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
);
185 init_srcu_struct_nodes(ssp
, is_static
);
186 ssp
->srcu_gp_seq_needed_exp
= 0;
187 ssp
->srcu_last_gp_end
= ktime_get_mono_fast_ns();
188 smp_store_release(&ssp
->srcu_gp_seq_needed
, 0); /* Init done. */
192 #ifdef CONFIG_DEBUG_LOCK_ALLOC
194 int __init_srcu_struct(struct srcu_struct
*ssp
, const char *name
,
195 struct lock_class_key
*key
)
197 /* Don't re-initialize a lock while it is held. */
198 debug_check_no_locks_freed((void *)ssp
, sizeof(*ssp
));
199 lockdep_init_map(&ssp
->dep_map
, name
, key
, 0);
200 spin_lock_init(&ACCESS_PRIVATE(ssp
, lock
));
201 return init_srcu_struct_fields(ssp
, false);
203 EXPORT_SYMBOL_GPL(__init_srcu_struct
);
205 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
208 * init_srcu_struct - initialize a sleep-RCU structure
209 * @ssp: structure to initialize.
211 * Must invoke this on a given srcu_struct before passing that srcu_struct
212 * to any other function. Each srcu_struct represents a separate domain
213 * of SRCU protection.
215 int init_srcu_struct(struct srcu_struct
*ssp
)
217 spin_lock_init(&ACCESS_PRIVATE(ssp
, lock
));
218 return init_srcu_struct_fields(ssp
, false);
220 EXPORT_SYMBOL_GPL(init_srcu_struct
);
222 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
225 * First-use initialization of statically allocated srcu_struct
226 * structure. Wiring up the combining tree is more than can be
227 * done with compile-time initialization, so this check is added
228 * to each update-side SRCU primitive. Use ssp->lock, which -is-
229 * compile-time initialized, to resolve races involving multiple
230 * CPUs trying to garner first-use privileges.
232 static void check_init_srcu_struct(struct srcu_struct
*ssp
)
236 /* The smp_load_acquire() pairs with the smp_store_release(). */
237 if (!rcu_seq_state(smp_load_acquire(&ssp
->srcu_gp_seq_needed
))) /*^^^*/
238 return; /* Already initialized. */
239 spin_lock_irqsave_rcu_node(ssp
, flags
);
240 if (!rcu_seq_state(ssp
->srcu_gp_seq_needed
)) {
241 spin_unlock_irqrestore_rcu_node(ssp
, flags
);
244 init_srcu_struct_fields(ssp
, true);
245 spin_unlock_irqrestore_rcu_node(ssp
, flags
);
249 * Returns approximate total of the readers' ->srcu_lock_count[] values
250 * for the rank of per-CPU counters specified by idx.
252 static unsigned long srcu_readers_lock_idx(struct srcu_struct
*ssp
, int idx
)
255 unsigned long sum
= 0;
257 for_each_possible_cpu(cpu
) {
258 struct srcu_data
*cpuc
= per_cpu_ptr(ssp
->sda
, cpu
);
260 sum
+= READ_ONCE(cpuc
->srcu_lock_count
[idx
]);
266 * Returns approximate total of the readers' ->srcu_unlock_count[] values
267 * for the rank of per-CPU counters specified by idx.
269 static unsigned long srcu_readers_unlock_idx(struct srcu_struct
*ssp
, int idx
)
272 unsigned long sum
= 0;
274 for_each_possible_cpu(cpu
) {
275 struct srcu_data
*cpuc
= per_cpu_ptr(ssp
->sda
, cpu
);
277 sum
+= READ_ONCE(cpuc
->srcu_unlock_count
[idx
]);
283 * Return true if the number of pre-existing readers is determined to
286 static bool srcu_readers_active_idx_check(struct srcu_struct
*ssp
, int idx
)
288 unsigned long unlocks
;
290 unlocks
= srcu_readers_unlock_idx(ssp
, idx
);
293 * Make sure that a lock is always counted if the corresponding
294 * unlock is counted. Needs to be a smp_mb() as the read side may
295 * contain a read from a variable that is written to before the
296 * synchronize_srcu() in the write side. In this case smp_mb()s
297 * A and B act like the store buffering pattern.
299 * This smp_mb() also pairs with smp_mb() C to prevent accesses
300 * after the synchronize_srcu() from being executed before the
306 * If the locks are the same as the unlocks, then there must have
307 * been no readers on this index at some time in between. This does
308 * not mean that there are no more readers, as one could have read
309 * the current index but not have incremented the lock counter yet.
311 * So suppose that the updater is preempted here for so long
312 * that more than ULONG_MAX non-nested readers come and go in
313 * the meantime. It turns out that this cannot result in overflow
314 * because if a reader modifies its unlock count after we read it
315 * above, then that reader's next load of ->srcu_idx is guaranteed
316 * to get the new value, which will cause it to operate on the
317 * other bank of counters, where it cannot contribute to the
318 * overflow of these counters. This means that there is a maximum
319 * of 2*NR_CPUS increments, which cannot overflow given current
320 * systems, especially not on 64-bit systems.
322 * OK, how about nesting? This does impose a limit on nesting
323 * of floor(ULONG_MAX/NR_CPUS/2), which should be sufficient,
324 * especially on 64-bit systems.
326 return srcu_readers_lock_idx(ssp
, idx
) == unlocks
;
330 * srcu_readers_active - returns true if there are readers. and false
332 * @ssp: which srcu_struct to count active readers (holding srcu_read_lock).
334 * Note that this is not an atomic primitive, and can therefore suffer
335 * severe errors when invoked on an active srcu_struct. That said, it
336 * can be useful as an error check at cleanup time.
338 static bool srcu_readers_active(struct srcu_struct
*ssp
)
341 unsigned long sum
= 0;
343 for_each_possible_cpu(cpu
) {
344 struct srcu_data
*cpuc
= per_cpu_ptr(ssp
->sda
, cpu
);
346 sum
+= READ_ONCE(cpuc
->srcu_lock_count
[0]);
347 sum
+= READ_ONCE(cpuc
->srcu_lock_count
[1]);
348 sum
-= READ_ONCE(cpuc
->srcu_unlock_count
[0]);
349 sum
-= READ_ONCE(cpuc
->srcu_unlock_count
[1]);
354 #define SRCU_INTERVAL 1
357 * Return grace-period delay, zero if there are expedited grace
358 * periods pending, SRCU_INTERVAL otherwise.
