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

TOMOYO Linux Cross Reference
Linux/kernel/rcutree.c

Version: ~ [ linux-5.1-rc1 ] ~ [ linux-5.0.3 ] ~ [ linux-4.20.17 ] ~ [ linux-4.19.30 ] ~ [ linux-4.18.20 ] ~ [ linux-4.17.19 ] ~ [ linux-4.16.18 ] ~ [ linux-4.15.18 ] ~ [ linux-4.14.107 ] ~ [ linux-4.13.16 ] ~ [ linux-4.12.14 ] ~ [ linux-4.11.12 ] ~ [ linux-4.10.17 ] ~ [ linux-4.9.164 ] ~ [ linux-4.8.17 ] ~ [ linux-4.7.10 ] ~ [ linux-4.6.7 ] ~ [ linux-4.5.7 ] ~ [ linux-4.4.176 ] ~ [ linux-4.3.6 ] ~ [ linux-4.2.8 ] ~ [ linux-4.1.52 ] ~ [ linux-4.0.9 ] ~ [ linux-3.19.8 ] ~ [ linux-3.18.136 ] ~ [ linux-3.17.8 ] ~ [ linux-3.16.63 ] ~ [ linux-3.15.10 ] ~ [ linux-3.14.79 ] ~ [ linux-3.13.11 ] ~ [ linux-3.12.74 ] ~ [ linux-3.11.10 ] ~ [ linux-3.10.108 ] ~ [ linux-3.9.11 ] ~ [ linux-3.8.13 ] ~ [ linux-3.7.10 ] ~ [ linux-3.6.11 ] ~ [ linux-3.5.7 ] ~ [ linux-3.4.113 ] ~ [ linux-3.3.8 ] ~ [ linux-3.2.102 ] ~ [ linux-3.1.10 ] ~ [ linux-3.0.101 ] ~ [ linux-2.6.39.4 ] ~ [ linux-2.6.38.8 ] ~ [ linux-2.6.37.6 ] ~ [ linux-2.6.36.4 ] ~ [ linux-2.6.35.14 ] ~ [ linux-2.6.34.15 ] ~ [ linux-2.6.33.20 ] ~ [ linux-2.6.32.71 ] ~ [ linux-2.6.0 ] ~ [ linux-2.4.37.11 ] ~ [ unix-v6-master ] ~ [ ccs-tools-1.8.5 ] ~ [ policy-sample ] ~
Architecture: ~ [ i386 ] ~ [ alpha ] ~ [ m68k ] ~ [ mips ] ~ [ ppc ] ~ [ sparc ] ~ [ sparc64 ] ~

  1 /*
  2  * Read-Copy Update mechanism for mutual exclusion
  3  *
  4  * This program is free software; you can redistribute it and/or modify
  5  * it under the terms of the GNU General Public License as published by
  6  * the Free Software Foundation; either version 2 of the License, or
  7  * (at your option) any later version.
  8  *
  9  * This program is distributed in the hope that it will be useful,
 10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
 11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 12  * GNU General Public License for more details.
 13  *
 14  * You should have received a copy of the GNU General Public License
 15  * along with this program; if not, write to the Free Software
 16  * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
 17  *
 18  * Copyright IBM Corporation, 2008
 19  *
 20  * Authors: Dipankar Sarma <dipankar@in.ibm.com>
 21  *          Manfred Spraul <manfred@colorfullife.com>
 22  *          Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
 23  *
 24  * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
 25  * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
 26  *
 27  * For detailed explanation of Read-Copy Update mechanism see -
 28  *      Documentation/RCU
 29  */
 30 #include <linux/types.h>
 31 #include <linux/kernel.h>
 32 #include <linux/init.h>
 33 #include <linux/spinlock.h>
 34 #include <linux/smp.h>
 35 #include <linux/rcupdate.h>
 36 #include <linux/interrupt.h>
 37 #include <linux/sched.h>
 38 #include <linux/nmi.h>
 39 #include <linux/atomic.h>
 40 #include <linux/bitops.h>
 41 #include <linux/module.h>
 42 #include <linux/completion.h>
 43 #include <linux/moduleparam.h>
 44 #include <linux/percpu.h>
 45 #include <linux/notifier.h>
 46 #include <linux/cpu.h>
 47 #include <linux/mutex.h>
 48 #include <linux/time.h>
 49 #include <linux/kernel_stat.h>
 50 #include <linux/wait.h>
 51 #include <linux/kthread.h>
 52 #include <linux/prefetch.h>
 53 
 54 #include "rcutree.h"
 55 
 56 /* Data structures. */
 57 
 58 static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
 59 
 60 #define RCU_STATE_INITIALIZER(structname) { \
 61         .level = { &structname.node[0] }, \
 62         .levelcnt = { \
 63                 NUM_RCU_LVL_0,  /* root of hierarchy. */ \
 64                 NUM_RCU_LVL_1, \
 65                 NUM_RCU_LVL_2, \
 66                 NUM_RCU_LVL_3, \
 67                 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
 68         }, \
 69         .signaled = RCU_GP_IDLE, \
 70         .gpnum = -300, \
 71         .completed = -300, \
 72         .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
 73         .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
 74         .n_force_qs = 0, \
 75         .n_force_qs_ngp = 0, \
 76         .name = #structname, \
 77 }
 78 
 79 struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
 80 DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
 81 
 82 struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
 83 DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
 84 
 85 static struct rcu_state *rcu_state;
 86 
 87 /*
 88  * The rcu_scheduler_active variable transitions from zero to one just
 89  * before the first task is spawned.  So when this variable is zero, RCU
 90  * can assume that there is but one task, allowing RCU to (for example)
 91  * optimized synchronize_sched() to a simple barrier().  When this variable
 92  * is one, RCU must actually do all the hard work required to detect real
 93  * grace periods.  This variable is also used to suppress boot-time false
 94  * positives from lockdep-RCU error checking.
 95  */
 96 int rcu_scheduler_active __read_mostly;
 97 EXPORT_SYMBOL_GPL(rcu_scheduler_active);
 98 
 99 /*
100  * The rcu_scheduler_fully_active variable transitions from zero to one
101  * during the early_initcall() processing, which is after the scheduler
102  * is capable of creating new tasks.  So RCU processing (for example,
103  * creating tasks for RCU priority boosting) must be delayed until after
104  * rcu_scheduler_fully_active transitions from zero to one.  We also
105  * currently delay invocation of any RCU callbacks until after this point.
106  *
107  * It might later prove better for people registering RCU callbacks during
108  * early boot to take responsibility for these callbacks, but one step at
109  * a time.
110  */
111 static int rcu_scheduler_fully_active __read_mostly;
112 
113 #ifdef CONFIG_RCU_BOOST
114 
115 /*
116  * Control variables for per-CPU and per-rcu_node kthreads.  These
117  * handle all flavors of RCU.
118  */
119 static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
120 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
121 DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
122 DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
123 DEFINE_PER_CPU(char, rcu_cpu_has_work);
124 
125 #endif /* #ifdef CONFIG_RCU_BOOST */
126 
127 static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
128 static void invoke_rcu_core(void);
129 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
130 
131 #define RCU_KTHREAD_PRIO 1      /* RT priority for per-CPU kthreads. */
132 
133 /*
134  * Track the rcutorture test sequence number and the update version
135  * number within a given test.  The rcutorture_testseq is incremented
136  * on every rcutorture module load and unload, so has an odd value
137  * when a test is running.  The rcutorture_vernum is set to zero
138  * when rcutorture starts and is incremented on each rcutorture update.
139  * These variables enable correlating rcutorture output with the
140  * RCU tracing information.
141  */
142 unsigned long rcutorture_testseq;
143 unsigned long rcutorture_vernum;
144 
145 /*
146  * Return true if an RCU grace period is in progress.  The ACCESS_ONCE()s
147  * permit this function to be invoked without holding the root rcu_node
148  * structure's ->lock, but of course results can be subject to change.
149  */
150 static int rcu_gp_in_progress(struct rcu_state *rsp)
151 {
152         return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
153 }
154 
155 /*
156  * Note a quiescent state.  Because we do not need to know
157  * how many quiescent states passed, just if there was at least
158  * one since the start of the grace period, this just sets a flag.
159  */
160 void rcu_sched_qs(int cpu)
161 {
162         struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
163 
164         rdp->passed_quiesc_completed = rdp->gpnum - 1;
165         barrier();
166         rdp->passed_quiesc = 1;
167 }
168 
169 void rcu_bh_qs(int cpu)
170 {
171         struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
172 
173         rdp->passed_quiesc_completed = rdp->gpnum - 1;
174         barrier();
175         rdp->passed_quiesc = 1;
176 }
177 
178 /*
179  * Note a context switch.  This is a quiescent state for RCU-sched,
180  * and requires special handling for preemptible RCU.
