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/srcu.h>
30 #include "rcu_segcblist.h"
32 /* Holdoff in nanoseconds for auto-expediting. */
33 #define DEFAULT_SRCU_EXP_HOLDOFF (25 * 1000)
34 static ulong exp_holdoff = DEFAULT_SRCU_EXP_HOLDOFF;
35 module_param(exp_holdoff, ulong, 0444);
37 /* Overflow-check frequency. N bits roughly says every 2**N grace periods. */
38 static ulong counter_wrap_check = (ULONG_MAX >> 2);
39 module_param(counter_wrap_check, ulong, 0444);
41 /* Early-boot callback-management, so early that no lock is required! */
42 static LIST_HEAD(srcu_boot_list);
43 static bool __read_mostly srcu_init_done;
45 static void srcu_invoke_callbacks(struct work_struct *work);
46 static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay);
47 static void process_srcu(struct work_struct *work);
48 static void srcu_delay_timer(struct timer_list *t);
50 /* Wrappers for lock acquisition and release, see raw_spin_lock_rcu_node(). */
51 #define spin_lock_rcu_node(p) \
53 spin_lock(&ACCESS_PRIVATE(p, lock)); \
54 smp_mb__after_unlock_lock(); \
57 #define spin_unlock_rcu_node(p) spin_unlock(&ACCESS_PRIVATE(p, lock))
59 #define spin_lock_irq_rcu_node(p) \
61 spin_lock_irq(&ACCESS_PRIVATE(p, lock)); \
62 smp_mb__after_unlock_lock(); \
65 #define spin_unlock_irq_rcu_node(p) \
66 spin_unlock_irq(&ACCESS_PRIVATE(p, lock))
68 #define spin_lock_irqsave_rcu_node(p, flags) \
70 spin_lock_irqsave(&ACCESS_PRIVATE(p, lock), flags); \
71 smp_mb__after_unlock_lock(); \
74 #define spin_unlock_irqrestore_rcu_node(p, flags) \
75 spin_unlock_irqrestore(&ACCESS_PRIVATE(p, lock), flags) \
78 * Initialize SRCU combining tree. Note that statically allocated
79 * srcu_struct structures might already have srcu_read_lock() and
80 * srcu_read_unlock() running against them. So if the is_static parameter
81 * is set, don't initialize ->srcu_lock_count[] and ->srcu_unlock_count[].
83 static void init_srcu_struct_nodes(struct srcu_struct *ssp, bool is_static)
88 int levelspread[RCU_NUM_LVLS];
89 struct srcu_data *sdp;
90 struct srcu_node *snp;
91 struct srcu_node *snp_first;
93 /* Initialize geometry if it has not already been initialized. */
96 /* Work out the overall tree geometry. */
97 ssp->level[0] = &ssp->node[0];
98 for (i = 1; i < rcu_num_lvls; i++)
99 ssp->level[i] = ssp->level[i - 1] + num_rcu_lvl[i - 1];
100 rcu_init_levelspread(levelspread, num_rcu_lvl);
102 /* Each pass through this loop initializes one srcu_node structure. */
103 srcu_for_each_node_breadth_first(ssp, snp) {
104 spin_lock_init(&ACCESS_PRIVATE(snp, lock));
105 WARN_ON_ONCE(ARRAY_SIZE(snp->srcu_have_cbs) !=
106 ARRAY_SIZE(snp->srcu_data_have_cbs));
107 for (i = 0; i < ARRAY_SIZE(snp->srcu_have_cbs); i++) {
108 snp->srcu_have_cbs[i] = 0;
109 snp->srcu_data_have_cbs[i] = 0;
111 snp->srcu_gp_seq_needed_exp = 0;
114 if (snp == &ssp->node[0]) {
115 /* Root node, special case. */
116 snp->srcu_parent = NULL;
121 if (snp == ssp->level[level + 1])
123 snp->srcu_parent = ssp->level[level - 1] +
124 (snp - ssp->level[level]) /
125 levelspread[level - 1];
129 * Initialize the per-CPU srcu_data array, which feeds into the
130 * leaves of the srcu_node tree.
132 WARN_ON_ONCE(ARRAY_SIZE(sdp->srcu_lock_count) !=
133 ARRAY_SIZE(sdp->srcu_unlock_count));
134 level = rcu_num_lvls - 1;
135 snp_first = ssp->level[level];
136 for_each_possible_cpu(cpu) {
137 sdp = per_cpu_ptr(ssp->sda, cpu);
138 spin_lock_init(&ACCESS_PRIVATE(sdp, lock));
139 rcu_segcblist_init(&sdp->srcu_cblist);
140 sdp->srcu_cblist_invoking = false;
141 sdp->srcu_gp_seq_needed = ssp->srcu_gp_seq;
142 sdp->srcu_gp_seq_needed_exp = ssp->srcu_gp_seq;
143 sdp->mynode = &snp_first[cpu / levelspread[level]];
144 for (snp = sdp->mynode; snp != NULL; snp = snp->srcu_parent) {
150 INIT_WORK(&sdp->work, srcu_invoke_callbacks);
151 timer_setup(&sdp->delay_work, srcu_delay_timer, 0);
153 sdp->grpmask = 1 << (cpu - sdp->mynode->grplo);
157 /* Dynamically allocated, better be no srcu_read_locks()! */
158 for (i = 0; i < ARRAY_SIZE(sdp->srcu_lock_count); i++) {
159 sdp->srcu_lock_count[i] = 0;
160 sdp->srcu_unlock_count[i] = 0;
166 * Initialize non-compile-time initialized fields, including the
167 * associated srcu_node and srcu_data structures. The is_static
168 * parameter is passed through to init_srcu_struct_nodes(), and
169 * also tells us that ->sda has already been wired up to srcu_data.
