1 // SPDX-License-Identifier: GPL-2.0+
2 /*
3 * Sleepable Read-Copy Update mechanism for mutual exclusion.
4 *
5 * Copyright (C) IBM Corporation, 2006
6 * Copyright (C) Fujitsu, 2012
7 *
8 * Author: Paul McKenney <paulmck@linux.ibm.com>
9 * Lai Jiangshan <laijs@cn.fujitsu.com>
10 *
11 * For detailed explanation of Read-Copy Update mechanism see -
12 * Documentation/RCU/ *.txt
13 *
14 */
15
16 #define pr_fmt(fmt) "rcu: " fmt
17
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>
28
29 #include "rcu.h"
30 #include "rcu_segcblist.h"
31
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);
36
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);
40
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;
44
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);
49
50 /* Wrappers for lock acquisition and release, see raw_spin_lock_rcu_node(). */
51 #define spin_lock_rcu_node(p) \
52 do { \
53 spin_lock(&ACCESS_PRIVATE(p, lock)); \
54 smp_mb__after_unlock_lock(); \
55 } while (0)
56
57 #define spin_unlock_rcu_node(p) spin_unlock(&ACCESS_PRIVATE(p, lock))
58
59 #define spin_lock_irq_rcu_node(p) \
60 do { \
61 spin_lock_irq(&ACCESS_PRIVATE(p, lock)); \
62 smp_mb__after_unlock_lock(); \
63 } while (0)
64
65 #define spin_unlock_irq_rcu_node(p) \
66 spin_unlock_irq(&ACCESS_PRIVATE(p, lock))
67
68 #define spin_lock_irqsave_rcu_node(p, flags) \
69 do { \
70 spin_lock_irqsave(&ACCESS_PRIVATE(p, lock), flags); \
71 smp_mb__after_unlock_lock(); \
72 } while (0)
73
74 #define spin_unlock_irqrestore_rcu_node(p, flags) \
75 spin_unlock_irqrestore(&ACCESS_PRIVATE(p, lock), flags) \
76
77 /*
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[].
82 */
83 static void init_srcu_struct_nodes(struct srcu_struct *ssp, bool is_static)
84 {
85 int cpu;
86 int i;
87 int level = 0;
88 int levelspread[RCU_NUM_LVLS];
89 struct srcu_data *sdp;
90 struct srcu_node *snp;
91 struct srcu_node *snp_first;
92
93 /* Work out the overall tree geometry. */
94 ssp->level[0] = &ssp->node[0];
95 for (i = 1; i < rcu_num_lvls; i++)
96 ssp->level[i] = ssp->level[i - 1] + num_rcu_lvl[i - 1];
97 rcu_init_levelspread(levelspread, num_rcu_lvl);
98
99 /* Each pass through this loop initializes one srcu_node structure. */
100 srcu_for_each_node_breadth_first(ssp, snp) {
101 spin_lock_init(&ACCESS_PRIVATE(snp, lock));
102 WARN_ON_ONCE(ARRAY_SIZE(snp->srcu_have_cbs) !=
103 ARRAY_SIZE(snp->srcu_data_have_cbs));
104 for (i = 0; i < ARRAY_SIZE(snp->srcu_have_cbs); i++) {
105 snp->srcu_have_cbs[i] = 0;
106 snp->srcu_data_have_cbs[i] = 0;
107 }
108 snp->srcu_gp_seq_needed_exp = 0;
109 snp->grplo = -1;
110 snp->grphi = -1;
111 if (snp == &ssp->node[0]) {
112 /* Root node, special case. */
113 snp->srcu_parent = NULL;
114 continue;
115 }
116
117 /* Non-root node. */
118 if (snp == ssp->level[level + 1])
119 level++;
120 snp->srcu_parent = ssp->level[level - 1] +
121 (snp - ssp->level[level]) /
122 levelspread[level - 1];
123 }
124
125 /*
126 * Initialize the per-CPU srcu_data array, which feeds into the
127 * leaves of the srcu_node tree.
128 */
129 WARN_ON_ONCE(ARRAY_SIZE(sdp->srcu_lock_count) !=
130 ARRAY_SIZE(sdp->srcu_unlock_count));
131 level = rcu_num_lvls - 1;
132 snp_first = ssp->level[level];
133 for_each_possible_cpu(cpu) {
134 sdp = per_cpu_ptr(ssp->sda, cpu);
135 spin_lock_init(&ACCESS_PRIVATE(sdp, lock));
136 rcu_segcblist_init(&sdp->srcu_cblist);
137 sdp->srcu_cblist_invoking = false;
138 sdp->srcu_gp_seq_needed = ssp->srcu_gp_seq;
139 sdp->srcu_gp_seq_needed_exp = ssp->srcu_gp_seq;
140 sdp->mynode = &snp_first[cpu / levelspread[level]];
141 for (snp = sdp->mynode; snp != NULL; snp = snp->srcu_parent) {
142 if (snp->grplo < 0)
143 snp->grplo = cpu;
144 snp->grphi = cpu;
145 }
146 sdp->cpu = cpu;
147 INIT_WORK(&sdp->work, srcu_invoke_callbacks);
148 timer_setup(&sdp->delay_work, srcu_delay_timer, 0);
149 sdp->ssp = ssp;
150 sdp->grpmask = 1 << (cpu - sdp->mynode->grplo);
151 if (is_static)
152 continue;
153
154 /* Dynamically allocated, better be no srcu_read_locks()! */
155 for (i = 0; i < ARRAY_SIZE(sdp->srcu_lock_count); i++) {
156 sdp->srcu_lock_count[i] = 0;
157 sdp->srcu_unlock_count[i] = 0;
158 }
159 }
160 }
161
162 /*
163 * Initialize non-compile-time initialized fields, including the
164 * associated srcu_node and srcu_data structures. The is_static
165 * parameter is passed through to init_srcu_struct_nodes(), and
166 * also tells us that ->sda has already been wired up to srcu_data.
167 */
168 static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static)
169 {
170 mutex_init(&ssp->srcu_cb_mutex);
171 mutex_init(&ssp->srcu_gp_mutex);
172 ssp->srcu_idx = 0;
173 ssp->srcu_gp_seq = 0;
174 ssp->srcu_barrier_seq = 0;
175 mutex_init(&ssp->srcu_barrier_mutex);
176 atomic_set(&ssp->srcu_barrier_cpu_cnt, 0);
177 INIT_DELAYED_WORK(&ssp->work, process_srcu);
178 if (!is_static)
179 ssp->sda = alloc_percpu(struct srcu_data);
180 init_srcu_struct_nodes(ssp, is_static);
181 ssp->srcu_gp_seq_needed_exp = 0;
182 ssp->srcu_last_gp_end = ktime_get_mono_fast_ns();
183 smp_store_release(&ssp->srcu_gp_seq_needed, 0); /* Init done. */
184 return ssp->sda ? 0 : -ENOMEM;
185 }
186
187 #ifdef CONFIG_DEBUG_LOCK_ALLOC
188
189 int __init_srcu_struct(struct srcu_struct *ssp, const char *name,
190 struct lock_class_key *key)
191 {
192 /* Don't re-initialize a lock while it is held. */
193 debug_check_no_locks_freed((void *)ssp, sizeof(*ssp));
194 lockdep_init_map(&ssp->dep_map, name, key, 0);
195 spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
196 return init_srcu_struct_fields(ssp, false);
197 }
198 EXPORT_SYMBOL_GPL(__init_srcu_struct);
199
200 #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */
201
202 /**
203 * init_srcu_struct - initialize a sleep-RCU structure
204 * @ssp: structure to initialize.
