blob: 32618b3fe4e6aa375b3cd3bde1acd12bdd4237b3 [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. McKenney661a85d2012-07-07 05:57:03 -070055#include <linux/random.h>
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -040056#include <linux/ftrace_event.h>
Borislav Petkovd1d74d12013-04-22 00:12:42 +020057#include <linux/suspend.h>
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010058
Paul E. McKenney9f77da92009-08-22 13:56:45 -070059#include "rcutree.h"
Paul E. McKenney29c00b42011-06-17 15:53:19 -070060#include <trace/events/rcu.h>
61
62#include "rcu.h"
Paul E. McKenney9f77da92009-08-22 13:56:45 -070063
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -040064/*
65 * Strings used in tracepoints need to be exported via the
66 * tracing system such that tools like perf and trace-cmd can
67 * translate the string address pointers to actual text.
68 */
69#define TPS(x) tracepoint_string(x)
70
Paul E. McKenney64db4cf2008-12-18 21:55:32 +010071/* Data structures. */
72
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -070073static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
Paul E. McKenney394f2762012-06-26 17:00:35 -070074static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
Peter Zijlstra88b91c72009-10-26 10:24:31 -070075
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -040076/*
77 * In order to export the rcu_state name to the tracing tools, it
78 * needs to be added in the __tracepoint_string section.
79 * This requires defining a separate variable tp_<sname>_varname
80 * that points to the string being used, and this will allow
81 * the tracing userspace tools to be able to decipher the string
82 * address to the matching string.
83 */
Steven Rostedt (Red Hat)a41bfeb2013-07-12 17:00:28 -040084#define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -040085static char sname##_varname[] = #sname; \
86static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
Steven Rostedt (Red Hat)a41bfeb2013-07-12 17:00:28 -040087struct rcu_state sname##_state = { \
Paul E. McKenney6c90cc72012-05-28 19:41:41 -070088 .level = { &sname##_state.node[0] }, \
Paul E. McKenney037b64e2012-05-28 23:26:01 -070089 .call = cr, \
Paul E. McKenneyaf446b72011-09-10 21:54:08 -070090 .fqs_state = RCU_GP_IDLE, \
Paul E. McKenney42c35332012-09-28 10:49:58 -070091 .gpnum = 0UL - 300UL, \
92 .completed = 0UL - 300UL, \
Paul E. McKenney7b2e6012012-10-08 10:54:03 -070093 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
Paul E. McKenney6c90cc72012-05-28 19:41:41 -070094 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
95 .orphan_donetail = &sname##_state.orphan_donelist, \
Paul E. McKenney7be7f0b2012-05-29 05:18:53 -070096 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
Paul E. McKenneya4fbe352012-10-07 08:36:12 -070097 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -040098 .name = sname##_varname, \
Paul E. McKenneya4889852012-12-03 08:16:28 -080099 .abbr = sabbr, \
Steven Rostedt (Red Hat)a41bfeb2013-07-12 17:00:28 -0400100}; \
101DEFINE_PER_CPU(struct rcu_data, sname##_data)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100102
Steven Rostedt (Red Hat)a41bfeb2013-07-12 17:00:28 -0400103RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
104RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100105
Paul E. McKenney27f4d282011-02-07 12:47:15 -0800106static struct rcu_state *rcu_state;
Paul E. McKenney6ce75a22012-06-12 11:01:13 -0700107LIST_HEAD(rcu_struct_flavors);
Paul E. McKenney27f4d282011-02-07 12:47:15 -0800108
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -0700109/* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
110static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
Paul E. McKenney7e5c2df2012-07-01 15:42:33 -0700111module_param(rcu_fanout_leaf, int, 0444);
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -0700112int rcu_num_lvls __read_mostly = RCU_NUM_LVLS;
113static int num_rcu_lvl[] = { /* Number of rcu_nodes at specified level. */
114 NUM_RCU_LVL_0,
115 NUM_RCU_LVL_1,
116 NUM_RCU_LVL_2,
117 NUM_RCU_LVL_3,
118 NUM_RCU_LVL_4,
119};
120int rcu_num_nodes __read_mostly = NUM_RCU_NODES; /* Total # rcu_nodes in use. */
121
Paul E. McKenneyb0d30412011-07-10 15:57:35 -0700122/*
123 * The rcu_scheduler_active variable transitions from zero to one just
124 * before the first task is spawned. So when this variable is zero, RCU
125 * can assume that there is but one task, allowing RCU to (for example)
Cody P Schaferb44f6652013-01-04 12:59:40 -0500126 * optimize synchronize_sched() to a simple barrier(). When this variable
Paul E. McKenneyb0d30412011-07-10 15:57:35 -0700127 * is one, RCU must actually do all the hard work required to detect real
128 * grace periods. This variable is also used to suppress boot-time false
129 * positives from lockdep-RCU error checking.
130 */
Paul E. McKenneybbad9372010-04-02 16:17:17 -0700131int rcu_scheduler_active __read_mostly;
132EXPORT_SYMBOL_GPL(rcu_scheduler_active);
133
Paul E. McKenneyb0d30412011-07-10 15:57:35 -0700134/*
135 * The rcu_scheduler_fully_active variable transitions from zero to one
136 * during the early_initcall() processing, which is after the scheduler
137 * is capable of creating new tasks. So RCU processing (for example,
138 * creating tasks for RCU priority boosting) must be delayed until after
139 * rcu_scheduler_fully_active transitions from zero to one. We also
140 * currently delay invocation of any RCU callbacks until after this point.
141 *
142 * It might later prove better for people registering RCU callbacks during
143 * early boot to take responsibility for these callbacks, but one step at
144 * a time.
145 */
146static int rcu_scheduler_fully_active __read_mostly;
147
Paul E. McKenneya46e0892011-06-15 15:47:09 -0700148#ifdef CONFIG_RCU_BOOST
149
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100150/*
Paul E. McKenneya26ac242011-01-12 14:10:23 -0800151 * Control variables for per-CPU and per-rcu_node kthreads. These
152 * handle all flavors of RCU.
153 */
154static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task);
Paul E. McKenneyd71df902011-03-29 17:48:28 -0700155DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
Paul E. McKenney5ece5ba2011-04-22 18:08:51 -0700156DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
Paul E. McKenneyd71df902011-03-29 17:48:28 -0700157DEFINE_PER_CPU(char, rcu_cpu_has_work);
Paul E. McKenneya26ac242011-01-12 14:10:23 -0800158
Paul E. McKenneya46e0892011-06-15 15:47:09 -0700159#endif /* #ifdef CONFIG_RCU_BOOST */
160
Thomas Gleixner5d01bbd2012-07-16 10:42:35 +0000161static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
Paul E. McKenneya46e0892011-06-15 15:47:09 -0700162static void invoke_rcu_core(void);
163static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
Paul E. McKenneya26ac242011-01-12 14:10:23 -0800164
Paul E. McKenneya26ac242011-01-12 14:10:23 -0800165/*
Paul E. McKenney4a298652011-04-03 21:33:51 -0700166 * Track the rcutorture test sequence number and the update version
167 * number within a given test. The rcutorture_testseq is incremented
168 * on every rcutorture module load and unload, so has an odd value
169 * when a test is running. The rcutorture_vernum is set to zero
170 * when rcutorture starts and is incremented on each rcutorture update.
171 * These variables enable correlating rcutorture output with the
172 * RCU tracing information.
173 */
174unsigned long rcutorture_testseq;
175unsigned long rcutorture_vernum;
176
177/*
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -0700178 * Return true if an RCU grace period is in progress. The ACCESS_ONCE()s
179 * permit this function to be invoked without holding the root rcu_node
180 * structure's ->lock, but of course results can be subject to change.
181 */
182static int rcu_gp_in_progress(struct rcu_state *rsp)
183{
184 return ACCESS_ONCE(rsp->completed) != ACCESS_ONCE(rsp->gpnum);
185}
186
187/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700188 * Note a quiescent state. Because we do not need to know
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100189 * how many quiescent states passed, just if there was at least
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700190 * one since the start of the grace period, this just sets a flag.
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700191 * The caller must have disabled preemption.
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100192 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700193void rcu_sched_qs(int cpu)
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100194{
Paul E. McKenney25502a62010-04-01 17:37:01 -0700195 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700196
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700197 if (rdp->passed_quiesce == 0)
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400198 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700199 rdp->passed_quiesce = 1;
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100200}
201
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700202void rcu_bh_qs(int cpu)
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100203{
Paul E. McKenney25502a62010-04-01 17:37:01 -0700204 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -0700205
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700206 if (rdp->passed_quiesce == 0)
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400207 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700208 rdp->passed_quiesce = 1;
Ingo Molnarb1f77b02009-03-13 03:20:49 +0100209}
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100210
Paul E. McKenney25502a62010-04-01 17:37:01 -0700211/*
212 * Note a context switch. This is a quiescent state for RCU-sched,
213 * and requires special handling for preemptible RCU.
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -0700214 * The caller must have disabled preemption.
Paul E. McKenney25502a62010-04-01 17:37:01 -0700215 */
216void rcu_note_context_switch(int cpu)
217{
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400218 trace_rcu_utilization(TPS("Start context switch"));
Paul E. McKenney25502a62010-04-01 17:37:01 -0700219 rcu_sched_qs(cpu);
Paul E. McKenneycba6d0d2012-07-02 07:08:42 -0700220 rcu_preempt_note_context_switch(cpu);
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400221 trace_rcu_utilization(TPS("End context switch"));
Paul E. McKenney25502a62010-04-01 17:37:01 -0700222}
Gleb Natapov29ce8312011-05-04 16:31:03 +0300223EXPORT_SYMBOL_GPL(rcu_note_context_switch);
Paul E. McKenney25502a62010-04-01 17:37:01 -0700224
Paul E. McKenney90a4d2c2009-01-04 11:41:11 -0800225DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
Paul E. McKenney29e37d82011-11-17 16:55:56 -0800226 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700227 .dynticks = ATOMIC_INIT(1),
Paul E. McKenney23332102013-06-21 12:34:33 -0700228#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
229 .dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE,
230 .dynticks_idle = ATOMIC_INIT(1),
231#endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
Paul E. McKenney90a4d2c2009-01-04 11:41:11 -0800232};
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100233
Eric Dumazet878d7432012-10-18 04:55:36 -0700234static long blimit = 10; /* Maximum callbacks per rcu_do_batch. */
235static long qhimark = 10000; /* If this many pending, ignore blimit. */
236static long qlowmark = 100; /* Once only this many pending, use blimit. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100237
Eric Dumazet878d7432012-10-18 04:55:36 -0700238module_param(blimit, long, 0444);
239module_param(qhimark, long, 0444);
240module_param(qlowmark, long, 0444);
Paul E. McKenney3d76c082009-09-28 07:46:32 -0700241
Paul E. McKenney026ad282013-04-03 22:14:11 -0700242static ulong jiffies_till_first_fqs = ULONG_MAX;
243static ulong jiffies_till_next_fqs = ULONG_MAX;
Paul E. McKenneyd40011f2012-06-26 20:45:57 -0700244
245module_param(jiffies_till_first_fqs, ulong, 0644);
246module_param(jiffies_till_next_fqs, ulong, 0644);
247
Paul E. McKenney910ee452012-12-31 02:24:21 -0800248static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
249 struct rcu_data *rdp);
Paul E. McKenney217af2a2013-06-21 15:39:06 -0700250static void force_qs_rnp(struct rcu_state *rsp,
251 int (*f)(struct rcu_data *rsp, bool *isidle,
252 unsigned long *maxj),
253 bool *isidle, unsigned long *maxj);
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -0700254static void force_quiescent_state(struct rcu_state *rsp);
Paul E. McKenneya1572292009-08-22 13:56:51 -0700255static int rcu_pending(int cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100256
257/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700258 * Return the number of RCU-sched batches processed thus far for debug & stats.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100259 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700260long rcu_batches_completed_sched(void)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100261{
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700262 return rcu_sched_state.completed;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100263}
Paul E. McKenneyd6714c22009-08-22 13:56:46 -0700264EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100265
266/*
267 * Return the number of RCU BH batches processed thus far for debug & stats.
268 */
269long rcu_batches_completed_bh(void)
270{
271 return rcu_bh_state.completed;
272}
273EXPORT_SYMBOL_GPL(rcu_batches_completed_bh);
274
275/*
Paul E. McKenneybf66f182010-01-04 15:09:10 -0800276 * Force a quiescent state for RCU BH.
277 */
278void rcu_bh_force_quiescent_state(void)
279{
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -0700280 force_quiescent_state(&rcu_bh_state);
Paul E. McKenneybf66f182010-01-04 15:09:10 -0800281}
282EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
283
284/*
Paul E. McKenney4a298652011-04-03 21:33:51 -0700285 * Record the number of times rcutorture tests have been initiated and
286 * terminated. This information allows the debugfs tracing stats to be
287 * correlated to the rcutorture messages, even when the rcutorture module
288 * is being repeatedly loaded and unloaded. In other words, we cannot
289 * store this state in rcutorture itself.
290 */
291void rcutorture_record_test_transition(void)
292{
293 rcutorture_testseq++;
294 rcutorture_vernum = 0;
295}
296EXPORT_SYMBOL_GPL(rcutorture_record_test_transition);
297
298/*
299 * Record the number of writer passes through the current rcutorture test.
300 * This is also used to correlate debugfs tracing stats with the rcutorture
301 * messages.
302 */
303void rcutorture_record_progress(unsigned long vernum)
304{
305 rcutorture_vernum++;
306}
307EXPORT_SYMBOL_GPL(rcutorture_record_progress);
308
309/*
Paul E. McKenneybf66f182010-01-04 15:09:10 -0800310 * Force a quiescent state for RCU-sched.
311 */
312void rcu_sched_force_quiescent_state(void)
313{
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -0700314 force_quiescent_state(&rcu_sched_state);
Paul E. McKenneybf66f182010-01-04 15:09:10 -0800315}
316EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
317
318/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100319 * Does the CPU have callbacks ready to be invoked?
320 */
321static int
322cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
323{
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -0700324 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
325 rdp->nxttail[RCU_DONE_TAIL] != NULL;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100326}
327
328/*
Paul E. McKenneydc35c892012-12-03 13:52:00 -0800329 * Does the current CPU require a not-yet-started grace period?
330 * The caller must have disabled interrupts to prevent races with
331 * normal callback registry.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100332 */
333static int
334cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
335{
Paul E. McKenneydc35c892012-12-03 13:52:00 -0800336 int i;
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -0700337
Paul E. McKenneydc35c892012-12-03 13:52:00 -0800338 if (rcu_gp_in_progress(rsp))
339 return 0; /* No, a grace period is already in progress. */
Paul E. McKenneydae6e642013-02-10 20:48:58 -0800340 if (rcu_nocb_needs_gp(rsp))
Paul E. McKenney34ed62462013-01-07 13:37:42 -0800341 return 1; /* Yes, a no-CBs CPU needs one. */
Paul E. McKenneydc35c892012-12-03 13:52:00 -0800342 if (!rdp->nxttail[RCU_NEXT_TAIL])
343 return 0; /* No, this is a no-CBs (or offline) CPU. */
344 if (*rdp->nxttail[RCU_NEXT_READY_TAIL])
345 return 1; /* Yes, this CPU has newly registered callbacks. */
346 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++)
347 if (rdp->nxttail[i - 1] != rdp->nxttail[i] &&
348 ULONG_CMP_LT(ACCESS_ONCE(rsp->completed),
349 rdp->nxtcompleted[i]))
350 return 1; /* Yes, CBs for future grace period. */
351 return 0; /* No grace period needed. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100352}
353
354/*
355 * Return the root node of the specified rcu_state structure.
356 */
357static struct rcu_node *rcu_get_root(struct rcu_state *rsp)
358{
359 return &rsp->node[0];
360}
361
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100362/*
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700363 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100364 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700365 * If the new value of the ->dynticks_nesting counter now is zero,
366 * we really have entered idle, and must do the appropriate accounting.
367 * The caller must have disabled interrupts.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100368 */
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700369static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
370 bool user)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100371{
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400372 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
Paul E. McKenneycb349ca2012-09-04 17:35:31 -0700373 if (!user && !is_idle_task(current)) {
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700374 struct task_struct *idle = idle_task(smp_processor_id());
375
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400376 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
Paul E. McKenneybf1304e2012-05-09 15:55:39 -0700377 ftrace_dump(DUMP_ORIG);
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700378 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
379 current->pid, current->comm,
380 idle->pid, idle->comm); /* must be idle task! */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700381 }
Paul E. McKenneyaea1b352011-11-02 06:54:54 -0700382 rcu_prepare_for_idle(smp_processor_id());
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700383 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
384 smp_mb__before_atomic_inc(); /* See above. */
385 atomic_inc(&rdtp->dynticks);
386 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
387 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
Paul E. McKenneyc44e2cd2012-01-12 13:08:18 -0800388
389 /*
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700390 * It is illegal to enter an extended quiescent state while
Paul E. McKenneyc44e2cd2012-01-12 13:08:18 -0800391 * in an RCU read-side critical section.
392 */
393 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
394 "Illegal idle entry in RCU read-side critical section.");
395 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
396 "Illegal idle entry in RCU-bh read-side critical section.");
397 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
398 "Illegal idle entry in RCU-sched read-side critical section.");
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100399}
400
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700401/*
402 * Enter an RCU extended quiescent state, which can be either the
403 * idle loop or adaptive-tickless usermode execution.
404 */
405static void rcu_eqs_enter(bool user)
406{
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700407 long long oldval;
408 struct rcu_dynticks *rdtp;
409
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700410 rdtp = &__get_cpu_var(rcu_dynticks);
411 oldval = rdtp->dynticks_nesting;
412 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
413 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
414 rdtp->dynticks_nesting = 0;
415 else
416 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
417 rcu_eqs_enter_common(rdtp, oldval, user);
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700418}
419
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700420/**
421 * rcu_idle_enter - inform RCU that current CPU is entering idle
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100422 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700423 * Enter idle mode, in other words, -leave- the mode in which RCU
424 * read-side critical sections can occur. (Though RCU read-side
425 * critical sections can occur in irq handlers in idle, a possibility
426 * handled by irq_enter() and irq_exit().)
