aboutsummaryrefslogtreecommitdiffstats
path: root/kernel/rcu
diff options
context:
space:
mode:
Diffstat (limited to 'kernel/rcu')
-rw-r--r--kernel/rcu/Makefile6
-rw-r--r--kernel/rcu/rcu.h132
-rw-r--r--kernel/rcu/srcu.c651
-rw-r--r--kernel/rcu/tiny.c388
-rw-r--r--kernel/rcu/tiny_plugin.h174
-rw-r--r--kernel/rcu/torture.c2145
-rw-r--r--kernel/rcu/tree.c3416
-rw-r--r--kernel/rcu/tree.h585
-rw-r--r--kernel/rcu/tree_plugin.h2831
-rw-r--r--kernel/rcu/tree_trace.c500
-rw-r--r--kernel/rcu/update.c347
11 files changed, 11175 insertions, 0 deletions
diff --git a/kernel/rcu/Makefile b/kernel/rcu/Makefile
new file mode 100644
index 000000000000..01e9ec37a3e3
--- /dev/null
+++ b/kernel/rcu/Makefile
@@ -0,0 +1,6 @@
1obj-y += update.o srcu.o
2obj-$(CONFIG_RCU_TORTURE_TEST) += torture.o
3obj-$(CONFIG_TREE_RCU) += tree.o
4obj-$(CONFIG_TREE_PREEMPT_RCU) += tree.o
5obj-$(CONFIG_TREE_RCU_TRACE) += tree_trace.o
6obj-$(CONFIG_TINY_RCU) += tiny.o
diff --git a/kernel/rcu/rcu.h b/kernel/rcu/rcu.h
new file mode 100644
index 000000000000..7859a0a3951e
--- /dev/null
+++ b/kernel/rcu/rcu.h
@@ -0,0 +1,132 @@
1/*
2 * Read-Copy Update definitions shared among RCU implementations.
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, 2011
19 *
20 * Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
21 */
22
23#ifndef __LINUX_RCU_H
24#define __LINUX_RCU_H
25
26#ifdef CONFIG_RCU_TRACE
27#define RCU_TRACE(stmt) stmt
28#else /* #ifdef CONFIG_RCU_TRACE */
29#define RCU_TRACE(stmt)
30#endif /* #else #ifdef CONFIG_RCU_TRACE */
31
32/*
33 * Process-level increment to ->dynticks_nesting field. This allows for
34 * architectures that use half-interrupts and half-exceptions from
35 * process context.
36 *
37 * DYNTICK_TASK_NEST_MASK defines a field of width DYNTICK_TASK_NEST_WIDTH
38 * that counts the number of process-based reasons why RCU cannot
39 * consider the corresponding CPU to be idle, and DYNTICK_TASK_NEST_VALUE
40 * is the value used to increment or decrement this field.
41 *
42 * The rest of the bits could in principle be used to count interrupts,
43 * but this would mean that a negative-one value in the interrupt
44 * field could incorrectly zero out the DYNTICK_TASK_NEST_MASK field.
45 * We therefore provide a two-bit guard field defined by DYNTICK_TASK_MASK
46 * that is set to DYNTICK_TASK_FLAG upon initial exit from idle.
47 * The DYNTICK_TASK_EXIT_IDLE value is thus the combined value used upon
48 * initial exit from idle.
49 */
50#define DYNTICK_TASK_NEST_WIDTH 7
51#define DYNTICK_TASK_NEST_VALUE ((LLONG_MAX >> DYNTICK_TASK_NEST_WIDTH) + 1)
52#define DYNTICK_TASK_NEST_MASK (LLONG_MAX - DYNTICK_TASK_NEST_VALUE + 1)
53#define DYNTICK_TASK_FLAG ((DYNTICK_TASK_NEST_VALUE / 8) * 2)
54#define DYNTICK_TASK_MASK ((DYNTICK_TASK_NEST_VALUE / 8) * 3)
55#define DYNTICK_TASK_EXIT_IDLE (DYNTICK_TASK_NEST_VALUE + \
56 DYNTICK_TASK_FLAG)
57
58/*
59 * debug_rcu_head_queue()/debug_rcu_head_unqueue() are used internally
60 * by call_rcu() and rcu callback execution, and are therefore not part of the
61 * RCU API. Leaving in rcupdate.h because they are used by all RCU flavors.
62 */
63
64#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
65# define STATE_RCU_HEAD_READY 0
66# define STATE_RCU_HEAD_QUEUED 1
67
68extern struct debug_obj_descr rcuhead_debug_descr;
69
70static inline int debug_rcu_head_queue(struct rcu_head *head)
71{
72 int r1;
73
74 r1 = debug_object_activate(head, &rcuhead_debug_descr);
75 debug_object_active_state(head, &rcuhead_debug_descr,
76 STATE_RCU_HEAD_READY,
77 STATE_RCU_HEAD_QUEUED);
78 return r1;
79}
80
81static inline void debug_rcu_head_unqueue(struct rcu_head *head)
82{
83 debug_object_active_state(head, &rcuhead_debug_descr,
84 STATE_RCU_HEAD_QUEUED,
85 STATE_RCU_HEAD_READY);
86 debug_object_deactivate(head, &rcuhead_debug_descr);
87}
88#else /* !CONFIG_DEBUG_OBJECTS_RCU_HEAD */
89static inline int debug_rcu_head_queue(struct rcu_head *head)
90{
91 return 0;
92}
93
94static inline void debug_rcu_head_unqueue(struct rcu_head *head)
95{
96}
97#endif /* #else !CONFIG_DEBUG_OBJECTS_RCU_HEAD */
98
99extern void kfree(const void *);
100
101static inline bool __rcu_reclaim(const char *rn, struct rcu_head *head)
102{
103 unsigned long offset = (unsigned long)head->func;
104
105 if (__is_kfree_rcu_offset(offset)) {
106 RCU_TRACE(trace_rcu_invoke_kfree_callback(rn, head, offset));
107 kfree((void *)head - offset);
108 return 1;
109 } else {
110 RCU_TRACE(trace_rcu_invoke_callback(rn, head));
111 head->func(head);
112 return 0;
113 }
114}
115
116extern int rcu_expedited;
117
118#ifdef CONFIG_RCU_STALL_COMMON
119
120extern int rcu_cpu_stall_suppress;
121int rcu_jiffies_till_stall_check(void);
122
123#endif /* #ifdef CONFIG_RCU_STALL_COMMON */
124
125/*
126 * Strings used in tracepoints need to be exported via the
127 * tracing system such that tools like perf and trace-cmd can
128 * translate the string address pointers to actual text.
129 */
130#define TPS(x) tracepoint_string(x)
131
132#endif /* __LINUX_RCU_H */
diff --git a/kernel/rcu/srcu.c b/kernel/rcu/srcu.c
new file mode 100644
index 000000000000..01d5ccb8bfe3
--- /dev/null
+++ b/kernel/rcu/srcu.c
@@ -0,0 +1,651 @@
1/*
2 * Sleepable 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 (C) IBM Corporation, 2006
19 * Copyright (C) Fujitsu, 2012
20 *
21 * Author: Paul McKenney <paulmck@us.ibm.com>
22 * Lai Jiangshan <laijs@cn.fujitsu.com>
23 *
24 * For detailed explanation of Read-Copy Update mechanism see -
25 * Documentation/RCU/ *.txt
26 *
27 */
28
29#include <linux/export.h>
30#include <linux/mutex.h>
31#include <linux/percpu.h>
32#include <linux/preempt.h>
33#include <linux/rcupdate.h>
34#include <linux/sched.h>
35#include <linux/smp.h>
36#include <linux/delay.h>
37#include <linux/srcu.h>
38
39#include <trace/events/rcu.h>
40
41#include "rcu.h"
42
43/*
44 * Initialize an rcu_batch structure to empty.
45 */
46static inline void rcu_batch_init(struct rcu_batch *b)
47{
48 b->head = NULL;
49 b->tail = &b->head;
50}
51
52/*
53 * Enqueue a callback onto the tail of the specified rcu_batch structure.
54 */
55static inline void rcu_batch_queue(struct rcu_batch *b, struct rcu_head *head)
56{
57 *b->tail = head;
58 b->tail = &head->next;
59}
60
61/*
62 * Is the specified rcu_batch structure empty?
63 */
64static inline bool rcu_batch_empty(struct rcu_batch *b)
65{
66 return b->tail == &b->head;
67}
68
69/*
70 * Remove the callback at the head of the specified rcu_batch structure
71 * and return a pointer to it, or return NULL if the structure is empty.
72 */
73static inline struct rcu_head *rcu_batch_dequeue(struct rcu_batch *b)
74{
75 struct rcu_head *head;
76
77 if (rcu_batch_empty(b))
78 return NULL;
79
80 head = b->head;
81 b->head = head->next;
82 if (b->tail == &head->next)
83 rcu_batch_init(b);
84
85 return head;
86}
87
88/*
89 * Move all callbacks from the rcu_batch structure specified by "from" to
90 * the structure specified by "to".
91 */
92static inline void rcu_batch_move(struct rcu_batch *to, struct rcu_batch *from)
93{
94 if (!rcu_batch_empty(from)) {
95 *to->tail = from->head;
96 to->tail = from->tail;
97 rcu_batch_init(from);
98 }
99}
100
101static int init_srcu_struct_fields(struct srcu_struct *sp)
102{
103 sp->completed = 0;
104 spin_lock_init(&sp->queue_lock);
105 sp->running = false;
106 rcu_batch_init(&sp->batch_queue);
107 rcu_batch_init(&sp->batch_check0);
108 rcu_batch_init(&sp->batch_check1);
109 rcu_batch_init(&sp->batch_done);
110 INIT_DELAYED_WORK(&sp->work, process_srcu);
111 sp->per_cpu_ref = alloc_percpu(struct srcu_struct_array);
112 return sp->per_cpu_ref ? 0 : -ENOMEM;
113}
114
115#ifdef CONFIG_DEBUG_LOCK_ALLOC
116
117int __init_srcu_struct(struct srcu_struct *sp, const char *name,
118 struct lock_class_key *key)
119{
120 /* Don't re-initialize a lock while it is held. */
121 debug_check_no_locks_freed((void *)sp, sizeof(*sp));
122 lockdep_init_map(&sp->dep_map, name, key, 0);
123 return init_srcu_struct_fields(sp);
124}
125EXPORT_SYMBOL_GPL(__init_srcu_struct);
126
127#else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
128
129/**
130 * init_srcu_struct - initialize a sleep-RCU structure
131 * @sp: structure to initialize.
132 *
133 * Must invoke this on a given srcu_struct before passing that srcu_struct
134 * to any other function. Each srcu_struct represents a separate domain
135 * of SRCU protection.
136 */
137int init_srcu_struct(struct srcu_struct *sp)
138{
139 return init_srcu_struct_fields(sp);
140}
141EXPORT_SYMBOL_GPL(init_srcu_struct);
142
143#endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
144
145/*
146 * Returns approximate total of the readers' ->seq[] values for the
147 * rank of per-CPU counters specified by idx.
148 */
149static unsigned long srcu_readers_seq_idx(struct srcu_struct *sp, int idx)
150{
151 int cpu;
152 unsigned long sum = 0;
153 unsigned long t;
154
155 for_each_possible_cpu(cpu) {
156 t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->seq[idx]);
157 sum += t;
158 }
159 return sum;
160}
161
162/*
163 * Returns approximate number of readers active on the specified rank
164 * of the per-CPU ->c[] counters.
165 */
166static unsigned long srcu_readers_active_idx(struct srcu_struct *sp, int idx)
167{
168 int cpu;
169 unsigned long sum = 0;
170 unsigned long t;
171
172 for_each_possible_cpu(cpu) {
173 t = ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[idx]);
174 sum += t;
175 }
176 return sum;
177}
178
179/*
180 * Return true if the number of pre-existing readers is determined to
181 * be stably zero. An example unstable zero can occur if the call
182 * to srcu_readers_active_idx() misses an __srcu_read_lock() increment,
183 * but due to task migration, sees the corresponding __srcu_read_unlock()
184 * decrement. This can happen because srcu_readers_active_idx() takes
185 * time to sum the array, and might in fact be interrupted or preempted
186 * partway through the summation.
187 */
188static bool srcu_readers_active_idx_check(struct srcu_struct *sp, int idx)
189{
190 unsigned long seq;
191
192 seq = srcu_readers_seq_idx(sp, idx);
193
194 /*
195 * The following smp_mb() A pairs with the smp_mb() B located in
196 * __srcu_read_lock(). This pairing ensures that if an
197 * __srcu_read_lock() increments its counter after the summation
198 * in srcu_readers_active_idx(), then the corresponding SRCU read-side
199 * critical section will see any changes made prior to the start
200 * of the current SRCU grace period.
201 *
202 * Also, if the above call to srcu_readers_seq_idx() saw the
203 * increment of ->seq[], then the call to srcu_readers_active_idx()
204 * must see the increment of ->c[].
205 */
206 smp_mb(); /* A */
207
208 /*
209 * Note that srcu_readers_active_idx() can incorrectly return
210 * zero even though there is a pre-existing reader throughout.
211 * To see this, suppose that task A is in a very long SRCU
212 * read-side critical section that started on CPU 0, and that
213 * no other reader exists, so that the sum of the counters
214 * is equal to one. Then suppose that task B starts executing
215 * srcu_readers_active_idx(), summing up to CPU 1, and then that
216 * task C starts reading on CPU 0, so that its increment is not
217 * summed, but finishes reading on CPU 2, so that its decrement
218 * -is- summed. Then when task B completes its sum, it will
219 * incorrectly get zero, despite the fact that task A has been
220 * in its SRCU read-side critical section the whole time.
221 *
222 * We therefore do a validation step should srcu_readers_active_idx()
223 * return zero.
224 */
225 if (srcu_readers_active_idx(sp, idx) != 0)
226 return false;
227
228 /*
229 * The remainder of this function is the validation step.
230 * The following smp_mb() D pairs with the smp_mb() C in
231 * __srcu_read_unlock(). If the __srcu_read_unlock() was seen
232 * by srcu_readers_active_idx() above, then any destructive
233 * operation performed after the grace period will happen after
234 * the corresponding SRCU read-side critical section.
235 *
236 * Note that there can be at most NR_CPUS worth of readers using
237 * the old index, which is not enough to overflow even a 32-bit
238 * integer. (Yes, this does mean that systems having more than
239 * a billion or so CPUs need to be 64-bit systems.) Therefore,
240 * the sum of the ->seq[] counters cannot possibly overflow.
241 * Therefore, the only way that the return values of the two
242 * calls to srcu_readers_seq_idx() can be equal is if there were
243 * no increments of the corresponding rank of ->seq[] counts
244 * in the interim. But the missed-increment scenario laid out
245 * above includes an increment of the ->seq[] counter by
246 * the corresponding __srcu_read_lock(). Therefore, if this
247 * scenario occurs, the return values from the two calls to
248 * srcu_readers_seq_idx() will differ, and thus the validation
249 * step below suffices.
250 */
251 smp_mb(); /* D */
252
253 return srcu_readers_seq_idx(sp, idx) == seq;
254}
255
256/**
257 * srcu_readers_active - returns approximate number of readers.
258 * @sp: which srcu_struct to count active readers (holding srcu_read_lock).
259 *
260 * Note that this is not an atomic primitive, and can therefore suffer
261 * severe errors when invoked on an active srcu_struct. That said, it
262 * can be useful as an error check at cleanup time.
263 */
264static int srcu_readers_active(struct srcu_struct *sp)
265{
266 int cpu;
267 unsigned long sum = 0;
268
269 for_each_possible_cpu(cpu) {
270 sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[0]);
271 sum += ACCESS_ONCE(per_cpu_ptr(sp->per_cpu_ref, cpu)->c[1]);
272 }
273 return sum;
274}
275
276/**
277 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
278 * @sp: structure to clean up.
279 *
280 * Must invoke this after you are finished using a given srcu_struct that
281 * was initialized via init_srcu_struct(), else you leak memory.
282 */
283void cleanup_srcu_struct(struct srcu_struct *sp)
284{
285 if (WARN_ON(srcu_readers_active(sp)))
286 return; /* Leakage unless caller handles error. */
287 free_percpu(sp->per_cpu_ref);
288 sp->per_cpu_ref = NULL;
289}
290EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
291
292/*
293 * Counts the new reader in the appropriate per-CPU element of the
294 * srcu_struct. Must be called from process context.
295 * Returns an index that must be passed to the matching srcu_read_unlock().
296 */
297int __srcu_read_lock(struct srcu_struct *sp)
298{
299 int idx;
300
301 idx = ACCESS_ONCE(sp->completed) & 0x1;
302 preempt_disable();
303 ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->c[idx]) += 1;
304 smp_mb(); /* B */ /* Avoid leaking the critical section. */
305 ACCESS_ONCE(this_cpu_ptr(sp->per_cpu_ref)->seq[idx]) += 1;
306 preempt_enable();
307 return idx;
308}
309EXPORT_SYMBOL_GPL(__srcu_read_lock);
310
311/*
312 * Removes the count for the old reader from the appropriate per-CPU
313 * element of the srcu_struct. Note that this may well be a different
314 * CPU than that which was incremented by the corresponding srcu_read_lock().
315 * Must be called from process context.
316 */
317void __srcu_read_unlock(struct srcu_struct *sp, int idx)
318{
319 smp_mb(); /* C */ /* Avoid leaking the critical section. */
320 this_cpu_dec(sp->per_cpu_ref->c[idx]);
321}
322EXPORT_SYMBOL_GPL(__srcu_read_unlock);
323
324/*
325 * We use an adaptive strategy for synchronize_srcu() and especially for
326 * synchronize_srcu_expedited(). We spin for a fixed time period
327 * (defined below) to allow SRCU readers to exit their read-side critical
328 * sections. If there are still some readers after 10 microseconds,
329 * we repeatedly block for 1-millisecond time periods. This approach
330 * has done well in testing, so there is no need for a config parameter.
331 */
332#define SRCU_RETRY_CHECK_DELAY 5
333#define SYNCHRONIZE_SRCU_TRYCOUNT 2
334#define SYNCHRONIZE_SRCU_EXP_TRYCOUNT 12
335
336/*
337 * @@@ Wait until all pre-existing readers complete. Such readers
338 * will have used the index specified by "idx".
339 * the caller should ensures the ->completed is not changed while checking
340 * and idx = (->completed & 1) ^ 1
341 */
342static bool try_check_zero(struct srcu_struct *sp, int idx, int trycount)
343{
344 for (;;) {
345 if (srcu_readers_active_idx_check(sp, idx))
346 return true;
347 if (--trycount <= 0)
348 return false;
349 udelay(SRCU_RETRY_CHECK_DELAY);
350 }
351}
352
353/*
354 * Increment the ->completed counter so that future SRCU readers will
355 * use the other rank of the ->c[] and ->seq[] arrays. This allows
356 * us to wait for pre-existing readers in a starvation-free manner.
357 */
358static void srcu_flip(struct srcu_struct *sp)
359{
360 sp->completed++;
361}
362
363/*
364 * Enqueue an SRCU callback on the specified srcu_struct structure,
365 * initiating grace-period processing if it is not already running.
366 */
367void call_srcu(struct srcu_struct *sp, struct rcu_head *head,
368 void (*func)(struct rcu_head *head))
369{
370 unsigned long flags;
371
372 head->next = NULL;
373 head->func = func;
374 spin_lock_irqsave(&sp->queue_lock, flags);
375 rcu_batch_queue(&sp->batch_queue, head);
376 if (!sp->running) {
377 sp->running = true;
378 schedule_delayed_work(&sp->work, 0);
379 }
380 spin_unlock_irqrestore(&sp->queue_lock, flags);
381}
382EXPORT_SYMBOL_GPL(call_srcu);
383
384struct rcu_synchronize {
385 struct rcu_head head;
386 struct completion completion;
387};
388
389/*
390 * Awaken the corresponding synchronize_srcu() instance now that a
391 * grace period has elapsed.
392 */
393static void wakeme_after_rcu(struct rcu_head *head)
394{
395 struct rcu_synchronize *rcu;
396
397 rcu = container_of(head, struct rcu_synchronize, head);
398 complete(&rcu->completion);
399}
400
401static void srcu_advance_batches(struct srcu_struct *sp, int trycount);
402static void srcu_reschedule(struct srcu_struct *sp);
403
404/*
405 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
406 */
407static void __synchronize_srcu(struct srcu_struct *sp, int trycount)
408{
409 struct rcu_synchronize rcu;
410 struct rcu_head *head = &rcu.head;
411 bool done = false;
412
413 rcu_lockdep_assert(!lock_is_held(&sp->dep_map) &&
414 !lock_is_held(&rcu_bh_lock_map) &&
415 !lock_is_held(&rcu_lock_map) &&
416 !lock_is_held(&rcu_sched_lock_map),
417 "Illegal synchronize_srcu() in same-type SRCU (or RCU) read-side critical section");
418
419 might_sleep();
420 init_completion(&rcu.completion);
421
422 head->next = NULL;
423 head->func = wakeme_after_rcu;
424 spin_lock_irq(&sp->queue_lock);
425 if (!sp->running) {
426 /* steal the processing owner */
427 sp->running = true;
428 rcu_batch_queue(&sp->batch_check0, head);
429 spin_unlock_irq(&sp->queue_lock);
430
431 srcu_advance_batches(sp, trycount);
432 if (!rcu_batch_empty(&sp->batch_done)) {
433 BUG_ON(sp->batch_done.head != head);
434 rcu_batch_dequeue(&sp->batch_done);
435 done = true;
436 }
437 /* give the processing owner to work_struct */
438 srcu_reschedule(sp);
439 } else {
440 rcu_batch_queue(&sp->batch_queue, head);
441 spin_unlock_irq(&sp->queue_lock);
442 }
443
444 if (!done)
445 wait_for_completion(&rcu.completion);
446}
447
448/**
449 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
450 * @sp: srcu_struct with which to synchronize.
451 *
452 * Wait for the count to drain to zero of both indexes. To avoid the
453 * possible starvation of synchronize_srcu(), it waits for the count of
454 * the index=((->completed & 1) ^ 1) to drain to zero at first,
455 * and then flip the completed and wait for the count of the other index.
456 *
457 * Can block; must be called from process context.
458 *
459 * Note that it is illegal to call synchronize_srcu() from the corresponding
460 * SRCU read-side critical section; doing so will result in deadlock.
461 * However, it is perfectly legal to call synchronize_srcu() on one
462 * srcu_struct from some other srcu_struct's read-side critical section.
463 */
464void synchronize_srcu(struct srcu_struct *sp)
465{
466 __synchronize_srcu(sp, rcu_expedited
467 ? SYNCHRONIZE_SRCU_EXP_TRYCOUNT
468 : SYNCHRONIZE_SRCU_TRYCOUNT);
469}
470EXPORT_SYMBOL_GPL(synchronize_srcu);
471
472/**
473 * synchronize_srcu_expedited - Brute-force SRCU grace period
474 * @sp: srcu_struct with which to synchronize.
475 *
476 * Wait for an SRCU grace period to elapse, but be more aggressive about
477 * spinning rather than blocking when waiting.
478 *
479 * Note that it is also illegal to call synchronize_srcu_expedited()
480 * from the corresponding SRCU read-side critical section;
481 * doing so will result in deadlock. However, it is perfectly legal
482 * to call synchronize_srcu_expedited() on one srcu_struct from some
483 * other srcu_struct's read-side critical section, as long as
484 * the resulting graph of srcu_structs is acyclic.
485 */
486void synchronize_srcu_expedited(struct srcu_struct *sp)
487{
488 __synchronize_srcu(sp, SYNCHRONIZE_SRCU_EXP_TRYCOUNT);
489}
490EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
491
492/**
493 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
494 */
495void srcu_barrier(struct srcu_struct *sp)
496{
497 synchronize_srcu(sp);
498}
499EXPORT_SYMBOL_GPL(srcu_barrier);
500
501/**
502 * srcu_batches_completed - return batches completed.
503 * @sp: srcu_struct on which to report batch completion.
504 *
505 * Report the number of batches, correlated with, but not necessarily
506 * precisely the same as, the number of grace periods that have elapsed.
507 */
508long srcu_batches_completed(struct srcu_struct *sp)
509{
510 return sp->completed;
511}
512EXPORT_SYMBOL_GPL(srcu_batches_completed);
513
514#define SRCU_CALLBACK_BATCH 10
515#define SRCU_INTERVAL 1
516
517/*
518 * Move any new SRCU callbacks to the first stage of the SRCU grace
519 * period pipeline.
520 */
521static void srcu_collect_new(struct srcu_struct *sp)
522{
523 if (!rcu_batch_empty(&sp->batch_queue)) {
524 spin_lock_irq(&sp->queue_lock);
525 rcu_batch_move(&sp->batch_check0, &sp->batch_queue);
526 spin_unlock_irq(&sp->queue_lock);
527 }
528}
529
530/*
531 * Core SRCU state machine. Advance callbacks from ->batch_check0 to
532 * ->batch_check1 and then to ->batch_done as readers drain.
533 */
534static void srcu_advance_batches(struct srcu_struct *sp, int trycount)
535{
536 int idx = 1 ^ (sp->completed & 1);
537
538 /*
539 * Because readers might be delayed for an extended period after
540 * fetching ->completed for their index, at any point in time there
541 * might well be readers using both idx=0 and idx=1. We therefore
542 * need to wait for readers to clear from both index values before
543 * invoking a callback.
544 */
545
546 if (rcu_batch_empty(&sp->batch_check0) &&
547 rcu_batch_empty(&sp->batch_check1))
548 return; /* no callbacks need to be advanced */
549
550 if (!try_check_zero(sp, idx, trycount))
551 return; /* failed to advance, will try after SRCU_INTERVAL */
552
553 /*
554 * The callbacks in ->batch_check1 have already done with their
555 * first zero check and flip back when they were enqueued on
556 * ->batch_check0 in a previous invocation of srcu_advance_batches().
557 * (Presumably try_check_zero() returned false during that
558 * invocation, leaving the callbacks stranded on ->batch_check1.)
559 * They are therefore ready to invoke, so move them to ->batch_done.
560 */
561 rcu_batch_move(&sp->batch_done, &sp->batch_check1);
562
563 if (rcu_batch_empty(&sp->batch_check0))
564 return; /* no callbacks need to be advanced */
565 srcu_flip(sp);
566
567 /*
568 * The callbacks in ->batch_check0 just finished their
569 * first check zero and flip, so move them to ->batch_check1
570 * for future checking on the other idx.
571 */
572 rcu_batch_move(&sp->batch_check1, &sp->batch_check0);
573
574 /*
575 * SRCU read-side critical sections are normally short, so check
576 * at least twice in quick succession after a flip.
577 */
578 trycount = trycount < 2 ? 2 : trycount;
579 if (!try_check_zero(sp, idx^1, trycount))
580 return; /* failed to advance, will try after SRCU_INTERVAL */
581
582 /*
583 * The callbacks in ->batch_check1 have now waited for all
584 * pre-existing readers using both idx values. They are therefore
585 * ready to invoke, so move them to ->batch_done.
586 */
587 rcu_batch_move(&sp->batch_done, &sp->batch_check1);
588}
589
590/*
591 * Invoke a limited number of SRCU callbacks that have passed through
592 * their grace period. If there are more to do, SRCU will reschedule
593 * the workqueue.
594 */
595static void srcu_invoke_callbacks(struct srcu_struct *sp)
596{
597 int i;
598 struct rcu_head *head;
599
600 for (i = 0; i < SRCU_CALLBACK_BATCH; i++) {
601 head = rcu_batch_dequeue(&sp->batch_done);
602 if (!head)
603 break;
604 local_bh_disable();
605 head->func(head);
606 local_bh_enable();
607 }
608}
609
610/*
611 * Finished one round of SRCU grace period. Start another if there are
612 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
613 */
614static void srcu_reschedule(struct srcu_struct *sp)
615{
616 bool pending = true;
617
618 if (rcu_batch_empty(&sp->batch_done) &&
619 rcu_batch_empty(&sp->batch_check1) &&
620 rcu_batch_empty(&sp->batch_check0) &&
621 rcu_batch_empty(&sp->batch_queue)) {
622 spin_lock_irq(&sp->queue_lock);
623 if (rcu_batch_empty(&sp->batch_done) &&
624 rcu_batch_empty(&sp->batch_check1) &&
625 rcu_batch_empty(&sp->batch_check0) &&
626 rcu_batch_empty(&sp->batch_queue)) {
627 sp->running = false;
628 pending = false;
629 }
630 spin_unlock_irq(&sp->queue_lock);
631 }
632
633 if (pending)
634 schedule_delayed_work(&sp->work, SRCU_INTERVAL);
635}
636
637/*
638 * This is the work-queue function that handles SRCU grace periods.
639 */
640void process_srcu(struct work_struct *work)
641{
642 struct srcu_struct *sp;
643
644 sp = container_of(work, struct srcu_struct, work.work);
645
646 srcu_collect_new(sp);
647 srcu_advance_batches(sp, 1);
648 srcu_invoke_callbacks(sp);
649 srcu_reschedule(sp);
650}
651EXPORT_SYMBOL_GPL(process_srcu);
diff --git a/kernel/rcu/tiny.c b/kernel/rcu/tiny.c
new file mode 100644
index 000000000000..0c9a934cfec1
--- /dev/null
+++ b/kernel/rcu/tiny.c
@@ -0,0 +1,388 @@
1/*
2 * Read-Copy Update mechanism for mutual exclusion, the Bloatwatch edition.
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 * Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
21 *
22 * For detailed explanation of Read-Copy Update mechanism see -
23 * Documentation/RCU
24 */
25#include <linux/completion.h>
26#include <linux/interrupt.h>
27#include <linux/notifier.h>
28#include <linux/rcupdate.h>
29#include <linux/kernel.h>
30#include <linux/export.h>
31#include <linux/mutex.h>
32#include <linux/sched.h>
33#include <linux/types.h>
34#include <linux/init.h>
35#include <linux/time.h>
36#include <linux/cpu.h>
37#include <linux/prefetch.h>
38#include <linux/ftrace_event.h>
39
40#ifdef CONFIG_RCU_TRACE
41#include <trace/events/rcu.h>
42#endif /* #else #ifdef CONFIG_RCU_TRACE */
43
44#include "rcu.h"
45
46/* Forward declarations for tiny_plugin.h. */
47struct rcu_ctrlblk;
48static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp);
49static void rcu_process_callbacks(struct softirq_action *unused);
50static void __call_rcu(struct rcu_head *head,
51 void (*func)(struct rcu_head *rcu),
52 struct rcu_ctrlblk *rcp);
53
54static long long rcu_dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
55
56#include "tiny_plugin.h"
57
58/* Common code for rcu_idle_enter() and rcu_irq_exit(), see kernel/rcutree.c. */
59static void rcu_idle_enter_common(long long newval)
60{
61 if (newval) {
62 RCU_TRACE(trace_rcu_dyntick(TPS("--="),
63 rcu_dynticks_nesting, newval));
64 rcu_dynticks_nesting = newval;
65 return;
66 }
67 RCU_TRACE(trace_rcu_dyntick(TPS("Start"),
68 rcu_dynticks_nesting, newval));
69 if (!is_idle_task(current)) {
70 struct task_struct *idle __maybe_unused = idle_task(smp_processor_id());
71
72 RCU_TRACE(trace_rcu_dyntick(TPS("Entry error: not idle task"),
73 rcu_dynticks_nesting, newval));
74 ftrace_dump(DUMP_ALL);
75 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
76 current->pid, current->comm,
77 idle->pid, idle->comm); /* must be idle task! */
78 }
79 rcu_sched_qs(0); /* implies rcu_bh_qsctr_inc(0) */
80 barrier();
81 rcu_dynticks_nesting = newval;
82}
83
84/*
85 * Enter idle, which is an extended quiescent state if we have fully
86 * entered that mode (i.e., if the new value of dynticks_nesting is zero).
87 */
88void rcu_idle_enter(void)
89{
90 unsigned long flags;
91 long long newval;
92
93 local_irq_save(flags);
94 WARN_ON_ONCE((rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK) == 0);
95 if ((rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK) ==
96 DYNTICK_TASK_NEST_VALUE)
97 newval = 0;
98 else
99 newval = rcu_dynticks_nesting - DYNTICK_TASK_NEST_VALUE;
100 rcu_idle_enter_common(newval);
101 local_irq_restore(flags);
102}
103EXPORT_SYMBOL_GPL(rcu_idle_enter);
104
105/*
106 * Exit an interrupt handler towards idle.
107 */
108void rcu_irq_exit(void)
109{
110 unsigned long flags;
111 long long newval;
112
113 local_irq_save(flags);
114 newval = rcu_dynticks_nesting - 1;
115 WARN_ON_ONCE(newval < 0);
116 rcu_idle_enter_common(newval);
117 local_irq_restore(flags);
118}
119EXPORT_SYMBOL_GPL(rcu_irq_exit);
120
121/* Common code for rcu_idle_exit() and rcu_irq_enter(), see kernel/rcutree.c. */
122static void rcu_idle_exit_common(long long oldval)
123{
124 if (oldval) {
125 RCU_TRACE(trace_rcu_dyntick(TPS("++="),
126 oldval, rcu_dynticks_nesting));
127 return;
128 }
129 RCU_TRACE(trace_rcu_dyntick(TPS("End"), oldval, rcu_dynticks_nesting));
130 if (!is_idle_task(current)) {
131 struct task_struct *idle __maybe_unused = idle_task(smp_processor_id());
132
133 RCU_TRACE(trace_rcu_dyntick(TPS("Exit error: not idle task"),
134 oldval, rcu_dynticks_nesting));
135 ftrace_dump(DUMP_ALL);
136 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
137 current->pid, current->comm,
138 idle->pid, idle->comm); /* must be idle task! */
139 }
140}
141
142/*
143 * Exit idle, so that we are no longer in an extended quiescent state.
144 */
145void rcu_idle_exit(void)
146{
147 unsigned long flags;
148 long long oldval;
149
150 local_irq_save(flags);
151 oldval = rcu_dynticks_nesting;
152 WARN_ON_ONCE(rcu_dynticks_nesting < 0);
153 if (rcu_dynticks_nesting & DYNTICK_TASK_NEST_MASK)
154 rcu_dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
155 else
156 rcu_dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
157 rcu_idle_exit_common(oldval);
158 local_irq_restore(flags);
159}
160EXPORT_SYMBOL_GPL(rcu_idle_exit);
161
162/*
163 * Enter an interrupt handler, moving away from idle.
164 */
165void rcu_irq_enter(void)
166{
167 unsigned long flags;
168 long long oldval;
169
170 local_irq_save(flags);
171 oldval = rcu_dynticks_nesting;
172 rcu_dynticks_nesting++;
173 WARN_ON_ONCE(rcu_dynticks_nesting == 0);
174 rcu_idle_exit_common(oldval);
175 local_irq_restore(flags);
176}
177EXPORT_SYMBOL_GPL(rcu_irq_enter);
178
179#if defined(CONFIG_DEBUG_LOCK_ALLOC) || defined(CONFIG_RCU_TRACE)
180
181/*
182 * Test whether RCU thinks that the current CPU is idle.
183 */
184bool __rcu_is_watching(void)
185{
186 return rcu_dynticks_nesting;
187}
188EXPORT_SYMBOL(__rcu_is_watching);
189
190#endif /* defined(CONFIG_DEBUG_LOCK_ALLOC) || defined(CONFIG_RCU_TRACE) */
191
192/*
193 * Test whether the current CPU was interrupted from idle. Nested
194 * interrupts don't count, we must be running at the first interrupt
195 * level.
196 */
197static int rcu_is_cpu_rrupt_from_idle(void)
198{
199 return rcu_dynticks_nesting <= 1;
200}
201
202/*
203 * Helper function for rcu_sched_qs() and rcu_bh_qs().
204 * Also irqs are disabled to avoid confusion due to interrupt handlers
205 * invoking call_rcu().
206 */
207static int rcu_qsctr_help(struct rcu_ctrlblk *rcp)
208{
209 RCU_TRACE(reset_cpu_stall_ticks(rcp));
210 if (rcp->rcucblist != NULL &&
211 rcp->donetail != rcp->curtail) {
212 rcp->donetail = rcp->curtail;
213 return 1;
214 }
215
216 return 0;
217}
218
219/*
220 * Record an rcu quiescent state. And an rcu_bh quiescent state while we
221 * are at it, given that any rcu quiescent state is also an rcu_bh
222 * quiescent state. Use "+" instead of "||" to defeat short circuiting.
223 */
224void rcu_sched_qs(int cpu)
225{
226 unsigned long flags;
227
228 local_irq_save(flags);
229 if (rcu_qsctr_help(&rcu_sched_ctrlblk) +
230 rcu_qsctr_help(&rcu_bh_ctrlblk))
231 raise_softirq(RCU_SOFTIRQ);
232 local_irq_restore(flags);
233}
234
235/*
236 * Record an rcu_bh quiescent state.
237 */
238void rcu_bh_qs(int cpu)
239{
240 unsigned long flags;
241
242 local_irq_save(flags);
243 if (rcu_qsctr_help(&rcu_bh_ctrlblk))
244 raise_softirq(RCU_SOFTIRQ);
245 local_irq_restore(flags);
246}
247
248/*
249 * Check to see if the scheduling-clock interrupt came from an extended
250 * quiescent state, and, if so, tell RCU about it. This function must
251 * be called from hardirq context. It is normally called from the
252 * scheduling-clock interrupt.
253 */
254void rcu_check_callbacks(int cpu, int user)
255{
256 RCU_TRACE(check_cpu_stalls());
257 if (user || rcu_is_cpu_rrupt_from_idle())
258 rcu_sched_qs(cpu);
259 else if (!in_softirq())
260 rcu_bh_qs(cpu);
261}
262
263/*
264 * Invoke the RCU callbacks on the specified rcu_ctrlkblk structure
265 * whose grace period has elapsed.
266 */
267static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp)
268{
269 const char *rn = NULL;
270 struct rcu_head *next, *list;
271 unsigned long flags;
272 RCU_TRACE(int cb_count = 0);
273
274 /* If no RCU callbacks ready to invoke, just return. */
275 if (&rcp->rcucblist == rcp->donetail) {
276 RCU_TRACE(trace_rcu_batch_start(rcp->name, 0, 0, -1));
277 RCU_TRACE(trace_rcu_batch_end(rcp->name, 0,
278 !!ACCESS_ONCE(rcp->rcucblist),
279 need_resched(),
280 is_idle_task(current),
281 false));
282 return;
283 }
284
285 /* Move the ready-to-invoke callbacks to a local list. */
286 local_irq_save(flags);
287 RCU_TRACE(trace_rcu_batch_start(rcp->name, 0, rcp->qlen, -1));
288 list = rcp->rcucblist;
289 rcp->rcucblist = *rcp->donetail;
290 *rcp->donetail = NULL;
291 if (rcp->curtail == rcp->donetail)
292 rcp->curtail = &rcp->rcucblist;
293 rcp->donetail = &rcp->rcucblist;
294 local_irq_restore(flags);
295
296 /* Invoke the callbacks on the local list. */
297 RCU_TRACE(rn = rcp->name);
298 while (list) {
299 next = list->next;
300 prefetch(next);
301 debug_rcu_head_unqueue(list);
302 local_bh_disable();
303 __rcu_reclaim(rn, list);
304 local_bh_enable();
305 list = next;
306 RCU_TRACE(cb_count++);
307 }
308 RCU_TRACE(rcu_trace_sub_qlen(rcp, cb_count));
309 RCU_TRACE(trace_rcu_batch_end(rcp->name,
310 cb_count, 0, need_resched(),
311 is_idle_task(current),
312 false));
313}
314
315static void rcu_process_callbacks(struct softirq_action *unused)
316{
317 __rcu_process_callbacks(&rcu_sched_ctrlblk);
318 __rcu_process_callbacks(&rcu_bh_ctrlblk);
319}
320
321/*
322 * Wait for a grace period to elapse. But it is illegal to invoke
323 * synchronize_sched() from within an RCU read-side critical section.
324 * Therefore, any legal call to synchronize_sched() is a quiescent
325 * state, and so on a UP system, synchronize_sched() need do nothing.
326 * Ditto for synchronize_rcu_bh(). (But Lai Jiangshan points out the
327 * benefits of doing might_sleep() to reduce latency.)
328 *
329 * Cool, huh? (Due to Josh Triplett.)
330 *
331 * But we want to make this a static inline later. The cond_resched()
332 * currently makes this problematic.
333 */
334void synchronize_sched(void)
335{
336 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
337 !lock_is_held(&rcu_lock_map) &&
338 !lock_is_held(&rcu_sched_lock_map),
339 "Illegal synchronize_sched() in RCU read-side critical section");
340 cond_resched();
341}
342EXPORT_SYMBOL_GPL(synchronize_sched);
343
344/*
345 * Helper function for call_rcu() and call_rcu_bh().
346 */
347static void __call_rcu(struct rcu_head *head,
348 void (*func)(struct rcu_head *rcu),
349 struct rcu_ctrlblk *rcp)
350{
351 unsigned long flags;
352
353 debug_rcu_head_queue(head);
354 head->func = func;
355 head->next = NULL;
356
357 local_irq_save(flags);
358 *rcp->curtail = head;
359 rcp->curtail = &head->next;
360 RCU_TRACE(rcp->qlen++);
361 local_irq_restore(flags);
362}
363
364/*
365 * Post an RCU callback to be invoked after the end of an RCU-sched grace
366 * period. But since we have but one CPU, that would be after any
367 * quiescent state.
368 */
369void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
370{
371 __call_rcu(head, func, &rcu_sched_ctrlblk);
372}
373EXPORT_SYMBOL_GPL(call_rcu_sched);
374
375/*
376 * Post an RCU bottom-half callback to be invoked after any subsequent
377 * quiescent state.
378 */
379void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
380{
381 __call_rcu(head, func, &rcu_bh_ctrlblk);
382}
383EXPORT_SYMBOL_GPL(call_rcu_bh);
384
385void rcu_init(void)
386{
387 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
388}
diff --git a/kernel/rcu/tiny_plugin.h b/kernel/rcu/tiny_plugin.h
new file mode 100644
index 000000000000..280d06cae352
--- /dev/null
+++ b/kernel/rcu/tiny_plugin.h
@@ -0,0 +1,174 @@
1/*
2 * Read-Copy Update mechanism for mutual exclusion, the Bloatwatch edition
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
19 *
20 * Copyright (c) 2010 Linaro
21 *
22 * Author: Paul E. McKenney <paulmck@linux.vnet.ibm.com>
23 */
24
25#include <linux/kthread.h>
26#include <linux/module.h>
27#include <linux/debugfs.h>
28#include <linux/seq_file.h>
29
30/* Global control variables for rcupdate callback mechanism. */
31struct rcu_ctrlblk {
32 struct rcu_head *rcucblist; /* List of pending callbacks (CBs). */
33 struct rcu_head **donetail; /* ->next pointer of last "done" CB. */
34 struct rcu_head **curtail; /* ->next pointer of last CB. */
35 RCU_TRACE(long qlen); /* Number of pending CBs. */
36 RCU_TRACE(unsigned long gp_start); /* Start time for stalls. */
37 RCU_TRACE(unsigned long ticks_this_gp); /* Statistic for stalls. */
38 RCU_TRACE(unsigned long jiffies_stall); /* Jiffies at next stall. */
39 RCU_TRACE(const char *name); /* Name of RCU type. */
40};
41
42/* Definition for rcupdate control block. */
43static struct rcu_ctrlblk rcu_sched_ctrlblk = {
44 .donetail = &rcu_sched_ctrlblk.rcucblist,
45 .curtail = &rcu_sched_ctrlblk.rcucblist,
46 RCU_TRACE(.name = "rcu_sched")
47};
48
49static struct rcu_ctrlblk rcu_bh_ctrlblk = {
50 .donetail = &rcu_bh_ctrlblk.rcucblist,
51 .curtail = &rcu_bh_ctrlblk.rcucblist,
52 RCU_TRACE(.name = "rcu_bh")
53};
54
55#ifdef CONFIG_DEBUG_LOCK_ALLOC
56#include <linux/kernel_stat.h>
57
58int rcu_scheduler_active __read_mostly;
59EXPORT_SYMBOL_GPL(rcu_scheduler_active);
60
61/*
62 * During boot, we forgive RCU lockdep issues. After this function is
63 * invoked, we start taking RCU lockdep issues seriously.
64 */
65void __init rcu_scheduler_starting(void)
66{
67 WARN_ON(nr_context_switches() > 0);
68 rcu_scheduler_active = 1;
69}
70
71#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
72
73#ifdef CONFIG_RCU_TRACE
74
75static void rcu_trace_sub_qlen(struct rcu_ctrlblk *rcp, int n)
76{
77 unsigned long flags;
78
79 local_irq_save(flags);
80 rcp->qlen -= n;
81 local_irq_restore(flags);
82}
83
84/*
85 * Dump statistics for TINY_RCU, such as they are.
86 */
87static int show_tiny_stats(struct seq_file *m, void *unused)
88{
89 seq_printf(m, "rcu_sched: qlen: %ld\n", rcu_sched_ctrlblk.qlen);
90 seq_printf(m, "rcu_bh: qlen: %ld\n", rcu_bh_ctrlblk.qlen);
91 return 0;
92}
93
94static int show_tiny_stats_open(struct inode *inode, struct file *file)
95{
96 return single_open(file, show_tiny_stats, NULL);
97}
98
99static const struct file_operations show_tiny_stats_fops = {
100 .owner = THIS_MODULE,
101 .open = show_tiny_stats_open,
102 .read = seq_read,
103 .llseek = seq_lseek,
104 .release = single_release,
105};
106
107static struct dentry *rcudir;
108
109static int __init rcutiny_trace_init(void)
110{
111 struct dentry *retval;
112
113 rcudir = debugfs_create_dir("rcu", NULL);
114 if (!rcudir)
115 goto free_out;
116 retval = debugfs_create_file("rcudata", 0444, rcudir,
117 NULL, &show_tiny_stats_fops);
118 if (!retval)
119 goto free_out;
120 return 0;
121free_out:
122 debugfs_remove_recursive(rcudir);
123 return 1;
124}
125
126static void __exit rcutiny_trace_cleanup(void)
127{
128 debugfs_remove_recursive(rcudir);
129}
130
131module_init(rcutiny_trace_init);
132module_exit(rcutiny_trace_cleanup);
133
134MODULE_AUTHOR("Paul E. McKenney");
135MODULE_DESCRIPTION("Read-Copy Update tracing for tiny implementation");
136MODULE_LICENSE("GPL");
137
138static void check_cpu_stall(struct rcu_ctrlblk *rcp)
139{
140 unsigned long j;
141 unsigned long js;
142
143 if (rcu_cpu_stall_suppress)
144 return;
145 rcp->ticks_this_gp++;
146 j = jiffies;
147 js = rcp->jiffies_stall;
148 if (*rcp->curtail && ULONG_CMP_GE(j, js)) {
149 pr_err("INFO: %s stall on CPU (%lu ticks this GP) idle=%llx (t=%lu jiffies q=%ld)\n",
150 rcp->name, rcp->ticks_this_gp, rcu_dynticks_nesting,
151 jiffies - rcp->gp_start, rcp->qlen);
152 dump_stack();
153 }
154 if (*rcp->curtail && ULONG_CMP_GE(j, js))
155 rcp->jiffies_stall = jiffies +
156 3 * rcu_jiffies_till_stall_check() + 3;
157 else if (ULONG_CMP_GE(j, js))
158 rcp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
159}
160
161static void reset_cpu_stall_ticks(struct rcu_ctrlblk *rcp)
162{
163 rcp->ticks_this_gp = 0;
164 rcp->gp_start = jiffies;
165 rcp->jiffies_stall = jiffies + rcu_jiffies_till_stall_check();
166}
167
168static void check_cpu_stalls(void)
169{
170 RCU_TRACE(check_cpu_stall(&rcu_bh_ctrlblk));
171 RCU_TRACE(check_cpu_stall(&rcu_sched_ctrlblk));
172}
173
174#endif /* #ifdef CONFIG_RCU_TRACE */
diff --git a/kernel/rcu/torture.c b/kernel/rcu/torture.c
new file mode 100644
index 000000000000..3929cd451511
--- /dev/null
+++ b/kernel/rcu/torture.c
@@ -0,0 +1,2145 @@
1/*
2 * Read-Copy Update module-based torture test facility
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 (C) IBM Corporation, 2005, 2006
19 *
20 * Authors: Paul E. McKenney <paulmck@us.ibm.com>
21 * Josh Triplett <josh@freedesktop.org>
22 *
23 * See also: Documentation/RCU/torture.txt
24 */
25#include <linux/types.h>
26#include <linux/kernel.h>
27#include <linux/init.h>
28#include <linux/module.h>
29#include <linux/kthread.h>
30#include <linux/err.h>
31#include <linux/spinlock.h>
32#include <linux/smp.h>
33#include <linux/rcupdate.h>
34#include <linux/interrupt.h>
35#include <linux/sched.h>
36#include <linux/atomic.h>
37#include <linux/bitops.h>
38#include <linux/completion.h>
39#include <linux/moduleparam.h>
40#include <linux/percpu.h>
41#include <linux/notifier.h>
42#include <linux/reboot.h>
43#include <linux/freezer.h>
44#include <linux/cpu.h>
45#include <linux/delay.h>
46#include <linux/stat.h>
47#include <linux/srcu.h>
48#include <linux/slab.h>
49#include <linux/trace_clock.h>
50#include <asm/byteorder.h>
51
52MODULE_LICENSE("GPL");
53MODULE_AUTHOR("Paul E. McKenney <paulmck@us.ibm.com> and Josh Triplett <josh@freedesktop.org>");
54
55MODULE_ALIAS("rcutorture");
56#ifdef MODULE_PARAM_PREFIX
57#undef MODULE_PARAM_PREFIX
58#endif
59#define MODULE_PARAM_PREFIX "rcutorture."
