blob: 1050d6d3922c182f09bfe602ee07e3a185f4be80 [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. McKenney8826f3b2011-05-11 05:41:41 -070039#include <linux/atomic.h>
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010040#include <linux/bitops.h>
Paul Gortmaker9984de12011-05-23 14:51:41 -040041#include <linux/export.h>
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010042#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. McKenneya26ac242011-01-12 14:10:23 -080050#include <linux/wait.h>
51#include <linux/kthread.h>
Linus Torvalds268bb0c2011-05-20 12:50:29 -070052#include <linux/prefetch.h>
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -080053#include <linux/delay.h>
54#include <linux/stop_machine.h>
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010055
Paul E. McKenney9f77da92009-08-22 13:56:45 -070056#include "rcutree.h"
Paul E. McKenney29c00b42011-06-17 15:53:19 -070057#include <trace/events/rcu.h>
58
59#include "rcu.h"
Paul E. McKenney9f77da92009-08-22 13:56:45 -070060
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010061/* Data structures. */
62
Paul E. McKenneyb668c9c2009-11-22 08:53:48 -080063static struct lock_class_key rcu_node_class[NUM_RCU_LVLS];
Peter Zijlstra88b91c72009-10-26 10:24:31 -070064
Paul E. McKenney4300aa62010-04-13 16:18:22 -070065#define RCU_STATE_INITIALIZER(structname) { \
Paul E. McKenneye99033c2011-06-21 00:13:44 -070066 .level = { &structname##_state.node[0] }, \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010067 .levelcnt = { \
68 NUM_RCU_LVL_0, /* root of hierarchy. */ \
69 NUM_RCU_LVL_1, \
70 NUM_RCU_LVL_2, \
Paul E. McKenneycf244dc2009-12-02 12:10:14 -080071 NUM_RCU_LVL_3, \
72 NUM_RCU_LVL_4, /* == MAX_RCU_LVLS */ \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010073 }, \
Paul E. McKenneyaf446b72011-09-10 21:54:08 -070074 .fqs_state = RCU_GP_IDLE, \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010075 .gpnum = -300, \
76 .completed = -300, \
Paul E. McKenneye99033c2011-06-21 00:13:44 -070077 .onofflock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.onofflock), \
78 .fqslock = __RAW_SPIN_LOCK_UNLOCKED(&structname##_state.fqslock), \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010079 .n_force_qs = 0, \
80 .n_force_qs_ngp = 0, \
Paul E. McKenney4300aa62010-04-13 16:18:22 -070081 .name = #structname, \
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010082}
83
Paul E. McKenneye99033c2011-06-21 00:13:44 -070084struct rcu_state rcu_sched_state = RCU_STATE_INITIALIZER(rcu_sched);
Paul E. McKenneyd6714c22009-08-22 13:56:46 -070085DEFINE_PER_CPU(struct rcu_data, rcu_sched_data);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010086
Paul E. McKenneye99033c2011-06-21 00:13:44 -070087struct rcu_state rcu_bh_state = RCU_STATE_INITIALIZER(rcu_bh);
Ingo Molnar6258c4f2009-03-25 16:42:24 +010088DEFINE_PER_CPU(struct rcu_data, rcu_bh_data);
Ingo Molnarb1f77b02009-03-13 03:20:49 +010089
Paul E. McKenney27f4d282011-02-07 12:47:15 -080090static struct rcu_state *rcu_state;
91
Paul E. McKenneyb0d30412011-07-10 15:57:35 -070092/*
93 * The rcu_scheduler_active variable transitions from zero to one just
94 * before the first task is spawned. So when this variable is zero, RCU
95 * can assume that there is but one task, allowing RCU to (for example)
96 * optimized synchronize_sched() to a simple barrier(). When this variable
97 * is one, RCU must actually do all the hard work required to detect real
98 * grace periods. This variable is also used to suppress boot-time false
99 * positives from lockdep-RCU error checking.
100 */
Paul E. McKenneybbad9372010-04-02 16:17:17 -0700101int rcu_scheduler_active __read_mostly;
102EXPORT_SYMBOL_GPL(rcu_scheduler_active);
103
Paul E. McKenneyb0d30412011-07-10 15:57:35 -0700104/*
105 * The rcu_scheduler_fully_active variable transitions from zero to one
106 * during the early_initcall() processing, which is after the scheduler
107 * is capable of creating new tasks. So RCU processing (for example,
108 * creating tasks for RCU priority boosting) must be delayed until after
109 * rcu_scheduler_fully_active transitions from zero to one. We also
110 * currently delay invocation of any RCU callbacks until after this point.
111 *
112 * It might later prove better for people registering RCU callbacks during
113 * early boot to take responsibility for these callbacks, but one step at
114 * a time.
115 */
116static int rcu_scheduler_fully_active __read_mostly;
117
Paul E. McKenneya46e0892011-06-15 15:47:09 -0700118#ifdef CONFIG_RCU_BOOST
119
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100120/*
Paul E. McKenneya26ac242011-01-12 14:10:23 -0800121 * Control variables for per-CPU and per-rcu_node kthreads. These
122 * handle all flavors of RCU.
123 */
124static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
Paul E. McKenneyd71df902011-03-29 17:48:28 -0700125DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
Paul E. McKenney15ba0ba2011-04-06 16:01:16 -0700126DEFINE_PER_CPU(int, rcu_cpu_kthread_cpu);
Paul E. McKenney5ece5ba2011-04-22 18:08:51 -0700127DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
Paul E. McKenneyd71df902011-03-29 17:48:28 -0700128DEFINE_PER_CPU(char, rcu_cpu_has_work);
Paul E. McKenneya26ac242011-01-12 14:10:23 -0800129
Paul E. McKenneya46e0892011-06-15 15:47:09 -0700130#endif /* #ifdef CONFIG_RCU_BOOST */
131
Paul E. McKenney0f962a52011-04-14 12:13:53 -0700132static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
Paul E. McKenneya46e0892011-06-15 15:47:09 -0700133static void invoke_rcu_core(void);
134static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
Paul E. McKenneya26ac242011-01-12 14:10:23 -0800135
Paul E. McKenneya26ac242011-01-12 14:10:23 -0800136/*
Paul E. McKenney4a298652011-04-03 21:33:51 -0700137 * Track the rcutorture test sequence number and the update version
138 * number within a given test. The rcutorture_testseq is incremented
139 * on every rcutorture module load and unload, so has an odd value
140 * when a test is running. The rcutorture_vernum is set to zero
141 * when rcutorture starts and is incremented on each rcutorture update.
142 * These variables enable correlating rcutorture output with the
143 * RCU tracing information.
144 */
145unsigned long rcutorture_testseq;
146unsigned long rcutorture_vernum;
147
148/*
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -0700149 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
150 * permit this function to be invoked without holding the root rcu_node
151 * structure's ->lock, but of course results can be subject to change.
152 */
153static int rcu_gp_in_progress(struct rcu_state *rsp)
154{
155 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
156}
157
158/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700159 * Note a quiescent state. Because we do not need to know
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100160 * how many quiescent states passed, just if there was at least
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700161 * one since the start of the grace period, this just sets a flag.
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700162 * The caller must have disabled preemption.
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100163 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700164void rcu_sched_qs(int cpu)
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100165{
Paul E. McKenney25502a62010-04-01 17:37:01 -0700166 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700167
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700168 rdp->passed_quiesce_gpnum = rdp->gpnum;
Paul E. McKenneyc3422be2009-09-13 09:15:10 -0700169 barrier();
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700170 if (rdp->passed_quiesce == 0)
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700171 trace_rcu_grace_period("rcu_sched", rdp->gpnum, "cpuqs");
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700172 rdp->passed_quiesce = 1;
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100173}
174
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700175void rcu_bh_qs(int cpu)
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100176{
Paul E. McKenney25502a62010-04-01 17:37:01 -0700177 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700178
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700179 rdp->passed_quiesce_gpnum = rdp->gpnum;
Paul E. McKenneyc3422be2009-09-13 09:15:10 -0700180 barrier();
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700181 if (rdp->passed_quiesce == 0)
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700182 trace_rcu_grace_period("rcu_bh", rdp->gpnum, "cpuqs");
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700183 rdp->passed_quiesce = 1;
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100184}
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100185
Paul E. McKenney25502a62010-04-01 17:37:01 -0700186/*
187 * Note a context switch. This is a quiescent state for RCU-sched,
188 * and requires special handling for preemptible RCU.
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700189 * The caller must have disabled preemption.
Paul E. McKenney25502a62010-04-01 17:37:01 -0700190 */
191void rcu_note_context_switch(int cpu)
192{
Paul E. McKenney300df912011-06-18 22:26:31 -0700193 trace_rcu_utilization("Start context switch");
Paul E. McKenney25502a62010-04-01 17:37:01 -0700194 rcu_sched_qs(cpu);
195 rcu_preempt_note_context_switch(cpu);
Paul E. McKenney300df912011-06-18 22:26:31 -0700196 trace_rcu_utilization("End context switch");
Paul E. McKenney25502a62010-04-01 17:37:01 -0700197}
Gleb Natapov29ce8312011-05-04 16:31:03 +0300198EXPORT_SYMBOL_GPL(rcu_note_context_switch);
Paul E. McKenney25502a62010-04-01 17:37:01 -0700199
Paul E. McKenney90a4d2c2009-01-04 11:41:11 -0800200DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
Paul E. McKenney29e37d82011-11-17 16:55:56 -0800201 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700202 .dynticks = ATOMIC_INIT(1),
Paul E. McKenney90a4d2c2009-01-04 11:41:11 -0800203};
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100204
Paul E. McKenneye0f23062011-06-21 01:29:39 -0700205static int blimit = 10; /* Maximum callbacks per rcu_do_batch. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100206static int qhimark = 10000; /* If this many pending, ignore blimit. */
207static int qlowmark = 100; /* Once only this many pending, use blimit. */
208
Paul E. McKenney3d76c082009-09-28 07:46:32 -0700209module_param(blimit, int, 0);
210module_param(qhimark, int, 0);
211module_param(qlowmark, int, 0);
212
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800213int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
214int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
215
Paul E. McKenneyf2e0dd72010-07-14 14:38:30 -0700216module_param(rcu_cpu_stall_suppress, int, 0644);
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800217module_param(rcu_cpu_stall_timeout, int, 0644);
Paul E. McKenney742734e2010-06-30 11:43:52 -0700218
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100219static void force_quiescent_state(struct rcu_state *rsp, int relaxed);
Paul E. McKenneya1572292009-08-22 13:56:51 -0700220static int rcu_pending(int cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100221
222/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700223 * Return the number of RCU-sched batches processed thus far for debug & stats.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100224 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700225long rcu_batches_completed_sched(void)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100226{
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700227 return rcu_sched_state.completed;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100228}
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700229EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100230
231/*
232 * Return the number of RCU BH batches processed thus far for debug & stats.
233 */
234long rcu_batches_completed_bh(void)
235{
236 return rcu_bh_state.completed;
237}
238EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
239
240/*
Paul E. McKenneybf66f182010-01-04 15:09:10 -0800241 * Force a quiescent state for RCU BH.
242 */
243void rcu_bh_force_quiescent_state(void)
244{
245 force_quiescent_state(&rcu_bh_state, 0);
246}
247EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
248
249/*
Paul E. McKenney4a298652011-04-03 21:33:51 -0700250 * Record the number of times rcutorture tests have been initiated and
251 * terminated. This information allows the debugfs tracing stats to be
252 * correlated to the rcutorture messages, even when the rcutorture module
253 * is being repeatedly loaded and unloaded. In other words, we cannot
254 * store this state in rcutorture itself.
255 */
256void rcutorture_record_test_transition(void)
257{
258 rcutorture_testseq++;
259 rcutorture_vernum = 0;
260}
261EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
262
263/*
264 * Record the number of writer passes through the current rcutorture test.
265 * This is also used to correlate debugfs tracing stats with the rcutorture
266 * messages.
267 */
268void rcutorture_record_progress(unsigned long vernum)
269{
270 rcutorture_vernum++;
271}
272EXPORT_SYMBOL_GPL(rcutorture_record_progress);
273
274/*
Paul E. McKenneybf66f182010-01-04 15:09:10 -0800275 * Force a quiescent state for RCU-sched.