427 *
428 * We crowbar the ->dynticks_nesting field to zero to allow for
429 * the possibility of usermode upcalls having messed up our count
430 * of interrupt nesting level during the prior busy period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100431 */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700432void rcu_idle_enter(void)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100433{
Frederic Weisbeckerc5d900b2012-07-11 20:26:31 +0200434 unsigned long flags;
435
436 local_irq_save(flags);
Paul E. McKenneycb349ca2012-09-04 17:35:31 -0700437 rcu_eqs_enter(false);
Paul E. McKenneyeb348b82013-06-21 13:00:57 -0700438 rcu_sysidle_enter(&__get_cpu_var(rcu_dynticks), 0);
Frederic Weisbeckerc5d900b2012-07-11 20:26:31 +0200439 local_irq_restore(flags);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700440}
Paul E. McKenney8a2ecf42012-02-02 15:42:04 -0800441EXPORT_SYMBOL_GPL(rcu_idle_enter);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700442
Frederic Weisbecker2b1d5022012-07-11 20:26:30 +0200443#ifdef CONFIG_RCU_USER_QS
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700444/**
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700445 * rcu_user_enter - inform RCU that we are resuming userspace.
446 *
447 * Enter RCU idle mode right before resuming userspace. No use of RCU
448 * is permitted between this call and rcu_user_exit(). This way the
449 * CPU doesn't need to maintain the tick for RCU maintenance purposes
450 * when the CPU runs in userspace.
451 */
452void rcu_user_enter(void)
453{
Frederic Weisbecker91d1aa432012-11-27 19:33:25 +0100454 rcu_eqs_enter(1);
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700455}
Frederic Weisbecker2b1d5022012-07-11 20:26:30 +0200456#endif /* CONFIG_RCU_USER_QS */
Frederic Weisbecker19dd15912012-06-04 16:42:35 -0700457
458/**
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700459 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
460 *
461 * Exit from an interrupt handler, which might possibly result in entering
462 * idle mode, in other words, leaving the mode in which read-side critical
463 * sections can occur.
464 *
465 * This code assumes that the idle loop never does anything that might
466 * result in unbalanced calls to irq_enter() and irq_exit(). If your
467 * architecture violates this assumption, RCU will give you what you
468 * deserve, good and hard. But very infrequently and irreproducibly.
469 *
470 * Use things like work queues to work around this limitation.
471 *
472 * You have been warned.
473 */
474void rcu_irq_exit(void)
475{
476 unsigned long flags;
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700477 long long oldval;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700478 struct rcu_dynticks *rdtp;
479
480 local_irq_save(flags);
481 rdtp = &__get_cpu_var(rcu_dynticks);
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700482 oldval = rdtp->dynticks_nesting;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700483 rdtp->dynticks_nesting--;
484 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
Frederic Weisbeckerb6fc6022011-11-28 16:18:56 -0800485 if (rdtp->dynticks_nesting)
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400486 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
Frederic Weisbeckerb6fc6022011-11-28 16:18:56 -0800487 else
Paul E. McKenneycb349ca2012-09-04 17:35:31 -0700488 rcu_eqs_enter_common(rdtp, oldval, true);
Paul E. McKenneyeb348b82013-06-21 13:00:57 -0700489 rcu_sysidle_enter(rdtp, 1);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700490 local_irq_restore(flags);
491}
492
493/*
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700494 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700495 *
496 * If the new value of the ->dynticks_nesting counter was previously zero,
497 * we really have exited idle, and must do the appropriate accounting.
498 * The caller must have disabled interrupts.
499 */
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700500static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
501 int user)
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700502{
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700503 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
504 atomic_inc(&rdtp->dynticks);
505 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
506 smp_mb__after_atomic_inc(); /* See above. */
507 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
Paul E. McKenney7cb92492011-11-28 12:28:34 -0800508 rcu_cleanup_after_idle(smp_processor_id());
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400509 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
Paul E. McKenneycb349ca2012-09-04 17:35:31 -0700510 if (!user && !is_idle_task(current)) {
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700511 struct task_struct *idle = idle_task(smp_processor_id());
512
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400513 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700514 oldval, rdtp->dynticks_nesting);
Paul E. McKenneybf1304e2012-05-09 15:55:39 -0700515 ftrace_dump(DUMP_ORIG);
Paul E. McKenney0989cb42011-11-01 08:57:21 -0700516 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
517 current->pid, current->comm,
518 idle->pid, idle->comm); /* must be idle task! */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700519 }
520}
521
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700522/*
523 * Exit an RCU extended quiescent state, which can be either the
524 * idle loop or adaptive-tickless usermode execution.
525 */
526static void rcu_eqs_exit(bool user)
527{
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700528 struct rcu_dynticks *rdtp;
529 long long oldval;
530
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700531 rdtp = &__get_cpu_var(rcu_dynticks);
532 oldval = rdtp->dynticks_nesting;
533 WARN_ON_ONCE(oldval < 0);
534 if (oldval & DYNTICK_TASK_NEST_MASK)
535 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
536 else
537 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
538 rcu_eqs_exit_common(rdtp, oldval, user);
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700539}
540
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700541/**
542 * rcu_idle_exit - inform RCU that current CPU is leaving idle
543 *
544 * Exit idle mode, in other words, -enter- the mode in which RCU
545 * read-side critical sections can occur.
546 *
Paul E. McKenney29e37d82011-11-17 16:55:56 -0800547 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
Paul E. McKenney4145fa72011-10-31 15:01:54 -0700548 * allow for the possibility of usermode upcalls messing up our count
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700549 * of interrupt nesting level during the busy period that is just
550 * now starting.
551 */
552void rcu_idle_exit(void)
553{
Frederic Weisbeckerc5d900b2012-07-11 20:26:31 +0200554 unsigned long flags;
555
556 local_irq_save(flags);
Paul E. McKenneycb349ca2012-09-04 17:35:31 -0700557 rcu_eqs_exit(false);
Paul E. McKenneyeb348b82013-06-21 13:00:57 -0700558 rcu_sysidle_exit(&__get_cpu_var(rcu_dynticks), 0);
Frederic Weisbeckerc5d900b2012-07-11 20:26:31 +0200559 local_irq_restore(flags);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700560}
Paul E. McKenney8a2ecf42012-02-02 15:42:04 -0800561EXPORT_SYMBOL_GPL(rcu_idle_exit);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700562
Frederic Weisbecker2b1d5022012-07-11 20:26:30 +0200563#ifdef CONFIG_RCU_USER_QS
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700564/**
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700565 * rcu_user_exit - inform RCU that we are exiting userspace.
566 *
567 * Exit RCU idle mode while entering the kernel because it can
568 * run a RCU read side critical section anytime.
569 */
570void rcu_user_exit(void)
571{
Frederic Weisbecker91d1aa432012-11-27 19:33:25 +0100572 rcu_eqs_exit(1);
Frederic Weisbeckeradf50912012-06-28 11:20:21 -0700573}
Frederic Weisbecker2b1d5022012-07-11 20:26:30 +0200574#endif /* CONFIG_RCU_USER_QS */
Frederic Weisbecker19dd15912012-06-04 16:42:35 -0700575
576/**
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700577 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
578 *
579 * Enter an interrupt handler, which might possibly result in exiting
580 * idle mode, in other words, entering the mode in which read-side critical
581 * sections can occur.
582 *
583 * Note that the Linux kernel is fully capable of entering an interrupt
584 * handler that it never exits, for example when doing upcalls to
585 * user mode! This code assumes that the idle loop never does upcalls to
586 * user mode. If your architecture does do upcalls from the idle loop (or
587 * does anything else that results in unbalanced calls to the irq_enter()
588 * and irq_exit() functions), RCU will give you what you deserve, good
589 * and hard. But very infrequently and irreproducibly.
590 *
591 * Use things like work queues to work around this limitation.
592 *
593 * You have been warned.
594 */
595void rcu_irq_enter(void)
596{
597 unsigned long flags;
598 struct rcu_dynticks *rdtp;
599 long long oldval;
600
601 local_irq_save(flags);
602 rdtp = &__get_cpu_var(rcu_dynticks);
603 oldval = rdtp->dynticks_nesting;
604 rdtp->dynticks_nesting++;
605 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
Frederic Weisbeckerb6fc6022011-11-28 16:18:56 -0800606 if (oldval)
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400607 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
Frederic Weisbeckerb6fc6022011-11-28 16:18:56 -0800608 else
Paul E. McKenneycb349ca2012-09-04 17:35:31 -0700609 rcu_eqs_exit_common(rdtp, oldval, true);
Paul E. McKenneyeb348b82013-06-21 13:00:57 -0700610 rcu_sysidle_exit(rdtp, 1);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100611 local_irq_restore(flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100612}
613
614/**
615 * rcu_nmi_enter - inform RCU of entry to NMI context
616 *
617 * If the CPU was idle with dynamic ticks active, and there is no
618 * irq handler running, this updates rdtp->dynticks_nmi to let the
619 * RCU grace-period handling know that the CPU is active.
620 */
621void rcu_nmi_enter(void)
622{
623 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
624
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700625 if (rdtp->dynticks_nmi_nesting == 0 &&
626 (atomic_read(&rdtp->dynticks) & 0x1))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100627 return;
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700628 rdtp->dynticks_nmi_nesting++;
629 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
630 atomic_inc(&rdtp->dynticks);
631 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
632 smp_mb__after_atomic_inc(); /* See above. */
633 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100634}
635
636/**
637 * rcu_nmi_exit - inform RCU of exit from NMI context
638 *
639 * If the CPU was idle with dynamic ticks active, and there is no
640 * irq handler running, this updates rdtp->dynticks_nmi to let the
641 * RCU grace-period handling know that the CPU is no longer active.
642 */
643void rcu_nmi_exit(void)
644{
645 struct rcu_dynticks *rdtp = &__get_cpu_var(rcu_dynticks);
646
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700647 if (rdtp->dynticks_nmi_nesting == 0 ||
648 --rdtp->dynticks_nmi_nesting != 0)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100649 return;
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700650 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
651 smp_mb__before_atomic_inc(); /* See above. */
652 atomic_inc(&rdtp->dynticks);
653 smp_mb__after_atomic_inc(); /* Force delay to next write. */
654 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100655}
656
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100657/**
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700658 * rcu_is_cpu_idle - see if RCU thinks that the current CPU is idle
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100659 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700660 * If the current CPU is in its idle loop and is neither in an interrupt
Paul E. McKenney34240692011-10-03 11:38:52 -0700661 * or NMI handler, return true.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100662 */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700663int rcu_is_cpu_idle(void)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100664{
Paul E. McKenney34240692011-10-03 11:38:52 -0700665 int ret;
666
667 preempt_disable();
668 ret = (atomic_read(&__get_cpu_var(rcu_dynticks).dynticks) & 0x1) == 0;
669 preempt_enable();
670 return ret;
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700671}
Frederic Weisbeckere6b80a32011-10-07 16:25:18 -0700672EXPORT_SYMBOL(rcu_is_cpu_idle);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700673
Paul E. McKenney62fde6e2012-05-22 22:10:24 -0700674#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800675
676/*
677 * Is the current CPU online? Disable preemption to avoid false positives
678 * that could otherwise happen due to the current CPU number being sampled,
679 * this task being preempted, its old CPU being taken offline, resuming
680 * on some other CPU, then determining that its old CPU is now offline.
681 * It is OK to use RCU on an offline processor during initial boot, hence
Paul E. McKenney2036d942012-01-30 17:02:47 -0800682 * the check for rcu_scheduler_fully_active. Note also that it is OK
683 * for a CPU coming online to use RCU for one jiffy prior to marking itself
684 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
685 * offline to continue to use RCU for one jiffy after marking itself
686 * offline in the cpu_online_mask. This leniency is necessary given the
687 * non-atomic nature of the online and offline processing, for example,
688 * the fact that a CPU enters the scheduler after completing the CPU_DYING
689 * notifiers.
690 *
691 * This is also why RCU internally marks CPUs online during the
692 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800693 *
694 * Disable checking if in an NMI handler because we cannot safely report
695 * errors from NMI handlers anyway.
696 */
697bool rcu_lockdep_current_cpu_online(void)
698{
Paul E. McKenney2036d942012-01-30 17:02:47 -0800699 struct rcu_data *rdp;
700 struct rcu_node *rnp;
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800701 bool ret;
702
703 if (in_nmi())
704 return 1;
705 preempt_disable();
Paul E. McKenney2036d942012-01-30 17:02:47 -0800706 rdp = &__get_cpu_var(rcu_sched_data);
707 rnp = rdp->mynode;
708 ret = (rdp->grpmask & rnp->qsmaskinit) ||
Paul E. McKenneyc0d6d012012-01-23 12:41:26 -0800709 !rcu_scheduler_fully_active;
710 preempt_enable();
711 return ret;
712}
713EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
714
Paul E. McKenney62fde6e2012-05-22 22:10:24 -0700715#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700716
717/**
718 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
719 *
720 * If the current CPU is idle or running at a first-level (not nested)
721 * interrupt from idle, return true. The caller must have at least
722 * disabled preemption.
723 */
Josh Triplett62e3cb12012-11-20 09:55:26 -0800724static int rcu_is_cpu_rrupt_from_idle(void)
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -0700725{
726 return __get_cpu_var(rcu_dynticks).dynticks_nesting <= 1;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100727}
728
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100729/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100730 * Snapshot the specified CPU's dynticks counter so that we can later
731 * credit them with an implicit quiescent state. Return 1 if this CPU
Paul E. McKenney1eba8f82009-09-23 09:50:42 -0700732 * is in dynticks idle mode, which is an extended quiescent state.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100733 */
Paul E. McKenney217af2a2013-06-21 15:39:06 -0700734static int dyntick_save_progress_counter(struct rcu_data *rdp,
735 bool *isidle, unsigned long *maxj)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100736{
Paul E. McKenney23b5c8f2010-09-07 10:38:22 -0700737 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
Paul E. McKenney0edd1b12013-06-21 16:37:22 -0700738 rcu_sysidle_check_cpu(rdp, isidle, maxj);
Paul E. McKenneyf0e7c192011-11-23 15:02:05 -0800739 return (rdp->dynticks_snap & 0x1) == 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100740}
741
742/*
743 * Return true if the specified CPU has passed through a quiescent
744 * state by virtue of being in or having passed through an dynticks
745 * idle state since the last call to dyntick_save_progress_counter()
Paul E. McKenneya82dcc72012-08-01 14:29:20 -0700746 * for this same CPU, or by virtue of having been offline.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100747 */
Paul E. McKenney217af2a2013-06-21 15:39:06 -0700748static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
749 bool *isidle, unsigned long *maxj)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100750{
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700751 unsigned int curr;
752 unsigned int snap;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100753
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700754 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
755 snap = (unsigned int)rdp->dynticks_snap;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100756
757 /*
758 * If the CPU passed through or entered a dynticks idle phase with
759 * no active irq/NMI handlers, then we can safely pretend that the CPU
760 * already acknowledged the request to pass through a quiescent
761 * state. Either way, that CPU cannot possibly be in an RCU
762 * read-side critical section that started before the beginning
763 * of the current RCU grace period.
764 */
Paul E. McKenney7eb4f452011-07-30 07:32:48 -0700765 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400766 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100767 rdp->dynticks_fqs++;
768 return 1;
769 }
770
Paul E. McKenneya82dcc72012-08-01 14:29:20 -0700771 /*
772 * Check for the CPU being offline, but only if the grace period
773 * is old enough. We don't need to worry about the CPU changing
774 * state: If we see it offline even once, it has been through a
775 * quiescent state.
776 *
777 * The reason for insisting that the grace period be at least
778 * one jiffy old is that CPUs that are not quite online and that
779 * have just gone offline can still execute RCU read-side critical
780 * sections.
781 */
782 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
783 return 0; /* Grace period is not old enough. */
784 barrier();
785 if (cpu_is_offline(rdp->cpu)) {
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -0400786 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
Paul E. McKenneya82dcc72012-08-01 14:29:20 -0700787 rdp->offline_fqs++;
788 return 1;
789 }
Paul E. McKenney65d798f2013-04-12 16:19:10 -0700790
791 /*
792 * There is a possibility that a CPU in adaptive-ticks state
793 * might run in the kernel with the scheduling-clock tick disabled
794 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
795 * force the CPU to restart the scheduling-clock tick in this
796 * CPU is in this state.
797 */
798 rcu_kick_nohz_cpu(rdp->cpu);
799
Paul E. McKenneya82dcc72012-08-01 14:29:20 -0700800 return 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100801}
802
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100803static void record_gp_stall_check_time(struct rcu_state *rsp)
804{
805 rsp->gp_start = jiffies;
Paul E. McKenney6bfc09e2012-10-19 12:49:17 -0700806 rsp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100807}
808
Paul E. McKenneyb637a322012-09-19 16:58:38 -0700809/*
810 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
811 * for architectures that do not implement trigger_all_cpu_backtrace().
812 * The NMI-triggered stack traces are more accurate because they are
813 * printed by the target CPU.
814 */
815static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
816{
817 int cpu;
818 unsigned long flags;
819 struct rcu_node *rnp;
820
821 rcu_for_each_leaf_node(rsp, rnp) {
822 raw_spin_lock_irqsave(&rnp->lock, flags);
823 if (rnp->qsmask != 0) {
824 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
825 if (rnp->qsmask & (1UL << cpu))
826 dump_cpu_task(rnp->grplo + cpu);
827 }
828 raw_spin_unlock_irqrestore(&rnp->lock, flags);
829 }
830}
831
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100832static void print_other_cpu_stall(struct rcu_state *rsp)
833{
834 int cpu;
835 long delta;
836 unsigned long flags;
Paul E. McKenney285fe292012-05-09 08:45:12 -0700837 int ndetected = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100838 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney53bb8572012-09-21 16:35:25 -0700839 long totqlen = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100840
841 /* Only let one CPU complain about others per time interval. */
842
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800843 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100844 delta = jiffies - rsp->jiffies_stall;
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -0700845 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800846 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100847 return;
848 }
Paul E. McKenney6bfc09e2012-10-19 12:49:17 -0700849 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800850 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100851
Paul E. McKenney8cdd32a2010-08-09 14:23:03 -0700852 /*
853 * OK, time to rat on our buddy...
