1 // SPDX-License-Identifier: GPL-2.0+
3 * Sleepable Read-Copy Update mechanism for mutual exclusion.
5 * Copyright (C) IBM Corporation, 2006
6 * Copyright (C) Fujitsu, 2012
8 * Authors: Paul McKenney <paulmck@linux.ibm.com>
9 * Lai Jiangshan <laijs@cn.fujitsu.com>
11 * For detailed explanation of Read-Copy Update mechanism see -
12 * Documentation/RCU/ *.txt
16 #define pr_fmt(fmt) "rcu: " fmt
18 #include <linux/export.h>
19 #include <linux/mutex.h>
20 #include <linux/percpu.h>
21 #include <linux/preempt.h>
22 #include <linux/rcupdate_wait.h>
23 #include <linux/sched.h>
24 #include <linux/smp.h>
25 #include <linux/delay.h>
26 #include <linux/module.h>
27 #include <linux/slab.h>
28 #include <linux/srcu.h>
31 #include "rcu_segcblist.h"
33 /* Holdoff in nanoseconds for auto-expediting. */
34 #define DEFAULT_SRCU_EXP_HOLDOFF (25 * 1000)
35 static ulong exp_holdoff = DEFAULT_SRCU_EXP_HOLDOFF;
36 module_param(exp_holdoff, ulong, 0444);
38 /* Overflow-check frequency. N bits roughly says every 2**N grace periods. */
39 static ulong counter_wrap_check = (ULONG_MAX >> 2);
40 module_param(counter_wrap_check, ulong, 0444);
43 * Control conversion to SRCU_SIZE_BIG:
44 * 0: Don't convert at all.
45 * 1: Convert at init_srcu_struct() time.
46 * 2: Convert when rcutorture invokes srcu_torture_stats_print().
47 * 3: Decide at boot time based on system shape (default).
48 * 0x1x: Convert when excessive contention encountered.
50 #define SRCU_SIZING_NONE 0
51 #define SRCU_SIZING_INIT 1
52 #define SRCU_SIZING_TORTURE 2
53 #define SRCU_SIZING_AUTO 3
54 #define SRCU_SIZING_CONTEND 0x10
55 #define SRCU_SIZING_IS(x) ((convert_to_big & ~SRCU_SIZING_CONTEND) == x)
56 #define SRCU_SIZING_IS_NONE() (SRCU_SIZING_IS(SRCU_SIZING_NONE))
57 #define SRCU_SIZING_IS_INIT() (SRCU_SIZING_IS(SRCU_SIZING_INIT))
58 #define SRCU_SIZING_IS_TORTURE() (SRCU_SIZING_IS(SRCU_SIZING_TORTURE))
59 #define SRCU_SIZING_IS_CONTEND() (convert_to_big & SRCU_SIZING_CONTEND)
60 static int convert_to_big = SRCU_SIZING_AUTO;
61 module_param(convert_to_big, int, 0444);
63 /* Number of CPUs to trigger init_srcu_struct()-time transition to big. */
64 static int big_cpu_lim __read_mostly = 128;
65 module_param(big_cpu_lim, int, 0444);
67 /* Contention events per jiffy to initiate transition to big. */
68 static int small_contention_lim __read_mostly = 100;
69 module_param(small_contention_lim, int, 0444);
71 /* Early-boot callback-management, so early that no lock is required! */
72 static LIST_HEAD(srcu_boot_list);
73 static bool __read_mostly srcu_init_done;
75 static void srcu_invoke_callbacks(struct work_struct *work);
76 static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay);
77 static void process_srcu(struct work_struct *work);
78 static void srcu_delay_timer(struct timer_list *t);
80 /* Wrappers for lock acquisition and release, see raw_spin_lock_rcu_node(). */
81 #define spin_lock_rcu_node(p) \
83 spin_lock(&ACCESS_PRIVATE(p, lock)); \
84 smp_mb__after_unlock_lock(); \
87 #define spin_unlock_rcu_node(p) spin_unlock(&ACCESS_PRIVATE(p, lock))
89 #define spin_lock_irq_rcu_node(p) \
91 spin_lock_irq(&ACCESS_PRIVATE(p, lock)); \
92 smp_mb__after_unlock_lock(); \
95 #define spin_unlock_irq_rcu_node(p) \
96 spin_unlock_irq(&ACCESS_PRIVATE(p, lock))
98 #define spin_lock_irqsave_rcu_node(p, flags) \
100 spin_lock_irqsave(&ACCESS_PRIVATE(p, lock), flags); \
101 smp_mb__after_unlock_lock(); \
104 #define spin_trylock_irqsave_rcu_node(p, flags) \
106 bool ___locked = spin_trylock_irqsave(&ACCESS_PRIVATE(p, lock), flags); \
109 smp_mb__after_unlock_lock(); \
113 #define spin_unlock_irqrestore_rcu_node(p, flags) \
114 spin_unlock_irqrestore(&ACCESS_PRIVATE(p, lock), flags) \
117 * Initialize SRCU per-CPU data. Note that statically allocated
118 * srcu_struct structures might already have srcu_read_lock() and
119 * srcu_read_unlock() running against them. So if the is_static parameter
120 * is set, don't initialize ->srcu_lock_count[] and ->srcu_unlock_count[].
122 static void init_srcu_struct_data(struct srcu_struct *ssp)
125 struct srcu_data *sdp;
128 * Initialize the per-CPU srcu_data array, which feeds into the
129 * leaves of the srcu_node tree.
131 WARN_ON_ONCE(ARRAY_SIZE(sdp->srcu_lock_count) !=
132 ARRAY_SIZE(sdp->srcu_unlock_count));
133 for_each_possible_cpu(cpu) {
134 sdp = per_cpu_ptr(ssp->sda, cpu);
135 spin_lock_init(&ACCESS_PRIVATE(sdp, lock));
136 rcu_segcblist_init(&sdp->srcu_cblist);
137 sdp->srcu_cblist_invoking = false;
138 sdp->srcu_gp_seq_needed = ssp->srcu_gp_seq;
139 sdp->srcu_gp_seq_needed_exp = ssp->srcu_gp_seq;
142 INIT_WORK(&sdp->work, srcu_invoke_callbacks);
143 timer_setup(&sdp->delay_work, srcu_delay_timer, 0);
148 /* Invalid seq state, used during snp node initialization */
149 #define SRCU_SNP_INIT_SEQ 0x2
152 * Check whether sequence number corresponding to snp node,
155 static inline bool srcu_invl_snp_seq(unsigned long s)
157 return rcu_seq_state(s) == SRCU_SNP_INIT_SEQ;
161 * Allocated and initialize SRCU combining tree. Returns @true if
162 * allocation succeeded and @false otherwise.
164 static bool init_srcu_struct_nodes(struct srcu_struct *ssp, gfp_t gfp_flags)
169 int levelspread[RCU_NUM_LVLS];
170 struct srcu_data *sdp;
171 struct srcu_node *snp;
172 struct srcu_node *snp_first;
174 /* Initialize geometry if it has not already been initialized. */
176 ssp->node = kcalloc(rcu_num_nodes, sizeof(*ssp->node), gfp_flags);
180 /* Work out the overall tree geometry. */
181 ssp->level[0] = &ssp->node[0];
182 for (i = 1; i < rcu_num_lvls; i++)
183 ssp->level[i] = ssp->level[i - 1] + num_rcu_lvl[i - 1];
184 rcu_init_levelspread(levelspread, num_rcu_lvl);
186 /* Each pass through this loop initializes one srcu_node structure. */
187 srcu_for_each_node_breadth_first(ssp, snp) {
188 spin_lock_init(&ACCESS_PRIVATE(snp, lock));
189 WARN_ON_ONCE(ARRAY_SIZE(snp->srcu_have_cbs) !=
190 ARRAY_SIZE(snp->srcu_data_have_cbs));
191 for (i = 0; i < ARRAY_SIZE(snp->srcu_have_cbs); i++) {
192 snp->srcu_have_cbs[i] = SRCU_SNP_INIT_SEQ;
193 snp->srcu_data_have_cbs[i] = 0;
195 snp->srcu_gp_seq_needed_exp = SRCU_SNP_INIT_SEQ;
198 if (snp == &ssp->node[0]) {
199 /* Root node, special case. */
200 snp->srcu_parent = NULL;
205 if (snp == ssp->level[level + 1])
207 snp->srcu_parent = ssp->level[level - 1] +
208 (snp - ssp->level[level]) /
209 levelspread[level - 1];
213 * Initialize the per-CPU srcu_data array, which feeds into the
214 * leaves of the srcu_node tree.
