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

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

Version: ~ [ linux-4.19-rc5 ] ~ [ linux-4.18.9 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.71 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.128 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.157 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.122 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.57 ] ~ [ 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-2.6.31.14 ] ~ [ linux-2.6.30.10 ] ~ [ linux-2.6.29.6 ] ~ [ linux-2.6.28.10 ] ~ [ linux-2.6.27.62 ] ~ [ 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.  These
 42  * handle all flavors of RCU.
 43  */
 44 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
 45 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
 46 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
 47 DEFINE_PER_CPU(char, rcu_cpu_has_work);
 48 
 49 #else /* #ifdef CONFIG_RCU_BOOST */
 50 
 51 /*
 52  * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST,
 53  * all uses are in dead code.  Provide a definition to keep the compiler
 54  * happy, but add WARN_ON_ONCE() to complain if used in the wrong place.
 55  * This probably needs to be excluded from -rt builds.
 56  */
 57 #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; })
 58 #define rt_mutex_futex_unlock(x) WARN_ON_ONCE(1)
 59 
 60 #endif /* #else #ifdef CONFIG_RCU_BOOST */
 61 
 62 #ifdef CONFIG_RCU_NOCB_CPU
 63 static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
 64 static bool __read_mostly rcu_nocb_poll;    /* Offload kthread are to poll. */
 65 #endif /* #ifdef CONFIG_RCU_NOCB_CPU */
 66 
 67 /*
 68  * Check the RCU kernel configuration parameters and print informative
 69  * messages about anything out of the ordinary.
 70  */
 71 static void __init rcu_bootup_announce_oddness(void)
 72 {
 73         if (IS_ENABLED(CONFIG_RCU_TRACE))
 74                 pr_info("\tRCU event tracing is enabled.\n");
 75         if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) ||
 76             (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32))
 77                 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d.\n",
 78                         RCU_FANOUT);
 79         if (rcu_fanout_exact)
 80                 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
 81         if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ))
 82                 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
 83         if (IS_ENABLED(CONFIG_PROVE_RCU))
 84                 pr_info("\tRCU lockdep checking is enabled.\n");
 85         if (RCU_NUM_LVLS >= 4)
 86                 pr_info("\tFour(or more)-level hierarchy is enabled.\n");
 87         if (RCU_FANOUT_LEAF != 16)
 88                 pr_info("\tBuild-time adjustment of leaf fanout to %d.\n",
 89                         RCU_FANOUT_LEAF);
 90         if (rcu_fanout_leaf != RCU_FANOUT_LEAF)
 91                 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n",
 92                         rcu_fanout_leaf);
 93         if (nr_cpu_ids != NR_CPUS)
 94                 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%u.\n", NR_CPUS, nr_cpu_ids);
 95 #ifdef CONFIG_RCU_BOOST
 96         pr_info("\tRCU priority boosting: priority %d delay %d ms.\n",
 97                 kthread_prio, CONFIG_RCU_BOOST_DELAY);
 98 #endif
 99         if (blimit != DEFAULT_RCU_BLIMIT)
100                 pr_info("\tBoot-time adjustment of callback invocation limit to %ld.\n", blimit);
101         if (qhimark != DEFAULT_RCU_QHIMARK)
102                 pr_info("\tBoot-time adjustment of callback high-water mark to %ld.\n", qhimark);
103         if (qlowmark != DEFAULT_RCU_QLOMARK)
104                 pr_info("\tBoot-time adjustment of callback low-water mark to %ld.\n", qlowmark);
105         if (jiffies_till_first_fqs != ULONG_MAX)
106                 pr_info("\tBoot-time adjustment of first FQS scan delay to %ld jiffies.\n", jiffies_till_first_fqs);
107         if (jiffies_till_next_fqs != ULONG_MAX)
108                 pr_info("\tBoot-time adjustment of subsequent FQS scan delay to %ld jiffies.\n", jiffies_till_next_fqs);
109         if (rcu_kick_kthreads)
110                 pr_info("\tKick kthreads if too-long grace period.\n");
111         if (IS_ENABLED(CONFIG_DEBUG_OBJECTS_RCU_HEAD))
112                 pr_info("\tRCU callback double-/use-after-free debug enabled.\n");
113         if (gp_preinit_delay)
114                 pr_info("\tRCU debug GP pre-init slowdown %d jiffies.\n", gp_preinit_delay);
115         if (gp_init_delay)
116                 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_init_delay);
117         if (gp_cleanup_delay)
118                 pr_info("\tRCU debug GP init slowdown %d jiffies.\n", gp_cleanup_delay);
119         if (IS_ENABLED(CONFIG_RCU_EQS_DEBUG))
120                 pr_info("\tRCU debug extended QS entry/exit.\n");
121         rcupdate_announce_bootup_oddness();
122 }
123 
124 #ifdef CONFIG_PREEMPT_RCU
125 
126 RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
127 static struct rcu_state *const rcu_state_p = &rcu_preempt_state;
128 static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data;
129 
130 static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
131                                bool wake);
132 static void rcu_read_unlock_special(struct task_struct *t);
133 
134 /*
135  * Tell them what RCU they are running.
136  */
137 static void __init rcu_bootup_announce(void)
138 {
139         pr_info("Preemptible hierarchical RCU implementation.\n");
140         rcu_bootup_announce_oddness();
141 }
142 
143 /* Flags for rcu_preempt_ctxt_queue() decision table. */
144 #define RCU_GP_TASKS    0x8
145 #define RCU_EXP_TASKS   0x4
146 #define RCU_GP_BLKD     0x2
147 #define RCU_EXP_BLKD    0x1
148 
149 /*
150  * Queues a task preempted within an RCU-preempt read-side critical
151  * section into the appropriate location within the ->blkd_tasks list,
152  * depending on the states of any ongoing normal and expedited grace
153  * periods.  The ->gp_tasks pointer indicates which element the normal
154  * grace period is waiting on (NULL if none), and the ->exp_tasks pointer
155  * indicates which element the expedited grace period is waiting on (again,
156  * NULL if none).  If a grace period is waiting on a given element in the
157  * ->blkd_tasks list, it also waits on all subsequent elements.  Thus,
158  * adding a task to the tail of the list blocks any grace period that is
159  * already waiting on one of the elements.  In contrast, adding a task
160  * to the head of the list won't block any grace period that is already
161  * waiting on one of the elements.
162  *
163  * This queuing is imprecise, and can sometimes make an ongoing grace
164  * period wait for a task that is not strictly speaking blocking it.
165  * Given the choice, we needlessly block a normal grace period rather than
166  * blocking an expedited grace period.
167  *
168  * Note that an endless sequence of expedited grace periods still cannot
169  * indefinitely postpone a normal grace period.  Eventually, all of the
170  * fixed number of preempted tasks blocking the normal grace period that are
171  * not also blocking the expedited grace period will resume and complete
172  * their RCU read-side critical sections.  At that point, the ->gp_tasks
173  * pointer will equal the ->exp_tasks pointer, at which point the end of
174  * the corresponding expedited grace period will also be the end of the
175  * normal grace period.
176  */
177 static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp)
178         __releases(rnp->lock) /* But leaves rrupts disabled. */
179 {
180         int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) +
181                          (rnp->exp_tasks ? RCU_EXP_TASKS : 0) +
182                          (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) +
183                          (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0);
184         struct task_struct *t = current;
185 
186         raw_lockdep_assert_held_rcu_node(rnp);
187         WARN_ON_ONCE(rdp->mynode != rnp);
188         WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
189         /* RCU better not be waiting on newly onlined CPUs! */
190         WARN_ON_ONCE(rnp->qsmaskinitnext & ~rnp->qsmaskinit & rnp->qsmask &
191                      rdp->grpmask);
192 
193         /*
194          * Decide where to queue the newly blocked task.  In theory,
195          * this could be an if-statement.  In practice, when I tried
196          * that, it was quite messy.
197          */
198         switch (blkd_state) {
199         case 0:
200         case                RCU_EXP_TASKS:
201         case                RCU_EXP_TASKS + RCU_GP_BLKD:
202         case RCU_GP_TASKS:
203         case RCU_GP_TASKS + RCU_EXP_TASKS:
204 
205                 /*
206                  * Blocking neither GP, or first task blocking the normal
207                  * GP but not blocking the already-waiting expedited GP.
208                  * Queue at the head of the list to avoid unnecessarily
209                  * blocking the already-waiting GPs.
210                  */
211                 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
212                 break;
213 
214         case                                              RCU_EXP_BLKD:
215         case                                RCU_GP_BLKD:
216         case                                RCU_GP_BLKD + RCU_EXP_BLKD:
217         case RCU_GP_TASKS +                               RCU_EXP_BLKD:
218         case RCU_GP_TASKS +                 RCU_GP_BLKD + RCU_EXP_BLKD:
219         case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
220 
221                 /*
222                  * First task arriving that blocks either GP, or first task
223                  * arriving that blocks the expedited GP (with the normal
224                  * GP already waiting), or a task arriving that blocks
225                  * both GPs with both GPs already waiting.  Queue at the
226                  * tail of the list to avoid any GP waiting on any of the
227                  * already queued tasks that are not blocking it.
228                  */
229                 list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks);
230                 break;
231 
232         case                RCU_EXP_TASKS +               RCU_EXP_BLKD:
233         case                RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD:
234         case RCU_GP_TASKS + RCU_EXP_TASKS +               RCU_EXP_BLKD:
235 
236                 /*
237                  * Second or subsequent task blocking the expedited GP.
238                  * The task either does not block the normal GP, or is the
239                  * first task blocking the normal GP.  Queue just after
240                  * the first task blocking the expedited GP.
241                  */
242                 list_add(&t->rcu_node_entry, rnp->exp_tasks);
243                 break;
244 
245         case RCU_GP_TASKS +                 RCU_GP_BLKD:
246         case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD:
247 
248                 /*
249                  * Second or subsequent task blocking the normal GP.
250                  * The task does not block the expedited GP. Queue just
251                  * after the first task blocking the normal GP.
252                  */
253                 list_add(&t->rcu_node_entry, rnp->gp_tasks);
254                 break;
255 
256         default:
257 
258                 /* Yet another exercise in excessive paranoia. */
259                 WARN_ON_ONCE(1);
260                 break;
261         }
262 
263         /*
264          * We have now queued the task.  If it was the first one to
265          * block either grace period, update the ->gp_tasks and/or
266          * ->exp_tasks pointers, respectively, to reference the newly
267          * blocked tasks.
268          */
269         if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) {
270                 rnp->gp_tasks = &t->rcu_node_entry;
271                 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq);
272         }
273         if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD))
274                 rnp->exp_tasks = &t->rcu_node_entry;
275         WARN_ON_ONCE(!(blkd_state & RCU_GP_BLKD) !=
276                      !(rnp->qsmask & rdp->grpmask));
277         WARN_ON_ONCE(!(blkd_state & RCU_EXP_BLKD) !=
278                      !(rnp->expmask & rdp->grpmask));
279         raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */
280 
281         /*
282          * Report the quiescent state for the expedited GP.  This expedited
283          * GP should not be able to end until we report, so there should be
284          * no need to check for a subsequent expedited GP.  (Though we are
285          * still in a quiescent state in any case.)
286          */
287         if (blkd_state & RCU_EXP_BLKD &&
288             t->rcu_read_unlock_special.b.exp_need_qs) {
289                 t->rcu_read_unlock_special.b.exp_need_qs = false;
290                 rcu_report_exp_rdp(rdp->rsp, rdp, true);
291         } else {
292                 WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs);
293         }
294 }
295 
296 /*
297  * Record a preemptible-RCU quiescent state for the specified CPU.
298  * Note that this does not necessarily mean that the task currently running
299  * on the CPU is in a quiescent state:  Instead, it means that the current
300  * grace period need not wait on any RCU read-side critical section that
301  * starts later on this CPU.  It also means that if the current task is
302  * in an RCU read-side critical section, it has already added itself to
303  * some leaf rcu_node structure's ->blkd_tasks list.  In addition to the
304  * current task, there might be any number of other tasks blocked while
305  * in an RCU read-side critical section.
306  *
307  * Callers to this function must disable preemption.
