blob: df0e3c1bb68e9c814bb149ecdf02ac939b788377 [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. McKenney4145fa72011-10-31 15:01:54 -0700201 .dynticks_nesting = DYNTICK_TASK_NESTING,
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. McKenney9b2e4f12011-09-30 12:10:22 -0700397 rdtp->dynticks_nesting = 0;
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700398 rcu_idle_enter_common(rdtp, oldval);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700399 local_irq_restore(flags);
400}
401
402/**
403 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
404 *
405 * Exit from an interrupt handler, which might possibly result in entering
406 * idle mode, in other words, leaving the mode in which read-side critical
407 * sections can occur.
408 *
409 * This code assumes that the idle loop never does anything that might
410 * result in unbalanced calls to irq_enter() and irq_exit(). If your
411 * architecture violates this assumption, RCU will give you what you
412 * deserve, good and hard. But very infrequently and irreproducibly.
413 *
414 * Use things like work queues to work around this limitation.
415 *
416 * You have been warned.
417 */
418void rcu_irq_exit(void)
419{
420 unsigned long flags;
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700421 long long oldval;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700422 struct rcu_dynticks *rdtp;
423
424 local_irq_save(flags);
425 rdtp = &__get_cpu_var(rcu_dynticks);
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700426 oldval = rdtp->dynticks_nesting;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700427 rdtp->dynticks_nesting--;
428 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
Frederic Weisbeckerb6fc6022011-11-28 16:18:56 -0800429 if (rdtp->dynticks_nesting)
430 trace_rcu_dyntick("--=", oldval, rdtp->dynticks_nesting);
431 else
432 rcu_idle_enter_common(rdtp, oldval);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700433 local_irq_restore(flags);
434}
435
436/*
437 * rcu_idle_exit_common - inform RCU that current CPU is moving away from idle
438 *
439 * If the new value of the ->dynticks_nesting counter was previously zero,
440 * we really have exited idle, and must do the appropriate accounting.
441 * The caller must have disabled interrupts.
442 */
443static void rcu_idle_exit_common(struct rcu_dynticks *rdtp, long long oldval)
444{
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700445 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
446 atomic_inc(&rdtp->dynticks);
447 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
448 smp_mb__after_atomic_inc(); /* See above. */
449 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
Paul E. McKenney7cb92492011-11-28 12:28:34 -0800450 rcu_cleanup_after_idle(smp_processor_id());
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700451 trace_rcu_dyntick("End", oldval, rdtp->dynticks_nesting);
Paul E. McKenney99745b62011-11-10 15:48:45 -0800452 if (!is_idle_task(current)) {
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700453 struct task_struct *idle = idle_task(smp_processor_id());
454
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700455 trace_rcu_dyntick("Error on exit: not idle task",
456 oldval, rdtp->dynticks_nesting);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700457 ftrace_dump(DUMP_ALL);
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700458 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
459 current->pid, current->comm,
460 idle->pid, idle->comm); /* must be idle task! */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700461 }
462}
463
464/**
465 * rcu_idle_exit - inform RCU that current CPU is leaving idle
466 *
467 * Exit idle mode, in other words, -enter- the mode in which RCU
468 * read-side critical sections can occur.
469 *
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700470 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NESTING to
471 * allow for the possibility of usermode upcalls messing up our count
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700472 * of interrupt nesting level during the busy period that is just
473 * now starting.
474 */
475void rcu_idle_exit(void)
476{
477 unsigned long flags;
478 struct rcu_dynticks *rdtp;
479 long long oldval;
480
481 local_irq_save(flags);
482 rdtp = &__get_cpu_var(rcu_dynticks);
483 oldval = rdtp->dynticks_nesting;
484 WARN_ON_ONCE(oldval != 0);
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700485 rdtp->dynticks_nesting = DYNTICK_TASK_NESTING;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700486 rcu_idle_exit_common(rdtp, oldval);
487 local_irq_restore(flags);
488}
489
490/**
491 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
492 *
493 * Enter an interrupt handler, which might possibly result in exiting
494 * idle mode, in other words, entering the mode in which read-side critical
495 * sections can occur.
496 *
497 * Note that the Linux kernel is fully capable of entering an interrupt
498 * handler that it never exits, for example when doing upcalls to
499 * user mode! This code assumes that the idle loop never does upcalls to
500 * user mode. If your architecture does do upcalls from the idle loop (or
501 * does anything else that results in unbalanced calls to the irq_enter()
502 * and irq_exit() functions), RCU will give you what you deserve, good
503 * and hard. But very infrequently and irreproducibly.
504 *
505 * Use things like work queues to work around this limitation.
506 *
507 * You have been warned.
508 */
509void rcu_irq_enter(void)
510{
511 unsigned long flags;
512 struct rcu_dynticks *rdtp;
513 long long oldval;
514
515 local_irq_save(flags);
516 rdtp = &__get_cpu_var(rcu_dynticks);
517 oldval = rdtp->dynticks_nesting;
518 rdtp->dynticks_nesting++;
519 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
Frederic Weisbeckerb6fc6022011-11-28 16:18:56 -0800520 if (oldval)
521 trace_rcu_dyntick("++=", oldval, rdtp->dynticks_nesting);
522 else
523 rcu_idle_exit_common(rdtp, oldval);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100524 local_irq_restore(flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100525}
526
527/**
528 * rcu_nmi_enter - inform RCU of entry to NMI context
529 *
530 * If the CPU was idle with dynamic ticks active, and there is no
531 * irq handler running, this updates rdtp->dynticks_nmi to let the
532 * RCU grace-period handling know that the CPU is active.
533 */
534void rcu_nmi_enter(void)
535{
536 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
537
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700538 if (rdtp->dynticks_nmi_nesting == 0 &&
539 (atomic_read(&rdtp->dynticks) & 0x1))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100540 return;
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700541 rdtp->dynticks_nmi_nesting++;
542 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
543 atomic_inc(&rdtp->dynticks);
544 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
545 smp_mb__after_atomic_inc(); /* See above. */
546 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100547}
548
549/**
550 * rcu_nmi_exit - inform RCU of exit from NMI context
551 *
552 * If the CPU was idle with dynamic ticks active, and there is no
553 * irq handler running, this updates rdtp->dynticks_nmi to let the
554 * RCU grace-period handling know that the CPU is no longer active.
555 */
556void rcu_nmi_exit(void)
557{
558 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
559
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700560 if (rdtp->dynticks_nmi_nesting == 0 ||
561 --rdtp->dynticks_nmi_nesting != 0)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100562 return;
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700563 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
564 smp_mb__before_atomic_inc(); /* See above. */
565 atomic_inc(&rdtp->dynticks);
566 smp_mb__after_atomic_inc(); /* Force delay to next write. */
567 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100568}
569
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700570#ifdef CONFIG_PROVE_RCU
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100571
572/**
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700573 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100574 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700575 * If the current CPU is in its idle loop and is neither in an interrupt
Paul E. McKenney34240692011-10-03 11:38:52 -0700576 * or NMI handler, return true.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100577 */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700578int rcu_is_cpu_idle(void)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100579{
Paul E. McKenney34240692011-10-03 11:38:52 -0700580 int ret;
581
582 preempt_disable();
583 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
584 preempt_enable();
585 return ret;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700586}
Frederic Weisbeckere6b80a32011-10-07 16:25:18 -0700587EXPORT_SYMBOL(rcu_is_cpu_idle);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700588
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800589#ifdef CONFIG_HOTPLUG_CPU
590
591/*
592 * Is the current CPU online? Disable preemption to avoid false positives
593 * that could otherwise happen due to the current CPU number being sampled,
594 * this task being preempted, its old CPU being taken offline, resuming
595 * on some other CPU, then determining that its old CPU is now offline.
596 * It is OK to use RCU on an offline processor during initial boot, hence
Paul E. McKenney2036d942012-01-30 17:02:47 -0800597 * the check for rcu_scheduler_fully_active. Note also that it is OK
598 * for a CPU coming online to use RCU for one jiffy prior to marking itself
599 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
600 * offline to continue to use RCU for one jiffy after marking itself
601 * offline in the cpu_online_mask. This leniency is necessary given the
602 * non-atomic nature of the online and offline processing, for example,
603 * the fact that a CPU enters the scheduler after completing the CPU_DYING
604 * notifiers.
605 *
606 * This is also why RCU internally marks CPUs online during the
607 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800608 *
609 * Disable checking if in an NMI handler because we cannot safely report
610 * errors from NMI handlers anyway.
611 */
612bool rcu_lockdep_current_cpu_online(void)
613{
Paul E. McKenney2036d942012-01-30 17:02:47 -0800614 struct rcu_data *rdp;
615 struct rcu_node *rnp;
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800616 bool ret;
617
618 if (in_nmi())
619 return 1;
620 preempt_disable();
Paul E. McKenney2036d942012-01-30 17:02:47 -0800621 rdp = &__get_cpu_var(rcu_sched_data);
622 rnp = rdp->mynode;
623 ret = (rdp->grpmask & rnp->qsmaskinit) ||
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800624 !rcu_scheduler_fully_active;
625 preempt_enable();
626 return ret;
627}
628EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
629
630#endif /* #ifdef CONFIG_HOTPLUG_CPU */
631
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700632#endif /* #ifdef CONFIG_PROVE_RCU */
633
634/**
635 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
636 *
637 * If the current CPU is idle or running at a first-level (not nested)
638 * interrupt from idle, return true. The caller must have at least
639 * disabled preemption.
640 */
641int rcu_is_cpu_rrupt_from_idle(void)
642{
643 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100644}
645
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100646/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100647 * Snapshot the specified CPU's dynticks counter so that we can later
648 * credit them with an implicit quiescent state. Return 1 if this CPU
Paul E. McKenney1eba8f82009-09-23 09:50:42 -0700649 * is in dynticks idle mode, which is an extended quiescent state.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100650 */
651static int dyntick_save_progress_counter(struct rcu_data *rdp)
652{
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700653 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
Paul E. McKenneyf0e7c192011-11-23 15:02:05 -0800654 return (rdp->dynticks_snap & 0x1) == 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100655}
656
657/*
658 * Return true if the specified CPU has passed through a quiescent
659 * state by virtue of being in or having passed through an dynticks
660 * idle state since the last call to dyntick_save_progress_counter()
661 * for this same CPU.
662 */
663static int rcu_implicit_dynticks_qs(struct rcu_data *rdp)
664{
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700665 unsigned int curr;
666 unsigned int snap;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100667
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700668 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
669 snap = (unsigned int)rdp->dynticks_snap;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100670
671 /*
672 * If the CPU passed through or entered a dynticks idle phase with
673 * no active irq/NMI handlers, then we can safely pretend that the CPU
674 * already acknowledged the request to pass through a quiescent
675 * state. Either way, that CPU cannot possibly be in an RCU
676 * read-side critical section that started before the beginning
677 * of the current RCU grace period.