276 */
277void rcu_sched_force_quiescent_state(void)
278{
279 force_quiescent_state(&rcu_sched_state, 0);
280}
281EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
282
283/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100284 * Does the CPU have callbacks ready to be invoked?
285 */
286static int
287cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
288{
289 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL];
290}
291
292/*
293 * Does the current CPU require a yet-as-unscheduled grace period?
294 */
295static int
296cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
297{
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -0700298 return *rdp->nxttail[RCU_DONE_TAIL] && !rcu_gp_in_progress(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100299}
300
301/*
302 * Return the root node of the specified rcu_state structure.
303 */
304static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
305{
306 return &rsp->node[0];
307}
308
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100309/*
310 * If the specified CPU is offline, tell the caller that it is in
311 * a quiescent state. Otherwise, whack it with a reschedule IPI.
312 * Grace periods can end up waiting on an offline CPU when that
313 * CPU is in the process of coming online -- it will be added to the
314 * rcu_node bitmasks before it actually makes it online. The same thing
315 * can happen while a CPU is in the process of coming online. Because this
316 * race is quite rare, we check for it after detecting that the grace
317 * period has been delayed rather than checking each and every CPU
318 * each and every time we start a new grace period.
319 */
320static int rcu_implicit_offline_qs(struct rcu_data *rdp)
321{
322 /*
Paul E. McKenney2036d942012-01-30 17:02:47 -0800323 * If the CPU is offline for more than a jiffy, it is in a quiescent
324 * state. We can trust its state not to change because interrupts
325 * are disabled. The reason for the jiffy's worth of slack is to
326 * handle CPUs initializing on the way up and finding their way
327 * to the idle loop on the way down.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100328 */
Paul E. McKenney2036d942012-01-30 17:02:47 -0800329 if (cpu_is_offline(rdp->cpu) &&
330 ULONG_CMP_LT(rdp->rsp->gp_start + 2, jiffies)) {
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700331 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "ofl");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100332 rdp->offline_fqs++;
333 return 1;
334 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100335 return 0;
336}
337
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700338/*
339 * rcu_idle_enter_common - inform RCU that current CPU is moving towards idle
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100340 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700341 * If the new value of the ->dynticks_nesting counter now is zero,
342 * we really have entered idle, and must do the appropriate accounting.
343 * The caller must have disabled interrupts.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100344 */
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700345static void rcu_idle_enter_common(struct rcu_dynticks *rdtp, long long oldval)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100346{
Frederic Weisbeckerfacc4e12011-11-28 16:26:56 -0800347 trace_rcu_dyntick("Start", oldval, 0);
Paul E. McKenney99745b62011-11-10 15:48:45 -0800348 if (!is_idle_task(current)) {
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700349 struct task_struct *idle = idle_task(smp_processor_id());
350
Frederic Weisbeckerfacc4e12011-11-28 16:26:56 -0800351 trace_rcu_dyntick("Error on entry: not idle task", oldval, 0);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700352 ftrace_dump(DUMP_ALL);
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700353 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
354 current->pid, current->comm,
355 idle->pid, idle->comm); /* must be idle task! */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700356 }
Paul E. McKenneyaea1b352011-11-02 06:54:54 -0700357 rcu_prepare_for_idle(smp_processor_id());
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700358 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
359 smp_mb__before_atomic_inc(); /* See above. */
360 atomic_inc(&rdtp->dynticks);
361 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
362 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
Paul E. McKenneyc44e2cd2012-01-12 13:08:18 -0800363
364 /*
365 * The idle task is not permitted to enter the idle loop while
366 * in an RCU read-side critical section.
367 */
368 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
369 "Illegal idle entry in RCU read-side critical section.");
370 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
371 "Illegal idle entry in RCU-bh read-side critical section.");
372 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
373 "Illegal idle entry in RCU-sched read-side critical section.");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100374}
375
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700376/**
377 * rcu_idle_enter - inform RCU that current CPU is entering idle
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100378 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700379 * Enter idle mode, in other words, -leave- the mode in which RCU
380 * read-side critical sections can occur. (Though RCU read-side
381 * critical sections can occur in irq handlers in idle, a possibility
382 * handled by irq_enter() and irq_exit().)
383 *
384 * We crowbar the ->dynticks_nesting field to zero to allow for
385 * the possibility of usermode upcalls having messed up our count
386 * of interrupt nesting level during the prior busy period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100387 */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700388void rcu_idle_enter(void)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100389{
390 unsigned long flags;
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700391 long long oldval;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100392 struct rcu_dynticks *rdtp;
393
394 local_irq_save(flags);
395 rdtp = &__get_cpu_var(rcu_dynticks);
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700396 oldval = rdtp->dynticks_nesting;
Paul E. McKenney29e37d82011-11-17 16:55:56 -0800397 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
398 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
399 rdtp->dynticks_nesting = 0;
400 else
401 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700402 rcu_idle_enter_common(rdtp, oldval);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700403 local_irq_restore(flags);
404}
Paul E. McKenney8a2ecf42012-02-02 15:42:04 -0800405EXPORT_SYMBOL_GPL(rcu_idle_enter);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700406
407/**
408 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
409 *
410 * Exit from an interrupt handler, which might possibly result in entering
411 * idle mode, in other words, leaving the mode in which read-side critical
412 * sections can occur.
413 *
414 * This code assumes that the idle loop never does anything that might
415 * result in unbalanced calls to irq_enter() and irq_exit(). If your
416 * architecture violates this assumption, RCU will give you what you
417 * deserve, good and hard. But very infrequently and irreproducibly.
418 *
419 * Use things like work queues to work around this limitation.
420 *
421 * You have been warned.
422 */
423void rcu_irq_exit(void)
424{
425 unsigned long flags;
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700426 long long oldval;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700427 struct rcu_dynticks *rdtp;
428
429 local_irq_save(flags);
430 rdtp = &__get_cpu_var(rcu_dynticks);
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700431 oldval = rdtp->dynticks_nesting;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700432 rdtp->dynticks_nesting--;
433 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
Frederic Weisbeckerb6fc6022011-11-28 16:18:56 -0800434 if (rdtp->dynticks_nesting)
435 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
436 else
437 rcu_idle_enter_common(rdtp, oldval);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700438 local_irq_restore(flags);
439}
440
441/*
442 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
443 *
444 * If the new value of the ->dynticks_nesting counter was previously zero,
445 * we really have exited idle, and must do the appropriate accounting.
446 * The caller must have disabled interrupts.
447 */
448static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
449{
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700450 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
451 atomic_inc(&rdtp->dynticks);
452 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
453 smp_mb__after_atomic_inc(); /* See above. */
454 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
Paul E. McKenney7cb92492011-11-28 12:28:34 -0800455 rcu_cleanup_after_idle(smp_processor_id());
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700456 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
Paul E. McKenney99745b62011-11-10 15:48:45 -0800457 if (!is_idle_task(current)) {
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700458 struct task_struct *idle = idle_task(smp_processor_id());
459
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700460 trace_rcu_dyntick("Error on exit: not idle task",
461 oldval, rdtp->dynticks_nesting);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700462 ftrace_dump(DUMP_ALL);
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700463 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
464 current->pid, current->comm,
465 idle->pid, idle->comm); /* must be idle task! */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700466 }
467}
468
469/**
470 * rcu_idle_exit - inform RCU that current CPU is leaving idle
471 *
472 * Exit idle mode, in other words, -enter- the mode in which RCU
473 * read-side critical sections can occur.
474 *
Paul E. McKenney29e37d82011-11-17 16:55:56 -0800475 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700476 * allow for the possibility of usermode upcalls messing up our count
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700477 * of interrupt nesting level during the busy period that is just
478 * now starting.
479 */
480void rcu_idle_exit(void)
481{
482 unsigned long flags;
483 struct rcu_dynticks *rdtp;
484 long long oldval;
485
486 local_irq_save(flags);
487 rdtp = &__get_cpu_var(rcu_dynticks);
488 oldval = rdtp->dynticks_nesting;
Paul E. McKenney29e37d82011-11-17 16:55:56 -0800489 WARN_ON_ONCE(oldval < 0);
490 if (oldval & DYNTICK_TASK_NEST_MASK)
491 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
492 else
493 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700494 rcu_idle_exit_common(rdtp, oldval);
495 local_irq_restore(flags);
496}
Paul E. McKenney8a2ecf42012-02-02 15:42:04 -0800497EXPORT_SYMBOL_GPL(rcu_idle_exit);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700498
499/**
500 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
501 *
502 * Enter an interrupt handler, which might possibly result in exiting
503 * idle mode, in other words, entering the mode in which read-side critical
504 * sections can occur.
505 *
506 * Note that the Linux kernel is fully capable of entering an interrupt
507 * handler that it never exits, for example when doing upcalls to
508 * user mode! This code assumes that the idle loop never does upcalls to
509 * user mode. If your architecture does do upcalls from the idle loop (or
510 * does anything else that results in unbalanced calls to the irq_enter()
511 * and irq_exit() functions), RCU will give you what you deserve, good
512 * and hard. But very infrequently and irreproducibly.
513 *
514 * Use things like work queues to work around this limitation.
515 *
516 * You have been warned.
517 */
518void rcu_irq_enter(void)
519{
520 unsigned long flags;
521 struct rcu_dynticks *rdtp;
522 long long oldval;
523
524 local_irq_save(flags);
525 rdtp = &__get_cpu_var(rcu_dynticks);
526 oldval = rdtp->dynticks_nesting;
527 rdtp->dynticks_nesting++;
528 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
Frederic Weisbeckerb6fc6022011-11-28 16:18:56 -0800529 if (oldval)
530 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
531 else
532 rcu_idle_exit_common(rdtp, oldval);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100533 local_irq_restore(flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100534}
535
536/**
537 * rcu_nmi_enter - inform RCU of entry to NMI context
538 *
539 * If the CPU was idle with dynamic ticks active, and there is no
540 * irq handler running, this updates rdtp->dynticks_nmi to let the
541 * RCU grace-period handling know that the CPU is active.
542 */
543void rcu_nmi_enter(void)
544{
545 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
546
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700547 if (rdtp->dynticks_nmi_nesting == 0 &&
548 (atomic_read(&rdtp->dynticks) & 0x1))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100549 return;
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700550 rdtp->dynticks_nmi_nesting++;
551 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
552 atomic_inc(&rdtp->dynticks);
553 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
554 smp_mb__after_atomic_inc(); /* See above. */
555 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100556}
557
558/**
559 * rcu_nmi_exit - inform RCU of exit from NMI context
560 *
561 * If the CPU was idle with dynamic ticks active, and there is no
562 * irq handler running, this updates rdtp->dynticks_nmi to let the
563 * RCU grace-period handling know that the CPU is no longer active.
564 */
565void rcu_nmi_exit(void)
566{
567 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
568
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700569 if (rdtp->dynticks_nmi_nesting == 0 ||
570 --rdtp->dynticks_nmi_nesting != 0)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100571 return;
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700572 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
573 smp_mb__before_atomic_inc(); /* See above. */
574 atomic_inc(&rdtp->dynticks);
575 smp_mb__after_atomic_inc(); /* Force delay to next write. */
576 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100577}
578
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700579#ifdef CONFIG_PROVE_RCU
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100580
581/**
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700582 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100583 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700584 * If the current CPU is in its idle loop and is neither in an interrupt
Paul E. McKenney34240692011-10-03 11:38:52 -0700585 * or NMI handler, return true.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100586 */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700587int rcu_is_cpu_idle(void)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100588{
Paul E. McKenney34240692011-10-03 11:38:52 -0700589 int ret;
590
591 preempt_disable();
592 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
593 preempt_enable();
594 return ret;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700595}
Frederic Weisbeckere6b80a32011-10-07 16:25:18 -0700596EXPORT_SYMBOL(rcu_is_cpu_idle);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700597
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800598#ifdef CONFIG_HOTPLUG_CPU
599
600/*
601 * Is the current CPU online? Disable preemption to avoid false positives
602 * that could otherwise happen due to the current CPU number being sampled,
603 * this task being preempted, its old CPU being taken offline, resuming
604 * on some other CPU, then determining that its old CPU is now offline.