308  */
309 static void rcu_preempt_qs(void)
310 {
311         RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_qs() invoked with preemption enabled!!!\n");
312         if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) {
313                 trace_rcu_grace_period(TPS("rcu_preempt"),
314                                        __this_cpu_read(rcu_data_p->gp_seq),
315                                        TPS("cpuqs"));
316                 __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false);
317                 barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */
318                 current->rcu_read_unlock_special.b.need_qs = false;
319         }
320 }
321 
322 /*
323  * We have entered the scheduler, and the current task might soon be
324  * context-switched away from.  If this task is in an RCU read-side
325  * critical section, we will no longer be able to rely on the CPU to
326  * record that fact, so we enqueue the task on the blkd_tasks list.
327  * The task will dequeue itself when it exits the outermost enclosing
328  * RCU read-side critical section.  Therefore, the current grace period
329  * cannot be permitted to complete until the blkd_tasks list entries
330  * predating the current grace period drain, in other words, until
331  * rnp->gp_tasks becomes NULL.
332  *
333  * Caller must disable interrupts.
334  */
335 static void rcu_preempt_note_context_switch(bool preempt)
336 {
337         struct task_struct *t = current;
338         struct rcu_data *rdp;
339         struct rcu_node *rnp;
340 
341         lockdep_assert_irqs_disabled();
342         WARN_ON_ONCE(!preempt && t->rcu_read_lock_nesting > 0);
343         if (t->rcu_read_lock_nesting > 0 &&
344             !t->rcu_read_unlock_special.b.blocked) {
345 
346                 /* Possibly blocking in an RCU read-side critical section. */
347                 rdp = this_cpu_ptr(rcu_state_p->rda);
348                 rnp = rdp->mynode;
349                 raw_spin_lock_rcu_node(rnp);
350                 t->rcu_read_unlock_special.b.blocked = true;
351                 t->rcu_blocked_node = rnp;
352 
353                 /*
354                  * Verify the CPU's sanity, trace the preemption, and
355                  * then queue the task as required based on the states
356                  * of any ongoing and expedited grace periods.
357                  */
358                 WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0);
359                 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
360                 trace_rcu_preempt_task(rdp->rsp->name,
361                                        t->pid,
362                                        (rnp->qsmask & rdp->grpmask)
363                                        ? rnp->gp_seq
364                                        : rcu_seq_snap(&rnp->gp_seq));
365                 rcu_preempt_ctxt_queue(rnp, rdp);
366         } else if (t->rcu_read_lock_nesting < 0 &&
367                    t->rcu_read_unlock_special.s) {
368 
369                 /*
370                  * Complete exit from RCU read-side critical section on
371                  * behalf of preempted instance of __rcu_read_unlock().
372                  */
373                 rcu_read_unlock_special(t);
374         }
375 
376         /*
377          * Either we were not in an RCU read-side critical section to
378          * begin with, or we have now recorded that critical section
379          * globally.  Either way, we can now note a quiescent state
380          * for this CPU.  Again, if we were in an RCU read-side critical
381          * section, and if that critical section was blocking the current
382          * grace period, then the fact that the task has been enqueued
383          * means that we continue to block the current grace period.
384          */
385         rcu_preempt_qs();
386 }
387 
388 /*
389  * Check for preempted RCU readers blocking the current grace period
390  * for the specified rcu_node structure.  If the caller needs a reliable
391  * answer, it must hold the rcu_node's ->lock.
392  */
393 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
394 {
395         return rnp->gp_tasks != NULL;
396 }
397 
398 /*
399  * Preemptible RCU implementation for rcu_read_lock().
400  * Just increment ->rcu_read_lock_nesting, shared state will be updated
401  * if we block.
402  */
403 void __rcu_read_lock(void)
404 {
405         current->rcu_read_lock_nesting++;
406         barrier();  /* critical section after entry code. */
407 }
408 EXPORT_SYMBOL_GPL(__rcu_read_lock);
409 
410 /*
411  * Preemptible RCU implementation for rcu_read_unlock().
412  * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
413  * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
414  * invoke rcu_read_unlock_special() to clean up after a context switch
415  * in an RCU read-side critical section and other special cases.
416  */
417 void __rcu_read_unlock(void)
418 {
419         struct task_struct *t = current;
420 
421         if (t->rcu_read_lock_nesting != 1) {
422                 --t->rcu_read_lock_nesting;
423         } else {
424                 barrier();  /* critical section before exit code. */
425                 t->rcu_read_lock_nesting = INT_MIN;
426                 barrier();  /* assign before ->rcu_read_unlock_special load */
427                 if (unlikely(READ_ONCE(t->rcu_read_unlock_special.s)))
428                         rcu_read_unlock_special(t);
429                 barrier();  /* ->rcu_read_unlock_special load before assign */
430                 t->rcu_read_lock_nesting = 0;
431         }
432 #ifdef CONFIG_PROVE_LOCKING
433         {
434                 int rrln = READ_ONCE(t->rcu_read_lock_nesting);
435 
436                 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
437         }
438 #endif /* #ifdef CONFIG_PROVE_LOCKING */
439 }
440 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
441 
442 /*
443  * Advance a ->blkd_tasks-list pointer to the next entry, instead
444  * returning NULL if at the end of the list.
445  */
446 static struct list_head *rcu_next_node_entry(struct task_struct *t,
447                                              struct rcu_node *rnp)
448 {
449         struct list_head *np;
450 
451         np = t->rcu_node_entry.next;
452         if (np == &rnp->blkd_tasks)
453                 np = NULL;
454         return np;
455 }
456 
457 /*
458  * Return true if the specified rcu_node structure has tasks that were
459  * preempted within an RCU read-side critical section.
460  */
461 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
462 {
463         return !list_empty(&rnp->blkd_tasks);
464 }
465 
466 /*
467  * Handle special cases during rcu_read_unlock(), such as needing to
468  * notify RCU core processing or task having blocked during the RCU
469  * read-side critical section.
470  */
471 static void rcu_read_unlock_special(struct task_struct *t)
472 {
473         bool empty_exp;
474         bool empty_norm;
475         bool empty_exp_now;
476         unsigned long flags;
477         struct list_head *np;
478         bool drop_boost_mutex = false;
479         struct rcu_data *rdp;
480         struct rcu_node *rnp;
481         union rcu_special special;
482 
483         /* NMI handlers cannot block and cannot safely manipulate state. */
484         if (in_nmi())
485                 return;
486 
487         local_irq_save(flags);
488 
489         /*
490          * If RCU core is waiting for this CPU to exit its critical section,
491          * report the fact that it has exited.  Because irqs are disabled,
492          * t->rcu_read_unlock_special cannot change.
493          */
494         special = t->rcu_read_unlock_special;
495         if (special.b.need_qs) {
496                 rcu_preempt_qs();
497                 t->rcu_read_unlock_special.b.need_qs = false;
498                 if (!t->rcu_read_unlock_special.s) {
499                         local_irq_restore(flags);
500                         return;
501                 }
502         }
503 
504         /*
505          * Respond to a request for an expedited grace period, but only if
506          * we were not preempted, meaning that we were running on the same
507          * CPU throughout.  If we were preempted, the exp_need_qs flag
508          * would have been cleared at the time of the first preemption,
509          * and the quiescent state would be reported when we were dequeued.
510          */
511         if (special.b.exp_need_qs) {
512                 WARN_ON_ONCE(special.b.blocked);
513                 t->rcu_read_unlock_special.b.exp_need_qs = false;
514                 rdp = this_cpu_ptr(rcu_state_p->rda);
515                 rcu_report_exp_rdp(rcu_state_p, rdp, true);
516                 if (!t->rcu_read_unlock_special.s) {
517                         local_irq_restore(flags);
518                         return;
519                 }
520         }
521 
522         /* Hardware IRQ handlers cannot block, complain if they get here. */
523         if (in_irq() || in_serving_softirq()) {
524                 lockdep_rcu_suspicious(__FILE__, __LINE__,
525                                        "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n");
526                 pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n",
527                          t->rcu_read_unlock_special.s,
528                          t->rcu_read_unlock_special.b.blocked,
529                          t->rcu_read_unlock_special.b.exp_need_qs,
530                          t->rcu_read_unlock_special.b.need_qs);
531                 local_irq_restore(flags);
532                 return;
533         }
534 
535         /* Clean up if blocked during RCU read-side critical section. */
536         if (special.b.blocked) {
537                 t->rcu_read_unlock_special.b.blocked = false;
538 
539                 /*
540                  * Remove this task from the list it blocked on.  The task
541                  * now remains queued on the rcu_node corresponding to the
542                  * CPU it first blocked on, so there is no longer any need
543                  * to loop.  Retain a WARN_ON_ONCE() out of sheer paranoia.
544                  */
545                 rnp = t->rcu_blocked_node;
546                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
547                 WARN_ON_ONCE(rnp != t->rcu_blocked_node);
548                 WARN_ON_ONCE(!rcu_is_leaf_node(rnp));
549                 empty_norm = !rcu_preempt_blocked_readers_cgp(rnp);
550                 WARN_ON_ONCE(rnp->completedqs == rnp->gp_seq &&
551                              (!empty_norm || rnp->qsmask));
552                 empty_exp = sync_rcu_preempt_exp_done(rnp);
553                 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
554                 np = rcu_next_node_entry(t, rnp);
555                 list_del_init(&t->rcu_node_entry);
556                 t->rcu_blocked_node = NULL;
557                 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
558                                                 rnp->gp_seq, t->pid);
559                 if (&t->rcu_node_entry == rnp->gp_tasks)
560                         rnp->gp_tasks = np;
561                 if (&t->rcu_node_entry == rnp->exp_tasks)
562                         rnp->exp_tasks = np;
563                 if (IS_ENABLED(CONFIG_RCU_BOOST)) {
564                         /* Snapshot ->boost_mtx ownership w/rnp->lock held. */
565                         drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t;
566                         if (&t->rcu_node_entry == rnp->boost_tasks)
567                                 rnp->boost_tasks = np;
568                 }
569 
570                 /*
571                  * If this was the last task on the current list, and if
572                  * we aren't waiting on any CPUs, report the quiescent state.
573                  * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
574                  * so we must take a snapshot of the expedited state.
575                  */
576                 empty_exp_now = sync_rcu_preempt_exp_done(rnp);
577                 if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) {
578                         trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
579                                                          rnp->gp_seq,
580                                                          0, rnp->qsmask,
581                                                          rnp->level,
582                                                          rnp->grplo,
583                                                          rnp->grphi,
584                                                          !!rnp->gp_tasks);
585                         rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags);
586                 } else {
587                         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
588                 }
589 
590                 /* Unboost if we were boosted. */
591                 if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex)
592                         rt_mutex_futex_unlock(&rnp->boost_mtx);
593 
594                 /*
595                  * If this was the last task on the expedited lists,
596                  * then we need to report up the rcu_node hierarchy.
597                  */
598                 if (!empty_exp && empty_exp_now)
599                         rcu_report_exp_rnp(rcu_state_p, rnp, true);
600         } else {
601                 local_irq_restore(flags);
602         }
603 }
604 
605 /*
606  * Dump detailed information for all tasks blocking the current RCU
607  * grace period on the specified rcu_node structure.
608  */
609 static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
610 {
611         unsigned long flags;
612         struct task_struct *t;
613 
614         raw_spin_lock_irqsave_rcu_node(rnp, flags);
615         if (!rcu_preempt_blocked_readers_cgp(rnp)) {
616                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
617                 return;
618         }
619         t = list_entry(rnp->gp_tasks->prev,
620                        struct task_struct, rcu_node_entry);
621         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
622                 /*
623                  * We could be printing a lot while holding a spinlock.
624                  * Avoid triggering hard lockup.
625                  */
626                 touch_nmi_watchdog();
627                 sched_show_task(t);
628         }
629         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
630 }
631 
632 /*
633  * Dump detailed information for all tasks blocking the current RCU
634  * grace period.
635  */
636 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
637 {
638         struct rcu_node *rnp = rcu_get_root(rsp);
639 
640         rcu_print_detail_task_stall_rnp(rnp);
641         rcu_for_each_leaf_node(rsp, rnp)
642                 rcu_print_detail_task_stall_rnp(rnp);
643 }
644 
645 static void rcu_print_task_stall_begin(struct rcu_node *rnp)
646 {
647         pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
648                rnp->level, rnp->grplo, rnp->grphi);
649 }
650 
651 static void rcu_print_task_stall_end(void)
652 {
653         pr_cont("\n");
654 }
655 
656 /*
657  * Scan the current list of tasks blocked within RCU read-side critical
658  * sections, printing out the tid of each.