678 */
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700679 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700680 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, "dti");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100681 rdp->dynticks_fqs++;
682 return 1;
683 }
684
685 /* Go check for the CPU being offline. */
686 return rcu_implicit_offline_qs(rdp);
687}
688
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800689static int jiffies_till_stall_check(void)
690{
691 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
692
693 /*
694 * Limit check must be consistent with the Kconfig limits
695 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
696 */
697 if (till_stall_check < 3) {
698 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
699 till_stall_check = 3;
700 } else if (till_stall_check > 300) {
701 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
702 till_stall_check = 300;
703 }
704 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
705}
706
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100707static void record_gp_stall_check_time(struct rcu_state *rsp)
708{
709 rsp->gp_start = jiffies;
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800710 rsp->jiffies_stall = jiffies + jiffies_till_stall_check();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100711}
712
713static void print_other_cpu_stall(struct rcu_state *rsp)
714{
715 int cpu;
716 long delta;
717 unsigned long flags;
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700718 int ndetected;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100719 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100720
721 /* Only let one CPU complain about others per time interval. */
722
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800723 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100724 delta = jiffies - rsp->jiffies_stall;
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -0700725 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800726 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100727 return;
728 }
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800729 rsp->jiffies_stall = jiffies + 3 * jiffies_till_stall_check() + 3;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800730 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100731
Paul E. McKenney8cdd32a2010-08-09 14:23:03 -0700732 /*
733 * OK, time to rat on our buddy...
734 * See Documentation/RCU/stallwarn.txt for info on how to debug
735 * RCU CPU stall warnings.
736 */
Paul E. McKenneya858af22012-01-16 13:29:10 -0800737 printk(KERN_ERR "INFO: %s detected stalls on CPUs/tasks:",
Paul E. McKenney4300aa62010-04-13 16:18:22 -0700738 rsp->name);
Paul E. McKenneya858af22012-01-16 13:29:10 -0800739 print_cpu_stall_info_begin();
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700740 rcu_for_each_leaf_node(rsp, rnp) {
Paul E. McKenney3acd9eb2010-02-22 17:05:03 -0800741 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700742 ndetected += rcu_print_task_stall(rnp);
Paul E. McKenney3acd9eb2010-02-22 17:05:03 -0800743 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700744 if (rnp->qsmask == 0)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100745 continue;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700746 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700747 if (rnp->qsmask & (1UL << cpu)) {
Paul E. McKenneya858af22012-01-16 13:29:10 -0800748 print_cpu_stall_info(rsp, rnp->grplo + cpu);
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700749 ndetected++;
750 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100751 }
Paul E. McKenneya858af22012-01-16 13:29:10 -0800752
753 /*
754 * Now rat on any tasks that got kicked up to the root rcu_node
755 * due to CPU offlining.
756 */
757 rnp = rcu_get_root(rsp);
758 raw_spin_lock_irqsave(&rnp->lock, flags);
759 ndetected = rcu_print_task_stall(rnp);
760 raw_spin_unlock_irqrestore(&rnp->lock, flags);
761
762 print_cpu_stall_info_end();
763 printk(KERN_CONT "(detected by %d, t=%ld jiffies)\n",
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100764 smp_processor_id(), (long)(jiffies - rsp->gp_start));
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700765 if (ndetected == 0)
766 printk(KERN_ERR "INFO: Stall ended before state dump start\n");
767 else if (!trigger_all_cpu_backtrace())
Paul E. McKenney4627e242011-08-03 03:34:24 -0700768 dump_stack();
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200769
Paul E. McKenney1ed509a2010-02-22 17:05:05 -0800770 /* If so configured, complain about tasks blocking the grace period. */
771
772 rcu_print_detail_task_stall(rsp);
773
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100774 force_quiescent_state(rsp, 0); /* Kick them all. */
775}
776
777static void print_cpu_stall(struct rcu_state *rsp)
778{
779 unsigned long flags;
780 struct rcu_node *rnp = rcu_get_root(rsp);
781
Paul E. McKenney8cdd32a2010-08-09 14:23:03 -0700782 /*
783 * OK, time to rat on ourselves...
784 * See Documentation/RCU/stallwarn.txt for info on how to debug
785 * RCU CPU stall warnings.
786 */
Paul E. McKenneya858af22012-01-16 13:29:10 -0800787 printk(KERN_ERR "INFO: %s self-detected stall on CPU", rsp->name);
788 print_cpu_stall_info_begin();
789 print_cpu_stall_info(rsp, smp_processor_id());
790 print_cpu_stall_info_end();
791 printk(KERN_CONT " (t=%lu jiffies)\n", jiffies - rsp->gp_start);
Paul E. McKenney4627e242011-08-03 03:34:24 -0700792 if (!trigger_all_cpu_backtrace())
793 dump_stack();
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200794
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800795 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney20133cf2010-02-22 17:05:01 -0800796 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800797 rsp->jiffies_stall = jiffies +
798 3 * jiffies_till_stall_check() + 3;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800799 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200800
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100801 set_need_resched(); /* kick ourselves to get things going. */
802}
803
804static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
805{
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700806 unsigned long j;
807 unsigned long js;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100808 struct rcu_node *rnp;
809
Paul E. McKenney742734e2010-06-30 11:43:52 -0700810 if (rcu_cpu_stall_suppress)
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700811 return;
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700812 j = ACCESS_ONCE(jiffies);
813 js = ACCESS_ONCE(rsp->jiffies_stall);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100814 rnp = rdp->mynode;
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700815 if ((ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100816
817 /* We haven't checked in, so go dump stack. */
818 print_cpu_stall(rsp);
819
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700820 } else if (rcu_gp_in_progress(rsp) &&
821 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100822
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700823 /* They had a few time units to dump stack, so complain. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100824 print_other_cpu_stall(rsp);
825 }
826}
827
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700828static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
829{
Paul E. McKenney742734e2010-06-30 11:43:52 -0700830 rcu_cpu_stall_suppress = 1;
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700831 return NOTIFY_DONE;
832}
833
Paul E. McKenney53d84e02010-08-10 14:28:53 -0700834/**
835 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
836 *
837 * Set the stall-warning timeout way off into the future, thus preventing
838 * any RCU CPU stall-warning messages from appearing in the current set of
839 * RCU grace periods.
840 *
841 * The caller must disable hard irqs.
842 */
843void rcu_cpu_stall_reset(void)
844{
845 rcu_sched_state.jiffies_stall = jiffies + ULONG_MAX / 2;
846 rcu_bh_state.jiffies_stall = jiffies + ULONG_MAX / 2;
847 rcu_preempt_stall_reset();
848}
849
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700850static struct notifier_block rcu_panic_block = {
851 .notifier_call = rcu_panic,
852};
853
854static void __init check_cpu_stall_init(void)
855{
856 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
857}
858
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100859/*
860 * Update CPU-local rcu_data state to record the newly noticed grace period.
861 * This is used both when we started the grace period and when we notice
Paul E. McKenney91603062009-11-02 13:52:29 -0800862 * that someone else started the grace period. The caller must hold the
863 * ->lock of the leaf rcu_node structure corresponding to the current CPU,
864 * and must have irqs disabled.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100865 */
Paul E. McKenney91603062009-11-02 13:52:29 -0800866static void __note_new_gpnum(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
867{
868 if (rdp->gpnum != rnp->gpnum) {
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800869 /*
870 * If the current grace period is waiting for this CPU,
871 * set up to detect a quiescent state, otherwise don't
872 * go looking for one.
873 */
Paul E. McKenney91603062009-11-02 13:52:29 -0800874 rdp->gpnum = rnp->gpnum;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700875 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpustart");
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800876 if (rnp->qsmask & rdp->grpmask) {
877 rdp->qs_pending = 1;
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700878 rdp->passed_quiesce = 0;
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800879 } else
880 rdp->qs_pending = 0;
Paul E. McKenneya858af22012-01-16 13:29:10 -0800881 zero_cpu_stall_ticks(rdp);
Paul E. McKenney91603062009-11-02 13:52:29 -0800882 }
883}
884
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100885static void note_new_gpnum(struct rcu_state *rsp, struct rcu_data *rdp)
886{
Paul E. McKenney91603062009-11-02 13:52:29 -0800887 unsigned long flags;
888 struct rcu_node *rnp;
889
890 local_irq_save(flags);
891 rnp = rdp->mynode;
892 if (rdp->gpnum == ACCESS_ONCE(rnp->gpnum) || /* outside lock. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800893 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
Paul E. McKenney91603062009-11-02 13:52:29 -0800894 local_irq_restore(flags);
895 return;
896 }
897 __note_new_gpnum(rsp, rnp, rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800898 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100899}
900
901/*
902 * Did someone else start a new RCU grace period start since we last
903 * checked? Update local state appropriately if so. Must be called
904 * on the CPU corresponding to rdp.
905 */
906static int
907check_for_new_grace_period(struct rcu_state *rsp, struct rcu_data *rdp)
908{
909 unsigned long flags;
910 int ret = 0;
911
912 local_irq_save(flags);
913 if (rdp->gpnum != rsp->gpnum) {
914 note_new_gpnum(rsp, rdp);
915 ret = 1;
916 }
917 local_irq_restore(flags);
918 return ret;
919}
920
921/*
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800922 * Advance this CPU's callbacks, but only if the current grace period
923 * has ended. This may be called only from the CPU to whom the rdp
924 * belongs. In addition, the corresponding leaf rcu_node structure's
925 * ->lock must be held by the caller, with irqs disabled.
926 */
927static void
928__rcu_process_gp_end(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
929{
930 /* Did another grace period end? */
931 if (rdp->completed != rnp->completed) {
932
933 /* Advance callbacks. No harm if list empty. */
934 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[RCU_WAIT_TAIL];
935 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_READY_TAIL];
936 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
937
938 /* Remember that we saw this grace-period completion. */
939 rdp->completed = rnp->completed;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -0700940 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuend");
Frederic Weisbecker20377f32010-12-10 22:11:10 +0100941
942 /*
Frederic Weisbecker5ff8e6f2010-12-10 22:11:11 +0100943 * If we were in an extended quiescent state, we may have
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800944 * missed some grace periods that others CPUs handled on
Frederic Weisbecker5ff8e6f2010-12-10 22:11:11 +0100945 * our behalf. Catch up with this state to avoid noting
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800946 * spurious new grace periods. If another grace period
947 * has started, then rnp->gpnum will have advanced, so
948 * we will detect this later on.
Frederic Weisbecker5ff8e6f2010-12-10 22:11:11 +0100949 */
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800950 if (ULONG_CMP_LT(rdp->gpnum, rdp->completed))
Frederic Weisbecker5ff8e6f2010-12-10 22:11:11 +0100951 rdp->gpnum = rdp->completed;
952
953 /*
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800954 * If RCU does not need a quiescent state from this CPU,
955 * then make sure that this CPU doesn't go looking for one.
Frederic Weisbecker20377f32010-12-10 22:11:10 +0100956 */
Paul E. McKenney121dfc42010-12-10 15:02:47 -0800957 if ((rnp->qsmask & rdp->grpmask) == 0)
Frederic Weisbecker20377f32010-12-10 22:11:10 +0100958 rdp->qs_pending = 0;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800959 }
960}
961
962/*
963 * Advance this CPU's callbacks, but only if the current grace period
964 * has ended. This may be called only from the CPU to whom the rdp
965 * belongs.