181  */
182 void rcu_note_context_switch(int cpu)
183 {
184         rcu_sched_qs(cpu);
185         rcu_preempt_note_context_switch(cpu);
186 }
187 EXPORT_SYMBOL_GPL(rcu_note_context_switch);
188 
189 #ifdef CONFIG_NO_HZ
190 DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
191         .dynticks_nesting = 1,
192         .dynticks = ATOMIC_INIT(1),
193 };
194 #endif /* #ifdef CONFIG_NO_HZ */
195 
196 static int blimit = 10;         /* Maximum callbacks per softirq. */
197 static int qhimark = 10000;     /* If this many pending, ignore blimit. */
198 static int qlowmark = 100;      /* Once only this many pending, use blimit. */
199 
200 module_param(blimit, int, 0);
201 module_param(qhimark, int, 0);
202 module_param(qlowmark, int, 0);
203 
204 int rcu_cpu_stall_suppress __read_mostly;
205 module_param(rcu_cpu_stall_suppress, int, 0644);
206 
207 static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
208 static int rcu_pending(int cpu);
209 
210 /*
211  * Return the number of RCU-sched batches processed thus far for debug & stats.
212  */
213 long rcu_batches_completed_sched(void)
214 {
215         return rcu_sched_state.completed;
216 }
217 EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
218 
219 /*
220  * Return the number of RCU BH batches processed thus far for debug & stats.
221  */
222 long rcu_batches_completed_bh(void)
223 {
224         return rcu_bh_state.completed;
225 }
226 EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
227 
228 /*
229  * Force a quiescent state for RCU BH.
230  */
231 void rcu_bh_force_quiescent_state(void)
232 {
233         force_quiescent_state(&rcu_bh_state, 0);
234 }
235 EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
236 
237 /*
238  * Record the number of times rcutorture tests have been initiated and
239  * terminated.  This information allows the debugfs tracing stats to be
240  * correlated to the rcutorture messages, even when the rcutorture module
241  * is being repeatedly loaded and unloaded.  In other words, we cannot
242  * store this state in rcutorture itself.
243  */
244 void rcutorture_record_test_transition(void)
245 {
246         rcutorture_testseq++;
247         rcutorture_vernum = 0;
248 }
249 EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
250 
251 /*
252  * Record the number of writer passes through the current rcutorture test.
253  * This is also used to correlate debugfs tracing stats with the rcutorture
254  * messages.
255  */
256 void rcutorture_record_progress(unsigned long vernum)
257 {
258         rcutorture_vernum++;
259 }
260 EXPORT_SYMBOL_GPL(rcutorture_record_progress);
261 
262 /*
263  * Force a quiescent state for RCU-sched.
264  */
265 void rcu_sched_force_quiescent_state(void)
266 {
267         force_quiescent_state(&rcu_sched_state, 0);
268 }
269 EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
270 
271 /*
272  * Does the CPU have callbacks ready to be invoked?
273  */
274 static int
275 cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
276 {
277         return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
278 }
279 
280 /*
281  * Does the current CPU require a yet-as-unscheduled grace period?
282  */
283 static int
284 cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
285 {
286         return *rdp->nxttail[RCU_DONE_TAIL +
287                              ACCESS_ONCE(rsp->completed) != rdp->completed] &&
288                !rcu_gp_in_progress(rsp);
289 }
290 
291 /*
292  * Return the root node of the specified rcu_state structure.
293  */
294 static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
295 {
296         return &rsp->node[0];
297 }
298 
299 #ifdef CONFIG_SMP
300 
301 /*
302  * If the specified CPU is offline, tell the caller that it is in
303  * a quiescent state.  Otherwise, whack it with a reschedule IPI.
304  * Grace periods can end up waiting on an offline CPU when that
305  * CPU is in the process of coming online -- it will be added to the
306  * rcu_node bitmasks before it actually makes it online.  The same thing
307  * can happen while a CPU is in the process of coming online.  Because this
308  * race is quite rare, we check for it after detecting that the grace
309  * period has been delayed rather than checking each and every CPU
310  * each and every time we start a new grace period.
311  */
312 static int rcu_implicit_offline_qs(struct rcu_data *rdp)
313 {
314         /*
315          * If the CPU is offline, it is in a quiescent state.  We can
316          * trust its state not to change because interrupts are disabled.
317          */
318         if (cpu_is_offline(rdp->cpu)) {
319                 rdp->offline_fqs++;
320                 return 1;
321         }
322 
323         /* If preemptible RCU, no point in sending reschedule IPI. */
324         if (rdp->preemptible)
325                 return 0;
326 
327         /* The CPU is online, so send it a reschedule IPI. */
328         if (rdp->cpu != smp_processor_id())
329                 smp_send_reschedule(rdp->cpu);
330         else
331                 set_need_resched();
332         rdp->resched_ipi++;
333         return 0;
334 }
335 
336 #endif /* #ifdef CONFIG_SMP */
337 
338 #ifdef CONFIG_NO_HZ
339 
340 /**
341  * rcu_enter_nohz - inform RCU that current CPU is entering nohz
342  *
343  * Enter nohz mode, in other words, -leave- the mode in which RCU
344  * read-side critical sections can occur.  (Though RCU read-side
345  * critical sections can occur in irq handlers in nohz mode, a possibility
346  * handled by rcu_irq_enter() and rcu_irq_exit()).
347  */
348 void rcu_enter_nohz(void)
349 {
350         unsigned long flags;
351         struct rcu_dynticks *rdtp;
352 
353         local_irq_save(flags);
354         rdtp = &__get_cpu_var(rcu_dynticks);
355         if (--rdtp->dynticks_nesting) {
356                 local_irq_restore(flags);
357                 return;
358         }
359         /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
360         smp_mb__before_atomic_inc();  /* See above. */
361         atomic_inc(&rdtp->dynticks);
362         smp_mb__after_atomic_inc();  /* Force ordering with next sojourn. */
363         WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
364         local_irq_restore(flags);
365 
366         /* If the interrupt queued a callback, get out of dyntick mode. */
367         if (in_irq() &&
368             (__get_cpu_var(rcu_sched_data).nxtlist ||
369              __get_cpu_var(rcu_bh_data).nxtlist ||
370              rcu_preempt_needs_cpu(smp_processor_id())))
371                 set_need_resched();
372 }
373 
374 /*
375  * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
376  *
377  * Exit nohz mode, in other words, -enter- the mode in which RCU
378  * read-side critical sections normally occur.
379  */
380 void rcu_exit_nohz(void)
381 {
382         unsigned long flags;
383         struct rcu_dynticks *rdtp;
384 
385         local_irq_save(flags);
386         rdtp = &__get_cpu_var(rcu_dynticks);
387         if (rdtp->dynticks_nesting++) {
388                 local_irq_restore(flags);
389                 return;
390         }
391         smp_mb__before_atomic_inc();  /* Force ordering w/previous sojourn. */
392         atomic_inc(&rdtp->dynticks);
393         /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
394         smp_mb__after_atomic_inc();  /* See above. */
395         WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
396         local_irq_restore(flags);
397 }
398 
399 /**
400  * rcu_nmi_enter - inform RCU of entry to NMI context
401  *
402  * If the CPU was idle with dynamic ticks active, and there is no
403  * irq handler running, this updates rdtp->dynticks_nmi to let the
404  * RCU grace-period handling know that the CPU is active.
405  */
406 void rcu_nmi_enter(void)
407 {
408         struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
409 
410         if (rdtp->dynticks_nmi_nesting == 0 &&
411             (atomic_read(&rdtp->dynticks) & 0x1))
412                 return;
413         rdtp->dynticks_nmi_nesting++;
414         smp_mb__before_atomic_inc();  /* Force delay from prior write. */
415         atomic_inc(&rdtp->dynticks);
416         /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
417         smp_mb__after_atomic_inc();  /* See above. */
418         WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
419 }
420 
421 /**
422  * rcu_nmi_exit - inform RCU of exit from NMI context
423  *
424  * If the CPU was idle with dynamic ticks active, and there is no
425  * irq handler running, this updates rdtp->dynticks_nmi to let the
426  * RCU grace-period handling know that the CPU is no longer active.
427  */
428 void rcu_nmi_exit(void)
429 {
430         struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
431 
432         if (rdtp->dynticks_nmi_nesting == 0 ||
433             --rdtp->dynticks_nmi_nesting != 0)
434                 return;
435         /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
436         smp_mb__before_atomic_inc();  /* See above. */
437         atomic_inc(&rdtp->dynticks);
438         smp_mb__after_atomic_inc();  /* Force delay to next write. */
439         WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
440 }
441 
442 /**
443  * rcu_irq_enter - inform RCU of entry to hard irq context
444  *
445  * If the CPU was idle with dynamic ticks active, this updates the
446  * rdtp->dynticks to let the RCU handling know that the CPU is active.
447  */
448 void rcu_irq_enter(void)
449 {
450         rcu_exit_nohz();
451 }
452 
453 /**
454  * rcu_irq_exit - inform RCU of exit from hard irq context
455  *
456  * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
457  * to put let the RCU handling be aware that the CPU is going back to idle
458  * with no ticks.
459  */
460 void rcu_irq_exit(void)
461 {
462         rcu_enter_nohz();
463 }
464 
465 #ifdef CONFIG_SMP
466 
467 /*
468  * Snapshot the specified CPU's dynticks counter so that we can later
469  * credit them with an implicit quiescent state.  Return 1 if this CPU
470  * is in dynticks idle mode, which is an extended quiescent state.
471  */
472 static int dyntick_save_progress_counter(struct rcu_data *rdp)
473 {
474         rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
475         return 0;
476 }
477 
478 /*
479  * Return true if the specified CPU has passed through a quiescent
480  * state by virtue of being in or having passed through an dynticks
481  * idle state since the last call to dyntick_save_progress_counter()
482  * for this same CPU.