171 static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
173 mutex_init(&ssp->srcu_cb_mutex);
174 mutex_init(&ssp->srcu_gp_mutex);
176 ssp->srcu_gp_seq = 0;
177 ssp->srcu_barrier_seq = 0;
178 mutex_init(&ssp->srcu_barrier_mutex);
179 atomic_set(&ssp->srcu_barrier_cpu_cnt, 0);
180 INIT_DELAYED_WORK(&ssp->work, process_srcu);
182 ssp->sda = alloc_percpu(struct srcu_data);
183 init_srcu_struct_nodes(ssp, is_static);
184 ssp->srcu_gp_seq_needed_exp = 0;
185 ssp->srcu_last_gp_end = ktime_get_mono_fast_ns();
186 smp_store_release(&ssp->srcu_gp_seq_needed, 0); /* Init done. */
187 return ssp->sda ? 0 : -ENOMEM;
190 #ifdef CONFIG_DEBUG_LOCK_ALLOC
192 int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
193 struct lock_class_key *key)
195 /* Don't re-initialize a lock while it is held. */
196 debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
197 lockdep_init_map(&ssp->dep_map, name, key, 0);
198 spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
199 return init_srcu_struct_fields(ssp, false);
201 EXPORT_SYMBOL_GPL(__init_srcu_struct);
203 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
206 * init_srcu_struct - initialize a sleep-RCU structure
207 * @ssp: structure to initialize.
209 * Must invoke this on a given srcu_struct before passing that srcu_struct
210 * to any other function. Each srcu_struct represents a separate domain
211 * of SRCU protection.
213 int init_srcu_struct(struct srcu_struct *ssp)
215 spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
216 return init_srcu_struct_fields(ssp, false);
218 EXPORT_SYMBOL_GPL(init_srcu_struct);
220 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
223 * First-use initialization of statically allocated srcu_struct
224 * structure. Wiring up the combining tree is more than can be
225 * done with compile-time initialization, so this check is added
226 * to each update-side SRCU primitive. Use ssp->lock, which -is-
227 * compile-time initialized, to resolve races involving multiple
228 * CPUs trying to garner first-use privileges.
230 static void check_init_srcu_struct(struct srcu_struct *ssp)
234 /* The smp_load_acquire() pairs with the smp_store_release(). */
235 if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq_needed))) /*^^^*/
236 return; /* Already initialized. */
237 spin_lock_irqsave_rcu_node(ssp, flags);
238 if (!rcu_seq_state(ssp->srcu_gp_seq_needed)) {
239 spin_unlock_irqrestore_rcu_node(ssp, flags);
242 init_srcu_struct_fields(ssp, true);
243 spin_unlock_irqrestore_rcu_node(ssp, flags);
247 * Returns approximate total of the readers' ->srcu_lock_count[] values
248 * for the rank of per-CPU counters specified by idx.
250 static unsigned long srcu_readers_lock_idx(struct srcu_struct *ssp, int idx)
253 unsigned long sum = 0;
255 for_each_possible_cpu(cpu) {
256 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
258 sum += READ_ONCE(cpuc->srcu_lock_count[idx]);
264 * Returns approximate total of the readers' ->srcu_unlock_count[] values
265 * for the rank of per-CPU counters specified by idx.
267 static unsigned long srcu_readers_unlock_idx(struct srcu_struct *ssp, int idx)
270 unsigned long sum = 0;
272 for_each_possible_cpu(cpu) {
273 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
275 sum += READ_ONCE(cpuc->srcu_unlock_count[idx]);
281 * Return true if the number of pre-existing readers is determined to
284 static bool srcu_readers_active_idx_check(struct srcu_struct *ssp, int idx)
286 unsigned long unlocks;
288 unlocks = srcu_readers_unlock_idx(ssp, idx);
291 * Make sure that a lock is always counted if the corresponding
292 * unlock is counted. Needs to be a smp_mb() as the read side may
293 * contain a read from a variable that is written to before the
294 * synchronize_srcu() in the write side. In this case smp_mb()s
295 * A and B act like the store buffering pattern.
297 * This smp_mb() also pairs with smp_mb() C to prevent accesses
298 * after the synchronize_srcu() from being executed before the
304 * If the locks are the same as the unlocks, then there must have
305 * been no readers on this index at some time in between. This does
306 * not mean that there are no more readers, as one could have read
307 * the current index but not have incremented the lock counter yet.
309 * So suppose that the updater is preempted here for so long
310 * that more than ULONG_MAX non-nested readers come and go in
311 * the meantime. It turns out that this cannot result in overflow
312 * because if a reader modifies its unlock count after we read it
313 * above, then that reader's next load of ->srcu_idx is guaranteed
314 * to get the new value, which will cause it to operate on the
315 * other bank of counters, where it cannot contribute to the
316 * overflow of these counters. This means that there is a maximum
317 * of 2*NR_CPUS increments, which cannot overflow given current
318 * systems, especially not on 64-bit systems.
320 * OK, how about nesting? This does impose a limit on nesting
321 * of floor(ULONG_MAX/NR_CPUS/2), which should be sufficient,
322 * especially on 64-bit systems.
324 return srcu_readers_lock_idx(ssp, idx) == unlocks;
328 * srcu_readers_active - returns true if there are readers. and false
330 * @ssp: which srcu_struct to count active readers (holding srcu_read_lock).
332 * Note that this is not an atomic primitive, and can therefore suffer
333 * severe errors when invoked on an active srcu_struct. That said, it
334 * can be useful as an error check at cleanup time.
336 static bool srcu_readers_active(struct srcu_struct *ssp)
339 unsigned long sum = 0;
341 for_each_possible_cpu(cpu) {
342 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
344 sum += READ_ONCE(cpuc->srcu_lock_count[0]);
345 sum += READ_ONCE(cpuc->srcu_lock_count[1]);
346 sum -= READ_ONCE(cpuc->srcu_unlock_count[0]);
347 sum -= READ_ONCE(cpuc->srcu_unlock_count[1]);
352 #define SRCU_INTERVAL 1
355 * Return grace-period delay, zero if there are expedited grace
356 * periods pending, SRCU_INTERVAL otherwise.
358 static unsigned long srcu_get_delay(struct srcu_struct *ssp)
360 if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq),
361 READ_ONCE(ssp->srcu_gp_seq_needed_exp)))
363 return SRCU_INTERVAL;
367 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
368 * @ssp: structure to clean up.
370 * Must invoke this after you are finished using a given srcu_struct that
371 * was initialized via init_srcu_struct(), else you leak memory.