966 */
967static void
968rcu_process_gp_end(struct rcu_state *rsp, struct rcu_data *rdp)
969{
970 unsigned long flags;
971 struct rcu_node *rnp;
972
973 local_irq_save(flags);
974 rnp = rdp->mynode;
975 if (rdp->completed == ACCESS_ONCE(rnp->completed) || /* outside lock. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800976 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800977 local_irq_restore(flags);
978 return;
979 }
980 __rcu_process_gp_end(rsp, rnp, rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800981 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -0800982}
983
984/*
985 * Do per-CPU grace-period initialization for running CPU. The caller
986 * must hold the lock of the leaf rcu_node structure corresponding to
987 * this CPU.
988 */
989static void
990rcu_start_gp_per_cpu(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
991{
992 /* Prior grace period ended, so advance callbacks for current CPU. */
993 __rcu_process_gp_end(rsp, rnp, rdp);
994
995 /*
996 * Because this CPU just now started the new grace period, we know
997 * that all of its callbacks will be covered by this upcoming grace
998 * period, even the ones that were registered arbitrarily recently.
999 * Therefore, advance all outstanding callbacks to RCU_WAIT_TAIL.
1000 *
1001 * Other CPUs cannot be sure exactly when the grace period started.
1002 * Therefore, their recently registered callbacks must pass through
1003 * an additional RCU_NEXT_READY stage, so that they will be handled
1004 * by the next RCU grace period.
1005 */
1006 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1007 rdp->nxttail[RCU_WAIT_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
Paul E. McKenney91603062009-11-02 13:52:29 -08001008
1009 /* Set state so that this CPU will detect the next quiescent state. */
1010 __note_new_gpnum(rsp, rnp, rdp);
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001011}
1012
1013/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001014 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1015 * in preparation for detecting the next grace period. The caller must hold
1016 * the root node's ->lock, which is released before return. Hard irqs must
1017 * be disabled.
Paul E. McKenneye5601402012-01-07 11:03:57 -08001018 *
1019 * Note that it is legal for a dying CPU (which is marked as offline) to
1020 * invoke this function. This can happen when the dying CPU reports its
1021 * quiescent state.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001022 */
1023static void
1024rcu_start_gp(struct rcu_state *rsp, unsigned long flags)
1025 __releases(rcu_get_root(rsp)->lock)
1026{
Lai Jiangshan394f99a2010-06-28 16:25:04 +08001027 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001028 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001029
Paul E. McKenney037067a2011-08-07 20:26:31 -07001030 if (!rcu_scheduler_fully_active ||
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001031 !cpu_needs_another_gp(rsp, rdp)) {
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -08001032 /*
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001033 * Either the scheduler hasn't yet spawned the first
1034 * non-idle task or this CPU does not need another
1035 * grace period. Either way, don't start a new grace
1036 * period.
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -08001037 */
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001038 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1039 return;
1040 }
1041
1042 if (rsp->fqs_active) {
1043 /*
1044 * This CPU needs a grace period, but force_quiescent_state()
1045 * is running. Tell it to start one on this CPU's behalf.
1046 */
1047 rsp->fqs_need_gp = 1;
1048 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001049 return;
1050 }
1051
1052 /* Advance to a new grace period and initialize state. */
1053 rsp->gpnum++;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001054 trace_rcu_grace_period(rsp->name, rsp->gpnum, "start");
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001055 WARN_ON_ONCE(rsp->fqs_state == RCU_GP_INIT);
1056 rsp->fqs_state = RCU_GP_INIT; /* Hold off force_quiescent_state. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001057 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001058 record_gp_stall_check_time(rsp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001059 raw_spin_unlock(&rnp->lock); /* leave irqs disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001060
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001061 /* Exclude any concurrent CPU-hotplug operations. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001062 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001063
1064 /*
Paul E. McKenneyb835db12009-09-08 15:54:37 -07001065 * Set the quiescent-state-needed bits in all the rcu_node
1066 * structures for all currently online CPUs in breadth-first
1067 * order, starting from the root rcu_node structure. This
1068 * operation relies on the layout of the hierarchy within the
1069 * rsp->node[] array. Note that other CPUs will access only
1070 * the leaves of the hierarchy, which still indicate that no
1071 * grace period is in progress, at least until the corresponding
1072 * leaf node has been initialized. In addition, we have excluded
1073 * CPU-hotplug operations.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001074 *
1075 * Note that the grace period cannot complete until we finish
1076 * the initialization process, as there will be at least one
1077 * qsmask bit set in the root node until that time, namely the
Paul E. McKenneyb835db12009-09-08 15:54:37 -07001078 * one corresponding to this CPU, due to the fact that we have
1079 * irqs disabled.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001080 */
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001081 rcu_for_each_node_breadth_first(rsp, rnp) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001082 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenneyb0e165c02009-09-13 09:15:09 -07001083 rcu_preempt_check_blocked_tasks(rnp);
Paul E. McKenney49e29122009-09-18 09:50:19 -07001084 rnp->qsmask = rnp->qsmaskinit;
Paul E. McKenneyde078d82009-09-08 15:54:36 -07001085 rnp->gpnum = rsp->gpnum;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001086 rnp->completed = rsp->completed;
1087 if (rnp == rdp->mynode)
1088 rcu_start_gp_per_cpu(rsp, rnp, rdp);
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001089 rcu_preempt_boost_start_gp(rnp);
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001090 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1091 rnp->level, rnp->grplo,
1092 rnp->grphi, rnp->qsmask);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001093 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001094 }
1095
Paul E. McKenney83f5b012009-10-28 08:14:49 -07001096 rnp = rcu_get_root(rsp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001097 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001098 rsp->fqs_state = RCU_SIGNAL_INIT; /* force_quiescent_state now OK. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001099 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1100 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001101}
1102
1103/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001104 * Report a full set of quiescent states to the specified rcu_state
1105 * data structure. This involves cleaning up after the prior grace
1106 * period and letting rcu_start_gp() start up the next grace period
1107 * if one is needed. Note that the caller must hold rnp->lock, as
1108 * required by rcu_start_gp(), which will release it.
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001109 */
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001110static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -07001111 __releases(rcu_get_root(rsp)->lock)
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001112{
Paul E. McKenney15ba0ba2011-04-06 16:01:16 -07001113 unsigned long gp_duration;
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001114 struct rcu_node *rnp = rcu_get_root(rsp);
1115 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
Paul E. McKenney15ba0ba2011-04-06 16:01:16 -07001116
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -07001117 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
Paul E. McKenney0bbcc522011-05-16 02:24:04 -07001118
1119 /*
1120 * Ensure that all grace-period and pre-grace-period activity
1121 * is seen before the assignment to rsp->completed.
1122 */
1123 smp_mb(); /* See above block comment. */
Paul E. McKenney15ba0ba2011-04-06 16:01:16 -07001124 gp_duration = jiffies - rsp->gp_start;
1125 if (gp_duration > rsp->gp_max)
1126 rsp->gp_max = gp_duration;
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001127
1128 /*
1129 * We know the grace period is complete, but to everyone else
1130 * it appears to still be ongoing. But it is also the case
1131 * that to everyone else it looks like there is nothing that
1132 * they can do to advance the grace period. It is therefore
1133 * safe for us to drop the lock in order to mark the grace
1134 * period as completed in all of the rcu_node structures.
1135 *
1136 * But if this CPU needs another grace period, it will take
1137 * care of this while initializing the next grace period.
1138 * We use RCU_WAIT_TAIL instead of the usual RCU_DONE_TAIL
1139 * because the callbacks have not yet been advanced: Those
1140 * callbacks are waiting on the grace period that just now
1141 * completed.
1142 */
1143 if (*rdp->nxttail[RCU_WAIT_TAIL] == NULL) {
1144 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1145
1146 /*
1147 * Propagate new ->completed value to rcu_node structures
1148 * so that other CPUs don't have to wait until the start
1149 * of the next grace period to process their callbacks.
1150 */
1151 rcu_for_each_node_breadth_first(rsp, rnp) {
1152 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1153 rnp->completed = rsp->gpnum;
1154 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1155 }
1156 rnp = rcu_get_root(rsp);
1157 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1158 }
1159
1160 rsp->completed = rsp->gpnum; /* Declare the grace period complete. */
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001161 trace_rcu_grace_period(rsp->name, rsp->completed, "end");
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001162 rsp->fqs_state = RCU_GP_IDLE;
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001163 rcu_start_gp(rsp, flags); /* releases root node's rnp->lock. */
1164}
1165
1166/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001167 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1168 * Allows quiescent states for a group of CPUs to be reported at one go
1169 * to the specified rcu_node structure, though all the CPUs in the group
1170 * must be represented by the same rcu_node structure (which need not be
1171 * a leaf rcu_node structure, though it often will be). That structure's
1172 * lock must be held upon entry, and it is released before return.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001173 */
1174static void
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001175rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1176 struct rcu_node *rnp, unsigned long flags)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001177 __releases(rnp->lock)
1178{
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001179 struct rcu_node *rnp_c;
1180
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001181 /* Walk up the rcu_node hierarchy. */
1182 for (;;) {
1183 if (!(rnp->qsmask & mask)) {
1184
1185 /* Our bit has already been cleared, so done. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001186 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001187 return;
1188 }
1189 rnp->qsmask &= ~mask;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001190 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1191 mask, rnp->qsmask, rnp->level,
1192 rnp->grplo, rnp->grphi,
1193 !!rnp->gp_tasks);
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001194 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001195
1196 /* Other bits still set at this level, so done. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001197 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001198 return;
1199 }
1200 mask = rnp->grpmask;
1201 if (rnp->parent == NULL) {
1202
1203 /* No more levels. Exit loop holding root lock. */
1204
1205 break;
1206 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001207 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001208 rnp_c = rnp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001209 rnp = rnp->parent;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001210 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001211 WARN_ON_ONCE(rnp_c->qsmask);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001212 }
1213
1214 /*
1215 * Get here if we are the last CPU to pass through a quiescent
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001216 * state for this grace period. Invoke rcu_report_qs_rsp()
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001217 * to clean up and start the next grace period if one is needed.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001218 */
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001219 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001220}
1221
1222/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001223 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1224 * structure. This must be either called from the specified CPU, or
1225 * called when the specified CPU is known to be offline (and when it is
1226 * also known that no other CPU is concurrently trying to help the offline
1227 * CPU). The lastcomp argument is used to make sure we are still in the
1228 * grace period of interest. We don't want to end the current grace period
1229 * based on quiescent states detected in an earlier grace period!
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001230 */
1231static void
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001232rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp, long lastgp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001233{
1234 unsigned long flags;
1235 unsigned long mask;
1236 struct rcu_node *rnp;
1237
1238 rnp = rdp->mynode;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001239 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001240 if (lastgp != rnp->gpnum || rnp->completed == rnp->gpnum) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001241
1242 /*
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001243 * The grace period in which this quiescent state was
1244 * recorded has ended, so don't report it upwards.
