blob: d4437345706f8a539eee6dba201662fd791978ef [file] [log] [blame]
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001/*
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 -
Paul E. McKenneya71fca52009-09-18 10:28:19 -070028 * Documentation/RCU
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010029 */
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>
Ingo Molnarc1dc0b92009-08-02 11:28:21 +020038#include <linux/nmi.h>
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010039#include <asm/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>
Paul E. McKenneybbad9372010-04-02 16:17:17 -070049#include <linux/kernel_stat.h>
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010050
Paul E. McKenney9f77da92009-08-22 13:56:45 -070051#include "rcutree.h"
52
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010053/* Data structures. */
54
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -080055static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
Peter Zijlstra88b91c72009-10-26 10:24:31 -070056
Paul E. McKenney4300aa62010-04-13 16:18:22 -070057#define RCU_STATE_INITIALIZER(structname) { \
58 .level = { &structname.node[0] }, \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010059 .levelcnt = { \
60 NUM_RCU_LVL_0, /* root of hierarchy. */ \
61 NUM_RCU_LVL_1, \
62 NUM_RCU_LVL_2, \
Paul E. McKenneycf244dc2009-12-02 12:10:14 -080063 NUM_RCU_LVL_3, \
64 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010065 }, \
Paul E. McKenney83f5b012009-10-28 08:14:49 -070066 .signaled = RCU_GP_IDLE, \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010067 .gpnum = -300, \
68 .completed = -300, \
Paul E. McKenney4300aa62010-04-13 16:18:22 -070069 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname.onofflock), \
Paul E. McKenneye74f4c42009-10-06 21:48:17 -070070 .orphan_cbs_list = NULL, \
Paul E. McKenney4300aa62010-04-13 16:18:22 -070071 .orphan_cbs_tail = &structname.orphan_cbs_list, \
Paul E. McKenneye74f4c42009-10-06 21:48:17 -070072 .orphan_qlen = 0, \
Paul E. McKenney4300aa62010-04-13 16:18:22 -070073 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname.fqslock), \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010074 .n_force_qs = 0, \
75 .n_force_qs_ngp = 0, \
Paul E. McKenney4300aa62010-04-13 16:18:22 -070076 .name = #structname, \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010077}
78
Paul E. McKenneyd6714c22009-08-22 13:56:46 -070079struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched_state);
80DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010081
Ingo Molnar6258c4f2009-03-25 16:42:24 +010082struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh_state);
83DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
Ingo Molnarb1f77b02009-03-13 03:20:49 +010084
Paul E. McKenneybbad9372010-04-02 16:17:17 -070085int rcu_scheduler_active __read_mostly;
86EXPORT_SYMBOL_GPL(rcu_scheduler_active);
87
Ingo Molnarb1f77b02009-03-13 03:20:49 +010088/*
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -070089 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
90 * permit this function to be invoked without holding the root rcu_node
91 * structure's ->lock, but of course results can be subject to change.
92 */
93static int rcu_gp_in_progress(struct rcu_state *rsp)
94{
95 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
96}
97
98/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -070099 * Note a quiescent state. Because we do not need to know
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100100 * how many quiescent states passed, just if there was at least
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700101 * one since the start of the grace period, this just sets a flag.
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100102 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700103void rcu_sched_qs(int cpu)
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100104{
Paul E. McKenney25502a62010-04-01 17:37:01 -0700105 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700106
Paul E. McKenneyc64ac3c2009-11-10 13:37:22 -0800107 rdp->passed_quiesc_completed = rdp->gpnum - 1;
Paul E. McKenneyc3422be2009-09-13 09:15:10 -0700108 barrier();
109 rdp->passed_quiesc = 1;
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100110}
111
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700112void rcu_bh_qs(int cpu)
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100113{
Paul E. McKenney25502a62010-04-01 17:37:01 -0700114 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700115
Paul E. McKenneyc64ac3c2009-11-10 13:37:22 -0800116 rdp->passed_quiesc_completed = rdp->gpnum - 1;
Paul E. McKenneyc3422be2009-09-13 09:15:10 -0700117 barrier();
118 rdp->passed_quiesc = 1;
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100119}
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100120
Paul E. McKenney25502a62010-04-01 17:37:01 -0700121/*
122 * Note a context switch. This is a quiescent state for RCU-sched,
123 * and requires special handling for preemptible RCU.
124 */
125void rcu_note_context_switch(int cpu)
126{
127 rcu_sched_qs(cpu);
128 rcu_preempt_note_context_switch(cpu);
129}
130
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100131#ifdef CONFIG_NO_HZ
Paul E. McKenney90a4d2c2009-01-04 11:41:11 -0800132DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
133 .dynticks_nesting = 1,
134 .dynticks = 1,
135};
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100136#endif /* #ifdef CONFIG_NO_HZ */
137
138static int blimit = 10; /* Maximum callbacks per softirq. */
139static int qhimark = 10000; /* If this many pending, ignore blimit. */
140static int qlowmark = 100; /* Once only this many pending, use blimit. */
141
Paul E. McKenney3d76c082009-09-28 07:46:32 -0700142module_param(blimit, int, 0);
143module_param(qhimark, int, 0);
144module_param(qlowmark, int, 0);
145
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100146static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
Paul E. McKenneya1572292009-08-22 13:56:51 -0700147static int rcu_pending(int cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100148
149/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700150 * Return the number of RCU-sched batches processed thus far for debug & stats.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100151 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700152long rcu_batches_completed_sched(void)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100153{
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700154 return rcu_sched_state.completed;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100155}
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700156EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100157
158/*
159 * Return the number of RCU BH batches processed thus far for debug & stats.
160 */
161long rcu_batches_completed_bh(void)
162{
163 return rcu_bh_state.completed;
164}
165EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
166
167/*
Paul E. McKenneybf66f182010-01-04 15:09:10 -0800168 * Force a quiescent state for RCU BH.
169 */
170void rcu_bh_force_quiescent_state(void)
171{
172 force_quiescent_state(&rcu_bh_state, 0);
173}
174EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
175
176/*
177 * Force a quiescent state for RCU-sched.
178 */
179void rcu_sched_force_quiescent_state(void)
180{
181 force_quiescent_state(&rcu_sched_state, 0);
182}
183EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
184
185/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100186 * Does the CPU have callbacks ready to be invoked?
187 */
188static int
189cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
190{
191 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
192}
193
194/*
195 * Does the current CPU require a yet-as-unscheduled grace period?
196 */
197static int
198cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
199{
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -0700200 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100201}
202
203/*
204 * Return the root node of the specified rcu_state structure.
205 */
206static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
207{
208 return &rsp->node[0];
209}
210
211#ifdef CONFIG_SMP
212
213/*
214 * If the specified CPU is offline, tell the caller that it is in
215 * a quiescent state. Otherwise, whack it with a reschedule IPI.
216 * Grace periods can end up waiting on an offline CPU when that
217 * CPU is in the process of coming online -- it will be added to the
218 * rcu_node bitmasks before it actually makes it online. The same thing
219 * can happen while a CPU is in the process of coming online. Because this
220 * race is quite rare, we check for it after detecting that the grace
221 * period has been delayed rather than checking each and every CPU
222 * each and every time we start a new grace period.
223 */
224static int rcu_implicit_offline_qs(struct rcu_data *rdp)
225{
226 /*
227 * If the CPU is offline, it is in a quiescent state. We can
228 * trust its state not to change because interrupts are disabled.
229 */
230 if (cpu_is_offline(rdp->cpu)) {
231 rdp->offline_fqs++;
232 return 1;
233 }
234
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700235 /* If preemptable RCU, no point in sending reschedule IPI. */
236 if (rdp->preemptable)
237 return 0;
238
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100239 /* The CPU is online, so send it a reschedule IPI. */
240 if (rdp->cpu != smp_processor_id())
241 smp_send_reschedule(rdp->cpu);
242 else
243 set_need_resched();
244 rdp->resched_ipi++;
245 return 0;
246}
247
248#endif /* #ifdef CONFIG_SMP */
249
250#ifdef CONFIG_NO_HZ
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100251
252/**
253 * rcu_enter_nohz - inform RCU that current CPU is entering nohz
254 *
255 * Enter nohz mode, in other words, -leave- the mode in which RCU
256 * read-side critical sections can occur. (Though RCU read-side
257 * critical sections can occur in irq handlers in nohz mode, a possibility
258 * handled by rcu_irq_enter() and rcu_irq_exit()).