60
61static int fqs_duration;
62module_param(fqs_duration, int, 0444);
63MODULE_PARM_DESC(fqs_duration, "Duration of fqs bursts (us), 0 to disable");
64static int fqs_holdoff;
65module_param(fqs_holdoff, int, 0444);
66MODULE_PARM_DESC(fqs_holdoff, "Holdoff time within fqs bursts (us)");
67static int fqs_stutter = 3;
68module_param(fqs_stutter, int, 0444);
69MODULE_PARM_DESC(fqs_stutter, "Wait time between fqs bursts (s)");
70static bool gp_exp;
71module_param(gp_exp, bool, 0444);
72MODULE_PARM_DESC(gp_exp, "Use expedited GP wait primitives");
73static bool gp_normal;
74module_param(gp_normal, bool, 0444);
75MODULE_PARM_DESC(gp_normal, "Use normal (non-expedited) GP wait primitives");
76static int irqreader = 1;
77module_param(irqreader, int, 0444);
78MODULE_PARM_DESC(irqreader, "Allow RCU readers from irq handlers");
79static int n_barrier_cbs;
80module_param(n_barrier_cbs, int, 0444);
81MODULE_PARM_DESC(n_barrier_cbs, "# of callbacks/kthreads for barrier testing");
82static int nfakewriters = 4;
83module_param(nfakewriters, int, 0444);
84MODULE_PARM_DESC(nfakewriters, "Number of RCU fake writer threads");
85static int nreaders = -1;
86module_param(nreaders, int, 0444);
87MODULE_PARM_DESC(nreaders, "Number of RCU reader threads");
88static int object_debug;
89module_param(object_debug, int, 0444);
90MODULE_PARM_DESC(object_debug, "Enable debug-object double call_rcu() testing");
91static int onoff_holdoff;
92module_param(onoff_holdoff, int, 0444);
93MODULE_PARM_DESC(onoff_holdoff, "Time after boot before CPU hotplugs (s)");
94static int onoff_interval;
95module_param(onoff_interval, int, 0444);
96MODULE_PARM_DESC(onoff_interval, "Time between CPU hotplugs (s), 0=disable");
97static int shuffle_interval = 3;
98module_param(shuffle_interval, int, 0444);
99MODULE_PARM_DESC(shuffle_interval, "Number of seconds between shuffles");
100static int shutdown_secs;
101module_param(shutdown_secs, int, 0444);
102MODULE_PARM_DESC(shutdown_secs, "Shutdown time (s), <= zero to disable.");
103static int stall_cpu;
104module_param(stall_cpu, int, 0444);
105MODULE_PARM_DESC(stall_cpu, "Stall duration (s), zero to disable.");
106static int stall_cpu_holdoff = 10;
107module_param(stall_cpu_holdoff, int, 0444);
108MODULE_PARM_DESC(stall_cpu_holdoff, "Time to wait before starting stall (s).");
109static int stat_interval = 60;
110module_param(stat_interval, int, 0644);
111MODULE_PARM_DESC(stat_interval, "Number of seconds between stats printk()s");
112static int stutter = 5;
113module_param(stutter, int, 0444);
114MODULE_PARM_DESC(stutter, "Number of seconds to run/halt test");
115static int test_boost = 1;
116module_param(test_boost, int, 0444);
117MODULE_PARM_DESC(test_boost, "Test RCU prio boost: 0=no, 1=maybe, 2=yes.");
118static int test_boost_duration = 4;
119module_param(test_boost_duration, int, 0444);
120MODULE_PARM_DESC(test_boost_duration, "Duration of each boost test, seconds.");
121static int test_boost_interval = 7;
122module_param(test_boost_interval, int, 0444);
123MODULE_PARM_DESC(test_boost_interval, "Interval between boost tests, seconds.");
124static bool test_no_idle_hz = true;
125module_param(test_no_idle_hz, bool, 0444);
126MODULE_PARM_DESC(test_no_idle_hz, "Test support for tickless idle CPUs");
127static char *torture_type = "rcu";
128module_param(torture_type, charp, 0444);
129MODULE_PARM_DESC(torture_type, "Type of RCU to torture (rcu, rcu_bh, ...)");
130static bool verbose;
131module_param(verbose, bool, 0444);
132MODULE_PARM_DESC(verbose, "Enable verbose debugging printk()s");
133
134#define TORTURE_FLAG "-torture:"
135#define PRINTK_STRING(s) \
136 do { pr_alert("%s" TORTURE_FLAG s "\n", torture_type); } while (0)
137#define VERBOSE_PRINTK_STRING(s) \
138 do { if (verbose) pr_alert("%s" TORTURE_FLAG s "\n", torture_type); } while (0)
139#define VERBOSE_PRINTK_ERRSTRING(s) \
140 do { if (verbose) pr_alert("%s" TORTURE_FLAG "!!! " s "\n", torture_type); } while (0)
141
142static char printk_buf[4096];
143
144static int nrealreaders;
145static struct task_struct *writer_task;
146static struct task_struct **fakewriter_tasks;
147static struct task_struct **reader_tasks;
148static struct task_struct *stats_task;
149static struct task_struct *shuffler_task;
150static struct task_struct *stutter_task;
151static struct task_struct *fqs_task;
152static struct task_struct *boost_tasks[NR_CPUS];
153static struct task_struct *shutdown_task;
154#ifdef CONFIG_HOTPLUG_CPU
155static struct task_struct *onoff_task;
156#endif /* #ifdef CONFIG_HOTPLUG_CPU */
157static struct task_struct *stall_task;
158static struct task_struct **barrier_cbs_tasks;
159static struct task_struct *barrier_task;
160
161#define RCU_TORTURE_PIPE_LEN 10
162
163struct rcu_torture {
164 struct rcu_head rtort_rcu;
165 int rtort_pipe_count;
166 struct list_head rtort_free;
167 int rtort_mbtest;
168};
169
170static LIST_HEAD(rcu_torture_freelist);
171static struct rcu_torture __rcu *rcu_torture_current;
172static unsigned long rcu_torture_current_version;
173static struct rcu_torture rcu_tortures[10 * RCU_TORTURE_PIPE_LEN];
174static DEFINE_SPINLOCK(rcu_torture_lock);
175static DEFINE_PER_CPU(long [RCU_TORTURE_PIPE_LEN + 1], rcu_torture_count) =
176 { 0 };
177static DEFINE_PER_CPU(long [RCU_TORTURE_PIPE_LEN + 1], rcu_torture_batch) =
178 { 0 };
179static atomic_t rcu_torture_wcount[RCU_TORTURE_PIPE_LEN + 1];
180static atomic_t n_rcu_torture_alloc;
181static atomic_t n_rcu_torture_alloc_fail;
182static atomic_t n_rcu_torture_free;
183static atomic_t n_rcu_torture_mberror;
184static atomic_t n_rcu_torture_error;
185static long n_rcu_torture_barrier_error;
186static long n_rcu_torture_boost_ktrerror;
187static long n_rcu_torture_boost_rterror;
188static long n_rcu_torture_boost_failure;
189static long n_rcu_torture_boosts;
190static long n_rcu_torture_timers;
191static long n_offline_attempts;
192static long n_offline_successes;
193static unsigned long sum_offline;
194static int min_offline = -1;
195static int max_offline;
196static long n_online_attempts;
197static long n_online_successes;
198static unsigned long sum_online;
199static int min_online = -1;
200static int max_online;
201static long n_barrier_attempts;
202static long n_barrier_successes;
203static struct list_head rcu_torture_removed;
204static cpumask_var_t shuffle_tmp_mask;
205
206static int stutter_pause_test;
207
208#if defined(MODULE) || defined(CONFIG_RCU_TORTURE_TEST_RUNNABLE)
209#define RCUTORTURE_RUNNABLE_INIT 1
210#else
211#define RCUTORTURE_RUNNABLE_INIT 0
212#endif
213int rcutorture_runnable = RCUTORTURE_RUNNABLE_INIT;
214module_param(rcutorture_runnable, int, 0444);
215MODULE_PARM_DESC(rcutorture_runnable, "Start rcutorture at boot");
216
217#if defined(CONFIG_RCU_BOOST) && !defined(CONFIG_HOTPLUG_CPU)
218#define rcu_can_boost() 1
219#else /* #if defined(CONFIG_RCU_BOOST) && !defined(CONFIG_HOTPLUG_CPU) */
220#define rcu_can_boost() 0
221#endif /* #else #if defined(CONFIG_RCU_BOOST) && !defined(CONFIG_HOTPLUG_CPU) */
222
223#ifdef CONFIG_RCU_TRACE
224static u64 notrace rcu_trace_clock_local(void)
225{
226 u64 ts = trace_clock_local();
227 unsigned long __maybe_unused ts_rem = do_div(ts, NSEC_PER_USEC);
228 return ts;
229}
230#else /* #ifdef CONFIG_RCU_TRACE */
231static u64 notrace rcu_trace_clock_local(void)
232{
233 return 0ULL;
234}
235#endif /* #else #ifdef CONFIG_RCU_TRACE */
236
237static unsigned long shutdown_time; /* jiffies to system shutdown. */
238static unsigned long boost_starttime; /* jiffies of next boost test start. */
239DEFINE_MUTEX(boost_mutex); /* protect setting boost_starttime */
240 /* and boost task create/destroy. */
241static atomic_t barrier_cbs_count; /* Barrier callbacks registered. */
242static bool barrier_phase; /* Test phase. */
243static atomic_t barrier_cbs_invoked; /* Barrier callbacks invoked. */
244static wait_queue_head_t *barrier_cbs_wq; /* Coordinate barrier testing. */
245static DECLARE_WAIT_QUEUE_HEAD(barrier_wq);
246
247/* Mediate rmmod and system shutdown. Concurrent rmmod & shutdown illegal! */
248
249#define FULLSTOP_DONTSTOP 0 /* Normal operation. */
250#define FULLSTOP_SHUTDOWN 1 /* System shutdown with rcutorture running. */
251#define FULLSTOP_RMMOD 2 /* Normal rmmod of rcutorture. */
252static int fullstop = FULLSTOP_RMMOD;
253/*
254 * Protect fullstop transitions and spawning of kthreads.
255 */
256static DEFINE_MUTEX(fullstop_mutex);
257
258/* Forward reference. */
259static void rcu_torture_cleanup(void);
260
261/*
262 * Detect and respond to a system shutdown.
263 */
264static int
265rcutorture_shutdown_notify(struct notifier_block *unused1,
266 unsigned long unused2, void *unused3)
267{
268 mutex_lock(&fullstop_mutex);
269 if (fullstop == FULLSTOP_DONTSTOP)
270 fullstop = FULLSTOP_SHUTDOWN;
271 else
272 pr_warn(/* but going down anyway, so... */
273 "Concurrent 'rmmod rcutorture' and shutdown illegal!\n");
274 mutex_unlock(&fullstop_mutex);
275 return NOTIFY_DONE;
276}
277
278/*
279 * Absorb kthreads into a kernel function that won't return, so that
280 * they won't ever access module text or data again.
281 */
282static void rcutorture_shutdown_absorb(const char *title)
283{
284 if (ACCESS_ONCE(fullstop) == FULLSTOP_SHUTDOWN) {
285 pr_notice(
286 "rcutorture thread %s parking due to system shutdown\n",
287 title);
288 schedule_timeout_uninterruptible(MAX_SCHEDULE_TIMEOUT);
289 }
290}
291
292/*
293 * Allocate an element from the rcu_tortures pool.
294 */
295static struct rcu_torture *
296rcu_torture_alloc(void)
297{
298 struct list_head *p;
299
300 spin_lock_bh(&rcu_torture_lock);
301 if (list_empty(&rcu_torture_freelist)) {
302 atomic_inc(&n_rcu_torture_alloc_fail);
303 spin_unlock_bh(&rcu_torture_lock);
304 return NULL;
305 }
306 atomic_inc(&n_rcu_torture_alloc);
307 p = rcu_torture_freelist.next;
308 list_del_init(p);
309 spin_unlock_bh(&rcu_torture_lock);
310 return container_of(p, struct rcu_torture, rtort_free);
311}
312
313/*
314 * Free an element to the rcu_tortures pool.
315 */
316static void
317rcu_torture_free(struct rcu_torture *p)
318{
319 atomic_inc(&n_rcu_torture_free);
320 spin_lock_bh(&rcu_torture_lock);
321 list_add_tail(&p->rtort_free, &rcu_torture_freelist);
322 spin_unlock_bh(&rcu_torture_lock);
323}
324
325struct rcu_random_state {
326 unsigned long rrs_state;
327 long rrs_count;
328};
329
330#define RCU_RANDOM_MULT 39916801 /* prime */
331#define RCU_RANDOM_ADD 479001701 /* prime */
332#define RCU_RANDOM_REFRESH 10000
333
334#define DEFINE_RCU_RANDOM(name) struct rcu_random_state name = { 0, 0 }
335
336/*
337 * Crude but fast random-number generator. Uses a linear congruential
338 * generator, with occasional help from cpu_clock().
339 */
340static unsigned long
341rcu_random(struct rcu_random_state *rrsp)
342{
343 if (--rrsp->rrs_count < 0) {
344 rrsp->rrs_state += (unsigned long)local_clock();
345 rrsp->rrs_count = RCU_RANDOM_REFRESH;
346 }
347 rrsp->rrs_state = rrsp->rrs_state * RCU_RANDOM_MULT + RCU_RANDOM_ADD;
348 return swahw32(rrsp->rrs_state);
349}
350
351static void
352rcu_stutter_wait(const char *title)
353{
354 while (stutter_pause_test || !rcutorture_runnable) {
355 if (rcutorture_runnable)
356 schedule_timeout_interruptible(1);
357 else
358 schedule_timeout_interruptible(round_jiffies_relative(HZ));
359 rcutorture_shutdown_absorb(title);
360 }
361}
362
363/*
364 * Operations vector for selecting different types of tests.
365 */
366
367struct rcu_torture_ops {
368 void (*init)(void);
369 int (*readlock)(void);
370 void (*read_delay)(struct rcu_random_state *rrsp);
371 void (*readunlock)(int idx);
372 int (*completed)(void);
373 void (*deferred_free)(struct rcu_torture *p);
374 void (*sync)(void);
375 void (*exp_sync)(void);
376 void (*call)(struct rcu_head *head, void (*func)(struct rcu_head *rcu));
377 void (*cb_barrier)(void);
378 void (*fqs)(void);
379 int (*stats)(char *page);
380 int irq_capable;
381 int can_boost;
382 const char *name;
383};
384
385static struct rcu_torture_ops *cur_ops;
386
387/*
388 * Definitions for rcu torture testing.
389 */
390
391static int rcu_torture_read_lock(void) __acquires(RCU)
392{
393 rcu_read_lock();
394 return 0;
395}
396
397static void rcu_read_delay(struct rcu_random_state *rrsp)
398{
399 const unsigned long shortdelay_us = 200;
400 const unsigned long longdelay_ms = 50;
401
402 /* We want a short delay sometimes to make a reader delay the grace
403 * period, and we want a long delay occasionally to trigger
404 * force_quiescent_state. */
405
406 if (!(rcu_random(rrsp) % (nrealreaders * 2000 * longdelay_ms)))
407 mdelay(longdelay_ms);
408 if (!(rcu_random(rrsp) % (nrealreaders * 2 * shortdelay_us)))
409 udelay(shortdelay_us);
410#ifdef CONFIG_PREEMPT
411 if (!preempt_count() && !(rcu_random(rrsp) % (nrealreaders * 20000)))
412 preempt_schedule(); /* No QS if preempt_disable() in effect */
413#endif
414}
415
416static void rcu_torture_read_unlock(int idx) __releases(RCU)
417{
418 rcu_read_unlock();
419}
420
421static int rcu_torture_completed(void)
422{
423 return rcu_batches_completed();
424}
425
426static void
427rcu_torture_cb(struct rcu_head *p)
428{
429 int i;
430 struct rcu_torture *rp = container_of(p, struct rcu_torture, rtort_rcu);
431
432 if (fullstop != FULLSTOP_DONTSTOP) {
433 /* Test is ending, just drop callbacks on the floor. */
434 /* The next initialization will pick up the pieces. */
435 return;
436 }
437 i = rp->rtort_pipe_count;
438 if (i > RCU_TORTURE_PIPE_LEN)
439 i = RCU_TORTURE_PIPE_LEN;
440 atomic_inc(&rcu_torture_wcount[i]);
441 if (++rp->rtort_pipe_count >= RCU_TORTURE_PIPE_LEN) {
442 rp->rtort_mbtest = 0;
443 rcu_torture_free(rp);
444 } else {
445 cur_ops->deferred_free(rp);
446 }
447}
448
449static int rcu_no_completed(void)
450{
451 return 0;
452}
453
454static void rcu_torture_deferred_free(struct rcu_torture *p)
455{
456 call_rcu(&p->rtort_rcu, rcu_torture_cb);
457}
458
459static void rcu_sync_torture_init(void)
460{
461 INIT_LIST_HEAD(&rcu_torture_removed);
462}
463
464static struct rcu_torture_ops rcu_ops = {
465 .init = rcu_sync_torture_init,
466 .readlock = rcu_torture_read_lock,
467 .read_delay = rcu_read_delay,
468 .readunlock = rcu_torture_read_unlock,
469 .completed = rcu_torture_completed,
470 .deferred_free = rcu_torture_deferred_free,
471 .sync = synchronize_rcu,
472 .exp_sync = synchronize_rcu_expedited,
473 .call = call_rcu,
474 .cb_barrier = rcu_barrier,
475 .fqs = rcu_force_quiescent_state,
476 .stats = NULL,
477 .irq_capable = 1,
478 .can_boost = rcu_can_boost(),
479 .name = "rcu"
480};
481
482/*
483 * Definitions for rcu_bh torture testing.
484 */
485
486static int rcu_bh_torture_read_lock(void) __acquires(RCU_BH)
487{
488 rcu_read_lock_bh();
489 return 0;
490}
491
492static void rcu_bh_torture_read_unlock(int idx) __releases(RCU_BH)
493{
494 rcu_read_unlock_bh();
495}
496
497static int rcu_bh_torture_completed(void)
498{
499 return rcu_batches_completed_bh();
500}
501
502static void rcu_bh_torture_deferred_free(struct rcu_torture *p)
503{
504 call_rcu_bh(&p->rtort_rcu, rcu_torture_cb);
505}
506
507static struct rcu_torture_ops rcu_bh_ops = {
508 .init = rcu_sync_torture_init,
509 .readlock = rcu_bh_torture_read_lock,
510 .read_delay = rcu_read_delay, /* just reuse rcu's version. */
511 .readunlock = rcu_bh_torture_read_unlock,
512 .completed = rcu_bh_torture_completed,
513 .deferred_free = rcu_bh_torture_deferred_free,
514 .sync = synchronize_rcu_bh,
515 .exp_sync = synchronize_rcu_bh_expedited,
516 .call = call_rcu_bh,
517 .cb_barrier = rcu_barrier_bh,
518 .fqs = rcu_bh_force_quiescent_state,
519 .stats = NULL,
520 .irq_capable = 1,
521 .name = "rcu_bh"
522};
523
524/*
525 * Definitions for srcu torture testing.
526 */
527
528DEFINE_STATIC_SRCU(srcu_ctl);
529
530static int srcu_torture_read_lock(void) __acquires(&srcu_ctl)
531{
532 return srcu_read_lock(&srcu_ctl);
533}
534
535static void srcu_read_delay(struct rcu_random_state *rrsp)
536{
537 long delay;
538 const long uspertick = 1000000 / HZ;
539 const long longdelay = 10;
540
541 /* We want there to be long-running readers, but not all the time. */
542
543 delay = rcu_random(rrsp) % (nrealreaders * 2 * longdelay * uspertick);
544 if (!delay)
545 schedule_timeout_interruptible(longdelay);
546 else
547 rcu_read_delay(rrsp);
548}
549
550static void srcu_torture_read_unlock(int idx) __releases(&srcu_ctl)
551{
552 srcu_read_unlock(&srcu_ctl, idx);
553}
554
555static int srcu_torture_completed(void)
556{
557 return srcu_batches_completed(&srcu_ctl);
558}
559
560static void srcu_torture_deferred_free(struct rcu_torture *rp)
561{
562 call_srcu(&srcu_ctl, &rp->rtort_rcu, rcu_torture_cb);
563}
564
565static void srcu_torture_synchronize(void)
566{
567 synchronize_srcu(&srcu_ctl);
568}
569
570static void srcu_torture_call(struct rcu_head *head,
571 void (*func)(struct rcu_head *head))
572{
573 call_srcu(&srcu_ctl, head, func);
574}
575
576static void srcu_torture_barrier(void)
577{
578 srcu_barrier(&srcu_ctl);
579}
580
581static int srcu_torture_stats(char *page)
582{
583 int cnt = 0;
584 int cpu;
585 int idx = srcu_ctl.completed & 0x1;
586
587 cnt += sprintf(&page[cnt], "%s%s per-CPU(idx=%d):",
588 torture_type, TORTURE_FLAG, idx);
589 for_each_possible_cpu(cpu) {
590 cnt += sprintf(&page[cnt], " %d(%lu,%lu)", cpu,
591 per_cpu_ptr(srcu_ctl.per_cpu_ref, cpu)->c[!idx],
592 per_cpu_ptr(srcu_ctl.per_cpu_ref, cpu)->c[idx]);
593 }
594 cnt += sprintf(&page[cnt], "\n");
595 return cnt;
596}
597
598static void srcu_torture_synchronize_expedited(void)
599{
600 synchronize_srcu_expedited(&srcu_ctl);
601}
602
603static struct rcu_torture_ops srcu_ops = {
604 .init = rcu_sync_torture_init,
605 .readlock = srcu_torture_read_lock,
606 .read_delay = srcu_read_delay,
607 .readunlock = srcu_torture_read_unlock,
608 .completed = srcu_torture_completed,
609 .deferred_free = srcu_torture_deferred_free,
610 .sync = srcu_torture_synchronize,
611 .exp_sync = srcu_torture_synchronize_expedited,
612 .call = srcu_torture_call,
613 .cb_barrier = srcu_torture_barrier,
614 .stats = srcu_torture_stats,
615 .name = "srcu"
616};
617
618/*
619 * Definitions for sched torture testing.
620 */
621
622static int sched_torture_read_lock(void)
623{
624 preempt_disable();
625 return 0;
626}
627
628static void sched_torture_read_unlock(int idx)
629{
630 preempt_enable();
631}
632
633static void rcu_sched_torture_deferred_free(struct rcu_torture *p)
634{
635 call_rcu_sched(&p->rtort_rcu, rcu_torture_cb);
636}
637
638static struct rcu_torture_ops sched_ops = {
639 .init = rcu_sync_torture_init,
640 .readlock = sched_torture_read_lock,
641 .read_delay = rcu_read_delay, /* just reuse rcu's version. */
642 .readunlock = sched_torture_read_unlock,
643 .completed = rcu_no_completed,
644 .deferred_free = rcu_sched_torture_deferred_free,
645 .sync = synchronize_sched,
646 .exp_sync = synchronize_sched_expedited,
647 .call = call_rcu_sched,
648 .cb_barrier = rcu_barrier_sched,
649 .fqs = rcu_sched_force_quiescent_state,
650 .stats = NULL,
651 .irq_capable = 1,
652 .name = "sched"
653};
654
655/*
656 * RCU torture priority-boost testing. Runs one real-time thread per
657 * CPU for moderate bursts, repeatedly registering RCU callbacks and
658 * spinning waiting for them to be invoked. If a given callback takes
659 * too long to be invoked, we assume that priority inversion has occurred.
660 */
661
662struct rcu_boost_inflight {
663 struct rcu_head rcu;
664 int inflight;
665};
666
667static void rcu_torture_boost_cb(struct rcu_head *head)
668{
669 struct rcu_boost_inflight *rbip =
670 container_of(head, struct rcu_boost_inflight, rcu);
671
672 smp_mb(); /* Ensure RCU-core accesses precede clearing ->inflight */
673 rbip->inflight = 0;
674}
675
676static int rcu_torture_boost(void *arg)
677{
678 unsigned long call_rcu_time;
679 unsigned long endtime;
680 unsigned long oldstarttime;
681 struct rcu_boost_inflight rbi = { .inflight = 0 };
682 struct sched_param sp;
683
684 VERBOSE_PRINTK_STRING("rcu_torture_boost started");
685
686 /* Set real-time priority. */
687 sp.sched_priority = 1;
688 if (sched_setscheduler(current, SCHED_FIFO, &sp) < 0) {
689 VERBOSE_PRINTK_STRING("rcu_torture_boost RT prio failed!");
690 n_rcu_torture_boost_rterror++;
691 }
692
693 init_rcu_head_on_stack(&rbi.rcu);
694 /* Each pass through the following loop does one boost-test cycle. */
695 do {
696 /* Wait for the next test interval. */
697 oldstarttime = boost_starttime;
698 while (ULONG_CMP_LT(jiffies, oldstarttime)) {
699 schedule_timeout_interruptible(oldstarttime - jiffies);
700 rcu_stutter_wait("rcu_torture_boost");
701 if (kthread_should_stop() ||
702 fullstop != FULLSTOP_DONTSTOP)
703 goto checkwait;
704 }
705
706 /* Do one boost-test interval. */
707 endtime = oldstarttime + test_boost_duration * HZ;
708 call_rcu_time = jiffies;
709 while (ULONG_CMP_LT(jiffies, endtime)) {
710 /* If we don't have a callback in flight, post one. */
711 if (!rbi.inflight) {
712 smp_mb(); /* RCU core before ->inflight = 1. */
713 rbi.inflight = 1;
714 call_rcu(&rbi.rcu, rcu_torture_boost_cb);
715 if (jiffies - call_rcu_time >
716 test_boost_duration * HZ - HZ / 2) {
717 VERBOSE_PRINTK_STRING("rcu_torture_boost boosting failed");
718 n_rcu_torture_boost_failure++;
719 }
720 call_rcu_time = jiffies;
721 }
722 cond_resched();
723 rcu_stutter_wait("rcu_torture_boost");
724 if (kthread_should_stop() ||
725 fullstop != FULLSTOP_DONTSTOP)
726 goto checkwait;
727 }
728
729 /*
730 * Set the start time of the next test interval.
731 * Yes, this is vulnerable to long delays, but such
732 * delays simply cause a false negative for the next
733 * interval. Besides, we are running at RT priority,
734 * so delays should be relatively rare.
735 */
736 while (oldstarttime == boost_starttime &&
737 !kthread_should_stop()) {
738 if (mutex_trylock(&boost_mutex)) {
739 boost_starttime = jiffies +
740 test_boost_interval * HZ;
741 n_rcu_torture_boosts++;
742 mutex_unlock(&boost_mutex);
743 break;
744 }
745 schedule_timeout_uninterruptible(1);
746 }
747
748 /* Go do the stutter. */
749checkwait: rcu_stutter_wait("rcu_torture_boost");
750 } while (!kthread_should_stop() && fullstop == FULLSTOP_DONTSTOP);
751
752 /* Clean up and exit. */
753 VERBOSE_PRINTK_STRING("rcu_torture_boost task stopping");
754 rcutorture_shutdown_absorb("rcu_torture_boost");
755 while (!kthread_should_stop() || rbi.inflight)
756 schedule_timeout_uninterruptible(1);
757 smp_mb(); /* order accesses to ->inflight before stack-frame death. */
758 destroy_rcu_head_on_stack(&rbi.rcu);
759 return 0;
760}
761
762/*
763 * RCU torture force-quiescent-state kthread. Repeatedly induces
764 * bursts of calls to force_quiescent_state(), increasing the probability
765 * of occurrence of some important types of race conditions.
766 */
767static int
768rcu_torture_fqs(void *arg)
769{
770 unsigned long fqs_resume_time;
771 int fqs_burst_remaining;
772
773 VERBOSE_PRINTK_STRING("rcu_torture_fqs task started");
774 do {
775 fqs_resume_time = jiffies + fqs_stutter * HZ;
776 while (ULONG_CMP_LT(jiffies, fqs_resume_time) &&
777 !kthread_should_stop()) {
778 schedule_timeout_interruptible(1);
779 }
780 fqs_burst_remaining = fqs_duration;
781 while (fqs_burst_remaining > 0 &&
782 !kthread_should_stop()) {
783 cur_ops->fqs();
784 udelay(fqs_holdoff);
785 fqs_burst_remaining -= fqs_holdoff;
786 }
787 rcu_stutter_wait("rcu_torture_fqs");
788 } while (!kthread_should_stop() && fullstop == FULLSTOP_DONTSTOP);
789 VERBOSE_PRINTK_STRING("rcu_torture_fqs task stopping");
790 rcutorture_shutdown_absorb("rcu_torture_fqs");
791 while (!kthread_should_stop())
792 schedule_timeout_uninterruptible(1);
793 return 0;
794}
795
796/*
797 * RCU torture writer kthread. Repeatedly substitutes a new structure
798 * for that pointed to by rcu_torture_current, freeing the old structure
799 * after a series of grace periods (the "pipeline").
800 */
801static int
802rcu_torture_writer(void *arg)
803{
804 bool exp;
805 int i;
806 struct rcu_torture *rp;
807 struct rcu_torture *rp1;
808 struct rcu_torture *old_rp;
809 static DEFINE_RCU_RANDOM(rand);
810
811 VERBOSE_PRINTK_STRING("rcu_torture_writer task started");
812 set_user_nice(current, 19);
813
814 do {
815 schedule_timeout_uninterruptible(1);
816 rp = rcu_torture_alloc();
817 if (rp == NULL)
818 continue;
819 rp->rtort_pipe_count = 0;
820 udelay(rcu_random(&rand) & 0x3ff);
821 old_rp = rcu_dereference_check(rcu_torture_current,
822 current == writer_task);
823 rp->rtort_mbtest = 1;
824 rcu_assign_pointer(rcu_torture_current, rp);
825 smp_wmb(); /* Mods to old_rp must follow rcu_assign_pointer() */
826 if (old_rp) {
827 i = old_rp->rtort_pipe_count;
828 if (i > RCU_TORTURE_PIPE_LEN)
829 i = RCU_TORTURE_PIPE_LEN;
830 atomic_inc(&rcu_torture_wcount[i]);
831 old_rp->rtort_pipe_count++;
832 if (gp_normal == gp_exp)
833 exp = !!(rcu_random(&rand) & 0x80);
834 else
835 exp = gp_exp;
836 if (!exp) {
837 cur_ops->deferred_free(old_rp);
838 } else {
839 cur_ops->exp_sync();
840 list_add(&old_rp->rtort_free,
841 &rcu_torture_removed);
842 list_for_each_entry_safe(rp, rp1,
843 &rcu_torture_removed,
844 rtort_free) {
845 i = rp->rtort_pipe_count;
846 if (i > RCU_TORTURE_PIPE_LEN)
847 i = RCU_TORTURE_PIPE_LEN;
848 atomic_inc(&rcu_torture_wcount[i]);
849 if (++rp->rtort_pipe_count >=
850 RCU_TORTURE_PIPE_LEN) {
851 rp->rtort_mbtest = 0;
852 list_del(&rp->rtort_free);
853 rcu_torture_free(rp);
854 }
855 }
856 }
857 }
858 rcutorture_record_progress(++rcu_torture_current_version);
859 rcu_stutter_wait("rcu_torture_writer");
860 } while (!kthread_should_stop() && fullstop == FULLSTOP_DONTSTOP);
861 VERBOSE_PRINTK_STRING("rcu_torture_writer task stopping");
862 rcutorture_shutdown_absorb("rcu_torture_writer");
863 while (!kthread_should_stop())
864 schedule_timeout_uninterruptible(1);
865 return 0;
866}
867
868/*
869 * RCU torture fake writer kthread. Repeatedly calls sync, with a random
870 * delay between calls.
871 */
872static int
873rcu_torture_fakewriter(void *arg)
874{
875 DEFINE_RCU_RANDOM(rand);
876
877 VERBOSE_PRINTK_STRING("rcu_torture_fakewriter task started");
878 set_user_nice(current, 19);
879
880 do {
881 schedule_timeout_uninterruptible(1 + rcu_random(&rand)%10);
882 udelay(rcu_random(&rand) & 0x3ff);
883 if (cur_ops->cb_barrier != NULL &&
884 rcu_random(&rand) % (nfakewriters * 8) == 0) {
885 cur_ops->cb_barrier();
886 } else if (gp_normal == gp_exp) {
887 if (rcu_random(&rand) & 0x80)
888 cur_ops->sync();
889 else
890 cur_ops->exp_sync();
891 } else if (gp_normal) {
892 cur_ops->sync();
893 } else {
894 cur_ops->exp_sync();
895 }
896 rcu_stutter_wait("rcu_torture_fakewriter");
897 } while (!kthread_should_stop() && fullstop == FULLSTOP_DONTSTOP);
898
899 VERBOSE_PRINTK_STRING("rcu_torture_fakewriter task stopping");
900 rcutorture_shutdown_absorb("rcu_torture_fakewriter");
901 while (!kthread_should_stop())
902 schedule_timeout_uninterruptible(1);
903 return 0;
904}
905
906void rcutorture_trace_dump(void)
907{
908 static atomic_t beenhere = ATOMIC_INIT(0);
909
910 if (atomic_read(&beenhere))
911 return;
912 if (atomic_xchg(&beenhere, 1) != 0)
913 return;
914 ftrace_dump(DUMP_ALL);
915}
916
917/*
918 * RCU torture reader from timer handler. Dereferences rcu_torture_current,
919 * incrementing the corresponding element of the pipeline array. The
920 * counter in the element should never be greater than 1, otherwise, the
921 * RCU implementation is broken.
922 */
923static void rcu_torture_timer(unsigned long unused)
924{
925 int idx;
926 int completed;
927 int completed_end;
928 static DEFINE_RCU_RANDOM(rand);
929 static DEFINE_SPINLOCK(rand_lock);
930 struct rcu_torture *p;
931 int pipe_count;
932 unsigned long long ts;
933
934 idx = cur_ops->readlock();
935 completed = cur_ops->completed();
936 ts = rcu_trace_clock_local();
937 p = rcu_dereference_check(rcu_torture_current,
938 rcu_read_lock_bh_held() ||
939 rcu_read_lock_sched_held() ||
940 srcu_read_lock_held(&srcu_ctl));
941 if (p == NULL) {
942 /* Leave because rcu_torture_writer is not yet underway */
943 cur_ops->readunlock(idx);
944 return;
945 }
946 if (p->rtort_mbtest == 0)
947 atomic_inc(&n_rcu_torture_mberror);
948 spin_lock(&rand_lock);
949 cur_ops->read_delay(&rand);
950 n_rcu_torture_timers++;
951 spin_unlock(&rand_lock);
952 preempt_disable();
953 pipe_count = p->rtort_pipe_count;
954 if (pipe_count > RCU_TORTURE_PIPE_LEN) {
955 /* Should not happen, but... */
956 pipe_count = RCU_TORTURE_PIPE_LEN;
957 }
958 completed_end = cur_ops->completed();
959 if (pipe_count > 1) {
960 do_trace_rcu_torture_read(cur_ops->name, &p->rtort_rcu, ts,
961 completed, completed_end);
962 rcutorture_trace_dump();
963 }
964 __this_cpu_inc(rcu_torture_count[pipe_count]);
965 completed = completed_end - completed;
966 if (completed > RCU_TORTURE_PIPE_LEN) {
967 /* Should not happen, but... */
968 completed = RCU_TORTURE_PIPE_LEN;
969 }
970 __this_cpu_inc(rcu_torture_batch[completed]);
971 preempt_enable();
972 cur_ops->readunlock(idx);
973}
974
975/*
976 * RCU torture reader kthread. Repeatedly dereferences rcu_torture_current,
977 * incrementing the corresponding element of the pipeline array. The
978 * counter in the element should never be greater than 1, otherwise, the
979 * RCU implementation is broken.
980 */
981static int
982rcu_torture_reader(void *arg)
983{
984 int completed;
985 int completed_end;
986 int idx;
987 DEFINE_RCU_RANDOM(rand);
988 struct rcu_torture *p;
989 int pipe_count;
990 struct timer_list t;
991 unsigned long long ts;
992
993 VERBOSE_PRINTK_STRING("rcu_torture_reader task started");
994 set_user_nice(current, 19);
995 if (irqreader && cur_ops->irq_capable)
996 setup_timer_on_stack(&t, rcu_torture_timer, 0);
997
998 do {
999 if (irqreader && cur_ops->irq_capable) {
1000 if (!timer_pending(&t))
1001 mod_timer(&t, jiffies + 1);
1002 }
1003 idx = cur_ops->readlock();
1004 completed = cur_ops->completed();
1005 ts = rcu_trace_clock_local();
1006 p = rcu_dereference_check(rcu_torture_current,
1007 rcu_read_lock_bh_held() ||
1008 rcu_read_lock_sched_held() ||
1009 srcu_read_lock_held(&srcu_ctl));
1010 if (p == NULL) {
1011 /* Wait for rcu_torture_writer to get underway */
1012 cur_ops->readunlock(idx);
1013 schedule_timeout_interruptible(HZ);
1014 continue;
1015 }
1016 if (p->rtort_mbtest == 0)
1017 atomic_inc(&n_rcu_torture_mberror);
1018 cur_ops->read_delay(&rand);
1019 preempt_disable();
1020 pipe_count = p->rtort_pipe_count;
1021 if (pipe_count > RCU_TORTURE_PIPE_LEN) {
1022 /* Should not happen, but... */
1023 pipe_count = RCU_TORTURE_PIPE_LEN;
1024 }
1025 completed_end = cur_ops->completed();
1026 if (pipe_count > 1) {
1027 do_trace_rcu_torture_read(cur_ops->name, &p->rtort_rcu,
1028 ts, completed, completed_end);
1029 rcutorture_trace_dump();
1030 }
1031 __this_cpu_inc(rcu_torture_count[pipe_count]);
1032 completed = completed_end - completed;
1033 if (completed > RCU_TORTURE_PIPE_LEN) {
1034 /* Should not happen, but... */
1035 completed = RCU_TORTURE_PIPE_LEN;
1036 }
1037 __this_cpu_inc(rcu_torture_batch[completed]);
1038 preempt_enable();
1039 cur_ops->readunlock(idx);
1040 schedule();
1041 rcu_stutter_wait("rcu_torture_reader");
1042 } while (!kthread_should_stop() && fullstop == FULLSTOP_DONTSTOP);
1043 VERBOSE_PRINTK_STRING("rcu_torture_reader task stopping");
1044 rcutorture_shutdown_absorb("rcu_torture_reader");
1045 if (irqreader && cur_ops->irq_capable)
1046 del_timer_sync(&t);
1047 while (!kthread_should_stop())
1048 schedule_timeout_uninterruptible(1);
1049 return 0;
1050}
1051
1052/*
1053 * Create an RCU-torture statistics message in the specified buffer.
1054 */
1055static int
1056rcu_torture_printk(char *page)
1057{
1058 int cnt = 0;
1059 int cpu;
1060 int i;
1061 long pipesummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
1062 long batchsummary[RCU_TORTURE_PIPE_LEN + 1] = { 0 };
1063
1064 for_each_possible_cpu(cpu) {
1065 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
1066 pipesummary[i] += per_cpu(rcu_torture_count, cpu)[i];
1067 batchsummary[i] += per_cpu(rcu_torture_batch, cpu)[i];
1068 }
1069 }
1070 for (i = RCU_TORTURE_PIPE_LEN - 1; i >= 0; i--) {
1071 if (pipesummary[i] != 0)
1072 break;
1073 }
1074 cnt += sprintf(&page[cnt], "%s%s ", torture_type, TORTURE_FLAG);
1075 cnt += sprintf(&page[cnt],
1076 "rtc: %p ver: %lu tfle: %d rta: %d rtaf: %d rtf: %d ",
1077 rcu_torture_current,
1078 rcu_torture_current_version,
1079 list_empty(&rcu_torture_freelist),
1080 atomic_read(&n_rcu_torture_alloc),
1081 atomic_read(&n_rcu_torture_alloc_fail),
1082 atomic_read(&n_rcu_torture_free));
1083 cnt += sprintf(&page[cnt], "rtmbe: %d rtbke: %ld rtbre: %ld ",
1084 atomic_read(&n_rcu_torture_mberror),
1085 n_rcu_torture_boost_ktrerror,
1086 n_rcu_torture_boost_rterror);
1087 cnt += sprintf(&page[cnt], "rtbf: %ld rtb: %ld nt: %ld ",
1088 n_rcu_torture_boost_failure,
1089 n_rcu_torture_boosts,
1090 n_rcu_torture_timers);
1091 cnt += sprintf(&page[cnt],
1092 "onoff: %ld/%ld:%ld/%ld %d,%d:%d,%d %lu:%lu (HZ=%d) ",
1093 n_online_successes, n_online_attempts,
1094 n_offline_successes, n_offline_attempts,
1095 min_online, max_online,
1096 min_offline, max_offline,
1097 sum_online, sum_offline, HZ);
1098 cnt += sprintf(&page[cnt], "barrier: %ld/%ld:%ld",
1099 n_barrier_successes,
1100 n_barrier_attempts,
1101 n_rcu_torture_barrier_error);
1102 cnt += sprintf(&page[cnt], "\n%s%s ", torture_type, TORTURE_FLAG);
1103 if (atomic_read(&n_rcu_torture_mberror) != 0 ||
1104 n_rcu_torture_barrier_error != 0 ||
1105 n_rcu_torture_boost_ktrerror != 0 ||
1106 n_rcu_torture_boost_rterror != 0 ||
1107 n_rcu_torture_boost_failure != 0 ||
1108 i > 1) {
1109 cnt += sprintf(&page[cnt], "!!! ");
1110 atomic_inc(&n_rcu_torture_error);
1111 WARN_ON_ONCE(1);
1112 }
1113 cnt += sprintf(&page[cnt], "Reader Pipe: ");
1114 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
1115 cnt += sprintf(&page[cnt], " %ld", pipesummary[i]);
1116 cnt += sprintf(&page[cnt], "\n%s%s ", torture_type, TORTURE_FLAG);
1117 cnt += sprintf(&page[cnt], "Reader Batch: ");
1118 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
1119 cnt += sprintf(&page[cnt], " %ld", batchsummary[i]);
1120 cnt += sprintf(&page[cnt], "\n%s%s ", torture_type, TORTURE_FLAG);
1121 cnt += sprintf(&page[cnt], "Free-Block Circulation: ");
1122 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
1123 cnt += sprintf(&page[cnt], " %d",
1124 atomic_read(&rcu_torture_wcount[i]));
1125 }
1126 cnt += sprintf(&page[cnt], "\n");
1127 if (cur_ops->stats)
1128 cnt += cur_ops->stats(&page[cnt]);
1129 return cnt;
1130}
1131
1132/*
1133 * Print torture statistics. Caller must ensure that there is only
1134 * one call to this function at a given time!!! This is normally
1135 * accomplished by relying on the module system to only have one copy
1136 * of the module loaded, and then by giving the rcu_torture_stats
1137 * kthread full control (or the init/cleanup functions when rcu_torture_stats
1138 * thread is not running).
1139 */
1140static void
1141rcu_torture_stats_print(void)
1142{
1143 int cnt;
1144
1145 cnt = rcu_torture_printk(printk_buf);
1146 pr_alert("%s", printk_buf);
1147}
1148
1149/*
1150 * Periodically prints torture statistics, if periodic statistics printing
1151 * was specified via the stat_interval module parameter.
1152 *
1153 * No need to worry about fullstop here, since this one doesn't reference
1154 * volatile state or register callbacks.
1155 */
1156static int
1157rcu_torture_stats(void *arg)
1158{
1159 VERBOSE_PRINTK_STRING("rcu_torture_stats task started");
1160 do {
1161 schedule_timeout_interruptible(stat_interval * HZ);
1162 rcu_torture_stats_print();
1163 rcutorture_shutdown_absorb("rcu_torture_stats");
1164 } while (!kthread_should_stop());
1165 VERBOSE_PRINTK_STRING("rcu_torture_stats task stopping");
1166 return 0;
1167}
1168
1169static int rcu_idle_cpu; /* Force all torture tasks off this CPU */
1170
1171/* Shuffle tasks such that we allow @rcu_idle_cpu to become idle. A special case
1172 * is when @rcu_idle_cpu = -1, when we allow the tasks to run on all CPUs.
1173 */
1174static void rcu_torture_shuffle_tasks(void)
1175{
1176 int i;
1177
1178 cpumask_setall(shuffle_tmp_mask);
1179 get_online_cpus();
1180
1181 /* No point in shuffling if there is only one online CPU (ex: UP) */
1182 if (num_online_cpus() == 1) {
1183 put_online_cpus();
1184 return;
1185 }
1186
1187 if (rcu_idle_cpu != -1)
1188 cpumask_clear_cpu(rcu_idle_cpu, shuffle_tmp_mask);
1189
1190 set_cpus_allowed_ptr(current, shuffle_tmp_mask);
1191
1192 if (reader_tasks) {
1193 for (i = 0; i < nrealreaders; i++)
1194 if (reader_tasks[i])
1195 set_cpus_allowed_ptr(reader_tasks[i],
1196 shuffle_tmp_mask);
1197 }
1198 if (fakewriter_tasks) {
1199 for (i = 0; i < nfakewriters; i++)
1200 if (fakewriter_tasks[i])
1201 set_cpus_allowed_ptr(fakewriter_tasks[i],
1202 shuffle_tmp_mask);
1203 }
1204 if (writer_task)
1205 set_cpus_allowed_ptr(writer_task, shuffle_tmp_mask);
1206 if (stats_task)
1207 set_cpus_allowed_ptr(stats_task, shuffle_tmp_mask);
1208 if (stutter_task)
1209 set_cpus_allowed_ptr(stutter_task, shuffle_tmp_mask);
1210 if (fqs_task)
1211 set_cpus_allowed_ptr(fqs_task, shuffle_tmp_mask);
1212 if (shutdown_task)
1213 set_cpus_allowed_ptr(shutdown_task, shuffle_tmp_mask);
1214#ifdef CONFIG_HOTPLUG_CPU
1215 if (onoff_task)
1216 set_cpus_allowed_ptr(onoff_task, shuffle_tmp_mask);
1217#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1218 if (stall_task)
1219 set_cpus_allowed_ptr(stall_task, shuffle_tmp_mask);
1220 if (barrier_cbs_tasks)
1221 for (i = 0; i < n_barrier_cbs; i++)
1222 if (barrier_cbs_tasks[i])
1223 set_cpus_allowed_ptr(barrier_cbs_tasks[i],
1224 shuffle_tmp_mask);
1225 if (barrier_task)
1226 set_cpus_allowed_ptr(barrier_task, shuffle_tmp_mask);
1227
1228 if (rcu_idle_cpu == -1)
1229 rcu_idle_cpu = num_online_cpus() - 1;
1230 else
1231 rcu_idle_cpu--;
1232
1233 put_online_cpus();
1234}
1235
1236/* Shuffle tasks across CPUs, with the intent of allowing each CPU in the
1237 * system to become idle at a time and cut off its timer ticks. This is meant
1238 * to test the support for such tickless idle CPU in RCU.
1239 */
1240static int
1241rcu_torture_shuffle(void *arg)
1242{
1243 VERBOSE_PRINTK_STRING("rcu_torture_shuffle task started");
1244 do {
1245 schedule_timeout_interruptible(shuffle_interval * HZ);
1246 rcu_torture_shuffle_tasks();
1247 rcutorture_shutdown_absorb("rcu_torture_shuffle");
1248 } while (!kthread_should_stop());
1249 VERBOSE_PRINTK_STRING("rcu_torture_shuffle task stopping");
1250 return 0;
1251}
1252
1253/* Cause the rcutorture test to "stutter", starting and stopping all
1254 * threads periodically.
1255 */
1256static int
1257rcu_torture_stutter(void *arg)
1258{
1259 VERBOSE_PRINTK_STRING("rcu_torture_stutter task started");
1260 do {
1261 schedule_timeout_interruptible(stutter * HZ);
1262 stutter_pause_test = 1;
1263 if (!kthread_should_stop())
1264 schedule_timeout_interruptible(stutter * HZ);
1265 stutter_pause_test = 0;
1266 rcutorture_shutdown_absorb("rcu_torture_stutter");
1267 } while (!kthread_should_stop());
1268 VERBOSE_PRINTK_STRING("rcu_torture_stutter task stopping");
1269 return 0;
1270}
1271
1272static inline void
1273rcu_torture_print_module_parms(struct rcu_torture_ops *cur_ops, const char *tag)
1274{
1275 pr_alert("%s" TORTURE_FLAG
1276 "--- %s: nreaders=%d nfakewriters=%d "
1277 "stat_interval=%d verbose=%d test_no_idle_hz=%d "
1278 "shuffle_interval=%d stutter=%d irqreader=%d "
1279 "fqs_duration=%d fqs_holdoff=%d fqs_stutter=%d "
1280 "test_boost=%d/%d test_boost_interval=%d "
1281 "test_boost_duration=%d shutdown_secs=%d "
1282 "stall_cpu=%d stall_cpu_holdoff=%d "
1283 "n_barrier_cbs=%d "
1284 "onoff_interval=%d onoff_holdoff=%d\n",
1285 torture_type, tag, nrealreaders, nfakewriters,
1286 stat_interval, verbose, test_no_idle_hz, shuffle_interval,
1287 stutter, irqreader, fqs_duration, fqs_holdoff, fqs_stutter,
1288 test_boost, cur_ops->can_boost,
1289 test_boost_interval, test_boost_duration, shutdown_secs,
1290 stall_cpu, stall_cpu_holdoff,
1291 n_barrier_cbs,
1292 onoff_interval, onoff_holdoff);
1293}
1294
1295static struct notifier_block rcutorture_shutdown_nb = {
1296 .notifier_call = rcutorture_shutdown_notify,
1297};
1298
1299static void rcutorture_booster_cleanup(int cpu)
1300{
1301 struct task_struct *t;
1302
1303 if (boost_tasks[cpu] == NULL)
1304 return;
1305 mutex_lock(&boost_mutex);
1306 VERBOSE_PRINTK_STRING("Stopping rcu_torture_boost task");
1307 t = boost_tasks[cpu];
1308 boost_tasks[cpu] = NULL;
1309 mutex_unlock(&boost_mutex);
1310
1311 /* This must be outside of the mutex, otherwise deadlock! */
1312 kthread_stop(t);
1313 boost_tasks[cpu] = NULL;
1314}
1315
1316static int rcutorture_booster_init(int cpu)
1317{
1318 int retval;
1319
1320 if (boost_tasks[cpu] != NULL)
1321 return 0; /* Already created, nothing more to do. */
1322
1323 /* Don't allow time recalculation while creating a new task. */
1324 mutex_lock(&boost_mutex);
1325 VERBOSE_PRINTK_STRING("Creating rcu_torture_boost task");
1326 boost_tasks[cpu] = kthread_create_on_node(rcu_torture_boost, NULL,
1327 cpu_to_node(cpu),
1328 "rcu_torture_boost");
1329 if (IS_ERR(boost_tasks[cpu])) {
1330 retval = PTR_ERR(boost_tasks[cpu]);
1331 VERBOSE_PRINTK_STRING("rcu_torture_boost task create failed");
1332 n_rcu_torture_boost_ktrerror++;
1333 boost_tasks[cpu] = NULL;
1334 mutex_unlock(&boost_mutex);
1335 return retval;
1336 }
1337 kthread_bind(boost_tasks[cpu], cpu);
1338 wake_up_process(boost_tasks[cpu]);
1339 mutex_unlock(&boost_mutex);
1340 return 0;
1341}
1342
1343/*
1344 * Cause the rcutorture test to shutdown the system after the test has
1345 * run for the time specified by the shutdown_secs module parameter.
1346 */
1347static int
1348rcu_torture_shutdown(void *arg)
1349{
1350 long delta;
1351 unsigned long jiffies_snap;
1352
1353 VERBOSE_PRINTK_STRING("rcu_torture_shutdown task started");
1354 jiffies_snap = ACCESS_ONCE(jiffies);
1355 while (ULONG_CMP_LT(jiffies_snap, shutdown_time) &&
1356 !kthread_should_stop()) {
1357 delta = shutdown_time - jiffies_snap;
1358 if (verbose)
1359 pr_alert("%s" TORTURE_FLAG
1360 "rcu_torture_shutdown task: %lu jiffies remaining\n",
1361 torture_type, delta);
1362 schedule_timeout_interruptible(delta);
1363 jiffies_snap = ACCESS_ONCE(jiffies);
1364 }
1365 if (kthread_should_stop()) {
1366 VERBOSE_PRINTK_STRING("rcu_torture_shutdown task stopping");
1367 return 0;
1368 }
1369
1370 /* OK, shut down the system. */
1371
1372 VERBOSE_PRINTK_STRING("rcu_torture_shutdown task shutting down system");
1373 shutdown_task = NULL; /* Avoid self-kill deadlock. */
1374 rcu_torture_cleanup(); /* Get the success/failure message. */
1375 kernel_power_off(); /* Shut down the system. */
1376 return 0;
1377}
1378
1379#ifdef CONFIG_HOTPLUG_CPU
1380
1381/*
1382 * Execute random CPU-hotplug operations at the interval specified
1383 * by the onoff_interval.