1245 * We will instead need a new quiescent state that lies
1246 * within the current grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001247 */
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001248 rdp->passed_quiesce = 0; /* need qs for new gp. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001249 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001250 return;
1251 }
1252 mask = rdp->grpmask;
1253 if ((rnp->qsmask & mask) == 0) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001254 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001255 } else {
1256 rdp->qs_pending = 0;
1257
1258 /*
1259 * This GP can't end until cpu checks in, so all of our
1260 * callbacks can be processed during the next GP.
1261 */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001262 rdp->nxttail[RCU_NEXT_READY_TAIL] = rdp->nxttail[RCU_NEXT_TAIL];
1263
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001264 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001265 }
1266}
1267
1268/*
1269 * Check to see if there is a new grace period of which this CPU
1270 * is not yet aware, and if so, set up local rcu_data state for it.
1271 * Otherwise, see if this CPU has just passed through its first
1272 * quiescent state for this grace period, and record that fact if so.
1273 */
1274static void
1275rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1276{
1277 /* If there is now a new grace period, record and return. */
1278 if (check_for_new_grace_period(rsp, rdp))
1279 return;
1280
1281 /*
1282 * Does this CPU still need to do its part for current grace period?
1283 * If no, return and let the other CPUs do their part as well.
1284 */
1285 if (!rdp->qs_pending)
1286 return;
1287
1288 /*
1289 * Was there a quiescent state since the beginning of the grace
1290 * period? If no, then exit and wait for the next call.
1291 */
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001292 if (!rdp->passed_quiesce)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001293 return;
1294
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001295 /*
1296 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1297 * judge of that).
1298 */
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001299 rcu_report_qs_rdp(rdp->cpu, rsp, rdp, rdp->passed_quiesce_gpnum);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001300}
1301
1302#ifdef CONFIG_HOTPLUG_CPU
1303
1304/*
Lai Jiangshan29494be2010-10-20 14:13:06 +08001305 * Move a dying CPU's RCU callbacks to online CPU's callback list.
Paul E. McKenneye5601402012-01-07 11:03:57 -08001306 * Also record a quiescent state for this CPU for the current grace period.
1307 * Synchronization and interrupt disabling are not required because
1308 * this function executes in stop_machine() context. Therefore, cleanup
1309 * operations that might block must be done later from the CPU_DEAD
1310 * notifier.
1311 *
1312 * Note that the outgoing CPU's bit has already been cleared in the
1313 * cpu_online_mask. This allows us to randomly pick a callback
1314 * destination from the bits set in that mask.
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001315 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001316static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001317{
1318 int i;
Paul E. McKenneye5601402012-01-07 11:03:57 -08001319 unsigned long mask;
Lai Jiangshan29494be2010-10-20 14:13:06 +08001320 int receive_cpu = cpumask_any(cpu_online_mask);
Lai Jiangshan394f99a2010-06-28 16:25:04 +08001321 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
Lai Jiangshan29494be2010-10-20 14:13:06 +08001322 struct rcu_data *receive_rdp = per_cpu_ptr(rsp->rda, receive_cpu);
Paul E. McKenney2036d942012-01-30 17:02:47 -08001323 RCU_TRACE(struct rcu_node *rnp = rdp->mynode); /* For dying CPU. */
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001324
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001325 /* First, adjust the counts. */
1326 if (rdp->nxtlist != NULL) {
1327 receive_rdp->qlen_lazy += rdp->qlen_lazy;
1328 receive_rdp->qlen += rdp->qlen;
1329 rdp->qlen_lazy = 0;
1330 rdp->qlen = 0;
1331 }
1332
1333 /*
1334 * Next, move ready-to-invoke callbacks to be invoked on some
1335 * other CPU. These will not be required to pass through another
1336 * grace period: They are done, regardless of CPU.
1337 */
1338 if (rdp->nxtlist != NULL &&
1339 rdp->nxttail[RCU_DONE_TAIL] != &rdp->nxtlist) {
1340 struct rcu_head *oldhead;
1341 struct rcu_head **oldtail;
1342 struct rcu_head **newtail;
1343
1344 oldhead = rdp->nxtlist;
1345 oldtail = receive_rdp->nxttail[RCU_DONE_TAIL];
1346 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1347 *rdp->nxttail[RCU_DONE_TAIL] = *oldtail;
1348 *receive_rdp->nxttail[RCU_DONE_TAIL] = oldhead;
1349 newtail = rdp->nxttail[RCU_DONE_TAIL];
1350 for (i = RCU_DONE_TAIL; i < RCU_NEXT_SIZE; i++) {
1351 if (receive_rdp->nxttail[i] == oldtail)
1352 receive_rdp->nxttail[i] = newtail;
1353 if (rdp->nxttail[i] == newtail)
1354 rdp->nxttail[i] = &rdp->nxtlist;
1355 }
1356 }
1357
1358 /*
1359 * Finally, put the rest of the callbacks at the end of the list.
1360 * The ones that made it partway through get to start over: We
1361 * cannot assume that grace periods are synchronized across CPUs.
1362 * (We could splice RCU_WAIT_TAIL into RCU_NEXT_READY_TAIL, but
1363 * this does not seem compelling. Not yet, anyway.)
1364 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001365 if (rdp->nxtlist != NULL) {
1366 *receive_rdp->nxttail[RCU_NEXT_TAIL] = rdp->nxtlist;
1367 receive_rdp->nxttail[RCU_NEXT_TAIL] =
1368 rdp->nxttail[RCU_NEXT_TAIL];
Paul E. McKenneye5601402012-01-07 11:03:57 -08001369 receive_rdp->n_cbs_adopted += rdp->qlen;
1370 rdp->n_cbs_orphaned += rdp->qlen;
Lai Jiangshan29494be2010-10-20 14:13:06 +08001371
Paul E. McKenneye5601402012-01-07 11:03:57 -08001372 rdp->nxtlist = NULL;
1373 for (i = 0; i < RCU_NEXT_SIZE; i++)
1374 rdp->nxttail[i] = &rdp->nxtlist;
Paul E. McKenneye5601402012-01-07 11:03:57 -08001375 }
Lai Jiangshan29494be2010-10-20 14:13:06 +08001376
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001377 /*
1378 * Record a quiescent state for the dying CPU. This is safe
1379 * only because we have already cleared out the callbacks.
1380 * (Otherwise, the RCU core might try to schedule the invocation
1381 * of callbacks on this now-offline CPU, which would be bad.)
1382 */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001383 mask = rdp->grpmask; /* rnp->grplo is constant. */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001384 trace_rcu_grace_period(rsp->name,
1385 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1386 "cpuofl");
1387 rcu_report_qs_rdp(smp_processor_id(), rsp, rdp, rsp->gpnum);
1388 /* Note that rcu_report_qs_rdp() might call trace_rcu_grace_period(). */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001389}
1390
1391/*
Paul E. McKenneye5601402012-01-07 11:03:57 -08001392 * The CPU has been completely removed, and some other CPU is reporting
1393 * this fact from process context. Do the remainder of the cleanup.
1394 * There can only be one CPU hotplug operation at a time, so no other
1395 * CPU can be attempting to update rcu_cpu_kthread_task.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001396 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001397static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001398{
Paul E. McKenney2036d942012-01-30 17:02:47 -08001399 unsigned long flags;
1400 unsigned long mask;
1401 int need_report = 0;
Paul E. McKenneye5601402012-01-07 11:03:57 -08001402 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenney2036d942012-01-30 17:02:47 -08001403 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rnp. */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001404
Paul E. McKenney2036d942012-01-30 17:02:47 -08001405 /* Adjust any no-longer-needed kthreads. */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001406 rcu_stop_cpu_kthread(cpu);
1407 rcu_node_kthread_setaffinity(rnp, -1);
Paul E. McKenney2036d942012-01-30 17:02:47 -08001408
1409 /* Remove the dying CPU from the bitmasks in the rcu_node hierarchy. */
1410
1411 /* Exclude any attempts to start a new grace period. */
1412 raw_spin_lock_irqsave(&rsp->onofflock, flags);
1413
1414 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1415 mask = rdp->grpmask; /* rnp->grplo is constant. */
1416 do {
1417 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1418 rnp->qsmaskinit &= ~mask;
1419 if (rnp->qsmaskinit != 0) {
1420 if (rnp != rdp->mynode)
1421 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1422 break;
1423 }
1424 if (rnp == rdp->mynode)
1425 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1426 else
1427 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1428 mask = rnp->grpmask;
1429 rnp = rnp->parent;
1430 } while (rnp != NULL);
1431
1432 /*
1433 * We still hold the leaf rcu_node structure lock here, and
1434 * irqs are still disabled. The reason for this subterfuge is
1435 * because invoking rcu_report_unblock_qs_rnp() with ->onofflock
1436 * held leads to deadlock.
1437 */
1438 raw_spin_unlock(&rsp->onofflock); /* irqs remain disabled. */
1439 rnp = rdp->mynode;
1440 if (need_report & RCU_OFL_TASKS_NORM_GP)
1441 rcu_report_unblock_qs_rnp(rnp, flags);
1442 else
1443 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1444 if (need_report & RCU_OFL_TASKS_EXP_GP)
1445 rcu_report_exp_rnp(rsp, rnp, true);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001446}
1447
1448#else /* #ifdef CONFIG_HOTPLUG_CPU */
1449
Paul E. McKenneye5601402012-01-07 11:03:57 -08001450static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001451{
1452}
1453
Paul E. McKenneye5601402012-01-07 11:03:57 -08001454static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001455{
1456}
1457
1458#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1459
1460/*
1461 * Invoke any RCU callbacks that have made it to the end of their grace
1462 * period. Thottle as specified by rdp->blimit.
1463 */
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001464static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001465{
1466 unsigned long flags;
1467 struct rcu_head *next, *list, **tail;
Paul E. McKenney486e2592012-01-06 14:11:30 -08001468 int bl, count, count_lazy;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001469
1470 /* If no callbacks are ready, just return.*/
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001471 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
Paul E. McKenney486e2592012-01-06 14:11:30 -08001472 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
Paul E. McKenney4968c302011-12-07 16:32:40 -08001473 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1474 need_resched(), is_idle_task(current),
1475 rcu_is_callbacks_kthread());
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001476 return;
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001477 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001478
1479 /*
1480 * Extract the list of ready callbacks, disabling to prevent
1481 * races with call_rcu() from interrupt handlers.