205 *
206 * Must invoke this on a given srcu_struct before passing that srcu_struct
207 * to any other function. Each srcu_struct represents a separate domain
208 * of SRCU protection.
209 */
210 int init_srcu_struct(struct srcu_struct *ssp)
211 {
212 spin_lock_init(&ACCESS_PRIVATE(ssp, lock));
213 return init_srcu_struct_fields(ssp, false);
214 }
215 EXPORT_SYMBOL_GPL(init_srcu_struct);
216
217 #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */
218
219 /*
220 * First-use initialization of statically allocated srcu_struct
221 * structure. Wiring up the combining tree is more than can be
222 * done with compile-time initialization, so this check is added
223 * to each update-side SRCU primitive. Use ssp->lock, which -is-
224 * compile-time initialized, to resolve races involving multiple
225 * CPUs trying to garner first-use privileges.
226 */
227 static void check_init_srcu_struct(struct srcu_struct *ssp)
228 {
229 unsigned long flags;
230
231 /* The smp_load_acquire() pairs with the smp_store_release(). */
232 if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq_needed))) /*^^^*/
233 return; /* Already initialized. */
234 spin_lock_irqsave_rcu_node(ssp, flags);
235 if (!rcu_seq_state(ssp->srcu_gp_seq_needed)) {
236 spin_unlock_irqrestore_rcu_node(ssp, flags);
237 return;
238 }
239 init_srcu_struct_fields(ssp, true);
240 spin_unlock_irqrestore_rcu_node(ssp, flags);
241 }
242
243 /*
244 * Returns approximate total of the readers' ->srcu_lock_count[] values
245 * for the rank of per-CPU counters specified by idx.
246 */
247 static unsigned long srcu_readers_lock_idx(struct srcu_struct *ssp, int idx)
248 {
249 int cpu;
250 unsigned long sum = 0;
251
252 for_each_possible_cpu(cpu) {
253 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
254
255 sum += READ_ONCE(cpuc->srcu_lock_count[idx]);
256 }
257 return sum;
258 }
259
260 /*
261 * Returns approximate total of the readers' ->srcu_unlock_count[] values
262 * for the rank of per-CPU counters specified by idx.
263 */
264 static unsigned long srcu_readers_unlock_idx(struct srcu_struct *ssp, int idx)
265 {
266 int cpu;
267 unsigned long sum = 0;
268
269 for_each_possible_cpu(cpu) {
270 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
271
272 sum += READ_ONCE(cpuc->srcu_unlock_count[idx]);
273 }
274 return sum;
275 }
276
277 /*
278 * Return true if the number of pre-existing readers is determined to
279 * be zero.
280 */
281 static bool srcu_readers_active_idx_check(struct srcu_struct *ssp, int idx)
282 {
283 unsigned long unlocks;
284
285 unlocks = srcu_readers_unlock_idx(ssp, idx);
286
287 /*
288 * Make sure that a lock is always counted if the corresponding
289 * unlock is counted. Needs to be a smp_mb() as the read side may
290 * contain a read from a variable that is written to before the
291 * synchronize_srcu() in the write side. In this case smp_mb()s
292 * A and B act like the store buffering pattern.
293 *
294 * This smp_mb() also pairs with smp_mb() C to prevent accesses
295 * after the synchronize_srcu() from being executed before the
296 * grace period ends.
297 */
298 smp_mb(); /* A */
299
300 /*
301 * If the locks are the same as the unlocks, then there must have
302 * been no readers on this index at some time in between. This does
303 * not mean that there are no more readers, as one could have read
304 * the current index but not have incremented the lock counter yet.
305 *
306 * So suppose that the updater is preempted here for so long
307 * that more than ULONG_MAX non-nested readers come and go in
308 * the meantime. It turns out that this cannot result in overflow
309 * because if a reader modifies its unlock count after we read it
310 * above, then that reader's next load of ->srcu_idx is guaranteed
311 * to get the new value, which will cause it to operate on the
312 * other bank of counters, where it cannot contribute to the
313 * overflow of these counters. This means that there is a maximum
314 * of 2*NR_CPUS increments, which cannot overflow given current
315 * systems, especially not on 64-bit systems.
316 *
317 * OK, how about nesting? This does impose a limit on nesting
318 * of floor(ULONG_MAX/NR_CPUS/2), which should be sufficient,
319 * especially on 64-bit systems.
320 */
321 return srcu_readers_lock_idx(ssp, idx) == unlocks;
322 }
323
324 /**
325 * srcu_readers_active - returns true if there are readers. and false
326 * otherwise
327 * @ssp: which srcu_struct to count active readers (holding srcu_read_lock).
328 *
329 * Note that this is not an atomic primitive, and can therefore suffer
330 * severe errors when invoked on an active srcu_struct. That said, it
331 * can be useful as an error check at cleanup time.
332 */
333 static bool srcu_readers_active(struct srcu_struct *ssp)
334 {
335 int cpu;
336 unsigned long sum = 0;
337
338 for_each_possible_cpu(cpu) {
339 struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu);
340
341 sum += READ_ONCE(cpuc->srcu_lock_count[0]);
342 sum += READ_ONCE(cpuc->srcu_lock_count[1]);
343 sum -= READ_ONCE(cpuc->srcu_unlock_count[0]);
344 sum -= READ_ONCE(cpuc->srcu_unlock_count[1]);
345 }
346 return sum;
347 }
348
349 #define SRCU_INTERVAL 1
350
351 /*
352 * Return grace-period delay, zero if there are expedited grace
353 * periods pending, SRCU_INTERVAL otherwise.
354 */
355 static unsigned long srcu_get_delay(struct srcu_struct *ssp)
356 {
357 if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq),
358 READ_ONCE(ssp->srcu_gp_seq_needed_exp)))
359 return 0;
360 return SRCU_INTERVAL;
361 }
362
363 /**
364 * cleanup_srcu_struct - deconstruct a sleep-RCU structure
365 * @ssp: structure to clean up.