483  */
484 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
485 {
486         unsigned long curr;
487         unsigned long snap;
488 
489         curr = (unsigned long)atomic_add_return(0, &rdp->dynticks->dynticks);
490         snap = (unsigned long)rdp->dynticks_snap;
491 
492         /*
493          * If the CPU passed through or entered a dynticks idle phase with
494          * no active irq/NMI handlers, then we can safely pretend that the CPU
495          * already acknowledged the request to pass through a quiescent
496          * state.  Either way, that CPU cannot possibly be in an RCU
497          * read-side critical section that started before the beginning
498          * of the current RCU grace period.
499          */
500         if ((curr & 0x1) == 0 || ULONG_CMP_GE(curr, snap + 2)) {
501                 rdp->dynticks_fqs++;
502                 return 1;
503         }
504 
505         /* Go check for the CPU being offline. */
506         return rcu_implicit_offline_qs(rdp);
507 }
508 
509 #endif /* #ifdef CONFIG_SMP */
510 
511 #else /* #ifdef CONFIG_NO_HZ */
512 
513 #ifdef CONFIG_SMP
514 
515 static int dyntick_save_progress_counter(struct rcu_data *rdp)
516 {
517         return 0;
518 }
519 
520 static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
521 {
522         return rcu_implicit_offline_qs(rdp);
523 }
524 
525 #endif /* #ifdef CONFIG_SMP */
526 
527 #endif /* #else #ifdef CONFIG_NO_HZ */
528 
529 int rcu_cpu_stall_suppress __read_mostly;
530 
531 static void record_gp_stall_check_time(struct rcu_state *rsp)
532 {
533         rsp->gp_start = jiffies;
534         rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
535 }
536 
537 static void print_other_cpu_stall(struct rcu_state *rsp)
538 {
539         int cpu;
540         long delta;
541         unsigned long flags;
542         struct rcu_node *rnp = rcu_get_root(rsp);
543 
544         /* Only let one CPU complain about others per time interval. */
545 
546         raw_spin_lock_irqsave(&rnp->lock, flags);
547         delta = jiffies - rsp->jiffies_stall;
548         if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
549                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
550                 return;
551         }
552         rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
553 
554         /*
555          * Now rat on any tasks that got kicked up to the root rcu_node
556          * due to CPU offlining.
557          */
558         rcu_print_task_stall(rnp);
559         raw_spin_unlock_irqrestore(&rnp->lock, flags);
560 
561         /*
562          * OK, time to rat on our buddy...
563          * See Documentation/RCU/stallwarn.txt for info on how to debug
564          * RCU CPU stall warnings.
565          */
566         printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
567                rsp->name);
568         rcu_for_each_leaf_node(rsp, rnp) {
569                 raw_spin_lock_irqsave(&rnp->lock, flags);
570                 rcu_print_task_stall(rnp);
571                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
572                 if (rnp->qsmask == 0)
573                         continue;
574                 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
575                         if (rnp->qsmask & (1UL << cpu))
576                                 printk(" %d", rnp->grplo + cpu);
577         }
578         printk("} (detected by %d, t=%ld jiffies)\n",
579                smp_processor_id(), (long)(jiffies - rsp->gp_start));
580         trigger_all_cpu_backtrace();
581 
582         /* If so configured, complain about tasks blocking the grace period. */
583 
584         rcu_print_detail_task_stall(rsp);
585 
586         force_quiescent_state(rsp, 0);  /* Kick them all. */
587 }
588 
589 static void print_cpu_stall(struct rcu_state *rsp)
590 {
591         unsigned long flags;
592         struct rcu_node *rnp = rcu_get_root(rsp);
593 
594         /*
595          * OK, time to rat on ourselves...
596          * See Documentation/RCU/stallwarn.txt for info on how to debug
597          * RCU CPU stall warnings.
598          */
599         printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
600                rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
601         trigger_all_cpu_backtrace();
602 
603         raw_spin_lock_irqsave(&rnp->lock, flags);
604         if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
605                 rsp->jiffies_stall =
606                         jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
607         raw_spin_unlock_irqrestore(&rnp->lock, flags);
608 
609         set_need_resched();  /* kick ourselves to get things going. */
610 }
611 
612 static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
613 {
614         unsigned long j;
615         unsigned long js;
616         struct rcu_node *rnp;
617 
618         if (rcu_cpu_stall_suppress)
619                 return;
620         j = ACCESS_ONCE(jiffies);
621         js = ACCESS_ONCE(rsp->jiffies_stall);
622         rnp = rdp->mynode;
623         if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
624 
625                 /* We haven't checked in, so go dump stack. */
626                 print_cpu_stall(rsp);
627 
628         } else if (rcu_gp_in_progress(rsp) &&
629                    ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
630 
631                 /* They had a few time units to dump stack, so complain. */
632                 print_other_cpu_stall(rsp);
633         }
634 }
635 
636 static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
637 {
638         rcu_cpu_stall_suppress = 1;
639         return NOTIFY_DONE;
640 }
641 
642 /**
643  * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
644  *
645  * Set the stall-warning timeout way off into the future, thus preventing
646  * any RCU CPU stall-warning messages from appearing in the current set of
647  * RCU grace periods.
648  *
649  * The caller must disable hard irqs.
650  */
651 void rcu_cpu_stall_reset(void)
652 {
653         rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
654         rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
655         rcu_preempt_stall_reset();
656 }
657 
658 static struct notifier_block rcu_panic_block = {
659         .notifier_call = rcu_panic,
660 };
661 
662 static void __init check_cpu_stall_init(void)
663 {
664         atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
665 }
666 
667 /*
668  * Update CPU-local rcu_data state to record the newly noticed grace period.
669  * This is used both when we started the grace period and when we notice
670  * that someone else started the grace period.  The caller must hold the
671  * ->lock of the leaf rcu_node structure corresponding to the current CPU,
672  *  and must have irqs disabled.
673  */
674 static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
675 {
676         if (rdp->gpnum != rnp->gpnum) {
677                 /*
678                  * If the current grace period is waiting for this CPU,
679                  * set up to detect a quiescent state, otherwise don't
680                  * go looking for one.
681                  */
682                 rdp->gpnum = rnp->gpnum;
683                 if (rnp->qsmask & rdp->grpmask) {
684                         rdp->qs_pending = 1;
685                         rdp->passed_quiesc = 0;
686                 } else
687                         rdp->qs_pending = 0;
688         }
689 }
690 
691 static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
692 {
693         unsigned long flags;
694         struct rcu_node *rnp;
695 
696         local_irq_save(flags);
697         rnp = rdp->mynode;
698         if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
699             !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
700                 local_irq_restore(flags);
701                 return;
702         }
703         __note_new_gpnum(rsp, rnp, rdp);
704         raw_spin_unlock_irqrestore(&rnp->lock, flags);
705 }
706 
707 /*
708  * Did someone else start a new RCU grace period start since we last
709  * checked?  Update local state appropriately if so.  Must be called
710  * on the CPU corresponding to rdp.
711  */
712 static int
713 check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
714 {
715         unsigned long flags;
716         int ret = 0;
717 
718         local_irq_save(flags);
719         if (rdp->gpnum != rsp->gpnum) {
720                 note_new_gpnum(rsp, rdp);
721                 ret = 1;
722         }
723         local_irq_restore(flags);
724         return ret;
725 }
726 
727 /*
728  * Advance this CPU's callbacks, but only if the current grace period
729  * has ended.  This may be called only from the CPU to whom the rdp
730  * belongs.  In addition, the corresponding leaf rcu_node structure's
731  * ->lock must be held by the caller, with irqs disabled.
732  */
733 static void
734 __rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
735 {
736         /* Did another grace period end? */
737         if (rdp->completed != rnp->completed) {
738 
739                 /* Advance callbacks.  No harm if list empty. */
740                 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
741                 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
742                 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
743 
744                 /* Remember that we saw this grace-period completion. */
745                 rdp->completed = rnp->completed;
746 
747                 /*
748                  * If we were in an extended quiescent state, we may have
749                  * missed some grace periods that others CPUs handled on
750                  * our behalf. Catch up with this state to avoid noting
751                  * spurious new grace periods.  If another grace period
752                  * has started, then rnp->gpnum will have advanced, so
753                  * we will detect this later on.
754                  */
755                 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
756                         rdp->gpnum = rdp->completed;
757 
758                 /*
759                  * If RCU does not need a quiescent state from this CPU,
760                  * then make sure that this CPU doesn't go looking for one.
761                  */
762                 if ((rnp->qsmask & rdp->grpmask) == 0)
763                         rdp->qs_pending = 0;
764         }
765 }
766 
767 /*
768  * Advance this CPU's callbacks, but only if the current grace period
769  * has ended.  This may be called only from the CPU to whom the rdp
770  * belongs.
771  */
772 static void
773 rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
774 {
775         unsigned long flags;
776         struct rcu_node *rnp;
777 
778         local_irq_save(flags);
779         rnp = rdp->mynode;
780         if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
781             !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
782                 local_irq_restore(flags);
783                 return;
784         }
785         __rcu_process_gp_end(rsp, rnp, rdp);
786         raw_spin_unlock_irqrestore(&rnp->lock, flags);
787 }
788 
789 /*
790  * Do per-CPU grace-period initialization for running CPU.  The caller
791  * must hold the lock of the leaf rcu_node structure corresponding to
792  * this CPU.