1384 */
1385static int
1386rcu_torture_onoff(void *arg)
1387{
1388 int cpu;
1389 unsigned long delta;
1390 int maxcpu = -1;
1391 DEFINE_RCU_RANDOM(rand);
1392 int ret;
1393 unsigned long starttime;
1394
1395 VERBOSE_PRINTK_STRING("rcu_torture_onoff task started");
1396 for_each_online_cpu(cpu)
1397 maxcpu = cpu;
1398 WARN_ON(maxcpu < 0);
1399 if (onoff_holdoff > 0) {
1400 VERBOSE_PRINTK_STRING("rcu_torture_onoff begin holdoff");
1401 schedule_timeout_interruptible(onoff_holdoff * HZ);
1402 VERBOSE_PRINTK_STRING("rcu_torture_onoff end holdoff");
1403 }
1404 while (!kthread_should_stop()) {
1405 cpu = (rcu_random(&rand) >> 4) % (maxcpu + 1);
1406 if (cpu_online(cpu) && cpu_is_hotpluggable(cpu)) {
1407 if (verbose)
1408 pr_alert("%s" TORTURE_FLAG
1409 "rcu_torture_onoff task: offlining %d\n",
1410 torture_type, cpu);
1411 starttime = jiffies;
1412 n_offline_attempts++;
1413 ret = cpu_down(cpu);
1414 if (ret) {
1415 if (verbose)
1416 pr_alert("%s" TORTURE_FLAG
1417 "rcu_torture_onoff task: offline %d failed: errno %d\n",
1418 torture_type, cpu, ret);
1419 } else {
1420 if (verbose)
1421 pr_alert("%s" TORTURE_FLAG
1422 "rcu_torture_onoff task: offlined %d\n",
1423 torture_type, cpu);
1424 n_offline_successes++;
1425 delta = jiffies - starttime;
1426 sum_offline += delta;
1427 if (min_offline < 0) {
1428 min_offline = delta;
1429 max_offline = delta;
1430 }
1431 if (min_offline > delta)
1432 min_offline = delta;
1433 if (max_offline < delta)
1434 max_offline = delta;
1435 }
1436 } else if (cpu_is_hotpluggable(cpu)) {
1437 if (verbose)
1438 pr_alert("%s" TORTURE_FLAG
1439 "rcu_torture_onoff task: onlining %d\n",
1440 torture_type, cpu);
1441 starttime = jiffies;
1442 n_online_attempts++;
1443 ret = cpu_up(cpu);
1444 if (ret) {
1445 if (verbose)
1446 pr_alert("%s" TORTURE_FLAG
1447 "rcu_torture_onoff task: online %d failed: errno %d\n",
1448 torture_type, cpu, ret);
1449 } else {
1450 if (verbose)
1451 pr_alert("%s" TORTURE_FLAG
1452 "rcu_torture_onoff task: onlined %d\n",
1453 torture_type, cpu);
1454 n_online_successes++;
1455 delta = jiffies - starttime;
1456 sum_online += delta;
1457 if (min_online < 0) {
1458 min_online = delta;
1459 max_online = delta;
1460 }
1461 if (min_online > delta)
1462 min_online = delta;
1463 if (max_online < delta)
1464 max_online = delta;
1465 }
1466 }
1467 schedule_timeout_interruptible(onoff_interval * HZ);
1468 }
1469 VERBOSE_PRINTK_STRING("rcu_torture_onoff task stopping");
1470 return 0;
1471}
1472
1473static int
1474rcu_torture_onoff_init(void)
1475{
1476 int ret;
1477
1478 if (onoff_interval <= 0)
1479 return 0;
1480 onoff_task = kthread_run(rcu_torture_onoff, NULL, "rcu_torture_onoff");
1481 if (IS_ERR(onoff_task)) {
1482 ret = PTR_ERR(onoff_task);
1483 onoff_task = NULL;
1484 return ret;
1485 }
1486 return 0;
1487}
1488
1489static void rcu_torture_onoff_cleanup(void)
1490{
1491 if (onoff_task == NULL)
1492 return;
1493 VERBOSE_PRINTK_STRING("Stopping rcu_torture_onoff task");
1494 kthread_stop(onoff_task);
1495 onoff_task = NULL;
1496}
1497
1498#else /* #ifdef CONFIG_HOTPLUG_CPU */
1499
1500static int
1501rcu_torture_onoff_init(void)
1502{
1503 return 0;
1504}
1505
1506static void rcu_torture_onoff_cleanup(void)
1507{
1508}
1509
1510#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
1511
1512/*
1513 * CPU-stall kthread. It waits as specified by stall_cpu_holdoff, then
1514 * induces a CPU stall for the time specified by stall_cpu.
1515 */
1516static int rcu_torture_stall(void *args)
1517{
1518 unsigned long stop_at;
1519
1520 VERBOSE_PRINTK_STRING("rcu_torture_stall task started");
1521 if (stall_cpu_holdoff > 0) {
1522 VERBOSE_PRINTK_STRING("rcu_torture_stall begin holdoff");
1523 schedule_timeout_interruptible(stall_cpu_holdoff * HZ);
1524 VERBOSE_PRINTK_STRING("rcu_torture_stall end holdoff");
1525 }
1526 if (!kthread_should_stop()) {
1527 stop_at = get_seconds() + stall_cpu;
1528 /* RCU CPU stall is expected behavior in following code. */
1529 pr_alert("rcu_torture_stall start.\n");
1530 rcu_read_lock();
1531 preempt_disable();
1532 while (ULONG_CMP_LT(get_seconds(), stop_at))
1533 continue; /* Induce RCU CPU stall warning. */
1534 preempt_enable();
1535 rcu_read_unlock();
1536 pr_alert("rcu_torture_stall end.\n");
1537 }
1538 rcutorture_shutdown_absorb("rcu_torture_stall");
1539 while (!kthread_should_stop())
1540 schedule_timeout_interruptible(10 * HZ);
1541 return 0;
1542}
1543
1544/* Spawn CPU-stall kthread, if stall_cpu specified. */
1545static int __init rcu_torture_stall_init(void)
1546{
1547 int ret;
1548
1549 if (stall_cpu <= 0)
1550 return 0;
1551 stall_task = kthread_run(rcu_torture_stall, NULL, "rcu_torture_stall");
1552 if (IS_ERR(stall_task)) {
1553 ret = PTR_ERR(stall_task);
1554 stall_task = NULL;
1555 return ret;
1556 }
1557 return 0;
1558}
1559
1560/* Clean up after the CPU-stall kthread, if one was spawned. */
1561static void rcu_torture_stall_cleanup(void)
1562{
1563 if (stall_task == NULL)
1564 return;
1565 VERBOSE_PRINTK_STRING("Stopping rcu_torture_stall_task.");
1566 kthread_stop(stall_task);
1567 stall_task = NULL;
1568}
1569
1570/* Callback function for RCU barrier testing. */
1571void rcu_torture_barrier_cbf(struct rcu_head *rcu)
1572{
1573 atomic_inc(&barrier_cbs_invoked);
1574}
1575
1576/* kthread function to register callbacks used to test RCU barriers. */
1577static int rcu_torture_barrier_cbs(void *arg)
1578{
1579 long myid = (long)arg;
1580 bool lastphase = 0;
1581 struct rcu_head rcu;
1582
1583 init_rcu_head_on_stack(&rcu);
1584 VERBOSE_PRINTK_STRING("rcu_torture_barrier_cbs task started");
1585 set_user_nice(current, 19);
1586 do {
1587 wait_event(barrier_cbs_wq[myid],
1588 barrier_phase != lastphase ||
1589 kthread_should_stop() ||
1590 fullstop != FULLSTOP_DONTSTOP);
1591 lastphase = barrier_phase;
1592 smp_mb(); /* ensure barrier_phase load before ->call(). */
1593 if (kthread_should_stop() || fullstop != FULLSTOP_DONTSTOP)
1594 break;
1595 cur_ops->call(&rcu, rcu_torture_barrier_cbf);
1596 if (atomic_dec_and_test(&barrier_cbs_count))
1597 wake_up(&barrier_wq);
1598 } while (!kthread_should_stop() && fullstop == FULLSTOP_DONTSTOP);
1599 VERBOSE_PRINTK_STRING("rcu_torture_barrier_cbs task stopping");
1600 rcutorture_shutdown_absorb("rcu_torture_barrier_cbs");
1601 while (!kthread_should_stop())
1602 schedule_timeout_interruptible(1);
1603 cur_ops->cb_barrier();
1604 destroy_rcu_head_on_stack(&rcu);
1605 return 0;
1606}
1607
1608/* kthread function to drive and coordinate RCU barrier testing. */
1609static int rcu_torture_barrier(void *arg)
1610{
1611 int i;
1612
1613 VERBOSE_PRINTK_STRING("rcu_torture_barrier task starting");
1614 do {
1615 atomic_set(&barrier_cbs_invoked, 0);
1616 atomic_set(&barrier_cbs_count, n_barrier_cbs);
1617 smp_mb(); /* Ensure barrier_phase after prior assignments. */
1618 barrier_phase = !barrier_phase;
1619 for (i = 0; i < n_barrier_cbs; i++)
1620 wake_up(&barrier_cbs_wq[i]);
1621 wait_event(barrier_wq,
1622 atomic_read(&barrier_cbs_count) == 0 ||
1623 kthread_should_stop() ||
1624 fullstop != FULLSTOP_DONTSTOP);
1625 if (kthread_should_stop() || fullstop != FULLSTOP_DONTSTOP)
1626 break;
1627 n_barrier_attempts++;
1628 cur_ops->cb_barrier();
1629 if (atomic_read(&barrier_cbs_invoked) != n_barrier_cbs) {
1630 n_rcu_torture_barrier_error++;
1631 WARN_ON_ONCE(1);
1632 }
1633 n_barrier_successes++;
1634 schedule_timeout_interruptible(HZ / 10);
1635 } while (!kthread_should_stop() && fullstop == FULLSTOP_DONTSTOP);
1636 VERBOSE_PRINTK_STRING("rcu_torture_barrier task stopping");
1637 rcutorture_shutdown_absorb("rcu_torture_barrier");
1638 while (!kthread_should_stop())
1639 schedule_timeout_interruptible(1);
1640 return 0;
1641}
1642
1643/* Initialize RCU barrier testing. */
1644static int rcu_torture_barrier_init(void)
1645{
1646 int i;
1647 int ret;
1648
1649 if (n_barrier_cbs == 0)
1650 return 0;
1651 if (cur_ops->call == NULL || cur_ops->cb_barrier == NULL) {
1652 pr_alert("%s" TORTURE_FLAG
1653 " Call or barrier ops missing for %s,\n",
1654 torture_type, cur_ops->name);
1655 pr_alert("%s" TORTURE_FLAG
1656 " RCU barrier testing omitted from run.\n",
1657 torture_type);
1658 return 0;
1659 }
1660 atomic_set(&barrier_cbs_count, 0);
1661 atomic_set(&barrier_cbs_invoked, 0);
1662 barrier_cbs_tasks =
1663 kzalloc(n_barrier_cbs * sizeof(barrier_cbs_tasks[0]),
1664 GFP_KERNEL);
1665 barrier_cbs_wq =
1666 kzalloc(n_barrier_cbs * sizeof(barrier_cbs_wq[0]),
1667 GFP_KERNEL);
1668 if (barrier_cbs_tasks == NULL || !barrier_cbs_wq)
1669 return -ENOMEM;
1670 for (i = 0; i < n_barrier_cbs; i++) {
1671 init_waitqueue_head(&barrier_cbs_wq[i]);
1672 barrier_cbs_tasks[i] = kthread_run(rcu_torture_barrier_cbs,
1673 (void *)(long)i,
1674 "rcu_torture_barrier_cbs");
1675 if (IS_ERR(barrier_cbs_tasks[i])) {
1676 ret = PTR_ERR(barrier_cbs_tasks[i]);
1677 VERBOSE_PRINTK_ERRSTRING("Failed to create rcu_torture_barrier_cbs");
1678 barrier_cbs_tasks[i] = NULL;
1679 return ret;
1680 }
1681 }
1682 barrier_task = kthread_run(rcu_torture_barrier, NULL,
1683 "rcu_torture_barrier");
1684 if (IS_ERR(barrier_task)) {
1685 ret = PTR_ERR(barrier_task);
1686 VERBOSE_PRINTK_ERRSTRING("Failed to create rcu_torture_barrier");
1687 barrier_task = NULL;
1688 }
1689 return 0;
1690}
1691
1692/* Clean up after RCU barrier testing. */
1693static void rcu_torture_barrier_cleanup(void)
1694{
1695 int i;
1696
1697 if (barrier_task != NULL) {
1698 VERBOSE_PRINTK_STRING("Stopping rcu_torture_barrier task");
1699 kthread_stop(barrier_task);
1700 barrier_task = NULL;
1701 }
1702 if (barrier_cbs_tasks != NULL) {
1703 for (i = 0; i < n_barrier_cbs; i++) {
1704 if (barrier_cbs_tasks[i] != NULL) {
1705 VERBOSE_PRINTK_STRING("Stopping rcu_torture_barrier_cbs task");
1706 kthread_stop(barrier_cbs_tasks[i]);
1707 barrier_cbs_tasks[i] = NULL;
1708 }
1709 }
1710 kfree(barrier_cbs_tasks);
1711 barrier_cbs_tasks = NULL;
1712 }
1713 if (barrier_cbs_wq != NULL) {
1714 kfree(barrier_cbs_wq);
1715 barrier_cbs_wq = NULL;
1716 }
1717}
1718
1719static int rcutorture_cpu_notify(struct notifier_block *self,
1720 unsigned long action, void *hcpu)
1721{
1722 long cpu = (long)hcpu;
1723
1724 switch (action) {
1725 case CPU_ONLINE:
1726 case CPU_DOWN_FAILED:
1727 (void)rcutorture_booster_init(cpu);
1728 break;
1729 case CPU_DOWN_PREPARE:
1730 rcutorture_booster_cleanup(cpu);
1731 break;
1732 default:
1733 break;
1734 }
1735 return NOTIFY_OK;
1736}
1737
1738static struct notifier_block rcutorture_cpu_nb = {
1739 .notifier_call = rcutorture_cpu_notify,
1740};
1741
1742static void
1743rcu_torture_cleanup(void)
1744{
1745 int i;
1746
1747 mutex_lock(&fullstop_mutex);
1748 rcutorture_record_test_transition();
1749 if (fullstop == FULLSTOP_SHUTDOWN) {
1750 pr_warn(/* but going down anyway, so... */
1751 "Concurrent 'rmmod rcutorture' and shutdown illegal!\n");
1752 mutex_unlock(&fullstop_mutex);
1753 schedule_timeout_uninterruptible(10);
1754 if (cur_ops->cb_barrier != NULL)
1755 cur_ops->cb_barrier();
1756 return;
1757 }
1758 fullstop = FULLSTOP_RMMOD;
1759 mutex_unlock(&fullstop_mutex);
1760 unregister_reboot_notifier(&rcutorture_shutdown_nb);
1761 rcu_torture_barrier_cleanup();
1762 rcu_torture_stall_cleanup();
1763 if (stutter_task) {
1764 VERBOSE_PRINTK_STRING("Stopping rcu_torture_stutter task");
1765 kthread_stop(stutter_task);
1766 }
1767 stutter_task = NULL;
1768 if (shuffler_task) {
1769 VERBOSE_PRINTK_STRING("Stopping rcu_torture_shuffle task");
1770 kthread_stop(shuffler_task);
1771 free_cpumask_var(shuffle_tmp_mask);
1772 }
1773 shuffler_task = NULL;
1774
1775 if (writer_task) {
1776 VERBOSE_PRINTK_STRING("Stopping rcu_torture_writer task");
1777 kthread_stop(writer_task);
1778 }
1779 writer_task = NULL;
1780
1781 if (reader_tasks) {
1782 for (i = 0; i < nrealreaders; i++) {
1783 if (reader_tasks[i]) {
1784 VERBOSE_PRINTK_STRING(
1785 "Stopping rcu_torture_reader task");
1786 kthread_stop(reader_tasks[i]);
1787 }
1788 reader_tasks[i] = NULL;
1789 }
1790 kfree(reader_tasks);
1791 reader_tasks = NULL;
1792 }
1793 rcu_torture_current = NULL;
1794
1795 if (fakewriter_tasks) {
1796 for (i = 0; i < nfakewriters; i++) {
1797 if (fakewriter_tasks[i]) {
1798 VERBOSE_PRINTK_STRING(
1799 "Stopping rcu_torture_fakewriter task");
1800 kthread_stop(fakewriter_tasks[i]);
1801 }
1802 fakewriter_tasks[i] = NULL;
1803 }
1804 kfree(fakewriter_tasks);
1805 fakewriter_tasks = NULL;
1806 }
1807
1808 if (stats_task) {
1809 VERBOSE_PRINTK_STRING("Stopping rcu_torture_stats task");
1810 kthread_stop(stats_task);
1811 }
1812 stats_task = NULL;
1813
1814 if (fqs_task) {
1815 VERBOSE_PRINTK_STRING("Stopping rcu_torture_fqs task");
1816 kthread_stop(fqs_task);
1817 }
1818 fqs_task = NULL;
1819 if ((test_boost == 1 && cur_ops->can_boost) ||
1820 test_boost == 2) {
1821 unregister_cpu_notifier(&rcutorture_cpu_nb);
1822 for_each_possible_cpu(i)
1823 rcutorture_booster_cleanup(i);
1824 }
1825 if (shutdown_task != NULL) {
1826 VERBOSE_PRINTK_STRING("Stopping rcu_torture_shutdown task");
1827 kthread_stop(shutdown_task);
1828 }
1829 shutdown_task = NULL;
1830 rcu_torture_onoff_cleanup();
1831
1832 /* Wait for all RCU callbacks to fire. */
1833
1834 if (cur_ops->cb_barrier != NULL)
1835 cur_ops->cb_barrier();
1836
1837 rcu_torture_stats_print(); /* -After- the stats thread is stopped! */
1838
1839 if (atomic_read(&n_rcu_torture_error) || n_rcu_torture_barrier_error)
1840 rcu_torture_print_module_parms(cur_ops, "End of test: FAILURE");
1841 else if (n_online_successes != n_online_attempts ||
1842 n_offline_successes != n_offline_attempts)
1843 rcu_torture_print_module_parms(cur_ops,
1844 "End of test: RCU_HOTPLUG");
1845 else
1846 rcu_torture_print_module_parms(cur_ops, "End of test: SUCCESS");
1847}
1848
1849#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
1850static void rcu_torture_leak_cb(struct rcu_head *rhp)
1851{
1852}
1853
1854static void rcu_torture_err_cb(struct rcu_head *rhp)
1855{
1856 /*
1857 * This -might- happen due to race conditions, but is unlikely.
1858 * The scenario that leads to this happening is that the
1859 * first of the pair of duplicate callbacks is queued,
1860 * someone else starts a grace period that includes that
1861 * callback, then the second of the pair must wait for the
1862 * next grace period. Unlikely, but can happen. If it
1863 * does happen, the debug-objects subsystem won't have splatted.
1864 */
1865 pr_alert("rcutorture: duplicated callback was invoked.\n");
1866}
1867#endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
1868
1869/*
1870 * Verify that double-free causes debug-objects to complain, but only
1871 * if CONFIG_DEBUG_OBJECTS_RCU_HEAD=y. Otherwise, say that the test
1872 * cannot be carried out.
1873 */
1874static void rcu_test_debug_objects(void)
1875{
1876#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
1877 struct rcu_head rh1;
1878 struct rcu_head rh2;
1879
1880 init_rcu_head_on_stack(&rh1);
1881 init_rcu_head_on_stack(&rh2);
1882 pr_alert("rcutorture: WARN: Duplicate call_rcu() test starting.\n");
1883
1884 /* Try to queue the rh2 pair of callbacks for the same grace period. */
1885 preempt_disable(); /* Prevent preemption from interrupting test. */
1886 rcu_read_lock(); /* Make it impossible to finish a grace period. */
1887 call_rcu(&rh1, rcu_torture_leak_cb); /* Start grace period. */
1888 local_irq_disable(); /* Make it harder to start a new grace period. */
1889 call_rcu(&rh2, rcu_torture_leak_cb);
1890 call_rcu(&rh2, rcu_torture_err_cb); /* Duplicate callback. */
1891 local_irq_enable();
1892 rcu_read_unlock();
1893 preempt_enable();
1894
1895 /* Wait for them all to get done so we can safely return. */
1896 rcu_barrier();
1897 pr_alert("rcutorture: WARN: Duplicate call_rcu() test complete.\n");
1898 destroy_rcu_head_on_stack(&rh1);
1899 destroy_rcu_head_on_stack(&rh2);
1900#else /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
1901 pr_alert("rcutorture: !CONFIG_DEBUG_OBJECTS_RCU_HEAD, not testing duplicate call_rcu()\n");
1902#endif /* #else #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
1903}
1904
1905static int __init
1906rcu_torture_init(void)
1907{
1908 int i;
1909 int cpu;
1910 int firsterr = 0;
1911 int retval;
1912 static struct rcu_torture_ops *torture_ops[] = {
1913 &rcu_ops, &rcu_bh_ops, &srcu_ops, &sched_ops,
1914 };
1915
1916 mutex_lock(&fullstop_mutex);
1917
1918 /* Process args and tell the world that the torturer is on the job. */
1919 for (i = 0; i < ARRAY_SIZE(torture_ops); i++) {
1920 cur_ops = torture_ops[i];
1921 if (strcmp(torture_type, cur_ops->name) == 0)
1922 break;
1923 }
1924 if (i == ARRAY_SIZE(torture_ops)) {
1925 pr_alert("rcu-torture: invalid torture type: \"%s\"\n",
1926 torture_type);
1927 pr_alert("rcu-torture types:");
1928 for (i = 0; i < ARRAY_SIZE(torture_ops); i++)
1929 pr_alert(" %s", torture_ops[i]->name);
1930 pr_alert("\n");
1931 mutex_unlock(&fullstop_mutex);
1932 return -EINVAL;
1933 }
1934 if (cur_ops->fqs == NULL && fqs_duration != 0) {
1935 pr_alert("rcu-torture: ->fqs NULL and non-zero fqs_duration, fqs disabled.\n");
1936 fqs_duration = 0;
1937 }
1938 if (cur_ops->init)
1939 cur_ops->init(); /* no "goto unwind" prior to this point!!! */
1940
1941 if (nreaders >= 0)
1942 nrealreaders = nreaders;
1943 else
1944 nrealreaders = 2 * num_online_cpus();
1945 rcu_torture_print_module_parms(cur_ops, "Start of test");
1946 fullstop = FULLSTOP_DONTSTOP;
1947
1948 /* Set up the freelist. */
1949
1950 INIT_LIST_HEAD(&rcu_torture_freelist);
1951 for (i = 0; i < ARRAY_SIZE(rcu_tortures); i++) {
1952 rcu_tortures[i].rtort_mbtest = 0;
1953 list_add_tail(&rcu_tortures[i].rtort_free,
1954 &rcu_torture_freelist);
1955 }
1956
1957 /* Initialize the statistics so that each run gets its own numbers. */
1958
1959 rcu_torture_current = NULL;
1960 rcu_torture_current_version = 0;
1961 atomic_set(&n_rcu_torture_alloc, 0);
1962 atomic_set(&n_rcu_torture_alloc_fail, 0);
1963 atomic_set(&n_rcu_torture_free, 0);
1964 atomic_set(&n_rcu_torture_mberror, 0);
1965 atomic_set(&n_rcu_torture_error, 0);
1966 n_rcu_torture_barrier_error = 0;
1967 n_rcu_torture_boost_ktrerror = 0;
1968 n_rcu_torture_boost_rterror = 0;
1969 n_rcu_torture_boost_failure = 0;
1970 n_rcu_torture_boosts = 0;
1971 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++)
1972 atomic_set(&rcu_torture_wcount[i], 0);
1973 for_each_possible_cpu(cpu) {
1974 for (i = 0; i < RCU_TORTURE_PIPE_LEN + 1; i++) {
1975 per_cpu(rcu_torture_count, cpu)[i] = 0;
1976 per_cpu(rcu_torture_batch, cpu)[i] = 0;
1977 }
1978 }
1979
1980 /* Start up the kthreads. */
1981
1982 VERBOSE_PRINTK_STRING("Creating rcu_torture_writer task");
1983 writer_task = kthread_create(rcu_torture_writer, NULL,
1984 "rcu_torture_writer");
1985 if (IS_ERR(writer_task)) {
1986 firsterr = PTR_ERR(writer_task);
1987 VERBOSE_PRINTK_ERRSTRING("Failed to create writer");
1988 writer_task = NULL;
1989 goto unwind;
1990 }
1991 wake_up_process(writer_task);
1992 fakewriter_tasks = kzalloc(nfakewriters * sizeof(fakewriter_tasks[0]),
1993 GFP_KERNEL);
1994 if (fakewriter_tasks == NULL) {
1995 VERBOSE_PRINTK_ERRSTRING("out of memory");
1996 firsterr = -ENOMEM;
1997 goto unwind;
1998 }
1999 for (i = 0; i < nfakewriters; i++) {
2000 VERBOSE_PRINTK_STRING("Creating rcu_torture_fakewriter task");
2001 fakewriter_tasks[i] = kthread_run(rcu_torture_fakewriter, NULL,
2002 "rcu_torture_fakewriter");
2003 if (IS_ERR(fakewriter_tasks[i])) {
2004 firsterr = PTR_ERR(fakewriter_tasks[i]);
2005 VERBOSE_PRINTK_ERRSTRING("Failed to create fakewriter");
2006 fakewriter_tasks[i] = NULL;
2007 goto unwind;
2008 }
2009 }
2010 reader_tasks = kzalloc(nrealreaders * sizeof(reader_tasks[0]),
2011 GFP_KERNEL);
2012 if (reader_tasks == NULL) {
2013 VERBOSE_PRINTK_ERRSTRING("out of memory");
2014 firsterr = -ENOMEM;
2015 goto unwind;
2016 }
2017 for (i = 0; i < nrealreaders; i++) {
2018 VERBOSE_PRINTK_STRING("Creating rcu_torture_reader task");
2019 reader_tasks[i] = kthread_run(rcu_torture_reader, NULL,
2020 "rcu_torture_reader");
2021 if (IS_ERR(reader_tasks[i])) {
2022 firsterr = PTR_ERR(reader_tasks[i]);
2023 VERBOSE_PRINTK_ERRSTRING("Failed to create reader");
2024 reader_tasks[i] = NULL;
2025 goto unwind;
2026 }
2027 }
2028 if (stat_interval > 0) {
2029 VERBOSE_PRINTK_STRING("Creating rcu_torture_stats task");
2030 stats_task = kthread_run(rcu_torture_stats, NULL,
2031 "rcu_torture_stats");
2032 if (IS_ERR(stats_task)) {
2033 firsterr = PTR_ERR(stats_task);
2034 VERBOSE_PRINTK_ERRSTRING("Failed to create stats");
2035 stats_task = NULL;
2036 goto unwind;
2037 }
2038 }
2039 if (test_no_idle_hz) {
2040 rcu_idle_cpu = num_online_cpus() - 1;
2041
2042 if (!alloc_cpumask_var(&shuffle_tmp_mask, GFP_KERNEL)) {
2043 firsterr = -ENOMEM;
2044 VERBOSE_PRINTK_ERRSTRING("Failed to alloc mask");
2045 goto unwind;
2046 }
2047
2048 /* Create the shuffler thread */
2049 shuffler_task = kthread_run(rcu_torture_shuffle, NULL,
2050 "rcu_torture_shuffle");
2051 if (IS_ERR(shuffler_task)) {
2052 free_cpumask_var(shuffle_tmp_mask);
2053 firsterr = PTR_ERR(shuffler_task);
2054 VERBOSE_PRINTK_ERRSTRING("Failed to create shuffler");
2055 shuffler_task = NULL;
2056 goto unwind;
2057 }
2058 }
2059 if (stutter < 0)
2060 stutter = 0;
2061 if (stutter) {
2062 /* Create the stutter thread */
2063 stutter_task = kthread_run(rcu_torture_stutter, NULL,
2064 "rcu_torture_stutter");
2065 if (IS_ERR(stutter_task)) {
2066 firsterr = PTR_ERR(stutter_task);
2067 VERBOSE_PRINTK_ERRSTRING("Failed to create stutter");
2068 stutter_task = NULL;
2069 goto unwind;
2070 }
2071 }
2072 if (fqs_duration < 0)
2073 fqs_duration = 0;
2074 if (fqs_duration) {
2075 /* Create the stutter thread */
2076 fqs_task = kthread_run(rcu_torture_fqs, NULL,
2077 "rcu_torture_fqs");
2078 if (IS_ERR(fqs_task)) {
2079 firsterr = PTR_ERR(fqs_task);
2080 VERBOSE_PRINTK_ERRSTRING("Failed to create fqs");
2081 fqs_task = NULL;
2082 goto unwind;
2083 }
2084 }
2085 if (test_boost_interval < 1)
2086 test_boost_interval = 1;
2087 if (test_boost_duration < 2)
2088 test_boost_duration = 2;
2089 if ((test_boost == 1 && cur_ops->can_boost) ||
2090 test_boost == 2) {
2091
2092 boost_starttime = jiffies + test_boost_interval * HZ;
2093 register_cpu_notifier(&rcutorture_cpu_nb);
2094 for_each_possible_cpu(i) {
2095 if (cpu_is_offline(i))
2096 continue; /* Heuristic: CPU can go offline. */
2097 retval = rcutorture_booster_init(i);
2098 if (retval < 0) {
2099 firsterr = retval;
2100 goto unwind;
2101 }
2102 }
2103 }
2104 if (shutdown_secs > 0) {
2105 shutdown_time = jiffies + shutdown_secs * HZ;
2106 shutdown_task = kthread_create(rcu_torture_shutdown, NULL,
2107 "rcu_torture_shutdown");
2108 if (IS_ERR(shutdown_task)) {
2109 firsterr = PTR_ERR(shutdown_task);
2110 VERBOSE_PRINTK_ERRSTRING("Failed to create shutdown");
2111 shutdown_task = NULL;
2112 goto unwind;
2113 }
2114 wake_up_process(shutdown_task);
2115 }
2116 i = rcu_torture_onoff_init();
2117 if (i != 0) {
2118 firsterr = i;
2119 goto unwind;
2120 }
2121 register_reboot_notifier(&rcutorture_shutdown_nb);
2122 i = rcu_torture_stall_init();
2123 if (i != 0) {
2124 firsterr = i;
2125 goto unwind;
2126 }
2127 retval = rcu_torture_barrier_init();
2128 if (retval != 0) {
2129 firsterr = retval;
2130 goto unwind;
2131 }
2132 if (object_debug)
2133 rcu_test_debug_objects();
2134 rcutorture_record_test_transition();
2135 mutex_unlock(&fullstop_mutex);
2136 return 0;
2137
2138unwind:
2139 mutex_unlock(&fullstop_mutex);
2140 rcu_torture_cleanup();
2141 return firsterr;
2142}
2143
2144module_init(rcu_torture_init);
2145module_exit(rcu_torture_cleanup);
diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
new file mode 100644
index 000000000000..4c06ddfea7cd
--- /dev/null
+++ b/kernel/rcu/tree.c
@@ -0,0 +1,3416 @@
1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
17 *
18 * Copyright IBM Corporation, 2008
19 *
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com> Hierarchical version
23 *
24 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
25 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
26 *
27 * For detailed explanation of Read-Copy Update mechanism see -
28 * Documentation/RCU
29 */
30#include <linux/types.h>
31#include <linux/kernel.h>
32#include <linux/init.h>
33#include <linux/spinlock.h>
34#include <linux/smp.h>
35#include <linux/rcupdate.h>
36#include <linux/interrupt.h>
37#include <linux/sched.h>
38#include <linux/nmi.h>
39#include <linux/atomic.h>
40#include <linux/bitops.h>
41#include <linux/export.h>
42#include <linux/completion.h>
43#include <linux/moduleparam.h>
44#include <linux/module.h>
45#include <linux/percpu.h>
46#include <linux/notifier.h>
47#include <linux/cpu.h>
48#include <linux/mutex.h>
49#include <linux/time.h>
50#include <linux/kernel_stat.h>
51#include <linux/wait.h>
52#include <linux/kthread.h>
53#include <linux/prefetch.h>
54#include <linux/delay.h>
55#include <linux/stop_machine.h>
56#include <linux/random.h>
57#include <linux/ftrace_event.h>
58#include <linux/suspend.h>
59
60#include "tree.h"
61#include <trace/events/rcu.h>
62
63#include "rcu.h"
64
65MODULE_ALIAS("rcutree");
66#ifdef MODULE_PARAM_PREFIX
67#undef MODULE_PARAM_PREFIX
68#endif
69#define MODULE_PARAM_PREFIX "rcutree."
70
71/* Data structures. */
72
73static struct lock_class_key rcu_node_class[RCU_NUM_LVLS];
74static struct lock_class_key rcu_fqs_class[RCU_NUM_LVLS];
75
76/*
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 */
84#define RCU_STATE_INITIALIZER(sname, sabbr, cr) \
85static char sname##_varname[] = #sname; \
86static const char *tp_##sname##_varname __used __tracepoint_string = sname##_varname; \
87struct rcu_state sname##_state = { \
88 .level = { &sname##_state.node[0] }, \
89 .call = cr, \
90 .fqs_state = RCU_GP_IDLE, \
91 .gpnum = 0UL - 300UL, \
92 .completed = 0UL - 300UL, \
93 .orphan_lock = __RAW_SPIN_LOCK_UNLOCKED(&sname##_state.orphan_lock), \
94 .orphan_nxttail = &sname##_state.orphan_nxtlist, \
95 .orphan_donetail = &sname##_state.orphan_donelist, \
96 .barrier_mutex = __MUTEX_INITIALIZER(sname##_state.barrier_mutex), \
97 .onoff_mutex = __MUTEX_INITIALIZER(sname##_state.onoff_mutex), \
98 .name = sname##_varname, \
99 .abbr = sabbr, \
100}; \
101DEFINE_PER_CPU(struct rcu_data, sname##_data)
102
103RCU_STATE_INITIALIZER(rcu_sched, 's', call_rcu_sched);
104RCU_STATE_INITIALIZER(rcu_bh, 'b', call_rcu_bh);
105
106static struct rcu_state *rcu_state;
107LIST_HEAD(rcu_struct_flavors);
108
109/* Increase (but not decrease) the CONFIG_RCU_FANOUT_LEAF at boot time. */
110static int rcu_fanout_leaf = CONFIG_RCU_FANOUT_LEAF;
111module_param(rcu_fanout_leaf, int, 0444);
112int 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
122/*
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)
126 * optimize synchronize_sched() to a simple barrier(). When this variable
127 * 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 */
131int rcu_scheduler_active __read_mostly;
132EXPORT_SYMBOL_GPL(rcu_scheduler_active);
133
134/*
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
148#ifdef CONFIG_RCU_BOOST
149
150/*
151 * 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);
155DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
156DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
157DEFINE_PER_CPU(char, rcu_cpu_has_work);
158
159#endif /* #ifdef CONFIG_RCU_BOOST */
160
161static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu);
162static void invoke_rcu_core(void);
163static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp);
164
165/*
166 * 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/*
178 * 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/*
188 * Note a quiescent state. Because we do not need to know
189 * how many quiescent states passed, just if there was at least
190 * one since the start of the grace period, this just sets a flag.
191 * The caller must have disabled preemption.
192 */
193void rcu_sched_qs(int cpu)
194{
195 struct rcu_data *rdp = &per_cpu(rcu_sched_data, cpu);
196
197 if (rdp->passed_quiesce == 0)
198 trace_rcu_grace_period(TPS("rcu_sched"), rdp->gpnum, TPS("cpuqs"));
199 rdp->passed_quiesce = 1;
200}
201
202void rcu_bh_qs(int cpu)
203{
204 struct rcu_data *rdp = &per_cpu(rcu_bh_data, cpu);
205
206 if (rdp->passed_quiesce == 0)
207 trace_rcu_grace_period(TPS("rcu_bh"), rdp->gpnum, TPS("cpuqs"));
208 rdp->passed_quiesce = 1;
209}
210
211/*
212 * Note a context switch. This is a quiescent state for RCU-sched,
213 * and requires special handling for preemptible RCU.
214 * The caller must have disabled preemption.
215 */
216void rcu_note_context_switch(int cpu)
217{
218 trace_rcu_utilization(TPS("Start context switch"));
219 rcu_sched_qs(cpu);
220 rcu_preempt_note_context_switch(cpu);
221 trace_rcu_utilization(TPS("End context switch"));
222}
223EXPORT_SYMBOL_GPL(rcu_note_context_switch);
224
225static DEFINE_PER_CPU(struct rcu_dynticks, rcu_dynticks) = {
226 .dynticks_nesting = DYNTICK_TASK_EXIT_IDLE,
227 .dynticks = ATOMIC_INIT(1),
228#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 */
232};
233
234static 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. */
237
238module_param(blimit, long, 0444);
239module_param(qhimark, long, 0444);
240module_param(qlowmark, long, 0444);
241
242static ulong jiffies_till_first_fqs = ULONG_MAX;
243static ulong jiffies_till_next_fqs = ULONG_MAX;
244
245module_param(jiffies_till_first_fqs, ulong, 0644);
246module_param(jiffies_till_next_fqs, ulong, 0644);
247
248static void rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
249 struct rcu_data *rdp);
250static 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);
254static void force_quiescent_state(struct rcu_state *rsp);
255static int rcu_pending(int cpu);
256
257/*
258 * Return the number of RCU-sched batches processed thus far for debug & stats.
259 */
260long rcu_batches_completed_sched(void)
261{
262 return rcu_sched_state.completed;
263}
264EXPORT_SYMBOL_GPL(rcu_batches_completed_sched);
265
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/*
276 * Force a quiescent state for RCU BH.
277 */
278void rcu_bh_force_quiescent_state(void)
279{
280 force_quiescent_state(&rcu_bh_state);
281}
282EXPORT_SYMBOL_GPL(rcu_bh_force_quiescent_state);
283
284/*
285 * 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/*
310 * Force a quiescent state for RCU-sched.
311 */
312void rcu_sched_force_quiescent_state(void)
313{
314 force_quiescent_state(&rcu_sched_state);
315}
316EXPORT_SYMBOL_GPL(rcu_sched_force_quiescent_state);
317
318/*
319 * Does the CPU have callbacks ready to be invoked?
320 */
321static int
322cpu_has_callbacks_ready_to_invoke(struct rcu_data *rdp)
323{
324 return &rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL] &&
325 rdp->nxttail[RCU_DONE_TAIL] != NULL;
326}
327
328/*
329 * 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.
332 */
333static int
334cpu_needs_another_gp(struct rcu_state *rsp, struct rcu_data *rdp)
335{
336 int i;
337
338 if (rcu_gp_in_progress(rsp))
339 return 0; /* No, a grace period is already in progress. */
340 if (rcu_nocb_needs_gp(rsp))
341 return 1; /* Yes, a no-CBs CPU needs one. */
342 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. */
352}
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
362/*
363 * rcu_eqs_enter_common - current CPU is moving towards extended quiescent state
364 *
365 * 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.
368 */
369static void rcu_eqs_enter_common(struct rcu_dynticks *rdtp, long long oldval,
370 bool user)
371{
372 trace_rcu_dyntick(TPS("Start"), oldval, rdtp->dynticks_nesting);
373 if (!user && !is_idle_task(current)) {
374 struct task_struct *idle __maybe_unused =
375 idle_task(smp_processor_id());
376
377 trace_rcu_dyntick(TPS("Error on entry: not idle task"), oldval, 0);
378 ftrace_dump(DUMP_ORIG);
379 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
380 current->pid, current->comm,
381 idle->pid, idle->comm); /* must be idle task! */
382 }
383 rcu_prepare_for_idle(smp_processor_id());
384 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
385 smp_mb__before_atomic_inc(); /* See above. */
386 atomic_inc(&rdtp->dynticks);
387 smp_mb__after_atomic_inc(); /* Force ordering with next sojourn. */
388 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
389
390 /*
391 * It is illegal to enter an extended quiescent state while
392 * in an RCU read-side critical section.
393 */
394 rcu_lockdep_assert(!lock_is_held(&rcu_lock_map),
395 "Illegal idle entry in RCU read-side critical section.");
396 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map),
397 "Illegal idle entry in RCU-bh read-side critical section.");
398 rcu_lockdep_assert(!lock_is_held(&rcu_sched_lock_map),
399 "Illegal idle entry in RCU-sched read-side critical section.");
400}
401
402/*
403 * Enter an RCU extended quiescent state, which can be either the
404 * idle loop or adaptive-tickless usermode execution.
405 */
406static void rcu_eqs_enter(bool user)
407{
408 long long oldval;
409 struct rcu_dynticks *rdtp;
410
411 rdtp = this_cpu_ptr(&rcu_dynticks);
412 oldval = rdtp->dynticks_nesting;
413 WARN_ON_ONCE((oldval & DYNTICK_TASK_NEST_MASK) == 0);
414 if ((oldval & DYNTICK_TASK_NEST_MASK) == DYNTICK_TASK_NEST_VALUE)
415 rdtp->dynticks_nesting = 0;
416 else
417 rdtp->dynticks_nesting -= DYNTICK_TASK_NEST_VALUE;
418 rcu_eqs_enter_common(rdtp, oldval, user);
419}
420
421/**
422 * rcu_idle_enter - inform RCU that current CPU is entering idle
423 *
424 * Enter idle mode, in other words, -leave- the mode in which RCU
425 * read-side critical sections can occur. (Though RCU read-side
426 * critical sections can occur in irq handlers in idle, a possibility
427 * handled by irq_enter() and irq_exit().)
428 *
429 * We crowbar the ->dynticks_nesting field to zero to allow for
430 * the possibility of usermode upcalls having messed up our count
431 * of interrupt nesting level during the prior busy period.
432 */
433void rcu_idle_enter(void)
434{
435 unsigned long flags;
436
437 local_irq_save(flags);
438 rcu_eqs_enter(false);
439 rcu_sysidle_enter(this_cpu_ptr(&rcu_dynticks), 0);
440 local_irq_restore(flags);
441}
442EXPORT_SYMBOL_GPL(rcu_idle_enter);
443
444#ifdef CONFIG_RCU_USER_QS
445/**
446 * rcu_user_enter - inform RCU that we are resuming userspace.
447 *
448 * Enter RCU idle mode right before resuming userspace. No use of RCU
449 * is permitted between this call and rcu_user_exit(). This way the
450 * CPU doesn't need to maintain the tick for RCU maintenance purposes
451 * when the CPU runs in userspace.
452 */
453void rcu_user_enter(void)
454{
455 rcu_eqs_enter(1);
456}
457#endif /* CONFIG_RCU_USER_QS */
458
459/**
460 * rcu_irq_exit - inform RCU that current CPU is exiting irq towards idle
461 *
462 * Exit from an interrupt handler, which might possibly result in entering
463 * idle mode, in other words, leaving the mode in which read-side critical
464 * sections can occur.
465 *
466 * This code assumes that the idle loop never does anything that might
467 * result in unbalanced calls to irq_enter() and irq_exit(). If your
468 * architecture violates this assumption, RCU will give you what you
469 * deserve, good and hard. But very infrequently and irreproducibly.
470 *
471 * Use things like work queues to work around this limitation.
472 *
473 * You have been warned.
474 */
475void rcu_irq_exit(void)
476{
477 unsigned long flags;
478 long long oldval;
479 struct rcu_dynticks *rdtp;
480
481 local_irq_save(flags);
482 rdtp = this_cpu_ptr(&rcu_dynticks);
483 oldval = rdtp->dynticks_nesting;
484 rdtp->dynticks_nesting--;
485 WARN_ON_ONCE(rdtp->dynticks_nesting < 0);
486 if (rdtp->dynticks_nesting)
487 trace_rcu_dyntick(TPS("--="), oldval, rdtp->dynticks_nesting);
488 else
489 rcu_eqs_enter_common(rdtp, oldval, true);
490 rcu_sysidle_enter(rdtp, 1);
491 local_irq_restore(flags);
492}
493
494/*
495 * rcu_eqs_exit_common - current CPU moving away from extended quiescent state
496 *
497 * If the new value of the ->dynticks_nesting counter was previously zero,
498 * we really have exited idle, and must do the appropriate accounting.
499 * The caller must have disabled interrupts.
500 */
501static void rcu_eqs_exit_common(struct rcu_dynticks *rdtp, long long oldval,
502 int user)
503{
504 smp_mb__before_atomic_inc(); /* Force ordering w/previous sojourn. */
505 atomic_inc(&rdtp->dynticks);
506 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
507 smp_mb__after_atomic_inc(); /* See above. */
508 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
509 rcu_cleanup_after_idle(smp_processor_id());
510 trace_rcu_dyntick(TPS("End"), oldval, rdtp->dynticks_nesting);
511 if (!user && !is_idle_task(current)) {
512 struct task_struct *idle __maybe_unused =
513 idle_task(smp_processor_id());
514
515 trace_rcu_dyntick(TPS("Error on exit: not idle task"),
516 oldval, rdtp->dynticks_nesting);
517 ftrace_dump(DUMP_ORIG);
518 WARN_ONCE(1, "Current pid: %d comm: %s / Idle pid: %d comm: %s",
519 current->pid, current->comm,
520 idle->pid, idle->comm); /* must be idle task! */
521 }
522}
523
524/*
525 * Exit an RCU extended quiescent state, which can be either the
526 * idle loop or adaptive-tickless usermode execution.
527 */
528static void rcu_eqs_exit(bool user)
529{
530 struct rcu_dynticks *rdtp;
531 long long oldval;
532
533 rdtp = this_cpu_ptr(&rcu_dynticks);
534 oldval = rdtp->dynticks_nesting;
535 WARN_ON_ONCE(oldval < 0);
536 if (oldval & DYNTICK_TASK_NEST_MASK)
537 rdtp->dynticks_nesting += DYNTICK_TASK_NEST_VALUE;
538 else
539 rdtp->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
540 rcu_eqs_exit_common(rdtp, oldval, user);
541}
542
543/**
544 * rcu_idle_exit - inform RCU that current CPU is leaving idle
545 *
546 * Exit idle mode, in other words, -enter- the mode in which RCU
547 * read-side critical sections can occur.
548 *
549 * We crowbar the ->dynticks_nesting field to DYNTICK_TASK_NEST to
550 * allow for the possibility of usermode upcalls messing up our count
551 * of interrupt nesting level during the busy period that is just
552 * now starting.
553 */
554void rcu_idle_exit(void)
555{
556 unsigned long flags;
557
558 local_irq_save(flags);
559 rcu_eqs_exit(false);
560 rcu_sysidle_exit(this_cpu_ptr(&rcu_dynticks), 0);
561 local_irq_restore(flags);
562}
563EXPORT_SYMBOL_GPL(rcu_idle_exit);
564
565#ifdef CONFIG_RCU_USER_QS
566/**
567 * rcu_user_exit - inform RCU that we are exiting userspace.
568 *
569 * Exit RCU idle mode while entering the kernel because it can
570 * run a RCU read side critical section anytime.
571 */
572void rcu_user_exit(void)
573{
574 rcu_eqs_exit(1);
575}
576#endif /* CONFIG_RCU_USER_QS */
577
578/**
579 * rcu_irq_enter - inform RCU that current CPU is entering irq away from idle
580 *
581 * Enter an interrupt handler, which might possibly result in exiting
582 * idle mode, in other words, entering the mode in which read-side critical
583 * sections can occur.
584 *
585 * Note that the Linux kernel is fully capable of entering an interrupt
586 * handler that it never exits, for example when doing upcalls to
587 * user mode! This code assumes that the idle loop never does upcalls to
588 * user mode. If your architecture does do upcalls from the idle loop (or
589 * does anything else that results in unbalanced calls to the irq_enter()
590 * and irq_exit() functions), RCU will give you what you deserve, good
591 * and hard. But very infrequently and irreproducibly.
592 *
593 * Use things like work queues to work around this limitation.
594 *
595 * You have been warned.
596 */
597void rcu_irq_enter(void)
598{
599 unsigned long flags;
600 struct rcu_dynticks *rdtp;
601 long long oldval;
602
603 local_irq_save(flags);
604 rdtp = this_cpu_ptr(&rcu_dynticks);
605 oldval = rdtp->dynticks_nesting;
606 rdtp->dynticks_nesting++;
607 WARN_ON_ONCE(rdtp->dynticks_nesting == 0);
608 if (oldval)
609 trace_rcu_dyntick(TPS("++="), oldval, rdtp->dynticks_nesting);
610 else
611 rcu_eqs_exit_common(rdtp, oldval, true);
612 rcu_sysidle_exit(rdtp, 1);
613 local_irq_restore(flags);
614}
615
616/**
617 * rcu_nmi_enter - inform RCU of entry to NMI context
618 *
619 * If the CPU was idle with dynamic ticks active, and there is no
620 * irq handler running, this updates rdtp->dynticks_nmi to let the
621 * RCU grace-period handling know that the CPU is active.
622 */
623void rcu_nmi_enter(void)
624{
625 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
626
627 if (rdtp->dynticks_nmi_nesting == 0 &&
628 (atomic_read(&rdtp->dynticks) & 0x1))
629 return;
630 rdtp->dynticks_nmi_nesting++;
631 smp_mb__before_atomic_inc(); /* Force delay from prior write. */
632 atomic_inc(&rdtp->dynticks);
633 /* CPUs seeing atomic_inc() must see later RCU read-side crit sects */
634 smp_mb__after_atomic_inc(); /* See above. */
635 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks) & 0x1));
636}
637
638/**
639 * rcu_nmi_exit - inform RCU of exit from NMI context
640 *
641 * If the CPU was idle with dynamic ticks active, and there is no
642 * irq handler running, this updates rdtp->dynticks_nmi to let the
643 * RCU grace-period handling know that the CPU is no longer active.
644 */
645void rcu_nmi_exit(void)
646{
647 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
648
649 if (rdtp->dynticks_nmi_nesting == 0 ||
650 --rdtp->dynticks_nmi_nesting != 0)
651 return;
652 /* CPUs seeing atomic_inc() must see prior RCU read-side crit sects */
653 smp_mb__before_atomic_inc(); /* See above. */
654 atomic_inc(&rdtp->dynticks);
655 smp_mb__after_atomic_inc(); /* Force delay to next write. */
656 WARN_ON_ONCE(atomic_read(&rdtp->dynticks) & 0x1);
657}
658
659/**
660 * __rcu_is_watching - are RCU read-side critical sections safe?
661 *
662 * Return true if RCU is watching the running CPU, which means that
663 * this CPU can safely enter RCU read-side critical sections. Unlike
664 * rcu_is_watching(), the caller of __rcu_is_watching() must have at
665 * least disabled preemption.
666 */
667bool __rcu_is_watching(void)
668{
669 return atomic_read(this_cpu_ptr(&rcu_dynticks.dynticks)) & 0x1;
670}
671
672/**
673 * rcu_is_watching - see if RCU thinks that the current CPU is idle
674 *
675 * If the current CPU is in its idle loop and is neither in an interrupt
676 * or NMI handler, return true.
677 */
678bool rcu_is_watching(void)
679{
680 int ret;
681
682 preempt_disable();
683 ret = __rcu_is_watching();
684 preempt_enable();
685 return ret;
686}
687EXPORT_SYMBOL_GPL(rcu_is_watching);
688
689#if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU)
690
691/*
692 * Is the current CPU online? Disable preemption to avoid false positives
693 * that could otherwise happen due to the current CPU number being sampled,
694 * this task being preempted, its old CPU being taken offline, resuming
695 * on some other CPU, then determining that its old CPU is now offline.
696 * It is OK to use RCU on an offline processor during initial boot, hence
697 * the check for rcu_scheduler_fully_active. Note also that it is OK
698 * for a CPU coming online to use RCU for one jiffy prior to marking itself
699 * online in the cpu_online_mask. Similarly, it is OK for a CPU going
700 * offline to continue to use RCU for one jiffy after marking itself
701 * offline in the cpu_online_mask. This leniency is necessary given the
702 * non-atomic nature of the online and offline processing, for example,
703 * the fact that a CPU enters the scheduler after completing the CPU_DYING
704 * notifiers.
705 *
706 * This is also why RCU internally marks CPUs online during the
707 * CPU_UP_PREPARE phase and offline during the CPU_DEAD phase.
708 *
709 * Disable checking if in an NMI handler because we cannot safely report
710 * errors from NMI handlers anyway.
711 */
712bool rcu_lockdep_current_cpu_online(void)
713{
714 struct rcu_data *rdp;
715 struct rcu_node *rnp;
716 bool ret;
717
718 if (in_nmi())
719 return 1;
720 preempt_disable();
721 rdp = this_cpu_ptr(&rcu_sched_data);
722 rnp = rdp->mynode;
723 ret = (rdp->grpmask & rnp->qsmaskinit) ||
724 !rcu_scheduler_fully_active;
725 preempt_enable();
726 return ret;
727}
728EXPORT_SYMBOL_GPL(rcu_lockdep_current_cpu_online);
729
730#endif /* #if defined(CONFIG_PROVE_RCU) && defined(CONFIG_HOTPLUG_CPU) */
731
732/**
733 * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
734 *
735 * If the current CPU is idle or running at a first-level (not nested)
736 * interrupt from idle, return true. The caller must have at least
737 * disabled preemption.
738 */
739static int rcu_is_cpu_rrupt_from_idle(void)
740{
741 return __this_cpu_read(rcu_dynticks.dynticks_nesting) <= 1;
742}
743
744/*
745 * Snapshot the specified CPU's dynticks counter so that we can later
746 * credit them with an implicit quiescent state. Return 1 if this CPU
747 * is in dynticks idle mode, which is an extended quiescent state.
748 */
749static int dyntick_save_progress_counter(struct rcu_data *rdp,
750 bool *isidle, unsigned long *maxj)
751{
752 rdp->dynticks_snap = atomic_add_return(0, &rdp->dynticks->dynticks);
753 rcu_sysidle_check_cpu(rdp, isidle, maxj);
754 return (rdp->dynticks_snap & 0x1) == 0;
755}
756
757/*
758 * Return true if the specified CPU has passed through a quiescent
759 * state by virtue of being in or having passed through an dynticks
760 * idle state since the last call to dyntick_save_progress_counter()
761 * for this same CPU, or by virtue of having been offline.