216 level = rcu_num_lvls - 1;
217 snp_first = ssp->level[level];
218 for_each_possible_cpu(cpu) {
219 sdp = per_cpu_ptr(ssp->sda, cpu);
220 sdp->mynode = &snp_first[cpu / levelspread[level]];
221 for (snp = sdp->mynode; snp != NULL; snp = snp->srcu_parent) {
226 sdp->grpmask = 1 << (cpu - sdp->mynode->grplo);
228 smp_store_release(&ssp->srcu_size_state, SRCU_SIZE_WAIT_BARRIER);
233 * Initialize non-compile-time initialized fields, including the
234 * associated srcu_node and srcu_data structures. The is_static parameter
235 * tells us that ->sda has already been wired up to srcu_data.
237 static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
239 ssp->srcu_size_state = SRCU_SIZE_SMALL;
241 mutex_init(&ssp->srcu_cb_mutex);
242 mutex_init(&ssp->srcu_gp_mutex);
244 ssp->srcu_gp_seq = 0;
245 ssp->srcu_barrier_seq = 0;
246 mutex_init(&ssp->srcu_barrier_mutex);
247 atomic_set(&ssp->srcu_barrier_cpu_cnt, 0);
248 INIT_DELAYED_WORK(&ssp->work, process_srcu);
249 ssp->sda_is_static = is_static;
251 ssp->sda = alloc_percpu(struct srcu_data);
254 init_srcu_struct_data(ssp);
255 ssp->srcu_gp_seq_needed_exp = 0;
256 ssp->srcu_last_gp_end = ktime_get_mono_fast_ns();
257 if (READ_ONCE(ssp->srcu_size_state) == SRCU_SIZE_SMALL && SRCU_SIZING_IS_INIT()) {
258 if (!init_srcu_struct_nodes(ssp, GFP_ATOMIC)) {
259 if (!ssp->sda_is_static) {
260 free_percpu(ssp->sda);
265 WRITE_ONCE(ssp->srcu_size_state, SRCU_SIZE_BIG);
268 smp_store_release(&ssp->srcu_gp_seq_needed, 0); /* Init done. */
272 #ifdef CONFIG_DEBUG_LOCK_ALLOC
274 int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
275 struct lock_class_key *key)
277 /* Don't re-initialize a lock while it is held. */
278 debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
279 lockdep_init_map(&ssp->dep_map, name, key, 0);
280 spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
281 return init_srcu_struct_fields(ssp, false);
283 EXPORT_SYMBOL_GPL(__init_srcu_struct);
285 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
288 * init_srcu_struct - initialize a sleep-RCU structure
289 * @ssp: structure to initialize.
291 * Must invoke this on a given srcu_struct before passing that srcu_struct
292 * to any other function. Each srcu_struct represents a separate domain
293 * of SRCU protection.
295 int init_srcu_struct(struct srcu_struct *ssp)
297 spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
298 return init_srcu_struct_fields(ssp, false);
300 EXPORT_SYMBOL_GPL(init_srcu_struct);
302 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
305 * Initiate a transition to SRCU_SIZE_BIG with lock held.
307 static void __srcu_transition_to_big(struct srcu_struct *ssp)
309 lockdep_assert_held(&ACCESS_PRIVATE(ssp, lock));
310 smp_store_release(&ssp->srcu_size_state, SRCU_SIZE_ALLOC);
314 * Initiate an idempotent transition to SRCU_SIZE_BIG.
316 static void srcu_transition_to_big(struct srcu_struct *ssp)
320 /* Double-checked locking on ->srcu_size-state. */
321 if (smp_load_acquire(&ssp->srcu_size_state) != SRCU_SIZE_SMALL)
323 spin_lock_irqsave_rcu_node(ssp, flags);
324 if (smp_load_acquire(&ssp->srcu_size_state) != SRCU_SIZE_SMALL) {
325 spin_unlock_irqrestore_rcu_node(ssp, flags);
328 __srcu_transition_to_big(ssp);
329 spin_unlock_irqrestore_rcu_node(ssp, flags);
333 * Check to see if the just-encountered contention event justifies
334 * a transition to SRCU_SIZE_BIG.
336 static void spin_lock_irqsave_check_contention(struct srcu_struct *ssp)
340 if (!SRCU_SIZING_IS_CONTEND() || ssp->srcu_size_state)
343 if (ssp->srcu_size_jiffies != j) {
344 ssp->srcu_size_jiffies = j;
345 ssp->srcu_n_lock_retries = 0;
347 if (++ssp->srcu_n_lock_retries <= small_contention_lim)
349 __srcu_transition_to_big(ssp);
353 * Acquire the specified srcu_data structure's ->lock, but check for
354 * excessive contention, which results in initiation of a transition
355 * to SRCU_SIZE_BIG. But only if the srcutree.convert_to_big module
356 * parameter permits this.
358 static void spin_lock_irqsave_sdp_contention(struct srcu_data *sdp, unsigned long *flags)
360 struct srcu_struct *ssp = sdp->ssp;
362 if (spin_trylock_irqsave_rcu_node(sdp, *flags))
364 spin_lock_irqsave_rcu_node(ssp, *flags);
365 spin_lock_irqsave_check_contention(ssp);
366 spin_unlock_irqrestore_rcu_node(ssp, *flags);
367 spin_lock_irqsave_rcu_node(sdp, *flags);
371 * Acquire the specified srcu_struct structure's ->lock, but check for
372 * excessive contention, which results in initiation of a transition
373 * to SRCU_SIZE_BIG. But only if the srcutree.convert_to_big module
374 * parameter permits this.
376 static void spin_lock_irqsave_ssp_contention(struct srcu_struct *ssp, unsigned long *flags)
378 if (spin_trylock_irqsave_rcu_node(ssp, *flags))
380 spin_lock_irqsave_rcu_node(ssp, *flags);
381 spin_lock_irqsave_check_contention(ssp);
385 * First-use initialization of statically allocated srcu_struct
386 * structure. Wiring up the combining tree is more than can be
387 * done with compile-time initialization, so this check is added
388 * to each update-side SRCU primitive. Use ssp->lock, which -is-
389 * compile-time initialized, to resolve races involving multiple
390 * CPUs trying to garner first-use privileges.
392 static void check_init_srcu_struct(struct srcu_struct *ssp)
396 /* The smp_load_acquire() pairs with the smp_store_release(). */
397 if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq_needed))) /*^^^*/
398 return; /* Already initialized. */
399 spin_lock_irqsave_rcu_node(ssp, flags);
400 if (!rcu_seq_state(ssp->srcu_gp_seq_needed)) {
401 spin_unlock_irqrestore_rcu_node(ssp, flags);
404 init_srcu_struct_fields(ssp, true);
405 spin_unlock_irqrestore_rcu_node(ssp, flags);
409 * Returns approximate total of the readers' ->srcu_lock_count[] values
410 * for the rank of per-CPU counters specified by idx.
412 static unsigned long srcu_readers_lock_idx(struct srcu_struct *ssp, int idx)
415 unsigned long sum = 0;
417 for_each_possible_cpu(cpu) {
418 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
420 sum += READ_ONCE(cpuc->srcu_lock_count[idx]);
426 * Returns approximate total of the readers' ->srcu_unlock_count[] values
427 * for the rank of per-CPU counters specified by idx.