360 static unsigned long srcu_get_delay(struct srcu_struct
*ssp
)
362 if (ULONG_CMP_LT(READ_ONCE(ssp
->srcu_gp_seq
),
363 READ_ONCE(ssp
->srcu_gp_seq_needed_exp
)))
365 return SRCU_INTERVAL
;
369 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
370 * @ssp: structure to clean up.
372 * Must invoke this after you are finished using a given srcu_struct that
373 * was initialized via init_srcu_struct(), else you leak memory.
375 void cleanup_srcu_struct(struct srcu_struct
*ssp
)
379 if (WARN_ON(!srcu_get_delay(ssp
)))
380 return; /* Just leak it! */
381 if (WARN_ON(srcu_readers_active(ssp
)))
382 return; /* Just leak it! */
383 flush_delayed_work(&ssp
->work
);
384 for_each_possible_cpu(cpu
) {
385 struct srcu_data
*sdp
= per_cpu_ptr(ssp
->sda
, cpu
);
387 del_timer_sync(&sdp
->delay_work
);
388 flush_work(&sdp
->work
);
389 if (WARN_ON(rcu_segcblist_n_cbs(&sdp
->srcu_cblist
)))
390 return; /* Forgot srcu_barrier(), so just leak it! */
392 if (WARN_ON(rcu_seq_state(READ_ONCE(ssp
->srcu_gp_seq
)) != SRCU_STATE_IDLE
) ||
393 WARN_ON(srcu_readers_active(ssp
))) {
394 pr_info("%s: Active srcu_struct %p state: %d\n",
395 __func__
, ssp
, rcu_seq_state(READ_ONCE(ssp
->srcu_gp_seq
)));
396 return; /* Caller forgot to stop doing call_srcu()? */
398 free_percpu(ssp
->sda
);
401 EXPORT_SYMBOL_GPL(cleanup_srcu_struct
);
404 * Counts the new reader in the appropriate per-CPU element of the
406 * Returns an index that must be passed to the matching srcu_read_unlock().
408 int __srcu_read_lock(struct srcu_struct
*ssp
)
412 idx
= READ_ONCE(ssp
->srcu_idx
) & 0x1;
413 this_cpu_inc(ssp
->sda
->srcu_lock_count
[idx
]);
414 smp_mb(); /* B */ /* Avoid leaking the critical section. */
417 EXPORT_SYMBOL_GPL(__srcu_read_lock
);
420 * Removes the count for the old reader from the appropriate per-CPU
421 * element of the srcu_struct. Note that this may well be a different
422 * CPU than that which was incremented by the corresponding srcu_read_lock().
424 void __srcu_read_unlock(struct srcu_struct
*ssp
, int idx
)
426 smp_mb(); /* C */ /* Avoid leaking the critical section. */
427 this_cpu_inc(ssp
->sda
->srcu_unlock_count
[idx
]);
429 EXPORT_SYMBOL_GPL(__srcu_read_unlock
);
432 * We use an adaptive strategy for synchronize_srcu() and especially for
433 * synchronize_srcu_expedited(). We spin for a fixed time period
434 * (defined below) to allow SRCU readers to exit their read-side critical
435 * sections. If there are still some readers after a few microseconds,
436 * we repeatedly block for 1-millisecond time periods.
438 #define SRCU_RETRY_CHECK_DELAY 5
441 * Start an SRCU grace period.
443 static void srcu_gp_start(struct srcu_struct
*ssp
)
445 struct srcu_data
*sdp
= this_cpu_ptr(ssp
->sda
);
448 lockdep_assert_held(&ACCESS_PRIVATE(ssp
, lock
));
449 WARN_ON_ONCE(ULONG_CMP_GE(ssp
->srcu_gp_seq
, ssp
->srcu_gp_seq_needed
));
450 spin_lock_rcu_node(sdp
); /* Interrupts already disabled. */
451 rcu_segcblist_advance(&sdp
->srcu_cblist
,
452 rcu_seq_current(&ssp
->srcu_gp_seq
));
453 (void)rcu_segcblist_accelerate(&sdp
->srcu_cblist
,
454 rcu_seq_snap(&ssp
->srcu_gp_seq
));
455 spin_unlock_rcu_node(sdp
); /* Interrupts remain disabled. */
456 smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
457 rcu_seq_start(&ssp
->srcu_gp_seq
);
458 state
= rcu_seq_state(ssp
->srcu_gp_seq
);
459 WARN_ON_ONCE(state
!= SRCU_STATE_SCAN1
);
463 static void srcu_delay_timer(struct timer_list
*t
)
465 struct srcu_data
*sdp
= container_of(t
, struct srcu_data
, delay_work
);
467 queue_work_on(sdp
->cpu
, rcu_gp_wq
, &sdp
->work
);
470 static void srcu_queue_delayed_work_on(struct srcu_data
*sdp
,
474 queue_work_on(sdp
->cpu
, rcu_gp_wq
, &sdp
->work
);
478 timer_reduce(&sdp
->delay_work
, jiffies
+ delay
);
482 * Schedule callback invocation for the specified srcu_data structure,
483 * if possible, on the corresponding CPU.
485 static void srcu_schedule_cbs_sdp(struct srcu_data
*sdp
, unsigned long delay
)
487 srcu_queue_delayed_work_on(sdp
, delay
);
491 * Schedule callback invocation for all srcu_data structures associated
492 * with the specified srcu_node structure that have callbacks for the
493 * just-completed grace period, the one corresponding to idx. If possible,
494 * schedule this invocation on the corresponding CPUs.
496 static void srcu_schedule_cbs_snp(struct srcu_struct
*ssp
, struct srcu_node
*snp
,
497 unsigned long mask
, unsigned long delay
)
501 for (cpu
= snp
->grplo
; cpu
<= snp
->grphi
; cpu
++) {
502 if (!(mask
& (1 << (cpu
- snp
->grplo
))))
504 srcu_schedule_cbs_sdp(per_cpu_ptr(ssp
->sda
, cpu
), delay
);
509 * Note the end of an SRCU grace period. Initiates callback invocation
510 * and starts a new grace period if needed.
512 * The ->srcu_cb_mutex acquisition does not protect any data, but
513 * instead prevents more than one grace period from starting while we
514 * are initiating callback invocation. This allows the ->srcu_have_cbs[]
515 * array to have a finite number of elements.