366 *
367 * Must invoke this after you are finished using a given srcu_struct that
368 * was initialized via init_srcu_struct(), else you leak memory.
369 */
370 void cleanup_srcu_struct(struct srcu_struct *ssp)
371 {
372 int cpu;
373
374 if (WARN_ON(!srcu_get_delay(ssp)))
375 return; /* Just leak it! */
376 if (WARN_ON(srcu_readers_active(ssp)))
377 return; /* Just leak it! */
378 flush_delayed_work(&ssp->work);
379 for_each_possible_cpu(cpu) {
380 struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu);
381
382 del_timer_sync(&sdp->delay_work);
383 flush_work(&sdp->work);
384 if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist)))
385 return; /* Forgot srcu_barrier(), so just leak it! */
386 }
387 if (WARN_ON(rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) != SRCU_STATE_IDLE) ||
388 WARN_ON(srcu_readers_active(ssp))) {
389 pr_info("%s: Active srcu_struct %p state: %d\n",
390 __func__, ssp, rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)));
391 return; /* Caller forgot to stop doing call_srcu()? */
392 }
393 free_percpu(ssp->sda);
394 ssp->sda = NULL;
395 }
396 EXPORT_SYMBOL_GPL(cleanup_srcu_struct);
397
398 /*
399 * Counts the new reader in the appropriate per-CPU element of the
400 * srcu_struct.
401 * Returns an index that must be passed to the matching srcu_read_unlock().
402 */
403 int __srcu_read_lock(struct srcu_struct *ssp)
404 {
405 int idx;
406
407 idx = READ_ONCE(ssp->srcu_idx) & 0x1;
408 this_cpu_inc(ssp->sda->srcu_lock_count[idx]);
409 smp_mb(); /* B */ /* Avoid leaking the critical section. */
410 return idx;
411 }
412 EXPORT_SYMBOL_GPL(__srcu_read_lock);
413
414 /*
415 * Removes the count for the old reader from the appropriate per-CPU
416 * element of the srcu_struct. Note that this may well be a different
417 * CPU than that which was incremented by the corresponding srcu_read_lock().
418 */
419 void __srcu_read_unlock(struct srcu_struct *ssp, int idx)
420 {
421 smp_mb(); /* C */ /* Avoid leaking the critical section. */
422 this_cpu_inc(ssp->sda->srcu_unlock_count[idx]);
423 }
424 EXPORT_SYMBOL_GPL(__srcu_read_unlock);
425
426 /*
427 * We use an adaptive strategy for synchronize_srcu() and especially for
428 * synchronize_srcu_expedited(). We spin for a fixed time period
429 * (defined below) to allow SRCU readers to exit their read-side critical
430 * sections. If there are still some readers after a few microseconds,
431 * we repeatedly block for 1-millisecond time periods.
432 */
433 #define SRCU_RETRY_CHECK_DELAY 5
434
435 /*
436 * Start an SRCU grace period.
437 */
438 static void srcu_gp_start(struct srcu_struct *ssp)
439 {
440 struct srcu_data *sdp = this_cpu_ptr(ssp->sda);
441 int state;
442
443 lockdep_assert_held(&ACCESS_PRIVATE(ssp, lock));
444 WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
445 spin_lock_rcu_node(sdp); /* Interrupts already disabled. */
446 rcu_segcblist_advance(&sdp->srcu_cblist,
447 rcu_seq_current(&ssp->srcu_gp_seq));
448 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
449 rcu_seq_snap(&ssp->srcu_gp_seq));
450 spin_unlock_rcu_node(sdp); /* Interrupts remain disabled. */
451 smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */
452 rcu_seq_start(&ssp->srcu_gp_seq);
453 state = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq));
454 WARN_ON_ONCE(state != SRCU_STATE_SCAN1);
455 }
456
457
458 static void srcu_delay_timer(struct timer_list *t)
459 {
460 struct srcu_data *sdp = container_of(t, struct srcu_data, delay_work);
461
462 queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
463 }
464
465 static void srcu_queue_delayed_work_on(struct srcu_data *sdp,
466 unsigned long delay)
467 {
468 if (!delay) {
469 queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work);
470 return;
471 }
472
473 timer_reduce(&sdp->delay_work, jiffies + delay);
474 }
475
476 /*
477 * Schedule callback invocation for the specified srcu_data structure,
478 * if possible, on the corresponding CPU.
479 */
480 static void srcu_schedule_cbs_sdp(struct srcu_data *sdp, unsigned long delay)
481 {
482 srcu_queue_delayed_work_on(sdp, delay);
483 }
484
485 /*
486 * Schedule callback invocation for all srcu_data structures associated
487 * with the specified srcu_node structure that have callbacks for the
488 * just-completed grace period, the one corresponding to idx. If possible,
489 * schedule this invocation on the corresponding CPUs.
490 */
491 static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp,
492 unsigned long mask, unsigned long delay)
493 {
494 int cpu;
495
496 for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
497 if (!(mask & (1 << (cpu - snp->grplo))))
498 continue;
499 srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay);
500 }
501 }
502
503 /*
504 * Note the end of an SRCU grace period. Initiates callback invocation
505 * and starts a new grace period if needed.
506 *
507 * The ->srcu_cb_mutex acquisition does not protect any data, but
508 * instead prevents more than one grace period from starting while we
509 * are initiating callback invocation. This allows the ->srcu_have_cbs[]
510 * array to have a finite number of elements.