659  */
660 static int rcu_print_task_stall(struct rcu_node *rnp)
661 {
662         struct task_struct *t;
663         int ndetected = 0;
664 
665         if (!rcu_preempt_blocked_readers_cgp(rnp))
666                 return 0;
667         rcu_print_task_stall_begin(rnp);
668         t = list_entry(rnp->gp_tasks->prev,
669                        struct task_struct, rcu_node_entry);
670         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
671                 pr_cont(" P%d", t->pid);
672                 ndetected++;
673         }
674         rcu_print_task_stall_end();
675         return ndetected;
676 }
677 
678 /*
679  * Scan the current list of tasks blocked within RCU read-side critical
680  * sections, printing out the tid of each that is blocking the current
681  * expedited grace period.
682  */
683 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
684 {
685         struct task_struct *t;
686         int ndetected = 0;
687 
688         if (!rnp->exp_tasks)
689                 return 0;
690         t = list_entry(rnp->exp_tasks->prev,
691                        struct task_struct, rcu_node_entry);
692         list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
693                 pr_cont(" P%d", t->pid);
694                 ndetected++;
695         }
696         return ndetected;
697 }
698 
699 /*
700  * Check that the list of blocked tasks for the newly completed grace
701  * period is in fact empty.  It is a serious bug to complete a grace
702  * period that still has RCU readers blocked!  This function must be
703  * invoked -before- updating this rnp's ->gp_seq, and the rnp's ->lock
704  * must be held by the caller.
705  *
706  * Also, if there are blocked tasks on the list, they automatically
707  * block the newly created grace period, so set up ->gp_tasks accordingly.
708  */
709 static void
710 rcu_preempt_check_blocked_tasks(struct rcu_state *rsp, struct rcu_node *rnp)
711 {
712         struct task_struct *t;
713 
714         RCU_LOCKDEP_WARN(preemptible(), "rcu_preempt_check_blocked_tasks() invoked with preemption enabled!!!\n");
715         if (WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)))
716                 dump_blkd_tasks(rsp, rnp, 10);
717         if (rcu_preempt_has_tasks(rnp) &&
718             (rnp->qsmaskinit || rnp->wait_blkd_tasks)) {
719                 rnp->gp_tasks = rnp->blkd_tasks.next;
720                 t = container_of(rnp->gp_tasks, struct task_struct,
721                                  rcu_node_entry);
722                 trace_rcu_unlock_preempted_task(TPS("rcu_preempt-GPS"),
723                                                 rnp->gp_seq, t->pid);
724         }
725         WARN_ON_ONCE(rnp->qsmask);
726 }
727 
728 /*
729  * Check for a quiescent state from the current CPU.  When a task blocks,
730  * the task is recorded in the corresponding CPU's rcu_node structure,
731  * which is checked elsewhere.
732  *
733  * Caller must disable hard irqs.
734  */
735 static void rcu_preempt_check_callbacks(void)
736 {
737         struct rcu_state *rsp = &rcu_preempt_state;
738         struct task_struct *t = current;
739 
740         if (t->rcu_read_lock_nesting == 0) {
741                 rcu_preempt_qs();
742                 return;
743         }
744         if (t->rcu_read_lock_nesting > 0 &&
745             __this_cpu_read(rcu_data_p->core_needs_qs) &&
746             __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm) &&
747             !t->rcu_read_unlock_special.b.need_qs &&
748             time_after(jiffies, rsp->gp_start + HZ))
749                 t->rcu_read_unlock_special.b.need_qs = true;
750 }
751 
752 /**
753  * call_rcu() - Queue an RCU callback for invocation after a grace period.
754  * @head: structure to be used for queueing the RCU updates.
755  * @func: actual callback function to be invoked after the grace period
756  *
757  * The callback function will be invoked some time after a full grace
758  * period elapses, in other words after all pre-existing RCU read-side
759  * critical sections have completed.  However, the callback function
760  * might well execute concurrently with RCU read-side critical sections
761  * that started after call_rcu() was invoked.  RCU read-side critical
762  * sections are delimited by rcu_read_lock() and rcu_read_unlock(),
763  * and may be nested.
764  *
765  * Note that all CPUs must agree that the grace period extended beyond
766  * all pre-existing RCU read-side critical section.  On systems with more
767  * than one CPU, this means that when "func()" is invoked, each CPU is
768  * guaranteed to have executed a full memory barrier since the end of its
769  * last RCU read-side critical section whose beginning preceded the call
770  * to call_rcu().  It also means that each CPU executing an RCU read-side
771  * critical section that continues beyond the start of "func()" must have
772  * executed a memory barrier after the call_rcu() but before the beginning
773  * of that RCU read-side critical section.  Note that these guarantees
774  * include CPUs that are offline, idle, or executing in user mode, as
775  * well as CPUs that are executing in the kernel.
776  *
777  * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
778  * resulting RCU callback function "func()", then both CPU A and CPU B are
779  * guaranteed to execute a full memory barrier during the time interval
780  * between the call to call_rcu() and the invocation of "func()" -- even
781  * if CPU A and CPU B are the same CPU (but again only if the system has
782  * more than one CPU).
783  */
784 void call_rcu(struct rcu_head *head, rcu_callback_t func)
785 {
786         __call_rcu(head, func, rcu_state_p, -1, 0);
787 }
788 EXPORT_SYMBOL_GPL(call_rcu);
789 
790 /**
791  * synchronize_rcu - wait until a grace period has elapsed.
792  *
793  * Control will return to the caller some time after a full grace
794  * period has elapsed, in other words after all currently executing RCU
795  * read-side critical sections have completed.  Note, however, that
796  * upon return from synchronize_rcu(), the caller might well be executing
797  * concurrently with new RCU read-side critical sections that began while
798  * synchronize_rcu() was waiting.  RCU read-side critical sections are
799  * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
800  *
801  * See the description of synchronize_sched() for more detailed
802  * information on memory-ordering guarantees.  However, please note
803  * that -only- the memory-ordering guarantees apply.  For example,
804  * synchronize_rcu() is -not- guaranteed to wait on things like code
805  * protected by preempt_disable(), instead, synchronize_rcu() is -only-
806  * guaranteed to wait on RCU read-side critical sections, that is, sections
807  * of code protected by rcu_read_lock().
808  */
809 void synchronize_rcu(void)
810 {
811         RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) ||
812                          lock_is_held(&rcu_lock_map) ||
813                          lock_is_held(&rcu_sched_lock_map),
814                          "Illegal synchronize_rcu() in RCU read-side critical section");
815         if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
816                 return;
817         if (rcu_gp_is_expedited())
818                 synchronize_rcu_expedited();
819         else
820                 wait_rcu_gp(call_rcu);
821 }
822 EXPORT_SYMBOL_GPL(synchronize_rcu);
823 
824 /**
825  * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
826  *
827  * Note that this primitive does not necessarily wait for an RCU grace period
828  * to complete.  For example, if there are no RCU callbacks queued anywhere
829  * in the system, then rcu_barrier() is within its rights to return
830  * immediately, without waiting for anything, much less an RCU grace period.
831  */
832 void rcu_barrier(void)
833 {
834         _rcu_barrier(rcu_state_p);
835 }
836 EXPORT_SYMBOL_GPL(rcu_barrier);
837 
838 /*
839  * Initialize preemptible RCU's state structures.
840  */
841 static void __init __rcu_init_preempt(void)
842 {
843         rcu_init_one(rcu_state_p);
844 }
845 
846 /*
847  * Check for a task exiting while in a preemptible-RCU read-side
848  * critical section, clean up if so.  No need to issue warnings,
849  * as debug_check_no_locks_held() already does this if lockdep
850  * is enabled.
851  */
852 void exit_rcu(void)
853 {
854         struct task_struct *t = current;
855 
856         if (likely(list_empty(&current->rcu_node_entry)))
857                 return;
858         t->rcu_read_lock_nesting = 1;
859         barrier();
860         t->rcu_read_unlock_special.b.blocked = true;
861         __rcu_read_unlock();
862 }
863 
864 /*
865  * Dump the blocked-tasks state, but limit the list dump to the
866  * specified number of elements.
867  */
868 static void
869 dump_blkd_tasks(struct rcu_state *rsp, struct rcu_node *rnp, int ncheck)
870 {
871         int cpu;
872         int i;
873         struct list_head *lhp;
874         bool onl;
875         struct rcu_data *rdp;
876         struct rcu_node *rnp1;
877 
878         raw_lockdep_assert_held_rcu_node(rnp);
879         pr_info("%s: grp: %d-%d level: %d ->gp_seq %ld ->completedqs %ld\n",
880                 __func__, rnp->grplo, rnp->grphi, rnp->level,
881                 (long)rnp->gp_seq, (long)rnp->completedqs);
882         for (rnp1 = rnp; rnp1; rnp1 = rnp1->parent)
883                 pr_info("%s: %d:%d ->qsmask %#lx ->qsmaskinit %#lx ->qsmaskinitnext %#lx\n",
884                         __func__, rnp1->grplo, rnp1->grphi, rnp1->qsmask, rnp1->qsmaskinit, rnp1->qsmaskinitnext);
885         pr_info("%s: ->gp_tasks %p ->boost_tasks %p ->exp_tasks %p\n",
886                 __func__, rnp->gp_tasks, rnp->boost_tasks, rnp->exp_tasks);
887         pr_info("%s: ->blkd_tasks", __func__);
888         i = 0;
889         list_for_each(lhp, &rnp->blkd_tasks) {
890                 pr_cont(" %p", lhp);
891                 if (++i >= 10)
892                         break;
893         }
894         pr_cont("\n");
895         for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) {
896                 rdp = per_cpu_ptr(rsp->rda, cpu);
897                 onl = !!(rdp->grpmask & rcu_rnp_online_cpus(rnp));
898                 pr_info("\t%d: %c online: %ld(%d) offline: %ld(%d)\n",
899                         cpu, ".o"[onl],
900                         (long)rdp->rcu_onl_gp_seq, rdp->rcu_onl_gp_flags,
901                         (long)rdp->rcu_ofl_gp_seq, rdp->rcu_ofl_gp_flags);
902         }
903 }
904 
905 #else /* #ifdef CONFIG_PREEMPT_RCU */
906 
907 static struct rcu_state *const rcu_state_p = &rcu_sched_state;
908 
909 /*
910  * Tell them what RCU they are running.
911  */
912 static void __init rcu_bootup_announce(void)
913 {
914         pr_info("Hierarchical RCU implementation.\n");
915         rcu_bootup_announce_oddness();
916 }
917 
918 /*
919  * Because preemptible RCU does not exist, we never have to check for
920  * CPUs being in quiescent states.
921  */
922 static void rcu_preempt_note_context_switch(bool preempt)
923 {
924 }
925 
926 /*
927  * Because preemptible RCU does not exist, there are never any preempted
928  * RCU readers.
929  */
930 static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
931 {
932         return 0;
933 }
934 
935 /*
936  * Because there is no preemptible RCU, there can be no readers blocked.
937  */
938 static bool rcu_preempt_has_tasks(struct rcu_node *rnp)
939 {
940         return false;
941 }
942 
943 /*
944  * Because preemptible RCU does not exist, we never have to check for
945  * tasks blocked within RCU read-side critical sections.
946  */
947 static void rcu_print_detail_task_stall(struct rcu_state *rsp)
948 {
949 }
950 
951 /*
952  * Because preemptible RCU does not exist, we never have to check for
953  * tasks blocked within RCU read-side critical sections.
954  */
955 static int rcu_print_task_stall(struct rcu_node *rnp)
956 {
957         return 0;
958 }
959 
960 /*
961  * Because preemptible RCU does not exist, we never have to check for
962  * tasks blocked within RCU read-side critical sections that are
963  * blocking the current expedited grace period.
964  */
965 static int rcu_print_task_exp_stall(struct rcu_node *rnp)
966 {
967         return 0;
968 }
969 
970 /*
971  * Because there is no preemptible RCU, there can be no readers blocked,
972  * so there is no need to check for blocked tasks.  So check only for
973  * bogus qsmask values.