517 static void srcu_gp_end(struct srcu_struct
*ssp
)
519 unsigned long cbdelay
;
527 struct srcu_data
*sdp
;
528 struct srcu_node
*snp
;
530 /* Prevent more than one additional grace period. */
531 mutex_lock(&ssp
->srcu_cb_mutex
);
533 /* End the current grace period. */
534 spin_lock_irq_rcu_node(ssp
);
535 idx
= rcu_seq_state(ssp
->srcu_gp_seq
);
536 WARN_ON_ONCE(idx
!= SRCU_STATE_SCAN2
);
537 cbdelay
= srcu_get_delay(ssp
);
538 WRITE_ONCE(ssp
->srcu_last_gp_end
, ktime_get_mono_fast_ns());
539 rcu_seq_end(&ssp
->srcu_gp_seq
);
540 gpseq
= rcu_seq_current(&ssp
->srcu_gp_seq
);
541 if (ULONG_CMP_LT(ssp
->srcu_gp_seq_needed_exp
, gpseq
))
542 WRITE_ONCE(ssp
->srcu_gp_seq_needed_exp
, gpseq
);
543 spin_unlock_irq_rcu_node(ssp
);
544 mutex_unlock(&ssp
->srcu_gp_mutex
);
545 /* A new grace period can start at this point. But only one. */
547 /* Initiate callback invocation as needed. */
548 idx
= rcu_seq_ctr(gpseq
) % ARRAY_SIZE(snp
->srcu_have_cbs
);
549 srcu_for_each_node_breadth_first(ssp
, snp
) {
550 spin_lock_irq_rcu_node(snp
);
552 last_lvl
= snp
>= ssp
->level
[rcu_num_lvls
- 1];
554 cbs
= snp
->srcu_have_cbs
[idx
] == gpseq
;
555 snp
->srcu_have_cbs
[idx
] = gpseq
;
556 rcu_seq_set_state(&snp
->srcu_have_cbs
[idx
], 1);
557 if (ULONG_CMP_LT(snp
->srcu_gp_seq_needed_exp
, gpseq
))
558 WRITE_ONCE(snp
->srcu_gp_seq_needed_exp
, gpseq
);
559 mask
= snp
->srcu_data_have_cbs
[idx
];
560 snp
->srcu_data_have_cbs
[idx
] = 0;
561 spin_unlock_irq_rcu_node(snp
);
563 srcu_schedule_cbs_snp(ssp
, snp
, mask
, cbdelay
);
565 /* Occasionally prevent srcu_data counter wrap. */
566 if (!(gpseq
& counter_wrap_check
) && last_lvl
)
567 for (cpu
= snp
->grplo
; cpu
<= snp
->grphi
; cpu
++) {
568 sdp
= per_cpu_ptr(ssp
->sda
, cpu
);
569 spin_lock_irqsave_rcu_node(sdp
, flags
);
570 if (ULONG_CMP_GE(gpseq
,
571 sdp
->srcu_gp_seq_needed
+ 100))
572 sdp
->srcu_gp_seq_needed
= gpseq
;
573 if (ULONG_CMP_GE(gpseq
,
574 sdp
->srcu_gp_seq_needed_exp
+ 100))
575 sdp
->srcu_gp_seq_needed_exp
= gpseq
;
576 spin_unlock_irqrestore_rcu_node(sdp
, flags
);
580 /* Callback initiation done, allow grace periods after next. */
581 mutex_unlock(&ssp
->srcu_cb_mutex
);
583 /* Start a new grace period if needed. */
584 spin_lock_irq_rcu_node(ssp
);
585 gpseq
= rcu_seq_current(&ssp
->srcu_gp_seq
);
586 if (!rcu_seq_state(gpseq
) &&
587 ULONG_CMP_LT(gpseq
, ssp
->srcu_gp_seq_needed
)) {
589 spin_unlock_irq_rcu_node(ssp
);
590 srcu_reschedule(ssp
, 0);
592 spin_unlock_irq_rcu_node(ssp
);
597 * Funnel-locking scheme to scalably mediate many concurrent expedited
598 * grace-period requests. This function is invoked for the first known
599 * expedited request for a grace period that has already been requested,
600 * but without expediting. To start a completely new grace period,
601 * whether expedited or not, use srcu_funnel_gp_start() instead.
603 static void srcu_funnel_exp_start(struct srcu_struct
*ssp
, struct srcu_node
*snp
,
608 for (; snp
!= NULL
; snp
= snp
->srcu_parent
) {
609 if (rcu_seq_done(&ssp
->srcu_gp_seq
, s
) ||
610 ULONG_CMP_GE(READ_ONCE(snp
->srcu_gp_seq_needed_exp
), s
))
612 spin_lock_irqsave_rcu_node(snp
, flags
);
613 if (ULONG_CMP_GE(snp
->srcu_gp_seq_needed_exp
, s
)) {
614 spin_unlock_irqrestore_rcu_node(snp
, flags
);
617 WRITE_ONCE(snp
->srcu_gp_seq_needed_exp
, s
);
618 spin_unlock_irqrestore_rcu_node(snp
, flags
);
620 spin_lock_irqsave_rcu_node(ssp
, flags
);
621 if (ULONG_CMP_LT(ssp
->srcu_gp_seq_needed_exp
, s
))
622 WRITE_ONCE(ssp
->srcu_gp_seq_needed_exp
, s
);
623 spin_unlock_irqrestore_rcu_node(ssp
, flags
);
627 * Funnel-locking scheme to scalably mediate many concurrent grace-period
628 * requests. The winner has to do the work of actually starting grace
629 * period s. Losers must either ensure that their desired grace-period
630 * number is recorded on at least their leaf srcu_node structure, or they
631 * must take steps to invoke their own callbacks.
633 * Note that this function also does the work of srcu_funnel_exp_start(),
634 * in some cases by directly invoking it.