605 * It is OK to use RCU on an offline processor during initial boot, hence
Paul E. McKenney2036d942012-01-30 17:02:47 -0800606 * the check for rcu_scheduler_fully_active. Note also that it is OK
607 * for a CPU coming online to use RCU for one jiffy prior to marking itself
608 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
609 * offline to continue to use RCU for one jiffy after marking itself
610 * offline in the cpu_online_mask. This leniency is necessary given the
611 * non-atomic nature of the online and offline processing, for example,
612 * the fact that a CPU enters the scheduler after completing the CPU_DYING
613 * notifiers.
614 *
615 * This is also why RCU internally marks CPUs online during the
616 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800617 *
618 * Disable checking if in an NMI handler because we cannot safely report
619 * errors from NMI handlers anyway.
620 */
621bool rcu_lockdep_current_cpu_online(void)
622{
Paul E. McKenney2036d942012-01-30 17:02:47 -0800623 struct rcu_data *rdp;
624 struct rcu_node *rnp;
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800625 bool ret;
626
627 if (in_nmi())
628 return 1;
629 preempt_disable();
Paul E. McKenney2036d942012-01-30 17:02:47 -0800630 rdp = &__get_cpu_var(rcu_sched_data);
631 rnp = rdp->mynode;
632 ret = (rdp->grpmask & rnp->qsmaskinit) ||
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800633 !rcu_scheduler_fully_active;
634 preempt_enable();
635 return ret;
636}
637EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
638
639#endif /* #ifdef CONFIG_HOTPLUG_CPU */
640
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700641#endif /* #ifdef CONFIG_PROVE_RCU */
642
643/**
644 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
645 *
646 * If the current CPU is idle or running at a first-level (not nested)
647 * interrupt from idle, return true. The caller must have at least
648 * disabled preemption.
649 */
650int rcu_is_cpu_rrupt_from_idle(void)
651{
652 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100653}
654
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100655/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100656 * Snapshot the specified CPU's dynticks counter so that we can later
657 * credit them with an implicit quiescent state. Return 1 if this CPU
Paul E. McKenney1eba8f82009-09-23 09:50:42 -0700658 * is in dynticks idle mode, which is an extended quiescent state.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100659 */
660static int dyntick_save_progress_counter(struct rcu_data *rdp)
661{
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700662 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
Paul E. McKenneyf0e7c192011-11-23 15:02:05 -0800663 return (rdp->dynticks_snap & 0x1) == 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100664}
665
666/*
667 * Return true if the specified CPU has passed through a quiescent
668 * state by virtue of being in or having passed through an dynticks
669 * idle state since the last call to dyntick_save_progress_counter()
670 * for this same CPU.
671 */
672static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
673{
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700674 unsigned int curr;
675 unsigned int snap;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100676
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700677 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
678 snap = (unsigned int)rdp->dynticks_snap;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100679
680 /*
681 * If the CPU passed through or entered a dynticks idle phase with
682 * no active irq/NMI handlers, then we can safely pretend that the CPU
683 * already acknowledged the request to pass through a quiescent
684 * state. Either way, that CPU cannot possibly be in an RCU
685 * read-side critical section that started before the beginning
686 * of the current RCU grace period.
687 */
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700688 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700689 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100690 rdp->dynticks_fqs++;
691 return 1;
692 }
693
694 /* Go check for the CPU being offline. */
695 return rcu_implicit_offline_qs(rdp);
696}
697
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800698static int jiffies_till_stall_check(void)
699{
700 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
701
702 /*
703 * Limit check must be consistent with the Kconfig limits
704 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
705 */
706 if (till_stall_check < 3) {
707 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
708 till_stall_check = 3;
709 } else if (till_stall_check > 300) {
710 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
711 till_stall_check = 300;
712 }
713 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
714}
715
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100716static void record_gp_stall_check_time(struct rcu_state *rsp)
717{
718 rsp->gp_start = jiffies;
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800719 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100720}
721
722static void print_other_cpu_stall(struct rcu_state *rsp)
723{
724 int cpu;
725 long delta;
726 unsigned long flags;
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700727 int ndetected;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100728 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100729
730 /* Only let one CPU complain about others per time interval. */
731
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800732 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100733 delta = jiffies - rsp->jiffies_stall;
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -0700734 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800735 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100736 return;
737 }
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800738 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800739 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100740
Paul E. McKenney8cdd32a2010-08-09 14:23:03 -0700741 /*
742 * OK, time to rat on our buddy...
743 * See Documentation/RCU/stallwarn.txt for info on how to debug
744 * RCU CPU stall warnings.
745 */
Paul E. McKenneya858af22012-01-16 13:29:10 -0800746 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
Paul E. McKenney4300aa62010-04-13 16:18:22 -0700747 rsp->name);
Paul E. McKenneya858af22012-01-16 13:29:10 -0800748 print_cpu_stall_info_begin();
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700749 rcu_for_each_leaf_node(rsp, rnp) {
Paul E. McKenney3acd9eb2010-02-22 17:05:03 -0800750 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700751 ndetected += rcu_print_task_stall(rnp);
Paul E. McKenney3acd9eb2010-02-22 17:05:03 -0800752 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700753 if (rnp->qsmask == 0)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100754 continue;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700755 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700756 if (rnp->qsmask & (1UL << cpu)) {
Paul E. McKenneya858af22012-01-16 13:29:10 -0800757 print_cpu_stall_info(rsp, rnp->grplo + cpu);
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700758 ndetected++;
759 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100760 }
Paul E. McKenneya858af22012-01-16 13:29:10 -0800761
762 /*
763 * Now rat on any tasks that got kicked up to the root rcu_node
764 * due to CPU offlining.
765 */
766 rnp = rcu_get_root(rsp);
767 raw_spin_lock_irqsave(&rnp->lock, flags);
768 ndetected = rcu_print_task_stall(rnp);
769 raw_spin_unlock_irqrestore(&rnp->lock, flags);
770
771 print_cpu_stall_info_end();
772 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100773 smp_processor_id(), (long)(jiffies - rsp->gp_start));
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700774 if (ndetected == 0)
775 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
776 else if (!trigger_all_cpu_backtrace())
Paul E. McKenney4627e242011-08-03 03:34:24 -0700777 dump_stack();
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200778
Paul E. McKenney1ed509a2010-02-22 17:05:05 -0800779 /* If so configured, complain about tasks blocking the grace period. */
780
781 rcu_print_detail_task_stall(rsp);
782
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100783 force_quiescent_state(rsp, 0); /* Kick them all. */
784}
785
786static void print_cpu_stall(struct rcu_state *rsp)
787{
788 unsigned long flags;
789 struct rcu_node *rnp = rcu_get_root(rsp);
790
Paul E. McKenney8cdd32a2010-08-09 14:23:03 -0700791 /*
792 * OK, time to rat on ourselves...
793 * See Documentation/RCU/stallwarn.txt for info on how to debug
794 * RCU CPU stall warnings.
795 */
Paul E. McKenneya858af22012-01-16 13:29:10 -0800796 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
797 print_cpu_stall_info_begin();
798 print_cpu_stall_info(rsp, smp_processor_id());
799 print_cpu_stall_info_end();
800 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
Paul E. McKenney4627e242011-08-03 03:34:24 -0700801 if (!trigger_all_cpu_backtrace())
802 dump_stack();
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200803
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800804 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney20133cf2010-02-22 17:05:01 -0800805 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800806 rsp->jiffies_stall = jiffies +
807 3 * jiffies_till_stall_check() + 3;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800808 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200809
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100810 set_need_resched(); /* kick ourselves to get things going. */
811}
812
813static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
814{
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700815 unsigned long j;
816 unsigned long js;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100817 struct rcu_node *rnp;
818
Paul E. McKenney742734e2010-06-30 11:43:52 -0700819 if (rcu_cpu_stall_suppress)
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700820 return;
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700821 j = ACCESS_ONCE(jiffies);
822 js = ACCESS_ONCE(rsp->jiffies_stall);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100823 rnp = rdp->mynode;
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700824 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100825
826 /* We haven't checked in, so go dump stack. */
827 print_cpu_stall(rsp);
828
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700829 } else if (rcu_gp_in_progress(rsp) &&
830 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100831
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700832 /* They had a few time units to dump stack, so complain. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100833 print_other_cpu_stall(rsp);
834 }
835}
836
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700837static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
838{
Paul E. McKenney742734e2010-06-30 11:43:52 -0700839 rcu_cpu_stall_suppress = 1;
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700840 return NOTIFY_DONE;
841}
842
Paul E. McKenney53d84e02010-08-10 14:28:53 -0700843/**
844 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
845 *
846 * Set the stall-warning timeout way off into the future, thus preventing
847 * any RCU CPU stall-warning messages from appearing in the current set of
848 * RCU grace periods.
849 *
850 * The caller must disable hard irqs.
851 */
852void rcu_cpu_stall_reset(void)
853{
854 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
855 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
856 rcu_preempt_stall_reset();
857}
858
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700859static struct notifier_block rcu_panic_block = {
860 .notifier_call = rcu_panic,
861};
862
863static void __init check_cpu_stall_init(void)
864{
865 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
866}
867
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100868/*
869 * Update CPU-local rcu_data state to record the newly noticed grace period.
870 * This is used both when we started the grace period and when we notice
Paul E. McKenney91603062009-11-02 13:52:29 -0800871 * that someone else started the grace period. The caller must hold the
872 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
873 * and must have irqs disabled.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100874 */
Paul E. McKenney91603062009-11-02 13:52:29 -0800875static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
876{
877 if (rdp->gpnum != rnp->gpnum) {
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800878 /*
879 * If the current grace period is waiting for this CPU,
880 * set up to detect a quiescent state, otherwise don't
881 * go looking for one.
882 */
Paul E. McKenney91603062009-11-02 13:52:29 -0800883 rdp->gpnum = rnp->gpnum;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700884 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800885 if (rnp->qsmask & rdp->grpmask) {
886 rdp->qs_pending = 1;
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700887 rdp->passed_quiesce = 0;
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800888 } else
889 rdp->qs_pending = 0;
Paul E. McKenneya858af22012-01-16 13:29:10 -0800890 zero_cpu_stall_ticks(rdp);
Paul E. McKenney91603062009-11-02 13:52:29 -0800891 }
892}
893
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100894static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
895{
Paul E. McKenney91603062009-11-02 13:52:29 -0800896 unsigned long flags;
897 struct rcu_node *rnp;
898
899 local_irq_save(flags);
900 rnp = rdp->mynode;
901 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800902 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
Paul E. McKenney91603062009-11-02 13:52:29 -0800903 local_irq_restore(flags);
904 return;
905 }
906 __note_new_gpnum(rsp, rnp, rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800907 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100908}
909
910/*
911 * Did someone else start a new RCU grace period start since we last
912 * checked? Update local state appropriately if so. Must be called
913 * on the CPU corresponding to rdp.
914 */
915static int
916check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
917{
918 unsigned long flags;
919 int ret = 0;
920
921 local_irq_save(flags);
922 if (rdp->gpnum != rsp->gpnum) {
923 note_new_gpnum(rsp, rdp);
924 ret = 1;
925 }
926 local_irq_restore(flags);
927 return ret;
928}
929
930/*
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800931 * Advance this CPU's callbacks, but only if the current grace period
932 * has ended. This may be called only from the CPU to whom the rdp
933 * belongs. In addition, the corresponding leaf rcu_node structure's
934 * ->lock must be held by the caller, with irqs disabled.
935 */
936static void
937__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
938{
939 /* Did another grace period end? */
940 if (rdp->completed != rnp->completed) {
941
942 /* Advance callbacks. No harm if list empty. */
943 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
944 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
945 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
946
947 /* Remember that we saw this grace-period completion. */
948 rdp->completed = rnp->completed;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700949 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
Frederic Weisbecker20377f32010-12-10 22:11:10 +0100950
951 /*
Frederic Weisbecker5ff8e6f2010-12-10 22:11:11 +0100952 * If we were in an extended quiescent state, we may have
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800953 * missed some grace periods that others CPUs handled on
Frederic Weisbecker5ff8e6f2010-12-10 22:11:11 +0100954 * our behalf. Catch up with this state to avoid noting
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800955 * spurious new grace periods. If another grace period
956 * has started, then rnp->gpnum will have advanced, so
957 * we will detect this later on.
Frederic Weisbecker5ff8e6f2010-12-10 22:11:11 +0100958 */
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800959 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
Frederic Weisbecker5ff8e6f2010-12-10 22:11:11 +0100960 rdp->gpnum = rdp->completed;
961
962 /*
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800963 * If RCU does not need a quiescent state from this CPU,
964 * then make sure that this CPU doesn't go looking for one.