762 */
763static int rcu_implicit_dynticks_qs(struct rcu_data *rdp,
764 bool *isidle, unsigned long *maxj)
765{
766 unsigned int curr;
767 unsigned int snap;
768
769 curr = (unsigned int)atomic_add_return(0, &rdp->dynticks->dynticks);
770 snap = (unsigned int)rdp->dynticks_snap;
771
772 /*
773 * If the CPU passed through or entered a dynticks idle phase with
774 * no active irq/NMI handlers, then we can safely pretend that the CPU
775 * already acknowledged the request to pass through a quiescent
776 * state. Either way, that CPU cannot possibly be in an RCU
777 * read-side critical section that started before the beginning
778 * of the current RCU grace period.
779 */
780 if ((curr & 0x1) == 0 || UINT_CMP_GE(curr, snap + 2)) {
781 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("dti"));
782 rdp->dynticks_fqs++;
783 return 1;
784 }
785
786 /*
787 * Check for the CPU being offline, but only if the grace period
788 * is old enough. We don't need to worry about the CPU changing
789 * state: If we see it offline even once, it has been through a
790 * quiescent state.
791 *
792 * The reason for insisting that the grace period be at least
793 * one jiffy old is that CPUs that are not quite online and that
794 * have just gone offline can still execute RCU read-side critical
795 * sections.
796 */
797 if (ULONG_CMP_GE(rdp->rsp->gp_start + 2, jiffies))
798 return 0; /* Grace period is not old enough. */
799 barrier();
800 if (cpu_is_offline(rdp->cpu)) {
801 trace_rcu_fqs(rdp->rsp->name, rdp->gpnum, rdp->cpu, TPS("ofl"));
802 rdp->offline_fqs++;
803 return 1;
804 }
805
806 /*
807 * There is a possibility that a CPU in adaptive-ticks state
808 * might run in the kernel with the scheduling-clock tick disabled
809 * for an extended time period. Invoke rcu_kick_nohz_cpu() to
810 * force the CPU to restart the scheduling-clock tick in this
811 * CPU is in this state.
812 */
813 rcu_kick_nohz_cpu(rdp->cpu);
814
815 return 0;
816}
817
818static void record_gp_stall_check_time(struct rcu_state *rsp)
819{
820 unsigned long j = ACCESS_ONCE(jiffies);
821
822 rsp->gp_start = j;
823 smp_wmb(); /* Record start time before stall time. */
824 rsp->jiffies_stall = j + rcu_jiffies_till_stall_check();
825}
826
827/*
828 * Dump stacks of all tasks running on stalled CPUs. This is a fallback
829 * for architectures that do not implement trigger_all_cpu_backtrace().
830 * The NMI-triggered stack traces are more accurate because they are
831 * printed by the target CPU.
832 */
833static void rcu_dump_cpu_stacks(struct rcu_state *rsp)
834{
835 int cpu;
836 unsigned long flags;
837 struct rcu_node *rnp;
838
839 rcu_for_each_leaf_node(rsp, rnp) {
840 raw_spin_lock_irqsave(&rnp->lock, flags);
841 if (rnp->qsmask != 0) {
842 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
843 if (rnp->qsmask & (1UL << cpu))
844 dump_cpu_task(rnp->grplo + cpu);
845 }
846 raw_spin_unlock_irqrestore(&rnp->lock, flags);
847 }
848}
849
850static void print_other_cpu_stall(struct rcu_state *rsp)
851{
852 int cpu;
853 long delta;
854 unsigned long flags;
855 int ndetected = 0;
856 struct rcu_node *rnp = rcu_get_root(rsp);
857 long totqlen = 0;
858
859 /* Only let one CPU complain about others per time interval. */
860
861 raw_spin_lock_irqsave(&rnp->lock, flags);
862 delta = jiffies - rsp->jiffies_stall;
863 if (delta < RCU_STALL_RAT_DELAY || !rcu_gp_in_progress(rsp)) {
864 raw_spin_unlock_irqrestore(&rnp->lock, flags);
865 return;
866 }
867 rsp->jiffies_stall = jiffies + 3 * rcu_jiffies_till_stall_check() + 3;
868 raw_spin_unlock_irqrestore(&rnp->lock, flags);
869
870 /*
871 * OK, time to rat on our buddy...
872 * See Documentation/RCU/stallwarn.txt for info on how to debug
873 * RCU CPU stall warnings.
874 */
875 pr_err("INFO: %s detected stalls on CPUs/tasks:",
876 rsp->name);
877 print_cpu_stall_info_begin();
878 rcu_for_each_leaf_node(rsp, rnp) {
879 raw_spin_lock_irqsave(&rnp->lock, flags);
880 ndetected += rcu_print_task_stall(rnp);
881 if (rnp->qsmask != 0) {
882 for (cpu = 0; cpu <= rnp->grphi - rnp->grplo; cpu++)
883 if (rnp->qsmask & (1UL << cpu)) {
884 print_cpu_stall_info(rsp,
885 rnp->grplo + cpu);
886 ndetected++;
887 }
888 }
889 raw_spin_unlock_irqrestore(&rnp->lock, flags);
890 }
891
892 /*
893 * Now rat on any tasks that got kicked up to the root rcu_node
894 * due to CPU offlining.
895 */
896 rnp = rcu_get_root(rsp);
897 raw_spin_lock_irqsave(&rnp->lock, flags);
898 ndetected += rcu_print_task_stall(rnp);
899 raw_spin_unlock_irqrestore(&rnp->lock, flags);
900
901 print_cpu_stall_info_end();
902 for_each_possible_cpu(cpu)
903 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
904 pr_cont("(detected by %d, t=%ld jiffies, g=%lu, c=%lu, q=%lu)\n",
905 smp_processor_id(), (long)(jiffies - rsp->gp_start),
906 rsp->gpnum, rsp->completed, totqlen);
907 if (ndetected == 0)
908 pr_err("INFO: Stall ended before state dump start\n");
909 else if (!trigger_all_cpu_backtrace())
910 rcu_dump_cpu_stacks(rsp);
911
912 /* Complain about tasks blocking the grace period. */
913
914 rcu_print_detail_task_stall(rsp);
915
916 force_quiescent_state(rsp); /* Kick them all. */
917}
918
919/*
920 * This function really isn't for public consumption, but RCU is special in
921 * that context switches can allow the state machine to make progress.
922 */
923extern void resched_cpu(int cpu);
924
925static void print_cpu_stall(struct rcu_state *rsp)
926{
927 int cpu;
928 unsigned long flags;
929 struct rcu_node *rnp = rcu_get_root(rsp);
930 long totqlen = 0;
931
932 /*
933 * OK, time to rat on ourselves...
934 * See Documentation/RCU/stallwarn.txt for info on how to debug
935 * RCU CPU stall warnings.
936 */
937 pr_err("INFO: %s self-detected stall on CPU", rsp->name);
938 print_cpu_stall_info_begin();
939 print_cpu_stall_info(rsp, smp_processor_id());
940 print_cpu_stall_info_end();
941 for_each_possible_cpu(cpu)
942 totqlen += per_cpu_ptr(rsp->rda, cpu)->qlen;
943 pr_cont(" (t=%lu jiffies g=%lu c=%lu q=%lu)\n",
944 jiffies - rsp->gp_start, rsp->gpnum, rsp->completed, totqlen);
945 if (!trigger_all_cpu_backtrace())
946 dump_stack();
947
948 raw_spin_lock_irqsave(&rnp->lock, flags);
949 if (ULONG_CMP_GE(jiffies, rsp->jiffies_stall))
950 rsp->jiffies_stall = jiffies +
951 3 * rcu_jiffies_till_stall_check() + 3;
952 raw_spin_unlock_irqrestore(&rnp->lock, flags);
953
954 /*
955 * Attempt to revive the RCU machinery by forcing a context switch.
956 *
957 * A context switch would normally allow the RCU state machine to make
958 * progress and it could be we're stuck in kernel space without context
959 * switches for an entirely unreasonable amount of time.
960 */
961 resched_cpu(smp_processor_id());
962}
963
964static void check_cpu_stall(struct rcu_state *rsp, struct rcu_data *rdp)
965{
966 unsigned long completed;
967 unsigned long gpnum;
968 unsigned long gps;
969 unsigned long j;
970 unsigned long js;
971 struct rcu_node *rnp;
972
973 if (rcu_cpu_stall_suppress || !rcu_gp_in_progress(rsp))
974 return;
975 j = ACCESS_ONCE(jiffies);
976
977 /*
978 * Lots of memory barriers to reject false positives.
979 *
980 * The idea is to pick up rsp->gpnum, then rsp->jiffies_stall,
981 * then rsp->gp_start, and finally rsp->completed. These values
982 * are updated in the opposite order with memory barriers (or
983 * equivalent) during grace-period initialization and cleanup.
984 * Now, a false positive can occur if we get an new value of
985 * rsp->gp_start and a old value of rsp->jiffies_stall. But given
986 * the memory barriers, the only way that this can happen is if one
987 * grace period ends and another starts between these two fetches.
988 * Detect this by comparing rsp->completed with the previous fetch
989 * from rsp->gpnum.
990 *
991 * Given this check, comparisons of jiffies, rsp->jiffies_stall,
992 * and rsp->gp_start suffice to forestall false positives.
993 */
994 gpnum = ACCESS_ONCE(rsp->gpnum);
995 smp_rmb(); /* Pick up ->gpnum first... */
996 js = ACCESS_ONCE(rsp->jiffies_stall);
997 smp_rmb(); /* ...then ->jiffies_stall before the rest... */
998 gps = ACCESS_ONCE(rsp->gp_start);
999 smp_rmb(); /* ...and finally ->gp_start before ->completed. */
1000 completed = ACCESS_ONCE(rsp->completed);
1001 if (ULONG_CMP_GE(completed, gpnum) ||
1002 ULONG_CMP_LT(j, js) ||
1003 ULONG_CMP_GE(gps, js))
1004 return; /* No stall or GP completed since entering function. */
1005 rnp = rdp->mynode;
1006 if (rcu_gp_in_progress(rsp) &&
1007 (ACCESS_ONCE(rnp->qsmask) & rdp->grpmask)) {
1008
1009 /* We haven't checked in, so go dump stack. */
1010 print_cpu_stall(rsp);
1011
1012 } else if (rcu_gp_in_progress(rsp) &&
1013 ULONG_CMP_GE(j, js + RCU_STALL_RAT_DELAY)) {
1014
1015 /* They had a few time units to dump stack, so complain. */
1016 print_other_cpu_stall(rsp);
1017 }
1018}
1019
1020/**
1021 * rcu_cpu_stall_reset - prevent further stall warnings in current grace period
1022 *
1023 * Set the stall-warning timeout way off into the future, thus preventing
1024 * any RCU CPU stall-warning messages from appearing in the current set of
1025 * RCU grace periods.
1026 *
1027 * The caller must disable hard irqs.
1028 */
1029void rcu_cpu_stall_reset(void)
1030{
1031 struct rcu_state *rsp;
1032
1033 for_each_rcu_flavor(rsp)
1034 rsp->jiffies_stall = jiffies + ULONG_MAX / 2;
1035}
1036
1037/*
1038 * Initialize the specified rcu_data structure's callback list to empty.
1039 */
1040static void init_callback_list(struct rcu_data *rdp)
1041{
1042 int i;
1043
1044 if (init_nocb_callback_list(rdp))
1045 return;
1046 rdp->nxtlist = NULL;
1047 for (i = 0; i < RCU_NEXT_SIZE; i++)
1048 rdp->nxttail[i] = &rdp->nxtlist;
1049}
1050
1051/*
1052 * Determine the value that ->completed will have at the end of the
1053 * next subsequent grace period. This is used to tag callbacks so that
1054 * a CPU can invoke callbacks in a timely fashion even if that CPU has
1055 * been dyntick-idle for an extended period with callbacks under the
1056 * influence of RCU_FAST_NO_HZ.
1057 *
1058 * The caller must hold rnp->lock with interrupts disabled.
1059 */
1060static unsigned long rcu_cbs_completed(struct rcu_state *rsp,
1061 struct rcu_node *rnp)
1062{
1063 /*
1064 * If RCU is idle, we just wait for the next grace period.
1065 * But we can only be sure that RCU is idle if we are looking
1066 * at the root rcu_node structure -- otherwise, a new grace
1067 * period might have started, but just not yet gotten around
1068 * to initializing the current non-root rcu_node structure.
1069 */
1070 if (rcu_get_root(rsp) == rnp && rnp->gpnum == rnp->completed)
1071 return rnp->completed + 1;
1072
1073 /*
1074 * Otherwise, wait for a possible partial grace period and
1075 * then the subsequent full grace period.
1076 */
1077 return rnp->completed + 2;
1078}
1079
1080/*
1081 * Trace-event helper function for rcu_start_future_gp() and
1082 * rcu_nocb_wait_gp().
1083 */
1084static void trace_rcu_future_gp(struct rcu_node *rnp, struct rcu_data *rdp,
1085 unsigned long c, const char *s)
1086{
1087 trace_rcu_future_grace_period(rdp->rsp->name, rnp->gpnum,
1088 rnp->completed, c, rnp->level,
1089 rnp->grplo, rnp->grphi, s);
1090}
1091
1092/*
1093 * Start some future grace period, as needed to handle newly arrived
1094 * callbacks. The required future grace periods are recorded in each
1095 * rcu_node structure's ->need_future_gp field.
1096 *
1097 * The caller must hold the specified rcu_node structure's ->lock.
1098 */
1099static unsigned long __maybe_unused
1100rcu_start_future_gp(struct rcu_node *rnp, struct rcu_data *rdp)
1101{
1102 unsigned long c;
1103 int i;
1104 struct rcu_node *rnp_root = rcu_get_root(rdp->rsp);
1105
1106 /*
1107 * Pick up grace-period number for new callbacks. If this
1108 * grace period is already marked as needed, return to the caller.
1109 */
1110 c = rcu_cbs_completed(rdp->rsp, rnp);
1111 trace_rcu_future_gp(rnp, rdp, c, TPS("Startleaf"));
1112 if (rnp->need_future_gp[c & 0x1]) {
1113 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartleaf"));
1114 return c;
1115 }
1116
1117 /*
1118 * If either this rcu_node structure or the root rcu_node structure
1119 * believe that a grace period is in progress, then we must wait
1120 * for the one following, which is in "c". Because our request
1121 * will be noticed at the end of the current grace period, we don't
1122 * need to explicitly start one.
1123 */
1124 if (rnp->gpnum != rnp->completed ||
1125 ACCESS_ONCE(rnp->gpnum) != ACCESS_ONCE(rnp->completed)) {
1126 rnp->need_future_gp[c & 0x1]++;
1127 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleaf"));
1128 return c;
1129 }
1130
1131 /*
1132 * There might be no grace period in progress. If we don't already
1133 * hold it, acquire the root rcu_node structure's lock in order to
1134 * start one (if needed).
1135 */
1136 if (rnp != rnp_root)
1137 raw_spin_lock(&rnp_root->lock);
1138
1139 /*
1140 * Get a new grace-period number. If there really is no grace
1141 * period in progress, it will be smaller than the one we obtained
1142 * earlier. Adjust callbacks as needed. Note that even no-CBs
1143 * CPUs have a ->nxtcompleted[] array, so no no-CBs checks needed.
1144 */
1145 c = rcu_cbs_completed(rdp->rsp, rnp_root);
1146 for (i = RCU_DONE_TAIL; i < RCU_NEXT_TAIL; i++)
1147 if (ULONG_CMP_LT(c, rdp->nxtcompleted[i]))
1148 rdp->nxtcompleted[i] = c;
1149
1150 /*
1151 * If the needed for the required grace period is already
1152 * recorded, trace and leave.
1153 */
1154 if (rnp_root->need_future_gp[c & 0x1]) {
1155 trace_rcu_future_gp(rnp, rdp, c, TPS("Prestartedroot"));
1156 goto unlock_out;
1157 }
1158
1159 /* Record the need for the future grace period. */
1160 rnp_root->need_future_gp[c & 0x1]++;
1161
1162 /* If a grace period is not already in progress, start one. */
1163 if (rnp_root->gpnum != rnp_root->completed) {
1164 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedleafroot"));
1165 } else {
1166 trace_rcu_future_gp(rnp, rdp, c, TPS("Startedroot"));
1167 rcu_start_gp_advanced(rdp->rsp, rnp_root, rdp);
1168 }
1169unlock_out:
1170 if (rnp != rnp_root)
1171 raw_spin_unlock(&rnp_root->lock);
1172 return c;
1173}
1174
1175/*
1176 * Clean up any old requests for the just-ended grace period. Also return
1177 * whether any additional grace periods have been requested. Also invoke
1178 * rcu_nocb_gp_cleanup() in order to wake up any no-callbacks kthreads
1179 * waiting for this grace period to complete.
1180 */
1181static int rcu_future_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
1182{
1183 int c = rnp->completed;
1184 int needmore;
1185 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1186
1187 rcu_nocb_gp_cleanup(rsp, rnp);
1188 rnp->need_future_gp[c & 0x1] = 0;
1189 needmore = rnp->need_future_gp[(c + 1) & 0x1];
1190 trace_rcu_future_gp(rnp, rdp, c,
1191 needmore ? TPS("CleanupMore") : TPS("Cleanup"));
1192 return needmore;
1193}
1194
1195/*
1196 * If there is room, assign a ->completed number to any callbacks on
1197 * this CPU that have not already been assigned. Also accelerate any
1198 * callbacks that were previously assigned a ->completed number that has
1199 * since proven to be too conservative, which can happen if callbacks get
1200 * assigned a ->completed number while RCU is idle, but with reference to
1201 * a non-root rcu_node structure. This function is idempotent, so it does
1202 * not hurt to call it repeatedly.
1203 *
1204 * The caller must hold rnp->lock with interrupts disabled.
1205 */
1206static void rcu_accelerate_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1207 struct rcu_data *rdp)
1208{
1209 unsigned long c;
1210 int i;
1211
1212 /* If the CPU has no callbacks, nothing to do. */
1213 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1214 return;
1215
1216 /*
1217 * Starting from the sublist containing the callbacks most
1218 * recently assigned a ->completed number and working down, find the
1219 * first sublist that is not assignable to an upcoming grace period.
1220 * Such a sublist has something in it (first two tests) and has
1221 * a ->completed number assigned that will complete sooner than
1222 * the ->completed number for newly arrived callbacks (last test).
1223 *
1224 * The key point is that any later sublist can be assigned the
1225 * same ->completed number as the newly arrived callbacks, which
1226 * means that the callbacks in any of these later sublist can be
1227 * grouped into a single sublist, whether or not they have already
1228 * been assigned a ->completed number.
1229 */
1230 c = rcu_cbs_completed(rsp, rnp);
1231 for (i = RCU_NEXT_TAIL - 1; i > RCU_DONE_TAIL; i--)
1232 if (rdp->nxttail[i] != rdp->nxttail[i - 1] &&
1233 !ULONG_CMP_GE(rdp->nxtcompleted[i], c))
1234 break;
1235
1236 /*
1237 * If there are no sublist for unassigned callbacks, leave.
1238 * At the same time, advance "i" one sublist, so that "i" will
1239 * index into the sublist where all the remaining callbacks should
1240 * be grouped into.
1241 */
1242 if (++i >= RCU_NEXT_TAIL)
1243 return;
1244
1245 /*
1246 * Assign all subsequent callbacks' ->completed number to the next
1247 * full grace period and group them all in the sublist initially
1248 * indexed by "i".
1249 */
1250 for (; i <= RCU_NEXT_TAIL; i++) {
1251 rdp->nxttail[i] = rdp->nxttail[RCU_NEXT_TAIL];
1252 rdp->nxtcompleted[i] = c;
1253 }
1254 /* Record any needed additional grace periods. */
1255 rcu_start_future_gp(rnp, rdp);
1256
1257 /* Trace depending on how much we were able to accelerate. */
1258 if (!*rdp->nxttail[RCU_WAIT_TAIL])
1259 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccWaitCB"));
1260 else
1261 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("AccReadyCB"));
1262}
1263
1264/*
1265 * Move any callbacks whose grace period has completed to the
1266 * RCU_DONE_TAIL sublist, then compact the remaining sublists and
1267 * assign ->completed numbers to any callbacks in the RCU_NEXT_TAIL
1268 * sublist. This function is idempotent, so it does not hurt to
1269 * invoke it repeatedly. As long as it is not invoked -too- often...
1270 *
1271 * The caller must hold rnp->lock with interrupts disabled.
1272 */
1273static void rcu_advance_cbs(struct rcu_state *rsp, struct rcu_node *rnp,
1274 struct rcu_data *rdp)
1275{
1276 int i, j;
1277
1278 /* If the CPU has no callbacks, nothing to do. */
1279 if (!rdp->nxttail[RCU_NEXT_TAIL] || !*rdp->nxttail[RCU_DONE_TAIL])
1280 return;
1281
1282 /*
1283 * Find all callbacks whose ->completed numbers indicate that they
1284 * are ready to invoke, and put them into the RCU_DONE_TAIL sublist.
1285 */
1286 for (i = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++) {
1287 if (ULONG_CMP_LT(rnp->completed, rdp->nxtcompleted[i]))
1288 break;
1289 rdp->nxttail[RCU_DONE_TAIL] = rdp->nxttail[i];
1290 }
1291 /* Clean up any sublist tail pointers that were misordered above. */
1292 for (j = RCU_WAIT_TAIL; j < i; j++)
1293 rdp->nxttail[j] = rdp->nxttail[RCU_DONE_TAIL];
1294
1295 /* Copy down callbacks to fill in empty sublists. */
1296 for (j = RCU_WAIT_TAIL; i < RCU_NEXT_TAIL; i++, j++) {
1297 if (rdp->nxttail[j] == rdp->nxttail[RCU_NEXT_TAIL])
1298 break;
1299 rdp->nxttail[j] = rdp->nxttail[i];
1300 rdp->nxtcompleted[j] = rdp->nxtcompleted[i];
1301 }
1302
1303 /* Classify any remaining callbacks. */
1304 rcu_accelerate_cbs(rsp, rnp, rdp);
1305}
1306
1307/*
1308 * Update CPU-local rcu_data state to record the beginnings and ends of
1309 * grace periods. The caller must hold the ->lock of the leaf rcu_node
1310 * structure corresponding to the current CPU, and must have irqs disabled.
1311 */
1312static void __note_gp_changes(struct rcu_state *rsp, struct rcu_node *rnp, struct rcu_data *rdp)
1313{
1314 /* Handle the ends of any preceding grace periods first. */
1315 if (rdp->completed == rnp->completed) {
1316
1317 /* No grace period end, so just accelerate recent callbacks. */
1318 rcu_accelerate_cbs(rsp, rnp, rdp);
1319
1320 } else {
1321
1322 /* Advance callbacks. */
1323 rcu_advance_cbs(rsp, rnp, rdp);
1324
1325 /* Remember that we saw this grace-period completion. */
1326 rdp->completed = rnp->completed;
1327 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuend"));
1328 }
1329
1330 if (rdp->gpnum != rnp->gpnum) {
1331 /*
1332 * If the current grace period is waiting for this CPU,
1333 * set up to detect a quiescent state, otherwise don't
1334 * go looking for one.
1335 */
1336 rdp->gpnum = rnp->gpnum;
1337 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpustart"));
1338 rdp->passed_quiesce = 0;
1339 rdp->qs_pending = !!(rnp->qsmask & rdp->grpmask);
1340 zero_cpu_stall_ticks(rdp);
1341 }
1342}
1343
1344static void note_gp_changes(struct rcu_state *rsp, struct rcu_data *rdp)
1345{
1346 unsigned long flags;
1347 struct rcu_node *rnp;
1348
1349 local_irq_save(flags);
1350 rnp = rdp->mynode;
1351 if ((rdp->gpnum == ACCESS_ONCE(rnp->gpnum) &&
1352 rdp->completed == ACCESS_ONCE(rnp->completed)) || /* w/out lock. */
1353 !raw_spin_trylock(&rnp->lock)) { /* irqs already off, so later. */
1354 local_irq_restore(flags);
1355 return;
1356 }
1357 __note_gp_changes(rsp, rnp, rdp);
1358 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1359}
1360
1361/*
1362 * Initialize a new grace period. Return 0 if no grace period required.
1363 */
1364static int rcu_gp_init(struct rcu_state *rsp)
1365{
1366 struct rcu_data *rdp;
1367 struct rcu_node *rnp = rcu_get_root(rsp);
1368
1369 rcu_bind_gp_kthread();
1370 raw_spin_lock_irq(&rnp->lock);
1371 if (rsp->gp_flags == 0) {
1372 /* Spurious wakeup, tell caller to go back to sleep. */
1373 raw_spin_unlock_irq(&rnp->lock);
1374 return 0;
1375 }
1376 rsp->gp_flags = 0; /* Clear all flags: New grace period. */
1377
1378 if (WARN_ON_ONCE(rcu_gp_in_progress(rsp))) {
1379 /*
1380 * Grace period already in progress, don't start another.
1381 * Not supposed to be able to happen.
1382 */
1383 raw_spin_unlock_irq(&rnp->lock);
1384 return 0;
1385 }
1386
1387 /* Advance to a new grace period and initialize state. */
1388 record_gp_stall_check_time(rsp);
1389 smp_wmb(); /* Record GP times before starting GP. */
1390 rsp->gpnum++;
1391 trace_rcu_grace_period(rsp->name, rsp->gpnum, TPS("start"));
1392 raw_spin_unlock_irq(&rnp->lock);
1393
1394 /* Exclude any concurrent CPU-hotplug operations. */
1395 mutex_lock(&rsp->onoff_mutex);
1396
1397 /*
1398 * Set the quiescent-state-needed bits in all the rcu_node
1399 * structures for all currently online CPUs in breadth-first order,
1400 * starting from the root rcu_node structure, relying on the layout
1401 * of the tree within the rsp->node[] array. Note that other CPUs
1402 * will access only the leaves of the hierarchy, thus seeing that no
1403 * grace period is in progress, at least until the corresponding
1404 * leaf node has been initialized. In addition, we have excluded
1405 * CPU-hotplug operations.
1406 *
1407 * The grace period cannot complete until the initialization
1408 * process finishes, because this kthread handles both.
1409 */
1410 rcu_for_each_node_breadth_first(rsp, rnp) {
1411 raw_spin_lock_irq(&rnp->lock);
1412 rdp = this_cpu_ptr(rsp->rda);
1413 rcu_preempt_check_blocked_tasks(rnp);
1414 rnp->qsmask = rnp->qsmaskinit;
1415 ACCESS_ONCE(rnp->gpnum) = rsp->gpnum;
1416 WARN_ON_ONCE(rnp->completed != rsp->completed);
1417 ACCESS_ONCE(rnp->completed) = rsp->completed;
1418 if (rnp == rdp->mynode)
1419 __note_gp_changes(rsp, rnp, rdp);
1420 rcu_preempt_boost_start_gp(rnp);
1421 trace_rcu_grace_period_init(rsp->name, rnp->gpnum,
1422 rnp->level, rnp->grplo,
1423 rnp->grphi, rnp->qsmask);
1424 raw_spin_unlock_irq(&rnp->lock);
1425#ifdef CONFIG_PROVE_RCU_DELAY
1426 if ((prandom_u32() % (rcu_num_nodes + 1)) == 0 &&
1427 system_state == SYSTEM_RUNNING)
1428 udelay(200);
1429#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
1430 cond_resched();
1431 }
1432
1433 mutex_unlock(&rsp->onoff_mutex);
1434 return 1;
1435}
1436
1437/*
1438 * Do one round of quiescent-state forcing.
1439 */
1440static int rcu_gp_fqs(struct rcu_state *rsp, int fqs_state_in)
1441{
1442 int fqs_state = fqs_state_in;
1443 bool isidle = false;
1444 unsigned long maxj;
1445 struct rcu_node *rnp = rcu_get_root(rsp);
1446
1447 rsp->n_force_qs++;
1448 if (fqs_state == RCU_SAVE_DYNTICK) {
1449 /* Collect dyntick-idle snapshots. */
1450 if (is_sysidle_rcu_state(rsp)) {
1451 isidle = 1;
1452 maxj = jiffies - ULONG_MAX / 4;
1453 }
1454 force_qs_rnp(rsp, dyntick_save_progress_counter,
1455 &isidle, &maxj);
1456 rcu_sysidle_report_gp(rsp, isidle, maxj);
1457 fqs_state = RCU_FORCE_QS;
1458 } else {
1459 /* Handle dyntick-idle and offline CPUs. */
1460 isidle = 0;
1461 force_qs_rnp(rsp, rcu_implicit_dynticks_qs, &isidle, &maxj);
1462 }
1463 /* Clear flag to prevent immediate re-entry. */
1464 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
1465 raw_spin_lock_irq(&rnp->lock);
1466 rsp->gp_flags &= ~RCU_GP_FLAG_FQS;
1467 raw_spin_unlock_irq(&rnp->lock);
1468 }
1469 return fqs_state;
1470}
1471
1472/*
1473 * Clean up after the old grace period.
1474 */
1475static void rcu_gp_cleanup(struct rcu_state *rsp)
1476{
1477 unsigned long gp_duration;
1478 int nocb = 0;
1479 struct rcu_data *rdp;
1480 struct rcu_node *rnp = rcu_get_root(rsp);
1481
1482 raw_spin_lock_irq(&rnp->lock);
1483 gp_duration = jiffies - rsp->gp_start;
1484 if (gp_duration > rsp->gp_max)
1485 rsp->gp_max = gp_duration;
1486
1487 /*
1488 * We know the grace period is complete, but to everyone else
1489 * it appears to still be ongoing. But it is also the case
1490 * that to everyone else it looks like there is nothing that
1491 * they can do to advance the grace period. It is therefore
1492 * safe for us to drop the lock in order to mark the grace
1493 * period as completed in all of the rcu_node structures.
1494 */
1495 raw_spin_unlock_irq(&rnp->lock);
1496
1497 /*
1498 * Propagate new ->completed value to rcu_node structures so
1499 * that other CPUs don't have to wait until the start of the next
1500 * grace period to process their callbacks. This also avoids
1501 * some nasty RCU grace-period initialization races by forcing
1502 * the end of the current grace period to be completely recorded in
1503 * all of the rcu_node structures before the beginning of the next
1504 * grace period is recorded in any of the rcu_node structures.
1505 */
1506 rcu_for_each_node_breadth_first(rsp, rnp) {
1507 raw_spin_lock_irq(&rnp->lock);
1508 ACCESS_ONCE(rnp->completed) = rsp->gpnum;
1509 rdp = this_cpu_ptr(rsp->rda);
1510 if (rnp == rdp->mynode)
1511 __note_gp_changes(rsp, rnp, rdp);
1512 nocb += rcu_future_gp_cleanup(rsp, rnp);
1513 raw_spin_unlock_irq(&rnp->lock);
1514 cond_resched();
1515 }
1516 rnp = rcu_get_root(rsp);
1517 raw_spin_lock_irq(&rnp->lock);
1518 rcu_nocb_gp_set(rnp, nocb);
1519
1520 rsp->completed = rsp->gpnum; /* Declare grace period done. */
1521 trace_rcu_grace_period(rsp->name, rsp->completed, TPS("end"));
1522 rsp->fqs_state = RCU_GP_IDLE;
1523 rdp = this_cpu_ptr(rsp->rda);
1524 rcu_advance_cbs(rsp, rnp, rdp); /* Reduce false positives below. */
1525 if (cpu_needs_another_gp(rsp, rdp)) {
1526 rsp->gp_flags = RCU_GP_FLAG_INIT;
1527 trace_rcu_grace_period(rsp->name,
1528 ACCESS_ONCE(rsp->gpnum),
1529 TPS("newreq"));
1530 }
1531 raw_spin_unlock_irq(&rnp->lock);
1532}
1533
1534/*
1535 * Body of kthread that handles grace periods.
1536 */
1537static int __noreturn rcu_gp_kthread(void *arg)
1538{
1539 int fqs_state;
1540 int gf;
1541 unsigned long j;
1542 int ret;
1543 struct rcu_state *rsp = arg;
1544 struct rcu_node *rnp = rcu_get_root(rsp);
1545
1546 for (;;) {
1547
1548 /* Handle grace-period start. */
1549 for (;;) {
1550 trace_rcu_grace_period(rsp->name,
1551 ACCESS_ONCE(rsp->gpnum),
1552 TPS("reqwait"));
1553 wait_event_interruptible(rsp->gp_wq,
1554 ACCESS_ONCE(rsp->gp_flags) &
1555 RCU_GP_FLAG_INIT);
1556 if (rcu_gp_init(rsp))
1557 break;
1558 cond_resched();
1559 flush_signals(current);
1560 trace_rcu_grace_period(rsp->name,
1561 ACCESS_ONCE(rsp->gpnum),
1562 TPS("reqwaitsig"));
1563 }
1564
1565 /* Handle quiescent-state forcing. */
1566 fqs_state = RCU_SAVE_DYNTICK;
1567 j = jiffies_till_first_fqs;
1568 if (j > HZ) {
1569 j = HZ;
1570 jiffies_till_first_fqs = HZ;
1571 }
1572 ret = 0;
1573 for (;;) {
1574 if (!ret)
1575 rsp->jiffies_force_qs = jiffies + j;
1576 trace_rcu_grace_period(rsp->name,
1577 ACCESS_ONCE(rsp->gpnum),
1578 TPS("fqswait"));
1579 ret = wait_event_interruptible_timeout(rsp->gp_wq,
1580 ((gf = ACCESS_ONCE(rsp->gp_flags)) &
1581 RCU_GP_FLAG_FQS) ||
1582 (!ACCESS_ONCE(rnp->qsmask) &&
1583 !rcu_preempt_blocked_readers_cgp(rnp)),
1584 j);
1585 /* If grace period done, leave loop. */
1586 if (!ACCESS_ONCE(rnp->qsmask) &&
1587 !rcu_preempt_blocked_readers_cgp(rnp))
1588 break;
1589 /* If time for quiescent-state forcing, do it. */
1590 if (ULONG_CMP_GE(jiffies, rsp->jiffies_force_qs) ||
1591 (gf & RCU_GP_FLAG_FQS)) {
1592 trace_rcu_grace_period(rsp->name,
1593 ACCESS_ONCE(rsp->gpnum),
1594 TPS("fqsstart"));
1595 fqs_state = rcu_gp_fqs(rsp, fqs_state);
1596 trace_rcu_grace_period(rsp->name,
1597 ACCESS_ONCE(rsp->gpnum),
1598 TPS("fqsend"));
1599 cond_resched();
1600 } else {
1601 /* Deal with stray signal. */
1602 cond_resched();
1603 flush_signals(current);
1604 trace_rcu_grace_period(rsp->name,
1605 ACCESS_ONCE(rsp->gpnum),
1606 TPS("fqswaitsig"));
1607 }
1608 j = jiffies_till_next_fqs;
1609 if (j > HZ) {
1610 j = HZ;
1611 jiffies_till_next_fqs = HZ;
1612 } else if (j < 1) {
1613 j = 1;
1614 jiffies_till_next_fqs = 1;
1615 }
1616 }
1617
1618 /* Handle grace-period end. */
1619 rcu_gp_cleanup(rsp);
1620 }
1621}
1622
1623static void rsp_wakeup(struct irq_work *work)
1624{
1625 struct rcu_state *rsp = container_of(work, struct rcu_state, wakeup_work);
1626
1627 /* Wake up rcu_gp_kthread() to start the grace period. */
1628 wake_up(&rsp->gp_wq);
1629}
1630
1631/*
1632 * Start a new RCU grace period if warranted, re-initializing the hierarchy
1633 * in preparation for detecting the next grace period. The caller must hold
1634 * the root node's ->lock and hard irqs must be disabled.
1635 *
1636 * Note that it is legal for a dying CPU (which is marked as offline) to
1637 * invoke this function. This can happen when the dying CPU reports its
1638 * quiescent state.
1639 */
1640static void
1641rcu_start_gp_advanced(struct rcu_state *rsp, struct rcu_node *rnp,
1642 struct rcu_data *rdp)
1643{
1644 if (!rsp->gp_kthread || !cpu_needs_another_gp(rsp, rdp)) {
1645 /*
1646 * Either we have not yet spawned the grace-period
1647 * task, this CPU does not need another grace period,
1648 * or a grace period is already in progress.
1649 * Either way, don't start a new grace period.
1650 */
1651 return;
1652 }
1653 rsp->gp_flags = RCU_GP_FLAG_INIT;
1654 trace_rcu_grace_period(rsp->name, ACCESS_ONCE(rsp->gpnum),
1655 TPS("newreq"));
1656
1657 /*
1658 * We can't do wakeups while holding the rnp->lock, as that
1659 * could cause possible deadlocks with the rq->lock. Defer
1660 * the wakeup to interrupt context. And don't bother waking
1661 * up the running kthread.
1662 */
1663 if (current != rsp->gp_kthread)
1664 irq_work_queue(&rsp->wakeup_work);
1665}
1666
1667/*
1668 * Similar to rcu_start_gp_advanced(), but also advance the calling CPU's
1669 * callbacks. Note that rcu_start_gp_advanced() cannot do this because it
1670 * is invoked indirectly from rcu_advance_cbs(), which would result in
1671 * endless recursion -- or would do so if it wasn't for the self-deadlock
1672 * that is encountered beforehand.
1673 */
1674static void
1675rcu_start_gp(struct rcu_state *rsp)
1676{
1677 struct rcu_data *rdp = this_cpu_ptr(rsp->rda);
1678 struct rcu_node *rnp = rcu_get_root(rsp);
1679
1680 /*
1681 * If there is no grace period in progress right now, any
1682 * callbacks we have up to this point will be satisfied by the
1683 * next grace period. Also, advancing the callbacks reduces the
1684 * probability of false positives from cpu_needs_another_gp()
1685 * resulting in pointless grace periods. So, advance callbacks
1686 * then start the grace period!
1687 */
1688 rcu_advance_cbs(rsp, rnp, rdp);
1689 rcu_start_gp_advanced(rsp, rnp, rdp);
1690}
1691
1692/*
1693 * Report a full set of quiescent states to the specified rcu_state
1694 * data structure. This involves cleaning up after the prior grace
1695 * period and letting rcu_start_gp() start up the next grace period
1696 * if one is needed. Note that the caller must hold rnp->lock, which
1697 * is released before return.
1698 */
1699static void rcu_report_qs_rsp(struct rcu_state *rsp, unsigned long flags)
1700 __releases(rcu_get_root(rsp)->lock)
1701{
1702 WARN_ON_ONCE(!rcu_gp_in_progress(rsp));
1703 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
1704 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
1705}
1706
1707/*
1708 * Similar to rcu_report_qs_rdp(), for which it is a helper function.
1709 * Allows quiescent states for a group of CPUs to be reported at one go
1710 * to the specified rcu_node structure, though all the CPUs in the group
1711 * must be represented by the same rcu_node structure (which need not be
1712 * a leaf rcu_node structure, though it often will be). That structure's
1713 * lock must be held upon entry, and it is released before return.
1714 */
1715static void
1716rcu_report_qs_rnp(unsigned long mask, struct rcu_state *rsp,
1717 struct rcu_node *rnp, unsigned long flags)
1718 __releases(rnp->lock)
1719{
1720 struct rcu_node *rnp_c;
1721
1722 /* Walk up the rcu_node hierarchy. */
1723 for (;;) {
1724 if (!(rnp->qsmask & mask)) {
1725
1726 /* Our bit has already been cleared, so done. */
1727 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1728 return;
1729 }
1730 rnp->qsmask &= ~mask;
1731 trace_rcu_quiescent_state_report(rsp->name, rnp->gpnum,
1732 mask, rnp->qsmask, rnp->level,
1733 rnp->grplo, rnp->grphi,
1734 !!rnp->gp_tasks);
1735 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
1736
1737 /* Other bits still set at this level, so done. */
1738 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1739 return;
1740 }
1741 mask = rnp->grpmask;
1742 if (rnp->parent == NULL) {
1743
1744 /* No more levels. Exit loop holding root lock. */
1745
1746 break;
1747 }
1748 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1749 rnp_c = rnp;
1750 rnp = rnp->parent;
1751 raw_spin_lock_irqsave(&rnp->lock, flags);
1752 WARN_ON_ONCE(rnp_c->qsmask);
1753 }
1754
1755 /*
1756 * Get here if we are the last CPU to pass through a quiescent
1757 * state for this grace period. Invoke rcu_report_qs_rsp()
1758 * to clean up and start the next grace period if one is needed.
1759 */
1760 rcu_report_qs_rsp(rsp, flags); /* releases rnp->lock. */
1761}
1762
1763/*
1764 * Record a quiescent state for the specified CPU to that CPU's rcu_data
1765 * structure. This must be either called from the specified CPU, or
1766 * called when the specified CPU is known to be offline (and when it is
1767 * also known that no other CPU is concurrently trying to help the offline
1768 * CPU). The lastcomp argument is used to make sure we are still in the
1769 * grace period of interest. We don't want to end the current grace period
1770 * based on quiescent states detected in an earlier grace period!
1771 */
1772static void
1773rcu_report_qs_rdp(int cpu, struct rcu_state *rsp, struct rcu_data *rdp)
1774{
1775 unsigned long flags;
1776 unsigned long mask;
1777 struct rcu_node *rnp;
1778
1779 rnp = rdp->mynode;
1780 raw_spin_lock_irqsave(&rnp->lock, flags);
1781 if (rdp->passed_quiesce == 0 || rdp->gpnum != rnp->gpnum ||
1782 rnp->completed == rnp->gpnum) {
1783
1784 /*
1785 * The grace period in which this quiescent state was
1786 * recorded has ended, so don't report it upwards.
1787 * We will instead need a new quiescent state that lies
1788 * within the current grace period.
1789 */
1790 rdp->passed_quiesce = 0; /* need qs for new gp. */
1791 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1792 return;
1793 }
1794 mask = rdp->grpmask;
1795 if ((rnp->qsmask & mask) == 0) {
1796 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1797 } else {
1798 rdp->qs_pending = 0;
1799
1800 /*
1801 * This GP can't end until cpu checks in, so all of our
1802 * callbacks can be processed during the next GP.
1803 */
1804 rcu_accelerate_cbs(rsp, rnp, rdp);
1805
1806 rcu_report_qs_rnp(mask, rsp, rnp, flags); /* rlses rnp->lock */
1807 }
1808}
1809
1810/*
1811 * Check to see if there is a new grace period of which this CPU
1812 * is not yet aware, and if so, set up local rcu_data state for it.
1813 * Otherwise, see if this CPU has just passed through its first
1814 * quiescent state for this grace period, and record that fact if so.
1815 */
1816static void
1817rcu_check_quiescent_state(struct rcu_state *rsp, struct rcu_data *rdp)
1818{
1819 /* Check for grace-period ends and beginnings. */
1820 note_gp_changes(rsp, rdp);
1821
1822 /*
1823 * Does this CPU still need to do its part for current grace period?
1824 * If no, return and let the other CPUs do their part as well.
1825 */
1826 if (!rdp->qs_pending)
1827 return;
1828
1829 /*
1830 * Was there a quiescent state since the beginning of the grace
1831 * period? If no, then exit and wait for the next call.
1832 */
1833 if (!rdp->passed_quiesce)
1834 return;
1835
1836 /*
1837 * Tell RCU we are done (but rcu_report_qs_rdp() will be the
1838 * judge of that).
1839 */
1840 rcu_report_qs_rdp(rdp->cpu, rsp, rdp);
1841}
1842
1843#ifdef CONFIG_HOTPLUG_CPU
1844
1845/*
1846 * Send the specified CPU's RCU callbacks to the orphanage. The
1847 * specified CPU must be offline, and the caller must hold the
1848 * ->orphan_lock.
1849 */
1850static void
1851rcu_send_cbs_to_orphanage(int cpu, struct rcu_state *rsp,
1852 struct rcu_node *rnp, struct rcu_data *rdp)
1853{
1854 /* No-CBs CPUs do not have orphanable callbacks. */
1855 if (rcu_is_nocb_cpu(rdp->cpu))
1856 return;
1857
1858 /*
1859 * Orphan the callbacks. First adjust the counts. This is safe
1860 * because _rcu_barrier() excludes CPU-hotplug operations, so it
1861 * cannot be running now. Thus no memory barrier is required.
1862 */
1863 if (rdp->nxtlist != NULL) {
1864 rsp->qlen_lazy += rdp->qlen_lazy;
1865 rsp->qlen += rdp->qlen;
1866 rdp->n_cbs_orphaned += rdp->qlen;
1867 rdp->qlen_lazy = 0;
1868 ACCESS_ONCE(rdp->qlen) = 0;
1869 }
1870
1871 /*
1872 * Next, move those callbacks still needing a grace period to
1873 * the orphanage, where some other CPU will pick them up.
1874 * Some of the callbacks might have gone partway through a grace
1875 * period, but that is too bad. They get to start over because we
1876 * cannot assume that grace periods are synchronized across CPUs.
1877 * We don't bother updating the ->nxttail[] array yet, instead
1878 * we just reset the whole thing later on.
1879 */
1880 if (*rdp->nxttail[RCU_DONE_TAIL] != NULL) {
1881 *rsp->orphan_nxttail = *rdp->nxttail[RCU_DONE_TAIL];
1882 rsp->orphan_nxttail = rdp->nxttail[RCU_NEXT_TAIL];
1883 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
1884 }
1885
1886 /*
1887 * Then move the ready-to-invoke callbacks to the orphanage,
1888 * where some other CPU will pick them up. These will not be
1889 * required to pass though another grace period: They are done.
1890 */
1891 if (rdp->nxtlist != NULL) {
1892 *rsp->orphan_donetail = rdp->nxtlist;
1893 rsp->orphan_donetail = rdp->nxttail[RCU_DONE_TAIL];
1894 }
1895
1896 /* Finally, initialize the rcu_data structure's list to empty. */
1897 init_callback_list(rdp);
1898}
1899
1900/*
1901 * Adopt the RCU callbacks from the specified rcu_state structure's
1902 * orphanage. The caller must hold the ->orphan_lock.
1903 */
1904static void rcu_adopt_orphan_cbs(struct rcu_state *rsp)
1905{
1906 int i;
1907 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
1908
1909 /* No-CBs CPUs are handled specially. */
1910 if (rcu_nocb_adopt_orphan_cbs(rsp, rdp))
1911 return;
1912
1913 /* Do the accounting first. */
1914 rdp->qlen_lazy += rsp->qlen_lazy;
1915 rdp->qlen += rsp->qlen;
1916 rdp->n_cbs_adopted += rsp->qlen;
1917 if (rsp->qlen_lazy != rsp->qlen)
1918 rcu_idle_count_callbacks_posted();
1919 rsp->qlen_lazy = 0;
1920 rsp->qlen = 0;
1921
1922 /*
1923 * We do not need a memory barrier here because the only way we
1924 * can get here if there is an rcu_barrier() in flight is if
1925 * we are the task doing the rcu_barrier().
1926 */
1927
1928 /* First adopt the ready-to-invoke callbacks. */
1929 if (rsp->orphan_donelist != NULL) {
1930 *rsp->orphan_donetail = *rdp->nxttail[RCU_DONE_TAIL];
1931 *rdp->nxttail[RCU_DONE_TAIL] = rsp->orphan_donelist;
1932 for (i = RCU_NEXT_SIZE - 1; i >= RCU_DONE_TAIL; i--)
1933 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
1934 rdp->nxttail[i] = rsp->orphan_donetail;
1935 rsp->orphan_donelist = NULL;
1936 rsp->orphan_donetail = &rsp->orphan_donelist;
1937 }
1938
1939 /* And then adopt the callbacks that still need a grace period. */
1940 if (rsp->orphan_nxtlist != NULL) {
1941 *rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxtlist;
1942 rdp->nxttail[RCU_NEXT_TAIL] = rsp->orphan_nxttail;
1943 rsp->orphan_nxtlist = NULL;
1944 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
1945 }
1946}
1947
1948/*
1949 * Trace the fact that this CPU is going offline.
1950 */
1951static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
1952{
1953 RCU_TRACE(unsigned long mask);
1954 RCU_TRACE(struct rcu_data *rdp = this_cpu_ptr(rsp->rda));
1955 RCU_TRACE(struct rcu_node *rnp = rdp->mynode);
1956
1957 RCU_TRACE(mask = rdp->grpmask);
1958 trace_rcu_grace_period(rsp->name,
1959 rnp->gpnum + 1 - !!(rnp->qsmask & mask),
1960 TPS("cpuofl"));
1961}
1962
1963/*
1964 * The CPU has been completely removed, and some other CPU is reporting
1965 * this fact from process context. Do the remainder of the cleanup,
1966 * including orphaning the outgoing CPU's RCU callbacks, and also
1967 * adopting them. There can only be one CPU hotplug operation at a time,
1968 * so no other CPU can be attempting to update rcu_cpu_kthread_task.
1969 */
1970static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
1971{
1972 unsigned long flags;
1973 unsigned long mask;
1974 int need_report = 0;
1975 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1976 struct rcu_node *rnp = rdp->mynode; /* Outgoing CPU's rdp & rnp. */
1977
1978 /* Adjust any no-longer-needed kthreads. */
1979 rcu_boost_kthread_setaffinity(rnp, -1);
1980
1981 /* Remove the dead CPU from the bitmasks in the rcu_node hierarchy. */
1982
1983 /* Exclude any attempts to start a new grace period. */
1984 mutex_lock(&rsp->onoff_mutex);
1985 raw_spin_lock_irqsave(&rsp->orphan_lock, flags);
1986
1987 /* Orphan the dead CPU's callbacks, and adopt them if appropriate. */
1988 rcu_send_cbs_to_orphanage(cpu, rsp, rnp, rdp);
1989 rcu_adopt_orphan_cbs(rsp);
1990
1991 /* Remove the outgoing CPU from the masks in the rcu_node hierarchy. */
1992 mask = rdp->grpmask; /* rnp->grplo is constant. */
1993 do {
1994 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1995 rnp->qsmaskinit &= ~mask;
1996 if (rnp->qsmaskinit != 0) {
1997 if (rnp != rdp->mynode)
1998 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1999 break;
2000 }
2001 if (rnp == rdp->mynode)
2002 need_report = rcu_preempt_offline_tasks(rsp, rnp, rdp);
2003 else
2004 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
2005 mask = rnp->grpmask;
2006 rnp = rnp->parent;
2007 } while (rnp != NULL);
2008
2009 /*
2010 * We still hold the leaf rcu_node structure lock here, and
2011 * irqs are still disabled. The reason for this subterfuge is
2012 * because invoking rcu_report_unblock_qs_rnp() with ->orphan_lock
2013 * held leads to deadlock.
2014 */
2015 raw_spin_unlock(&rsp->orphan_lock); /* irqs remain disabled. */
2016 rnp = rdp->mynode;
2017 if (need_report & RCU_OFL_TASKS_NORM_GP)
2018 rcu_report_unblock_qs_rnp(rnp, flags);
2019 else
2020 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2021 if (need_report & RCU_OFL_TASKS_EXP_GP)
2022 rcu_report_exp_rnp(rsp, rnp, true);
2023 WARN_ONCE(rdp->qlen != 0 || rdp->nxtlist != NULL,
2024 "rcu_cleanup_dead_cpu: Callbacks on offline CPU %d: qlen=%lu, nxtlist=%p\n",
2025 cpu, rdp->qlen, rdp->nxtlist);
2026 init_callback_list(rdp);
2027 /* Disallow further callbacks on this CPU. */
2028 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2029 mutex_unlock(&rsp->onoff_mutex);
2030}
2031
2032#else /* #ifdef CONFIG_HOTPLUG_CPU */
2033
2034static void rcu_cleanup_dying_cpu(struct rcu_state *rsp)
2035{
2036}
2037
2038static void rcu_cleanup_dead_cpu(int cpu, struct rcu_state *rsp)
2039{
2040}
2041
2042#endif /* #else #ifdef CONFIG_HOTPLUG_CPU */
2043
2044/*
2045 * Invoke any RCU callbacks that have made it to the end of their grace
2046 * period. Thottle as specified by rdp->blimit.