259 */
260void rcu_enter_nohz(void)
261{
262 unsigned long flags;
263 struct rcu_dynticks *rdtp;
264
265 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
266 local_irq_save(flags);
267 rdtp = &__get_cpu_var(rcu_dynticks);
268 rdtp->dynticks++;
269 rdtp->dynticks_nesting--;
Paul E. McKenney86848962009-08-27 15:00:12 -0700270 WARN_ON_ONCE(rdtp->dynticks & 0x1);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100271 local_irq_restore(flags);
272}
273
274/*
275 * rcu_exit_nohz - inform RCU that current CPU is leaving nohz
276 *
277 * Exit nohz mode, in other words, -enter- the mode in which RCU
278 * read-side critical sections normally occur.
279 */
280void rcu_exit_nohz(void)
281{
282 unsigned long flags;
283 struct rcu_dynticks *rdtp;
284
285 local_irq_save(flags);
286 rdtp = &__get_cpu_var(rcu_dynticks);
287 rdtp->dynticks++;
288 rdtp->dynticks_nesting++;
Paul E. McKenney86848962009-08-27 15:00:12 -0700289 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100290 local_irq_restore(flags);
291 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
292}
293
294/**
295 * rcu_nmi_enter - inform RCU of entry to NMI context
296 *
297 * If the CPU was idle with dynamic ticks active, and there is no
298 * irq handler running, this updates rdtp->dynticks_nmi to let the
299 * RCU grace-period handling know that the CPU is active.
300 */
301void rcu_nmi_enter(void)
302{
303 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
304
305 if (rdtp->dynticks & 0x1)
306 return;
307 rdtp->dynticks_nmi++;
Paul E. McKenney86848962009-08-27 15:00:12 -0700308 WARN_ON_ONCE(!(rdtp->dynticks_nmi & 0x1));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100309 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
310}
311
312/**
313 * rcu_nmi_exit - inform RCU of exit from NMI context
314 *
315 * If the CPU was idle with dynamic ticks active, and there is no
316 * irq handler running, this updates rdtp->dynticks_nmi to let the
317 * RCU grace-period handling know that the CPU is no longer active.
318 */
319void rcu_nmi_exit(void)
320{
321 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
322
323 if (rdtp->dynticks & 0x1)
324 return;
325 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
326 rdtp->dynticks_nmi++;
Paul E. McKenney86848962009-08-27 15:00:12 -0700327 WARN_ON_ONCE(rdtp->dynticks_nmi & 0x1);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100328}
329
330/**
331 * rcu_irq_enter - inform RCU of entry to hard irq context
332 *
333 * If the CPU was idle with dynamic ticks active, this updates the
334 * rdtp->dynticks to let the RCU handling know that the CPU is active.
335 */
336void rcu_irq_enter(void)
337{
338 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
339
340 if (rdtp->dynticks_nesting++)
341 return;
342 rdtp->dynticks++;
Paul E. McKenney86848962009-08-27 15:00:12 -0700343 WARN_ON_ONCE(!(rdtp->dynticks & 0x1));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100344 smp_mb(); /* CPUs seeing ++ must see later RCU read-side crit sects */
345}
346
347/**
348 * rcu_irq_exit - inform RCU of exit from hard irq context
349 *
350 * If the CPU was idle with dynamic ticks active, update the rdp->dynticks
351 * to put let the RCU handling be aware that the CPU is going back to idle
352 * with no ticks.
353 */
354void rcu_irq_exit(void)
355{
356 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
357
358 if (--rdtp->dynticks_nesting)
359 return;
360 smp_mb(); /* CPUs seeing ++ must see prior RCU read-side crit sects */
361 rdtp->dynticks++;
Paul E. McKenney86848962009-08-27 15:00:12 -0700362 WARN_ON_ONCE(rdtp->dynticks & 0x1);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100363
364 /* If the interrupt queued a callback, get out of dyntick mode. */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700365 if (__get_cpu_var(rcu_sched_data).nxtlist ||
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100366 __get_cpu_var(rcu_bh_data).nxtlist)
367 set_need_resched();
368}
369
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100370#ifdef CONFIG_SMP
371
372/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100373 * Snapshot the specified CPU's dynticks counter so that we can later
374 * credit them with an implicit quiescent state. Return 1 if this CPU
Paul E. McKenney1eba8f82009-09-23 09:50:42 -0700375 * is in dynticks idle mode, which is an extended quiescent state.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100376 */
377static int dyntick_save_progress_counter(struct rcu_data *rdp)
378{
379 int ret;
380 int snap;
381 int snap_nmi;
382
383 snap = rdp->dynticks->dynticks;
384 snap_nmi = rdp->dynticks->dynticks_nmi;
385 smp_mb(); /* Order sampling of snap with end of grace period. */
386 rdp->dynticks_snap = snap;
387 rdp->dynticks_nmi_snap = snap_nmi;
388 ret = ((snap & 0x1) == 0) && ((snap_nmi & 0x1) == 0);
389 if (ret)
390 rdp->dynticks_fqs++;
391 return ret;
392}
393
394/*
395 * Return true if the specified CPU has passed through a quiescent
396 * state by virtue of being in or having passed through an dynticks
397 * idle state since the last call to dyntick_save_progress_counter()
398 * for this same CPU.
399 */
400static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
401{
402 long curr;
403 long curr_nmi;
404 long snap;
405 long snap_nmi;
406
407 curr = rdp->dynticks->dynticks;
408 snap = rdp->dynticks_snap;
409 curr_nmi = rdp->dynticks->dynticks_nmi;
410 snap_nmi = rdp->dynticks_nmi_snap;
411 smp_mb(); /* force ordering with cpu entering/leaving dynticks. */
412
413 /*
414 * If the CPU passed through or entered a dynticks idle phase with
415 * no active irq/NMI handlers, then we can safely pretend that the CPU
416 * already acknowledged the request to pass through a quiescent
417 * state. Either way, that CPU cannot possibly be in an RCU
418 * read-side critical section that started before the beginning
419 * of the current RCU grace period.
420 */
421 if ((curr != snap || (curr & 0x1) == 0) &&
422 (curr_nmi != snap_nmi || (curr_nmi & 0x1) == 0)) {
423 rdp->dynticks_fqs++;
424 return 1;
425 }
426
427 /* Go check for the CPU being offline. */
428 return rcu_implicit_offline_qs(rdp);
429}
430
431#endif /* #ifdef CONFIG_SMP */
432
433#else /* #ifdef CONFIG_NO_HZ */
434
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100435#ifdef CONFIG_SMP
436
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100437static int dyntick_save_progress_counter(struct rcu_data *rdp)
438{
439 return 0;
440}
441
442static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
443{
444 return rcu_implicit_offline_qs(rdp);
445}
446
447#endif /* #ifdef CONFIG_SMP */
448
449#endif /* #else #ifdef CONFIG_NO_HZ */
450
451#ifdef CONFIG_RCU_CPU_STALL_DETECTOR
452
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700453int rcu_cpu_stall_panicking __read_mostly;
454
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100455static void record_gp_stall_check_time(struct rcu_state *rsp)
456{
457 rsp->gp_start = jiffies;
458 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK;
459}
460
461static void print_other_cpu_stall(struct rcu_state *rsp)
462{
463 int cpu;
464 long delta;
465 unsigned long flags;
466 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100467
468 /* Only let one CPU complain about others per time interval. */
469
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800470 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100471 delta = jiffies - rsp->jiffies_stall;
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -0700472 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800473 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100474 return;
475 }
476 rsp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700477
478 /*
479 * Now rat on any tasks that got kicked up to the root rcu_node
480 * due to CPU offlining.
