~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

TOMOYO Linux Cross Reference
Linux/kernel/rcu/tree_plugin.h

Version: ~ [ linux-5.1-rc5 ] ~ [ linux-5.0.7 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.34 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.111 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.168 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.178 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.138 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.65 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Read-Copy Update mechanism for mutual exclusion (tree-based version)
  3  * Internal non-public definitions that provide either classic
  4  * or preemptible semantics.
  5  *
  6  * This program is free software; you can redistribute it and/or modify
  7  * it under the terms of the GNU General Public License as published by
  8  * the Free Software Foundation; either version 2 of the License, or
  9  * (at your option) any later version.
 10  *
 11  * This program is distributed in the hope that it will be useful,
 12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 14  * GNU General Public License for more details.
 15  *
 16  * You should have received a copy of the GNU General Public License
 17  * along with this program; if not, you can access it online at
 18  * http://www.gnu.org/licenses/gpl-2.0.html.
 19  *
 20  * Copyright Red Hat, 2009
 21  * Copyright IBM Corporation, 2009
 22  *
 23  * Author: Ingo Molnar <mingo@elte.hu>
 24  *         Paul E. McKenney <paulmck@linux.vnet.ibm.com>
 25  */
 26 
 27 #include <linux/delay.h>
 28 #include <linux/gfp.h>
 29 #include <linux/oom.h>
 30 #include <linux/sched/debug.h>
 31 #include <linux/smpboot.h>
 32 #include <linux/sched/isolation.h>
 33 #include <uapi/linux/sched/types.h>
 34 #include "../time/tick-internal.h"
 35 
 36 #ifdef CONFIG_RCU_BOOST
 37 
 38 #include "../locking/rtmutex_common.h"
 39 
 40 /*
 41  * Control variables for per-CPU and per-rcu_node kthreads.
 42  */
 43 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
 44 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
 45 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
 46 DEFINE_PER_CPU(char, rcu_cpu_has_work);
 47 
 48 #else /* #ifdef CONFIG_RCU_BOOST */
 49 
 50 /*
 51  * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
 52  * all uses are in dead code.  Provide a definition to keep the compiler
 53  * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
 54  * This probably needs to be excluded from -rt builds.
 55  */
 56 #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
 57 #define rt_mutex_futex_unlock(x) WARN_ON_ONCE(1)
 58 
 59 #endif /* #else #ifdef CONFIG_RCU_BOOST */
 60 
 61 #ifdef CONFIG_RCU_NOCB_CPU
 62 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
 63 static bool __read_mostly rcu_nocb_poll;    /* Offload kthread are to poll. */
 64 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
 65 
 66 /*
 67  * Check the RCU kernel configuration parameters and print informative
 68  * messages about anything out of the ordinary.
 69  */
 70 static void __init rcu_bootup_announce_oddness(void)
 71 {
 72         if (IS_ENABLED(CONFIG_RCU_TRACE))
 73                 pr_info("\tRCU event tracing is enabled.\n");
 74         if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
 75             (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
 76                 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
 77                         RCU_FANOUT);
 78         if (rcu_fanout_exact)
 79                 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
 80         if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
 81                 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
 82         if (IS_ENABLED(CONFIG_PROVE_RCU))
 83                 pr_info("\tRCU lockdep checking is enabled.\n");
 84         if (RCU_NUM_LVLS >= 4)
 85                 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
 86         if (RCU_FANOUT_LEAF != 16)
 87                 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
 88                         RCU_FANOUT_LEAF);
 89         if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
 90                 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
 91                         rcu_fanout_leaf);
 92         if (nr_cpu_ids != NR_CPUS)
 93                 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
 94 #ifdef CONFIG_RCU_BOOST
 95         pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
 96                 kthread_prio, CONFIG_RCU_BOOST_DELAY);
 97 #endif
 98         if (blimit != DEFAULT_RCU_BLIMIT)
 99                 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
100         if (qhimark != DEFAULT_RCU_QHIMARK)
101                 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
102         if (qlowmark != DEFAULT_RCU_QLOMARK)
103                 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
104         if (jiffies_till_first_fqs != ULONG_MAX)
105                 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
106         if (jiffies_till_next_fqs != ULONG_MAX)
107                 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
108         if (jiffies_till_sched_qs != ULONG_MAX)
109                 pr_info("\tBoot-time adjustment of scheduler-enlistment delay to %ld jiffies.\n", jiffies_till_sched_qs);
110         if (rcu_kick_kthreads)
111                 pr_info("\tKick kthreads if too-long grace period.\n");
112         if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
113                 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
114         if (gp_preinit_delay)
115                 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
116         if (gp_init_delay)
117                 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
118         if (gp_cleanup_delay)
119                 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
120         if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
121                 pr_info("\tRCU debug extended QS entry/exit.\n");
122         rcupdate_announce_bootup_oddness();
123 }
124 
125 #ifdef CONFIG_PREEMPT_RCU
126 
127 static void rcu_report_exp_rnp(struct rcu_node *rnp, bool wake);
128 static void rcu_read_unlock_special(struct task_struct *t);
129 
130 /*
131  * Tell them what RCU they are running.
132  */
133 static void __init rcu_bootup_announce(void)
134 {
135         pr_info("Preemptible hierarchical RCU implementation.\n");
136         rcu_bootup_announce_oddness();
137 }
138 
139 /* Flags for rcu_preempt_ctxt_queue() decision table. */
140 #define RCU_GP_TASKS    0x8
141 #define RCU_EXP_TASKS   0x4
142 #define RCU_GP_BLKD     0x2
143 #define RCU_EXP_BLKD    0x1
144 
145 /*
146  * Queues a task preempted within an RCU-preempt read-side critical
147  * section into the appropriate location within the ->blkd_tasks list,
148  * depending on the states of any ongoing normal and expedited grace
149  * periods.  The ->gp_tasks pointer indicates which element the normal
150  * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
151  * indicates which element the expedited grace period is waiting on (again,
152  * NULL if none).  If a grace period is waiting on a given element in the
153  * ->blkd_tasks list, it also waits on all subsequent elements.  Thus,
154  * adding a task to the tail of the list blocks any grace period that is
155  * already waiting on one of the elements.  In contrast, adding a task
156  * to the head of the list won't block any grace period that is already
157  * waiting on one of the elements.
158  *
159  * This queuing is imprecise, and can sometimes make an ongoing grace
160  * period wait for a task that is not strictly speaking blocking it.
161  * Given the choice, we needlessly block a normal grace period rather than
162  * blocking an expedited grace period.
163  *
164  * Note that an endless sequence of expedited grace periods still cannot
165  * indefinitely postpone a normal grace period.  Eventually, all of the
166  * fixed number of preempted tasks blocking the normal grace period that are
167  * not also blocking the expedited grace period will resume and complete
168  * their RCU read-side critical sections.  At that point, the ->gp_tasks
169  * pointer will equal the ->exp_tasks pointer, at which point the end of
170  * the corresponding expedited grace period will also be the end of the
171  * normal grace period.
172  */
173 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
174         __releases(rnp->lock) /* But leaves rrupts disabled. */
175 {
176         int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
177                          (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
178                          (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
179                          (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
180         struct task_struct *t = current;
181 
182         raw_lockdep_assert_held_rcu_node(rnp);
183         WARN_ON_ONCE(rdp->mynode != rnp);
184         WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
185         /* RCU better not be waiting on newly onlined CPUs! */
186         WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
187                      rdp->grpmask);
188 
189         /*
190          * Decide where to queue the newly blocked task.  In theory,
191          * this could be an if-statement.  In practice, when I tried
192          * that, it was quite messy.
193          */
194         switch (blkd_state) {
195         case 0:
196         case                RCU_EXP_TASKS:
197         case                RCU_EXP_TASKS + RCU_GP_BLKD:
198         case RCU_GP_TASKS:
199         case RCU_GP_TASKS + RCU_EXP_TASKS:
200 
201                 /*
202                  * Blocking neither GP, or first task blocking the normal
203                  * GP but not blocking the already-waiting expedited GP.
204                  * Queue at the head of the list to avoid unnecessarily
205                  * blocking the already-waiting GPs.
206                  */
207                 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
208                 break;
209 
210         case                                              RCU_EXP_BLKD:
211         case                                RCU_GP_BLKD:
212         case                                RCU_GP_BLKD + RCU_EXP_BLKD:
213         case RCU_GP_TASKS +                               RCU_EXP_BLKD:
214         case RCU_GP_TASKS +                 RCU_GP_BLKD + RCU_EXP_BLKD:
215         case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
216 
217                 /*
218                  * First task arriving that blocks either GP, or first task
219                  * arriving that blocks the expedited GP (with the normal
220                  * GP already waiting), or a task arriving that blocks
221                  * both GPs with both GPs already waiting.  Queue at the
222                  * tail of the list to avoid any GP waiting on any of the
223                  * already queued tasks that are not blocking it.
224                  */
225                 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
226                 break;
227 
228         case                RCU_EXP_TASKS +               RCU_EXP_BLKD:
229         case                RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
230         case RCU_GP_TASKS + RCU_EXP_TASKS +               RCU_EXP_BLKD:
231 
232                 /*
233                  * Second or subsequent task blocking the expedited GP.
234                  * The task either does not block the normal GP, or is the
235                  * first task blocking the normal GP.  Queue just after
236                  * the first task blocking the expedited GP.
237                  */
238                 list_add(&t->rcu_node_entry, rnp->exp_tasks);
239                 break;
240 
241         case RCU_GP_TASKS +                 RCU_GP_BLKD:
242         case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
243 
244                 /*
245                  * Second or subsequent task blocking the normal GP.
246                  * The task does not block the expedited GP. Queue just
247                  * after the first task blocking the normal GP.
248                  */
249                 list_add(&t->rcu_node_entry, rnp->gp_tasks);
250                 break;
251 
252         default:
253 
254                 /* Yet another exercise in excessive paranoia. */
255                 WARN_ON_ONCE(1);
256                 break;
257         }
258 
259         /*
260          * We have now queued the task.  If it was the first one to
261          * block either grace period, update the ->gp_tasks and/or
262          * ->exp_tasks pointers, respectively, to reference the newly
263          * blocked tasks.
264          */
265         if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
266                 rnp->gp_tasks = &t->rcu_node_entry;
267                 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
268         }
269         if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
270                 rnp->exp_tasks = &t->rcu_node_entry;
271         WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
272                      !(rnp->qsmask & rdp->grpmask));
273         WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
274                      !(rnp->expmask & rdp->grpmask));
275         raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
276 
277         /*
278          * Report the quiescent state for the expedited GP.  This expedited
279          * GP should not be able to end until we report, so there should be
280          * no need to check for a subsequent expedited GP.  (Though we are
281          * still in a quiescent state in any case.)
282          */
283         if (blkd_state & RCU_EXP_BLKD && rdp->deferred_qs)
284                 rcu_report_exp_rdp(rdp);
285         else
286                 WARN_ON_ONCE(rdp->deferred_qs);
287 }
288 
289 /*
290  * Record a preemptible-RCU quiescent state for the specified CPU.
291  * Note that this does not necessarily mean that the task currently running
292  * on the CPU is in a quiescent state:  Instead, it means that the current
293  * grace period need not wait on any RCU read-side critical section that
294  * starts later on this CPU.  It also means that if the current task is
295  * in an RCU read-side critical section, it has already added itself to
296  * some leaf rcu_node structure's ->blkd_tasks list.  In addition to the
297  * current task, there might be any number of other tasks blocked while
298  * in an RCU read-side critical section.
299  *
300  * Callers to this function must disable preemption.