2047 */
2048static void rcu_do_batch(struct rcu_state *rsp, struct rcu_data *rdp)
2049{
2050 unsigned long flags;
2051 struct rcu_head *next, *list, **tail;
2052 long bl, count, count_lazy;
2053 int i;
2054
2055 /* If no callbacks are ready, just return. */
2056 if (!cpu_has_callbacks_ready_to_invoke(rdp)) {
2057 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, 0);
2058 trace_rcu_batch_end(rsp->name, 0, !!ACCESS_ONCE(rdp->nxtlist),
2059 need_resched(), is_idle_task(current),
2060 rcu_is_callbacks_kthread());
2061 return;
2062 }
2063
2064 /*
2065 * Extract the list of ready callbacks, disabling to prevent
2066 * races with call_rcu() from interrupt handlers.
2067 */
2068 local_irq_save(flags);
2069 WARN_ON_ONCE(cpu_is_offline(smp_processor_id()));
2070 bl = rdp->blimit;
2071 trace_rcu_batch_start(rsp->name, rdp->qlen_lazy, rdp->qlen, bl);
2072 list = rdp->nxtlist;
2073 rdp->nxtlist = *rdp->nxttail[RCU_DONE_TAIL];
2074 *rdp->nxttail[RCU_DONE_TAIL] = NULL;
2075 tail = rdp->nxttail[RCU_DONE_TAIL];
2076 for (i = RCU_NEXT_SIZE - 1; i >= 0; i--)
2077 if (rdp->nxttail[i] == rdp->nxttail[RCU_DONE_TAIL])
2078 rdp->nxttail[i] = &rdp->nxtlist;
2079 local_irq_restore(flags);
2080
2081 /* Invoke callbacks. */
2082 count = count_lazy = 0;
2083 while (list) {
2084 next = list->next;
2085 prefetch(next);
2086 debug_rcu_head_unqueue(list);
2087 if (__rcu_reclaim(rsp->name, list))
2088 count_lazy++;
2089 list = next;
2090 /* Stop only if limit reached and CPU has something to do. */
2091 if (++count >= bl &&
2092 (need_resched() ||
2093 (!is_idle_task(current) && !rcu_is_callbacks_kthread())))
2094 break;
2095 }
2096
2097 local_irq_save(flags);
2098 trace_rcu_batch_end(rsp->name, count, !!list, need_resched(),
2099 is_idle_task(current),
2100 rcu_is_callbacks_kthread());
2101
2102 /* Update count, and requeue any remaining callbacks. */
2103 if (list != NULL) {
2104 *tail = rdp->nxtlist;
2105 rdp->nxtlist = list;
2106 for (i = 0; i < RCU_NEXT_SIZE; i++)
2107 if (&rdp->nxtlist == rdp->nxttail[i])
2108 rdp->nxttail[i] = tail;
2109 else
2110 break;
2111 }
2112 smp_mb(); /* List handling before counting for rcu_barrier(). */
2113 rdp->qlen_lazy -= count_lazy;
2114 ACCESS_ONCE(rdp->qlen) -= count;
2115 rdp->n_cbs_invoked += count;
2116
2117 /* Reinstate batch limit if we have worked down the excess. */
2118 if (rdp->blimit == LONG_MAX && rdp->qlen <= qlowmark)
2119 rdp->blimit = blimit;
2120
2121 /* Reset ->qlen_last_fqs_check trigger if enough CBs have drained. */
2122 if (rdp->qlen == 0 && rdp->qlen_last_fqs_check != 0) {
2123 rdp->qlen_last_fqs_check = 0;
2124 rdp->n_force_qs_snap = rsp->n_force_qs;
2125 } else if (rdp->qlen < rdp->qlen_last_fqs_check - qhimark)
2126 rdp->qlen_last_fqs_check = rdp->qlen;
2127 WARN_ON_ONCE((rdp->nxtlist == NULL) != (rdp->qlen == 0));
2128
2129 local_irq_restore(flags);
2130
2131 /* Re-invoke RCU core processing if there are callbacks remaining. */
2132 if (cpu_has_callbacks_ready_to_invoke(rdp))
2133 invoke_rcu_core();
2134}
2135
2136/*
2137 * Check to see if this CPU is in a non-context-switch quiescent state
2138 * (user mode or idle loop for rcu, non-softirq execution for rcu_bh).
2139 * Also schedule RCU core processing.
2140 *
2141 * This function must be called from hardirq context. It is normally
2142 * invoked from the scheduling-clock interrupt. If rcu_pending returns
2143 * false, there is no point in invoking rcu_check_callbacks().
2144 */
2145void rcu_check_callbacks(int cpu, int user)
2146{
2147 trace_rcu_utilization(TPS("Start scheduler-tick"));
2148 increment_cpu_stall_ticks();
2149 if (user || rcu_is_cpu_rrupt_from_idle()) {
2150
2151 /*
2152 * Get here if this CPU took its interrupt from user
2153 * mode or from the idle loop, and if this is not a
2154 * nested interrupt. In this case, the CPU is in
2155 * a quiescent state, so note it.
2156 *
2157 * No memory barrier is required here because both
2158 * rcu_sched_qs() and rcu_bh_qs() reference only CPU-local
2159 * variables that other CPUs neither access nor modify,
2160 * at least not while the corresponding CPU is online.
2161 */
2162
2163 rcu_sched_qs(cpu);
2164 rcu_bh_qs(cpu);
2165
2166 } else if (!in_softirq()) {
2167
2168 /*
2169 * Get here if this CPU did not take its interrupt from
2170 * softirq, in other words, if it is not interrupting
2171 * a rcu_bh read-side critical section. This is an _bh
2172 * critical section, so note it.
2173 */
2174
2175 rcu_bh_qs(cpu);
2176 }
2177 rcu_preempt_check_callbacks(cpu);
2178 if (rcu_pending(cpu))
2179 invoke_rcu_core();
2180 trace_rcu_utilization(TPS("End scheduler-tick"));
2181}
2182
2183/*
2184 * Scan the leaf rcu_node structures, processing dyntick state for any that
2185 * have not yet encountered a quiescent state, using the function specified.
2186 * Also initiate boosting for any threads blocked on the root rcu_node.
2187 *
2188 * The caller must have suppressed start of new grace periods.
2189 */
2190static void force_qs_rnp(struct rcu_state *rsp,
2191 int (*f)(struct rcu_data *rsp, bool *isidle,
2192 unsigned long *maxj),
2193 bool *isidle, unsigned long *maxj)
2194{
2195 unsigned long bit;
2196 int cpu;
2197 unsigned long flags;
2198 unsigned long mask;
2199 struct rcu_node *rnp;
2200
2201 rcu_for_each_leaf_node(rsp, rnp) {
2202 cond_resched();
2203 mask = 0;
2204 raw_spin_lock_irqsave(&rnp->lock, flags);
2205 if (!rcu_gp_in_progress(rsp)) {
2206 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2207 return;
2208 }
2209 if (rnp->qsmask == 0) {
2210 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
2211 continue;
2212 }
2213 cpu = rnp->grplo;
2214 bit = 1;
2215 for (; cpu <= rnp->grphi; cpu++, bit <<= 1) {
2216 if ((rnp->qsmask & bit) != 0) {
2217 if ((rnp->qsmaskinit & bit) != 0)
2218 *isidle = 0;
2219 if (f(per_cpu_ptr(rsp->rda, cpu), isidle, maxj))
2220 mask |= bit;
2221 }
2222 }
2223 if (mask != 0) {
2224
2225 /* rcu_report_qs_rnp() releases rnp->lock. */
2226 rcu_report_qs_rnp(mask, rsp, rnp, flags);
2227 continue;
2228 }
2229 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2230 }
2231 rnp = rcu_get_root(rsp);
2232 if (rnp->qsmask == 0) {
2233 raw_spin_lock_irqsave(&rnp->lock, flags);
2234 rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */
2235 }
2236}
2237
2238/*
2239 * Force quiescent states on reluctant CPUs, and also detect which
2240 * CPUs are in dyntick-idle mode.
2241 */
2242static void force_quiescent_state(struct rcu_state *rsp)
2243{
2244 unsigned long flags;
2245 bool ret;
2246 struct rcu_node *rnp;
2247 struct rcu_node *rnp_old = NULL;
2248
2249 /* Funnel through hierarchy to reduce memory contention. */
2250 rnp = per_cpu_ptr(rsp->rda, raw_smp_processor_id())->mynode;
2251 for (; rnp != NULL; rnp = rnp->parent) {
2252 ret = (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) ||
2253 !raw_spin_trylock(&rnp->fqslock);
2254 if (rnp_old != NULL)
2255 raw_spin_unlock(&rnp_old->fqslock);
2256 if (ret) {
2257 rsp->n_force_qs_lh++;
2258 return;
2259 }
2260 rnp_old = rnp;
2261 }
2262 /* rnp_old == rcu_get_root(rsp), rnp == NULL. */
2263
2264 /* Reached the root of the rcu_node tree, acquire lock. */
2265 raw_spin_lock_irqsave(&rnp_old->lock, flags);
2266 raw_spin_unlock(&rnp_old->fqslock);
2267 if (ACCESS_ONCE(rsp->gp_flags) & RCU_GP_FLAG_FQS) {
2268 rsp->n_force_qs_lh++;
2269 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2270 return; /* Someone beat us to it. */
2271 }
2272 rsp->gp_flags |= RCU_GP_FLAG_FQS;
2273 raw_spin_unlock_irqrestore(&rnp_old->lock, flags);
2274 wake_up(&rsp->gp_wq); /* Memory barrier implied by wake_up() path. */
2275}
2276
2277/*
2278 * This does the RCU core processing work for the specified rcu_state
2279 * and rcu_data structures. This may be called only from the CPU to
2280 * whom the rdp belongs.
2281 */
2282static void
2283__rcu_process_callbacks(struct rcu_state *rsp)
2284{
2285 unsigned long flags;
2286 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2287
2288 WARN_ON_ONCE(rdp->beenonline == 0);
2289
2290 /* Update RCU state based on any recent quiescent states. */
2291 rcu_check_quiescent_state(rsp, rdp);
2292
2293 /* Does this CPU require a not-yet-started grace period? */
2294 local_irq_save(flags);
2295 if (cpu_needs_another_gp(rsp, rdp)) {
2296 raw_spin_lock(&rcu_get_root(rsp)->lock); /* irqs disabled. */
2297 rcu_start_gp(rsp);
2298 raw_spin_unlock_irqrestore(&rcu_get_root(rsp)->lock, flags);
2299 } else {
2300 local_irq_restore(flags);
2301 }
2302
2303 /* If there are callbacks ready, invoke them. */
2304 if (cpu_has_callbacks_ready_to_invoke(rdp))
2305 invoke_rcu_callbacks(rsp, rdp);
2306}
2307
2308/*
2309 * Do RCU core processing for the current CPU.
2310 */
2311static void rcu_process_callbacks(struct softirq_action *unused)
2312{
2313 struct rcu_state *rsp;
2314
2315 if (cpu_is_offline(smp_processor_id()))
2316 return;
2317 trace_rcu_utilization(TPS("Start RCU core"));
2318 for_each_rcu_flavor(rsp)
2319 __rcu_process_callbacks(rsp);
2320 trace_rcu_utilization(TPS("End RCU core"));
2321}
2322
2323/*
2324 * Schedule RCU callback invocation. If the specified type of RCU
2325 * does not support RCU priority boosting, just do a direct call,
2326 * otherwise wake up the per-CPU kernel kthread. Note that because we
2327 * are running on the current CPU with interrupts disabled, the
2328 * rcu_cpu_kthread_task cannot disappear out from under us.
2329 */
2330static void invoke_rcu_callbacks(struct rcu_state *rsp, struct rcu_data *rdp)
2331{
2332 if (unlikely(!ACCESS_ONCE(rcu_scheduler_fully_active)))
2333 return;
2334 if (likely(!rsp->boost)) {
2335 rcu_do_batch(rsp, rdp);
2336 return;
2337 }
2338 invoke_rcu_callbacks_kthread();
2339}
2340
2341static void invoke_rcu_core(void)
2342{
2343 if (cpu_online(smp_processor_id()))
2344 raise_softirq(RCU_SOFTIRQ);
2345}
2346
2347/*
2348 * Handle any core-RCU processing required by a call_rcu() invocation.
2349 */
2350static void __call_rcu_core(struct rcu_state *rsp, struct rcu_data *rdp,
2351 struct rcu_head *head, unsigned long flags)
2352{
2353 /*
2354 * If called from an extended quiescent state, invoke the RCU
2355 * core in order to force a re-evaluation of RCU's idleness.
2356 */
2357 if (!rcu_is_watching() && cpu_online(smp_processor_id()))
2358 invoke_rcu_core();
2359
2360 /* If interrupts were disabled or CPU offline, don't invoke RCU core. */
2361 if (irqs_disabled_flags(flags) || cpu_is_offline(smp_processor_id()))
2362 return;
2363
2364 /*
2365 * Force the grace period if too many callbacks or too long waiting.
2366 * Enforce hysteresis, and don't invoke force_quiescent_state()
2367 * if some other CPU has recently done so. Also, don't bother
2368 * invoking force_quiescent_state() if the newly enqueued callback
2369 * is the only one waiting for a grace period to complete.
2370 */
2371 if (unlikely(rdp->qlen > rdp->qlen_last_fqs_check + qhimark)) {
2372
2373 /* Are we ignoring a completed grace period? */
2374 note_gp_changes(rsp, rdp);
2375
2376 /* Start a new grace period if one not already started. */
2377 if (!rcu_gp_in_progress(rsp)) {
2378 struct rcu_node *rnp_root = rcu_get_root(rsp);
2379
2380 raw_spin_lock(&rnp_root->lock);
2381 rcu_start_gp(rsp);
2382 raw_spin_unlock(&rnp_root->lock);
2383 } else {
2384 /* Give the grace period a kick. */
2385 rdp->blimit = LONG_MAX;
2386 if (rsp->n_force_qs == rdp->n_force_qs_snap &&
2387 *rdp->nxttail[RCU_DONE_TAIL] != head)
2388 force_quiescent_state(rsp);
2389 rdp->n_force_qs_snap = rsp->n_force_qs;
2390 rdp->qlen_last_fqs_check = rdp->qlen;
2391 }
2392 }
2393}
2394
2395/*
2396 * RCU callback function to leak a callback.
2397 */
2398static void rcu_leak_callback(struct rcu_head *rhp)
2399{
2400}
2401
2402/*
2403 * Helper function for call_rcu() and friends. The cpu argument will
2404 * normally be -1, indicating "currently running CPU". It may specify
2405 * a CPU only if that CPU is a no-CBs CPU. Currently, only _rcu_barrier()
2406 * is expected to specify a CPU.
2407 */
2408static void
2409__call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu),
2410 struct rcu_state *rsp, int cpu, bool lazy)
2411{
2412 unsigned long flags;
2413 struct rcu_data *rdp;
2414
2415 WARN_ON_ONCE((unsigned long)head & 0x3); /* Misaligned rcu_head! */
2416 if (debug_rcu_head_queue(head)) {
2417 /* Probable double call_rcu(), so leak the callback. */
2418 ACCESS_ONCE(head->func) = rcu_leak_callback;
2419 WARN_ONCE(1, "__call_rcu(): Leaked duplicate callback\n");
2420 return;
2421 }
2422 head->func = func;
2423 head->next = NULL;
2424
2425 /*
2426 * Opportunistically note grace-period endings and beginnings.
2427 * Note that we might see a beginning right after we see an
2428 * end, but never vice versa, since this CPU has to pass through
2429 * a quiescent state betweentimes.
2430 */
2431 local_irq_save(flags);
2432 rdp = this_cpu_ptr(rsp->rda);
2433
2434 /* Add the callback to our list. */
2435 if (unlikely(rdp->nxttail[RCU_NEXT_TAIL] == NULL) || cpu != -1) {
2436 int offline;
2437
2438 if (cpu != -1)
2439 rdp = per_cpu_ptr(rsp->rda, cpu);
2440 offline = !__call_rcu_nocb(rdp, head, lazy);
2441 WARN_ON_ONCE(offline);
2442 /* _call_rcu() is illegal on offline CPU; leak the callback. */
2443 local_irq_restore(flags);
2444 return;
2445 }
2446 ACCESS_ONCE(rdp->qlen)++;
2447 if (lazy)
2448 rdp->qlen_lazy++;
2449 else
2450 rcu_idle_count_callbacks_posted();
2451 smp_mb(); /* Count before adding callback for rcu_barrier(). */
2452 *rdp->nxttail[RCU_NEXT_TAIL] = head;
2453 rdp->nxttail[RCU_NEXT_TAIL] = &head->next;
2454
2455 if (__is_kfree_rcu_offset((unsigned long)func))
2456 trace_rcu_kfree_callback(rsp->name, head, (unsigned long)func,
2457 rdp->qlen_lazy, rdp->qlen);
2458 else
2459 trace_rcu_callback(rsp->name, head, rdp->qlen_lazy, rdp->qlen);
2460
2461 /* Go handle any RCU core processing required. */
2462 __call_rcu_core(rsp, rdp, head, flags);
2463 local_irq_restore(flags);
2464}
2465
2466/*
2467 * Queue an RCU-sched callback for invocation after a grace period.
2468 */
2469void call_rcu_sched(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2470{
2471 __call_rcu(head, func, &rcu_sched_state, -1, 0);
2472}
2473EXPORT_SYMBOL_GPL(call_rcu_sched);
2474
2475/*
2476 * Queue an RCU callback for invocation after a quicker grace period.
2477 */
2478void call_rcu_bh(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
2479{
2480 __call_rcu(head, func, &rcu_bh_state, -1, 0);
2481}
2482EXPORT_SYMBOL_GPL(call_rcu_bh);
2483
2484/*
2485 * Because a context switch is a grace period for RCU-sched and RCU-bh,
2486 * any blocking grace-period wait automatically implies a grace period
2487 * if there is only one CPU online at any point time during execution
2488 * of either synchronize_sched() or synchronize_rcu_bh(). It is OK to
2489 * occasionally incorrectly indicate that there are multiple CPUs online
2490 * when there was in fact only one the whole time, as this just adds
2491 * some overhead: RCU still operates correctly.
2492 */
2493static inline int rcu_blocking_is_gp(void)
2494{
2495 int ret;
2496
2497 might_sleep(); /* Check for RCU read-side critical section. */
2498 preempt_disable();
2499 ret = num_online_cpus() <= 1;
2500 preempt_enable();
2501 return ret;
2502}
2503
2504/**
2505 * synchronize_sched - wait until an rcu-sched grace period has elapsed.
2506 *
2507 * Control will return to the caller some time after a full rcu-sched
2508 * grace period has elapsed, in other words after all currently executing
2509 * rcu-sched read-side critical sections have completed. These read-side
2510 * critical sections are delimited by rcu_read_lock_sched() and
2511 * rcu_read_unlock_sched(), and may be nested. Note that preempt_disable(),
2512 * local_irq_disable(), and so on may be used in place of
2513 * rcu_read_lock_sched().
2514 *
2515 * This means that all preempt_disable code sequences, including NMI and
2516 * non-threaded hardware-interrupt handlers, in progress on entry will
2517 * have completed before this primitive returns. However, this does not
2518 * guarantee that softirq handlers will have completed, since in some
2519 * kernels, these handlers can run in process context, and can block.
2520 *
2521 * Note that this guarantee implies further memory-ordering guarantees.
2522 * On systems with more than one CPU, when synchronize_sched() returns,
2523 * each CPU is guaranteed to have executed a full memory barrier since the
2524 * end of its last RCU-sched read-side critical section whose beginning
2525 * preceded the call to synchronize_sched(). In addition, each CPU having
2526 * an RCU read-side critical section that extends beyond the return from
2527 * synchronize_sched() is guaranteed to have executed a full memory barrier
2528 * after the beginning of synchronize_sched() and before the beginning of
2529 * that RCU read-side critical section. Note that these guarantees include
2530 * CPUs that are offline, idle, or executing in user mode, as well as CPUs
2531 * that are executing in the kernel.
2532 *
2533 * Furthermore, if CPU A invoked synchronize_sched(), which returned
2534 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
2535 * to have executed a full memory barrier during the execution of
2536 * synchronize_sched() -- even if CPU A and CPU B are the same CPU (but
2537 * again only if the system has more than one CPU).
2538 *
2539 * This primitive provides the guarantees made by the (now removed)
2540 * synchronize_kernel() API. In contrast, synchronize_rcu() only
2541 * guarantees that rcu_read_lock() sections will have completed.
2542 * In "classic RCU", these two guarantees happen to be one and
2543 * the same, but can differ in realtime RCU implementations.
2544 */
2545void synchronize_sched(void)
2546{
2547 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2548 !lock_is_held(&rcu_lock_map) &&
2549 !lock_is_held(&rcu_sched_lock_map),
2550 "Illegal synchronize_sched() in RCU-sched read-side critical section");
2551 if (rcu_blocking_is_gp())
2552 return;
2553 if (rcu_expedited)
2554 synchronize_sched_expedited();
2555 else
2556 wait_rcu_gp(call_rcu_sched);
2557}
2558EXPORT_SYMBOL_GPL(synchronize_sched);
2559
2560/**
2561 * synchronize_rcu_bh - wait until an rcu_bh grace period has elapsed.
2562 *
2563 * Control will return to the caller some time after a full rcu_bh grace
2564 * period has elapsed, in other words after all currently executing rcu_bh
2565 * read-side critical sections have completed. RCU read-side critical
2566 * sections are delimited by rcu_read_lock_bh() and rcu_read_unlock_bh(),
2567 * and may be nested.
2568 *
2569 * See the description of synchronize_sched() for more detailed information
2570 * on memory ordering guarantees.
2571 */
2572void synchronize_rcu_bh(void)
2573{
2574 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
2575 !lock_is_held(&rcu_lock_map) &&
2576 !lock_is_held(&rcu_sched_lock_map),
2577 "Illegal synchronize_rcu_bh() in RCU-bh read-side critical section");
2578 if (rcu_blocking_is_gp())
2579 return;
2580 if (rcu_expedited)
2581 synchronize_rcu_bh_expedited();
2582 else
2583 wait_rcu_gp(call_rcu_bh);
2584}
2585EXPORT_SYMBOL_GPL(synchronize_rcu_bh);
2586
2587static int synchronize_sched_expedited_cpu_stop(void *data)
2588{
2589 /*
2590 * There must be a full memory barrier on each affected CPU
2591 * between the time that try_stop_cpus() is called and the
2592 * time that it returns.
2593 *
2594 * In the current initial implementation of cpu_stop, the
2595 * above condition is already met when the control reaches
2596 * this point and the following smp_mb() is not strictly
2597 * necessary. Do smp_mb() anyway for documentation and
2598 * robustness against future implementation changes.
2599 */
2600 smp_mb(); /* See above comment block. */
2601 return 0;
2602}
2603
2604/**
2605 * synchronize_sched_expedited - Brute-force RCU-sched grace period
2606 *
2607 * Wait for an RCU-sched grace period to elapse, but use a "big hammer"
2608 * approach to force the grace period to end quickly. This consumes
2609 * significant time on all CPUs and is unfriendly to real-time workloads,
2610 * so is thus not recommended for any sort of common-case code. In fact,
2611 * if you are using synchronize_sched_expedited() in a loop, please
2612 * restructure your code to batch your updates, and then use a single
2613 * synchronize_sched() instead.
2614 *
2615 * Note that it is illegal to call this function while holding any lock
2616 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
2617 * to call this function from a CPU-hotplug notifier. Failing to observe
2618 * these restriction will result in deadlock.
2619 *
2620 * This implementation can be thought of as an application of ticket
2621 * locking to RCU, with sync_sched_expedited_started and
2622 * sync_sched_expedited_done taking on the roles of the halves
2623 * of the ticket-lock word. Each task atomically increments
2624 * sync_sched_expedited_started upon entry, snapshotting the old value,
2625 * then attempts to stop all the CPUs. If this succeeds, then each
2626 * CPU will have executed a context switch, resulting in an RCU-sched
2627 * grace period. We are then done, so we use atomic_cmpxchg() to
2628 * update sync_sched_expedited_done to match our snapshot -- but
2629 * only if someone else has not already advanced past our snapshot.
2630 *
2631 * On the other hand, if try_stop_cpus() fails, we check the value
2632 * of sync_sched_expedited_done. If it has advanced past our
2633 * initial snapshot, then someone else must have forced a grace period
2634 * some time after we took our snapshot. In this case, our work is
2635 * done for us, and we can simply return. Otherwise, we try again,
2636 * but keep our initial snapshot for purposes of checking for someone
2637 * doing our work for us.
2638 *
2639 * If we fail too many times in a row, we fall back to synchronize_sched().
2640 */
2641void synchronize_sched_expedited(void)
2642{
2643 long firstsnap, s, snap;
2644 int trycount = 0;
2645 struct rcu_state *rsp = &rcu_sched_state;
2646
2647 /*
2648 * If we are in danger of counter wrap, just do synchronize_sched().
2649 * By allowing sync_sched_expedited_started to advance no more than
2650 * ULONG_MAX/8 ahead of sync_sched_expedited_done, we are ensuring
2651 * that more than 3.5 billion CPUs would be required to force a
2652 * counter wrap on a 32-bit system. Quite a few more CPUs would of
2653 * course be required on a 64-bit system.
2654 */
2655 if (ULONG_CMP_GE((ulong)atomic_long_read(&rsp->expedited_start),
2656 (ulong)atomic_long_read(&rsp->expedited_done) +
2657 ULONG_MAX / 8)) {
2658 synchronize_sched();
2659 atomic_long_inc(&rsp->expedited_wrap);
2660 return;
2661 }
2662
2663 /*
2664 * Take a ticket. Note that atomic_inc_return() implies a
2665 * full memory barrier.
2666 */
2667 snap = atomic_long_inc_return(&rsp->expedited_start);
2668 firstsnap = snap;
2669 get_online_cpus();
2670 WARN_ON_ONCE(cpu_is_offline(raw_smp_processor_id()));
2671
2672 /*
2673 * Each pass through the following loop attempts to force a
2674 * context switch on each CPU.
2675 */
2676 while (try_stop_cpus(cpu_online_mask,
2677 synchronize_sched_expedited_cpu_stop,
2678 NULL) == -EAGAIN) {
2679 put_online_cpus();
2680 atomic_long_inc(&rsp->expedited_tryfail);
2681
2682 /* Check to see if someone else did our work for us. */
2683 s = atomic_long_read(&rsp->expedited_done);
2684 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2685 /* ensure test happens before caller kfree */
2686 smp_mb__before_atomic_inc(); /* ^^^ */
2687 atomic_long_inc(&rsp->expedited_workdone1);
2688 return;
2689 }
2690
2691 /* No joy, try again later. Or just synchronize_sched(). */
2692 if (trycount++ < 10) {
2693 udelay(trycount * num_online_cpus());
2694 } else {
2695 wait_rcu_gp(call_rcu_sched);
2696 atomic_long_inc(&rsp->expedited_normal);
2697 return;
2698 }
2699
2700 /* Recheck to see if someone else did our work for us. */
2701 s = atomic_long_read(&rsp->expedited_done);
2702 if (ULONG_CMP_GE((ulong)s, (ulong)firstsnap)) {
2703 /* ensure test happens before caller kfree */
2704 smp_mb__before_atomic_inc(); /* ^^^ */
2705 atomic_long_inc(&rsp->expedited_workdone2);
2706 return;
2707 }
2708
2709 /*
2710 * Refetching sync_sched_expedited_started allows later
2711 * callers to piggyback on our grace period. We retry
2712 * after they started, so our grace period works for them,
2713 * and they started after our first try, so their grace
2714 * period works for us.
2715 */
2716 get_online_cpus();
2717 snap = atomic_long_read(&rsp->expedited_start);
2718 smp_mb(); /* ensure read is before try_stop_cpus(). */
2719 }
2720 atomic_long_inc(&rsp->expedited_stoppedcpus);
2721
2722 /*
2723 * Everyone up to our most recent fetch is covered by our grace
2724 * period. Update the counter, but only if our work is still
2725 * relevant -- which it won't be if someone who started later
2726 * than we did already did their update.
2727 */
2728 do {
2729 atomic_long_inc(&rsp->expedited_done_tries);
2730 s = atomic_long_read(&rsp->expedited_done);
2731 if (ULONG_CMP_GE((ulong)s, (ulong)snap)) {
2732 /* ensure test happens before caller kfree */
2733 smp_mb__before_atomic_inc(); /* ^^^ */
2734 atomic_long_inc(&rsp->expedited_done_lost);
2735 break;
2736 }
2737 } while (atomic_long_cmpxchg(&rsp->expedited_done, s, snap) != s);
2738 atomic_long_inc(&rsp->expedited_done_exit);
2739
2740 put_online_cpus();
2741}
2742EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
2743
2744/*
2745 * Check to see if there is any immediate RCU-related work to be done
2746 * by the current CPU, for the specified type of RCU, returning 1 if so.
2747 * The checks are in order of increasing expense: checks that can be
2748 * carried out against CPU-local state are performed first. However,
2749 * we must check for CPU stalls first, else we might not get a chance.
2750 */
2751static int __rcu_pending(struct rcu_state *rsp, struct rcu_data *rdp)
2752{
2753 struct rcu_node *rnp = rdp->mynode;
2754
2755 rdp->n_rcu_pending++;
2756
2757 /* Check for CPU stalls, if enabled. */
2758 check_cpu_stall(rsp, rdp);
2759
2760 /* Is the RCU core waiting for a quiescent state from this CPU? */
2761 if (rcu_scheduler_fully_active &&
2762 rdp->qs_pending && !rdp->passed_quiesce) {
2763 rdp->n_rp_qs_pending++;
2764 } else if (rdp->qs_pending && rdp->passed_quiesce) {
2765 rdp->n_rp_report_qs++;
2766 return 1;
2767 }
2768
2769 /* Does this CPU have callbacks ready to invoke? */
2770 if (cpu_has_callbacks_ready_to_invoke(rdp)) {
2771 rdp->n_rp_cb_ready++;
2772 return 1;
2773 }
2774
2775 /* Has RCU gone idle with this CPU needing another grace period? */
2776 if (cpu_needs_another_gp(rsp, rdp)) {
2777 rdp->n_rp_cpu_needs_gp++;
2778 return 1;
2779 }
2780
2781 /* Has another RCU grace period completed? */
2782 if (ACCESS_ONCE(rnp->completed) != rdp->completed) { /* outside lock */
2783 rdp->n_rp_gp_completed++;
2784 return 1;
2785 }
2786
2787 /* Has a new RCU grace period started? */
2788 if (ACCESS_ONCE(rnp->gpnum) != rdp->gpnum) { /* outside lock */
2789 rdp->n_rp_gp_started++;
2790 return 1;
2791 }
2792
2793 /* nothing to do */
2794 rdp->n_rp_need_nothing++;
2795 return 0;
2796}
2797
2798/*
2799 * Check to see if there is any immediate RCU-related work to be done
2800 * by the current CPU, returning 1 if so. This function is part of the
2801 * RCU implementation; it is -not- an exported member of the RCU API.
2802 */
2803static int rcu_pending(int cpu)
2804{
2805 struct rcu_state *rsp;
2806
2807 for_each_rcu_flavor(rsp)
2808 if (__rcu_pending(rsp, per_cpu_ptr(rsp->rda, cpu)))
2809 return 1;
2810 return 0;
2811}
2812
2813/*
2814 * Return true if the specified CPU has any callback. If all_lazy is
2815 * non-NULL, store an indication of whether all callbacks are lazy.
2816 * (If there are no callbacks, all of them are deemed to be lazy.)
2817 */
2818static int rcu_cpu_has_callbacks(int cpu, bool *all_lazy)
2819{
2820 bool al = true;
2821 bool hc = false;
2822 struct rcu_data *rdp;
2823 struct rcu_state *rsp;
2824
2825 for_each_rcu_flavor(rsp) {
2826 rdp = per_cpu_ptr(rsp->rda, cpu);
2827 if (!rdp->nxtlist)
2828 continue;
2829 hc = true;
2830 if (rdp->qlen != rdp->qlen_lazy || !all_lazy) {
2831 al = false;
2832 break;
2833 }
2834 }
2835 if (all_lazy)
2836 *all_lazy = al;
2837 return hc;
2838}
2839
2840/*
2841 * Helper function for _rcu_barrier() tracing. If tracing is disabled,
2842 * the compiler is expected to optimize this away.
2843 */
2844static void _rcu_barrier_trace(struct rcu_state *rsp, const char *s,
2845 int cpu, unsigned long done)
2846{
2847 trace_rcu_barrier(rsp->name, s, cpu,
2848 atomic_read(&rsp->barrier_cpu_count), done);
2849}
2850
2851/*
2852 * RCU callback function for _rcu_barrier(). If we are last, wake
2853 * up the task executing _rcu_barrier().
2854 */
2855static void rcu_barrier_callback(struct rcu_head *rhp)
2856{
2857 struct rcu_data *rdp = container_of(rhp, struct rcu_data, barrier_head);
2858 struct rcu_state *rsp = rdp->rsp;
2859
2860 if (atomic_dec_and_test(&rsp->barrier_cpu_count)) {
2861 _rcu_barrier_trace(rsp, "LastCB", -1, rsp->n_barrier_done);
2862 complete(&rsp->barrier_completion);
2863 } else {
2864 _rcu_barrier_trace(rsp, "CB", -1, rsp->n_barrier_done);
2865 }
2866}
2867
2868/*
2869 * Called with preemption disabled, and from cross-cpu IRQ context.
2870 */
2871static void rcu_barrier_func(void *type)
2872{
2873 struct rcu_state *rsp = type;
2874 struct rcu_data *rdp = __this_cpu_ptr(rsp->rda);
2875
2876 _rcu_barrier_trace(rsp, "IRQ", -1, rsp->n_barrier_done);
2877 atomic_inc(&rsp->barrier_cpu_count);
2878 rsp->call(&rdp->barrier_head, rcu_barrier_callback);
2879}
2880
2881/*
2882 * Orchestrate the specified type of RCU barrier, waiting for all
2883 * RCU callbacks of the specified type to complete.
2884 */
2885static void _rcu_barrier(struct rcu_state *rsp)
2886{
2887 int cpu;
2888 struct rcu_data *rdp;
2889 unsigned long snap = ACCESS_ONCE(rsp->n_barrier_done);
2890 unsigned long snap_done;
2891
2892 _rcu_barrier_trace(rsp, "Begin", -1, snap);
2893
2894 /* Take mutex to serialize concurrent rcu_barrier() requests. */
2895 mutex_lock(&rsp->barrier_mutex);
2896
2897 /*
2898 * Ensure that all prior references, including to ->n_barrier_done,
2899 * are ordered before the _rcu_barrier() machinery.
2900 */
2901 smp_mb(); /* See above block comment. */
2902
2903 /*
2904 * Recheck ->n_barrier_done to see if others did our work for us.
2905 * This means checking ->n_barrier_done for an even-to-odd-to-even
2906 * transition. The "if" expression below therefore rounds the old
2907 * value up to the next even number and adds two before comparing.
2908 */
2909 snap_done = rsp->n_barrier_done;
2910 _rcu_barrier_trace(rsp, "Check", -1, snap_done);
2911
2912 /*
2913 * If the value in snap is odd, we needed to wait for the current
2914 * rcu_barrier() to complete, then wait for the next one, in other
2915 * words, we need the value of snap_done to be three larger than
2916 * the value of snap. On the other hand, if the value in snap is
2917 * even, we only had to wait for the next rcu_barrier() to complete,
2918 * in other words, we need the value of snap_done to be only two
2919 * greater than the value of snap. The "(snap + 3) & ~0x1" computes
2920 * this for us (thank you, Linus!).
2921 */
2922 if (ULONG_CMP_GE(snap_done, (snap + 3) & ~0x1)) {
2923 _rcu_barrier_trace(rsp, "EarlyExit", -1, snap_done);
2924 smp_mb(); /* caller's subsequent code after above check. */
2925 mutex_unlock(&rsp->barrier_mutex);
2926 return;
2927 }
2928
2929 /*
2930 * Increment ->n_barrier_done to avoid duplicate work. Use
2931 * ACCESS_ONCE() to prevent the compiler from speculating
2932 * the increment to precede the early-exit check.
2933 */
2934 ACCESS_ONCE(rsp->n_barrier_done)++;
2935 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 1);
2936 _rcu_barrier_trace(rsp, "Inc1", -1, rsp->n_barrier_done);
2937 smp_mb(); /* Order ->n_barrier_done increment with below mechanism. */
2938
2939 /*
2940 * Initialize the count to one rather than to zero in order to
2941 * avoid a too-soon return to zero in case of a short grace period
2942 * (or preemption of this task). Exclude CPU-hotplug operations
2943 * to ensure that no offline CPU has callbacks queued.
2944 */
2945 init_completion(&rsp->barrier_completion);
2946 atomic_set(&rsp->barrier_cpu_count, 1);
2947 get_online_cpus();
2948
2949 /*
2950 * Force each CPU with callbacks to register a new callback.
2951 * When that callback is invoked, we will know that all of the
2952 * corresponding CPU's preceding callbacks have been invoked.
2953 */
2954 for_each_possible_cpu(cpu) {
2955 if (!cpu_online(cpu) && !rcu_is_nocb_cpu(cpu))
2956 continue;
2957 rdp = per_cpu_ptr(rsp->rda, cpu);
2958 if (rcu_is_nocb_cpu(cpu)) {
2959 _rcu_barrier_trace(rsp, "OnlineNoCB", cpu,
2960 rsp->n_barrier_done);
2961 atomic_inc(&rsp->barrier_cpu_count);
2962 __call_rcu(&rdp->barrier_head, rcu_barrier_callback,
2963 rsp, cpu, 0);
2964 } else if (ACCESS_ONCE(rdp->qlen)) {
2965 _rcu_barrier_trace(rsp, "OnlineQ", cpu,
2966 rsp->n_barrier_done);
2967 smp_call_function_single(cpu, rcu_barrier_func, rsp, 1);
2968 } else {
2969 _rcu_barrier_trace(rsp, "OnlineNQ", cpu,
2970 rsp->n_barrier_done);
2971 }
2972 }
2973 put_online_cpus();
2974
2975 /*
2976 * Now that we have an rcu_barrier_callback() callback on each
2977 * CPU, and thus each counted, remove the initial count.
2978 */
2979 if (atomic_dec_and_test(&rsp->barrier_cpu_count))
2980 complete(&rsp->barrier_completion);
2981
2982 /* Increment ->n_barrier_done to prevent duplicate work. */
2983 smp_mb(); /* Keep increment after above mechanism. */
2984 ACCESS_ONCE(rsp->n_barrier_done)++;
2985 WARN_ON_ONCE((rsp->n_barrier_done & 0x1) != 0);
2986 _rcu_barrier_trace(rsp, "Inc2", -1, rsp->n_barrier_done);
2987 smp_mb(); /* Keep increment before caller's subsequent code. */
2988
2989 /* Wait for all rcu_barrier_callback() callbacks to be invoked. */
2990 wait_for_completion(&rsp->barrier_completion);
2991
2992 /* Other rcu_barrier() invocations can now safely proceed. */
2993 mutex_unlock(&rsp->barrier_mutex);
2994}
2995
2996/**
2997 * rcu_barrier_bh - Wait until all in-flight call_rcu_bh() callbacks complete.
2998 */
2999void rcu_barrier_bh(void)
3000{
3001 _rcu_barrier(&rcu_bh_state);
3002}
3003EXPORT_SYMBOL_GPL(rcu_barrier_bh);
3004
3005/**
3006 * rcu_barrier_sched - Wait for in-flight call_rcu_sched() callbacks.
3007 */
3008void rcu_barrier_sched(void)
3009{
3010 _rcu_barrier(&rcu_sched_state);
3011}
3012EXPORT_SYMBOL_GPL(rcu_barrier_sched);
3013
3014/*
3015 * Do boot-time initialization of a CPU's per-CPU RCU data.
3016 */
3017static void __init
3018rcu_boot_init_percpu_data(int cpu, struct rcu_state *rsp)
3019{
3020 unsigned long flags;
3021 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3022 struct rcu_node *rnp = rcu_get_root(rsp);
3023
3024 /* Set up local state, ensuring consistent view of global state. */
3025 raw_spin_lock_irqsave(&rnp->lock, flags);
3026 rdp->grpmask = 1UL << (cpu - rdp->mynode->grplo);
3027 init_callback_list(rdp);
3028 rdp->qlen_lazy = 0;
3029 ACCESS_ONCE(rdp->qlen) = 0;
3030 rdp->dynticks = &per_cpu(rcu_dynticks, cpu);
3031 WARN_ON_ONCE(rdp->dynticks->dynticks_nesting != DYNTICK_TASK_EXIT_IDLE);
3032 WARN_ON_ONCE(atomic_read(&rdp->dynticks->dynticks) != 1);
3033 rdp->cpu = cpu;
3034 rdp->rsp = rsp;
3035 rcu_boot_init_nocb_percpu_data(rdp);
3036 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3037}
3038
3039/*
3040 * Initialize a CPU's per-CPU RCU data. Note that only one online or
3041 * offline event can be happening at a given time. Note also that we
3042 * can accept some slop in the rsp->completed access due to the fact
3043 * that this CPU cannot possibly have any RCU callbacks in flight yet.
3044 */
3045static void
3046rcu_init_percpu_data(int cpu, struct rcu_state *rsp, int preemptible)
3047{
3048 unsigned long flags;
3049 unsigned long mask;
3050 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
3051 struct rcu_node *rnp = rcu_get_root(rsp);
3052
3053 /* Exclude new grace periods. */
3054 mutex_lock(&rsp->onoff_mutex);
3055
3056 /* Set up local state, ensuring consistent view of global state. */
3057 raw_spin_lock_irqsave(&rnp->lock, flags);
3058 rdp->beenonline = 1; /* We have now been online. */
3059 rdp->preemptible = preemptible;
3060 rdp->qlen_last_fqs_check = 0;
3061 rdp->n_force_qs_snap = rsp->n_force_qs;
3062 rdp->blimit = blimit;
3063 init_callback_list(rdp); /* Re-enable callbacks on this CPU. */
3064 rdp->dynticks->dynticks_nesting = DYNTICK_TASK_EXIT_IDLE;
3065 rcu_sysidle_init_percpu_data(rdp->dynticks);
3066 atomic_set(&rdp->dynticks->dynticks,
3067 (atomic_read(&rdp->dynticks->dynticks) & ~0x1) + 1);
3068 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
3069
3070 /* Add CPU to rcu_node bitmasks. */
3071 rnp = rdp->mynode;
3072 mask = rdp->grpmask;
3073 do {
3074 /* Exclude any attempts to start a new GP on small systems. */
3075 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
3076 rnp->qsmaskinit |= mask;
3077 mask = rnp->grpmask;
3078 if (rnp == rdp->mynode) {
3079 /*
3080 * If there is a grace period in progress, we will
3081 * set up to wait for it next time we run the
3082 * RCU core code.
3083 */
3084 rdp->gpnum = rnp->completed;
3085 rdp->completed = rnp->completed;
3086 rdp->passed_quiesce = 0;
3087 rdp->qs_pending = 0;
3088 trace_rcu_grace_period(rsp->name, rdp->gpnum, TPS("cpuonl"));
3089 }
3090 raw_spin_unlock(&rnp->lock); /* irqs already disabled. */
3091 rnp = rnp->parent;
3092 } while (rnp != NULL && !(rnp->qsmaskinit & mask));
3093 local_irq_restore(flags);
3094
3095 mutex_unlock(&rsp->onoff_mutex);
3096}
3097
3098static void rcu_prepare_cpu(int cpu)
3099{
3100 struct rcu_state *rsp;
3101
3102 for_each_rcu_flavor(rsp)
3103 rcu_init_percpu_data(cpu, rsp,
3104 strcmp(rsp->name, "rcu_preempt") == 0);
3105}
3106
3107/*
3108 * Handle CPU online/offline notification events.
3109 */
3110static int rcu_cpu_notify(struct notifier_block *self,
3111 unsigned long action, void *hcpu)
3112{
3113 long cpu = (long)hcpu;
3114 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
3115 struct rcu_node *rnp = rdp->mynode;
3116 struct rcu_state *rsp;
3117
3118 trace_rcu_utilization(TPS("Start CPU hotplug"));
3119 switch (action) {
3120 case CPU_UP_PREPARE:
3121 case CPU_UP_PREPARE_FROZEN:
3122 rcu_prepare_cpu(cpu);
3123 rcu_prepare_kthreads(cpu);
3124 break;
3125 case CPU_ONLINE:
3126 case CPU_DOWN_FAILED:
3127 rcu_boost_kthread_setaffinity(rnp, -1);
3128 break;
3129 case CPU_DOWN_PREPARE:
3130 rcu_boost_kthread_setaffinity(rnp, cpu);
3131 break;
3132 case CPU_DYING:
3133 case CPU_DYING_FROZEN:
3134 for_each_rcu_flavor(rsp)
3135 rcu_cleanup_dying_cpu(rsp);
3136 break;
3137 case CPU_DEAD:
3138 case CPU_DEAD_FROZEN:
3139 case CPU_UP_CANCELED:
3140 case CPU_UP_CANCELED_FROZEN:
3141 for_each_rcu_flavor(rsp)
3142 rcu_cleanup_dead_cpu(cpu, rsp);
3143 break;
3144 default:
3145 break;
3146 }
3147 trace_rcu_utilization(TPS("End CPU hotplug"));
3148 return NOTIFY_OK;
3149}
3150
3151static int rcu_pm_notify(struct notifier_block *self,
3152 unsigned long action, void *hcpu)
3153{
3154 switch (action) {
3155 case PM_HIBERNATION_PREPARE:
3156 case PM_SUSPEND_PREPARE:
3157 if (nr_cpu_ids <= 256) /* Expediting bad for large systems. */
3158 rcu_expedited = 1;
3159 break;
3160 case PM_POST_HIBERNATION:
3161 case PM_POST_SUSPEND:
3162 rcu_expedited = 0;
3163 break;
3164 default:
3165 break;
3166 }
3167 return NOTIFY_OK;
3168}
3169
3170/*
3171 * Spawn the kthread that handles this RCU flavor's grace periods.
3172 */
3173static int __init rcu_spawn_gp_kthread(void)
3174{
3175 unsigned long flags;
3176 struct rcu_node *rnp;
3177 struct rcu_state *rsp;
3178 struct task_struct *t;
3179
3180 for_each_rcu_flavor(rsp) {
3181 t = kthread_run(rcu_gp_kthread, rsp, "%s", rsp->name);
3182 BUG_ON(IS_ERR(t));
3183 rnp = rcu_get_root(rsp);
3184 raw_spin_lock_irqsave(&rnp->lock, flags);
3185 rsp->gp_kthread = t;
3186 raw_spin_unlock_irqrestore(&rnp->lock, flags);
3187 rcu_spawn_nocb_kthreads(rsp);
3188 }
3189 return 0;
3190}
3191early_initcall(rcu_spawn_gp_kthread);
3192
3193/*
3194 * This function is invoked towards the end of the scheduler's initialization
3195 * process. Before this is called, the idle task might contain
3196 * RCU read-side critical sections (during which time, this idle
3197 * task is booting the system). After this function is called, the
3198 * idle tasks are prohibited from containing RCU read-side critical
3199 * sections. This function also enables RCU lockdep checking.
3200 */
3201void rcu_scheduler_starting(void)
3202{
3203 WARN_ON(num_online_cpus() != 1);
3204 WARN_ON(nr_context_switches() > 0);
3205 rcu_scheduler_active = 1;
3206}
3207
3208/*
3209 * Compute the per-level fanout, either using the exact fanout specified
3210 * or balancing the tree, depending on CONFIG_RCU_FANOUT_EXACT.
3211 */
3212#ifdef CONFIG_RCU_FANOUT_EXACT
3213static void __init rcu_init_levelspread(struct rcu_state *rsp)
3214{
3215 int i;
3216
3217 for (i = rcu_num_lvls - 1; i > 0; i--)
3218 rsp->levelspread[i] = CONFIG_RCU_FANOUT;
3219 rsp->levelspread[0] = rcu_fanout_leaf;
3220}
3221#else /* #ifdef CONFIG_RCU_FANOUT_EXACT */
3222static void __init rcu_init_levelspread(struct rcu_state *rsp)
3223{
3224 int ccur;
3225 int cprv;
3226 int i;
3227
3228 cprv = nr_cpu_ids;
3229 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3230 ccur = rsp->levelcnt[i];
3231 rsp->levelspread[i] = (cprv + ccur - 1) / ccur;
3232 cprv = ccur;
3233 }
3234}
3235#endif /* #else #ifdef CONFIG_RCU_FANOUT_EXACT */
3236
3237/*
3238 * Helper function for rcu_init() that initializes one rcu_state structure.
3239 */
3240static void __init rcu_init_one(struct rcu_state *rsp,
3241 struct rcu_data __percpu *rda)
3242{
3243 static char *buf[] = { "rcu_node_0",
3244 "rcu_node_1",
3245 "rcu_node_2",
3246 "rcu_node_3" }; /* Match MAX_RCU_LVLS */
3247 static char *fqs[] = { "rcu_node_fqs_0",
3248 "rcu_node_fqs_1",
3249 "rcu_node_fqs_2",
3250 "rcu_node_fqs_3" }; /* Match MAX_RCU_LVLS */
3251 int cpustride = 1;
3252 int i;
3253 int j;
3254 struct rcu_node *rnp;
3255
3256 BUILD_BUG_ON(MAX_RCU_LVLS > ARRAY_SIZE(buf)); /* Fix buf[] init! */
3257
3258 /* Silence gcc 4.8 warning about array index out of range. */
3259 if (rcu_num_lvls > RCU_NUM_LVLS)
3260 panic("rcu_init_one: rcu_num_lvls overflow");
3261
3262 /* Initialize the level-tracking arrays. */
3263
3264 for (i = 0; i < rcu_num_lvls; i++)
3265 rsp->levelcnt[i] = num_rcu_lvl[i];
3266 for (i = 1; i < rcu_num_lvls; i++)
3267 rsp->level[i] = rsp->level[i - 1] + rsp->levelcnt[i - 1];
3268 rcu_init_levelspread(rsp);
3269
3270 /* Initialize the elements themselves, starting from the leaves. */
3271
3272 for (i = rcu_num_lvls - 1; i >= 0; i--) {
3273 cpustride *= rsp->levelspread[i];
3274 rnp = rsp->level[i];
3275 for (j = 0; j < rsp->levelcnt[i]; j++, rnp++) {
3276 raw_spin_lock_init(&rnp->lock);
3277 lockdep_set_class_and_name(&rnp->lock,
3278 &rcu_node_class[i], buf[i]);
3279 raw_spin_lock_init(&rnp->fqslock);
3280 lockdep_set_class_and_name(&rnp->fqslock,
3281 &rcu_fqs_class[i], fqs[i]);
3282 rnp->gpnum = rsp->gpnum;
3283 rnp->completed = rsp->completed;
3284 rnp->qsmask = 0;
3285 rnp->qsmaskinit = 0;
3286 rnp->grplo = j * cpustride;
3287 rnp->grphi = (j + 1) * cpustride - 1;
3288 if (rnp->grphi >= NR_CPUS)
3289 rnp->grphi = NR_CPUS - 1;
3290 if (i == 0) {
3291 rnp->grpnum = 0;
3292 rnp->grpmask = 0;
3293 rnp->parent = NULL;
3294 } else {
3295 rnp->grpnum = j % rsp->levelspread[i - 1];
3296 rnp->grpmask = 1UL << rnp->grpnum;
3297 rnp->parent = rsp->level[i - 1] +
3298 j / rsp->levelspread[i - 1];
3299 }
3300 rnp->level = i;
3301 INIT_LIST_HEAD(&rnp->blkd_tasks);
3302 rcu_init_one_nocb(rnp);
3303 }
3304 }
3305
3306 rsp->rda = rda;
3307 init_waitqueue_head(&rsp->gp_wq);
3308 init_irq_work(&rsp->wakeup_work, rsp_wakeup);
3309 rnp = rsp->level[rcu_num_lvls - 1];
3310 for_each_possible_cpu(i) {
3311 while (i > rnp->grphi)
3312 rnp++;
3313 per_cpu_ptr(rsp->rda, i)->mynode = rnp;
3314 rcu_boot_init_percpu_data(i, rsp);
3315 }
3316 list_add(&rsp->flavors, &rcu_struct_flavors);
3317}
3318
3319/*
3320 * Compute the rcu_node tree geometry from kernel parameters. This cannot
3321 * replace the definitions in tree.h because those are needed to size
3322 * the ->node array in the rcu_state structure.