511 */
512 static void srcu_gp_end(struct srcu_struct *ssp)
513 {
514 unsigned long cbdelay;
515 bool cbs;
516 bool last_lvl;
517 int cpu;
518 unsigned long flags;
519 unsigned long gpseq;
520 int idx;
521 unsigned long mask;
522 struct srcu_data *sdp;
523 struct srcu_node *snp;
524
525 /* Prevent more than one additional grace period. */
526 mutex_lock(&ssp->srcu_cb_mutex);
527
528 /* End the current grace period. */
529 spin_lock_irq_rcu_node(ssp);
530 idx = rcu_seq_state(ssp->srcu_gp_seq);
531 WARN_ON_ONCE(idx != SRCU_STATE_SCAN2);
532 cbdelay = srcu_get_delay(ssp);
533 WRITE_ONCE(ssp->srcu_last_gp_end, ktime_get_mono_fast_ns());
534 rcu_seq_end(&ssp->srcu_gp_seq);
535 gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
536 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, gpseq))
537 ssp->srcu_gp_seq_needed_exp = gpseq;
538 spin_unlock_irq_rcu_node(ssp);
539 mutex_unlock(&ssp->srcu_gp_mutex);
540 /* A new grace period can start at this point. But only one. */
541
542 /* Initiate callback invocation as needed. */
543 idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs);
544 srcu_for_each_node_breadth_first(ssp, snp) {
545 spin_lock_irq_rcu_node(snp);
546 cbs = false;
547 last_lvl = snp >= ssp->level[rcu_num_lvls - 1];
548 if (last_lvl)
549 cbs = snp->srcu_have_cbs[idx] == gpseq;
550 snp->srcu_have_cbs[idx] = gpseq;
551 rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1);
552 if (ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, gpseq))
553 snp->srcu_gp_seq_needed_exp = gpseq;
554 mask = snp->srcu_data_have_cbs[idx];
555 snp->srcu_data_have_cbs[idx] = 0;
556 spin_unlock_irq_rcu_node(snp);
557 if (cbs)
558 srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay);
559
560 /* Occasionally prevent srcu_data counter wrap. */
561 if (!(gpseq & counter_wrap_check) && last_lvl)
562 for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) {
563 sdp = per_cpu_ptr(ssp->sda, cpu);
564 spin_lock_irqsave_rcu_node(sdp, flags);
565 if (ULONG_CMP_GE(gpseq,
566 sdp->srcu_gp_seq_needed + 100))
567 sdp->srcu_gp_seq_needed = gpseq;
568 if (ULONG_CMP_GE(gpseq,
569 sdp->srcu_gp_seq_needed_exp + 100))
570 sdp->srcu_gp_seq_needed_exp = gpseq;
571 spin_unlock_irqrestore_rcu_node(sdp, flags);
572 }
573 }
574
575 /* Callback initiation done, allow grace periods after next. */
576 mutex_unlock(&ssp->srcu_cb_mutex);
577
578 /* Start a new grace period if needed. */
579 spin_lock_irq_rcu_node(ssp);
580 gpseq = rcu_seq_current(&ssp->srcu_gp_seq);
581 if (!rcu_seq_state(gpseq) &&
582 ULONG_CMP_LT(gpseq, ssp->srcu_gp_seq_needed)) {
583 srcu_gp_start(ssp);
584 spin_unlock_irq_rcu_node(ssp);
585 srcu_reschedule(ssp, 0);
586 } else {
587 spin_unlock_irq_rcu_node(ssp);
588 }
589 }
590
591 /*
592 * Funnel-locking scheme to scalably mediate many concurrent expedited
593 * grace-period requests. This function is invoked for the first known
594 * expedited request for a grace period that has already been requested,
595 * but without expediting. To start a completely new grace period,
596 * whether expedited or not, use srcu_funnel_gp_start() instead.
597 */
598 static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp,
599 unsigned long s)
600 {
601 unsigned long flags;
602
603 for (; snp != NULL; snp = snp->srcu_parent) {
604 if (rcu_seq_done(&ssp->srcu_gp_seq, s) ||
605 ULONG_CMP_GE(READ_ONCE(snp->srcu_gp_seq_needed_exp), s))
606 return;
607 spin_lock_irqsave_rcu_node(snp, flags);
608 if (ULONG_CMP_GE(snp->srcu_gp_seq_needed_exp, s)) {
609 spin_unlock_irqrestore_rcu_node(snp, flags);
610 return;
611 }
612 WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s);
613 spin_unlock_irqrestore_rcu_node(snp, flags);
614 }
615 spin_lock_irqsave_rcu_node(ssp, flags);
616 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
617 ssp->srcu_gp_seq_needed_exp = s;
618 spin_unlock_irqrestore_rcu_node(ssp, flags);
619 }
620
621 /*
622 * Funnel-locking scheme to scalably mediate many concurrent grace-period
623 * requests. The winner has to do the work of actually starting grace
624 * period s. Losers must either ensure that their desired grace-period
625 * number is recorded on at least their leaf srcu_node structure, or they
626 * must take steps to invoke their own callbacks.
627 *
628 * Note that this function also does the work of srcu_funnel_exp_start(),
629 * in some cases by directly invoking it.
630 */
631 static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp,
632 unsigned long s, bool do_norm)
633 {
634 unsigned long flags;
635 int idx = rcu_seq_ctr(s) % ARRAY_SIZE(sdp->mynode->srcu_have_cbs);
636 struct srcu_node *snp = sdp->mynode;
637 unsigned long snp_seq;
638
639 /* Each pass through the loop does one level of the srcu_node tree. */
640 for (; snp != NULL; snp = snp->srcu_parent) {
641 if (rcu_seq_done(&ssp->srcu_gp_seq, s) && snp != sdp->mynode)
642 return; /* GP already done and CBs recorded. */
643 spin_lock_irqsave_rcu_node(snp, flags);
644 if (ULONG_CMP_GE(snp->srcu_have_cbs[idx], s)) {
645 snp_seq = snp->srcu_have_cbs[idx];
646 if (snp == sdp->mynode && snp_seq == s)
647 snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
648 spin_unlock_irqrestore_rcu_node(snp, flags);
649 if (snp == sdp->mynode && snp_seq != s) {
650 srcu_schedule_cbs_sdp(sdp, do_norm
651 ? SRCU_INTERVAL
652 : 0);
653 return;
654 }
655 if (!do_norm)
656 srcu_funnel_exp_start(ssp, snp, s);
657 return;
658 }
659 snp->srcu_have_cbs[idx] = s;
660 if (snp == sdp->mynode)
661 snp->srcu_data_have_cbs[idx] |= sdp->grpmask;
662 if (!do_norm && ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, s))
663 snp->srcu_gp_seq_needed_exp = s;
664 spin_unlock_irqrestore_rcu_node(snp, flags);
665 }
666
667 /* Top of tree, must ensure the grace period will be started. */
668 spin_lock_irqsave_rcu_node(ssp, flags);
669 if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed, s)) {
670 /*
671 * Record need for grace period s. Pair with load
672 * acquire setting up for initialization.
673 */
674 smp_store_release(&ssp->srcu_gp_seq_needed, s); /*^^^*/
675 }
676 if (!do_norm && ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s))
677 ssp->srcu_gp_seq_needed_exp = s;
678
679 /* If grace period not already done and none in progress, start it. */
680 if (!rcu_seq_done(&ssp->srcu_gp_seq, s) &&
681 rcu_seq_state(ssp->srcu_gp_seq) == SRCU_STATE_IDLE) {
682 WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed));
683 srcu_gp_start(ssp);
684 if (likely(srcu_init_done))
685 queue_delayed_work(rcu_gp_wq, &ssp->work,
686 srcu_get_delay(ssp));
687 else if (list_empty(&ssp->work.work.entry))
688 list_add(&ssp->work.work.entry, &srcu_boot_list);
689 }
690 spin_unlock_irqrestore_rcu_node(ssp, flags);
691 }
692
693 /*
694 * Wait until all readers counted by array index idx complete, but
695 * loop an additional time if there is an expedited grace period pending.
696 * The caller must ensure that ->srcu_idx is not changed while checking.
