]>
Commit | Line | Data |
---|---|---|
dad81a20 PM |
1 | /* |
2 | * Sleepable Read-Copy Update mechanism for mutual exclusion. | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public License | |
15 | * along with this program; if not, you can access it online at | |
16 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
17 | * | |
18 | * Copyright (C) IBM Corporation, 2006 | |
19 | * Copyright (C) Fujitsu, 2012 | |
20 | * | |
21 | * Author: Paul McKenney <paulmck@us.ibm.com> | |
22 | * Lai Jiangshan <laijs@cn.fujitsu.com> | |
23 | * | |
24 | * For detailed explanation of Read-Copy Update mechanism see - | |
25 | * Documentation/RCU/ *.txt | |
26 | * | |
27 | */ | |
28 | ||
29 | #include <linux/export.h> | |
30 | #include <linux/mutex.h> | |
31 | #include <linux/percpu.h> | |
32 | #include <linux/preempt.h> | |
33 | #include <linux/rcupdate_wait.h> | |
34 | #include <linux/sched.h> | |
35 | #include <linux/smp.h> | |
36 | #include <linux/delay.h> | |
22607d66 | 37 | #include <linux/module.h> |
dad81a20 PM |
38 | #include <linux/srcu.h> |
39 | ||
dad81a20 | 40 | #include "rcu.h" |
45753c5f | 41 | #include "rcu_segcblist.h" |
dad81a20 | 42 | |
0c8e0e3c PM |
43 | /* Holdoff in nanoseconds for auto-expediting. */ |
44 | #define DEFAULT_SRCU_EXP_HOLDOFF (25 * 1000) | |
45 | static ulong exp_holdoff = DEFAULT_SRCU_EXP_HOLDOFF; | |
22607d66 PM |
46 | module_param(exp_holdoff, ulong, 0444); |
47 | ||
da915ad5 PM |
48 | static void srcu_invoke_callbacks(struct work_struct *work); |
49 | static void srcu_reschedule(struct srcu_struct *sp, unsigned long delay); | |
50 | ||
51 | /* | |
52 | * Initialize SRCU combining tree. Note that statically allocated | |
53 | * srcu_struct structures might already have srcu_read_lock() and | |
54 | * srcu_read_unlock() running against them. So if the is_static parameter | |
55 | * is set, don't initialize ->srcu_lock_count[] and ->srcu_unlock_count[]. | |
56 | */ | |
57 | static void init_srcu_struct_nodes(struct srcu_struct *sp, bool is_static) | |
dad81a20 | 58 | { |
da915ad5 PM |
59 | int cpu; |
60 | int i; | |
61 | int level = 0; | |
62 | int levelspread[RCU_NUM_LVLS]; | |
63 | struct srcu_data *sdp; | |
64 | struct srcu_node *snp; | |
65 | struct srcu_node *snp_first; | |
66 | ||
67 | /* Work out the overall tree geometry. */ | |
68 | sp->level[0] = &sp->node[0]; | |
69 | for (i = 1; i < rcu_num_lvls; i++) | |
70 | sp->level[i] = sp->level[i - 1] + num_rcu_lvl[i - 1]; | |
71 | rcu_init_levelspread(levelspread, num_rcu_lvl); | |
72 | ||
73 | /* Each pass through this loop initializes one srcu_node structure. */ | |
74 | rcu_for_each_node_breadth_first(sp, snp) { | |
75 | spin_lock_init(&snp->lock); | |
c7e88067 PM |
76 | WARN_ON_ONCE(ARRAY_SIZE(snp->srcu_have_cbs) != |
77 | ARRAY_SIZE(snp->srcu_data_have_cbs)); | |
78 | for (i = 0; i < ARRAY_SIZE(snp->srcu_have_cbs); i++) { | |
da915ad5 | 79 | snp->srcu_have_cbs[i] = 0; |
c7e88067 PM |
80 | snp->srcu_data_have_cbs[i] = 0; |
81 | } | |
1e9a038b | 82 | snp->srcu_gp_seq_needed_exp = 0; |
da915ad5 PM |
83 | snp->grplo = -1; |
84 | snp->grphi = -1; | |
85 | if (snp == &sp->node[0]) { | |
86 | /* Root node, special case. */ | |
87 | snp->srcu_parent = NULL; | |
88 | continue; | |
89 | } | |
90 | ||
91 | /* Non-root node. */ | |
92 | if (snp == sp->level[level + 1]) | |
93 | level++; | |
94 | snp->srcu_parent = sp->level[level - 1] + | |
95 | (snp - sp->level[level]) / | |
96 | levelspread[level - 1]; | |
97 | } | |
98 | ||
99 | /* | |
100 | * Initialize the per-CPU srcu_data array, which feeds into the | |
101 | * leaves of the srcu_node tree. | |
102 | */ | |
103 | WARN_ON_ONCE(ARRAY_SIZE(sdp->srcu_lock_count) != | |
104 | ARRAY_SIZE(sdp->srcu_unlock_count)); | |
105 | level = rcu_num_lvls - 1; | |
106 | snp_first = sp->level[level]; | |
107 | for_each_possible_cpu(cpu) { | |
108 | sdp = per_cpu_ptr(sp->sda, cpu); | |
109 | spin_lock_init(&sdp->lock); | |
110 | rcu_segcblist_init(&sdp->srcu_cblist); | |
111 | sdp->srcu_cblist_invoking = false; | |
112 | sdp->srcu_gp_seq_needed = sp->srcu_gp_seq; | |
1e9a038b | 113 | sdp->srcu_gp_seq_needed_exp = sp->srcu_gp_seq; |
da915ad5 PM |
114 | sdp->mynode = &snp_first[cpu / levelspread[level]]; |
115 | for (snp = sdp->mynode; snp != NULL; snp = snp->srcu_parent) { | |
116 | if (snp->grplo < 0) | |
117 | snp->grplo = cpu; | |
118 | snp->grphi = cpu; | |
119 | } | |
120 | sdp->cpu = cpu; | |
121 | INIT_DELAYED_WORK(&sdp->work, srcu_invoke_callbacks); | |
122 | sdp->sp = sp; | |
c7e88067 | 123 | sdp->grpmask = 1 << (cpu - sdp->mynode->grplo); |
da915ad5 PM |
124 | if (is_static) |
125 | continue; | |
126 | ||