974  */
975 static void
976 rcu_preempt_check_blocked_tasks(struct rcu_state *rsp, struct rcu_node *rnp)
977 {
978         WARN_ON_ONCE(rnp->qsmask);
979 }
980 
981 /*
982  * Because preemptible RCU does not exist, it never has any callbacks
983  * to check.
984  */
985 static void rcu_preempt_check_callbacks(void)
986 {
987 }
988 
989 /*
990  * Because preemptible RCU does not exist, rcu_barrier() is just
991  * another name for rcu_barrier_sched().
992  */
993 void rcu_barrier(void)
994 {
995         rcu_barrier_sched();
996 }
997 EXPORT_SYMBOL_GPL(rcu_barrier);
998 
999 /*
1000  * Because preemptible RCU does not exist, it need not be initialized.
1001  */
1002 static void __init __rcu_init_preempt(void)
1003 {
1004 }
1005 
1006 /*
1007  * Because preemptible RCU does not exist, tasks cannot possibly exit
1008  * while in preemptible RCU read-side critical sections.
1009  */
1010 void exit_rcu(void)
1011 {
1012 }
1013 
1014 /*
1015  * Dump the guaranteed-empty blocked-tasks state.  Trust but verify.
1016  */
1017 static void
1018 dump_blkd_tasks(struct rcu_state *rsp, struct rcu_node *rnp, int ncheck)
1019 {
1020         WARN_ON_ONCE(!list_empty(&rnp->blkd_tasks));
1021 }
1022 
1023 #endif /* #else #ifdef CONFIG_PREEMPT_RCU */
1024 
1025 #ifdef CONFIG_RCU_BOOST
1026 
1027 static void rcu_wake_cond(struct task_struct *t, int status)
1028 {
1029         /*
1030          * If the thread is yielding, only wake it when this
1031          * is invoked from idle
1032          */
1033         if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1034                 wake_up_process(t);
1035 }
1036 
1037 /*
1038  * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1039  * or ->boost_tasks, advancing the pointer to the next task in the
1040  * ->blkd_tasks list.
1041  *
1042  * Note that irqs must be enabled: boosting the task can block.
1043  * Returns 1 if there are more tasks needing to be boosted.
1044  */
1045 static int rcu_boost(struct rcu_node *rnp)
1046 {
1047         unsigned long flags;
1048         struct task_struct *t;
1049         struct list_head *tb;
1050 
1051         if (READ_ONCE(rnp->exp_tasks) == NULL &&
1052             READ_ONCE(rnp->boost_tasks) == NULL)
1053                 return 0;  /* Nothing left to boost. */
1054 
1055         raw_spin_lock_irqsave_rcu_node(rnp, flags);
1056 
1057         /*
1058          * Recheck under the lock: all tasks in need of boosting
1059          * might exit their RCU read-side critical sections on their own.
1060          */
1061         if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1062                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1063                 return 0;
1064         }
1065 
1066         /*
1067          * Preferentially boost tasks blocking expedited grace periods.
1068          * This cannot starve the normal grace periods because a second
1069          * expedited grace period must boost all blocked tasks, including
1070          * those blocking the pre-existing normal grace period.
1071          */
1072         if (rnp->exp_tasks != NULL)
1073                 tb = rnp->exp_tasks;
1074         else
1075                 tb = rnp->boost_tasks;
1076 
1077         /*
1078          * We boost task t by manufacturing an rt_mutex that appears to
1079          * be held by task t.  We leave a pointer to that rt_mutex where
1080          * task t can find it, and task t will release the mutex when it
1081          * exits its outermost RCU read-side critical section.  Then
1082          * simply acquiring this artificial rt_mutex will boost task
1083          * t's priority.  (Thanks to tglx for suggesting this approach!)
1084          *
1085          * Note that task t must acquire rnp->lock to remove itself from
1086          * the ->blkd_tasks list, which it will do from exit() if from
1087          * nowhere else.  We therefore are guaranteed that task t will
1088          * stay around at least until we drop rnp->lock.  Note that
1089          * rnp->lock also resolves races between our priority boosting
1090          * and task t's exiting its outermost RCU read-side critical
1091          * section.
1092          */
1093         t = container_of(tb, struct task_struct, rcu_node_entry);
1094         rt_mutex_init_proxy_locked(&rnp->boost_mtx, t);
1095         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1096         /* Lock only for side effect: boosts task t's priority. */
1097         rt_mutex_lock(&rnp->boost_mtx);
1098         rt_mutex_unlock(&rnp->boost_mtx);  /* Then keep lockdep happy. */
1099 
1100         return READ_ONCE(rnp->exp_tasks) != NULL ||
1101                READ_ONCE(rnp->boost_tasks) != NULL;
1102 }
1103 
1104 /*
1105  * Priority-boosting kthread, one per leaf rcu_node.
1106  */
1107 static int rcu_boost_kthread(void *arg)
1108 {
1109         struct rcu_node *rnp = (struct rcu_node *)arg;
1110         int spincnt = 0;
1111         int more2boost;
1112 
1113         trace_rcu_utilization(TPS("Start boost kthread@init"));
1114         for (;;) {
1115                 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1116                 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1117                 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1118                 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1119                 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1120                 more2boost = rcu_boost(rnp);
1121                 if (more2boost)
1122                         spincnt++;
1123                 else
1124                         spincnt = 0;
1125                 if (spincnt > 10) {
1126                         rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1127                         trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1128                         schedule_timeout_interruptible(2);
1129                         trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1130                         spincnt = 0;
1131                 }
1132         }
1133         /* NOTREACHED */
1134         trace_rcu_utilization(TPS("End boost kthread@notreached"));
1135         return 0;
1136 }
1137 
1138 /*
1139  * Check to see if it is time to start boosting RCU readers that are
1140  * blocking the current grace period, and, if so, tell the per-rcu_node
1141  * kthread to start boosting them.  If there is an expedited grace
1142  * period in progress, it is always time to boost.
1143  *
1144  * The caller must hold rnp->lock, which this function releases.
1145  * The ->boost_kthread_task is immortal, so we don't need to worry
1146  * about it going away.
1147  */
1148 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1149         __releases(rnp->lock)
1150 {
1151         struct task_struct *t;
1152 
1153         raw_lockdep_assert_held_rcu_node(rnp);
1154         if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1155                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1156                 return;
1157         }
1158         if (rnp->exp_tasks != NULL ||
1159             (rnp->gp_tasks != NULL &&
1160              rnp->boost_tasks == NULL &&
1161              rnp->qsmask == 0 &&
1162              ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1163                 if (rnp->exp_tasks == NULL)
1164                         rnp->boost_tasks = rnp->gp_tasks;
1165                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1166                 t = rnp->boost_kthread_task;
1167                 if (t)
1168                         rcu_wake_cond(t, rnp->boost_kthread_status);
1169         } else {
1170                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1171         }
1172 }
1173 
1174 /*
1175  * Wake up the per-CPU kthread to invoke RCU callbacks.
1176  */
1177 static void invoke_rcu_callbacks_kthread(void)
1178 {
1179         unsigned long flags;
1180 
1181         local_irq_save(flags);
1182         __this_cpu_write(rcu_cpu_has_work, 1);
1183         if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1184             current != __this_cpu_read(rcu_cpu_kthread_task)) {
1185                 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1186                               __this_cpu_read(rcu_cpu_kthread_status));
1187         }
1188         local_irq_restore(flags);
1189 }
1190 
1191 /*
1192  * Is the current CPU running the RCU-callbacks kthread?
1193  * Caller must have preemption disabled.
1194  */
1195 static bool rcu_is_callbacks_kthread(void)
1196 {
1197         return __this_cpu_read(rcu_cpu_kthread_task) == current;
1198 }
1199 
1200 #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1201 
1202 /*
1203  * Do priority-boost accounting for the start of a new grace period.
1204  */
1205 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1206 {
1207         rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1208 }
1209 
1210 /*
1211  * Create an RCU-boost kthread for the specified node if one does not
1212  * already exist.  We only create this kthread for preemptible RCU.
1213  * Returns zero if all is well, a negated errno otherwise.
1214  */
1215 static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1216                                        struct rcu_node *rnp)
1217 {
1218         int rnp_index = rnp - &rsp->node[0];
1219         unsigned long flags;
1220         struct sched_param sp;
1221         struct task_struct *t;
1222 
1223         if (rcu_state_p != rsp)
1224                 return 0;
1225 
1226         if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0)
1227                 return 0;
1228 
1229         rsp->boost = 1;
1230         if (rnp->boost_kthread_task != NULL)
1231                 return 0;
1232         t = kthread_create(rcu_boost_kthread, (void *)rnp,
1233                            "rcub/%d", rnp_index);
1234         if (IS_ERR(t))
1235                 return PTR_ERR(t);
1236         raw_spin_lock_irqsave_rcu_node(rnp, flags);
1237         rnp->boost_kthread_task = t;
1238         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1239         sp.sched_priority = kthread_prio;
1240         sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1241         wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1242         return 0;
1243 }
1244 
1245 static void rcu_kthread_do_work(void)
1246 {
1247         rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1248         rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1249         rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
1250 }
1251 
1252 static void rcu_cpu_kthread_setup(unsigned int cpu)
1253 {
1254         struct sched_param sp;
1255 
1256         sp.sched_priority = kthread_prio;
1257         sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1258 }
1259 
1260 static void rcu_cpu_kthread_park(unsigned int cpu)
1261 {
1262         per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1263 }
1264 
1265 static int rcu_cpu_kthread_should_run(unsigned int cpu)
1266 {
1267         return __this_cpu_read(rcu_cpu_has_work);
1268 }
1269 
1270 /*
1271  * Per-CPU kernel thread that invokes RCU callbacks.  This replaces the
1272  * RCU softirq used in flavors and configurations of RCU that do not
1273  * support RCU priority boosting.
1274  */
1275 static void rcu_cpu_kthread(unsigned int cpu)
1276 {
1277         unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1278         char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1279         int spincnt;
1280 
1281         for (spincnt = 0; spincnt < 10; spincnt++) {
1282                 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1283                 local_bh_disable();
1284                 *statusp = RCU_KTHREAD_RUNNING;
1285                 this_cpu_inc(rcu_cpu_kthread_loops);
1286                 local_irq_disable();
1287                 work = *workp;
1288                 *workp = 0;
1289                 local_irq_enable();
1290                 if (work)
1291                         rcu_kthread_do_work();
1292                 local_bh_enable();
1293                 if (*workp == 0) {
1294                         trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1295                         *statusp = RCU_KTHREAD_WAITING;
1296                         return;
1297                 }
1298         }
1299         *statusp = RCU_KTHREAD_YIELDING;
1300         trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1301         schedule_timeout_interruptible(2);
1302         trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1303         *statusp = RCU_KTHREAD_WAITING;
1304 }
1305 
1306 /*
1307  * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1308  * served by the rcu_node in question.  The CPU hotplug lock is still
1309  * held, so the value of rnp->qsmaskinit will be stable.
1310  *
1311  * We don't include outgoingcpu in the affinity set, use -1 if there is
1312  * no outgoing CPU.  If there are no CPUs left in the affinity set,
1313  * this function allows the kthread to execute on any CPU.
1314  */
1315 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1316 {
1317         struct task_struct *t = rnp->boost_kthread_task;
1318         unsigned long mask = rcu_rnp_online_cpus(rnp);
1319         cpumask_var_t cm;
1320         int cpu;
1321 
1322         if (!t)
1323                 return;
1324         if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1325                 return;
1326         for_each_leaf_node_possible_cpu(rnp, cpu)
1327                 if ((mask & leaf_node_cpu_bit(rnp, cpu)) &&
1328                     cpu != outgoingcpu)
1329                         cpumask_set_cpu(cpu, cm);
1330         if (cpumask_weight(cm) == 0)
1331                 cpumask_setall(cm);
1332         set_cpus_allowed_ptr(t, cm);
1333         free_cpumask_var(cm);
1334 }
1335 
1336 static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1337         .store                  = &rcu_cpu_kthread_task,
1338         .thread_should_run      = rcu_cpu_kthread_should_run,
1339         .thread_fn              = rcu_cpu_kthread,
1340         .thread_comm            = "rcuc/%u",
1341         .setup                  = rcu_cpu_kthread_setup,
1342         .park                   = rcu_cpu_kthread_park,
1343 };
1344 
1345 /*
1346  * Spawn boost kthreads -- called as soon as the scheduler is running.
