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_sup->srcu_gp_seq;
139 sdp->srcu_gp_seq_needed_exp = ssp->srcu_sup->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 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->srcu_sup->node = kcalloc(rcu_num_nodes, sizeof(*ssp->srcu_sup->node), gfp_flags);
177 if (!ssp->srcu_sup->node)
180 /* Work out the overall tree geometry. */
181 ssp->srcu_sup->level[0] = &ssp->srcu_sup->node[0];
182 for (i = 1; i < rcu_num_lvls; i++)
183 ssp->srcu_sup->level[i] = ssp->srcu_sup->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->srcu_sup->node[0]) {
199 /* Root node, special case. */
200 snp->srcu_parent = NULL;
205 if (snp == ssp->srcu_sup->level[level + 1])
207 snp->srcu_parent = ssp->srcu_sup->level[level - 1] +
208 (snp - ssp->srcu_sup->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->srcu_sup->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 = 1UL << (cpu - sdp->mynode->grplo);
228 smp_store_release(&ssp->srcu_sup->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)
240 ssp->srcu_sup = kzalloc(sizeof(*ssp->srcu_sup), GFP_KERNEL);
244 spin_lock_init(&ACCESS_PRIVATE(ssp->srcu_sup, lock));
245 ssp->srcu_sup->srcu_size_state = SRCU_SIZE_SMALL;
246 ssp->srcu_sup->node = NULL;
247 mutex_init(&ssp->srcu_sup->srcu_cb_mutex);
248 mutex_init(&ssp->srcu_sup->srcu_gp_mutex);
250 ssp->srcu_sup->srcu_gp_seq = 0;
251 ssp->srcu_sup->srcu_barrier_seq = 0;
252 mutex_init(&ssp->srcu_sup->srcu_barrier_mutex);
253 atomic_set(&ssp->srcu_sup->srcu_barrier_cpu_cnt, 0);
254 INIT_DELAYED_WORK(&ssp->srcu_sup->work, process_srcu);
255 ssp->srcu_sup->sda_is_static = is_static;
257 ssp->sda = alloc_percpu(struct srcu_data);
260 kfree(ssp->srcu_sup);
263 init_srcu_struct_data(ssp);
264 ssp->srcu_sup->srcu_gp_seq_needed_exp = 0;
265 ssp->srcu_sup->srcu_last_gp_end = ktime_get_mono_fast_ns();
266 if (READ_ONCE(ssp->srcu_sup->srcu_size_state) == SRCU_SIZE_SMALL && SRCU_SIZING_IS_INIT()) {
267 if (!init_srcu_struct_nodes(ssp, GFP_ATOMIC)) {
268 if (!ssp->srcu_sup->sda_is_static) {
269 free_percpu(ssp->sda);
271 kfree(ssp->srcu_sup);
275 WRITE_ONCE(ssp->srcu_sup->srcu_size_state, SRCU_SIZE_BIG);
278 ssp->srcu_sup->srcu_ssp = ssp;
279 smp_store_release(&ssp->srcu_sup->srcu_gp_seq_needed, 0); /* Init done. */
283 #ifdef CONFIG_DEBUG_LOCK_ALLOC
285 int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
286 struct lock_class_key *key)
288 /* Don't re-initialize a lock while it is held. */
289 debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
290 lockdep_init_map(&ssp->dep_map, name, key, 0);
291 return init_srcu_struct_fields(ssp, false);
293 EXPORT_SYMBOL_GPL(__init_srcu_struct);
295 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
298 * init_srcu_struct - initialize a sleep-RCU structure
299 * @ssp: structure to initialize.
301 * Must invoke this on a given srcu_struct before passing that srcu_struct
302 * to any other function. Each srcu_struct represents a separate domain
303 * of SRCU protection.
305 int init_srcu_struct(struct srcu_struct *ssp)
307 return init_srcu_struct_fields(ssp, false);
309 EXPORT_SYMBOL_GPL(init_srcu_struct);
311 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
314 * Initiate a transition to SRCU_SIZE_BIG with lock held.
316 static void __srcu_transition_to_big(struct srcu_struct *ssp)
318 lockdep_assert_held(&ACCESS_PRIVATE(ssp->srcu_sup, lock));
319 smp_store_release(&ssp->srcu_sup->srcu_size_state, SRCU_SIZE_ALLOC);
323 * Initiate an idempotent transition to SRCU_SIZE_BIG.
325 static void srcu_transition_to_big(struct srcu_struct *ssp)
329 /* Double-checked locking on ->srcu_size-state. */
330 if (smp_load_acquire(&ssp->srcu_sup->srcu_size_state) != SRCU_SIZE_SMALL)
332 spin_lock_irqsave_rcu_node(ssp->srcu_sup, flags);
333 if (smp_load_acquire(&ssp->srcu_sup->srcu_size_state) != SRCU_SIZE_SMALL) {
334 spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
337 __srcu_transition_to_big(ssp);
338 spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
342 * Check to see if the just-encountered contention event justifies
343 * a transition to SRCU_SIZE_BIG.
345 static void spin_lock_irqsave_check_contention(struct srcu_struct *ssp)
349 if (!SRCU_SIZING_IS_CONTEND() || ssp->srcu_sup->srcu_size_state)
352 if (ssp->srcu_sup->srcu_size_jiffies != j) {
353 ssp->srcu_sup->srcu_size_jiffies = j;
354 ssp->srcu_sup->srcu_n_lock_retries = 0;
356 if (++ssp->srcu_sup->srcu_n_lock_retries <= small_contention_lim)
358 __srcu_transition_to_big(ssp);
362 * Acquire the specified srcu_data structure's ->lock, but check for
363 * excessive contention, which results in initiation of a transition
364 * to SRCU_SIZE_BIG. But only if the srcutree.convert_to_big module
365 * parameter permits this.
367 static void spin_lock_irqsave_sdp_contention(struct srcu_data *sdp, unsigned long *flags)
369 struct srcu_struct *ssp = sdp->ssp;
371 if (spin_trylock_irqsave_rcu_node(sdp, *flags))
373 spin_lock_irqsave_rcu_node(ssp->srcu_sup, *flags);
374 spin_lock_irqsave_check_contention(ssp);
375 spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, *flags);
376 spin_lock_irqsave_rcu_node(sdp, *flags);
380 * Acquire the specified srcu_struct structure's ->lock, but check for
381 * excessive contention, which results in initiation of a transition
382 * to SRCU_SIZE_BIG. But only if the srcutree.convert_to_big module
383 * parameter permits this.
385 static void spin_lock_irqsave_ssp_contention(struct srcu_struct *ssp, unsigned long *flags)
387 if (spin_trylock_irqsave_rcu_node(ssp->srcu_sup, *flags))
389 spin_lock_irqsave_rcu_node(ssp->srcu_sup, *flags);
390 spin_lock_irqsave_check_contention(ssp);
394 * First-use initialization of statically allocated srcu_struct
395 * structure. Wiring up the combining tree is more than can be
396 * done with compile-time initialization, so this check is added
397 * to each update-side SRCU primitive. Use ssp->lock, which -is-
398 * compile-time initialized, to resolve races involving multiple
399 * CPUs trying to garner first-use privileges.
401 static void check_init_srcu_struct(struct srcu_struct *ssp)
405 /* The smp_load_acquire() pairs with the smp_store_release(). */
406 if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq_needed))) /*^^^*/
407 return; /* Already initialized. */
408 spin_lock_irqsave_rcu_node(ssp->srcu_sup, flags);
409 if (!rcu_seq_state(ssp->srcu_sup->srcu_gp_seq_needed)) {
410 spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
413 init_srcu_struct_fields(ssp, true);
414 spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
418 * Returns approximate total of the readers' ->srcu_lock_count[] values
419 * for the rank of per-CPU counters specified by idx.
421 static unsigned long srcu_readers_lock_idx(struct srcu_struct *ssp, int idx)
424 unsigned long sum = 0;
426 for_each_possible_cpu(cpu) {
427 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
429 sum += atomic_long_read(&cpuc->srcu_lock_count[idx]);
435 * Returns approximate total of the readers' ->srcu_unlock_count[] values
436 * for the rank of per-CPU counters specified by idx.
438 static unsigned long srcu_readers_unlock_idx(struct srcu_struct *ssp, int idx)
441 unsigned long mask = 0;
442 unsigned long sum = 0;
444 for_each_possible_cpu(cpu) {
445 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
447 sum += atomic_long_read(&cpuc->srcu_unlock_count[idx]);
448 if (IS_ENABLED(CONFIG_PROVE_RCU))
449 mask = mask | READ_ONCE(cpuc->srcu_nmi_safety);
451 WARN_ONCE(IS_ENABLED(CONFIG_PROVE_RCU) && (mask & (mask >> 1)),
452 "Mixed NMI-safe readers for srcu_struct at %ps.\n", ssp);
457 * Return true if the number of pre-existing readers is determined to
460 static bool srcu_readers_active_idx_check(struct srcu_struct *ssp, int idx)
462 unsigned long unlocks;
464 unlocks = srcu_readers_unlock_idx(ssp, idx);
467 * Make sure that a lock is always counted if the corresponding
468 * unlock is counted. Needs to be a smp_mb() as the read side may
469 * contain a read from a variable that is written to before the
470 * synchronize_srcu() in the write side. In this case smp_mb()s
471 * A and B act like the store buffering pattern.
