3761 * the end of its last RCU read-side critical section whose beginning
3762 * preceded the call to synchronize_rcu(). In addition, each CPU having
3763 * an RCU read-side critical section that extends beyond the return from
3764 * synchronize_rcu() is guaranteed to have executed a full memory barrier
3765 * after the beginning of synchronize_rcu() and before the beginning of
3766 * that RCU read-side critical section. Note that these guarantees include
3767 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
3768 * that are executing in the kernel.
3769 *
3770 * Furthermore, if CPU A invoked synchronize_rcu(), which returned
3771 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
3772 * to have executed a full memory barrier during the execution of
3773 * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
3774 * again only if the system has more than one CPU).
3775 *
3776 * Implementation of these memory-ordering guarantees is described here:
3777 * Documentation/RCU/Design/Memory-Ordering/Tree-RCU-Memory-Ordering.rst.
3778 */
3779void synchronize_rcu(void)
3780{
3781 RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
3782 lock_is_held(&rcu_lock_map) ||
3783 lock_is_held(&rcu_sched_lock_map),
3784 "Illegal synchronize_rcu() in RCU read-side critical section");
3785 if (rcu_blocking_is_gp())
3786 return; // Context allows vacuous grace periods.
3787 if (rcu_gp_is_expedited())
3788 synchronize_rcu_expedited();
3789 else
3790 wait_rcu_gp(call_rcu);
3791}
3792EXPORT_SYMBOL_GPL(synchronize_rcu);
3793
3794/**
3795 * get_state_synchronize_rcu - Snapshot current RCU state
3796 *
3797 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3798 * or poll_state_synchronize_rcu() to determine whether or not a full
3799 * grace period has elapsed in the meantime.
3800 */
3801unsigned long get_state_synchronize_rcu(void)
3802{
3803 /*
3804 * Any prior manipulation of RCU-protected data must happen
3805 * before the load from ->gp_seq.
3806 */
3807 smp_mb(); /* ^^^ */
3808 return rcu_seq_snap(&rcu_state.gp_seq);
3809}
3810EXPORT_SYMBOL_GPL(get_state_synchronize_rcu);
3811
3812/**
3813 * start_poll_synchronize_rcu - Snapshot and start RCU grace period
3814 *
3815 * Returns a cookie that is used by a later call to cond_synchronize_rcu()
3816 * or poll_state_synchronize_rcu() to determine whether or not a full
3817 * grace period has elapsed in the meantime. If the needed grace period
3818 * is not already slated to start, notifies RCU core of the need for that
3819 * grace period.
3820 *
3821 * Interrupts must be enabled for the case where it is necessary to awaken
3822 * the grace-period kthread.
3823 */
3824unsigned long start_poll_synchronize_rcu(void)
3825{
3826 unsigned long flags;
3827 unsigned long gp_seq = get_state_synchronize_rcu();
3828 bool needwake;
3829 struct rcu_data *rdp;
3830 struct rcu_node *rnp;
3831
3832 lockdep_assert_irqs_enabled();
3833 local_irq_save(flags);
3834 rdp = this_cpu_ptr(&rcu_data);
3835 rnp = rdp->mynode;
3836 raw_spin_lock_rcu_node(rnp); // irqs already disabled.
3837 needwake = rcu_start_this_gp(rnp, rdp, gp_seq);
3838 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
3839 if (needwake)
3840 rcu_gp_kthread_wake();
3841 return gp_seq;
3842}
3843EXPORT_SYMBOL_GPL(start_poll_synchronize_rcu);
3844
3845/**
3846 * poll_state_synchronize_rcu - Conditionally wait for an RCU grace period
3847 *
3848 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3849 *
3850 * If a full RCU grace period has elapsed since the earlier call from
3851 * which oldstate was obtained, return @true, otherwise return @false.
3852 * If @false is returned, it is the caller's responsibility to invoke this
3853 * function later on until it does return @true. Alternatively, the caller
3854 * can explicitly wait for a grace period, for example, by passing @oldstate
3855 * to cond_synchronize_rcu() or by directly invoking synchronize_rcu().
3856 *
3857 * Yes, this function does not take counter wrap into account.
3858 * But counter wrap is harmless. If the counter wraps, we have waited for
3859 * more than 2 billion grace periods (and way more on a 64-bit system!).
3860 * Those needing to keep oldstate values for very long time periods
3861 * (many hours even on 32-bit systems) should check them occasionally
3862 * and either refresh them or set a flag indicating that the grace period
3863 * has completed.
3864 *
3865 * This function provides the same memory-ordering guarantees that
3866 * would be provided by a synchronize_rcu() that was invoked at the call
3867 * to the function that provided @oldstate, and that returned at the end
3868 * of this function.
3869 */
3870bool poll_state_synchronize_rcu(unsigned long oldstate)
3871{
3872 if (rcu_seq_done(&rcu_state.gp_seq, oldstate)) {
3873 smp_mb(); /* Ensure GP ends before subsequent accesses. */
3874 return true;
3875 }
3876 return false;
3877}
3878EXPORT_SYMBOL_GPL(poll_state_synchronize_rcu);
3879
3880/**
3881 * cond_synchronize_rcu - Conditionally wait for an RCU grace period
3882 *
3883 * @oldstate: value from get_state_synchronize_rcu() or start_poll_synchronize_rcu()
3884 *
3885 * If a full RCU grace period has elapsed since the earlier call to
3886 * get_state_synchronize_rcu() or start_poll_synchronize_rcu(), just return.
3887 * Otherwise, invoke synchronize_rcu() to wait for a full grace period.