697 */
698 static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount)
699 {
700 for (;;) {
701 if (srcu_readers_active_idx_check(ssp, idx))
702 return true;
703 if (--trycount + !srcu_get_delay(ssp) <= 0)
704 return false;
705 udelay(SRCU_RETRY_CHECK_DELAY);
706 }
707 }
708
709 /*
710 * Increment the ->srcu_idx counter so that future SRCU readers will
711 * use the other rank of the ->srcu_(un)lock_count[] arrays. This allows
712 * us to wait for pre-existing readers in a starvation-free manner.
713 */
714 static void srcu_flip(struct srcu_struct *ssp)
715 {
716 /*
717 * Ensure that if this updater saw a given reader's increment
718 * from __srcu_read_lock(), that reader was using an old value
719 * of ->srcu_idx. Also ensure that if a given reader sees the
720 * new value of ->srcu_idx, this updater's earlier scans cannot
721 * have seen that reader's increments (which is OK, because this
722 * grace period need not wait on that reader).
723 */
724 smp_mb(); /* E */ /* Pairs with B and C. */
725
726 WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1);
727
728 /*
729 * Ensure that if the updater misses an __srcu_read_unlock()
730 * increment, that task's next __srcu_read_lock() will see the
731 * above counter update. Note that both this memory barrier
732 * and the one in srcu_readers_active_idx_check() provide the
733 * guarantee for __srcu_read_lock().
734 */
735 smp_mb(); /* D */ /* Pairs with C. */
736 }
737
738 /*
739 * If SRCU is likely idle, return true, otherwise return false.
740 *
741 * Note that it is OK for several current from-idle requests for a new
742 * grace period from idle to specify expediting because they will all end
743 * up requesting the same grace period anyhow. So no loss.
744 *
745 * Note also that if any CPU (including the current one) is still invoking
746 * callbacks, this function will nevertheless say "idle". This is not
747 * ideal, but the overhead of checking all CPUs' callback lists is even
748 * less ideal, especially on large systems. Furthermore, the wakeup
749 * can happen before the callback is fully removed, so we have no choice
750 * but to accept this type of error.
751 *
752 * This function is also subject to counter-wrap errors, but let's face
753 * it, if this function was preempted for enough time for the counters
754 * to wrap, it really doesn't matter whether or not we expedite the grace
755 * period. The extra overhead of a needlessly expedited grace period is
756 * negligible when amoritized over that time period, and the extra latency
757 * of a needlessly non-expedited grace period is similarly negligible.
758 */
759 static bool srcu_might_be_idle(struct srcu_struct *ssp)
760 {
761 unsigned long curseq;
762 unsigned long flags;
763 struct srcu_data *sdp;
764 unsigned long t;
765 unsigned long tlast;
766
767 /* If the local srcu_data structure has callbacks, not idle. */
768 local_irq_save(flags);
769 sdp = this_cpu_ptr(ssp->sda);
770 if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) {
771 local_irq_restore(flags);
772 return false; /* Callbacks already present, so not idle. */
773 }
774 local_irq_restore(flags);
775
776 /*
777 * No local callbacks, so probabalistically probe global state.
778 * Exact information would require acquiring locks, which would
779 * kill scalability, hence the probabalistic nature of the probe.
780 */
781
782 /* First, see if enough time has passed since the last GP. */
783 t = ktime_get_mono_fast_ns();
784 tlast = READ_ONCE(ssp->srcu_last_gp_end);
785 if (exp_holdoff == 0 ||
786 time_in_range_open(t, tlast, tlast + exp_holdoff))
787 return false; /* Too soon after last GP. */
788
789 /* Next, check for probable idleness. */
790 curseq = rcu_seq_current(&ssp->srcu_gp_seq);
791 smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */
792 if (ULONG_CMP_LT(curseq, READ_ONCE(ssp->srcu_gp_seq_needed)))
793 return false; /* Grace period in progress, so not idle. */
794 smp_mb(); /* Order ->srcu_gp_seq with prior access. */
795 if (curseq != rcu_seq_current(&ssp->srcu_gp_seq))
796 return false; /* GP # changed, so not idle. */
797 return true; /* With reasonable probability, idle! */
798 }
799
800 /*
801 * SRCU callback function to leak a callback.
802 */
803 static void srcu_leak_callback(struct rcu_head *rhp)
804 {
805 }
806
807 /*
808 * Enqueue an SRCU callback on the srcu_data structure associated with
809 * the current CPU and the specified srcu_struct structure, initiating
810 * grace-period processing if it is not already running.
811 *
812 * Note that all CPUs must agree that the grace period extended beyond
813 * all pre-existing SRCU read-side critical section. On systems with
814 * more than one CPU, this means that when "func()" is invoked, each CPU
815 * is guaranteed to have executed a full memory barrier since the end of
816 * its last corresponding SRCU read-side critical section whose beginning
817 * preceded the call to call_srcu(). It also means that each CPU executing
818 * an SRCU read-side critical section that continues beyond the start of
819 * "func()" must have executed a memory barrier after the call_srcu()
820 * but before the beginning of that SRCU read-side critical section.
821 * Note that these guarantees include CPUs that are offline, idle, or
822 * executing in user mode, as well as CPUs that are executing in the kernel.
823 *
824 * Furthermore, if CPU A invoked call_srcu() and CPU B invoked the
825 * resulting SRCU callback function "func()", then both CPU A and CPU
826 * B are guaranteed to execute a full memory barrier during the time
827 * interval between the call to call_srcu() and the invocation of "func()".
828 * This guarantee applies even if CPU A and CPU B are the same CPU (but
829 * again only if the system has more than one CPU).
830 *
831 * Of course, these guarantees apply only for invocations of call_srcu(),
832 * srcu_read_lock(), and srcu_read_unlock() that are all passed the same
833 * srcu_struct structure.
834 */
835 static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
836 rcu_callback_t func, bool do_norm)
837 {
838 unsigned long flags;
839 int idx;
840 bool needexp = false;
841 bool needgp = false;
842 unsigned long s;
843 struct srcu_data *sdp;
844
845 check_init_srcu_struct(ssp);
846 if (debug_rcu_head_queue(rhp)) {
847 /* Probable double call_srcu(), so leak the callback. */
848 WRITE_ONCE(rhp->func, srcu_leak_callback);
849 WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n");
850 return;
851 }
852 rhp->func = func;
853 idx = srcu_read_lock(ssp);
854 local_irq_save(flags);
855 sdp = this_cpu_ptr(ssp->sda);
856 spin_lock_rcu_node(sdp);
857 rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp, false);
858 rcu_segcblist_advance(&sdp->srcu_cblist,
859 rcu_seq_current(&ssp->srcu_gp_seq));
860 s = rcu_seq_snap(&ssp->srcu_gp_seq);
861 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, s);
862 if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) {
863 sdp->srcu_gp_seq_needed = s;
864 needgp = true;
865 }
866 if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) {
867 sdp->srcu_gp_seq_needed_exp = s;
868 needexp = true;
869 }
870 spin_unlock_irqrestore_rcu_node(sdp, flags);
871 if (needgp)
872 srcu_funnel_gp_start(ssp, sdp, s, do_norm);
873 else if (needexp)
874 srcu_funnel_exp_start(ssp, sdp->mynode, s);
875 srcu_read_unlock(ssp, idx);
876 }
877
878 /**
879 * call_srcu() - Queue a callback for invocation after an SRCU grace period
880 * @ssp: srcu_struct in queue the callback
881 * @rhp: structure to be used for queueing the SRCU callback.