1482 */
1483 local_irq_save(flags);
Paul E. McKenney8146c4e22012-01-10 14:23:29 -08001484 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001485 bl = rdp->blimit;
Paul E. McKenney486e2592012-01-06 14:11:30 -08001486 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001487 list = rdp->nxtlist;
1488 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1489 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1490 tail = rdp->nxttail[RCU_DONE_TAIL];
1491 for (count = RCU_NEXT_SIZE - 1; count >= 0; count--)
1492 if (rdp->nxttail[count] == rdp->nxttail[RCU_DONE_TAIL])
1493 rdp->nxttail[count] = &rdp->nxtlist;
1494 local_irq_restore(flags);
1495
1496 /* Invoke callbacks. */
Paul E. McKenney486e2592012-01-06 14:11:30 -08001497 count = count_lazy = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001498 while (list) {
1499 next = list->next;
1500 prefetch(next);
Mathieu Desnoyers551d55a2010-04-17 08:48:42 -04001501 debug_rcu_head_unqueue(list);
Paul E. McKenney486e2592012-01-06 14:11:30 -08001502 if (__rcu_reclaim(rsp->name, list))
1503 count_lazy++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001504 list = next;
Paul E. McKenneydff16722011-11-29 15:57:13 -08001505 /* Stop only if limit reached and CPU has something to do. */
1506 if (++count >= bl &&
1507 (need_resched() ||
1508 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001509 break;
1510 }
1511
1512 local_irq_save(flags);
Paul E. McKenney4968c302011-12-07 16:32:40 -08001513 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
1514 is_idle_task(current),
1515 rcu_is_callbacks_kthread());
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001516
1517 /* Update count, and requeue any remaining callbacks. */
Paul E. McKenney486e2592012-01-06 14:11:30 -08001518 rdp->qlen_lazy -= count_lazy;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001519 rdp->qlen -= count;
Paul E. McKenney269dcc12010-09-07 14:23:09 -07001520 rdp->n_cbs_invoked += count;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001521 if (list != NULL) {
1522 *tail = rdp->nxtlist;
1523 rdp->nxtlist = list;
1524 for (count = 0; count < RCU_NEXT_SIZE; count++)
1525 if (&rdp->nxtlist == rdp->nxttail[count])
1526 rdp->nxttail[count] = tail;
1527 else
1528 break;
1529 }
1530
1531 /* Reinstate batch limit if we have worked down the excess. */
1532 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
1533 rdp->blimit = blimit;
1534
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001535 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
1536 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
1537 rdp->qlen_last_fqs_check = 0;
1538 rdp->n_force_qs_snap = rsp->n_force_qs;
1539 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
1540 rdp->qlen_last_fqs_check = rdp->qlen;
1541
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001542 local_irq_restore(flags);
1543
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001544 /* Re-invoke RCU core processing if there are callbacks remaining. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001545 if (cpu_has_callbacks_ready_to_invoke(rdp))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001546 invoke_rcu_core();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001547}
1548
1549/*
1550 * Check to see if this CPU is in a non-context-switch quiescent state
1551 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001552 * Also schedule RCU core processing.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001553 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07001554 * This function must be called from hardirq context. It is normally
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001555 * invoked from the scheduling-clock interrupt. If rcu_pending returns
1556 * false, there is no point in invoking rcu_check_callbacks().
1557 */
1558void rcu_check_callbacks(int cpu, int user)
1559{
Paul E. McKenney300df912011-06-18 22:26:31 -07001560 trace_rcu_utilization("Start scheduler-tick");
Paul E. McKenneya858af22012-01-16 13:29:10 -08001561 increment_cpu_stall_ticks();
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07001562 if (user || rcu_is_cpu_rrupt_from_idle()) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001563
1564 /*
1565 * Get here if this CPU took its interrupt from user
1566 * mode or from the idle loop, and if this is not a
1567 * nested interrupt. In this case, the CPU is in
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001568 * a quiescent state, so note it.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001569 *
1570 * No memory barrier is required here because both
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001571 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
1572 * variables that other CPUs neither access nor modify,
1573 * at least not while the corresponding CPU is online.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001574 */
1575
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001576 rcu_sched_qs(cpu);
1577 rcu_bh_qs(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001578
1579 } else if (!in_softirq()) {
1580
1581 /*
1582 * Get here if this CPU did not take its interrupt from
1583 * softirq, in other words, if it is not interrupting
1584 * a rcu_bh read-side critical section. This is an _bh
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001585 * critical section, so note it.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001586 */
1587
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001588 rcu_bh_qs(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001589 }
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001590 rcu_preempt_check_callbacks(cpu);
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07001591 if (rcu_pending(cpu))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001592 invoke_rcu_core();
Paul E. McKenney300df912011-06-18 22:26:31 -07001593 trace_rcu_utilization("End scheduler-tick");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001594}
1595
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001596/*
1597 * Scan the leaf rcu_node structures, processing dyntick state for any that
1598 * have not yet encountered a quiescent state, using the function specified.
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001599 * Also initiate boosting for any threads blocked on the root rcu_node.
1600 *
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001601 * The caller must have suppressed start of new grace periods.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001602 */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001603static void force_qs_rnp(struct rcu_state *rsp, int (*f)(struct rcu_data *))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001604{
1605 unsigned long bit;
1606 int cpu;
1607 unsigned long flags;
1608 unsigned long mask;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001609 struct rcu_node *rnp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001610
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001611 rcu_for_each_leaf_node(rsp, rnp) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001612 mask = 0;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001613 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001614 if (!rcu_gp_in_progress(rsp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001615 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney0f10dc822010-01-04 15:09:06 -08001616 return;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001617 }
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001618 if (rnp->qsmask == 0) {
Paul E. McKenney1217ed12011-05-04 21:43:49 -07001619 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001620 continue;
1621 }
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001622 cpu = rnp->grplo;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001623 bit = 1;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07001624 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
Lai Jiangshan394f99a2010-06-28 16:25:04 +08001625 if ((rnp->qsmask & bit) != 0 &&
1626 f(per_cpu_ptr(rsp->rda, cpu)))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001627 mask |= bit;
1628 }
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001629 if (mask != 0) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001630
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001631 /* rcu_report_qs_rnp() releases rnp->lock. */
1632 rcu_report_qs_rnp(mask, rsp, rnp, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001633 continue;
1634 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001635 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001636 }
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001637 rnp = rcu_get_root(rsp);
Paul E. McKenney1217ed12011-05-04 21:43:49 -07001638 if (rnp->qsmask == 0) {
1639 raw_spin_lock_irqsave(&rnp->lock, flags);
1640 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
1641 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001642}
1643
1644/*
1645 * Force quiescent states on reluctant CPUs, and also detect which
1646 * CPUs are in dyntick-idle mode.
1647 */
1648static void force_quiescent_state(struct rcu_state *rsp, int relaxed)
1649{
1650 unsigned long flags;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001651 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001652
Paul E. McKenney300df912011-06-18 22:26:31 -07001653 trace_rcu_utilization("Start fqs");
1654 if (!rcu_gp_in_progress(rsp)) {
1655 trace_rcu_utilization("End fqs");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001656 return; /* No grace period in progress, nothing to force. */
Paul E. McKenney300df912011-06-18 22:26:31 -07001657 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001658 if (!raw_spin_trylock_irqsave(&rsp->fqslock, flags)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001659 rsp->n_force_qs_lh++; /* Inexact, can lose counts. Tough! */
Paul E. McKenney300df912011-06-18 22:26:31 -07001660 trace_rcu_utilization("End fqs");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001661 return; /* Someone else is already on the job. */
1662 }
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001663 if (relaxed && ULONG_CMP_GE(rsp->jiffies_force_qs, jiffies))
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001664 goto unlock_fqs_ret; /* no emergency and done recently. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001665 rsp->n_force_qs++;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001666 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001667 rsp->jiffies_force_qs = jiffies + RCU_JIFFIES_TILL_FORCE_QS;
Paul E. McKenney560d4bc2009-11-13 19:51:38 -08001668 if(!rcu_gp_in_progress(rsp)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001669 rsp->n_force_qs_ngp++;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001670 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001671 goto unlock_fqs_ret; /* no GP in progress, time updated. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001672 }
Paul E. McKenney07079d52010-01-04 15:09:02 -08001673 rsp->fqs_active = 1;
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001674 switch (rsp->fqs_state) {
Paul E. McKenney83f5b012009-10-28 08:14:49 -07001675 case RCU_GP_IDLE:
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001676 case RCU_GP_INIT:
1677
Paul E. McKenney83f5b012009-10-28 08:14:49 -07001678 break; /* grace period idle or initializing, ignore. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001679
1680 case RCU_SAVE_DYNTICK:
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001681 if (RCU_SIGNAL_INIT != RCU_SAVE_DYNTICK)
1682 break; /* So gcc recognizes the dead code. */
1683
Lai Jiangshanf2614142010-03-28 11:15:20 +08001684 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
1685
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001686 /* Record dyntick-idle state. */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001687 force_qs_rnp(rsp, dyntick_save_progress_counter);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001688 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001689 if (rcu_gp_in_progress(rsp))
Paul E. McKenneyaf446b72011-09-10 21:54:08 -07001690 rsp->fqs_state = RCU_FORCE_QS;
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08001691 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001692
1693 case RCU_FORCE_QS:
1694
1695 /* Check dyntick-idle state, send IPI to laggarts. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001696 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08001697 force_qs_rnp(rsp, rcu_implicit_dynticks_qs);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001698
1699 /* Leave state in case more forcing is required. */
1700
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001701 raw_spin_lock(&rnp->lock); /* irqs already disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001702 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001703 }
Paul E. McKenney07079d52010-01-04 15:09:02 -08001704 rsp->fqs_active = 0;
Paul E. McKenney46a1e342010-01-04 15:09:09 -08001705 if (rsp->fqs_need_gp) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001706 raw_spin_unlock(&rsp->fqslock); /* irqs remain disabled */
Paul E. McKenney46a1e342010-01-04 15:09:09 -08001707 rsp->fqs_need_gp = 0;
1708 rcu_start_gp(rsp, flags); /* releases rnp->lock */
Paul E. McKenney300df912011-06-18 22:26:31 -07001709 trace_rcu_utilization("End fqs");
Paul E. McKenney46a1e342010-01-04 15:09:09 -08001710 return;
1711 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001712 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
Paul E. McKenneyf96e9232010-01-04 15:09:00 -08001713unlock_fqs_ret:
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001714 raw_spin_unlock_irqrestore(&rsp->fqslock, flags);
Paul E. McKenney300df912011-06-18 22:26:31 -07001715 trace_rcu_utilization("End fqs");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001716}
1717
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001718/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001719 * This does the RCU core processing work for the specified rcu_state
1720 * and rcu_data structures. This may be called only from the CPU to
1721 * whom the rdp belongs.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001722 */
1723static void
1724__rcu_process_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
1725{
1726 unsigned long flags;
1727
Paul E. McKenney2e597552009-08-15 09:53:48 -07001728 WARN_ON_ONCE(rdp->beenonline == 0);
1729
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001730 /*
1731 * If an RCU GP has gone long enough, go check for dyntick
1732 * idle CPUs and, if needed, send resched IPIs.
1733 */
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001734 if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001735 force_quiescent_state(rsp, 1);
1736
1737 /*
1738 * Advance callbacks in response to end of earlier grace
1739 * period that some other CPU ended.