3888 *
3889 * Yes, this function does not take counter wrap into account. But
3890 * counter wrap is harmless. If the counter wraps, we have waited for
3891 * more than 2 billion grace periods (and way more on a 64-bit system!),
3892 * so waiting for one additional grace period should be just fine.
3893 *
3894 * This function provides the same memory-ordering guarantees that
3895 * would be provided by a synchronize_rcu() that was invoked at the call
3896 * to the function that provided @oldstate, and that returned at the end
3897 * of this function.
3898 */
3899void cond_synchronize_rcu(unsigned long oldstate)
3900{
3901 if (!poll_state_synchronize_rcu(oldstate))
3902 synchronize_rcu();
3903}
3904EXPORT_SYMBOL_GPL(cond_synchronize_rcu);
3905
3906/*
3907 * Check to see if there is any immediate RCU-related work to be done by
3908 * the current CPU, returning 1 if so and zero otherwise. The checks are
3909 * in order of increasing expense: checks that can be carried out against
3910 * CPU-local state are performed first. However, we must check for CPU
3911 * stalls first, else we might not get a chance.
3912 */
3913static int rcu_pending(int user)
3914{
3915 bool gp_in_progress;
3916 struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
3917 struct rcu_node *rnp = rdp->mynode;
3918
3919 lockdep_assert_irqs_disabled();
3920
3921 /* Check for CPU stalls, if enabled. */
3922 check_cpu_stall(rdp);
3923
3924 /* Does this CPU need a deferred NOCB wakeup? */
3925 if (rcu_nocb_need_deferred_wakeup(rdp, RCU_NOCB_WAKE))
3926 return 1;
3927
3928 /* Is this a nohz_full CPU in userspace or idle? (Ignore RCU if so.) */
3929 if ((user || rcu_is_cpu_rrupt_from_idle()) && rcu_nohz_full_cpu())
3930 return 0;
3931
3932 /* Is the RCU core waiting for a quiescent state from this CPU? */
3933 gp_in_progress = rcu_gp_in_progress();
3934 if (rdp->core_needs_qs && !rdp->cpu_no_qs.b.norm && gp_in_progress)
3935 return 1;
3936
3937 /* Does this CPU have callbacks ready to invoke? */
3938 if (!rcu_rdp_is_offloaded(rdp) &&
3939 rcu_segcblist_ready_cbs(&rdp->cblist))
3940 return 1;
3941
3942 /* Has RCU gone idle with this CPU needing another grace period? */
3943 if (!gp_in_progress && rcu_segcblist_is_enabled(&rdp->cblist) &&
3944 !rcu_rdp_is_offloaded(rdp) &&
3945 !rcu_segcblist_restempty(&rdp->cblist, RCU_NEXT_READY_TAIL))
3946 return 1;
3947
3948 /* Have RCU grace period completed or started? */
3949 if (rcu_seq_current(&rnp->gp_seq) != rdp->gp_seq ||
3950 unlikely(READ_ONCE(rdp->gpwrap))) /* outside lock */
3951 return 1;
3952
3953 /* nothing to do */
3954 return 0;
3955}
3956
3957/*
3958 * Helper function for rcu_barrier() tracing. If tracing is disabled,
3959 * the compiler is expected to optimize this away.
3960 */
3961static void rcu_barrier_trace(const char *s, int cpu, unsigned long done)
3962{
3963 trace_rcu_barrier(rcu_state.name, s, cpu,
3964 atomic_read(&rcu_state.barrier_cpu_count), done);
3965}
3966
3967/*
3968 * RCU callback function for rcu_barrier(). If we are last, wake
3969 * up the task executing rcu_barrier().
3970 *
3971 * Note that the value of rcu_state.barrier_sequence must be captured
3972 * before the atomic_dec_and_test(). Otherwise, if this CPU is not last,
3973 * other CPUs might count the value down to zero before this CPU gets
3974 * around to invoking rcu_barrier_trace(), which might result in bogus
3975 * data from the next instance of rcu_barrier().
3976 */
3977static void rcu_barrier_callback(struct rcu_head *rhp)
3978{
3979 unsigned long __maybe_unused s = rcu_state.barrier_sequence;
3980
3981 if (atomic_dec_and_test(&rcu_state.barrier_cpu_count)) {
3982 rcu_barrier_trace(TPS("LastCB"), -1, s);
3983 complete(&rcu_state.barrier_completion);
3984 } else {
3985 rcu_barrier_trace(TPS("CB"), -1, s);
3986 }
3987}
3988
3989/*
3990 * Called with preemption disabled, and from cross-cpu IRQ context.
3991 */
3992static void rcu_barrier_func(void *cpu_in)
3993{
3994 uintptr_t cpu = (uintptr_t)cpu_in;
3995 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
3996
3997 rcu_barrier_trace(TPS("IRQ"), -1, rcu_state.barrier_sequence);
3998 rdp->barrier_head.func = rcu_barrier_callback;
3999 debug_rcu_head_queue(&rdp->barrier_head);
4000 rcu_nocb_lock(rdp);
4001 WARN_ON_ONCE(!rcu_nocb_flush_bypass(rdp, NULL, jiffies));
4002 if (rcu_segcblist_entrain(&rdp->cblist, &rdp->barrier_head)) {
4003 atomic_inc(&rcu_state.barrier_cpu_count);
4004 } else {
4005 debug_rcu_head_unqueue(&rdp->barrier_head);
4006 rcu_barrier_trace(TPS("IRQNQ"), -1,
4007 rcu_state.barrier_sequence);
4008 }
4009 rcu_nocb_unlock(rdp);
4010}
4011
4012/**
4013 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
4014 *
4015 * Note that this primitive does not necessarily wait for an RCU grace period
4016 * to complete. For example, if there are no RCU callbacks queued anywhere
4017 * in the system, then rcu_barrier() is within its rights to return
4018 * immediately, without waiting for anything, much less an RCU grace period.