882 * @func: function to be invoked after the SRCU grace period
883 *
884 * The callback function will be invoked some time after a full SRCU
885 * grace period elapses, in other words after all pre-existing SRCU
886 * read-side critical sections have completed. However, the callback
887 * function might well execute concurrently with other SRCU read-side
888 * critical sections that started after call_srcu() was invoked. SRCU
889 * read-side critical sections are delimited by srcu_read_lock() and
890 * srcu_read_unlock(), and may be nested.
891 *
892 * The callback will be invoked from process context, but must nevertheless
893 * be fast and must not block.
894 */
895 void call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp,
896 rcu_callback_t func)
897 {
898 __call_srcu(ssp, rhp, func, true);
899 }
900 EXPORT_SYMBOL_GPL(call_srcu);
901
902 /*
903 * Helper function for synchronize_srcu() and synchronize_srcu_expedited().
904 */
905 static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm)
906 {
907 struct rcu_synchronize rcu;
908
909 RCU_LOCKDEP_WARN(lock_is_held(&ssp->dep_map) ||
910 lock_is_held(&rcu_bh_lock_map) ||
911 lock_is_held(&rcu_lock_map) ||
912 lock_is_held(&rcu_sched_lock_map),
913 "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section");
914
915 if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE)
916 return;
917 might_sleep();
918 check_init_srcu_struct(ssp);
919 init_completion(&rcu.completion);
920 init_rcu_head_on_stack(&rcu.head);
921 __call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm);
922 wait_for_completion(&rcu.completion);
923 destroy_rcu_head_on_stack(&rcu.head);
924
925 /*
926 * Make sure that later code is ordered after the SRCU grace
927 * period. This pairs with the spin_lock_irq_rcu_node()
928 * in srcu_invoke_callbacks(). Unlike Tree RCU, this is needed
929 * because the current CPU might have been totally uninvolved with
930 * (and thus unordered against) that grace period.
931 */
932 smp_mb();
933 }
934
935 /**
936 * synchronize_srcu_expedited - Brute-force SRCU grace period
937 * @ssp: srcu_struct with which to synchronize.
938 *
939 * Wait for an SRCU grace period to elapse, but be more aggressive about
940 * spinning rather than blocking when waiting.
941 *
942 * Note that synchronize_srcu_expedited() has the same deadlock and
943 * memory-ordering properties as does synchronize_srcu().
944 */
945 void synchronize_srcu_expedited(struct srcu_struct *ssp)
946 {
947 __synchronize_srcu(ssp, rcu_gp_is_normal());
948 }
949 EXPORT_SYMBOL_GPL(synchronize_srcu_expedited);
950
951 /**
952 * synchronize_srcu - wait for prior SRCU read-side critical-section completion
953 * @ssp: srcu_struct with which to synchronize.
954 *
955 * Wait for the count to drain to zero of both indexes. To avoid the
956 * possible starvation of synchronize_srcu(), it waits for the count of
957 * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first,
958 * and then flip the srcu_idx and wait for the count of the other index.
959 *
960 * Can block; must be called from process context.
961 *
962 * Note that it is illegal to call synchronize_srcu() from the corresponding
963 * SRCU read-side critical section; doing so will result in deadlock.
964 * However, it is perfectly legal to call synchronize_srcu() on one
965 * srcu_struct from some other srcu_struct's read-side critical section,
966 * as long as the resulting graph of srcu_structs is acyclic.
967 *
968 * There are memory-ordering constraints implied by synchronize_srcu().
969 * On systems with more than one CPU, when synchronize_srcu() returns,
970 * each CPU is guaranteed to have executed a full memory barrier since
971 * the end of its last corresponding SRCU read-side critical section
972 * whose beginning preceded the call to synchronize_srcu(). In addition,
973 * each CPU having an SRCU read-side critical section that extends beyond
974 * the return from synchronize_srcu() is guaranteed to have executed a
975 * full memory barrier after the beginning of synchronize_srcu() and before
976 * the beginning of that SRCU read-side critical section. Note that these
977 * guarantees include CPUs that are offline, idle, or executing in user mode,
978 * as well as CPUs that are executing in the kernel.
979 *
980 * Furthermore, if CPU A invoked synchronize_srcu(), which returned
981 * to its caller on CPU B, then both CPU A and CPU B are guaranteed
982 * to have executed a full memory barrier during the execution of
983 * synchronize_srcu(). This guarantee applies even if CPU A and CPU B
984 * are the same CPU, but again only if the system has more than one CPU.
985 *
986 * Of course, these memory-ordering guarantees apply only when
987 * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are
988 * passed the same srcu_struct structure.
989 *
990 * If SRCU is likely idle, expedite the first request. This semantic
991 * was provided by Classic SRCU, and is relied upon by its users, so TREE
992 * SRCU must also provide it. Note that detecting idleness is heuristic
993 * and subject to both false positives and negatives.
994 */
995 void synchronize_srcu(struct srcu_struct *ssp)
996 {
997 if (srcu_might_be_idle(ssp) || rcu_gp_is_expedited())
998 synchronize_srcu_expedited(ssp);
999 else
1000 __synchronize_srcu(ssp, true);
1001 }
1002 EXPORT_SYMBOL_GPL(synchronize_srcu);
1003
1004 /*
1005 * Callback function for srcu_barrier() use.
1006 */
1007 static void srcu_barrier_cb(struct rcu_head *rhp)
1008 {
1009 struct srcu_data *sdp;
1010 struct srcu_struct *ssp;
1011
1012 sdp = container_of(rhp, struct srcu_data, srcu_barrier_head);
1013 ssp = sdp->ssp;
1014 if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
1015 complete(&ssp->srcu_barrier_completion);
1016 }
1017
1018 /**
1019 * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete.
1020 * @ssp: srcu_struct on which to wait for in-flight callbacks.