429 static unsigned long srcu_readers_unlock_idx(struct srcu_struct *ssp, int idx)
432 unsigned long sum = 0;
434 for_each_possible_cpu(cpu) {
435 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
437 sum += READ_ONCE(cpuc->srcu_unlock_count[idx]);
443 * Return true if the number of pre-existing readers is determined to
446 static bool srcu_readers_active_idx_check(struct srcu_struct *ssp, int idx)
448 unsigned long unlocks;
450 unlocks = srcu_readers_unlock_idx(ssp, idx);
453 * Make sure that a lock is always counted if the corresponding
454 * unlock is counted. Needs to be a smp_mb() as the read side may
455 * contain a read from a variable that is written to before the
456 * synchronize_srcu() in the write side. In this case smp_mb()s
457 * A and B act like the store buffering pattern.
459 * This smp_mb() also pairs with smp_mb() C to prevent accesses
460 * after the synchronize_srcu() from being executed before the
466 * If the locks are the same as the unlocks, then there must have
467 * been no readers on this index at some time in between. This does
468 * not mean that there are no more readers, as one could have read
469 * the current index but not have incremented the lock counter yet.
471 * So suppose that the updater is preempted here for so long
472 * that more than ULONG_MAX non-nested readers come and go in
473 * the meantime. It turns out that this cannot result in overflow
474 * because if a reader modifies its unlock count after we read it
475 * above, then that reader's next load of ->srcu_idx is guaranteed
476 * to get the new value, which will cause it to operate on the
477 * other bank of counters, where it cannot contribute to the
478 * overflow of these counters. This means that there is a maximum
479 * of 2*NR_CPUS increments, which cannot overflow given current
480 * systems, especially not on 64-bit systems.
482 * OK, how about nesting? This does impose a limit on nesting
483 * of floor(ULONG_MAX/NR_CPUS/2), which should be sufficient,
484 * especially on 64-bit systems.
486 return srcu_readers_lock_idx(ssp, idx) == unlocks;
490 * srcu_readers_active - returns true if there are readers. and false
492 * @ssp: which srcu_struct to count active readers (holding srcu_read_lock).
494 * Note that this is not an atomic primitive, and can therefore suffer
495 * severe errors when invoked on an active srcu_struct. That said, it
496 * can be useful as an error check at cleanup time.
498 static bool srcu_readers_active(struct srcu_struct *ssp)
501 unsigned long sum = 0;
503 for_each_possible_cpu(cpu) {
504 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
506 sum += READ_ONCE(cpuc->srcu_lock_count[0]);
507 sum += READ_ONCE(cpuc->srcu_lock_count[1]);
508 sum -= READ_ONCE(cpuc->srcu_unlock_count[0]);
509 sum -= READ_ONCE(cpuc->srcu_unlock_count[1]);
514 #define SRCU_INTERVAL 1 // Base delay if no expedited GPs pending.
515 #define SRCU_MAX_INTERVAL 10 // Maximum incremental delay from slow readers.
516 #define SRCU_MAX_NODELAY_PHASE 1 // Maximum per-GP-phase consecutive no-delay instances.
517 #define SRCU_MAX_NODELAY 100 // Maximum consecutive no-delay instances.
520 * Return grace-period delay, zero if there are expedited grace
521 * periods pending, SRCU_INTERVAL otherwise.
523 static unsigned long srcu_get_delay(struct srcu_struct *ssp)
525 unsigned long jbase = SRCU_INTERVAL;
527 if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq), READ_ONCE(ssp->srcu_gp_seq_needed_exp)))
529 if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)))
530 jbase += jiffies - READ_ONCE(ssp->srcu_gp_start);
532 WRITE_ONCE(ssp->srcu_n_exp_nodelay, READ_ONCE(ssp->srcu_n_exp_nodelay) + 1);
533 if (READ_ONCE(ssp->srcu_n_exp_nodelay) > SRCU_MAX_NODELAY_PHASE)
536 return jbase > SRCU_MAX_INTERVAL ? SRCU_MAX_INTERVAL : jbase;
540 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
541 * @ssp: structure to clean up.
543 * Must invoke this after you are finished using a given srcu_struct that
544 * was initialized via init_srcu_struct(), else you leak memory.
546 void cleanup_srcu_struct(struct srcu_struct *ssp)
550 if (WARN_ON(!srcu_get_delay(ssp)))
551 return; /* Just leak it! */
552 if (WARN_ON(srcu_readers_active(ssp)))
553 return; /* Just leak it! */
554 flush_delayed_work(&ssp->work);
555 for_each_possible_cpu(cpu) {
556 struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
558 del_timer_sync(&sdp->delay_work);
559 flush_work(&sdp->work);
560 if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist)))
561 return; /* Forgot srcu_barrier(), so just leak it! */
563 if (WARN_ON(rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) != SRCU_STATE_IDLE) ||
564 WARN_ON(rcu_seq_current(&ssp->srcu_gp_seq) != ssp->srcu_gp_seq_needed) ||
565 WARN_ON(srcu_readers_active(ssp))) {
566 pr_info("%s: Active srcu_struct %p read state: %d gp state: %lu/%lu\n",
567 __func__, ssp, rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)),
568 rcu_seq_current(&ssp->srcu_gp_seq), ssp->srcu_gp_seq_needed);
569 return; /* Caller forgot to stop doing call_srcu()? */
571 if (!ssp->sda_is_static) {
572 free_percpu(ssp->sda);
577 ssp->srcu_size_state = SRCU_SIZE_SMALL;
579 EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
582 * Counts the new reader in the appropriate per-CPU element of the
584 * Returns an index that must be passed to the matching srcu_read_unlock().
586 int __srcu_read_lock(struct srcu_struct *ssp)
590 idx = READ_ONCE(ssp->srcu_idx) & 0x1;
591 this_cpu_inc(ssp->sda->srcu_lock_count[idx]);
592 smp_mb(); /* B */ /* Avoid leaking the critical section. */
595 EXPORT_SYMBOL_GPL(__srcu_read_lock);
598 * Removes the count for the old reader from the appropriate per-CPU
599 * element of the srcu_struct. Note that this may well be a different
600 * CPU than that which was incremented by the corresponding srcu_read_lock().
602 void __srcu_read_unlock(struct srcu_struct *ssp, int idx)
604 smp_mb(); /* C */ /* Avoid leaking the critical section. */
605 this_cpu_inc(ssp->sda->srcu_unlock_count[idx]);
607 EXPORT_SYMBOL_GPL(__srcu_read_unlock);
610 * We use an adaptive strategy for synchronize_srcu() and especially for
611 * synchronize_srcu_expedited(). We spin for a fixed time period
612 * (defined below) to allow SRCU readers to exit their read-side critical
613 * sections. If there are still some readers after a few microseconds,
614 * we repeatedly block for 1-millisecond time periods.
616 #define SRCU_RETRY_CHECK_DELAY 5
619 * Start an SRCU grace period.
621 static void srcu_gp_start(struct srcu_struct *ssp)
623 struct srcu_data *sdp;
626 if (smp_load_acquire(&ssp->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
627 sdp = per_cpu_ptr(ssp->sda, 0);
629 sdp = this_cpu_ptr(ssp->sda);
630 lockdep_assert_held(&ACCESS_PRIVATE(ssp, lock));
631 WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
632 spin_lock_rcu_node(sdp); /* Interrupts already disabled. */
633 rcu_segcblist_advance(&sdp->srcu_cblist,
634 rcu_seq_current(&ssp->srcu_gp_seq));
635 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
636 rcu_seq_snap(&ssp->srcu_gp_seq));
637 spin_unlock_rcu_node(sdp); /* Interrupts remain disabled. */
638 WRITE_ONCE(ssp->srcu_gp_start, jiffies);
639 WRITE_ONCE(ssp->srcu_n_exp_nodelay, 0);
640 smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
641 rcu_seq_start(&ssp->srcu_gp_seq);
642 state = rcu_seq_state(ssp->srcu_gp_seq);
643 WARN_ON_ONCE(state != SRCU_STATE_SCAN1);
647 static void srcu_delay_timer(struct timer_list *t)
649 struct srcu_data *sdp = container_of(t, struct srcu_data, delay_work);
651 queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
654 static void srcu_queue_delayed_work_on(struct srcu_data *sdp,
658 queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
662 timer_reduce(&sdp->delay_work, jiffies + delay);
666 * Schedule callback invocation for the specified srcu_data structure,
667 * if possible, on the corresponding CPU.