127 | /* Dynamically allocated, better be no srcu_read_locks()! */ | |
128 | for (i = 0; i < ARRAY_SIZE(sdp->srcu_lock_count); i++) { | |
129 | sdp->srcu_lock_count[i] = 0; | |
130 | sdp->srcu_unlock_count[i] = 0; | |
131 | } | |
132 | } | |
133 | } | |
134 | ||
135 | /* | |
136 | * Initialize non-compile-time initialized fields, including the | |
137 | * associated srcu_node and srcu_data structures. The is_static | |
138 | * parameter is passed through to init_srcu_struct_nodes(), and | |
139 | * also tells us that ->sda has already been wired up to srcu_data. | |
140 | */ | |
141 | static int init_srcu_struct_fields(struct srcu_struct *sp, bool is_static) | |
142 | { | |
143 | mutex_init(&sp->srcu_cb_mutex); | |
144 | mutex_init(&sp->srcu_gp_mutex); | |
145 | sp->srcu_idx = 0; | |
dad81a20 | 146 | sp->srcu_gp_seq = 0; |
da915ad5 PM |
147 | sp->srcu_barrier_seq = 0; |
148 | mutex_init(&sp->srcu_barrier_mutex); | |
149 | atomic_set(&sp->srcu_barrier_cpu_cnt, 0); | |
dad81a20 | 150 | INIT_DELAYED_WORK(&sp->work, process_srcu); |
da915ad5 PM |
151 | if (!is_static) |
152 | sp->sda = alloc_percpu(struct srcu_data); | |
153 | init_srcu_struct_nodes(sp, is_static); | |
1e9a038b | 154 | sp->srcu_gp_seq_needed_exp = 0; |
22607d66 | 155 | sp->srcu_last_gp_end = ktime_get_mono_fast_ns(); |
da915ad5 PM |
156 | smp_store_release(&sp->srcu_gp_seq_needed, 0); /* Init done. */ |
157 | return sp->sda ? 0 : -ENOMEM; | |
dad81a20 PM |
158 | } |
159 | ||
160 | #ifdef CONFIG_DEBUG_LOCK_ALLOC | |
161 | ||
162 | int __init_srcu_struct(struct srcu_struct *sp, const char *name, | |
163 | struct lock_class_key *key) | |
164 | { | |
165 | /* Don't re-initialize a lock while it is held. */ | |
166 | debug_check_no_locks_freed((void *)sp, sizeof(*sp)); | |
167 | lockdep_init_map(&sp->dep_map, name, key, 0); | |
da915ad5 PM |
168 | spin_lock_init(&sp->gp_lock); |
169 | return init_srcu_struct_fields(sp, false); | |
dad81a20 PM |
170 | } |
171 | EXPORT_SYMBOL_GPL(__init_srcu_struct); | |
172 | ||
173 | #else /* #ifdef CONFIG_DEBUG_LOCK_ALLOC */ | |
174 | ||
175 | /** | |
176 | * init_srcu_struct - initialize a sleep-RCU structure | |
177 | * @sp: structure to initialize. | |
178 | * | |
179 | * Must invoke this on a given srcu_struct before passing that srcu_struct | |
180 | * to any other function. Each srcu_struct represents a separate domain | |
181 | * of SRCU protection. | |
182 | */ | |
183 | int init_srcu_struct(struct srcu_struct *sp) | |
184 | { | |
da915ad5 PM |
185 | spin_lock_init(&sp->gp_lock); |
186 | return init_srcu_struct_fields(sp, false); | |
dad81a20 PM |
187 | } |
188 | EXPORT_SYMBOL_GPL(init_srcu_struct); | |
189 | ||
190 | #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */ | |
191 | ||
192 | /* | |
da915ad5 PM |
193 | * First-use initialization of statically allocated srcu_struct |
194 | * structure. Wiring up the combining tree is more than can be | |
195 | * done with compile-time initialization, so this check is added | |
196 | * to each update-side SRCU primitive. Use ->gp_lock, which -is- | |
197 | * compile-time initialized, to resolve races involving multiple | |
198 | * CPUs trying to garner first-use privileges. | |
199 | */ | |
200 | static void check_init_srcu_struct(struct srcu_struct *sp) | |
201 | { | |
202 | unsigned long flags; | |
203 | ||
204 | WARN_ON_ONCE(rcu_scheduler_active == RCU_SCHEDULER_INIT); | |
205 | /* The smp_load_acquire() pairs with the smp_store_release(). */ | |
206 | if (!rcu_seq_state(smp_load_acquire(&sp->srcu_gp_seq_needed))) /*^^^*/ | |
207 | return; /* Already initialized. */ | |
208 | spin_lock_irqsave(&sp->gp_lock, flags); | |
209 | if (!rcu_seq_state(sp->srcu_gp_seq_needed)) { | |
210 | spin_unlock_irqrestore(&sp->gp_lock, flags); | |
211 | return; | |
212 | } | |
213 | init_srcu_struct_fields(sp, true); | |
214 | spin_unlock_irqrestore(&sp->gp_lock, flags); | |
215 | } | |
216 | ||
217 | /* | |
218 | * Returns approximate total of the readers' ->srcu_lock_count[] values | |
219 | * for the rank of per-CPU counters specified by idx. | |
dad81a20 PM |
220 | */ |
221 | static unsigned long srcu_readers_lock_idx(struct srcu_struct *sp, int idx) | |
222 | { | |
223 | int cpu; | |
224 | unsigned long sum = 0; | |
225 | ||
226 | for_each_possible_cpu(cpu) { | |
da915ad5 | 227 | struct srcu_data *cpuc = per_cpu_ptr(sp->sda, cpu); |
dad81a20 | 228 | |
da915ad5 | 229 | sum += READ_ONCE(cpuc->srcu_lock_count[idx]); |
dad81a20 PM |
230 | } |
231 | return sum; | |
232 | } | |
233 | ||
234 | /* | |
da915ad5 PM |
235 | * Returns approximate total of the readers' ->srcu_unlock_count[] values |
236 | * for the rank of per-CPU counters specified by idx. | |