1740 */
1741 rcu_process_gp_end(rsp, rdp);
1742
1743 /* Update RCU state based on any recent quiescent states. */
1744 rcu_check_quiescent_state(rsp, rdp);
1745
1746 /* Does this CPU require a not-yet-started grace period? */
1747 if (cpu_needs_another_gp(rsp, rdp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001748 raw_spin_lock_irqsave(&rcu_get_root(rsp)->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001749 rcu_start_gp(rsp, flags); /* releases above lock */
1750 }
1751
1752 /* If there are callbacks ready, invoke them. */
Shaohua Li09223372011-06-14 13:26:25 +08001753 if (cpu_has_callbacks_ready_to_invoke(rdp))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001754 invoke_rcu_callbacks(rsp, rdp);
Shaohua Li09223372011-06-14 13:26:25 +08001755}
1756
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001757/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001758 * Do RCU core processing for the current CPU.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001759 */
Shaohua Li09223372011-06-14 13:26:25 +08001760static void rcu_process_callbacks(struct softirq_action *unused)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001761{
Paul E. McKenney300df912011-06-18 22:26:31 -07001762 trace_rcu_utilization("Start RCU core");
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001763 __rcu_process_callbacks(&rcu_sched_state,
1764 &__get_cpu_var(rcu_sched_data));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001765 __rcu_process_callbacks(&rcu_bh_state, &__get_cpu_var(rcu_bh_data));
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001766 rcu_preempt_process_callbacks();
Paul E. McKenney300df912011-06-18 22:26:31 -07001767 trace_rcu_utilization("End RCU core");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001768}
1769
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001770/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07001771 * Schedule RCU callback invocation. If the specified type of RCU
1772 * does not support RCU priority boosting, just do a direct call,
1773 * otherwise wake up the per-CPU kernel kthread. Note that because we
1774 * are running on the current CPU with interrupts disabled, the
1775 * rcu_cpu_kthread_task cannot disappear out from under us.
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001776 */
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001777static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001778{
Paul E. McKenneyb0d30412011-07-10 15:57:35 -07001779 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
1780 return;
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001781 if (likely(!rsp->boost)) {
1782 rcu_do_batch(rsp, rdp);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001783 return;
1784 }
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001785 invoke_rcu_callbacks_kthread();
Paul E. McKenneya26ac242011-01-12 14:10:23 -08001786}
1787
Paul E. McKenneya46e0892011-06-15 15:47:09 -07001788static void invoke_rcu_core(void)
Shaohua Li09223372011-06-14 13:26:25 +08001789{
1790 raise_softirq(RCU_SOFTIRQ);
1791}
1792
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001793static void
1794__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
Paul E. McKenney486e2592012-01-06 14:11:30 -08001795 struct rcu_state *rsp, bool lazy)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001796{
1797 unsigned long flags;
1798 struct rcu_data *rdp;
1799
Paul E. McKenney0bb7b592012-01-05 14:44:39 -08001800 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
Mathieu Desnoyers551d55a2010-04-17 08:48:42 -04001801 debug_rcu_head_queue(head);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001802 head->func = func;
1803 head->next = NULL;
1804
1805 smp_mb(); /* Ensure RCU update seen before callback registry. */
1806
1807 /*
1808 * Opportunistically note grace-period endings and beginnings.
1809 * Note that we might see a beginning right after we see an
1810 * end, but never vice versa, since this CPU has to pass through
1811 * a quiescent state betweentimes.
1812 */
1813 local_irq_save(flags);
Paul E. McKenneye5601402012-01-07 11:03:57 -08001814 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
Lai Jiangshan394f99a2010-06-28 16:25:04 +08001815 rdp = this_cpu_ptr(rsp->rda);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001816
1817 /* Add the callback to our list. */
1818 *rdp->nxttail[RCU_NEXT_TAIL] = head;
1819 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
Paul E. McKenney2655d572011-04-07 22:47:23 -07001820 rdp->qlen++;
Paul E. McKenney486e2592012-01-06 14:11:30 -08001821 if (lazy)
1822 rdp->qlen_lazy++;
Paul E. McKenney2655d572011-04-07 22:47:23 -07001823
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001824 if (__is_kfree_rcu_offset((unsigned long)func))
1825 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
Paul E. McKenney486e2592012-01-06 14:11:30 -08001826 rdp->qlen_lazy, rdp->qlen);
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001827 else
Paul E. McKenney486e2592012-01-06 14:11:30 -08001828 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001829
Paul E. McKenney2655d572011-04-07 22:47:23 -07001830 /* If interrupts were disabled, don't dive into RCU core. */
1831 if (irqs_disabled_flags(flags)) {
1832 local_irq_restore(flags);
1833 return;
1834 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001835
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001836 /*
1837 * Force the grace period if too many callbacks or too long waiting.
1838 * Enforce hysteresis, and don't invoke force_quiescent_state()
1839 * if some other CPU has recently done so. Also, don't bother
1840 * invoking force_quiescent_state() if the newly enqueued callback
1841 * is the only one waiting for a grace period to complete.
1842 */
Paul E. McKenney2655d572011-04-07 22:47:23 -07001843 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
Paul E. McKenneyb52573d2010-12-14 17:36:02 -08001844
1845 /* Are we ignoring a completed grace period? */
1846 rcu_process_gp_end(rsp, rdp);
1847 check_for_new_grace_period(rsp, rdp);
1848
1849 /* Start a new grace period if one not already started. */
1850 if (!rcu_gp_in_progress(rsp)) {
1851 unsigned long nestflag;
1852 struct rcu_node *rnp_root = rcu_get_root(rsp);
1853
1854 raw_spin_lock_irqsave(&rnp_root->lock, nestflag);
1855 rcu_start_gp(rsp, nestflag); /* rlses rnp_root->lock */
1856 } else {
1857 /* Give the grace period a kick. */
1858 rdp->blimit = LONG_MAX;
1859 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
1860 *rdp->nxttail[RCU_DONE_TAIL] != head)
1861 force_quiescent_state(rsp, 0);
1862 rdp->n_force_qs_snap = rsp->n_force_qs;
1863 rdp->qlen_last_fqs_check = rdp->qlen;
1864 }
Paul E. McKenney20133cf2010-02-22 17:05:01 -08001865 } else if (ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001866 force_quiescent_state(rsp, 1);
1867 local_irq_restore(flags);
1868}
1869
1870/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001871 * Queue an RCU-sched callback for invocation after a grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001872 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001873void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001874{
Paul E. McKenney486e2592012-01-06 14:11:30 -08001875 __call_rcu(head, func, &rcu_sched_state, 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001876}
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07001877EXPORT_SYMBOL_GPL(call_rcu_sched);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001878
1879/*
Paul E. McKenney486e2592012-01-06 14:11:30 -08001880 * Queue an RCU callback for invocation after a quicker grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001881 */
1882void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
1883{
Paul E. McKenney486e2592012-01-06 14:11:30 -08001884 __call_rcu(head, func, &rcu_bh_state, 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001885}
1886EXPORT_SYMBOL_GPL(call_rcu_bh);
1887
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001888/**
1889 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
1890 *
1891 * Control will return to the caller some time after a full rcu-sched
1892 * grace period has elapsed, in other words after all currently executing
1893 * rcu-sched read-side critical sections have completed. These read-side
1894 * critical sections are delimited by rcu_read_lock_sched() and
1895 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
1896 * local_irq_disable(), and so on may be used in place of
1897 * rcu_read_lock_sched().
1898 *
1899 * This means that all preempt_disable code sequences, including NMI and
1900 * hardware-interrupt handlers, in progress on entry will have completed
1901 * before this primitive returns. However, this does not guarantee that
1902 * softirq handlers will have completed, since in some kernels, these
1903 * handlers can run in process context, and can block.
1904 *
1905 * This primitive provides the guarantees made by the (now removed)
1906 * synchronize_kernel() API. In contrast, synchronize_rcu() only
1907 * guarantees that rcu_read_lock() sections will have completed.
1908 * In "classic RCU", these two guarantees happen to be one and
1909 * the same, but can differ in realtime RCU implementations.
1910 */
1911void synchronize_sched(void)
1912{
Paul E. McKenneyfe15d702012-01-04 13:30:33 -08001913 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
1914 !lock_is_held(&rcu_lock_map) &&
1915 !lock_is_held(&rcu_sched_lock_map),
1916 "Illegal synchronize_sched() in RCU-sched read-side critical section");
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001917 if (rcu_blocking_is_gp())
1918 return;
Paul E. McKenney2c428182011-05-26 22:14:36 -07001919 wait_rcu_gp(call_rcu_sched);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001920}
1921EXPORT_SYMBOL_GPL(synchronize_sched);
1922
1923/**
1924 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
1925 *
1926 * Control will return to the caller some time after a full rcu_bh grace
1927 * period has elapsed, in other words after all currently executing rcu_bh
1928 * read-side critical sections have completed. RCU read-side critical
1929 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
1930 * and may be nested.
1931 */
1932void synchronize_rcu_bh(void)
1933{
Paul E. McKenneyfe15d702012-01-04 13:30:33 -08001934 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
1935 !lock_is_held(&rcu_lock_map) &&
1936 !lock_is_held(&rcu_sched_lock_map),
1937 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001938 if (rcu_blocking_is_gp())
1939 return;
Paul E. McKenney2c428182011-05-26 22:14:36 -07001940 wait_rcu_gp(call_rcu_bh);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08001941}
1942EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
1943
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08001944static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0);
1945static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0);
1946
1947static int synchronize_sched_expedited_cpu_stop(void *data)
1948{
1949 /*
1950 * There must be a full memory barrier on each affected CPU
1951 * between the time that try_stop_cpus() is called and the
1952 * time that it returns.
1953 *
1954 * In the current initial implementation of cpu_stop, the
1955 * above condition is already met when the control reaches
1956 * this point and the following smp_mb() is not strictly
1957 * necessary. Do smp_mb() anyway for documentation and
1958 * robustness against future implementation changes.
1959 */
1960 smp_mb(); /* See above comment block. */
1961 return 0;
1962}
1963
Paul E. McKenney236fefa2012-01-31 14:00:41 -08001964/**
1965 * synchronize_sched_expedited - Brute-force RCU-sched grace period
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08001966 *
Paul E. McKenney236fefa2012-01-31 14:00:41 -08001967 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
1968 * approach to force the grace period to end quickly. This consumes
1969 * significant time on all CPUs and is unfriendly to real-time workloads,
1970 * so is thus not recommended for any sort of common-case code. In fact,
1971 * if you are using synchronize_sched_expedited() in a loop, please
1972 * restructure your code to batch your updates, and then use a single
1973 * synchronize_sched() instead.
1974 *
1975 * Note that it is illegal to call this function while holding any lock
1976 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
1977 * to call this function from a CPU-hotplug notifier. Failing to observe
1978 * these restriction will result in deadlock.
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08001979 *
1980 * This implementation can be thought of as an application of ticket
1981 * locking to RCU, with sync_sched_expedited_started and
1982 * sync_sched_expedited_done taking on the roles of the halves
1983 * of the ticket-lock word. Each task atomically increments
1984 * sync_sched_expedited_started upon entry, snapshotting the old value,
1985 * then attempts to stop all the CPUs. If this succeeds, then each
1986 * CPU will have executed a context switch, resulting in an RCU-sched
1987 * grace period. We are then done, so we use atomic_cmpxchg() to
1988 * update sync_sched_expedited_done to match our snapshot -- but
1989 * only if someone else has not already advanced past our snapshot.
1990 *
1991 * On the other hand, if try_stop_cpus() fails, we check the value
1992 * of sync_sched_expedited_done. If it has advanced past our
1993 * initial snapshot, then someone else must have forced a grace period
1994 * some time after we took our snapshot. In this case, our work is
1995 * done for us, and we can simply return. Otherwise, we try again,
1996 * but keep our initial snapshot for purposes of checking for someone
1997 * doing our work for us.