481 */
482 rcu_print_task_stall(rnp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800483 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100484
485 /* OK, time to rat on our buddy... */
486
Paul E. McKenney4300aa62010-04-13 16:18:22 -0700487 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks: {",
488 rsp->name);
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700489 rcu_for_each_leaf_node(rsp, rnp) {
Paul E. McKenney3acd9eb2010-02-22 17:05:03 -0800490 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700491 rcu_print_task_stall(rnp);
Paul E. McKenney3acd9eb2010-02-22 17:05:03 -0800492 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700493 if (rnp->qsmask == 0)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100494 continue;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700495 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
496 if (rnp->qsmask & (1UL << cpu))
497 printk(" %d", rnp->grplo + cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100498 }
Paul E. McKenney4300aa62010-04-13 16:18:22 -0700499 printk("} (detected by %d, t=%ld jiffies)\n",
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100500 smp_processor_id(), (long)(jiffies - rsp->gp_start));
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200501 trigger_all_cpu_backtrace();
502
Paul E. McKenney1ed509a2010-02-22 17:05:05 -0800503 /* If so configured, complain about tasks blocking the grace period. */
504
505 rcu_print_detail_task_stall(rsp);
506
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100507 force_quiescent_state(rsp, 0); /* Kick them all. */
508}
509
510static void print_cpu_stall(struct rcu_state *rsp)
511{
512 unsigned long flags;
513 struct rcu_node *rnp = rcu_get_root(rsp);
514
Paul E. McKenney4300aa62010-04-13 16:18:22 -0700515 printk(KERN_ERR "INFO: %s detected stall on CPU %d (t=%lu jiffies)\n",
516 rsp->name, smp_processor_id(), jiffies - rsp->gp_start);
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200517 trigger_all_cpu_backtrace();
518
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800519 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney20133cf2010-02-22 17:05:01 -0800520 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100521 rsp->jiffies_stall =
522 jiffies + RCU_SECONDS_TILL_STALL_RECHECK;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800523 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200524
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100525 set_need_resched(); /* kick ourselves to get things going. */
526}
527
528static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
529{
530 long delta;
531 struct rcu_node *rnp;
532
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700533 if (rcu_cpu_stall_panicking)
534 return;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100535 delta = jiffies - rsp->jiffies_stall;
536 rnp = rdp->mynode;
537 if ((rnp->qsmask & rdp->grpmask) && delta >= 0) {
538
539 /* We haven't checked in, so go dump stack. */
540 print_cpu_stall(rsp);
541
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -0700542 } else if (rcu_gp_in_progress(rsp) && delta >= RCU_STALL_RAT_DELAY) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100543
544 /* They had two time units to dump stack, so complain. */
545 print_other_cpu_stall(rsp);
546 }
547}
548
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700549static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
550{
551 rcu_cpu_stall_panicking = 1;
552 return NOTIFY_DONE;
553}
554
555static struct notifier_block rcu_panic_block = {
556 .notifier_call = rcu_panic,
557};
558
559static void __init check_cpu_stall_init(void)
560{
561 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
562}
563
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100564#else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
565
566static void record_gp_stall_check_time(struct rcu_state *rsp)
567{
568}
569
570static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
571{
572}
573
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700574static void __init check_cpu_stall_init(void)
575{
576}
577
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100578#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
579
580/*
581 * Update CPU-local rcu_data state to record the newly noticed grace period.
582 * This is used both when we started the grace period and when we notice
Paul E. McKenney91603062009-11-02 13:52:29 -0800583 * that someone else started the grace period. The caller must hold the
584 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
585 * and must have irqs disabled.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100586 */
Paul E. McKenney91603062009-11-02 13:52:29 -0800587static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
588{
589 if (rdp->gpnum != rnp->gpnum) {
590 rdp->qs_pending = 1;
591 rdp->passed_quiesc = 0;
592 rdp->gpnum = rnp->gpnum;
593 }
594}
595
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100596static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
597{
Paul E. McKenney91603062009-11-02 13:52:29 -0800598 unsigned long flags;
599 struct rcu_node *rnp;
600
601 local_irq_save(flags);
602 rnp = rdp->mynode;
603 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800604 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
Paul E. McKenney91603062009-11-02 13:52:29 -0800605 local_irq_restore(flags);
606 return;
607 }
608 __note_new_gpnum(rsp, rnp, rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800609 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100610}
611
612/*
613 * Did someone else start a new RCU grace period start since we last
614 * checked? Update local state appropriately if so. Must be called
615 * on the CPU corresponding to rdp.
616 */
617static int
618check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
619{
620 unsigned long flags;
621 int ret = 0;
622
623 local_irq_save(flags);
624 if (rdp->gpnum != rsp->gpnum) {
625 note_new_gpnum(rsp, rdp);
626 ret = 1;
627 }
628 local_irq_restore(flags);
629 return ret;
630}
631
632/*
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800633 * Advance this CPU's callbacks, but only if the current grace period
634 * has ended. This may be called only from the CPU to whom the rdp
635 * belongs. In addition, the corresponding leaf rcu_node structure's
636 * ->lock must be held by the caller, with irqs disabled.
637 */
638static void
639__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
640{
641 /* Did another grace period end? */
642 if (rdp->completed != rnp->completed) {
643
644 /* Advance callbacks. No harm if list empty. */
645 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
646 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
647 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
648
649 /* Remember that we saw this grace-period completion. */
650 rdp->completed = rnp->completed;
651 }
652}
653
654/*
655 * Advance this CPU's callbacks, but only if the current grace period
656 * has ended. This may be called only from the CPU to whom the rdp
657 * belongs.
658 */
659static void
660rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
661{
662 unsigned long flags;
663 struct rcu_node *rnp;
664
665 local_irq_save(flags);
666 rnp = rdp->mynode;
667 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800668 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800669 local_irq_restore(flags);
670 return;
671 }
672 __rcu_process_gp_end(rsp, rnp, rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800673 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800674}
675
676/*
677 * Do per-CPU grace-period initialization for running CPU. The caller
678 * must hold the lock of the leaf rcu_node structure corresponding to
679 * this CPU.
680 */
681static void
682rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
683{
684 /* Prior grace period ended, so advance callbacks for current CPU. */
685 __rcu_process_gp_end(rsp, rnp, rdp);
686
687 /*
688 * Because this CPU just now started the new grace period, we know
689 * that all of its callbacks will be covered by this upcoming grace
690 * period, even the ones that were registered arbitrarily recently.
691 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
692 *
693 * Other CPUs cannot be sure exactly when the grace period started.
694 * Therefore, their recently registered callbacks must pass through
695 * an additional RCU_NEXT_READY stage, so that they will be handled
696 * by the next RCU grace period.
697 */
698 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
699 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
Paul E. McKenney91603062009-11-02 13:52:29 -0800700
701 /* Set state so that this CPU will detect the next quiescent state. */
702 __note_new_gpnum(rsp, rnp, rdp);
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800703}
704
705/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100706 * Start a new RCU grace period if warranted, re-initializing the hierarchy
707 * in preparation for detecting the next grace period. The caller must hold
708 * the root node's ->lock, which is released before return. Hard irqs must
709 * be disabled.
710 */
711static void
712rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
713 __releases(rcu_get_root(rsp)->lock)
714{
715 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
716 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100717
Paul E. McKenney07079d52010-01-04 15:09:02 -0800718 if (!cpu_needs_another_gp(rsp, rdp) || rsp->fqs_active) {
Paul E. McKenney46a1e342010-01-04 15:09:09 -0800719 if (cpu_needs_another_gp(rsp, rdp))
720 rsp->fqs_need_gp = 1;
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -0800721 if (rnp->completed == rsp->completed) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800722 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -0800723 return;
724 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800725 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -0800726
727 /*
728 * Propagate new ->completed value to rcu_node structures
729 * so that other CPUs don't have to wait until the start
730 * of the next grace period to process their callbacks.
731 */
732 rcu_for_each_node_breadth_first(rsp, rnp) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800733 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -0800734 rnp->completed = rsp->completed;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800735 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -0800736 }
737 local_irq_restore(flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100738 return;
739 }
740
741 /* Advance to a new grace period and initialize state. */
742 rsp->gpnum++;
Paul E. McKenneyc3422be2009-09-13 09:15:10 -0700743 WARN_ON_ONCE(rsp->signaled == RCU_GP_INIT);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100744 rsp->signaled = RCU_GP_INIT; /* Hold off force_quiescent_state. */
745 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100746 record_gp_stall_check_time(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100747
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100748 /* Special-case the common single-level case. */
749 if (NUM_RCU_NODES == 1) {
Paul E. McKenneyb0e165c2009-09-13 09:15:09 -0700750 rcu_preempt_check_blocked_tasks(rnp);
Paul E. McKenney28ecd582009-09-18 09:50:17 -0700751 rnp->qsmask = rnp->qsmaskinit;
Paul E. McKenneyde078d82009-09-08 15:54:36 -0700752 rnp->gpnum = rsp->gpnum;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800753 rnp->completed = rsp->completed;
Paul E. McKenneyc12172c2009-01-04 20:30:06 -0800754 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state OK. */
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800755 rcu_start_gp_per_cpu(rsp, rnp, rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800756 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100757 return;
758 }
759
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800760 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100761
762
763 /* Exclude any concurrent CPU-hotplug operations. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800764 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100765
766 /*
Paul E. McKenneyb835db12009-09-08 15:54:37 -0700767 * Set the quiescent-state-needed bits in all the rcu_node
768 * structures for all currently online CPUs in breadth-first
769 * order, starting from the root rcu_node structure. This
770 * operation relies on the layout of the hierarchy within the
771 * rsp->node[] array. Note that other CPUs will access only
772 * the leaves of the hierarchy, which still indicate that no
773 * grace period is in progress, at least until the corresponding
774 * leaf node has been initialized. In addition, we have excluded
775 * CPU-hotplug operations.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100776 *
777 * Note that the grace period cannot complete until we finish
778 * the initialization process, as there will be at least one
779 * qsmask bit set in the root node until that time, namely the
Paul E. McKenneyb835db12009-09-08 15:54:37 -0700780 * one corresponding to this CPU, due to the fact that we have
781 * irqs disabled.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100782 */
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700783 rcu_for_each_node_breadth_first(rsp, rnp) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800784 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenneyb0e165c2009-09-13 09:15:09 -0700785 rcu_preempt_check_blocked_tasks(rnp);
Paul E. McKenney49e29122009-09-18 09:50:19 -0700786 rnp->qsmask = rnp->qsmaskinit;
Paul E. McKenneyde078d82009-09-08 15:54:36 -0700787 rnp->gpnum = rsp->gpnum;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800788 rnp->completed = rsp->completed;
789 if (rnp == rdp->mynode)
790 rcu_start_gp_per_cpu(rsp, rnp, rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800791 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100792 }
793
Paul E. McKenney83f5b012009-10-28 08:14:49 -0700794 rnp = rcu_get_root(rsp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800795 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100796 rsp->signaled = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800797 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
798 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100799}
800
801/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800802 * Report a full set of quiescent states to the specified rcu_state
803 * data structure. This involves cleaning up after the prior grace
804 * period and letting rcu_start_gp() start up the next grace period
805 * if one is needed. Note that the caller must hold rnp->lock, as
806 * required by rcu_start_gp(), which will release it.