854 * See Documentation/RCU/stallwarn.txt for info on how to debug
855 * RCU CPU stall warnings.
856 */
Paul E. McKenneyd7f3e202013-03-18 16:19:52 -0700857 pr_err("INFO: %s detected stalls on CPUs/tasks:",
Paul E. McKenney4300aa62010-04-13 16:18:22 -0700858 rsp->name);
Paul E. McKenneya858af22012-01-16 13:29:10 -0800859 print_cpu_stall_info_begin();
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -0700860 rcu_for_each_leaf_node(rsp, rnp) {
Paul E. McKenney3acd9eb2010-02-22 17:05:03 -0800861 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700862 ndetected += rcu_print_task_stall(rnp);
Paul E. McKenneyc8020a62012-08-10 16:55:59 -0700863 if (rnp->qsmask != 0) {
864 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
865 if (rnp->qsmask & (1UL << cpu)) {
866 print_cpu_stall_info(rsp,
867 rnp->grplo + cpu);
868 ndetected++;
869 }
870 }
Paul E. McKenney3acd9eb2010-02-22 17:05:03 -0800871 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100872 }
Paul E. McKenneya858af22012-01-16 13:29:10 -0800873
874 /*
875 * Now rat on any tasks that got kicked up to the root rcu_node
876 * due to CPU offlining.
877 */
878 rnp = rcu_get_root(rsp);
879 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney285fe292012-05-09 08:45:12 -0700880 ndetected += rcu_print_task_stall(rnp);
Paul E. McKenneya858af22012-01-16 13:29:10 -0800881 raw_spin_unlock_irqrestore(&rnp->lock, flags);
882
883 print_cpu_stall_info_end();
Paul E. McKenney53bb8572012-09-21 16:35:25 -0700884 for_each_possible_cpu(cpu)
885 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
886 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
Paul E. McKenneyeee05882012-09-21 14:15:05 -0700887 smp_processor_id(), (long)(jiffies - rsp->gp_start),
Paul E. McKenney53bb8572012-09-21 16:35:25 -0700888 rsp->gpnum, rsp->completed, totqlen);
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700889 if (ndetected == 0)
Paul E. McKenneyd7f3e202013-03-18 16:19:52 -0700890 pr_err("INFO: Stall ended before state dump start\n");
Paul E. McKenney9bc8b552011-08-13 13:31:47 -0700891 else if (!trigger_all_cpu_backtrace())
Paul E. McKenneyb637a322012-09-19 16:58:38 -0700892 rcu_dump_cpu_stacks(rsp);
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200893
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -0700894 /* Complain about tasks blocking the grace period. */
Paul E. McKenney1ed509a2010-02-22 17:05:05 -0800895
896 rcu_print_detail_task_stall(rsp);
897
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -0700898 force_quiescent_state(rsp); /* Kick them all. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100899}
900
901static void print_cpu_stall(struct rcu_state *rsp)
902{
Paul E. McKenney53bb8572012-09-21 16:35:25 -0700903 int cpu;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100904 unsigned long flags;
905 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenney53bb8572012-09-21 16:35:25 -0700906 long totqlen = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100907
Paul E. McKenney8cdd32a2010-08-09 14:23:03 -0700908 /*
909 * OK, time to rat on ourselves...
910 * See Documentation/RCU/stallwarn.txt for info on how to debug
911 * RCU CPU stall warnings.
912 */
Paul E. McKenneyd7f3e202013-03-18 16:19:52 -0700913 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
Paul E. McKenneya858af22012-01-16 13:29:10 -0800914 print_cpu_stall_info_begin();
915 print_cpu_stall_info(rsp, smp_processor_id());
916 print_cpu_stall_info_end();
Paul E. McKenney53bb8572012-09-21 16:35:25 -0700917 for_each_possible_cpu(cpu)
918 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
919 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
920 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
Paul E. McKenney4627e242011-08-03 03:34:24 -0700921 if (!trigger_all_cpu_backtrace())
922 dump_stack();
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200923
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800924 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney20133cf2010-02-22 17:05:01 -0800925 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
Paul E. McKenney13cfcca2012-01-13 15:32:18 -0800926 rsp->jiffies_stall = jiffies +
Paul E. McKenney6bfc09e2012-10-19 12:49:17 -0700927 3 * rcu_jiffies_till_stall_check() + 3;
Paul E. McKenney1304afb2010-02-22 17:05:02 -0800928 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Ingo Molnarc1dc0b92009-08-02 11:28:21 +0200929
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100930 set_need_resched(); /* kick ourselves to get things going. */
931}
932
933static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
934{
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700935 unsigned long j;
936 unsigned long js;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100937 struct rcu_node *rnp;
938
Paul E. McKenney742734e2010-06-30 11:43:52 -0700939 if (rcu_cpu_stall_suppress)
Paul E. McKenneyc68de202010-04-15 10:12:40 -0700940 return;
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700941 j = ACCESS_ONCE(jiffies);
942 js = ACCESS_ONCE(rsp->jiffies_stall);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100943 rnp = rdp->mynode;
Paul E. McKenneyc96ea7c2012-08-13 11:17:06 -0700944 if (rcu_gp_in_progress(rsp) &&
945 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask) && ULONG_CMP_GE(j, js)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100946
947 /* We haven't checked in, so go dump stack. */
948 print_cpu_stall(rsp);
949
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700950 } else if (rcu_gp_in_progress(rsp) &&
951 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100952
Paul E. McKenneybad6e132011-05-02 23:40:04 -0700953 /* They had a few time units to dump stack, so complain. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100954 print_other_cpu_stall(rsp);
955 }
956}
957
Paul E. McKenney53d84e02010-08-10 14:28:53 -0700958/**
959 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
960 *
961 * Set the stall-warning timeout way off into the future, thus preventing
962 * any RCU CPU stall-warning messages from appearing in the current set of
963 * RCU grace periods.
964 *
965 * The caller must disable hard irqs.
966 */
967void rcu_cpu_stall_reset(void)
968{
Paul E. McKenney6ce75a22012-06-12 11:01:13 -0700969 struct rcu_state *rsp;
970
971 for_each_rcu_flavor(rsp)
972 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
Paul E. McKenney53d84e02010-08-10 14:28:53 -0700973}
974
Paul E. McKenney64db4cf2008-12-18 21:55:32 +0100975/*
Paul E. McKenney3f5d3ea2012-05-09 15:39:56 -0700976 * Initialize the specified rcu_data structure's callback list to empty.
977 */
978static void init_callback_list(struct rcu_data *rdp)
979{
980 int i;
981
Paul E. McKenney34ed62462013-01-07 13:37:42 -0800982 if (init_nocb_callback_list(rdp))
983 return;
Paul E. McKenney3f5d3ea2012-05-09 15:39:56 -0700984 rdp->nxtlist = NULL;
985 for (i = 0; i < RCU_NEXT_SIZE; i++)
986 rdp->nxttail[i] = &rdp->nxtlist;
987}
988
989/*
Paul E. McKenneydc35c892012-12-03 13:52:00 -0800990 * Determine the value that ->completed will have at the end of the
991 * next subsequent grace period. This is used to tag callbacks so that
992 * a CPU can invoke callbacks in a timely fashion even if that CPU has
993 * been dyntick-idle for an extended period with callbacks under the
994 * influence of RCU_FAST_NO_HZ.
995 *
996 * The caller must hold rnp->lock with interrupts disabled.
997 */
998static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
999 struct rcu_node *rnp)
1000{
1001 /*
1002 * If RCU is idle, we just wait for the next grace period.
1003 * But we can only be sure that RCU is idle if we are looking
1004 * at the root rcu_node structure -- otherwise, a new grace
1005 * period might have started, but just not yet gotten around
1006 * to initializing the current non-root rcu_node structure.
1007 */
1008 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1009 return rnp->completed + 1;
1010
1011 /*
1012 * Otherwise, wait for a possible partial grace period and
1013 * then the subsequent full grace period.
1014 */
1015 return rnp->completed + 2;
1016}
1017
1018/*
Paul E. McKenney0446be42012-12-30 15:21:01 -08001019 * Trace-event helper function for rcu_start_future_gp() and
1020 * rcu_nocb_wait_gp().
1021 */
1022static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
Steven Rostedt (Red Hat)e66c33d2013-07-12 16:50:28 -04001023 unsigned long c, const char *s)
Paul E. McKenney0446be42012-12-30 15:21:01 -08001024{
1025 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1026 rnp->completed, c, rnp->level,
1027 rnp->grplo, rnp->grphi, s);
1028}
1029
1030/*
1031 * Start some future grace period, as needed to handle newly arrived
1032 * callbacks. The required future grace periods are recorded in each
1033 * rcu_node structure's ->need_future_gp field.
1034 *
1035 * The caller must hold the specified rcu_node structure's ->lock.
1036 */
1037static unsigned long __maybe_unused
1038rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1039{
1040 unsigned long c;
1041 int i;
1042 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1043
1044 /*
1045 * Pick up grace-period number for new callbacks. If this
1046 * grace period is already marked as needed, return to the caller.
1047 */
1048 c = rcu_cbs_completed(rdp->rsp, rnp);
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001049 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
Paul E. McKenney0446be42012-12-30 15:21:01 -08001050 if (rnp->need_future_gp[c & 0x1]) {
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001051 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
Paul E. McKenney0446be42012-12-30 15:21:01 -08001052 return c;
1053 }
1054
1055 /*
1056 * If either this rcu_node structure or the root rcu_node structure
1057 * believe that a grace period is in progress, then we must wait
1058 * for the one following, which is in "c". Because our request
1059 * will be noticed at the end of the current grace period, we don't
1060 * need to explicitly start one.
1061 */
1062 if (rnp->gpnum != rnp->completed ||
1063 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1064 rnp->need_future_gp[c & 0x1]++;
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001065 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
Paul E. McKenney0446be42012-12-30 15:21:01 -08001066 return c;
1067 }
1068
1069 /*
1070 * There might be no grace period in progress. If we don't already
1071 * hold it, acquire the root rcu_node structure's lock in order to
1072 * start one (if needed).
1073 */
1074 if (rnp != rnp_root)
1075 raw_spin_lock(&rnp_root->lock);
1076
1077 /*
1078 * Get a new grace-period number. If there really is no grace
1079 * period in progress, it will be smaller than the one we obtained
1080 * earlier. Adjust callbacks as needed. Note that even no-CBs
1081 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1082 */
1083 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1084 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1085 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1086 rdp->nxtcompleted[i] = c;
1087
1088 /*
1089 * If the needed for the required grace period is already
1090 * recorded, trace and leave.
1091 */
1092 if (rnp_root->need_future_gp[c & 0x1]) {
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001093 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
Paul E. McKenney0446be42012-12-30 15:21:01 -08001094 goto unlock_out;
1095 }
1096
1097 /* Record the need for the future grace period. */
1098 rnp_root->need_future_gp[c & 0x1]++;
1099
1100 /* If a grace period is not already in progress, start one. */
1101 if (rnp_root->gpnum != rnp_root->completed) {
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001102 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
Paul E. McKenney0446be42012-12-30 15:21:01 -08001103 } else {
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001104 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
Paul E. McKenney910ee452012-12-31 02:24:21 -08001105 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
Paul E. McKenney0446be42012-12-30 15:21:01 -08001106 }
1107unlock_out:
1108 if (rnp != rnp_root)
1109 raw_spin_unlock(&rnp_root->lock);
1110 return c;
1111}
1112
1113/*
1114 * Clean up any old requests for the just-ended grace period. Also return
1115 * whether any additional grace periods have been requested. Also invoke
1116 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1117 * waiting for this grace period to complete.
1118 */
1119static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1120{
1121 int c = rnp->completed;
1122 int needmore;
1123 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1124
1125 rcu_nocb_gp_cleanup(rsp, rnp);
1126 rnp->need_future_gp[c & 0x1] = 0;
1127 needmore = rnp->need_future_gp[(c + 1) & 0x1];
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001128 trace_rcu_future_gp(rnp, rdp, c,
1129 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
Paul E. McKenney0446be42012-12-30 15:21:01 -08001130 return needmore;
1131}
1132
1133/*
Paul E. McKenneydc35c892012-12-03 13:52:00 -08001134 * If there is room, assign a ->completed number to any callbacks on
1135 * this CPU that have not already been assigned. Also accelerate any
1136 * callbacks that were previously assigned a ->completed number that has
1137 * since proven to be too conservative, which can happen if callbacks get
1138 * assigned a ->completed number while RCU is idle, but with reference to
1139 * a non-root rcu_node structure. This function is idempotent, so it does
1140 * not hurt to call it repeatedly.
1141 *
1142 * The caller must hold rnp->lock with interrupts disabled.
1143 */
1144static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1145 struct rcu_data *rdp)
1146{
1147 unsigned long c;
1148 int i;
1149
1150 /* If the CPU has no callbacks, nothing to do. */
1151 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1152 return;
1153
1154 /*
1155 * Starting from the sublist containing the callbacks most
1156 * recently assigned a ->completed number and working down, find the
1157 * first sublist that is not assignable to an upcoming grace period.
1158 * Such a sublist has something in it (first two tests) and has
1159 * a ->completed number assigned that will complete sooner than
1160 * the ->completed number for newly arrived callbacks (last test).
1161 *
1162 * The key point is that any later sublist can be assigned the
1163 * same ->completed number as the newly arrived callbacks, which
1164 * means that the callbacks in any of these later sublist can be
1165 * grouped into a single sublist, whether or not they have already
1166 * been assigned a ->completed number.
1167 */
1168 c = rcu_cbs_completed(rsp, rnp);
1169 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1170 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1171 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1172 break;
1173
1174 /*
1175 * If there are no sublist for unassigned callbacks, leave.
1176 * At the same time, advance "i" one sublist, so that "i" will
1177 * index into the sublist where all the remaining callbacks should
1178 * be grouped into.
1179 */
1180 if (++i >= RCU_NEXT_TAIL)
1181 return;
1182
1183 /*
1184 * Assign all subsequent callbacks' ->completed number to the next
1185 * full grace period and group them all in the sublist initially
1186 * indexed by "i".
1187 */
1188 for (; i <= RCU_NEXT_TAIL; i++) {
1189 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1190 rdp->nxtcompleted[i] = c;
1191 }
Paul E. McKenney910ee452012-12-31 02:24:21 -08001192 /* Record any needed additional grace periods. */
1193 rcu_start_future_gp(rnp, rdp);
Paul E. McKenney6d4b4182012-11-27 16:55:44 -08001194
1195 /* Trace depending on how much we were able to accelerate. */
1196 if (!*rdp->nxttail[RCU_WAIT_TAIL])
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001197 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
Paul E. McKenney6d4b4182012-11-27 16:55:44 -08001198 else
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001199 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
Paul E. McKenneydc35c892012-12-03 13:52:00 -08001200}
1201
1202/*
1203 * Move any callbacks whose grace period has completed to the
1204 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1205 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1206 * sublist. This function is idempotent, so it does not hurt to
1207 * invoke it repeatedly. As long as it is not invoked -too- often...
1208 *
1209 * The caller must hold rnp->lock with interrupts disabled.
1210 */
1211static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1212 struct rcu_data *rdp)
1213{
1214 int i, j;
1215
1216 /* If the CPU has no callbacks, nothing to do. */
1217 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1218 return;
1219
1220 /*
1221 * Find all callbacks whose ->completed numbers indicate that they
1222 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1223 */
1224 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1225 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1226 break;
1227 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1228 }
1229 /* Clean up any sublist tail pointers that were misordered above. */
1230 for (j = RCU_WAIT_TAIL; j < i; j++)
1231 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1232
1233 /* Copy down callbacks to fill in empty sublists. */
1234 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1235 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1236 break;
1237 rdp->nxttail[j] = rdp->nxttail[i];
1238 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1239 }
1240
1241 /* Classify any remaining callbacks. */
1242 rcu_accelerate_cbs(rsp, rnp, rdp);
1243}
1244
1245/*
Paul E. McKenneyba9fbe92013-03-19 11:53:31 -07001246 * Update CPU-local rcu_data state to record the beginnings and ends of
1247 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1248 * structure corresponding to the current CPU, and must have irqs disabled.
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001249 */
Paul E. McKenneyba9fbe92013-03-19 11:53:31 -07001250static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001251{
Paul E. McKenneyba9fbe92013-03-19 11:53:31 -07001252 /* Handle the ends of any preceding grace periods first. */
Paul E. McKenneydc35c892012-12-03 13:52:00 -08001253 if (rdp->completed == rnp->completed) {
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001254
Paul E. McKenneyba9fbe92013-03-19 11:53:31 -07001255 /* No grace period end, so just accelerate recent callbacks. */
Paul E. McKenneydc35c892012-12-03 13:52:00 -08001256 rcu_accelerate_cbs(rsp, rnp, rdp);
1257
1258 } else {
1259
1260 /* Advance callbacks. */
1261 rcu_advance_cbs(rsp, rnp, rdp);
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001262
1263 /* Remember that we saw this grace-period completion. */
1264 rdp->completed = rnp->completed;
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001265 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08001266 }
Paul E. McKenney398ebe62013-03-19 10:53:14 -07001267
Paul E. McKenney6eaef632013-03-19 10:08:37 -07001268 if (rdp->gpnum != rnp->gpnum) {
1269 /*
1270 * If the current grace period is waiting for this CPU,
1271 * set up to detect a quiescent state, otherwise don't
1272 * go looking for one.
1273 */
1274 rdp->gpnum = rnp->gpnum;
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001275 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
Paul E. McKenney6eaef632013-03-19 10:08:37 -07001276 rdp->passed_quiesce = 0;
1277 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1278 zero_cpu_stall_ticks(rdp);
1279 }
1280}
1281
Paul E. McKenneyd34ea322013-03-19 11:10:43 -07001282static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenney6eaef632013-03-19 10:08:37 -07001283{
1284 unsigned long flags;
1285 struct rcu_node *rnp;
1286
1287 local_irq_save(flags);
1288 rnp = rdp->mynode;
Paul E. McKenneyd34ea322013-03-19 11:10:43 -07001289 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1290 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
Paul E. McKenney6eaef632013-03-19 10:08:37 -07001291 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1292 local_irq_restore(flags);
1293 return;
1294 }
Paul E. McKenneyd34ea322013-03-19 11:10:43 -07001295 __note_gp_changes(rsp, rnp, rdp);
Paul E. McKenney6eaef632013-03-19 10:08:37 -07001296 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1297}
1298
1299/*
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001300 * Initialize a new grace period.