473 * This smp_mb() also pairs with smp_mb() C to prevent accesses
474 * after the synchronize_srcu() from being executed before the
480 * If the locks are the same as the unlocks, then there must have
481 * been no readers on this index at some point in this function.
482 * But there might be more readers, as a task might have read
483 * the current ->srcu_idx but not yet have incremented its CPU's
484 * ->srcu_lock_count[idx] counter. In fact, it is possible
485 * that most of the tasks have been preempted between fetching
486 * ->srcu_idx and incrementing ->srcu_lock_count[idx]. And there
487 * could be almost (ULONG_MAX / sizeof(struct task_struct)) tasks
488 * in a system whose address space was fully populated with memory.
489 * Call this quantity Nt.
491 * So suppose that the updater is preempted at this point in the
492 * code for a long time. That now-preempted updater has already
493 * flipped ->srcu_idx (possibly during the preceding grace period),
494 * done an smp_mb() (again, possibly during the preceding grace
495 * period), and summed up the ->srcu_unlock_count[idx] counters.
496 * How many times can a given one of the aforementioned Nt tasks
497 * increment the old ->srcu_idx value's ->srcu_lock_count[idx]
498 * counter, in the absence of nesting?
500 * It can clearly do so once, given that it has already fetched
501 * the old value of ->srcu_idx and is just about to use that value
502 * to index its increment of ->srcu_lock_count[idx]. But as soon as
503 * it leaves that SRCU read-side critical section, it will increment
504 * ->srcu_unlock_count[idx], which must follow the updater's above
505 * read from that same value. Thus, as soon the reading task does
506 * an smp_mb() and a later fetch from ->srcu_idx, that task will be
507 * guaranteed to get the new index. Except that the increment of
508 * ->srcu_unlock_count[idx] in __srcu_read_unlock() is after the
509 * smp_mb(), and the fetch from ->srcu_idx in __srcu_read_lock()
510 * is before the smp_mb(). Thus, that task might not see the new
511 * value of ->srcu_idx until the -second- __srcu_read_lock(),
512 * which in turn means that this task might well increment
513 * ->srcu_lock_count[idx] for the old value of ->srcu_idx twice,
516 * However, it is important to note that a given smp_mb() takes
517 * effect not just for the task executing it, but also for any
518 * later task running on that same CPU.
520 * That is, there can be almost Nt + Nc further increments of
521 * ->srcu_lock_count[idx] for the old index, where Nc is the number
522 * of CPUs. But this is OK because the size of the task_struct
523 * structure limits the value of Nt and current systems limit Nc
526 * OK, but what about nesting? This does impose a limit on
527 * nesting of half of the size of the task_struct structure
528 * (measured in bytes), which should be sufficient. A late 2022
529 * TREE01 rcutorture run reported this size to be no less than
530 * 9408 bytes, allowing up to 4704 levels of nesting, which is
531 * comfortably beyond excessive. Especially on 64-bit systems,
532 * which are unlikely to be configured with an address space fully
533 * populated with memory, at least not anytime soon.
535 return srcu_readers_lock_idx(ssp, idx) == unlocks;
539 * srcu_readers_active - returns true if there are readers. and false
541 * @ssp: which srcu_struct to count active readers (holding srcu_read_lock).
543 * Note that this is not an atomic primitive, and can therefore suffer
544 * severe errors when invoked on an active srcu_struct. That said, it
545 * can be useful as an error check at cleanup time.
547 static bool srcu_readers_active(struct srcu_struct *ssp)
550 unsigned long sum = 0;
552 for_each_possible_cpu(cpu) {
553 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
555 sum += atomic_long_read(&cpuc->srcu_lock_count[0]);
556 sum += atomic_long_read(&cpuc->srcu_lock_count[1]);
557 sum -= atomic_long_read(&cpuc->srcu_unlock_count[0]);
558 sum -= atomic_long_read(&cpuc->srcu_unlock_count[1]);
564 * We use an adaptive strategy for synchronize_srcu() and especially for
565 * synchronize_srcu_expedited(). We spin for a fixed time period
566 * (defined below, boot time configurable) to allow SRCU readers to exit
567 * their read-side critical sections. If there are still some readers
568 * after one jiffy, we repeatedly block for one jiffy time periods.
569 * The blocking time is increased as the grace-period age increases,
570 * with max blocking time capped at 10 jiffies.
572 #define SRCU_DEFAULT_RETRY_CHECK_DELAY 5
574 static ulong srcu_retry_check_delay = SRCU_DEFAULT_RETRY_CHECK_DELAY;
575 module_param(srcu_retry_check_delay, ulong, 0444);
577 #define SRCU_INTERVAL 1 // Base delay if no expedited GPs pending.
578 #define SRCU_MAX_INTERVAL 10 // Maximum incremental delay from slow readers.
580 #define SRCU_DEFAULT_MAX_NODELAY_PHASE_LO 3UL // Lowmark on default per-GP-phase
581 // no-delay instances.
582 #define SRCU_DEFAULT_MAX_NODELAY_PHASE_HI 1000UL // Highmark on default per-GP-phase
583 // no-delay instances.
585 #define SRCU_UL_CLAMP_LO(val, low) ((val) > (low) ? (val) : (low))
586 #define SRCU_UL_CLAMP_HI(val, high) ((val) < (high) ? (val) : (high))
587 #define SRCU_UL_CLAMP(val, low, high) SRCU_UL_CLAMP_HI(SRCU_UL_CLAMP_LO((val), (low)), (high))
588 // per-GP-phase no-delay instances adjusted to allow non-sleeping poll upto
589 // one jiffies time duration. Mult by 2 is done to factor in the srcu_get_delay()
590 // called from process_srcu().
591 #define SRCU_DEFAULT_MAX_NODELAY_PHASE_ADJUSTED \
592 (2UL * USEC_PER_SEC / HZ / SRCU_DEFAULT_RETRY_CHECK_DELAY)
594 // Maximum per-GP-phase consecutive no-delay instances.
595 #define SRCU_DEFAULT_MAX_NODELAY_PHASE \
596 SRCU_UL_CLAMP(SRCU_DEFAULT_MAX_NODELAY_PHASE_ADJUSTED, \
597 SRCU_DEFAULT_MAX_NODELAY_PHASE_LO, \
598 SRCU_DEFAULT_MAX_NODELAY_PHASE_HI)
600 static ulong srcu_max_nodelay_phase = SRCU_DEFAULT_MAX_NODELAY_PHASE;
601 module_param(srcu_max_nodelay_phase, ulong, 0444);
603 // Maximum consecutive no-delay instances.
604 #define SRCU_DEFAULT_MAX_NODELAY (SRCU_DEFAULT_MAX_NODELAY_PHASE > 100 ? \
605 SRCU_DEFAULT_MAX_NODELAY_PHASE : 100)
607 static ulong srcu_max_nodelay = SRCU_DEFAULT_MAX_NODELAY;
608 module_param(srcu_max_nodelay, ulong, 0444);
611 * Return grace-period delay, zero if there are expedited grace
612 * periods pending, SRCU_INTERVAL otherwise.
614 static unsigned long srcu_get_delay(struct srcu_struct *ssp)
616 unsigned long gpstart;
618 unsigned long jbase = SRCU_INTERVAL;
619 struct srcu_usage *sup = ssp->srcu_sup;
621 if (ULONG_CMP_LT(READ_ONCE(sup->srcu_gp_seq), READ_ONCE(sup->srcu_gp_seq_needed_exp)))
623 if (rcu_seq_state(READ_ONCE(sup->srcu_gp_seq))) {
625 gpstart = READ_ONCE(sup->srcu_gp_start);
626 if (time_after(j, gpstart))
627 jbase += j - gpstart;
629 WRITE_ONCE(sup->srcu_n_exp_nodelay, READ_ONCE(sup->srcu_n_exp_nodelay) + 1);
630 if (READ_ONCE(sup->srcu_n_exp_nodelay) > srcu_max_nodelay_phase)
634 return jbase > SRCU_MAX_INTERVAL ? SRCU_MAX_INTERVAL : jbase;
638 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
639 * @ssp: structure to clean up.
641 * Must invoke this after you are finished using a given srcu_struct that
642 * was initialized via init_srcu_struct(), else you leak memory.