301  */
302 static void rcu_qs(void)
303 {
304         RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!\n");
305         if (__this_cpu_read(rcu_data.cpu_no_qs.s)) {
306                 trace_rcu_grace_period(TPS("rcu_preempt"),
307                                        __this_cpu_read(rcu_data.gp_seq),
308                                        TPS("cpuqs"));
309                 __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
310                 barrier(); /* Coordinate with rcu_flavor_check_callbacks(). */
311                 current->rcu_read_unlock_special.b.need_qs = false;
312         }
313 }
314 
315 /*
316  * We have entered the scheduler, and the current task might soon be
317  * context-switched away from.  If this task is in an RCU read-side
318  * critical section, we will no longer be able to rely on the CPU to
319  * record that fact, so we enqueue the task on the blkd_tasks list.
320  * The task will dequeue itself when it exits the outermost enclosing
321  * RCU read-side critical section.  Therefore, the current grace period
322  * cannot be permitted to complete until the blkd_tasks list entries
323  * predating the current grace period drain, in other words, until
324  * rnp->gp_tasks becomes NULL.
325  *
326  * Caller must disable interrupts.
327  */
328 void rcu_note_context_switch(bool preempt)
329 {
330         struct task_struct *t = current;
331         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
332         struct rcu_node *rnp;
333 
334         barrier(); /* Avoid RCU read-side critical sections leaking down. */
335         trace_rcu_utilization(TPS("Start context switch"));
336         lockdep_assert_irqs_disabled();
337         WARN_ON_ONCE(!preempt && t->rcu_read_lock_nesting > 0);
338         if (t->rcu_read_lock_nesting > 0 &&
339             !t->rcu_read_unlock_special.b.blocked) {
340 
341                 /* Possibly blocking in an RCU read-side critical section. */
342                 rnp = rdp->mynode;
343                 raw_spin_lock_rcu_node(rnp);
344                 t->rcu_read_unlock_special.b.blocked = true;
345                 t->rcu_blocked_node = rnp;
346 
347                 /*
348                  * Verify the CPU's sanity, trace the preemption, and
349                  * then queue the task as required based on the states
350                  * of any ongoing and expedited grace periods.
351                  */
352                 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
353                 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
354                 trace_rcu_preempt_task(rcu_state.name,
355                                        t->pid,
356                                        (rnp->qsmask & rdp->grpmask)
357                                        ? rnp->gp_seq
358                                        : rcu_seq_snap(&rnp->gp_seq));
359                 rcu_preempt_ctxt_queue(rnp, rdp);
360         } else if (t->rcu_read_lock_nesting < 0 &&
361                    t->rcu_read_unlock_special.s) {
362 
363                 /*
364                  * Complete exit from RCU read-side critical section on
365                  * behalf of preempted instance of __rcu_read_unlock().
366                  */
367                 rcu_read_unlock_special(t);
368                 rcu_preempt_deferred_qs(t);
369         } else {
370                 rcu_preempt_deferred_qs(t);
371         }
372 
373         /*
374          * Either we were not in an RCU read-side critical section to
375          * begin with, or we have now recorded that critical section
376          * globally.  Either way, we can now note a quiescent state
377          * for this CPU.  Again, if we were in an RCU read-side critical
378          * section, and if that critical section was blocking the current
379          * grace period, then the fact that the task has been enqueued
380          * means that we continue to block the current grace period.
381          */
382         rcu_qs();
383         if (rdp->deferred_qs)
384                 rcu_report_exp_rdp(rdp);
385         trace_rcu_utilization(TPS("End context switch"));
386         barrier(); /* Avoid RCU read-side critical sections leaking up. */
387 }
388 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
389 
390 /*
391  * Check for preempted RCU readers blocking the current grace period
392  * for the specified rcu_node structure.  If the caller needs a reliable
393  * answer, it must hold the rcu_node's ->lock.
394  */
395 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
396 {
397         return rnp->gp_tasks != NULL;
398 }
399 
400 /*
401  * Preemptible RCU implementation for rcu_read_lock().
402  * Just increment ->rcu_read_lock_nesting, shared state will be updated
403  * if we block.
404  */
405 void __rcu_read_lock(void)
406 {
407         current->rcu_read_lock_nesting++;
408         barrier();  /* critical section after entry code. */
409 }
410 EXPORT_SYMBOL_GPL(__rcu_read_lock);
411 
412 /*
413  * Preemptible RCU implementation for rcu_read_unlock().
414  * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
415  * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
416  * invoke rcu_read_unlock_special() to clean up after a context switch
417  * in an RCU read-side critical section and other special cases.
418  */
419 void __rcu_read_unlock(void)
420 {
421         struct task_struct *t = current;
422 
423         if (t->rcu_read_lock_nesting != 1) {
424                 --t->rcu_read_lock_nesting;
425         } else {
426                 barrier();  /* critical section before exit code. */
427                 t->rcu_read_lock_nesting = INT_MIN;
428                 barrier();  /* assign before ->rcu_read_unlock_special load */
429                 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
430                         rcu_read_unlock_special(t);
431                 barrier();  /* ->rcu_read_unlock_special load before assign */
432                 t->rcu_read_lock_nesting = 0;
433         }
434 #ifdef CONFIG_PROVE_LOCKING
435         {
436                 int rrln = READ_ONCE(t->rcu_read_lock_nesting);
437 
438                 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
439         }
440 #endif /* #ifdef CONFIG_PROVE_LOCKING */
441 }
442 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
443 
444 /*
445  * Advance a ->blkd_tasks-list pointer to the next entry, instead
446  * returning NULL if at the end of the list.
447  */
448 static struct list_head *rcu_next_node_entry(struct task_struct *t,
449                                              struct rcu_node *rnp)
450 {
451         struct list_head *np;
452 
453         np = t->rcu_node_entry.next;
454         if (np == &rnp->blkd_tasks)
455                 np = NULL;
456         return np;
457 }
458 
459 /*
460  * Return true if the specified rcu_node structure has tasks that were
461  * preempted within an RCU read-side critical section.
462  */
463 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
464 {
465         return !list_empty(&rnp->blkd_tasks);
466 }
467 
468 /*
469  * Report deferred quiescent states.  The deferral time can
470  * be quite short, for example, in the case of the call from
471  * rcu_read_unlock_special().
472  */
473 static void
474 rcu_preempt_deferred_qs_irqrestore(struct task_struct *t, unsigned long flags)
475 {
476         bool empty_exp;
477         bool empty_norm;
478         bool empty_exp_now;
479         struct list_head *np;
480         bool drop_boost_mutex = false;
481         struct rcu_data *rdp;
482         struct rcu_node *rnp;
483         union rcu_special special;
484 
485         /*
486          * If RCU core is waiting for this CPU to exit its critical section,
487          * report the fact that it has exited.  Because irqs are disabled,
488          * t->rcu_read_unlock_special cannot change.
489          */
490         special = t->rcu_read_unlock_special;
491         rdp = this_cpu_ptr(&rcu_data);
492         if (!special.s && !rdp->deferred_qs) {
493                 local_irq_restore(flags);
494                 return;
495         }
496         if (special.b.need_qs) {
497                 rcu_qs();
498                 t->rcu_read_unlock_special.b.need_qs = false;
499                 if (!t->rcu_read_unlock_special.s && !rdp->deferred_qs) {
500                         local_irq_restore(flags);
501                         return;
502                 }
503         }
504 
505         /*
506          * Respond to a request by an expedited grace period for a
507          * quiescent state from this CPU.  Note that requests from
508          * tasks are handled when removing the task from the
509          * blocked-tasks list below.
510          */
511         if (rdp->deferred_qs) {
512                 rcu_report_exp_rdp(rdp);
513                 if (!t->rcu_read_unlock_special.s) {
514                         local_irq_restore(flags);
515                         return;
516                 }
517         }
518 
519         /* Clean up if blocked during RCU read-side critical section. */
520         if (special.b.blocked) {
521                 t->rcu_read_unlock_special.b.blocked = false;
522 
523                 /*
524                  * Remove this task from the list it blocked on.  The task
525                  * now remains queued on the rcu_node corresponding to the
526                  * CPU it first blocked on, so there is no longer any need
527                  * to loop.  Retain a WARN_ON_ONCE() out of sheer paranoia.
528                  */
529                 rnp = t->rcu_blocked_node;
530                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
531                 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
532                 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
533                 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
534                 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
535                              (!empty_norm || rnp->qsmask));
536                 empty_exp = sync_rcu_preempt_exp_done(rnp);
537                 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
538                 np = rcu_next_node_entry(t, rnp);
539                 list_del_init(&t->rcu_node_entry);
540                 t->rcu_blocked_node = NULL;
541                 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
542                                                 rnp->gp_seq, t->pid);
543                 if (&t->rcu_node_entry == rnp->gp_tasks)
544                         rnp->gp_tasks = np;
545                 if (&t->rcu_node_entry == rnp->exp_tasks)
546                         rnp->exp_tasks = np;
547                 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
548                         /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
549                         drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
550                         if (&t->rcu_node_entry == rnp->boost_tasks)
551                                 rnp->boost_tasks = np;
552                 }
553 
554                 /*
555                  * If this was the last task on the current list, and if
556                  * we aren't waiting on any CPUs, report the quiescent state.
557                  * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
558                  * so we must take a snapshot of the expedited state.
559                  */
560                 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
561                 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
562                         trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
563                                                          rnp->gp_seq,
564                                                          0, rnp->qsmask,
565                                                          rnp->level,
566                                                          rnp->grplo,
567                                                          rnp->grphi,
568                                                          !!rnp->gp_tasks);
569                         rcu_report_unblock_qs_rnp(rnp, flags);
570                 } else {
571                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
572                 }
573 
574                 /* Unboost if we were boosted. */
575                 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
576                         rt_mutex_futex_unlock(&rnp->boost_mtx);
577 
578                 /*
579                  * If this was the last task on the expedited lists,
580                  * then we need to report up the rcu_node hierarchy.
581                  */
582                 if (!empty_exp && empty_exp_now)
583                         rcu_report_exp_rnp(rnp, true);
584         } else {
585                 local_irq_restore(flags);
586         }
587 }
588 
589 /*
590  * Is a deferred quiescent-state pending, and are we also not in
591  * an RCU read-side critical section?  It is the caller's responsibility
592  * to ensure it is otherwise safe to report any deferred quiescent
593  * states.  The reason for this is that it is safe to report a
594  * quiescent state during context switch even though preemption
595  * is disabled.  This function cannot be expected to understand these
596  * nuances, so the caller must handle them.
597  */
598 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
599 {
600         return (this_cpu_ptr(&rcu_data)->deferred_qs ||
601                 READ_ONCE(t->rcu_read_unlock_special.s)) &&
602                t->rcu_read_lock_nesting <= 0;
603 }
604 
605 /*
606  * Report a deferred quiescent state if needed and safe to do so.
607  * As with rcu_preempt_need_deferred_qs(), "safe" involves only
608  * not being in an RCU read-side critical section.  The caller must
609  * evaluate safety in terms of interrupt, softirq, and preemption
610  * disabling.
611  */
612 static void rcu_preempt_deferred_qs(struct task_struct *t)
613 {
614         unsigned long flags;
615         bool couldrecurse = t->rcu_read_lock_nesting >= 0;
616 
617         if (!rcu_preempt_need_deferred_qs(t))
618                 return;
619         if (couldrecurse)
620                 t->rcu_read_lock_nesting -= INT_MIN;
621         local_irq_save(flags);
622         rcu_preempt_deferred_qs_irqrestore(t, flags);
623         if (couldrecurse)
624                 t->rcu_read_lock_nesting += INT_MIN;
625 }
626 
627 /*
628  * Handle special cases during rcu_read_unlock(), such as needing to
629  * notify RCU core processing or task having blocked during the RCU
630  * read-side critical section.