4019 */
4020void rcu_barrier(void)
4021{
4022 uintptr_t cpu;
4023 struct rcu_data *rdp;
4024 unsigned long s = rcu_seq_snap(&rcu_state.barrier_sequence);
4025
4026 rcu_barrier_trace(TPS("Begin"), -1, s);
4027
4028 /* Take mutex to serialize concurrent rcu_barrier() requests. */
4029 mutex_lock(&rcu_state.barrier_mutex);
4030
4031 /* Did someone else do our work for us? */
4032 if (rcu_seq_done(&rcu_state.barrier_sequence, s)) {
4033 rcu_barrier_trace(TPS("EarlyExit"), -1,
4034 rcu_state.barrier_sequence);
4035 smp_mb(); /* caller's subsequent code after above check. */
4036 mutex_unlock(&rcu_state.barrier_mutex);
4037 return;
4038 }
4039
4040 /* Mark the start of the barrier operation. */
4041 rcu_seq_start(&rcu_state.barrier_sequence);
4042 rcu_barrier_trace(TPS("Inc1"), -1, rcu_state.barrier_sequence);
4043
4044 /*
4045 * Initialize the count to two rather than to zero in order
4046 * to avoid a too-soon return to zero in case of an immediate
4047 * invocation of the just-enqueued callback (or preemption of
4048 * this task). Exclude CPU-hotplug operations to ensure that no
4049 * offline non-offloaded CPU has callbacks queued.
4050 */
4051 init_completion(&rcu_state.barrier_completion);
4052 atomic_set(&rcu_state.barrier_cpu_count, 2);
4053 get_online_cpus();
4054
4055 /*
4056 * Force each CPU with callbacks to register a new callback.
4057 * When that callback is invoked, we will know that all of the
4058 * corresponding CPU's preceding callbacks have been invoked.
4059 */
4060 for_each_possible_cpu(cpu) {
4061 rdp = per_cpu_ptr(&rcu_data, cpu);
4062 if (cpu_is_offline(cpu) &&
4063 !rcu_rdp_is_offloaded(rdp))
4064 continue;
4065 if (rcu_segcblist_n_cbs(&rdp->cblist) && cpu_online(cpu)) {
4066 rcu_barrier_trace(TPS("OnlineQ"), cpu,
4067 rcu_state.barrier_sequence);
4068 smp_call_function_single(cpu, rcu_barrier_func, (void *)cpu, 1);
4069 } else if (rcu_segcblist_n_cbs(&rdp->cblist) &&
4070 cpu_is_offline(cpu)) {
4071 rcu_barrier_trace(TPS("OfflineNoCBQ"), cpu,
4072 rcu_state.barrier_sequence);
4073 local_irq_disable();
4074 rcu_barrier_func((void *)cpu);
4075 local_irq_enable();
4076 } else if (cpu_is_offline(cpu)) {
4077 rcu_barrier_trace(TPS("OfflineNoCBNoQ"), cpu,
4078 rcu_state.barrier_sequence);
4079 } else {
4080 rcu_barrier_trace(TPS("OnlineNQ"), cpu,
4081 rcu_state.barrier_sequence);
4082 }
4083 }
4084 put_online_cpus();
4085
4086 /*
4087 * Now that we have an rcu_barrier_callback() callback on each
4088 * CPU, and thus each counted, remove the initial count.
4089 */
4090 if (atomic_sub_and_test(2, &rcu_state.barrier_cpu_count))
4091 complete(&rcu_state.barrier_completion);
4092
4093 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
4094 wait_for_completion(&rcu_state.barrier_completion);
4095
4096 /* Mark the end of the barrier operation. */
4097 rcu_barrier_trace(TPS("Inc2"), -1, rcu_state.barrier_sequence);
4098 rcu_seq_end(&rcu_state.barrier_sequence);
4099
4100 /* Other rcu_barrier() invocations can now safely proceed. */
4101 mutex_unlock(&rcu_state.barrier_mutex);
4102}
4103EXPORT_SYMBOL_GPL(rcu_barrier);
4104
4105/*
4106 * Propagate ->qsinitmask bits up the rcu_node tree to account for the
4107 * first CPU in a given leaf rcu_node structure coming online. The caller
4108 * must hold the corresponding leaf rcu_node ->lock with interrupts
4109 * disabled.
4110 */
4111static void rcu_init_new_rnp(struct rcu_node *rnp_leaf)
4112{
4113 long mask;
4114 long oldmask;
4115 struct rcu_node *rnp = rnp_leaf;
4116
4117 raw_lockdep_assert_held_rcu_node(rnp_leaf);
4118 WARN_ON_ONCE(rnp->wait_blkd_tasks);
4119 for (;;) {
4120 mask = rnp->grpmask;
4121 rnp = rnp->parent;
4122 if (rnp == NULL)
4123 return;
4124 raw_spin_lock_rcu_node(rnp); /* Interrupts already disabled. */
4125 oldmask = rnp->qsmaskinit;
4126 rnp->qsmaskinit |= mask;
4127 raw_spin_unlock_rcu_node(rnp); /* Interrupts remain disabled. */
4128 if (oldmask)
4129 return;
4130 }
4131}
4132
4133/*
4134 * Do boot-time initialization of a CPU's per-CPU RCU data.