1301 */
1302static int rcu_gp_init(struct rcu_state *rsp)
1303{
1304 struct rcu_data *rdp;
1305 struct rcu_node *rnp = rcu_get_root(rsp);
1306
Paul E. McKenneyeb757672013-06-21 17:10:40 -07001307 rcu_bind_gp_kthread();
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001308 raw_spin_lock_irq(&rnp->lock);
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001309 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001310
1311 if (rcu_gp_in_progress(rsp)) {
1312 /* Grace period already in progress, don't start another. */
1313 raw_spin_unlock_irq(&rnp->lock);
1314 return 0;
1315 }
1316
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001317 /* Advance to a new grace period and initialize state. */
1318 rsp->gpnum++;
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001319 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001320 record_gp_stall_check_time(rsp);
1321 raw_spin_unlock_irq(&rnp->lock);
1322
1323 /* Exclude any concurrent CPU-hotplug operations. */
Paul E. McKenneya4fbe352012-10-07 08:36:12 -07001324 mutex_lock(&rsp->onoff_mutex);
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001325
1326 /*
1327 * Set the quiescent-state-needed bits in all the rcu_node
1328 * structures for all currently online CPUs in breadth-first order,
1329 * starting from the root rcu_node structure, relying on the layout
1330 * of the tree within the rsp->node[] array. Note that other CPUs
1331 * will access only the leaves of the hierarchy, thus seeing that no
1332 * grace period is in progress, at least until the corresponding
1333 * leaf node has been initialized. In addition, we have excluded
1334 * CPU-hotplug operations.
1335 *
1336 * The grace period cannot complete until the initialization
1337 * process finishes, because this kthread handles both.
1338 */
1339 rcu_for_each_node_breadth_first(rsp, rnp) {
1340 raw_spin_lock_irq(&rnp->lock);
1341 rdp = this_cpu_ptr(rsp->rda);
1342 rcu_preempt_check_blocked_tasks(rnp);
1343 rnp->qsmask = rnp->qsmaskinit;
Paul E. McKenney0446be42012-12-30 15:21:01 -08001344 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
Paul E. McKenney25d30cf2012-07-11 05:23:18 -07001345 WARN_ON_ONCE(rnp->completed != rsp->completed);
Paul E. McKenney0446be42012-12-30 15:21:01 -08001346 ACCESS_ONCE(rnp->completed) = rsp->completed;
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001347 if (rnp == rdp->mynode)
Paul E. McKenneyce3d9c02013-03-19 12:27:50 -07001348 __note_gp_changes(rsp, rnp, rdp);
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001349 rcu_preempt_boost_start_gp(rnp);
1350 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1351 rnp->level, rnp->grplo,
1352 rnp->grphi, rnp->qsmask);
1353 raw_spin_unlock_irq(&rnp->lock);
Paul E. McKenney661a85d2012-07-07 05:57:03 -07001354#ifdef CONFIG_PROVE_RCU_DELAY
Paul E. McKenney971394f2013-06-02 07:13:57 -07001355 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
Paul E. McKenney81e59492013-03-10 15:44:52 -07001356 system_state == SYSTEM_RUNNING)
Paul E. McKenney971394f2013-06-02 07:13:57 -07001357 udelay(200);
Paul E. McKenney661a85d2012-07-07 05:57:03 -07001358#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001359 cond_resched();
1360 }
1361
Paul E. McKenneya4fbe352012-10-07 08:36:12 -07001362 mutex_unlock(&rsp->onoff_mutex);
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001363 return 1;
1364}
1365
1366/*
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001367 * Do one round of quiescent-state forcing.
1368 */
1369int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1370{
1371 int fqs_state = fqs_state_in;
Paul E. McKenney217af2a2013-06-21 15:39:06 -07001372 bool isidle = false;
1373 unsigned long maxj;
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001374 struct rcu_node *rnp = rcu_get_root(rsp);
1375
1376 rsp->n_force_qs++;
1377 if (fqs_state == RCU_SAVE_DYNTICK) {
1378 /* Collect dyntick-idle snapshots. */
Paul E. McKenney0edd1b12013-06-21 16:37:22 -07001379 if (is_sysidle_rcu_state(rsp)) {
1380 isidle = 1;
1381 maxj = jiffies - ULONG_MAX / 4;
1382 }
Paul E. McKenney217af2a2013-06-21 15:39:06 -07001383 force_qs_rnp(rsp, dyntick_save_progress_counter,
1384 &isidle, &maxj);
Paul E. McKenney0edd1b12013-06-21 16:37:22 -07001385 rcu_sysidle_report_gp(rsp, isidle, maxj);
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001386 fqs_state = RCU_FORCE_QS;
1387 } else {
1388 /* Handle dyntick-idle and offline CPUs. */
Paul E. McKenney0edd1b12013-06-21 16:37:22 -07001389 isidle = 0;
Paul E. McKenney217af2a2013-06-21 15:39:06 -07001390 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001391 }
1392 /* Clear flag to prevent immediate re-entry. */
1393 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1394 raw_spin_lock_irq(&rnp->lock);
1395 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1396 raw_spin_unlock_irq(&rnp->lock);
1397 }
1398 return fqs_state;
1399}
1400
1401/*
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001402 * Clean up after the old grace period.
1403 */
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001404static void rcu_gp_cleanup(struct rcu_state *rsp)
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001405{
1406 unsigned long gp_duration;
Paul E. McKenneydae6e642013-02-10 20:48:58 -08001407 int nocb = 0;
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001408 struct rcu_data *rdp;
1409 struct rcu_node *rnp = rcu_get_root(rsp);
1410
1411 raw_spin_lock_irq(&rnp->lock);
1412 gp_duration = jiffies - rsp->gp_start;
1413 if (gp_duration > rsp->gp_max)
1414 rsp->gp_max = gp_duration;
1415
1416 /*
1417 * We know the grace period is complete, but to everyone else
1418 * it appears to still be ongoing. But it is also the case
1419 * that to everyone else it looks like there is nothing that
1420 * they can do to advance the grace period. It is therefore
1421 * safe for us to drop the lock in order to mark the grace
1422 * period as completed in all of the rcu_node structures.
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001423 */
Paul E. McKenney5d4b8652012-07-07 07:56:57 -07001424 raw_spin_unlock_irq(&rnp->lock);
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001425
Paul E. McKenney5d4b8652012-07-07 07:56:57 -07001426 /*
1427 * Propagate new ->completed value to rcu_node structures so
1428 * that other CPUs don't have to wait until the start of the next
1429 * grace period to process their callbacks. This also avoids
1430 * some nasty RCU grace-period initialization races by forcing
1431 * the end of the current grace period to be completely recorded in
1432 * all of the rcu_node structures before the beginning of the next
1433 * grace period is recorded in any of the rcu_node structures.
1434 */
1435 rcu_for_each_node_breadth_first(rsp, rnp) {
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001436 raw_spin_lock_irq(&rnp->lock);
Paul E. McKenney0446be42012-12-30 15:21:01 -08001437 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
Paul E. McKenneyb11cc572012-12-17 14:21:14 -08001438 rdp = this_cpu_ptr(rsp->rda);
1439 if (rnp == rdp->mynode)
Paul E. McKenney470716f2013-03-19 11:32:11 -07001440 __note_gp_changes(rsp, rnp, rdp);
Paul E. McKenney0446be42012-12-30 15:21:01 -08001441 nocb += rcu_future_gp_cleanup(rsp, rnp);
Paul E. McKenney5d4b8652012-07-07 07:56:57 -07001442 raw_spin_unlock_irq(&rnp->lock);
1443 cond_resched();
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001444 }
Paul E. McKenney5d4b8652012-07-07 07:56:57 -07001445 rnp = rcu_get_root(rsp);
1446 raw_spin_lock_irq(&rnp->lock);
Paul E. McKenneydae6e642013-02-10 20:48:58 -08001447 rcu_nocb_gp_set(rnp, nocb);
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001448
1449 rsp->completed = rsp->gpnum; /* Declare grace period done. */
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001450 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001451 rsp->fqs_state = RCU_GP_IDLE;
Paul E. McKenney5d4b8652012-07-07 07:56:57 -07001452 rdp = this_cpu_ptr(rsp->rda);
Paul E. McKenneyb11cc572012-12-17 14:21:14 -08001453 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001454 if (cpu_needs_another_gp(rsp, rdp))
1455 rsp->gp_flags = 1;
1456 raw_spin_unlock_irq(&rnp->lock);
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001457}
1458
1459/*
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001460 * Body of kthread that handles grace periods.
1461 */
Paul E. McKenney755609a2012-06-19 17:18:20 -07001462static int __noreturn rcu_gp_kthread(void *arg)
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001463{
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001464 int fqs_state;
Paul E. McKenneyd40011f2012-06-26 20:45:57 -07001465 unsigned long j;
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001466 int ret;
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001467 struct rcu_state *rsp = arg;
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001468 struct rcu_node *rnp = rcu_get_root(rsp);
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001469
1470 for (;;) {
1471
1472 /* Handle grace-period start. */
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001473 for (;;) {
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001474 wait_event_interruptible(rsp->gp_wq,
1475 rsp->gp_flags &
1476 RCU_GP_FLAG_INIT);
1477 if ((rsp->gp_flags & RCU_GP_FLAG_INIT) &&
1478 rcu_gp_init(rsp))
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001479 break;
Paul E. McKenney7fdefc12012-06-22 11:08:41 -07001480 cond_resched();
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001481 flush_signals(current);
1482 }
Paul E. McKenneycabc49c2012-06-20 17:07:14 -07001483
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001484 /* Handle quiescent-state forcing. */
1485 fqs_state = RCU_SAVE_DYNTICK;
Paul E. McKenneyd40011f2012-06-26 20:45:57 -07001486 j = jiffies_till_first_fqs;
1487 if (j > HZ) {
1488 j = HZ;
1489 jiffies_till_first_fqs = HZ;
1490 }
Paul E. McKenneycabc49c2012-06-20 17:07:14 -07001491 for (;;) {
Paul E. McKenneyd40011f2012-06-26 20:45:57 -07001492 rsp->jiffies_force_qs = jiffies + j;
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001493 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1494 (rsp->gp_flags & RCU_GP_FLAG_FQS) ||
1495 (!ACCESS_ONCE(rnp->qsmask) &&
1496 !rcu_preempt_blocked_readers_cgp(rnp)),
Paul E. McKenneyd40011f2012-06-26 20:45:57 -07001497 j);
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001498 /* If grace period done, leave loop. */
Paul E. McKenneycabc49c2012-06-20 17:07:14 -07001499 if (!ACCESS_ONCE(rnp->qsmask) &&
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001500 !rcu_preempt_blocked_readers_cgp(rnp))
Paul E. McKenneycabc49c2012-06-20 17:07:14 -07001501 break;
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001502 /* If time for quiescent-state forcing, do it. */
1503 if (ret == 0 || (rsp->gp_flags & RCU_GP_FLAG_FQS)) {
1504 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1505 cond_resched();
1506 } else {
1507 /* Deal with stray signal. */
1508 cond_resched();
1509 flush_signals(current);
1510 }
Paul E. McKenneyd40011f2012-06-26 20:45:57 -07001511 j = jiffies_till_next_fqs;
1512 if (j > HZ) {
1513 j = HZ;
1514 jiffies_till_next_fqs = HZ;
1515 } else if (j < 1) {
1516 j = 1;
1517 jiffies_till_next_fqs = 1;
1518 }
Paul E. McKenneycabc49c2012-06-20 17:07:14 -07001519 }
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001520
1521 /* Handle grace-period end. */
1522 rcu_gp_cleanup(rsp);
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001523 }
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001524}
1525
Steven Rostedt016a8d52013-05-28 17:32:53 -04001526static void rsp_wakeup(struct irq_work *work)
1527{
1528 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1529
1530 /* Wake up rcu_gp_kthread() to start the grace period. */
1531 wake_up(&rsp->gp_wq);
1532}
1533
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001534/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001535 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1536 * in preparation for detecting the next grace period. The caller must hold
Paul E. McKenneyb8462082012-12-29 22:04:18 -08001537 * the root node's ->lock and hard irqs must be disabled.
Paul E. McKenneye5601402012-01-07 11:03:57 -08001538 *
1539 * Note that it is legal for a dying CPU (which is marked as offline) to
1540 * invoke this function. This can happen when the dying CPU reports its
1541 * quiescent state.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001542 */
1543static void
Paul E. McKenney910ee452012-12-31 02:24:21 -08001544rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1545 struct rcu_data *rdp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001546{
Paul E. McKenneyb8462082012-12-29 22:04:18 -08001547 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -08001548 /*
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001549 * Either we have not yet spawned the grace-period
Paul E. McKenney62da1922012-09-20 16:02:49 -07001550 * task, this CPU does not need another grace period,
1551 * or a grace period is already in progress.
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07001552 * Either way, don't start a new grace period.
Paul E. McKenneyb32e9eb2009-11-12 22:35:03 -08001553 */
Paul E. McKenneyafe24b12011-08-24 16:52:09 -07001554 return;
1555 }
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07001556 rsp->gp_flags = RCU_GP_FLAG_INIT;
Paul E. McKenney62da1922012-09-20 16:02:49 -07001557
Steven Rostedt016a8d52013-05-28 17:32:53 -04001558 /*
1559 * We can't do wakeups while holding the rnp->lock, as that
Paul E. McKenney1eafd312013-06-20 13:50:40 -07001560 * could cause possible deadlocks with the rq->lock. Defer
1561 * the wakeup to interrupt context. And don't bother waking
1562 * up the running kthread.
Steven Rostedt016a8d52013-05-28 17:32:53 -04001563 */
Paul E. McKenney1eafd312013-06-20 13:50:40 -07001564 if (current != rsp->gp_kthread)
1565 irq_work_queue(&rsp->wakeup_work);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001566}
1567
1568/*
Paul E. McKenney910ee452012-12-31 02:24:21 -08001569 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1570 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1571 * is invoked indirectly from rcu_advance_cbs(), which would result in
1572 * endless recursion -- or would do so if it wasn't for the self-deadlock
1573 * that is encountered beforehand.
1574 */
1575static void
1576rcu_start_gp(struct rcu_state *rsp)
1577{
1578 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1579 struct rcu_node *rnp = rcu_get_root(rsp);
1580
1581 /*
1582 * If there is no grace period in progress right now, any
1583 * callbacks we have up to this point will be satisfied by the
1584 * next grace period. Also, advancing the callbacks reduces the
1585 * probability of false positives from cpu_needs_another_gp()
1586 * resulting in pointless grace periods. So, advance callbacks
1587 * then start the grace period!
1588 */
1589 rcu_advance_cbs(rsp, rnp, rdp);
1590 rcu_start_gp_advanced(rsp, rnp, rdp);
1591}
1592
1593/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001594 * Report a full set of quiescent states to the specified rcu_state
1595 * data structure. This involves cleaning up after the prior grace
1596 * period and letting rcu_start_gp() start up the next grace period
Paul E. McKenneyb8462082012-12-29 22:04:18 -08001597 * if one is needed. Note that the caller must hold rnp->lock, which
1598 * is released before return.
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001599 */
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001600static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -07001601 __releases(rcu_get_root(rsp)->lock)
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001602{
Paul E. McKenneyfc2219d42009-09-23 09:50:41 -07001603 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
Paul E. McKenneycabc49c2012-06-20 17:07:14 -07001604 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1605 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001606}
1607
1608/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001609 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1610 * Allows quiescent states for a group of CPUs to be reported at one go
1611 * to the specified rcu_node structure, though all the CPUs in the group
1612 * must be represented by the same rcu_node structure (which need not be
1613 * a leaf rcu_node structure, though it often will be). That structure's
1614 * lock must be held upon entry, and it is released before return.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001615 */
1616static void
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001617rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1618 struct rcu_node *rnp, unsigned long flags)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001619 __releases(rnp->lock)
1620{
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001621 struct rcu_node *rnp_c;
1622
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001623 /* Walk up the rcu_node hierarchy. */
1624 for (;;) {
1625 if (!(rnp->qsmask & mask)) {
1626
1627 /* Our bit has already been cleared, so done. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001628 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001629 return;
1630 }
1631 rnp->qsmask &= ~mask;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07001632 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1633 mask, rnp->qsmask, rnp->level,
1634 rnp->grplo, rnp->grphi,
1635 !!rnp->gp_tasks);
Paul E. McKenney27f4d282011-02-07 12:47:15 -08001636 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001637
1638 /* Other bits still set at this level, so done. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001639 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001640 return;
1641 }
1642 mask = rnp->grpmask;
1643 if (rnp->parent == NULL) {
1644
1645 /* No more levels. Exit loop holding root lock. */
1646
1647 break;
1648 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001649 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001650 rnp_c = rnp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001651 rnp = rnp->parent;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001652 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney28ecd582009-09-18 09:50:17 -07001653 WARN_ON_ONCE(rnp_c->qsmask);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001654 }
1655
1656 /*
1657 * Get here if we are the last CPU to pass through a quiescent
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001658 * state for this grace period. Invoke rcu_report_qs_rsp()
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07001659 * to clean up and start the next grace period if one is needed.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001660 */
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001661 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001662}
1663
1664/*
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001665 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1666 * structure. This must be either called from the specified CPU, or
1667 * called when the specified CPU is known to be offline (and when it is
1668 * also known that no other CPU is concurrently trying to help the offline
1669 * CPU). The lastcomp argument is used to make sure we are still in the
1670 * grace period of interest. We don't want to end the current grace period
1671 * based on quiescent states detected in an earlier grace period!
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001672 */
1673static void
Paul E. McKenneyd7d6a112012-08-21 15:00:05 -07001674rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001675{
1676 unsigned long flags;
1677 unsigned long mask;
1678 struct rcu_node *rnp;
1679
1680 rnp = rdp->mynode;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001681 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenneyd7d6a112012-08-21 15:00:05 -07001682 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1683 rnp->completed == rnp->gpnum) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001684
1685 /*
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001686 * The grace period in which this quiescent state was
1687 * recorded has ended, so don't report it upwards.