Frederic Weisbecker20377f32010-12-10 22:11:10 +0100965 */
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800966 if ((rnp->qsmask & rdp->grpmask) == 0)
Frederic Weisbecker20377f32010-12-10 22:11:10 +0100967 rdp->qs_pending = 0;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800968 }
969}
970
971/*
972 * Advance this CPU's callbacks, but only if the current grace period
973 * has ended. This may be called only from the CPU to whom the rdp
974 * belongs.
975 */
976static void
977rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
978{
979 unsigned long flags;
980 struct rcu_node *rnp;
981
982 local_irq_save(flags);
983 rnp = rdp->mynode;
984 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800985 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800986 local_irq_restore(flags);
987 return;
988 }
989 __rcu_process_gp_end(rsp, rnp, rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800990 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800991}
992
993/*
994 * Do per-CPU grace-period initialization for running CPU. The caller
995 * must hold the lock of the leaf rcu_node structure corresponding to
996 * this CPU.
997 */
998static void
999rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1000{
1001 /* Prior grace period ended, so advance callbacks for current CPU. */
1002 __rcu_process_gp_end(rsp, rnp, rdp);
1003
1004 /*
1005 * Because this CPU just now started the new grace period, we know
1006 * that all of its callbacks will be covered by this upcoming grace
1007 * period, even the ones that were registered arbitrarily recently.
1008 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1009 *
1010 * Other CPUs cannot be sure exactly when the grace period started.
1011 * Therefore, their recently registered callbacks must pass through
1012 * an additional RCU_NEXT_READY stage, so that they will be handled
1013 * by the next RCU grace period.
1014 */
1015 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1016 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
Paul E. McKenney91603062009-11-02 13:52:29 -08001017
1018 /* Set state so that this CPU will detect the next quiescent state. */
1019 __note_new_gpnum(rsp, rnp, rdp);
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001020}
1021
1022/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001023 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1024 * in preparation for detecting the next grace period. The caller must hold
1025 * the root node's ->lock, which is released before return. Hard irqs must
1026 * be disabled.
Paul E. McKenneye5601402012-01-07 11:03:57 -08001027 *
1028 * Note that it is legal for a dying CPU (which is marked as offline) to
1029 * invoke this function. This can happen when the dying CPU reports its
1030 * quiescent state.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001031 */
1032static void
1033rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1034 __releases(rcu_get_root(rsp)->lock)
1035{
Lai Jiangshan394f99a2010-06-28 16:25:04 +08001036 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001037 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001038
Paul E. McKenney037067a2011-08-07 20:26:31 -07001039 if (!rcu_scheduler_fully_active ||
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001040 !cpu_needs_another_gp(rsp, rdp)) {
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -08001041 /*
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001042 * Either the scheduler hasn't yet spawned the first
1043 * non-idle task or this CPU does not need another
1044 * grace period. Either way, don't start a new grace
1045 * period.
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -08001046 */
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001047 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1048 return;
1049 }
1050
1051 if (rsp->fqs_active) {
1052 /*
1053 * This CPU needs a grace period, but force_quiescent_state()
1054 * is running. Tell it to start one on this CPU's behalf.
1055 */
1056 rsp->fqs_need_gp = 1;
1057 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001058 return;
1059 }
1060
1061 /* Advance to a new grace period and initialize state. */
1062 rsp->gpnum++;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001063 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001064 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1065 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001066 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001067 record_gp_stall_check_time(rsp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001068 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001069
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001070 /* Exclude any concurrent CPU-hotplug operations. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001071 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001072
1073 /*
Paul E. McKenneyb835db12009-09-08 15:54:37 -07001074 * Set the quiescent-state-needed bits in all the rcu_node
1075 * structures for all currently online CPUs in breadth-first
1076 * order, starting from the root rcu_node structure. This
1077 * operation relies on the layout of the hierarchy within the
1078 * rsp->node[] array. Note that other CPUs will access only
1079 * the leaves of the hierarchy, which still indicate that no
1080 * grace period is in progress, at least until the corresponding
1081 * leaf node has been initialized. In addition, we have excluded
1082 * CPU-hotplug operations.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001083 *
1084 * Note that the grace period cannot complete until we finish
1085 * the initialization process, as there will be at least one
1086 * qsmask bit set in the root node until that time, namely the
Paul E. McKenneyb835db12009-09-08 15:54:37 -07001087 * one corresponding to this CPU, due to the fact that we have
1088 * irqs disabled.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001089 */
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001090 rcu_for_each_node_breadth_first(rsp, rnp) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001091 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenneyb0e165c02009-09-13 09:15:09 -07001092 rcu_preempt_check_blocked_tasks(rnp);
Paul E. McKenney49e29122009-09-18 09:50:19 -07001093 rnp->qsmask = rnp->qsmaskinit;
Paul E. McKenneyde078d82009-09-08 15:54:36 -07001094 rnp->gpnum = rsp->gpnum;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001095 rnp->completed = rsp->completed;
1096 if (rnp == rdp->mynode)
1097 rcu_start_gp_per_cpu(rsp, rnp, rdp);
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001098 rcu_preempt_boost_start_gp(rnp);
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001099 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1100 rnp->level, rnp->grplo,
1101 rnp->grphi, rnp->qsmask);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001102 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001103 }
1104
Paul E. McKenney83f5b012009-10-28 08:14:49 -07001105 rnp = rcu_get_root(rsp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001106 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001107 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001108 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1109 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001110}
1111
1112/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001113 * Report a full set of quiescent states to the specified rcu_state
1114 * data structure. This involves cleaning up after the prior grace
1115 * period and letting rcu_start_gp() start up the next grace period
1116 * if one is needed. Note that the caller must hold rnp->lock, as
1117 * required by rcu_start_gp(), which will release it.
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001118 */
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001119static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -07001120 __releases(rcu_get_root(rsp)->lock)
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001121{
Paul E. McKenney15ba0ba2011-04-06 16:01:16 -07001122 unsigned long gp_duration;
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001123 struct rcu_node *rnp = rcu_get_root(rsp);
1124 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
Paul E. McKenney15ba0ba2011-04-06 16:01:16 -07001125
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -07001126 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
Paul E. McKenney0bbcc522011-05-16 02:24:04 -07001127
1128 /*
1129 * Ensure that all grace-period and pre-grace-period activity
1130 * is seen before the assignment to rsp->completed.
1131 */
1132 smp_mb(); /* See above block comment. */
Paul E. McKenney15ba0ba2011-04-06 16:01:16 -07001133 gp_duration = jiffies - rsp->gp_start;
1134 if (gp_duration > rsp->gp_max)
1135 rsp->gp_max = gp_duration;
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001136
1137 /*
1138 * We know the grace period is complete, but to everyone else
1139 * it appears to still be ongoing. But it is also the case
1140 * that to everyone else it looks like there is nothing that
1141 * they can do to advance the grace period. It is therefore
1142 * safe for us to drop the lock in order to mark the grace
1143 * period as completed in all of the rcu_node structures.
1144 *
1145 * But if this CPU needs another grace period, it will take
1146 * care of this while initializing the next grace period.
1147 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1148 * because the callbacks have not yet been advanced: Those
1149 * callbacks are waiting on the grace period that just now
1150 * completed.
1151 */
1152 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1153 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1154
1155 /*
1156 * Propagate new ->completed value to rcu_node structures
1157 * so that other CPUs don't have to wait until the start
1158 * of the next grace period to process their callbacks.
1159 */
1160 rcu_for_each_node_breadth_first(rsp, rnp) {
1161 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1162 rnp->completed = rsp->gpnum;
1163 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1164 }
1165 rnp = rcu_get_root(rsp);
1166 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1167 }
1168
1169 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001170 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001171 rsp->fqs_state = RCU_GP_IDLE;
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001172 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1173}
1174
1175/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001176 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1177 * Allows quiescent states for a group of CPUs to be reported at one go
1178 * to the specified rcu_node structure, though all the CPUs in the group
1179 * must be represented by the same rcu_node structure (which need not be
1180 * a leaf rcu_node structure, though it often will be). That structure's
1181 * lock must be held upon entry, and it is released before return.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001182 */
1183static void
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001184rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1185 struct rcu_node *rnp, unsigned long flags)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001186 __releases(rnp->lock)
1187{
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001188 struct rcu_node *rnp_c;
1189
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001190 /* Walk up the rcu_node hierarchy. */
1191 for (;;) {
1192 if (!(rnp->qsmask & mask)) {
1193
1194 /* Our bit has already been cleared, so done. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001195 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001196 return;
1197 }
1198 rnp->qsmask &= ~mask;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001199 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1200 mask, rnp->qsmask, rnp->level,
1201 rnp->grplo, rnp->grphi,
1202 !!rnp->gp_tasks);
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001203 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001204
1205 /* Other bits still set at this level, so done. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001206 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001207 return;
1208 }
1209 mask = rnp->grpmask;
1210 if (rnp->parent == NULL) {
1211
1212 /* No more levels. Exit loop holding root lock. */
1213
1214 break;
1215 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001216 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001217 rnp_c = rnp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001218 rnp = rnp->parent;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001219 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001220 WARN_ON_ONCE(rnp_c->qsmask);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001221 }
1222
1223 /*
1224 * Get here if we are the last CPU to pass through a quiescent
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001225 * state for this grace period. Invoke rcu_report_qs_rsp()
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001226 * to clean up and start the next grace period if one is needed.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001227 */
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001228 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001229}
1230
1231/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001232 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1233 * structure. This must be either called from the specified CPU, or
1234 * called when the specified CPU is known to be offline (and when it is
1235 * also known that no other CPU is concurrently trying to help the offline
1236 * CPU). The lastcomp argument is used to make sure we are still in the
1237 * grace period of interest. We don't want to end the current grace period
1238 * based on quiescent states detected in an earlier grace period!
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001239 */
1240static void
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001241rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001242{
1243 unsigned long flags;
1244 unsigned long mask;
1245 struct rcu_node *rnp;
1246
1247 rnp = rdp->mynode;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001248 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001249 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001250
1251 /*
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001252 * The grace period in which this quiescent state was
1253 * recorded has ended, so don't report it upwards.
1254 * We will instead need a new quiescent state that lies
1255 * within the current grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001256 */
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001257 rdp->passed_quiesce = 0; /* need qs for new gp. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001258 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001259 return;
1260 }
1261 mask = rdp->grpmask;
1262 if ((rnp->qsmask & mask) == 0) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001263 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001264 } else {
1265 rdp->qs_pending = 0;
1266
1267 /*
1268 * This GP can't end until cpu checks in, so all of our
1269 * callbacks can be processed during the next GP.
1270 */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001271 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1272
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001273 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001274 }
1275}
1276
1277/*
1278 * Check to see if there is a new grace period of which this CPU
1279 * is not yet aware, and if so, set up local rcu_data state for it.
1280 * Otherwise, see if this CPU has just passed through its first
1281 * quiescent state for this grace period, and record that fact if so.
1282 */
1283static void
1284rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1285{
1286 /* If there is now a new grace period, record and return. */
1287 if (check_for_new_grace_period(rsp, rdp))
1288 return;
1289
1290 /*
1291 * Does this CPU still need to do its part for current grace period?
1292 * If no, return and let the other CPUs do their part as well.
1293 */
1294 if (!rdp->qs_pending)
1295 return;
1296
1297 /*
1298 * Was there a quiescent state since the beginning of the grace
1299 * period? If no, then exit and wait for the next call.
1300 */
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001301 if (!rdp->passed_quiesce)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001302 return;
1303
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001304 /*
1305 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1306 * judge of that).
1307 */
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001308 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001309}
1310
1311#ifdef CONFIG_HOTPLUG_CPU
1312
1313/*
Lai Jiangshan29494be2010-10-20 14:13:06 +08001314 * Move a dying CPU's RCU callbacks to online CPU's callback list.
Paul E. McKenneye5601402012-01-07 11:03:57 -08001315 * Also record a quiescent state for this CPU for the current grace period.
1316 * Synchronization and interrupt disabling are not required because
1317 * this function executes in stop_machine() context. Therefore, cleanup
1318 * operations that might block must be done later from the CPU_DEAD
1319 * notifier.
1320 *
1321 * Note that the outgoing CPU's bit has already been cleared in the
1322 * cpu_online_mask. This allows us to randomly pick a callback
1323 * destination from the bits set in that mask.