4135 */
4136static void __init
4137rcu_boot_init_percpu_data(int cpu)
4138{
4139 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4140
4141 /* Set up local state, ensuring consistent view of global state. */
4142 rdp->grpmask = leaf_node_cpu_bit(rdp->mynode, cpu);
4143 INIT_WORK(&rdp->strict_work, strict_work_handler);
4144 WARN_ON_ONCE(rdp->dynticks_nesting != 1);
4145 WARN_ON_ONCE(rcu_dynticks_in_eqs(rcu_dynticks_snap(rdp)));
4146 rdp->rcu_ofl_gp_seq = rcu_state.gp_seq;
4147 rdp->rcu_ofl_gp_flags = RCU_GP_CLEANED;
4148 rdp->rcu_onl_gp_seq = rcu_state.gp_seq;
4149 rdp->rcu_onl_gp_flags = RCU_GP_CLEANED;
4150 rdp->cpu = cpu;
4151 rcu_boot_init_nocb_percpu_data(rdp);
4152}
4153
4154/*
4155 * Invoked early in the CPU-online process, when pretty much all services
4156 * are available. The incoming CPU is not present.
4157 *
4158 * Initializes a CPU's per-CPU RCU data. Note that only one online or
4159 * offline event can be happening at a given time. Note also that we can
4160 * accept some slop in the rsp->gp_seq access due to the fact that this
4161 * CPU cannot possibly have any non-offloaded RCU callbacks in flight yet.
4162 * And any offloaded callbacks are being numbered elsewhere.
4163 */
4164int rcutree_prepare_cpu(unsigned int cpu)
4165{
4166 unsigned long flags;
4167 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4168 struct rcu_node *rnp = rcu_get_root();
4169
4170 /* Set up local state, ensuring consistent view of global state. */
4171 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4172 rdp->qlen_last_fqs_check = 0;
4173 rdp->n_force_qs_snap = rcu_state.n_force_qs;
4174 rdp->blimit = blimit;
4175 rdp->dynticks_nesting = 1; /* CPU not up, no tearing. */
4176 rcu_dynticks_eqs_online();
4177 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
4178
4179 /*
4180 * Only non-NOCB CPUs that didn't have early-boot callbacks need to be
4181 * (re-)initialized.
4182 */
4183 if (!rcu_segcblist_is_enabled(&rdp->cblist))
4184 rcu_segcblist_init(&rdp->cblist); /* Re-enable callbacks. */
4185
4186 /*
4187 * Add CPU to leaf rcu_node pending-online bitmask. Any needed
4188 * propagation up the rcu_node tree will happen at the beginning
4189 * of the next grace period.
4190 */
4191 rnp = rdp->mynode;
4192 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
4193 rdp->beenonline = true; /* We have now been online. */
4194 rdp->gp_seq = READ_ONCE(rnp->gp_seq);
4195 rdp->gp_seq_needed = rdp->gp_seq;
4196 rdp->cpu_no_qs.b.norm = true;
4197 rdp->core_needs_qs = false;
4198 rdp->rcu_iw_pending = false;
4199 rdp->rcu_iw = IRQ_WORK_INIT_HARD(rcu_iw_handler);
4200 rdp->rcu_iw_gp_seq = rdp->gp_seq - 1;
4201 trace_rcu_grace_period(rcu_state.name, rdp->gp_seq, TPS("cpuonl"));
4202 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4203 rcu_spawn_one_boost_kthread(rnp);
4204 rcu_spawn_cpu_nocb_kthread(cpu);
4205 WRITE_ONCE(rcu_state.n_online_cpus, rcu_state.n_online_cpus + 1);
4206
4207 return 0;
4208}
4209
4210/*
4211 * Update RCU priority boot kthread affinity for CPU-hotplug changes.
4212 */
4213static void rcutree_affinity_setting(unsigned int cpu, int outgoing)
4214{
4215 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4216
4217 rcu_boost_kthread_setaffinity(rdp->mynode, outgoing);
4218}
4219
4220/*
4221 * Near the end of the CPU-online process. Pretty much all services
4222 * enabled, and the CPU is now very much alive.
4223 */
4224int rcutree_online_cpu(unsigned int cpu)
4225{
4226 unsigned long flags;
4227 struct rcu_data *rdp;
4228 struct rcu_node *rnp;
4229
4230 rdp = per_cpu_ptr(&rcu_data, cpu);
4231 rnp = rdp->mynode;
4232 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4233 rnp->ffmask |= rdp->grpmask;
4234 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4235 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
4236 return 0; /* Too early in boot for scheduler work. */
4237 sync_sched_exp_online_cleanup(cpu);
4238 rcutree_affinity_setting(cpu, -1);
4239
4240 // Stop-machine done, so allow nohz_full to disable tick.
4241 tick_dep_clear(TICK_DEP_BIT_RCU);
4242 return 0;
4243}
4244
4245/*
4246 * Near the beginning of the process. The CPU is still very much alive
4247 * with pretty much all services enabled.
4248 */
4249int rcutree_offline_cpu(unsigned int cpu)
4250{
4251 unsigned long flags;
4252 struct rcu_data *rdp;
4253 struct rcu_node *rnp;
4254
4255 rdp = per_cpu_ptr(&rcu_data, cpu);
4256 rnp = rdp->mynode;
4257 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4258 rnp->ffmask &= ~rdp->grpmask;
4259 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4260
4261 rcutree_affinity_setting(cpu, cpu);
4262
4263 // nohz_full CPUs need the tick for stop-machine to work quickly
4264 tick_dep_set(TICK_DEP_BIT_RCU);
4265 return 0;
4266}
4267
4268/*
4269 * Mark the specified CPU as being online so that subsequent grace periods
4270 * (both expedited and normal) will wait on it. Note that this means that
4271 * incoming CPUs are not allowed to use RCU read-side critical sections
4272 * until this function is called. Failing to observe this restriction
4273 * will result in lockdep splats.