636 static void srcu_funnel_gp_start(struct srcu_struct
*ssp
, struct srcu_data
*sdp
,
637 unsigned long s
, bool do_norm
)
640 int idx
= rcu_seq_ctr(s
) % ARRAY_SIZE(sdp
->mynode
->srcu_have_cbs
);
641 struct srcu_node
*snp
= sdp
->mynode
;
642 unsigned long snp_seq
;
644 /* Each pass through the loop does one level of the srcu_node tree. */
645 for (; snp
!= NULL
; snp
= snp
->srcu_parent
) {
646 if (rcu_seq_done(&ssp
->srcu_gp_seq
, s
) && snp
!= sdp
->mynode
)
647 return; /* GP already done and CBs recorded. */
648 spin_lock_irqsave_rcu_node(snp
, flags
);
649 if (ULONG_CMP_GE(snp
->srcu_have_cbs
[idx
], s
)) {
650 snp_seq
= snp
->srcu_have_cbs
[idx
];
651 if (snp
== sdp
->mynode
&& snp_seq
== s
)
652 snp
->srcu_data_have_cbs
[idx
] |= sdp
->grpmask
;
653 spin_unlock_irqrestore_rcu_node(snp
, flags
);
654 if (snp
== sdp
->mynode
&& snp_seq
!= s
) {
655 srcu_schedule_cbs_sdp(sdp
, do_norm
661 srcu_funnel_exp_start(ssp
, snp
, s
);
664 snp
->srcu_have_cbs
[idx
] = s
;
665 if (snp
== sdp
->mynode
)
666 snp
->srcu_data_have_cbs
[idx
] |= sdp
->grpmask
;
667 if (!do_norm
&& ULONG_CMP_LT(snp
->srcu_gp_seq_needed_exp
, s
))
668 WRITE_ONCE(snp
->srcu_gp_seq_needed_exp
, s
);
669 spin_unlock_irqrestore_rcu_node(snp
, flags
);
672 /* Top of tree, must ensure the grace period will be started. */
673 spin_lock_irqsave_rcu_node(ssp
, flags
);
674 if (ULONG_CMP_LT(ssp
->srcu_gp_seq_needed
, s
)) {
676 * Record need for grace period s. Pair with load
677 * acquire setting up for initialization.
679 smp_store_release(&ssp
->srcu_gp_seq_needed
, s
); /*^^^*/
681 if (!do_norm
&& ULONG_CMP_LT(ssp
->srcu_gp_seq_needed_exp
, s
))
682 WRITE_ONCE(ssp
->srcu_gp_seq_needed_exp
, s
);
684 /* If grace period not already done and none in progress, start it. */
685 if (!rcu_seq_done(&ssp
->srcu_gp_seq
, s
) &&
686 rcu_seq_state(ssp
->srcu_gp_seq
) == SRCU_STATE_IDLE
) {
687 WARN_ON_ONCE(ULONG_CMP_GE(ssp
->srcu_gp_seq
, ssp
->srcu_gp_seq_needed
));
689 if (likely(srcu_init_done
))
690 queue_delayed_work(rcu_gp_wq
, &ssp
->work
,
691 srcu_get_delay(ssp
));
692 else if (list_empty(&ssp
->work
.work
.entry
))
693 list_add(&ssp
->work
.work
.entry
, &srcu_boot_list
);
695 spin_unlock_irqrestore_rcu_node(ssp
, flags
);
699 * Wait until all readers counted by array index idx complete, but
700 * loop an additional time if there is an expedited grace period pending.
701 * The caller must ensure that ->srcu_idx is not changed while checking.
703 static bool try_check_zero(struct srcu_struct
*ssp
, int idx
, int trycount
)
706 if (srcu_readers_active_idx_check(ssp
, idx
))
708 if (--trycount
+ !srcu_get_delay(ssp
) <= 0)
710 udelay(SRCU_RETRY_CHECK_DELAY
);
715 * Increment the ->srcu_idx counter so that future SRCU readers will
716 * use the other rank of the ->srcu_(un)lock_count[] arrays. This allows
717 * us to wait for pre-existing readers in a starvation-free manner.
719 static void srcu_flip(struct srcu_struct
*ssp
)
722 * Ensure that if this updater saw a given reader's increment
723 * from __srcu_read_lock(), that reader was using an old value
724 * of ->srcu_idx. Also ensure that if a given reader sees the
725 * new value of ->srcu_idx, this updater's earlier scans cannot
726 * have seen that reader's increments (which is OK, because this
727 * grace period need not wait on that reader).
729 smp_mb(); /* E */ /* Pairs with B and C. */
731 WRITE_ONCE(ssp
->srcu_idx
, ssp
->srcu_idx
+ 1);
734 * Ensure that if the updater misses an __srcu_read_unlock()
735 * increment, that task's next __srcu_read_lock() will see the
736 * above counter update. Note that both this memory barrier
737 * and the one in srcu_readers_active_idx_check() provide the
738 * guarantee for __srcu_read_lock().
740 smp_mb(); /* D */ /* Pairs with C. */
744 * If SRCU is likely idle, return true, otherwise return false.
746 * Note that it is OK for several current from-idle requests for a new
747 * grace period from idle to specify expediting because they will all end
748 * up requesting the same grace period anyhow. So no loss.
750 * Note also that if any CPU (including the current one) is still invoking
751 * callbacks, this function will nevertheless say "idle". This is not
752 * ideal, but the overhead of checking all CPUs' callback lists is even
753 * less ideal, especially on large systems. Furthermore, the wakeup
754 * can happen before the callback is fully removed, so we have no choice
755 * but to accept this type of error.
757 * This function is also subject to counter-wrap errors, but let's face
758 * it, if this function was preempted for enough time for the counters
759 * to wrap, it really doesn't matter whether or not we expedite the grace
760 * period. The extra overhead of a needlessly expedited grace period is
761 * negligible when amortized over that time period, and the extra latency
762 * of a needlessly non-expedited grace period is similarly negligible.
764 static bool srcu_might_be_idle(struct srcu_struct
*ssp
)
766 unsigned long curseq
;
768 struct srcu_data
*sdp
;
772 check_init_srcu_struct(ssp
);
773 /* If the local srcu_data structure has callbacks, not idle. */
774 sdp
= raw_cpu_ptr(ssp
->sda
);
775 spin_lock_irqsave_rcu_node(sdp
, flags
);
776 if (rcu_segcblist_pend_cbs(&sdp
->srcu_cblist
)) {
777 spin_unlock_irqrestore_rcu_node(sdp
, flags
);
778 return false; /* Callbacks already present, so not idle. */
780 spin_unlock_irqrestore_rcu_node(sdp
, flags
);
783 * No local callbacks, so probabalistically probe global state.
784 * Exact information would require acquiring locks, which would
785 * kill scalability, hence the probabalistic nature of the probe.
788 /* First, see if enough time has passed since the last GP. */
789 t
= ktime_get_mono_fast_ns();
790 tlast
= READ_ONCE(ssp
->srcu_last_gp_end
);
791 if (exp_holdoff
== 0 ||
792 time_in_range_open(t
, tlast
, tlast
+ exp_holdoff
))
793 return false; /* Too soon after last GP. */
795 /* Next, check for probable idleness. */
796 curseq
= rcu_seq_current(&ssp
->srcu_gp_seq
);
797 smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
798 if (ULONG_CMP_LT(curseq
, READ_ONCE(ssp
->srcu_gp_seq_needed
)))
799 return false; /* Grace period in progress, so not idle. */
800 smp_mb(); /* Order ->srcu_gp_seq with prior access. */
801 if (curseq
!= rcu_seq_current(&ssp
->srcu_gp_seq
))
802 return false; /* GP # changed, so not idle. */
803 return true; /* With reasonable probability, idle! */
807 * SRCU callback function to leak a callback.