644 void cleanup_srcu_struct(struct srcu_struct *ssp)
647 struct srcu_usage *sup = ssp->srcu_sup;
649 if (WARN_ON(!srcu_get_delay(ssp)))
650 return; /* Just leak it! */
651 if (WARN_ON(srcu_readers_active(ssp)))
652 return; /* Just leak it! */
653 flush_delayed_work(&sup->work);
654 for_each_possible_cpu(cpu) {
655 struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
657 del_timer_sync(&sdp->delay_work);
658 flush_work(&sdp->work);
659 if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist)))
660 return; /* Forgot srcu_barrier(), so just leak it! */
662 if (WARN_ON(rcu_seq_state(READ_ONCE(sup->srcu_gp_seq)) != SRCU_STATE_IDLE) ||
663 WARN_ON(rcu_seq_current(&sup->srcu_gp_seq) != sup->srcu_gp_seq_needed) ||
664 WARN_ON(srcu_readers_active(ssp))) {
665 pr_info("%s: Active srcu_struct %p read state: %d gp state: %lu/%lu\n",
666 __func__, ssp, rcu_seq_state(READ_ONCE(sup->srcu_gp_seq)),
667 rcu_seq_current(&sup->srcu_gp_seq), sup->srcu_gp_seq_needed);
668 return; /* Caller forgot to stop doing call_srcu()? */
672 sup->srcu_size_state = SRCU_SIZE_SMALL;
673 if (!sup->sda_is_static) {
674 free_percpu(ssp->sda);
677 ssp->srcu_sup = NULL;
680 EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
682 #ifdef CONFIG_PROVE_RCU
684 * Check for consistent NMI safety.
686 void srcu_check_nmi_safety(struct srcu_struct *ssp, bool nmi_safe)
688 int nmi_safe_mask = 1 << nmi_safe;
689 int old_nmi_safe_mask;
690 struct srcu_data *sdp;
692 /* NMI-unsafe use in NMI is a bad sign */
693 WARN_ON_ONCE(!nmi_safe && in_nmi());
694 sdp = raw_cpu_ptr(ssp->sda);
695 old_nmi_safe_mask = READ_ONCE(sdp->srcu_nmi_safety);
696 if (!old_nmi_safe_mask) {
697 WRITE_ONCE(sdp->srcu_nmi_safety, nmi_safe_mask);
700 WARN_ONCE(old_nmi_safe_mask != nmi_safe_mask, "CPU %d old state %d new state %d\n", sdp->cpu, old_nmi_safe_mask, nmi_safe_mask);
702 EXPORT_SYMBOL_GPL(srcu_check_nmi_safety);
703 #endif /* CONFIG_PROVE_RCU */
706 * Counts the new reader in the appropriate per-CPU element of the
708 * Returns an index that must be passed to the matching srcu_read_unlock().
710 int __srcu_read_lock(struct srcu_struct *ssp)
714 idx = READ_ONCE(ssp->srcu_idx) & 0x1;
715 this_cpu_inc(ssp->sda->srcu_lock_count[idx].counter);
716 smp_mb(); /* B */ /* Avoid leaking the critical section. */
719 EXPORT_SYMBOL_GPL(__srcu_read_lock);
722 * Removes the count for the old reader from the appropriate per-CPU
723 * element of the srcu_struct. Note that this may well be a different
724 * CPU than that which was incremented by the corresponding srcu_read_lock().
726 void __srcu_read_unlock(struct srcu_struct *ssp, int idx)
728 smp_mb(); /* C */ /* Avoid leaking the critical section. */
729 this_cpu_inc(ssp->sda->srcu_unlock_count[idx].counter);
731 EXPORT_SYMBOL_GPL(__srcu_read_unlock);
733 #ifdef CONFIG_NEED_SRCU_NMI_SAFE
736 * Counts the new reader in the appropriate per-CPU element of the
737 * srcu_struct, but in an NMI-safe manner using RMW atomics.
738 * Returns an index that must be passed to the matching srcu_read_unlock().
740 int __srcu_read_lock_nmisafe(struct srcu_struct *ssp)
743 struct srcu_data *sdp = raw_cpu_ptr(ssp->sda);
745 idx = READ_ONCE(ssp->srcu_idx) & 0x1;
746 atomic_long_inc(&sdp->srcu_lock_count[idx]);
747 smp_mb__after_atomic(); /* B */ /* Avoid leaking the critical section. */
750 EXPORT_SYMBOL_GPL(__srcu_read_lock_nmisafe);
753 * Removes the count for the old reader from the appropriate per-CPU
754 * element of the srcu_struct. Note that this may well be a different
755 * CPU than that which was incremented by the corresponding srcu_read_lock().
757 void __srcu_read_unlock_nmisafe(struct srcu_struct *ssp, int idx)
759 struct srcu_data *sdp = raw_cpu_ptr(ssp->sda);
761 smp_mb__before_atomic(); /* C */ /* Avoid leaking the critical section. */
762 atomic_long_inc(&sdp->srcu_unlock_count[idx]);
764 EXPORT_SYMBOL_GPL(__srcu_read_unlock_nmisafe);
766 #endif // CONFIG_NEED_SRCU_NMI_SAFE
769 * Start an SRCU grace period.
771 static void srcu_gp_start(struct srcu_struct *ssp)
773 struct srcu_data *sdp;
776 if (smp_load_acquire(&ssp->srcu_sup->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
777 sdp = per_cpu_ptr(ssp->sda, get_boot_cpu_id());
779 sdp = this_cpu_ptr(ssp->sda);
780 lockdep_assert_held(&ACCESS_PRIVATE(ssp->srcu_sup, lock));
781 WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_sup->srcu_gp_seq, ssp->srcu_sup->srcu_gp_seq_needed));
782 spin_lock_rcu_node(sdp); /* Interrupts already disabled. */
783 rcu_segcblist_advance(&sdp->srcu_cblist,
784 rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq));
785 WARN_ON_ONCE(!rcu_segcblist_segempty(&sdp->srcu_cblist, RCU_NEXT_TAIL));
786 spin_unlock_rcu_node(sdp); /* Interrupts remain disabled. */
787 WRITE_ONCE(ssp->srcu_sup->srcu_gp_start, jiffies);
788 WRITE_ONCE(ssp->srcu_sup->srcu_n_exp_nodelay, 0);
789 smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
790 rcu_seq_start(&ssp->srcu_sup->srcu_gp_seq);
791 state = rcu_seq_state(ssp->srcu_sup->srcu_gp_seq);
792 WARN_ON_ONCE(state != SRCU_STATE_SCAN1);
796 static void srcu_delay_timer(struct timer_list *t)
798 struct srcu_data *sdp = container_of(t, struct srcu_data, delay_work);
800 queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
803 static void srcu_queue_delayed_work_on(struct srcu_data *sdp,
807 queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
811 timer_reduce(&sdp->delay_work, jiffies + delay);
815 * Schedule callback invocation for the specified srcu_data structure,
816 * if possible, on the corresponding CPU.
818 static void srcu_schedule_cbs_sdp(struct srcu_data *sdp, unsigned long delay)
820 srcu_queue_delayed_work_on(sdp, delay);
824 * Schedule callback invocation for all srcu_data structures associated
825 * with the specified srcu_node structure that have callbacks for the
826 * just-completed grace period, the one corresponding to idx. If possible,
827 * schedule this invocation on the corresponding CPUs.
829 static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp,
830 unsigned long mask, unsigned long delay)
834 for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
835 if (!(mask & (1UL << (cpu - snp->grplo))))
837 srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
842 * Note the end of an SRCU grace period. Initiates callback invocation
843 * and starts a new grace period if needed.
845 * The ->srcu_cb_mutex acquisition does not protect any data, but
846 * instead prevents more than one grace period from starting while we
847 * are initiating callback invocation. This allows the ->srcu_have_cbs[]
848 * array to have a finite number of elements.