669 static void srcu_schedule_cbs_sdp(struct srcu_data *sdp, unsigned long delay)
671 srcu_queue_delayed_work_on(sdp, delay);
675 * Schedule callback invocation for all srcu_data structures associated
676 * with the specified srcu_node structure that have callbacks for the
677 * just-completed grace period, the one corresponding to idx. If possible,
678 * schedule this invocation on the corresponding CPUs.
680 static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp,
681 unsigned long mask, unsigned long delay)
685 for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
686 if (!(mask & (1 << (cpu - snp->grplo))))
688 srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
693 * Note the end of an SRCU grace period. Initiates callback invocation
694 * and starts a new grace period if needed.
696 * The ->srcu_cb_mutex acquisition does not protect any data, but
697 * instead prevents more than one grace period from starting while we
698 * are initiating callback invocation. This allows the ->srcu_have_cbs[]
699 * array to have a finite number of elements.
701 static void srcu_gp_end(struct srcu_struct *ssp)
703 unsigned long cbdelay;
711 struct srcu_data *sdp;
713 struct srcu_node *snp;
716 /* Prevent more than one additional grace period. */
717 mutex_lock(&ssp->srcu_cb_mutex);
719 /* End the current grace period. */
720 spin_lock_irq_rcu_node(ssp);
721 idx = rcu_seq_state(ssp->srcu_gp_seq);
722 WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
723 cbdelay = !!srcu_get_delay(ssp);
724 WRITE_ONCE(ssp->srcu_last_gp_end, ktime_get_mono_fast_ns());
725 rcu_seq_end(&ssp->srcu_gp_seq);
726 gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
727 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, gpseq))
728 WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, gpseq);
729 spin_unlock_irq_rcu_node(ssp);
730 mutex_unlock(&ssp->srcu_gp_mutex);
731 /* A new grace period can start at this point. But only one. */
733 /* Initiate callback invocation as needed. */
734 ss_state = smp_load_acquire(&ssp->srcu_size_state);
735 if (ss_state < SRCU_SIZE_WAIT_BARRIER) {
736 srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, 0), cbdelay);
738 idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs);
739 srcu_for_each_node_breadth_first(ssp, snp) {
740 spin_lock_irq_rcu_node(snp);
742 last_lvl = snp >= ssp->level[rcu_num_lvls - 1];
744 cbs = ss_state < SRCU_SIZE_BIG || snp->srcu_have_cbs[idx] == gpseq;
745 snp->srcu_have_cbs[idx] = gpseq;
746 rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1);
747 sgsne = snp->srcu_gp_seq_needed_exp;
748 if (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, gpseq))
749 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, gpseq);
750 if (ss_state < SRCU_SIZE_BIG)
753 mask = snp->srcu_data_have_cbs[idx];
754 snp->srcu_data_have_cbs[idx] = 0;
755 spin_unlock_irq_rcu_node(snp);
757 srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay);
761 /* Occasionally prevent srcu_data counter wrap. */
762 if (!(gpseq & counter_wrap_check))
763 for_each_possible_cpu(cpu) {
764 sdp = per_cpu_ptr(ssp->sda, cpu);
765 spin_lock_irqsave_rcu_node(sdp, flags);
766 if (ULONG_CMP_GE(gpseq, sdp->srcu_gp_seq_needed + 100))
767 sdp->srcu_gp_seq_needed = gpseq;
768 if (ULONG_CMP_GE(gpseq, sdp->srcu_gp_seq_needed_exp + 100))
769 sdp->srcu_gp_seq_needed_exp = gpseq;
770 spin_unlock_irqrestore_rcu_node(sdp, flags);
773 /* Callback initiation done, allow grace periods after next. */
774 mutex_unlock(&ssp->srcu_cb_mutex);
776 /* Start a new grace period if needed. */
777 spin_lock_irq_rcu_node(ssp);
778 gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
779 if (!rcu_seq_state(gpseq) &&
780 ULONG_CMP_LT(gpseq, ssp->srcu_gp_seq_needed)) {
782 spin_unlock_irq_rcu_node(ssp);
783 srcu_reschedule(ssp, 0);
785 spin_unlock_irq_rcu_node(ssp);
788 /* Transition to big if needed. */
789 if (ss_state != SRCU_SIZE_SMALL && ss_state != SRCU_SIZE_BIG) {
790 if (ss_state == SRCU_SIZE_ALLOC)
791 init_srcu_struct_nodes(ssp, GFP_KERNEL);
793 smp_store_release(&ssp->srcu_size_state, ss_state + 1);
798 * Funnel-locking scheme to scalably mediate many concurrent expedited
799 * grace-period requests. This function is invoked for the first known
800 * expedited request for a grace period that has already been requested,
801 * but without expediting. To start a completely new grace period,
802 * whether expedited or not, use srcu_funnel_gp_start() instead.
804 static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp,
811 for (; snp != NULL; snp = snp->srcu_parent) {
812 sgsne = READ_ONCE(snp->srcu_gp_seq_needed_exp);
813 if (rcu_seq_done(&ssp->srcu_gp_seq, s) ||
814 (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)))
816 spin_lock_irqsave_rcu_node(snp, flags);
817 sgsne = snp->srcu_gp_seq_needed_exp;
818 if (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)) {
819 spin_unlock_irqrestore_rcu_node(snp, flags);
822 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
823 spin_unlock_irqrestore_rcu_node(snp, flags);
825 spin_lock_irqsave_ssp_contention(ssp, &flags);
826 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
827 WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, s);
828 spin_unlock_irqrestore_rcu_node(ssp, flags);
832 * Funnel-locking scheme to scalably mediate many concurrent grace-period
833 * requests. The winner has to do the work of actually starting grace
834 * period s. Losers must either ensure that their desired grace-period
835 * number is recorded on at least their leaf srcu_node structure, or they
836 * must take steps to invoke their own callbacks.
838 * Note that this function also does the work of srcu_funnel_exp_start(),
839 * in some cases by directly invoking it.
841 static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
842 unsigned long s, bool do_norm)
845 int idx = rcu_seq_ctr(s) % ARRAY_SIZE(sdp->mynode->srcu_have_cbs);
847 struct srcu_node *snp;
848 struct srcu_node *snp_leaf;
849 unsigned long snp_seq;
851 /* Ensure that snp node tree is fully initialized before traversing it */
852 if (smp_load_acquire(&ssp->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
855 snp_leaf = sdp->mynode;
858 /* Each pass through the loop does one level of the srcu_node tree. */
859 for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
860 if (rcu_seq_done(&ssp->srcu_gp_seq, s) && snp != snp_leaf)
861 return; /* GP already done and CBs recorded. */
862 spin_lock_irqsave_rcu_node(snp, flags);
863 snp_seq = snp->srcu_have_cbs[idx];
864 if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
865 if (snp == snp_leaf && snp_seq == s)
866 snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
867 spin_unlock_irqrestore_rcu_node(snp, flags);
868 if (snp == snp_leaf && snp_seq != s) {
869 srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
873 srcu_funnel_exp_start(ssp, snp, s);
876 snp->srcu_have_cbs[idx] = s;
878 snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
879 sgsne = snp->srcu_gp_seq_needed_exp;
880 if (!do_norm && (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, s)))
881 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
882 spin_unlock_irqrestore_rcu_node(snp, flags);
885 /* Top of tree, must ensure the grace period will be started. */
886 spin_lock_irqsave_ssp_contention(ssp, &flags);
887 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed, s)) {
889 * Record need for grace period s. Pair with load
890 * acquire setting up for initialization.