809 static void srcu_leak_callback(struct rcu_head
*rhp
)
814 * Start an SRCU grace period, and also queue the callback if non-NULL.
816 static unsigned long srcu_gp_start_if_needed(struct srcu_struct
*ssp
,
817 struct rcu_head
*rhp
, bool do_norm
)
821 bool needexp
= false;
824 struct srcu_data
*sdp
;
826 check_init_srcu_struct(ssp
);
827 idx
= srcu_read_lock(ssp
);
828 sdp
= raw_cpu_ptr(ssp
->sda
);
829 spin_lock_irqsave_rcu_node(sdp
, flags
);
831 rcu_segcblist_enqueue(&sdp
->srcu_cblist
, rhp
);
832 rcu_segcblist_advance(&sdp
->srcu_cblist
,
833 rcu_seq_current(&ssp
->srcu_gp_seq
));
834 s
= rcu_seq_snap(&ssp
->srcu_gp_seq
);
835 (void)rcu_segcblist_accelerate(&sdp
->srcu_cblist
, s
);
836 if (ULONG_CMP_LT(sdp
->srcu_gp_seq_needed
, s
)) {
837 sdp
->srcu_gp_seq_needed
= s
;
840 if (!do_norm
&& ULONG_CMP_LT(sdp
->srcu_gp_seq_needed_exp
, s
)) {
841 sdp
->srcu_gp_seq_needed_exp
= s
;
844 spin_unlock_irqrestore_rcu_node(sdp
, flags
);
846 srcu_funnel_gp_start(ssp
, sdp
, s
, do_norm
);
848 srcu_funnel_exp_start(ssp
, sdp
->mynode
, s
);
849 srcu_read_unlock(ssp
, idx
);
854 * Enqueue an SRCU callback on the srcu_data structure associated with
855 * the current CPU and the specified srcu_struct structure, initiating
856 * grace-period processing if it is not already running.
858 * Note that all CPUs must agree that the grace period extended beyond
859 * all pre-existing SRCU read-side critical section. On systems with
860 * more than one CPU, this means that when "func()" is invoked, each CPU
861 * is guaranteed to have executed a full memory barrier since the end of
862 * its last corresponding SRCU read-side critical section whose beginning
863 * preceded the call to call_srcu(). It also means that each CPU executing
864 * an SRCU read-side critical section that continues beyond the start of
865 * "func()" must have executed a memory barrier after the call_srcu()
866 * but before the beginning of that SRCU read-side critical section.
867 * Note that these guarantees include CPUs that are offline, idle, or
868 * executing in user mode, as well as CPUs that are executing in the kernel.
870 * Furthermore, if CPU A invoked call_srcu() and CPU B invoked the
871 * resulting SRCU callback function "func()", then both CPU A and CPU
872 * B are guaranteed to execute a full memory barrier during the time
873 * interval between the call to call_srcu() and the invocation of "func()".
874 * This guarantee applies even if CPU A and CPU B are the same CPU (but
875 * again only if the system has more than one CPU).
877 * Of course, these guarantees apply only for invocations of call_srcu(),
878 * srcu_read_lock(), and srcu_read_unlock() that are all passed the same
879 * srcu_struct structure.
881 static void __call_srcu(struct srcu_struct
*ssp
, struct rcu_head
*rhp
,
882 rcu_callback_t func
, bool do_norm
)
884 if (debug_rcu_head_queue(rhp
)) {
885 /* Probable double call_srcu(), so leak the callback. */
886 WRITE_ONCE(rhp
->func
, srcu_leak_callback
);
887 WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n");
891 (void)srcu_gp_start_if_needed(ssp
, rhp
, do_norm
);
895 * call_srcu() - Queue a callback for invocation after an SRCU grace period
896 * @ssp: srcu_struct in queue the callback
897 * @rhp: structure to be used for queueing the SRCU callback.
898 * @func: function to be invoked after the SRCU grace period
900 * The callback function will be invoked some time after a full SRCU
901 * grace period elapses, in other words after all pre-existing SRCU
902 * read-side critical sections have completed. However, the callback
903 * function might well execute concurrently with other SRCU read-side
904 * critical sections that started after call_srcu() was invoked. SRCU
905 * read-side critical sections are delimited by srcu_read_lock() and
906 * srcu_read_unlock(), and may be nested.
908 * The callback will be invoked from process context, but must nevertheless
909 * be fast and must not block.
911 void call_srcu(struct srcu_struct
*ssp
, struct rcu_head
*rhp
,
914 __call_srcu(ssp
, rhp
, func
, true);
916 EXPORT_SYMBOL_GPL(call_srcu
);
919 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
921 static void __synchronize_srcu(struct srcu_struct
*ssp
, bool do_norm
)
923 struct rcu_synchronize rcu
;
925 RCU_LOCKDEP_WARN(lockdep_is_held(ssp
) ||
926 lock_is_held(&rcu_bh_lock_map
) ||
927 lock_is_held(&rcu_lock_map
) ||
928 lock_is_held(&rcu_sched_lock_map
),
929 "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
931 if (rcu_scheduler_active
== RCU_SCHEDULER_INACTIVE
)
934 check_init_srcu_struct(ssp
);
935 init_completion(&rcu
.completion
);
936 init_rcu_head_on_stack(&rcu
.head
);
937 __call_srcu(ssp
, &rcu
.head
, wakeme_after_rcu
, do_norm
);
938 wait_for_completion(&rcu
.completion
);
939 destroy_rcu_head_on_stack(&rcu
.head
);
942 * Make sure that later code is ordered after the SRCU grace
943 * period. This pairs with the spin_lock_irq_rcu_node()
944 * in srcu_invoke_callbacks(). Unlike Tree RCU, this is needed
945 * because the current CPU might have been totally uninvolved with
946 * (and thus unordered against) that grace period.
952 * synchronize_srcu_expedited - Brute-force SRCU grace period
953 * @ssp: srcu_struct with which to synchronize.
955 * Wait for an SRCU grace period to elapse, but be more aggressive about
956 * spinning rather than blocking when waiting.
958 * Note that synchronize_srcu_expedited() has the same deadlock and
959 * memory-ordering properties as does synchronize_srcu().