4274 *
4275 * Note that this function is special in that it is invoked directly
4276 * from the incoming CPU rather than from the cpuhp_step mechanism.
4277 * This is because this function must be invoked at a precise location.
4278 */
4279void rcu_cpu_starting(unsigned int cpu)
4280{
4281 unsigned long flags;
4282 unsigned long mask;
4283 struct rcu_data *rdp;
4284 struct rcu_node *rnp;
4285 bool newcpu;
4286
4287 rdp = per_cpu_ptr(&rcu_data, cpu);
4288 if (rdp->cpu_started)
4289 return;
4290 rdp->cpu_started = true;
4291
4292 rnp = rdp->mynode;
4293 mask = rdp->grpmask;
4294 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4295 WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
4296 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4297 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4298 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext | mask);
4299 newcpu = !(rnp->expmaskinitnext & mask);
4300 rnp->expmaskinitnext |= mask;
4301 /* Allow lockless access for expedited grace periods. */
4302 smp_store_release(&rcu_state.ncpus, rcu_state.ncpus + newcpu); /* ^^^ */
4303 ASSERT_EXCLUSIVE_WRITER(rcu_state.ncpus);
4304 rcu_gpnum_ovf(rnp, rdp); /* Offline-induced counter wrap? */
4305 rdp->rcu_onl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4306 rdp->rcu_onl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4307
4308 /* An incoming CPU should never be blocking a grace period. */
4309 if (WARN_ON_ONCE(rnp->qsmask & mask)) { /* RCU waiting on incoming CPU? */
4310 rcu_disable_urgency_upon_qs(rdp);
4311 /* Report QS -after- changing ->qsmaskinitnext! */
4312 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4313 } else {
4314 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4315 }
4316 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4317 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4318 WARN_ON_ONCE(rnp->ofl_seq & 0x1);
4319 smp_mb(); /* Ensure RCU read-side usage follows above initialization. */
4320}
4321
4322/*
4323 * The outgoing function has no further need of RCU, so remove it from
4324 * the rcu_node tree's ->qsmaskinitnext bit masks.
4325 *
4326 * Note that this function is special in that it is invoked directly
4327 * from the outgoing CPU rather than from the cpuhp_step mechanism.
4328 * This is because this function must be invoked at a precise location.
4329 */
4330void rcu_report_dead(unsigned int cpu)
4331{
4332 unsigned long flags;
4333 unsigned long mask;
4334 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4335 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
4336
4337 // Do any dangling deferred wakeups.
4338 do_nocb_deferred_wakeup(rdp);
4339
4340 /* QS for any half-done expedited grace period. */
4341 preempt_disable();
4342 rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
4343 preempt_enable();
4344 rcu_preempt_deferred_qs(current);
4345
4346 /* Remove outgoing CPU from mask in the leaf rcu_node structure. */
4347 mask = rdp->grpmask;
4348 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4349 WARN_ON_ONCE(!(rnp->ofl_seq & 0x1));
4350 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4351 raw_spin_lock(&rcu_state.ofl_lock);
4352 raw_spin_lock_irqsave_rcu_node(rnp, flags); /* Enforce GP memory-order guarantee. */
4353 rdp->rcu_ofl_gp_seq = READ_ONCE(rcu_state.gp_seq);
4354 rdp->rcu_ofl_gp_flags = READ_ONCE(rcu_state.gp_flags);
4355 if (rnp->qsmask & mask) { /* RCU waiting on outgoing CPU? */
4356 /* Report quiescent state -before- changing ->qsmaskinitnext! */
4357 rcu_report_qs_rnp(mask, rnp, rnp->gp_seq, flags);
4358 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4359 }
4360 WRITE_ONCE(rnp->qsmaskinitnext, rnp->qsmaskinitnext & ~mask);
4361 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4362 raw_spin_unlock(&rcu_state.ofl_lock);
4363 smp_mb(); // Pair with rcu_gp_cleanup()'s ->ofl_seq barrier().
4364 WRITE_ONCE(rnp->ofl_seq, rnp->ofl_seq + 1);
4365 WARN_ON_ONCE(rnp->ofl_seq & 0x1);
4366
4367 rdp->cpu_started = false;
4368}
4369
4370#ifdef CONFIG_HOTPLUG_CPU
4371/*
4372 * The outgoing CPU has just passed through the dying-idle state, and we
4373 * are being invoked from the CPU that was IPIed to continue the offline
4374 * operation. Migrate the outgoing CPU's callbacks to the current CPU.