892 smp_store_release(&ssp->srcu_gp_seq_needed, s); /*^^^*/
894 if (!do_norm && ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
895 WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, s);
897 /* If grace period not already done and none in progress, start it. */
898 if (!rcu_seq_done(&ssp->srcu_gp_seq, s) &&
899 rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
900 WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
903 // And how can that list_add() in the "else" clause
904 // possibly be safe for concurrent execution? Well,
905 // it isn't. And it does not have to be. After all, it
906 // can only be executed during early boot when there is only
907 // the one boot CPU running with interrupts still disabled.
908 if (likely(srcu_init_done))
909 queue_delayed_work(rcu_gp_wq, &ssp->work,
910 !!srcu_get_delay(ssp));
911 else if (list_empty(&ssp->work.work.entry))
912 list_add(&ssp->work.work.entry, &srcu_boot_list);
914 spin_unlock_irqrestore_rcu_node(ssp, flags);
918 * Wait until all readers counted by array index idx complete, but
919 * loop an additional time if there is an expedited grace period pending.
920 * The caller must ensure that ->srcu_idx is not changed while checking.
922 static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount)
925 if (srcu_readers_active_idx_check(ssp, idx))
927 if (--trycount + !srcu_get_delay(ssp) <= 0)
929 udelay(SRCU_RETRY_CHECK_DELAY);
934 * Increment the ->srcu_idx counter so that future SRCU readers will
935 * use the other rank of the ->srcu_(un)lock_count[] arrays. This allows
936 * us to wait for pre-existing readers in a starvation-free manner.
938 static void srcu_flip(struct srcu_struct *ssp)
941 * Ensure that if this updater saw a given reader's increment
942 * from __srcu_read_lock(), that reader was using an old value
943 * of ->srcu_idx. Also ensure that if a given reader sees the
944 * new value of ->srcu_idx, this updater's earlier scans cannot
945 * have seen that reader's increments (which is OK, because this
946 * grace period need not wait on that reader).
948 smp_mb(); /* E */ /* Pairs with B and C. */
950 WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1);
953 * Ensure that if the updater misses an __srcu_read_unlock()
954 * increment, that task's next __srcu_read_lock() will see the
955 * above counter update. Note that both this memory barrier
956 * and the one in srcu_readers_active_idx_check() provide the
957 * guarantee for __srcu_read_lock().
959 smp_mb(); /* D */ /* Pairs with C. */
963 * If SRCU is likely idle, return true, otherwise return false.
965 * Note that it is OK for several current from-idle requests for a new
966 * grace period from idle to specify expediting because they will all end
967 * up requesting the same grace period anyhow. So no loss.
969 * Note also that if any CPU (including the current one) is still invoking
970 * callbacks, this function will nevertheless say "idle". This is not
971 * ideal, but the overhead of checking all CPUs' callback lists is even
972 * less ideal, especially on large systems. Furthermore, the wakeup
973 * can happen before the callback is fully removed, so we have no choice
974 * but to accept this type of error.
976 * This function is also subject to counter-wrap errors, but let's face
977 * it, if this function was preempted for enough time for the counters
978 * to wrap, it really doesn't matter whether or not we expedite the grace
979 * period. The extra overhead of a needlessly expedited grace period is
980 * negligible when amortized over that time period, and the extra latency
981 * of a needlessly non-expedited grace period is similarly negligible.
983 static bool srcu_might_be_idle(struct srcu_struct *ssp)
985 unsigned long curseq;
987 struct srcu_data *sdp;
991 check_init_srcu_struct(ssp);
992 /* If the local srcu_data structure has callbacks, not idle. */
993 sdp = raw_cpu_ptr(ssp->sda);
994 spin_lock_irqsave_rcu_node(sdp, flags);
995 if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) {
996 spin_unlock_irqrestore_rcu_node(sdp, flags);
997 return false; /* Callbacks already present, so not idle. */
999 spin_unlock_irqrestore_rcu_node(sdp, flags);
1002 * No local callbacks, so probabilistically probe global state.
1003 * Exact information would require acquiring locks, which would
1004 * kill scalability, hence the probabilistic nature of the probe.
1007 /* First, see if enough time has passed since the last GP. */
1008 t = ktime_get_mono_fast_ns();
1009 tlast = READ_ONCE(ssp->srcu_last_gp_end);
1010 if (exp_holdoff == 0 ||
1011 time_in_range_open(t, tlast, tlast + exp_holdoff))
1012 return false; /* Too soon after last GP. */
1014 /* Next, check for probable idleness. */
1015 curseq = rcu_seq_current(&ssp->srcu_gp_seq);
1016 smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
1017 if (ULONG_CMP_LT(curseq, READ_ONCE(ssp->srcu_gp_seq_needed)))
1018 return false; /* Grace period in progress, so not idle. */
1019 smp_mb(); /* Order ->srcu_gp_seq with prior access. */
1020 if (curseq != rcu_seq_current(&ssp->srcu_gp_seq))
1021 return false; /* GP # changed, so not idle. */
1022 return true; /* With reasonable probability, idle! */
1026 * SRCU callback function to leak a callback.
1028 static void srcu_leak_callback(struct rcu_head *rhp)
1033 * Start an SRCU grace period, and also queue the callback if non-NULL.
1035 static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
1036 struct rcu_head *rhp, bool do_norm)
1038 unsigned long flags;
1040 bool needexp = false;
1041 bool needgp = false;
1043 struct srcu_data *sdp;
1044 struct srcu_node *sdp_mynode;
1047 check_init_srcu_struct(ssp);
1048 idx = srcu_read_lock(ssp);
1049 ss_state = smp_load_acquire(&ssp->srcu_size_state);
1050 if (ss_state < SRCU_SIZE_WAIT_CALL)
1051 sdp = per_cpu_ptr(ssp->sda, 0);
1053 sdp = raw_cpu_ptr(ssp->sda);
1054 spin_lock_irqsave_sdp_contention(sdp, &flags);
1056 rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp);
1057 rcu_segcblist_advance(&sdp->srcu_cblist,
1058 rcu_seq_current(&ssp->srcu_gp_seq));
1059 s = rcu_seq_snap(&ssp->srcu_gp_seq);
1060 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, s);
1061 if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) {
1062 sdp->srcu_gp_seq_needed = s;
1065 if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) {
1066 sdp->srcu_gp_seq_needed_exp = s;
1069 spin_unlock_irqrestore_rcu_node(sdp, flags);
1071 /* Ensure that snp node tree is fully initialized before traversing it */
1072 if (ss_state < SRCU_SIZE_WAIT_BARRIER)
1075 sdp_mynode = sdp->mynode;
1078 srcu_funnel_gp_start(ssp, sdp, s, do_norm);
1080 srcu_funnel_exp_start(ssp, sdp_mynode, s);
1081 srcu_read_unlock(ssp, idx);
1086 * Enqueue an SRCU callback on the srcu_data structure associated with
1087 * the current CPU and the specified srcu_struct structure, initiating
1088 * grace-period processing if it is not already running.
1090 * Note that all CPUs must agree that the grace period extended beyond
1091 * all pre-existing SRCU read-side critical section. On systems with
1092 * more than one CPU, this means that when "func()" is invoked, each CPU
1093 * is guaranteed to have executed a full memory barrier since the end of
1094 * its last corresponding SRCU read-side critical section whose beginning
1095 * preceded the call to call_srcu(). It also means that each CPU executing
1096 * an SRCU read-side critical section that continues beyond the start of
1097 * "func()" must have executed a memory barrier after the call_srcu()
1098 * but before the beginning of that SRCU read-side critical section.
1099 * Note that these guarantees include CPUs that are offline, idle, or
1100 * executing in user mode, as well as CPUs that are executing in the kernel.
1102 * Furthermore, if CPU A invoked call_srcu() and CPU B invoked the
1103 * resulting SRCU callback function "func()", then both CPU A and CPU
1104 * B are guaranteed to execute a full memory barrier during the time
1105 * interval between the call to call_srcu() and the invocation of "func()".
1106 * This guarantee applies even if CPU A and CPU B are the same CPU (but
1107 * again only if the system has more than one CPU).
