2 * Sleepable Read-Copy Update mechanism for mutual exclusion.
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.
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.
14 * You should have received a copy of the GNU General Public License
15 * along with this program; if not, you can access it online at
16 * http://www.gnu.org/licenses/gpl-2.0.html.
18 * Copyright (C) IBM Corporation, 2006
19 * Copyright (C) Fujitsu, 2012
21 * Author: Paul McKenney <paulmck@us.ibm.com>
22 * Lai Jiangshan <laijs@cn.fujitsu.com>
24 * For detailed explanation of Read-Copy Update mechanism see -
25 * Documentation/RCU/ *.txt
29 #include <linux/export.h>
30 #include <linux/mutex.h>
31 #include <linux/percpu.h>
32 #include <linux/preempt.h>
33 #include <linux/rcupdate_wait.h>
34 #include <linux/sched.h>
35 #include <linux/smp.h>
36 #include <linux/delay.h>
37 #include <linux/module.h>
38 #include <linux/srcu.h>
41 #include "rcu_segcblist.h"
43 ulong exp_holdoff
= 25 * 1000; /* Holdoff (ns) for auto-expediting. */
44 module_param(exp_holdoff
, ulong
, 0444);
46 static void srcu_invoke_callbacks(struct work_struct
*work
);
47 static void srcu_reschedule(struct srcu_struct
*sp
, unsigned long delay
);
50 * Initialize SRCU combining tree. Note that statically allocated
51 * srcu_struct structures might already have srcu_read_lock() and
52 * srcu_read_unlock() running against them. So if the is_static parameter
53 * is set, don't initialize ->srcu_lock_count[] and ->srcu_unlock_count[].
55 static void init_srcu_struct_nodes(struct srcu_struct
*sp
, bool is_static
)
60 int levelspread
[RCU_NUM_LVLS
];
61 struct srcu_data
*sdp
;
62 struct srcu_node
*snp
;
63 struct srcu_node
*snp_first
;
65 /* Work out the overall tree geometry. */
66 sp
->level
[0] = &sp
->node
[0];
67 for (i
= 1; i
< rcu_num_lvls
; i
++)
68 sp
->level
[i
] = sp
->level
[i
- 1] + num_rcu_lvl
[i
- 1];
69 rcu_init_levelspread(levelspread
, num_rcu_lvl
);
71 /* Each pass through this loop initializes one srcu_node structure. */
72 rcu_for_each_node_breadth_first(sp
, snp
) {
73 spin_lock_init(&snp
->lock
);
74 WARN_ON_ONCE(ARRAY_SIZE(snp
->srcu_have_cbs
) !=
75 ARRAY_SIZE(snp
->srcu_data_have_cbs
));
76 for (i
= 0; i
< ARRAY_SIZE(snp
->srcu_have_cbs
); i
++) {
77 snp
->srcu_have_cbs
[i
] = 0;
78 snp
->srcu_data_have_cbs
[i
] = 0;
80 snp
->srcu_gp_seq_needed_exp
= 0;
83 if (snp
== &sp
->node
[0]) {
84 /* Root node, special case. */
85 snp
->srcu_parent
= NULL
;
90 if (snp
== sp
->level
[level
+ 1])
92 snp
->srcu_parent
= sp
->level
[level
- 1] +
93 (snp
- sp
->level
[level
]) /
94 levelspread
[level
- 1];
98 * Initialize the per-CPU srcu_data array, which feeds into the
99 * leaves of the srcu_node tree.
101 WARN_ON_ONCE(ARRAY_SIZE(sdp
->srcu_lock_count
) !=
102 ARRAY_SIZE(sdp
->srcu_unlock_count
));
103 level
= rcu_num_lvls
- 1;
104 snp_first
= sp
->level
[level
];
105 for_each_possible_cpu(cpu
) {
106 sdp
= per_cpu_ptr(sp
->sda
, cpu
);
107 spin_lock_init(&sdp
->lock
);
108 rcu_segcblist_init(&sdp
->srcu_cblist
);
109 sdp
->srcu_cblist_invoking
= false;
110 sdp
->srcu_gp_seq_needed
= sp
->srcu_gp_seq
;
111 sdp
->srcu_gp_seq_needed_exp
= sp
->srcu_gp_seq
;
112 sdp
->mynode
= &snp_first
[cpu
/ levelspread
[level
]];
113 for (snp
= sdp
->mynode
; snp
!= NULL
; snp
= snp
->srcu_parent
) {
119 INIT_DELAYED_WORK(&sdp
->work
, srcu_invoke_callbacks
);
121 sdp
->grpmask
= 1 << (cpu
- sdp
->mynode
->grplo
);
125 /* Dynamically allocated, better be no srcu_read_locks()! */
126 for (i
= 0; i
< ARRAY_SIZE(sdp
->srcu_lock_count
); i
++) {
127 sdp
->srcu_lock_count
[i
] = 0;
128 sdp
->srcu_unlock_count
[i
] = 0;
134 * Initialize non-compile-time initialized fields, including the
135 * associated srcu_node and srcu_data structures. The is_static
136 * parameter is passed through to init_srcu_struct_nodes(), and
137 * also tells us that ->sda has already been wired up to srcu_data.