3323 */
3324static void __init rcu_init_geometry(void)
3325{
3326 ulong d;
3327 int i;
3328 int j;
3329 int n = nr_cpu_ids;
3330 int rcu_capacity[MAX_RCU_LVLS + 1];
3331
3332 /*
3333 * Initialize any unspecified boot parameters.
3334 * The default values of jiffies_till_first_fqs and
3335 * jiffies_till_next_fqs are set to the RCU_JIFFIES_TILL_FORCE_QS
3336 * value, which is a function of HZ, then adding one for each
3337 * RCU_JIFFIES_FQS_DIV CPUs that might be on the system.
3338 */
3339 d = RCU_JIFFIES_TILL_FORCE_QS + nr_cpu_ids / RCU_JIFFIES_FQS_DIV;
3340 if (jiffies_till_first_fqs == ULONG_MAX)
3341 jiffies_till_first_fqs = d;
3342 if (jiffies_till_next_fqs == ULONG_MAX)
3343 jiffies_till_next_fqs = d;
3344
3345 /* If the compile-time values are accurate, just leave. */
3346 if (rcu_fanout_leaf == CONFIG_RCU_FANOUT_LEAF &&
3347 nr_cpu_ids == NR_CPUS)
3348 return;
3349
3350 /*
3351 * Compute number of nodes that can be handled an rcu_node tree
3352 * with the given number of levels. Setting rcu_capacity[0] makes
3353 * some of the arithmetic easier.
3354 */
3355 rcu_capacity[0] = 1;
3356 rcu_capacity[1] = rcu_fanout_leaf;
3357 for (i = 2; i <= MAX_RCU_LVLS; i++)
3358 rcu_capacity[i] = rcu_capacity[i - 1] * CONFIG_RCU_FANOUT;
3359
3360 /*
3361 * The boot-time rcu_fanout_leaf parameter is only permitted
3362 * to increase the leaf-level fanout, not decrease it. Of course,
3363 * the leaf-level fanout cannot exceed the number of bits in
3364 * the rcu_node masks. Finally, the tree must be able to accommodate
3365 * the configured number of CPUs. Complain and fall back to the
3366 * compile-time values if these limits are exceeded.
3367 */
3368 if (rcu_fanout_leaf < CONFIG_RCU_FANOUT_LEAF ||
3369 rcu_fanout_leaf > sizeof(unsigned long) * 8 ||
3370 n > rcu_capacity[MAX_RCU_LVLS]) {
3371 WARN_ON(1);
3372 return;
3373 }
3374
3375 /* Calculate the number of rcu_nodes at each level of the tree. */
3376 for (i = 1; i <= MAX_RCU_LVLS; i++)
3377 if (n <= rcu_capacity[i]) {
3378 for (j = 0; j <= i; j++)
3379 num_rcu_lvl[j] =
3380 DIV_ROUND_UP(n, rcu_capacity[i - j]);
3381 rcu_num_lvls = i;
3382 for (j = i + 1; j <= MAX_RCU_LVLS; j++)
3383 num_rcu_lvl[j] = 0;
3384 break;
3385 }
3386
3387 /* Calculate the total number of rcu_node structures. */
3388 rcu_num_nodes = 0;
3389 for (i = 0; i <= MAX_RCU_LVLS; i++)
3390 rcu_num_nodes += num_rcu_lvl[i];
3391 rcu_num_nodes -= n;
3392}
3393
3394void __init rcu_init(void)
3395{
3396 int cpu;
3397
3398 rcu_bootup_announce();
3399 rcu_init_geometry();
3400 rcu_init_one(&rcu_bh_state, &rcu_bh_data);
3401 rcu_init_one(&rcu_sched_state, &rcu_sched_data);
3402 __rcu_init_preempt();
3403 open_softirq(RCU_SOFTIRQ, rcu_process_callbacks);
3404
3405 /*
3406 * We don't need protection against CPU-hotplug here because
3407 * this is called early in boot, before either interrupts
3408 * or the scheduler are operational.
3409 */
3410 cpu_notifier(rcu_cpu_notify, 0);
3411 pm_notifier(rcu_pm_notify, 0);
3412 for_each_online_cpu(cpu)
3413 rcu_cpu_notify(NULL, CPU_UP_PREPARE, (void *)(long)cpu);
3414}
3415
3416#include "tree_plugin.h"
diff --git a/kernel/rcu/tree.h b/kernel/rcu/tree.h
new file mode 100644
index 000000000000..52be957c9fe2
--- /dev/null
+++ b/kernel/rcu/tree.h
@@ -0,0 +1,585 @@
1/*
2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions.
4 *
5 * This program is free software; you can redistribute it and/or modify
6 * it under the terms of the GNU General Public License as published by
7 * the Free Software Foundation; either version 2 of the License, or
8 * (at your option) any later version.
9 *
10 * This program is distributed in the hope that it will be useful,
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
13 * GNU General Public License for more details.
14 *
15 * You should have received a copy of the GNU General Public License
16 * along with this program; if not, write to the Free Software
17 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
18 *
19 * Copyright IBM Corporation, 2008
20 *
21 * Author: Ingo Molnar <mingo@elte.hu>
22 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
23 */
24
25#include <linux/cache.h>
26#include <linux/spinlock.h>
27#include <linux/threads.h>
28#include <linux/cpumask.h>
29#include <linux/seqlock.h>
30#include <linux/irq_work.h>
31
32/*
33 * Define shape of hierarchy based on NR_CPUS, CONFIG_RCU_FANOUT, and
34 * CONFIG_RCU_FANOUT_LEAF.
35 * In theory, it should be possible to add more levels straightforwardly.
36 * In practice, this did work well going from three levels to four.
37 * Of course, your mileage may vary.
38 */
39#define MAX_RCU_LVLS 4
40#define RCU_FANOUT_1 (CONFIG_RCU_FANOUT_LEAF)
41#define RCU_FANOUT_2 (RCU_FANOUT_1 * CONFIG_RCU_FANOUT)
42#define RCU_FANOUT_3 (RCU_FANOUT_2 * CONFIG_RCU_FANOUT)
43#define RCU_FANOUT_4 (RCU_FANOUT_3 * CONFIG_RCU_FANOUT)
44
45#if NR_CPUS <= RCU_FANOUT_1
46# define RCU_NUM_LVLS 1
47# define NUM_RCU_LVL_0 1
48# define NUM_RCU_LVL_1 (NR_CPUS)
49# define NUM_RCU_LVL_2 0
50# define NUM_RCU_LVL_3 0
51# define NUM_RCU_LVL_4 0
52#elif NR_CPUS <= RCU_FANOUT_2
53# define RCU_NUM_LVLS 2
54# define NUM_RCU_LVL_0 1
55# define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
56# define NUM_RCU_LVL_2 (NR_CPUS)
57# define NUM_RCU_LVL_3 0
58# define NUM_RCU_LVL_4 0
59#elif NR_CPUS <= RCU_FANOUT_3
60# define RCU_NUM_LVLS 3
61# define NUM_RCU_LVL_0 1
62# define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
63# define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
64# define NUM_RCU_LVL_3 (NR_CPUS)
65# define NUM_RCU_LVL_4 0
66#elif NR_CPUS <= RCU_FANOUT_4
67# define RCU_NUM_LVLS 4
68# define NUM_RCU_LVL_0 1
69# define NUM_RCU_LVL_1 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_3)
70# define NUM_RCU_LVL_2 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_2)
71# define NUM_RCU_LVL_3 DIV_ROUND_UP(NR_CPUS, RCU_FANOUT_1)
72# define NUM_RCU_LVL_4 (NR_CPUS)
73#else
74# error "CONFIG_RCU_FANOUT insufficient for NR_CPUS"
75#endif /* #if (NR_CPUS) <= RCU_FANOUT_1 */
76
77#define RCU_SUM (NUM_RCU_LVL_0 + NUM_RCU_LVL_1 + NUM_RCU_LVL_2 + NUM_RCU_LVL_3 + NUM_RCU_LVL_4)
78#define NUM_RCU_NODES (RCU_SUM - NR_CPUS)
79
80extern int rcu_num_lvls;
81extern int rcu_num_nodes;
82
83/*
84 * Dynticks per-CPU state.
85 */
86struct rcu_dynticks {
87 long long dynticks_nesting; /* Track irq/process nesting level. */
88 /* Process level is worth LLONG_MAX/2. */
89 int dynticks_nmi_nesting; /* Track NMI nesting level. */
90 atomic_t dynticks; /* Even value for idle, else odd. */
91#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
92 long long dynticks_idle_nesting;
93 /* irq/process nesting level from idle. */
94 atomic_t dynticks_idle; /* Even value for idle, else odd. */
95 /* "Idle" excludes userspace execution. */
96 unsigned long dynticks_idle_jiffies;
97 /* End of last non-NMI non-idle period. */
98#endif /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
99#ifdef CONFIG_RCU_FAST_NO_HZ
100 bool all_lazy; /* Are all CPU's CBs lazy? */
101 unsigned long nonlazy_posted;
102 /* # times non-lazy CBs posted to CPU. */
103 unsigned long nonlazy_posted_snap;
104 /* idle-period nonlazy_posted snapshot. */
105 unsigned long last_accelerate;
106 /* Last jiffy CBs were accelerated. */
107 unsigned long last_advance_all;
108 /* Last jiffy CBs were all advanced. */
109 int tick_nohz_enabled_snap; /* Previously seen value from sysfs. */
110#endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
111};
112
113/* RCU's kthread states for tracing. */
114#define RCU_KTHREAD_STOPPED 0
115#define RCU_KTHREAD_RUNNING 1
116#define RCU_KTHREAD_WAITING 2
117#define RCU_KTHREAD_OFFCPU 3
118#define RCU_KTHREAD_YIELDING 4
119#define RCU_KTHREAD_MAX 4
120
121/*
122 * Definition for node within the RCU grace-period-detection hierarchy.
123 */
124struct rcu_node {
125 raw_spinlock_t lock; /* Root rcu_node's lock protects some */
126 /* rcu_state fields as well as following. */
127 unsigned long gpnum; /* Current grace period for this node. */
128 /* This will either be equal to or one */
129 /* behind the root rcu_node's gpnum. */
130 unsigned long completed; /* Last GP completed for this node. */
131 /* This will either be equal to or one */
132 /* behind the root rcu_node's gpnum. */
133 unsigned long qsmask; /* CPUs or groups that need to switch in */
134 /* order for current grace period to proceed.*/
135 /* In leaf rcu_node, each bit corresponds to */
136 /* an rcu_data structure, otherwise, each */
137 /* bit corresponds to a child rcu_node */
138 /* structure. */
139 unsigned long expmask; /* Groups that have ->blkd_tasks */
140 /* elements that need to drain to allow the */
141 /* current expedited grace period to */
142 /* complete (only for TREE_PREEMPT_RCU). */
143 unsigned long qsmaskinit;
144 /* Per-GP initial value for qsmask & expmask. */
145 unsigned long grpmask; /* Mask to apply to parent qsmask. */
146 /* Only one bit will be set in this mask. */
147 int grplo; /* lowest-numbered CPU or group here. */
148 int grphi; /* highest-numbered CPU or group here. */
149 u8 grpnum; /* CPU/group number for next level up. */
150 u8 level; /* root is at level 0. */
151 struct rcu_node *parent;
152 struct list_head blkd_tasks;
153 /* Tasks blocked in RCU read-side critical */
154 /* section. Tasks are placed at the head */
155 /* of this list and age towards the tail. */
156 struct list_head *gp_tasks;
157 /* Pointer to the first task blocking the */
158 /* current grace period, or NULL if there */
159 /* is no such task. */
160 struct list_head *exp_tasks;
161 /* Pointer to the first task blocking the */
162 /* current expedited grace period, or NULL */
163 /* if there is no such task. If there */
164 /* is no current expedited grace period, */
165 /* then there can cannot be any such task. */
166#ifdef CONFIG_RCU_BOOST
167 struct list_head *boost_tasks;
168 /* Pointer to first task that needs to be */
169 /* priority boosted, or NULL if no priority */
170 /* boosting is needed for this rcu_node */
171 /* structure. If there are no tasks */
172 /* queued on this rcu_node structure that */
173 /* are blocking the current grace period, */
174 /* there can be no such task. */
175 unsigned long boost_time;
176 /* When to start boosting (jiffies). */
177 struct task_struct *boost_kthread_task;
178 /* kthread that takes care of priority */
179 /* boosting for this rcu_node structure. */
180 unsigned int boost_kthread_status;
181 /* State of boost_kthread_task for tracing. */
182 unsigned long n_tasks_boosted;
183 /* Total number of tasks boosted. */
184 unsigned long n_exp_boosts;
185 /* Number of tasks boosted for expedited GP. */
186 unsigned long n_normal_boosts;
187 /* Number of tasks boosted for normal GP. */
188 unsigned long n_balk_blkd_tasks;
189 /* Refused to boost: no blocked tasks. */
190 unsigned long n_balk_exp_gp_tasks;
191 /* Refused to boost: nothing blocking GP. */
192 unsigned long n_balk_boost_tasks;
193 /* Refused to boost: already boosting. */
194 unsigned long n_balk_notblocked;
195 /* Refused to boost: RCU RS CS still running. */
196 unsigned long n_balk_notyet;
197 /* Refused to boost: not yet time. */
198 unsigned long n_balk_nos;
199 /* Refused to boost: not sure why, though. */
200 /* This can happen due to race conditions. */
201#endif /* #ifdef CONFIG_RCU_BOOST */
202#ifdef CONFIG_RCU_NOCB_CPU
203 wait_queue_head_t nocb_gp_wq[2];
204 /* Place for rcu_nocb_kthread() to wait GP. */
205#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
206 int need_future_gp[2];
207 /* Counts of upcoming no-CB GP requests. */
208 raw_spinlock_t fqslock ____cacheline_internodealigned_in_smp;
209} ____cacheline_internodealigned_in_smp;
210
211/*
212 * Do a full breadth-first scan of the rcu_node structures for the
213 * specified rcu_state structure.
214 */
215#define rcu_for_each_node_breadth_first(rsp, rnp) \
216 for ((rnp) = &(rsp)->node[0]; \
217 (rnp) < &(rsp)->node[rcu_num_nodes]; (rnp)++)
218
219/*
220 * Do a breadth-first scan of the non-leaf rcu_node structures for the
221 * specified rcu_state structure. Note that if there is a singleton
222 * rcu_node tree with but one rcu_node structure, this loop is a no-op.
223 */
224#define rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) \
225 for ((rnp) = &(rsp)->node[0]; \
226 (rnp) < (rsp)->level[rcu_num_lvls - 1]; (rnp)++)
227
228/*
229 * Scan the leaves of the rcu_node hierarchy for the specified rcu_state
230 * structure. Note that if there is a singleton rcu_node tree with but
231 * one rcu_node structure, this loop -will- visit the rcu_node structure.
232 * It is still a leaf node, even if it is also the root node.
233 */
234#define rcu_for_each_leaf_node(rsp, rnp) \
235 for ((rnp) = (rsp)->level[rcu_num_lvls - 1]; \
236 (rnp) < &(rsp)->node[rcu_num_nodes]; (rnp)++)
237
238/* Index values for nxttail array in struct rcu_data. */
239#define RCU_DONE_TAIL 0 /* Also RCU_WAIT head. */
240#define RCU_WAIT_TAIL 1 /* Also RCU_NEXT_READY head. */
241#define RCU_NEXT_READY_TAIL 2 /* Also RCU_NEXT head. */
242#define RCU_NEXT_TAIL 3
243#define RCU_NEXT_SIZE 4
244
245/* Per-CPU data for read-copy update. */
246struct rcu_data {
247 /* 1) quiescent-state and grace-period handling : */
248 unsigned long completed; /* Track rsp->completed gp number */
249 /* in order to detect GP end. */
250 unsigned long gpnum; /* Highest gp number that this CPU */
251 /* is aware of having started. */
252 bool passed_quiesce; /* User-mode/idle loop etc. */
253 bool qs_pending; /* Core waits for quiesc state. */
254 bool beenonline; /* CPU online at least once. */
255 bool preemptible; /* Preemptible RCU? */
256 struct rcu_node *mynode; /* This CPU's leaf of hierarchy */
257 unsigned long grpmask; /* Mask to apply to leaf qsmask. */
258#ifdef CONFIG_RCU_CPU_STALL_INFO
259 unsigned long ticks_this_gp; /* The number of scheduling-clock */
260 /* ticks this CPU has handled */
261 /* during and after the last grace */
262 /* period it is aware of. */
263#endif /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
264
265 /* 2) batch handling */
266 /*
267 * If nxtlist is not NULL, it is partitioned as follows.
268 * Any of the partitions might be empty, in which case the
269 * pointer to that partition will be equal to the pointer for
270 * the following partition. When the list is empty, all of
271 * the nxttail elements point to the ->nxtlist pointer itself,
272 * which in that case is NULL.
273 *
274 * [nxtlist, *nxttail[RCU_DONE_TAIL]):
275 * Entries that batch # <= ->completed
276 * The grace period for these entries has completed, and
277 * the other grace-period-completed entries may be moved
278 * here temporarily in rcu_process_callbacks().
279 * [*nxttail[RCU_DONE_TAIL], *nxttail[RCU_WAIT_TAIL]):
280 * Entries that batch # <= ->completed - 1: waiting for current GP
281 * [*nxttail[RCU_WAIT_TAIL], *nxttail[RCU_NEXT_READY_TAIL]):
282 * Entries known to have arrived before current GP ended
283 * [*nxttail[RCU_NEXT_READY_TAIL], *nxttail[RCU_NEXT_TAIL]):
284 * Entries that might have arrived after current GP ended
285 * Note that the value of *nxttail[RCU_NEXT_TAIL] will
286 * always be NULL, as this is the end of the list.
287 */
288 struct rcu_head *nxtlist;
289 struct rcu_head **nxttail[RCU_NEXT_SIZE];
290 unsigned long nxtcompleted[RCU_NEXT_SIZE];
291 /* grace periods for sublists. */
292 long qlen_lazy; /* # of lazy queued callbacks */
293 long qlen; /* # of queued callbacks, incl lazy */
294 long qlen_last_fqs_check;
295 /* qlen at last check for QS forcing */
296 unsigned long n_cbs_invoked; /* count of RCU cbs invoked. */
297 unsigned long n_nocbs_invoked; /* count of no-CBs RCU cbs invoked. */
298 unsigned long n_cbs_orphaned; /* RCU cbs orphaned by dying CPU */
299 unsigned long n_cbs_adopted; /* RCU cbs adopted from dying CPU */
300 unsigned long n_force_qs_snap;
301 /* did other CPU force QS recently? */
302 long blimit; /* Upper limit on a processed batch */
303
304 /* 3) dynticks interface. */
305 struct rcu_dynticks *dynticks; /* Shared per-CPU dynticks state. */
306 int dynticks_snap; /* Per-GP tracking for dynticks. */
307
308 /* 4) reasons this CPU needed to be kicked by force_quiescent_state */
309 unsigned long dynticks_fqs; /* Kicked due to dynticks idle. */
310 unsigned long offline_fqs; /* Kicked due to being offline. */
311
312 /* 5) __rcu_pending() statistics. */
313 unsigned long n_rcu_pending; /* rcu_pending() calls since boot. */
314 unsigned long n_rp_qs_pending;
315 unsigned long n_rp_report_qs;
316 unsigned long n_rp_cb_ready;
317 unsigned long n_rp_cpu_needs_gp;
318 unsigned long n_rp_gp_completed;
319 unsigned long n_rp_gp_started;
320 unsigned long n_rp_need_nothing;
321
322 /* 6) _rcu_barrier() and OOM callbacks. */
323 struct rcu_head barrier_head;
324#ifdef CONFIG_RCU_FAST_NO_HZ
325 struct rcu_head oom_head;
326#endif /* #ifdef CONFIG_RCU_FAST_NO_HZ */
327
328 /* 7) Callback offloading. */
329#ifdef CONFIG_RCU_NOCB_CPU
330 struct rcu_head *nocb_head; /* CBs waiting for kthread. */
331 struct rcu_head **nocb_tail;
332 atomic_long_t nocb_q_count; /* # CBs waiting for kthread */
333 atomic_long_t nocb_q_count_lazy; /* (approximate). */
334 int nocb_p_count; /* # CBs being invoked by kthread */
335 int nocb_p_count_lazy; /* (approximate). */
336 wait_queue_head_t nocb_wq; /* For nocb kthreads to sleep on. */
337 struct task_struct *nocb_kthread;
338#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
339
340 /* 8) RCU CPU stall data. */
341#ifdef CONFIG_RCU_CPU_STALL_INFO
342 unsigned int softirq_snap; /* Snapshot of softirq activity. */
343#endif /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
344
345 int cpu;
346 struct rcu_state *rsp;
347};
348
349/* Values for fqs_state field in struct rcu_state. */
350#define RCU_GP_IDLE 0 /* No grace period in progress. */
351#define RCU_GP_INIT 1 /* Grace period being initialized. */
352#define RCU_SAVE_DYNTICK 2 /* Need to scan dyntick state. */
353#define RCU_FORCE_QS 3 /* Need to force quiescent state. */
354#define RCU_SIGNAL_INIT RCU_SAVE_DYNTICK
355
356#define RCU_JIFFIES_TILL_FORCE_QS (1 + (HZ > 250) + (HZ > 500))
357 /* For jiffies_till_first_fqs and */
358 /* and jiffies_till_next_fqs. */
359
360#define RCU_JIFFIES_FQS_DIV 256 /* Very large systems need more */
361 /* delay between bouts of */
362 /* quiescent-state forcing. */
363
364#define RCU_STALL_RAT_DELAY 2 /* Allow other CPUs time to take */
365 /* at least one scheduling clock */
366 /* irq before ratting on them. */
367
368#define rcu_wait(cond) \
369do { \
370 for (;;) { \
371 set_current_state(TASK_INTERRUPTIBLE); \
372 if (cond) \
373 break; \
374 schedule(); \
375 } \
376 __set_current_state(TASK_RUNNING); \
377} while (0)
378
379/*
380 * RCU global state, including node hierarchy. This hierarchy is
381 * represented in "heap" form in a dense array. The root (first level)
382 * of the hierarchy is in ->node[0] (referenced by ->level[0]), the second
383 * level in ->node[1] through ->node[m] (->node[1] referenced by ->level[1]),
384 * and the third level in ->node[m+1] and following (->node[m+1] referenced
385 * by ->level[2]). The number of levels is determined by the number of
386 * CPUs and by CONFIG_RCU_FANOUT. Small systems will have a "hierarchy"
387 * consisting of a single rcu_node.
388 */
389struct rcu_state {
390 struct rcu_node node[NUM_RCU_NODES]; /* Hierarchy. */
391 struct rcu_node *level[RCU_NUM_LVLS]; /* Hierarchy levels. */
392 u32 levelcnt[MAX_RCU_LVLS + 1]; /* # nodes in each level. */
393 u8 levelspread[RCU_NUM_LVLS]; /* kids/node in each level. */
394 struct rcu_data __percpu *rda; /* pointer of percu rcu_data. */
395 void (*call)(struct rcu_head *head, /* call_rcu() flavor. */
396 void (*func)(struct rcu_head *head));
397
398 /* The following fields are guarded by the root rcu_node's lock. */
399
400 u8 fqs_state ____cacheline_internodealigned_in_smp;
401 /* Force QS state. */
402 u8 boost; /* Subject to priority boost. */
403 unsigned long gpnum; /* Current gp number. */
404 unsigned long completed; /* # of last completed gp. */
405 struct task_struct *gp_kthread; /* Task for grace periods. */
406 wait_queue_head_t gp_wq; /* Where GP task waits. */
407 int gp_flags; /* Commands for GP task. */
408
409 /* End of fields guarded by root rcu_node's lock. */
410
411 raw_spinlock_t orphan_lock ____cacheline_internodealigned_in_smp;
412 /* Protect following fields. */
413 struct rcu_head *orphan_nxtlist; /* Orphaned callbacks that */
414 /* need a grace period. */
415 struct rcu_head **orphan_nxttail; /* Tail of above. */
416 struct rcu_head *orphan_donelist; /* Orphaned callbacks that */
417 /* are ready to invoke. */
418 struct rcu_head **orphan_donetail; /* Tail of above. */
419 long qlen_lazy; /* Number of lazy callbacks. */
420 long qlen; /* Total number of callbacks. */
421 /* End of fields guarded by orphan_lock. */
422
423 struct mutex onoff_mutex; /* Coordinate hotplug & GPs. */
424
425 struct mutex barrier_mutex; /* Guards barrier fields. */
426 atomic_t barrier_cpu_count; /* # CPUs waiting on. */
427 struct completion barrier_completion; /* Wake at barrier end. */
428 unsigned long n_barrier_done; /* ++ at start and end of */
429 /* _rcu_barrier(). */
430 /* End of fields guarded by barrier_mutex. */
431
432 atomic_long_t expedited_start; /* Starting ticket. */
433 atomic_long_t expedited_done; /* Done ticket. */
434 atomic_long_t expedited_wrap; /* # near-wrap incidents. */
435 atomic_long_t expedited_tryfail; /* # acquisition failures. */
436 atomic_long_t expedited_workdone1; /* # done by others #1. */
437 atomic_long_t expedited_workdone2; /* # done by others #2. */
438 atomic_long_t expedited_normal; /* # fallbacks to normal. */
439 atomic_long_t expedited_stoppedcpus; /* # successful stop_cpus. */
440 atomic_long_t expedited_done_tries; /* # tries to update _done. */
441 atomic_long_t expedited_done_lost; /* # times beaten to _done. */
442 atomic_long_t expedited_done_exit; /* # times exited _done loop. */
443
444 unsigned long jiffies_force_qs; /* Time at which to invoke */
445 /* force_quiescent_state(). */
446 unsigned long n_force_qs; /* Number of calls to */
447 /* force_quiescent_state(). */
448 unsigned long n_force_qs_lh; /* ~Number of calls leaving */
449 /* due to lock unavailable. */
450 unsigned long n_force_qs_ngp; /* Number of calls leaving */
451 /* due to no GP active. */
452 unsigned long gp_start; /* Time at which GP started, */
453 /* but in jiffies. */
454 unsigned long jiffies_stall; /* Time at which to check */
455 /* for CPU stalls. */
456 unsigned long gp_max; /* Maximum GP duration in */
457 /* jiffies. */
458 const char *name; /* Name of structure. */
459 char abbr; /* Abbreviated name. */
460 struct list_head flavors; /* List of RCU flavors. */
461 struct irq_work wakeup_work; /* Postponed wakeups */
462};
463
464/* Values for rcu_state structure's gp_flags field. */
465#define RCU_GP_FLAG_INIT 0x1 /* Need grace-period initialization. */
466#define RCU_GP_FLAG_FQS 0x2 /* Need grace-period quiescent-state forcing. */
467
468extern struct list_head rcu_struct_flavors;
469
470/* Sequence through rcu_state structures for each RCU flavor. */
471#define for_each_rcu_flavor(rsp) \
472 list_for_each_entry((rsp), &rcu_struct_flavors, flavors)
473
474/* Return values for rcu_preempt_offline_tasks(). */
475
476#define RCU_OFL_TASKS_NORM_GP 0x1 /* Tasks blocking normal */
477 /* GP were moved to root. */
478#define RCU_OFL_TASKS_EXP_GP 0x2 /* Tasks blocking expedited */
479 /* GP were moved to root. */
480
481/*
482 * RCU implementation internal declarations:
483 */
484extern struct rcu_state rcu_sched_state;
485DECLARE_PER_CPU(struct rcu_data, rcu_sched_data);
486
487extern struct rcu_state rcu_bh_state;
488DECLARE_PER_CPU(struct rcu_data, rcu_bh_data);
489
490#ifdef CONFIG_TREE_PREEMPT_RCU
491extern struct rcu_state rcu_preempt_state;
492DECLARE_PER_CPU(struct rcu_data, rcu_preempt_data);
493#endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */
494
495#ifdef CONFIG_RCU_BOOST
496DECLARE_PER_CPU(unsigned int, rcu_cpu_kthread_status);
497DECLARE_PER_CPU(int, rcu_cpu_kthread_cpu);
498DECLARE_PER_CPU(unsigned int, rcu_cpu_kthread_loops);
499DECLARE_PER_CPU(char, rcu_cpu_has_work);
500#endif /* #ifdef CONFIG_RCU_BOOST */
501
502#ifndef RCU_TREE_NONCORE
503
504/* Forward declarations for rcutree_plugin.h */
505static void rcu_bootup_announce(void);
506long rcu_batches_completed(void);
507static void rcu_preempt_note_context_switch(int cpu);
508static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp);
509#ifdef CONFIG_HOTPLUG_CPU
510static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp,
511 unsigned long flags);
512#endif /* #ifdef CONFIG_HOTPLUG_CPU */
513static void rcu_print_detail_task_stall(struct rcu_state *rsp);
514static int rcu_print_task_stall(struct rcu_node *rnp);
515static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp);
516#ifdef CONFIG_HOTPLUG_CPU
517static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
518 struct rcu_node *rnp,
519 struct rcu_data *rdp);
520#endif /* #ifdef CONFIG_HOTPLUG_CPU */
521static void rcu_preempt_check_callbacks(int cpu);
522void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu));
523#if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_TREE_PREEMPT_RCU)
524static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
525 bool wake);
526#endif /* #if defined(CONFIG_HOTPLUG_CPU) || defined(CONFIG_TREE_PREEMPT_RCU) */
527static void __init __rcu_init_preempt(void);
528static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags);
529static void rcu_preempt_boost_start_gp(struct rcu_node *rnp);
530static void invoke_rcu_callbacks_kthread(void);
531static bool rcu_is_callbacks_kthread(void);
532#ifdef CONFIG_RCU_BOOST
533static void rcu_preempt_do_callbacks(void);
534static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
535 struct rcu_node *rnp);
536#endif /* #ifdef CONFIG_RCU_BOOST */
537static void rcu_prepare_kthreads(int cpu);
538static void rcu_cleanup_after_idle(int cpu);
539static void rcu_prepare_for_idle(int cpu);
540static void rcu_idle_count_callbacks_posted(void);
541static void print_cpu_stall_info_begin(void);
542static void print_cpu_stall_info(struct rcu_state *rsp, int cpu);
543static void print_cpu_stall_info_end(void);
544static void zero_cpu_stall_ticks(struct rcu_data *rdp);
545static void increment_cpu_stall_ticks(void);
546static int rcu_nocb_needs_gp(struct rcu_state *rsp);
547static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq);
548static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp);
549static void rcu_init_one_nocb(struct rcu_node *rnp);
550static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
551 bool lazy);
552static bool rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
553 struct rcu_data *rdp);
554static void rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp);
555static void rcu_spawn_nocb_kthreads(struct rcu_state *rsp);
556static void rcu_kick_nohz_cpu(int cpu);
557static bool init_nocb_callback_list(struct rcu_data *rdp);
558static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq);
559static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq);
560static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
561 unsigned long *maxj);
562static bool is_sysidle_rcu_state(struct rcu_state *rsp);
563static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
564 unsigned long maxj);
565static void rcu_bind_gp_kthread(void);
566static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp);
567
568#endif /* #ifndef RCU_TREE_NONCORE */
569
570#ifdef CONFIG_RCU_TRACE
571#ifdef CONFIG_RCU_NOCB_CPU
572/* Sum up queue lengths for tracing. */
573static inline void rcu_nocb_q_lengths(struct rcu_data *rdp, long *ql, long *qll)
574{
575 *ql = atomic_long_read(&rdp->nocb_q_count) + rdp->nocb_p_count;
576 *qll = atomic_long_read(&rdp->nocb_q_count_lazy) + rdp->nocb_p_count_lazy;
577}
578#else /* #ifdef CONFIG_RCU_NOCB_CPU */
579static inline void rcu_nocb_q_lengths(struct rcu_data *rdp, long *ql, long *qll)
580{
581 *ql = 0;
582 *qll = 0;
583}
584#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
585#endif /* #ifdef CONFIG_RCU_TRACE */
diff --git a/kernel/rcu/tree_plugin.h b/kernel/rcu/tree_plugin.h
new file mode 100644
index 000000000000..3822ac0c4b27
--- /dev/null
+++ b/kernel/rcu/tree_plugin.h
@@ -0,0 +1,2831 @@
1/*
2 * Read-Copy Update mechanism for mutual exclusion (tree-based version)
3 * Internal non-public definitions that provide either classic
4 * or preemptible semantics.
5 *
6 * This program is free software; you can redistribute it and/or modify
7 * it under the terms of the GNU General Public License as published by
8 * the Free Software Foundation; either version 2 of the License, or
9 * (at your option) any later version.
10 *
11 * This program is distributed in the hope that it will be useful,
12 * but WITHOUT ANY WARRANTY; without even the implied warranty of
13 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
14 * GNU General Public License for more details.
15 *
16 * You should have received a copy of the GNU General Public License
17 * along with this program; if not, write to the Free Software
18 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
19 *
20 * Copyright Red Hat, 2009
21 * Copyright IBM Corporation, 2009
22 *
23 * Author: Ingo Molnar <mingo@elte.hu>
24 * Paul E. McKenney <paulmck@linux.vnet.ibm.com>
25 */
26
27#include <linux/delay.h>
28#include <linux/gfp.h>
29#include <linux/oom.h>
30#include <linux/smpboot.h>
31#include "../time/tick-internal.h"
32
33#define RCU_KTHREAD_PRIO 1
34
35#ifdef CONFIG_RCU_BOOST
36#define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO
37#else
38#define RCU_BOOST_PRIO RCU_KTHREAD_PRIO
39#endif
40
41#ifdef CONFIG_RCU_NOCB_CPU
42static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */
43static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */
44static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */
45static char __initdata nocb_buf[NR_CPUS * 5];
46#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
47
48/*
49 * Check the RCU kernel configuration parameters and print informative
50 * messages about anything out of the ordinary. If you like #ifdef, you
51 * will love this function.
52 */
53static void __init rcu_bootup_announce_oddness(void)
54{
55#ifdef CONFIG_RCU_TRACE
56 pr_info("\tRCU debugfs-based tracing is enabled.\n");
57#endif
58#if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32)
59 pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n",
60 CONFIG_RCU_FANOUT);
61#endif
62#ifdef CONFIG_RCU_FANOUT_EXACT
63 pr_info("\tHierarchical RCU autobalancing is disabled.\n");
64#endif
65#ifdef CONFIG_RCU_FAST_NO_HZ
66 pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n");
67#endif
68#ifdef CONFIG_PROVE_RCU
69 pr_info("\tRCU lockdep checking is enabled.\n");
70#endif
71#ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE
72 pr_info("\tRCU torture testing starts during boot.\n");
73#endif
74#if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE)
75 pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n");
76#endif
77#if defined(CONFIG_RCU_CPU_STALL_INFO)
78 pr_info("\tAdditional per-CPU info printed with stalls.\n");
79#endif
80#if NUM_RCU_LVL_4 != 0
81 pr_info("\tFour-level hierarchy is enabled.\n");
82#endif
83 if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF)
84 pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf);
85 if (nr_cpu_ids != NR_CPUS)
86 pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids);
87#ifdef CONFIG_RCU_NOCB_CPU
88#ifndef CONFIG_RCU_NOCB_CPU_NONE
89 if (!have_rcu_nocb_mask) {
90 zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL);
91 have_rcu_nocb_mask = true;
92 }
93#ifdef CONFIG_RCU_NOCB_CPU_ZERO
94 pr_info("\tOffload RCU callbacks from CPU 0\n");
95 cpumask_set_cpu(0, rcu_nocb_mask);
96#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */
97#ifdef CONFIG_RCU_NOCB_CPU_ALL
98 pr_info("\tOffload RCU callbacks from all CPUs\n");
99 cpumask_copy(rcu_nocb_mask, cpu_possible_mask);
100#endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */
101#endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */
102 if (have_rcu_nocb_mask) {
103 if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) {
104 pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n");
105 cpumask_and(rcu_nocb_mask, cpu_possible_mask,
106 rcu_nocb_mask);
107 }
108 cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask);
109 pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf);
110 if (rcu_nocb_poll)
111 pr_info("\tPoll for callbacks from no-CBs CPUs.\n");
112 }
113#endif /* #ifdef CONFIG_RCU_NOCB_CPU */
114}
115
116#ifdef CONFIG_TREE_PREEMPT_RCU
117
118RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu);
119static struct rcu_state *rcu_state = &rcu_preempt_state;
120
121static int rcu_preempted_readers_exp(struct rcu_node *rnp);
122
123/*
124 * Tell them what RCU they are running.
125 */
126static void __init rcu_bootup_announce(void)
127{
128 pr_info("Preemptible hierarchical RCU implementation.\n");
129 rcu_bootup_announce_oddness();
130}
131
132/*
133 * Return the number of RCU-preempt batches processed thus far
134 * for debug and statistics.
135 */
136long rcu_batches_completed_preempt(void)
137{
138 return rcu_preempt_state.completed;
139}
140EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
141
142/*
143 * Return the number of RCU batches processed thus far for debug & stats.
144 */
145long rcu_batches_completed(void)
146{
147 return rcu_batches_completed_preempt();
148}
149EXPORT_SYMBOL_GPL(rcu_batches_completed);
150
151/*
152 * Force a quiescent state for preemptible RCU.
153 */
154void rcu_force_quiescent_state(void)
155{
156 force_quiescent_state(&rcu_preempt_state);
157}
158EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
159
160/*
161 * Record a preemptible-RCU quiescent state for the specified CPU. Note
162 * that this just means that the task currently running on the CPU is
163 * not in a quiescent state. There might be any number of tasks blocked
164 * while in an RCU read-side critical section.
165 *
166 * Unlike the other rcu_*_qs() functions, callers to this function
167 * must disable irqs in order to protect the assignment to
168 * ->rcu_read_unlock_special.
169 */
170static void rcu_preempt_qs(int cpu)
171{
172 struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
173
174 if (rdp->passed_quiesce == 0)
175 trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs"));
176 rdp->passed_quiesce = 1;
177 current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
178}
179
180/*
181 * We have entered the scheduler, and the current task might soon be
182 * context-switched away from. If this task is in an RCU read-side
183 * critical section, we will no longer be able to rely on the CPU to
184 * record that fact, so we enqueue the task on the blkd_tasks list.
185 * The task will dequeue itself when it exits the outermost enclosing
186 * RCU read-side critical section. Therefore, the current grace period
187 * cannot be permitted to complete until the blkd_tasks list entries
188 * predating the current grace period drain, in other words, until
189 * rnp->gp_tasks becomes NULL.
190 *
191 * Caller must disable preemption.
192 */
193static void rcu_preempt_note_context_switch(int cpu)
194{
195 struct task_struct *t = current;
196 unsigned long flags;
197 struct rcu_data *rdp;
198 struct rcu_node *rnp;
199
200 if (t->rcu_read_lock_nesting > 0 &&
201 (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
202
203 /* Possibly blocking in an RCU read-side critical section. */
204 rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu);
205 rnp = rdp->mynode;
206 raw_spin_lock_irqsave(&rnp->lock, flags);
207 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
208 t->rcu_blocked_node = rnp;
209
210 /*
211 * If this CPU has already checked in, then this task
212 * will hold up the next grace period rather than the
213 * current grace period. Queue the task accordingly.
214 * If the task is queued for the current grace period
215 * (i.e., this CPU has not yet passed through a quiescent
216 * state for the current grace period), then as long
217 * as that task remains queued, the current grace period
218 * cannot end. Note that there is some uncertainty as
219 * to exactly when the current grace period started.
220 * We take a conservative approach, which can result
221 * in unnecessarily waiting on tasks that started very
222 * slightly after the current grace period began. C'est
223 * la vie!!!
224 *
225 * But first, note that the current CPU must still be
226 * on line!
227 */
228 WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0);
229 WARN_ON_ONCE(!list_empty(&t->rcu_node_entry));
230 if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) {
231 list_add(&t->rcu_node_entry, rnp->gp_tasks->prev);
232 rnp->gp_tasks = &t->rcu_node_entry;
233#ifdef CONFIG_RCU_BOOST
234 if (rnp->boost_tasks != NULL)
235 rnp->boost_tasks = rnp->gp_tasks;
236#endif /* #ifdef CONFIG_RCU_BOOST */
237 } else {
238 list_add(&t->rcu_node_entry, &rnp->blkd_tasks);
239 if (rnp->qsmask & rdp->grpmask)
240 rnp->gp_tasks = &t->rcu_node_entry;
241 }
242 trace_rcu_preempt_task(rdp->rsp->name,
243 t->pid,
244 (rnp->qsmask & rdp->grpmask)
245 ? rnp->gpnum
246 : rnp->gpnum + 1);
247 raw_spin_unlock_irqrestore(&rnp->lock, flags);
248 } else if (t->rcu_read_lock_nesting < 0 &&
249 t->rcu_read_unlock_special) {
250
251 /*
252 * Complete exit from RCU read-side critical section on
253 * behalf of preempted instance of __rcu_read_unlock().
254 */
255 rcu_read_unlock_special(t);
256 }
257
258 /*
259 * Either we were not in an RCU read-side critical section to
260 * begin with, or we have now recorded that critical section
261 * globally. Either way, we can now note a quiescent state
262 * for this CPU. Again, if we were in an RCU read-side critical
263 * section, and if that critical section was blocking the current
264 * grace period, then the fact that the task has been enqueued
265 * means that we continue to block the current grace period.
266 */
267 local_irq_save(flags);
268 rcu_preempt_qs(cpu);
269 local_irq_restore(flags);
270}
271
272/*
273 * Check for preempted RCU readers blocking the current grace period
274 * for the specified rcu_node structure. If the caller needs a reliable
275 * answer, it must hold the rcu_node's ->lock.
276 */
277static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
278{
279 return rnp->gp_tasks != NULL;
280}
281
282/*
283 * Record a quiescent state for all tasks that were previously queued
284 * on the specified rcu_node structure and that were blocking the current
285 * RCU grace period. The caller must hold the specified rnp->lock with
286 * irqs disabled, and this lock is released upon return, but irqs remain
287 * disabled.
288 */
289static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
290 __releases(rnp->lock)
291{
292 unsigned long mask;
293 struct rcu_node *rnp_p;
294
295 if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) {
296 raw_spin_unlock_irqrestore(&rnp->lock, flags);
297 return; /* Still need more quiescent states! */
298 }
299
300 rnp_p = rnp->parent;
301 if (rnp_p == NULL) {
302 /*
303 * Either there is only one rcu_node in the tree,
304 * or tasks were kicked up to root rcu_node due to
305 * CPUs going offline.
306 */
307 rcu_report_qs_rsp(&rcu_preempt_state, flags);
308 return;
309 }
310
311 /* Report up the rest of the hierarchy. */
312 mask = rnp->grpmask;
313 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
314 raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */
315 rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags);
316}
317
318/*
319 * Advance a ->blkd_tasks-list pointer to the next entry, instead
320 * returning NULL if at the end of the list.
321 */
322static struct list_head *rcu_next_node_entry(struct task_struct *t,
323 struct rcu_node *rnp)
324{
325 struct list_head *np;
326
327 np = t->rcu_node_entry.next;
328 if (np == &rnp->blkd_tasks)
329 np = NULL;
330 return np;
331}
332
333/*
334 * Handle special cases during rcu_read_unlock(), such as needing to
335 * notify RCU core processing or task having blocked during the RCU
336 * read-side critical section.
337 */
338void rcu_read_unlock_special(struct task_struct *t)
339{
340 int empty;
341 int empty_exp;
342 int empty_exp_now;
343 unsigned long flags;
344 struct list_head *np;
345#ifdef CONFIG_RCU_BOOST
346 struct rt_mutex *rbmp = NULL;
347#endif /* #ifdef CONFIG_RCU_BOOST */
348 struct rcu_node *rnp;
349 int special;
350
351 /* NMI handlers cannot block and cannot safely manipulate state. */
352 if (in_nmi())
353 return;
354
355 local_irq_save(flags);
356
357 /*
358 * If RCU core is waiting for this CPU to exit critical section,
359 * let it know that we have done so.
360 */
361 special = t->rcu_read_unlock_special;
362 if (special & RCU_READ_UNLOCK_NEED_QS) {
363 rcu_preempt_qs(smp_processor_id());
364 }
365
366 /* Hardware IRQ handlers cannot block. */
367 if (in_irq() || in_serving_softirq()) {
368 local_irq_restore(flags);
369 return;
370 }
371
372 /* Clean up if blocked during RCU read-side critical section. */
373 if (special & RCU_READ_UNLOCK_BLOCKED) {
374 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
375
376 /*
377 * Remove this task from the list it blocked on. The
378 * task can migrate while we acquire the lock, but at
379 * most one time. So at most two passes through loop.
380 */
381 for (;;) {
382 rnp = t->rcu_blocked_node;
383 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
384 if (rnp == t->rcu_blocked_node)
385 break;
386 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
387 }
388 empty = !rcu_preempt_blocked_readers_cgp(rnp);
389 empty_exp = !rcu_preempted_readers_exp(rnp);
390 smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */
391 np = rcu_next_node_entry(t, rnp);
392 list_del_init(&t->rcu_node_entry);
393 t->rcu_blocked_node = NULL;
394 trace_rcu_unlock_preempted_task(TPS("rcu_preempt"),
395 rnp->gpnum, t->pid);
396 if (&t->rcu_node_entry == rnp->gp_tasks)
397 rnp->gp_tasks = np;
398 if (&t->rcu_node_entry == rnp->exp_tasks)
399 rnp->exp_tasks = np;
400#ifdef CONFIG_RCU_BOOST
401 if (&t->rcu_node_entry == rnp->boost_tasks)
402 rnp->boost_tasks = np;
403 /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */
404 if (t->rcu_boost_mutex) {
405 rbmp = t->rcu_boost_mutex;
406 t->rcu_boost_mutex = NULL;
407 }
408#endif /* #ifdef CONFIG_RCU_BOOST */
409
410 /*
411 * If this was the last task on the current list, and if
412 * we aren't waiting on any CPUs, report the quiescent state.
413 * Note that rcu_report_unblock_qs_rnp() releases rnp->lock,
414 * so we must take a snapshot of the expedited state.
415 */
416 empty_exp_now = !rcu_preempted_readers_exp(rnp);
417 if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) {
418 trace_rcu_quiescent_state_report(TPS("preempt_rcu"),
419 rnp->gpnum,
420 0, rnp->qsmask,
421 rnp->level,
422 rnp->grplo,
423 rnp->grphi,
424 !!rnp->gp_tasks);
425 rcu_report_unblock_qs_rnp(rnp, flags);
426 } else {
427 raw_spin_unlock_irqrestore(&rnp->lock, flags);
428 }
429
430#ifdef CONFIG_RCU_BOOST
431 /* Unboost if we were boosted. */
432 if (rbmp)
433 rt_mutex_unlock(rbmp);
434#endif /* #ifdef CONFIG_RCU_BOOST */
435
436 /*
437 * If this was the last task on the expedited lists,
438 * then we need to report up the rcu_node hierarchy.
439 */
440 if (!empty_exp && empty_exp_now)
441 rcu_report_exp_rnp(&rcu_preempt_state, rnp, true);
442 } else {
443 local_irq_restore(flags);
444 }
445}
446
447#ifdef CONFIG_RCU_CPU_STALL_VERBOSE
448
449/*
450 * Dump detailed information for all tasks blocking the current RCU
451 * grace period on the specified rcu_node structure.
452 */
453static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp)
454{
455 unsigned long flags;
456 struct task_struct *t;
457
458 raw_spin_lock_irqsave(&rnp->lock, flags);
459 if (!rcu_preempt_blocked_readers_cgp(rnp)) {
460 raw_spin_unlock_irqrestore(&rnp->lock, flags);
461 return;
462 }
463 t = list_entry(rnp->gp_tasks,
464 struct task_struct, rcu_node_entry);
465 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry)
466 sched_show_task(t);
467 raw_spin_unlock_irqrestore(&rnp->lock, flags);
468}
469
470/*
471 * Dump detailed information for all tasks blocking the current RCU
472 * grace period.
473 */
474static void rcu_print_detail_task_stall(struct rcu_state *rsp)
475{
476 struct rcu_node *rnp = rcu_get_root(rsp);
477
478 rcu_print_detail_task_stall_rnp(rnp);
479 rcu_for_each_leaf_node(rsp, rnp)
480 rcu_print_detail_task_stall_rnp(rnp);
481}
482
483#else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
484
485static void rcu_print_detail_task_stall(struct rcu_state *rsp)
486{
487}
488
489#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */
490
491#ifdef CONFIG_RCU_CPU_STALL_INFO
492
493static void rcu_print_task_stall_begin(struct rcu_node *rnp)
494{
495 pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):",
496 rnp->level, rnp->grplo, rnp->grphi);
497}
498
499static void rcu_print_task_stall_end(void)
500{
501 pr_cont("\n");
502}
503
504#else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
505
506static void rcu_print_task_stall_begin(struct rcu_node *rnp)
507{
508}
509
510static void rcu_print_task_stall_end(void)
511{
512}
513
514#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
515
516/*
517 * Scan the current list of tasks blocked within RCU read-side critical
518 * sections, printing out the tid of each.
519 */
520static int rcu_print_task_stall(struct rcu_node *rnp)
521{
522 struct task_struct *t;
523 int ndetected = 0;
524
525 if (!rcu_preempt_blocked_readers_cgp(rnp))
526 return 0;
527 rcu_print_task_stall_begin(rnp);
528 t = list_entry(rnp->gp_tasks,
529 struct task_struct, rcu_node_entry);
530 list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) {
531 pr_cont(" P%d", t->pid);
532 ndetected++;
533 }
534 rcu_print_task_stall_end();
535 return ndetected;
536}
537
538/*
539 * Check that the list of blocked tasks for the newly completed grace
540 * period is in fact empty. It is a serious bug to complete a grace
541 * period that still has RCU readers blocked! This function must be
542 * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock
543 * must be held by the caller.
544 *
545 * Also, if there are blocked tasks on the list, they automatically
546 * block the newly created grace period, so set up ->gp_tasks accordingly.
547 */
548static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
549{
550 WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp));
551 if (!list_empty(&rnp->blkd_tasks))
552 rnp->gp_tasks = rnp->blkd_tasks.next;
553 WARN_ON_ONCE(rnp->qsmask);
554}
555
556#ifdef CONFIG_HOTPLUG_CPU
557
558/*
559 * Handle tasklist migration for case in which all CPUs covered by the
560 * specified rcu_node have gone offline. Move them up to the root
561 * rcu_node. The reason for not just moving them to the immediate
562 * parent is to remove the need for rcu_read_unlock_special() to
563 * make more than two attempts to acquire the target rcu_node's lock.
564 * Returns true if there were tasks blocking the current RCU grace
565 * period.
566 *
567 * Returns 1 if there was previously a task blocking the current grace
568 * period on the specified rcu_node structure.
569 *
570 * The caller must hold rnp->lock with irqs disabled.
571 */
572static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
573 struct rcu_node *rnp,
574 struct rcu_data *rdp)
575{
576 struct list_head *lp;
577 struct list_head *lp_root;
578 int retval = 0;
579 struct rcu_node *rnp_root = rcu_get_root(rsp);
580 struct task_struct *t;
581
582 if (rnp == rnp_root) {
583 WARN_ONCE(1, "Last CPU thought to be offlined?");
584 return 0; /* Shouldn't happen: at least one CPU online. */
585 }
586
587 /* If we are on an internal node, complain bitterly. */
588 WARN_ON_ONCE(rnp != rdp->mynode);
589
590 /*
591 * Move tasks up to root rcu_node. Don't try to get fancy for
592 * this corner-case operation -- just put this node's tasks
593 * at the head of the root node's list, and update the root node's
594 * ->gp_tasks and ->exp_tasks pointers to those of this node's,
595 * if non-NULL. This might result in waiting for more tasks than
596 * absolutely necessary, but this is a good performance/complexity
597 * tradeoff.