373 void cleanup_srcu_struct(struct srcu_struct *ssp)
377 if (WARN_ON(!srcu_get_delay(ssp)))
378 return; /* Just leak it! */
379 if (WARN_ON(srcu_readers_active(ssp)))
380 return; /* Just leak it! */
381 flush_delayed_work(&ssp->work);
382 for_each_possible_cpu(cpu) {
383 struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
385 del_timer_sync(&sdp->delay_work);
386 flush_work(&sdp->work);
387 if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist)))
388 return; /* Forgot srcu_barrier(), so just leak it! */
390 if (WARN_ON(rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) != SRCU_STATE_IDLE) ||
391 WARN_ON(srcu_readers_active(ssp))) {
392 pr_info("%s: Active srcu_struct %p state: %d\n",
393 __func__, ssp, rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)));
394 return; /* Caller forgot to stop doing call_srcu()? */
396 free_percpu(ssp->sda);
399 EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
402 * Counts the new reader in the appropriate per-CPU element of the
404 * Returns an index that must be passed to the matching srcu_read_unlock().
406 int __srcu_read_lock(struct srcu_struct *ssp)
410 idx = READ_ONCE(ssp->srcu_idx) & 0x1;
411 this_cpu_inc(ssp->sda->srcu_lock_count[idx]);
412 smp_mb(); /* B */ /* Avoid leaking the critical section. */
415 EXPORT_SYMBOL_GPL(__srcu_read_lock);
418 * Removes the count for the old reader from the appropriate per-CPU
419 * element of the srcu_struct. Note that this may well be a different
420 * CPU than that which was incremented by the corresponding srcu_read_lock().
422 void __srcu_read_unlock(struct srcu_struct *ssp, int idx)
424 smp_mb(); /* C */ /* Avoid leaking the critical section. */
425 this_cpu_inc(ssp->sda->srcu_unlock_count[idx]);
427 EXPORT_SYMBOL_GPL(__srcu_read_unlock);
430 * We use an adaptive strategy for synchronize_srcu() and especially for
431 * synchronize_srcu_expedited(). We spin for a fixed time period
432 * (defined below) to allow SRCU readers to exit their read-side critical
433 * sections. If there are still some readers after a few microseconds,
434 * we repeatedly block for 1-millisecond time periods.
436 #define SRCU_RETRY_CHECK_DELAY 5
439 * Start an SRCU grace period.
441 static void srcu_gp_start(struct srcu_struct *ssp)
443 struct srcu_data *sdp = this_cpu_ptr(ssp->sda);
446 lockdep_assert_held(&ACCESS_PRIVATE(ssp, lock));
447 WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
448 spin_lock_rcu_node(sdp); /* Interrupts already disabled. */
449 rcu_segcblist_advance(&sdp->srcu_cblist,
450 rcu_seq_current(&ssp->srcu_gp_seq));
451 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
452 rcu_seq_snap(&ssp->srcu_gp_seq));
453 spin_unlock_rcu_node(sdp); /* Interrupts remain disabled. */
454 smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
455 rcu_seq_start(&ssp->srcu_gp_seq);
456 state = rcu_seq_state(ssp->srcu_gp_seq);
457 WARN_ON_ONCE(state != SRCU_STATE_SCAN1);
461 static void srcu_delay_timer(struct timer_list *t)
463 struct srcu_data *sdp = container_of(t, struct srcu_data, delay_work);
465 queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
468 static void srcu_queue_delayed_work_on(struct srcu_data *sdp,
472 queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
476 timer_reduce(&sdp->delay_work, jiffies + delay);
480 * Schedule callback invocation for the specified srcu_data structure,
481 * if possible, on the corresponding CPU.
483 static void srcu_schedule_cbs_sdp(struct srcu_data *sdp, unsigned long delay)
485 srcu_queue_delayed_work_on(sdp, delay);
489 * Schedule callback invocation for all srcu_data structures associated
490 * with the specified srcu_node structure that have callbacks for the
491 * just-completed grace period, the one corresponding to idx. If possible,
492 * schedule this invocation on the corresponding CPUs.
494 static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp,
495 unsigned long mask, unsigned long delay)
499 for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
500 if (!(mask & (1 << (cpu - snp->grplo))))
502 srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
507 * Note the end of an SRCU grace period. Initiates callback invocation
508 * and starts a new grace period if needed.
510 * The ->srcu_cb_mutex acquisition does not protect any data, but
511 * instead prevents more than one grace period from starting while we
512 * are initiating callback invocation. This allows the ->srcu_have_cbs[]
513 * array to have a finite number of elements.
515 static void srcu_gp_end(struct srcu_struct *ssp)
517 unsigned long cbdelay;
525 struct srcu_data *sdp;
526 struct srcu_node *snp;
528 /* Prevent more than one additional grace period. */
529 mutex_lock(&ssp->srcu_cb_mutex);
531 /* End the current grace period. */
532 spin_lock_irq_rcu_node(ssp);
533 idx = rcu_seq_state(ssp->srcu_gp_seq);
534 WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
535 cbdelay = srcu_get_delay(ssp);
536 WRITE_ONCE(ssp->srcu_last_gp_end, ktime_get_mono_fast_ns());
537 rcu_seq_end(&ssp->srcu_gp_seq);
538 gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
539 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, gpseq))
540 WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, gpseq);
541 spin_unlock_irq_rcu_node(ssp);
542 mutex_unlock(&ssp->srcu_gp_mutex);
543 /* A new grace period can start at this point. But only one. */
545 /* Initiate callback invocation as needed. */
546 idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs);
547 srcu_for_each_node_breadth_first(ssp, snp) {
548 spin_lock_irq_rcu_node(snp);
550 last_lvl = snp >= ssp->level[rcu_num_lvls - 1];
552 cbs = snp->srcu_have_cbs[idx] == gpseq;
553 snp->srcu_have_cbs[idx] = gpseq;
554 rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1);
555 if (ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, gpseq))
556 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, gpseq);
557 mask = snp->srcu_data_have_cbs[idx];
558 snp->srcu_data_have_cbs[idx] = 0;
559 spin_unlock_irq_rcu_node(snp);
561 srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay);
563 /* Occasionally prevent srcu_data counter wrap. */
564 if (!(gpseq & counter_wrap_check) && last_lvl)
565 for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
566 sdp = per_cpu_ptr(ssp->sda, cpu);
567 spin_lock_irqsave_rcu_node(sdp, flags);
568 if (ULONG_CMP_GE(gpseq,
569 sdp->srcu_gp_seq_needed + 100))
570 sdp->srcu_gp_seq_needed = gpseq;
571 if (ULONG_CMP_GE(gpseq,
572 sdp->srcu_gp_seq_needed_exp + 100))
573 sdp->srcu_gp_seq_needed_exp = gpseq;
574 spin_unlock_irqrestore_rcu_node(sdp, flags);
578 /* Callback initiation done, allow grace periods after next. */
579 mutex_unlock(&ssp->srcu_cb_mutex);
581 /* Start a new grace period if needed. */
582 spin_lock_irq_rcu_node(ssp);
583 gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
584 if (!rcu_seq_state(gpseq) &&
585 ULONG_CMP_LT(gpseq, ssp->srcu_gp_seq_needed)) {
587 spin_unlock_irq_rcu_node(ssp);
588 srcu_reschedule(ssp, 0);
590 spin_unlock_irq_rcu_node(ssp);
595 * Funnel-locking scheme to scalably mediate many concurrent expedited
596 * grace-period requests. This function is invoked for the first known
597 * expedited request for a grace period that has already been requested,
598 * but without expediting. To start a completely new grace period,
599 * whether expedited or not, use srcu_funnel_gp_start() instead.