1109 * Of course, these guarantees apply only for invocations of call_srcu(),
1110 * srcu_read_lock(), and srcu_read_unlock() that are all passed the same
1111 * srcu_struct structure.
1113 static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
1114 rcu_callback_t func, bool do_norm)
1116 if (debug_rcu_head_queue(rhp)) {
1117 /* Probable double call_srcu(), so leak the callback. */
1118 WRITE_ONCE(rhp->func, srcu_leak_callback);
1119 WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n");
1123 (void)srcu_gp_start_if_needed(ssp, rhp, do_norm);
1127 * call_srcu() - Queue a callback for invocation after an SRCU grace period
1128 * @ssp: srcu_struct in queue the callback
1129 * @rhp: structure to be used for queueing the SRCU callback.
1130 * @func: function to be invoked after the SRCU grace period
1132 * The callback function will be invoked some time after a full SRCU
1133 * grace period elapses, in other words after all pre-existing SRCU
1134 * read-side critical sections have completed. However, the callback
1135 * function might well execute concurrently with other SRCU read-side
1136 * critical sections that started after call_srcu() was invoked. SRCU
1137 * read-side critical sections are delimited by srcu_read_lock() and
1138 * srcu_read_unlock(), and may be nested.
1140 * The callback will be invoked from process context, but must nevertheless
1141 * be fast and must not block.
1143 void call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
1144 rcu_callback_t func)
1146 __call_srcu(ssp, rhp, func, true);
1148 EXPORT_SYMBOL_GPL(call_srcu);
1151 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
1153 static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
1155 struct rcu_synchronize rcu;
1157 RCU_LOCKDEP_WARN(lockdep_is_held(ssp) ||
1158 lock_is_held(&rcu_bh_lock_map) ||
1159 lock_is_held(&rcu_lock_map) ||
1160 lock_is_held(&rcu_sched_lock_map),
1161 "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
1163 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
1166 check_init_srcu_struct(ssp);
1167 init_completion(&rcu.completion);
1168 init_rcu_head_on_stack(&rcu.head);
1169 __call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm);
1170 wait_for_completion(&rcu.completion);
1171 destroy_rcu_head_on_stack(&rcu.head);
1174 * Make sure that later code is ordered after the SRCU grace
1175 * period. This pairs with the spin_lock_irq_rcu_node()
1176 * in srcu_invoke_callbacks(). Unlike Tree RCU, this is needed
1177 * because the current CPU might have been totally uninvolved with
1178 * (and thus unordered against) that grace period.
1184 * synchronize_srcu_expedited - Brute-force SRCU grace period
1185 * @ssp: srcu_struct with which to synchronize.
1187 * Wait for an SRCU grace period to elapse, but be more aggressive about
1188 * spinning rather than blocking when waiting.
1190 * Note that synchronize_srcu_expedited() has the same deadlock and
1191 * memory-ordering properties as does synchronize_srcu().
1193 void synchronize_srcu_expedited(struct srcu_struct *ssp)
1195 __synchronize_srcu(ssp, rcu_gp_is_normal());
1197 EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
1200 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
1201 * @ssp: srcu_struct with which to synchronize.
1203 * Wait for the count to drain to zero of both indexes. To avoid the
1204 * possible starvation of synchronize_srcu(), it waits for the count of
1205 * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
1206 * and then flip the srcu_idx and wait for the count of the other index.
1208 * Can block; must be called from process context.
1210 * Note that it is illegal to call synchronize_srcu() from the corresponding
1211 * SRCU read-side critical section; doing so will result in deadlock.
1212 * However, it is perfectly legal to call synchronize_srcu() on one
1213 * srcu_struct from some other srcu_struct's read-side critical section,
1214 * as long as the resulting graph of srcu_structs is acyclic.
1216 * There are memory-ordering constraints implied by synchronize_srcu().
1217 * On systems with more than one CPU, when synchronize_srcu() returns,
1218 * each CPU is guaranteed to have executed a full memory barrier since
1219 * the end of its last corresponding SRCU read-side critical section
1220 * whose beginning preceded the call to synchronize_srcu(). In addition,
1221 * each CPU having an SRCU read-side critical section that extends beyond
1222 * the return from synchronize_srcu() is guaranteed to have executed a
1223 * full memory barrier after the beginning of synchronize_srcu() and before
1224 * the beginning of that SRCU read-side critical section. Note that these
1225 * guarantees include CPUs that are offline, idle, or executing in user mode,
1226 * as well as CPUs that are executing in the kernel.
1228 * Furthermore, if CPU A invoked synchronize_srcu(), which returned
1229 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
1230 * to have executed a full memory barrier during the execution of
1231 * synchronize_srcu(). This guarantee applies even if CPU A and CPU B
1232 * are the same CPU, but again only if the system has more than one CPU.
1234 * Of course, these memory-ordering guarantees apply only when
1235 * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
1236 * passed the same srcu_struct structure.
1238 * Implementation of these memory-ordering guarantees is similar to
1239 * that of synchronize_rcu().
1241 * If SRCU is likely idle, expedite the first request. This semantic
1242 * was provided by Classic SRCU, and is relied upon by its users, so TREE
1243 * SRCU must also provide it. Note that detecting idleness is heuristic
1244 * and subject to both false positives and negatives.
1246 void synchronize_srcu(struct srcu_struct *ssp)
1248 if (srcu_might_be_idle(ssp) || rcu_gp_is_expedited())
1249 synchronize_srcu_expedited(ssp);
1251 __synchronize_srcu(ssp, true);
1253 EXPORT_SYMBOL_GPL(synchronize_srcu);
1256 * get_state_synchronize_srcu - Provide an end-of-grace-period cookie
1257 * @ssp: srcu_struct to provide cookie for.
1259 * This function returns a cookie that can be passed to
1260 * poll_state_synchronize_srcu(), which will return true if a full grace
1261 * period has elapsed in the meantime. It is the caller's responsibility
1262 * to make sure that grace period happens, for example, by invoking
1263 * call_srcu() after return from get_state_synchronize_srcu().
1265 unsigned long get_state_synchronize_srcu(struct srcu_struct *ssp)
1267 // Any prior manipulation of SRCU-protected data must happen
1268 // before the load from ->srcu_gp_seq.
1270 return rcu_seq_snap(&ssp->srcu_gp_seq);
1272 EXPORT_SYMBOL_GPL(get_state_synchronize_srcu);
1275 * start_poll_synchronize_srcu - Provide cookie and start grace period
1276 * @ssp: srcu_struct to provide cookie for.
1278 * This function returns a cookie that can be passed to
1279 * poll_state_synchronize_srcu(), which will return true if a full grace
1280 * period has elapsed in the meantime. Unlike get_state_synchronize_srcu(),
1281 * this function also ensures that any needed SRCU grace period will be
1282 * started. This convenience does come at a cost in terms of CPU overhead.
1284 unsigned long start_poll_synchronize_srcu(struct srcu_struct *ssp)
1286 return srcu_gp_start_if_needed(ssp, NULL, true);
1288 EXPORT_SYMBOL_GPL(start_poll_synchronize_srcu);
1291 * poll_state_synchronize_srcu - Has cookie's grace period ended?
1292 * @ssp: srcu_struct to provide cookie for.
1293 * @cookie: Return value from get_state_synchronize_srcu() or start_poll_synchronize_srcu().
1295 * This function takes the cookie that was returned from either
1296 * get_state_synchronize_srcu() or start_poll_synchronize_srcu(), and
1297 * returns @true if an SRCU grace period elapsed since the time that the
1298 * cookie was created.
1300 * Because cookies are finite in size, wrapping/overflow is possible.
1301 * This is more pronounced on 32-bit systems where cookies are 32 bits,
1302 * where in theory wrapping could happen in about 14 hours assuming
1303 * 25-microsecond expedited SRCU grace periods. However, a more likely
1304 * overflow lower bound is on the order of 24 days in the case of
1305 * one-millisecond SRCU grace periods. Of course, wrapping in a 64-bit
1306 * system requires geologic timespans, as in more than seven million years
1307 * even for expedited SRCU grace periods.