598 */
599 if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0)
600 retval |= RCU_OFL_TASKS_NORM_GP;
601 if (rcu_preempted_readers_exp(rnp))
602 retval |= RCU_OFL_TASKS_EXP_GP;
603 lp = &rnp->blkd_tasks;
604 lp_root = &rnp_root->blkd_tasks;
605 while (!list_empty(lp)) {
606 t = list_entry(lp->next, typeof(*t), rcu_node_entry);
607 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
608 list_del(&t->rcu_node_entry);
609 t->rcu_blocked_node = rnp_root;
610 list_add(&t->rcu_node_entry, lp_root);
611 if (&t->rcu_node_entry == rnp->gp_tasks)
612 rnp_root->gp_tasks = rnp->gp_tasks;
613 if (&t->rcu_node_entry == rnp->exp_tasks)
614 rnp_root->exp_tasks = rnp->exp_tasks;
615#ifdef CONFIG_RCU_BOOST
616 if (&t->rcu_node_entry == rnp->boost_tasks)
617 rnp_root->boost_tasks = rnp->boost_tasks;
618#endif /* #ifdef CONFIG_RCU_BOOST */
619 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
620 }
621
622 rnp->gp_tasks = NULL;
623 rnp->exp_tasks = NULL;
624#ifdef CONFIG_RCU_BOOST
625 rnp->boost_tasks = NULL;
626 /*
627 * In case root is being boosted and leaf was not. Make sure
628 * that we boost the tasks blocking the current grace period
629 * in this case.
630 */
631 raw_spin_lock(&rnp_root->lock); /* irqs already disabled */
632 if (rnp_root->boost_tasks != NULL &&
633 rnp_root->boost_tasks != rnp_root->gp_tasks &&
634 rnp_root->boost_tasks != rnp_root->exp_tasks)
635 rnp_root->boost_tasks = rnp_root->gp_tasks;
636 raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */
637#endif /* #ifdef CONFIG_RCU_BOOST */
638
639 return retval;
640}
641
642#endif /* #ifdef CONFIG_HOTPLUG_CPU */
643
644/*
645 * Check for a quiescent state from the current CPU. When a task blocks,
646 * the task is recorded in the corresponding CPU's rcu_node structure,
647 * which is checked elsewhere.
648 *
649 * Caller must disable hard irqs.
650 */
651static void rcu_preempt_check_callbacks(int cpu)
652{
653 struct task_struct *t = current;
654
655 if (t->rcu_read_lock_nesting == 0) {
656 rcu_preempt_qs(cpu);
657 return;
658 }
659 if (t->rcu_read_lock_nesting > 0 &&
660 per_cpu(rcu_preempt_data, cpu).qs_pending)
661 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
662}
663
664#ifdef CONFIG_RCU_BOOST
665
666static void rcu_preempt_do_callbacks(void)
667{
668 rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data));
669}
670
671#endif /* #ifdef CONFIG_RCU_BOOST */
672
673/*
674 * Queue a preemptible-RCU callback for invocation after a grace period.
675 */
676void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
677{
678 __call_rcu(head, func, &rcu_preempt_state, -1, 0);
679}
680EXPORT_SYMBOL_GPL(call_rcu);
681
682/*
683 * Queue an RCU callback for lazy invocation after a grace period.
684 * This will likely be later named something like "call_rcu_lazy()",
685 * but this change will require some way of tagging the lazy RCU
686 * callbacks in the list of pending callbacks. Until then, this
687 * function may only be called from __kfree_rcu().
688 */
689void kfree_call_rcu(struct rcu_head *head,
690 void (*func)(struct rcu_head *rcu))
691{
692 __call_rcu(head, func, &rcu_preempt_state, -1, 1);
693}
694EXPORT_SYMBOL_GPL(kfree_call_rcu);
695
696/**
697 * synchronize_rcu - wait until a grace period has elapsed.
698 *
699 * Control will return to the caller some time after a full grace
700 * period has elapsed, in other words after all currently executing RCU
701 * read-side critical sections have completed. Note, however, that
702 * upon return from synchronize_rcu(), the caller might well be executing
703 * concurrently with new RCU read-side critical sections that began while
704 * synchronize_rcu() was waiting. RCU read-side critical sections are
705 * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested.
706 *
707 * See the description of synchronize_sched() for more detailed information
708 * on memory ordering guarantees.
709 */
710void synchronize_rcu(void)
711{
712 rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) &&
713 !lock_is_held(&rcu_lock_map) &&
714 !lock_is_held(&rcu_sched_lock_map),
715 "Illegal synchronize_rcu() in RCU read-side critical section");
716 if (!rcu_scheduler_active)
717 return;
718 if (rcu_expedited)
719 synchronize_rcu_expedited();
720 else
721 wait_rcu_gp(call_rcu);
722}
723EXPORT_SYMBOL_GPL(synchronize_rcu);
724
725static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq);
726static unsigned long sync_rcu_preempt_exp_count;
727static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex);
728
729/*
730 * Return non-zero if there are any tasks in RCU read-side critical
731 * sections blocking the current preemptible-RCU expedited grace period.
732 * If there is no preemptible-RCU expedited grace period currently in
733 * progress, returns zero unconditionally.
734 */
735static int rcu_preempted_readers_exp(struct rcu_node *rnp)
736{
737 return rnp->exp_tasks != NULL;
738}
739
740/*
741 * return non-zero if there is no RCU expedited grace period in progress
742 * for the specified rcu_node structure, in other words, if all CPUs and
743 * tasks covered by the specified rcu_node structure have done their bit
744 * for the current expedited grace period. Works only for preemptible
745 * RCU -- other RCU implementation use other means.
746 *
747 * Caller must hold sync_rcu_preempt_exp_mutex.
748 */
749static int sync_rcu_preempt_exp_done(struct rcu_node *rnp)
750{
751 return !rcu_preempted_readers_exp(rnp) &&
752 ACCESS_ONCE(rnp->expmask) == 0;
753}
754
755/*
756 * Report the exit from RCU read-side critical section for the last task
757 * that queued itself during or before the current expedited preemptible-RCU
758 * grace period. This event is reported either to the rcu_node structure on
759 * which the task was queued or to one of that rcu_node structure's ancestors,
760 * recursively up the tree. (Calm down, calm down, we do the recursion
761 * iteratively!)
762 *
763 * Most callers will set the "wake" flag, but the task initiating the
764 * expedited grace period need not wake itself.
765 *
766 * Caller must hold sync_rcu_preempt_exp_mutex.
767 */
768static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
769 bool wake)
770{
771 unsigned long flags;
772 unsigned long mask;
773
774 raw_spin_lock_irqsave(&rnp->lock, flags);
775 for (;;) {
776 if (!sync_rcu_preempt_exp_done(rnp)) {
777 raw_spin_unlock_irqrestore(&rnp->lock, flags);
778 break;
779 }
780 if (rnp->parent == NULL) {
781 raw_spin_unlock_irqrestore(&rnp->lock, flags);
782 if (wake)
783 wake_up(&sync_rcu_preempt_exp_wq);
784 break;
785 }
786 mask = rnp->grpmask;
787 raw_spin_unlock(&rnp->lock); /* irqs remain disabled */
788 rnp = rnp->parent;
789 raw_spin_lock(&rnp->lock); /* irqs already disabled */
790 rnp->expmask &= ~mask;
791 }
792}
793
794/*
795 * Snapshot the tasks blocking the newly started preemptible-RCU expedited
796 * grace period for the specified rcu_node structure. If there are no such
797 * tasks, report it up the rcu_node hierarchy.
798 *
799 * Caller must hold sync_rcu_preempt_exp_mutex and must exclude
800 * CPU hotplug operations.
801 */
802static void
803sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp)
804{
805 unsigned long flags;
806 int must_wait = 0;
807
808 raw_spin_lock_irqsave(&rnp->lock, flags);
809 if (list_empty(&rnp->blkd_tasks)) {
810 raw_spin_unlock_irqrestore(&rnp->lock, flags);
811 } else {
812 rnp->exp_tasks = rnp->blkd_tasks.next;
813 rcu_initiate_boost(rnp, flags); /* releases rnp->lock */
814 must_wait = 1;
815 }
816 if (!must_wait)
817 rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */
818}
819
820/**
821 * synchronize_rcu_expedited - Brute-force RCU grace period
822 *
823 * Wait for an RCU-preempt grace period, but expedite it. The basic
824 * idea is to invoke synchronize_sched_expedited() to push all the tasks to
825 * the ->blkd_tasks lists and wait for this list to drain. This consumes
826 * significant time on all CPUs and is unfriendly to real-time workloads,
827 * so is thus not recommended for any sort of common-case code.
828 * In fact, if you are using synchronize_rcu_expedited() in a loop,
829 * please restructure your code to batch your updates, and then Use a
830 * single synchronize_rcu() instead.
831 *
832 * Note that it is illegal to call this function while holding any lock
833 * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal
834 * to call this function from a CPU-hotplug notifier. Failing to observe
835 * these restriction will result in deadlock.
836 */
837void synchronize_rcu_expedited(void)
838{
839 unsigned long flags;
840 struct rcu_node *rnp;
841 struct rcu_state *rsp = &rcu_preempt_state;
842 unsigned long snap;
843 int trycount = 0;
844
845 smp_mb(); /* Caller's modifications seen first by other CPUs. */
846 snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1;
847 smp_mb(); /* Above access cannot bleed into critical section. */
848
849 /*
850 * Block CPU-hotplug operations. This means that any CPU-hotplug
851 * operation that finds an rcu_node structure with tasks in the
852 * process of being boosted will know that all tasks blocking
853 * this expedited grace period will already be in the process of
854 * being boosted. This simplifies the process of moving tasks
855 * from leaf to root rcu_node structures.
856 */
857 get_online_cpus();
858
859 /*
860 * Acquire lock, falling back to synchronize_rcu() if too many
861 * lock-acquisition failures. Of course, if someone does the
862 * expedited grace period for us, just leave.
863 */
864 while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) {
865 if (ULONG_CMP_LT(snap,
866 ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
867 put_online_cpus();
868 goto mb_ret; /* Others did our work for us. */
869 }
870 if (trycount++ < 10) {
871 udelay(trycount * num_online_cpus());
872 } else {
873 put_online_cpus();
874 wait_rcu_gp(call_rcu);
875 return;
876 }
877 }
878 if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) {
879 put_online_cpus();
880 goto unlock_mb_ret; /* Others did our work for us. */
881 }
882
883 /* force all RCU readers onto ->blkd_tasks lists. */
884 synchronize_sched_expedited();
885
886 /* Initialize ->expmask for all non-leaf rcu_node structures. */
887 rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) {
888 raw_spin_lock_irqsave(&rnp->lock, flags);
889 rnp->expmask = rnp->qsmaskinit;
890 raw_spin_unlock_irqrestore(&rnp->lock, flags);
891 }
892
893 /* Snapshot current state of ->blkd_tasks lists. */
894 rcu_for_each_leaf_node(rsp, rnp)
895 sync_rcu_preempt_exp_init(rsp, rnp);
896 if (NUM_RCU_NODES > 1)
897 sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp));
898
899 put_online_cpus();
900
901 /* Wait for snapshotted ->blkd_tasks lists to drain. */
902 rnp = rcu_get_root(rsp);
903 wait_event(sync_rcu_preempt_exp_wq,
904 sync_rcu_preempt_exp_done(rnp));
905
906 /* Clean up and exit. */
907 smp_mb(); /* ensure expedited GP seen before counter increment. */
908 ACCESS_ONCE(sync_rcu_preempt_exp_count)++;
909unlock_mb_ret:
910 mutex_unlock(&sync_rcu_preempt_exp_mutex);
911mb_ret:
912 smp_mb(); /* ensure subsequent action seen after grace period. */
913}
914EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
915
916/**
917 * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete.
918 *
919 * Note that this primitive does not necessarily wait for an RCU grace period
920 * to complete. For example, if there are no RCU callbacks queued anywhere
921 * in the system, then rcu_barrier() is within its rights to return
922 * immediately, without waiting for anything, much less an RCU grace period.
923 */
924void rcu_barrier(void)
925{
926 _rcu_barrier(&rcu_preempt_state);
927}
928EXPORT_SYMBOL_GPL(rcu_barrier);
929
930/*
931 * Initialize preemptible RCU's state structures.
932 */
933static void __init __rcu_init_preempt(void)
934{
935 rcu_init_one(&rcu_preempt_state, &rcu_preempt_data);
936}
937
938/*
939 * Check for a task exiting while in a preemptible-RCU read-side
940 * critical section, clean up if so. No need to issue warnings,
941 * as debug_check_no_locks_held() already does this if lockdep
942 * is enabled.
943 */
944void exit_rcu(void)
945{
946 struct task_struct *t = current;
947
948 if (likely(list_empty(&current->rcu_node_entry)))
949 return;
950 t->rcu_read_lock_nesting = 1;
951 barrier();
952 t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED;
953 __rcu_read_unlock();
954}
955
956#else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
957
958static struct rcu_state *rcu_state = &rcu_sched_state;
959
960/*
961 * Tell them what RCU they are running.
962 */
963static void __init rcu_bootup_announce(void)
964{
965 pr_info("Hierarchical RCU implementation.\n");
966 rcu_bootup_announce_oddness();
967}
968
969/*
970 * Return the number of RCU batches processed thus far for debug & stats.
971 */
972long rcu_batches_completed(void)
973{
974 return rcu_batches_completed_sched();
975}
976EXPORT_SYMBOL_GPL(rcu_batches_completed);
977
978/*
979 * Force a quiescent state for RCU, which, because there is no preemptible
980 * RCU, becomes the same as rcu-sched.
981 */
982void rcu_force_quiescent_state(void)
983{
984 rcu_sched_force_quiescent_state();
985}
986EXPORT_SYMBOL_GPL(rcu_force_quiescent_state);
987
988/*
989 * Because preemptible RCU does not exist, we never have to check for
990 * CPUs being in quiescent states.
991 */
992static void rcu_preempt_note_context_switch(int cpu)
993{
994}
995
996/*
997 * Because preemptible RCU does not exist, there are never any preempted
998 * RCU readers.
999 */
1000static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp)
1001{
1002 return 0;
1003}
1004
1005#ifdef CONFIG_HOTPLUG_CPU
1006
1007/* Because preemptible RCU does not exist, no quieting of tasks. */
1008static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags)
1009{
1010 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1011}
1012
1013#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1014
1015/*
1016 * Because preemptible RCU does not exist, we never have to check for
1017 * tasks blocked within RCU read-side critical sections.
1018 */
1019static void rcu_print_detail_task_stall(struct rcu_state *rsp)
1020{
1021}
1022
1023/*
1024 * Because preemptible RCU does not exist, we never have to check for
1025 * tasks blocked within RCU read-side critical sections.
1026 */
1027static int rcu_print_task_stall(struct rcu_node *rnp)
1028{
1029 return 0;
1030}
1031
1032/*
1033 * Because there is no preemptible RCU, there can be no readers blocked,
1034 * so there is no need to check for blocked tasks. So check only for
1035 * bogus qsmask values.
1036 */
1037static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp)
1038{
1039 WARN_ON_ONCE(rnp->qsmask);
1040}
1041
1042#ifdef CONFIG_HOTPLUG_CPU
1043
1044/*
1045 * Because preemptible RCU does not exist, it never needs to migrate
1046 * tasks that were blocked within RCU read-side critical sections, and
1047 * such non-existent tasks cannot possibly have been blocking the current
1048 * grace period.
1049 */
1050static int rcu_preempt_offline_tasks(struct rcu_state *rsp,
1051 struct rcu_node *rnp,
1052 struct rcu_data *rdp)
1053{
1054 return 0;
1055}
1056
1057#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1058
1059/*
1060 * Because preemptible RCU does not exist, it never has any callbacks
1061 * to check.
1062 */
1063static void rcu_preempt_check_callbacks(int cpu)
1064{
1065}
1066
1067/*
1068 * Queue an RCU callback for lazy invocation after a grace period.
1069 * This will likely be later named something like "call_rcu_lazy()",
1070 * but this change will require some way of tagging the lazy RCU
1071 * callbacks in the list of pending callbacks. Until then, this
1072 * function may only be called from __kfree_rcu().
1073 *
1074 * Because there is no preemptible RCU, we use RCU-sched instead.
1075 */
1076void kfree_call_rcu(struct rcu_head *head,
1077 void (*func)(struct rcu_head *rcu))
1078{
1079 __call_rcu(head, func, &rcu_sched_state, -1, 1);
1080}
1081EXPORT_SYMBOL_GPL(kfree_call_rcu);
1082
1083/*
1084 * Wait for an rcu-preempt grace period, but make it happen quickly.
1085 * But because preemptible RCU does not exist, map to rcu-sched.
1086 */
1087void synchronize_rcu_expedited(void)
1088{
1089 synchronize_sched_expedited();
1090}
1091EXPORT_SYMBOL_GPL(synchronize_rcu_expedited);
1092
1093#ifdef CONFIG_HOTPLUG_CPU
1094
1095/*
1096 * Because preemptible RCU does not exist, there is never any need to
1097 * report on tasks preempted in RCU read-side critical sections during
1098 * expedited RCU grace periods.
1099 */
1100static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp,
1101 bool wake)
1102{
1103}
1104
1105#endif /* #ifdef CONFIG_HOTPLUG_CPU */
1106
1107/*
1108 * Because preemptible RCU does not exist, rcu_barrier() is just
1109 * another name for rcu_barrier_sched().
1110 */
1111void rcu_barrier(void)
1112{
1113 rcu_barrier_sched();
1114}
1115EXPORT_SYMBOL_GPL(rcu_barrier);
1116
1117/*
1118 * Because preemptible RCU does not exist, it need not be initialized.
1119 */
1120static void __init __rcu_init_preempt(void)
1121{
1122}
1123
1124/*
1125 * Because preemptible RCU does not exist, tasks cannot possibly exit
1126 * while in preemptible RCU read-side critical sections.
1127 */
1128void exit_rcu(void)
1129{
1130}
1131
1132#endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */
1133
1134#ifdef CONFIG_RCU_BOOST
1135
1136#include "../rtmutex_common.h"
1137
1138#ifdef CONFIG_RCU_TRACE
1139
1140static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1141{
1142 if (list_empty(&rnp->blkd_tasks))
1143 rnp->n_balk_blkd_tasks++;
1144 else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL)
1145 rnp->n_balk_exp_gp_tasks++;
1146 else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL)
1147 rnp->n_balk_boost_tasks++;
1148 else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
1149 rnp->n_balk_notblocked++;
1150 else if (rnp->gp_tasks != NULL &&
1151 ULONG_CMP_LT(jiffies, rnp->boost_time))
1152 rnp->n_balk_notyet++;
1153 else
1154 rnp->n_balk_nos++;
1155}
1156
1157#else /* #ifdef CONFIG_RCU_TRACE */
1158
1159static void rcu_initiate_boost_trace(struct rcu_node *rnp)
1160{
1161}
1162
1163#endif /* #else #ifdef CONFIG_RCU_TRACE */
1164
1165static void rcu_wake_cond(struct task_struct *t, int status)
1166{
1167 /*
1168 * If the thread is yielding, only wake it when this
1169 * is invoked from idle
1170 */
1171 if (status != RCU_KTHREAD_YIELDING || is_idle_task(current))
1172 wake_up_process(t);
1173}
1174
1175/*
1176 * Carry out RCU priority boosting on the task indicated by ->exp_tasks
1177 * or ->boost_tasks, advancing the pointer to the next task in the
1178 * ->blkd_tasks list.
1179 *
1180 * Note that irqs must be enabled: boosting the task can block.
1181 * Returns 1 if there are more tasks needing to be boosted.
1182 */
1183static int rcu_boost(struct rcu_node *rnp)
1184{
1185 unsigned long flags;
1186 struct rt_mutex mtx;
1187 struct task_struct *t;
1188 struct list_head *tb;
1189
1190 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL)
1191 return 0; /* Nothing left to boost. */
1192
1193 raw_spin_lock_irqsave(&rnp->lock, flags);
1194
1195 /*
1196 * Recheck under the lock: all tasks in need of boosting
1197 * might exit their RCU read-side critical sections on their own.
1198 */
1199 if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) {
1200 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1201 return 0;
1202 }
1203
1204 /*
1205 * Preferentially boost tasks blocking expedited grace periods.
1206 * This cannot starve the normal grace periods because a second
1207 * expedited grace period must boost all blocked tasks, including
1208 * those blocking the pre-existing normal grace period.
1209 */
1210 if (rnp->exp_tasks != NULL) {
1211 tb = rnp->exp_tasks;
1212 rnp->n_exp_boosts++;
1213 } else {
1214 tb = rnp->boost_tasks;
1215 rnp->n_normal_boosts++;
1216 }
1217 rnp->n_tasks_boosted++;
1218
1219 /*
1220 * We boost task t by manufacturing an rt_mutex that appears to
1221 * be held by task t. We leave a pointer to that rt_mutex where
1222 * task t can find it, and task t will release the mutex when it
1223 * exits its outermost RCU read-side critical section. Then
1224 * simply acquiring this artificial rt_mutex will boost task
1225 * t's priority. (Thanks to tglx for suggesting this approach!)
1226 *
1227 * Note that task t must acquire rnp->lock to remove itself from
1228 * the ->blkd_tasks list, which it will do from exit() if from
1229 * nowhere else. We therefore are guaranteed that task t will
1230 * stay around at least until we drop rnp->lock. Note that
1231 * rnp->lock also resolves races between our priority boosting
1232 * and task t's exiting its outermost RCU read-side critical
1233 * section.
1234 */
1235 t = container_of(tb, struct task_struct, rcu_node_entry);
1236 rt_mutex_init_proxy_locked(&mtx, t);
1237 t->rcu_boost_mutex = &mtx;
1238 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1239 rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */
1240 rt_mutex_unlock(&mtx); /* Keep lockdep happy. */
1241
1242 return ACCESS_ONCE(rnp->exp_tasks) != NULL ||
1243 ACCESS_ONCE(rnp->boost_tasks) != NULL;
1244}
1245
1246/*
1247 * Priority-boosting kthread. One per leaf rcu_node and one for the
1248 * root rcu_node.
1249 */
1250static int rcu_boost_kthread(void *arg)
1251{
1252 struct rcu_node *rnp = (struct rcu_node *)arg;
1253 int spincnt = 0;
1254 int more2boost;
1255
1256 trace_rcu_utilization(TPS("Start boost kthread@init"));
1257 for (;;) {
1258 rnp->boost_kthread_status = RCU_KTHREAD_WAITING;
1259 trace_rcu_utilization(TPS("End boost kthread@rcu_wait"));
1260 rcu_wait(rnp->boost_tasks || rnp->exp_tasks);
1261 trace_rcu_utilization(TPS("Start boost kthread@rcu_wait"));
1262 rnp->boost_kthread_status = RCU_KTHREAD_RUNNING;
1263 more2boost = rcu_boost(rnp);
1264 if (more2boost)
1265 spincnt++;
1266 else
1267 spincnt = 0;
1268 if (spincnt > 10) {
1269 rnp->boost_kthread_status = RCU_KTHREAD_YIELDING;
1270 trace_rcu_utilization(TPS("End boost kthread@rcu_yield"));
1271 schedule_timeout_interruptible(2);
1272 trace_rcu_utilization(TPS("Start boost kthread@rcu_yield"));
1273 spincnt = 0;
1274 }
1275 }
1276 /* NOTREACHED */
1277 trace_rcu_utilization(TPS("End boost kthread@notreached"));
1278 return 0;
1279}
1280
1281/*
1282 * Check to see if it is time to start boosting RCU readers that are
1283 * blocking the current grace period, and, if so, tell the per-rcu_node
1284 * kthread to start boosting them. If there is an expedited grace
1285 * period in progress, it is always time to boost.
1286 *
1287 * The caller must hold rnp->lock, which this function releases.
1288 * The ->boost_kthread_task is immortal, so we don't need to worry
1289 * about it going away.
1290 */
1291static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1292{
1293 struct task_struct *t;
1294
1295 if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) {
1296 rnp->n_balk_exp_gp_tasks++;
1297 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1298 return;
1299 }
1300 if (rnp->exp_tasks != NULL ||
1301 (rnp->gp_tasks != NULL &&
1302 rnp->boost_tasks == NULL &&
1303 rnp->qsmask == 0 &&
1304 ULONG_CMP_GE(jiffies, rnp->boost_time))) {
1305 if (rnp->exp_tasks == NULL)
1306 rnp->boost_tasks = rnp->gp_tasks;
1307 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1308 t = rnp->boost_kthread_task;
1309 if (t)
1310 rcu_wake_cond(t, rnp->boost_kthread_status);
1311 } else {
1312 rcu_initiate_boost_trace(rnp);
1313 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1314 }
1315}
1316
1317/*
1318 * Wake up the per-CPU kthread to invoke RCU callbacks.
1319 */
1320static void invoke_rcu_callbacks_kthread(void)
1321{
1322 unsigned long flags;
1323
1324 local_irq_save(flags);
1325 __this_cpu_write(rcu_cpu_has_work, 1);
1326 if (__this_cpu_read(rcu_cpu_kthread_task) != NULL &&
1327 current != __this_cpu_read(rcu_cpu_kthread_task)) {
1328 rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task),
1329 __this_cpu_read(rcu_cpu_kthread_status));
1330 }
1331 local_irq_restore(flags);
1332}
1333
1334/*
1335 * Is the current CPU running the RCU-callbacks kthread?
1336 * Caller must have preemption disabled.
1337 */
1338static bool rcu_is_callbacks_kthread(void)
1339{
1340 return __this_cpu_read(rcu_cpu_kthread_task) == current;
1341}
1342
1343#define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000)
1344
1345/*
1346 * Do priority-boost accounting for the start of a new grace period.
1347 */
1348static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1349{
1350 rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES;
1351}
1352
1353/*
1354 * Create an RCU-boost kthread for the specified node if one does not
1355 * already exist. We only create this kthread for preemptible RCU.
1356 * Returns zero if all is well, a negated errno otherwise.
1357 */
1358static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp,
1359 struct rcu_node *rnp)
1360{
1361 int rnp_index = rnp - &rsp->node[0];
1362 unsigned long flags;
1363 struct sched_param sp;
1364 struct task_struct *t;
1365
1366 if (&rcu_preempt_state != rsp)
1367 return 0;
1368
1369 if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0)
1370 return 0;
1371
1372 rsp->boost = 1;
1373 if (rnp->boost_kthread_task != NULL)
1374 return 0;
1375 t = kthread_create(rcu_boost_kthread, (void *)rnp,
1376 "rcub/%d", rnp_index);
1377 if (IS_ERR(t))
1378 return PTR_ERR(t);
1379 raw_spin_lock_irqsave(&rnp->lock, flags);
1380 rnp->boost_kthread_task = t;
1381 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1382 sp.sched_priority = RCU_BOOST_PRIO;
1383 sched_setscheduler_nocheck(t, SCHED_FIFO, &sp);
1384 wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */
1385 return 0;
1386}
1387
1388static void rcu_kthread_do_work(void)
1389{
1390 rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data));
1391 rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data));
1392 rcu_preempt_do_callbacks();
1393}
1394
1395static void rcu_cpu_kthread_setup(unsigned int cpu)
1396{
1397 struct sched_param sp;
1398
1399 sp.sched_priority = RCU_KTHREAD_PRIO;
1400 sched_setscheduler_nocheck(current, SCHED_FIFO, &sp);
1401}
1402
1403static void rcu_cpu_kthread_park(unsigned int cpu)
1404{
1405 per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU;
1406}
1407
1408static int rcu_cpu_kthread_should_run(unsigned int cpu)
1409{
1410 return __this_cpu_read(rcu_cpu_has_work);
1411}
1412
1413/*
1414 * Per-CPU kernel thread that invokes RCU callbacks. This replaces the
1415 * RCU softirq used in flavors and configurations of RCU that do not
1416 * support RCU priority boosting.
1417 */
1418static void rcu_cpu_kthread(unsigned int cpu)
1419{
1420 unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status);
1421 char work, *workp = this_cpu_ptr(&rcu_cpu_has_work);
1422 int spincnt;
1423
1424 for (spincnt = 0; spincnt < 10; spincnt++) {
1425 trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait"));
1426 local_bh_disable();
1427 *statusp = RCU_KTHREAD_RUNNING;
1428 this_cpu_inc(rcu_cpu_kthread_loops);
1429 local_irq_disable();
1430 work = *workp;
1431 *workp = 0;
1432 local_irq_enable();
1433 if (work)
1434 rcu_kthread_do_work();
1435 local_bh_enable();
1436 if (*workp == 0) {
1437 trace_rcu_utilization(TPS("End CPU kthread@rcu_wait"));
1438 *statusp = RCU_KTHREAD_WAITING;
1439 return;
1440 }
1441 }
1442 *statusp = RCU_KTHREAD_YIELDING;
1443 trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield"));
1444 schedule_timeout_interruptible(2);
1445 trace_rcu_utilization(TPS("End CPU kthread@rcu_yield"));
1446 *statusp = RCU_KTHREAD_WAITING;
1447}
1448
1449/*
1450 * Set the per-rcu_node kthread's affinity to cover all CPUs that are
1451 * served by the rcu_node in question. The CPU hotplug lock is still
1452 * held, so the value of rnp->qsmaskinit will be stable.
1453 *
1454 * We don't include outgoingcpu in the affinity set, use -1 if there is
1455 * no outgoing CPU. If there are no CPUs left in the affinity set,
1456 * this function allows the kthread to execute on any CPU.
1457 */
1458static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1459{
1460 struct task_struct *t = rnp->boost_kthread_task;
1461 unsigned long mask = rnp->qsmaskinit;
1462 cpumask_var_t cm;
1463 int cpu;
1464
1465 if (!t)
1466 return;
1467 if (!zalloc_cpumask_var(&cm, GFP_KERNEL))
1468 return;
1469 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1)
1470 if ((mask & 0x1) && cpu != outgoingcpu)
1471 cpumask_set_cpu(cpu, cm);
1472 if (cpumask_weight(cm) == 0) {
1473 cpumask_setall(cm);
1474 for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++)
1475 cpumask_clear_cpu(cpu, cm);
1476 WARN_ON_ONCE(cpumask_weight(cm) == 0);
1477 }
1478 set_cpus_allowed_ptr(t, cm);
1479 free_cpumask_var(cm);
1480}
1481
1482static struct smp_hotplug_thread rcu_cpu_thread_spec = {
1483 .store = &rcu_cpu_kthread_task,
1484 .thread_should_run = rcu_cpu_kthread_should_run,
1485 .thread_fn = rcu_cpu_kthread,
1486 .thread_comm = "rcuc/%u",
1487 .setup = rcu_cpu_kthread_setup,
1488 .park = rcu_cpu_kthread_park,
1489};
1490
1491/*
1492 * Spawn all kthreads -- called as soon as the scheduler is running.
1493 */
1494static int __init rcu_spawn_kthreads(void)
1495{
1496 struct rcu_node *rnp;
1497 int cpu;
1498
1499 rcu_scheduler_fully_active = 1;
1500 for_each_possible_cpu(cpu)
1501 per_cpu(rcu_cpu_has_work, cpu) = 0;
1502 BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec));
1503 rnp = rcu_get_root(rcu_state);
1504 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1505 if (NUM_RCU_NODES > 1) {
1506 rcu_for_each_leaf_node(rcu_state, rnp)
1507 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1508 }
1509 return 0;
1510}
1511early_initcall(rcu_spawn_kthreads);
1512
1513static void rcu_prepare_kthreads(int cpu)
1514{
1515 struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu);
1516 struct rcu_node *rnp = rdp->mynode;
1517
1518 /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */
1519 if (rcu_scheduler_fully_active)
1520 (void)rcu_spawn_one_boost_kthread(rcu_state, rnp);
1521}
1522
1523#else /* #ifdef CONFIG_RCU_BOOST */
1524
1525static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags)
1526{
1527 raw_spin_unlock_irqrestore(&rnp->lock, flags);
1528}
1529
1530static void invoke_rcu_callbacks_kthread(void)
1531{
1532 WARN_ON_ONCE(1);
1533}
1534
1535static bool rcu_is_callbacks_kthread(void)
1536{
1537 return false;
1538}
1539
1540static void rcu_preempt_boost_start_gp(struct rcu_node *rnp)
1541{
1542}
1543
1544static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu)
1545{
1546}
1547
1548static int __init rcu_scheduler_really_started(void)
1549{
1550 rcu_scheduler_fully_active = 1;
1551 return 0;
1552}
1553early_initcall(rcu_scheduler_really_started);
1554
1555static void rcu_prepare_kthreads(int cpu)
1556{
1557}
1558
1559#endif /* #else #ifdef CONFIG_RCU_BOOST */
1560
1561#if !defined(CONFIG_RCU_FAST_NO_HZ)
1562
1563/*
1564 * Check to see if any future RCU-related work will need to be done
1565 * by the current CPU, even if none need be done immediately, returning
1566 * 1 if so. This function is part of the RCU implementation; it is -not-
1567 * an exported member of the RCU API.
1568 *
1569 * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs
1570 * any flavor of RCU.
1571 */
1572int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies)
1573{
1574 *delta_jiffies = ULONG_MAX;
1575 return rcu_cpu_has_callbacks(cpu, NULL);
1576}
1577
1578/*
1579 * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up
1580 * after it.
1581 */
1582static void rcu_cleanup_after_idle(int cpu)
1583{
1584}
1585
1586/*
1587 * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n,
1588 * is nothing.
1589 */
1590static void rcu_prepare_for_idle(int cpu)
1591{
1592}
1593
1594/*
1595 * Don't bother keeping a running count of the number of RCU callbacks
1596 * posted because CONFIG_RCU_FAST_NO_HZ=n.
1597 */
1598static void rcu_idle_count_callbacks_posted(void)
1599{
1600}
1601
1602#else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1603
1604/*
1605 * This code is invoked when a CPU goes idle, at which point we want
1606 * to have the CPU do everything required for RCU so that it can enter
1607 * the energy-efficient dyntick-idle mode. This is handled by a
1608 * state machine implemented by rcu_prepare_for_idle() below.
1609 *
1610 * The following three proprocessor symbols control this state machine:
1611 *
1612 * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted
1613 * to sleep in dyntick-idle mode with RCU callbacks pending. This
1614 * is sized to be roughly one RCU grace period. Those energy-efficiency
1615 * benchmarkers who might otherwise be tempted to set this to a large
1616 * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your
1617 * system. And if you are -that- concerned about energy efficiency,
1618 * just power the system down and be done with it!
1619 * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is
1620 * permitted to sleep in dyntick-idle mode with only lazy RCU
1621 * callbacks pending. Setting this too high can OOM your system.
1622 *
1623 * The values below work well in practice. If future workloads require
1624 * adjustment, they can be converted into kernel config parameters, though
1625 * making the state machine smarter might be a better option.
1626 */
1627#define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */
1628#define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */
1629
1630static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY;
1631module_param(rcu_idle_gp_delay, int, 0644);
1632static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY;
1633module_param(rcu_idle_lazy_gp_delay, int, 0644);
1634
1635extern int tick_nohz_enabled;
1636
1637/*
1638 * Try to advance callbacks for all flavors of RCU on the current CPU, but
1639 * only if it has been awhile since the last time we did so. Afterwards,
1640 * if there are any callbacks ready for immediate invocation, return true.
1641 */
1642static bool rcu_try_advance_all_cbs(void)
1643{
1644 bool cbs_ready = false;
1645 struct rcu_data *rdp;
1646 struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks);
1647 struct rcu_node *rnp;
1648 struct rcu_state *rsp;
1649
1650 /* Exit early if we advanced recently. */
1651 if (jiffies == rdtp->last_advance_all)
1652 return 0;
1653 rdtp->last_advance_all = jiffies;
1654
1655 for_each_rcu_flavor(rsp) {
1656 rdp = this_cpu_ptr(rsp->rda);
1657 rnp = rdp->mynode;
1658
1659 /*
1660 * Don't bother checking unless a grace period has
1661 * completed since we last checked and there are
1662 * callbacks not yet ready to invoke.
1663 */
1664 if (rdp->completed != rnp->completed &&
1665 rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL])
1666 note_gp_changes(rsp, rdp);
1667
1668 if (cpu_has_callbacks_ready_to_invoke(rdp))
1669 cbs_ready = true;
1670 }
1671 return cbs_ready;
1672}
1673
1674/*
1675 * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready
1676 * to invoke. If the CPU has callbacks, try to advance them. Tell the
1677 * caller to set the timeout based on whether or not there are non-lazy
1678 * callbacks.
1679 *
1680 * The caller must have disabled interrupts.
1681 */
1682int rcu_needs_cpu(int cpu, unsigned long *dj)
1683{
1684 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1685
1686 /* Snapshot to detect later posting of non-lazy callback. */
1687 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1688
1689 /* If no callbacks, RCU doesn't need the CPU. */
1690 if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) {
1691 *dj = ULONG_MAX;
1692 return 0;
1693 }
1694
1695 /* Attempt to advance callbacks. */
1696 if (rcu_try_advance_all_cbs()) {
1697 /* Some ready to invoke, so initiate later invocation. */
1698 invoke_rcu_core();
1699 return 1;
1700 }
1701 rdtp->last_accelerate = jiffies;
1702
1703 /* Request timer delay depending on laziness, and round. */
1704 if (!rdtp->all_lazy) {
1705 *dj = round_up(rcu_idle_gp_delay + jiffies,
1706 rcu_idle_gp_delay) - jiffies;
1707 } else {
1708 *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies;
1709 }
1710 return 0;
1711}
1712
1713/*
1714 * Prepare a CPU for idle from an RCU perspective. The first major task
1715 * is to sense whether nohz mode has been enabled or disabled via sysfs.
1716 * The second major task is to check to see if a non-lazy callback has
1717 * arrived at a CPU that previously had only lazy callbacks. The third
1718 * major task is to accelerate (that is, assign grace-period numbers to)
1719 * any recently arrived callbacks.
1720 *
1721 * The caller must have disabled interrupts.
1722 */
1723static void rcu_prepare_for_idle(int cpu)
1724{
1725 struct rcu_data *rdp;
1726 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1727 struct rcu_node *rnp;
1728 struct rcu_state *rsp;
1729 int tne;
1730
1731 /* Handle nohz enablement switches conservatively. */
1732 tne = ACCESS_ONCE(tick_nohz_enabled);
1733 if (tne != rdtp->tick_nohz_enabled_snap) {
1734 if (rcu_cpu_has_callbacks(cpu, NULL))
1735 invoke_rcu_core(); /* force nohz to see update. */
1736 rdtp->tick_nohz_enabled_snap = tne;
1737 return;
1738 }
1739 if (!tne)
1740 return;
1741
1742 /* If this is a no-CBs CPU, no callbacks, just return. */
1743 if (rcu_is_nocb_cpu(cpu))
1744 return;
1745
1746 /*
1747 * If a non-lazy callback arrived at a CPU having only lazy
1748 * callbacks, invoke RCU core for the side-effect of recalculating
1749 * idle duration on re-entry to idle.
1750 */
1751 if (rdtp->all_lazy &&
1752 rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) {
1753 rdtp->all_lazy = false;
1754 rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted;
1755 invoke_rcu_core();
1756 return;
1757 }
1758
1759 /*
1760 * If we have not yet accelerated this jiffy, accelerate all
1761 * callbacks on this CPU.
1762 */
1763 if (rdtp->last_accelerate == jiffies)
1764 return;
1765 rdtp->last_accelerate = jiffies;
1766 for_each_rcu_flavor(rsp) {
1767 rdp = per_cpu_ptr(rsp->rda, cpu);
1768 if (!*rdp->nxttail[RCU_DONE_TAIL])
1769 continue;
1770 rnp = rdp->mynode;
1771 raw_spin_lock(&rnp->lock); /* irqs already disabled. */
1772 rcu_accelerate_cbs(rsp, rnp, rdp);
1773 raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */
1774 }
1775}
1776
1777/*
1778 * Clean up for exit from idle. Attempt to advance callbacks based on
1779 * any grace periods that elapsed while the CPU was idle, and if any
1780 * callbacks are now ready to invoke, initiate invocation.
1781 */
1782static void rcu_cleanup_after_idle(int cpu)
1783{
1784
1785 if (rcu_is_nocb_cpu(cpu))
1786 return;
1787 if (rcu_try_advance_all_cbs())
1788 invoke_rcu_core();
1789}
1790
1791/*
1792 * Keep a running count of the number of non-lazy callbacks posted
1793 * on this CPU. This running counter (which is never decremented) allows
1794 * rcu_prepare_for_idle() to detect when something out of the idle loop
1795 * posts a callback, even if an equal number of callbacks are invoked.
1796 * Of course, callbacks should only be posted from within a trace event
1797 * designed to be called from idle or from within RCU_NONIDLE().
1798 */
1799static void rcu_idle_count_callbacks_posted(void)
1800{
1801 __this_cpu_add(rcu_dynticks.nonlazy_posted, 1);
1802}
1803
1804/*
1805 * Data for flushing lazy RCU callbacks at OOM time.
1806 */
1807static atomic_t oom_callback_count;
1808static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq);
1809
1810/*
1811 * RCU OOM callback -- decrement the outstanding count and deliver the
1812 * wake-up if we are the last one.
1813 */
1814static void rcu_oom_callback(struct rcu_head *rhp)
1815{
1816 if (atomic_dec_and_test(&oom_callback_count))
1817 wake_up(&oom_callback_wq);
1818}
1819
1820/*
1821 * Post an rcu_oom_notify callback on the current CPU if it has at
1822 * least one lazy callback. This will unnecessarily post callbacks
1823 * to CPUs that already have a non-lazy callback at the end of their
1824 * callback list, but this is an infrequent operation, so accept some
1825 * extra overhead to keep things simple.
1826 */
1827static void rcu_oom_notify_cpu(void *unused)
1828{
1829 struct rcu_state *rsp;
1830 struct rcu_data *rdp;
1831
1832 for_each_rcu_flavor(rsp) {
1833 rdp = __this_cpu_ptr(rsp->rda);
1834 if (rdp->qlen_lazy != 0) {
1835 atomic_inc(&oom_callback_count);
1836 rsp->call(&rdp->oom_head, rcu_oom_callback);
1837 }
1838 }
1839}
1840
1841/*
1842 * If low on memory, ensure that each CPU has a non-lazy callback.
1843 * This will wake up CPUs that have only lazy callbacks, in turn
1844 * ensuring that they free up the corresponding memory in a timely manner.
1845 * Because an uncertain amount of memory will be freed in some uncertain
1846 * timeframe, we do not claim to have freed anything.
1847 */
1848static int rcu_oom_notify(struct notifier_block *self,
1849 unsigned long notused, void *nfreed)
1850{
1851 int cpu;
1852
1853 /* Wait for callbacks from earlier instance to complete. */
1854 wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0);
1855
1856 /*
1857 * Prevent premature wakeup: ensure that all increments happen
1858 * before there is a chance of the counter reaching zero.
1859 */
1860 atomic_set(&oom_callback_count, 1);
1861
1862 get_online_cpus();
1863 for_each_online_cpu(cpu) {
1864 smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1);
1865 cond_resched();
1866 }
1867 put_online_cpus();
1868
1869 /* Unconditionally decrement: no need to wake ourselves up. */
1870 atomic_dec(&oom_callback_count);
1871
1872 return NOTIFY_OK;
1873}
1874
1875static struct notifier_block rcu_oom_nb = {
1876 .notifier_call = rcu_oom_notify
1877};
1878
1879static int __init rcu_register_oom_notifier(void)
1880{
1881 register_oom_notifier(&rcu_oom_nb);
1882 return 0;
1883}
1884early_initcall(rcu_register_oom_notifier);
1885
1886#endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */
1887
1888#ifdef CONFIG_RCU_CPU_STALL_INFO
1889
1890#ifdef CONFIG_RCU_FAST_NO_HZ
1891
1892static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1893{
1894 struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu);
1895 unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap;
1896
1897 sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c",
1898 rdtp->last_accelerate & 0xffff, jiffies & 0xffff,
1899 ulong2long(nlpd),
1900 rdtp->all_lazy ? 'L' : '.',
1901 rdtp->tick_nohz_enabled_snap ? '.' : 'D');
1902}
1903
1904#else /* #ifdef CONFIG_RCU_FAST_NO_HZ */
1905
1906static void print_cpu_stall_fast_no_hz(char *cp, int cpu)
1907{
1908 *cp = '\0';
1909}
1910
1911#endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */
1912
1913/* Initiate the stall-info list. */
1914static void print_cpu_stall_info_begin(void)
1915{
1916 pr_cont("\n");
1917}
1918
1919/*
1920 * Print out diagnostic information for the specified stalled CPU.
1921 *
1922 * If the specified CPU is aware of the current RCU grace period
1923 * (flavor specified by rsp), then print the number of scheduling
1924 * clock interrupts the CPU has taken during the time that it has
1925 * been aware. Otherwise, print the number of RCU grace periods
1926 * that this CPU is ignorant of, for example, "1" if the CPU was
1927 * aware of the previous grace period.
1928 *
1929 * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info.
1930 */
1931static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1932{
1933 char fast_no_hz[72];
1934 struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu);
1935 struct rcu_dynticks *rdtp = rdp->dynticks;
1936 char *ticks_title;
1937 unsigned long ticks_value;
1938
1939 if (rsp->gpnum == rdp->gpnum) {
1940 ticks_title = "ticks this GP";
1941 ticks_value = rdp->ticks_this_gp;
1942 } else {
1943 ticks_title = "GPs behind";
1944 ticks_value = rsp->gpnum - rdp->gpnum;
1945 }
1946 print_cpu_stall_fast_no_hz(fast_no_hz, cpu);
1947 pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n",
1948 cpu, ticks_value, ticks_title,
1949 atomic_read(&rdtp->dynticks) & 0xfff,
1950 rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting,
1951 rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu),
1952 fast_no_hz);
1953}
1954
1955/* Terminate the stall-info list. */
1956static void print_cpu_stall_info_end(void)
1957{
1958 pr_err("\t");
1959}
1960
1961/* Zero ->ticks_this_gp for all flavors of RCU. */
1962static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1963{
1964 rdp->ticks_this_gp = 0;
1965 rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id());
1966}
1967
1968/* Increment ->ticks_this_gp for all flavors of RCU. */
1969static void increment_cpu_stall_ticks(void)
1970{
1971 struct rcu_state *rsp;
1972
1973 for_each_rcu_flavor(rsp)
1974 __this_cpu_ptr(rsp->rda)->ticks_this_gp++;
1975}
1976
1977#else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */
1978
1979static void print_cpu_stall_info_begin(void)
1980{
1981 pr_cont(" {");
1982}
1983
1984static void print_cpu_stall_info(struct rcu_state *rsp, int cpu)
1985{
1986 pr_cont(" %d", cpu);
1987}
1988
1989static void print_cpu_stall_info_end(void)
1990{
1991 pr_cont("} ");
1992}
1993
1994static void zero_cpu_stall_ticks(struct rcu_data *rdp)
1995{
1996}
1997
1998static void increment_cpu_stall_ticks(void)
1999{
2000}
2001
2002#endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */
2003
2004#ifdef CONFIG_RCU_NOCB_CPU
2005
2006/*
2007 * Offload callback processing from the boot-time-specified set of CPUs
2008 * specified by rcu_nocb_mask. For each CPU in the set, there is a
2009 * kthread created that pulls the callbacks from the corresponding CPU,
2010 * waits for a grace period to elapse, and invokes the callbacks.
2011 * The no-CBs CPUs do a wake_up() on their kthread when they insert
2012 * a callback into any empty list, unless the rcu_nocb_poll boot parameter
2013 * has been specified, in which case each kthread actively polls its
2014 * CPU. (Which isn't so great for energy efficiency, but which does
2015 * reduce RCU's overhead on that CPU.)
2016 *
2017 * This is intended to be used in conjunction with Frederic Weisbecker's
2018 * adaptive-idle work, which would seriously reduce OS jitter on CPUs
2019 * running CPU-bound user-mode computations.
2020 *
2021 * Offloading of callback processing could also in theory be used as
2022 * an energy-efficiency measure because CPUs with no RCU callbacks
2023 * queued are more aggressive about entering dyntick-idle mode.
2024 */
2025
2026
2027/* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */
2028static int __init rcu_nocb_setup(char *str)
2029{
2030 alloc_bootmem_cpumask_var(&rcu_nocb_mask);
2031 have_rcu_nocb_mask = true;
2032 cpulist_parse(str, rcu_nocb_mask);
2033 return 1;
2034}
2035__setup("rcu_nocbs=", rcu_nocb_setup);
2036
2037static int __init parse_rcu_nocb_poll(char *arg)
2038{
2039 rcu_nocb_poll = 1;
2040 return 0;
2041}
2042early_param("rcu_nocb_poll", parse_rcu_nocb_poll);
2043
2044/*
2045 * Do any no-CBs CPUs need another grace period?
2046 *
2047 * Interrupts must be disabled. If the caller does not hold the root
2048 * rnp_node structure's ->lock, the results are advisory only.
2049 */
2050static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2051{
2052 struct rcu_node *rnp = rcu_get_root(rsp);
2053
2054 return rnp->need_future_gp[(ACCESS_ONCE(rnp->completed) + 1) & 0x1];
2055}
2056
2057/*
2058 * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended
2059 * grace period.
2060 */
2061static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2062{
2063 wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]);
2064}
2065
2066/*
2067 * Set the root rcu_node structure's ->need_future_gp field
2068 * based on the sum of those of all rcu_node structures. This does
2069 * double-count the root rcu_node structure's requests, but this
2070 * is necessary to handle the possibility of a rcu_nocb_kthread()
2071 * having awakened during the time that the rcu_node structures
2072 * were being updated for the end of the previous grace period.
2073 */
2074static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2075{
2076 rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq;
2077}
2078
2079static void rcu_init_one_nocb(struct rcu_node *rnp)
2080{
2081 init_waitqueue_head(&rnp->nocb_gp_wq[0]);
2082 init_waitqueue_head(&rnp->nocb_gp_wq[1]);
2083}
2084
2085/* Is the specified CPU a no-CPUs CPU? */
2086bool rcu_is_nocb_cpu(int cpu)
2087{
2088 if (have_rcu_nocb_mask)
2089 return cpumask_test_cpu(cpu, rcu_nocb_mask);
2090 return false;
2091}
2092
2093/*
2094 * Enqueue the specified string of rcu_head structures onto the specified
2095 * CPU's no-CBs lists. The CPU is specified by rdp, the head of the
2096 * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy
2097 * counts are supplied by rhcount and rhcount_lazy.
2098 *
2099 * If warranted, also wake up the kthread servicing this CPUs queues.
2100 */
2101static void __call_rcu_nocb_enqueue(struct rcu_data *rdp,
2102 struct rcu_head *rhp,
2103 struct rcu_head **rhtp,
2104 int rhcount, int rhcount_lazy)
2105{
2106 int len;
2107 struct rcu_head **old_rhpp;
2108 struct task_struct *t;
2109
2110 /* Enqueue the callback on the nocb list and update counts. */
2111 old_rhpp = xchg(&rdp->nocb_tail, rhtp);
2112 ACCESS_ONCE(*old_rhpp) = rhp;
2113 atomic_long_add(rhcount, &rdp->nocb_q_count);
2114 atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy);
2115
2116 /* If we are not being polled and there is a kthread, awaken it ... */
2117 t = ACCESS_ONCE(rdp->nocb_kthread);
2118 if (rcu_nocb_poll || !t) {
2119 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2120 TPS("WakeNotPoll"));
2121 return;
2122 }
2123 len = atomic_long_read(&rdp->nocb_q_count);
2124 if (old_rhpp == &rdp->nocb_head) {
2125 wake_up(&rdp->nocb_wq); /* ... only if queue was empty ... */
2126 rdp->qlen_last_fqs_check = 0;
2127 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeEmpty"));
2128 } else if (len > rdp->qlen_last_fqs_check + qhimark) {
2129 wake_up_process(t); /* ... or if many callbacks queued. */
2130 rdp->qlen_last_fqs_check = LONG_MAX / 2;
2131 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf"));
2132 } else {
2133 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot"));
2134 }
2135 return;
2136}
2137
2138/*
2139 * This is a helper for __call_rcu(), which invokes this when the normal
2140 * callback queue is inoperable. If this is not a no-CBs CPU, this
2141 * function returns failure back to __call_rcu(), which can complain
2142 * appropriately.
2143 *
2144 * Otherwise, this function queues the callback where the corresponding
2145 * "rcuo" kthread can find it.
2146 */
2147static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2148 bool lazy)
2149{
2150
2151 if (!rcu_is_nocb_cpu(rdp->cpu))
2152 return 0;
2153 __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy);
2154 if (__is_kfree_rcu_offset((unsigned long)rhp->func))
2155 trace_rcu_kfree_callback(rdp->rsp->name, rhp,
2156 (unsigned long)rhp->func,
2157 -atomic_long_read(&rdp->nocb_q_count_lazy),
2158 -atomic_long_read(&rdp->nocb_q_count));
2159 else
2160 trace_rcu_callback(rdp->rsp->name, rhp,
2161 -atomic_long_read(&rdp->nocb_q_count_lazy),
2162 -atomic_long_read(&rdp->nocb_q_count));
2163 return 1;
2164}
2165
2166/*
2167 * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is
2168 * not a no-CBs CPU.