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001324 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001325static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001326{
1327 int i;
Paul E. McKenneye5601402012-01-07 11:03:57 -08001328 unsigned long mask;
Lai Jiangshan29494be2010-10-20 14:13:06 +08001329 int receive_cpu = cpumask_any(cpu_online_mask);
Lai Jiangshan394f99a2010-06-28 16:25:04 +08001330 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
Lai Jiangshan29494be2010-10-20 14:13:06 +08001331 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
Paul E. McKenney2036d942012-01-30 17:02:47 -08001332 RCU_TRACE(struct rcu_node *rnp = rdp->mynode); /* For dying CPU. */
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001333
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001334 /* First, adjust the counts. */
1335 if (rdp->nxtlist != NULL) {
1336 receive_rdp->qlen_lazy += rdp->qlen_lazy;
1337 receive_rdp->qlen += rdp->qlen;
1338 rdp->qlen_lazy = 0;
1339 rdp->qlen = 0;
1340 }
1341
1342 /*
1343 * Next, move ready-to-invoke callbacks to be invoked on some
1344 * other CPU. These will not be required to pass through another
1345 * grace period: They are done, regardless of CPU.
1346 */
1347 if (rdp->nxtlist != NULL &&
1348 rdp->nxttail[RCU_DONE_TAIL] != &rdp->nxtlist) {
1349 struct rcu_head *oldhead;
1350 struct rcu_head **oldtail;
1351 struct rcu_head **newtail;
1352
1353 oldhead = rdp->nxtlist;
1354 oldtail = receive_rdp->nxttail[RCU_DONE_TAIL];
1355 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1356 *rdp->nxttail[RCU_DONE_TAIL] = *oldtail;
1357 *receive_rdp->nxttail[RCU_DONE_TAIL] = oldhead;
1358 newtail = rdp->nxttail[RCU_DONE_TAIL];
1359 for (i = RCU_DONE_TAIL; i < RCU_NEXT_SIZE; i++) {
1360 if (receive_rdp->nxttail[i] == oldtail)
1361 receive_rdp->nxttail[i] = newtail;
1362 if (rdp->nxttail[i] == newtail)
1363 rdp->nxttail[i] = &rdp->nxtlist;
1364 }
1365 }
1366
1367 /*
1368 * Finally, put the rest of the callbacks at the end of the list.
1369 * The ones that made it partway through get to start over: We
1370 * cannot assume that grace periods are synchronized across CPUs.
1371 * (We could splice RCU_WAIT_TAIL into RCU_NEXT_READY_TAIL, but
1372 * this does not seem compelling. Not yet, anyway.)
1373 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001374 if (rdp->nxtlist != NULL) {
1375 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1376 receive_rdp->nxttail[RCU_NEXT_TAIL] =
1377 rdp->nxttail[RCU_NEXT_TAIL];
Paul E. McKenneye5601402012-01-07 11:03:57 -08001378 receive_rdp->n_cbs_adopted += rdp->qlen;
1379 rdp->n_cbs_orphaned += rdp->qlen;
Lai Jiangshan29494be2010-10-20 14:13:06 +08001380
Paul E. McKenneye5601402012-01-07 11:03:57 -08001381 rdp->nxtlist = NULL;
1382 for (i = 0; i < RCU_NEXT_SIZE; i++)
1383 rdp->nxttail[i] = &rdp->nxtlist;
Paul E. McKenneye5601402012-01-07 11:03:57 -08001384 }
Lai Jiangshan29494be2010-10-20 14:13:06 +08001385
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001386 /*
1387 * Record a quiescent state for the dying CPU. This is safe
1388 * only because we have already cleared out the callbacks.
1389 * (Otherwise, the RCU core might try to schedule the invocation
1390 * of callbacks on this now-offline CPU, which would be bad.)
1391 */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001392 mask = rdp->grpmask; /* rnp->grplo is constant. */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001393 trace_rcu_grace_period(rsp->name,
1394 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1395 "cpuofl");
1396 rcu_report_qs_rdp(smp_processor_id(), rsp, rdp, rsp->gpnum);
1397 /* Note that rcu_report_qs_rdp() might call trace_rcu_grace_period(). */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001398}
1399
1400/*
Paul E. McKenneye5601402012-01-07 11:03:57 -08001401 * The CPU has been completely removed, and some other CPU is reporting
1402 * this fact from process context. Do the remainder of the cleanup.
1403 * There can only be one CPU hotplug operation at a time, so no other
1404 * CPU can be attempting to update rcu_cpu_kthread_task.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001405 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001406static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001407{
Paul E. McKenney2036d942012-01-30 17:02:47 -08001408 unsigned long flags;
1409 unsigned long mask;
1410 int need_report = 0;
Paul E. McKenneye5601402012-01-07 11:03:57 -08001411 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenney2036d942012-01-30 17:02:47 -08001412 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rnp. */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001413
Paul E. McKenney2036d942012-01-30 17:02:47 -08001414 /* Adjust any no-longer-needed kthreads. */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001415 rcu_stop_cpu_kthread(cpu);
1416 rcu_node_kthread_setaffinity(rnp, -1);
Paul E. McKenney2036d942012-01-30 17:02:47 -08001417
1418 /* Remove the dying CPU from the bitmasks in the rcu_node hierarchy. */
1419
1420 /* Exclude any attempts to start a new grace period. */
1421 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1422
1423 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1424 mask = rdp->grpmask; /* rnp->grplo is constant. */
1425 do {
1426 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1427 rnp->qsmaskinit &= ~mask;
1428 if (rnp->qsmaskinit != 0) {
1429 if (rnp != rdp->mynode)
1430 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1431 break;
1432 }
1433 if (rnp == rdp->mynode)
1434 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1435 else
1436 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1437 mask = rnp->grpmask;
1438 rnp = rnp->parent;
1439 } while (rnp != NULL);
1440
1441 /*
1442 * We still hold the leaf rcu_node structure lock here, and
1443 * irqs are still disabled. The reason for this subterfuge is
1444 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1445 * held leads to deadlock.
1446 */
1447 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1448 rnp = rdp->mynode;
1449 if (need_report & RCU_OFL_TASKS_NORM_GP)
1450 rcu_report_unblock_qs_rnp(rnp, flags);
1451 else
1452 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1453 if (need_report & RCU_OFL_TASKS_EXP_GP)
1454 rcu_report_exp_rnp(rsp, rnp, true);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001455}
1456
1457#else /* #ifdef CONFIG_HOTPLUG_CPU */
1458
Paul E. McKenneye5601402012-01-07 11:03:57 -08001459static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001460{
1461}
1462
Paul E. McKenneye5601402012-01-07 11:03:57 -08001463static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001464{
1465}
1466
1467#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1468
1469/*
1470 * Invoke any RCU callbacks that have made it to the end of their grace
1471 * period. Thottle as specified by rdp->blimit.
1472 */
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001473static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001474{
1475 unsigned long flags;
1476 struct rcu_head *next, *list, **tail;
Paul E. McKenney486e2592012-01-06 14:11:30 -08001477 int bl, count, count_lazy;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001478
1479 /* If no callbacks are ready, just return.*/
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001480 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
Paul E. McKenney486e2592012-01-06 14:11:30 -08001481 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
Paul E. McKenney4968c302011-12-07 16:32:40 -08001482 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1483 need_resched(), is_idle_task(current),
1484 rcu_is_callbacks_kthread());
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001485 return;
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001486 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001487
1488 /*
1489 * Extract the list of ready callbacks, disabling to prevent
1490 * races with call_rcu() from interrupt handlers.
1491 */
1492 local_irq_save(flags);
Paul E. McKenney8146c4e22012-01-10 14:23:29 -08001493 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001494 bl = rdp->blimit;
Paul E. McKenney486e2592012-01-06 14:11:30 -08001495 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001496 list = rdp->nxtlist;
1497 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1498 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1499 tail = rdp->nxttail[RCU_DONE_TAIL];
1500 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1501 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1502 rdp->nxttail[count] = &rdp->nxtlist;
1503 local_irq_restore(flags);
1504
1505 /* Invoke callbacks. */
Paul E. McKenney486e2592012-01-06 14:11:30 -08001506 count = count_lazy = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001507 while (list) {
1508 next = list->next;
1509 prefetch(next);
Mathieu Desnoyers551d55a2010-04-17 08:48:42 -04001510 debug_rcu_head_unqueue(list);
Paul E. McKenney486e2592012-01-06 14:11:30 -08001511 if (__rcu_reclaim(rsp->name, list))
1512 count_lazy++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001513 list = next;
Paul E. McKenneydff16722011-11-29 15:57:13 -08001514 /* Stop only if limit reached and CPU has something to do. */
1515 if (++count >= bl &&
1516 (need_resched() ||
1517 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001518 break;
1519 }
1520
1521 local_irq_save(flags);
Paul E. McKenney4968c302011-12-07 16:32:40 -08001522 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1523 is_idle_task(current),
1524 rcu_is_callbacks_kthread());
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001525
1526 /* Update count, and requeue any remaining callbacks. */
Paul E. McKenney486e2592012-01-06 14:11:30 -08001527 rdp->qlen_lazy -= count_lazy;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001528 rdp->qlen -= count;
Paul E. McKenney269dcc12010-09-07 14:23:09 -07001529 rdp->n_cbs_invoked += count;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001530 if (list != NULL) {
1531 *tail = rdp->nxtlist;
1532 rdp->nxtlist = list;
1533 for (count = 0; count < RCU_NEXT_SIZE; count++)
1534 if (&rdp->nxtlist == rdp->nxttail[count])
1535 rdp->nxttail[count] = tail;
1536 else
1537 break;
1538 }
1539
1540 /* Reinstate batch limit if we have worked down the excess. */
1541 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1542 rdp->blimit = blimit;
1543
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001544 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1545 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1546 rdp->qlen_last_fqs_check = 0;
1547 rdp->n_force_qs_snap = rsp->n_force_qs;
1548 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1549 rdp->qlen_last_fqs_check = rdp->qlen;
1550
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001551 local_irq_restore(flags);
1552
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001553 /* Re-invoke RCU core processing if there are callbacks remaining. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001554 if (cpu_has_callbacks_ready_to_invoke(rdp))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001555 invoke_rcu_core();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001556}
1557
1558/*
1559 * Check to see if this CPU is in a non-context-switch quiescent state
1560 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001561 * Also schedule RCU core processing.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001562 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07001563 * This function must be called from hardirq context. It is normally
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001564 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1565 * false, there is no point in invoking rcu_check_callbacks().
1566 */
1567void rcu_check_callbacks(int cpu, int user)
1568{
Paul E. McKenney300df912011-06-18 22:26:31 -07001569 trace_rcu_utilization("Start scheduler-tick");
Paul E. McKenneya858af22012-01-16 13:29:10 -08001570 increment_cpu_stall_ticks();
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07001571 if (user || rcu_is_cpu_rrupt_from_idle()) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001572
1573 /*
1574 * Get here if this CPU took its interrupt from user
1575 * mode or from the idle loop, and if this is not a
1576 * nested interrupt. In this case, the CPU is in
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001577 * a quiescent state, so note it.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001578 *
1579 * No memory barrier is required here because both
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001580 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1581 * variables that other CPUs neither access nor modify,
1582 * at least not while the corresponding CPU is online.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001583 */
1584
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001585 rcu_sched_qs(cpu);
1586 rcu_bh_qs(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001587
1588 } else if (!in_softirq()) {
1589
1590 /*
1591 * Get here if this CPU did not take its interrupt from
1592 * softirq, in other words, if it is not interrupting
1593 * a rcu_bh read-side critical section. This is an _bh
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001594 * critical section, so note it.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001595 */
1596
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001597 rcu_bh_qs(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001598 }
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001599 rcu_preempt_check_callbacks(cpu);
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07001600 if (rcu_pending(cpu))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001601 invoke_rcu_core();
Paul E. McKenney300df912011-06-18 22:26:31 -07001602 trace_rcu_utilization("End scheduler-tick");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001603}
1604
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001605/*
1606 * Scan the leaf rcu_node structures, processing dyntick state for any that
1607 * have not yet encountered a quiescent state, using the function specified.
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001608 * Also initiate boosting for any threads blocked on the root rcu_node.