4375 */
4376void rcutree_migrate_callbacks(int cpu)
4377{
4378 unsigned long flags;
4379 struct rcu_data *my_rdp;
4380 struct rcu_node *my_rnp;
4381 struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
4382 bool needwake;
4383
4384 if (rcu_rdp_is_offloaded(rdp) ||
4385 rcu_segcblist_empty(&rdp->cblist))
4386 return; /* No callbacks to migrate. */
4387
4388 local_irq_save(flags);
4389 my_rdp = this_cpu_ptr(&rcu_data);
4390 my_rnp = my_rdp->mynode;
4391 rcu_nocb_lock(my_rdp); /* irqs already disabled. */
4392 WARN_ON_ONCE(!rcu_nocb_flush_bypass(my_rdp, NULL, jiffies));
4393 raw_spin_lock_rcu_node(my_rnp); /* irqs already disabled. */
4394 /* Leverage recent GPs and set GP for new callbacks. */
4395 needwake = rcu_advance_cbs(my_rnp, rdp) ||
4396 rcu_advance_cbs(my_rnp, my_rdp);
4397 rcu_segcblist_merge(&my_rdp->cblist, &rdp->cblist);
4398 needwake = needwake || rcu_advance_cbs(my_rnp, my_rdp);
4399 rcu_segcblist_disable(&rdp->cblist);
4400 WARN_ON_ONCE(rcu_segcblist_empty(&my_rdp->cblist) !=
4401 !rcu_segcblist_n_cbs(&my_rdp->cblist));
4402 if (rcu_rdp_is_offloaded(my_rdp)) {
4403 raw_spin_unlock_rcu_node(my_rnp); /* irqs remain disabled. */
4404 __call_rcu_nocb_wake(my_rdp, true, flags);
4405 } else {
4406 rcu_nocb_unlock(my_rdp); /* irqs remain disabled. */
4407 raw_spin_unlock_irqrestore_rcu_node(my_rnp, flags);
4408 }
4409 if (needwake)
4410 rcu_gp_kthread_wake();
4411 lockdep_assert_irqs_enabled();
4412 WARN_ONCE(rcu_segcblist_n_cbs(&rdp->cblist) != 0 ||
4413 !rcu_segcblist_empty(&rdp->cblist),
4414 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, 1stCB=%p\n",
4415 cpu, rcu_segcblist_n_cbs(&rdp->cblist),
4416 rcu_segcblist_first_cb(&rdp->cblist));
4417}
4418#endif
4419
4420/*
4421 * On non-huge systems, use expedited RCU grace periods to make suspend
4422 * and hibernation run faster.
4423 */
4424static int rcu_pm_notify(struct notifier_block *self,
4425 unsigned long action, void *hcpu)
4426{
4427 switch (action) {
4428 case PM_HIBERNATION_PREPARE:
4429 case PM_SUSPEND_PREPARE:
4430 rcu_expedite_gp();
4431 break;
4432 case PM_POST_HIBERNATION:
4433 case PM_POST_SUSPEND:
4434 rcu_unexpedite_gp();
4435 break;
4436 default:
4437 break;
4438 }
4439 return NOTIFY_OK;
4440}
4441
4442/*
4443 * Spawn the kthreads that handle RCU's grace periods.
4444 */
4445static int __init rcu_spawn_gp_kthread(void)
4446{
4447 unsigned long flags;
4448 int kthread_prio_in = kthread_prio;
4449 struct rcu_node *rnp;
4450 struct sched_param sp;
4451 struct task_struct *t;
4452
4453 /* Force priority into range. */
4454 if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 2
4455 && IS_BUILTIN(CONFIG_RCU_TORTURE_TEST))
4456 kthread_prio = 2;
4457 else if (IS_ENABLED(CONFIG_RCU_BOOST) && kthread_prio < 1)
4458 kthread_prio = 1;
4459 else if (kthread_prio < 0)
4460 kthread_prio = 0;
4461 else if (kthread_prio > 99)
4462 kthread_prio = 99;
4463
4464 if (kthread_prio != kthread_prio_in)
4465 pr_alert("rcu_spawn_gp_kthread(): Limited prio to %d from %d\n",
4466 kthread_prio, kthread_prio_in);
4467
4468 rcu_scheduler_fully_active = 1;
4469 t = kthread_create(rcu_gp_kthread, NULL, "%s", rcu_state.name);
4470 if (WARN_ONCE(IS_ERR(t), "%s: Could not start grace-period kthread, OOM is now expected behavior\n", __func__))
4471 return 0;
4472 if (kthread_prio) {
4473 sp.sched_priority = kthread_prio;
4474 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
4475 }
4476 rnp = rcu_get_root();
4477 raw_spin_lock_irqsave_rcu_node(rnp, flags);
4478 WRITE_ONCE(rcu_state.gp_activity, jiffies);
4479 WRITE_ONCE(rcu_state.gp_req_activity, jiffies);
4480 // Reset .gp_activity and .gp_req_activity before setting .gp_kthread.
4481 smp_store_release(&rcu_state.gp_kthread, t); /* ^^^ */
4482 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
4483 wake_up_process(t);
4484 rcu_spawn_nocb_kthreads();
4485 rcu_spawn_boost_kthreads();
4486 rcu_spawn_core_kthreads();
4487 return 0;
4488}
4489early_initcall(rcu_spawn_gp_kthread);
4490
4491/*
4492 * This function is invoked towards the end of the scheduler's
4493 * initialization process. Before this is called, the idle task might
4494 * contain synchronous grace-period primitives (during which time, this idle
4495 * task is booting the system, and such primitives are no-ops). After this
4496 * function is called, any synchronous grace-period primitives are run as
4497 * expedited, with the requesting task driving the grace period forward.
4498 * A later core_initcall() rcu_set_runtime_mode() will switch to full
4499 * runtime RCU functionality.
4500 */
4501void rcu_scheduler_starting(void)
4502{
4503 WARN_ON(num_online_cpus() != 1);
4504 WARN_ON(nr_context_switches() > 0);
4505 rcu_test_sync_prims();
4506 rcu_scheduler_active = RCU_SCHEDULER_INIT;
4507 rcu_test_sync_prims();
4508}
4509
4510/*
4511 * Helper function for rcu_init() that initializes the rcu_state structure.