601 static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp,
606 for (; snp != NULL; snp = snp->srcu_parent) {
607 if (rcu_seq_done(&ssp->srcu_gp_seq, s) ||
608 ULONG_CMP_GE(READ_ONCE(snp->srcu_gp_seq_needed_exp), s))
610 spin_lock_irqsave_rcu_node(snp, flags);
611 if (ULONG_CMP_GE(snp->srcu_gp_seq_needed_exp, s)) {
612 spin_unlock_irqrestore_rcu_node(snp, flags);
615 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
616 spin_unlock_irqrestore_rcu_node(snp, flags);
618 spin_lock_irqsave_rcu_node(ssp, flags);
619 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
620 WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, s);
621 spin_unlock_irqrestore_rcu_node(ssp, flags);
625 * Funnel-locking scheme to scalably mediate many concurrent grace-period
626 * requests. The winner has to do the work of actually starting grace
627 * period s. Losers must either ensure that their desired grace-period
628 * number is recorded on at least their leaf srcu_node structure, or they
629 * must take steps to invoke their own callbacks.
631 * Note that this function also does the work of srcu_funnel_exp_start(),
632 * in some cases by directly invoking it.
634 static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
635 unsigned long s, bool do_norm)
638 int idx = rcu_seq_ctr(s) % ARRAY_SIZE(sdp->mynode->srcu_have_cbs);
639 struct srcu_node *snp = sdp->mynode;
640 unsigned long snp_seq;
642 /* Each pass through the loop does one level of the srcu_node tree. */
643 for (; snp != NULL; snp = snp->srcu_parent) {
644 if (rcu_seq_done(&ssp->srcu_gp_seq, s) && snp != sdp->mynode)
645 return; /* GP already done and CBs recorded. */
646 spin_lock_irqsave_rcu_node(snp, flags);
647 if (ULONG_CMP_GE(snp->srcu_have_cbs[idx], s)) {
648 snp_seq = snp->srcu_have_cbs[idx];
649 if (snp == sdp->mynode && snp_seq == s)
650 snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
651 spin_unlock_irqrestore_rcu_node(snp, flags);
652 if (snp == sdp->mynode && snp_seq != s) {
653 srcu_schedule_cbs_sdp(sdp, do_norm
659 srcu_funnel_exp_start(ssp, snp, s);
662 snp->srcu_have_cbs[idx] = s;
663 if (snp == sdp->mynode)
664 snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
665 if (!do_norm && ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, s))
666 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
667 spin_unlock_irqrestore_rcu_node(snp, flags);
670 /* Top of tree, must ensure the grace period will be started. */
671 spin_lock_irqsave_rcu_node(ssp, flags);
672 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed, s)) {
674 * Record need for grace period s. Pair with load
675 * acquire setting up for initialization.
677 smp_store_release(&ssp->srcu_gp_seq_needed, s); /*^^^*/
679 if (!do_norm && ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
680 WRITE_ONCE(ssp->srcu_gp_seq_needed_exp, s);
682 /* If grace period not already done and none in progress, start it. */
683 if (!rcu_seq_done(&ssp->srcu_gp_seq, s) &&
684 rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
685 WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
687 if (likely(srcu_init_done))
688 queue_delayed_work(rcu_gp_wq, &ssp->work,
689 srcu_get_delay(ssp));
690 else if (list_empty(&ssp->work.work.entry))
691 list_add(&ssp->work.work.entry, &srcu_boot_list);
693 spin_unlock_irqrestore_rcu_node(ssp, flags);
697 * Wait until all readers counted by array index idx complete, but
698 * loop an additional time if there is an expedited grace period pending.
699 * The caller must ensure that ->srcu_idx is not changed while checking.
701 static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount)
704 if (srcu_readers_active_idx_check(ssp, idx))
706 if (--trycount + !srcu_get_delay(ssp) <= 0)
708 udelay(SRCU_RETRY_CHECK_DELAY);
713 * Increment the ->srcu_idx counter so that future SRCU readers will
714 * use the other rank of the ->srcu_(un)lock_count[] arrays. This allows
715 * us to wait for pre-existing readers in a starvation-free manner.
717 static void srcu_flip(struct srcu_struct *ssp)
720 * Ensure that if this updater saw a given reader's increment
721 * from __srcu_read_lock(), that reader was using an old value
722 * of ->srcu_idx. Also ensure that if a given reader sees the
723 * new value of ->srcu_idx, this updater's earlier scans cannot
724 * have seen that reader's increments (which is OK, because this
725 * grace period need not wait on that reader).
727 smp_mb(); /* E */ /* Pairs with B and C. */
729 WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1);
732 * Ensure that if the updater misses an __srcu_read_unlock()
733 * increment, that task's next __srcu_read_lock() will see the
734 * above counter update. Note that both this memory barrier
735 * and the one in srcu_readers_active_idx_check() provide the
736 * guarantee for __srcu_read_lock().