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700807 */
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800808static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -0700809 __releases(rcu_get_root(rsp)->lock)
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700810{
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -0700811 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700812 rsp->completed = rsp->gpnum;
Paul E. McKenney83f5b012009-10-28 08:14:49 -0700813 rsp->signaled = RCU_GP_IDLE;
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700814 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
815}
816
817/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800818 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
819 * Allows quiescent states for a group of CPUs to be reported at one go
820 * to the specified rcu_node structure, though all the CPUs in the group
821 * must be represented by the same rcu_node structure (which need not be
822 * a leaf rcu_node structure, though it often will be). That structure's
823 * lock must be held upon entry, and it is released before return.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100824 */
825static void
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800826rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
827 struct rcu_node *rnp, unsigned long flags)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100828 __releases(rnp->lock)
829{
Paul E. McKenney28ecd582009-09-18 09:50:17 -0700830 struct rcu_node *rnp_c;
831
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100832 /* Walk up the rcu_node hierarchy. */
833 for (;;) {
834 if (!(rnp->qsmask & mask)) {
835
836 /* Our bit has already been cleared, so done. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800837 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100838 return;
839 }
840 rnp->qsmask &= ~mask;
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700841 if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100842
843 /* Other bits still set at this level, so done. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800844 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100845 return;
846 }
847 mask = rnp->grpmask;
848 if (rnp->parent == NULL) {
849
850 /* No more levels. Exit loop holding root lock. */
851
852 break;
853 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800854 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney28ecd582009-09-18 09:50:17 -0700855 rnp_c = rnp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100856 rnp = rnp->parent;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800857 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney28ecd582009-09-18 09:50:17 -0700858 WARN_ON_ONCE(rnp_c->qsmask);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100859 }
860
861 /*
862 * Get here if we are the last CPU to pass through a quiescent
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800863 * state for this grace period. Invoke rcu_report_qs_rsp()
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700864 * to clean up and start the next grace period if one is needed.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100865 */
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800866 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100867}
868
869/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800870 * Record a quiescent state for the specified CPU to that CPU's rcu_data
871 * structure. This must be either called from the specified CPU, or
872 * called when the specified CPU is known to be offline (and when it is
873 * also known that no other CPU is concurrently trying to help the offline
874 * CPU). The lastcomp argument is used to make sure we are still in the
875 * grace period of interest. We don't want to end the current grace period
876 * based on quiescent states detected in an earlier grace period!
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100877 */
878static void
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800879rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastcomp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100880{
881 unsigned long flags;
882 unsigned long mask;
883 struct rcu_node *rnp;
884
885 rnp = rdp->mynode;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800886 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney560d4bc2009-11-13 19:51:38 -0800887 if (lastcomp != rnp->completed) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100888
889 /*
890 * Someone beat us to it for this grace period, so leave.
891 * The race with GP start is resolved by the fact that we
892 * hold the leaf rcu_node lock, so that the per-CPU bits
893 * cannot yet be initialized -- so we would simply find our
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800894 * CPU's bit already cleared in rcu_report_qs_rnp() if this
895 * race occurred.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100896 */
897 rdp->passed_quiesc = 0; /* try again later! */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800898 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100899 return;
900 }
901 mask = rdp->grpmask;
902 if ((rnp->qsmask & mask) == 0) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800903 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100904 } else {
905 rdp->qs_pending = 0;
906
907 /*
908 * This GP can't end until cpu checks in, so all of our
909 * callbacks can be processed during the next GP.
910 */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100911 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
912
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800913 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100914 }
915}
916
917/*
918 * Check to see if there is a new grace period of which this CPU
919 * is not yet aware, and if so, set up local rcu_data state for it.
920 * Otherwise, see if this CPU has just passed through its first
921 * quiescent state for this grace period, and record that fact if so.
922 */
923static void
924rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
925{
926 /* If there is now a new grace period, record and return. */
927 if (check_for_new_grace_period(rsp, rdp))
928 return;
929
930 /*
931 * Does this CPU still need to do its part for current grace period?
932 * If no, return and let the other CPUs do their part as well.
933 */
934 if (!rdp->qs_pending)
935 return;
936
937 /*
938 * Was there a quiescent state since the beginning of the grace
939 * period? If no, then exit and wait for the next call.
940 */
941 if (!rdp->passed_quiesc)
942 return;
943
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -0800944 /*
945 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
946 * judge of that).
947 */
948 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesc_completed);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100949}
950
951#ifdef CONFIG_HOTPLUG_CPU
952
953/*
Paul E. McKenneye74f4c42009-10-06 21:48:17 -0700954 * Move a dying CPU's RCU callbacks to the ->orphan_cbs_list for the
955 * specified flavor of RCU. The callbacks will be adopted by the next
956 * _rcu_barrier() invocation or by the CPU_DEAD notifier, whichever
957 * comes first. Because this is invoked from the CPU_DYING notifier,
958 * irqs are already disabled.
959 */
960static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
961{
962 int i;
963 struct rcu_data *rdp = rsp->rda[smp_processor_id()];
964
965 if (rdp->nxtlist == NULL)
966 return; /* irqs disabled, so comparison is stable. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800967 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
Paul E. McKenneye74f4c42009-10-06 21:48:17 -0700968 *rsp->orphan_cbs_tail = rdp->nxtlist;
969 rsp->orphan_cbs_tail = rdp->nxttail[RCU_NEXT_TAIL];
970 rdp->nxtlist = NULL;
971 for (i = 0; i < RCU_NEXT_SIZE; i++)
972 rdp->nxttail[i] = &rdp->nxtlist;
973 rsp->orphan_qlen += rdp->qlen;
974 rdp->qlen = 0;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800975 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
Paul E. McKenneye74f4c42009-10-06 21:48:17 -0700976}
977
978/*
979 * Adopt previously orphaned RCU callbacks.
980 */
981static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
982{
983 unsigned long flags;
984 struct rcu_data *rdp;
985
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800986 raw_spin_lock_irqsave(&rsp->onofflock, flags);
Paul E. McKenneye74f4c42009-10-06 21:48:17 -0700987 rdp = rsp->rda[smp_processor_id()];
988 if (rsp->orphan_cbs_list == NULL) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800989 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
Paul E. McKenneye74f4c42009-10-06 21:48:17 -0700990 return;
991 }
992 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_list;
993 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_cbs_tail;
994 rdp->qlen += rsp->orphan_qlen;
995 rsp->orphan_cbs_list = NULL;
996 rsp->orphan_cbs_tail = &rsp->orphan_cbs_list;
997 rsp->orphan_qlen = 0;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800998 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
Paul E. McKenneye74f4c42009-10-06 21:48:17 -0700999}
1000
1001/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001002 * Remove the outgoing CPU from the bitmasks in the rcu_node hierarchy
1003 * and move all callbacks from the outgoing CPU to the current one.
1004 */
1005static void __rcu_offline_cpu(int cpu, struct rcu_state *rsp)
1006{
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001007 unsigned long flags;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001008 unsigned long mask;
Paul E. McKenneyd9a3da02009-12-02 12:10:15 -08001009 int need_report = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001010 struct rcu_data *rdp = rsp->rda[cpu];
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001011 struct rcu_node *rnp;
1012
1013 /* Exclude any attempts to start a new grace period. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001014 raw_spin_lock_irqsave(&rsp->onofflock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001015
1016 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001017 rnp = rdp->mynode; /* this is the outgoing CPU's rnp. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001018 mask = rdp->grpmask; /* rnp->grplo is constant. */
1019 do {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001020 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001021 rnp->qsmaskinit &= ~mask;
1022 if (rnp->qsmaskinit != 0) {
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -08001023 if (rnp != rdp->mynode)
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001024 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001025 break;
1026 }
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -08001027 if (rnp == rdp->mynode)
Paul E. McKenneyd9a3da02009-12-02 12:10:15 -08001028 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -08001029 else
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001030 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001031 mask = rnp->grpmask;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001032 rnp = rnp->parent;
1033 } while (rnp != NULL);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001034
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -08001035 /*
1036 * We still hold the leaf rcu_node structure lock here, and
1037 * irqs are still disabled. The reason for this subterfuge is
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001038 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1039 * held leads to deadlock.