2169 */
2170static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2171 struct rcu_data *rdp)
2172{
2173 long ql = rsp->qlen;
2174 long qll = rsp->qlen_lazy;
2175
2176 /* If this is not a no-CBs CPU, tell the caller to do it the old way. */
2177 if (!rcu_is_nocb_cpu(smp_processor_id()))
2178 return 0;
2179 rsp->qlen = 0;
2180 rsp->qlen_lazy = 0;
2181
2182 /* First, enqueue the donelist, if any. This preserves CB ordering. */
2183 if (rsp->orphan_donelist != NULL) {
2184 __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist,
2185 rsp->orphan_donetail, ql, qll);
2186 ql = qll = 0;
2187 rsp->orphan_donelist = NULL;
2188 rsp->orphan_donetail = &rsp->orphan_donelist;
2189 }
2190 if (rsp->orphan_nxtlist != NULL) {
2191 __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist,
2192 rsp->orphan_nxttail, ql, qll);
2193 ql = qll = 0;
2194 rsp->orphan_nxtlist = NULL;
2195 rsp->orphan_nxttail = &rsp->orphan_nxtlist;
2196 }
2197 return 1;
2198}
2199
2200/*
2201 * If necessary, kick off a new grace period, and either way wait
2202 * for a subsequent grace period to complete.
2203 */
2204static void rcu_nocb_wait_gp(struct rcu_data *rdp)
2205{
2206 unsigned long c;
2207 bool d;
2208 unsigned long flags;
2209 struct rcu_node *rnp = rdp->mynode;
2210
2211 raw_spin_lock_irqsave(&rnp->lock, flags);
2212 c = rcu_start_future_gp(rnp, rdp);
2213 raw_spin_unlock_irqrestore(&rnp->lock, flags);
2214
2215 /*
2216 * Wait for the grace period. Do so interruptibly to avoid messing
2217 * up the load average.
2218 */
2219 trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait"));
2220 for (;;) {
2221 wait_event_interruptible(
2222 rnp->nocb_gp_wq[c & 0x1],
2223 (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c)));
2224 if (likely(d))
2225 break;
2226 flush_signals(current);
2227 trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait"));
2228 }
2229 trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait"));
2230 smp_mb(); /* Ensure that CB invocation happens after GP end. */
2231}
2232
2233/*
2234 * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes
2235 * callbacks queued by the corresponding no-CBs CPU.
2236 */
2237static int rcu_nocb_kthread(void *arg)
2238{
2239 int c, cl;
2240 bool firsttime = 1;
2241 struct rcu_head *list;
2242 struct rcu_head *next;
2243 struct rcu_head **tail;
2244 struct rcu_data *rdp = arg;
2245
2246 /* Each pass through this loop invokes one batch of callbacks */
2247 for (;;) {
2248 /* If not polling, wait for next batch of callbacks. */
2249 if (!rcu_nocb_poll) {
2250 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2251 TPS("Sleep"));
2252 wait_event_interruptible(rdp->nocb_wq, rdp->nocb_head);
2253 } else if (firsttime) {
2254 firsttime = 0;
2255 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2256 TPS("Poll"));
2257 }
2258 list = ACCESS_ONCE(rdp->nocb_head);
2259 if (!list) {
2260 if (!rcu_nocb_poll)
2261 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2262 TPS("WokeEmpty"));
2263 schedule_timeout_interruptible(1);
2264 flush_signals(current);
2265 continue;
2266 }
2267 firsttime = 1;
2268 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2269 TPS("WokeNonEmpty"));
2270
2271 /*
2272 * Extract queued callbacks, update counts, and wait
2273 * for a grace period to elapse.
2274 */
2275 ACCESS_ONCE(rdp->nocb_head) = NULL;
2276 tail = xchg(&rdp->nocb_tail, &rdp->nocb_head);
2277 c = atomic_long_xchg(&rdp->nocb_q_count, 0);
2278 cl = atomic_long_xchg(&rdp->nocb_q_count_lazy, 0);
2279 ACCESS_ONCE(rdp->nocb_p_count) += c;
2280 ACCESS_ONCE(rdp->nocb_p_count_lazy) += cl;
2281 rcu_nocb_wait_gp(rdp);
2282
2283 /* Each pass through the following loop invokes a callback. */
2284 trace_rcu_batch_start(rdp->rsp->name, cl, c, -1);
2285 c = cl = 0;
2286 while (list) {
2287 next = list->next;
2288 /* Wait for enqueuing to complete, if needed. */
2289 while (next == NULL && &list->next != tail) {
2290 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2291 TPS("WaitQueue"));
2292 schedule_timeout_interruptible(1);
2293 trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu,
2294 TPS("WokeQueue"));
2295 next = list->next;
2296 }
2297 debug_rcu_head_unqueue(list);
2298 local_bh_disable();
2299 if (__rcu_reclaim(rdp->rsp->name, list))
2300 cl++;
2301 c++;
2302 local_bh_enable();
2303 list = next;
2304 }
2305 trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1);
2306 ACCESS_ONCE(rdp->nocb_p_count) -= c;
2307 ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl;
2308 rdp->n_nocbs_invoked += c;
2309 }
2310 return 0;
2311}
2312
2313/* Initialize per-rcu_data variables for no-CBs CPUs. */
2314static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2315{
2316 rdp->nocb_tail = &rdp->nocb_head;
2317 init_waitqueue_head(&rdp->nocb_wq);
2318}
2319
2320/* Create a kthread for each RCU flavor for each no-CBs CPU. */
2321static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2322{
2323 int cpu;
2324 struct rcu_data *rdp;
2325 struct task_struct *t;
2326
2327 if (rcu_nocb_mask == NULL)
2328 return;
2329 for_each_cpu(cpu, rcu_nocb_mask) {
2330 rdp = per_cpu_ptr(rsp->rda, cpu);
2331 t = kthread_run(rcu_nocb_kthread, rdp,
2332 "rcuo%c/%d", rsp->abbr, cpu);
2333 BUG_ON(IS_ERR(t));
2334 ACCESS_ONCE(rdp->nocb_kthread) = t;
2335 }
2336}
2337
2338/* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */
2339static bool init_nocb_callback_list(struct rcu_data *rdp)
2340{
2341 if (rcu_nocb_mask == NULL ||
2342 !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask))
2343 return false;
2344 rdp->nxttail[RCU_NEXT_TAIL] = NULL;
2345 return true;
2346}
2347
2348#else /* #ifdef CONFIG_RCU_NOCB_CPU */
2349
2350static int rcu_nocb_needs_gp(struct rcu_state *rsp)
2351{
2352 return 0;
2353}
2354
2355static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp)
2356{
2357}
2358
2359static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq)
2360{
2361}
2362
2363static void rcu_init_one_nocb(struct rcu_node *rnp)
2364{
2365}
2366
2367static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp,
2368 bool lazy)
2369{
2370 return 0;
2371}
2372
2373static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp,
2374 struct rcu_data *rdp)
2375{
2376 return 0;
2377}
2378
2379static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp)
2380{
2381}
2382
2383static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp)
2384{
2385}
2386
2387static bool init_nocb_callback_list(struct rcu_data *rdp)
2388{
2389 return false;
2390}
2391
2392#endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */
2393
2394/*
2395 * An adaptive-ticks CPU can potentially execute in kernel mode for an
2396 * arbitrarily long period of time with the scheduling-clock tick turned
2397 * off. RCU will be paying attention to this CPU because it is in the
2398 * kernel, but the CPU cannot be guaranteed to be executing the RCU state
2399 * machine because the scheduling-clock tick has been disabled. Therefore,
2400 * if an adaptive-ticks CPU is failing to respond to the current grace
2401 * period and has not be idle from an RCU perspective, kick it.
2402 */
2403static void rcu_kick_nohz_cpu(int cpu)
2404{
2405#ifdef CONFIG_NO_HZ_FULL
2406 if (tick_nohz_full_cpu(cpu))
2407 smp_send_reschedule(cpu);
2408#endif /* #ifdef CONFIG_NO_HZ_FULL */
2409}
2410
2411
2412#ifdef CONFIG_NO_HZ_FULL_SYSIDLE
2413
2414/*
2415 * Define RCU flavor that holds sysidle state. This needs to be the
2416 * most active flavor of RCU.
2417 */
2418#ifdef CONFIG_PREEMPT_RCU
2419static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state;
2420#else /* #ifdef CONFIG_PREEMPT_RCU */
2421static struct rcu_state *rcu_sysidle_state = &rcu_sched_state;
2422#endif /* #else #ifdef CONFIG_PREEMPT_RCU */
2423
2424static int full_sysidle_state; /* Current system-idle state. */
2425#define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */
2426#define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */
2427#define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */
2428#define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */
2429#define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */
2430
2431/*
2432 * Invoked to note exit from irq or task transition to idle. Note that
2433 * usermode execution does -not- count as idle here! After all, we want
2434 * to detect full-system idle states, not RCU quiescent states and grace
2435 * periods. The caller must have disabled interrupts.
2436 */
2437static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2438{
2439 unsigned long j;
2440
2441 /* Adjust nesting, check for fully idle. */
2442 if (irq) {
2443 rdtp->dynticks_idle_nesting--;
2444 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2445 if (rdtp->dynticks_idle_nesting != 0)
2446 return; /* Still not fully idle. */
2447 } else {
2448 if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) ==
2449 DYNTICK_TASK_NEST_VALUE) {
2450 rdtp->dynticks_idle_nesting = 0;
2451 } else {
2452 rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE;
2453 WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0);
2454 return; /* Still not fully idle. */
2455 }
2456 }
2457
2458 /* Record start of fully idle period. */
2459 j = jiffies;
2460 ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j;
2461 smp_mb__before_atomic_inc();
2462 atomic_inc(&rdtp->dynticks_idle);
2463 smp_mb__after_atomic_inc();
2464 WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1);
2465}
2466
2467/*
2468 * Unconditionally force exit from full system-idle state. This is
2469 * invoked when a normal CPU exits idle, but must be called separately
2470 * for the timekeeping CPU (tick_do_timer_cpu). The reason for this
2471 * is that the timekeeping CPU is permitted to take scheduling-clock
2472 * interrupts while the system is in system-idle state, and of course
2473 * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock
2474 * interrupt from any other type of interrupt.
2475 */
2476void rcu_sysidle_force_exit(void)
2477{
2478 int oldstate = ACCESS_ONCE(full_sysidle_state);
2479 int newoldstate;
2480
2481 /*
2482 * Each pass through the following loop attempts to exit full
2483 * system-idle state. If contention proves to be a problem,
2484 * a trylock-based contention tree could be used here.
2485 */
2486 while (oldstate > RCU_SYSIDLE_SHORT) {
2487 newoldstate = cmpxchg(&full_sysidle_state,
2488 oldstate, RCU_SYSIDLE_NOT);
2489 if (oldstate == newoldstate &&
2490 oldstate == RCU_SYSIDLE_FULL_NOTED) {
2491 rcu_kick_nohz_cpu(tick_do_timer_cpu);
2492 return; /* We cleared it, done! */
2493 }
2494 oldstate = newoldstate;
2495 }
2496 smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */
2497}
2498
2499/*
2500 * Invoked to note entry to irq or task transition from idle. Note that
2501 * usermode execution does -not- count as idle here! The caller must
2502 * have disabled interrupts.
2503 */
2504static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2505{
2506 /* Adjust nesting, check for already non-idle. */
2507 if (irq) {
2508 rdtp->dynticks_idle_nesting++;
2509 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2510 if (rdtp->dynticks_idle_nesting != 1)
2511 return; /* Already non-idle. */
2512 } else {
2513 /*
2514 * Allow for irq misnesting. Yes, it really is possible
2515 * to enter an irq handler then never leave it, and maybe
2516 * also vice versa. Handle both possibilities.
2517 */
2518 if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) {
2519 rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE;
2520 WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0);
2521 return; /* Already non-idle. */
2522 } else {
2523 rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE;
2524 }
2525 }
2526
2527 /* Record end of idle period. */
2528 smp_mb__before_atomic_inc();
2529 atomic_inc(&rdtp->dynticks_idle);
2530 smp_mb__after_atomic_inc();
2531 WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1));
2532
2533 /*
2534 * If we are the timekeeping CPU, we are permitted to be non-idle
2535 * during a system-idle state. This must be the case, because
2536 * the timekeeping CPU has to take scheduling-clock interrupts
2537 * during the time that the system is transitioning to full
2538 * system-idle state. This means that the timekeeping CPU must
2539 * invoke rcu_sysidle_force_exit() directly if it does anything
2540 * more than take a scheduling-clock interrupt.
2541 */
2542 if (smp_processor_id() == tick_do_timer_cpu)
2543 return;
2544
2545 /* Update system-idle state: We are clearly no longer fully idle! */
2546 rcu_sysidle_force_exit();
2547}
2548
2549/*
2550 * Check to see if the current CPU is idle. Note that usermode execution
2551 * does not count as idle. The caller must have disabled interrupts.
2552 */
2553static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2554 unsigned long *maxj)
2555{
2556 int cur;
2557 unsigned long j;
2558 struct rcu_dynticks *rdtp = rdp->dynticks;
2559
2560 /*
2561 * If some other CPU has already reported non-idle, if this is
2562 * not the flavor of RCU that tracks sysidle state, or if this
2563 * is an offline or the timekeeping CPU, nothing to do.
2564 */
2565 if (!*isidle || rdp->rsp != rcu_sysidle_state ||
2566 cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu)
2567 return;
2568 if (rcu_gp_in_progress(rdp->rsp))
2569 WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu);
2570
2571 /* Pick up current idle and NMI-nesting counter and check. */
2572 cur = atomic_read(&rdtp->dynticks_idle);
2573 if (cur & 0x1) {
2574 *isidle = false; /* We are not idle! */
2575 return;
2576 }
2577 smp_mb(); /* Read counters before timestamps. */
2578
2579 /* Pick up timestamps. */
2580 j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies);
2581 /* If this CPU entered idle more recently, update maxj timestamp. */
2582 if (ULONG_CMP_LT(*maxj, j))
2583 *maxj = j;
2584}
2585
2586/*
2587 * Is this the flavor of RCU that is handling full-system idle?
2588 */
2589static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2590{
2591 return rsp == rcu_sysidle_state;
2592}
2593
2594/*
2595 * Bind the grace-period kthread for the sysidle flavor of RCU to the
2596 * timekeeping CPU.
2597 */
2598static void rcu_bind_gp_kthread(void)
2599{
2600 int cpu = ACCESS_ONCE(tick_do_timer_cpu);
2601
2602 if (cpu < 0 || cpu >= nr_cpu_ids)
2603 return;
2604 if (raw_smp_processor_id() != cpu)
2605 set_cpus_allowed_ptr(current, cpumask_of(cpu));
2606}
2607
2608/*
2609 * Return a delay in jiffies based on the number of CPUs, rcu_node
2610 * leaf fanout, and jiffies tick rate. The idea is to allow larger
2611 * systems more time to transition to full-idle state in order to
2612 * avoid the cache thrashing that otherwise occur on the state variable.
2613 * Really small systems (less than a couple of tens of CPUs) should
2614 * instead use a single global atomically incremented counter, and later
2615 * versions of this will automatically reconfigure themselves accordingly.
2616 */
2617static unsigned long rcu_sysidle_delay(void)
2618{
2619 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2620 return 0;
2621 return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000);
2622}
2623
2624/*
2625 * Advance the full-system-idle state. This is invoked when all of
2626 * the non-timekeeping CPUs are idle.
2627 */
2628static void rcu_sysidle(unsigned long j)
2629{
2630 /* Check the current state. */
2631 switch (ACCESS_ONCE(full_sysidle_state)) {
2632 case RCU_SYSIDLE_NOT:
2633
2634 /* First time all are idle, so note a short idle period. */
2635 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT;
2636 break;
2637
2638 case RCU_SYSIDLE_SHORT:
2639
2640 /*
2641 * Idle for a bit, time to advance to next state?
2642 * cmpxchg failure means race with non-idle, let them win.
2643 */
2644 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2645 (void)cmpxchg(&full_sysidle_state,
2646 RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG);
2647 break;
2648
2649 case RCU_SYSIDLE_LONG:
2650
2651 /*
2652 * Do an additional check pass before advancing to full.
2653 * cmpxchg failure means race with non-idle, let them win.
2654 */
2655 if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay()))
2656 (void)cmpxchg(&full_sysidle_state,
2657 RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL);
2658 break;
2659
2660 default:
2661 break;
2662 }
2663}
2664
2665/*
2666 * Found a non-idle non-timekeeping CPU, so kick the system-idle state
2667 * back to the beginning.
2668 */
2669static void rcu_sysidle_cancel(void)
2670{
2671 smp_mb();
2672 ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT;
2673}
2674
2675/*
2676 * Update the sysidle state based on the results of a force-quiescent-state
2677 * scan of the CPUs' dyntick-idle state.
2678 */
2679static void rcu_sysidle_report(struct rcu_state *rsp, int isidle,
2680 unsigned long maxj, bool gpkt)
2681{
2682 if (rsp != rcu_sysidle_state)
2683 return; /* Wrong flavor, ignore. */
2684 if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL)
2685 return; /* Running state machine from timekeeping CPU. */
2686 if (isidle)
2687 rcu_sysidle(maxj); /* More idle! */
2688 else
2689 rcu_sysidle_cancel(); /* Idle is over. */
2690}
2691
2692/*
2693 * Wrapper for rcu_sysidle_report() when called from the grace-period
2694 * kthread's context.
2695 */
2696static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2697 unsigned long maxj)
2698{
2699 rcu_sysidle_report(rsp, isidle, maxj, true);
2700}
2701
2702/* Callback and function for forcing an RCU grace period. */
2703struct rcu_sysidle_head {
2704 struct rcu_head rh;
2705 int inuse;
2706};
2707
2708static void rcu_sysidle_cb(struct rcu_head *rhp)
2709{
2710 struct rcu_sysidle_head *rshp;
2711
2712 /*
2713 * The following memory barrier is needed to replace the
2714 * memory barriers that would normally be in the memory
2715 * allocator.
2716 */
2717 smp_mb(); /* grace period precedes setting inuse. */
2718
2719 rshp = container_of(rhp, struct rcu_sysidle_head, rh);
2720 ACCESS_ONCE(rshp->inuse) = 0;
2721}
2722
2723/*
2724 * Check to see if the system is fully idle, other than the timekeeping CPU.
2725 * The caller must have disabled interrupts.
2726 */
2727bool rcu_sys_is_idle(void)
2728{
2729 static struct rcu_sysidle_head rsh;
2730 int rss = ACCESS_ONCE(full_sysidle_state);
2731
2732 if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu))
2733 return false;
2734
2735 /* Handle small-system case by doing a full scan of CPUs. */
2736 if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) {
2737 int oldrss = rss - 1;
2738
2739 /*
2740 * One pass to advance to each state up to _FULL.
2741 * Give up if any pass fails to advance the state.
2742 */
2743 while (rss < RCU_SYSIDLE_FULL && oldrss < rss) {
2744 int cpu;
2745 bool isidle = true;
2746 unsigned long maxj = jiffies - ULONG_MAX / 4;
2747 struct rcu_data *rdp;
2748
2749 /* Scan all the CPUs looking for nonidle CPUs. */
2750 for_each_possible_cpu(cpu) {
2751 rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu);
2752 rcu_sysidle_check_cpu(rdp, &isidle, &maxj);
2753 if (!isidle)
2754 break;
2755 }
2756 rcu_sysidle_report(rcu_sysidle_state,
2757 isidle, maxj, false);
2758 oldrss = rss;
2759 rss = ACCESS_ONCE(full_sysidle_state);
2760 }
2761 }
2762
2763 /* If this is the first observation of an idle period, record it. */
2764 if (rss == RCU_SYSIDLE_FULL) {
2765 rss = cmpxchg(&full_sysidle_state,
2766 RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED);
2767 return rss == RCU_SYSIDLE_FULL;
2768 }
2769
2770 smp_mb(); /* ensure rss load happens before later caller actions. */
2771
2772 /* If already fully idle, tell the caller (in case of races). */
2773 if (rss == RCU_SYSIDLE_FULL_NOTED)
2774 return true;
2775
2776 /*
2777 * If we aren't there yet, and a grace period is not in flight,
2778 * initiate a grace period. Either way, tell the caller that
2779 * we are not there yet. We use an xchg() rather than an assignment
2780 * to make up for the memory barriers that would otherwise be
2781 * provided by the memory allocator.
2782 */
2783 if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL &&
2784 !rcu_gp_in_progress(rcu_sysidle_state) &&
2785 !rsh.inuse && xchg(&rsh.inuse, 1) == 0)
2786 call_rcu(&rsh.rh, rcu_sysidle_cb);
2787 return false;
2788}
2789
2790/*
2791 * Initialize dynticks sysidle state for CPUs coming online.
2792 */
2793static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2794{
2795 rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE;
2796}
2797
2798#else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
2799
2800static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq)
2801{
2802}
2803
2804static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq)
2805{
2806}
2807
2808static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle,
2809 unsigned long *maxj)
2810{
2811}
2812
2813static bool is_sysidle_rcu_state(struct rcu_state *rsp)
2814{
2815 return false;
2816}
2817
2818static void rcu_bind_gp_kthread(void)
2819{
2820}
2821
2822static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle,
2823 unsigned long maxj)
2824{
2825}
2826
2827static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp)
2828{
2829}
2830
2831#endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */
diff --git a/kernel/rcu/tree_trace.c b/kernel/rcu/tree_trace.c
new file mode 100644
index 000000000000..3596797b7e46
--- /dev/null
+++ b/kernel/rcu/tree_trace.c
@@ -0,0 +1,500 @@
1/*
2 * Read-Copy Update tracing for classic implementation
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 * Papers: http://www.rdrop.com/users/paulmck/RCU
21 *
22 * For detailed explanation of Read-Copy Update mechanism see -
23 * Documentation/RCU
24 *
25 */
26#include <linux/types.h>
27#include <linux/kernel.h>
28#include <linux/init.h>
29#include <linux/spinlock.h>
30#include <linux/smp.h>
31#include <linux/rcupdate.h>
32#include <linux/interrupt.h>
33#include <linux/sched.h>
34#include <linux/atomic.h>
35#include <linux/bitops.h>
36#include <linux/module.h>
37#include <linux/completion.h>
38#include <linux/moduleparam.h>
39#include <linux/percpu.h>
40#include <linux/notifier.h>
41#include <linux/cpu.h>
42#include <linux/mutex.h>
43#include <linux/debugfs.h>
44#include <linux/seq_file.h>
45
46#define RCU_TREE_NONCORE
47#include "tree.h"
48
49static int r_open(struct inode *inode, struct file *file,
50 const struct seq_operations *op)
51{
52 int ret = seq_open(file, op);
53 if (!ret) {
54 struct seq_file *m = (struct seq_file *)file->private_data;
55 m->private = inode->i_private;
56 }
57 return ret;
58}
59
60static void *r_start(struct seq_file *m, loff_t *pos)
61{
62 struct rcu_state *rsp = (struct rcu_state *)m->private;
63 *pos = cpumask_next(*pos - 1, cpu_possible_mask);
64 if ((*pos) < nr_cpu_ids)
65 return per_cpu_ptr(rsp->rda, *pos);
66 return NULL;
67}
68
69static void *r_next(struct seq_file *m, void *v, loff_t *pos)
70{
71 (*pos)++;
72 return r_start(m, pos);
73}
74
75static void r_stop(struct seq_file *m, void *v)
76{
77}
78
79static int show_rcubarrier(struct seq_file *m, void *v)
80{
81 struct rcu_state *rsp = (struct rcu_state *)m->private;
82 seq_printf(m, "bcc: %d nbd: %lu\n",
83 atomic_read(&rsp->barrier_cpu_count),
84 rsp->n_barrier_done);
85 return 0;
86}
87
88static int rcubarrier_open(struct inode *inode, struct file *file)
89{
90 return single_open(file, show_rcubarrier, inode->i_private);
91}
92
93static const struct file_operations rcubarrier_fops = {
94 .owner = THIS_MODULE,
95 .open = rcubarrier_open,
96 .read = seq_read,
97 .llseek = no_llseek,
98 .release = single_release,
99};
100
101#ifdef CONFIG_RCU_BOOST
102
103static char convert_kthread_status(unsigned int kthread_status)
104{
105 if (kthread_status > RCU_KTHREAD_MAX)
106 return '?';
107 return "SRWOY"[kthread_status];
108}
109
110#endif /* #ifdef CONFIG_RCU_BOOST */
111
112static void print_one_rcu_data(struct seq_file *m, struct rcu_data *rdp)
113{
114 long ql, qll;
115
116 if (!rdp->beenonline)
117 return;
118 seq_printf(m, "%3d%cc=%ld g=%ld pq=%d qp=%d",
119 rdp->cpu,
120 cpu_is_offline(rdp->cpu) ? '!' : ' ',
121 ulong2long(rdp->completed), ulong2long(rdp->gpnum),
122 rdp->passed_quiesce, rdp->qs_pending);
123 seq_printf(m, " dt=%d/%llx/%d df=%lu",
124 atomic_read(&rdp->dynticks->dynticks),
125 rdp->dynticks->dynticks_nesting,
126 rdp->dynticks->dynticks_nmi_nesting,
127 rdp->dynticks_fqs);
128 seq_printf(m, " of=%lu", rdp->offline_fqs);
129 rcu_nocb_q_lengths(rdp, &ql, &qll);
130 qll += rdp->qlen_lazy;
131 ql += rdp->qlen;
132 seq_printf(m, " ql=%ld/%ld qs=%c%c%c%c",
133 qll, ql,
134 ".N"[rdp->nxttail[RCU_NEXT_READY_TAIL] !=
135 rdp->nxttail[RCU_NEXT_TAIL]],
136 ".R"[rdp->nxttail[RCU_WAIT_TAIL] !=
137 rdp->nxttail[RCU_NEXT_READY_TAIL]],
138 ".W"[rdp->nxttail[RCU_DONE_TAIL] !=
139 rdp->nxttail[RCU_WAIT_TAIL]],
140 ".D"[&rdp->nxtlist != rdp->nxttail[RCU_DONE_TAIL]]);
141#ifdef CONFIG_RCU_BOOST
142 seq_printf(m, " kt=%d/%c ktl=%x",
143 per_cpu(rcu_cpu_has_work, rdp->cpu),
144 convert_kthread_status(per_cpu(rcu_cpu_kthread_status,
145 rdp->cpu)),
146 per_cpu(rcu_cpu_kthread_loops, rdp->cpu) & 0xffff);
147#endif /* #ifdef CONFIG_RCU_BOOST */
148 seq_printf(m, " b=%ld", rdp->blimit);
149 seq_printf(m, " ci=%lu nci=%lu co=%lu ca=%lu\n",
150 rdp->n_cbs_invoked, rdp->n_nocbs_invoked,
151 rdp->n_cbs_orphaned, rdp->n_cbs_adopted);
152}
153
154static int show_rcudata(struct seq_file *m, void *v)
155{
156 print_one_rcu_data(m, (struct rcu_data *)v);
157 return 0;
158}
159
160static const struct seq_operations rcudate_op = {
161 .start = r_start,
162 .next = r_next,
163 .stop = r_stop,
164 .show = show_rcudata,
165};
166
167static int rcudata_open(struct inode *inode, struct file *file)
168{
169 return r_open(inode, file, &rcudate_op);
170}
171
172static const struct file_operations rcudata_fops = {
173 .owner = THIS_MODULE,
174 .open = rcudata_open,
175 .read = seq_read,
176 .llseek = no_llseek,
177 .release = seq_release,
178};
179
180static int show_rcuexp(struct seq_file *m, void *v)
181{
182 struct rcu_state *rsp = (struct rcu_state *)m->private;
183
184 seq_printf(m, "s=%lu d=%lu w=%lu tf=%lu wd1=%lu wd2=%lu n=%lu sc=%lu dt=%lu dl=%lu dx=%lu\n",
185 atomic_long_read(&rsp->expedited_start),
186 atomic_long_read(&rsp->expedited_done),
187 atomic_long_read(&rsp->expedited_wrap),
188 atomic_long_read(&rsp->expedited_tryfail),
189 atomic_long_read(&rsp->expedited_workdone1),
190 atomic_long_read(&rsp->expedited_workdone2),
191 atomic_long_read(&rsp->expedited_normal),
192 atomic_long_read(&rsp->expedited_stoppedcpus),
193 atomic_long_read(&rsp->expedited_done_tries),
194 atomic_long_read(&rsp->expedited_done_lost),
195 atomic_long_read(&rsp->expedited_done_exit));
196 return 0;
197}
198
199static int rcuexp_open(struct inode *inode, struct file *file)
200{
201 return single_open(file, show_rcuexp, inode->i_private);
202}
203
204static const struct file_operations rcuexp_fops = {
205 .owner = THIS_MODULE,
206 .open = rcuexp_open,
207 .read = seq_read,
208 .llseek = no_llseek,
209 .release = single_release,
210};
211
212#ifdef CONFIG_RCU_BOOST
213
214static void print_one_rcu_node_boost(struct seq_file *m, struct rcu_node *rnp)
215{
216 seq_printf(m, "%d:%d tasks=%c%c%c%c kt=%c ntb=%lu neb=%lu nnb=%lu ",
217 rnp->grplo, rnp->grphi,
218 "T."[list_empty(&rnp->blkd_tasks)],
219 "N."[!rnp->gp_tasks],
220 "E."[!rnp->exp_tasks],
221 "B."[!rnp->boost_tasks],
222 convert_kthread_status(rnp->boost_kthread_status),
223 rnp->n_tasks_boosted, rnp->n_exp_boosts,
224 rnp->n_normal_boosts);
225 seq_printf(m, "j=%04x bt=%04x\n",
226 (int)(jiffies & 0xffff),
227 (int)(rnp->boost_time & 0xffff));
228 seq_printf(m, " balk: nt=%lu egt=%lu bt=%lu nb=%lu ny=%lu nos=%lu\n",
229 rnp->n_balk_blkd_tasks,
230 rnp->n_balk_exp_gp_tasks,
231 rnp->n_balk_boost_tasks,
232 rnp->n_balk_notblocked,
233 rnp->n_balk_notyet,
234 rnp->n_balk_nos);
235}
236
237static int show_rcu_node_boost(struct seq_file *m, void *unused)
238{
239 struct rcu_node *rnp;
240
241 rcu_for_each_leaf_node(&rcu_preempt_state, rnp)
242 print_one_rcu_node_boost(m, rnp);
243 return 0;
244}
245
246static int rcu_node_boost_open(struct inode *inode, struct file *file)
247{
248 return single_open(file, show_rcu_node_boost, NULL);
249}
250
251static const struct file_operations rcu_node_boost_fops = {
252 .owner = THIS_MODULE,
253 .open = rcu_node_boost_open,
254 .read = seq_read,
255 .llseek = no_llseek,
256 .release = single_release,
257};
258
259#endif /* #ifdef CONFIG_RCU_BOOST */
260
261static void print_one_rcu_state(struct seq_file *m, struct rcu_state *rsp)
262{
263 unsigned long gpnum;
264 int level = 0;
265 struct rcu_node *rnp;
266
267 gpnum = rsp->gpnum;
268 seq_printf(m, "c=%ld g=%ld s=%d jfq=%ld j=%x ",
269 ulong2long(rsp->completed), ulong2long(gpnum),
270 rsp->fqs_state,
271 (long)(rsp->jiffies_force_qs - jiffies),
272 (int)(jiffies & 0xffff));
273 seq_printf(m, "nfqs=%lu/nfqsng=%lu(%lu) fqlh=%lu oqlen=%ld/%ld\n",
274 rsp->n_force_qs, rsp->n_force_qs_ngp,
275 rsp->n_force_qs - rsp->n_force_qs_ngp,
276 rsp->n_force_qs_lh, rsp->qlen_lazy, rsp->qlen);
277 for (rnp = &rsp->node[0]; rnp - &rsp->node[0] < rcu_num_nodes; rnp++) {
278 if (rnp->level != level) {
279 seq_puts(m, "\n");
280 level = rnp->level;
281 }
282 seq_printf(m, "%lx/%lx %c%c>%c %d:%d ^%d ",
283 rnp->qsmask, rnp->qsmaskinit,
284 ".G"[rnp->gp_tasks != NULL],
285 ".E"[rnp->exp_tasks != NULL],
286 ".T"[!list_empty(&rnp->blkd_tasks)],
287 rnp->grplo, rnp->grphi, rnp->grpnum);
288 }
289 seq_puts(m, "\n");
290}
291
292static int show_rcuhier(struct seq_file *m, void *v)
293{
294 struct rcu_state *rsp = (struct rcu_state *)m->private;
295 print_one_rcu_state(m, rsp);
296 return 0;
297}
298
299static int rcuhier_open(struct inode *inode, struct file *file)
300{
301 return single_open(file, show_rcuhier, inode->i_private);
302}
303
304static const struct file_operations rcuhier_fops = {
305 .owner = THIS_MODULE,
306 .open = rcuhier_open,
307 .read = seq_read,
308 .llseek = no_llseek,
309 .release = single_release,
310};
311
312static void show_one_rcugp(struct seq_file *m, struct rcu_state *rsp)
313{
314 unsigned long flags;
315 unsigned long completed;
316 unsigned long gpnum;
317 unsigned long gpage;
318 unsigned long gpmax;
319 struct rcu_node *rnp = &rsp->node[0];
320
321 raw_spin_lock_irqsave(&rnp->lock, flags);
322 completed = ACCESS_ONCE(rsp->completed);
323 gpnum = ACCESS_ONCE(rsp->gpnum);
324 if (completed == gpnum)
325 gpage = 0;
326 else
327 gpage = jiffies - rsp->gp_start;
328 gpmax = rsp->gp_max;
329 raw_spin_unlock_irqrestore(&rnp->lock, flags);
330 seq_printf(m, "completed=%ld gpnum=%ld age=%ld max=%ld\n",
331 ulong2long(completed), ulong2long(gpnum), gpage, gpmax);
332}
333
334static int show_rcugp(struct seq_file *m, void *v)
335{
336 struct rcu_state *rsp = (struct rcu_state *)m->private;
337 show_one_rcugp(m, rsp);
338 return 0;
339}
340
341static int rcugp_open(struct inode *inode, struct file *file)
342{
343 return single_open(file, show_rcugp, inode->i_private);
344}
345
346static const struct file_operations rcugp_fops = {
347 .owner = THIS_MODULE,
348 .open = rcugp_open,
349 .read = seq_read,
350 .llseek = no_llseek,
351 .release = single_release,
352};
353
354static void print_one_rcu_pending(struct seq_file *m, struct rcu_data *rdp)
355{
356 if (!rdp->beenonline)
357 return;
358 seq_printf(m, "%3d%cnp=%ld ",
359 rdp->cpu,
360 cpu_is_offline(rdp->cpu) ? '!' : ' ',
361 rdp->n_rcu_pending);
362 seq_printf(m, "qsp=%ld rpq=%ld cbr=%ld cng=%ld ",
363 rdp->n_rp_qs_pending,
364 rdp->n_rp_report_qs,
365 rdp->n_rp_cb_ready,
366 rdp->n_rp_cpu_needs_gp);
367 seq_printf(m, "gpc=%ld gps=%ld nn=%ld\n",
368 rdp->n_rp_gp_completed,
369 rdp->n_rp_gp_started,
370 rdp->n_rp_need_nothing);
371}
372
373static int show_rcu_pending(struct seq_file *m, void *v)
374{
375 print_one_rcu_pending(m, (struct rcu_data *)v);
376 return 0;
377}
378
379static const struct seq_operations rcu_pending_op = {
380 .start = r_start,
381 .next = r_next,
382 .stop = r_stop,
383 .show = show_rcu_pending,
384};
385
386static int rcu_pending_open(struct inode *inode, struct file *file)
387{
388 return r_open(inode, file, &rcu_pending_op);
389}
390
391static const struct file_operations rcu_pending_fops = {
392 .owner = THIS_MODULE,
393 .open = rcu_pending_open,
394 .read = seq_read,
395 .llseek = no_llseek,
396 .release = seq_release,
397};
398
399static int show_rcutorture(struct seq_file *m, void *unused)
400{
401 seq_printf(m, "rcutorture test sequence: %lu %s\n",
402 rcutorture_testseq >> 1,
403 (rcutorture_testseq & 0x1) ? "(test in progress)" : "");
404 seq_printf(m, "rcutorture update version number: %lu\n",
405 rcutorture_vernum);
406 return 0;
407}
408
409static int rcutorture_open(struct inode *inode, struct file *file)
410{
411 return single_open(file, show_rcutorture, NULL);
412}
413
414static const struct file_operations rcutorture_fops = {
415 .owner = THIS_MODULE,
416 .open = rcutorture_open,
417 .read = seq_read,
418 .llseek = seq_lseek,
419 .release = single_release,
420};
421
422static struct dentry *rcudir;
423
424static int __init rcutree_trace_init(void)
425{
426 struct rcu_state *rsp;
427 struct dentry *retval;
428 struct dentry *rspdir;
429
430 rcudir = debugfs_create_dir("rcu", NULL);
431 if (!rcudir)
432 goto free_out;
433
434 for_each_rcu_flavor(rsp) {
435 rspdir = debugfs_create_dir(rsp->name, rcudir);
436 if (!rspdir)
437 goto free_out;
438
439 retval = debugfs_create_file("rcudata", 0444,
440 rspdir, rsp, &rcudata_fops);
441 if (!retval)
442 goto free_out;
443
444 retval = debugfs_create_file("rcuexp", 0444,
445 rspdir, rsp, &rcuexp_fops);
446 if (!retval)
447 goto free_out;
448
449 retval = debugfs_create_file("rcu_pending", 0444,
450 rspdir, rsp, &rcu_pending_fops);
451 if (!retval)
452 goto free_out;
453
454 retval = debugfs_create_file("rcubarrier", 0444,
455 rspdir, rsp, &rcubarrier_fops);
456 if (!retval)
457 goto free_out;
458
459#ifdef CONFIG_RCU_BOOST
460 if (rsp == &rcu_preempt_state) {
461 retval = debugfs_create_file("rcuboost", 0444,
462 rspdir, NULL, &rcu_node_boost_fops);
463 if (!retval)
464 goto free_out;
465 }
466#endif
467
468 retval = debugfs_create_file("rcugp", 0444,
469 rspdir, rsp, &rcugp_fops);
470 if (!retval)
471 goto free_out;
472
473 retval = debugfs_create_file("rcuhier", 0444,
474 rspdir, rsp, &rcuhier_fops);
475 if (!retval)
476 goto free_out;
477 }
478
479 retval = debugfs_create_file("rcutorture", 0444, rcudir,
480 NULL, &rcutorture_fops);
481 if (!retval)
482 goto free_out;
483 return 0;
484free_out:
485 debugfs_remove_recursive(rcudir);
486 return 1;
487}
488
489static void __exit rcutree_trace_cleanup(void)
490{
491 debugfs_remove_recursive(rcudir);
492}
493
494
495module_init(rcutree_trace_init);
496module_exit(rcutree_trace_cleanup);
497
498MODULE_AUTHOR("Paul E. McKenney");
499MODULE_DESCRIPTION("Read-Copy Update tracing for hierarchical implementation");
500MODULE_LICENSE("GPL");
diff --git a/kernel/rcu/update.c b/kernel/rcu/update.c
new file mode 100644
index 000000000000..6cb3dff89e2b
--- /dev/null
+++ b/kernel/rcu/update.c
@@ -0,0 +1,347 @@
1/*
2 * Read-Copy Update mechanism for mutual exclusion
3 *
4 * This program is free software; you can redistribute it and/or modify
5 * it under the terms of the GNU General Public License as published by
6 * the Free Software Foundation; either version 2 of the License, or
7 * (at your option) any later version.
8 *
9 * This program is distributed in the hope that it will be useful,
10 * but WITHOUT ANY WARRANTY; without even the implied warranty of
11 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
12 * GNU General Public License for more details.
13 *
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, write to the Free Software
16 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA.
17 *
18 * Copyright IBM Corporation, 2001
19 *
20 * Authors: Dipankar Sarma <dipankar@in.ibm.com>
21 * Manfred Spraul <manfred@colorfullife.com>
22 *
23 * Based on the original work by Paul McKenney <paulmck@us.ibm.com>
24 * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen.
25 * Papers:
26 * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf
27 * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001)
28 *
29 * For detailed explanation of Read-Copy Update mechanism see -
30 * http://lse.sourceforge.net/locking/rcupdate.html
31 *
32 */
33#include <linux/types.h>
34#include <linux/kernel.h>
35#include <linux/init.h>
36#include <linux/spinlock.h>
37#include <linux/smp.h>
38#include <linux/interrupt.h>
39#include <linux/sched.h>
40#include <linux/atomic.h>
41#include <linux/bitops.h>
42#include <linux/percpu.h>
43#include <linux/notifier.h>
44#include <linux/cpu.h>
45#include <linux/mutex.h>
46#include <linux/export.h>
47#include <linux/hardirq.h>
48#include <linux/delay.h>
49#include <linux/module.h>
50
51#define CREATE_TRACE_POINTS
52#include <trace/events/rcu.h>
53
54#include "rcu.h"
55
56MODULE_ALIAS("rcupdate");
57#ifdef MODULE_PARAM_PREFIX
58#undef MODULE_PARAM_PREFIX
59#endif
60#define MODULE_PARAM_PREFIX "rcupdate."
61
62module_param(rcu_expedited, int, 0);
63
64#ifdef CONFIG_PREEMPT_RCU
65
66/*
67 * Preemptible RCU implementation for rcu_read_lock().
68 * Just increment ->rcu_read_lock_nesting, shared state will be updated
69 * if we block.
70 */
71void __rcu_read_lock(void)
72{
73 current->rcu_read_lock_nesting++;
74 barrier(); /* critical section after entry code. */
75}
76EXPORT_SYMBOL_GPL(__rcu_read_lock);
77
78/*
79 * Preemptible RCU implementation for rcu_read_unlock().
80 * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost
81 * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
82 * invoke rcu_read_unlock_special() to clean up after a context switch
83 * in an RCU read-side critical section and other special cases.
84 */
85void __rcu_read_unlock(void)
86{
87 struct task_struct *t = current;
88
89 if (t->rcu_read_lock_nesting != 1) {
90 --t->rcu_read_lock_nesting;
91 } else {
92 barrier(); /* critical section before exit code. */
93 t->rcu_read_lock_nesting = INT_MIN;
94#ifdef CONFIG_PROVE_RCU_DELAY
95 udelay(10); /* Make preemption more probable. */
96#endif /* #ifdef CONFIG_PROVE_RCU_DELAY */
97 barrier(); /* assign before ->rcu_read_unlock_special load */
98 if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
99 rcu_read_unlock_special(t);
100 barrier(); /* ->rcu_read_unlock_special load before assign */
101 t->rcu_read_lock_nesting = 0;
102 }
103#ifdef CONFIG_PROVE_LOCKING
104 {
105 int rrln = ACCESS_ONCE(t->rcu_read_lock_nesting);
106
107 WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2);
108 }
109#endif /* #ifdef CONFIG_PROVE_LOCKING */
110}
111EXPORT_SYMBOL_GPL(__rcu_read_unlock);
112
113#endif /* #ifdef CONFIG_PREEMPT_RCU */
114
115#ifdef CONFIG_DEBUG_LOCK_ALLOC
116static struct lock_class_key rcu_lock_key;
117struct lockdep_map rcu_lock_map =
118 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key);
119EXPORT_SYMBOL_GPL(rcu_lock_map);
120
121static struct lock_class_key rcu_bh_lock_key;
122struct lockdep_map rcu_bh_lock_map =
123 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_bh", &rcu_bh_lock_key);
124EXPORT_SYMBOL_GPL(rcu_bh_lock_map);
125
126static struct lock_class_key rcu_sched_lock_key;
127struct lockdep_map rcu_sched_lock_map =
128 STATIC_LOCKDEP_MAP_INIT("rcu_read_lock_sched", &rcu_sched_lock_key);
129EXPORT_SYMBOL_GPL(rcu_sched_lock_map);
130
131int notrace debug_lockdep_rcu_enabled(void)
132{
133 return rcu_scheduler_active && debug_locks &&
134 current->lockdep_recursion == 0;
135}
136EXPORT_SYMBOL_GPL(debug_lockdep_rcu_enabled);
137
138/**
139 * rcu_read_lock_bh_held() - might we be in RCU-bh read-side critical section?
140 *
141 * Check for bottom half being disabled, which covers both the
142 * CONFIG_PROVE_RCU and not cases. Note that if someone uses
143 * rcu_read_lock_bh(), but then later enables BH, lockdep (if enabled)
144 * will show the situation. This is useful for debug checks in functions
145 * that require that they be called within an RCU read-side critical
146 * section.
147 *
148 * Check debug_lockdep_rcu_enabled() to prevent false positives during boot.
149 *
150 * Note that rcu_read_lock() is disallowed if the CPU is either idle or
151 * offline from an RCU perspective, so check for those as well.
152 */
153int rcu_read_lock_bh_held(void)
154{
155 if (!debug_lockdep_rcu_enabled())
156 return 1;
157 if (!rcu_is_watching())
158 return 0;
159 if (!rcu_lockdep_current_cpu_online())
160 return 0;
161 return in_softirq() || irqs_disabled();
162}
163EXPORT_SYMBOL_GPL(rcu_read_lock_bh_held);
164
165#endif /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
166
167struct rcu_synchronize {
168 struct rcu_head head;
169 struct completion completion;
170};
171
172/*
173 * Awaken the corresponding synchronize_rcu() instance now that a
174 * grace period has elapsed.
175 */
176static void wakeme_after_rcu(struct rcu_head *head)
177{
178 struct rcu_synchronize *rcu;
179
180 rcu = container_of(head, struct rcu_synchronize, head);
181 complete(&rcu->completion);
182}
183
184void wait_rcu_gp(call_rcu_func_t crf)
185{
186 struct rcu_synchronize rcu;
187
188 init_rcu_head_on_stack(&rcu.head);
189 init_completion(&rcu.completion);
190 /* Will wake me after RCU finished. */
191 crf(&rcu.head, wakeme_after_rcu);
192 /* Wait for it. */
193 wait_for_completion(&rcu.completion);
194 destroy_rcu_head_on_stack(&rcu.head);
195}
196EXPORT_SYMBOL_GPL(wait_rcu_gp);
197
198#ifdef CONFIG_PROVE_RCU
199/*
200 * wrapper function to avoid #include problems.
201 */
202int rcu_my_thread_group_empty(void)
203{
204 return thread_group_empty(current);
205}
206EXPORT_SYMBOL_GPL(rcu_my_thread_group_empty);
207#endif /* #ifdef CONFIG_PROVE_RCU */
208
209#ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD
210static inline void debug_init_rcu_head(struct rcu_head *head)
211{
212 debug_object_init(head, &rcuhead_debug_descr);
213}
214
215static inline void debug_rcu_head_free(struct rcu_head *head)
216{
217 debug_object_free(head, &rcuhead_debug_descr);
218}
219
220/*
221 * fixup_activate is called when:
222 * - an active object is activated
223 * - an unknown object is activated (might be a statically initialized object)
224 * Activation is performed internally by call_rcu().
225 */
226static int rcuhead_fixup_activate(void *addr, enum debug_obj_state state)
227{
228 struct rcu_head *head = addr;
229
230 switch (state) {
231
232 case ODEBUG_STATE_NOTAVAILABLE:
233 /*
234 * This is not really a fixup. We just make sure that it is
235 * tracked in the object tracker.
236 */
237 debug_object_init(head, &rcuhead_debug_descr);
238 debug_object_activate(head, &rcuhead_debug_descr);
239 return 0;
240 default:
241 return 1;
242 }
243}
244
245/**
246 * init_rcu_head_on_stack() - initialize on-stack rcu_head for debugobjects
247 * @head: pointer to rcu_head structure to be initialized
248 *
249 * This function informs debugobjects of a new rcu_head structure that
250 * has been allocated as an auto variable on the stack. This function
251 * is not required for rcu_head structures that are statically defined or
252 * that are dynamically allocated on the heap. This function has no
253 * effect for !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
254 */
255void init_rcu_head_on_stack(struct rcu_head *head)
256{
257 debug_object_init_on_stack(head, &rcuhead_debug_descr);
258}
259EXPORT_SYMBOL_GPL(init_rcu_head_on_stack);
260
261/**
262 * destroy_rcu_head_on_stack() - destroy on-stack rcu_head for debugobjects
263 * @head: pointer to rcu_head structure to be initialized
264 *
265 * This function informs debugobjects that an on-stack rcu_head structure
266 * is about to go out of scope. As with init_rcu_head_on_stack(), this
267 * function is not required for rcu_head structures that are statically
268 * defined or that are dynamically allocated on the heap. Also as with
269 * init_rcu_head_on_stack(), this function has no effect for
270 * !CONFIG_DEBUG_OBJECTS_RCU_HEAD kernel builds.
271 */
272void destroy_rcu_head_on_stack(struct rcu_head *head)
273{
274 debug_object_free(head, &rcuhead_debug_descr);
275}
276EXPORT_SYMBOL_GPL(destroy_rcu_head_on_stack);
277
278struct debug_obj_descr rcuhead_debug_descr = {
279 .name = "rcu_head",
280 .fixup_activate = rcuhead_fixup_activate,
281};
282EXPORT_SYMBOL_GPL(rcuhead_debug_descr);
283#endif /* #ifdef CONFIG_DEBUG_OBJECTS_RCU_HEAD */
284
285#if defined(CONFIG_TREE_RCU) || defined(CONFIG_TREE_PREEMPT_RCU) || defined(CONFIG_RCU_TRACE)
286void do_trace_rcu_torture_read(const char *rcutorturename, struct rcu_head *rhp,
287 unsigned long secs,
288 unsigned long c_old, unsigned long c)
289{
290 trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c);
291}
292EXPORT_SYMBOL_GPL(do_trace_rcu_torture_read);
293#else
294#define do_trace_rcu_torture_read(rcutorturename, rhp, secs, c_old, c) \
295 do { } while (0)
296#endif
297
298#ifdef CONFIG_RCU_STALL_COMMON
299
300#ifdef CONFIG_PROVE_RCU
301#define RCU_STALL_DELAY_DELTA (5 * HZ)
302#else
303#define RCU_STALL_DELAY_DELTA 0
304#endif
305
306int rcu_cpu_stall_suppress __read_mostly; /* 1 = suppress stall warnings. */
307static int rcu_cpu_stall_timeout __read_mostly = CONFIG_RCU_CPU_STALL_TIMEOUT;
308
309module_param(rcu_cpu_stall_suppress, int, 0644);
310module_param(rcu_cpu_stall_timeout, int, 0644);
311
312int rcu_jiffies_till_stall_check(void)
313{
314 int till_stall_check = ACCESS_ONCE(rcu_cpu_stall_timeout);
315
316 /*
317 * Limit check must be consistent with the Kconfig limits
318 * for CONFIG_RCU_CPU_STALL_TIMEOUT.
319 */
320 if (till_stall_check < 3) {
321 ACCESS_ONCE(rcu_cpu_stall_timeout) = 3;
322 till_stall_check = 3;
323 } else if (till_stall_check > 300) {
324 ACCESS_ONCE(rcu_cpu_stall_timeout) = 300;
325 till_stall_check = 300;
326 }
327 return till_stall_check * HZ + RCU_STALL_DELAY_DELTA;
328}
329
330static int rcu_panic(struct notifier_block *this, unsigned long ev, void *ptr)
331{
332 rcu_cpu_stall_suppress = 1;
333 return NOTIFY_DONE;
334}
335
336static struct notifier_block rcu_panic_block = {
337 .notifier_call = rcu_panic,
338};
339
340static int __init check_cpu_stall_init(void)
341{
342 atomic_notifier_chain_register(&panic_notifier_list, &rcu_panic_block);
343 return 0;
344}
345early_initcall(check_cpu_stall_init);
346
347#endif /* #ifdef CONFIG_RCU_STALL_COMMON */
generated by cgit v1.2.2 (git 2.25.0) at 2025-12-27 11:07:58 -0500