1309 * Wrapping/overflow is much more of an issue for CONFIG_SMP=n systems
1310 * that also have CONFIG_PREEMPTION=n, which selects Tiny SRCU. This uses
1311 * a 16-bit cookie, which rcutorture routinely wraps in a matter of a
1312 * few minutes. If this proves to be a problem, this counter will be
1313 * expanded to the same size as for Tree SRCU.
1315 bool poll_state_synchronize_srcu(struct srcu_struct *ssp, unsigned long cookie)
1317 if (!rcu_seq_done(&ssp->srcu_gp_seq, cookie))
1319 // Ensure that the end of the SRCU grace period happens before
1320 // any subsequent code that the caller might execute.
1324 EXPORT_SYMBOL_GPL(poll_state_synchronize_srcu);
1327 * Callback function for srcu_barrier() use.
1329 static void srcu_barrier_cb(struct rcu_head *rhp)
1331 struct srcu_data *sdp;
1332 struct srcu_struct *ssp;
1334 sdp = container_of(rhp, struct srcu_data, srcu_barrier_head);
1336 if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
1337 complete(&ssp->srcu_barrier_completion);
1341 * Enqueue an srcu_barrier() callback on the specified srcu_data
1342 * structure's ->cblist. but only if that ->cblist already has at least one
1343 * callback enqueued. Note that if a CPU already has callbacks enqueue,
1344 * it must have already registered the need for a future grace period,
1345 * so all we need do is enqueue a callback that will use the same grace
1346 * period as the last callback already in the queue.
1348 static void srcu_barrier_one_cpu(struct srcu_struct *ssp, struct srcu_data *sdp)
1350 spin_lock_irq_rcu_node(sdp);
1351 atomic_inc(&ssp->srcu_barrier_cpu_cnt);
1352 sdp->srcu_barrier_head.func = srcu_barrier_cb;
1353 debug_rcu_head_queue(&sdp->srcu_barrier_head);
1354 if (!rcu_segcblist_entrain(&sdp->srcu_cblist,
1355 &sdp->srcu_barrier_head)) {
1356 debug_rcu_head_unqueue(&sdp->srcu_barrier_head);
1357 atomic_dec(&ssp->srcu_barrier_cpu_cnt);
1359 spin_unlock_irq_rcu_node(sdp);
1363 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
1364 * @ssp: srcu_struct on which to wait for in-flight callbacks.
1366 void srcu_barrier(struct srcu_struct *ssp)
1370 unsigned long s = rcu_seq_snap(&ssp->srcu_barrier_seq);
1372 check_init_srcu_struct(ssp);
1373 mutex_lock(&ssp->srcu_barrier_mutex);
1374 if (rcu_seq_done(&ssp->srcu_barrier_seq, s)) {
1375 smp_mb(); /* Force ordering following return. */
1376 mutex_unlock(&ssp->srcu_barrier_mutex);
1377 return; /* Someone else did our work for us. */
1379 rcu_seq_start(&ssp->srcu_barrier_seq);
1380 init_completion(&ssp->srcu_barrier_completion);
1382 /* Initial count prevents reaching zero until all CBs are posted. */
1383 atomic_set(&ssp->srcu_barrier_cpu_cnt, 1);
1385 idx = srcu_read_lock(ssp);
1386 if (smp_load_acquire(&ssp->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
1387 srcu_barrier_one_cpu(ssp, per_cpu_ptr(ssp->sda, 0));
1389 for_each_possible_cpu(cpu)
1390 srcu_barrier_one_cpu(ssp, per_cpu_ptr(ssp->sda, cpu));
1391 srcu_read_unlock(ssp, idx);
1393 /* Remove the initial count, at which point reaching zero can happen. */
1394 if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
1395 complete(&ssp->srcu_barrier_completion);
1396 wait_for_completion(&ssp->srcu_barrier_completion);
1398 rcu_seq_end(&ssp->srcu_barrier_seq);
1399 mutex_unlock(&ssp->srcu_barrier_mutex);
1401 EXPORT_SYMBOL_GPL(srcu_barrier);
1404 * srcu_batches_completed - return batches completed.
1405 * @ssp: srcu_struct on which to report batch completion.
1407 * Report the number of batches, correlated with, but not necessarily
1408 * precisely the same as, the number of grace periods that have elapsed.
1410 unsigned long srcu_batches_completed(struct srcu_struct *ssp)
1412 return READ_ONCE(ssp->srcu_idx);
1414 EXPORT_SYMBOL_GPL(srcu_batches_completed);
1417 * Core SRCU state machine. Push state bits of ->srcu_gp_seq
1418 * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
1419 * completed in that state.
1421 static void srcu_advance_state(struct srcu_struct *ssp)
1425 mutex_lock(&ssp->srcu_gp_mutex);
1428 * Because readers might be delayed for an extended period after
1429 * fetching ->srcu_idx for their index, at any point in time there
1430 * might well be readers using both idx=0 and idx=1. We therefore
1431 * need to wait for readers to clear from both index values before
1432 * invoking a callback.
1434 * The load-acquire ensures that we see the accesses performed
1435 * by the prior grace period.
1437 idx = rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq)); /* ^^^ */
1438 if (idx == SRCU_STATE_IDLE) {
1439 spin_lock_irq_rcu_node(ssp);
1440 if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
1441 WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq));
1442 spin_unlock_irq_rcu_node(ssp);
1443 mutex_unlock(&ssp->srcu_gp_mutex);
1446 idx = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq));
1447 if (idx == SRCU_STATE_IDLE)
1449 spin_unlock_irq_rcu_node(ssp);
1450 if (idx != SRCU_STATE_IDLE) {
1451 mutex_unlock(&ssp->srcu_gp_mutex);
1452 return; /* Someone else started the grace period. */
1456 if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN1) {
1457 idx = 1 ^ (ssp->srcu_idx & 1);
1458 if (!try_check_zero(ssp, idx, 1)) {
1459 mutex_unlock(&ssp->srcu_gp_mutex);
1460 return; /* readers present, retry later. */
1463 spin_lock_irq_rcu_node(ssp);
1464 rcu_seq_set_state(&ssp->srcu_gp_seq, SRCU_STATE_SCAN2);
1465 ssp->srcu_n_exp_nodelay = 0;
1466 spin_unlock_irq_rcu_node(ssp);
1469 if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN2) {
1472 * SRCU read-side critical sections are normally short,
1473 * so check at least twice in quick succession after a flip.
1475 idx = 1 ^ (ssp->srcu_idx & 1);
1476 if (!try_check_zero(ssp, idx, 2)) {
1477 mutex_unlock(&ssp->srcu_gp_mutex);
1478 return; /* readers present, retry later. */
1480 ssp->srcu_n_exp_nodelay = 0;
1481 srcu_gp_end(ssp); /* Releases ->srcu_gp_mutex. */
1486 * Invoke a limited number of SRCU callbacks that have passed through
1487 * their grace period. If there are more to do, SRCU will reschedule
1488 * the workqueue. Note that needed memory barriers have been executed
1489 * in this task's context by srcu_readers_active_idx_check().
1491 static void srcu_invoke_callbacks(struct work_struct *work)
1495 struct rcu_cblist ready_cbs;
1496 struct rcu_head *rhp;
1497 struct srcu_data *sdp;
1498 struct srcu_struct *ssp;
1500 sdp = container_of(work, struct srcu_data, work);
1503 rcu_cblist_init(&ready_cbs);
1504 spin_lock_irq_rcu_node(sdp);
1505 rcu_segcblist_advance(&sdp->srcu_cblist,
1506 rcu_seq_current(&ssp->srcu_gp_seq));
1507 if (sdp->srcu_cblist_invoking ||
1508 !rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) {
1509 spin_unlock_irq_rcu_node(sdp);
1510 return; /* Someone else on the job or nothing to do. */
1513 /* We are on the job! Extract and invoke ready callbacks. */
1514 sdp->srcu_cblist_invoking = true;
1515 rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs);
1516 len = ready_cbs.len;
1517 spin_unlock_irq_rcu_node(sdp);
1518 rhp = rcu_cblist_dequeue(&ready_cbs);
1519 for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) {
1520 debug_rcu_head_unqueue(rhp);
1525 WARN_ON_ONCE(ready_cbs.len);
1528 * Update counts, accelerate new callbacks, and if needed,
1529 * schedule another round of callback invocation.