dad81a20 PM |
237 | */ |
238 | static unsigned long srcu_readers_unlock_idx(struct srcu_struct *sp, int idx) | |
239 | { | |
240 | int cpu; | |
241 | unsigned long sum = 0; | |
242 | ||
243 | for_each_possible_cpu(cpu) { | |
da915ad5 | 244 | struct srcu_data *cpuc = per_cpu_ptr(sp->sda, cpu); |
dad81a20 | 245 | |
da915ad5 | 246 | sum += READ_ONCE(cpuc->srcu_unlock_count[idx]); |
dad81a20 PM |
247 | } |
248 | return sum; | |
249 | } | |
250 | ||
251 | /* | |
252 | * Return true if the number of pre-existing readers is determined to | |
253 | * be zero. | |
254 | */ | |
255 | static bool srcu_readers_active_idx_check(struct srcu_struct *sp, int idx) | |
256 | { | |
257 | unsigned long unlocks; | |
258 | ||
259 | unlocks = srcu_readers_unlock_idx(sp, idx); | |
260 | ||
261 | /* | |
262 | * Make sure that a lock is always counted if the corresponding | |
263 | * unlock is counted. Needs to be a smp_mb() as the read side may | |
264 | * contain a read from a variable that is written to before the | |
265 | * synchronize_srcu() in the write side. In this case smp_mb()s | |
266 | * A and B act like the store buffering pattern. | |
267 | * | |
268 | * This smp_mb() also pairs with smp_mb() C to prevent accesses | |
269 | * after the synchronize_srcu() from being executed before the | |
270 | * grace period ends. | |
271 | */ | |
272 | smp_mb(); /* A */ | |
273 | ||
274 | /* | |
275 | * If the locks are the same as the unlocks, then there must have | |
276 | * been no readers on this index at some time in between. This does | |
277 | * not mean that there are no more readers, as one could have read | |
278 | * the current index but not have incremented the lock counter yet. | |
279 | * | |
881ec9d2 PM |
280 | * So suppose that the updater is preempted here for so long |
281 | * that more than ULONG_MAX non-nested readers come and go in | |
282 | * the meantime. It turns out that this cannot result in overflow | |
283 | * because if a reader modifies its unlock count after we read it | |
284 | * above, then that reader's next load of ->srcu_idx is guaranteed | |
285 | * to get the new value, which will cause it to operate on the | |
286 | * other bank of counters, where it cannot contribute to the | |
287 | * overflow of these counters. This means that there is a maximum | |
288 | * of 2*NR_CPUS increments, which cannot overflow given current | |
289 | * systems, especially not on 64-bit systems. | |
290 | * | |
291 | * OK, how about nesting? This does impose a limit on nesting | |
292 | * of floor(ULONG_MAX/NR_CPUS/2), which should be sufficient, | |
293 | * especially on 64-bit systems. | |
dad81a20 PM |
294 | */ |
295 | return srcu_readers_lock_idx(sp, idx) == unlocks; | |
296 | } | |
297 | ||
298 | /** | |
299 | * srcu_readers_active - returns true if there are readers. and false | |
300 | * otherwise | |
301 | * @sp: which srcu_struct to count active readers (holding srcu_read_lock). | |
302 | * | |
303 | * Note that this is not an atomic primitive, and can therefore suffer | |
304 | * severe errors when invoked on an active srcu_struct. That said, it | |
305 | * can be useful as an error check at cleanup time. | |
306 | */ | |
307 | static bool srcu_readers_active(struct srcu_struct *sp) | |
308 | { | |
309 | int cpu; | |
310 | unsigned long sum = 0; | |
311 | ||
312 | for_each_possible_cpu(cpu) { | |
da915ad5 | 313 | struct srcu_data *cpuc = per_cpu_ptr(sp->sda, cpu); |
dad81a20 | 314 | |
da915ad5 PM |
315 | sum += READ_ONCE(cpuc->srcu_lock_count[0]); |
316 | sum += READ_ONCE(cpuc->srcu_lock_count[1]); | |
317 | sum -= READ_ONCE(cpuc->srcu_unlock_count[0]); | |
318 | sum -= READ_ONCE(cpuc->srcu_unlock_count[1]); | |
dad81a20 PM |
319 | } |
320 | return sum; | |
321 | } | |
322 | ||
323 | #define SRCU_INTERVAL 1 | |
324 | ||
1e9a038b PM |
325 | /* |
326 | * Return grace-period delay, zero if there are expedited grace | |
327 | * periods pending, SRCU_INTERVAL otherwise. | |
328 | */ | |
329 | static unsigned long srcu_get_delay(struct srcu_struct *sp) | |
330 | { | |
331 | if (ULONG_CMP_LT(READ_ONCE(sp->srcu_gp_seq), | |
332 | READ_ONCE(sp->srcu_gp_seq_needed_exp))) | |
333 | return 0; | |
334 | return SRCU_INTERVAL; | |
335 | } | |
336 | ||
dad81a20 PM |
337 | /** |
338 | * cleanup_srcu_struct - deconstruct a sleep-RCU structure | |
339 | * @sp: structure to clean up. | |
340 | * | |
341 | * Must invoke this after you are finished using a given srcu_struct that | |
342 | * was initialized via init_srcu_struct(), else you leak memory. | |
343 | */ | |
344 | void cleanup_srcu_struct(struct srcu_struct *sp) | |
345 | { | |
da915ad5 PM |