793  */
794 static void
795 rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
796 {
797         /* Prior grace period ended, so advance callbacks for current CPU. */
798         __rcu_process_gp_end(rsp, rnp, rdp);
799 
800         /*
801          * Because this CPU just now started the new grace period, we know
802          * that all of its callbacks will be covered by this upcoming grace
803          * period, even the ones that were registered arbitrarily recently.
804          * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
805          *
806          * Other CPUs cannot be sure exactly when the grace period started.
807          * Therefore, their recently registered callbacks must pass through
808          * an additional RCU_NEXT_READY stage, so that they will be handled
809          * by the next RCU grace period.
810          */
811         rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
812         rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
813 
814         /* Set state so that this CPU will detect the next quiescent state. */
815         __note_new_gpnum(rsp, rnp, rdp);
816 }
817 
818 /*
819  * Start a new RCU grace period if warranted, re-initializing the hierarchy
820  * in preparation for detecting the next grace period.  The caller must hold
821  * the root node's ->lock, which is released before return.  Hard irqs must
822  * be disabled.
823  */
824 static void
825 rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
826         __releases(rcu_get_root(rsp)->lock)
827 {
828         struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
829         struct rcu_node *rnp = rcu_get_root(rsp);
830 
831         if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
832                 if (cpu_needs_another_gp(rsp, rdp))
833                         rsp->fqs_need_gp = 1;
834                 if (rnp->completed == rsp->completed) {
835                         raw_spin_unlock_irqrestore(&rnp->lock, flags);
836                         return;
837                 }
838                 raw_spin_unlock(&rnp->lock);     /* irqs remain disabled. */
839 
840                 /*
841                  * Propagate new ->completed value to rcu_node structures
842                  * so that other CPUs don't have to wait until the start
843                  * of the next grace period to process their callbacks.
844                  */
845                 rcu_for_each_node_breadth_first(rsp, rnp) {
846                         raw_spin_lock(&rnp->lock); /* irqs already disabled. */
847                         rnp->completed = rsp->completed;
848                         raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
849                 }
850                 local_irq_restore(flags);
851                 return;
852         }
853 
854         /* Advance to a new grace period and initialize state. */
855         rsp->gpnum++;
856         WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
857         rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
858         rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
859         record_gp_stall_check_time(rsp);
860 
861         /* Special-case the common single-level case. */
862         if (NUM_RCU_NODES == 1) {
863                 rcu_preempt_check_blocked_tasks(rnp);
864                 rnp->qsmask = rnp->qsmaskinit;
865                 rnp->gpnum = rsp->gpnum;
866                 rnp->completed = rsp->completed;
867                 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
868                 rcu_start_gp_per_cpu(rsp, rnp, rdp);
869                 rcu_preempt_boost_start_gp(rnp);
870                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
871                 return;
872         }
873 
874         raw_spin_unlock(&rnp->lock);  /* leave irqs disabled. */
875 
876 
877         /* Exclude any concurrent CPU-hotplug operations. */
878         raw_spin_lock(&rsp->onofflock);  /* irqs already disabled. */
879 
880         /*
881          * Set the quiescent-state-needed bits in all the rcu_node
882          * structures for all currently online CPUs in breadth-first
883          * order, starting from the root rcu_node structure.  This
884          * operation relies on the layout of the hierarchy within the
885          * rsp->node[] array.  Note that other CPUs will access only
886          * the leaves of the hierarchy, which still indicate that no
887          * grace period is in progress, at least until the corresponding
888          * leaf node has been initialized.  In addition, we have excluded
889          * CPU-hotplug operations.
890          *
891          * Note that the grace period cannot complete until we finish
892          * the initialization process, as there will be at least one
893          * qsmask bit set in the root node until that time, namely the
894          * one corresponding to this CPU, due to the fact that we have
895          * irqs disabled.
896          */
897         rcu_for_each_node_breadth_first(rsp, rnp) {
898                 raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
899                 rcu_preempt_check_blocked_tasks(rnp);
900                 rnp->qsmask = rnp->qsmaskinit;
901                 rnp->gpnum = rsp->gpnum;
902                 rnp->completed = rsp->completed;
903                 if (rnp == rdp->mynode)
904                         rcu_start_gp_per_cpu(rsp, rnp, rdp);
905                 rcu_preempt_boost_start_gp(rnp);
906                 raw_spin_unlock(&rnp->lock);    /* irqs remain disabled. */
907         }
908 
909         rnp = rcu_get_root(rsp);
910         raw_spin_lock(&rnp->lock);              /* irqs already disabled. */
911         rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
912         raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
913         raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
914 }
915 
916 /*
917  * Report a full set of quiescent states to the specified rcu_state
918  * data structure.  This involves cleaning up after the prior grace
919  * period and letting rcu_start_gp() start up the next grace period
920  * if one is needed.  Note that the caller must hold rnp->lock, as
921  * required by rcu_start_gp(), which will release it.
922  */
923 static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
924         __releases(rcu_get_root(rsp)->lock)
925 {
926         unsigned long gp_duration;
927 
928         WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
929 
930         /*
931          * Ensure that all grace-period and pre-grace-period activity
932          * is seen before the assignment to rsp->completed.
933          */
934         smp_mb(); /* See above block comment. */
935         gp_duration = jiffies - rsp->gp_start;
936         if (gp_duration > rsp->gp_max)
937                 rsp->gp_max = gp_duration;
938         rsp->completed = rsp->gpnum;
939         rsp->signaled = RCU_GP_IDLE;
940         rcu_start_gp(rsp, flags);  /* releases root node's rnp->lock. */
941 }
942 
943 /*
944  * Similar to rcu_report_qs_rdp(), for which it is a helper function.
945  * Allows quiescent states for a group of CPUs to be reported at one go
946  * to the specified rcu_node structure, though all the CPUs in the group
947  * must be represented by the same rcu_node structure (which need not be
948  * a leaf rcu_node structure, though it often will be).  That structure's
949  * lock must be held upon entry, and it is released before return.
950  */
951 static void
952 rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
953                   struct rcu_node *rnp, unsigned long flags)
954         __releases(rnp->lock)
955 {
956         struct rcu_node *rnp_c;
957 
958         /* Walk up the rcu_node hierarchy. */
959         for (;;) {
960                 if (!(rnp->qsmask & mask)) {
961 
962                         /* Our bit has already been cleared, so done. */
963                         raw_spin_unlock_irqrestore(&rnp->lock, flags);
964                         return;
965                 }
966                 rnp->qsmask &= ~mask;
967                 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
968 
969                         /* Other bits still set at this level, so done. */
970                         raw_spin_unlock_irqrestore(&rnp->lock, flags);
971                         return;
972                 }
973                 mask = rnp->grpmask;
974                 if (rnp->parent == NULL) {
975 
976                         /* No more levels.  Exit loop holding root lock. */
977 
978                         break;
979                 }
980                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
981                 rnp_c = rnp;
982                 rnp = rnp->parent;
983                 raw_spin_lock_irqsave(&rnp->lock, flags);
984                 WARN_ON_ONCE(rnp_c->qsmask);
985         }
986 
987         /*
988          * Get here if we are the last CPU to pass through a quiescent
989          * state for this grace period.  Invoke rcu_report_qs_rsp()
990          * to clean up and start the next grace period if one is needed.
991          */
992         rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
993 }
994 
995 /*
996  * Record a quiescent state for the specified CPU to that CPU's rcu_data
997  * structure.  This must be either called from the specified CPU, or
998  * called when the specified CPU is known to be offline (and when it is
999  * also known that no other CPU is concurrently trying to help the offline
1000  * CPU).  The lastcomp argument is used to make sure we are still in the
1001  * grace period of interest.  We don't want to end the current grace period
1002  * based on quiescent states detected in an earlier grace period!
1003  */
1004 static void
1005 rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
1006 {
1007         unsigned long flags;
1008         unsigned long mask;
1009         struct rcu_node *rnp;
1010 
1011         rnp = rdp->mynode;
1012         raw_spin_lock_irqsave(&rnp->lock, flags);
1013         if (lastcomp != rnp->completed) {
1014 
1015                 /*
1016                  * Someone beat us to it for this grace period, so leave.
1017                  * The race with GP start is resolved by the fact that we
1018                  * hold the leaf rcu_node lock, so that the per-CPU bits
1019                  * cannot yet be initialized -- so we would simply find our
1020                  * CPU's bit already cleared in rcu_report_qs_rnp() if this
1021                  * race occurred.
1022                  */
1023                 rdp->passed_quiesc = 0; /* try again later! */
1024                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1025                 return;
1026         }
1027         mask = rdp->grpmask;
1028         if ((rnp->qsmask & mask) == 0) {
1029                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1030         } else {
1031                 rdp->qs_pending = 0;
1032 
1033                 /*
1034                  * This GP can't end until cpu checks in, so all of our
1035                  * callbacks can be processed during the next GP.
1036                  */
1037                 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1038 
1039                 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1040         }
1041 }
1042 
1043 /*
1044  * Check to see if there is a new grace period of which this CPU
1045  * is not yet aware, and if so, set up local rcu_data state for it.
1046  * Otherwise, see if this CPU has just passed through its first
1047  * quiescent state for this grace period, and record that fact if so.