850 static void srcu_gp_end(struct srcu_struct *ssp)
852 unsigned long cbdelay = 1;
860 struct srcu_data *sdp;
862 struct srcu_node *snp;
864 struct srcu_usage *sup = ssp->srcu_sup;
866 /* Prevent more than one additional grace period. */
867 mutex_lock(&sup->srcu_cb_mutex);
869 /* End the current grace period. */
870 spin_lock_irq_rcu_node(sup);
871 idx = rcu_seq_state(sup->srcu_gp_seq);
872 WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
873 if (ULONG_CMP_LT(READ_ONCE(sup->srcu_gp_seq), READ_ONCE(sup->srcu_gp_seq_needed_exp)))
876 WRITE_ONCE(sup->srcu_last_gp_end, ktime_get_mono_fast_ns());
877 rcu_seq_end(&sup->srcu_gp_seq);
878 gpseq = rcu_seq_current(&sup->srcu_gp_seq);
879 if (ULONG_CMP_LT(sup->srcu_gp_seq_needed_exp, gpseq))
880 WRITE_ONCE(sup->srcu_gp_seq_needed_exp, gpseq);
881 spin_unlock_irq_rcu_node(sup);
882 mutex_unlock(&sup->srcu_gp_mutex);
883 /* A new grace period can start at this point. But only one. */
885 /* Initiate callback invocation as needed. */
886 ss_state = smp_load_acquire(&sup->srcu_size_state);
887 if (ss_state < SRCU_SIZE_WAIT_BARRIER) {
888 srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, get_boot_cpu_id()),
891 idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs);
892 srcu_for_each_node_breadth_first(ssp, snp) {
893 spin_lock_irq_rcu_node(snp);
895 last_lvl = snp >= sup->level[rcu_num_lvls - 1];
897 cbs = ss_state < SRCU_SIZE_BIG || snp->srcu_have_cbs[idx] == gpseq;
898 snp->srcu_have_cbs[idx] = gpseq;
899 rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1);
900 sgsne = snp->srcu_gp_seq_needed_exp;
901 if (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, gpseq))
902 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, gpseq);
903 if (ss_state < SRCU_SIZE_BIG)
906 mask = snp->srcu_data_have_cbs[idx];
907 snp->srcu_data_have_cbs[idx] = 0;
908 spin_unlock_irq_rcu_node(snp);
910 srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay);
914 /* Occasionally prevent srcu_data counter wrap. */
915 if (!(gpseq & counter_wrap_check))
916 for_each_possible_cpu(cpu) {
917 sdp = per_cpu_ptr(ssp->sda, cpu);
918 spin_lock_irqsave_rcu_node(sdp, flags);
919 if (ULONG_CMP_GE(gpseq, sdp->srcu_gp_seq_needed + 100))
920 sdp->srcu_gp_seq_needed = gpseq;
921 if (ULONG_CMP_GE(gpseq, sdp->srcu_gp_seq_needed_exp + 100))
922 sdp->srcu_gp_seq_needed_exp = gpseq;
923 spin_unlock_irqrestore_rcu_node(sdp, flags);
926 /* Callback initiation done, allow grace periods after next. */
927 mutex_unlock(&sup->srcu_cb_mutex);
929 /* Start a new grace period if needed. */
930 spin_lock_irq_rcu_node(sup);
931 gpseq = rcu_seq_current(&sup->srcu_gp_seq);
932 if (!rcu_seq_state(gpseq) &&
933 ULONG_CMP_LT(gpseq, sup->srcu_gp_seq_needed)) {
935 spin_unlock_irq_rcu_node(sup);
936 srcu_reschedule(ssp, 0);
938 spin_unlock_irq_rcu_node(sup);
941 /* Transition to big if needed. */
942 if (ss_state != SRCU_SIZE_SMALL && ss_state != SRCU_SIZE_BIG) {
943 if (ss_state == SRCU_SIZE_ALLOC)
944 init_srcu_struct_nodes(ssp, GFP_KERNEL);
946 smp_store_release(&sup->srcu_size_state, ss_state + 1);
951 * Funnel-locking scheme to scalably mediate many concurrent expedited
952 * grace-period requests. This function is invoked for the first known
953 * expedited request for a grace period that has already been requested,
954 * but without expediting. To start a completely new grace period,
955 * whether expedited or not, use srcu_funnel_gp_start() instead.
957 static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp,
964 for (; snp != NULL; snp = snp->srcu_parent) {
965 sgsne = READ_ONCE(snp->srcu_gp_seq_needed_exp);
966 if (WARN_ON_ONCE(rcu_seq_done(&ssp->srcu_sup->srcu_gp_seq, s)) ||
967 (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)))
969 spin_lock_irqsave_rcu_node(snp, flags);
970 sgsne = snp->srcu_gp_seq_needed_exp;
971 if (!srcu_invl_snp_seq(sgsne) && ULONG_CMP_GE(sgsne, s)) {
972 spin_unlock_irqrestore_rcu_node(snp, flags);
975 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
976 spin_unlock_irqrestore_rcu_node(snp, flags);
978 spin_lock_irqsave_ssp_contention(ssp, &flags);
979 if (ULONG_CMP_LT(ssp->srcu_sup->srcu_gp_seq_needed_exp, s))
980 WRITE_ONCE(ssp->srcu_sup->srcu_gp_seq_needed_exp, s);
981 spin_unlock_irqrestore_rcu_node(ssp->srcu_sup, flags);
985 * Funnel-locking scheme to scalably mediate many concurrent grace-period
986 * requests. The winner has to do the work of actually starting grace
987 * period s. Losers must either ensure that their desired grace-period
988 * number is recorded on at least their leaf srcu_node structure, or they
989 * must take steps to invoke their own callbacks.
991 * Note that this function also does the work of srcu_funnel_exp_start(),
992 * in some cases by directly invoking it.
994 * The srcu read lock should be hold around this function. And s is a seq snap
995 * after holding that lock.
997 static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
998 unsigned long s, bool do_norm)
1000 unsigned long flags;
1001 int idx = rcu_seq_ctr(s) % ARRAY_SIZE(sdp->mynode->srcu_have_cbs);
1002 unsigned long sgsne;
1003 struct srcu_node *snp;
1004 struct srcu_node *snp_leaf;
1005 unsigned long snp_seq;
1006 struct srcu_usage *sup = ssp->srcu_sup;
1008 /* Ensure that snp node tree is fully initialized before traversing it */
1009 if (smp_load_acquire(&sup->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
1012 snp_leaf = sdp->mynode;
1015 /* Each pass through the loop does one level of the srcu_node tree. */
1016 for (snp = snp_leaf; snp != NULL; snp = snp->srcu_parent) {
1017 if (WARN_ON_ONCE(rcu_seq_done(&sup->srcu_gp_seq, s)) && snp != snp_leaf)
1018 return; /* GP already done and CBs recorded. */
1019 spin_lock_irqsave_rcu_node(snp, flags);
1020 snp_seq = snp->srcu_have_cbs[idx];
1021 if (!srcu_invl_snp_seq(snp_seq) && ULONG_CMP_GE(snp_seq, s)) {
1022 if (snp == snp_leaf && snp_seq == s)
1023 snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
1024 spin_unlock_irqrestore_rcu_node(snp, flags);
1025 if (snp == snp_leaf && snp_seq != s) {
1026 srcu_schedule_cbs_sdp(sdp, do_norm ? SRCU_INTERVAL : 0);
1030 srcu_funnel_exp_start(ssp, snp, s);
1033 snp->srcu_have_cbs[idx] = s;
1034 if (snp == snp_leaf)
1035 snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
1036 sgsne = snp->srcu_gp_seq_needed_exp;
1037 if (!do_norm && (srcu_invl_snp_seq(sgsne) || ULONG_CMP_LT(sgsne, s)))
1038 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
1039 spin_unlock_irqrestore_rcu_node(snp, flags);
1042 /* Top of tree, must ensure the grace period will be started. */
1043 spin_lock_irqsave_ssp_contention(ssp, &flags);
1044 if (ULONG_CMP_LT(sup->srcu_gp_seq_needed, s)) {
1046 * Record need for grace period s. Pair with load
1047 * acquire setting up for initialization.
1049 smp_store_release(&sup->srcu_gp_seq_needed, s); /*^^^*/
1051 if (!do_norm && ULONG_CMP_LT(sup->srcu_gp_seq_needed_exp, s))
1052 WRITE_ONCE(sup->srcu_gp_seq_needed_exp, s);
1054 /* If grace period not already in progress, start it. */
1055 if (!WARN_ON_ONCE(rcu_seq_done(&sup->srcu_gp_seq, s)) &&
1056 rcu_seq_state(sup->srcu_gp_seq) == SRCU_STATE_IDLE) {
1057 WARN_ON_ONCE(ULONG_CMP_GE(sup->srcu_gp_seq, sup->srcu_gp_seq_needed));
1060 // And how can that list_add() in the "else" clause
1061 // possibly be safe for concurrent execution? Well,
1062 // it isn't. And it does not have to be. After all, it
1063 // can only be executed during early boot when there is only
1064 // the one boot CPU running with interrupts still disabled.
1065 if (likely(srcu_init_done))
1066 queue_delayed_work(rcu_gp_wq, &sup->work,
1067 !!srcu_get_delay(ssp));
1068 else if (list_empty(&sup->work.work.entry))
1069 list_add(&sup->work.work.entry, &srcu_boot_list);
1071 spin_unlock_irqrestore_rcu_node(sup, flags);
1075 * Wait until all readers counted by array index idx complete, but
1076 * loop an additional time if there is an expedited grace period pending.
1077 * The caller must ensure that ->srcu_idx is not changed while checking.
1079 static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount)
1081 unsigned long curdelay;
1083 curdelay = !srcu_get_delay(ssp);
1086 if (srcu_readers_active_idx_check(ssp, idx))
1088 if ((--trycount + curdelay) <= 0)
1090 udelay(srcu_retry_check_delay);
1095 * Increment the ->srcu_idx counter so that future SRCU readers will
1096 * use the other rank of the ->srcu_(un)lock_count[] arrays. This allows
1097 * us to wait for pre-existing readers in a starvation-free manner.