346 | int cpu; |
347 | ||
1e9a038b PM |
348 | if (WARN_ON(!srcu_get_delay(sp))) |
349 | return; /* Leakage unless caller handles error. */ | |
dad81a20 PM |
350 | if (WARN_ON(srcu_readers_active(sp))) |
351 | return; /* Leakage unless caller handles error. */ | |
dad81a20 | 352 | flush_delayed_work(&sp->work); |
da915ad5 PM |
353 | for_each_possible_cpu(cpu) |
354 | flush_delayed_work(&per_cpu_ptr(sp->sda, cpu)->work); | |
355 | if (WARN_ON(rcu_seq_state(READ_ONCE(sp->srcu_gp_seq)) != SRCU_STATE_IDLE) || | |
356 | WARN_ON(srcu_readers_active(sp))) { | |
357 | pr_info("cleanup_srcu_struct: Active srcu_struct %p state: %d\n", sp, rcu_seq_state(READ_ONCE(sp->srcu_gp_seq))); | |
dad81a20 PM |
358 | return; /* Caller forgot to stop doing call_srcu()? */ |
359 | } | |
da915ad5 PM |
360 | free_percpu(sp->sda); |
361 | sp->sda = NULL; | |
dad81a20 PM |
362 | } |
363 | EXPORT_SYMBOL_GPL(cleanup_srcu_struct); | |
364 | ||
365 | /* | |
366 | * Counts the new reader in the appropriate per-CPU element of the | |
cdf7abc4 | 367 | * srcu_struct. |
dad81a20 PM |
368 | * Returns an index that must be passed to the matching srcu_read_unlock(). |
369 | */ | |
370 | int __srcu_read_lock(struct srcu_struct *sp) | |
371 | { | |
372 | int idx; | |
373 | ||
da915ad5 | 374 | idx = READ_ONCE(sp->srcu_idx) & 0x1; |
cdf7abc4 | 375 | this_cpu_inc(sp->sda->srcu_lock_count[idx]); |
dad81a20 PM |
376 | smp_mb(); /* B */ /* Avoid leaking the critical section. */ |
377 | return idx; | |
378 | } | |
379 | EXPORT_SYMBOL_GPL(__srcu_read_lock); | |
380 | ||
381 | /* | |
382 | * Removes the count for the old reader from the appropriate per-CPU | |
383 | * element of the srcu_struct. Note that this may well be a different | |
384 | * CPU than that which was incremented by the corresponding srcu_read_lock(). | |
dad81a20 PM |
385 | */ |
386 | void __srcu_read_unlock(struct srcu_struct *sp, int idx) | |
387 | { | |
388 | smp_mb(); /* C */ /* Avoid leaking the critical section. */ | |
da915ad5 | 389 | this_cpu_inc(sp->sda->srcu_unlock_count[idx]); |
dad81a20 PM |
390 | } |
391 | EXPORT_SYMBOL_GPL(__srcu_read_unlock); | |
392 | ||
393 | /* | |
394 | * We use an adaptive strategy for synchronize_srcu() and especially for | |
395 | * synchronize_srcu_expedited(). We spin for a fixed time period | |
396 | * (defined below) to allow SRCU readers to exit their read-side critical | |
397 | * sections. If there are still some readers after a few microseconds, | |
398 | * we repeatedly block for 1-millisecond time periods. | |
399 | */ | |
400 | #define SRCU_RETRY_CHECK_DELAY 5 | |
401 | ||
402 | /* | |
403 | * Start an SRCU grace period. | |
404 | */ | |
405 | static void srcu_gp_start(struct srcu_struct *sp) | |
406 | { | |
da915ad5 | 407 | struct srcu_data *sdp = this_cpu_ptr(sp->sda); |
dad81a20 PM |
408 | int state; |
409 | ||
da915ad5 PM |
410 | RCU_LOCKDEP_WARN(!lockdep_is_held(&sp->gp_lock), |
411 | "Invoked srcu_gp_start() without ->gp_lock!"); | |
412 | WARN_ON_ONCE(ULONG_CMP_GE(sp->srcu_gp_seq, sp->srcu_gp_seq_needed)); | |
413 | rcu_segcblist_advance(&sdp->srcu_cblist, | |
414 | rcu_seq_current(&sp->srcu_gp_seq)); | |
415 | (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, | |
416 | rcu_seq_snap(&sp->srcu_gp_seq)); | |
2da4b2a7 | 417 | smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */ |
dad81a20 PM |
418 | rcu_seq_start(&sp->srcu_gp_seq); |
419 | state = rcu_seq_state(READ_ONCE(sp->srcu_gp_seq)); | |
420 | WARN_ON_ONCE(state != SRCU_STATE_SCAN1); | |
421 | } | |
422 | ||
da915ad5 PM |
423 | /* |
424 | * Track online CPUs to guide callback workqueue placement. | |
425 | */ | |
426 | DEFINE_PER_CPU(bool, srcu_online); | |
427 | ||
428 | void srcu_online_cpu(unsigned int cpu) | |
429 | { | |
430 | WRITE_ONCE(per_cpu(srcu_online, cpu), true); | |
431 | } | |
432 | ||
433 | void srcu_offline_cpu(unsigned int cpu) | |
434 | { | |
435 | WRITE_ONCE(per_cpu(srcu_online, cpu), false); | |
436 | } | |
437 | ||
438 | /* | |
439 | * Place the workqueue handler on the specified CPU if online, otherwise | |
440 | * just run it whereever. This is useful for placing workqueue handlers | |
441 | * that are to invoke the specified CPU's callbacks. | |
442 | */ | |
443 | static bool srcu_queue_delayed_work_on(int cpu, struct workqueue_struct *wq, | |
444 | struct delayed_work *dwork, | |
445 | unsigned long delay) | |
446 | { | |
447 | bool ret; | |
448 | ||
449 | preempt_disable(); | |
450 | if (READ_ONCE(per_cpu(srcu_online, cpu))) | |
451 | ret = queue_delayed_work_on(cpu, wq, dwork, delay); | |
452 | else | |
453 | ret = queue_delayed_work(wq, dwork, delay); | |
454 | preempt_enable(); | |