961 void synchronize_srcu_expedited(struct srcu_struct
*ssp
)
963 __synchronize_srcu(ssp
, rcu_gp_is_normal());
965 EXPORT_SYMBOL_GPL(synchronize_srcu_expedited
);
968 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
969 * @ssp: srcu_struct with which to synchronize.
971 * Wait for the count to drain to zero of both indexes. To avoid the
972 * possible starvation of synchronize_srcu(), it waits for the count of
973 * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
974 * and then flip the srcu_idx and wait for the count of the other index.
976 * Can block; must be called from process context.
978 * Note that it is illegal to call synchronize_srcu() from the corresponding
979 * SRCU read-side critical section; doing so will result in deadlock.
980 * However, it is perfectly legal to call synchronize_srcu() on one
981 * srcu_struct from some other srcu_struct's read-side critical section,
982 * as long as the resulting graph of srcu_structs is acyclic.
984 * There are memory-ordering constraints implied by synchronize_srcu().
985 * On systems with more than one CPU, when synchronize_srcu() returns,
986 * each CPU is guaranteed to have executed a full memory barrier since
987 * the end of its last corresponding SRCU read-side critical section
988 * whose beginning preceded the call to synchronize_srcu(). In addition,
989 * each CPU having an SRCU read-side critical section that extends beyond
990 * the return from synchronize_srcu() is guaranteed to have executed a
991 * full memory barrier after the beginning of synchronize_srcu() and before
992 * the beginning of that SRCU read-side critical section. Note that these
993 * guarantees include CPUs that are offline, idle, or executing in user mode,
994 * as well as CPUs that are executing in the kernel.
996 * Furthermore, if CPU A invoked synchronize_srcu(), which returned
997 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
998 * to have executed a full memory barrier during the execution of
999 * synchronize_srcu(). This guarantee applies even if CPU A and CPU B
1000 * are the same CPU, but again only if the system has more than one CPU.
1002 * Of course, these memory-ordering guarantees apply only when
1003 * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
1004 * passed the same srcu_struct structure.
1006 * If SRCU is likely idle, expedite the first request. This semantic
1007 * was provided by Classic SRCU, and is relied upon by its users, so TREE
1008 * SRCU must also provide it. Note that detecting idleness is heuristic
1009 * and subject to both false positives and negatives.
1011 void synchronize_srcu(struct srcu_struct
*ssp
)
1013 if (srcu_might_be_idle(ssp
) || rcu_gp_is_expedited())
1014 synchronize_srcu_expedited(ssp
);
1016 __synchronize_srcu(ssp
, true);
1018 EXPORT_SYMBOL_GPL(synchronize_srcu
);
1021 * get_state_synchronize_srcu - Provide an end-of-grace-period cookie
1022 * @ssp: srcu_struct to provide cookie for.
1024 * This function returns a cookie that can be passed to
1025 * poll_state_synchronize_srcu(), which will return true if a full grace
1026 * period has elapsed in the meantime. It is the caller's responsibility
1027 * to make sure that grace period happens, for example, by invoking
1028 * call_srcu() after return from get_state_synchronize_srcu().
1030 unsigned long get_state_synchronize_srcu(struct srcu_struct
*ssp
)
1032 // Any prior manipulation of SRCU-protected data must happen
1033 // before the load from ->srcu_gp_seq.
1035 return rcu_seq_snap(&ssp
->srcu_gp_seq
);
1037 EXPORT_SYMBOL_GPL(get_state_synchronize_srcu
);
1040 * start_poll_synchronize_srcu - Provide cookie and start grace period
1041 * @ssp: srcu_struct to provide cookie for.
1043 * This function returns a cookie that can be passed to
1044 * poll_state_synchronize_srcu(), which will return true if a full grace
1045 * period has elapsed in the meantime. Unlike get_state_synchronize_srcu(),
1046 * this function also ensures that any needed SRCU grace period will be
1047 * started. This convenience does come at a cost in terms of CPU overhead.
1049 unsigned long start_poll_synchronize_srcu(struct srcu_struct
*ssp
)
1051 return srcu_gp_start_if_needed(ssp
, NULL
, true);
1053 EXPORT_SYMBOL_GPL(start_poll_synchronize_srcu
);
1056 * poll_state_synchronize_srcu - Has cookie's grace period ended?
1057 * @ssp: srcu_struct to provide cookie for.
1058 * @cookie: Return value from get_state_synchronize_srcu() or start_poll_synchronize_srcu().
1060 * This function takes the cookie that was returned from either
1061 * get_state_synchronize_srcu() or start_poll_synchronize_srcu(), and
1062 * returns @true if an SRCU grace period elapsed since the time that the
1063 * cookie was created.
1065 bool poll_state_synchronize_srcu(struct srcu_struct
*ssp
, unsigned long cookie
)
1067 if (!rcu_seq_done(&ssp
->srcu_gp_seq
, cookie
))
1069 // Ensure that the end of the SRCU grace period happens before
1070 // any subsequent code that the caller might execute.
1074 EXPORT_SYMBOL_GPL(poll_state_synchronize_srcu
);
1077 * Callback function for srcu_barrier() use.
1079 static void srcu_barrier_cb(struct rcu_head
*rhp
)
1081 struct srcu_data
*sdp
;
1082 struct srcu_struct
*ssp
;
1084 sdp
= container_of(rhp
, struct srcu_data
, srcu_barrier_head
);
1086 if (atomic_dec_and_test(&ssp
->srcu_barrier_cpu_cnt
))
1087 complete(&ssp
->srcu_barrier_completion
);
1091 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
1092 * @ssp: srcu_struct on which to wait for in-flight callbacks.
1094 void srcu_barrier(struct srcu_struct
*ssp
)
1097 struct srcu_data
*sdp
;
1098 unsigned long s
= rcu_seq_snap(&ssp
->srcu_barrier_seq
);
1100 check_init_srcu_struct(ssp
);
1101 mutex_lock(&ssp
->srcu_barrier_mutex
);
1102 if (rcu_seq_done(&ssp
->srcu_barrier_seq
, s
)) {
1103 smp_mb(); /* Force ordering following return. */
1104 mutex_unlock(&ssp
->srcu_barrier_mutex
);
1105 return; /* Someone else did our work for us. */
1107 rcu_seq_start(&ssp
->srcu_barrier_seq
);
1108 init_completion(&ssp
->srcu_barrier_completion
);
1110 /* Initial count prevents reaching zero until all CBs are posted. */
1111 atomic_set(&ssp
->srcu_barrier_cpu_cnt
, 1);
1114 * Each pass through this loop enqueues a callback, but only
1115 * on CPUs already having callbacks enqueued. Note that if
1116 * a CPU already has callbacks enqueue, it must have already
1117 * registered the need for a future grace period, so all we
1118 * need do is enqueue a callback that will use the same
1119 * grace period as the last callback already in the queue.