1998 *
1999 * If we fail too many times in a row, we fall back to synchronize_sched().
2000 */
2001void synchronize_sched_expedited(void)
2002{
2003 int firstsnap, s, snap, trycount = 0;
2004
2005 /* Note that atomic_inc_return() implies full memory barrier. */
2006 firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started);
2007 get_online_cpus();
2008 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2009
2010 /*
2011 * Each pass through the following loop attempts to force a
2012 * context switch on each CPU.
2013 */
2014 while (try_stop_cpus(cpu_online_mask,
2015 synchronize_sched_expedited_cpu_stop,
2016 NULL) == -EAGAIN) {
2017 put_online_cpus();
2018
2019 /* No joy, try again later. Or just synchronize_sched(). */
2020 if (trycount++ < 10)
2021 udelay(trycount * num_online_cpus());
2022 else {
2023 synchronize_sched();
2024 return;
2025 }
2026
2027 /* Check to see if someone else did our work for us. */
2028 s = atomic_read(&sync_sched_expedited_done);
2029 if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) {
2030 smp_mb(); /* ensure test happens before caller kfree */
2031 return;
2032 }
2033
2034 /*
2035 * Refetching sync_sched_expedited_started allows later
2036 * callers to piggyback on our grace period. We subtract
2037 * 1 to get the same token that the last incrementer got.
2038 * We retry after they started, so our grace period works
2039 * for them, and they started after our first try, so their
2040 * grace period works for us.
2041 */
2042 get_online_cpus();
2043 snap = atomic_read(&sync_sched_expedited_started);
2044 smp_mb(); /* ensure read is before try_stop_cpus(). */
2045 }
2046
2047 /*
2048 * Everyone up to our most recent fetch is covered by our grace
2049 * period. Update the counter, but only if our work is still
2050 * relevant -- which it won't be if someone who started later
2051 * than we did beat us to the punch.
2052 */
2053 do {
2054 s = atomic_read(&sync_sched_expedited_done);
2055 if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) {
2056 smp_mb(); /* ensure test happens before caller kfree */
2057 break;
2058 }
2059 } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s);
2060
2061 put_online_cpus();
2062}
2063EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2064
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002065/*
2066 * Check to see if there is any immediate RCU-related work to be done
2067 * by the current CPU, for the specified type of RCU, returning 1 if so.
2068 * The checks are in order of increasing expense: checks that can be
2069 * carried out against CPU-local state are performed first. However,
2070 * we must check for CPU stalls first, else we might not get a chance.
2071 */
2072static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2073{
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002074 struct rcu_node *rnp = rdp->mynode;
2075
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002076 rdp->n_rcu_pending++;
2077
2078 /* Check for CPU stalls, if enabled. */
2079 check_cpu_stall(rsp, rdp);
2080
2081 /* Is the RCU core waiting for a quiescent state from this CPU? */
Paul E. McKenney5c51dd72011-08-04 06:59:03 -07002082 if (rcu_scheduler_fully_active &&
2083 rdp->qs_pending && !rdp->passed_quiesce) {
Paul E. McKenneyd25eb942010-03-18 21:36:51 -07002084
2085 /*
2086 * If force_quiescent_state() coming soon and this CPU
2087 * needs a quiescent state, and this is either RCU-sched
2088 * or RCU-bh, force a local reschedule.
2089 */
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07002090 rdp->n_rp_qs_pending++;
Paul E. McKenney6cc68792011-03-02 13:15:15 -08002091 if (!rdp->preemptible &&
Paul E. McKenneyd25eb942010-03-18 21:36:51 -07002092 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs) - 1,
2093 jiffies))
2094 set_need_resched();
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07002095 } else if (rdp->qs_pending && rdp->passed_quiesce) {
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07002096 rdp->n_rp_report_qs++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002097 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002098 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002099
2100 /* Does this CPU have callbacks ready to invoke? */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002101 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2102 rdp->n_rp_cb_ready++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002103 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002104 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002105
2106 /* Has RCU gone idle with this CPU needing another grace period? */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002107 if (cpu_needs_another_gp(rsp, rdp)) {
2108 rdp->n_rp_cpu_needs_gp++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002109 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002110 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002111
2112 /* Has another RCU grace period completed? */
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002113 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002114 rdp->n_rp_gp_completed++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002115 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002116 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002117
2118 /* Has a new RCU grace period started? */
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002119 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002120 rdp->n_rp_gp_started++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002121 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002122 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002123
2124 /* Has an RCU GP gone long enough to send resched IPIs &c? */
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -07002125 if (rcu_gp_in_progress(rsp) &&
Paul E. McKenney20133cf2010-02-22 17:05:01 -08002126 ULONG_CMP_LT(ACCESS_ONCE(rsp->jiffies_force_qs), jiffies)) {
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002127 rdp->n_rp_need_fqs++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002128 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002129 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002130
2131 /* nothing to do */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002132 rdp->n_rp_need_nothing++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002133 return 0;
2134}
2135
2136/*
2137 * Check to see if there is any immediate RCU-related work to be done
2138 * by the current CPU, returning 1 if so. This function is part of the
2139 * RCU implementation; it is -not- an exported member of the RCU API.
2140 */
Paul E. McKenneya1572292009-08-22 13:56:51 -07002141static int rcu_pending(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002142{
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002143 return __rcu_pending(&rcu_sched_state, &per_cpu(rcu_sched_data, cpu)) ||
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002144 __rcu_pending(&rcu_bh_state, &per_cpu(rcu_bh_data, cpu)) ||
2145 rcu_preempt_pending(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002146}
2147
2148/*
2149 * Check to see if any future RCU-related work will need to be done
2150 * by the current CPU, even if none need be done immediately, returning
Paul E. McKenney8bd93a22010-02-22 17:04:59 -08002151 * 1 if so.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002152 */
Paul E. McKenneyaea1b352011-11-02 06:54:54 -07002153static int rcu_cpu_has_callbacks(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002154{
2155 /* RCU callbacks either ready or pending? */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002156 return per_cpu(rcu_sched_data, cpu).nxtlist ||
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002157 per_cpu(rcu_bh_data, cpu).nxtlist ||
Paul E. McKenney30fbcc92012-01-12 11:01:14 -08002158 rcu_preempt_cpu_has_callbacks(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002159}
2160
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002161static DEFINE_PER_CPU(struct rcu_head, rcu_barrier_head) = {NULL};
2162static atomic_t rcu_barrier_cpu_count;
2163static DEFINE_MUTEX(rcu_barrier_mutex);
2164static struct completion rcu_barrier_completion;
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002165
2166static void rcu_barrier_callback(struct rcu_head *notused)
2167{
2168 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2169 complete(&rcu_barrier_completion);
2170}
2171
2172/*
2173 * Called with preemption disabled, and from cross-cpu IRQ context.
2174 */
2175static void rcu_barrier_func(void *type)
2176{
2177 int cpu = smp_processor_id();
2178 struct rcu_head *head = &per_cpu(rcu_barrier_head, cpu);
2179 void (*call_rcu_func)(struct rcu_head *head,
2180 void (*func)(struct rcu_head *head));
2181
2182 atomic_inc(&rcu_barrier_cpu_count);
2183 call_rcu_func = type;
2184 call_rcu_func(head, rcu_barrier_callback);
2185}
2186
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002187/*
2188 * Orchestrate the specified type of RCU barrier, waiting for all
2189 * RCU callbacks of the specified type to complete.
2190 */
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002191static void _rcu_barrier(struct rcu_state *rsp,
2192 void (*call_rcu_func)(struct rcu_head *head,
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002193 void (*func)(struct rcu_head *head)))
2194{
2195 BUG_ON(in_interrupt());
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002196 /* Take mutex to serialize concurrent rcu_barrier() requests. */
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002197 mutex_lock(&rcu_barrier_mutex);
2198 init_completion(&rcu_barrier_completion);
2199 /*
2200 * Initialize rcu_barrier_cpu_count to 1, then invoke
2201 * rcu_barrier_func() on each CPU, so that each CPU also has
2202 * incremented rcu_barrier_cpu_count. Only then is it safe to
2203 * decrement rcu_barrier_cpu_count -- otherwise the first CPU
2204 * might complete its grace period before all of the other CPUs
2205 * did their increment, causing this function to return too
Paul E. McKenney2d999e02010-10-20 12:06:18 -07002206 * early. Note that on_each_cpu() disables irqs, which prevents
2207 * any CPUs from coming online or going offline until each online
2208 * CPU has queued its RCU-barrier callback.
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002209 */
2210 atomic_set(&rcu_barrier_cpu_count, 1);
2211 on_each_cpu(rcu_barrier_func, (void *)call_rcu_func, 1);
2212 if (atomic_dec_and_test(&rcu_barrier_cpu_count))
2213 complete(&rcu_barrier_completion);
2214 wait_for_completion(&rcu_barrier_completion);
2215 mutex_unlock(&rcu_barrier_mutex);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002216}
2217
2218/**
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002219 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2220 */
2221void rcu_barrier_bh(void)
2222{
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002223 _rcu_barrier(&rcu_bh_state, call_rcu_bh);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002224}
2225EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2226
2227/**
2228 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2229 */
2230void rcu_barrier_sched(void)
2231{
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002232 _rcu_barrier(&rcu_sched_state, call_rcu_sched);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002233}
2234EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2235
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002236/*
Paul E. McKenney27569622009-08-15 09:53:46 -07002237 * Do boot-time initialization of a CPU's per-CPU RCU data.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002238 */
Paul E. McKenney27569622009-08-15 09:53:46 -07002239static void __init
2240rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002241{
2242 unsigned long flags;
2243 int i;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002244 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenney27569622009-08-15 09:53:46 -07002245 struct rcu_node *rnp = rcu_get_root(rsp);
2246
2247 /* Set up local state, ensuring consistent view of global state. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002248 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney27569622009-08-15 09:53:46 -07002249 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
2250 rdp->nxtlist = NULL;
2251 for (i = 0; i < RCU_NEXT_SIZE; i++)
2252 rdp->nxttail[i] = &rdp->nxtlist;
Paul E. McKenney486e2592012-01-06 14:11:30 -08002253 rdp->qlen_lazy = 0;
Paul E. McKenney27569622009-08-15 09:53:46 -07002254 rdp->qlen = 0;
Paul E. McKenney27569622009-08-15 09:53:46 -07002255 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
Paul E. McKenney4145fa72011-10-31 15:01:54 -07002256 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_NESTING);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07002257 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
Paul E. McKenney27569622009-08-15 09:53:46 -07002258 rdp->cpu = cpu;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07002259 rdp->rsp = rsp;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002260 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney27569622009-08-15 09:53:46 -07002261}
2262
2263/*
2264 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2265 * offline event can be happening at a given time. Note also that we
2266 * can accept some slop in the rsp->completed access due to the fact
2267 * that this CPU cannot possibly have any RCU callbacks in flight yet.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002268 */
2269static void __cpuinit
Paul E. McKenney6cc68792011-03-02 13:15:15 -08002270rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002271{
2272 unsigned long flags;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002273 unsigned long mask;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002274 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002275 struct rcu_node *rnp = rcu_get_root(rsp);
2276
2277 /* Set up local state, ensuring consistent view of global state. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002278 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002279 rdp->beenonline = 1; /* We have now been online. */
Paul E. McKenney6cc68792011-03-02 13:15:15 -08002280 rdp->preemptible = preemptible;
Paul E. McKenney37c72e52009-10-14 10:15:55 -07002281 rdp->qlen_last_fqs_check = 0;
2282 rdp->n_force_qs_snap = rsp->n_force_qs;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002283 rdp->blimit = blimit;
Paul E. McKenneyc92b1312011-11-23 13:38:58 -08002284 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_NESTING;
2285 atomic_set(&rdp->dynticks->dynticks,
2286 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
Paul E. McKenney7cb92492011-11-28 12:28:34 -08002287 rcu_prepare_for_idle_init(cpu);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002288 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002289
2290 /*
2291 * A new grace period might start here. If so, we won't be part
2292 * of it, but that is OK, as we are currently in a quiescent state.