455 | return ret; | |
456 | } | |
457 | ||
458 | /* | |
459 | * Schedule callback invocation for the specified srcu_data structure, | |
460 | * if possible, on the corresponding CPU. | |
461 | */ | |
462 | static void srcu_schedule_cbs_sdp(struct srcu_data *sdp, unsigned long delay) | |
463 | { | |
464 | srcu_queue_delayed_work_on(sdp->cpu, system_power_efficient_wq, | |
465 | &sdp->work, delay); | |
466 | } | |
467 | ||
468 | /* | |
469 | * Schedule callback invocation for all srcu_data structures associated | |
c7e88067 PM |
470 | * with the specified srcu_node structure that have callbacks for the |
471 | * just-completed grace period, the one corresponding to idx. If possible, | |
472 | * schedule this invocation on the corresponding CPUs. | |
da915ad5 | 473 | */ |
c7e88067 | 474 | static void srcu_schedule_cbs_snp(struct srcu_struct *sp, struct srcu_node *snp, |
1e9a038b | 475 | unsigned long mask, unsigned long delay) |
da915ad5 PM |
476 | { |
477 | int cpu; | |
478 | ||
c7e88067 PM |
479 | for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) { |
480 | if (!(mask & (1 << (cpu - snp->grplo)))) | |
481 | continue; | |
1e9a038b | 482 | srcu_schedule_cbs_sdp(per_cpu_ptr(sp->sda, cpu), delay); |
c7e88067 | 483 | } |
da915ad5 PM |
484 | } |
485 | ||
486 | /* | |
487 | * Note the end of an SRCU grace period. Initiates callback invocation | |
488 | * and starts a new grace period if needed. | |
489 | * | |
490 | * The ->srcu_cb_mutex acquisition does not protect any data, but | |
491 | * instead prevents more than one grace period from starting while we | |
492 | * are initiating callback invocation. This allows the ->srcu_have_cbs[] | |
493 | * array to have a finite number of elements. | |
494 | */ | |
495 | static void srcu_gp_end(struct srcu_struct *sp) | |
496 | { | |
1e9a038b | 497 | unsigned long cbdelay; |
da915ad5 PM |
498 | bool cbs; |
499 | unsigned long gpseq; | |
500 | int idx; | |
501 | int idxnext; | |
c7e88067 | 502 | unsigned long mask; |
da915ad5 PM |
503 | struct srcu_node *snp; |
504 | ||
505 | /* Prevent more than one additional grace period. */ | |
506 | mutex_lock(&sp->srcu_cb_mutex); | |
507 | ||
508 | /* End the current grace period. */ | |
509 | spin_lock_irq(&sp->gp_lock); | |
510 | idx = rcu_seq_state(sp->srcu_gp_seq); | |
511 | WARN_ON_ONCE(idx != SRCU_STATE_SCAN2); | |
1e9a038b | 512 | cbdelay = srcu_get_delay(sp); |
22607d66 | 513 | sp->srcu_last_gp_end = ktime_get_mono_fast_ns(); |
da915ad5 PM |
514 | rcu_seq_end(&sp->srcu_gp_seq); |
515 | gpseq = rcu_seq_current(&sp->srcu_gp_seq); | |
1e9a038b PM |
516 | if (ULONG_CMP_LT(sp->srcu_gp_seq_needed_exp, gpseq)) |
517 | sp->srcu_gp_seq_needed_exp = gpseq; | |
da915ad5 PM |
518 | spin_unlock_irq(&sp->gp_lock); |
519 | mutex_unlock(&sp->srcu_gp_mutex); | |
520 | /* A new grace period can start at this point. But only one. */ | |
521 | ||
522 | /* Initiate callback invocation as needed. */ | |
523 | idx = rcu_seq_ctr(gpseq) % ARRAY_SIZE(snp->srcu_have_cbs); | |
524 | idxnext = (idx + 1) % ARRAY_SIZE(snp->srcu_have_cbs); | |
525 | rcu_for_each_node_breadth_first(sp, snp) { | |
526 | spin_lock_irq(&snp->lock); | |
527 | cbs = false; | |
528 | if (snp >= sp->level[rcu_num_lvls - 1]) | |
529 | cbs = snp->srcu_have_cbs[idx] == gpseq; | |
530 | snp->srcu_have_cbs[idx] = gpseq; | |
531 | rcu_seq_set_state(&snp->srcu_have_cbs[idx], 1); | |
1e9a038b PM |
532 | if (ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, gpseq)) |
533 | snp->srcu_gp_seq_needed_exp = gpseq; | |
c7e88067 PM |
534 | mask = snp->srcu_data_have_cbs[idx]; |
535 | snp->srcu_data_have_cbs[idx] = 0; | |
da915ad5 PM |
536 | spin_unlock_irq(&snp->lock); |
537 | if (cbs) { | |
538 | smp_mb(); /* GP end before CB invocation. */ | |
1e9a038b | 539 | srcu_schedule_cbs_snp(sp, snp, mask, cbdelay); |
da915ad5 PM |
540 | } |
541 | } | |
542 | ||
543 | /* Callback initiation done, allow grace periods after next. */ | |
544 | mutex_unlock(&sp->srcu_cb_mutex); | |
545 | ||
546 | /* Start a new grace period if needed. */ | |
547 | spin_lock_irq(&sp->gp_lock); | |
548 | gpseq = rcu_seq_current(&sp->srcu_gp_seq); | |
549 | if (!rcu_seq_state(gpseq) && | |
550 | ULONG_CMP_LT(gpseq, sp->srcu_gp_seq_needed)) { | |
551 | srcu_gp_start(sp); | |
552 | spin_unlock_irq(&sp->gp_lock); | |
553 | /* Throttle expedited grace periods: Should be rare! */ | |
1e9a038b PM |
554 | srcu_reschedule(sp, rcu_seq_ctr(gpseq) & 0x3ff |
555 | ? 0 : SRCU_INTERVAL); | |
da915ad5 PM |
556 | } else { |
557 | spin_unlock_irq(&sp->gp_lock); | |
558 | } | |
559 | } | |
560 | ||
1e9a038b PM |
561 | /* |
562 | * Funnel-locking scheme to scalably mediate many concurrent expedited | |