139 static int init_srcu_struct_fields(struct srcu_struct
*sp
, bool is_static
)
141 mutex_init(&sp
->srcu_cb_mutex
);
142 mutex_init(&sp
->srcu_gp_mutex
);
145 sp
->srcu_barrier_seq
= 0;
146 mutex_init(&sp
->srcu_barrier_mutex
);
147 atomic_set(&sp
->srcu_barrier_cpu_cnt
, 0);
148 INIT_DELAYED_WORK(&sp
->work
, process_srcu
);
150 sp
->sda
= alloc_percpu(struct srcu_data
);
151 init_srcu_struct_nodes(sp
, is_static
);
152 sp
->srcu_gp_seq_needed_exp
= 0;
153 sp
->srcu_last_gp_end
= ktime_get_mono_fast_ns();
154 smp_store_release(&sp
->srcu_gp_seq_needed
, 0); /* Init done. */
155 return sp
->sda
? 0 : -ENOMEM
;
158 #ifdef CONFIG_DEBUG_LOCK_ALLOC
160 int __init_srcu_struct(struct srcu_struct
*sp
, const char *name
,
161 struct lock_class_key
*key
)
163 /* Don't re-initialize a lock while it is held. */
164 debug_check_no_locks_freed((void *)sp
, sizeof(*sp
));
165 lockdep_init_map(&sp
->dep_map
, name
, key
, 0);
166 spin_lock_init(&sp
->gp_lock
);
167 return init_srcu_struct_fields(sp
, false);
169 EXPORT_SYMBOL_GPL(__init_srcu_struct
);
171 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
174 * init_srcu_struct - initialize a sleep-RCU structure
175 * @sp: structure to initialize.
177 * Must invoke this on a given srcu_struct before passing that srcu_struct
178 * to any other function. Each srcu_struct represents a separate domain
179 * of SRCU protection.
181 int init_srcu_struct(struct srcu_struct
*sp
)
183 spin_lock_init(&sp
->gp_lock
);
184 return init_srcu_struct_fields(sp
, false);
186 EXPORT_SYMBOL_GPL(init_srcu_struct
);
188 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
191 * First-use initialization of statically allocated srcu_struct
192 * structure. Wiring up the combining tree is more than can be
193 * done with compile-time initialization, so this check is added
194 * to each update-side SRCU primitive. Use ->gp_lock, which -is-
195 * compile-time initialized, to resolve races involving multiple
196 * CPUs trying to garner first-use privileges.
198 static void check_init_srcu_struct(struct srcu_struct
*sp
)
202 WARN_ON_ONCE(rcu_scheduler_active
== RCU_SCHEDULER_INIT
);
203 /* The smp_load_acquire() pairs with the smp_store_release(). */
204 if (!rcu_seq_state(smp_load_acquire(&sp
->srcu_gp_seq_needed
))) /*^^^*/
205 return; /* Already initialized. */
206 spin_lock_irqsave(&sp
->gp_lock
, flags
);
207 if (!rcu_seq_state(sp
->srcu_gp_seq_needed
)) {
208 spin_unlock_irqrestore(&sp
->gp_lock
, flags
);
211 init_srcu_struct_fields(sp
, true);
212 spin_unlock_irqrestore(&sp
->gp_lock
, flags
);
216 * Returns approximate total of the readers' ->srcu_lock_count[] values
217 * for the rank of per-CPU counters specified by idx.
219 static unsigned long srcu_readers_lock_idx(struct srcu_struct
*sp
, int idx
)
222 unsigned long sum
= 0;
224 for_each_possible_cpu(cpu
) {
225 struct srcu_data
*cpuc
= per_cpu_ptr(sp
->sda
, cpu
);
227 sum
+= READ_ONCE(cpuc
->srcu_lock_count
[idx
]);
233 * Returns approximate total of the readers' ->srcu_unlock_count[] values
234 * for the rank of per-CPU counters specified by idx.
236 static unsigned long srcu_readers_unlock_idx(struct srcu_struct
*sp
, int idx
)
239 unsigned long sum
= 0;
241 for_each_possible_cpu(cpu
) {
242 struct srcu_data
*cpuc
= per_cpu_ptr(sp
->sda
, cpu
);
244 sum
+= READ_ONCE(cpuc
->srcu_unlock_count
[idx
]);
250 * Return true if the number of pre-existing readers is determined to
253 static bool srcu_readers_active_idx_check(struct srcu_struct
*sp
, int idx
)
255 unsigned long unlocks
;
257 unlocks
= srcu_readers_unlock_idx(sp
, idx
);
260 * Make sure that a lock is always counted if the corresponding
261 * unlock is counted. Needs to be a smp_mb() as the read side may
262 * contain a read from a variable that is written to before the
263 * synchronize_srcu() in the write side. In this case smp_mb()s
264 * A and B act like the store buffering pattern.
266 * This smp_mb() also pairs with smp_mb() C to prevent accesses
267 * after the synchronize_srcu() from being executed before the
273 * If the locks are the same as the unlocks, then there must have
274 * been no readers on this index at some time in between. This does
275 * not mean that there are no more readers, as one could have read
276 * the current index but not have incremented the lock counter yet.
278 * Possible bug: There is no guarantee that there haven't been
279 * ULONG_MAX increments of ->srcu_lock_count[] since the unlocks were
280 * counted, meaning that this could return true even if there are
281 * still active readers. Since there are no memory barriers around
282 * srcu_flip(), the CPU is not required to increment ->srcu_idx
283 * before running srcu_readers_unlock_idx(), which means that there
284 * could be an arbitrarily large number of critical sections that
285 * execute after srcu_readers_unlock_idx() but use the old value
288 return srcu_readers_lock_idx(sp
, idx
) == unlocks
;
292 * srcu_readers_active - returns true if there are readers. and false
294 * @sp: which srcu_struct to count active readers (holding srcu_read_lock).
296 * Note that this is not an atomic primitive, and can therefore suffer
297 * severe errors when invoked on an active srcu_struct. That said, it
298 * can be useful as an error check at cleanup time.
300 static bool srcu_readers_active(struct srcu_struct
*sp
)
303 unsigned long sum
= 0;
305 for_each_possible_cpu(cpu
) {
306 struct srcu_data
*cpuc
= per_cpu_ptr(sp
->sda
, cpu
);
308 sum
+= READ_ONCE(cpuc
->srcu_lock_count
[0]);
309 sum
+= READ_ONCE(cpuc
->srcu_lock_count
[1]);
310 sum
-= READ_ONCE(cpuc
->srcu_unlock_count
[0]);
311 sum
-= READ_ONCE(cpuc
->srcu_unlock_count
[1]);
316 #define SRCU_INTERVAL 1
319 * Return grace-period delay, zero if there are expedited grace
320 * periods pending, SRCU_INTERVAL otherwise.