1688 * We will instead need a new quiescent state that lies
1689 * within the current grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001690 */
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001691 rdp->passed_quiesce = 0; /* need qs for new gp. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001692 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001693 return;
1694 }
1695 mask = rdp->grpmask;
1696 if ((rnp->qsmask & mask) == 0) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08001697 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001698 } else {
1699 rdp->qs_pending = 0;
1700
1701 /*
1702 * This GP can't end until cpu checks in, so all of our
1703 * callbacks can be processed during the next GP.
1704 */
Paul E. McKenneydc35c892012-12-03 13:52:00 -08001705 rcu_accelerate_cbs(rsp, rnp, rdp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001706
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001707 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001708 }
1709}
1710
1711/*
1712 * Check to see if there is a new grace period of which this CPU
1713 * is not yet aware, and if so, set up local rcu_data state for it.
1714 * Otherwise, see if this CPU has just passed through its first
1715 * quiescent state for this grace period, and record that fact if so.
1716 */
1717static void
1718rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1719{
Paul E. McKenney05eb5522013-03-19 12:38:24 -07001720 /* Check for grace-period ends and beginnings. */
1721 note_gp_changes(rsp, rdp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001722
1723 /*
1724 * Does this CPU still need to do its part for current grace period?
1725 * If no, return and let the other CPUs do their part as well.
1726 */
1727 if (!rdp->qs_pending)
1728 return;
1729
1730 /*
1731 * Was there a quiescent state since the beginning of the grace
1732 * period? If no, then exit and wait for the next call.
1733 */
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07001734 if (!rdp->passed_quiesce)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001735 return;
1736
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08001737 /*
1738 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1739 * judge of that).
1740 */
Paul E. McKenneyd7d6a112012-08-21 15:00:05 -07001741 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001742}
1743
1744#ifdef CONFIG_HOTPLUG_CPU
1745
1746/*
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001747 * Send the specified CPU's RCU callbacks to the orphanage. The
1748 * specified CPU must be offline, and the caller must hold the
Paul E. McKenney7b2e6012012-10-08 10:54:03 -07001749 * ->orphan_lock.
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001750 */
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001751static void
1752rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1753 struct rcu_node *rnp, struct rcu_data *rdp)
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001754{
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07001755 /* No-CBs CPUs do not have orphanable callbacks. */
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +01001756 if (rcu_is_nocb_cpu(rdp->cpu))
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07001757 return;
1758
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001759 /*
1760 * Orphan the callbacks. First adjust the counts. This is safe
Paul E. McKenneyabfd6e52012-09-20 16:59:47 -07001761 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1762 * cannot be running now. Thus no memory barrier is required.
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001763 */
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001764 if (rdp->nxtlist != NULL) {
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001765 rsp->qlen_lazy += rdp->qlen_lazy;
1766 rsp->qlen += rdp->qlen;
1767 rdp->n_cbs_orphaned += rdp->qlen;
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001768 rdp->qlen_lazy = 0;
Paul E. McKenney1d1fb392012-05-09 15:44:42 -07001769 ACCESS_ONCE(rdp->qlen) = 0;
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001770 }
1771
1772 /*
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001773 * Next, move those callbacks still needing a grace period to
1774 * the orphanage, where some other CPU will pick them up.
1775 * Some of the callbacks might have gone partway through a grace
1776 * period, but that is too bad. They get to start over because we
1777 * cannot assume that grace periods are synchronized across CPUs.
1778 * We don't bother updating the ->nxttail[] array yet, instead
1779 * we just reset the whole thing later on.
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001780 */
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001781 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1782 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1783 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1784 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001785 }
1786
1787 /*
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001788 * Then move the ready-to-invoke callbacks to the orphanage,
1789 * where some other CPU will pick them up. These will not be
1790 * required to pass though another grace period: They are done.
Paul E. McKenneya50c3af2012-01-10 17:52:31 -08001791 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001792 if (rdp->nxtlist != NULL) {
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001793 *rsp->orphan_donetail = rdp->nxtlist;
1794 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
Paul E. McKenneye5601402012-01-07 11:03:57 -08001795 }
Lai Jiangshan29494be2010-10-20 14:13:06 +08001796
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001797 /* Finally, initialize the rcu_data structure's list to empty. */
Paul E. McKenney3f5d3ea2012-05-09 15:39:56 -07001798 init_callback_list(rdp);
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001799}
1800
1801/*
1802 * Adopt the RCU callbacks from the specified rcu_state structure's
Paul E. McKenney7b2e6012012-10-08 10:54:03 -07001803 * orphanage. The caller must hold the ->orphan_lock.
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001804 */
1805static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1806{
1807 int i;
1808 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1809
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07001810 /* No-CBs CPUs are handled specially. */
1811 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1812 return;
1813
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001814 /* Do the accounting first. */
1815 rdp->qlen_lazy += rsp->qlen_lazy;
1816 rdp->qlen += rsp->qlen;
1817 rdp->n_cbs_adopted += rsp->qlen;
Paul E. McKenney8f5af6f2012-05-04 08:31:53 -07001818 if (rsp->qlen_lazy != rsp->qlen)
1819 rcu_idle_count_callbacks_posted();
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001820 rsp->qlen_lazy = 0;
1821 rsp->qlen = 0;
1822
1823 /*
1824 * We do not need a memory barrier here because the only way we
1825 * can get here if there is an rcu_barrier() in flight is if
1826 * we are the task doing the rcu_barrier().
1827 */
1828
1829 /* First adopt the ready-to-invoke callbacks. */
1830 if (rsp->orphan_donelist != NULL) {
1831 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1832 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1833 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1834 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1835 rdp->nxttail[i] = rsp->orphan_donetail;
1836 rsp->orphan_donelist = NULL;
1837 rsp->orphan_donetail = &rsp->orphan_donelist;
1838 }
1839
1840 /* And then adopt the callbacks that still need a grace period. */
1841 if (rsp->orphan_nxtlist != NULL) {
1842 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1843 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1844 rsp->orphan_nxtlist = NULL;
1845 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1846 }
1847}
1848
1849/*
1850 * Trace the fact that this CPU is going offline.
1851 */
1852static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1853{
1854 RCU_TRACE(unsigned long mask);
1855 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1856 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1857
1858 RCU_TRACE(mask = rdp->grpmask);
Paul E. McKenneye5601402012-01-07 11:03:57 -08001859 trace_rcu_grace_period(rsp->name,
1860 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04001861 TPS("cpuofl"));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001862}
1863
1864/*
Paul E. McKenneye5601402012-01-07 11:03:57 -08001865 * The CPU has been completely removed, and some other CPU is reporting
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001866 * this fact from process context. Do the remainder of the cleanup,
1867 * including orphaning the outgoing CPU's RCU callbacks, and also
Paul E. McKenney1331e7a2012-08-02 17:43:50 -07001868 * adopting them. There can only be one CPU hotplug operation at a time,
1869 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001870 */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001871static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001872{
Paul E. McKenney2036d942012-01-30 17:02:47 -08001873 unsigned long flags;
1874 unsigned long mask;
1875 int need_report = 0;
Paul E. McKenneye5601402012-01-07 11:03:57 -08001876 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001877 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
Paul E. McKenneye5601402012-01-07 11:03:57 -08001878
Paul E. McKenney2036d942012-01-30 17:02:47 -08001879 /* Adjust any no-longer-needed kthreads. */
Thomas Gleixner5d01bbd2012-07-16 10:42:35 +00001880 rcu_boost_kthread_setaffinity(rnp, -1);
Paul E. McKenney2036d942012-01-30 17:02:47 -08001881
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001882 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
Paul E. McKenney2036d942012-01-30 17:02:47 -08001883
1884 /* Exclude any attempts to start a new grace period. */
Paul E. McKenneya4fbe352012-10-07 08:36:12 -07001885 mutex_lock(&rsp->onoff_mutex);
Paul E. McKenney7b2e6012012-10-08 10:54:03 -07001886 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
Paul E. McKenney2036d942012-01-30 17:02:47 -08001887
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08001888 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1889 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1890 rcu_adopt_orphan_cbs(rsp);
1891
Paul E. McKenney2036d942012-01-30 17:02:47 -08001892 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1893 mask = rdp->grpmask; /* rnp->grplo is constant. */
1894 do {
1895 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1896 rnp->qsmaskinit &= ~mask;
1897 if (rnp->qsmaskinit != 0) {
1898 if (rnp != rdp->mynode)
1899 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1900 break;
1901 }
1902 if (rnp == rdp->mynode)
1903 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
1904 else
1905 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1906 mask = rnp->grpmask;
1907 rnp = rnp->parent;
1908 } while (rnp != NULL);
1909
1910 /*
1911 * We still hold the leaf rcu_node structure lock here, and
1912 * irqs are still disabled. The reason for this subterfuge is
Paul E. McKenney7b2e6012012-10-08 10:54:03 -07001913 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
Paul E. McKenney2036d942012-01-30 17:02:47 -08001914 * held leads to deadlock.
1915 */
Paul E. McKenney7b2e6012012-10-08 10:54:03 -07001916 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
Paul E. McKenney2036d942012-01-30 17:02:47 -08001917 rnp = rdp->mynode;
1918 if (need_report & RCU_OFL_TASKS_NORM_GP)
1919 rcu_report_unblock_qs_rnp(rnp, flags);
1920 else
1921 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1922 if (need_report & RCU_OFL_TASKS_EXP_GP)
1923 rcu_report_exp_rnp(rsp, rnp, true);
Paul E. McKenneycf015372012-06-21 11:26:42 -07001924 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
1925 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
1926 cpu, rdp->qlen, rdp->nxtlist);
Paul E. McKenney0d8ee372012-08-03 13:16:15 -07001927 init_callback_list(rdp);
1928 /* Disallow further callbacks on this CPU. */
1929 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
Paul E. McKenneya4fbe352012-10-07 08:36:12 -07001930 mutex_unlock(&rsp->onoff_mutex);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001931}
1932
1933#else /* #ifdef CONFIG_HOTPLUG_CPU */
1934
Paul E. McKenneye5601402012-01-07 11:03:57 -08001935static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07001936{
1937}
1938
Paul E. McKenneye5601402012-01-07 11:03:57 -08001939static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001940{
1941}
1942
1943#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1944
1945/*
1946 * Invoke any RCU callbacks that have made it to the end of their grace
1947 * period. Thottle as specified by rdp->blimit.
1948 */
Paul E. McKenney37c72e52009-10-14 10:15:55 -07001949static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001950{
1951 unsigned long flags;
1952 struct rcu_head *next, *list, **tail;
Eric Dumazet878d7432012-10-18 04:55:36 -07001953 long bl, count, count_lazy;
1954 int i;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001955
Paul E. McKenneydc35c892012-12-03 13:52:00 -08001956 /* If no callbacks are ready, just return. */
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001957 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
Paul E. McKenney486e2592012-01-06 14:11:30 -08001958 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
Paul E. McKenney4968c302011-12-07 16:32:40 -08001959 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
1960 need_resched(), is_idle_task(current),
1961 rcu_is_callbacks_kthread());
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001962 return;
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001963 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001964
1965 /*
1966 * Extract the list of ready callbacks, disabling to prevent
1967 * races with call_rcu() from interrupt handlers.
1968 */
1969 local_irq_save(flags);
Paul E. McKenney8146c4e22012-01-10 14:23:29 -08001970 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
Paul E. McKenney29c00b42011-06-17 15:53:19 -07001971 bl = rdp->blimit;
Paul E. McKenney486e2592012-01-06 14:11:30 -08001972 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001973 list = rdp->nxtlist;
1974 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
1975 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1976 tail = rdp->nxttail[RCU_DONE_TAIL];
Paul E. McKenneyb41772a2012-06-21 20:50:42 -07001977 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
1978 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1979 rdp->nxttail[i] = &rdp->nxtlist;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001980 local_irq_restore(flags);
1981
1982 /* Invoke callbacks. */
Paul E. McKenney486e2592012-01-06 14:11:30 -08001983 count = count_lazy = 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001984 while (list) {
1985 next = list->next;
1986 prefetch(next);
Mathieu Desnoyers551d55a2010-04-17 08:48:42 -04001987 debug_rcu_head_unqueue(list);
Paul E. McKenney486e2592012-01-06 14:11:30 -08001988 if (__rcu_reclaim(rsp->name, list))
1989 count_lazy++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001990 list = next;
Paul E. McKenneydff16722011-11-29 15:57:13 -08001991 /* Stop only if limit reached and CPU has something to do. */
1992 if (++count >= bl &&
1993 (need_resched() ||
1994 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01001995 break;
1996 }
1997
1998 local_irq_save(flags);
Paul E. McKenney4968c302011-12-07 16:32:40 -08001999 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2000 is_idle_task(current),
2001 rcu_is_callbacks_kthread());
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002002
2003 /* Update count, and requeue any remaining callbacks. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002004 if (list != NULL) {
2005 *tail = rdp->nxtlist;
2006 rdp->nxtlist = list;
Paul E. McKenneyb41772a2012-06-21 20:50:42 -07002007 for (i = 0; i < RCU_NEXT_SIZE; i++)
2008 if (&rdp->nxtlist == rdp->nxttail[i])
2009 rdp->nxttail[i] = tail;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002010 else
2011 break;
2012 }
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002013 smp_mb(); /* List handling before counting for rcu_barrier(). */
2014 rdp->qlen_lazy -= count_lazy;
Paul E. McKenney1d1fb392012-05-09 15:44:42 -07002015 ACCESS_ONCE(rdp->qlen) -= count;
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002016 rdp->n_cbs_invoked += count;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002017
2018 /* Reinstate batch limit if we have worked down the excess. */
2019 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2020 rdp->blimit = blimit;
2021
Paul E. McKenney37c72e52009-10-14 10:15:55 -07002022 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2023 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2024 rdp->qlen_last_fqs_check = 0;
2025 rdp->n_force_qs_snap = rsp->n_force_qs;
2026 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2027 rdp->qlen_last_fqs_check = rdp->qlen;
Paul E. McKenneycfca9272012-06-25 12:54:17 -07002028 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
Paul E. McKenney37c72e52009-10-14 10:15:55 -07002029
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002030 local_irq_restore(flags);
2031
Paul E. McKenneye0f23062011-06-21 01:29:39 -07002032 /* Re-invoke RCU core processing if there are callbacks remaining. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002033 if (cpu_has_callbacks_ready_to_invoke(rdp))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07002034 invoke_rcu_core();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002035}
2036
2037/*
2038 * Check to see if this CPU is in a non-context-switch quiescent state
2039 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
Paul E. McKenneye0f23062011-06-21 01:29:39 -07002040 * Also schedule RCU core processing.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002041 *
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07002042 * This function must be called from hardirq context. It is normally
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002043 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2044 * false, there is no point in invoking rcu_check_callbacks().
2045 */
2046void rcu_check_callbacks(int cpu, int user)
2047{
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04002048 trace_rcu_utilization(TPS("Start scheduler-tick"));
Paul E. McKenneya858af22012-01-16 13:29:10 -08002049 increment_cpu_stall_ticks();
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07002050 if (user || rcu_is_cpu_rrupt_from_idle()) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002051
2052 /*
2053 * Get here if this CPU took its interrupt from user
2054 * mode or from the idle loop, and if this is not a
2055 * nested interrupt. In this case, the CPU is in
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002056 * a quiescent state, so note it.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002057 *
2058 * No memory barrier is required here because both
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002059 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2060 * variables that other CPUs neither access nor modify,
2061 * at least not while the corresponding CPU is online.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002062 */
2063
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002064 rcu_sched_qs(cpu);
2065 rcu_bh_qs(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002066
2067 } else if (!in_softirq()) {
2068
2069 /*
2070 * Get here if this CPU did not take its interrupt from
2071 * softirq, in other words, if it is not interrupting
2072 * a rcu_bh read-side critical section. This is an _bh
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002073 * critical section, so note it.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002074 */
2075
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002076 rcu_bh_qs(cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002077 }
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07002078 rcu_preempt_check_callbacks(cpu);
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07002079 if (rcu_pending(cpu))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07002080 invoke_rcu_core();
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04002081 trace_rcu_utilization(TPS("End scheduler-tick"));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002082}
2083
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002084/*
2085 * Scan the leaf rcu_node structures, processing dyntick state for any that
2086 * have not yet encountered a quiescent state, using the function specified.
Paul E. McKenney27f4d282011-02-07 12:47:15 -08002087 * Also initiate boosting for any threads blocked on the root rcu_node.
2088 *
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08002089 * The caller must have suppressed start of new grace periods.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002090 */
Paul E. McKenney217af2a2013-06-21 15:39:06 -07002091static void force_qs_rnp(struct rcu_state *rsp,
2092 int (*f)(struct rcu_data *rsp, bool *isidle,
2093 unsigned long *maxj),
2094 bool *isidle, unsigned long *maxj)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002095{
2096 unsigned long bit;
2097 int cpu;
2098 unsigned long flags;
2099 unsigned long mask;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07002100 struct rcu_node *rnp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002101
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07002102 rcu_for_each_leaf_node(rsp, rnp) {
Paul E. McKenneyb4be093f2012-06-25 08:41:11 -07002103 cond_resched();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002104 mask = 0;
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002105 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenneyee47eb92010-01-04 15:09:07 -08002106 if (!rcu_gp_in_progress(rsp)) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002107 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney0f10dc822010-01-04 15:09:06 -08002108 return;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002109 }
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07002110 if (rnp->qsmask == 0) {
Paul E. McKenney1217ed12011-05-04 21:43:49 -07002111 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002112 continue;
2113 }
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07002114 cpu = rnp->grplo;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002115 bit = 1;
Paul E. McKenneya0b6c9a2009-09-28 07:46:33 -07002116 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
Paul E. McKenney0edd1b12013-06-21 16:37:22 -07002117 if ((rnp->qsmask & bit) != 0) {
2118 if ((rnp->qsmaskinit & bit) != 0)
2119 *isidle = 0;
2120 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2121 mask |= bit;
2122 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002123 }
Paul E. McKenney45f014c52e2010-01-04 15:09:08 -08002124 if (mask != 0) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002125
Paul E. McKenneyd3f6bad2009-12-02 12:10:13 -08002126 /* rcu_report_qs_rnp() releases rnp->lock. */
2127 rcu_report_qs_rnp(mask, rsp, rnp, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002128 continue;
2129 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002130 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002131 }
Paul E. McKenney27f4d282011-02-07 12:47:15 -08002132 rnp = rcu_get_root(rsp);
Paul E. McKenney1217ed12011-05-04 21:43:49 -07002133 if (rnp->qsmask == 0) {
2134 raw_spin_lock_irqsave(&rnp->lock, flags);
2135 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2136 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002137}
2138
2139/*
2140 * Force quiescent states on reluctant CPUs, and also detect which
2141 * CPUs are in dyntick-idle mode.