4512 */
4513static void __init rcu_init_one(void)
4514{
4515 static const char * const buf[] = RCU_NODE_NAME_INIT;
4516 static const char * const fqs[] = RCU_FQS_NAME_INIT;
4517 static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
4518 static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
4519
4520 int levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
4521 int cpustride = 1;
4522 int i;
4523 int j;
4524 struct rcu_node *rnp;
4525
4526 BUILD_BUG_ON(RCU_NUM_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
4527
4528 /* Silence gcc 4.8 false positive about array index out of range. */
4529 if (rcu_num_lvls <= 0 || rcu_num_lvls > RCU_NUM_LVLS)
4530 panic("rcu_init_one: rcu_num_lvls out of range");
4531
4532 /* Initialize the level-tracking arrays. */
4533
4534 for (i = 1; i < rcu_num_lvls; i++)
4535 rcu_state.level[i] =
4536 rcu_state.level[i - 1] + num_rcu_lvl[i - 1];
4537 rcu_init_levelspread(levelspread, num_rcu_lvl);
4538
4539 /* Initialize the elements themselves, starting from the leaves. */
4540
4541 for (i = rcu_num_lvls - 1; i >= 0; i--) {
4542 cpustride *= levelspread[i];
4543 rnp = rcu_state.level[i];
4544 for (j = 0; j < num_rcu_lvl[i]; j++, rnp++) {
4545 raw_spin_lock_init(&ACCESS_PRIVATE(rnp, lock));
4546 lockdep_set_class_and_name(&ACCESS_PRIVATE(rnp, lock),
4547 &rcu_node_class[i], buf[i]);
4548 raw_spin_lock_init(&rnp->fqslock);
4549 lockdep_set_class_and_name(&rnp->fqslock,
4550 &rcu_fqs_class[i], fqs[i]);
4551 rnp->gp_seq = rcu_state.gp_seq;
4552 rnp->gp_seq_needed = rcu_state.gp_seq;
4553 rnp->completedqs = rcu_state.gp_seq;
4554 rnp->qsmask = 0;
4555 rnp->qsmaskinit = 0;
4556 rnp->grplo = j * cpustride;
4557 rnp->grphi = (j + 1) * cpustride - 1;
4558 if (rnp->grphi >= nr_cpu_ids)
4559 rnp->grphi = nr_cpu_ids - 1;
4560 if (i == 0) {
4561 rnp->grpnum = 0;
4562 rnp->grpmask = 0;
4563 rnp->parent = NULL;
4564 } else {
4565 rnp->grpnum = j % levelspread[i - 1];
4566 rnp->grpmask = BIT(rnp->grpnum);
4567 rnp->parent = rcu_state.level[i - 1] +
4568 j / levelspread[i - 1];
4569 }
4570 rnp->level = i;
4571 INIT_LIST_HEAD(&rnp->blkd_tasks);
4572 rcu_init_one_nocb(rnp);
4573 init_waitqueue_head(&rnp->exp_wq[0]);
4574 init_waitqueue_head(&rnp->exp_wq[1]);
4575 init_waitqueue_head(&rnp->exp_wq[2]);
4576 init_waitqueue_head(&rnp->exp_wq[3]);
4577 spin_lock_init(&rnp->exp_lock);
4578 }
4579 }
4580
4581 init_swait_queue_head(&rcu_state.gp_wq);
4582 init_swait_queue_head(&rcu_state.expedited_wq);
4583 rnp = rcu_first_leaf_node();
4584 for_each_possible_cpu(i) {
4585 while (i > rnp->grphi)
4586 rnp++;
4587 per_cpu_ptr(&rcu_data, i)->mynode = rnp;
4588 rcu_boot_init_percpu_data(i);
4589 }
4590}
4591
4592/*
4593 * Compute the rcu_node tree geometry from kernel parameters. This cannot
4594 * replace the definitions in tree.h because those are needed to size
4595 * the ->node array in the rcu_state structure.
4596 */
4597void rcu_init_geometry(void)
4598{
4599 ulong d;
4600 int i;
4601 static unsigned long old_nr_cpu_ids;
4602 int rcu_capacity[RCU_NUM_LVLS];
4603 static bool initialized;
4604
4605 if (initialized) {
4606 /*
4607 * Warn if setup_nr_cpu_ids() had not yet been invoked,
4608 * unless nr_cpus_ids == NR_CPUS, in which case who cares?
4609 */
4610 WARN_ON_ONCE(old_nr_cpu_ids != nr_cpu_ids);
4611 return;
4612 }
4613
4614 old_nr_cpu_ids = nr_cpu_ids;
4615 initialized = true;
4616
4617 /*
4618 * Initialize any unspecified boot parameters.
4619 * The default values of jiffies_till_first_fqs and
4620 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
4621 * value, which is a function of HZ, then adding one for each
4622 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
4623 */
4624 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
4625 if (jiffies_till_first_fqs == ULONG_MAX)
4626 jiffies_till_first_fqs = d;
4627 if (jiffies_till_next_fqs == ULONG_MAX)
4628 jiffies_till_next_fqs = d;
4629 adjust_jiffies_till_sched_qs();
4630
4631 /* If the compile-time values are accurate, just leave. */
4632 if (rcu_fanout_leaf == RCU_FANOUT_LEAF &&
4633 nr_cpu_ids == NR_CPUS)
4634 return;
4635 pr_info("Adjusting geometry for rcu_fanout_leaf=%d, nr_cpu_ids=%u\n",
4636 rcu_fanout_leaf, nr_cpu_ids);
4637
4638 /*
4639 * The boot-time rcu_fanout_leaf parameter must be at least two
4640 * and cannot exceed the number of bits in the rcu_node masks.
4641 * Complain and fall back to the compile-time values if this
4642 * limit is exceeded.
4643 */
4644 if (rcu_fanout_leaf < 2 ||
4645 rcu_fanout_leaf > sizeof(unsigned long) * 8) {
4646 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4647 WARN_ON(1);
4648 return;
4649 }
4650
4651 /*
4652 * Compute number of nodes that can be handled an rcu_node tree
4653 * with the given number of levels.