1048  */
1049 static void
1050 rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1051 {
1052         /* If there is now a new grace period, record and return. */
1053         if (check_for_new_grace_period(rsp, rdp))
1054                 return;
1055 
1056         /*
1057          * Does this CPU still need to do its part for current grace period?
1058          * If no, return and let the other CPUs do their part as well.
1059          */
1060         if (!rdp->qs_pending)
1061                 return;
1062 
1063         /*
1064          * Was there a quiescent state since the beginning of the grace
1065          * period? If no, then exit and wait for the next call.
1066          */
1067         if (!rdp->passed_quiesc)
1068                 return;
1069 
1070         /*
1071          * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1072          * judge of that).
1073          */
1074         rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
1075 }
1076 
1077 #ifdef CONFIG_HOTPLUG_CPU
1078 
1079 /*
1080  * Move a dying CPU's RCU callbacks to online CPU's callback list.
1081  * Synchronization is not required because this function executes
1082  * in stop_machine() context.
1083  */
1084 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1085 {
1086         int i;
1087         /* current DYING CPU is cleared in the cpu_online_mask */
1088         int receive_cpu = cpumask_any(cpu_online_mask);
1089         struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1090         struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
1091 
1092         if (rdp->nxtlist == NULL)
1093                 return;  /* irqs disabled, so comparison is stable. */
1094 
1095         *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1096         receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1097         receive_rdp->qlen += rdp->qlen;
1098         receive_rdp->n_cbs_adopted += rdp->qlen;
1099         rdp->n_cbs_orphaned += rdp->qlen;
1100 
1101         rdp->nxtlist = NULL;
1102         for (i = 0; i < RCU_NEXT_SIZE; i++)
1103                 rdp->nxttail[i] = &rdp->nxtlist;
1104         rdp->qlen = 0;
1105 }
1106 
1107 /*
1108  * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1109  * and move all callbacks from the outgoing CPU to the current one.
1110  * There can only be one CPU hotplug operation at a time, so no other
1111  * CPU can be attempting to update rcu_cpu_kthread_task.
1112  */
1113 static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1114 {
1115         unsigned long flags;
1116         unsigned long mask;
1117         int need_report = 0;
1118         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1119         struct rcu_node *rnp;
1120 
1121         rcu_stop_cpu_kthread(cpu);
1122 
1123         /* Exclude any attempts to start a new grace period. */
1124         raw_spin_lock_irqsave(&rsp->onofflock, flags);
1125 
1126         /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1127         rnp = rdp->mynode;      /* this is the outgoing CPU's rnp. */
1128         mask = rdp->grpmask;    /* rnp->grplo is constant. */
1129         do {
1130                 raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1131                 rnp->qsmaskinit &= ~mask;
1132                 if (rnp->qsmaskinit != 0) {
1133                         if (rnp != rdp->mynode)
1134                                 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1135                         break;
1136                 }
1137                 if (rnp == rdp->mynode)
1138                         need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1139                 else
1140                         raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1141                 mask = rnp->grpmask;
1142                 rnp = rnp->parent;
1143         } while (rnp != NULL);
1144 
1145         /*
1146          * We still hold the leaf rcu_node structure lock here, and
1147          * irqs are still disabled.  The reason for this subterfuge is
1148          * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1149          * held leads to deadlock.
1150          */
1151         raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1152         rnp = rdp->mynode;
1153         if (need_report & RCU_OFL_TASKS_NORM_GP)
1154                 rcu_report_unblock_qs_rnp(rnp, flags);
1155         else
1156                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1157         if (need_report & RCU_OFL_TASKS_EXP_GP)
1158                 rcu_report_exp_rnp(rsp, rnp);
1159         rcu_node_kthread_setaffinity(rnp, -1);
1160 }
1161 
1162 /*
1163  * Remove the specified CPU from the RCU hierarchy and move any pending
1164  * callbacks that it might have to the current CPU.  This code assumes
1165  * that at least one CPU in the system will remain running at all times.
1166  * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1167  */
1168 static void rcu_offline_cpu(int cpu)
1169 {
1170         __rcu_offline_cpu(cpu, &rcu_sched_state);
1171         __rcu_offline_cpu(cpu, &rcu_bh_state);
1172         rcu_preempt_offline_cpu(cpu);
1173 }
1174 
1175 #else /* #ifdef CONFIG_HOTPLUG_CPU */
1176 
1177 static void rcu_send_cbs_to_online(struct rcu_state *rsp)
1178 {
1179 }
1180 
1181 static void rcu_offline_cpu(int cpu)
1182 {
1183 }
1184 
1185 #endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1186 
1187 /*
1188  * Invoke any RCU callbacks that have made it to the end of their grace
1189  * period.  Thottle as specified by rdp->blimit.
1190  */
1191 static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
1192 {
1193         unsigned long flags;
1194         struct rcu_head *next, *list, **tail;
1195         int count;
1196 
1197         /* If no callbacks are ready, just return.*/
1198         if (!cpu_has_callbacks_ready_to_invoke(rdp))
1199                 return;
1200 
1201         /*
1202          * Extract the list of ready callbacks, disabling to prevent
1203          * races with call_rcu() from interrupt handlers.
1204          */
1205         local_irq_save(flags);
1206         list = rdp->nxtlist;
1207         rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1208         *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1209         tail = rdp->nxttail[RCU_DONE_TAIL];
1210         for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1211                 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1212                         rdp->nxttail[count] = &rdp->nxtlist;
1213         local_irq_restore(flags);
1214 
1215         /* Invoke callbacks. */
1216         count = 0;
1217         while (list) {
1218                 next = list->next;
1219                 prefetch(next);
1220                 debug_rcu_head_unqueue(list);
1221                 __rcu_reclaim(list);
1222                 list = next;
1223                 if (++count >= rdp->blimit)
1224                         break;
1225         }
1226 
1227         local_irq_save(flags);
1228 
1229         /* Update count, and requeue any remaining callbacks. */
1230         rdp->qlen -= count;
1231         rdp->n_cbs_invoked += count;
1232         if (list != NULL) {
1233                 *tail = rdp->nxtlist;
1234                 rdp->nxtlist = list;
1235                 for (count = 0; count < RCU_NEXT_SIZE; count++)
1236                         if (&rdp->nxtlist == rdp->nxttail[count])
1237                                 rdp->nxttail[count] = tail;
1238                         else
1239                                 break;
1240         }
1241 
1242         /* Reinstate batch limit if we have worked down the excess. */
1243         if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1244                 rdp->blimit = blimit;
1245 
1246         /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1247         if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1248                 rdp->qlen_last_fqs_check = 0;
1249                 rdp->n_force_qs_snap = rsp->n_force_qs;
1250         } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1251                 rdp->qlen_last_fqs_check = rdp->qlen;
1252 
1253         local_irq_restore(flags);
1254 
1255         /* Re-raise the RCU softirq if there are callbacks remaining. */
1256         if (cpu_has_callbacks_ready_to_invoke(rdp))
1257                 invoke_rcu_core();
1258 }
1259 
1260 /*
1261  * Check to see if this CPU is in a non-context-switch quiescent state
1262  * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1263  * Also schedule the RCU softirq handler.
1264  *
1265  * This function must be called with hardirqs disabled.  It is normally
1266  * invoked from the scheduling-clock interrupt.  If rcu_pending returns
1267  * false, there is no point in invoking rcu_check_callbacks().
1268  */
1269 void rcu_check_callbacks(int cpu, int user)
1270 {
1271         if (user ||
1272             (idle_cpu(cpu) && rcu_scheduler_active &&
1273              !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
1274 
1275                 /*
1276                  * Get here if this CPU took its interrupt from user
1277                  * mode or from the idle loop, and if this is not a
1278                  * nested interrupt.  In this case, the CPU is in
1279                  * a quiescent state, so note it.
1280                  *
1281                  * No memory barrier is required here because both
1282                  * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1283                  * variables that other CPUs neither access nor modify,
1284                  * at least not while the corresponding CPU is online.
1285                  */
1286 
1287                 rcu_sched_qs(cpu);
1288                 rcu_bh_qs(cpu);
1289 
1290         } else if (!in_softirq()) {
1291 
1292                 /*
1293                  * Get here if this CPU did not take its interrupt from
1294                  * softirq, in other words, if it is not interrupting
1295                  * a rcu_bh read-side critical section.  This is an _bh
1296                  * critical section, so note it.
1297                  */
1298 
1299                 rcu_bh_qs(cpu);
1300         }
1301         rcu_preempt_check_callbacks(cpu);
1302         if (rcu_pending(cpu))
1303                 invoke_rcu_core();
1304 }
1305 
1306 #ifdef CONFIG_SMP
1307 
1308 /*
1309  * Scan the leaf rcu_node structures, processing dyntick state for any that
1310  * have not yet encountered a quiescent state, using the function specified.
1311  * Also initiate boosting for any threads blocked on the root rcu_node.
1312  *
1313  * The caller must have suppressed start of new grace periods.