631  */
632 static void rcu_read_unlock_special(struct task_struct *t)
633 {
634         unsigned long flags;
635         bool preempt_bh_were_disabled =
636                         !!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK));
637         bool irqs_were_disabled;
638 
639         /* NMI handlers cannot block and cannot safely manipulate state. */
640         if (in_nmi())
641                 return;
642 
643         local_irq_save(flags);
644         irqs_were_disabled = irqs_disabled_flags(flags);
645         if ((preempt_bh_were_disabled || irqs_were_disabled) &&
646             t->rcu_read_unlock_special.b.blocked) {
647                 /* Need to defer quiescent state until everything is enabled. */
648                 raise_softirq_irqoff(RCU_SOFTIRQ);
649                 local_irq_restore(flags);
650                 return;
651         }
652         rcu_preempt_deferred_qs_irqrestore(t, flags);
653 }
654 
655 /*
656  * Dump detailed information for all tasks blocking the current RCU
657  * grace period on the specified rcu_node structure.
658  */
659 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
660 {
661         unsigned long flags;
662         struct task_struct *t;
663 
664         raw_spin_lock_irqsave_rcu_node(rnp, flags);
665         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
666                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
667                 return;
668         }
669         t = list_entry(rnp->gp_tasks->prev,
670                        struct task_struct, rcu_node_entry);
671         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
672                 /*
673                  * We could be printing a lot while holding a spinlock.
674                  * Avoid triggering hard lockup.
675                  */
676                 touch_nmi_watchdog();
677                 sched_show_task(t);
678         }
679         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
680 }
681 
682 /*
683  * Dump detailed information for all tasks blocking the current RCU
684  * grace period.
685  */
686 static void rcu_print_detail_task_stall(void)
687 {
688         struct rcu_node *rnp = rcu_get_root();
689 
690         rcu_print_detail_task_stall_rnp(rnp);
691         rcu_for_each_leaf_node(rnp)
692                 rcu_print_detail_task_stall_rnp(rnp);
693 }
694 
695 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
696 {
697         pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
698                rnp->level, rnp->grplo, rnp->grphi);
699 }
700 
701 static void rcu_print_task_stall_end(void)
702 {
703         pr_cont("\n");
704 }
705 
706 /*
707  * Scan the current list of tasks blocked within RCU read-side critical
708  * sections, printing out the tid of each.
709  */
710 static int rcu_print_task_stall(struct rcu_node *rnp)
711 {
712         struct task_struct *t;
713         int ndetected = 0;
714 
715         if (!rcu_preempt_blocked_readers_cgp(rnp))
716                 return 0;
717         rcu_print_task_stall_begin(rnp);
718         t = list_entry(rnp->gp_tasks->prev,
719                        struct task_struct, rcu_node_entry);
720         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
721                 pr_cont(" P%d", t->pid);
722                 ndetected++;
723         }
724         rcu_print_task_stall_end();
725         return ndetected;
726 }
727 
728 /*
729  * Scan the current list of tasks blocked within RCU read-side critical
730  * sections, printing out the tid of each that is blocking the current
731  * expedited grace period.
732  */
733 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
734 {
735         struct task_struct *t;
736         int ndetected = 0;
737 
738         if (!rnp->exp_tasks)
739                 return 0;
740         t = list_entry(rnp->exp_tasks->prev,
741                        struct task_struct, rcu_node_entry);
742         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
743                 pr_cont(" P%d", t->pid);
744                 ndetected++;
745         }
746         return ndetected;
747 }
748 
749 /*
750  * Check that the list of blocked tasks for the newly completed grace
751  * period is in fact empty.  It is a serious bug to complete a grace
752  * period that still has RCU readers blocked!  This function must be
753  * invoked -before- updating this rnp's ->gp_seq, and the rnp's ->lock
754  * must be held by the caller.
755  *
756  * Also, if there are blocked tasks on the list, they automatically
757  * block the newly created grace period, so set up ->gp_tasks accordingly.
758  */
759 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
760 {
761         struct task_struct *t;
762 
763         RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
764         if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
765                 dump_blkd_tasks(rnp, 10);
766         if (rcu_preempt_has_tasks(rnp) &&
767             (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
768                 rnp->gp_tasks = rnp->blkd_tasks.next;
769                 t = container_of(rnp->gp_tasks, struct task_struct,
770                                  rcu_node_entry);
771                 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
772                                                 rnp->gp_seq, t->pid);
773         }
774         WARN_ON_ONCE(rnp->qsmask);
775 }
776 
777 /*
778  * Check for a quiescent state from the current CPU.  When a task blocks,
779  * the task is recorded in the corresponding CPU's rcu_node structure,
780  * which is checked elsewhere.
781  *
782  * Caller must disable hard irqs.
783  */
784 static void rcu_flavor_check_callbacks(int user)
785 {
786         struct task_struct *t = current;
787 
788         if (user || rcu_is_cpu_rrupt_from_idle()) {
789                 rcu_note_voluntary_context_switch(current);
790         }
791         if (t->rcu_read_lock_nesting > 0 ||
792             (preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK))) {
793                 /* No QS, force context switch if deferred. */
794                 if (rcu_preempt_need_deferred_qs(t)) {
795                         set_tsk_need_resched(t);
796                         set_preempt_need_resched();
797                 }
798         } else if (rcu_preempt_need_deferred_qs(t)) {
799                 rcu_preempt_deferred_qs(t); /* Report deferred QS. */
800                 return;
801         } else if (!t->rcu_read_lock_nesting) {
802                 rcu_qs(); /* Report immediate QS. */
803                 return;
804         }
805 
806         /* If GP is oldish, ask for help from rcu_read_unlock_special(). */
807         if (t->rcu_read_lock_nesting > 0 &&
808             __this_cpu_read(rcu_data.core_needs_qs) &&
809             __this_cpu_read(rcu_data.cpu_no_qs.b.norm) &&
810             !t->rcu_read_unlock_special.b.need_qs &&
811             time_after(jiffies, rcu_state.gp_start + HZ))
812                 t->rcu_read_unlock_special.b.need_qs = true;
813 }
814 
815 /**
816  * synchronize_rcu - wait until a grace period has elapsed.
817  *
818  * Control will return to the caller some time after a full grace
819  * period has elapsed, in other words after all currently executing RCU
820  * read-side critical sections have completed.  Note, however, that
821  * upon return from synchronize_rcu(), the caller might well be executing
822  * concurrently with new RCU read-side critical sections that began while
823  * synchronize_rcu() was waiting.  RCU read-side critical sections are
824  * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
825  * In addition, regions of code across which interrupts, preemption, or
826  * softirqs have been disabled also serve as RCU read-side critical
827  * sections.  This includes hardware interrupt handlers, softirq handlers,
828  * and NMI handlers.
829  *
830  * Note that this guarantee implies further memory-ordering guarantees.
831  * On systems with more than one CPU, when synchronize_rcu() returns,
832  * each CPU is guaranteed to have executed a full memory barrier since
833  * the end of its last RCU read-side critical section whose beginning
834  * preceded the call to synchronize_rcu().  In addition, each CPU having
835  * an RCU read-side critical section that extends beyond the return from
836  * synchronize_rcu() is guaranteed to have executed a full memory barrier
837  * after the beginning of synchronize_rcu() and before the beginning of
838  * that RCU read-side critical section.  Note that these guarantees include
839  * CPUs that are offline, idle, or executing in user mode, as well as CPUs
840  * that are executing in the kernel.
841  *
842  * Furthermore, if CPU A invoked synchronize_rcu(), which returned
843  * to its caller on CPU B, then both CPU A and CPU B are guaranteed
844  * to have executed a full memory barrier during the execution of
845  * synchronize_rcu() -- even if CPU A and CPU B are the same CPU (but
846  * again only if the system has more than one CPU).
847  */
848 void synchronize_rcu(void)
849 {
850         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
851                          lock_is_held(&rcu_lock_map) ||
852                          lock_is_held(&rcu_sched_lock_map),
853                          "Illegal synchronize_rcu() in RCU read-side critical section");
854         if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
855                 return;
856         if (rcu_gp_is_expedited())
857                 synchronize_rcu_expedited();
858         else
859                 wait_rcu_gp(call_rcu);
860 }
861 EXPORT_SYMBOL_GPL(synchronize_rcu);
862 
863 /*
864  * Check for a task exiting while in a preemptible-RCU read-side
865  * critical section, clean up if so.  No need to issue warnings,
866  * as debug_check_no_locks_held() already does this if lockdep
867  * is enabled.
868  */
869 void exit_rcu(void)
870 {
871         struct task_struct *t = current;
872 
873         if (likely(list_empty(&current->rcu_node_entry)))
874                 return;
875         t->rcu_read_lock_nesting = 1;
876         barrier();
877         t->rcu_read_unlock_special.b.blocked = true;
878         __rcu_read_unlock();
879         rcu_preempt_deferred_qs(current);
880 }
881 
882 /*
883  * Dump the blocked-tasks state, but limit the list dump to the
884  * specified number of elements.
885  */
886 static void
887 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
888 {
889         int cpu;
890         int i;
891         struct list_head *lhp;
892         bool onl;
893         struct rcu_data *rdp;
894         struct rcu_node *rnp1;
895 
896         raw_lockdep_assert_held_rcu_node(rnp);
897         pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
898                 __func__, rnp->grplo, rnp->grphi, rnp->level,
899                 (long)rnp->gp_seq, (long)rnp->completedqs);
900         for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
901                 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
902                         __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
903         pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
904                 __func__, rnp->gp_tasks, rnp->boost_tasks, rnp->exp_tasks);
905         pr_info("%s: ->blkd_tasks", __func__);
906         i = 0;
907         list_for_each(lhp, &rnp->blkd_tasks) {
908                 pr_cont(" %p", lhp);
909                 if (++i >= 10)
910                         break;
911         }
912         pr_cont("\n");
913         for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
914                 rdp = per_cpu_ptr(&rcu_data, cpu);
915                 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
916                 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
917                         cpu, ".o"[onl],
918                         (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
919                         (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
920         }
921 }
922 
923 #else /* #ifdef CONFIG_PREEMPT_RCU */
924 
925 /*
926  * Tell them what RCU they are running.
927  */
928 static void __init rcu_bootup_announce(void)
929 {
930         pr_info("Hierarchical RCU implementation.\n");
931         rcu_bootup_announce_oddness();
932 }
933 
934 /*
935  * Note a quiescent state for PREEMPT=n.  Because we do not need to know
936  * how many quiescent states passed, just if there was at least one since
937  * the start of the grace period, this just sets a flag.  The caller must
938  * have disabled preemption.
939  */
940 static void rcu_qs(void)
941 {
942         RCU_LOCKDEP_WARN(preemptible(), "rcu_qs() invoked with preemption enabled!!!");
943         if (!__this_cpu_read(rcu_data.cpu_no_qs.s))
944                 return;
945         trace_rcu_grace_period(TPS("rcu_sched"),
946                                __this_cpu_read(rcu_data.gp_seq), TPS("cpuqs"));
947         __this_cpu_write(rcu_data.cpu_no_qs.b.norm, false);
948         if (!__this_cpu_read(rcu_data.cpu_no_qs.b.exp))
949                 return;
950         __this_cpu_write(rcu_data.cpu_no_qs.b.exp, false);
951         rcu_report_exp_rdp(this_cpu_ptr(&rcu_data));
952 }
953 
954 /*
955  * Register an urgently needed quiescent state.  If there is an
956  * emergency, invoke rcu_momentary_dyntick_idle() to do a heavy-weight
957  * dyntick-idle quiescent state visible to other CPUs, which will in
958  * some cases serve for expedited as well as normal grace periods.
959  * Either way, register a lightweight quiescent state.
960  *
961  * The barrier() calls are redundant in the common case when this is
962  * called externally, but just in case this is called from within this
963  * file.