322 static unsigned long srcu_get_delay(struct srcu_struct
*sp
)
324 if (ULONG_CMP_LT(READ_ONCE(sp
->srcu_gp_seq
),
325 READ_ONCE(sp
->srcu_gp_seq_needed_exp
)))
327 return SRCU_INTERVAL
;
331 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
332 * @sp: structure to clean up.
334 * Must invoke this after you are finished using a given srcu_struct that
335 * was initialized via init_srcu_struct(), else you leak memory.
337 void cleanup_srcu_struct(struct srcu_struct
*sp
)
341 if (WARN_ON(!srcu_get_delay(sp
)))
342 return; /* Leakage unless caller handles error. */
343 if (WARN_ON(srcu_readers_active(sp
)))
344 return; /* Leakage unless caller handles error. */
345 flush_delayed_work(&sp
->work
);
346 for_each_possible_cpu(cpu
)
347 flush_delayed_work(&per_cpu_ptr(sp
->sda
, cpu
)->work
);
348 if (WARN_ON(rcu_seq_state(READ_ONCE(sp
->srcu_gp_seq
)) != SRCU_STATE_IDLE
) ||
349 WARN_ON(srcu_readers_active(sp
))) {
350 pr_info("cleanup_srcu_struct: Active srcu_struct %p state: %d\n", sp
, rcu_seq_state(READ_ONCE(sp
->srcu_gp_seq
)));
351 return; /* Caller forgot to stop doing call_srcu()? */
353 free_percpu(sp
->sda
);
356 EXPORT_SYMBOL_GPL(cleanup_srcu_struct
);
359 * Counts the new reader in the appropriate per-CPU element of the
361 * Returns an index that must be passed to the matching srcu_read_unlock().
363 int __srcu_read_lock(struct srcu_struct
*sp
)
367 idx
= READ_ONCE(sp
->srcu_idx
) & 0x1;
368 this_cpu_inc(sp
->sda
->srcu_lock_count
[idx
]);
369 smp_mb(); /* B */ /* Avoid leaking the critical section. */
372 EXPORT_SYMBOL_GPL(__srcu_read_lock
);
375 * Removes the count for the old reader from the appropriate per-CPU
376 * element of the srcu_struct. Note that this may well be a different
377 * CPU than that which was incremented by the corresponding srcu_read_lock().
379 void __srcu_read_unlock(struct srcu_struct
*sp
, int idx
)
381 smp_mb(); /* C */ /* Avoid leaking the critical section. */
382 this_cpu_inc(sp
->sda
->srcu_unlock_count
[idx
]);
384 EXPORT_SYMBOL_GPL(__srcu_read_unlock
);
387 * We use an adaptive strategy for synchronize_srcu() and especially for
388 * synchronize_srcu_expedited(). We spin for a fixed time period
389 * (defined below) to allow SRCU readers to exit their read-side critical
390 * sections. If there are still some readers after a few microseconds,
391 * we repeatedly block for 1-millisecond time periods.
393 #define SRCU_RETRY_CHECK_DELAY 5
396 * Start an SRCU grace period.
398 static void srcu_gp_start(struct srcu_struct
*sp
)
400 struct srcu_data
*sdp
= this_cpu_ptr(sp
->sda
);
403 RCU_LOCKDEP_WARN(!lockdep_is_held(&sp
->gp_lock
),
404 "Invoked srcu_gp_start() without ->gp_lock!");
405 WARN_ON_ONCE(ULONG_CMP_GE(sp
->srcu_gp_seq
, sp
->srcu_gp_seq_needed
));
406 rcu_segcblist_advance(&sdp
->srcu_cblist
,
407 rcu_seq_current(&sp
->srcu_gp_seq
));
408 (void)rcu_segcblist_accelerate(&sdp
->srcu_cblist
,
409 rcu_seq_snap(&sp
->srcu_gp_seq
));
410 smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
411 rcu_seq_start(&sp
->srcu_gp_seq
);
412 state
= rcu_seq_state(READ_ONCE(sp
->srcu_gp_seq
));
413 WARN_ON_ONCE(state
!= SRCU_STATE_SCAN1
);
417 * Track online CPUs to guide callback workqueue placement.
419 DEFINE_PER_CPU(bool, srcu_online
);
421 void srcu_online_cpu(unsigned int cpu
)
423 WRITE_ONCE(per_cpu(srcu_online
, cpu
), true);
426 void srcu_offline_cpu(unsigned int cpu
)
428 WRITE_ONCE(per_cpu(srcu_online
, cpu
), false);
432 * Place the workqueue handler on the specified CPU if online, otherwise
433 * just run it whereever. This is useful for placing workqueue handlers
434 * that are to invoke the specified CPU's callbacks.
436 static bool srcu_queue_delayed_work_on(int cpu
, struct workqueue_struct
*wq
,
437 struct delayed_work
*dwork
,
443 if (READ_ONCE(per_cpu(srcu_online
, cpu
)))
444 ret
= queue_delayed_work_on(cpu
, wq
, dwork
, delay
);
446 ret
= queue_delayed_work(wq
, dwork
, delay
);
452 * Schedule callback invocation for the specified srcu_data structure,
453 * if possible, on the corresponding CPU.
455 static void srcu_schedule_cbs_sdp(struct srcu_data
*sdp
, unsigned long delay
)
457 srcu_queue_delayed_work_on(sdp
->cpu
, system_power_efficient_wq
,
462 * Schedule callback invocation for all srcu_data structures associated
463 * with the specified srcu_node structure that have callbacks for the
464 * just-completed grace period, the one corresponding to idx. If possible,
465 * schedule this invocation on the corresponding CPUs.