1121 for_each_possible_cpu(cpu
) {
1122 sdp
= per_cpu_ptr(ssp
->sda
, cpu
);
1123 spin_lock_irq_rcu_node(sdp
);
1124 atomic_inc(&ssp
->srcu_barrier_cpu_cnt
);
1125 sdp
->srcu_barrier_head
.func
= srcu_barrier_cb
;
1126 debug_rcu_head_queue(&sdp
->srcu_barrier_head
);
1127 if (!rcu_segcblist_entrain(&sdp
->srcu_cblist
,
1128 &sdp
->srcu_barrier_head
)) {
1129 debug_rcu_head_unqueue(&sdp
->srcu_barrier_head
);
1130 atomic_dec(&ssp
->srcu_barrier_cpu_cnt
);
1132 spin_unlock_irq_rcu_node(sdp
);
1135 /* Remove the initial count, at which point reaching zero can happen. */
1136 if (atomic_dec_and_test(&ssp
->srcu_barrier_cpu_cnt
))
1137 complete(&ssp
->srcu_barrier_completion
);
1138 wait_for_completion(&ssp
->srcu_barrier_completion
);
1140 rcu_seq_end(&ssp
->srcu_barrier_seq
);
1141 mutex_unlock(&ssp
->srcu_barrier_mutex
);
1143 EXPORT_SYMBOL_GPL(srcu_barrier
);
1146 * srcu_batches_completed - return batches completed.
1147 * @ssp: srcu_struct on which to report batch completion.
1149 * Report the number of batches, correlated with, but not necessarily
1150 * precisely the same as, the number of grace periods that have elapsed.
1152 unsigned long srcu_batches_completed(struct srcu_struct
*ssp
)
1154 return READ_ONCE(ssp
->srcu_idx
);
1156 EXPORT_SYMBOL_GPL(srcu_batches_completed
);
1159 * Core SRCU state machine. Push state bits of ->srcu_gp_seq
1160 * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
1161 * completed in that state.
1163 static void srcu_advance_state(struct srcu_struct
*ssp
)
1167 mutex_lock(&ssp
->srcu_gp_mutex
);
1170 * Because readers might be delayed for an extended period after
1171 * fetching ->srcu_idx for their index, at any point in time there
1172 * might well be readers using both idx=0 and idx=1. We therefore
1173 * need to wait for readers to clear from both index values before
1174 * invoking a callback.
1176 * The load-acquire ensures that we see the accesses performed
1177 * by the prior grace period.
1179 idx
= rcu_seq_state(smp_load_acquire(&ssp
->srcu_gp_seq
)); /* ^^^ */
1180 if (idx
== SRCU_STATE_IDLE
) {
1181 spin_lock_irq_rcu_node(ssp
);
1182 if (ULONG_CMP_GE(ssp
->srcu_gp_seq
, ssp
->srcu_gp_seq_needed
)) {
1183 WARN_ON_ONCE(rcu_seq_state(ssp
->srcu_gp_seq
));
1184 spin_unlock_irq_rcu_node(ssp
);
1185 mutex_unlock(&ssp
->srcu_gp_mutex
);
1188 idx
= rcu_seq_state(READ_ONCE(ssp
->srcu_gp_seq
));
1189 if (idx
== SRCU_STATE_IDLE
)
1191 spin_unlock_irq_rcu_node(ssp
);
1192 if (idx
!= SRCU_STATE_IDLE
) {
1193 mutex_unlock(&ssp
->srcu_gp_mutex
);
1194 return; /* Someone else started the grace period. */
1198 if (rcu_seq_state(READ_ONCE(ssp
->srcu_gp_seq
)) == SRCU_STATE_SCAN1
) {
1199 idx
= 1 ^ (ssp
->srcu_idx
& 1);
1200 if (!try_check_zero(ssp
, idx
, 1)) {
1201 mutex_unlock(&ssp
->srcu_gp_mutex
);
1202 return; /* readers present, retry later. */
1205 spin_lock_irq_rcu_node(ssp
);
1206 rcu_seq_set_state(&ssp
->srcu_gp_seq
, SRCU_STATE_SCAN2
);
1207 spin_unlock_irq_rcu_node(ssp
);
1210 if (rcu_seq_state(READ_ONCE(ssp
->srcu_gp_seq
)) == SRCU_STATE_SCAN2
) {
1213 * SRCU read-side critical sections are normally short,
1214 * so check at least twice in quick succession after a flip.
1216 idx
= 1 ^ (ssp
->srcu_idx
& 1);
1217 if (!try_check_zero(ssp
, idx
, 2)) {
1218 mutex_unlock(&ssp
->srcu_gp_mutex
);
1219 return; /* readers present, retry later. */
1221 srcu_gp_end(ssp
); /* Releases ->srcu_gp_mutex. */
1226 * Invoke a limited number of SRCU callbacks that have passed through
1227 * their grace period. If there are more to do, SRCU will reschedule
1228 * the workqueue. Note that needed memory barriers have been executed
1229 * in this task's context by srcu_readers_active_idx_check().
1231 static void srcu_invoke_callbacks(struct work_struct
*work
)
1234 struct rcu_cblist ready_cbs
;
1235 struct rcu_head
*rhp
;
1236 struct srcu_data
*sdp
;
1237 struct srcu_struct
*ssp
;
1239 sdp
= container_of(work
, struct srcu_data
, work
);
1242 rcu_cblist_init(&ready_cbs
);
1243 spin_lock_irq_rcu_node(sdp
);
1244 rcu_segcblist_advance(&sdp
->srcu_cblist
,
1245 rcu_seq_current(&ssp
->srcu_gp_seq
));
1246 if (sdp
->srcu_cblist_invoking
||
1247 !rcu_segcblist_ready_cbs(&sdp
->srcu_cblist
)) {
1248 spin_unlock_irq_rcu_node(sdp
);
1249 return; /* Someone else on the job or nothing to do. */
1252 /* We are on the job! Extract and invoke ready callbacks. */
1253 sdp
->srcu_cblist_invoking
= true;
1254 rcu_segcblist_extract_done_cbs(&sdp
->srcu_cblist
, &ready_cbs
);
1255 spin_unlock_irq_rcu_node(sdp
);
1256 rhp
= rcu_cblist_dequeue(&ready_cbs
);
1257 for (; rhp
!= NULL
; rhp
= rcu_cblist_dequeue(&ready_cbs
)) {
1258 debug_rcu_head_unqueue(rhp
);
1265 * Update counts, accelerate new callbacks, and if needed,
1266 * schedule another round of callback invocation.