4654 */
4655 rcu_capacity[0] = rcu_fanout_leaf;
4656 for (i = 1; i < RCU_NUM_LVLS; i++)
4657 rcu_capacity[i] = rcu_capacity[i - 1] * RCU_FANOUT;
4658
4659 /*
4660 * The tree must be able to accommodate the configured number of CPUs.
4661 * If this limit is exceeded, fall back to the compile-time values.
4662 */
4663 if (nr_cpu_ids > rcu_capacity[RCU_NUM_LVLS - 1]) {
4664 rcu_fanout_leaf = RCU_FANOUT_LEAF;
4665 WARN_ON(1);
4666 return;
4667 }
4668
4669 /* Calculate the number of levels in the tree. */
4670 for (i = 0; nr_cpu_ids > rcu_capacity[i]; i++) {
4671 }
4672 rcu_num_lvls = i + 1;
4673
4674 /* Calculate the number of rcu_nodes at each level of the tree. */
4675 for (i = 0; i < rcu_num_lvls; i++) {
4676 int cap = rcu_capacity[(rcu_num_lvls - 1) - i];
4677 num_rcu_lvl[i] = DIV_ROUND_UP(nr_cpu_ids, cap);
4678 }
4679
4680 /* Calculate the total number of rcu_node structures. */
4681 rcu_num_nodes = 0;
4682 for (i = 0; i < rcu_num_lvls; i++)
4683 rcu_num_nodes += num_rcu_lvl[i];
4684}
4685
4686/*
4687 * Dump out the structure of the rcu_node combining tree associated
4688 * with the rcu_state structure.
4689 */
4690static void __init rcu_dump_rcu_node_tree(void)
4691{
4692 int level = 0;
4693 struct rcu_node *rnp;
4694
4695 pr_info("rcu_node tree layout dump\n");
4696 pr_info(" ");
4697 rcu_for_each_node_breadth_first(rnp) {
4698 if (rnp->level != level) {
4699 pr_cont("\n");
4700 pr_info(" ");
4701 level = rnp->level;
4702 }
4703 pr_cont("%d:%d ^%d ", rnp->grplo, rnp->grphi, rnp->grpnum);
4704 }
4705 pr_cont("\n");
4706}
4707
4708struct workqueue_struct *rcu_gp_wq;
4709struct workqueue_struct *rcu_par_gp_wq;
4710
4711static void __init kfree_rcu_batch_init(void)
4712{
4713 int cpu;
4714 int i;
4715
4716 /* Clamp it to [0:100] seconds interval. */
4717 if (rcu_delay_page_cache_fill_msec < 0 ||
4718 rcu_delay_page_cache_fill_msec > 100 * MSEC_PER_SEC) {
4719
4720 rcu_delay_page_cache_fill_msec =
4721 clamp(rcu_delay_page_cache_fill_msec, 0,
4722 (int) (100 * MSEC_PER_SEC));
4723
4724 pr_info("Adjusting rcutree.rcu_delay_page_cache_fill_msec to %d ms.\n",
4725 rcu_delay_page_cache_fill_msec);
4726 }
4727
4728 for_each_possible_cpu(cpu) {
4729 struct kfree_rcu_cpu *krcp = per_cpu_ptr(&krc, cpu);
4730
4731 for (i = 0; i < KFREE_N_BATCHES; i++) {
4732 INIT_RCU_WORK(&krcp->krw_arr[i].rcu_work, kfree_rcu_work);
4733 krcp->krw_arr[i].krcp = krcp;
4734 }
4735
4736 INIT_DELAYED_WORK(&krcp->monitor_work, kfree_rcu_monitor);
4737 INIT_DELAYED_WORK(&krcp->page_cache_work, fill_page_cache_func);
4738 krcp->initialized = true;
4739 }
4740 if (register_shrinker(&kfree_rcu_shrinker))
4741 pr_err("Failed to register kfree_rcu() shrinker!\n");
4742}
4743
4744void __init rcu_init(void)
4745{
4746 int cpu;
4747
4748 rcu_early_boot_tests();
4749
4750 kfree_rcu_batch_init();
4751 rcu_bootup_announce();
4752 rcu_init_geometry();
4753 rcu_init_one();
4754 if (dump_tree)
4755 rcu_dump_rcu_node_tree();
4756 if (use_softirq)
4757 open_softirq(RCU_SOFTIRQ, rcu_core_si);
4758
4759 /*
4760 * We don't need protection against CPU-hotplug here because
4761 * this is called early in boot, before either interrupts
4762 * or the scheduler are operational.
4763 */
4764 pm_notifier(rcu_pm_notify, 0);
4765 for_each_online_cpu(cpu) {
4766 rcutree_prepare_cpu(cpu);
4767 rcu_cpu_starting(cpu);
4768 rcutree_online_cpu(cpu);
4769 }
4770
4771 /* Create workqueue for Tree SRCU and for expedited GPs. */
4772 rcu_gp_wq = alloc_workqueue("rcu_gp", WQ_MEM_RECLAIM, 0);
4773 WARN_ON(!rcu_gp_wq);
4774 rcu_par_gp_wq = alloc_workqueue("rcu_par_gp", WQ_MEM_RECLAIM, 0);
4775 WARN_ON(!rcu_par_gp_wq);
4776
4777 /* Fill in default value for rcutree.qovld boot parameter. */
4778 /* -After- the rcu_node ->lock fields are initialized! */
4779 if (qovld < 0)
4780 qovld_calc = DEFAULT_RCU_QOVLD_MULT * qhimark;
4781 else
4782 qovld_calc = qovld;
4783}
4784
4785#include "tree_stall.h"
4786#include "tree_exp.h"
4787#include "tree_plugin.h"