964  *
965  */
966 void rcu_all_qs(void)
967 {
968         unsigned long flags;
969 
970         if (!raw_cpu_read(rcu_data.rcu_urgent_qs))
971                 return;
972         preempt_disable();
973         /* Load rcu_urgent_qs before other flags. */
974         if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs))) {
975                 preempt_enable();
976                 return;
977         }
978         this_cpu_write(rcu_data.rcu_urgent_qs, false);
979         barrier(); /* Avoid RCU read-side critical sections leaking down. */
980         if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs))) {
981                 local_irq_save(flags);
982                 rcu_momentary_dyntick_idle();
983                 local_irq_restore(flags);
984         }
985         rcu_qs();
986         barrier(); /* Avoid RCU read-side critical sections leaking up. */
987         preempt_enable();
988 }
989 EXPORT_SYMBOL_GPL(rcu_all_qs);
990 
991 /*
992  * Note a PREEMPT=n context switch.  The caller must have disabled interrupts.
993  */
994 void rcu_note_context_switch(bool preempt)
995 {
996         barrier(); /* Avoid RCU read-side critical sections leaking down. */
997         trace_rcu_utilization(TPS("Start context switch"));
998         rcu_qs();
999         /* Load rcu_urgent_qs before other flags. */
1000         if (!smp_load_acquire(this_cpu_ptr(&rcu_data.rcu_urgent_qs)))
1001                 goto out;
1002         this_cpu_write(rcu_data.rcu_urgent_qs, false);
1003         if (unlikely(raw_cpu_read(rcu_data.rcu_need_heavy_qs)))
1004                 rcu_momentary_dyntick_idle();
1005         if (!preempt)
1006                 rcu_tasks_qs(current);
1007 out:
1008         trace_rcu_utilization(TPS("End context switch"));
1009         barrier(); /* Avoid RCU read-side critical sections leaking up. */
1010 }
1011 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
1012 
1013 /*
1014  * Because preemptible RCU does not exist, there are never any preempted
1015  * RCU readers.
1016  */
1017 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
1018 {
1019         return 0;
1020 }
1021 
1022 /*
1023  * Because there is no preemptible RCU, there can be no readers blocked.
1024  */
1025 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
1026 {
1027         return false;
1028 }
1029 
1030 /*
1031  * Because there is no preemptible RCU, there can be no deferred quiescent
1032  * states.
1033  */
1034 static bool rcu_preempt_need_deferred_qs(struct task_struct *t)
1035 {
1036         return false;
1037 }
1038 static void rcu_preempt_deferred_qs(struct task_struct *t) { }
1039 
1040 /*
1041  * Because preemptible RCU does not exist, we never have to check for
1042  * tasks blocked within RCU read-side critical sections.
1043  */
1044 static void rcu_print_detail_task_stall(void)
1045 {
1046 }
1047 
1048 /*
1049  * Because preemptible RCU does not exist, we never have to check for
1050  * tasks blocked within RCU read-side critical sections.
1051  */
1052 static int rcu_print_task_stall(struct rcu_node *rnp)
1053 {
1054         return 0;
1055 }
1056 
1057 /*
1058  * Because preemptible RCU does not exist, we never have to check for
1059  * tasks blocked within RCU read-side critical sections that are
1060  * blocking the current expedited grace period.
1061  */
1062 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
1063 {
1064         return 0;
1065 }
1066 
1067 /*
1068  * Because there is no preemptible RCU, there can be no readers blocked,
1069  * so there is no need to check for blocked tasks.  So check only for
1070  * bogus qsmask values.
1071  */
1072 static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1073 {
1074         WARN_ON_ONCE(rnp->qsmask);
1075 }
1076 
1077 /*
1078  * Check to see if this CPU is in a non-context-switch quiescent state
1079  * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1080  * Also schedule RCU core processing.
1081  *
1082  * This function must be called from hardirq context.  It is normally
1083  * invoked from the scheduling-clock interrupt.
1084  */
1085 static void rcu_flavor_check_callbacks(int user)
1086 {
1087         if (user || rcu_is_cpu_rrupt_from_idle()) {
1088 
1089                 /*
1090                  * Get here if this CPU took its interrupt from user
1091                  * mode or from the idle loop, and if this is not a
1092                  * nested interrupt.  In this case, the CPU is in
1093                  * a quiescent state, so note it.
1094                  *
1095                  * No memory barrier is required here because rcu_qs()
1096                  * references only CPU-local variables that other CPUs
1097                  * neither access nor modify, at least not while the
1098                  * corresponding CPU is online.
1099                  */
1100 
1101                 rcu_qs();
1102         }
1103 }
1104 
1105 /* PREEMPT=n implementation of synchronize_rcu(). */
1106 void synchronize_rcu(void)
1107 {
1108         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
1109                          lock_is_held(&rcu_lock_map) ||
1110                          lock_is_held(&rcu_sched_lock_map),
1111                          "Illegal synchronize_rcu() in RCU read-side critical section");
1112         if (rcu_blocking_is_gp())
1113                 return;
1114         if (rcu_gp_is_expedited())
1115                 synchronize_rcu_expedited();
1116         else
1117                 wait_rcu_gp(call_rcu);
1118 }
1119 EXPORT_SYMBOL_GPL(synchronize_rcu);
1120 
1121 /*
1122  * Because preemptible RCU does not exist, tasks cannot possibly exit
1123  * while in preemptible RCU read-side critical sections.
1124  */
1125 void exit_rcu(void)
1126 {
1127 }
1128 
1129 /*
1130  * Dump the guaranteed-empty blocked-tasks state.  Trust but verify.
1131  */
1132 static void
1133 dump_blkd_tasks(struct rcu_node *rnp, int ncheck)
1134 {
1135         WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
1136 }
1137 
1138 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
1139 
1140 #ifdef CONFIG_RCU_BOOST
1141 
1142 static void rcu_wake_cond(struct task_struct *t, int status)
1143 {
1144         /*
1145          * If the thread is yielding, only wake it when this
1146          * is invoked from idle
1147          */
1148         if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1149                 wake_up_process(t);
1150 }
1151 
1152 /*
1153  * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1154  * or ->boost_tasks, advancing the pointer to the next task in the
1155  * ->blkd_tasks list.
1156  *
1157  * Note that irqs must be enabled: boosting the task can block.
1158  * Returns 1 if there are more tasks needing to be boosted.
1159  */
1160 static int rcu_boost(struct rcu_node *rnp)
1161 {
1162         unsigned long flags;
1163         struct task_struct *t;
1164         struct list_head *tb;
1165 
1166         if (READ_ONCE(rnp->exp_tasks) == NULL &&
1167             READ_ONCE(rnp->boost_tasks) == NULL)
1168                 return 0;  /* Nothing left to boost. */
1169 
1170         raw_spin_lock_irqsave_rcu_node(rnp, flags);
1171 
1172         /*
1173          * Recheck under the lock: all tasks in need of boosting
1174          * might exit their RCU read-side critical sections on their own.
1175          */
1176         if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1177                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1178                 return 0;
1179         }
1180 
1181         /*
1182          * Preferentially boost tasks blocking expedited grace periods.
1183          * This cannot starve the normal grace periods because a second
1184          * expedited grace period must boost all blocked tasks, including
1185          * those blocking the pre-existing normal grace period.
1186          */
1187         if (rnp->exp_tasks != NULL)
1188                 tb = rnp->exp_tasks;
1189         else
1190                 tb = rnp->boost_tasks;
1191 
1192         /*
1193          * We boost task t by manufacturing an rt_mutex that appears to
1194          * be held by task t.  We leave a pointer to that rt_mutex where
1195          * task t can find it, and task t will release the mutex when it
1196          * exits its outermost RCU read-side critical section.  Then
1197          * simply acquiring this artificial rt_mutex will boost task
1198          * t's priority.  (Thanks to tglx for suggesting this approach!)
1199          *
1200          * Note that task t must acquire rnp->lock to remove itself from
1201          * the ->blkd_tasks list, which it will do from exit() if from
1202          * nowhere else.  We therefore are guaranteed that task t will
1203          * stay around at least until we drop rnp->lock.  Note that
1204          * rnp->lock also resolves races between our priority boosting
1205          * and task t's exiting its outermost RCU read-side critical
1206          * section.
1207          */
1208         t = container_of(tb, struct task_struct, rcu_node_entry);
1209         rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1210         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1211         /* Lock only for side effect: boosts task t's priority. */
1212         rt_mutex_lock(&rnp->boost_mtx);
1213         rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
1214 
1215         return READ_ONCE(rnp->exp_tasks) != NULL ||
1216                READ_ONCE(rnp->boost_tasks) != NULL;
1217 }
1218 
1219 /*
1220  * Priority-boosting kthread, one per leaf rcu_node.
1221  */
1222 static int rcu_boost_kthread(void *arg)
1223 {
1224         struct rcu_node *rnp = (struct rcu_node *)arg;
1225         int spincnt = 0;
1226         int more2boost;
1227 
1228         trace_rcu_utilization(TPS("Start boost kthread@init"));
1229         for (;;) {
1230                 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1231                 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1232                 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1233                 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1234                 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1235                 more2boost = rcu_boost(rnp);
1236                 if (more2boost)
1237                         spincnt++;
1238                 else
1239                         spincnt = 0;
1240                 if (spincnt > 10) {
1241                         rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1242                         trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1243                         schedule_timeout_interruptible(2);
1244                         trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1245                         spincnt = 0;
1246                 }
1247         }
1248         /* NOTREACHED */
1249         trace_rcu_utilization(TPS("End boost kthread@notreached"));
1250         return 0;
1251 }
1252 
1253 /*
1254  * Check to see if it is time to start boosting RCU readers that are
1255  * blocking the current grace period, and, if so, tell the per-rcu_node
1256  * kthread to start boosting them.  If there is an expedited grace
1257  * period in progress, it is always time to boost.
1258  *
1259  * The caller must hold rnp->lock, which this function releases.
1260  * The ->boost_kthread_task is immortal, so we don't need to worry
1261  * about it going away.
1262  */
1263 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1264         __releases(rnp->lock)
1265 {
1266         struct task_struct *t;
1267 
1268         raw_lockdep_assert_held_rcu_node(rnp);
1269         if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1270                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1271                 return;
1272         }
1273         if (rnp->exp_tasks != NULL ||
1274             (rnp->gp_tasks != NULL &&
1275              rnp->boost_tasks == NULL &&
1276              rnp->qsmask == 0 &&
1277              ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1278                 if (rnp->exp_tasks == NULL)
1279                         rnp->boost_tasks = rnp->gp_tasks;
1280                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1281                 t = rnp->boost_kthread_task;
1282                 if (t)
1283                         rcu_wake_cond(t, rnp->boost_kthread_status);
1284         } else {
1285                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1286         }
1287 }
1288 
1289 /*
1290  * Wake up the per-CPU kthread to invoke RCU callbacks.
1291  */
1292 static void invoke_rcu_callbacks_kthread(void)
1293 {
1294         unsigned long flags;
1295 
1296         local_irq_save(flags);
1297         __this_cpu_write(rcu_cpu_has_work, 1);
1298         if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1299             current != __this_cpu_read(rcu_cpu_kthread_task)) {
1300                 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1301                               __this_cpu_read(rcu_cpu_kthread_status));
1302         }
1303         local_irq_restore(flags);
1304 }
1305 
1306 /*
1307  * Is the current CPU running the RCU-callbacks kthread?
1308  * Caller must have preemption disabled.
1309  */
1310 static bool rcu_is_callbacks_kthread(void)
1311 {
1312         return __this_cpu_read(rcu_cpu_kthread_task) == current;
1313 }
1314 
1315 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1316 
1317 /*
1318  * Do priority-boost accounting for the start of a new grace period.
1319  */
1320 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1321 {
1322         rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1323 }
1324 
1325 /*
1326  * Create an RCU-boost kthread for the specified node if one does not
1327  * already exist.  We only create this kthread for preemptible RCU.
1328  * Returns zero if all is well, a negated errno otherwise.