1347  */
1348 static void __init rcu_spawn_boost_kthreads(void)
1349 {
1350         struct rcu_node *rnp;
1351         int cpu;
1352 
1353         for_each_possible_cpu(cpu)
1354                 per_cpu(rcu_cpu_has_work, cpu) = 0;
1355         BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1356         rcu_for_each_leaf_node(rcu_state_p, rnp)
1357                 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1358 }
1359 
1360 static void rcu_prepare_kthreads(int cpu)
1361 {
1362         struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu);
1363         struct rcu_node *rnp = rdp->mynode;
1364 
1365         /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1366         if (rcu_scheduler_fully_active)
1367                 (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp);
1368 }
1369 
1370 #else /* #ifdef CONFIG_RCU_BOOST */
1371 
1372 static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1373         __releases(rnp->lock)
1374 {
1375         raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
1376 }
1377 
1378 static void invoke_rcu_callbacks_kthread(void)
1379 {
1380         WARN_ON_ONCE(1);
1381 }
1382 
1383 static bool rcu_is_callbacks_kthread(void)
1384 {
1385         return false;
1386 }
1387 
1388 static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1389 {
1390 }
1391 
1392 static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1393 {
1394 }
1395 
1396 static void __init rcu_spawn_boost_kthreads(void)
1397 {
1398 }
1399 
1400 static void rcu_prepare_kthreads(int cpu)
1401 {
1402 }
1403 
1404 #endif /* #else #ifdef CONFIG_RCU_BOOST */
1405 
1406 #if !defined(CONFIG_RCU_FAST_NO_HZ)
1407 
1408 /*
1409  * Check to see if any future RCU-related work will need to be done
1410  * by the current CPU, even if none need be done immediately, returning
1411  * 1 if so.  This function is part of the RCU implementation; it is -not-
1412  * an exported member of the RCU API.
1413  *
1414  * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1415  * any flavor of RCU.
1416  */
1417 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1418 {
1419         *nextevt = KTIME_MAX;
1420         return rcu_cpu_has_callbacks(NULL);
1421 }
1422 
1423 /*
1424  * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1425  * after it.
1426  */
1427 static void rcu_cleanup_after_idle(void)
1428 {
1429 }
1430 
1431 /*
1432  * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1433  * is nothing.
1434  */
1435 static void rcu_prepare_for_idle(void)
1436 {
1437 }
1438 
1439 /*
1440  * Don't bother keeping a running count of the number of RCU callbacks
1441  * posted because CONFIG_RCU_FAST_NO_HZ=n.
1442  */
1443 static void rcu_idle_count_callbacks_posted(void)
1444 {
1445 }
1446 
1447 #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1448 
1449 /*
1450  * This code is invoked when a CPU goes idle, at which point we want
1451  * to have the CPU do everything required for RCU so that it can enter
1452  * the energy-efficient dyntick-idle mode.  This is handled by a
1453  * state machine implemented by rcu_prepare_for_idle() below.
1454  *
1455  * The following three proprocessor symbols control this state machine:
1456  *
1457  * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1458  *      to sleep in dyntick-idle mode with RCU callbacks pending.  This
1459  *      is sized to be roughly one RCU grace period.  Those energy-efficiency
1460  *      benchmarkers who might otherwise be tempted to set this to a large
1461  *      number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1462  *      system.  And if you are -that- concerned about energy efficiency,
1463  *      just power the system down and be done with it!
1464  * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1465  *      permitted to sleep in dyntick-idle mode with only lazy RCU
1466  *      callbacks pending.  Setting this too high can OOM your system.
1467  *
1468  * The values below work well in practice.  If future workloads require
1469  * adjustment, they can be converted into kernel config parameters, though
1470  * making the state machine smarter might be a better option.
1471  */
1472 #define RCU_IDLE_GP_DELAY 4             /* Roughly one grace period. */
1473 #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1474 
1475 static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1476 module_param(rcu_idle_gp_delay, int, 0644);
1477 static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1478 module_param(rcu_idle_lazy_gp_delay, int, 0644);
1479 
1480 /*
1481  * Try to advance callbacks for all flavors of RCU on the current CPU, but
1482  * only if it has been awhile since the last time we did so.  Afterwards,
1483  * if there are any callbacks ready for immediate invocation, return true.
1484  */
1485 static bool __maybe_unused rcu_try_advance_all_cbs(void)
1486 {
1487         bool cbs_ready = false;
1488         struct rcu_data *rdp;
1489         struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1490         struct rcu_node *rnp;
1491         struct rcu_state *rsp;
1492 
1493         /* Exit early if we advanced recently. */
1494         if (jiffies == rdtp->last_advance_all)
1495                 return false;
1496         rdtp->last_advance_all = jiffies;
1497 
1498         for_each_rcu_flavor(rsp) {
1499                 rdp = this_cpu_ptr(rsp->rda);
1500                 rnp = rdp->mynode;
1501 
1502                 /*
1503                  * Don't bother checking unless a grace period has
1504                  * completed since we last checked and there are
1505                  * callbacks not yet ready to invoke.
1506                  */
1507                 if ((rcu_seq_completed_gp(rdp->gp_seq,
1508                                           rcu_seq_current(&rnp->gp_seq)) ||
1509                      unlikely(READ_ONCE(rdp->gpwrap))) &&
1510                     rcu_segcblist_pend_cbs(&rdp->cblist))
1511                         note_gp_changes(rsp, rdp);
1512 
1513                 if (rcu_segcblist_ready_cbs(&rdp->cblist))
1514                         cbs_ready = true;
1515         }
1516         return cbs_ready;
1517 }
1518 
1519 /*
1520  * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1521  * to invoke.  If the CPU has callbacks, try to advance them.  Tell the
1522  * caller to set the timeout based on whether or not there are non-lazy
1523  * callbacks.
1524  *
1525  * The caller must have disabled interrupts.
1526  */
1527 int rcu_needs_cpu(u64 basemono, u64 *nextevt)
1528 {
1529         struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1530         unsigned long dj;
1531 
1532         lockdep_assert_irqs_disabled();
1533 
1534         /* Snapshot to detect later posting of non-lazy callback. */
1535         rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1536 
1537         /* If no callbacks, RCU doesn't need the CPU. */
1538         if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) {
1539                 *nextevt = KTIME_MAX;
1540                 return 0;
1541         }
1542 
1543         /* Attempt to advance callbacks. */
1544         if (rcu_try_advance_all_cbs()) {
1545                 /* Some ready to invoke, so initiate later invocation. */
1546                 invoke_rcu_core();
1547                 return 1;
1548         }
1549         rdtp->last_accelerate = jiffies;
1550 
1551         /* Request timer delay depending on laziness, and round. */
1552         if (!rdtp->all_lazy) {
1553                 dj = round_up(rcu_idle_gp_delay + jiffies,
1554                                rcu_idle_gp_delay) - jiffies;
1555         } else {
1556                 dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1557         }
1558         *nextevt = basemono + dj * TICK_NSEC;
1559         return 0;
1560 }
1561 
1562 /*
1563  * Prepare a CPU for idle from an RCU perspective.  The first major task
1564  * is to sense whether nohz mode has been enabled or disabled via sysfs.
1565  * The second major task is to check to see if a non-lazy callback has
1566  * arrived at a CPU that previously had only lazy callbacks.  The third
1567  * major task is to accelerate (that is, assign grace-period numbers to)
1568  * any recently arrived callbacks.
1569  *
1570  * The caller must have disabled interrupts.
1571  */
1572 static void rcu_prepare_for_idle(void)
1573 {
1574         bool needwake;
1575         struct rcu_data *rdp;
1576         struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1577         struct rcu_node *rnp;
1578         struct rcu_state *rsp;
1579         int tne;
1580 
1581         lockdep_assert_irqs_disabled();
1582         if (rcu_is_nocb_cpu(smp_processor_id()))
1583                 return;
1584 
1585         /* Handle nohz enablement switches conservatively. */
1586         tne = READ_ONCE(tick_nohz_active);
1587         if (tne != rdtp->tick_nohz_enabled_snap) {
1588                 if (rcu_cpu_has_callbacks(NULL))
1589                         invoke_rcu_core(); /* force nohz to see update. */
1590                 rdtp->tick_nohz_enabled_snap = tne;
1591                 return;
1592         }
1593         if (!tne)
1594                 return;
1595 
1596         /*
1597          * If a non-lazy callback arrived at a CPU having only lazy
1598          * callbacks, invoke RCU core for the side-effect of recalculating
1599          * idle duration on re-entry to idle.
1600          */
1601         if (rdtp->all_lazy &&
1602             rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1603                 rdtp->all_lazy = false;
1604                 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1605                 invoke_rcu_core();
1606                 return;
1607         }
1608 
1609         /*
1610          * If we have not yet accelerated this jiffy, accelerate all
1611          * callbacks on this CPU.
1612          */
1613         if (rdtp->last_accelerate == jiffies)
1614                 return;
1615         rdtp->last_accelerate = jiffies;
1616         for_each_rcu_flavor(rsp) {
1617                 rdp = this_cpu_ptr(rsp->rda);
1618                 if (!rcu_segcblist_pend_cbs(&rdp->cblist))
1619                         continue;
1620                 rnp = rdp->mynode;
1621                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
1622                 needwake = rcu_accelerate_cbs(rsp, rnp, rdp);
1623                 raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */
1624                 if (needwake)
1625                         rcu_gp_kthread_wake(rsp);
1626         }
1627 }
1628 
1629 /*
1630  * Clean up for exit from idle.  Attempt to advance callbacks based on
1631  * any grace periods that elapsed while the CPU was idle, and if any
1632  * callbacks are now ready to invoke, initiate invocation.
1633  */
1634 static void rcu_cleanup_after_idle(void)
1635 {
1636         lockdep_assert_irqs_disabled();
1637         if (rcu_is_nocb_cpu(smp_processor_id()))
1638                 return;
1639         if (rcu_try_advance_all_cbs())
1640                 invoke_rcu_core();
1641 }
1642 
1643 /*
1644  * Keep a running count of the number of non-lazy callbacks posted
1645  * on this CPU.  This running counter (which is never decremented) allows
1646  * rcu_prepare_for_idle() to detect when something out of the idle loop
1647  * posts a callback, even if an equal number of callbacks are invoked.
1648  * Of course, callbacks should only be posted from within a trace event
1649  * designed to be called from idle or from within RCU_NONIDLE().
1650  */
1651 static void rcu_idle_count_callbacks_posted(void)
1652 {
1653         __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1654 }
1655 
1656 /*
1657  * Data for flushing lazy RCU callbacks at OOM time.
1658  */
1659 static atomic_t oom_callback_count;
1660 static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1661 
1662 /*
1663  * RCU OOM callback -- decrement the outstanding count and deliver the
1664  * wake-up if we are the last one.
1665  */
1666 static void rcu_oom_callback(struct rcu_head *rhp)
1667 {
1668         if (atomic_dec_and_test(&oom_callback_count))
1669                 wake_up(&oom_callback_wq);
1670 }
1671 
1672 /*
1673  * Post an rcu_oom_notify callback on the current CPU if it has at
1674  * least one lazy callback.  This will unnecessarily post callbacks
1675  * to CPUs that already have a non-lazy callback at the end of their
1676  * callback list, but this is an infrequent operation, so accept some
1677  * extra overhead to keep things simple.
1678  */
1679 static void rcu_oom_notify_cpu(void *unused)
1680 {
1681         struct rcu_state *rsp;
1682         struct rcu_data *rdp;
1683 
1684         for_each_rcu_flavor(rsp) {
1685                 rdp = raw_cpu_ptr(rsp->rda);
1686                 if (rcu_segcblist_n_lazy_cbs(&rdp->cblist)) {
1687                         atomic_inc(&oom_callback_count);
1688                         rsp->call(&rdp->oom_head, rcu_oom_callback);
1689                 }
1690         }
1691 }
1692 
1693 /*
1694  * If low on memory, ensure that each CPU has a non-lazy callback.
1695  * This will wake up CPUs that have only lazy callbacks, in turn
1696  * ensuring that they free up the corresponding memory in a timely manner.
1697  * Because an uncertain amount of memory will be freed in some uncertain
1698  * timeframe, we do not claim to have freed anything.