1021 */
1022 void srcu_barrier(struct srcu_struct *ssp)
1023 {
1024 int cpu;
1025 struct srcu_data *sdp;
1026 unsigned long s = rcu_seq_snap(&ssp->srcu_barrier_seq);
1027
1028 check_init_srcu_struct(ssp);
1029 mutex_lock(&ssp->srcu_barrier_mutex);
1030 if (rcu_seq_done(&ssp->srcu_barrier_seq, s)) {
1031 smp_mb(); /* Force ordering following return. */
1032 mutex_unlock(&ssp->srcu_barrier_mutex);
1033 return; /* Someone else did our work for us. */
1034 }
1035 rcu_seq_start(&ssp->srcu_barrier_seq);
1036 init_completion(&ssp->srcu_barrier_completion);
1037
1038 /* Initial count prevents reaching zero until all CBs are posted. */
1039 atomic_set(&ssp->srcu_barrier_cpu_cnt, 1);
1040
1041 /*
1042 * Each pass through this loop enqueues a callback, but only
1043 * on CPUs already having callbacks enqueued. Note that if
1044 * a CPU already has callbacks enqueue, it must have already
1045 * registered the need for a future grace period, so all we
1046 * need do is enqueue a callback that will use the same
1047 * grace period as the last callback already in the queue.
1048 */
1049 for_each_possible_cpu(cpu) {
1050 sdp = per_cpu_ptr(ssp->sda, cpu);
1051 spin_lock_irq_rcu_node(sdp);
1052 atomic_inc(&ssp->srcu_barrier_cpu_cnt);
1053 sdp->srcu_barrier_head.func = srcu_barrier_cb;
1054 debug_rcu_head_queue(&sdp->srcu_barrier_head);
1055 if (!rcu_segcblist_entrain(&sdp->srcu_cblist,
1056 &sdp->srcu_barrier_head, 0)) {
1057 debug_rcu_head_unqueue(&sdp->srcu_barrier_head);
1058 atomic_dec(&ssp->srcu_barrier_cpu_cnt);
1059 }
1060 spin_unlock_irq_rcu_node(sdp);
1061 }
1062
1063 /* Remove the initial count, at which point reaching zero can happen. */
1064 if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt))
1065 complete(&ssp->srcu_barrier_completion);
1066 wait_for_completion(&ssp->srcu_barrier_completion);
1067
1068 rcu_seq_end(&ssp->srcu_barrier_seq);
1069 mutex_unlock(&ssp->srcu_barrier_mutex);
1070 }
1071 EXPORT_SYMBOL_GPL(srcu_barrier);
1072
1073 /**
1074 * srcu_batches_completed - return batches completed.
1075 * @ssp: srcu_struct on which to report batch completion.
1076 *
1077 * Report the number of batches, correlated with, but not necessarily
1078 * precisely the same as, the number of grace periods that have elapsed.
1079 */
1080 unsigned long srcu_batches_completed(struct srcu_struct *ssp)
1081 {
1082 return ssp->srcu_idx;
1083 }
1084 EXPORT_SYMBOL_GPL(srcu_batches_completed);
1085
1086 /*
1087 * Core SRCU state machine. Push state bits of ->srcu_gp_seq
1088 * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has
1089 * completed in that state.
1090 */
1091 static void srcu_advance_state(struct srcu_struct *ssp)
1092 {
1093 int idx;
1094
1095 mutex_lock(&ssp->srcu_gp_mutex);
1096
1097 /*
1098 * Because readers might be delayed for an extended period after
1099 * fetching ->srcu_idx for their index, at any point in time there
1100 * might well be readers using both idx=0 and idx=1. We therefore
1101 * need to wait for readers to clear from both index values before
1102 * invoking a callback.
1103 *
1104 * The load-acquire ensures that we see the accesses performed
1105 * by the prior grace period.
1106 */
1107 idx = rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq)); /* ^^^ */
1108 if (idx == SRCU_STATE_IDLE) {
1109 spin_lock_irq_rcu_node(ssp);
1110 if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
1111 WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq));
1112 spin_unlock_irq_rcu_node(ssp);
1113 mutex_unlock(&ssp->srcu_gp_mutex);
1114 return;
1115 }
1116 idx = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq));
1117 if (idx == SRCU_STATE_IDLE)
1118 srcu_gp_start(ssp);
1119 spin_unlock_irq_rcu_node(ssp);
1120 if (idx != SRCU_STATE_IDLE) {
1121 mutex_unlock(&ssp->srcu_gp_mutex);
1122 return; /* Someone else started the grace period. */
1123 }
1124 }
1125
1126 if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN1) {
1127 idx = 1 ^ (ssp->srcu_idx & 1);
1128 if (!try_check_zero(ssp, idx, 1)) {
1129 mutex_unlock(&ssp->srcu_gp_mutex);
1130 return; /* readers present, retry later. */
1131 }
1132 srcu_flip(ssp);
1133 rcu_seq_set_state(&ssp->srcu_gp_seq, SRCU_STATE_SCAN2);
1134 }
1135
1136 if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN2) {
1137
1138 /*
1139 * SRCU read-side critical sections are normally short,
1140 * so check at least twice in quick succession after a flip.
1141 */
1142 idx = 1 ^ (ssp->srcu_idx & 1);
1143 if (!try_check_zero(ssp, idx, 2)) {
1144 mutex_unlock(&ssp->srcu_gp_mutex);
1145 return; /* readers present, retry later. */
1146 }
1147 srcu_gp_end(ssp); /* Releases ->srcu_gp_mutex. */
1148 }
1149 }
1150
1151 /*
1152 * Invoke a limited number of SRCU callbacks that have passed through
1153 * their grace period. If there are more to do, SRCU will reschedule
1154 * the workqueue. Note that needed memory barriers have been executed
1155 * in this task's context by srcu_readers_active_idx_check().
1156 */
1157 static void srcu_invoke_callbacks(struct work_struct *work)
1158 {
1159 bool more;
1160 struct rcu_cblist ready_cbs;
1161 struct rcu_head *rhp;
1162 struct srcu_data *sdp;
1163 struct srcu_struct *ssp;
1164
1165 sdp = container_of(work, struct srcu_data, work);
1166
1167 ssp = sdp->ssp;
1168 rcu_cblist_init(&ready_cbs);
1169 spin_lock_irq_rcu_node(sdp);
1170 rcu_segcblist_advance(&sdp->srcu_cblist,
1171 rcu_seq_current(&ssp->srcu_gp_seq));
1172 if (sdp->srcu_cblist_invoking ||
1173 !rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) {
1174 spin_unlock_irq_rcu_node(sdp);
1175 return; /* Someone else on the job or nothing to do. */
1176 }
1177
1178 /* We are on the job! Extract and invoke ready callbacks. */
1179 sdp->srcu_cblist_invoking = true;
1180 rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs);
1181 spin_unlock_irq_rcu_node(sdp);
1182 rhp = rcu_cblist_dequeue(&ready_cbs);
1183 for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) {
1184 debug_rcu_head_unqueue(rhp);
1185 local_bh_disable();
1186 rhp->func(rhp);
1187 local_bh_enable();
1188 }
1189
1190 /*
1191 * Update counts, accelerate new callbacks, and if needed,
1192 * schedule another round of callback invocation.