1329  */
1330 static int rcu_spawn_one_boost_kthread(struct rcu_node *rnp)
1331 {
1332         int rnp_index = rnp - rcu_get_root();
1333         unsigned long flags;
1334         struct sched_param sp;
1335         struct task_struct *t;
1336 
1337         if (!IS_ENABLED(CONFIG_PREEMPT_RCU))
1338                 return 0;
1339 
1340         if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1341                 return 0;
1342 
1343         rcu_state.boost = 1;
1344         if (rnp->boost_kthread_task != NULL)
1345                 return 0;
1346         t = kthread_create(rcu_boost_kthread, (void *)rnp,
1347                            "rcub/%d", rnp_index);
1348         if (IS_ERR(t))
1349                 return PTR_ERR(t);
1350         raw_spin_lock_irqsave_rcu_node(rnp, flags);
1351         rnp->boost_kthread_task = t;
1352         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1353         sp.sched_priority = kthread_prio;
1354         sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1355         wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1356         return 0;
1357 }
1358 
1359 static void rcu_kthread_do_work(void)
1360 {
1361         rcu_do_batch(this_cpu_ptr(&rcu_data));
1362 }
1363 
1364 static void rcu_cpu_kthread_setup(unsigned int cpu)
1365 {
1366         struct sched_param sp;
1367 
1368         sp.sched_priority = kthread_prio;
1369         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1370 }
1371 
1372 static void rcu_cpu_kthread_park(unsigned int cpu)
1373 {
1374         per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1375 }
1376 
1377 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1378 {
1379         return __this_cpu_read(rcu_cpu_has_work);
1380 }
1381 
1382 /*
1383  * Per-CPU kernel thread that invokes RCU callbacks.  This replaces
1384  * the RCU softirq used in configurations of RCU that do not support RCU
1385  * priority boosting.
1386  */
1387 static void rcu_cpu_kthread(unsigned int cpu)
1388 {
1389         unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1390         char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1391         int spincnt;
1392 
1393         for (spincnt = 0; spincnt < 10; spincnt++) {
1394                 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1395                 local_bh_disable();
1396                 *statusp = RCU_KTHREAD_RUNNING;
1397                 this_cpu_inc(rcu_cpu_kthread_loops);
1398                 local_irq_disable();
1399                 work = *workp;
1400                 *workp = 0;
1401                 local_irq_enable();
1402                 if (work)
1403                         rcu_kthread_do_work();
1404                 local_bh_enable();
1405                 if (*workp == 0) {
1406                         trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1407                         *statusp = RCU_KTHREAD_WAITING;
1408                         return;
1409                 }
1410         }
1411         *statusp = RCU_KTHREAD_YIELDING;
1412         trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1413         schedule_timeout_interruptible(2);
1414         trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1415         *statusp = RCU_KTHREAD_WAITING;
1416 }
1417 
1418 /*
1419  * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1420  * served by the rcu_node in question.  The CPU hotplug lock is still
1421  * held, so the value of rnp->qsmaskinit will be stable.
1422  *
1423  * We don't include outgoingcpu in the affinity set, use -1 if there is
1424  * no outgoing CPU.  If there are no CPUs left in the affinity set,
1425  * this function allows the kthread to execute on any CPU.
1426  */
1427 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1428 {
1429         struct task_struct *t = rnp->boost_kthread_task;
1430         unsigned long mask = rcu_rnp_online_cpus(rnp);
1431         cpumask_var_t cm;
1432         int cpu;
1433 
1434         if (!t)
1435                 return;
1436         if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1437                 return;
1438         for_each_leaf_node_possible_cpu(rnp, cpu)
1439                 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1440                     cpu != outgoingcpu)
1441                         cpumask_set_cpu(cpu, cm);
1442         if (cpumask_weight(cm) == 0)
1443                 cpumask_setall(cm);
1444         set_cpus_allowed_ptr(t, cm);
1445         free_cpumask_var(cm);
1446 }
1447 
1448 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1449         .store                  = &rcu_cpu_kthread_task,
1450         .thread_should_run      = rcu_cpu_kthread_should_run,
1451         .thread_fn              = rcu_cpu_kthread,
1452         .thread_comm            = "rcuc/%u",
1453         .setup                  = rcu_cpu_kthread_setup,
1454         .park                   = rcu_cpu_kthread_park,
1455 };
1456 
1457 /*
1458  * Spawn boost kthreads -- called as soon as the scheduler is running.
1459  */
1460 static void __init rcu_spawn_boost_kthreads(void)
1461 {
1462         struct rcu_node *rnp;
1463         int cpu;
1464 
1465         for_each_possible_cpu(cpu)
1466                 per_cpu(rcu_cpu_has_work, cpu) = 0;
1467         BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1468         rcu_for_each_leaf_node(rnp)
1469                 (void)rcu_spawn_one_boost_kthread(rnp);
1470 }
1471 
1472 static void rcu_prepare_kthreads(int cpu)
1473 {
1474         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1475         struct rcu_node *rnp = rdp->mynode;
1476 
1477         /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1478         if (rcu_scheduler_fully_active)
1479                 (void)rcu_spawn_one_boost_kthread(rnp);
1480 }
1481 
1482 #else /* #ifdef CONFIG_RCU_BOOST */
1483 
1484 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1485         __releases(rnp->lock)
1486 {
1487         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1488 }
1489 
1490 static void invoke_rcu_callbacks_kthread(void)
1491 {
1492         WARN_ON_ONCE(1);
1493 }
1494 
1495 static bool rcu_is_callbacks_kthread(void)
1496 {
1497         return false;
1498 }
1499 
1500 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1501 {
1502 }
1503 
1504 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1505 {
1506 }
1507 
1508 static void __init rcu_spawn_boost_kthreads(void)
1509 {
1510 }
1511 
1512 static void rcu_prepare_kthreads(int cpu)
1513 {
1514 }
1515 
1516 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1517 
1518 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1519 
1520 /*
1521  * Check to see if any future RCU-related work will need to be done
1522  * by the current CPU, even if none need be done immediately, returning
1523  * 1 if so.  This function is part of the RCU implementation; it is -not-
1524  * an exported member of the RCU API.
1525  *
1526  * Because we not have RCU_FAST_NO_HZ, just check whether or not this
1527  * CPU has RCU callbacks queued.
1528  */
1529 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1530 {
1531         *nextevt = KTIME_MAX;
1532         return rcu_cpu_has_callbacks(NULL);
1533 }
1534 
1535 /*
1536  * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1537  * after it.
1538  */
1539 static void rcu_cleanup_after_idle(void)
1540 {
1541 }
1542 
1543 /*
1544  * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1545  * is nothing.
1546  */
1547 static void rcu_prepare_for_idle(void)
1548 {
1549 }
1550 
1551 /*
1552  * Don't bother keeping a running count of the number of RCU callbacks
1553  * posted because CONFIG_RCU_FAST_NO_HZ=n.
1554  */
1555 static void rcu_idle_count_callbacks_posted(void)
1556 {
1557 }
1558 
1559 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1560 
1561 /*
1562  * This code is invoked when a CPU goes idle, at which point we want
1563  * to have the CPU do everything required for RCU so that it can enter
1564  * the energy-efficient dyntick-idle mode.  This is handled by a
1565  * state machine implemented by rcu_prepare_for_idle() below.
1566  *
1567  * The following three proprocessor symbols control this state machine:
1568  *
1569  * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1570  *      to sleep in dyntick-idle mode with RCU callbacks pending.  This
1571  *      is sized to be roughly one RCU grace period.  Those energy-efficiency
1572  *      benchmarkers who might otherwise be tempted to set this to a large
1573  *      number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1574  *      system.  And if you are -that- concerned about energy efficiency,
1575  *      just power the system down and be done with it!
1576  * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1577  *      permitted to sleep in dyntick-idle mode with only lazy RCU
1578  *      callbacks pending.  Setting this too high can OOM your system.
1579  *
1580  * The values below work well in practice.  If future workloads require
1581  * adjustment, they can be converted into kernel config parameters, though
1582  * making the state machine smarter might be a better option.
1583  */
1584 #define RCU_IDLE_GP_DELAY 4             /* Roughly one grace period. */
1585 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1586 
1587 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1588 module_param(rcu_idle_gp_delay, int, 0644);
1589 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1590 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1591 
1592 /*
1593  * Try to advance callbacks on the current CPU, but only if it has been
1594  * awhile since the last time we did so.  Afterwards, if there are any
1595  * callbacks ready for immediate invocation, return true.
1596  */
1597 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1598 {
1599         bool cbs_ready = false;
1600         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1601         struct rcu_node *rnp;
1602 
1603         /* Exit early if we advanced recently. */
1604         if (jiffies == rdp->last_advance_all)
1605                 return false;
1606         rdp->last_advance_all = jiffies;
1607 
1608         rnp = rdp->mynode;
1609 
1610         /*
1611          * Don't bother checking unless a grace period has
1612          * completed since we last checked and there are
1613          * callbacks not yet ready to invoke.
1614          */
1615         if ((rcu_seq_completed_gp(rdp->gp_seq,
1616                                   rcu_seq_current(&rnp->gp_seq)) ||
1617              unlikely(READ_ONCE(rdp->gpwrap))) &&
1618             rcu_segcblist_pend_cbs(&rdp->cblist))
1619                 note_gp_changes(rdp);
1620 
1621         if (rcu_segcblist_ready_cbs(&rdp->cblist))
1622                 cbs_ready = true;
1623         return cbs_ready;
1624 }
1625 
1626 /*
1627  * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1628  * to invoke.  If the CPU has callbacks, try to advance them.  Tell the
1629  * caller to set the timeout based on whether or not there are non-lazy
1630  * callbacks.
1631  *
1632  * The caller must have disabled interrupts.
1633  */
1634 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1635 {
1636         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1637         unsigned long dj;
1638 
1639         lockdep_assert_irqs_disabled();
1640 
1641         /* Snapshot to detect later posting of non-lazy callback. */
1642         rdp->nonlazy_posted_snap = rdp->nonlazy_posted;
1643 
1644         /* If no callbacks, RCU doesn't need the CPU. */
1645         if (!rcu_cpu_has_callbacks(&rdp->all_lazy)) {
1646                 *nextevt = KTIME_MAX;
1647                 return 0;
1648         }
1649 
1650         /* Attempt to advance callbacks. */
1651         if (rcu_try_advance_all_cbs()) {
1652                 /* Some ready to invoke, so initiate later invocation. */
1653                 invoke_rcu_core();
1654                 return 1;
1655         }
1656         rdp->last_accelerate = jiffies;
1657 
1658         /* Request timer delay depending on laziness, and round. */
1659         if (!rdp->all_lazy) {
1660                 dj = round_up(rcu_idle_gp_delay + jiffies,
1661                                rcu_idle_gp_delay) - jiffies;
1662         } else {
1663                 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1664         }
1665         *nextevt = basemono + dj * TICK_NSEC;
1666         return 0;
1667 }
1668 
1669 /*
1670  * Prepare a CPU for idle from an RCU perspective.  The first major task
1671  * is to sense whether nohz mode has been enabled or disabled via sysfs.
1672  * The second major task is to check to see if a non-lazy callback has
1673  * arrived at a CPU that previously had only lazy callbacks.  The third
1674  * major task is to accelerate (that is, assign grace-period numbers to)
1675  * any recently arrived callbacks.
1676  *
1677  * The caller must have disabled interrupts.