2142 */
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07002143static void force_quiescent_state(struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002144{
2145 unsigned long flags;
Paul E. McKenney394f2762012-06-26 17:00:35 -07002146 bool ret;
2147 struct rcu_node *rnp;
2148 struct rcu_node *rnp_old = NULL;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002149
Paul E. McKenney394f2762012-06-26 17:00:35 -07002150 /* Funnel through hierarchy to reduce memory contention. */
2151 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2152 for (; rnp != NULL; rnp = rnp->parent) {
2153 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2154 !raw_spin_trylock(&rnp->fqslock);
2155 if (rnp_old != NULL)
2156 raw_spin_unlock(&rnp_old->fqslock);
2157 if (ret) {
2158 rsp->n_force_qs_lh++;
2159 return;
2160 }
2161 rnp_old = rnp;
2162 }
2163 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2164
2165 /* Reached the root of the rcu_node tree, acquire lock. */
2166 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2167 raw_spin_unlock(&rnp_old->fqslock);
2168 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2169 rsp->n_force_qs_lh++;
2170 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07002171 return; /* Someone beat us to it. */
Paul E. McKenney46a1e342010-01-04 15:09:09 -08002172 }
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07002173 rsp->gp_flags |= RCU_GP_FLAG_FQS;
Paul E. McKenney394f2762012-06-26 17:00:35 -07002174 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07002175 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002176}
2177
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002178/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07002179 * This does the RCU core processing work for the specified rcu_state
2180 * and rcu_data structures. This may be called only from the CPU to
2181 * whom the rdp belongs.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002182 */
2183static void
Paul E. McKenney1bca8cf2012-06-12 09:40:38 -07002184__rcu_process_callbacks(struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002185{
2186 unsigned long flags;
Paul E. McKenney1bca8cf2012-06-12 09:40:38 -07002187 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002188
Paul E. McKenney2e597552009-08-15 09:53:48 -07002189 WARN_ON_ONCE(rdp->beenonline == 0);
2190
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002191 /* Update RCU state based on any recent quiescent states. */
2192 rcu_check_quiescent_state(rsp, rdp);
2193
2194 /* Does this CPU require a not-yet-started grace period? */
Paul E. McKenneydc35c892012-12-03 13:52:00 -08002195 local_irq_save(flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002196 if (cpu_needs_another_gp(rsp, rdp)) {
Paul E. McKenneydc35c892012-12-03 13:52:00 -08002197 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
Paul E. McKenneyb8462082012-12-29 22:04:18 -08002198 rcu_start_gp(rsp);
2199 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
Paul E. McKenneydc35c892012-12-03 13:52:00 -08002200 } else {
2201 local_irq_restore(flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002202 }
2203
2204 /* If there are callbacks ready, invoke them. */
Shaohua Li09223372011-06-14 13:26:25 +08002205 if (cpu_has_callbacks_ready_to_invoke(rdp))
Paul E. McKenneya46e0892011-06-15 15:47:09 -07002206 invoke_rcu_callbacks(rsp, rdp);
Shaohua Li09223372011-06-14 13:26:25 +08002207}
2208
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002209/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07002210 * Do RCU core processing for the current CPU.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002211 */
Shaohua Li09223372011-06-14 13:26:25 +08002212static void rcu_process_callbacks(struct softirq_action *unused)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002213{
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07002214 struct rcu_state *rsp;
2215
Paul E. McKenneybfa00b42012-06-21 09:54:10 -07002216 if (cpu_is_offline(smp_processor_id()))
2217 return;
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04002218 trace_rcu_utilization(TPS("Start RCU core"));
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07002219 for_each_rcu_flavor(rsp)
2220 __rcu_process_callbacks(rsp);
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04002221 trace_rcu_utilization(TPS("End RCU core"));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002222}
2223
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002224/*
Paul E. McKenneye0f23062011-06-21 01:29:39 -07002225 * Schedule RCU callback invocation. If the specified type of RCU
2226 * does not support RCU priority boosting, just do a direct call,
2227 * otherwise wake up the per-CPU kernel kthread. Note that because we
2228 * are running on the current CPU with interrupts disabled, the
2229 * rcu_cpu_kthread_task cannot disappear out from under us.
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002230 */
Paul E. McKenneya46e0892011-06-15 15:47:09 -07002231static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002232{
Paul E. McKenneyb0d30412011-07-10 15:57:35 -07002233 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2234 return;
Paul E. McKenneya46e0892011-06-15 15:47:09 -07002235 if (likely(!rsp->boost)) {
2236 rcu_do_batch(rsp, rdp);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002237 return;
2238 }
Paul E. McKenneya46e0892011-06-15 15:47:09 -07002239 invoke_rcu_callbacks_kthread();
Paul E. McKenneya26ac242011-01-12 14:10:23 -08002240}
2241
Paul E. McKenneya46e0892011-06-15 15:47:09 -07002242static void invoke_rcu_core(void)
Shaohua Li09223372011-06-14 13:26:25 +08002243{
Paul E. McKenneyb0f74032013-02-04 12:14:24 -08002244 if (cpu_online(smp_processor_id()))
2245 raise_softirq(RCU_SOFTIRQ);
Shaohua Li09223372011-06-14 13:26:25 +08002246}
2247
Paul E. McKenney29154c52012-05-30 03:21:48 -07002248/*
2249 * Handle any core-RCU processing required by a call_rcu() invocation.
2250 */
2251static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2252 struct rcu_head *head, unsigned long flags)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002253{
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002254 /*
Paul E. McKenney29154c52012-05-30 03:21:48 -07002255 * If called from an extended quiescent state, invoke the RCU
2256 * core in order to force a re-evaluation of RCU's idleness.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002257 */
Paul E. McKenney29154c52012-05-30 03:21:48 -07002258 if (rcu_is_cpu_idle() && cpu_online(smp_processor_id()))
2259 invoke_rcu_core();
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002260
Paul E. McKenney29154c52012-05-30 03:21:48 -07002261 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2262 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
Paul E. McKenney2655d572011-04-07 22:47:23 -07002263 return;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002264
Paul E. McKenney37c72e52009-10-14 10:15:55 -07002265 /*
2266 * Force the grace period if too many callbacks or too long waiting.
2267 * Enforce hysteresis, and don't invoke force_quiescent_state()
2268 * if some other CPU has recently done so. Also, don't bother
2269 * invoking force_quiescent_state() if the newly enqueued callback
2270 * is the only one waiting for a grace period to complete.
2271 */
Paul E. McKenney2655d572011-04-07 22:47:23 -07002272 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
Paul E. McKenneyb52573d2010-12-14 17:36:02 -08002273
2274 /* Are we ignoring a completed grace period? */
Paul E. McKenney470716f2013-03-19 11:32:11 -07002275 note_gp_changes(rsp, rdp);
Paul E. McKenneyb52573d2010-12-14 17:36:02 -08002276
2277 /* Start a new grace period if one not already started. */
2278 if (!rcu_gp_in_progress(rsp)) {
Paul E. McKenneyb52573d2010-12-14 17:36:02 -08002279 struct rcu_node *rnp_root = rcu_get_root(rsp);
2280
Paul E. McKenneyb8462082012-12-29 22:04:18 -08002281 raw_spin_lock(&rnp_root->lock);
2282 rcu_start_gp(rsp);
2283 raw_spin_unlock(&rnp_root->lock);
Paul E. McKenneyb52573d2010-12-14 17:36:02 -08002284 } else {
2285 /* Give the grace period a kick. */
2286 rdp->blimit = LONG_MAX;
2287 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2288 *rdp->nxttail[RCU_DONE_TAIL] != head)
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07002289 force_quiescent_state(rsp);
Paul E. McKenneyb52573d2010-12-14 17:36:02 -08002290 rdp->n_force_qs_snap = rsp->n_force_qs;
2291 rdp->qlen_last_fqs_check = rdp->qlen;
2292 }
Paul E. McKenney4cdfc1752012-06-22 17:06:26 -07002293 }
Paul E. McKenney29154c52012-05-30 03:21:48 -07002294}
2295
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002296/*
Paul E. McKenneyae150182013-04-23 13:20:57 -07002297 * RCU callback function to leak a callback.
2298 */
2299static void rcu_leak_callback(struct rcu_head *rhp)
2300{
2301}
2302
2303/*
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002304 * Helper function for call_rcu() and friends. The cpu argument will
2305 * normally be -1, indicating "currently running CPU". It may specify
2306 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2307 * is expected to specify a CPU.
2308 */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002309static void
2310__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002311 struct rcu_state *rsp, int cpu, bool lazy)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002312{
2313 unsigned long flags;
2314 struct rcu_data *rdp;
2315
2316 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
Paul E. McKenneyae150182013-04-23 13:20:57 -07002317 if (debug_rcu_head_queue(head)) {
2318 /* Probable double call_rcu(), so leak the callback. */
2319 ACCESS_ONCE(head->func) = rcu_leak_callback;
2320 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2321 return;
2322 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002323 head->func = func;
2324 head->next = NULL;
2325
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002326 /*
2327 * Opportunistically note grace-period endings and beginnings.
2328 * Note that we might see a beginning right after we see an
2329 * end, but never vice versa, since this CPU has to pass through
2330 * a quiescent state betweentimes.
2331 */
2332 local_irq_save(flags);
2333 rdp = this_cpu_ptr(rsp->rda);
2334
2335 /* Add the callback to our list. */
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002336 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2337 int offline;
2338
2339 if (cpu != -1)
2340 rdp = per_cpu_ptr(rsp->rda, cpu);
2341 offline = !__call_rcu_nocb(rdp, head, lazy);
2342 WARN_ON_ONCE(offline);
Paul E. McKenney0d8ee372012-08-03 13:16:15 -07002343 /* _call_rcu() is illegal on offline CPU; leak the callback. */
Paul E. McKenney0d8ee372012-08-03 13:16:15 -07002344 local_irq_restore(flags);
2345 return;
2346 }
Paul E. McKenney1d1fb392012-05-09 15:44:42 -07002347 ACCESS_ONCE(rdp->qlen)++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002348 if (lazy)
2349 rdp->qlen_lazy++;
2350 else
2351 rcu_idle_count_callbacks_posted();
2352 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2353 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2354 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2355
2356 if (__is_kfree_rcu_offset((unsigned long)func))
2357 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2358 rdp->qlen_lazy, rdp->qlen);
2359 else
2360 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2361
Paul E. McKenney29154c52012-05-30 03:21:48 -07002362 /* Go handle any RCU core processing required. */
2363 __call_rcu_core(rsp, rdp, head, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002364 local_irq_restore(flags);
2365}
2366
2367/*
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002368 * Queue an RCU-sched callback for invocation after a grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002369 */
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002370void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002371{
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002372 __call_rcu(head, func, &rcu_sched_state, -1, 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002373}
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07002374EXPORT_SYMBOL_GPL(call_rcu_sched);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002375
2376/*
Paul E. McKenney486e2592012-01-06 14:11:30 -08002377 * Queue an RCU callback for invocation after a quicker grace period.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002378 */
2379void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2380{
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002381 __call_rcu(head, func, &rcu_bh_state, -1, 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002382}
2383EXPORT_SYMBOL_GPL(call_rcu_bh);
2384
Paul E. McKenney6d813392012-02-23 13:30:16 -08002385/*
2386 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2387 * any blocking grace-period wait automatically implies a grace period
2388 * if there is only one CPU online at any point time during execution
2389 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2390 * occasionally incorrectly indicate that there are multiple CPUs online
2391 * when there was in fact only one the whole time, as this just adds
2392 * some overhead: RCU still operates correctly.
Paul E. McKenney6d813392012-02-23 13:30:16 -08002393 */
2394static inline int rcu_blocking_is_gp(void)
2395{
Paul E. McKenney95f0c1d2012-06-19 11:58:27 -07002396 int ret;
2397
Paul E. McKenney6d813392012-02-23 13:30:16 -08002398 might_sleep(); /* Check for RCU read-side critical section. */
Paul E. McKenney95f0c1d2012-06-19 11:58:27 -07002399 preempt_disable();
2400 ret = num_online_cpus() <= 1;
2401 preempt_enable();
2402 return ret;
Paul E. McKenney6d813392012-02-23 13:30:16 -08002403}
2404
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08002405/**
2406 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2407 *
2408 * Control will return to the caller some time after a full rcu-sched
2409 * grace period has elapsed, in other words after all currently executing
2410 * rcu-sched read-side critical sections have completed. These read-side
2411 * critical sections are delimited by rcu_read_lock_sched() and
2412 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2413 * local_irq_disable(), and so on may be used in place of
2414 * rcu_read_lock_sched().
2415 *
2416 * This means that all preempt_disable code sequences, including NMI and
Paul E. McKenneyf0a0e6f2012-10-23 13:47:01 -07002417 * non-threaded hardware-interrupt handlers, in progress on entry will
2418 * have completed before this primitive returns. However, this does not
2419 * guarantee that softirq handlers will have completed, since in some
2420 * kernels, these handlers can run in process context, and can block.
2421 *
2422 * Note that this guarantee implies further memory-ordering guarantees.
2423 * On systems with more than one CPU, when synchronize_sched() returns,
2424 * each CPU is guaranteed to have executed a full memory barrier since the
2425 * end of its last RCU-sched read-side critical section whose beginning
2426 * preceded the call to synchronize_sched(). In addition, each CPU having
2427 * an RCU read-side critical section that extends beyond the return from
2428 * synchronize_sched() is guaranteed to have executed a full memory barrier
2429 * after the beginning of synchronize_sched() and before the beginning of
2430 * that RCU read-side critical section. Note that these guarantees include
2431 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2432 * that are executing in the kernel.
2433 *
2434 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2435 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2436 * to have executed a full memory barrier during the execution of
2437 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2438 * again only if the system has more than one CPU).
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08002439 *
2440 * This primitive provides the guarantees made by the (now removed)
2441 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2442 * guarantees that rcu_read_lock() sections will have completed.
2443 * In "classic RCU", these two guarantees happen to be one and
2444 * the same, but can differ in realtime RCU implementations.
2445 */
2446void synchronize_sched(void)
2447{
Paul E. McKenneyfe15d702012-01-04 13:30:33 -08002448 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2449 !lock_is_held(&rcu_lock_map) &&
2450 !lock_is_held(&rcu_sched_lock_map),
2451 "Illegal synchronize_sched() in RCU-sched read-side critical section");
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08002452 if (rcu_blocking_is_gp())
2453 return;
Antti P Miettinen3705b882012-10-05 09:59:15 +03002454 if (rcu_expedited)
2455 synchronize_sched_expedited();
2456 else
2457 wait_rcu_gp(call_rcu_sched);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08002458}
2459EXPORT_SYMBOL_GPL(synchronize_sched);
2460
2461/**
2462 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2463 *
2464 * Control will return to the caller some time after a full rcu_bh grace
2465 * period has elapsed, in other words after all currently executing rcu_bh
2466 * read-side critical sections have completed. RCU read-side critical
2467 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2468 * and may be nested.
Paul E. McKenneyf0a0e6f2012-10-23 13:47:01 -07002469 *
2470 * See the description of synchronize_sched() for more detailed information
2471 * on memory ordering guarantees.
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08002472 */
2473void synchronize_rcu_bh(void)
2474{
Paul E. McKenneyfe15d702012-01-04 13:30:33 -08002475 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2476 !lock_is_held(&rcu_lock_map) &&
2477 !lock_is_held(&rcu_sched_lock_map),
2478 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08002479 if (rcu_blocking_is_gp())
2480 return;
Antti P Miettinen3705b882012-10-05 09:59:15 +03002481 if (rcu_expedited)
2482 synchronize_rcu_bh_expedited();
2483 else
2484 wait_rcu_gp(call_rcu_bh);
Paul E. McKenney6ebb2372009-11-22 08:53:50 -08002485}
2486EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2487
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002488static int synchronize_sched_expedited_cpu_stop(void *data)
2489{
2490 /*
2491 * There must be a full memory barrier on each affected CPU
2492 * between the time that try_stop_cpus() is called and the
2493 * time that it returns.
2494 *
2495 * In the current initial implementation of cpu_stop, the
2496 * above condition is already met when the control reaches
2497 * this point and the following smp_mb() is not strictly
2498 * necessary. Do smp_mb() anyway for documentation and
2499 * robustness against future implementation changes.
2500 */
2501 smp_mb(); /* See above comment block. */
2502 return 0;
2503}
2504
Paul E. McKenney236fefa2012-01-31 14:00:41 -08002505/**
2506 * synchronize_sched_expedited - Brute-force RCU-sched grace period
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002507 *
Paul E. McKenney236fefa2012-01-31 14:00:41 -08002508 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2509 * approach to force the grace period to end quickly. This consumes
2510 * significant time on all CPUs and is unfriendly to real-time workloads,
2511 * so is thus not recommended for any sort of common-case code. In fact,
2512 * if you are using synchronize_sched_expedited() in a loop, please
2513 * restructure your code to batch your updates, and then use a single
2514 * synchronize_sched() instead.
2515 *
2516 * Note that it is illegal to call this function while holding any lock
2517 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2518 * to call this function from a CPU-hotplug notifier. Failing to observe
2519 * these restriction will result in deadlock.
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002520 *
2521 * This implementation can be thought of as an application of ticket
2522 * locking to RCU, with sync_sched_expedited_started and
2523 * sync_sched_expedited_done taking on the roles of the halves
2524 * of the ticket-lock word. Each task atomically increments
2525 * sync_sched_expedited_started upon entry, snapshotting the old value,
2526 * then attempts to stop all the CPUs. If this succeeds, then each
2527 * CPU will have executed a context switch, resulting in an RCU-sched
2528 * grace period. We are then done, so we use atomic_cmpxchg() to
2529 * update sync_sched_expedited_done to match our snapshot -- but
2530 * only if someone else has not already advanced past our snapshot.