1609 *
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001610 * The caller must have suppressed start of new grace periods.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001611 */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001612static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001613{
1614 unsigned long bit;
1615 int cpu;
1616 unsigned long flags;
1617 unsigned long mask;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001618 struct rcu_node *rnp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001619
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001620 rcu_for_each_leaf_node(rsp, rnp) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001621 mask = 0;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001622 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001623 if (!rcu_gp_in_progress(rsp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001624 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney0f10dc822010-01-04 15:09:06 -08001625 return;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001626 }
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001627 if (rnp->qsmask == 0) {
Paul E. McKenney1217ed12011-05-04 21:43:49 -07001628 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001629 continue;
1630 }
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001631 cpu = rnp->grplo;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001632 bit = 1;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001633 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
Lai Jiangshan394f99a2010-06-28 16:25:04 +08001634 if ((rnp->qsmask & bit) != 0 &&
1635 f(per_cpu_ptr(rsp->rda, cpu)))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001636 mask |= bit;
1637 }
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001638 if (mask != 0) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001639
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001640 /* rcu_report_qs_rnp() releases rnp->lock. */
1641 rcu_report_qs_rnp(mask, rsp, rnp, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001642 continue;
1643 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001644 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001645 }
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001646 rnp = rcu_get_root(rsp);
Paul E. McKenney1217ed12011-05-04 21:43:49 -07001647 if (rnp->qsmask == 0) {
1648 raw_spin_lock_irqsave(&rnp->lock, flags);
1649 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1650 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001651}
1652
1653/*
1654 * Force quiescent states on reluctant CPUs, and also detect which
1655 * CPUs are in dyntick-idle mode.
1656 */
1657static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1658{
1659 unsigned long flags;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001660 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001661
Paul E. McKenney300df912011-06-18 22:26:31 -07001662 trace_rcu_utilization("Start fqs");
1663 if (!rcu_gp_in_progress(rsp)) {
1664 trace_rcu_utilization("End fqs");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001665 return; /* No grace period in progress, nothing to force. */
Paul E. McKenney300df912011-06-18 22:26:31 -07001666 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001667 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001668 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
Paul E. McKenney300df912011-06-18 22:26:31 -07001669 trace_rcu_utilization("End fqs");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001670 return; /* Someone else is already on the job. */
1671 }
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001672 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001673 goto unlock_fqs_ret; /* no emergency and done recently. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001674 rsp->n_force_qs++;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001675 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001676 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
Paul E. McKenney560d4bc2009-11-13 19:51:38 -08001677 if(!rcu_gp_in_progress(rsp)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001678 rsp->n_force_qs_ngp++;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001679 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001680 goto unlock_fqs_ret; /* no GP in progress, time updated. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001681 }
Paul E. McKenney07079d52010-01-04 15:09:02 -08001682 rsp->fqs_active = 1;
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001683 switch (rsp->fqs_state) {
Paul E. McKenney83f5b012009-10-28 08:14:49 -07001684 case RCU_GP_IDLE:
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001685 case RCU_GP_INIT:
1686
Paul E. McKenney83f5b012009-10-28 08:14:49 -07001687 break; /* grace period idle or initializing, ignore. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001688
1689 case RCU_SAVE_DYNTICK:
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001690 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1691 break; /* So gcc recognizes the dead code. */
1692
Lai Jiangshanf2614142010-03-28 11:15:20 +08001693 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1694
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001695 /* Record dyntick-idle state. */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001696 force_qs_rnp(rsp, dyntick_save_progress_counter);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001697 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001698 if (rcu_gp_in_progress(rsp))
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001699 rsp->fqs_state = RCU_FORCE_QS;
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001700 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001701
1702 case RCU_FORCE_QS:
1703
1704 /* Check dyntick-idle state, send IPI to laggarts. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001705 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001706 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001707
1708 /* Leave state in case more forcing is required. */
1709
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001710 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001711 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001712 }
Paul E. McKenney07079d52010-01-04 15:09:02 -08001713 rsp->fqs_active = 0;
Paul E. McKenney46a1e342010-01-04 15:09:09 -08001714 if (rsp->fqs_need_gp) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001715 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
Paul E. McKenney46a1e342010-01-04 15:09:09 -08001716 rsp->fqs_need_gp = 0;
1717 rcu_start_gp(rsp, flags); /* releases rnp->lock */
Paul E. McKenney300df912011-06-18 22:26:31 -07001718 trace_rcu_utilization("End fqs");
Paul E. McKenney46a1e342010-01-04 15:09:09 -08001719 return;
1720 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001721 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001722unlock_fqs_ret:
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001723 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
Paul E. McKenney300df912011-06-18 22:26:31 -07001724 trace_rcu_utilization("End fqs");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001725}
1726
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001727/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001728 * This does the RCU core processing work for the specified rcu_state
1729 * and rcu_data structures. This may be called only from the CPU to
1730 * whom the rdp belongs.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001731 */
1732static void
1733__rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1734{
1735 unsigned long flags;
1736
Paul E. McKenney2e597552009-08-15 09:53:48 -07001737 WARN_ON_ONCE(rdp->beenonline == 0);
1738
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001739 /*
1740 * If an RCU GP has gone long enough, go check for dyntick
1741 * idle CPUs and, if needed, send resched IPIs.
1742 */
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001743 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001744 force_quiescent_state(rsp, 1);
1745
1746 /*
1747 * Advance callbacks in response to end of earlier grace
1748 * period that some other CPU ended.
1749 */
1750 rcu_process_gp_end(rsp, rdp);
1751
1752 /* Update RCU state based on any recent quiescent states. */
1753 rcu_check_quiescent_state(rsp, rdp);
1754
1755 /* Does this CPU require a not-yet-started grace period? */
1756 if (cpu_needs_another_gp(rsp, rdp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001757 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001758 rcu_start_gp(rsp, flags); /* releases above lock */
1759 }
1760
1761 /* If there are callbacks ready, invoke them. */
Shaohua Li09223372011-06-14 13:26:25 +08001762 if (cpu_has_callbacks_ready_to_invoke(rdp))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001763 invoke_rcu_callbacks(rsp, rdp);
Shaohua Li09223372011-06-14 13:26:25 +08001764}
1765
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001766/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001767 * Do RCU core processing for the current CPU.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001768 */
Shaohua Li09223372011-06-14 13:26:25 +08001769static void rcu_process_callbacks(struct softirq_action *unused)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001770{
Paul E. McKenney300df912011-06-18 22:26:31 -07001771 trace_rcu_utilization("Start RCU core");
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001772 __rcu_process_callbacks(&rcu_sched_state,
1773 &__get_cpu_var(rcu_sched_data));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001774 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001775 rcu_preempt_process_callbacks();
Paul E. McKenney300df912011-06-18 22:26:31 -07001776 trace_rcu_utilization("End RCU core");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001777}
1778
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001779/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001780 * Schedule RCU callback invocation. If the specified type of RCU
1781 * does not support RCU priority boosting, just do a direct call,
1782 * otherwise wake up the per-CPU kernel kthread. Note that because we
1783 * are running on the current CPU with interrupts disabled, the
1784 * rcu_cpu_kthread_task cannot disappear out from under us.
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001785 */
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001786static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001787{
Paul E. McKenneyb0d30412011-07-10 15:57:35 -07001788 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1789 return;
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001790 if (likely(!rsp->boost)) {
1791 rcu_do_batch(rsp, rdp);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001792 return;
1793 }
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001794 invoke_rcu_callbacks_kthread();
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001795}
1796
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001797static void invoke_rcu_core(void)
Shaohua Li09223372011-06-14 13:26:25 +08001798{
1799 raise_softirq(RCU_SOFTIRQ);
1800}
1801
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001802static void
1803__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
Paul E. McKenney486e2592012-01-06 14:11:30 -08001804 struct rcu_state *rsp, bool lazy)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001805{
1806 unsigned long flags;
1807 struct rcu_data *rdp;
1808
Paul E. McKenney0bb7b592012-01-05 14:44:39 -08001809 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
Mathieu Desnoyers551d55a2010-04-17 08:48:42 -04001810 debug_rcu_head_queue(head);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001811 head->func = func;
1812 head->next = NULL;
1813
1814 smp_mb(); /* Ensure RCU update seen before callback registry. */
1815
1816 /*
1817 * Opportunistically note grace-period endings and beginnings.
1818 * Note that we might see a beginning right after we see an
1819 * end, but never vice versa, since this CPU has to pass through
1820 * a quiescent state betweentimes.
1821 */
1822 local_irq_save(flags);
Paul E. McKenneye5601402012-01-07 11:03:57 -08001823 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
Lai Jiangshan394f99a2010-06-28 16:25:04 +08001824 rdp = this_cpu_ptr(rsp->rda);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001825
1826 /* Add the callback to our list. */
1827 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1828 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
Paul E. McKenney2655d572011-04-07 22:47:23 -07001829 rdp->qlen++;
Paul E. McKenney486e2592012-01-06 14:11:30 -08001830 if (lazy)
1831 rdp->qlen_lazy++;
Paul E. McKenney2655d572011-04-07 22:47:23 -07001832
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001833 if (__is_kfree_rcu_offset((unsigned long)func))
1834 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
Paul E. McKenney486e2592012-01-06 14:11:30 -08001835 rdp->qlen_lazy, rdp->qlen);
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001836 else
Paul E. McKenney486e2592012-01-06 14:11:30 -08001837 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001838
Paul E. McKenney2655d572011-04-07 22:47:23 -07001839 /* If interrupts were disabled, don't dive into RCU core. */
1840 if (irqs_disabled_flags(flags)) {
1841 local_irq_restore(flags);
1842 return;
1843 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001844
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001845 /*
1846 * Force the grace period if too many callbacks or too long waiting.
1847 * Enforce hysteresis, and don't invoke force_quiescent_state()
1848 * if some other CPU has recently done so. Also, don't bother
1849 * invoking force_quiescent_state() if the newly enqueued callback
1850 * is the only one waiting for a grace period to complete.
1851 */
Paul E. McKenney2655d572011-04-07 22:47:23 -07001852 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
Paul E. McKenneyb52573d2010-12-14 17:36:02 -08001853
1854 /* Are we ignoring a completed grace period? */
1855 rcu_process_gp_end(rsp, rdp);
1856 check_for_new_grace_period(rsp, rdp);
1857
1858 /* Start a new grace period if one not already started. */
1859 if (!rcu_gp_in_progress(rsp)) {
1860 unsigned long nestflag;
1861 struct rcu_node *rnp_root = rcu_get_root(rsp);
1862
1863 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1864 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1865 } else {
1866 /* Give the grace period a kick. */
1867 rdp->blimit = LONG_MAX;
1868 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1869 *rdp->nxttail[RCU_DONE_TAIL] != head)
1870 force_quiescent_state(rsp, 0);
1871 rdp->n_force_qs_snap = rsp->n_force_qs;
1872 rdp->qlen_last_fqs_check = rdp->qlen;
1873 }
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001874 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001875 force_quiescent_state(rsp, 1);
1876 local_irq_restore(flags);
1877}
1878
1879/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001880 * Queue an RCU-sched callback for invocation after a grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001881 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001882void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001883{
Paul E. McKenney486e2592012-01-06 14:11:30 -08001884 __call_rcu(head, func, &rcu_sched_state, 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001885}
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001886EXPORT_SYMBOL_GPL(call_rcu_sched);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001887
1888/*
Paul E. McKenney486e2592012-01-06 14:11:30 -08001889 * Queue an RCU callback for invocation after a quicker grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001890 */
1891void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1892{
Paul E. McKenney486e2592012-01-06 14:11:30 -08001893 __call_rcu(head, func, &rcu_bh_state, 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001894}
1895EXPORT_SYMBOL_GPL(call_rcu_bh);
1896
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001897/**
1898 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1899 *
1900 * Control will return to the caller some time after a full rcu-sched
1901 * grace period has elapsed, in other words after all currently executing
1902 * rcu-sched read-side critical sections have completed. These read-side
1903 * critical sections are delimited by rcu_read_lock_sched() and
1904 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1905 * local_irq_disable(), and so on may be used in place of
1906 * rcu_read_lock_sched().
1907 *
1908 * This means that all preempt_disable code sequences, including NMI and
1909 * hardware-interrupt handlers, in progress on entry will have completed
1910 * before this primitive returns. However, this does not guarantee that
1911 * softirq handlers will have completed, since in some kernels, these
1912 * handlers can run in process context, and can block.