1099 static void srcu_flip(struct srcu_struct *ssp)
1102 * Because the flip of ->srcu_idx is executed only if the
1103 * preceding call to srcu_readers_active_idx_check() found that
1104 * the ->srcu_unlock_count[] and ->srcu_lock_count[] sums matched
1105 * and because that summing uses atomic_long_read(), there is
1106 * ordering due to a control dependency between that summing and
1107 * the WRITE_ONCE() in this call to srcu_flip(). This ordering
1108 * ensures that if this updater saw a given reader's increment from
1109 * __srcu_read_lock(), that reader was using a value of ->srcu_idx
1110 * from before the previous call to srcu_flip(), which should be
1111 * quite rare. This ordering thus helps forward progress because
1112 * the grace period could otherwise be delayed by additional
1113 * calls to __srcu_read_lock() using that old (soon to be new)
1114 * value of ->srcu_idx.
1116 * This sum-equality check and ordering also ensures that if
1117 * a given call to __srcu_read_lock() uses the new value of
1118 * ->srcu_idx, this updater's earlier scans cannot have seen
1119 * that reader's increments, which is all to the good, because
1120 * this grace period need not wait on that reader. After all,
1121 * if those earlier scans had seen that reader, there would have
1122 * been a sum mismatch and this code would not be reached.
1124 * This means that the following smp_mb() is redundant, but
1125 * it stays until either (1) Compilers learn about this sort of
1126 * control dependency or (2) Some production workload running on
1127 * a production system is unduly delayed by this slowpath smp_mb().
1129 smp_mb(); /* E */ /* Pairs with B and C. */
1131 WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1); // Flip the counter.
1134 * Ensure that if the updater misses an __srcu_read_unlock()
1135 * increment, that task's __srcu_read_lock() following its next
1136 * __srcu_read_lock() or __srcu_read_unlock() will see the above
1137 * counter update. Note that both this memory barrier and the
1138 * one in srcu_readers_active_idx_check() provide the guarantee
1139 * for __srcu_read_lock().
1141 smp_mb(); /* D */ /* Pairs with C. */
1145 * If SRCU is likely idle, return true, otherwise return false.
1147 * Note that it is OK for several current from-idle requests for a new
1148 * grace period from idle to specify expediting because they will all end
1149 * up requesting the same grace period anyhow. So no loss.
1151 * Note also that if any CPU (including the current one) is still invoking
1152 * callbacks, this function will nevertheless say "idle". This is not
1153 * ideal, but the overhead of checking all CPUs' callback lists is even
1154 * less ideal, especially on large systems. Furthermore, the wakeup
1155 * can happen before the callback is fully removed, so we have no choice
1156 * but to accept this type of error.
1158 * This function is also subject to counter-wrap errors, but let's face
1159 * it, if this function was preempted for enough time for the counters
1160 * to wrap, it really doesn't matter whether or not we expedite the grace
1161 * period. The extra overhead of a needlessly expedited grace period is
1162 * negligible when amortized over that time period, and the extra latency
1163 * of a needlessly non-expedited grace period is similarly negligible.
1165 static bool srcu_might_be_idle(struct srcu_struct *ssp)
1167 unsigned long curseq;
1168 unsigned long flags;
1169 struct srcu_data *sdp;
1171 unsigned long tlast;
1173 check_init_srcu_struct(ssp);
1174 /* If the local srcu_data structure has callbacks, not idle. */
1175 sdp = raw_cpu_ptr(ssp->sda);
1176 spin_lock_irqsave_rcu_node(sdp, flags);
1177 if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) {
1178 spin_unlock_irqrestore_rcu_node(sdp, flags);
1179 return false; /* Callbacks already present, so not idle. */
1181 spin_unlock_irqrestore_rcu_node(sdp, flags);
1184 * No local callbacks, so probabilistically probe global state.
1185 * Exact information would require acquiring locks, which would
1186 * kill scalability, hence the probabilistic nature of the probe.
1189 /* First, see if enough time has passed since the last GP. */
1190 t = ktime_get_mono_fast_ns();
1191 tlast = READ_ONCE(ssp->srcu_sup->srcu_last_gp_end);
1192 if (exp_holdoff == 0 ||
1193 time_in_range_open(t, tlast, tlast + exp_holdoff))
1194 return false; /* Too soon after last GP. */
1196 /* Next, check for probable idleness. */
1197 curseq = rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq);
1198 smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
1199 if (ULONG_CMP_LT(curseq, READ_ONCE(ssp->srcu_sup->srcu_gp_seq_needed)))
1200 return false; /* Grace period in progress, so not idle. */
1201 smp_mb(); /* Order ->srcu_gp_seq with prior access. */
1202 if (curseq != rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq))
1203 return false; /* GP # changed, so not idle. */
1204 return true; /* With reasonable probability, idle! */
1208 * SRCU callback function to leak a callback.
1210 static void srcu_leak_callback(struct rcu_head *rhp)
1215 * Start an SRCU grace period, and also queue the callback if non-NULL.
1217 static unsigned long srcu_gp_start_if_needed(struct srcu_struct *ssp,
1218 struct rcu_head *rhp, bool do_norm)
1220 unsigned long flags;
1222 bool needexp = false;
1223 bool needgp = false;
1225 struct srcu_data *sdp;
1226 struct srcu_node *sdp_mynode;
1229 check_init_srcu_struct(ssp);
1231 * While starting a new grace period, make sure we are in an
1232 * SRCU read-side critical section so that the grace-period
1233 * sequence number cannot wrap around in the meantime.
1235 idx = __srcu_read_lock_nmisafe(ssp);
1236 ss_state = smp_load_acquire(&ssp->srcu_sup->srcu_size_state);
1237 if (ss_state < SRCU_SIZE_WAIT_CALL)
1238 sdp = per_cpu_ptr(ssp->sda, get_boot_cpu_id());
1240 sdp = raw_cpu_ptr(ssp->sda);
1241 spin_lock_irqsave_sdp_contention(sdp, &flags);
1243 rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp);
1245 * The snapshot for acceleration must be taken _before_ the read of the
1246 * current gp sequence used for advancing, otherwise advancing may fail
1247 * and acceleration may then fail too.
1249 * This could happen if:
1251 * 1) The RCU_WAIT_TAIL segment has callbacks (gp_num = X + 4) and the
1252 * RCU_NEXT_READY_TAIL also has callbacks (gp_num = X + 8).
1254 * 2) The grace period for RCU_WAIT_TAIL is seen as started but not
1255 * completed so rcu_seq_current() returns X + SRCU_STATE_SCAN1.
1257 * 3) This value is passed to rcu_segcblist_advance() which can't move
1258 * any segment forward and fails.
1260 * 4) srcu_gp_start_if_needed() still proceeds with callback acceleration.
1261 * But then the call to rcu_seq_snap() observes the grace period for the
1262 * RCU_WAIT_TAIL segment as completed and the subsequent one for the
1263 * RCU_NEXT_READY_TAIL segment as started (ie: X + 4 + SRCU_STATE_SCAN1)
1264 * so it returns a snapshot of the next grace period, which is X + 12.
1266 * 5) The value of X + 12 is passed to rcu_segcblist_accelerate() but the
1267 * freshly enqueued callback in RCU_NEXT_TAIL can't move to
1268 * RCU_NEXT_READY_TAIL which already has callbacks for a previous grace
1269 * period (gp_num = X + 8). So acceleration fails.
1271 s = rcu_seq_snap(&ssp->srcu_sup->srcu_gp_seq);
1272 rcu_segcblist_advance(&sdp->srcu_cblist,
1273 rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq));
1274 WARN_ON_ONCE(!rcu_segcblist_accelerate(&sdp->srcu_cblist, s) && rhp);
1275 if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) {
1276 sdp->srcu_gp_seq_needed = s;
1279 if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) {
1280 sdp->srcu_gp_seq_needed_exp = s;
1283 spin_unlock_irqrestore_rcu_node(sdp, flags);
1285 /* Ensure that snp node tree is fully initialized before traversing it */
1286 if (ss_state < SRCU_SIZE_WAIT_BARRIER)
1289 sdp_mynode = sdp->mynode;
1292 srcu_funnel_gp_start(ssp, sdp, s, do_norm);
1294 srcu_funnel_exp_start(ssp, sdp_mynode, s);
1295 __srcu_read_unlock_nmisafe(ssp, idx);
1300 * Enqueue an SRCU callback on the srcu_data structure associated with
1301 * the current CPU and the specified srcu_struct structure, initiating
1302 * grace-period processing if it is not already running.
1304 * Note that all CPUs must agree that the grace period extended beyond
1305 * all pre-existing SRCU read-side critical section. On systems with
1306 * more than one CPU, this means that when "func()" is invoked, each CPU
1307 * is guaranteed to have executed a full memory barrier since the end of
1308 * its last corresponding SRCU read-side critical section whose beginning
1309 * preceded the call to call_srcu(). It also means that each CPU executing
1310 * an SRCU read-side critical section that continues beyond the start of
1311 * "func()" must have executed a memory barrier after the call_srcu()
1312 * but before the beginning of that SRCU read-side critical section.