1678  */
1679 static void rcu_prepare_for_idle(void)
1680 {
1681         bool needwake;
1682         struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
1683         struct rcu_node *rnp;
1684         int tne;
1685 
1686         lockdep_assert_irqs_disabled();
1687         if (rcu_is_nocb_cpu(smp_processor_id()))
1688                 return;
1689 
1690         /* Handle nohz enablement switches conservatively. */
1691         tne = READ_ONCE(tick_nohz_active);
1692         if (tne != rdp->tick_nohz_enabled_snap) {
1693                 if (rcu_cpu_has_callbacks(NULL))
1694                         invoke_rcu_core(); /* force nohz to see update. */
1695                 rdp->tick_nohz_enabled_snap = tne;
1696                 return;
1697         }
1698         if (!tne)
1699                 return;
1700 
1701         /*
1702          * If a non-lazy callback arrived at a CPU having only lazy
1703          * callbacks, invoke RCU core for the side-effect of recalculating
1704          * idle duration on re-entry to idle.
1705          */
1706         if (rdp->all_lazy &&
1707             rdp->nonlazy_posted != rdp->nonlazy_posted_snap) {
1708                 rdp->all_lazy = false;
1709                 rdp->nonlazy_posted_snap = rdp->nonlazy_posted;
1710                 invoke_rcu_core();
1711                 return;
1712         }
1713 
1714         /*
1715          * If we have not yet accelerated this jiffy, accelerate all
1716          * callbacks on this CPU.
1717          */
1718         if (rdp->last_accelerate == jiffies)
1719                 return;
1720         rdp->last_accelerate = jiffies;
1721         if (rcu_segcblist_pend_cbs(&rdp->cblist)) {
1722                 rnp = rdp->mynode;
1723                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1724                 needwake = rcu_accelerate_cbs(rnp, rdp);
1725                 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1726                 if (needwake)
1727                         rcu_gp_kthread_wake();
1728         }
1729 }
1730 
1731 /*
1732  * Clean up for exit from idle.  Attempt to advance callbacks based on
1733  * any grace periods that elapsed while the CPU was idle, and if any
1734  * callbacks are now ready to invoke, initiate invocation.
1735  */
1736 static void rcu_cleanup_after_idle(void)
1737 {
1738         lockdep_assert_irqs_disabled();
1739         if (rcu_is_nocb_cpu(smp_processor_id()))
1740                 return;
1741         if (rcu_try_advance_all_cbs())
1742                 invoke_rcu_core();
1743 }
1744 
1745 /*
1746  * Keep a running count of the number of non-lazy callbacks posted
1747  * on this CPU.  This running counter (which is never decremented) allows
1748  * rcu_prepare_for_idle() to detect when something out of the idle loop
1749  * posts a callback, even if an equal number of callbacks are invoked.
1750  * Of course, callbacks should only be posted from within a trace event
1751  * designed to be called from idle or from within RCU_NONIDLE().
1752  */
1753 static void rcu_idle_count_callbacks_posted(void)
1754 {
1755         __this_cpu_add(rcu_data.nonlazy_posted, 1);
1756 }
1757 
1758 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1759 
1760 #ifdef CONFIG_RCU_FAST_NO_HZ
1761 
1762 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1763 {
1764         struct rcu_data *rdp = &per_cpu(rcu_data, cpu);
1765         unsigned long nlpd = rdp->nonlazy_posted - rdp->nonlazy_posted_snap;
1766 
1767         sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1768                 rdp->last_accelerate & 0xffff, jiffies & 0xffff,
1769                 ulong2long(nlpd),
1770                 rdp->all_lazy ? 'L' : '.',
1771                 rdp->tick_nohz_enabled_snap ? '.' : 'D');
1772 }
1773 
1774 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1775 
1776 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1777 {
1778         *cp = '\0';
1779 }
1780 
1781 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1782 
1783 /* Initiate the stall-info list. */
1784 static void print_cpu_stall_info_begin(void)
1785 {
1786         pr_cont("\n");
1787 }
1788 
1789 /*
1790  * Print out diagnostic information for the specified stalled CPU.
1791  *
1792  * If the specified CPU is aware of the current RCU grace period, then
1793  * print the number of scheduling clock interrupts the CPU has taken
1794  * during the time that it has been aware.  Otherwise, print the number
1795  * of RCU grace periods that this CPU is ignorant of, for example, "1"
1796  * if the CPU was aware of the previous grace period.
1797  *
1798  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1799  */
1800 static void print_cpu_stall_info(int cpu)
1801 {
1802         unsigned long delta;
1803         char fast_no_hz[72];
1804         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1805         char *ticks_title;
1806         unsigned long ticks_value;
1807 
1808         /*
1809          * We could be printing a lot while holding a spinlock.  Avoid
1810          * triggering hard lockup.
1811          */
1812         touch_nmi_watchdog();
1813 
1814         ticks_value = rcu_seq_ctr(rcu_state.gp_seq - rdp->gp_seq);
1815         if (ticks_value) {
1816                 ticks_title = "GPs behind";
1817         } else {
1818                 ticks_title = "ticks this GP";
1819                 ticks_value = rdp->ticks_this_gp;
1820         }
1821         print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1822         delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
1823         pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s\n",
1824                cpu,
1825                "O."[!!cpu_online(cpu)],
1826                "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
1827                "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
1828                !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
1829                         rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '' :
1830                                 "!."[!delta],
1831                ticks_value, ticks_title,
1832                rcu_dynticks_snap(rdp) & 0xfff,
1833                rdp->dynticks_nesting, rdp->dynticks_nmi_nesting,
1834                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1835                READ_ONCE(rcu_state.n_force_qs) - rcu_state.n_force_qs_gpstart,
1836                fast_no_hz);
1837 }
1838 
1839 /* Terminate the stall-info list. */
1840 static void print_cpu_stall_info_end(void)
1841 {
1842         pr_err("\t");
1843 }
1844 
1845 /* Zero ->ticks_this_gp and snapshot the number of RCU softirq handlers. */
1846 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1847 {
1848         rdp->ticks_this_gp = 0;
1849         rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1850         WRITE_ONCE(rdp->last_fqs_resched, jiffies);
1851 }
1852 
1853 #ifdef CONFIG_RCU_NOCB_CPU
1854 
1855 /*
1856  * Offload callback processing from the boot-time-specified set of CPUs
1857  * specified by rcu_nocb_mask.  For each CPU in the set, there is a
1858  * kthread created that pulls the callbacks from the corresponding CPU,
1859  * waits for a grace period to elapse, and invokes the callbacks.
1860  * The no-CBs CPUs do a wake_up() on their kthread when they insert
1861  * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1862  * has been specified, in which case each kthread actively polls its
1863  * CPU.  (Which isn't so great for energy efficiency, but which does
1864  * reduce RCU's overhead on that CPU.)
1865  *
1866  * This is intended to be used in conjunction with Frederic Weisbecker's
1867  * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1868  * running CPU-bound user-mode computations.
1869  *
1870  * Offloading of callback processing could also in theory be used as
1871  * an energy-efficiency measure because CPUs with no RCU callbacks
1872  * queued are more aggressive about entering dyntick-idle mode.
1873  */
1874 
1875 
1876 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1877 static int __init rcu_nocb_setup(char *str)
1878 {
1879         alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1880         cpulist_parse(str, rcu_nocb_mask);
1881         return 1;
1882 }
1883 __setup("rcu_nocbs=", rcu_nocb_setup);
1884 
1885 static int __init parse_rcu_nocb_poll(char *arg)
1886 {
1887         rcu_nocb_poll = true;
1888         return 0;
1889 }
1890 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1891 
1892 /*
1893  * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1894  * grace period.
1895  */
1896 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1897 {
1898         swake_up_all(sq);
1899 }
1900 
1901 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1902 {
1903         return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1904 }
1905 
1906 static void rcu_init_one_nocb(struct rcu_node *rnp)
1907 {
1908         init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1909         init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1910 }
1911 
1912 /* Is the specified CPU a no-CBs CPU? */
1913 bool rcu_is_nocb_cpu(int cpu)
1914 {
1915         if (cpumask_available(rcu_nocb_mask))
1916                 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1917         return false;
1918 }
1919 
1920 /*
1921  * Kick the leader kthread for this NOCB group.  Caller holds ->nocb_lock
1922  * and this function releases it.
1923  */
1924 static void __wake_nocb_leader(struct rcu_data *rdp, bool force,
1925                                unsigned long flags)
1926         __releases(rdp->nocb_lock)
1927 {
1928         struct rcu_data *rdp_leader = rdp->nocb_leader;
1929 
1930         lockdep_assert_held(&rdp->nocb_lock);
1931         if (!READ_ONCE(rdp_leader->nocb_kthread)) {
1932                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1933                 return;
1934         }
1935         if (rdp_leader->nocb_leader_sleep || force) {
1936                 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1937                 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1938                 del_timer(&rdp->nocb_timer);
1939                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1940                 smp_mb(); /* ->nocb_leader_sleep before swake_up_one(). */
1941                 swake_up_one(&rdp_leader->nocb_wq);
1942         } else {
1943                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1944         }
1945 }
1946 
1947 /*
1948  * Kick the leader kthread for this NOCB group, but caller has not
1949  * acquired locks.
1950  */
1951 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1952 {
1953         unsigned long flags;
1954 
1955         raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1956         __wake_nocb_leader(rdp, force, flags);
1957 }
1958 
1959 /*
1960  * Arrange to wake the leader kthread for this NOCB group at some
1961  * future time when it is safe to do so.
1962  */
1963 static void wake_nocb_leader_defer(struct rcu_data *rdp, int waketype,
1964                                    const char *reason)
1965 {
1966         unsigned long flags;
1967 
1968         raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1969         if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1970                 mod_timer(&rdp->nocb_timer, jiffies + 1);
1971         WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1972         trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, reason);
1973         raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1974 }
1975 
1976 /*
1977  * Does the specified CPU need an RCU callback for this invocation
1978  * of rcu_barrier()?
1979  */
1980 static bool rcu_nocb_cpu_needs_barrier(int cpu)
1981 {
1982         struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu);
1983         unsigned long ret;
1984 #ifdef CONFIG_PROVE_RCU
1985         struct rcu_head *rhp;
1986 #endif /* #ifdef CONFIG_PROVE_RCU */
1987 
1988         /*
1989          * Check count of all no-CBs callbacks awaiting invocation.
1990          * There needs to be a barrier before this function is called,
1991          * but associated with a prior determination that no more
1992          * callbacks would be posted.  In the worst case, the first
1993          * barrier in rcu_barrier() suffices (but the caller cannot
1994          * necessarily rely on this, not a substitute for the caller
1995          * getting the concurrency design right!).  There must also be
1996          * a barrier between the following load an posting of a callback
1997          * (if a callback is in fact needed).  This is associated with an
1998          * atomic_inc() in the caller.
1999          */
2000         ret = atomic_long_read(&rdp->nocb_q_count);
2001 
2002 #ifdef CONFIG_PROVE_RCU
2003         rhp = READ_ONCE(rdp->nocb_head);
2004         if (!rhp)
2005                 rhp = READ_ONCE(rdp->nocb_gp_head);
2006         if (!rhp)
2007                 rhp = READ_ONCE(rdp->nocb_follower_head);
2008 
2009         /* Having no rcuo kthread but CBs after scheduler starts is bad! */
2010         if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
2011             rcu_scheduler_fully_active) {
2012                 /* RCU callback enqueued before CPU first came online??? */
2013                 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
2014                        cpu, rhp->func);
2015                 WARN_ON_ONCE(1);
2016         }
2017 #endif /* #ifdef CONFIG_PROVE_RCU */
2018 
2019         return !!ret;
2020 }
2021 
2022 /*
2023  * Enqueue the specified string of rcu_head structures onto the specified
2024  * CPU's no-CBs lists.  The CPU is specified by rdp, the head of the
2025  * string by rhp, and the tail of the string by rhtp.  The non-lazy/lazy
2026  * counts are supplied by rhcount and rhcount_lazy.
2027  *
2028  * If warranted, also wake up the kthread servicing this CPUs queues.