738 smp_mb(); /* D */ /* Pairs with C. */
742 * If SRCU is likely idle, return true, otherwise return false.
744 * Note that it is OK for several current from-idle requests for a new
745 * grace period from idle to specify expediting because they will all end
746 * up requesting the same grace period anyhow. So no loss.
748 * Note also that if any CPU (including the current one) is still invoking
749 * callbacks, this function will nevertheless say "idle". This is not
750 * ideal, but the overhead of checking all CPUs' callback lists is even
751 * less ideal, especially on large systems. Furthermore, the wakeup
752 * can happen before the callback is fully removed, so we have no choice
753 * but to accept this type of error.
755 * This function is also subject to counter-wrap errors, but let's face
756 * it, if this function was preempted for enough time for the counters
757 * to wrap, it really doesn't matter whether or not we expedite the grace
758 * period. The extra overhead of a needlessly expedited grace period is
759 * negligible when amortized over that time period, and the extra latency
760 * of a needlessly non-expedited grace period is similarly negligible.
762 static bool srcu_might_be_idle(struct srcu_struct *ssp)
764 unsigned long curseq;
766 struct srcu_data *sdp;
770 check_init_srcu_struct(ssp);
771 /* If the local srcu_data structure has callbacks, not idle. */
772 sdp = raw_cpu_ptr(ssp->sda);
773 spin_lock_irqsave_rcu_node(sdp, flags);
774 if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) {
775 spin_unlock_irqrestore_rcu_node(sdp, flags);
776 return false; /* Callbacks already present, so not idle. */
778 spin_unlock_irqrestore_rcu_node(sdp, flags);
781 * No local callbacks, so probabalistically probe global state.
782 * Exact information would require acquiring locks, which would
783 * kill scalability, hence the probabalistic nature of the probe.
786 /* First, see if enough time has passed since the last GP. */
787 t = ktime_get_mono_fast_ns();
788 tlast = READ_ONCE(ssp->srcu_last_gp_end);
789 if (exp_holdoff == 0 ||
790 time_in_range_open(t, tlast, tlast + exp_holdoff))
791 return false; /* Too soon after last GP. */
793 /* Next, check for probable idleness. */
794 curseq = rcu_seq_current(&ssp->srcu_gp_seq);
795 smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
796 if (ULONG_CMP_LT(curseq, READ_ONCE(ssp->srcu_gp_seq_needed)))
797 return false; /* Grace period in progress, so not idle. */
798 smp_mb(); /* Order ->srcu_gp_seq with prior access. */
799 if (curseq != rcu_seq_current(&ssp->srcu_gp_seq))
800 return false; /* GP # changed, so not idle. */
801 return true; /* With reasonable probability, idle! */
805 * SRCU callback function to leak a callback.
807 static void srcu_leak_callback(struct rcu_head *rhp)
812 * Start an SRCU grace period, and also queue the callback if non-NULL.
814 static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
815 struct rcu_head *rhp, bool do_norm)
819 bool needexp = false;
822 struct srcu_data *sdp;
824 check_init_srcu_struct(ssp);
825 idx = srcu_read_lock(ssp);
826 sdp = raw_cpu_ptr(ssp->sda);
827 spin_lock_irqsave_rcu_node(sdp, flags);
829 rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp);
830 rcu_segcblist_advance(&sdp->srcu_cblist,
831 rcu_seq_current(&ssp->srcu_gp_seq));
832 s = rcu_seq_snap(&ssp->srcu_gp_seq);
833 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, s);
834 if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) {
835 sdp->srcu_gp_seq_needed = s;
838 if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) {
839 sdp->srcu_gp_seq_needed_exp = s;
842 spin_unlock_irqrestore_rcu_node(sdp, flags);
844 srcu_funnel_gp_start(ssp, sdp, s, do_norm);
846 srcu_funnel_exp_start(ssp, sdp->mynode, s);
847 srcu_read_unlock(ssp, idx);
852 * Enqueue an SRCU callback on the srcu_data structure associated with
853 * the current CPU and the specified srcu_struct structure, initiating
854 * grace-period processing if it is not already running.
856 * Note that all CPUs must agree that the grace period extended beyond
857 * all pre-existing SRCU read-side critical section. On systems with
858 * more than one CPU, this means that when "func()" is invoked, each CPU
859 * is guaranteed to have executed a full memory barrier since the end of
860 * its last corresponding SRCU read-side critical section whose beginning
861 * preceded the call to call_srcu(). It also means that each CPU executing
862 * an SRCU read-side critical section that continues beyond the start of
863 * "func()" must have executed a memory barrier after the call_srcu()
864 * but before the beginning of that SRCU read-side critical section.
865 * Note that these guarantees include CPUs that are offline, idle, or
866 * executing in user mode, as well as CPUs that are executing in the kernel.
868 * Furthermore, if CPU A invoked call_srcu() and CPU B invoked the
869 * resulting SRCU callback function "func()", then both CPU A and CPU
870 * B are guaranteed to execute a full memory barrier during the time
871 * interval between the call to call_srcu() and the invocation of "func()".
872 * This guarantee applies even if CPU A and CPU B are the same CPU (but
873 * again only if the system has more than one CPU).
875 * Of course, these guarantees apply only for invocations of call_srcu(),
876 * srcu_read_lock(), and srcu_read_unlock() that are all passed the same
877 * srcu_struct structure.
879 static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
880 rcu_callback_t func, bool do_norm)
882 if (debug_rcu_head_queue(rhp)) {
883 /* Probable double call_srcu(), so leak the callback. */
884 WRITE_ONCE(rhp->func, srcu_leak_callback);
885 WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n");
889 (void)srcu_gp_start_if_needed(ssp, rhp, do_norm);
893 * call_srcu() - Queue a callback for invocation after an SRCU grace period
894 * @ssp: srcu_struct in queue the callback
895 * @rhp: structure to be used for queueing the SRCU callback.
896 * @func: function to be invoked after the SRCU grace period
898 * The callback function will be invoked some time after a full SRCU
899 * grace period elapses, in other words after all pre-existing SRCU
900 * read-side critical sections have completed. However, the callback
901 * function might well execute concurrently with other SRCU read-side
902 * critical sections that started after call_srcu() was invoked. SRCU
903 * read-side critical sections are delimited by srcu_read_lock() and
904 * srcu_read_unlock(), and may be nested.