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -08001040 */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001041 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -08001042 rnp = rdp->mynode;
Paul E. McKenneyd9a3da02009-12-02 12:10:15 -08001043 if (need_report & RCU_OFL_TASKS_NORM_GP)
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001044 rcu_report_unblock_qs_rnp(rnp, flags);
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -08001045 else
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001046 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenneyd9a3da02009-12-02 12:10:15 -08001047 if (need_report & RCU_OFL_TASKS_EXP_GP)
1048 rcu_report_exp_rnp(rsp, rnp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001049
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001050 rcu_adopt_orphan_cbs(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001051}
1052
1053/*
1054 * Remove the specified CPU from the RCU hierarchy and move any pending
1055 * callbacks that it might have to the current CPU. This code assumes
1056 * that at least one CPU in the system will remain running at all times.
1057 * Any attempt to offline -all- CPUs is likely to strand RCU callbacks.
1058 */
1059static void rcu_offline_cpu(int cpu)
1060{
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001061 __rcu_offline_cpu(cpu, &rcu_sched_state);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001062 __rcu_offline_cpu(cpu, &rcu_bh_state);
Paul E. McKenney33f76142009-08-24 09:42:01 -07001063 rcu_preempt_offline_cpu(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001064}
1065
1066#else /* #ifdef CONFIG_HOTPLUG_CPU */
1067
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001068static void rcu_send_cbs_to_orphanage(struct rcu_state *rsp)
1069{
1070}
1071
1072static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1073{
1074}
1075
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001076static void rcu_offline_cpu(int cpu)
1077{
1078}
1079
1080#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1081
1082/*
1083 * Invoke any RCU callbacks that have made it to the end of their grace
1084 * period. Thottle as specified by rdp->blimit.
1085 */
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001086static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001087{
1088 unsigned long flags;
1089 struct rcu_head *next, *list, **tail;
1090 int count;
1091
1092 /* If no callbacks are ready, just return.*/
1093 if (!cpu_has_callbacks_ready_to_invoke(rdp))
1094 return;
1095
1096 /*
1097 * Extract the list of ready callbacks, disabling to prevent
1098 * races with call_rcu() from interrupt handlers.
1099 */
1100 local_irq_save(flags);
1101 list = rdp->nxtlist;
1102 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1103 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1104 tail = rdp->nxttail[RCU_DONE_TAIL];
1105 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1106 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1107 rdp->nxttail[count] = &rdp->nxtlist;
1108 local_irq_restore(flags);
1109
1110 /* Invoke callbacks. */
1111 count = 0;
1112 while (list) {
1113 next = list->next;
1114 prefetch(next);
1115 list->func(list);
1116 list = next;
1117 if (++count >= rdp->blimit)
1118 break;
1119 }
1120
1121 local_irq_save(flags);
1122
1123 /* Update count, and requeue any remaining callbacks. */
1124 rdp->qlen -= count;
1125 if (list != NULL) {
1126 *tail = rdp->nxtlist;
1127 rdp->nxtlist = list;
1128 for (count = 0; count < RCU_NEXT_SIZE; count++)
1129 if (&rdp->nxtlist == rdp->nxttail[count])
1130 rdp->nxttail[count] = tail;
1131 else
1132 break;
1133 }
1134
1135 /* Reinstate batch limit if we have worked down the excess. */
1136 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1137 rdp->blimit = blimit;
1138
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001139 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1140 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1141 rdp->qlen_last_fqs_check = 0;
1142 rdp->n_force_qs_snap = rsp->n_force_qs;
1143 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1144 rdp->qlen_last_fqs_check = rdp->qlen;
1145
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001146 local_irq_restore(flags);
1147
1148 /* Re-raise the RCU softirq if there are callbacks remaining. */
1149 if (cpu_has_callbacks_ready_to_invoke(rdp))
1150 raise_softirq(RCU_SOFTIRQ);
1151}
1152
1153/*
1154 * Check to see if this CPU is in a non-context-switch quiescent state
1155 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
1156 * Also schedule the RCU softirq handler.
1157 *
1158 * This function must be called with hardirqs disabled. It is normally
1159 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1160 * false, there is no point in invoking rcu_check_callbacks().
1161 */
1162void rcu_check_callbacks(int cpu, int user)
1163{
1164 if (user ||
Paul E. McKenneya6826042009-02-25 18:03:42 -08001165 (idle_cpu(cpu) && rcu_scheduler_active &&
1166 !in_softirq() && hardirq_count() <= (1 << HARDIRQ_SHIFT))) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001167
1168 /*
1169 * Get here if this CPU took its interrupt from user
1170 * mode or from the idle loop, and if this is not a
1171 * nested interrupt. In this case, the CPU is in
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001172 * a quiescent state, so note it.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001173 *
1174 * No memory barrier is required here because both
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001175 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1176 * variables that other CPUs neither access nor modify,
1177 * at least not while the corresponding CPU is online.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001178 */
1179
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001180 rcu_sched_qs(cpu);
1181 rcu_bh_qs(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001182
1183 } else if (!in_softirq()) {
1184
1185 /*
1186 * Get here if this CPU did not take its interrupt from
1187 * softirq, in other words, if it is not interrupting
1188 * a rcu_bh read-side critical section. This is an _bh
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001189 * critical section, so note it.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001190 */
1191
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001192 rcu_bh_qs(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001193 }
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001194 rcu_preempt_check_callbacks(cpu);
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07001195 if (rcu_pending(cpu))
1196 raise_softirq(RCU_SOFTIRQ);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001197}
1198
1199#ifdef CONFIG_SMP
1200
1201/*
1202 * Scan the leaf rcu_node structures, processing dyntick state for any that
1203 * have not yet encountered a quiescent state, using the function specified.
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001204 * The caller must have suppressed start of new grace periods.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001205 */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001206static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001207{
1208 unsigned long bit;
1209 int cpu;
1210 unsigned long flags;
1211 unsigned long mask;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001212 struct rcu_node *rnp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001213
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001214 rcu_for_each_leaf_node(rsp, rnp) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001215 mask = 0;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001216 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001217 if (!rcu_gp_in_progress(rsp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001218 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney0f10dc822010-01-04 15:09:06 -08001219 return;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001220 }
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001221 if (rnp->qsmask == 0) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001222 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001223 continue;
1224 }
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001225 cpu = rnp->grplo;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001226 bit = 1;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001227 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
1228 if ((rnp->qsmask & bit) != 0 && f(rsp->rda[cpu]))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001229 mask |= bit;
1230 }
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001231 if (mask != 0) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001232
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001233 /* rcu_report_qs_rnp() releases rnp->lock. */
1234 rcu_report_qs_rnp(mask, rsp, rnp, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001235 continue;
1236 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001237 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001238 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001239}
1240
1241/*
1242 * Force quiescent states on reluctant CPUs, and also detect which
1243 * CPUs are in dyntick-idle mode.