1699  */
1700 static int rcu_oom_notify(struct notifier_block *self,
1701                           unsigned long notused, void *nfreed)
1702 {
1703         int cpu;
1704 
1705         /* Wait for callbacks from earlier instance to complete. */
1706         wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1707         smp_mb(); /* Ensure callback reuse happens after callback invocation. */
1708 
1709         /*
1710          * Prevent premature wakeup: ensure that all increments happen
1711          * before there is a chance of the counter reaching zero.
1712          */
1713         atomic_set(&oom_callback_count, 1);
1714 
1715         for_each_online_cpu(cpu) {
1716                 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1717                 cond_resched_tasks_rcu_qs();
1718         }
1719 
1720         /* Unconditionally decrement: no need to wake ourselves up. */
1721         atomic_dec(&oom_callback_count);
1722 
1723         return NOTIFY_OK;
1724 }
1725 
1726 static struct notifier_block rcu_oom_nb = {
1727         .notifier_call = rcu_oom_notify
1728 };
1729 
1730 static int __init rcu_register_oom_notifier(void)
1731 {
1732         register_oom_notifier(&rcu_oom_nb);
1733         return 0;
1734 }
1735 early_initcall(rcu_register_oom_notifier);
1736 
1737 #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1738 
1739 #ifdef CONFIG_RCU_FAST_NO_HZ
1740 
1741 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1742 {
1743         struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1744         unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1745 
1746         sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1747                 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1748                 ulong2long(nlpd),
1749                 rdtp->all_lazy ? 'L' : '.',
1750                 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1751 }
1752 
1753 #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1754 
1755 static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1756 {
1757         *cp = '\0';
1758 }
1759 
1760 #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1761 
1762 /* Initiate the stall-info list. */
1763 static void print_cpu_stall_info_begin(void)
1764 {
1765         pr_cont("\n");
1766 }
1767 
1768 /*
1769  * Print out diagnostic information for the specified stalled CPU.
1770  *
1771  * If the specified CPU is aware of the current RCU grace period
1772  * (flavor specified by rsp), then print the number of scheduling
1773  * clock interrupts the CPU has taken during the time that it has
1774  * been aware.  Otherwise, print the number of RCU grace periods
1775  * that this CPU is ignorant of, for example, "1" if the CPU was
1776  * aware of the previous grace period.
1777  *
1778  * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1779  */
1780 static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1781 {
1782         unsigned long delta;
1783         char fast_no_hz[72];
1784         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1785         struct rcu_dynticks *rdtp = rdp->dynticks;
1786         char *ticks_title;
1787         unsigned long ticks_value;
1788 
1789         /*
1790          * We could be printing a lot while holding a spinlock.  Avoid
1791          * triggering hard lockup.
1792          */
1793         touch_nmi_watchdog();
1794 
1795         ticks_value = rcu_seq_ctr(rsp->gp_seq - rdp->gp_seq);
1796         if (ticks_value) {
1797                 ticks_title = "GPs behind";
1798         } else {
1799                 ticks_title = "ticks this GP";
1800                 ticks_value = rdp->ticks_this_gp;
1801         }
1802         print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1803         delta = rcu_seq_ctr(rdp->mynode->gp_seq - rdp->rcu_iw_gp_seq);
1804         pr_err("\t%d-%c%c%c%c: (%lu %s) idle=%03x/%ld/%#lx softirq=%u/%u fqs=%ld %s\n",
1805                cpu,
1806                "O."[!!cpu_online(cpu)],
1807                "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)],
1808                "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)],
1809                !IS_ENABLED(CONFIG_IRQ_WORK) ? '?' :
1810                         rdp->rcu_iw_pending ? (int)min(delta, 9UL) + '' :
1811                                 "!."[!delta],
1812                ticks_value, ticks_title,
1813                rcu_dynticks_snap(rdtp) & 0xfff,
1814                rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1815                rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1816                READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart,
1817                fast_no_hz);
1818 }
1819 
1820 /* Terminate the stall-info list. */
1821 static void print_cpu_stall_info_end(void)
1822 {
1823         pr_err("\t");
1824 }
1825 
1826 /* Zero ->ticks_this_gp for all flavors of RCU. */
1827 static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1828 {
1829         rdp->ticks_this_gp = 0;
1830         rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1831 }
1832 
1833 /* Increment ->ticks_this_gp for all flavors of RCU. */
1834 static void increment_cpu_stall_ticks(void)
1835 {
1836         struct rcu_state *rsp;
1837 
1838         for_each_rcu_flavor(rsp)
1839                 raw_cpu_inc(rsp->rda->ticks_this_gp);
1840 }
1841 
1842 #ifdef CONFIG_RCU_NOCB_CPU
1843 
1844 /*
1845  * Offload callback processing from the boot-time-specified set of CPUs
1846  * specified by rcu_nocb_mask.  For each CPU in the set, there is a
1847  * kthread created that pulls the callbacks from the corresponding CPU,
1848  * waits for a grace period to elapse, and invokes the callbacks.
1849  * The no-CBs CPUs do a wake_up() on their kthread when they insert
1850  * a callback into any empty list, unless the rcu_nocb_poll boot parameter
1851  * has been specified, in which case each kthread actively polls its
1852  * CPU.  (Which isn't so great for energy efficiency, but which does
1853  * reduce RCU's overhead on that CPU.)
1854  *
1855  * This is intended to be used in conjunction with Frederic Weisbecker's
1856  * adaptive-idle work, which would seriously reduce OS jitter on CPUs
1857  * running CPU-bound user-mode computations.
1858  *
1859  * Offloading of callback processing could also in theory be used as
1860  * an energy-efficiency measure because CPUs with no RCU callbacks
1861  * queued are more aggressive about entering dyntick-idle mode.
1862  */
1863 
1864 
1865 /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
1866 static int __init rcu_nocb_setup(char *str)
1867 {
1868         alloc_bootmem_cpumask_var(&rcu_nocb_mask);
1869         cpulist_parse(str, rcu_nocb_mask);
1870         return 1;
1871 }
1872 __setup("rcu_nocbs=", rcu_nocb_setup);
1873 
1874 static int __init parse_rcu_nocb_poll(char *arg)
1875 {
1876         rcu_nocb_poll = true;
1877         return 0;
1878 }
1879 early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
1880 
1881 /*
1882  * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
1883  * grace period.
1884  */
1885 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
1886 {
1887         swake_up_all(sq);
1888 }
1889 
1890 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
1891 {
1892         return &rnp->nocb_gp_wq[rcu_seq_ctr(rnp->gp_seq) & 0x1];
1893 }
1894 
1895 static void rcu_init_one_nocb(struct rcu_node *rnp)
1896 {
1897         init_swait_queue_head(&rnp->nocb_gp_wq[0]);
1898         init_swait_queue_head(&rnp->nocb_gp_wq[1]);
1899 }
1900 
1901 /* Is the specified CPU a no-CBs CPU? */
1902 bool rcu_is_nocb_cpu(int cpu)
1903 {
1904         if (cpumask_available(rcu_nocb_mask))
1905                 return cpumask_test_cpu(cpu, rcu_nocb_mask);
1906         return false;
1907 }
1908 
1909 /*
1910  * Kick the leader kthread for this NOCB group.  Caller holds ->nocb_lock
1911  * and this function releases it.
1912  */
1913 static void __wake_nocb_leader(struct rcu_data *rdp, bool force,
1914                                unsigned long flags)
1915         __releases(rdp->nocb_lock)
1916 {
1917         struct rcu_data *rdp_leader = rdp->nocb_leader;
1918 
1919         lockdep_assert_held(&rdp->nocb_lock);
1920         if (!READ_ONCE(rdp_leader->nocb_kthread)) {
1921                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1922                 return;
1923         }
1924         if (rdp_leader->nocb_leader_sleep || force) {
1925                 /* Prior smp_mb__after_atomic() orders against prior enqueue. */
1926                 WRITE_ONCE(rdp_leader->nocb_leader_sleep, false);
1927                 del_timer(&rdp->nocb_timer);
1928                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1929                 smp_mb(); /* ->nocb_leader_sleep before swake_up_one(). */
1930                 swake_up_one(&rdp_leader->nocb_wq);
1931         } else {
1932                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1933         }
1934 }
1935 
1936 /*
1937  * Kick the leader kthread for this NOCB group, but caller has not
1938  * acquired locks.
1939  */
1940 static void wake_nocb_leader(struct rcu_data *rdp, bool force)
1941 {
1942         unsigned long flags;
1943 
1944         raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1945         __wake_nocb_leader(rdp, force, flags);
1946 }
1947 
1948 /*
1949  * Arrange to wake the leader kthread for this NOCB group at some
1950  * future time when it is safe to do so.
1951  */
1952 static void wake_nocb_leader_defer(struct rcu_data *rdp, int waketype,
1953                                    const char *reason)
1954 {
1955         unsigned long flags;
1956 
1957         raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
1958         if (rdp->nocb_defer_wakeup == RCU_NOCB_WAKE_NOT)
1959                 mod_timer(&rdp->nocb_timer, jiffies + 1);
1960         WRITE_ONCE(rdp->nocb_defer_wakeup, waketype);
1961         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, reason);
1962         raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
1963 }
1964 
1965 /*
1966  * Does the specified CPU need an RCU callback for the specified flavor
1967  * of rcu_barrier()?
1968  */
1969 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
1970 {
1971         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1972         unsigned long ret;
1973 #ifdef CONFIG_PROVE_RCU
1974         struct rcu_head *rhp;
1975 #endif /* #ifdef CONFIG_PROVE_RCU */
1976 
1977         /*
1978          * Check count of all no-CBs callbacks awaiting invocation.
1979          * There needs to be a barrier before this function is called,
1980          * but associated with a prior determination that no more
1981          * callbacks would be posted.  In the worst case, the first
1982          * barrier in _rcu_barrier() suffices (but the caller cannot
1983          * necessarily rely on this, not a substitute for the caller
1984          * getting the concurrency design right!).  There must also be
1985          * a barrier between the following load an posting of a callback
1986          * (if a callback is in fact needed).  This is associated with an
1987          * atomic_inc() in the caller.
1988          */
1989         ret = atomic_long_read(&rdp->nocb_q_count);
1990 
1991 #ifdef CONFIG_PROVE_RCU
1992         rhp = READ_ONCE(rdp->nocb_head);
1993         if (!rhp)
1994                 rhp = READ_ONCE(rdp->nocb_gp_head);
1995         if (!rhp)
1996                 rhp = READ_ONCE(rdp->nocb_follower_head);
1997 
1998         /* Having no rcuo kthread but CBs after scheduler starts is bad! */
1999         if (!READ_ONCE(rdp->nocb_kthread) && rhp &&
2000             rcu_scheduler_fully_active) {
2001                 /* RCU callback enqueued before CPU first came online??? */
2002                 pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n",
2003                        cpu, rhp->func);
2004                 WARN_ON_ONCE(1);
2005         }
2006 #endif /* #ifdef CONFIG_PROVE_RCU */
2007 
2008         return !!ret;
2009 }
2010 
2011 /*
2012  * Enqueue the specified string of rcu_head structures onto the specified
2013  * CPU's no-CBs lists.  The CPU is specified by rdp, the head of the
2014  * string by rhp, and the tail of the string by rhtp.  The non-lazy/lazy
2015  * counts are supplied by rhcount and rhcount_lazy.
2016  *
2017  * If warranted, also wake up the kthread servicing this CPUs queues.