906 * The callback will be invoked from process context, but must nevertheless
907 * be fast and must not block.
909 void call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
912 __call_srcu(ssp, rhp, func, true);
914 EXPORT_SYMBOL_GPL(call_srcu);
917 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
919 static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
921 struct rcu_synchronize rcu;
923 RCU_LOCKDEP_WARN(lock_is_held(&ssp->dep_map) ||
924 lock_is_held(&rcu_bh_lock_map) ||
925 lock_is_held(&rcu_lock_map) ||
926 lock_is_held(&rcu_sched_lock_map),
927 "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
929 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
932 check_init_srcu_struct(ssp);
933 init_completion(&rcu.completion);
934 init_rcu_head_on_stack(&rcu.head);
935 __call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm);
936 wait_for_completion(&rcu.completion);
937 destroy_rcu_head_on_stack(&rcu.head);
940 * Make sure that later code is ordered after the SRCU grace
941 * period. This pairs with the spin_lock_irq_rcu_node()
942 * in srcu_invoke_callbacks(). Unlike Tree RCU, this is needed
943 * because the current CPU might have been totally uninvolved with
944 * (and thus unordered against) that grace period.
950 * synchronize_srcu_expedited - Brute-force SRCU grace period
951 * @ssp: srcu_struct with which to synchronize.
953 * Wait for an SRCU grace period to elapse, but be more aggressive about
954 * spinning rather than blocking when waiting.
956 * Note that synchronize_srcu_expedited() has the same deadlock and
957 * memory-ordering properties as does synchronize_srcu().
959 void synchronize_srcu_expedited(struct srcu_struct *ssp)
961 __synchronize_srcu(ssp, rcu_gp_is_normal());
963 EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
966 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
967 * @ssp: srcu_struct with which to synchronize.
969 * Wait for the count to drain to zero of both indexes. To avoid the
970 * possible starvation of synchronize_srcu(), it waits for the count of
971 * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
972 * and then flip the srcu_idx and wait for the count of the other index.
974 * Can block; must be called from process context.
976 * Note that it is illegal to call synchronize_srcu() from the corresponding
977 * SRCU read-side critical section; doing so will result in deadlock.
978 * However, it is perfectly legal to call synchronize_srcu() on one
979 * srcu_struct from some other srcu_struct's read-side critical section,
980 * as long as the resulting graph of srcu_structs is acyclic.
982 * There are memory-ordering constraints implied by synchronize_srcu().
983 * On systems with more than one CPU, when synchronize_srcu() returns,
984 * each CPU is guaranteed to have executed a full memory barrier since
985 * the end of its last corresponding SRCU read-side critical section
986 * whose beginning preceded the call to synchronize_srcu(). In addition,
987 * each CPU having an SRCU read-side critical section that extends beyond
988 * the return from synchronize_srcu() is guaranteed to have executed a
989 * full memory barrier after the beginning of synchronize_srcu() and before
990 * the beginning of that SRCU read-side critical section. Note that these
991 * guarantees include CPUs that are offline, idle, or executing in user mode,
992 * as well as CPUs that are executing in the kernel.
994 * Furthermore, if CPU A invoked synchronize_srcu(), which returned
995 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
996 * to have executed a full memory barrier during the execution of
997 * synchronize_srcu(). This guarantee applies even if CPU A and CPU B
998 * are the same CPU, but again only if the system has more than one CPU.
1000 * Of course, these memory-ordering guarantees apply only when
1001 * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
1002 * passed the same srcu_struct structure.
1004 * If SRCU is likely idle, expedite the first request. This semantic
1005 * was provided by Classic SRCU, and is relied upon by its users, so TREE
1006 * SRCU must also provide it. Note that detecting idleness is heuristic
1007 * and subject to both false positives and negatives.
1009 void synchronize_srcu(struct srcu_struct *ssp)
1011 if (srcu_might_be_idle(ssp) || rcu_gp_is_expedited())
1012 synchronize_srcu_expedited(ssp);
1014 __synchronize_srcu(ssp, true);
1016 EXPORT_SYMBOL_GPL(synchronize_srcu);
1019 * get_state_synchronize_srcu - Provide an end-of-grace-period cookie
1020 * @ssp: srcu_struct to provide cookie for.
1022 * This function returns a cookie that can be passed to
1023 * poll_state_synchronize_srcu(), which will return true if a full grace
1024 * period has elapsed in the meantime. It is the caller's responsibility
1025 * to make sure that grace period happens, for example, by invoking
1026 * call_srcu() after return from get_state_synchronize_srcu().
1028 unsigned long get_state_synchronize_srcu(struct srcu_struct *ssp)
1030 // Any prior manipulation of SRCU-protected data must happen
1031 // before the load from ->srcu_gp_seq.
1033 return rcu_seq_snap(&ssp->srcu_gp_seq);
1035 EXPORT_SYMBOL_GPL(get_state_synchronize_srcu);
1038 * start_poll_synchronize_srcu - Provide cookie and start grace period
1039 * @ssp: srcu_struct to provide cookie for.
1041 * This function returns a cookie that can be passed to
1042 * poll_state_synchronize_srcu(), which will return true if a full grace
1043 * period has elapsed in the meantime. Unlike get_state_synchronize_srcu(),
1044 * this function also ensures that any needed SRCU grace period will be
1045 * started. This convenience does come at a cost in terms of CPU overhead.
1047 unsigned long start_poll_synchronize_srcu(struct srcu_struct *ssp)
1049 return srcu_gp_start_if_needed(ssp, NULL, true);
1051 EXPORT_SYMBOL_GPL(start_poll_synchronize_srcu);
1054 * poll_state_synchronize_srcu - Has cookie's grace period ended?
1055 * @ssp: srcu_struct to provide cookie for.
1056 * @cookie: Return value from get_state_synchronize_srcu() or start_poll_synchronize_srcu().
1058 * This function takes the cookie that was returned from either
1059 * get_state_synchronize_srcu() or start_poll_synchronize_srcu(), and
1060 * returns @true if an SRCU grace period elapsed since the time that the
1061 * cookie was created.