1313 * Note that these guarantees include CPUs that are offline, idle, or
1314 * executing in user mode, as well as CPUs that are executing in the kernel.
1316 * Furthermore, if CPU A invoked call_srcu() and CPU B invoked the
1317 * resulting SRCU callback function "func()", then both CPU A and CPU
1318 * B are guaranteed to execute a full memory barrier during the time
1319 * interval between the call to call_srcu() and the invocation of "func()".
1320 * This guarantee applies even if CPU A and CPU B are the same CPU (but
1321 * again only if the system has more than one CPU).
1323 * Of course, these guarantees apply only for invocations of call_srcu(),
1324 * srcu_read_lock(), and srcu_read_unlock() that are all passed the same
1325 * srcu_struct structure.
1327 static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
1328 rcu_callback_t func, bool do_norm)
1330 if (debug_rcu_head_queue(rhp)) {
1331 /* Probable double call_srcu(), so leak the callback. */
1332 WRITE_ONCE(rhp->func, srcu_leak_callback);
1333 WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n");
1337 (void)srcu_gp_start_if_needed(ssp, rhp, do_norm);
1341 * call_srcu() - Queue a callback for invocation after an SRCU grace period
1342 * @ssp: srcu_struct in queue the callback
1343 * @rhp: structure to be used for queueing the SRCU callback.
1344 * @func: function to be invoked after the SRCU grace period
1346 * The callback function will be invoked some time after a full SRCU
1347 * grace period elapses, in other words after all pre-existing SRCU
1348 * read-side critical sections have completed. However, the callback
1349 * function might well execute concurrently with other SRCU read-side
1350 * critical sections that started after call_srcu() was invoked. SRCU
1351 * read-side critical sections are delimited by srcu_read_lock() and
1352 * srcu_read_unlock(), and may be nested.
1354 * The callback will be invoked from process context, but must nevertheless
1355 * be fast and must not block.
1357 void call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
1358 rcu_callback_t func)
1360 __call_srcu(ssp, rhp, func, true);
1362 EXPORT_SYMBOL_GPL(call_srcu);
1365 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
1367 static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
1369 struct rcu_synchronize rcu;
1371 srcu_lock_sync(&ssp->dep_map);
1373 RCU_LOCKDEP_WARN(lockdep_is_held(ssp) ||
1374 lock_is_held(&rcu_bh_lock_map) ||
1375 lock_is_held(&rcu_lock_map) ||
1376 lock_is_held(&rcu_sched_lock_map),
1377 "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
1379 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
1382 check_init_srcu_struct(ssp);
1383 init_completion(&rcu.completion);
1384 init_rcu_head_on_stack(&rcu.head);
1385 __call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm);
1386 wait_for_completion(&rcu.completion);
1387 destroy_rcu_head_on_stack(&rcu.head);
1390 * Make sure that later code is ordered after the SRCU grace
1391 * period. This pairs with the spin_lock_irq_rcu_node()
1392 * in srcu_invoke_callbacks(). Unlike Tree RCU, this is needed
1393 * because the current CPU might have been totally uninvolved with
1394 * (and thus unordered against) that grace period.
1400 * synchronize_srcu_expedited - Brute-force SRCU grace period
1401 * @ssp: srcu_struct with which to synchronize.
1403 * Wait for an SRCU grace period to elapse, but be more aggressive about
1404 * spinning rather than blocking when waiting.
1406 * Note that synchronize_srcu_expedited() has the same deadlock and
1407 * memory-ordering properties as does synchronize_srcu().
1409 void synchronize_srcu_expedited(struct srcu_struct *ssp)
1411 __synchronize_srcu(ssp, rcu_gp_is_normal());
1413 EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
1416 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
1417 * @ssp: srcu_struct with which to synchronize.
1419 * Wait for the count to drain to zero of both indexes. To avoid the
1420 * possible starvation of synchronize_srcu(), it waits for the count of
1421 * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
1422 * and then flip the srcu_idx and wait for the count of the other index.
1424 * Can block; must be called from process context.
1426 * Note that it is illegal to call synchronize_srcu() from the corresponding
1427 * SRCU read-side critical section; doing so will result in deadlock.
1428 * However, it is perfectly legal to call synchronize_srcu() on one
1429 * srcu_struct from some other srcu_struct's read-side critical section,
1430 * as long as the resulting graph of srcu_structs is acyclic.
1432 * There are memory-ordering constraints implied by synchronize_srcu().
1433 * On systems with more than one CPU, when synchronize_srcu() returns,
1434 * each CPU is guaranteed to have executed a full memory barrier since
1435 * the end of its last corresponding SRCU read-side critical section
1436 * whose beginning preceded the call to synchronize_srcu(). In addition,
1437 * each CPU having an SRCU read-side critical section that extends beyond
1438 * the return from synchronize_srcu() is guaranteed to have executed a
1439 * full memory barrier after the beginning of synchronize_srcu() and before
1440 * the beginning of that SRCU read-side critical section. Note that these
1441 * guarantees include CPUs that are offline, idle, or executing in user mode,
1442 * as well as CPUs that are executing in the kernel.
1444 * Furthermore, if CPU A invoked synchronize_srcu(), which returned
1445 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
1446 * to have executed a full memory barrier during the execution of
1447 * synchronize_srcu(). This guarantee applies even if CPU A and CPU B
1448 * are the same CPU, but again only if the system has more than one CPU.
1450 * Of course, these memory-ordering guarantees apply only when
1451 * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
1452 * passed the same srcu_struct structure.
1454 * Implementation of these memory-ordering guarantees is similar to
1455 * that of synchronize_rcu().
1457 * If SRCU is likely idle, expedite the first request. This semantic
1458 * was provided by Classic SRCU, and is relied upon by its users, so TREE
1459 * SRCU must also provide it. Note that detecting idleness is heuristic
1460 * and subject to both false positives and negatives.
1462 void synchronize_srcu(struct srcu_struct *ssp)
1464 if (srcu_might_be_idle(ssp) || rcu_gp_is_expedited())
1465 synchronize_srcu_expedited(ssp);
1467 __synchronize_srcu(ssp, true);
1469 EXPORT_SYMBOL_GPL(synchronize_srcu);
1472 * get_state_synchronize_srcu - Provide an end-of-grace-period cookie
1473 * @ssp: srcu_struct to provide cookie for.
1475 * This function returns a cookie that can be passed to
1476 * poll_state_synchronize_srcu(), which will return true if a full grace
1477 * period has elapsed in the meantime. It is the caller's responsibility
1478 * to make sure that grace period happens, for example, by invoking
1479 * call_srcu() after return from get_state_synchronize_srcu().
1481 unsigned long get_state_synchronize_srcu(struct srcu_struct *ssp)
1483 // Any prior manipulation of SRCU-protected data must happen
1484 // before the load from ->srcu_gp_seq.
1486 return rcu_seq_snap(&ssp->srcu_sup->srcu_gp_seq);
1488 EXPORT_SYMBOL_GPL(get_state_synchronize_srcu);
1491 * start_poll_synchronize_srcu - Provide cookie and start grace period
1492 * @ssp: srcu_struct to provide cookie for.
1494 * This function returns a cookie that can be passed to
1495 * poll_state_synchronize_srcu(), which will return true if a full grace
1496 * period has elapsed in the meantime. Unlike get_state_synchronize_srcu(),
1497 * this function also ensures that any needed SRCU grace period will be
1498 * started. This convenience does come at a cost in terms of CPU overhead.
1500 unsigned long start_poll_synchronize_srcu(struct srcu_struct *ssp)
1502 return srcu_gp_start_if_needed(ssp, NULL, true);
1504 EXPORT_SYMBOL_GPL(start_poll_synchronize_srcu);
1507 * poll_state_synchronize_srcu - Has cookie's grace period ended?
1508 * @ssp: srcu_struct to provide cookie for.
1509 * @cookie: Return value from get_state_synchronize_srcu() or start_poll_synchronize_srcu().
1511 * This function takes the cookie that was returned from either
1512 * get_state_synchronize_srcu() or start_poll_synchronize_srcu(), and
1513 * returns @true if an SRCU grace period elapsed since the time that the
1514 * cookie was created.
1516 * Because cookies are finite in size, wrapping/overflow is possible.
1517 * This is more pronounced on 32-bit systems where cookies are 32 bits,
1518 * where in theory wrapping could happen in about 14 hours assuming
1519 * 25-microsecond expedited SRCU grace periods. However, a more likely
1520 * overflow lower bound is on the order of 24 days in the case of
1521 * one-millisecond SRCU grace periods. Of course, wrapping in a 64-bit
1522 * system requires geologic timespans, as in more than seven million years
1523 * even for expedited SRCU grace periods.
1525 * Wrapping/overflow is much more of an issue for CONFIG_SMP=n systems
1526 * that also have CONFIG_PREEMPTION=n, which selects Tiny SRCU. This uses
1527 * a 16-bit cookie, which rcutorture routinely wraps in a matter of a
1528 * few minutes. If this proves to be a problem, this counter will be
1529 * expanded to the same size as for Tree SRCU.