1314  */
1315 static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
1316 {
1317         unsigned long bit;
1318         int cpu;
1319         unsigned long flags;
1320         unsigned long mask;
1321         struct rcu_node *rnp;
1322 
1323         rcu_for_each_leaf_node(rsp, rnp) {
1324                 mask = 0;
1325                 raw_spin_lock_irqsave(&rnp->lock, flags);
1326                 if (!rcu_gp_in_progress(rsp)) {
1327                         raw_spin_unlock_irqrestore(&rnp->lock, flags);
1328                         return;
1329                 }
1330                 if (rnp->qsmask == 0) {
1331                         rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
1332                         continue;
1333                 }
1334                 cpu = rnp->grplo;
1335                 bit = 1;
1336                 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1337                         if ((rnp->qsmask & bit) != 0 &&
1338                             f(per_cpu_ptr(rsp->rda, cpu)))
1339                                 mask |= bit;
1340                 }
1341                 if (mask != 0) {
1342 
1343                         /* rcu_report_qs_rnp() releases rnp->lock. */
1344                         rcu_report_qs_rnp(mask, rsp, rnp, flags);
1345                         continue;
1346                 }
1347                 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1348         }
1349         rnp = rcu_get_root(rsp);
1350         if (rnp->qsmask == 0) {
1351                 raw_spin_lock_irqsave(&rnp->lock, flags);
1352                 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1353         }
1354 }
1355 
1356 /*
1357  * Force quiescent states on reluctant CPUs, and also detect which
1358  * CPUs are in dyntick-idle mode.
1359  */
1360 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1361 {
1362         unsigned long flags;
1363         struct rcu_node *rnp = rcu_get_root(rsp);
1364 
1365         if (!rcu_gp_in_progress(rsp))
1366                 return;  /* No grace period in progress, nothing to force. */
1367         if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
1368                 rsp->n_force_qs_lh++; /* Inexact, can lose counts.  Tough! */
1369                 return; /* Someone else is already on the job. */
1370         }
1371         if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
1372                 goto unlock_fqs_ret; /* no emergency and done recently. */
1373         rsp->n_force_qs++;
1374         raw_spin_lock(&rnp->lock);  /* irqs already disabled */
1375         rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
1376         if(!rcu_gp_in_progress(rsp)) {
1377                 rsp->n_force_qs_ngp++;
1378                 raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
1379                 goto unlock_fqs_ret;  /* no GP in progress, time updated. */
1380         }
1381         rsp->fqs_active = 1;
1382         switch (rsp->signaled) {
1383         case RCU_GP_IDLE:
1384         case RCU_GP_INIT:
1385 
1386                 break; /* grace period idle or initializing, ignore. */
1387 
1388         case RCU_SAVE_DYNTICK:
1389                 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1390                         break; /* So gcc recognizes the dead code. */
1391 
1392                 raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
1393 
1394                 /* Record dyntick-idle state. */
1395                 force_qs_rnp(rsp, dyntick_save_progress_counter);
1396                 raw_spin_lock(&rnp->lock);  /* irqs already disabled */
1397                 if (rcu_gp_in_progress(rsp))
1398                         rsp->signaled = RCU_FORCE_QS;
1399                 break;
1400 
1401         case RCU_FORCE_QS:
1402 
1403                 /* Check dyntick-idle state, send IPI to laggarts. */
1404                 raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
1405                 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
1406 
1407                 /* Leave state in case more forcing is required. */
1408 
1409                 raw_spin_lock(&rnp->lock);  /* irqs already disabled */
1410                 break;
1411         }
1412         rsp->fqs_active = 0;
1413         if (rsp->fqs_need_gp) {
1414                 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
1415                 rsp->fqs_need_gp = 0;
1416                 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1417                 return;
1418         }
1419         raw_spin_unlock(&rnp->lock);  /* irqs remain disabled */
1420 unlock_fqs_ret:
1421         raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
1422 }
1423 
1424 #else /* #ifdef CONFIG_SMP */
1425 
1426 static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1427 {
1428         set_need_resched();
1429 }
1430 
1431 #endif /* #else #ifdef CONFIG_SMP */
1432 
1433 /*
1434  * This does the RCU processing work from softirq context for the
1435  * specified rcu_state and rcu_data structures.  This may be called
1436  * only from the CPU to whom the rdp belongs.
1437  */
1438 static void
1439 __rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1440 {
1441         unsigned long flags;
1442 
1443         WARN_ON_ONCE(rdp->beenonline == 0);
1444 
1445         /*
1446          * If an RCU GP has gone long enough, go check for dyntick
1447          * idle CPUs and, if needed, send resched IPIs.
1448          */
1449         if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1450                 force_quiescent_state(rsp, 1);
1451 
1452         /*
1453          * Advance callbacks in response to end of earlier grace
1454          * period that some other CPU ended.
1455          */
1456         rcu_process_gp_end(rsp, rdp);
1457 
1458         /* Update RCU state based on any recent quiescent states. */
1459         rcu_check_quiescent_state(rsp, rdp);
1460 
1461         /* Does this CPU require a not-yet-started grace period? */
1462         if (cpu_needs_another_gp(rsp, rdp)) {
1463                 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
1464                 rcu_start_gp(rsp, flags);  /* releases above lock */
1465         }
1466 
1467         /* If there are callbacks ready, invoke them. */
1468         if (cpu_has_callbacks_ready_to_invoke(rdp))
1469                 invoke_rcu_callbacks(rsp, rdp);
1470 }
1471 
1472 /*
1473  * Do softirq processing for the current CPU.
1474  */
1475 static void rcu_process_callbacks(struct softirq_action *unused)
1476 {
1477         __rcu_process_callbacks(&rcu_sched_state,
1478                                 &__get_cpu_var(rcu_sched_data));
1479         __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
1480         rcu_preempt_process_callbacks();
1481 
1482         /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1483         rcu_needs_cpu_flush();
1484 }
1485 
1486 /*
1487  * Wake up the current CPU's kthread.  This replaces raise_softirq()
1488  * in earlier versions of RCU.  Note that because we are running on
1489  * the current CPU with interrupts disabled, the rcu_cpu_kthread_task
1490  * cannot disappear out from under us.
1491  */
1492 static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1493 {
1494         if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1495                 return;
1496         if (likely(!rsp->boost)) {
1497                 rcu_do_batch(rsp, rdp);
1498                 return;
1499         }
1500         invoke_rcu_callbacks_kthread();
1501 }
1502 
1503 static void invoke_rcu_core(void)
1504 {
1505         raise_softirq(RCU_SOFTIRQ);
1506 }
1507 
1508 static void
1509 __call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1510            struct rcu_state *rsp)
1511 {
1512         unsigned long flags;
1513         struct rcu_data *rdp;
1514 
1515         debug_rcu_head_queue(head);
1516         head->func = func;
1517         head->next = NULL;
1518 
1519         smp_mb(); /* Ensure RCU update seen before callback registry. */
1520 
1521         /*
1522          * Opportunistically note grace-period endings and beginnings.
1523          * Note that we might see a beginning right after we see an
1524          * end, but never vice versa, since this CPU has to pass through
1525          * a quiescent state betweentimes.
1526          */
1527         local_irq_save(flags);
1528         rdp = this_cpu_ptr(rsp->rda);
1529 
1530         /* Add the callback to our list. */
1531         *rdp->nxttail[RCU_NEXT_TAIL] = head;
1532         rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1533         rdp->qlen++;
1534 
1535         /* If interrupts were disabled, don't dive into RCU core. */
1536         if (irqs_disabled_flags(flags)) {
1537                 local_irq_restore(flags);
1538                 return;
1539         }
1540 
1541         /*
1542          * Force the grace period if too many callbacks or too long waiting.
1543          * Enforce hysteresis, and don't invoke force_quiescent_state()
1544          * if some other CPU has recently done so.  Also, don't bother
1545          * invoking force_quiescent_state() if the newly enqueued callback
1546          * is the only one waiting for a grace period to complete.
1547          */
1548         if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
1549 
1550                 /* Are we ignoring a completed grace period? */
1551                 rcu_process_gp_end(rsp, rdp);
1552                 check_for_new_grace_period(rsp, rdp);
1553 
1554                 /* Start a new grace period if one not already started. */
1555                 if (!rcu_gp_in_progress(rsp)) {
1556                         unsigned long nestflag;
1557                         struct rcu_node *rnp_root = rcu_get_root(rsp);
1558 
1559                         raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1560                         rcu_start_gp(rsp, nestflag);  /* rlses rnp_root->lock */
1561                 } else {
1562                         /* Give the grace period a kick. */
1563                         rdp->blimit = LONG_MAX;
1564                         if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1565                             *rdp->nxttail[RCU_DONE_TAIL] != head)
1566                                 force_quiescent_state(rsp, 0);
1567                         rdp->n_force_qs_snap = rsp->n_force_qs;
1568                         rdp->qlen_last_fqs_check = rdp->qlen;
1569                 }
1570         } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
1571                 force_quiescent_state(rsp, 1);
1572         local_irq_restore(flags);
1573 }
1574 
1575 /*
1576  * Queue an RCU-sched callback for invocation after a grace period.
1577  */
1578 void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1579 {
1580         __call_rcu(head, func, &rcu_sched_state);
1581 }
1582 EXPORT_SYMBOL_GPL(call_rcu_sched);
1583 
1584 /*
1585  * Queue an RCU for invocation after a quicker grace period.