467 static void srcu_schedule_cbs_snp(struct srcu_struct
*sp
, struct srcu_node
*snp
,
468 unsigned long mask
, unsigned long delay
)
472 for (cpu
= snp
->grplo
; cpu
<= snp
->grphi
; cpu
++) {
473 if (!(mask
& (1 << (cpu
- snp
->grplo
))))
475 srcu_schedule_cbs_sdp(per_cpu_ptr(sp
->sda
, cpu
), delay
);
480 * Note the end of an SRCU grace period. Initiates callback invocation
481 * and starts a new grace period if needed.
483 * The ->srcu_cb_mutex acquisition does not protect any data, but
484 * instead prevents more than one grace period from starting while we
485 * are initiating callback invocation. This allows the ->srcu_have_cbs[]
486 * array to have a finite number of elements.
488 static void srcu_gp_end(struct srcu_struct
*sp
)
490 unsigned long cbdelay
;
496 struct srcu_node
*snp
;
498 /* Prevent more than one additional grace period. */
499 mutex_lock(&sp
->srcu_cb_mutex
);
501 /* End the current grace period. */
502 spin_lock_irq(&sp
->gp_lock
);
503 idx
= rcu_seq_state(sp
->srcu_gp_seq
);
504 WARN_ON_ONCE(idx
!= SRCU_STATE_SCAN2
);
505 cbdelay
= srcu_get_delay(sp
);
506 sp
->srcu_last_gp_end
= ktime_get_mono_fast_ns();
507 rcu_seq_end(&sp
->srcu_gp_seq
);
508 gpseq
= rcu_seq_current(&sp
->srcu_gp_seq
);
509 if (ULONG_CMP_LT(sp
->srcu_gp_seq_needed_exp
, gpseq
))
510 sp
->srcu_gp_seq_needed_exp
= gpseq
;
511 spin_unlock_irq(&sp
->gp_lock
);
512 mutex_unlock(&sp
->srcu_gp_mutex
);
513 /* A new grace period can start at this point. But only one. */
515 /* Initiate callback invocation as needed. */
516 idx
= rcu_seq_ctr(gpseq
) % ARRAY_SIZE(snp
->srcu_have_cbs
);
517 idxnext
= (idx
+ 1) % ARRAY_SIZE(snp
->srcu_have_cbs
);
518 rcu_for_each_node_breadth_first(sp
, snp
) {
519 spin_lock_irq(&snp
->lock
);
521 if (snp
>= sp
->level
[rcu_num_lvls
- 1])
522 cbs
= snp
->srcu_have_cbs
[idx
] == gpseq
;
523 snp
->srcu_have_cbs
[idx
] = gpseq
;
524 rcu_seq_set_state(&snp
->srcu_have_cbs
[idx
], 1);
525 if (ULONG_CMP_LT(snp
->srcu_gp_seq_needed_exp
, gpseq
))
526 snp
->srcu_gp_seq_needed_exp
= gpseq
;
527 mask
= snp
->srcu_data_have_cbs
[idx
];
528 snp
->srcu_data_have_cbs
[idx
] = 0;
529 spin_unlock_irq(&snp
->lock
);
531 smp_mb(); /* GP end before CB invocation. */
532 srcu_schedule_cbs_snp(sp
, snp
, mask
, cbdelay
);
536 /* Callback initiation done, allow grace periods after next. */
537 mutex_unlock(&sp
->srcu_cb_mutex
);
539 /* Start a new grace period if needed. */
540 spin_lock_irq(&sp
->gp_lock
);
541 gpseq
= rcu_seq_current(&sp
->srcu_gp_seq
);
542 if (!rcu_seq_state(gpseq
) &&
543 ULONG_CMP_LT(gpseq
, sp
->srcu_gp_seq_needed
)) {
545 spin_unlock_irq(&sp
->gp_lock
);
546 /* Throttle expedited grace periods: Should be rare! */
547 srcu_reschedule(sp
, rcu_seq_ctr(gpseq
) & 0x3ff
548 ? 0 : SRCU_INTERVAL
);
550 spin_unlock_irq(&sp
->gp_lock
);
555 * Funnel-locking scheme to scalably mediate many concurrent expedited
556 * grace-period requests. This function is invoked for the first known
557 * expedited request for a grace period that has already been requested,
558 * but without expediting. To start a completely new grace period,
559 * whether expedited or not, use srcu_funnel_gp_start() instead.
561 static void srcu_funnel_exp_start(struct srcu_struct
*sp
, struct srcu_node
*snp
,
566 for (; snp
!= NULL
; snp
= snp
->srcu_parent
) {
567 if (rcu_seq_done(&sp
->srcu_gp_seq
, s
) ||
568 ULONG_CMP_GE(READ_ONCE(snp
->srcu_gp_seq_needed_exp
), s
))
570 spin_lock_irqsave(&snp
->lock
, flags
);
571 if (ULONG_CMP_GE(snp
->srcu_gp_seq_needed_exp
, s
)) {
572 spin_unlock_irqrestore(&snp
->lock
, flags
);
575 WRITE_ONCE(snp
->srcu_gp_seq_needed_exp
, s
);
576 spin_unlock_irqrestore(&snp
->lock
, flags
);
578 spin_lock_irqsave(&sp
->gp_lock
, flags
);
579 if (!ULONG_CMP_LT(sp
->srcu_gp_seq_needed_exp
, s
))
580 sp
->srcu_gp_seq_needed_exp
= s
;
581 spin_unlock_irqrestore(&sp
->gp_lock
, flags
);
585 * Funnel-locking scheme to scalably mediate many concurrent grace-period
586 * requests. The winner has to do the work of actually starting grace
587 * period s. Losers must either ensure that their desired grace-period
588 * number is recorded on at least their leaf srcu_node structure, or they
589 * must take steps to invoke their own callbacks.