2018  */
2019 static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2020                                     struct rcu_head *rhp,
2021                                     struct rcu_head **rhtp,
2022                                     int rhcount, int rhcount_lazy,
2023                                     unsigned long flags)
2024 {
2025         int len;
2026         struct rcu_head **old_rhpp;
2027         struct task_struct *t;
2028 
2029         /* Enqueue the callback on the nocb list and update counts. */
2030         atomic_long_add(rhcount, &rdp->nocb_q_count);
2031         /* rcu_barrier() relies on ->nocb_q_count add before xchg. */
2032         old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2033         WRITE_ONCE(*old_rhpp, rhp);
2034         atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2035         smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */
2036 
2037         /* If we are not being polled and there is a kthread, awaken it ... */
2038         t = READ_ONCE(rdp->nocb_kthread);
2039         if (rcu_nocb_poll || !t) {
2040                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2041                                     TPS("WakeNotPoll"));
2042                 return;
2043         }
2044         len = atomic_long_read(&rdp->nocb_q_count);
2045         if (old_rhpp == &rdp->nocb_head) {
2046                 if (!irqs_disabled_flags(flags)) {
2047                         /* ... if queue was empty ... */
2048                         wake_nocb_leader(rdp, false);
2049                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2050                                             TPS("WakeEmpty"));
2051                 } else {
2052                         wake_nocb_leader_defer(rdp, RCU_NOCB_WAKE,
2053                                                TPS("WakeEmptyIsDeferred"));
2054                 }
2055                 rdp->qlen_last_fqs_check = 0;
2056         } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2057                 /* ... or if many callbacks queued. */
2058                 if (!irqs_disabled_flags(flags)) {
2059                         wake_nocb_leader(rdp, true);
2060                         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2061                                             TPS("WakeOvf"));
2062                 } else {
2063                         wake_nocb_leader_defer(rdp, RCU_NOCB_WAKE_FORCE,
2064                                                TPS("WakeOvfIsDeferred"));
2065                 }
2066                 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2067         } else {
2068                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2069         }
2070         return;
2071 }
2072 
2073 /*
2074  * This is a helper for __call_rcu(), which invokes this when the normal
2075  * callback queue is inoperable.  If this is not a no-CBs CPU, this
2076  * function returns failure back to __call_rcu(), which can complain
2077  * appropriately.
2078  *
2079  * Otherwise, this function queues the callback where the corresponding
2080  * "rcuo" kthread can find it.
2081  */
2082 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2083                             bool lazy, unsigned long flags)
2084 {
2085 
2086         if (!rcu_is_nocb_cpu(rdp->cpu))
2087                 return false;
2088         __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags);
2089         if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2090                 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2091                                          (unsigned long)rhp->func,
2092                                          -atomic_long_read(&rdp->nocb_q_count_lazy),
2093                                          -atomic_long_read(&rdp->nocb_q_count));
2094         else
2095                 trace_rcu_callback(rdp->rsp->name, rhp,
2096                                    -atomic_long_read(&rdp->nocb_q_count_lazy),
2097                                    -atomic_long_read(&rdp->nocb_q_count));
2098 
2099         /*
2100          * If called from an extended quiescent state with interrupts
2101          * disabled, invoke the RCU core in order to allow the idle-entry
2102          * deferred-wakeup check to function.
2103          */
2104         if (irqs_disabled_flags(flags) &&
2105             !rcu_is_watching() &&
2106             cpu_online(smp_processor_id()))
2107                 invoke_rcu_core();
2108 
2109         return true;
2110 }
2111 
2112 /*
2113  * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2114  * not a no-CBs CPU.
2115  */
2116 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
2117                                                      struct rcu_data *rdp,
2118                                                      unsigned long flags)
2119 {
2120         lockdep_assert_irqs_disabled();
2121         if (!rcu_is_nocb_cpu(smp_processor_id()))
2122                 return false; /* Not NOCBs CPU, caller must migrate CBs. */
2123         __call_rcu_nocb_enqueue(my_rdp, rcu_segcblist_head(&rdp->cblist),
2124                                 rcu_segcblist_tail(&rdp->cblist),
2125                                 rcu_segcblist_n_cbs(&rdp->cblist),
2126                                 rcu_segcblist_n_lazy_cbs(&rdp->cblist), flags);
2127         rcu_segcblist_init(&rdp->cblist);
2128         rcu_segcblist_disable(&rdp->cblist);
2129         return true;
2130 }
2131 
2132 /*
2133  * If necessary, kick off a new grace period, and either way wait
2134  * for a subsequent grace period to complete.
2135  */
2136 static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2137 {
2138         unsigned long c;
2139         bool d;
2140         unsigned long flags;
2141         bool needwake;
2142         struct rcu_node *rnp = rdp->mynode;
2143 
2144         local_irq_save(flags);
2145         c = rcu_seq_snap(&rdp->rsp->gp_seq);
2146         if (!rdp->gpwrap && ULONG_CMP_GE(rdp->gp_seq_needed, c)) {
2147                 local_irq_restore(flags);
2148         } else {
2149                 raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */
2150                 needwake = rcu_start_this_gp(rnp, rdp, c);
2151                 raw_spin_unlock_irqrestore_rcu_node(rnp, flags);
2152                 if (needwake)
2153                         rcu_gp_kthread_wake(rdp->rsp);
2154         }
2155 
2156         /*
2157          * Wait for the grace period.  Do so interruptibly to avoid messing
2158          * up the load average.
2159          */
2160         trace_rcu_this_gp(rnp, rdp, c, TPS("StartWait"));
2161         for (;;) {
2162                 swait_event_interruptible_exclusive(
2163                         rnp->nocb_gp_wq[rcu_seq_ctr(c) & 0x1],
2164                         (d = rcu_seq_done(&rnp->gp_seq, c)));
2165                 if (likely(d))
2166                         break;
2167                 WARN_ON(signal_pending(current));
2168                 trace_rcu_this_gp(rnp, rdp, c, TPS("ResumeWait"));
2169         }
2170         trace_rcu_this_gp(rnp, rdp, c, TPS("EndWait"));
2171         smp_mb(); /* Ensure that CB invocation happens after GP end. */
2172 }
2173 
2174 /*
2175  * Leaders come here to wait for additional callbacks to show up.
2176  * This function does not return until callbacks appear.
2177  */
2178 static void nocb_leader_wait(struct rcu_data *my_rdp)
2179 {
2180         bool firsttime = true;
2181         unsigned long flags;
2182         bool gotcbs;
2183         struct rcu_data *rdp;
2184         struct rcu_head **tail;
2185 
2186 wait_again:
2187 
2188         /* Wait for callbacks to appear. */
2189         if (!rcu_nocb_poll) {
2190                 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, TPS("Sleep"));
2191                 swait_event_interruptible_exclusive(my_rdp->nocb_wq,
2192                                 !READ_ONCE(my_rdp->nocb_leader_sleep));
2193                 raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
2194                 my_rdp->nocb_leader_sleep = true;
2195                 WRITE_ONCE(my_rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2196                 del_timer(&my_rdp->nocb_timer);
2197                 raw_spin_unlock_irqrestore(&my_rdp->nocb_lock, flags);
2198         } else if (firsttime) {
2199                 firsttime = false; /* Don't drown trace log with "Poll"! */
2200                 trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, TPS("Poll"));
2201         }
2202 
2203         /*
2204          * Each pass through the following loop checks a follower for CBs.
2205          * We are our own first follower.  Any CBs found are moved to
2206          * nocb_gp_head, where they await a grace period.
2207          */
2208         gotcbs = false;
2209         smp_mb(); /* wakeup and _sleep before ->nocb_head reads. */
2210         for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2211                 rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head);
2212                 if (!rdp->nocb_gp_head)
2213                         continue;  /* No CBs here, try next follower. */
2214 
2215                 /* Move callbacks to wait-for-GP list, which is empty. */
2216                 WRITE_ONCE(rdp->nocb_head, NULL);
2217                 rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2218                 gotcbs = true;
2219         }
2220 
2221         /* No callbacks?  Sleep a bit if polling, and go retry.  */
2222         if (unlikely(!gotcbs)) {
2223                 WARN_ON(signal_pending(current));
2224                 if (rcu_nocb_poll) {
2225                         schedule_timeout_interruptible(1);
2226                 } else {
2227                         trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu,
2228                                             TPS("WokeEmpty"));
2229                 }
2230                 goto wait_again;
2231         }
2232 
2233         /* Wait for one grace period. */
2234         rcu_nocb_wait_gp(my_rdp);
2235 
2236         /* Each pass through the following loop wakes a follower, if needed. */
2237         for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) {
2238                 if (!rcu_nocb_poll &&
2239                     READ_ONCE(rdp->nocb_head) &&
2240                     READ_ONCE(my_rdp->nocb_leader_sleep)) {
2241                         raw_spin_lock_irqsave(&my_rdp->nocb_lock, flags);
2242                         my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/
2243                         raw_spin_unlock_irqrestore(&my_rdp->nocb_lock, flags);
2244                 }
2245                 if (!rdp->nocb_gp_head)
2246                         continue; /* No CBs, so no need to wake follower. */
2247 
2248                 /* Append callbacks to follower's "done" list. */
2249                 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2250                 tail = rdp->nocb_follower_tail;
2251                 rdp->nocb_follower_tail = rdp->nocb_gp_tail;
2252                 *tail = rdp->nocb_gp_head;
2253                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2254                 if (rdp != my_rdp && tail == &rdp->nocb_follower_head) {
2255                         /* List was empty, so wake up the follower.  */
2256                         swake_up_one(&rdp->nocb_wq);
2257                 }
2258         }
2259 
2260         /* If we (the leader) don't have CBs, go wait some more. */
2261         if (!my_rdp->nocb_follower_head)
2262                 goto wait_again;
2263 }
2264 
2265 /*
2266  * Followers come here to wait for additional callbacks to show up.
2267  * This function does not return until callbacks appear.
2268  */
2269 static void nocb_follower_wait(struct rcu_data *rdp)
2270 {
2271         for (;;) {
2272                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("FollowerSleep"));
2273                 swait_event_interruptible_exclusive(rdp->nocb_wq,
2274                                          READ_ONCE(rdp->nocb_follower_head));
2275                 if (smp_load_acquire(&rdp->nocb_follower_head)) {
2276                         /* ^^^ Ensure CB invocation follows _head test. */
2277                         return;
2278                 }
2279                 WARN_ON(signal_pending(current));
2280                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WokeEmpty"));
2281         }
2282 }
2283 
2284 /*
2285  * Per-rcu_data kthread, but only for no-CBs CPUs.  Each kthread invokes
2286  * callbacks queued by the corresponding no-CBs CPU, however, there is
2287  * an optional leader-follower relationship so that the grace-period
2288  * kthreads don't have to do quite so many wakeups.
2289  */
2290 static int rcu_nocb_kthread(void *arg)
2291 {
2292         int c, cl;
2293         unsigned long flags;
2294         struct rcu_head *list;
2295         struct rcu_head *next;
2296         struct rcu_head **tail;
2297         struct rcu_data *rdp = arg;
2298 
2299         /* Each pass through this loop invokes one batch of callbacks */
2300         for (;;) {
2301                 /* Wait for callbacks. */
2302                 if (rdp->nocb_leader == rdp)
2303                         nocb_leader_wait(rdp);
2304                 else
2305                         nocb_follower_wait(rdp);
2306 
2307                 /* Pull the ready-to-invoke callbacks onto local list. */
2308                 raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2309                 list = rdp->nocb_follower_head;
2310                 rdp->nocb_follower_head = NULL;
2311                 tail = rdp->nocb_follower_tail;
2312                 rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2313                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2314                 BUG_ON(!list);
2315                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WokeNonEmpty"));
2316 
2317                 /* Each pass through the following loop invokes a callback. */
2318                 trace_rcu_batch_start(rdp->rsp->name,
2319                                       atomic_long_read(&rdp->nocb_q_count_lazy),
2320                                       atomic_long_read(&rdp->nocb_q_count), -1);
2321                 c = cl = 0;
2322                 while (list) {
2323                         next = list->next;
2324                         /* Wait for enqueuing to complete, if needed. */
2325                         while (next == NULL && &list->next != tail) {
2326                                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2327                                                     TPS("WaitQueue"));
2328                                 schedule_timeout_interruptible(1);
2329                                 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2330                                                     TPS("WokeQueue"));
2331                                 next = list->next;
2332                         }
2333                         debug_rcu_head_unqueue(list);
2334                         local_bh_disable();
2335                         if (__rcu_reclaim(rdp->rsp->name, list))
2336                                 cl++;
2337                         c++;
2338                         local_bh_enable();
2339                         cond_resched_tasks_rcu_qs();
2340                         list = next;
2341                 }
2342                 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2343                 smp_mb__before_atomic();  /* _add after CB invocation. */
2344                 atomic_long_add(-c, &rdp->nocb_q_count);
2345                 atomic_long_add(-cl, &rdp->nocb_q_count_lazy);
2346         }
2347         return 0;
2348 }
2349 
2350 /* Is a deferred wakeup of rcu_nocb_kthread() required? */
2351 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2352 {
2353         return READ_ONCE(rdp->nocb_defer_wakeup);
2354 }
2355 
2356 /* Do a deferred wakeup of rcu_nocb_kthread(). */
2357 static void do_nocb_deferred_wakeup_common(struct rcu_data *rdp)
2358 {
2359         unsigned long flags;
2360         int ndw;
2361 
2362         raw_spin_lock_irqsave(&rdp->nocb_lock, flags);
2363         if (!rcu_nocb_need_deferred_wakeup(rdp)) {
2364                 raw_spin_unlock_irqrestore(&rdp->nocb_lock, flags);
2365                 return;
2366         }
2367         ndw = READ_ONCE(rdp->nocb_defer_wakeup);
2368         WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOCB_WAKE_NOT);
2369         __wake_nocb_leader(rdp, ndw == RCU_NOCB_WAKE_FORCE, flags);
2370         trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake"));
2371 }
2372 
2373 /* Do a deferred wakeup of rcu_nocb_kthread() from a timer handler. */
2374 static void do_nocb_deferred_wakeup_timer(struct timer_list *t)
2375 {
2376         struct rcu_data *rdp = from_timer(rdp, t, nocb_timer);
2377 
2378         do_nocb_deferred_wakeup_common(rdp);
2379 }
2380 
2381 /*
2382  * Do a deferred wakeup of rcu_nocb_kthread() from fastpath.