1244 */
1245static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1246{
1247 unsigned long flags;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001248 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001249
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -07001250 if (!rcu_gp_in_progress(rsp))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001251 return; /* No grace period in progress, nothing to force. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001252 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001253 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
1254 return; /* Someone else is already on the job. */
1255 }
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001256 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001257 goto unlock_fqs_ret; /* no emergency and done recently. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001258 rsp->n_force_qs++;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001259 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001260 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
Paul E. McKenney560d4bc2009-11-13 19:51:38 -08001261 if(!rcu_gp_in_progress(rsp)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001262 rsp->n_force_qs_ngp++;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001263 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001264 goto unlock_fqs_ret; /* no GP in progress, time updated. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001265 }
Paul E. McKenney07079d52010-01-04 15:09:02 -08001266 rsp->fqs_active = 1;
Paul E. McKenneyf3a8b5c2010-01-04 15:09:03 -08001267 switch (rsp->signaled) {
Paul E. McKenney83f5b012009-10-28 08:14:49 -07001268 case RCU_GP_IDLE:
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001269 case RCU_GP_INIT:
1270
Paul E. McKenney83f5b012009-10-28 08:14:49 -07001271 break; /* grace period idle or initializing, ignore. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001272
1273 case RCU_SAVE_DYNTICK:
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001274 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1275 break; /* So gcc recognizes the dead code. */
1276
Lai Jiangshanf2614142010-03-28 11:15:20 +08001277 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1278
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001279 /* Record dyntick-idle state. */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001280 force_qs_rnp(rsp, dyntick_save_progress_counter);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001281 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001282 if (rcu_gp_in_progress(rsp))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001283 rsp->signaled = RCU_FORCE_QS;
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001284 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001285
1286 case RCU_FORCE_QS:
1287
1288 /* Check dyntick-idle state, send IPI to laggarts. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001289 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001290 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001291
1292 /* Leave state in case more forcing is required. */
1293
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001294 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001295 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001296 }
Paul E. McKenney07079d52010-01-04 15:09:02 -08001297 rsp->fqs_active = 0;
Paul E. McKenney46a1e342010-01-04 15:09:09 -08001298 if (rsp->fqs_need_gp) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001299 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
Paul E. McKenney46a1e342010-01-04 15:09:09 -08001300 rsp->fqs_need_gp = 0;
1301 rcu_start_gp(rsp, flags); /* releases rnp->lock */
1302 return;
1303 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001304 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001305unlock_fqs_ret:
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001306 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001307}
1308
1309#else /* #ifdef CONFIG_SMP */
1310
1311static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1312{
1313 set_need_resched();
1314}
1315
1316#endif /* #else #ifdef CONFIG_SMP */
1317
1318/*
1319 * This does the RCU processing work from softirq context for the
1320 * specified rcu_state and rcu_data structures. This may be called
1321 * only from the CPU to whom the rdp belongs.
1322 */
1323static void
1324__rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1325{
1326 unsigned long flags;
1327
Paul E. McKenney2e597552009-08-15 09:53:48 -07001328 WARN_ON_ONCE(rdp->beenonline == 0);
1329
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001330 /*
1331 * If an RCU GP has gone long enough, go check for dyntick
1332 * idle CPUs and, if needed, send resched IPIs.
1333 */
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001334 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001335 force_quiescent_state(rsp, 1);
1336
1337 /*
1338 * Advance callbacks in response to end of earlier grace
1339 * period that some other CPU ended.
1340 */
1341 rcu_process_gp_end(rsp, rdp);
1342
1343 /* Update RCU state based on any recent quiescent states. */
1344 rcu_check_quiescent_state(rsp, rdp);
1345
1346 /* Does this CPU require a not-yet-started grace period? */
1347 if (cpu_needs_another_gp(rsp, rdp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001348 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001349 rcu_start_gp(rsp, flags); /* releases above lock */
1350 }
1351
1352 /* If there are callbacks ready, invoke them. */
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001353 rcu_do_batch(rsp, rdp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001354}
1355
1356/*
1357 * Do softirq processing for the current CPU.
1358 */
1359static void rcu_process_callbacks(struct softirq_action *unused)
1360{
1361 /*
1362 * Memory references from any prior RCU read-side critical sections
1363 * executed by the interrupted code must be seen before any RCU
1364 * grace-period manipulations below.
1365 */
1366 smp_mb(); /* See above block comment. */
1367
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001368 __rcu_process_callbacks(&rcu_sched_state,
1369 &__get_cpu_var(rcu_sched_data));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001370 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001371 rcu_preempt_process_callbacks();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001372
1373 /*
1374 * Memory references from any later RCU read-side critical sections
1375 * executed by the interrupted code must be seen after any RCU
1376 * grace-period manipulations above.
1377 */
1378 smp_mb(); /* See above block comment. */
Paul E. McKenneya47cd882010-02-26 16:38:56 -08001379
1380 /* If we are last CPU on way to dyntick-idle mode, accelerate it. */
1381 rcu_needs_cpu_flush();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001382}
1383
1384static void
1385__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
1386 struct rcu_state *rsp)
1387{
1388 unsigned long flags;
1389 struct rcu_data *rdp;
1390
1391 head->func = func;
1392 head->next = NULL;
1393
1394 smp_mb(); /* Ensure RCU update seen before callback registry. */
1395
1396 /*
1397 * Opportunistically note grace-period endings and beginnings.
1398 * Note that we might see a beginning right after we see an
1399 * end, but never vice versa, since this CPU has to pass through
1400 * a quiescent state betweentimes.
1401 */
1402 local_irq_save(flags);
1403 rdp = rsp->rda[smp_processor_id()];
1404 rcu_process_gp_end(rsp, rdp);
1405 check_for_new_grace_period(rsp, rdp);
1406
1407 /* Add the callback to our list. */
1408 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1409 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
1410
1411 /* Start a new grace period if one not already started. */
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -07001412 if (!rcu_gp_in_progress(rsp)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001413 unsigned long nestflag;
1414 struct rcu_node *rnp_root = rcu_get_root(rsp);
1415
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001416 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001417 rcu_start_gp(rsp, nestflag); /* releases rnp_root->lock. */
1418 }
1419
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001420 /*
1421 * Force the grace period if too many callbacks or too long waiting.
1422 * Enforce hysteresis, and don't invoke force_quiescent_state()
1423 * if some other CPU has recently done so. Also, don't bother
1424 * invoking force_quiescent_state() if the newly enqueued callback
1425 * is the only one waiting for a grace period to complete.
1426 */
1427 if (unlikely(++rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001428 rdp->blimit = LONG_MAX;
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001429 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1430 *rdp->nxttail[RCU_DONE_TAIL] != head)
1431 force_quiescent_state(rsp, 0);
1432 rdp->n_force_qs_snap = rsp->n_force_qs;
1433 rdp->qlen_last_fqs_check = rdp->qlen;
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001434 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001435 force_quiescent_state(rsp, 1);
1436 local_irq_restore(flags);
1437}
1438
1439/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001440 * Queue an RCU-sched callback for invocation after a grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001441 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001442void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001443{
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001444 __call_rcu(head, func, &rcu_sched_state);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001445}
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001446EXPORT_SYMBOL_GPL(call_rcu_sched);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001447
1448/*
1449 * Queue an RCU for invocation after a quicker grace period.
1450 */
1451void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1452{
1453 __call_rcu(head, func, &rcu_bh_state);
1454}
1455EXPORT_SYMBOL_GPL(call_rcu_bh);
1456
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001457/**
1458 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1459 *
1460 * Control will return to the caller some time after a full rcu-sched
1461 * grace period has elapsed, in other words after all currently executing
1462 * rcu-sched read-side critical sections have completed. These read-side
1463 * critical sections are delimited by rcu_read_lock_sched() and
1464 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1465 * local_irq_disable(), and so on may be used in place of
1466 * rcu_read_lock_sched().
1467 *
1468 * This means that all preempt_disable code sequences, including NMI and
1469 * hardware-interrupt handlers, in progress on entry will have completed
1470 * before this primitive returns. However, this does not guarantee that
1471 * softirq handlers will have completed, since in some kernels, these
1472 * handlers can run in process context, and can block.
1473 *
1474 * This primitive provides the guarantees made by the (now removed)
1475 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1476 * guarantees that rcu_read_lock() sections will have completed.
1477 * In "classic RCU", these two guarantees happen to be one and
1478 * the same, but can differ in realtime RCU implementations.
1479 */
1480void synchronize_sched(void)
1481{
1482 struct rcu_synchronize rcu;
1483
1484 if (rcu_blocking_is_gp())
1485 return;
1486
Paul E. McKenney72d5a9f2010-05-10 17:12:17 -07001487 init_rcu_head_on_stack(&rcu.head);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001488 init_completion(&rcu.completion);
1489 /* Will wake me after RCU finished. */
1490 call_rcu_sched(&rcu.head, wakeme_after_rcu);
1491 /* Wait for it. */
1492 wait_for_completion(&rcu.completion);
Paul E. McKenney72d5a9f2010-05-10 17:12:17 -07001493 destroy_rcu_head_on_stack(&rcu.head);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001494}
1495EXPORT_SYMBOL_GPL(synchronize_sched);
1496
1497/**
1498 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1499 *
1500 * Control will return to the caller some time after a full rcu_bh grace
1501 * period has elapsed, in other words after all currently executing rcu_bh
1502 * read-side critical sections have completed. RCU read-side critical
1503 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1504 * and may be nested.
1505 */
1506void synchronize_rcu_bh(void)
1507{
1508 struct rcu_synchronize rcu;
1509
1510 if (rcu_blocking_is_gp())
1511 return;
1512
Paul E. McKenney72d5a9f2010-05-10 17:12:17 -07001513 init_rcu_head_on_stack(&rcu.head);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001514 init_completion(&rcu.completion);
1515 /* Will wake me after RCU finished. */
1516 call_rcu_bh(&rcu.head, wakeme_after_rcu);
1517 /* Wait for it. */
1518 wait_for_completion(&rcu.completion);
Paul E. McKenney72d5a9f2010-05-10 17:12:17 -07001519 destroy_rcu_head_on_stack(&rcu.head);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001520}
1521EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1522
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001523/*
1524 * Check to see if there is any immediate RCU-related work to be done
1525 * by the current CPU, for the specified type of RCU, returning 1 if so.