591 static void srcu_funnel_gp_start(struct srcu_struct
*sp
, struct srcu_data
*sdp
,
592 unsigned long s
, bool do_norm
)
595 int idx
= rcu_seq_ctr(s
) % ARRAY_SIZE(sdp
->mynode
->srcu_have_cbs
);
596 struct srcu_node
*snp
= sdp
->mynode
;
597 unsigned long snp_seq
;
599 /* Each pass through the loop does one level of the srcu_node tree. */
600 for (; snp
!= NULL
; snp
= snp
->srcu_parent
) {
601 if (rcu_seq_done(&sp
->srcu_gp_seq
, s
) && snp
!= sdp
->mynode
)
602 return; /* GP already done and CBs recorded. */
603 spin_lock_irqsave(&snp
->lock
, flags
);
604 if (ULONG_CMP_GE(snp
->srcu_have_cbs
[idx
], s
)) {
605 snp_seq
= snp
->srcu_have_cbs
[idx
];
606 if (snp
== sdp
->mynode
&& snp_seq
== s
)
607 snp
->srcu_data_have_cbs
[idx
] |= sdp
->grpmask
;
608 spin_unlock_irqrestore(&snp
->lock
, flags
);
609 if (snp
== sdp
->mynode
&& snp_seq
!= s
) {
610 smp_mb(); /* CBs after GP! */
611 srcu_schedule_cbs_sdp(sdp
, do_norm
617 srcu_funnel_exp_start(sp
, snp
, s
);
620 snp
->srcu_have_cbs
[idx
] = s
;
621 if (snp
== sdp
->mynode
)
622 snp
->srcu_data_have_cbs
[idx
] |= sdp
->grpmask
;
623 if (!do_norm
&& ULONG_CMP_LT(snp
->srcu_gp_seq_needed_exp
, s
))
624 snp
->srcu_gp_seq_needed_exp
= s
;
625 spin_unlock_irqrestore(&snp
->lock
, flags
);
628 /* Top of tree, must ensure the grace period will be started. */
629 spin_lock_irqsave(&sp
->gp_lock
, flags
);
630 if (ULONG_CMP_LT(sp
->srcu_gp_seq_needed
, s
)) {
632 * Record need for grace period s. Pair with load
633 * acquire setting up for initialization.
635 smp_store_release(&sp
->srcu_gp_seq_needed
, s
); /*^^^*/
637 if (!do_norm
&& ULONG_CMP_LT(sp
->srcu_gp_seq_needed_exp
, s
))
638 sp
->srcu_gp_seq_needed_exp
= s
;
640 /* If grace period not already done and none in progress, start it. */
641 if (!rcu_seq_done(&sp
->srcu_gp_seq
, s
) &&
642 rcu_seq_state(sp
->srcu_gp_seq
) == SRCU_STATE_IDLE
) {
643 WARN_ON_ONCE(ULONG_CMP_GE(sp
->srcu_gp_seq
, sp
->srcu_gp_seq_needed
));
645 queue_delayed_work(system_power_efficient_wq
, &sp
->work
,
648 spin_unlock_irqrestore(&sp
->gp_lock
, flags
);
652 * Wait until all readers counted by array index idx complete, but
653 * loop an additional time if there is an expedited grace period pending.
654 * The caller must ensure that ->srcu_idx is not changed while checking.
656 static bool try_check_zero(struct srcu_struct
*sp
, int idx
, int trycount
)
659 if (srcu_readers_active_idx_check(sp
, idx
))
661 if (--trycount
+ !srcu_get_delay(sp
) <= 0)
663 udelay(SRCU_RETRY_CHECK_DELAY
);
668 * Increment the ->srcu_idx counter so that future SRCU readers will
669 * use the other rank of the ->srcu_(un)lock_count[] arrays. This allows
670 * us to wait for pre-existing readers in a starvation-free manner.
672 static void srcu_flip(struct srcu_struct
*sp
)
674 WRITE_ONCE(sp
->srcu_idx
, sp
->srcu_idx
+ 1);
677 * Ensure that if the updater misses an __srcu_read_unlock()
678 * increment, that task's next __srcu_read_lock() will see the
679 * above counter update. Note that both this memory barrier
680 * and the one in srcu_readers_active_idx_check() provide the
681 * guarantee for __srcu_read_lock().
683 smp_mb(); /* D */ /* Pairs with C. */
687 * If SRCU is likely idle, return true, otherwise return false.
689 * Note that it is OK for several current from-idle requests for a new
690 * grace period from idle to specify expediting because they will all end
691 * up requesting the same grace period anyhow. So no loss.
693 * Note also that if any CPU (including the current one) is still invoking
694 * callbacks, this function will nevertheless say "idle". This is not
695 * ideal, but the overhead of checking all CPUs' callback lists is even
696 * less ideal, especially on large systems. Furthermore, the wakeup
697 * can happen before the callback is fully removed, so we have no choice
698 * but to accept this type of error.
700 * This function is also subject to counter-wrap errors, but let's face
701 * it, if this function was preempted for enough time for the counters
702 * to wrap, it really doesn't matter whether or not we expedite the grace
703 * period. The extra overhead of a needlessly expedited grace period is
704 * negligible when amoritized over that time period, and the extra latency
705 * of a needlessly non-expedited grace period is similarly negligible.
707 static bool srcu_might_be_idle(struct srcu_struct
*sp
)
709 unsigned long curseq
;
711 struct srcu_data
*sdp
;
714 /* If the local srcu_data structure has callbacks, not idle. */
715 local_irq_save(flags
);
716 sdp
= this_cpu_ptr(sp
->sda
);
717 if (rcu_segcblist_pend_cbs(&sdp
->srcu_cblist
)) {
718 local_irq_restore(flags
);
719 return false; /* Callbacks already present, so not idle. */
721 local_irq_restore(flags
);
724 * No local callbacks, so probabalistically probe global state.
725 * Exact information would require acquiring locks, which would
726 * kill scalability, hence the probabalistic nature of the probe.