1193 */
1194 spin_lock_irq_rcu_node(sdp);
1195 rcu_segcblist_insert_count(&sdp->srcu_cblist, &ready_cbs);
1196 (void)rcu_segcblist_accelerate(&sdp->srcu_cblist,
1197 rcu_seq_snap(&ssp->srcu_gp_seq));
1198 sdp->srcu_cblist_invoking = false;
1199 more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist);
1200 spin_unlock_irq_rcu_node(sdp);
1201 if (more)
1202 srcu_schedule_cbs_sdp(sdp, 0);
1203 }
1204
1205 /*
1206 * Finished one round of SRCU grace period. Start another if there are
1207 * more SRCU callbacks queued, otherwise put SRCU into not-running state.
1208 */
1209 static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay)
1210 {
1211 bool pushgp = true;
1212
1213 spin_lock_irq_rcu_node(ssp);
1214 if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) {
1215 if (!WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq))) {
1216 /* All requests fulfilled, time to go idle. */
1217 pushgp = false;
1218 }
1219 } else if (!rcu_seq_state(ssp->srcu_gp_seq)) {
1220 /* Outstanding request and no GP. Start one. */
1221 srcu_gp_start(ssp);
1222 }
1223 spin_unlock_irq_rcu_node(ssp);
1224
1225 if (pushgp)
1226 queue_delayed_work(rcu_gp_wq, &ssp->work, delay);
1227 }
1228
1229 /*
1230 * This is the work-queue function that handles SRCU grace periods.
1231 */
1232 static void process_srcu(struct work_struct *work)
1233 {
1234 struct srcu_struct *ssp;
1235
1236 ssp = container_of(work, struct srcu_struct, work.work);
1237
1238 srcu_advance_state(ssp);
1239 srcu_reschedule(ssp, srcu_get_delay(ssp));
1240 }
1241
1242 void srcutorture_get_gp_data(enum rcutorture_type test_type,
1243 struct srcu_struct *ssp, int *flags,
1244 unsigned long *gp_seq)
1245 {
1246 if (test_type != SRCU_FLAVOR)
1247 return;
1248 *flags = 0;
1249 *gp_seq = rcu_seq_current(&ssp->srcu_gp_seq);
1250 }
1251 EXPORT_SYMBOL_GPL(srcutorture_get_gp_data);
1252
1253 void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf)
1254 {
1255 int cpu;
1256 int idx;
1257 unsigned long s0 = 0, s1 = 0;
1258
1259 idx = ssp->srcu_idx & 0x1;
1260 pr_alert("%s%s Tree SRCU g%ld per-CPU(idx=%d):",
1261 tt, tf, rcu_seq_current(&ssp->srcu_gp_seq), idx);
1262 for_each_possible_cpu(cpu) {
1263 unsigned long l0, l1;
1264 unsigned long u0, u1;
1265 long c0, c1;
1266 struct srcu_data *sdp;
1267
1268 sdp = per_cpu_ptr(ssp->sda, cpu);
1269 u0 = sdp->srcu_unlock_count[!idx];
1270 u1 = sdp->srcu_unlock_count[idx];
1271
1272 /*
1273 * Make sure that a lock is always counted if the corresponding
1274 * unlock is counted.
1275 */
1276 smp_rmb();
1277
1278 l0 = sdp->srcu_lock_count[!idx];
1279 l1 = sdp->srcu_lock_count[idx];
1280
1281 c0 = l0 - u0;
1282 c1 = l1 - u1;
1283 pr_cont(" %d(%ld,%ld %c)",
1284 cpu, c0, c1,
1285 "C."[rcu_segcblist_empty(&sdp->srcu_cblist)]);
1286 s0 += c0;
1287 s1 += c1;
1288 }
1289 pr_cont(" T(%ld,%ld)\n", s0, s1);
1290 }
1291 EXPORT_SYMBOL_GPL(srcu_torture_stats_print);
1292
1293 static int __init srcu_bootup_announce(void)
1294 {
1295 pr_info("Hierarchical SRCU implementation.\n");
1296 if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF)
1297 pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff);
1298 return 0;
1299 }
1300 early_initcall(srcu_bootup_announce);
1301
1302 void __init srcu_init(void)
1303 {
1304 struct srcu_struct *ssp;
1305
1306 srcu_init_done = true;
1307 while (!list_empty(&srcu_boot_list)) {
1308 ssp = list_first_entry(&srcu_boot_list, struct srcu_struct,
1309 work.work.entry);
1310 check_init_srcu_struct(ssp);
1311 list_del_init(&ssp->work.work.entry);
1312 queue_work(rcu_gp_wq, &ssp->work.work);
1313 }
1314 }
1315
1316 #ifdef CONFIG_MODULES
1317
1318 /* Initialize any global-scope srcu_struct structures used by this module. */
1319 static int srcu_module_coming(struct module *mod)
1320 {
1321 int i;
1322 struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1323 int ret;
1324
1325 for (i = 0; i < mod->num_srcu_structs; i++) {
1326 ret = init_srcu_struct(*(sspp++));
1327 if (WARN_ON_ONCE(ret))
1328 return ret;
1329 }
1330 return 0;
1331 }
1332
1333 /* Clean up any global-scope srcu_struct structures used by this module. */
1334 static void srcu_module_going(struct module *mod)
1335 {
1336 int i;
1337 struct srcu_struct **sspp = mod->srcu_struct_ptrs;
1338
1339 for (i = 0; i < mod->num_srcu_structs; i++)
1340 cleanup_srcu_struct(*(sspp++));
1341 }
1342
1343 /* Handle one module, either coming or going. */
1344 static int srcu_module_notify(struct notifier_block *self,
1345 unsigned long val, void *data)
1346 {
1347 struct module *mod = data;
1348 int ret = 0;
1349
1350 switch (val) {
1351 case MODULE_STATE_COMING:
1352 ret = srcu_module_coming(mod);
1353 break;
1354 case MODULE_STATE_GOING:
1355 srcu_module_going(mod);
1356 break;
1357 default:
1358 break;
1359 }
1360 return ret;
1361 }
1362
1363 static struct notifier_block srcu_module_nb = {
1364 .notifier_call = srcu_module_notify,
1365 .priority = 0,
1366 };
1367
1368 static __init int init_srcu_module_notifier(void)
1369 {
1370 int ret;
1371
1372 ret = register_module_notifier(&srcu_module_nb);
1373 if (ret)
1374 pr_warn("Failed to register srcu module notifier\n");
1375 return ret;
1376 }
1377 late_initcall(init_srcu_module_notifier);
1378
1379 #endif /* #ifdef CONFIG_MODULES */