1531 bool poll_state_synchronize_srcu(struct srcu_struct *ssp, unsigned long cookie)
1533 if (!rcu_seq_done(&ssp->srcu_sup->srcu_gp_seq, cookie))
1535 // Ensure that the end of the SRCU grace period happens before
1536 // any subsequent code that the caller might execute.
1540 EXPORT_SYMBOL_GPL(poll_state_synchronize_srcu);
1543 * Callback function for srcu_barrier() use.
1545 static void srcu_barrier_cb(struct rcu_head *rhp)
1547 struct srcu_data *sdp;
1548 struct srcu_struct *ssp;
1550 sdp = container_of(rhp, struct srcu_data, srcu_barrier_head);
1552 if (atomic_dec_and_test(&ssp->srcu_sup->srcu_barrier_cpu_cnt))
1553 complete(&ssp->srcu_sup->srcu_barrier_completion);
1557 * Enqueue an srcu_barrier() callback on the specified srcu_data
1558 * structure's ->cblist. but only if that ->cblist already has at least one
1559 * callback enqueued. Note that if a CPU already has callbacks enqueue,
1560 * it must have already registered the need for a future grace period,
1561 * so all we need do is enqueue a callback that will use the same grace
1562 * period as the last callback already in the queue.
1564 static void srcu_barrier_one_cpu(struct srcu_struct *ssp, struct srcu_data *sdp)
1566 spin_lock_irq_rcu_node(sdp);
1567 atomic_inc(&ssp->srcu_sup->srcu_barrier_cpu_cnt);
1568 sdp->srcu_barrier_head.func = srcu_barrier_cb;
1569 debug_rcu_head_queue(&sdp->srcu_barrier_head);
1570 if (!rcu_segcblist_entrain(&sdp->srcu_cblist,
1571 &sdp->srcu_barrier_head)) {
1572 debug_rcu_head_unqueue(&sdp->srcu_barrier_head);
1573 atomic_dec(&ssp->srcu_sup->srcu_barrier_cpu_cnt);
1575 spin_unlock_irq_rcu_node(sdp);
1579 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
1580 * @ssp: srcu_struct on which to wait for in-flight callbacks.
1582 void srcu_barrier(struct srcu_struct *ssp)
1586 unsigned long s = rcu_seq_snap(&ssp->srcu_sup->srcu_barrier_seq);
1588 check_init_srcu_struct(ssp);
1589 mutex_lock(&ssp->srcu_sup->srcu_barrier_mutex);
1590 if (rcu_seq_done(&ssp->srcu_sup->srcu_barrier_seq, s)) {
1591 smp_mb(); /* Force ordering following return. */
1592 mutex_unlock(&ssp->srcu_sup->srcu_barrier_mutex);
1593 return; /* Someone else did our work for us. */
1595 rcu_seq_start(&ssp->srcu_sup->srcu_barrier_seq);
1596 init_completion(&ssp->srcu_sup->srcu_barrier_completion);
1598 /* Initial count prevents reaching zero until all CBs are posted. */
1599 atomic_set(&ssp->srcu_sup->srcu_barrier_cpu_cnt, 1);
1601 idx = __srcu_read_lock_nmisafe(ssp);
1602 if (smp_load_acquire(&ssp->srcu_sup->srcu_size_state) < SRCU_SIZE_WAIT_BARRIER)
1603 srcu_barrier_one_cpu(ssp, per_cpu_ptr(ssp->sda, get_boot_cpu_id()));
1605 for_each_possible_cpu(cpu)
1606 srcu_barrier_one_cpu(ssp, per_cpu_ptr(ssp->sda, cpu));
1607 __srcu_read_unlock_nmisafe(ssp, idx);
1609 /* Remove the initial count, at which point reaching zero can happen. */
1610 if (atomic_dec_and_test(&ssp->srcu_sup->srcu_barrier_cpu_cnt))
1611 complete(&ssp->srcu_sup->srcu_barrier_completion);
1612 wait_for_completion(&ssp->srcu_sup->srcu_barrier_completion);
1614 rcu_seq_end(&ssp->srcu_sup->srcu_barrier_seq);
1615 mutex_unlock(&ssp->srcu_sup->srcu_barrier_mutex);
1617 EXPORT_SYMBOL_GPL(srcu_barrier);
1620 * srcu_batches_completed - return batches completed.
1621 * @ssp: srcu_struct on which to report batch completion.
1623 * Report the number of batches, correlated with, but not necessarily
1624 * precisely the same as, the number of grace periods that have elapsed.
1626 unsigned long srcu_batches_completed(struct srcu_struct *ssp)
1628 return READ_ONCE(ssp->srcu_idx);
1630 EXPORT_SYMBOL_GPL(srcu_batches_completed);
1633 * Core SRCU state machine. Push state bits of ->srcu_gp_seq
1634 * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
1635 * completed in that state.
1637 static void srcu_advance_state(struct srcu_struct *ssp)
1641 mutex_lock(&ssp->srcu_sup->srcu_gp_mutex);
1644 * Because readers might be delayed for an extended period after
1645 * fetching ->srcu_idx for their index, at any point in time there
1646 * might well be readers using both idx=0 and idx=1. We therefore
1647 * need to wait for readers to clear from both index values before
1648 * invoking a callback.
1650 * The load-acquire ensures that we see the accesses performed
1651 * by the prior grace period.
1653 idx = rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq)); /* ^^^ */
1654 if (idx == SRCU_STATE_IDLE) {
1655 spin_lock_irq_rcu_node(ssp->srcu_sup);
1656 if (ULONG_CMP_GE(ssp->srcu_sup->srcu_gp_seq, ssp->srcu_sup->srcu_gp_seq_needed)) {
1657 WARN_ON_ONCE(rcu_seq_state(ssp->srcu_sup->srcu_gp_seq));
1658 spin_unlock_irq_rcu_node(ssp->srcu_sup);
1659 mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
1662 idx = rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq));
1663 if (idx == SRCU_STATE_IDLE)
1665 spin_unlock_irq_rcu_node(ssp->srcu_sup);
1666 if (idx != SRCU_STATE_IDLE) {
1667 mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
1668 return; /* Someone else started the grace period. */
1672 if (rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq)) == SRCU_STATE_SCAN1) {
1673 idx = 1 ^ (ssp->srcu_idx & 1);
1674 if (!try_check_zero(ssp, idx, 1)) {
1675 mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
1676 return; /* readers present, retry later. */
1679 spin_lock_irq_rcu_node(ssp->srcu_sup);
1680 rcu_seq_set_state(&ssp->srcu_sup->srcu_gp_seq, SRCU_STATE_SCAN2);
1681 ssp->srcu_sup->srcu_n_exp_nodelay = 0;
1682 spin_unlock_irq_rcu_node(ssp->srcu_sup);
1685 if (rcu_seq_state(READ_ONCE(ssp->srcu_sup->srcu_gp_seq)) == SRCU_STATE_SCAN2) {
1688 * SRCU read-side critical sections are normally short,
1689 * so check at least twice in quick succession after a flip.
1691 idx = 1 ^ (ssp->srcu_idx & 1);
1692 if (!try_check_zero(ssp, idx, 2)) {
1693 mutex_unlock(&ssp->srcu_sup->srcu_gp_mutex);
1694 return; /* readers present, retry later. */
1696 ssp->srcu_sup->srcu_n_exp_nodelay = 0;
1697 srcu_gp_end(ssp); /* Releases ->srcu_gp_mutex. */
1702 * Invoke a limited number of SRCU callbacks that have passed through
1703 * their grace period. If there are more to do, SRCU will reschedule
1704 * the workqueue. Note that needed memory barriers have been executed
1705 * in this task's context by srcu_readers_active_idx_check().
1707 static void srcu_invoke_callbacks(struct work_struct *work)
1711 struct rcu_cblist ready_cbs;
1712 struct rcu_head *rhp;
1713 struct srcu_data *sdp;
1714 struct srcu_struct *ssp;
1716 sdp = container_of(work, struct srcu_data, work);
1719 rcu_cblist_init(&ready_cbs);
1720 spin_lock_irq_rcu_node(sdp);
1721 WARN_ON_ONCE(!rcu_segcblist_segempty(&sdp->srcu_cblist, RCU_NEXT_TAIL));
1722 rcu_segcblist_advance(&sdp->srcu_cblist,
1723 rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq));
1724 if (sdp->srcu_cblist_invoking ||
1725 !rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) {
1726 spin_unlock_irq_rcu_node(sdp);
1727 return; /* Someone else on the job or nothing to do. */
1730 /* We are on the job! Extract and invoke ready callbacks. */
1731 sdp->srcu_cblist_invoking = true;
1732 rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs);
1733 len = ready_cbs.len;
1734 spin_unlock_irq_rcu_node(sdp);
1735 rhp = rcu_cblist_dequeue(&ready_cbs);
1736 for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) {
1737 debug_rcu_head_unqueue(rhp);
1742 WARN_ON_ONCE(ready_cbs.len);
1745 * Update counts, accelerate new callbacks, and if needed,
1746 * schedule another round of callback invocation.