2531 *
2532 * On the other hand, if try_stop_cpus() fails, we check the value
2533 * of sync_sched_expedited_done. If it has advanced past our
2534 * initial snapshot, then someone else must have forced a grace period
2535 * some time after we took our snapshot. In this case, our work is
2536 * done for us, and we can simply return. Otherwise, we try again,
2537 * but keep our initial snapshot for purposes of checking for someone
2538 * doing our work for us.
2539 *
2540 * If we fail too many times in a row, we fall back to synchronize_sched().
2541 */
2542void synchronize_sched_expedited(void)
2543{
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002544 long firstsnap, s, snap;
2545 int trycount = 0;
Paul E. McKenney40694d62012-10-11 15:24:03 -07002546 struct rcu_state *rsp = &rcu_sched_state;
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002547
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002548 /*
2549 * If we are in danger of counter wrap, just do synchronize_sched().
2550 * By allowing sync_sched_expedited_started to advance no more than
2551 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2552 * that more than 3.5 billion CPUs would be required to force a
2553 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2554 * course be required on a 64-bit system.
2555 */
Paul E. McKenney40694d62012-10-11 15:24:03 -07002556 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2557 (ulong)atomic_long_read(&rsp->expedited_done) +
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002558 ULONG_MAX / 8)) {
2559 synchronize_sched();
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002560 atomic_long_inc(&rsp->expedited_wrap);
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002561 return;
2562 }
2563
2564 /*
2565 * Take a ticket. Note that atomic_inc_return() implies a
2566 * full memory barrier.
2567 */
Paul E. McKenney40694d62012-10-11 15:24:03 -07002568 snap = atomic_long_inc_return(&rsp->expedited_start);
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002569 firstsnap = snap;
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002570 get_online_cpus();
Hugh Dickins1cc85962012-02-17 13:20:31 -08002571 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002572
2573 /*
2574 * Each pass through the following loop attempts to force a
2575 * context switch on each CPU.
2576 */
2577 while (try_stop_cpus(cpu_online_mask,
2578 synchronize_sched_expedited_cpu_stop,
2579 NULL) == -EAGAIN) {
2580 put_online_cpus();
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002581 atomic_long_inc(&rsp->expedited_tryfail);
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002582
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002583 /* Check to see if someone else did our work for us. */
Paul E. McKenney40694d62012-10-11 15:24:03 -07002584 s = atomic_long_read(&rsp->expedited_done);
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002585 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002586 /* ensure test happens before caller kfree */
2587 smp_mb__before_atomic_inc(); /* ^^^ */
2588 atomic_long_inc(&rsp->expedited_workdone1);
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002589 return;
2590 }
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002591
2592 /* No joy, try again later. Or just synchronize_sched(). */
Paul E. McKenneyc701d5d2012-06-28 08:08:25 -07002593 if (trycount++ < 10) {
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002594 udelay(trycount * num_online_cpus());
Paul E. McKenneyc701d5d2012-06-28 08:08:25 -07002595 } else {
Antti P Miettinen3705b882012-10-05 09:59:15 +03002596 wait_rcu_gp(call_rcu_sched);
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002597 atomic_long_inc(&rsp->expedited_normal);
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002598 return;
2599 }
2600
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002601 /* Recheck to see if someone else did our work for us. */
Paul E. McKenney40694d62012-10-11 15:24:03 -07002602 s = atomic_long_read(&rsp->expedited_done);
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002603 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002604 /* ensure test happens before caller kfree */
2605 smp_mb__before_atomic_inc(); /* ^^^ */
2606 atomic_long_inc(&rsp->expedited_workdone2);
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002607 return;
2608 }
2609
2610 /*
2611 * Refetching sync_sched_expedited_started allows later
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002612 * callers to piggyback on our grace period. We retry
2613 * after they started, so our grace period works for them,
2614 * and they started after our first try, so their grace
2615 * period works for us.
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002616 */
2617 get_online_cpus();
Paul E. McKenney40694d62012-10-11 15:24:03 -07002618 snap = atomic_long_read(&rsp->expedited_start);
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002619 smp_mb(); /* ensure read is before try_stop_cpus(). */
2620 }
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002621 atomic_long_inc(&rsp->expedited_stoppedcpus);
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002622
2623 /*
2624 * Everyone up to our most recent fetch is covered by our grace
2625 * period. Update the counter, but only if our work is still
2626 * relevant -- which it won't be if someone who started later
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002627 * than we did already did their update.
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002628 */
2629 do {
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002630 atomic_long_inc(&rsp->expedited_done_tries);
Paul E. McKenney40694d62012-10-11 15:24:03 -07002631 s = atomic_long_read(&rsp->expedited_done);
Paul E. McKenney1924bcb2012-10-11 12:30:37 -07002632 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002633 /* ensure test happens before caller kfree */
2634 smp_mb__before_atomic_inc(); /* ^^^ */
2635 atomic_long_inc(&rsp->expedited_done_lost);
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002636 break;
2637 }
Paul E. McKenney40694d62012-10-11 15:24:03 -07002638 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
Paul E. McKenneya30489c2012-10-11 16:18:09 -07002639 atomic_long_inc(&rsp->expedited_done_exit);
Paul E. McKenney3d3b7db2012-01-23 17:05:46 -08002640
2641 put_online_cpus();
2642}
2643EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2644
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002645/*
2646 * Check to see if there is any immediate RCU-related work to be done
2647 * by the current CPU, for the specified type of RCU, returning 1 if so.
2648 * The checks are in order of increasing expense: checks that can be
2649 * carried out against CPU-local state are performed first. However,
2650 * we must check for CPU stalls first, else we might not get a chance.
2651 */
2652static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2653{
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002654 struct rcu_node *rnp = rdp->mynode;
2655
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002656 rdp->n_rcu_pending++;
2657
2658 /* Check for CPU stalls, if enabled. */
2659 check_cpu_stall(rsp, rdp);
2660
2661 /* Is the RCU core waiting for a quiescent state from this CPU? */
Paul E. McKenney5c51dd72011-08-04 06:59:03 -07002662 if (rcu_scheduler_fully_active &&
2663 rdp->qs_pending && !rdp->passed_quiesce) {
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07002664 rdp->n_rp_qs_pending++;
Paul E. McKenneye4cc1f22011-06-27 00:17:43 -07002665 } else if (rdp->qs_pending && rdp->passed_quiesce) {
Paul E. McKenneyd21670a2010-04-14 17:39:26 -07002666 rdp->n_rp_report_qs++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002667 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002668 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002669
2670 /* Does this CPU have callbacks ready to invoke? */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002671 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2672 rdp->n_rp_cb_ready++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002673 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002674 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002675
2676 /* Has RCU gone idle with this CPU needing another grace period? */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002677 if (cpu_needs_another_gp(rsp, rdp)) {
2678 rdp->n_rp_cpu_needs_gp++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002679 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002680 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002681
2682 /* Has another RCU grace period completed? */
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002683 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002684 rdp->n_rp_gp_completed++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002685 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002686 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002687
2688 /* Has a new RCU grace period started? */
Paul E. McKenney2f51f982009-11-13 19:51:39 -08002689 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002690 rdp->n_rp_gp_started++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002691 return 1;
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002692 }
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002693
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002694 /* nothing to do */
Paul E. McKenney7ba5c842009-04-13 21:31:17 -07002695 rdp->n_rp_need_nothing++;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002696 return 0;
2697}
2698
2699/*
2700 * Check to see if there is any immediate RCU-related work to be done
2701 * by the current CPU, returning 1 if so. This function is part of the
2702 * RCU implementation; it is -not- an exported member of the RCU API.
2703 */
Paul E. McKenneya1572292009-08-22 13:56:51 -07002704static int rcu_pending(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002705{
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07002706 struct rcu_state *rsp;
2707
2708 for_each_rcu_flavor(rsp)
2709 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2710 return 1;
2711 return 0;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002712}
2713
2714/*
Paul E. McKenneyc0f4dfd2012-12-28 11:30:36 -08002715 * Return true if the specified CPU has any callback. If all_lazy is
2716 * non-NULL, store an indication of whether all callbacks are lazy.
2717 * (If there are no callbacks, all of them are deemed to be lazy.)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002718 */
Paul E. McKenneyc0f4dfd2012-12-28 11:30:36 -08002719static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002720{
Paul E. McKenneyc0f4dfd2012-12-28 11:30:36 -08002721 bool al = true;
2722 bool hc = false;
2723 struct rcu_data *rdp;
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07002724 struct rcu_state *rsp;
2725
Paul E. McKenneyc0f4dfd2012-12-28 11:30:36 -08002726 for_each_rcu_flavor(rsp) {
2727 rdp = per_cpu_ptr(rsp->rda, cpu);
2728 if (rdp->qlen != rdp->qlen_lazy)
2729 al = false;
2730 if (rdp->nxtlist)
2731 hc = true;
2732 }
2733 if (all_lazy)
2734 *all_lazy = al;
2735 return hc;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002736}
2737
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002738/*
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002739 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2740 * the compiler is expected to optimize this away.
2741 */
Steven Rostedt (Red Hat)e66c33d2013-07-12 16:50:28 -04002742static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002743 int cpu, unsigned long done)
2744{
2745 trace_rcu_barrier(rsp->name, s, cpu,
2746 atomic_read(&rsp->barrier_cpu_count), done);
2747}
2748
2749/*
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002750 * RCU callback function for _rcu_barrier(). If we are last, wake
2751 * up the task executing _rcu_barrier().
2752 */
Paul E. McKenney24ebbca2012-05-29 00:34:56 -07002753static void rcu_barrier_callback(struct rcu_head *rhp)
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002754{
Paul E. McKenney24ebbca2012-05-29 00:34:56 -07002755 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2756 struct rcu_state *rsp = rdp->rsp;
2757
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002758 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2759 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
Paul E. McKenney7db74df2012-05-29 03:03:37 -07002760 complete(&rsp->barrier_completion);
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002761 } else {
2762 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2763 }
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002764}
2765
2766/*
2767 * Called with preemption disabled, and from cross-cpu IRQ context.
2768 */
2769static void rcu_barrier_func(void *type)
2770{
Paul E. McKenney037b64e2012-05-28 23:26:01 -07002771 struct rcu_state *rsp = type;
Paul E. McKenney06668ef2012-05-28 23:57:46 -07002772 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002773
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002774 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
Paul E. McKenney24ebbca2012-05-29 00:34:56 -07002775 atomic_inc(&rsp->barrier_cpu_count);
Paul E. McKenney06668ef2012-05-28 23:57:46 -07002776 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002777}
2778
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002779/*
2780 * Orchestrate the specified type of RCU barrier, waiting for all
2781 * RCU callbacks of the specified type to complete.
2782 */
Paul E. McKenney037b64e2012-05-28 23:26:01 -07002783static void _rcu_barrier(struct rcu_state *rsp)
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002784{
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002785 int cpu;
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002786 struct rcu_data *rdp;
Paul E. McKenneycf3a9c42012-05-29 14:56:46 -07002787 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2788 unsigned long snap_done;
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002789
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002790 _rcu_barrier_trace(rsp, "Begin", -1, snap);
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002791
Paul E. McKenneye74f4c42009-10-06 21:48:17 -07002792 /* Take mutex to serialize concurrent rcu_barrier() requests. */
Paul E. McKenney7be7f0b2012-05-29 05:18:53 -07002793 mutex_lock(&rsp->barrier_mutex);
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002794
Paul E. McKenneycf3a9c42012-05-29 14:56:46 -07002795 /*
2796 * Ensure that all prior references, including to ->n_barrier_done,
2797 * are ordered before the _rcu_barrier() machinery.
2798 */
2799 smp_mb(); /* See above block comment. */
2800
2801 /*
2802 * Recheck ->n_barrier_done to see if others did our work for us.
2803 * This means checking ->n_barrier_done for an even-to-odd-to-even
2804 * transition. The "if" expression below therefore rounds the old
2805 * value up to the next even number and adds two before comparing.
2806 */
Paul E. McKenney458fb382013-07-26 20:47:42 -07002807 snap_done = rsp->n_barrier_done;
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002808 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
Paul E. McKenney458fb382013-07-26 20:47:42 -07002809
2810 /*
2811 * If the value in snap is odd, we needed to wait for the current
2812 * rcu_barrier() to complete, then wait for the next one, in other
2813 * words, we need the value of snap_done to be three larger than
2814 * the value of snap. On the other hand, if the value in snap is
2815 * even, we only had to wait for the next rcu_barrier() to complete,
2816 * in other words, we need the value of snap_done to be only two
2817 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
2818 * this for us (thank you, Linus!).
2819 */
2820 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002821 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
Paul E. McKenneycf3a9c42012-05-29 14:56:46 -07002822 smp_mb(); /* caller's subsequent code after above check. */
2823 mutex_unlock(&rsp->barrier_mutex);
2824 return;
2825 }
2826
2827 /*
2828 * Increment ->n_barrier_done to avoid duplicate work. Use
2829 * ACCESS_ONCE() to prevent the compiler from speculating
2830 * the increment to precede the early-exit check.
2831 */
2832 ACCESS_ONCE(rsp->n_barrier_done)++;
2833 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002834 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
Paul E. McKenneycf3a9c42012-05-29 14:56:46 -07002835 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002836
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002837 /*
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002838 * Initialize the count to one rather than to zero in order to
2839 * avoid a too-soon return to zero in case of a short grace period
Paul E. McKenney1331e7a2012-08-02 17:43:50 -07002840 * (or preemption of this task). Exclude CPU-hotplug operations
2841 * to ensure that no offline CPU has callbacks queued.
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002842 */
Paul E. McKenney7db74df2012-05-29 03:03:37 -07002843 init_completion(&rsp->barrier_completion);
Paul E. McKenney24ebbca2012-05-29 00:34:56 -07002844 atomic_set(&rsp->barrier_cpu_count, 1);
Paul E. McKenney1331e7a2012-08-02 17:43:50 -07002845 get_online_cpus();
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002846
2847 /*
Paul E. McKenney1331e7a2012-08-02 17:43:50 -07002848 * Force each CPU with callbacks to register a new callback.
2849 * When that callback is invoked, we will know that all of the
2850 * corresponding CPU's preceding callbacks have been invoked.
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002851 */
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002852 for_each_possible_cpu(cpu) {
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +01002853 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002854 continue;
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002855 rdp = per_cpu_ptr(rsp->rda, cpu);
Frederic Weisbeckerd1e43fa2013-03-26 23:47:24 +01002856 if (rcu_is_nocb_cpu(cpu)) {
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002857 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2858 rsp->n_barrier_done);
2859 atomic_inc(&rsp->barrier_cpu_count);
2860 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2861 rsp, cpu, 0);
2862 } else if (ACCESS_ONCE(rdp->qlen)) {
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002863 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2864 rsp->n_barrier_done);
Paul E. McKenney037b64e2012-05-28 23:26:01 -07002865 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002866 } else {
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002867 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2868 rsp->n_barrier_done);
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002869 }
2870 }
Paul E. McKenney1331e7a2012-08-02 17:43:50 -07002871 put_online_cpus();
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002872
2873 /*
2874 * Now that we have an rcu_barrier_callback() callback on each
2875 * CPU, and thus each counted, remove the initial count.
2876 */
Paul E. McKenney24ebbca2012-05-29 00:34:56 -07002877 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
Paul E. McKenney7db74df2012-05-29 03:03:37 -07002878 complete(&rsp->barrier_completion);
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002879
Paul E. McKenneycf3a9c42012-05-29 14:56:46 -07002880 /* Increment ->n_barrier_done to prevent duplicate work. */
2881 smp_mb(); /* Keep increment after above mechanism. */
2882 ACCESS_ONCE(rsp->n_barrier_done)++;
2883 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
Paul E. McKenneya83eff02012-05-23 18:47:05 -07002884 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
Paul E. McKenneycf3a9c42012-05-29 14:56:46 -07002885 smp_mb(); /* Keep increment before caller's subsequent code. */
2886
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002887 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
Paul E. McKenney7db74df2012-05-29 03:03:37 -07002888 wait_for_completion(&rsp->barrier_completion);
Paul E. McKenneyb1420f12012-03-01 13:18:08 -08002889
2890 /* Other rcu_barrier() invocations can now safely proceed. */
Paul E. McKenney7be7f0b2012-05-29 05:18:53 -07002891 mutex_unlock(&rsp->barrier_mutex);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002892}
2893
2894/**
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002895 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2896 */
2897void rcu_barrier_bh(void)
2898{
Paul E. McKenney037b64e2012-05-28 23:26:01 -07002899 _rcu_barrier(&rcu_bh_state);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002900}
2901EXPORT_SYMBOL_GPL(rcu_barrier_bh);
2902
2903/**
2904 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
2905 */
2906void rcu_barrier_sched(void)
2907{
Paul E. McKenney037b64e2012-05-28 23:26:01 -07002908 _rcu_barrier(&rcu_sched_state);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07002909}
2910EXPORT_SYMBOL_GPL(rcu_barrier_sched);
2911
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002912/*
Paul E. McKenney27569622009-08-15 09:53:46 -07002913 * Do boot-time initialization of a CPU's per-CPU RCU data.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002914 */
Paul E. McKenney27569622009-08-15 09:53:46 -07002915static void __init
2916rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002917{
2918 unsigned long flags;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002919 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenney27569622009-08-15 09:53:46 -07002920 struct rcu_node *rnp = rcu_get_root(rsp);
2921
2922 /* Set up local state, ensuring consistent view of global state. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002923 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney27569622009-08-15 09:53:46 -07002924 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
Paul E. McKenney3f5d3ea2012-05-09 15:39:56 -07002925 init_callback_list(rdp);
Paul E. McKenney486e2592012-01-06 14:11:30 -08002926 rdp->qlen_lazy = 0;
Paul E. McKenney1d1fb392012-05-09 15:44:42 -07002927 ACCESS_ONCE(rdp->qlen) = 0;
Paul E. McKenney27569622009-08-15 09:53:46 -07002928 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
Paul E. McKenney29e37d82011-11-17 16:55:56 -08002929 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
Paul E. McKenney9b2e4f12011-09-30 12:10:22 -07002930 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
Paul E. McKenney27569622009-08-15 09:53:46 -07002931 rdp->cpu = cpu;
Paul E. McKenneyd4c08f22011-06-25 06:36:56 -07002932 rdp->rsp = rsp;
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07002933 rcu_boot_init_nocb_percpu_data(rdp);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002934 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney27569622009-08-15 09:53:46 -07002935}
2936
2937/*
2938 * Initialize a CPU's per-CPU RCU data. Note that only one online or
2939 * offline event can be happening at a given time. Note also that we
2940 * can accept some slop in the rsp->completed access due to the fact
2941 * that this CPU cannot possibly have any RCU callbacks in flight yet.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002942 */
Paul Gortmaker49fb4c62013-06-19 14:52:21 -04002943static void
Paul E. McKenney6cc68792011-03-02 13:15:15 -08002944rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002945{
2946 unsigned long flags;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002947 unsigned long mask;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08002948 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002949 struct rcu_node *rnp = rcu_get_root(rsp);
2950
Paul E. McKenneya4fbe352012-10-07 08:36:12 -07002951 /* Exclude new grace periods. */
2952 mutex_lock(&rsp->onoff_mutex);
2953
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002954 /* Set up local state, ensuring consistent view of global state. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002955 raw_spin_lock_irqsave(&rnp->lock, flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002956 rdp->beenonline = 1; /* We have now been online. */
Paul E. McKenney6cc68792011-03-02 13:15:15 -08002957 rdp->preemptible = preemptible;
Paul E. McKenney37c72e52009-10-14 10:15:55 -07002958 rdp->qlen_last_fqs_check = 0;
2959 rdp->n_force_qs_snap = rsp->n_force_qs;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002960 rdp->blimit = blimit;
Paul E. McKenney0d8ee372012-08-03 13:16:15 -07002961 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
Paul E. McKenney29e37d82011-11-17 16:55:56 -08002962 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
Paul E. McKenney23332102013-06-21 12:34:33 -07002963 rcu_sysidle_init_percpu_data(rdp->dynticks);
Paul E. McKenneyc92b1312011-11-23 13:38:58 -08002964 atomic_set(&rdp->dynticks->dynticks,
2965 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002966 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002967
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002968 /* Add CPU to rcu_node bitmasks. */
2969 rnp = rdp->mynode;
2970 mask = rdp->grpmask;
2971 do {
2972 /* Exclude any attempts to start a new GP on small systems. */
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002973 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002974 rnp->qsmaskinit |= mask;
2975 mask = rnp->grpmask;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002976 if (rnp == rdp->mynode) {
Paul E. McKenney06ae115a2011-08-14 15:56:54 -07002977 /*
2978 * If there is a grace period in progress, we will
2979 * set up to wait for it next time we run the
2980 * RCU core code.