563 | * grace-period requests. This function is invoked for the first known | |
564 | * expedited request for a grace period that has already been requested, | |
565 | * but without expediting. To start a completely new grace period, | |
566 | * whether expedited or not, use srcu_funnel_gp_start() instead. | |
567 | */ | |
568 | static void srcu_funnel_exp_start(struct srcu_struct *sp, struct srcu_node *snp, | |
569 | unsigned long s) | |
570 | { | |
571 | unsigned long flags; | |
572 | ||
573 | for (; snp != NULL; snp = snp->srcu_parent) { | |
574 | if (rcu_seq_done(&sp->srcu_gp_seq, s) || | |
575 | ULONG_CMP_GE(READ_ONCE(snp->srcu_gp_seq_needed_exp), s)) | |
576 | return; | |
577 | spin_lock_irqsave(&snp->lock, flags); | |
578 | if (ULONG_CMP_GE(snp->srcu_gp_seq_needed_exp, s)) { | |
579 | spin_unlock_irqrestore(&snp->lock, flags); | |
580 | return; | |
581 | } | |
582 | WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s); | |
583 | spin_unlock_irqrestore(&snp->lock, flags); | |
584 | } | |
585 | spin_lock_irqsave(&sp->gp_lock, flags); | |
586 | if (!ULONG_CMP_LT(sp->srcu_gp_seq_needed_exp, s)) | |
587 | sp->srcu_gp_seq_needed_exp = s; | |
588 | spin_unlock_irqrestore(&sp->gp_lock, flags); | |
589 | } | |
590 | ||
da915ad5 PM |
591 | /* |
592 | * Funnel-locking scheme to scalably mediate many concurrent grace-period | |
593 | * requests. The winner has to do the work of actually starting grace | |
594 | * period s. Losers must either ensure that their desired grace-period | |
595 | * number is recorded on at least their leaf srcu_node structure, or they | |
596 | * must take steps to invoke their own callbacks. | |
597 | */ | |
1e9a038b PM |
598 | static void srcu_funnel_gp_start(struct srcu_struct *sp, struct srcu_data *sdp, |
599 | unsigned long s, bool do_norm) | |
da915ad5 PM |
600 | { |
601 | unsigned long flags; | |
602 | int idx = rcu_seq_ctr(s) % ARRAY_SIZE(sdp->mynode->srcu_have_cbs); | |
603 | struct srcu_node *snp = sdp->mynode; | |
604 | unsigned long snp_seq; | |
605 | ||
606 | /* Each pass through the loop does one level of the srcu_node tree. */ | |
607 | for (; snp != NULL; snp = snp->srcu_parent) { | |
608 | if (rcu_seq_done(&sp->srcu_gp_seq, s) && snp != sdp->mynode) | |
609 | return; /* GP already done and CBs recorded. */ | |
610 | spin_lock_irqsave(&snp->lock, flags); | |
611 | if (ULONG_CMP_GE(snp->srcu_have_cbs[idx], s)) { | |
612 | snp_seq = snp->srcu_have_cbs[idx]; | |
c7e88067 PM |
613 | if (snp == sdp->mynode && snp_seq == s) |
614 | snp->srcu_data_have_cbs[idx] |= sdp->grpmask; | |
da915ad5 PM |
615 | spin_unlock_irqrestore(&snp->lock, flags); |
616 | if (snp == sdp->mynode && snp_seq != s) { | |
617 | smp_mb(); /* CBs after GP! */ | |
1e9a038b PM |
618 | srcu_schedule_cbs_sdp(sdp, do_norm |
619 | ? SRCU_INTERVAL | |
620 | : 0); | |
621 | return; | |
da915ad5 | 622 | } |
1e9a038b PM |
623 | if (!do_norm) |
624 | srcu_funnel_exp_start(sp, snp, s); | |
da915ad5 PM |
625 | return; |
626 | } | |
627 | snp->srcu_have_cbs[idx] = s; | |
c7e88067 PM |
628 | if (snp == sdp->mynode) |
629 | snp->srcu_data_have_cbs[idx] |= sdp->grpmask; | |
1e9a038b PM |
630 | if (!do_norm && ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, s)) |
631 | snp->srcu_gp_seq_needed_exp = s; | |
da915ad5 PM |
632 | spin_unlock_irqrestore(&snp->lock, flags); |
633 | } | |
634 | ||
635 | /* Top of tree, must ensure the grace period will be started. */ | |
636 | spin_lock_irqsave(&sp->gp_lock, flags); | |
637 | if (ULONG_CMP_LT(sp->srcu_gp_seq_needed, s)) { | |
638 | /* | |
639 | * Record need for grace period s. Pair with load | |
640 | * acquire setting up for initialization. | |
641 | */ | |
642 | smp_store_release(&sp->srcu_gp_seq_needed, s); /*^^^*/ | |
643 | } | |
1e9a038b PM |
644 | if (!do_norm && ULONG_CMP_LT(sp->srcu_gp_seq_needed_exp, s)) |
645 | sp->srcu_gp_seq_needed_exp = s; | |
da915ad5 PM |
646 | |
647 | /* If grace period not already done and none in progress, start it. */ | |
648 | if (!rcu_seq_done(&sp->srcu_gp_seq, s) && | |
649 | rcu_seq_state(sp->srcu_gp_seq) == SRCU_STATE_IDLE) { | |
650 | WARN_ON_ONCE(ULONG_CMP_GE(sp->srcu_gp_seq, sp->srcu_gp_seq_needed)); | |
651 | srcu_gp_start(sp); | |
652 | queue_delayed_work(system_power_efficient_wq, &sp->work, | |
1e9a038b | 653 | srcu_get_delay(sp)); |
da915ad5 PM |
654 | } |
655 | spin_unlock_irqrestore(&sp->gp_lock, flags); | |
656 | } | |
657 | ||
dad81a20 PM |
658 | /* |
659 | * Wait until all readers counted by array index idx complete, but | |
660 | * loop an additional time if there is an expedited grace period pending. | |