1063 bool poll_state_synchronize_srcu(struct srcu_struct *ssp, unsigned long cookie)
1065 if (!rcu_seq_done(&ssp->srcu_gp_seq, cookie))
1067 // Ensure that the end of the SRCU grace period happens before
1068 // any subsequent code that the caller might execute.
1072 EXPORT_SYMBOL_GPL(poll_state_synchronize_srcu);
1075 * Callback function for srcu_barrier() use.
1077 static void srcu_barrier_cb(struct rcu_head *rhp)
1079 struct srcu_data *sdp;
1080 struct srcu_struct *ssp;
1082 sdp = container_of(rhp, struct srcu_data, srcu_barrier_head);
1084 if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
1085 complete(&ssp->srcu_barrier_completion);
1089 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
1090 * @ssp: srcu_struct on which to wait for in-flight callbacks.
1092 void srcu_barrier(struct srcu_struct *ssp)
1095 struct srcu_data *sdp;
1096 unsigned long s = rcu_seq_snap(&ssp->srcu_barrier_seq);
1098 check_init_srcu_struct(ssp);
1099 mutex_lock(&ssp->srcu_barrier_mutex);
1100 if (rcu_seq_done(&ssp->srcu_barrier_seq, s)) {
1101 smp_mb(); /* Force ordering following return. */
1102 mutex_unlock(&ssp->srcu_barrier_mutex);
1103 return; /* Someone else did our work for us. */
1105 rcu_seq_start(&ssp->srcu_barrier_seq);
1106 init_completion(&ssp->srcu_barrier_completion);
1108 /* Initial count prevents reaching zero until all CBs are posted. */
1109 atomic_set(&ssp->srcu_barrier_cpu_cnt, 1);
1112 * Each pass through this loop enqueues a callback, but only
1113 * on CPUs already having callbacks enqueued. Note that if
1114 * a CPU already has callbacks enqueue, it must have already
1115 * registered the need for a future grace period, so all we
1116 * need do is enqueue a callback that will use the same
1117 * grace period as the last callback already in the queue.
1119 for_each_possible_cpu(cpu) {
1120 sdp = per_cpu_ptr(ssp->sda, cpu);
1121 spin_lock_irq_rcu_node(sdp);
1122 atomic_inc(&ssp->srcu_barrier_cpu_cnt);
1123 sdp->srcu_barrier_head.func = srcu_barrier_cb;
1124 debug_rcu_head_queue(&sdp->srcu_barrier_head);
1125 if (!rcu_segcblist_entrain(&sdp->srcu_cblist,
1126 &sdp->srcu_barrier_head)) {
1127 debug_rcu_head_unqueue(&sdp->srcu_barrier_head);
1128 atomic_dec(&ssp->srcu_barrier_cpu_cnt);
1130 spin_unlock_irq_rcu_node(sdp);
1133 /* Remove the initial count, at which point reaching zero can happen. */
1134 if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
1135 complete(&ssp->srcu_barrier_completion);
1136 wait_for_completion(&ssp->srcu_barrier_completion);
1138 rcu_seq_end(&ssp->srcu_barrier_seq);
1139 mutex_unlock(&ssp->srcu_barrier_mutex);
1141 EXPORT_SYMBOL_GPL(srcu_barrier);
1144 * srcu_batches_completed - return batches completed.
1145 * @ssp: srcu_struct on which to report batch completion.
1147 * Report the number of batches, correlated with, but not necessarily
1148 * precisely the same as, the number of grace periods that have elapsed.
1150 unsigned long srcu_batches_completed(struct srcu_struct *ssp)
1152 return READ_ONCE(ssp->srcu_idx);
1154 EXPORT_SYMBOL_GPL(srcu_batches_completed);
1157 * Core SRCU state machine. Push state bits of ->srcu_gp_seq
1158 * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
1159 * completed in that state.
1161 static void srcu_advance_state(struct srcu_struct *ssp)
1165 mutex_lock(&ssp->srcu_gp_mutex);
1168 * Because readers might be delayed for an extended period after
1169 * fetching ->srcu_idx for their index, at any point in time there
1170 * might well be readers using both idx=0 and idx=1. We therefore
1171 * need to wait for readers to clear from both index values before
1172 * invoking a callback.
1174 * The load-acquire ensures that we see the accesses performed
1175 * by the prior grace period.
1177 idx = rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq)); /* ^^^ */
1178 if (idx == SRCU_STATE_IDLE) {
1179 spin_lock_irq_rcu_node(ssp);
1180 if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
1181 WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq));
1182 spin_unlock_irq_rcu_node(ssp);
1183 mutex_unlock(&ssp->srcu_gp_mutex);
1186 idx = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq));
1187 if (idx == SRCU_STATE_IDLE)
1189 spin_unlock_irq_rcu_node(ssp);
1190 if (idx != SRCU_STATE_IDLE) {
1191 mutex_unlock(&ssp->srcu_gp_mutex);
1192 return; /* Someone else started the grace period. */
1196 if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN1) {
1197 idx = 1 ^ (ssp->srcu_idx & 1);
1198 if (!try_check_zero(ssp, idx, 1)) {
1199 mutex_unlock(&ssp->srcu_gp_mutex);
1200 return; /* readers present, retry later. */
1203 spin_lock_irq_rcu_node(ssp);
1204 rcu_seq_set_state(&ssp->srcu_gp_seq, SRCU_STATE_SCAN2);
1205 spin_unlock_irq_rcu_node(ssp);
1208 if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN2) {
1211 * SRCU read-side critical sections are normally short,
1212 * so check at least twice in quick succession after a flip.
1214 idx = 1 ^ (ssp->srcu_idx & 1);
1215 if (!try_check_zero(ssp, idx, 2)) {
1216 mutex_unlock(&ssp->srcu_gp_mutex);
1217 return; /* readers present, retry later. */
1219 srcu_gp_end(ssp); /* Releases ->srcu_gp_mutex. */
1224 * Invoke a limited number of SRCU callbacks that have passed through
1225 * their grace period. If there are more to do, SRCU will reschedule
1226 * the workqueue. Note that needed memory barriers have been executed
1227 * in this task's context by srcu_readers_active_idx_check().