2981 */
2982 rdp->gpnum = rnp->completed;
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002983 rdp->completed = rnp->completed;
Paul E. McKenney06ae115a2011-08-14 15:56:54 -07002984 rdp->passed_quiesce = 0;
2985 rdp->qs_pending = 0;
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04002986 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
Paul E. McKenneyd09b62d2009-11-02 13:52:28 -08002987 }
Paul E. McKenney1304afb2010-02-22 17:05:02 -08002988 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002989 rnp = rnp->parent;
2990 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
Paul E. McKenneya4fbe352012-10-07 08:36:12 -07002991 local_irq_restore(flags);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002992
Paul E. McKenneya4fbe352012-10-07 08:36:12 -07002993 mutex_unlock(&rsp->onoff_mutex);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002994}
2995
Paul Gortmaker49fb4c62013-06-19 14:52:21 -04002996static void rcu_prepare_cpu(int cpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01002997{
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07002998 struct rcu_state *rsp;
2999
3000 for_each_rcu_flavor(rsp)
3001 rcu_init_percpu_data(cpu, rsp,
3002 strcmp(rsp->name, "rcu_preempt") == 0);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003003}
3004
3005/*
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07003006 * Handle CPU online/offline notification events.
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003007 */
Paul Gortmaker49fb4c62013-06-19 14:52:21 -04003008static int rcu_cpu_notify(struct notifier_block *self,
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08003009 unsigned long action, void *hcpu)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003010{
3011 long cpu = (long)hcpu;
Paul E. McKenney27f4d282011-02-07 12:47:15 -08003012 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08003013 struct rcu_node *rnp = rdp->mynode;
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07003014 struct rcu_state *rsp;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003015
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04003016 trace_rcu_utilization(TPS("Start CPU hotplug"));
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003017 switch (action) {
3018 case CPU_UP_PREPARE:
3019 case CPU_UP_PREPARE_FROZEN:
Peter Zijlstrad72bce02011-05-30 13:34:51 +02003020 rcu_prepare_cpu(cpu);
3021 rcu_prepare_kthreads(cpu);
Paul E. McKenneya26ac242011-01-12 14:10:23 -08003022 break;
3023 case CPU_ONLINE:
Paul E. McKenney0f962a52011-04-14 12:13:53 -07003024 case CPU_DOWN_FAILED:
Thomas Gleixner5d01bbd2012-07-16 10:42:35 +00003025 rcu_boost_kthread_setaffinity(rnp, -1);
Paul E. McKenney0f962a52011-04-14 12:13:53 -07003026 break;
3027 case CPU_DOWN_PREPARE:
Paul E. McKenney34ed62462013-01-07 13:37:42 -08003028 rcu_boost_kthread_setaffinity(rnp, cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003029 break;
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07003030 case CPU_DYING:
3031 case CPU_DYING_FROZEN:
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07003032 for_each_rcu_flavor(rsp)
3033 rcu_cleanup_dying_cpu(rsp);
Paul E. McKenneyd0ec7742009-10-06 21:48:16 -07003034 break;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003035 case CPU_DEAD:
3036 case CPU_DEAD_FROZEN:
3037 case CPU_UP_CANCELED:
3038 case CPU_UP_CANCELED_FROZEN:
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07003039 for_each_rcu_flavor(rsp)
3040 rcu_cleanup_dead_cpu(cpu, rsp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003041 break;
3042 default:
3043 break;
3044 }
Steven Rostedt (Red Hat)f7f7bac2013-07-12 17:18:47 -04003045 trace_rcu_utilization(TPS("End CPU hotplug"));
Paul E. McKenney34ed62462013-01-07 13:37:42 -08003046 return NOTIFY_OK;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003047}
3048
Borislav Petkovd1d74d12013-04-22 00:12:42 +02003049static int rcu_pm_notify(struct notifier_block *self,
3050 unsigned long action, void *hcpu)
3051{
3052 switch (action) {
3053 case PM_HIBERNATION_PREPARE:
3054 case PM_SUSPEND_PREPARE:
3055 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3056 rcu_expedited = 1;
3057 break;
3058 case PM_POST_HIBERNATION:
3059 case PM_POST_SUSPEND:
3060 rcu_expedited = 0;
3061 break;
3062 default:
3063 break;
3064 }
3065 return NOTIFY_OK;
3066}
3067
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003068/*
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07003069 * Spawn the kthread that handles this RCU flavor's grace periods.
3070 */
3071static int __init rcu_spawn_gp_kthread(void)
3072{
3073 unsigned long flags;
3074 struct rcu_node *rnp;
3075 struct rcu_state *rsp;
3076 struct task_struct *t;
3077
3078 for_each_rcu_flavor(rsp) {
Kees Cookf1701682013-07-03 15:04:58 -07003079 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07003080 BUG_ON(IS_ERR(t));
3081 rnp = rcu_get_root(rsp);
3082 raw_spin_lock_irqsave(&rnp->lock, flags);
3083 rsp->gp_kthread = t;
3084 raw_spin_unlock_irqrestore(&rnp->lock, flags);
Paul E. McKenney3fbfbf72012-08-19 21:35:53 -07003085 rcu_spawn_nocb_kthreads(rsp);
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07003086 }
3087 return 0;
3088}
3089early_initcall(rcu_spawn_gp_kthread);
3090
3091/*
Paul E. McKenneybbad9372010-04-02 16:17:17 -07003092 * This function is invoked towards the end of the scheduler's initialization
3093 * process. Before this is called, the idle task might contain
3094 * RCU read-side critical sections (during which time, this idle
3095 * task is booting the system). After this function is called, the
3096 * idle tasks are prohibited from containing RCU read-side critical
3097 * sections. This function also enables RCU lockdep checking.
3098 */
3099void rcu_scheduler_starting(void)
3100{
3101 WARN_ON(num_online_cpus() != 1);
3102 WARN_ON(nr_context_switches() > 0);
3103 rcu_scheduler_active = 1;
3104}
3105
3106/*
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003107 * Compute the per-level fanout, either using the exact fanout specified
3108 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3109 */
3110#ifdef CONFIG_RCU_FANOUT_EXACT
3111static void __init rcu_init_levelspread(struct rcu_state *rsp)
3112{
3113 int i;
3114
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003115 for (i = rcu_num_lvls - 1; i > 0; i--)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003116 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003117 rsp->levelspread[0] = rcu_fanout_leaf;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003118}
3119#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3120static void __init rcu_init_levelspread(struct rcu_state *rsp)
3121{
3122 int ccur;
3123 int cprv;
3124 int i;
3125
Paul E. McKenney4dbd6bb32012-09-05 21:43:57 -07003126 cprv = nr_cpu_ids;
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003127 for (i = rcu_num_lvls - 1; i >= 0; i--) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003128 ccur = rsp->levelcnt[i];
3129 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3130 cprv = ccur;
3131 }
3132}
3133#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3134
3135/*
3136 * Helper function for rcu_init() that initializes one rcu_state structure.
3137 */
Lai Jiangshan394f99a2010-06-28 16:25:04 +08003138static void __init rcu_init_one(struct rcu_state *rsp,
3139 struct rcu_data __percpu *rda)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003140{
Paul E. McKenney394f2762012-06-26 17:00:35 -07003141 static char *buf[] = { "rcu_node_0",
3142 "rcu_node_1",
3143 "rcu_node_2",
3144 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3145 static char *fqs[] = { "rcu_node_fqs_0",
3146 "rcu_node_fqs_1",
3147 "rcu_node_fqs_2",
3148 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003149 int cpustride = 1;
3150 int i;
3151 int j;
3152 struct rcu_node *rnp;
3153
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08003154 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3155
Paul E. McKenney49305212012-11-29 13:49:00 -08003156 /* Silence gcc 4.8 warning about array index out of range. */
3157 if (rcu_num_lvls > RCU_NUM_LVLS)
3158 panic("rcu_init_one: rcu_num_lvls overflow");
3159
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003160 /* Initialize the level-tracking arrays. */
3161
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003162 for (i = 0; i < rcu_num_lvls; i++)
3163 rsp->levelcnt[i] = num_rcu_lvl[i];
3164 for (i = 1; i < rcu_num_lvls; i++)
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003165 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3166 rcu_init_levelspread(rsp);
3167
3168 /* Initialize the elements themselves, starting from the leaves. */
3169
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003170 for (i = rcu_num_lvls - 1; i >= 0; i--) {
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003171 cpustride *= rsp->levelspread[i];
3172 rnp = rsp->level[i];
3173 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
Paul E. McKenney1304afb2010-02-22 17:05:02 -08003174 raw_spin_lock_init(&rnp->lock);
Paul E. McKenneyb6407e82010-01-04 16:04:02 -08003175 lockdep_set_class_and_name(&rnp->lock,
3176 &rcu_node_class[i], buf[i]);
Paul E. McKenney394f2762012-06-26 17:00:35 -07003177 raw_spin_lock_init(&rnp->fqslock);
3178 lockdep_set_class_and_name(&rnp->fqslock,
3179 &rcu_fqs_class[i], fqs[i]);
Paul E. McKenney25d30cf2012-07-11 05:23:18 -07003180 rnp->gpnum = rsp->gpnum;
3181 rnp->completed = rsp->completed;
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003182 rnp->qsmask = 0;
3183 rnp->qsmaskinit = 0;
3184 rnp->grplo = j * cpustride;
3185 rnp->grphi = (j + 1) * cpustride - 1;
3186 if (rnp->grphi >= NR_CPUS)
3187 rnp->grphi = NR_CPUS - 1;
3188 if (i == 0) {
3189 rnp->grpnum = 0;
3190 rnp->grpmask = 0;
3191 rnp->parent = NULL;
3192 } else {
3193 rnp->grpnum = j % rsp->levelspread[i - 1];
3194 rnp->grpmask = 1UL << rnp->grpnum;
3195 rnp->parent = rsp->level[i - 1] +
3196 j / rsp->levelspread[i - 1];
3197 }
3198 rnp->level = i;
Paul E. McKenney12f5f522010-11-29 21:56:39 -08003199 INIT_LIST_HEAD(&rnp->blkd_tasks);
Paul E. McKenneydae6e642013-02-10 20:48:58 -08003200 rcu_init_one_nocb(rnp);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003201 }
3202 }
Lai Jiangshan0c340292010-03-28 11:12:30 +08003203
Lai Jiangshan394f99a2010-06-28 16:25:04 +08003204 rsp->rda = rda;
Paul E. McKenneyb3dbec72012-06-18 18:36:08 -07003205 init_waitqueue_head(&rsp->gp_wq);
Steven Rostedt016a8d52013-05-28 17:32:53 -04003206 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003207 rnp = rsp->level[rcu_num_lvls - 1];
Lai Jiangshan0c340292010-03-28 11:12:30 +08003208 for_each_possible_cpu(i) {
Paul E. McKenney4a90a062010-04-14 16:48:11 -07003209 while (i > rnp->grphi)
Lai Jiangshan0c340292010-03-28 11:12:30 +08003210 rnp++;
Lai Jiangshan394f99a2010-06-28 16:25:04 +08003211 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
Lai Jiangshan0c340292010-03-28 11:12:30 +08003212 rcu_boot_init_percpu_data(i, rsp);
3213 }
Paul E. McKenney6ce75a22012-06-12 11:01:13 -07003214 list_add(&rsp->flavors, &rcu_struct_flavors);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003215}
3216
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003217/*
3218 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3219 * replace the definitions in rcutree.h because those are needed to size
3220 * the ->node array in the rcu_state structure.
3221 */
3222static void __init rcu_init_geometry(void)
3223{
Paul E. McKenney026ad282013-04-03 22:14:11 -07003224 ulong d;
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003225 int i;
3226 int j;
Paul E. McKenneycca6f392012-05-08 21:00:28 -07003227 int n = nr_cpu_ids;
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003228 int rcu_capacity[MAX_RCU_LVLS + 1];
3229
Paul E. McKenney026ad282013-04-03 22:14:11 -07003230 /*
3231 * Initialize any unspecified boot parameters.
3232 * The default values of jiffies_till_first_fqs and
3233 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3234 * value, which is a function of HZ, then adding one for each
3235 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3236 */
3237 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3238 if (jiffies_till_first_fqs == ULONG_MAX)
3239 jiffies_till_first_fqs = d;
3240 if (jiffies_till_next_fqs == ULONG_MAX)
3241 jiffies_till_next_fqs = d;
3242
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003243 /* If the compile-time values are accurate, just leave. */
Paul E. McKenneyb17c7032012-09-06 15:38:02 -07003244 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3245 nr_cpu_ids == NR_CPUS)
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003246 return;
3247
3248 /*
3249 * Compute number of nodes that can be handled an rcu_node tree
3250 * with the given number of levels. Setting rcu_capacity[0] makes
3251 * some of the arithmetic easier.
3252 */
3253 rcu_capacity[0] = 1;
3254 rcu_capacity[1] = rcu_fanout_leaf;
3255 for (i = 2; i <= MAX_RCU_LVLS; i++)
3256 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3257
3258 /*
3259 * The boot-time rcu_fanout_leaf parameter is only permitted
3260 * to increase the leaf-level fanout, not decrease it. Of course,
3261 * the leaf-level fanout cannot exceed the number of bits in
3262 * the rcu_node masks. Finally, the tree must be able to accommodate
3263 * the configured number of CPUs. Complain and fall back to the
3264 * compile-time values if these limits are exceeded.
3265 */
3266 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3267 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3268 n > rcu_capacity[MAX_RCU_LVLS]) {
3269 WARN_ON(1);
3270 return;
3271 }
3272
3273 /* Calculate the number of rcu_nodes at each level of the tree. */
3274 for (i = 1; i <= MAX_RCU_LVLS; i++)
3275 if (n <= rcu_capacity[i]) {
3276 for (j = 0; j <= i; j++)
3277 num_rcu_lvl[j] =
3278 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3279 rcu_num_lvls = i;
3280 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3281 num_rcu_lvl[j] = 0;
3282 break;
3283 }
3284
3285 /* Calculate the total number of rcu_node structures. */
3286 rcu_num_nodes = 0;
3287 for (i = 0; i <= MAX_RCU_LVLS; i++)
3288 rcu_num_nodes += num_rcu_lvl[i];
3289 rcu_num_nodes -= n;
3290}
3291
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08003292void __init rcu_init(void)
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07003293{
Paul E. McKenney017c4262010-01-14 16:10:58 -08003294 int cpu;
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08003295
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07003296 rcu_bootup_announce();
Paul E. McKenneyf885b7f2012-04-23 15:52:53 -07003297 rcu_init_geometry();
Lai Jiangshan394f99a2010-06-28 16:25:04 +08003298 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3299 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
Paul E. McKenneyf41d9112009-08-22 13:56:52 -07003300 __rcu_init_preempt();
Jiang Fangb5b39362013-02-02 14:13:42 -08003301 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
Paul E. McKenney9f680ab2009-11-22 08:53:49 -08003302
3303 /*
3304 * We don't need protection against CPU-hotplug here because
3305 * this is called early in boot, before either interrupts
3306 * or the scheduler are operational.
3307 */
3308 cpu_notifier(rcu_cpu_notify, 0);
Borislav Petkovd1d74d12013-04-22 00:12:42 +02003309 pm_notifier(rcu_pm_notify, 0);
Paul E. McKenney017c4262010-01-14 16:10:58 -08003310 for_each_online_cpu(cpu)
3311 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
Paul E. McKenney64db4cf2008-12-18 21:55:32 +01003312}
3313
Paul E. McKenney1eba8f82009-09-23 09:50:42 -07003314#include "rcutree_plugin.h"