1586  */
1587 void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1588 {
1589         __call_rcu(head, func, &rcu_bh_state);
1590 }
1591 EXPORT_SYMBOL_GPL(call_rcu_bh);
1592 
1593 /**
1594  * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1595  *
1596  * Control will return to the caller some time after a full rcu-sched
1597  * grace period has elapsed, in other words after all currently executing
1598  * rcu-sched read-side critical sections have completed.   These read-side
1599  * critical sections are delimited by rcu_read_lock_sched() and
1600  * rcu_read_unlock_sched(), and may be nested.  Note that preempt_disable(),
1601  * local_irq_disable(), and so on may be used in place of
1602  * rcu_read_lock_sched().
1603  *
1604  * This means that all preempt_disable code sequences, including NMI and
1605  * hardware-interrupt handlers, in progress on entry will have completed
1606  * before this primitive returns.  However, this does not guarantee that
1607  * softirq handlers will have completed, since in some kernels, these
1608  * handlers can run in process context, and can block.
1609  *
1610  * This primitive provides the guarantees made by the (now removed)
1611  * synchronize_kernel() API.  In contrast, synchronize_rcu() only
1612  * guarantees that rcu_read_lock() sections will have completed.
1613  * In "classic RCU", these two guarantees happen to be one and
1614  * the same, but can differ in realtime RCU implementations.
1615  */
1616 void synchronize_sched(void)
1617 {
1618         struct rcu_synchronize rcu;
1619 
1620         if (rcu_blocking_is_gp())
1621                 return;
1622 
1623         init_rcu_head_on_stack(&rcu.head);
1624         init_completion(&rcu.completion);
1625         /* Will wake me after RCU finished. */
1626         call_rcu_sched(&rcu.head, wakeme_after_rcu);
1627         /* Wait for it. */
1628         wait_for_completion(&rcu.completion);
1629         destroy_rcu_head_on_stack(&rcu.head);
1630 }
1631 EXPORT_SYMBOL_GPL(synchronize_sched);
1632 
1633 /**
1634  * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1635  *
1636  * Control will return to the caller some time after a full rcu_bh grace
1637  * period has elapsed, in other words after all currently executing rcu_bh
1638  * read-side critical sections have completed.  RCU read-side critical
1639  * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1640  * and may be nested.
1641  */
1642 void synchronize_rcu_bh(void)
1643 {
1644         struct rcu_synchronize rcu;
1645 
1646         if (rcu_blocking_is_gp())
1647                 return;
1648 
1649         init_rcu_head_on_stack(&rcu.head);
1650         init_completion(&rcu.completion);
1651         /* Will wake me after RCU finished. */
1652         call_rcu_bh(&rcu.head, wakeme_after_rcu);
1653         /* Wait for it. */
1654         wait_for_completion(&rcu.completion);
1655         destroy_rcu_head_on_stack(&rcu.head);
1656 }
1657 EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1658 
1659 /*
1660  * Check to see if there is any immediate RCU-related work to be done
1661  * by the current CPU, for the specified type of RCU, returning 1 if so.
1662  * The checks are in order of increasing expense: checks that can be
1663  * carried out against CPU-local state are performed first.  However,
1664  * we must check for CPU stalls first, else we might not get a chance.
1665  */
1666 static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1667 {
1668         struct rcu_node *rnp = rdp->mynode;
1669 
1670         rdp->n_rcu_pending++;
1671 
1672         /* Check for CPU stalls, if enabled. */
1673         check_cpu_stall(rsp, rdp);
1674 
1675         /* Is the RCU core waiting for a quiescent state from this CPU? */
1676         if (rdp->qs_pending && !rdp->passed_quiesc) {
1677 
1678                 /*
1679                  * If force_quiescent_state() coming soon and this CPU
1680                  * needs a quiescent state, and this is either RCU-sched
1681                  * or RCU-bh, force a local reschedule.
1682                  */
1683                 rdp->n_rp_qs_pending++;
1684                 if (!rdp->preemptible &&
1685                     ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1686                                  jiffies))
1687                         set_need_resched();
1688         } else if (rdp->qs_pending && rdp->passed_quiesc) {
1689                 rdp->n_rp_report_qs++;
1690                 return 1;
1691         }
1692 
1693         /* Does this CPU have callbacks ready to invoke? */
1694         if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1695                 rdp->n_rp_cb_ready++;
1696                 return 1;
1697         }
1698 
1699         /* Has RCU gone idle with this CPU needing another grace period? */
1700         if (cpu_needs_another_gp(rsp, rdp)) {
1701                 rdp->n_rp_cpu_needs_gp++;
1702                 return 1;
1703         }
1704 
1705         /* Has another RCU grace period completed?  */
1706         if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
1707                 rdp->n_rp_gp_completed++;
1708                 return 1;
1709         }
1710 
1711         /* Has a new RCU grace period started? */
1712         if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
1713                 rdp->n_rp_gp_started++;
1714                 return 1;
1715         }
1716 
1717         /* Has an RCU GP gone long enough to send resched IPIs &c? */
1718         if (rcu_gp_in_progress(rsp) &&
1719             ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
1720                 rdp->n_rp_need_fqs++;
1721                 return 1;
1722         }
1723 
1724         /* nothing to do */
1725         rdp->n_rp_need_nothing++;
1726         return 0;
1727 }
1728 
1729 /*
1730  * Check to see if there is any immediate RCU-related work to be done
1731  * by the current CPU, returning 1 if so.  This function is part of the
1732  * RCU implementation; it is -not- an exported member of the RCU API.
1733  */
1734 static int rcu_pending(int cpu)
1735 {
1736         return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
1737                __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1738                rcu_preempt_pending(cpu);
1739 }
1740 
1741 /*
1742  * Check to see if any future RCU-related work will need to be done
1743  * by the current CPU, even if none need be done immediately, returning
1744  * 1 if so.
1745  */
1746 static int rcu_needs_cpu_quick_check(int cpu)
1747 {
1748         /* RCU callbacks either ready or pending? */
1749         return per_cpu(rcu_sched_data, cpu).nxtlist ||
1750                per_cpu(rcu_bh_data, cpu).nxtlist ||
1751                rcu_preempt_needs_cpu(cpu);
1752 }
1753 
1754 static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1755 static atomic_t rcu_barrier_cpu_count;
1756 static DEFINE_MUTEX(rcu_barrier_mutex);
1757 static struct completion rcu_barrier_completion;
1758 
1759 static void rcu_barrier_callback(struct rcu_head *notused)
1760 {
1761         if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1762                 complete(&rcu_barrier_completion);
1763 }
1764 
1765 /*
1766  * Called with preemption disabled, and from cross-cpu IRQ context.
1767  */
1768 static void rcu_barrier_func(void *type)
1769 {
1770         int cpu = smp_processor_id();
1771         struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1772         void (*call_rcu_func)(struct rcu_head *head,
1773                               void (*func)(struct rcu_head *head));
1774 
1775         atomic_inc(&rcu_barrier_cpu_count);
1776         call_rcu_func = type;
1777         call_rcu_func(head, rcu_barrier_callback);
1778 }
1779 
1780 /*
1781  * Orchestrate the specified type of RCU barrier, waiting for all
1782  * RCU callbacks of the specified type to complete.
1783  */
1784 static void _rcu_barrier(struct rcu_state *rsp,
1785                          void (*call_rcu_func)(struct rcu_head *head,
1786                                                void (*func)(struct rcu_head *head)))
1787 {
1788         BUG_ON(in_interrupt());
1789         /* Take mutex to serialize concurrent rcu_barrier() requests. */
1790         mutex_lock(&rcu_barrier_mutex);
1791         init_completion(&rcu_barrier_completion);
1792         /*
1793          * Initialize rcu_barrier_cpu_count to 1, then invoke
1794          * rcu_barrier_func() on each CPU, so that each CPU also has
1795          * incremented rcu_barrier_cpu_count.  Only then is it safe to
1796          * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1797          * might complete its grace period before all of the other CPUs
1798          * did their increment, causing this function to return too
1799          * early.  Note that on_each_cpu() disables irqs, which prevents
1800          * any CPUs from coming online or going offline until each online
1801          * CPU has queued its RCU-barrier callback.
1802          */
1803         atomic_set(&rcu_barrier_cpu_count, 1);
1804         on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
1805         if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1806                 complete(&rcu_barrier_completion);
1807         wait_for_completion(&rcu_barrier_completion);
1808         mutex_unlock(&rcu_barrier_mutex);
1809 }
1810 
1811 /**
1812  * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1813  */
1814 void rcu_barrier_bh(void)
1815 {
1816         _rcu_barrier(&rcu_bh_state, call_rcu_bh);
1817 }
1818 EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1819 
1820 /**
1821  * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1822  */
1823 void rcu_barrier_sched(void)
1824 {
1825         _rcu_barrier(&rcu_sched_state, call_rcu_sched);
1826 }
1827 EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1828 
1829 /*
1830  * Do boot-time initialization of a CPU's per-CPU RCU data.