2383  * This means we do an inexact common-case check.  Note that if
2384  * we miss, ->nocb_timer will eventually clean things up.
2385  */
2386 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2387 {
2388         if (rcu_nocb_need_deferred_wakeup(rdp))
2389                 do_nocb_deferred_wakeup_common(rdp);
2390 }
2391 
2392 void __init rcu_init_nohz(void)
2393 {
2394         int cpu;
2395         bool need_rcu_nocb_mask = false;
2396         struct rcu_state *rsp;
2397 
2398 #if defined(CONFIG_NO_HZ_FULL)
2399         if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask))
2400                 need_rcu_nocb_mask = true;
2401 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2402 
2403         if (!cpumask_available(rcu_nocb_mask) && need_rcu_nocb_mask) {
2404                 if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) {
2405                         pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n");
2406                         return;
2407                 }
2408         }
2409         if (!cpumask_available(rcu_nocb_mask))
2410                 return;
2411 
2412 #if defined(CONFIG_NO_HZ_FULL)
2413         if (tick_nohz_full_running)
2414                 cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask);
2415 #endif /* #if defined(CONFIG_NO_HZ_FULL) */
2416 
2417         if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
2418                 pr_info("\tNote: kernel parameter 'rcu_nocbs=', 'nohz_full', or 'isolcpus=' contains nonexistent CPUs.\n");
2419                 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
2420                             rcu_nocb_mask);
2421         }
2422         if (cpumask_empty(rcu_nocb_mask))
2423                 pr_info("\tOffload RCU callbacks from CPUs: (none).\n");
2424         else
2425                 pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n",
2426                         cpumask_pr_args(rcu_nocb_mask));
2427         if (rcu_nocb_poll)
2428                 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
2429 
2430         for_each_rcu_flavor(rsp) {
2431                 for_each_cpu(cpu, rcu_nocb_mask)
2432                         init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu));
2433                 rcu_organize_nocb_kthreads(rsp);
2434         }
2435 }
2436 
2437 /* Initialize per-rcu_data variables for no-CBs CPUs. */
2438 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2439 {
2440         rdp->nocb_tail = &rdp->nocb_head;
2441         init_swait_queue_head(&rdp->nocb_wq);
2442         rdp->nocb_follower_tail = &rdp->nocb_follower_head;
2443         raw_spin_lock_init(&rdp->nocb_lock);
2444         timer_setup(&rdp->nocb_timer, do_nocb_deferred_wakeup_timer, 0);
2445 }
2446 
2447 /*
2448  * If the specified CPU is a no-CBs CPU that does not already have its
2449  * rcuo kthread for the specified RCU flavor, spawn it.  If the CPUs are
2450  * brought online out of order, this can require re-organizing the
2451  * leader-follower relationships.
2452  */
2453 static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu)
2454 {
2455         struct rcu_data *rdp;
2456         struct rcu_data *rdp_last;
2457         struct rcu_data *rdp_old_leader;
2458         struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu);
2459         struct task_struct *t;
2460 
2461         /*
2462          * If this isn't a no-CBs CPU or if it already has an rcuo kthread,
2463          * then nothing to do.
2464          */
2465         if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread)
2466                 return;
2467 
2468         /* If we didn't spawn the leader first, reorganize! */
2469         rdp_old_leader = rdp_spawn->nocb_leader;
2470         if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) {
2471                 rdp_last = NULL;
2472                 rdp = rdp_old_leader;
2473                 do {
2474                         rdp->nocb_leader = rdp_spawn;
2475                         if (rdp_last && rdp != rdp_spawn)
2476                                 rdp_last->nocb_next_follower = rdp;
2477                         if (rdp == rdp_spawn) {
2478                                 rdp = rdp->nocb_next_follower;
2479                         } else {
2480                                 rdp_last = rdp;
2481                                 rdp = rdp->nocb_next_follower;
2482                                 rdp_last->nocb_next_follower = NULL;
2483                         }
2484                 } while (rdp);
2485                 rdp_spawn->nocb_next_follower = rdp_old_leader;
2486         }
2487 
2488         /* Spawn the kthread for this CPU and RCU flavor. */
2489         t = kthread_run(rcu_nocb_kthread, rdp_spawn,
2490                         "rcuo%c/%d", rsp->abbr, cpu);
2491         BUG_ON(IS_ERR(t));
2492         WRITE_ONCE(rdp_spawn->nocb_kthread, t);
2493 }
2494 
2495 /*
2496  * If the specified CPU is a no-CBs CPU that does not already have its
2497  * rcuo kthreads, spawn them.
2498  */
2499 static void rcu_spawn_all_nocb_kthreads(int cpu)
2500 {
2501         struct rcu_state *rsp;
2502 
2503         if (rcu_scheduler_fully_active)
2504                 for_each_rcu_flavor(rsp)
2505                         rcu_spawn_one_nocb_kthread(rsp, cpu);
2506 }
2507 
2508 /*
2509  * Once the scheduler is running, spawn rcuo kthreads for all online
2510  * no-CBs CPUs.  This assumes that the early_initcall()s happen before
2511  * non-boot CPUs come online -- if this changes, we will need to add
2512  * some mutual exclusion.
2513  */
2514 static void __init rcu_spawn_nocb_kthreads(void)
2515 {
2516         int cpu;
2517 
2518         for_each_online_cpu(cpu)
2519                 rcu_spawn_all_nocb_kthreads(cpu);
2520 }
2521 
2522 /* How many follower CPU IDs per leader?  Default of -1 for sqrt(nr_cpu_ids). */
2523 static int rcu_nocb_leader_stride = -1;
2524 module_param(rcu_nocb_leader_stride, int, 0444);
2525 
2526 /*
2527  * Initialize leader-follower relationships for all no-CBs CPU.
2528  */
2529 static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp)
2530 {
2531         int cpu;
2532         int ls = rcu_nocb_leader_stride;
2533         int nl = 0;  /* Next leader. */
2534         struct rcu_data *rdp;
2535         struct rcu_data *rdp_leader = NULL;  /* Suppress misguided gcc warn. */
2536         struct rcu_data *rdp_prev = NULL;
2537 
2538         if (!cpumask_available(rcu_nocb_mask))
2539                 return;
2540         if (ls == -1) {
2541                 ls = int_sqrt(nr_cpu_ids);
2542                 rcu_nocb_leader_stride = ls;
2543         }
2544 
2545         /*
2546          * Each pass through this loop sets up one rcu_data structure.
2547          * Should the corresponding CPU come online in the future, then
2548          * we will spawn the needed set of rcu_nocb_kthread() kthreads.
2549          */
2550         for_each_cpu(cpu, rcu_nocb_mask) {
2551                 rdp = per_cpu_ptr(rsp->rda, cpu);
2552                 if (rdp->cpu >= nl) {
2553                         /* New leader, set up for followers & next leader. */
2554                         nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls;
2555                         rdp->nocb_leader = rdp;
2556                         rdp_leader = rdp;
2557                 } else {
2558                         /* Another follower, link to previous leader. */
2559                         rdp->nocb_leader = rdp_leader;
2560                         rdp_prev->nocb_next_follower = rdp;
2561                 }
2562                 rdp_prev = rdp;
2563         }
2564 }
2565 
2566 /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2567 static bool init_nocb_callback_list(struct rcu_data *rdp)
2568 {
2569         if (!rcu_is_nocb_cpu(rdp->cpu))
2570                 return false;
2571 
2572         /* If there are early-boot callbacks, move them to nocb lists. */
2573         if (!rcu_segcblist_empty(&rdp->cblist)) {
2574                 rdp->nocb_head = rcu_segcblist_head(&rdp->cblist);
2575                 rdp->nocb_tail = rcu_segcblist_tail(&rdp->cblist);
2576                 atomic_long_set(&rdp->nocb_q_count,
2577                                 rcu_segcblist_n_cbs(&rdp->cblist));
2578                 atomic_long_set(&rdp->nocb_q_count_lazy,
2579                                 rcu_segcblist_n_lazy_cbs(&rdp->cblist));
2580                 rcu_segcblist_init(&rdp->cblist);
2581         }
2582         rcu_segcblist_disable(&rdp->cblist);
2583         return true;
2584 }
2585 
2586 #else /* #ifdef CONFIG_RCU_NOCB_CPU */
2587 
2588 static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu)
2589 {
2590         WARN_ON_ONCE(1); /* Should be dead code. */
2591         return false;
2592 }
2593 
2594 static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq)
2595 {
2596 }
2597 
2598 static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp)
2599 {
2600         return NULL;
2601 }
2602 
2603 static void rcu_init_one_nocb(struct rcu_node *rnp)
2604 {
2605 }
2606 
2607 static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2608                             bool lazy, unsigned long flags)
2609 {
2610         return false;
2611 }
2612 
2613 static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_data *my_rdp,
2614                                                      struct rcu_data *rdp,
2615                                                      unsigned long flags)
2616 {
2617         return false;
2618 }
2619 
2620 static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2621 {
2622 }
2623 
2624 static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp)
2625 {
2626         return false;
2627 }
2628 
2629 static void do_nocb_deferred_wakeup(struct rcu_data *rdp)
2630 {
2631 }
2632 
2633 static void rcu_spawn_all_nocb_kthreads(int cpu)
2634 {
2635 }
2636 
2637 static void __init rcu_spawn_nocb_kthreads(void)
2638 {
2639 }
2640 
2641 static bool init_nocb_callback_list(struct rcu_data *rdp)
2642 {
2643         return false;
2644 }
2645 
2646 #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2647 
2648 /*
2649  * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the
2650  * grace-period kthread will do force_quiescent_state() processing?
2651  * The idea is to avoid waking up RCU core processing on such a
2652  * CPU unless the grace period has extended for too long.
2653  *
2654  * This code relies on the fact that all NO_HZ_FULL CPUs are also
2655  * CONFIG_RCU_NOCB_CPU CPUs.
2656  */
2657 static bool rcu_nohz_full_cpu(struct rcu_state *rsp)
2658 {
2659 #ifdef CONFIG_NO_HZ_FULL
2660         if (tick_nohz_full_cpu(smp_processor_id()) &&
2661             (!rcu_gp_in_progress(rsp) ||
2662              ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ)))
2663                 return true;
2664 #endif /* #ifdef CONFIG_NO_HZ_FULL */
2665         return false;
2666 }
2667 
2668 /*
2669  * Bind the RCU grace-period kthreads to the housekeeping CPU.
2670  */
2671 static void rcu_bind_gp_kthread(void)
2672 {
2673         if (!tick_nohz_full_enabled())
2674                 return;
2675         housekeeping_affine(current, HK_FLAG_RCU);
2676 }
2677 
2678 /* Record the current task on dyntick-idle entry. */
2679 static void rcu_dynticks_task_enter(void)
2680 {
2681 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2682         WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id());
2683 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2684 }
2685 
2686 /* Record no current task on dyntick-idle exit. */
2687 static void rcu_dynticks_task_exit(void)
2688 {
2689 #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL)
2690         WRITE_ONCE(current->rcu_tasks_idle_cpu, -1);
2691 #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */
2692 }
2693 

~ [ 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