1913 *
1914 * This primitive provides the guarantees made by the (now removed)
1915 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1916 * guarantees that rcu_read_lock() sections will have completed.
1917 * In "classic RCU", these two guarantees happen to be one and
1918 * the same, but can differ in realtime RCU implementations.
1919 */
1920void synchronize_sched(void)
1921{
Paul E. McKenneyfe15d702012-01-04 13:30:33 -08001922 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
1923 !lock_is_held(&rcu_lock_map) &&
1924 !lock_is_held(&rcu_sched_lock_map),
1925 "Illegal synchronize_sched() in RCU-sched read-side critical section");
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001926 if (rcu_blocking_is_gp())
1927 return;
Paul E. McKenney2c428182011-05-26 22:14:36 -07001928 wait_rcu_gp(call_rcu_sched);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001929}
1930EXPORT_SYMBOL_GPL(synchronize_sched);
1931
1932/**
1933 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1934 *
1935 * Control will return to the caller some time after a full rcu_bh grace
1936 * period has elapsed, in other words after all currently executing rcu_bh
1937 * read-side critical sections have completed. RCU read-side critical
1938 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1939 * and may be nested.
1940 */
1941void synchronize_rcu_bh(void)
1942{
Paul E. McKenneyfe15d702012-01-04 13:30:33 -08001943 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
1944 !lock_is_held(&rcu_lock_map) &&
1945 !lock_is_held(&rcu_sched_lock_map),
1946 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001947 if (rcu_blocking_is_gp())
1948 return;
Paul E. McKenney2c428182011-05-26 22:14:36 -07001949 wait_rcu_gp(call_rcu_bh);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001950}
1951EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1952
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08001953static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
1954static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
1955
1956static int synchronize_sched_expedited_cpu_stop(void *data)
1957{
1958 /*
1959 * There must be a full memory barrier on each affected CPU
1960 * between the time that try_stop_cpus() is called and the
1961 * time that it returns.
1962 *
1963 * In the current initial implementation of cpu_stop, the
1964 * above condition is already met when the control reaches
1965 * this point and the following smp_mb() is not strictly
1966 * necessary. Do smp_mb() anyway for documentation and
1967 * robustness against future implementation changes.
1968 */
1969 smp_mb(); /* See above comment block. */
1970 return 0;
1971}
1972
Paul E. McKenney236fefa2012-01-31 14:00:41 -08001973/**
1974 * synchronize_sched_expedited - Brute-force RCU-sched grace period
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08001975 *
Paul E. McKenney236fefa2012-01-31 14:00:41 -08001976 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
1977 * approach to force the grace period to end quickly. This consumes
1978 * significant time on all CPUs and is unfriendly to real-time workloads,
1979 * so is thus not recommended for any sort of common-case code. In fact,
1980 * if you are using synchronize_sched_expedited() in a loop, please
1981 * restructure your code to batch your updates, and then use a single
1982 * synchronize_sched() instead.
1983 *
1984 * Note that it is illegal to call this function while holding any lock
1985 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
1986 * to call this function from a CPU-hotplug notifier. Failing to observe
1987 * these restriction will result in deadlock.
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08001988 *
1989 * This implementation can be thought of as an application of ticket
1990 * locking to RCU, with sync_sched_expedited_started and
1991 * sync_sched_expedited_done taking on the roles of the halves
1992 * of the ticket-lock word. Each task atomically increments
1993 * sync_sched_expedited_started upon entry, snapshotting the old value,
1994 * then attempts to stop all the CPUs. If this succeeds, then each
1995 * CPU will have executed a context switch, resulting in an RCU-sched
1996 * grace period. We are then done, so we use atomic_cmpxchg() to
1997 * update sync_sched_expedited_done to match our snapshot -- but
1998 * only if someone else has not already advanced past our snapshot.
1999 *
2000 * On the other hand, if try_stop_cpus() fails, we check the value
2001 * of sync_sched_expedited_done. If it has advanced past our
2002 * initial snapshot, then someone else must have forced a grace period
2003 * some time after we took our snapshot. In this case, our work is
2004 * done for us, and we can simply return. Otherwise, we try again,
2005 * but keep our initial snapshot for purposes of checking for someone
2006 * doing our work for us.
2007 *
2008 * If we fail too many times in a row, we fall back to synchronize_sched().
2009 */
2010void synchronize_sched_expedited(void)
2011{
2012 int firstsnap, s, snap, trycount = 0;
2013
2014 /* Note that atomic_inc_return() implies full memory barrier. */
2015 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2016 get_online_cpus();
Hugh Dickins1cc85962012-02-17 13:20:31 -08002017 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002018
2019 /*
2020 * Each pass through the following loop attempts to force a
2021 * context switch on each CPU.
2022 */
2023 while (try_stop_cpus(cpu_online_mask,
2024 synchronize_sched_expedited_cpu_stop,
2025 NULL) == -EAGAIN) {
2026 put_online_cpus();
2027
2028 /* No joy, try again later. Or just synchronize_sched(). */
2029 if (trycount++ < 10)
2030 udelay(trycount * num_online_cpus());
2031 else {
2032 synchronize_sched();
2033 return;
2034 }
2035
2036 /* Check to see if someone else did our work for us. */
2037 s = atomic_read(&sync_sched_expedited_done);
2038 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2039 smp_mb(); /* ensure test happens before caller kfree */
2040 return;
2041 }
2042
2043 /*
2044 * Refetching sync_sched_expedited_started allows later
2045 * callers to piggyback on our grace period. We subtract
2046 * 1 to get the same token that the last incrementer got.
2047 * We retry after they started, so our grace period works
2048 * for them, and they started after our first try, so their
2049 * grace period works for us.
2050 */
2051 get_online_cpus();
2052 snap = atomic_read(&sync_sched_expedited_started);
2053 smp_mb(); /* ensure read is before try_stop_cpus(). */
2054 }
2055
2056 /*
2057 * Everyone up to our most recent fetch is covered by our grace
2058 * period. Update the counter, but only if our work is still
2059 * relevant -- which it won't be if someone who started later
2060 * than we did beat us to the punch.
2061 */
2062 do {
2063 s = atomic_read(&sync_sched_expedited_done);
2064 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2065 smp_mb(); /* ensure test happens before caller kfree */
2066 break;
2067 }
2068 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2069
2070 put_online_cpus();
2071}
2072EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2073
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002074/*
2075 * Check to see if there is any immediate RCU-related work to be done
2076 * by the current CPU, for the specified type of RCU, returning 1 if so.
2077 * The checks are in order of increasing expense: checks that can be
2078 * carried out against CPU-local state are performed first. However,
2079 * we must check for CPU stalls first, else we might not get a chance.
2080 */
2081static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2082{
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002083 struct rcu_node *rnp = rdp->mynode;
2084
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002085 rdp->n_rcu_pending++;
2086
2087 /* Check for CPU stalls, if enabled. */
2088 check_cpu_stall(rsp, rdp);
2089
2090 /* Is the RCU core waiting for a quiescent state from this CPU? */
Paul E. McKenney5c51dd72011-08-04 06:59:03 -07002091 if (rcu_scheduler_fully_active &&
2092 rdp->qs_pending && !rdp->passed_quiesce) {
Paul E. McKenneyd25eb942010-03-18 21:36:51 -07002093
2094 /*
2095 * If force_quiescent_state() coming soon and this CPU
2096 * needs a quiescent state, and this is either RCU-sched
2097 * or RCU-bh, force a local reschedule.
2098 */
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07002099 rdp->n_rp_qs_pending++;
Paul E. McKenney6cc68792011-03-02 13:15:15 -08002100 if (!rdp->preemptible &&
Paul E. McKenneyd25eb942010-03-18 21:36:51 -07002101 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2102 jiffies))
2103 set_need_resched();
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07002104 } else if (rdp->qs_pending && rdp->passed_quiesce) {
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07002105 rdp->n_rp_report_qs++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002106 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002107 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002108
2109 /* Does this CPU have callbacks ready to invoke? */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002110 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2111 rdp->n_rp_cb_ready++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002112 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002113 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002114
2115 /* Has RCU gone idle with this CPU needing another grace period? */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002116 if (cpu_needs_another_gp(rsp, rdp)) {
2117 rdp->n_rp_cpu_needs_gp++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002118 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002119 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002120
2121 /* Has another RCU grace period completed? */
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002122 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002123 rdp->n_rp_gp_completed++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002124 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002125 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002126
2127 /* Has a new RCU grace period started? */
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002128 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002129 rdp->n_rp_gp_started++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002130 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002131 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002132
2133 /* Has an RCU GP gone long enough to send resched IPIs &c? */
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -07002134 if (rcu_gp_in_progress(rsp) &&
Paul E. McKenney20133cf2010-02-22 17:05:01 -08002135 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002136 rdp->n_rp_need_fqs++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002137 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002138 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002139
2140 /* nothing to do */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002141 rdp->n_rp_need_nothing++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002142 return 0;
2143}
2144
2145/*
2146 * Check to see if there is any immediate RCU-related work to be done
2147 * by the current CPU, returning 1 if so. This function is part of the
2148 * RCU implementation; it is -not- an exported member of the RCU API.
2149 */
Paul E. McKenneya1572292009-08-22 13:56:51 -07002150static int rcu_pending(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002151{
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002152 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002153 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
2154 rcu_preempt_pending(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002155}
2156
2157/*
2158 * Check to see if any future RCU-related work will need to be done
2159 * by the current CPU, even if none need be done immediately, returning
Paul E. McKenney8bd93a22010-02-22 17:04:59 -08002160 * 1 if so.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002161 */
Paul E. McKenneyaea1b352011-11-02 06:54:54 -07002162static int rcu_cpu_has_callbacks(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002163{
2164 /* RCU callbacks either ready or pending? */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002165 return per_cpu(rcu_sched_data, cpu).nxtlist ||
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002166 per_cpu(rcu_bh_data, cpu).nxtlist ||
Paul E. McKenney30fbcc92012-01-12 11:01:14 -08002167 rcu_preempt_cpu_has_callbacks(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002168}
2169
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002170static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
2171static atomic_t rcu_barrier_cpu_count;
2172static DEFINE_MUTEX(rcu_barrier_mutex);
2173static struct completion rcu_barrier_completion;
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002174
2175static void rcu_barrier_callback(struct rcu_head *notused)
2176{
2177 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2178 complete(&rcu_barrier_completion);
2179}
2180
2181/*
2182 * Called with preemption disabled, and from cross-cpu IRQ context.
2183 */
2184static void rcu_barrier_func(void *type)
2185{
2186 int cpu = smp_processor_id();
2187 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
2188 void (*call_rcu_func)(struct rcu_head *head,
2189 void (*func)(struct rcu_head *head));
2190
2191 atomic_inc(&rcu_barrier_cpu_count);
2192 call_rcu_func = type;
2193 call_rcu_func(head, rcu_barrier_callback);
2194}
2195
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002196/*
2197 * Orchestrate the specified type of RCU barrier, waiting for all
2198 * RCU callbacks of the specified type to complete.
2199 */
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002200static void _rcu_barrier(struct rcu_state *rsp,
2201 void (*call_rcu_func)(struct rcu_head *head,
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002202 void (*func)(struct rcu_head *head)))
2203{
2204 BUG_ON(in_interrupt());
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002205 /* Take mutex to serialize concurrent rcu_barrier() requests. */
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002206 mutex_lock(&rcu_barrier_mutex);
2207 init_completion(&rcu_barrier_completion);
2208 /*
2209 * Initialize rcu_barrier_cpu_count to 1, then invoke
2210 * rcu_barrier_func() on each CPU, so that each CPU also has
2211 * incremented rcu_barrier_cpu_count. Only then is it safe to
2212 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2213 * might complete its grace period before all of the other CPUs
2214 * did their increment, causing this function to return too
Paul E. McKenney2d999e02010-10-20 12:06:18 -07002215 * early. Note that on_each_cpu() disables irqs, which prevents
2216 * any CPUs from coming online or going offline until each online
2217 * CPU has queued its RCU-barrier callback.