729 /* First, see if enough time has passed since the last GP. */
730 t
= ktime_get_mono_fast_ns();
731 if (exp_holdoff
== 0 ||
732 time_in_range_open(t
, sp
->srcu_last_gp_end
,
733 sp
->srcu_last_gp_end
+ exp_holdoff
))
734 return false; /* Too soon after last GP. */
736 /* Next, check for probable idleness. */
737 curseq
= rcu_seq_current(&sp
->srcu_gp_seq
);
738 smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
739 if (ULONG_CMP_LT(curseq
, READ_ONCE(sp
->srcu_gp_seq_needed
)))
740 return false; /* Grace period in progress, so not idle. */
741 smp_mb(); /* Order ->srcu_gp_seq with prior access. */
742 if (curseq
!= rcu_seq_current(&sp
->srcu_gp_seq
))
743 return false; /* GP # changed, so not idle. */
744 return true; /* With reasonable probability, idle! */
748 * Enqueue an SRCU callback on the srcu_data structure associated with
749 * the current CPU and the specified srcu_struct structure, initiating
750 * grace-period processing if it is not already running.
752 * Note that all CPUs must agree that the grace period extended beyond
753 * all pre-existing SRCU read-side critical section. On systems with
754 * more than one CPU, this means that when "func()" is invoked, each CPU
755 * is guaranteed to have executed a full memory barrier since the end of
756 * its last corresponding SRCU read-side critical section whose beginning
757 * preceded the call to call_rcu(). It also means that each CPU executing
758 * an SRCU read-side critical section that continues beyond the start of
759 * "func()" must have executed a memory barrier after the call_rcu()
760 * but before the beginning of that SRCU read-side critical section.
761 * Note that these guarantees include CPUs that are offline, idle, or
762 * executing in user mode, as well as CPUs that are executing in the kernel.
764 * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the
765 * resulting SRCU callback function "func()", then both CPU A and CPU
766 * B are guaranteed to execute a full memory barrier during the time
767 * interval between the call to call_rcu() and the invocation of "func()".
768 * This guarantee applies even if CPU A and CPU B are the same CPU (but
769 * again only if the system has more than one CPU).
771 * Of course, these guarantees apply only for invocations of call_srcu(),
772 * srcu_read_lock(), and srcu_read_unlock() that are all passed the same
773 * srcu_struct structure.
775 void __call_srcu(struct srcu_struct
*sp
, struct rcu_head
*rhp
,
776 rcu_callback_t func
, bool do_norm
)
779 bool needexp
= false;
782 struct srcu_data
*sdp
;
784 check_init_srcu_struct(sp
);
786 local_irq_save(flags
);
787 sdp
= this_cpu_ptr(sp
->sda
);
788 spin_lock(&sdp
->lock
);
789 rcu_segcblist_enqueue(&sdp
->srcu_cblist
, rhp
, false);
790 rcu_segcblist_advance(&sdp
->srcu_cblist
,
791 rcu_seq_current(&sp
->srcu_gp_seq
));
792 s
= rcu_seq_snap(&sp
->srcu_gp_seq
);
793 (void)rcu_segcblist_accelerate(&sdp
->srcu_cblist
, s
);
794 if (ULONG_CMP_LT(sdp
->srcu_gp_seq_needed
, s
)) {
795 sdp
->srcu_gp_seq_needed
= s
;
798 if (!do_norm
&& ULONG_CMP_LT(sdp
->srcu_gp_seq_needed_exp
, s
)) {
799 sdp
->srcu_gp_seq_needed_exp
= s
;
802 spin_unlock_irqrestore(&sdp
->lock
, flags
);
804 srcu_funnel_gp_start(sp
, sdp
, s
, do_norm
);
806 srcu_funnel_exp_start(sp
, sdp
->mynode
, s
);
809 void call_srcu(struct srcu_struct
*sp
, struct rcu_head
*rhp
,
812 __call_srcu(sp
, rhp
, func
, true);
814 EXPORT_SYMBOL_GPL(call_srcu
);
817 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
819 static void __synchronize_srcu(struct srcu_struct
*sp
, bool do_norm
)
821 struct rcu_synchronize rcu
;
823 RCU_LOCKDEP_WARN(lock_is_held(&sp
->dep_map
) ||
824 lock_is_held(&rcu_bh_lock_map
) ||
825 lock_is_held(&rcu_lock_map
) ||
826 lock_is_held(&rcu_sched_lock_map
),
827 "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
829 if (rcu_scheduler_active
== RCU_SCHEDULER_INACTIVE
)
832 check_init_srcu_struct(sp
);
833 init_completion(&rcu
.completion
);
834 init_rcu_head_on_stack(&rcu
.head
);
835 __call_srcu(sp
, &rcu
.head
, wakeme_after_rcu
, do_norm
);
836 wait_for_completion(&rcu
.completion
);
837 destroy_rcu_head_on_stack(&rcu
.head
);
841 * synchronize_srcu_expedited - Brute-force SRCU grace period
842 * @sp: srcu_struct with which to synchronize.
844 * Wait for an SRCU grace period to elapse, but be more aggressive about
845 * spinning rather than blocking when waiting.
847 * Note that synchronize_srcu_expedited() has the same deadlock and
848 * memory-ordering properties as does synchronize_srcu().
850 void synchronize_srcu_expedited(struct srcu_struct
*sp
)
852 __synchronize_srcu(sp
, rcu_gp_is_normal());
854 EXPORT_SYMBOL_GPL(synchronize_srcu_expedited
);
857 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
858 * @sp: srcu_struct with which to synchronize.
860 * Wait for the count to drain to zero of both indexes. To avoid the
861 * possible starvation of synchronize_srcu(), it waits for the count of
862 * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
863 * and then flip the srcu_idx and wait for the count of the other index.
865 * Can block; must be called from process context.
867 * Note that it is illegal to call synchronize_srcu() from the corresponding
868 * SRCU read-side critical section; doing so will result in deadlock.
869 * However, it is perfectly legal to call synchronize_srcu() on one
870 * srcu_struct from some other srcu_struct's read-side critical section,
871 * as long as the resulting graph of srcu_structs is acyclic.