1268 spin_lock_irq_rcu_node(sdp
);
1269 rcu_segcblist_insert_count(&sdp
->srcu_cblist
, &ready_cbs
);
1270 (void)rcu_segcblist_accelerate(&sdp
->srcu_cblist
,
1271 rcu_seq_snap(&ssp
->srcu_gp_seq
));
1272 sdp
->srcu_cblist_invoking
= false;
1273 more
= rcu_segcblist_ready_cbs(&sdp
->srcu_cblist
);
1274 spin_unlock_irq_rcu_node(sdp
);
1276 srcu_schedule_cbs_sdp(sdp
, 0);
1280 * Finished one round of SRCU grace period. Start another if there are
1281 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
1283 static void srcu_reschedule(struct srcu_struct
*ssp
, unsigned long delay
)
1287 spin_lock_irq_rcu_node(ssp
);
1288 if (ULONG_CMP_GE(ssp
->srcu_gp_seq
, ssp
->srcu_gp_seq_needed
)) {
1289 if (!WARN_ON_ONCE(rcu_seq_state(ssp
->srcu_gp_seq
))) {
1290 /* All requests fulfilled, time to go idle. */
1293 } else if (!rcu_seq_state(ssp
->srcu_gp_seq
)) {
1294 /* Outstanding request and no GP. Start one. */
1297 spin_unlock_irq_rcu_node(ssp
);
1300 queue_delayed_work(rcu_gp_wq
, &ssp
->work
, delay
);
1304 * This is the work-queue function that handles SRCU grace periods.
1306 static void process_srcu(struct work_struct
*work
)
1308 struct srcu_struct
*ssp
;
1310 ssp
= container_of(work
, struct srcu_struct
, work
.work
);
1312 srcu_advance_state(ssp
);
1313 srcu_reschedule(ssp
, srcu_get_delay(ssp
));
1316 void srcutorture_get_gp_data(enum rcutorture_type test_type
,
1317 struct srcu_struct
*ssp
, int *flags
,
1318 unsigned long *gp_seq
)
1320 if (test_type
!= SRCU_FLAVOR
)
1323 *gp_seq
= rcu_seq_current(&ssp
->srcu_gp_seq
);
1325 EXPORT_SYMBOL_GPL(srcutorture_get_gp_data
);
1327 void srcu_torture_stats_print(struct srcu_struct
*ssp
, char *tt
, char *tf
)
1331 unsigned long s0
= 0, s1
= 0;
1333 idx
= ssp
->srcu_idx
& 0x1;
1334 pr_alert("%s%s Tree SRCU g%ld per-CPU(idx=%d):",
1335 tt
, tf
, rcu_seq_current(&ssp
->srcu_gp_seq
), idx
);
1336 for_each_possible_cpu(cpu
) {
1337 unsigned long l0
, l1
;
1338 unsigned long u0
, u1
;
1340 struct srcu_data
*sdp
;
1342 sdp
= per_cpu_ptr(ssp
->sda
, cpu
);
1343 u0
= data_race(sdp
->srcu_unlock_count
[!idx
]);
1344 u1
= data_race(sdp
->srcu_unlock_count
[idx
]);
1347 * Make sure that a lock is always counted if the corresponding
1348 * unlock is counted.
1352 l0
= data_race(sdp
->srcu_lock_count
[!idx
]);
1353 l1
= data_race(sdp
->srcu_lock_count
[idx
]);
1357 pr_cont(" %d(%ld,%ld %c)",
1359 "C."[rcu_segcblist_empty(&sdp
->srcu_cblist
)]);
1363 pr_cont(" T(%ld,%ld)\n", s0
, s1
);
1365 EXPORT_SYMBOL_GPL(srcu_torture_stats_print
);
1367 static int __init
srcu_bootup_announce(void)
1369 pr_info("Hierarchical SRCU implementation.\n");
1370 if (exp_holdoff
!= DEFAULT_SRCU_EXP_HOLDOFF
)
1371 pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff
);
1374 early_initcall(srcu_bootup_announce
);
1376 void __init
srcu_init(void)
1378 struct srcu_struct
*ssp
;
1380 srcu_init_done
= true;
1381 while (!list_empty(&srcu_boot_list
)) {
1382 ssp
= list_first_entry(&srcu_boot_list
, struct srcu_struct
,
1384 check_init_srcu_struct(ssp
);
1385 list_del_init(&ssp
->work
.work
.entry
);
1386 queue_work(rcu_gp_wq
, &ssp
->work
.work
);
1390 #ifdef CONFIG_MODULES
1392 /* Initialize any global-scope srcu_struct structures used by this module. */
1393 static int srcu_module_coming(struct module
*mod
)
1396 struct srcu_struct
**sspp
= mod
->srcu_struct_ptrs
;
1399 for (i
= 0; i
< mod
->num_srcu_structs
; i
++) {
1400 ret
= init_srcu_struct(*(sspp
++));
1401 if (WARN_ON_ONCE(ret
))
1407 /* Clean up any global-scope srcu_struct structures used by this module. */
1408 static void srcu_module_going(struct module
*mod
)
1411 struct srcu_struct
**sspp
= mod
->srcu_struct_ptrs
;
1413 for (i
= 0; i
< mod
->num_srcu_structs
; i
++)
1414 cleanup_srcu_struct(*(sspp
++));
1417 /* Handle one module, either coming or going. */
1418 static int srcu_module_notify(struct notifier_block
*self
,
1419 unsigned long val
, void *data
)
1421 struct module
*mod
= data
;
1425 case MODULE_STATE_COMING
:
1426 ret
= srcu_module_coming(mod
);
1428 case MODULE_STATE_GOING
:
1429 srcu_module_going(mod
);
1437 static struct notifier_block srcu_module_nb
= {
1438 .notifier_call
= srcu_module_notify
,
1442 static __init
int init_srcu_module_notifier(void)
1446 ret
= register_module_notifier(&srcu_module_nb
);
1448 pr_warn("Failed to register srcu module notifier\n");
1451 late_initcall(init_srcu_module_notifier
);
1453 #endif /* #ifdef CONFIG_MODULES */