2293 */
2294
2295 /* Exclude any attempts to start a new GP on large systems. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002296 raw_spin_lock(&rsp->onofflock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002297
2298 /* Add CPU to rcu_node bitmasks. */
2299 rnp = rdp->mynode;
2300 mask = rdp->grpmask;
2301 do {
2302 /* Exclude any attempts to start a new GP on small systems. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002303 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002304 rnp->qsmaskinit |= mask;
2305 mask = rnp->grpmask;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002306 if (rnp == rdp->mynode) {
Paul E. McKenney06ae115a2011-08-14 15:56:54 -07002307 /*
2308 * If there is a grace period in progress, we will
2309 * set up to wait for it next time we run the
2310 * RCU core code.
2311 */
2312 rdp->gpnum = rnp->completed;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002313 rdp->completed = rnp->completed;
Paul E. McKenney06ae115a2011-08-14 15:56:54 -07002314 rdp->passed_quiesce = 0;
2315 rdp->qs_pending = 0;
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07002316 rdp->passed_quiesce_gpnum = rnp->gpnum - 1;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07002317 trace_rcu_grace_period(rsp->name, rdp->gpnum, "cpuonl");
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002318 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002319 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002320 rnp = rnp->parent;
2321 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
2322
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002323 raw_spin_unlock_irqrestore(&rsp->onofflock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002324}
2325
Peter Zijlstrad72bce02011-05-30 13:34:51 +02002326static void __cpuinit rcu_prepare_cpu(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002327{
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002328 rcu_init_percpu_data(cpu, &rcu_sched_state, 0);
2329 rcu_init_percpu_data(cpu, &rcu_bh_state, 0);
2330 rcu_preempt_init_percpu_data(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002331}
2332
2333/*
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002334 * Handle CPU online/offline notification events.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002335 */
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08002336static int __cpuinit rcu_cpu_notify(struct notifier_block *self,
2337 unsigned long action, void *hcpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002338{
2339 long cpu = (long)hcpu;
Paul E. McKenney27f4d282011-02-07 12:47:15 -08002340 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002341 struct rcu_node *rnp = rdp->mynode;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002342
Paul E. McKenney300df912011-06-18 22:26:31 -07002343 trace_rcu_utilization("Start CPU hotplug");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002344 switch (action) {
2345 case CPU_UP_PREPARE:
2346 case CPU_UP_PREPARE_FROZEN:
Peter Zijlstrad72bce02011-05-30 13:34:51 +02002347 rcu_prepare_cpu(cpu);
2348 rcu_prepare_kthreads(cpu);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002349 break;
2350 case CPU_ONLINE:
Paul E. McKenney0f962a52011-04-14 12:13:53 -07002351 case CPU_DOWN_FAILED:
2352 rcu_node_kthread_setaffinity(rnp, -1);
Paul E. McKenneye3995a22011-04-18 15:31:26 -07002353 rcu_cpu_kthread_setrt(cpu, 1);
Paul E. McKenney0f962a52011-04-14 12:13:53 -07002354 break;
2355 case CPU_DOWN_PREPARE:
2356 rcu_node_kthread_setaffinity(rnp, cpu);
Paul E. McKenneye3995a22011-04-18 15:31:26 -07002357 rcu_cpu_kthread_setrt(cpu, 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002358 break;
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002359 case CPU_DYING:
2360 case CPU_DYING_FROZEN:
2361 /*
Paul E. McKenney2d999e02010-10-20 12:06:18 -07002362 * The whole machine is "stopped" except this CPU, so we can
2363 * touch any data without introducing corruption. We send the
2364 * dying CPU's callbacks to an arbitrarily chosen online CPU.
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002365 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08002366 rcu_cleanup_dying_cpu(&rcu_bh_state);
2367 rcu_cleanup_dying_cpu(&rcu_sched_state);
2368 rcu_preempt_cleanup_dying_cpu();
Paul E. McKenney7cb92492011-11-28 12:28:34 -08002369 rcu_cleanup_after_idle(cpu);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002370 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002371 case CPU_DEAD:
2372 case CPU_DEAD_FROZEN:
2373 case CPU_UP_CANCELED:
2374 case CPU_UP_CANCELED_FROZEN:
Paul E. McKenneye5601402012-01-07 11:03:57 -08002375 rcu_cleanup_dead_cpu(cpu, &rcu_bh_state);
2376 rcu_cleanup_dead_cpu(cpu, &rcu_sched_state);
2377 rcu_preempt_cleanup_dead_cpu(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002378 break;
2379 default:
2380 break;
2381 }
Paul E. McKenney300df912011-06-18 22:26:31 -07002382 trace_rcu_utilization("End CPU hotplug");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002383 return NOTIFY_OK;
2384}
2385
2386/*
Paul E. McKenneybbad9372010-04-02 16:17:17 -07002387 * This function is invoked towards the end of the scheduler's initialization
2388 * process. Before this is called, the idle task might contain
2389 * RCU read-side critical sections (during which time, this idle
2390 * task is booting the system). After this function is called, the
2391 * idle tasks are prohibited from containing RCU read-side critical
2392 * sections. This function also enables RCU lockdep checking.
2393 */
2394void rcu_scheduler_starting(void)
2395{
2396 WARN_ON(num_online_cpus() != 1);
2397 WARN_ON(nr_context_switches() > 0);
2398 rcu_scheduler_active = 1;
2399}
2400
2401/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002402 * Compute the per-level fanout, either using the exact fanout specified
2403 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
2404 */
2405#ifdef CONFIG_RCU_FANOUT_EXACT
2406static void __init rcu_init_levelspread(struct rcu_state *rsp)
2407{
2408 int i;
2409
Paul E. McKenney0209f642010-12-14 16:07:52 -08002410 for (i = NUM_RCU_LVLS - 1; i > 0; i--)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002411 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
Paul E. McKenney0209f642010-12-14 16:07:52 -08002412 rsp->levelspread[0] = RCU_FANOUT_LEAF;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002413}
2414#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
2415static void __init rcu_init_levelspread(struct rcu_state *rsp)
2416{
2417 int ccur;
2418 int cprv;
2419 int i;
2420
2421 cprv = NR_CPUS;
2422 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2423 ccur = rsp->levelcnt[i];
2424 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
2425 cprv = ccur;
2426 }
2427}
2428#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
2429
2430/*
2431 * Helper function for rcu_init() that initializes one rcu_state structure.
2432 */
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002433static void __init rcu_init_one(struct rcu_state *rsp,
2434 struct rcu_data __percpu *rda)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002435{
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08002436 static char *buf[] = { "rcu_node_level_0",
2437 "rcu_node_level_1",
2438 "rcu_node_level_2",
2439 "rcu_node_level_3" }; /* Match MAX_RCU_LVLS */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002440 int cpustride = 1;
2441 int i;
2442 int j;
2443 struct rcu_node *rnp;
2444
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08002445 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
2446
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002447 /* Initialize the level-tracking arrays. */
2448
2449 for (i = 1; i < NUM_RCU_LVLS; i++)
2450 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
2451 rcu_init_levelspread(rsp);
2452
2453 /* Initialize the elements themselves, starting from the leaves. */
2454
2455 for (i = NUM_RCU_LVLS - 1; i >= 0; i--) {
2456 cpustride *= rsp->levelspread[i];
2457 rnp = rsp->level[i];
2458 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002459 raw_spin_lock_init(&rnp->lock);
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08002460 lockdep_set_class_and_name(&rnp->lock,
2461 &rcu_node_class[i], buf[i]);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002462 rnp->gpnum = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002463 rnp->qsmask = 0;
2464 rnp->qsmaskinit = 0;
2465 rnp->grplo = j * cpustride;
2466 rnp->grphi = (j + 1) * cpustride - 1;
2467 if (rnp->grphi >= NR_CPUS)
2468 rnp->grphi = NR_CPUS - 1;
2469 if (i == 0) {
2470 rnp->grpnum = 0;
2471 rnp->grpmask = 0;
2472 rnp->parent = NULL;
2473 } else {
2474 rnp->grpnum = j % rsp->levelspread[i - 1];
2475 rnp->grpmask = 1UL << rnp->grpnum;
2476 rnp->parent = rsp->level[i - 1] +
2477 j / rsp->levelspread[i - 1];
2478 }
2479 rnp->level = i;
Paul E. McKenney12f5f522010-11-29 21:56:39 -08002480 INIT_LIST_HEAD(&rnp->blkd_tasks);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002481 }
2482 }
Lai Jiangshan0c340292010-03-28 11:12:30 +08002483
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002484 rsp->rda = rda;
Lai Jiangshan0c340292010-03-28 11:12:30 +08002485 rnp = rsp->level[NUM_RCU_LVLS - 1];
2486 for_each_possible_cpu(i) {
Paul E. McKenney4a90a062010-04-14 16:48:11 -07002487 while (i > rnp->grphi)
Lai Jiangshan0c340292010-03-28 11:12:30 +08002488 rnp++;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002489 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
Lai Jiangshan0c340292010-03-28 11:12:30 +08002490 rcu_boot_init_percpu_data(i, rsp);
2491 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002492}
2493
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08002494void __init rcu_init(void)
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002495{
Paul E. McKenney017c4262010-01-14 16:10:58 -08002496 int cpu;
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08002497
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002498 rcu_bootup_announce();
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002499 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
2500 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002501 __rcu_init_preempt();
Shaohua Li09223372011-06-14 13:26:25 +08002502 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08002503
2504 /*
2505 * We don't need protection against CPU-hotplug here because
2506 * this is called early in boot, before either interrupts
2507 * or the scheduler are operational.
2508 */
2509 cpu_notifier(rcu_cpu_notify, 0);
Paul E. McKenney017c4262010-01-14 16:10:58 -08002510 for_each_online_cpu(cpu)
2511 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Paul E. McKenneyc68de202010-04-15 10:12:40 -07002512 check_cpu_stall_init();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002513}
2514
Paul E. McKenney1eba8f82009-09-23 09:50:42 -07002515#include "rcutree_plugin.h"