1526 * The checks are in order of increasing expense: checks that can be
1527 * carried out against CPU-local state are performed first. However,
1528 * we must check for CPU stalls first, else we might not get a chance.
1529 */
1530static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
1531{
Paul E. McKenney2f51f982009-11-13 19:51:39 -08001532 struct rcu_node *rnp = rdp->mynode;
1533
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001534 rdp->n_rcu_pending++;
1535
1536 /* Check for CPU stalls, if enabled. */
1537 check_cpu_stall(rsp, rdp);
1538
1539 /* Is the RCU core waiting for a quiescent state from this CPU? */
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07001540 if (rdp->qs_pending && !rdp->passed_quiesc) {
Paul E. McKenneyd25eb942010-03-18 21:36:51 -07001541
1542 /*
1543 * If force_quiescent_state() coming soon and this CPU
1544 * needs a quiescent state, and this is either RCU-sched
1545 * or RCU-bh, force a local reschedule.
1546 */
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07001547 rdp->n_rp_qs_pending++;
Paul E. McKenneyd25eb942010-03-18 21:36:51 -07001548 if (!rdp->preemptable &&
1549 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
1550 jiffies))
1551 set_need_resched();
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07001552 } else if (rdp->qs_pending && rdp->passed_quiesc) {
1553 rdp->n_rp_report_qs++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001554 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001555 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001556
1557 /* Does this CPU have callbacks ready to invoke? */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001558 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
1559 rdp->n_rp_cb_ready++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001560 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001561 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001562
1563 /* Has RCU gone idle with this CPU needing another grace period? */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001564 if (cpu_needs_another_gp(rsp, rdp)) {
1565 rdp->n_rp_cpu_needs_gp++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001566 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001567 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001568
1569 /* Has another RCU grace period completed? */
Paul E. McKenney2f51f982009-11-13 19:51:39 -08001570 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001571 rdp->n_rp_gp_completed++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001572 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001573 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001574
1575 /* Has a new RCU grace period started? */
Paul E. McKenney2f51f982009-11-13 19:51:39 -08001576 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001577 rdp->n_rp_gp_started++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001578 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001579 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001580
1581 /* Has an RCU GP gone long enough to send resched IPIs &c? */
Paul E. McKenneyfc2219d2009-09-23 09:50:41 -07001582 if (rcu_gp_in_progress(rsp) &&
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001583 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001584 rdp->n_rp_need_fqs++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001585 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001586 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001587
1588 /* nothing to do */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07001589 rdp->n_rp_need_nothing++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001590 return 0;
1591}
1592
1593/*
1594 * Check to see if there is any immediate RCU-related work to be done
1595 * by the current CPU, returning 1 if so. This function is part of the
1596 * RCU implementation; it is -not- an exported member of the RCU API.
1597 */
Paul E. McKenneya1572292009-08-22 13:56:51 -07001598static int rcu_pending(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001599{
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001600 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001601 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
1602 rcu_preempt_pending(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001603}
1604
1605/*
1606 * Check to see if any future RCU-related work will need to be done
1607 * by the current CPU, even if none need be done immediately, returning
Paul E. McKenney8bd93a22010-02-22 17:04:59 -08001608 * 1 if so.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001609 */
Paul E. McKenney8bd93a22010-02-22 17:04:59 -08001610static int rcu_needs_cpu_quick_check(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001611{
1612 /* RCU callbacks either ready or pending? */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001613 return per_cpu(rcu_sched_data, cpu).nxtlist ||
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001614 per_cpu(rcu_bh_data, cpu).nxtlist ||
1615 rcu_preempt_needs_cpu(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001616}
1617
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001618static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
1619static atomic_t rcu_barrier_cpu_count;
1620static DEFINE_MUTEX(rcu_barrier_mutex);
1621static struct completion rcu_barrier_completion;
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001622
1623static void rcu_barrier_callback(struct rcu_head *notused)
1624{
1625 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1626 complete(&rcu_barrier_completion);
1627}
1628
1629/*
1630 * Called with preemption disabled, and from cross-cpu IRQ context.
1631 */
1632static void rcu_barrier_func(void *type)
1633{
1634 int cpu = smp_processor_id();
1635 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
1636 void (*call_rcu_func)(struct rcu_head *head,
1637 void (*func)(struct rcu_head *head));
1638
1639 atomic_inc(&rcu_barrier_cpu_count);
1640 call_rcu_func = type;
1641 call_rcu_func(head, rcu_barrier_callback);
1642}
1643
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001644/*
1645 * Orchestrate the specified type of RCU barrier, waiting for all
1646 * RCU callbacks of the specified type to complete.
1647 */
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001648static void _rcu_barrier(struct rcu_state *rsp,
1649 void (*call_rcu_func)(struct rcu_head *head,
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001650 void (*func)(struct rcu_head *head)))
1651{
1652 BUG_ON(in_interrupt());
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001653 /* Take mutex to serialize concurrent rcu_barrier() requests. */
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001654 mutex_lock(&rcu_barrier_mutex);
1655 init_completion(&rcu_barrier_completion);
1656 /*
1657 * Initialize rcu_barrier_cpu_count to 1, then invoke
1658 * rcu_barrier_func() on each CPU, so that each CPU also has
1659 * incremented rcu_barrier_cpu_count. Only then is it safe to
1660 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
1661 * might complete its grace period before all of the other CPUs
1662 * did their increment, causing this function to return too
1663 * early.
1664 */
1665 atomic_set(&rcu_barrier_cpu_count, 1);
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001666 preempt_disable(); /* stop CPU_DYING from filling orphan_cbs_list */
1667 rcu_adopt_orphan_cbs(rsp);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001668 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001669 preempt_enable(); /* CPU_DYING can again fill orphan_cbs_list */
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001670 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
1671 complete(&rcu_barrier_completion);
1672 wait_for_completion(&rcu_barrier_completion);
1673 mutex_unlock(&rcu_barrier_mutex);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001674}
1675
1676/**
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001677 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
1678 */
1679void rcu_barrier_bh(void)
1680{
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001681 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001682}
1683EXPORT_SYMBOL_GPL(rcu_barrier_bh);
1684
1685/**
1686 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
1687 */
1688void rcu_barrier_sched(void)
1689{
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001690 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001691}
1692EXPORT_SYMBOL_GPL(rcu_barrier_sched);
1693
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001694/*
Paul E. McKenney27569622009-08-15 09:53:46 -07001695 * Do boot-time initialization of a CPU's per-CPU RCU data.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001696 */
Paul E. McKenney27569622009-08-15 09:53:46 -07001697static void __init
1698rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001699{
1700 unsigned long flags;
1701 int i;
Paul E. McKenney27569622009-08-15 09:53:46 -07001702 struct rcu_data *rdp = rsp->rda[cpu];
1703 struct rcu_node *rnp = rcu_get_root(rsp);
1704
1705 /* Set up local state, ensuring consistent view of global state. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001706 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney27569622009-08-15 09:53:46 -07001707 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
1708 rdp->nxtlist = NULL;
1709 for (i = 0; i < RCU_NEXT_SIZE; i++)
1710 rdp->nxttail[i] = &rdp->nxtlist;
1711 rdp->qlen = 0;
1712#ifdef CONFIG_NO_HZ
1713 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
1714#endif /* #ifdef CONFIG_NO_HZ */
1715 rdp->cpu = cpu;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001716 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney27569622009-08-15 09:53:46 -07001717}
1718
1719/*
1720 * Initialize a CPU's per-CPU RCU data. Note that only one online or
1721 * offline event can be happening at a given time. Note also that we
1722 * can accept some slop in the rsp->completed access due to the fact
1723 * that this CPU cannot possibly have any RCU callbacks in flight yet.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001724 */
1725static void __cpuinit
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001726rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptable)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001727{
1728 unsigned long flags;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001729 unsigned long mask;
1730 struct rcu_data *rdp = rsp->rda[cpu];
1731 struct rcu_node *rnp = rcu_get_root(rsp);
1732
1733 /* Set up local state, ensuring consistent view of global state. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001734 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001735 rdp->passed_quiesc = 0; /* We could be racing with new GP, */
1736 rdp->qs_pending = 1; /* so set up to respond to current GP. */
1737 rdp->beenonline = 1; /* We have now been online. */
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001738 rdp->preemptable = preemptable;
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001739 rdp->qlen_last_fqs_check = 0;
1740 rdp->n_force_qs_snap = rsp->n_force_qs;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001741 rdp->blimit = blimit;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001742 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001743
1744 /*
1745 * A new grace period might start here. If so, we won't be part
1746 * of it, but that is OK, as we are currently in a quiescent state.