1748 spin_lock_irq_rcu_node(sdp);
1749 rcu_segcblist_add_len(&sdp->srcu_cblist, -len);
1750 sdp->srcu_cblist_invoking = false;
1751 more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist);
1752 spin_unlock_irq_rcu_node(sdp);
1754 srcu_schedule_cbs_sdp(sdp, 0);
1758 * Finished one round of SRCU grace period. Start another if there are
1759 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
1761 static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
1765 spin_lock_irq_rcu_node(ssp->srcu_sup);
1766 if (ULONG_CMP_GE(ssp->srcu_sup->srcu_gp_seq, ssp->srcu_sup->srcu_gp_seq_needed)) {
1767 if (!WARN_ON_ONCE(rcu_seq_state(ssp->srcu_sup->srcu_gp_seq))) {
1768 /* All requests fulfilled, time to go idle. */
1771 } else if (!rcu_seq_state(ssp->srcu_sup->srcu_gp_seq)) {
1772 /* Outstanding request and no GP. Start one. */
1775 spin_unlock_irq_rcu_node(ssp->srcu_sup);
1778 queue_delayed_work(rcu_gp_wq, &ssp->srcu_sup->work, delay);
1782 * This is the work-queue function that handles SRCU grace periods.
1784 static void process_srcu(struct work_struct *work)
1786 unsigned long curdelay;
1788 struct srcu_struct *ssp;
1789 struct srcu_usage *sup;
1791 sup = container_of(work, struct srcu_usage, work.work);
1792 ssp = sup->srcu_ssp;
1794 srcu_advance_state(ssp);
1795 curdelay = srcu_get_delay(ssp);
1797 WRITE_ONCE(sup->reschedule_count, 0);
1800 if (READ_ONCE(sup->reschedule_jiffies) == j) {
1801 WRITE_ONCE(sup->reschedule_count, READ_ONCE(sup->reschedule_count) + 1);
1802 if (READ_ONCE(sup->reschedule_count) > srcu_max_nodelay)
1805 WRITE_ONCE(sup->reschedule_count, 1);
1806 WRITE_ONCE(sup->reschedule_jiffies, j);
1809 srcu_reschedule(ssp, curdelay);
1812 void srcutorture_get_gp_data(enum rcutorture_type test_type,
1813 struct srcu_struct *ssp, int *flags,
1814 unsigned long *gp_seq)
1816 if (test_type != SRCU_FLAVOR)
1819 *gp_seq = rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq);
1821 EXPORT_SYMBOL_GPL(srcutorture_get_gp_data);
1823 static const char * const srcu_size_state_name[] = {
1826 "SRCU_SIZE_WAIT_BARRIER",
1827 "SRCU_SIZE_WAIT_CALL",
1828 "SRCU_SIZE_WAIT_CBS1",
1829 "SRCU_SIZE_WAIT_CBS2",
1830 "SRCU_SIZE_WAIT_CBS3",
1831 "SRCU_SIZE_WAIT_CBS4",
1836 void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf)
1840 unsigned long s0 = 0, s1 = 0;
1841 int ss_state = READ_ONCE(ssp->srcu_sup->srcu_size_state);
1842 int ss_state_idx = ss_state;
1844 idx = ssp->srcu_idx & 0x1;
1845 if (ss_state < 0 || ss_state >= ARRAY_SIZE(srcu_size_state_name))
1846 ss_state_idx = ARRAY_SIZE(srcu_size_state_name) - 1;
1847 pr_alert("%s%s Tree SRCU g%ld state %d (%s)",
1848 tt, tf, rcu_seq_current(&ssp->srcu_sup->srcu_gp_seq), ss_state,
1849 srcu_size_state_name[ss_state_idx]);
1851 // Called after cleanup_srcu_struct(), perhaps.
1852 pr_cont(" No per-CPU srcu_data structures (->sda == NULL).\n");
1854 pr_cont(" per-CPU(idx=%d):", idx);
1855 for_each_possible_cpu(cpu) {
1856 unsigned long l0, l1;
1857 unsigned long u0, u1;
1859 struct srcu_data *sdp;
1861 sdp = per_cpu_ptr(ssp->sda, cpu);
1862 u0 = data_race(atomic_long_read(&sdp->srcu_unlock_count[!idx]));
1863 u1 = data_race(atomic_long_read(&sdp->srcu_unlock_count[idx]));
1866 * Make sure that a lock is always counted if the corresponding
1867 * unlock is counted.
1871 l0 = data_race(atomic_long_read(&sdp->srcu_lock_count[!idx]));
1872 l1 = data_race(atomic_long_read(&sdp->srcu_lock_count[idx]));
1876 pr_cont(" %d(%ld,%ld %c)",
1878 "C."[rcu_segcblist_empty(&sdp->srcu_cblist)]);
1882 pr_cont(" T(%ld,%ld)\n", s0, s1);
1884 if (SRCU_SIZING_IS_TORTURE())
1885 srcu_transition_to_big(ssp);
1887 EXPORT_SYMBOL_GPL(srcu_torture_stats_print);
1889 static int __init srcu_bootup_announce(void)
1891 pr_info("Hierarchical SRCU implementation.\n");
1892 if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF)
1893 pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff);
1894 if (srcu_retry_check_delay != SRCU_DEFAULT_RETRY_CHECK_DELAY)
1895 pr_info("\tNon-default retry check delay of %lu us.\n", srcu_retry_check_delay);
1896 if (srcu_max_nodelay != SRCU_DEFAULT_MAX_NODELAY)
1897 pr_info("\tNon-default max no-delay of %lu.\n", srcu_max_nodelay);
1898 pr_info("\tMax phase no-delay instances is %lu.\n", srcu_max_nodelay_phase);
1901 early_initcall(srcu_bootup_announce);
1903 void __init srcu_init(void)
1905 struct srcu_usage *sup;
1907 /* Decide on srcu_struct-size strategy. */
1908 if (SRCU_SIZING_IS(SRCU_SIZING_AUTO)) {
1909 if (nr_cpu_ids >= big_cpu_lim) {
1910 convert_to_big = SRCU_SIZING_INIT; // Don't bother waiting for contention.
1911 pr_info("%s: Setting srcu_struct sizes to big.\n", __func__);
1913 convert_to_big = SRCU_SIZING_NONE | SRCU_SIZING_CONTEND;
1914 pr_info("%s: Setting srcu_struct sizes based on contention.\n", __func__);
1919 * Once that is set, call_srcu() can follow the normal path and
1920 * queue delayed work. This must follow RCU workqueues creation
1921 * and timers initialization.
1923 srcu_init_done = true;
1924 while (!list_empty(&srcu_boot_list)) {
1925 sup = list_first_entry(&srcu_boot_list, struct srcu_usage,
1927 list_del_init(&sup->work.work.entry);
1928 if (SRCU_SIZING_IS(SRCU_SIZING_INIT) &&
1929 sup->srcu_size_state == SRCU_SIZE_SMALL)
1930 sup->srcu_size_state = SRCU_SIZE_ALLOC;
1931 queue_work(rcu_gp_wq, &sup->work.work);
1935 #ifdef CONFIG_MODULES
1937 /* Initialize any global-scope srcu_struct structures used by this module. */
1938 static int srcu_module_coming(struct module *mod)
1941 struct srcu_struct *ssp;
1942 struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1944 for (i = 0; i < mod->num_srcu_structs; i++) {
1946 ssp->sda = alloc_percpu(struct srcu_data);
1947 if (WARN_ON_ONCE(!ssp->sda))
1953 /* Clean up any global-scope srcu_struct structures used by this module. */
1954 static void srcu_module_going(struct module *mod)
1957 struct srcu_struct *ssp;
1958 struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1960 for (i = 0; i < mod->num_srcu_structs; i++) {
1962 if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_sup->srcu_gp_seq_needed)) &&
1963 !WARN_ON_ONCE(!ssp->srcu_sup->sda_is_static))
1964 cleanup_srcu_struct(ssp);
1965 if (!WARN_ON(srcu_readers_active(ssp)))
1966 free_percpu(ssp->sda);
1970 /* Handle one module, either coming or going. */
1971 static int srcu_module_notify(struct notifier_block *self,
1972 unsigned long val, void *data)
1974 struct module *mod = data;
1978 case MODULE_STATE_COMING:
1979 ret = srcu_module_coming(mod);
1981 case MODULE_STATE_GOING:
1982 srcu_module_going(mod);
1990 static struct notifier_block srcu_module_nb = {
1991 .notifier_call = srcu_module_notify,
1995 static __init int init_srcu_module_notifier(void)
1999 ret = register_module_notifier(&srcu_module_nb);
2001 pr_warn("Failed to register srcu module notifier\n");
2004 late_initcall(init_srcu_module_notifier);
2006 #endif /* #ifdef CONFIG_MODULES */