873 * There are memory-ordering constraints implied by synchronize_srcu().
874 * On systems with more than one CPU, when synchronize_srcu() returns,
875 * each CPU is guaranteed to have executed a full memory barrier since
876 * the end of its last corresponding SRCU-sched read-side critical section
877 * whose beginning preceded the call to synchronize_srcu(). In addition,
878 * each CPU having an SRCU read-side critical section that extends beyond
879 * the return from synchronize_srcu() is guaranteed to have executed a
880 * full memory barrier after the beginning of synchronize_srcu() and before
881 * the beginning of that SRCU read-side critical section. Note that these
882 * guarantees include CPUs that are offline, idle, or executing in user mode,
883 * as well as CPUs that are executing in the kernel.
885 * Furthermore, if CPU A invoked synchronize_srcu(), which returned
886 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
887 * to have executed a full memory barrier during the execution of
888 * synchronize_srcu(). This guarantee applies even if CPU A and CPU B
889 * are the same CPU, but again only if the system has more than one CPU.
891 * Of course, these memory-ordering guarantees apply only when
892 * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
893 * passed the same srcu_struct structure.
895 * If SRCU is likely idle, expedite the first request. This semantic
896 * was provided by Classic SRCU, and is relied upon by its users, so TREE
897 * SRCU must also provide it. Note that detecting idleness is heuristic
898 * and subject to both false positives and negatives.
900 void synchronize_srcu(struct srcu_struct
*sp
)
902 if (srcu_might_be_idle(sp
) || rcu_gp_is_expedited())
903 synchronize_srcu_expedited(sp
);
905 __synchronize_srcu(sp
, true);
907 EXPORT_SYMBOL_GPL(synchronize_srcu
);
910 * Callback function for srcu_barrier() use.
912 static void srcu_barrier_cb(struct rcu_head
*rhp
)
914 struct srcu_data
*sdp
;
915 struct srcu_struct
*sp
;
917 sdp
= container_of(rhp
, struct srcu_data
, srcu_barrier_head
);
919 if (atomic_dec_and_test(&sp
->srcu_barrier_cpu_cnt
))
920 complete(&sp
->srcu_barrier_completion
);
924 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
925 * @sp: srcu_struct on which to wait for in-flight callbacks.
927 void srcu_barrier(struct srcu_struct
*sp
)
930 struct srcu_data
*sdp
;
931 unsigned long s
= rcu_seq_snap(&sp
->srcu_barrier_seq
);
933 check_init_srcu_struct(sp
);
934 mutex_lock(&sp
->srcu_barrier_mutex
);
935 if (rcu_seq_done(&sp
->srcu_barrier_seq
, s
)) {
936 smp_mb(); /* Force ordering following return. */
937 mutex_unlock(&sp
->srcu_barrier_mutex
);
938 return; /* Someone else did our work for us. */
940 rcu_seq_start(&sp
->srcu_barrier_seq
);
941 init_completion(&sp
->srcu_barrier_completion
);
943 /* Initial count prevents reaching zero until all CBs are posted. */
944 atomic_set(&sp
->srcu_barrier_cpu_cnt
, 1);
947 * Each pass through this loop enqueues a callback, but only
948 * on CPUs already having callbacks enqueued. Note that if
949 * a CPU already has callbacks enqueue, it must have already
950 * registered the need for a future grace period, so all we
951 * need do is enqueue a callback that will use the same
952 * grace period as the last callback already in the queue.
954 for_each_possible_cpu(cpu
) {
955 sdp
= per_cpu_ptr(sp
->sda
, cpu
);
956 spin_lock_irq(&sdp
->lock
);
957 atomic_inc(&sp
->srcu_barrier_cpu_cnt
);
958 sdp
->srcu_barrier_head
.func
= srcu_barrier_cb
;
959 if (!rcu_segcblist_entrain(&sdp
->srcu_cblist
,
960 &sdp
->srcu_barrier_head
, 0))
961 atomic_dec(&sp
->srcu_barrier_cpu_cnt
);
962 spin_unlock_irq(&sdp
->lock
);
965 /* Remove the initial count, at which point reaching zero can happen. */
966 if (atomic_dec_and_test(&sp
->srcu_barrier_cpu_cnt
))
967 complete(&sp
->srcu_barrier_completion
);
968 wait_for_completion(&sp
->srcu_barrier_completion
);
970 rcu_seq_end(&sp
->srcu_barrier_seq
);
971 mutex_unlock(&sp
->srcu_barrier_mutex
);
973 EXPORT_SYMBOL_GPL(srcu_barrier
);
976 * srcu_batches_completed - return batches completed.
977 * @sp: srcu_struct on which to report batch completion.
979 * Report the number of batches, correlated with, but not necessarily
980 * precisely the same as, the number of grace periods that have elapsed.
982 unsigned long srcu_batches_completed(struct srcu_struct
*sp
)
986 EXPORT_SYMBOL_GPL(srcu_batches_completed
);
989 * Core SRCU state machine. Push state bits of ->srcu_gp_seq
990 * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
991 * completed in that state.
993 static void srcu_advance_state(struct srcu_struct
*sp
)
997 mutex_lock(&sp
->srcu_gp_mutex
);
1000 * Because readers might be delayed for an extended period after
1001 * fetching ->srcu_idx for their index, at any point in time there
1002 * might well be readers using both idx=0 and idx=1. We therefore
1003 * need to wait for readers to clear from both index values before
1004 * invoking a callback.
1006 * The load-acquire ensures that we see the accesses performed
1007 * by the prior grace period.