1229 static void srcu_invoke_callbacks(struct work_struct *work)
1232 struct rcu_cblist ready_cbs;
1233 struct rcu_head *rhp;
1234 struct srcu_data *sdp;
1235 struct srcu_struct *ssp;
1237 sdp = container_of(work, struct srcu_data, work);
1240 rcu_cblist_init(&ready_cbs);
1241 spin_lock_irq_rcu_node(sdp);
1242 rcu_segcblist_advance(&sdp->srcu_cblist,
1243 rcu_seq_current(&ssp->srcu_gp_seq));
1244 if (sdp->srcu_cblist_invoking ||
1245 !rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) {
1246 spin_unlock_irq_rcu_node(sdp);
1247 return; /* Someone else on the job or nothing to do. */
1250 /* We are on the job! Extract and invoke ready callbacks. */
1251 sdp->srcu_cblist_invoking = true;
1252 rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs);
1253 spin_unlock_irq_rcu_node(sdp);
1254 rhp = rcu_cblist_dequeue(&ready_cbs);
1255 for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) {
1256 debug_rcu_head_unqueue(rhp);
1263 * Update counts, accelerate new callbacks, and if needed,
1264 * schedule another round of callback invocation.
1266 spin_lock_irq_rcu_node(sdp);
1267 rcu_segcblist_insert_count(&sdp->srcu_cblist, &ready_cbs);
1268 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
1269 rcu_seq_snap(&ssp->srcu_gp_seq));
1270 sdp->srcu_cblist_invoking = false;
1271 more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist);
1272 spin_unlock_irq_rcu_node(sdp);
1274 srcu_schedule_cbs_sdp(sdp, 0);
1278 * Finished one round of SRCU grace period. Start another if there are
1279 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
1281 static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
1285 spin_lock_irq_rcu_node(ssp);
1286 if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
1287 if (!WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq))) {
1288 /* All requests fulfilled, time to go idle. */
1291 } else if (!rcu_seq_state(ssp->srcu_gp_seq)) {
1292 /* Outstanding request and no GP. Start one. */
1295 spin_unlock_irq_rcu_node(ssp);
1298 queue_delayed_work(rcu_gp_wq, &ssp->work, delay);
1302 * This is the work-queue function that handles SRCU grace periods.
1304 static void process_srcu(struct work_struct *work)
1306 struct srcu_struct *ssp;
1308 ssp = container_of(work, struct srcu_struct, work.work);
1310 srcu_advance_state(ssp);
1311 srcu_reschedule(ssp, srcu_get_delay(ssp));
1314 void srcutorture_get_gp_data(enum rcutorture_type test_type,
1315 struct srcu_struct *ssp, int *flags,
1316 unsigned long *gp_seq)
1318 if (test_type != SRCU_FLAVOR)
1321 *gp_seq = rcu_seq_current(&ssp->srcu_gp_seq);
1323 EXPORT_SYMBOL_GPL(srcutorture_get_gp_data);
1325 void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf)
1329 unsigned long s0 = 0, s1 = 0;
1331 idx = ssp->srcu_idx & 0x1;
1332 pr_alert("%s%s Tree SRCU g%ld per-CPU(idx=%d):",
1333 tt, tf, rcu_seq_current(&ssp->srcu_gp_seq), idx);
1334 for_each_possible_cpu(cpu) {
1335 unsigned long l0, l1;
1336 unsigned long u0, u1;
1338 struct srcu_data *sdp;
1340 sdp = per_cpu_ptr(ssp->sda, cpu);
1341 u0 = data_race(sdp->srcu_unlock_count[!idx]);
1342 u1 = data_race(sdp->srcu_unlock_count[idx]);
1345 * Make sure that a lock is always counted if the corresponding
1346 * unlock is counted.
1350 l0 = data_race(sdp->srcu_lock_count[!idx]);
1351 l1 = data_race(sdp->srcu_lock_count[idx]);
1355 pr_cont(" %d(%ld,%ld %c)",
1357 "C."[rcu_segcblist_empty(&sdp->srcu_cblist)]);
1361 pr_cont(" T(%ld,%ld)\n", s0, s1);
1363 EXPORT_SYMBOL_GPL(srcu_torture_stats_print);
1365 static int __init srcu_bootup_announce(void)
1367 pr_info("Hierarchical SRCU implementation.\n");
1368 if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF)
1369 pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff);
1372 early_initcall(srcu_bootup_announce);
1374 void __init srcu_init(void)
1376 struct srcu_struct *ssp;
1378 srcu_init_done = true;
1379 while (!list_empty(&srcu_boot_list)) {
1380 ssp = list_first_entry(&srcu_boot_list, struct srcu_struct,
1382 check_init_srcu_struct(ssp);
1383 list_del_init(&ssp->work.work.entry);
1384 queue_work(rcu_gp_wq, &ssp->work.work);
1388 #ifdef CONFIG_MODULES
1390 /* Initialize any global-scope srcu_struct structures used by this module. */
1391 static int srcu_module_coming(struct module *mod)
1394 struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1397 for (i = 0; i < mod->num_srcu_structs; i++) {
1398 ret = init_srcu_struct(*(sspp++));
1399 if (WARN_ON_ONCE(ret))
1405 /* Clean up any global-scope srcu_struct structures used by this module. */
1406 static void srcu_module_going(struct module *mod)
1409 struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1411 for (i = 0; i < mod->num_srcu_structs; i++)
1412 cleanup_srcu_struct(*(sspp++));
1415 /* Handle one module, either coming or going. */
1416 static int srcu_module_notify(struct notifier_block *self,
1417 unsigned long val, void *data)
1419 struct module *mod = data;
1423 case MODULE_STATE_COMING:
1424 ret = srcu_module_coming(mod);
1426 case MODULE_STATE_GOING:
1427 srcu_module_going(mod);
1435 static struct notifier_block srcu_module_nb = {
1436 .notifier_call = srcu_module_notify,
1440 static __init int init_srcu_module_notifier(void)
1444 ret = register_module_notifier(&srcu_module_nb);
1446 pr_warn("Failed to register srcu module notifier\n");
1449 late_initcall(init_srcu_module_notifier);
1451 #endif /* #ifdef CONFIG_MODULES */