2029  */
2030 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2031                                     struct rcu_head *rhp,
2032                                     struct rcu_head **rhtp,
2033                                     int rhcount, int rhcount_lazy,
2034                                     unsigned long flags)
2035 {
2036         int len;
2037         struct rcu_head **old_rhpp;
2038         struct task_struct *t;
2039 
2040         /* Enqueue the callback on the nocb list and update counts. */
2041         atomic_long_add(rhcount, &rdp->nocb_q_count);
2042         /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
2043         old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2044         WRITE_ONCE(*old_rhpp, rhp);
2045         atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2046         smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
2047 
2048         /* If we are not being polled and there is a kthread, awaken it ... */
2049         t = READ_ONCE(rdp->nocb_kthread);
2050         if (rcu_nocb_poll || !t) {
2051                 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2052                                     TPS("WakeNotPoll"));
2053                 return;
2054         }
2055         len = atomic_long_read(&rdp->nocb_q_count);
2056         if (old_rhpp == &rdp->nocb_head) {
2057                 if (!irqs_disabled_flags(flags)) {
2058                         /* ... if queue was empty ... */
2059                         wake_nocb_leader(rdp, false);
2060                         trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2061                                             TPS("WakeEmpty"));
2062                 } else {
2063                         wake_nocb_leader_defer(rdp, RCU_NOCB_WAKE,
2064                                                TPS("WakeEmptyIsDeferred"));
2065                 }
2066                 rdp->qlen_last_fqs_check = 0;
2067         } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2068                 /* ... or if many callbacks queued. */
2069                 if (!irqs_disabled_flags(flags)) {
2070                         wake_nocb_leader(rdp, true);
2071                         trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2072                                             TPS("WakeOvf"));
2073                 } else {
2074                         wake_nocb_leader_defer(rdp, RCU_NOCB_WAKE_FORCE,
2075                                                TPS("WakeOvfIsDeferred"));
2076                 }
2077                 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2078         } else {
2079                 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WakeNot"));
2080         }
2081         return;
2082 }
2083 
2084 /*
2085  * This is a helper for __call_rcu(), which invokes this when the normal
2086  * callback queue is inoperable.  If this is not a no-CBs CPU, this
2087  * function returns failure back to __call_rcu(), which can complain
2088  * appropriately.
2089  *
2090  * Otherwise, this function queues the callback where the corresponding
2091  * "rcuo" kthread can find it.
2092  */
2093 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2094                             bool lazy, unsigned long flags)
2095 {
2096 
2097         if (!rcu_is_nocb_cpu(rdp->cpu))
2098                 return false;
2099         __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2100         if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2101                 trace_rcu_kfree_callback(rcu_state.name, rhp,
2102                                          (unsigned long)rhp->func,
2103                                          -atomic_long_read(&rdp->nocb_q_count_lazy),
2104                                          -atomic_long_read(&rdp->nocb_q_count));
2105         else
2106                 trace_rcu_callback(rcu_state.name, rhp,
2107                                    -atomic_long_read(&rdp->nocb_q_count_lazy),
2108                                    -atomic_long_read(&rdp->nocb_q_count));
2109 
2110         /*
2111          * If called from an extended quiescent state with interrupts
2112          * disabled, invoke the RCU core in order to allow the idle-entry
2113          * deferred-wakeup check to function.
2114          */
2115         if (irqs_disabled_flags(flags) &&
2116             !rcu_is_watching() &&
2117             cpu_online(smp_processor_id()))
2118                 invoke_rcu_core();
2119 
2120         return true;
2121 }
2122 
2123 /*
2124  * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2125  * not a no-CBs CPU.
2126  */
2127 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
2128                                                      struct rcu_data *rdp,
2129                                                      unsigned long flags)
2130 {
2131         lockdep_assert_irqs_disabled();
2132         if (!rcu_is_nocb_cpu(smp_processor_id()))
2133                 return false; /* Not NOCBs CPU, caller must migrate CBs. */
2134         __call_rcu_nocb_enqueue(my_rdp, rcu_segcblist_head(&rdp->cblist),
2135                                 rcu_segcblist_tail(&rdp->cblist),
2136                                 rcu_segcblist_n_cbs(&rdp->cblist),
2137                                 rcu_segcblist_n_lazy_cbs(&rdp->cblist), flags);
2138         rcu_segcblist_init(&rdp->cblist);
2139         rcu_segcblist_disable(&rdp->cblist);
2140         return true;
2141 }
2142 
2143 /*
2144  * If necessary, kick off a new grace period, and either way wait
2145  * for a subsequent grace period to complete.
2146  */
2147 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2148 {
2149         unsigned long c;
2150         bool d;
2151         unsigned long flags;
2152         bool needwake;
2153         struct rcu_node *rnp = rdp->mynode;
2154 
2155         local_irq_save(flags);
2156         c = rcu_seq_snap(&rcu_state.gp_seq);
2157         if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
2158                 local_irq_restore(flags);
2159         } else {
2160                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2161                 needwake = rcu_start_this_gp(rnp, rdp, c);
2162                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2163                 if (needwake)
2164                         rcu_gp_kthread_wake();
2165         }
2166 
2167         /*
2168          * Wait for the grace period.  Do so interruptibly to avoid messing
2169          * up the load average.
2170          */
2171         trace_rcu_this_gp(rnp, rdp, c, TPS("StartWait"));
2172         for (;;) {
2173                 swait_event_interruptible_exclusive(
2174                         rnp->nocb_gp_wq[rcu_seq_ctr(c) & 0x1],
2175                         (d = rcu_seq_done(&rnp->gp_seq, c)));
2176                 if (likely(d))
2177                         break;
2178                 WARN_ON(signal_pending(current));
2179                 trace_rcu_this_gp(rnp, rdp, c, TPS("ResumeWait"));
2180         }
2181         trace_rcu_this_gp(rnp, rdp, c, TPS("EndWait"));
2182         smp_mb(); /* Ensure that CB invocation happens after GP end. */
2183 }
2184 
2185 /*
2186  * Leaders come here to wait for additional callbacks to show up.
2187  * This function does not return until callbacks appear.
2188  */
2189 static void nocb_leader_wait(struct rcu_data *my_rdp)
2190 {
2191         bool firsttime = true;
2192         unsigned long flags;
2193         bool gotcbs;
2194         struct rcu_data *rdp;
2195         struct rcu_head **tail;
2196 
2197 wait_again:
2198 
2199         /* Wait for callbacks to appear. */
2200         if (!rcu_nocb_poll) {
2201                 trace_rcu_nocb_wake(rcu_state.name, my_rdp->cpu, TPS("Sleep"));
2202                 swait_event_interruptible_exclusive(my_rdp->nocb_wq,
2203                                 !READ_ONCE(my_rdp->nocb_leader_sleep));
2204                 raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
2205                 my_rdp->nocb_leader_sleep = true;
2206                 WRITE_ONCE(my_rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2207                 del_timer(&my_rdp->nocb_timer);
2208                 raw_spin_unlock_irqrestore(&my_rdp->nocb_lock, flags);
2209         } else if (firsttime) {
2210                 firsttime = false; /* Don't drown trace log with "Poll"! */
2211                 trace_rcu_nocb_wake(rcu_state.name, my_rdp->cpu, TPS("Poll"));
2212         }
2213 
2214         /*
2215          * Each pass through the following loop checks a follower for CBs.
2216          * We are our own first follower.  Any CBs found are moved to
2217          * nocb_gp_head, where they await a grace period.
2218          */
2219         gotcbs = false;
2220         smp_mb(); /* wakeup and _sleep before ->nocb_head reads. */
2221         for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2222                 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2223                 if (!rdp->nocb_gp_head)
2224                         continue;  /* No CBs here, try next follower. */
2225 
2226                 /* Move callbacks to wait-for-GP list, which is empty. */
2227                 WRITE_ONCE(rdp->nocb_head, NULL);
2228                 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2229                 gotcbs = true;
2230         }
2231 
2232         /* No callbacks?  Sleep a bit if polling, and go retry.  */
2233         if (unlikely(!gotcbs)) {
2234                 WARN_ON(signal_pending(current));
2235                 if (rcu_nocb_poll) {
2236                         schedule_timeout_interruptible(1);
2237                 } else {
2238                         trace_rcu_nocb_wake(rcu_state.name, my_rdp->cpu,
2239                                             TPS("WokeEmpty"));
2240                 }
2241                 goto wait_again;
2242         }
2243 
2244         /* Wait for one grace period. */
2245         rcu_nocb_wait_gp(my_rdp);
2246 
2247         /* Each pass through the following loop wakes a follower, if needed. */
2248         for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2249                 if (!rcu_nocb_poll &&
2250                     READ_ONCE(rdp->nocb_head) &&
2251                     READ_ONCE(my_rdp->nocb_leader_sleep)) {
2252                         raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
2253                         my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2254                         raw_spin_unlock_irqrestore(&my_rdp->nocb_lock, flags);
2255                 }
2256                 if (!rdp->nocb_gp_head)
2257                         continue; /* No CBs, so no need to wake follower. */
2258 
2259                 /* Append callbacks to follower's "done" list. */
2260                 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2261                 tail = rdp->nocb_follower_tail;
2262                 rdp->nocb_follower_tail = rdp->nocb_gp_tail;
2263                 *tail = rdp->nocb_gp_head;
2264                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2265                 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2266                         /* List was empty, so wake up the follower.  */
2267                         swake_up_one(&rdp->nocb_wq);
2268                 }
2269         }
2270 
2271         /* If we (the leader) don't have CBs, go wait some more. */
2272         if (!my_rdp->nocb_follower_head)
2273                 goto wait_again;
2274 }
2275 
2276 /*
2277  * Followers come here to wait for additional callbacks to show up.
2278  * This function does not return until callbacks appear.
2279  */
2280 static void nocb_follower_wait(struct rcu_data *rdp)
2281 {
2282         for (;;) {
2283                 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("FollowerSleep"));
2284                 swait_event_interruptible_exclusive(rdp->nocb_wq,
2285                                          READ_ONCE(rdp->nocb_follower_head));
2286                 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2287                         /* ^^^ Ensure CB invocation follows _head test. */
2288                         return;
2289                 }
2290                 WARN_ON(signal_pending(current));
2291                 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeEmpty"));
2292         }
2293 }
2294 
2295 /*
2296  * Per-rcu_data kthread, but only for no-CBs CPUs.  Each kthread invokes
2297  * callbacks queued by the corresponding no-CBs CPU, however, there is
2298  * an optional leader-follower relationship so that the grace-period
2299  * kthreads don't have to do quite so many wakeups.