1009 idx
= rcu_seq_state(smp_load_acquire(&sp
->srcu_gp_seq
)); /* ^^^ */
1010 if (idx
== SRCU_STATE_IDLE
) {
1011 spin_lock_irq(&sp
->gp_lock
);
1012 if (ULONG_CMP_GE(sp
->srcu_gp_seq
, sp
->srcu_gp_seq_needed
)) {
1013 WARN_ON_ONCE(rcu_seq_state(sp
->srcu_gp_seq
));
1014 spin_unlock_irq(&sp
->gp_lock
);
1015 mutex_unlock(&sp
->srcu_gp_mutex
);
1018 idx
= rcu_seq_state(READ_ONCE(sp
->srcu_gp_seq
));
1019 if (idx
== SRCU_STATE_IDLE
)
1021 spin_unlock_irq(&sp
->gp_lock
);
1022 if (idx
!= SRCU_STATE_IDLE
) {
1023 mutex_unlock(&sp
->srcu_gp_mutex
);
1024 return; /* Someone else started the grace period. */
1028 if (rcu_seq_state(READ_ONCE(sp
->srcu_gp_seq
)) == SRCU_STATE_SCAN1
) {
1029 idx
= 1 ^ (sp
->srcu_idx
& 1);
1030 if (!try_check_zero(sp
, idx
, 1)) {
1031 mutex_unlock(&sp
->srcu_gp_mutex
);
1032 return; /* readers present, retry later. */
1035 rcu_seq_set_state(&sp
->srcu_gp_seq
, SRCU_STATE_SCAN2
);
1038 if (rcu_seq_state(READ_ONCE(sp
->srcu_gp_seq
)) == SRCU_STATE_SCAN2
) {
1041 * SRCU read-side critical sections are normally short,
1042 * so check at least twice in quick succession after a flip.
1044 idx
= 1 ^ (sp
->srcu_idx
& 1);
1045 if (!try_check_zero(sp
, idx
, 2)) {
1046 mutex_unlock(&sp
->srcu_gp_mutex
);
1047 return; /* readers present, retry later. */
1049 srcu_gp_end(sp
); /* Releases ->srcu_gp_mutex. */
1054 * Invoke a limited number of SRCU callbacks that have passed through
1055 * their grace period. If there are more to do, SRCU will reschedule
1056 * the workqueue. Note that needed memory barriers have been executed
1057 * in this task's context by srcu_readers_active_idx_check().
1059 static void srcu_invoke_callbacks(struct work_struct
*work
)
1062 struct rcu_cblist ready_cbs
;
1063 struct rcu_head
*rhp
;
1064 struct srcu_data
*sdp
;
1065 struct srcu_struct
*sp
;
1067 sdp
= container_of(work
, struct srcu_data
, work
.work
);
1069 rcu_cblist_init(&ready_cbs
);
1070 spin_lock_irq(&sdp
->lock
);
1071 smp_mb(); /* Old grace periods before callback invocation! */
1072 rcu_segcblist_advance(&sdp
->srcu_cblist
,
1073 rcu_seq_current(&sp
->srcu_gp_seq
));
1074 if (sdp
->srcu_cblist_invoking
||
1075 !rcu_segcblist_ready_cbs(&sdp
->srcu_cblist
)) {
1076 spin_unlock_irq(&sdp
->lock
);
1077 return; /* Someone else on the job or nothing to do. */
1080 /* We are on the job! Extract and invoke ready callbacks. */
1081 sdp
->srcu_cblist_invoking
= true;
1082 rcu_segcblist_extract_done_cbs(&sdp
->srcu_cblist
, &ready_cbs
);
1083 spin_unlock_irq(&sdp
->lock
);
1084 rhp
= rcu_cblist_dequeue(&ready_cbs
);
1085 for (; rhp
!= NULL
; rhp
= rcu_cblist_dequeue(&ready_cbs
)) {
1092 * Update counts, accelerate new callbacks, and if needed,
1093 * schedule another round of callback invocation.
1095 spin_lock_irq(&sdp
->lock
);
1096 rcu_segcblist_insert_count(&sdp
->srcu_cblist
, &ready_cbs
);
1097 (void)rcu_segcblist_accelerate(&sdp
->srcu_cblist
,
1098 rcu_seq_snap(&sp
->srcu_gp_seq
));
1099 sdp
->srcu_cblist_invoking
= false;
1100 more
= rcu_segcblist_ready_cbs(&sdp
->srcu_cblist
);
1101 spin_unlock_irq(&sdp
->lock
);
1103 srcu_schedule_cbs_sdp(sdp
, 0);
1107 * Finished one round of SRCU grace period. Start another if there are
1108 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
1110 static void srcu_reschedule(struct srcu_struct
*sp
, unsigned long delay
)
1114 spin_lock_irq(&sp
->gp_lock
);
1115 if (ULONG_CMP_GE(sp
->srcu_gp_seq
, sp
->srcu_gp_seq_needed
)) {
1116 if (!WARN_ON_ONCE(rcu_seq_state(sp
->srcu_gp_seq
))) {
1117 /* All requests fulfilled, time to go idle. */
1120 } else if (!rcu_seq_state(sp
->srcu_gp_seq
)) {
1121 /* Outstanding request and no GP. Start one. */
1124 spin_unlock_irq(&sp
->gp_lock
);
1127 queue_delayed_work(system_power_efficient_wq
, &sp
->work
, delay
);
1131 * This is the work-queue function that handles SRCU grace periods.
1133 void process_srcu(struct work_struct
*work
)
1135 struct srcu_struct
*sp
;
1137 sp
= container_of(work
, struct srcu_struct
, work
.work
);
1139 srcu_advance_state(sp
);
1140 srcu_reschedule(sp
, srcu_get_delay(sp
));
1142 EXPORT_SYMBOL_GPL(process_srcu
);
1144 void srcutorture_get_gp_data(enum rcutorture_type test_type
,
1145 struct srcu_struct
*sp
, int *flags
,
1146 unsigned long *gpnum
, unsigned long *completed
)
1148 if (test_type
!= SRCU_FLAVOR
)
1151 *completed
= rcu_seq_ctr(sp
->srcu_gp_seq
);
1152 *gpnum
= rcu_seq_ctr(sp
->srcu_gp_seq_needed
);
1154 EXPORT_SYMBOL_GPL(srcutorture_get_gp_data
);