1747 */
1748
1749 /* Exclude any attempts to start a new GP on large systems. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001750 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001751
1752 /* Add CPU to rcu_node bitmasks. */
1753 rnp = rdp->mynode;
1754 mask = rdp->grpmask;
1755 do {
1756 /* Exclude any attempts to start a new GP on small systems. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001757 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001758 rnp->qsmaskinit |= mask;
1759 mask = rnp->grpmask;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001760 if (rnp == rdp->mynode) {
1761 rdp->gpnum = rnp->completed; /* if GP in progress... */
1762 rdp->completed = rnp->completed;
1763 rdp->passed_quiesc_completed = rnp->completed - 1;
1764 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001765 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001766 rnp = rnp->parent;
1767 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
1768
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001769 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001770}
1771
1772static void __cpuinit rcu_online_cpu(int cpu)
1773{
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001774 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
1775 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
1776 rcu_preempt_init_percpu_data(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001777}
1778
1779/*
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001780 * Handle CPU online/offline notification events.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001781 */
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08001782static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
1783 unsigned long action, void *hcpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001784{
1785 long cpu = (long)hcpu;
1786
1787 switch (action) {
1788 case CPU_UP_PREPARE:
1789 case CPU_UP_PREPARE_FROZEN:
1790 rcu_online_cpu(cpu);
1791 break;
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001792 case CPU_DYING:
1793 case CPU_DYING_FROZEN:
1794 /*
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001795 * preempt_disable() in _rcu_barrier() prevents stop_machine(),
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001796 * so when "on_each_cpu(rcu_barrier_func, (void *)type, 1);"
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001797 * returns, all online cpus have queued rcu_barrier_func().
1798 * The dying CPU clears its cpu_online_mask bit and
1799 * moves all of its RCU callbacks to ->orphan_cbs_list
1800 * in the context of stop_machine(), so subsequent calls
1801 * to _rcu_barrier() will adopt these callbacks and only
1802 * then queue rcu_barrier_func() on all remaining CPUs.
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001803 */
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001804 rcu_send_cbs_to_orphanage(&rcu_bh_state);
1805 rcu_send_cbs_to_orphanage(&rcu_sched_state);
1806 rcu_preempt_send_cbs_to_orphanage();
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07001807 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001808 case CPU_DEAD:
1809 case CPU_DEAD_FROZEN:
1810 case CPU_UP_CANCELED:
1811 case CPU_UP_CANCELED_FROZEN:
1812 rcu_offline_cpu(cpu);
1813 break;
1814 default:
1815 break;
1816 }
1817 return NOTIFY_OK;
1818}
1819
1820/*
Paul E. McKenneybbad9372010-04-02 16:17:17 -07001821 * This function is invoked towards the end of the scheduler's initialization
1822 * process. Before this is called, the idle task might contain
1823 * RCU read-side critical sections (during which time, this idle
1824 * task is booting the system). After this function is called, the
1825 * idle tasks are prohibited from containing RCU read-side critical
1826 * sections. This function also enables RCU lockdep checking.
1827 */
1828void rcu_scheduler_starting(void)
1829{
1830 WARN_ON(num_online_cpus() != 1);
1831 WARN_ON(nr_context_switches() > 0);
1832 rcu_scheduler_active = 1;
1833}
1834
1835/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001836 * Compute the per-level fanout, either using the exact fanout specified
1837 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
1838 */
1839#ifdef CONFIG_RCU_FANOUT_EXACT
1840static void __init rcu_init_levelspread(struct rcu_state *rsp)
1841{
1842 int i;
1843
1844 for (i = NUM_RCU_LVLS - 1; i >= 0; i--)
1845 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
1846}
1847#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
1848static void __init rcu_init_levelspread(struct rcu_state *rsp)
1849{
1850 int ccur;
1851 int cprv;
1852 int i;
1853
1854 cprv = NR_CPUS;
1855 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1856 ccur = rsp->levelcnt[i];
1857 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
1858 cprv = ccur;
1859 }
1860}
1861#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
1862
1863/*
1864 * Helper function for rcu_init() that initializes one rcu_state structure.
1865 */
1866static void __init rcu_init_one(struct rcu_state *rsp)
1867{
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08001868 static char *buf[] = { "rcu_node_level_0",
1869 "rcu_node_level_1",
1870 "rcu_node_level_2",
1871 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001872 int cpustride = 1;
1873 int i;
1874 int j;
1875 struct rcu_node *rnp;
1876
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08001877 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
1878
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001879 /* Initialize the level-tracking arrays. */
1880
1881 for (i = 1; i < NUM_RCU_LVLS; i++)
1882 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
1883 rcu_init_levelspread(rsp);
1884
1885 /* Initialize the elements themselves, starting from the leaves. */
1886
1887 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
1888 cpustride *= rsp->levelspread[i];
1889 rnp = rsp->level[i];
1890 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001891 raw_spin_lock_init(&rnp->lock);
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08001892 lockdep_set_class_and_name(&rnp->lock,
1893 &rcu_node_class[i], buf[i]);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001894 rnp->gpnum = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001895 rnp->qsmask = 0;
1896 rnp->qsmaskinit = 0;
1897 rnp->grplo = j * cpustride;
1898 rnp->grphi = (j + 1) * cpustride - 1;
1899 if (rnp->grphi >= NR_CPUS)
1900 rnp->grphi = NR_CPUS - 1;
1901 if (i == 0) {
1902 rnp->grpnum = 0;
1903 rnp->grpmask = 0;
1904 rnp->parent = NULL;
1905 } else {
1906 rnp->grpnum = j % rsp->levelspread[i - 1];
1907 rnp->grpmask = 1UL << rnp->grpnum;
1908 rnp->parent = rsp->level[i - 1] +
1909 j / rsp->levelspread[i - 1];
1910 }
1911 rnp->level = i;
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001912 INIT_LIST_HEAD(&rnp->blocked_tasks[0]);
1913 INIT_LIST_HEAD(&rnp->blocked_tasks[1]);
Paul E. McKenneyd9a3da02009-12-02 12:10:15 -08001914 INIT_LIST_HEAD(&rnp->blocked_tasks[2]);
1915 INIT_LIST_HEAD(&rnp->blocked_tasks[3]);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001916 }
1917 }
Lai Jiangshan0c340292010-03-28 11:12:30 +08001918
1919 rnp = rsp->level[NUM_RCU_LVLS - 1];
1920 for_each_possible_cpu(i) {
Paul E. McKenney4a90a062010-04-14 16:48:11 -07001921 while (i > rnp->grphi)
Lai Jiangshan0c340292010-03-28 11:12:30 +08001922 rnp++;
1923 rsp->rda[i]->mynode = rnp;
1924 rcu_boot_init_percpu_data(i, rsp);
1925 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001926}
1927
1928/*
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001929 * Helper macro for __rcu_init() and __rcu_init_preempt(). To be used
1930 * nowhere else! Assigns leaf node pointers into each CPU's rcu_data
1931 * structure.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001932 */
Paul E. McKenney65cf8f82009-08-22 13:56:49 -07001933#define RCU_INIT_FLAVOR(rsp, rcu_data) \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001934do { \
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001935 int i; \
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001936 \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001937 for_each_possible_cpu(i) { \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001938 (rsp)->rda[i] = &per_cpu(rcu_data, i); \
1939 } \
Lai Jiangshan0c340292010-03-28 11:12:30 +08001940 rcu_init_one(rsp); \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001941} while (0)
1942
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08001943void __init rcu_init(void)
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001944{
Paul E. McKenney017c4262010-01-14 16:10:58 -08001945 int cpu;
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08001946
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001947 rcu_bootup_announce();
Paul E. McKenney65cf8f82009-08-22 13:56:49 -07001948 RCU_INIT_FLAVOR(&rcu_sched_state, rcu_sched_data);
1949 RCU_INIT_FLAVOR(&rcu_bh_state, rcu_bh_data);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001950 __rcu_init_preempt();
Paul E. McKenney2e597552009-08-15 09:53:48 -07001951 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08001952
1953 /*
1954 * We don't need protection against CPU-hotplug here because
1955 * this is called early in boot, before either interrupts
1956 * or the scheduler are operational.
1957 */
1958 cpu_notifier(rcu_cpu_notify, 0);
Paul E. McKenney017c4262010-01-14 16:10:58 -08001959 for_each_online_cpu(cpu)
1960 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Paul E. McKenneyc68de202010-04-15 10:12:40 -07001961 check_cpu_stall_init();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001962}
1963
Paul E. McKenney1eba8f82009-09-23 09:50:42 -07001964#include "rcutree_plugin.h"