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002218 */
2219 atomic_set(&rcu_barrier_cpu_count, 1);
2220 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
2221 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2222 complete(&rcu_barrier_completion);
2223 wait_for_completion(&rcu_barrier_completion);
2224 mutex_unlock(&rcu_barrier_mutex);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002225}
2226
2227/**
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002228 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2229 */
2230void rcu_barrier_bh(void)
2231{
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002232 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002233}
2234EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2235
2236/**
2237 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2238 */
2239void rcu_barrier_sched(void)
2240{
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002241 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002242}
2243EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2244
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002245/*
Paul E. McKenney27569622009-08-15 09:53:46 -07002246 * Do boot-time initialization of a CPU's per-CPU RCU data.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002247 */
Paul E. McKenney27569622009-08-15 09:53:46 -07002248static void __init
2249rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002250{
2251 unsigned long flags;
2252 int i;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002253 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenney27569622009-08-15 09:53:46 -07002254 struct rcu_node *rnp = rcu_get_root(rsp);
2255
2256 /* Set up local state, ensuring consistent view of global state. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002257 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney27569622009-08-15 09:53:46 -07002258 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2259 rdp->nxtlist = NULL;
2260 for (i = 0; i < RCU_NEXT_SIZE; i++)
2261 rdp->nxttail[i] = &rdp->nxtlist;
Paul E. McKenney486e2592012-01-06 14:11:30 -08002262 rdp->qlen_lazy = 0;
Paul E. McKenney27569622009-08-15 09:53:46 -07002263 rdp->qlen = 0;
Paul E. McKenney27569622009-08-15 09:53:46 -07002264 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
Paul E. McKenney29e37d82011-11-17 16:55:56 -08002265 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07002266 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
Paul E. McKenney27569622009-08-15 09:53:46 -07002267 rdp->cpu = cpu;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07002268 rdp->rsp = rsp;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002269 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney27569622009-08-15 09:53:46 -07002270}
2271
2272/*
2273 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2274 * offline event can be happening at a given time. Note also that we
2275 * can accept some slop in the rsp->completed access due to the fact
2276 * that this CPU cannot possibly have any RCU callbacks in flight yet.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002277 */
2278static void __cpuinit
Paul E. McKenney6cc68792011-03-02 13:15:15 -08002279rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002280{
2281 unsigned long flags;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002282 unsigned long mask;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002283 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002284 struct rcu_node *rnp = rcu_get_root(rsp);
2285
2286 /* Set up local state, ensuring consistent view of global state. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002287 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002288 rdp->beenonline = 1; /* We have now been online. */
Paul E. McKenney6cc68792011-03-02 13:15:15 -08002289 rdp->preemptible = preemptible;
Paul E. McKenney37c72e52009-10-14 10:15:55 -07002290 rdp->qlen_last_fqs_check = 0;
2291 rdp->n_force_qs_snap = rsp->n_force_qs;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002292 rdp->blimit = blimit;
Paul E. McKenney29e37d82011-11-17 16:55:56 -08002293 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
Paul E. McKenneyc92b1312011-11-23 13:38:58 -08002294 atomic_set(&rdp->dynticks->dynticks,
2295 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
Paul E. McKenney7cb92492011-11-28 12:28:34 -08002296 rcu_prepare_for_idle_init(cpu);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002297 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002298
2299 /*
2300 * A new grace period might start here. If so, we won't be part
2301 * of it, but that is OK, as we are currently in a quiescent state.
2302 */
2303
2304 /* Exclude any attempts to start a new GP on large systems. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002305 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002306
2307 /* Add CPU to rcu_node bitmasks. */
2308 rnp = rdp->mynode;
2309 mask = rdp->grpmask;
2310 do {
2311 /* Exclude any attempts to start a new GP on small systems. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002312 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002313 rnp->qsmaskinit |= mask;
2314 mask = rnp->grpmask;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002315 if (rnp == rdp->mynode) {
Paul E. McKenney06ae115a2011-08-14 15:56:54 -07002316 /*
2317 * If there is a grace period in progress, we will
2318 * set up to wait for it next time we run the
2319 * RCU core code.
2320 */
2321 rdp->gpnum = rnp->completed;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002322 rdp->completed = rnp->completed;
Paul E. McKenney06ae115a2011-08-14 15:56:54 -07002323 rdp->passed_quiesce = 0;
2324 rdp->qs_pending = 0;
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07002325 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07002326 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002327 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002328 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002329 rnp = rnp->parent;
2330 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2331
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002332 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002333}
2334
Peter Zijlstrad72bce02011-05-30 13:34:51 +02002335static void __cpuinit rcu_prepare_cpu(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002336{
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002337 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2338 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2339 rcu_preempt_init_percpu_data(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002340}
2341
2342/*
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002343 * Handle CPU online/offline notification events.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002344 */
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08002345static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2346 unsigned long action, void *hcpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002347{
2348 long cpu = (long)hcpu;
Paul E. McKenney27f4d282011-02-07 12:47:15 -08002349 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002350 struct rcu_node *rnp = rdp->mynode;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002351
Paul E. McKenney300df912011-06-18 22:26:31 -07002352 trace_rcu_utilization("Start CPU hotplug");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002353 switch (action) {
2354 case CPU_UP_PREPARE:
2355 case CPU_UP_PREPARE_FROZEN:
Peter Zijlstrad72bce02011-05-30 13:34:51 +02002356 rcu_prepare_cpu(cpu);
2357 rcu_prepare_kthreads(cpu);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002358 break;
2359 case CPU_ONLINE:
Paul E. McKenney0f962a52011-04-14 12:13:53 -07002360 case CPU_DOWN_FAILED:
2361 rcu_node_kthread_setaffinity(rnp, -1);
Paul E. McKenneye3995a22011-04-18 15:31:26 -07002362 rcu_cpu_kthread_setrt(cpu, 1);
Paul E. McKenney0f962a52011-04-14 12:13:53 -07002363 break;
2364 case CPU_DOWN_PREPARE:
2365 rcu_node_kthread_setaffinity(rnp, cpu);
Paul E. McKenneye3995a22011-04-18 15:31:26 -07002366 rcu_cpu_kthread_setrt(cpu, 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002367 break;
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002368 case CPU_DYING:
2369 case CPU_DYING_FROZEN:
2370 /*
Paul E. McKenney2d999e02010-10-20 12:06:18 -07002371 * The whole machine is "stopped" except this CPU, so we can
2372 * touch any data without introducing corruption. We send the
2373 * dying CPU's callbacks to an arbitrarily chosen online CPU.
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002374 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08002375 rcu_cleanup_dying_cpu(&rcu_bh_state);
2376 rcu_cleanup_dying_cpu(&rcu_sched_state);
2377 rcu_preempt_cleanup_dying_cpu();
Paul E. McKenney7cb92492011-11-28 12:28:34 -08002378 rcu_cleanup_after_idle(cpu);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002379 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002380 case CPU_DEAD:
2381 case CPU_DEAD_FROZEN:
2382 case CPU_UP_CANCELED:
2383 case CPU_UP_CANCELED_FROZEN:
Paul E. McKenneye5601402012-01-07 11:03:57 -08002384 rcu_cleanup_dead_cpu(cpu, &rcu_bh_state);
2385 rcu_cleanup_dead_cpu(cpu, &rcu_sched_state);
2386 rcu_preempt_cleanup_dead_cpu(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002387 break;
2388 default:
2389 break;
2390 }
Paul E. McKenney300df912011-06-18 22:26:31 -07002391 trace_rcu_utilization("End CPU hotplug");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002392 return NOTIFY_OK;
2393}
2394
2395/*
Paul E. McKenneybbad9372010-04-02 16:17:17 -07002396 * This function is invoked towards the end of the scheduler's initialization
2397 * process. Before this is called, the idle task might contain
2398 * RCU read-side critical sections (during which time, this idle
2399 * task is booting the system). After this function is called, the
2400 * idle tasks are prohibited from containing RCU read-side critical
2401 * sections. This function also enables RCU lockdep checking.
2402 */
2403void rcu_scheduler_starting(void)
2404{
2405 WARN_ON(num_online_cpus() != 1);
2406 WARN_ON(nr_context_switches() > 0);
2407 rcu_scheduler_active = 1;
2408}
2409
2410/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002411 * Compute the per-level fanout, either using the exact fanout specified
2412 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2413 */
2414#ifdef CONFIG_RCU_FANOUT_EXACT
2415static void __init rcu_init_levelspread(struct rcu_state *rsp)
2416{
2417 int i;
2418
Paul E. McKenney0209f642010-12-14 16:07:52 -08002419 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002420 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
Paul E. McKenney0209f642010-12-14 16:07:52 -08002421 rsp->levelspread[0] = RCU_FANOUT_LEAF;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002422}
2423#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2424static void __init rcu_init_levelspread(struct rcu_state *rsp)
2425{
2426 int ccur;
2427 int cprv;
2428 int i;
2429
2430 cprv = NR_CPUS;
2431 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2432 ccur = rsp->levelcnt[i];
2433 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2434 cprv = ccur;
2435 }
2436}
2437#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2438
2439/*
2440 * Helper function for rcu_init() that initializes one rcu_state structure.
2441 */
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002442static void __init rcu_init_one(struct rcu_state *rsp,
2443 struct rcu_data __percpu *rda)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002444{
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08002445 static char *buf[] = { "rcu_node_level_0",
2446 "rcu_node_level_1",
2447 "rcu_node_level_2",
2448 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002449 int cpustride = 1;
2450 int i;
2451 int j;
2452 struct rcu_node *rnp;
2453
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08002454 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2455
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002456 /* Initialize the level-tracking arrays. */
2457
2458 for (i = 1; i < NUM_RCU_LVLS; i++)
2459 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2460 rcu_init_levelspread(rsp);
2461
2462 /* Initialize the elements themselves, starting from the leaves. */
2463
2464 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2465 cpustride *= rsp->levelspread[i];
2466 rnp = rsp->level[i];
2467 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002468 raw_spin_lock_init(&rnp->lock);
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08002469 lockdep_set_class_and_name(&rnp->lock,
2470 &rcu_node_class[i], buf[i]);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002471 rnp->gpnum = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002472 rnp->qsmask = 0;
2473 rnp->qsmaskinit = 0;
2474 rnp->grplo = j * cpustride;
2475 rnp->grphi = (j + 1) * cpustride - 1;
2476 if (rnp->grphi >= NR_CPUS)
2477 rnp->grphi = NR_CPUS - 1;
2478 if (i == 0) {
2479 rnp->grpnum = 0;
2480 rnp->grpmask = 0;
2481 rnp->parent = NULL;
2482 } else {
2483 rnp->grpnum = j % rsp->levelspread[i - 1];
2484 rnp->grpmask = 1UL << rnp->grpnum;
2485 rnp->parent = rsp->level[i - 1] +
2486 j / rsp->levelspread[i - 1];
2487 }
2488 rnp->level = i;
Paul E. McKenney12f5f522010-11-29 21:56:39 -08002489 INIT_LIST_HEAD(&rnp->blkd_tasks);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002490 }
2491 }
Lai Jiangshan0c340292010-03-28 11:12:30 +08002492
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002493 rsp->rda = rda;
Lai Jiangshan0c340292010-03-28 11:12:30 +08002494 rnp = rsp->level[NUM_RCU_LVLS - 1];
2495 for_each_possible_cpu(i) {
Paul E. McKenney4a90a062010-04-14 16:48:11 -07002496 while (i > rnp->grphi)
Lai Jiangshan0c340292010-03-28 11:12:30 +08002497 rnp++;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002498 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
Lai Jiangshan0c340292010-03-28 11:12:30 +08002499 rcu_boot_init_percpu_data(i, rsp);
2500 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002501}
2502
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08002503void __init rcu_init(void)
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002504{
Paul E. McKenney017c4262010-01-14 16:10:58 -08002505 int cpu;
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08002506
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002507 rcu_bootup_announce();
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002508 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2509 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002510 __rcu_init_preempt();
Shaohua Li09223372011-06-14 13:26:25 +08002511 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08002512
2513 /*
2514 * We don't need protection against CPU-hotplug here because
2515 * this is called early in boot, before either interrupts
2516 * or the scheduler are operational.
2517 */
2518 cpu_notifier(rcu_cpu_notify, 0);
Paul E. McKenney017c4262010-01-14 16:10:58 -08002519 for_each_online_cpu(cpu)
2520 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Paul E. McKenneyc68de202010-04-15 10:12:40 -07002521 check_cpu_stall_init();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002522}
2523
Paul E. McKenney1eba8f82009-09-23 09:50:42 -07002524#include "rcutree_plugin.h"