1531 spin_lock_irq_rcu_node(sdp);
1532 rcu_segcblist_add_len(&sdp->srcu_cblist, -len);
1533 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
1534 rcu_seq_snap(&ssp->srcu_gp_seq));
1535 sdp->srcu_cblist_invoking = false;
1536 more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist);
1537 spin_unlock_irq_rcu_node(sdp);
1539 srcu_schedule_cbs_sdp(sdp, 0);
1543 * Finished one round of SRCU grace period. Start another if there are
1544 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
1546 static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
1550 spin_lock_irq_rcu_node(ssp);
1551 if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
1552 if (!WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq))) {
1553 /* All requests fulfilled, time to go idle. */
1556 } else if (!rcu_seq_state(ssp->srcu_gp_seq)) {
1557 /* Outstanding request and no GP. Start one. */
1560 spin_unlock_irq_rcu_node(ssp);
1563 queue_delayed_work(rcu_gp_wq, &ssp->work, delay);
1567 * This is the work-queue function that handles SRCU grace periods.
1569 static void process_srcu(struct work_struct *work)
1571 unsigned long curdelay;
1573 struct srcu_struct *ssp;
1575 ssp = container_of(work, struct srcu_struct, work.work);
1577 srcu_advance_state(ssp);
1578 curdelay = srcu_get_delay(ssp);
1580 WRITE_ONCE(ssp->reschedule_count, 0);
1583 if (READ_ONCE(ssp->reschedule_jiffies) == j) {
1584 WRITE_ONCE(ssp->reschedule_count, READ_ONCE(ssp->reschedule_count) + 1);
1585 if (READ_ONCE(ssp->reschedule_count) > SRCU_MAX_NODELAY)
1588 WRITE_ONCE(ssp->reschedule_count, 1);
1589 WRITE_ONCE(ssp->reschedule_jiffies, j);
1592 srcu_reschedule(ssp, curdelay);
1595 void srcutorture_get_gp_data(enum rcutorture_type test_type,
1596 struct srcu_struct *ssp, int *flags,
1597 unsigned long *gp_seq)
1599 if (test_type != SRCU_FLAVOR)
1602 *gp_seq = rcu_seq_current(&ssp->srcu_gp_seq);
1604 EXPORT_SYMBOL_GPL(srcutorture_get_gp_data);
1606 static const char * const srcu_size_state_name[] = {
1609 "SRCU_SIZE_WAIT_BARRIER",
1610 "SRCU_SIZE_WAIT_CALL",
1611 "SRCU_SIZE_WAIT_CBS1",
1612 "SRCU_SIZE_WAIT_CBS2",
1613 "SRCU_SIZE_WAIT_CBS3",
1614 "SRCU_SIZE_WAIT_CBS4",
1619 void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf)
1623 unsigned long s0 = 0, s1 = 0;
1624 int ss_state = READ_ONCE(ssp->srcu_size_state);
1625 int ss_state_idx = ss_state;
1627 idx = ssp->srcu_idx & 0x1;
1628 if (ss_state < 0 || ss_state >= ARRAY_SIZE(srcu_size_state_name))
1629 ss_state_idx = ARRAY_SIZE(srcu_size_state_name) - 1;
1630 pr_alert("%s%s Tree SRCU g%ld state %d (%s)",
1631 tt, tf, rcu_seq_current(&ssp->srcu_gp_seq), ss_state,
1632 srcu_size_state_name[ss_state_idx]);
1634 // Called after cleanup_srcu_struct(), perhaps.
1635 pr_cont(" No per-CPU srcu_data structures (->sda == NULL).\n");
1637 pr_cont(" per-CPU(idx=%d):", idx);
1638 for_each_possible_cpu(cpu) {
1639 unsigned long l0, l1;
1640 unsigned long u0, u1;
1642 struct srcu_data *sdp;
1644 sdp = per_cpu_ptr(ssp->sda, cpu);
1645 u0 = data_race(sdp->srcu_unlock_count[!idx]);
1646 u1 = data_race(sdp->srcu_unlock_count[idx]);
1649 * Make sure that a lock is always counted if the corresponding
1650 * unlock is counted.
1654 l0 = data_race(sdp->srcu_lock_count[!idx]);
1655 l1 = data_race(sdp->srcu_lock_count[idx]);
1659 pr_cont(" %d(%ld,%ld %c)",
1661 "C."[rcu_segcblist_empty(&sdp->srcu_cblist)]);
1665 pr_cont(" T(%ld,%ld)\n", s0, s1);
1667 if (SRCU_SIZING_IS_TORTURE())
1668 srcu_transition_to_big(ssp);
1670 EXPORT_SYMBOL_GPL(srcu_torture_stats_print);
1672 static int __init srcu_bootup_announce(void)
1674 pr_info("Hierarchical SRCU implementation.\n");
1675 if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF)
1676 pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff);
1679 early_initcall(srcu_bootup_announce);
1681 void __init srcu_init(void)
1683 struct srcu_struct *ssp;
1685 /* Decide on srcu_struct-size strategy. */
1686 if (SRCU_SIZING_IS(SRCU_SIZING_AUTO)) {
1687 if (nr_cpu_ids >= big_cpu_lim) {
1688 convert_to_big = SRCU_SIZING_INIT; // Don't bother waiting for contention.
1689 pr_info("%s: Setting srcu_struct sizes to big.\n", __func__);
1691 convert_to_big = SRCU_SIZING_NONE | SRCU_SIZING_CONTEND;
1692 pr_info("%s: Setting srcu_struct sizes based on contention.\n", __func__);
1697 * Once that is set, call_srcu() can follow the normal path and
1698 * queue delayed work. This must follow RCU workqueues creation
1699 * and timers initialization.
1701 srcu_init_done = true;
1702 while (!list_empty(&srcu_boot_list)) {
1703 ssp = list_first_entry(&srcu_boot_list, struct srcu_struct,
1705 list_del_init(&ssp->work.work.entry);
1706 if (SRCU_SIZING_IS(SRCU_SIZING_INIT) && ssp->srcu_size_state == SRCU_SIZE_SMALL)
1707 ssp->srcu_size_state = SRCU_SIZE_ALLOC;
1708 queue_work(rcu_gp_wq, &ssp->work.work);
1712 #ifdef CONFIG_MODULES
1714 /* Initialize any global-scope srcu_struct structures used by this module. */
1715 static int srcu_module_coming(struct module *mod)
1718 struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1721 for (i = 0; i < mod->num_srcu_structs; i++) {
1722 ret = init_srcu_struct(*(sspp++));
1723 if (WARN_ON_ONCE(ret))
1729 /* Clean up any global-scope srcu_struct structures used by this module. */
1730 static void srcu_module_going(struct module *mod)
1733 struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1735 for (i = 0; i < mod->num_srcu_structs; i++)
1736 cleanup_srcu_struct(*(sspp++));
1739 /* Handle one module, either coming or going. */
1740 static int srcu_module_notify(struct notifier_block *self,
1741 unsigned long val, void *data)
1743 struct module *mod = data;
1747 case MODULE_STATE_COMING:
1748 ret = srcu_module_coming(mod);
1750 case MODULE_STATE_GOING:
1751 srcu_module_going(mod);
1759 static struct notifier_block srcu_module_nb = {
1760 .notifier_call = srcu_module_notify,
1764 static __init int init_srcu_module_notifier(void)
1768 ret = register_module_notifier(&srcu_module_nb);
1770 pr_warn("Failed to register srcu module notifier\n");
1773 late_initcall(init_srcu_module_notifier);
1775 #endif /* #ifdef CONFIG_MODULES */