da915ad5 | 661 | * The caller must ensure that ->srcu_idx is not changed while checking. |
dad81a20 PM |
662 | */ |
663 | static bool try_check_zero(struct srcu_struct *sp, int idx, int trycount) | |
664 | { | |
665 | for (;;) { | |
666 | if (srcu_readers_active_idx_check(sp, idx)) | |
667 | return true; | |
1e9a038b | 668 | if (--trycount + !srcu_get_delay(sp) <= 0) |
dad81a20 PM |
669 | return false; |
670 | udelay(SRCU_RETRY_CHECK_DELAY); | |
671 | } | |
672 | } | |
673 | ||
674 | /* | |
da915ad5 PM |
675 | * Increment the ->srcu_idx counter so that future SRCU readers will |
676 | * use the other rank of the ->srcu_(un)lock_count[] arrays. This allows | |
dad81a20 PM |
677 | * us to wait for pre-existing readers in a starvation-free manner. |
678 | */ | |
679 | static void srcu_flip(struct srcu_struct *sp) | |
680 | { | |
881ec9d2 PM |
681 | /* |
682 | * Ensure that if this updater saw a given reader's increment | |
683 | * from __srcu_read_lock(), that reader was using an old value | |
684 | * of ->srcu_idx. Also ensure that if a given reader sees the | |
685 | * new value of ->srcu_idx, this updater's earlier scans cannot | |
686 | * have seen that reader's increments (which is OK, because this | |
687 | * grace period need not wait on that reader). | |
688 | */ | |
689 | smp_mb(); /* E */ /* Pairs with B and C. */ | |
690 | ||
da915ad5 | 691 | WRITE_ONCE(sp->srcu_idx, sp->srcu_idx + 1); |
dad81a20 PM |
692 | |
693 | /* | |
694 | * Ensure that if the updater misses an __srcu_read_unlock() | |
695 | * increment, that task's next __srcu_read_lock() will see the | |
696 | * above counter update. Note that both this memory barrier | |
697 | * and the one in srcu_readers_active_idx_check() provide the | |
698 | * guarantee for __srcu_read_lock(). | |
699 | */ | |
700 | smp_mb(); /* D */ /* Pairs with C. */ | |
701 | } | |
702 | ||
2da4b2a7 PM |
703 | /* |
704 | * If SRCU is likely idle, return true, otherwise return false. | |
705 | * | |
706 | * Note that it is OK for several current from-idle requests for a new | |
707 | * grace period from idle to specify expediting because they will all end | |
708 | * up requesting the same grace period anyhow. So no loss. | |
709 | * | |
710 | * Note also that if any CPU (including the current one) is still invoking | |
711 | * callbacks, this function will nevertheless say "idle". This is not | |
712 | * ideal, but the overhead of checking all CPUs' callback lists is even | |
713 | * less ideal, especially on large systems. Furthermore, the wakeup | |
714 | * can happen before the callback is fully removed, so we have no choice | |
715 | * but to accept this type of error. | |
716 | * | |
717 | * This function is also subject to counter-wrap errors, but let's face | |
718 | * it, if this function was preempted for enough time for the counters | |
719 | * to wrap, it really doesn't matter whether or not we expedite the grace | |
720 | * period. The extra overhead of a needlessly expedited grace period is | |
721 | * negligible when amoritized over that time period, and the extra latency | |
722 | * of a needlessly non-expedited grace period is similarly negligible. | |
723 | */ | |
724 | static bool srcu_might_be_idle(struct srcu_struct *sp) | |
725 | { | |
22607d66 | 726 | unsigned long curseq; |
2da4b2a7 PM |
727 | unsigned long flags; |
728 | struct srcu_data *sdp; | |
22607d66 | 729 | unsigned long t; |
2da4b2a7 PM |
730 | |
731 | /* If the local srcu_data structure has callbacks, not idle. */ | |
732 | local_irq_save(flags); | |
733 | sdp = this_cpu_ptr(sp->sda); | |
734 | if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) { | |
735 | local_irq_restore(flags); | |
736 | return false; /* Callbacks already present, so not idle. */ | |
737 | } | |
738 | local_irq_restore(flags); | |
739 | ||
740 | /* | |
741 | * No local callbacks, so probabalistically probe global state. | |
742 | * Exact information would require acquiring locks, which would | |
743 | * kill scalability, hence the probabalistic nature of the probe. | |
744 | */ | |
22607d66 PM |
745 | |
746 | /* First, see if enough time has passed since the last GP. */ | |
747 | t = ktime_get_mono_fast_ns(); | |
748 | if (exp_holdoff == 0 || | |
749 | time_in_range_open(t, sp->srcu_last_gp_end, | |
750 | sp->srcu_last_gp_end + exp_holdoff)) | |
751 | return false; /* Too soon after last GP. */ | |
752 | ||
753 | /* Next, check for probable idleness. */ | |
2da4b2a7 PM |
754 | curseq = rcu_seq_current(&sp->srcu_gp_seq); |
755 | smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */ | |
756 | if (ULONG_CMP_LT(curseq, READ_ONCE(sp->srcu_gp_seq_needed))) | |