1831  */
1832 static void __init
1833 rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
1834 {
1835         unsigned long flags;
1836         int i;
1837         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1838         struct rcu_node *rnp = rcu_get_root(rsp);
1839 
1840         /* Set up local state, ensuring consistent view of global state. */
1841         raw_spin_lock_irqsave(&rnp->lock, flags);
1842         rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1843         rdp->nxtlist = NULL;
1844         for (i = 0; i < RCU_NEXT_SIZE; i++)
1845                 rdp->nxttail[i] = &rdp->nxtlist;
1846         rdp->qlen = 0;
1847 #ifdef CONFIG_NO_HZ
1848         rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1849 #endif /* #ifdef CONFIG_NO_HZ */
1850         rdp->cpu = cpu;
1851         raw_spin_unlock_irqrestore(&rnp->lock, flags);
1852 }
1853 
1854 /*
1855  * Initialize a CPU's per-CPU RCU data.  Note that only one online or
1856  * offline event can be happening at a given time.  Note also that we
1857  * can accept some slop in the rsp->completed access due to the fact
1858  * that this CPU cannot possibly have any RCU callbacks in flight yet.
1859  */
1860 static void __cpuinit
1861 rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
1862 {
1863         unsigned long flags;
1864         unsigned long mask;
1865         struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1866         struct rcu_node *rnp = rcu_get_root(rsp);
1867 
1868         /* Set up local state, ensuring consistent view of global state. */
1869         raw_spin_lock_irqsave(&rnp->lock, flags);
1870         rdp->passed_quiesc = 0;  /* We could be racing with new GP, */
1871         rdp->qs_pending = 1;     /*  so set up to respond to current GP. */
1872         rdp->beenonline = 1;     /* We have now been online. */
1873         rdp->preemptible = preemptible;
1874         rdp->qlen_last_fqs_check = 0;
1875         rdp->n_force_qs_snap = rsp->n_force_qs;
1876         rdp->blimit = blimit;
1877         raw_spin_unlock(&rnp->lock);            /* irqs remain disabled. */
1878 
1879         /*
1880          * A new grace period might start here.  If so, we won't be part
1881          * of it, but that is OK, as we are currently in a quiescent state.
1882          */
1883 
1884         /* Exclude any attempts to start a new GP on large systems. */
1885         raw_spin_lock(&rsp->onofflock);         /* irqs already disabled. */
1886 
1887         /* Add CPU to rcu_node bitmasks. */
1888         rnp = rdp->mynode;
1889         mask = rdp->grpmask;
1890         do {
1891                 /* Exclude any attempts to start a new GP on small systems. */
1892                 raw_spin_lock(&rnp->lock);      /* irqs already disabled. */
1893                 rnp->qsmaskinit |= mask;
1894                 mask = rnp->grpmask;
1895                 if (rnp == rdp->mynode) {
1896                         rdp->gpnum = rnp->completed; /* if GP in progress... */
1897                         rdp->completed = rnp->completed;
1898                         rdp->passed_quiesc_completed = rnp->completed - 1;
1899                 }
1900                 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
1901                 rnp = rnp->parent;
1902         } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1903 
1904         raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
1905 }
1906 
1907 static void __cpuinit rcu_prepare_cpu(int cpu)
1908 {
1909         rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1910         rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1911         rcu_preempt_init_percpu_data(cpu);
1912 }
1913 
1914 /*
1915  * Handle CPU online/offline notification events.
1916  */
1917 static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1918                                     unsigned long action, void *hcpu)
1919 {
1920         long cpu = (long)hcpu;
1921         struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1922         struct rcu_node *rnp = rdp->mynode;
1923 
1924         switch (action) {
1925         case CPU_UP_PREPARE:
1926         case CPU_UP_PREPARE_FROZEN:
1927                 rcu_prepare_cpu(cpu);
1928                 rcu_prepare_kthreads(cpu);
1929                 break;
1930         case CPU_ONLINE:
1931         case CPU_DOWN_FAILED:
1932                 rcu_node_kthread_setaffinity(rnp, -1);
1933                 rcu_cpu_kthread_setrt(cpu, 1);
1934                 break;
1935         case CPU_DOWN_PREPARE:
1936                 rcu_node_kthread_setaffinity(rnp, cpu);
1937                 rcu_cpu_kthread_setrt(cpu, 0);
1938                 break;
1939         case CPU_DYING:
1940         case CPU_DYING_FROZEN:
1941                 /*
1942                  * The whole machine is "stopped" except this CPU, so we can
1943                  * touch any data without introducing corruption. We send the
1944                  * dying CPU's callbacks to an arbitrarily chosen online CPU.
1945                  */
1946                 rcu_send_cbs_to_online(&rcu_bh_state);
1947                 rcu_send_cbs_to_online(&rcu_sched_state);
1948                 rcu_preempt_send_cbs_to_online();
1949                 break;
1950         case CPU_DEAD:
1951         case CPU_DEAD_FROZEN:
1952         case CPU_UP_CANCELED:
1953         case CPU_UP_CANCELED_FROZEN:
1954                 rcu_offline_cpu(cpu);
1955                 break;
1956         default:
1957                 break;
1958         }
1959         return NOTIFY_OK;
1960 }
1961 
1962 /*
1963  * This function is invoked towards the end of the scheduler's initialization
1964  * process.  Before this is called, the idle task might contain
1965  * RCU read-side critical sections (during which time, this idle
1966  * task is booting the system).  After this function is called, the
1967  * idle tasks are prohibited from containing RCU read-side critical
1968  * sections.  This function also enables RCU lockdep checking.
1969  */
1970 void rcu_scheduler_starting(void)
1971 {
1972         WARN_ON(num_online_cpus() != 1);
1973         WARN_ON(nr_context_switches() > 0);
1974         rcu_scheduler_active = 1;
1975 }
1976 
1977 /*
1978  * Compute the per-level fanout, either using the exact fanout specified
1979  * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1980  */
1981 #ifdef CONFIG_RCU_FANOUT_EXACT
1982 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1983 {
1984         int i;
1985 
1986         for (i = NUM_RCU_LVLS - 1; i > 0; i--)
1987                 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1988         rsp->levelspread[0] = RCU_FANOUT_LEAF;
1989 }
1990 #else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1991 static void __init rcu_init_levelspread(struct rcu_state *rsp)
1992 {
1993         int ccur;
1994         int cprv;
1995         int i;
1996 
1997         cprv = NR_CPUS;
1998         for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1999                 ccur = rsp->levelcnt[i];
2000                 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2001                 cprv = ccur;
2002         }
2003 }
2004 #endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2005 
2006 /*
2007  * Helper function for rcu_init() that initializes one rcu_state structure.
2008  */
2009 static void __init rcu_init_one(struct rcu_state *rsp,
2010                 struct rcu_data __percpu *rda)
2011 {
2012         static char *buf[] = { "rcu_node_level_0",
2013                                "rcu_node_level_1",
2014                                "rcu_node_level_2",
2015                                "rcu_node_level_3" };  /* Match MAX_RCU_LVLS */
2016         int cpustride = 1;
2017         int i;
2018         int j;
2019         struct rcu_node *rnp;
2020 
2021         BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf));  /* Fix buf[] init! */
2022 
2023         /* Initialize the level-tracking arrays. */
2024 
2025         for (i = 1; i < NUM_RCU_LVLS; i++)
2026                 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2027         rcu_init_levelspread(rsp);
2028 
2029         /* Initialize the elements themselves, starting from the leaves. */
2030 
2031         for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2032                 cpustride *= rsp->levelspread[i];
2033                 rnp = rsp->level[i];
2034                 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
2035                         raw_spin_lock_init(&rnp->lock);
2036                         lockdep_set_class_and_name(&rnp->lock,
2037                                                    &rcu_node_class[i], buf[i]);
2038                         rnp->gpnum = 0;
2039                         rnp->qsmask = 0;
2040                         rnp->qsmaskinit = 0;
2041                         rnp->grplo = j * cpustride;
2042                         rnp->grphi = (j + 1) * cpustride - 1;
2043                         if (rnp->grphi >= NR_CPUS)
2044                                 rnp->grphi = NR_CPUS - 1;
2045                         if (i == 0) {
2046                                 rnp->grpnum = 0;
2047                                 rnp->grpmask = 0;
2048                                 rnp->parent = NULL;
2049                         } else {
2050                                 rnp->grpnum = j % rsp->levelspread[i - 1];
2051                                 rnp->grpmask = 1UL << rnp->grpnum;
2052                                 rnp->parent = rsp->level[i - 1] +
2053                                               j / rsp->levelspread[i - 1];
2054                         }
2055                         rnp->level = i;
2056                         INIT_LIST_HEAD(&rnp->blkd_tasks);
2057                 }
2058         }
2059 
2060         rsp->rda = rda;
2061         rnp = rsp->level[NUM_RCU_LVLS - 1];
2062         for_each_possible_cpu(i) {
2063                 while (i > rnp->grphi)
2064                         rnp++;
2065                 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
2066                 rcu_boot_init_percpu_data(i, rsp);
2067         }
2068 }
2069 
2070 void __init rcu_init(void)
2071 {
2072         int cpu;
2073 
2074         rcu_bootup_announce();
2075         rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2076         rcu_init_one(&rcu_bh_state, &rcu_bh_data);
2077         __rcu_init_preempt();
2078          open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
2079 
2080         /*
2081          * We don't need protection against CPU-hotplug here because
2082          * this is called early in boot, before either interrupts
2083          * or the scheduler are operational.
2084          */
2085         cpu_notifier(rcu_cpu_notify, 0);
2086         for_each_online_cpu(cpu)
2087                 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
2088         check_cpu_stall_init();
2089 }
2090 
2091 #include "rcutree_plugin.h"
2092 

~ [ 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