2300  */
2301 static int rcu_nocb_kthread(void *arg)
2302 {
2303         int c, cl;
2304         unsigned long flags;
2305         struct rcu_head *list;
2306         struct rcu_head *next;
2307         struct rcu_head **tail;
2308         struct rcu_data *rdp = arg;
2309 
2310         /* Each pass through this loop invokes one batch of callbacks */
2311         for (;;) {
2312                 /* Wait for callbacks. */
2313                 if (rdp->nocb_leader == rdp)
2314                         nocb_leader_wait(rdp);
2315                 else
2316                         nocb_follower_wait(rdp);
2317 
2318                 /* Pull the ready-to-invoke callbacks onto local list. */
2319                 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2320                 list = rdp->nocb_follower_head;
2321                 rdp->nocb_follower_head = NULL;
2322                 tail = rdp->nocb_follower_tail;
2323                 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2324                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2325                 BUG_ON(!list);
2326                 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("WokeNonEmpty"));
2327 
2328                 /* Each pass through the following loop invokes a callback. */
2329                 trace_rcu_batch_start(rcu_state.name,
2330                                       atomic_long_read(&rdp->nocb_q_count_lazy),
2331                                       atomic_long_read(&rdp->nocb_q_count), -1);
2332                 c = cl = 0;
2333                 while (list) {
2334                         next = list->next;
2335                         /* Wait for enqueuing to complete, if needed. */
2336                         while (next == NULL && &list->next != tail) {
2337                                 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2338                                                     TPS("WaitQueue"));
2339                                 schedule_timeout_interruptible(1);
2340                                 trace_rcu_nocb_wake(rcu_state.name, rdp->cpu,
2341                                                     TPS("WokeQueue"));
2342                                 next = list->next;
2343                         }
2344                         debug_rcu_head_unqueue(list);
2345                         local_bh_disable();
2346                         if (__rcu_reclaim(rcu_state.name, list))
2347                                 cl++;
2348                         c++;
2349                         local_bh_enable();
2350                         cond_resched_tasks_rcu_qs();
2351                         list = next;
2352                 }
2353                 trace_rcu_batch_end(rcu_state.name, c, !!list, 0, 0, 1);
2354                 smp_mb__before_atomic();  /* _add after CB invocation. */
2355                 atomic_long_add(-c, &rdp->nocb_q_count);
2356                 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2357         }
2358         return 0;
2359 }
2360 
2361 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2362 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2363 {
2364         return READ_ONCE(rdp->nocb_defer_wakeup);
2365 }
2366 
2367 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2368 static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2369 {
2370         unsigned long flags;
2371         int ndw;
2372 
2373         raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2374         if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2375                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2376                 return;
2377         }
2378         ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2379         WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2380         __wake_nocb_leader(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2381         trace_rcu_nocb_wake(rcu_state.name, rdp->cpu, TPS("DeferredWake"));
2382 }
2383 
2384 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2385 static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2386 {
2387         struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2388 
2389         do_nocb_deferred_wakeup_common(rdp);
2390 }
2391 
2392 /*
2393  * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2394  * This means we do an inexact common-case check.  Note that if
2395  * we miss, ->nocb_timer will eventually clean things up.
2396  */
2397 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2398 {
2399         if (rcu_nocb_need_deferred_wakeup(rdp))
2400                 do_nocb_deferred_wakeup_common(rdp);
2401 }
2402 
2403 void __init rcu_init_nohz(void)
2404 {
2405         int cpu;
2406         bool need_rcu_nocb_mask = false;
2407 
2408 #if defined(CONFIG_NO_HZ_FULL)
2409         if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2410                 need_rcu_nocb_mask = true;
2411 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2412 
2413         if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2414                 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2415                         pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2416                         return;
2417                 }
2418         }
2419         if (!cpumask_available(rcu_nocb_mask))
2420                 return;
2421 
2422 #if defined(CONFIG_NO_HZ_FULL)
2423         if (tick_nohz_full_running)
2424                 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2425 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2426 
2427         if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2428                 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2429                 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2430                             rcu_nocb_mask);
2431         }
2432         if (cpumask_empty(rcu_nocb_mask))
2433                 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2434         else
2435                 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2436                         cpumask_pr_args(rcu_nocb_mask));
2437         if (rcu_nocb_poll)
2438                 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2439 
2440         for_each_cpu(cpu, rcu_nocb_mask)
2441                 init_nocb_callback_list(per_cpu_ptr(&rcu_data, cpu));
2442         rcu_organize_nocb_kthreads();
2443 }
2444 
2445 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2446 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2447 {
2448         rdp->nocb_tail = &rdp->nocb_head;
2449         init_swait_queue_head(&rdp->nocb_wq);
2450         rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2451         raw_spin_lock_init(&rdp->nocb_lock);
2452         timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2453 }
2454 
2455 /*
2456  * If the specified CPU is a no-CBs CPU that does not already have its
2457  * rcuo kthread, spawn it.  If the CPUs are brought online out of order,
2458  * this can require re-organizing the leader-follower relationships.
2459  */
2460 static void rcu_spawn_one_nocb_kthread(int cpu)
2461 {
2462         struct rcu_data *rdp;
2463         struct rcu_data *rdp_last;
2464         struct rcu_data *rdp_old_leader;
2465         struct rcu_data *rdp_spawn = per_cpu_ptr(&rcu_data, cpu);
2466         struct task_struct *t;
2467 
2468         /*
2469          * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2470          * then nothing to do.
2471          */
2472         if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2473                 return;
2474 
2475         /* If we didn't spawn the leader first, reorganize! */
2476         rdp_old_leader = rdp_spawn->nocb_leader;
2477         if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2478                 rdp_last = NULL;
2479                 rdp = rdp_old_leader;
2480                 do {
2481                         rdp->nocb_leader = rdp_spawn;
2482                         if (rdp_last && rdp != rdp_spawn)
2483                                 rdp_last->nocb_next_follower = rdp;
2484                         if (rdp == rdp_spawn) {
2485                                 rdp = rdp->nocb_next_follower;
2486                         } else {
2487                                 rdp_last = rdp;
2488                                 rdp = rdp->nocb_next_follower;
2489                                 rdp_last->nocb_next_follower = NULL;
2490                         }
2491                 } while (rdp);
2492                 rdp_spawn->nocb_next_follower = rdp_old_leader;
2493         }
2494 
2495         /* Spawn the kthread for this CPU. */
2496         t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2497                         "rcuo%c/%d", rcu_state.abbr, cpu);
2498         BUG_ON(IS_ERR(t));
2499         WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2500 }
2501 
2502 /*
2503  * If the specified CPU is a no-CBs CPU that does not already have its
2504  * rcuo kthreads, spawn them.
2505  */
2506 static void rcu_spawn_all_nocb_kthreads(int cpu)
2507 {
2508         if (rcu_scheduler_fully_active)
2509                 rcu_spawn_one_nocb_kthread(cpu);
2510 }
2511 
2512 /*
2513  * Once the scheduler is running, spawn rcuo kthreads for all online
2514  * no-CBs CPUs.  This assumes that the early_initcall()s happen before
2515  * non-boot CPUs come online -- if this changes, we will need to add
2516  * some mutual exclusion.
2517  */
2518 static void __init rcu_spawn_nocb_kthreads(void)
2519 {
2520         int cpu;
2521 
2522         for_each_online_cpu(cpu)
2523                 rcu_spawn_all_nocb_kthreads(cpu);
2524 }
2525 
2526 /* How many follower CPU IDs per leader?  Default of -1 for sqrt(nr_cpu_ids). */
2527 static int rcu_nocb_leader_stride = -1;
2528 module_param(rcu_nocb_leader_stride, int, 0444);
2529 
2530 /*
2531  * Initialize leader-follower relationships for all no-CBs CPU.
2532  */
2533 static void __init rcu_organize_nocb_kthreads(void)
2534 {
2535         int cpu;
2536         int ls = rcu_nocb_leader_stride;
2537         int nl = 0;  /* Next leader. */
2538         struct rcu_data *rdp;
2539         struct rcu_data *rdp_leader = NULL;  /* Suppress misguided gcc warn. */
2540         struct rcu_data *rdp_prev = NULL;
2541 
2542         if (!cpumask_available(rcu_nocb_mask))
2543                 return;
2544         if (ls == -1) {
2545                 ls = int_sqrt(nr_cpu_ids);
2546                 rcu_nocb_leader_stride = ls;
2547         }
2548 
2549         /*
2550          * Each pass through this loop sets up one rcu_data structure.
2551          * Should the corresponding CPU come online in the future, then
2552          * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2553          */
2554         for_each_cpu(cpu, rcu_nocb_mask) {
2555                 rdp = per_cpu_ptr(&rcu_data, cpu);
2556                 if (rdp->cpu >= nl) {
2557                         /* New leader, set up for followers & next leader. */
2558                         nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2559                         rdp->nocb_leader = rdp;
2560                         rdp_leader = rdp;
2561                 } else {
2562                         /* Another follower, link to previous leader. */
2563                         rdp->nocb_leader = rdp_leader;
2564                         rdp_prev->nocb_next_follower = rdp;
2565                 }
2566                 rdp_prev = rdp;
2567         }
2568 }
2569 
2570 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2571 static bool init_nocb_callback_list(struct rcu_data *rdp)
2572 {
2573         if (!rcu_is_nocb_cpu(rdp->cpu))
2574                 return false;
2575 
2576         /* If there are early-boot callbacks, move them to nocb lists. */
2577         if (!rcu_segcblist_empty(&rdp->cblist)) {
2578                 rdp->nocb_head = rcu_segcblist_head(&rdp->cblist);
2579                 rdp->nocb_tail = rcu_segcblist_tail(&rdp->cblist);
2580                 atomic_long_set(&rdp->nocb_q_count,
2581                                 rcu_segcblist_n_cbs(&rdp->cblist));
2582                 atomic_long_set(&rdp->nocb_q_count_lazy,
2583                                 rcu_segcblist_n_lazy_cbs(&rdp->cblist));
2584                 rcu_segcblist_init(&rdp->cblist);
2585         }
2586         rcu_segcblist_disable(&rdp->cblist);
2587         return true;
2588 }
2589 
2590 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2591 
2592 static bool rcu_nocb_cpu_needs_barrier(int cpu)
2593 {
2594         WARN_ON_ONCE(1); /* Should be dead code. */
2595         return false;
2596 }
2597 
2598 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2599 {
2600 }
2601 
2602 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2603 {
2604         return NULL;
2605 }
2606 
2607 static void rcu_init_one_nocb(struct rcu_node *rnp)
2608 {
2609 }
2610 
2611 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2612                             bool lazy, unsigned long flags)
2613 {
2614         return false;
2615 }
2616 
2617 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
2618                                                      struct rcu_data *rdp,
2619                                                      unsigned long flags)
2620 {
2621         return false;
2622 }
2623 
2624 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2625 {
2626 }
2627 
2628 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2629 {
2630         return false;
2631 }
2632 
2633 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2634 {
2635 }
2636 
2637 static void rcu_spawn_all_nocb_kthreads(int cpu)
2638 {
2639 }
2640 
2641 static void __init rcu_spawn_nocb_kthreads(void)
2642 {
2643 }
2644 
2645 static bool init_nocb_callback_list(struct rcu_data *rdp)
2646 {
2647         return false;
2648 }
2649 
2650 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2651 
2652 /*
2653  * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2654  * grace-period kthread will do force_quiescent_state() processing?
2655  * The idea is to avoid waking up RCU core processing on such a
2656  * CPU unless the grace period has extended for too long.
2657  *
2658  * This code relies on the fact that all NO_HZ_FULL CPUs are also
2659  * CONFIG_RCU_NOCB_CPU CPUs.
2660  */
2661 static bool rcu_nohz_full_cpu(void)
2662 {
2663 #ifdef CONFIG_NO_HZ_FULL
2664         if (tick_nohz_full_cpu(smp_processor_id()) &&
2665             (!rcu_gp_in_progress() ||
2666              ULONG_CMP_LT(jiffies, READ_ONCE(rcu_state.gp_start) + HZ)))
2667                 return true;
2668 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2669         return false;
2670 }
2671 
2672 /*
2673  * Bind the RCU grace-period kthreads to the housekeeping CPU.
2674  */
2675 static void rcu_bind_gp_kthread(void)
2676 {
2677         if (!tick_nohz_full_enabled())
2678                 return;
2679         housekeeping_affine(current, HK_FLAG_RCU);
2680 }
2681 
2682 /* Record the current task on dyntick-idle entry. */
2683 static void rcu_dynticks_task_enter(void)
2684 {
2685 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2686         WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2687 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2688 }
2689 
2690 /* Record no current task on dyntick-idle exit. */
2691 static void rcu_dynticks_task_exit(void)
2692 {
2693 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2694         WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2695 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2696 }
2697 

~ [ source navigation ] ~ [ diff markup ] ~ [ identifier search ] ~

kernel.org | git.kernel.org | LWN.net | Project Home | Wiki (Japanese) | Wiki (English) | SVN repository | Mail admin

Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.

osdn.jp