757 | return false; /* Grace period in progress, so not idle. */ | |
758 | smp_mb(); /* Order ->srcu_gp_seq with prior access. */ | |
759 | if (curseq != rcu_seq_current(&sp->srcu_gp_seq)) | |
760 | return false; /* GP # changed, so not idle. */ | |
761 | return true; /* With reasonable probability, idle! */ | |
762 | } | |
763 | ||
dad81a20 | 764 | /* |
da915ad5 PM |
765 | * Enqueue an SRCU callback on the srcu_data structure associated with |
766 | * the current CPU and the specified srcu_struct structure, initiating | |
767 | * grace-period processing if it is not already running. | |
dad81a20 PM |
768 | * |
769 | * Note that all CPUs must agree that the grace period extended beyond | |
770 | * all pre-existing SRCU read-side critical section. On systems with | |
771 | * more than one CPU, this means that when "func()" is invoked, each CPU | |
772 | * is guaranteed to have executed a full memory barrier since the end of | |
773 | * its last corresponding SRCU read-side critical section whose beginning | |
774 | * preceded the call to call_rcu(). It also means that each CPU executing | |
775 | * an SRCU read-side critical section that continues beyond the start of | |
776 | * "func()" must have executed a memory barrier after the call_rcu() | |
777 | * but before the beginning of that SRCU read-side critical section. | |
778 | * Note that these guarantees include CPUs that are offline, idle, or | |
779 | * executing in user mode, as well as CPUs that are executing in the kernel. | |
780 | * | |
781 | * Furthermore, if CPU A invoked call_rcu() and CPU B invoked the | |
782 | * resulting SRCU callback function "func()", then both CPU A and CPU | |
783 | * B are guaranteed to execute a full memory barrier during the time | |
784 | * interval between the call to call_rcu() and the invocation of "func()". | |
785 | * This guarantee applies even if CPU A and CPU B are the same CPU (but | |
786 | * again only if the system has more than one CPU). | |
787 | * | |
788 | * Of course, these guarantees apply only for invocations of call_srcu(), | |
789 | * srcu_read_lock(), and srcu_read_unlock() that are all passed the same | |
790 | * srcu_struct structure. | |
791 | */ | |
1e9a038b PM |
792 | void __call_srcu(struct srcu_struct *sp, struct rcu_head *rhp, |
793 | rcu_callback_t func, bool do_norm) | |
dad81a20 PM |
794 | { |
795 | unsigned long flags; | |
1e9a038b | 796 | bool needexp = false; |
da915ad5 PM |
797 | bool needgp = false; |
798 | unsigned long s; | |
799 | struct srcu_data *sdp; | |
800 | ||
801 | check_init_srcu_struct(sp); | |
802 | rhp->func = func; | |
803 | local_irq_save(flags); | |
804 | sdp = this_cpu_ptr(sp->sda); | |
805 | spin_lock(&sdp->lock); | |
806 | rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp, false); | |
807 | rcu_segcblist_advance(&sdp->srcu_cblist, | |
808 | rcu_seq_current(&sp->srcu_gp_seq)); | |
809 | s = rcu_seq_snap(&sp->srcu_gp_seq); | |
810 | (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, s); | |
811 | if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) { | |
812 | sdp->srcu_gp_seq_needed = s; | |
813 | needgp = true; | |
dad81a20 | 814 | } |
1e9a038b PM |
815 | if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) { |
816 | sdp->srcu_gp_seq_needed_exp = s; | |
817 | needexp = true; | |
818 | } | |
da915ad5 PM |
819 | spin_unlock_irqrestore(&sdp->lock, flags); |
820 | if (needgp) | |
1e9a038b PM |
821 | srcu_funnel_gp_start(sp, sdp, s, do_norm); |
822 | else if (needexp) | |
823 | srcu_funnel_exp_start(sp, sdp->mynode, s); | |
824 | } | |
825 | ||
826 | void call_srcu(struct srcu_struct *sp, struct rcu_head *rhp, | |
827 | rcu_callback_t func) | |
828 | { | |
829 | __call_srcu(sp, rhp, func, true); | |
dad81a20 PM |
830 | } |
831 | EXPORT_SYMBOL_GPL(call_srcu); | |
832 | ||
dad81a20 PM |
833 | /* |
834 | * Helper function for synchronize_srcu() and synchronize_srcu_expedited(). | |
835 | */ | |
1e9a038b | 836 | static void __synchronize_srcu(struct srcu_struct *sp, bool do_norm) |
dad81a20 PM |
837 | { |
838 | struct rcu_synchronize rcu; | |
dad81a20 PM |
839 | |
840 | RCU_LOCKDEP_WARN(lock_is_held(&sp->dep_map) || | |
841 | lock_is_held(&rcu_bh_lock_map) || | |
842 | lock_is_held(&rcu_lock_map) || | |
843 | lock_is_held(&rcu_sched_lock_map), | |
844 | "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section"); | |
845 | ||
846 | if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE) | |
847 | return; | |
848 | might_sleep(); | |
da915ad5 | 849 | check_init_srcu_struct(sp); |
dad81a20 | 850 | init_completion(&rcu.completion); |
da915ad5 | 851 | init_rcu_head_on_stack(&rcu.head); |
1e9a038b | 852 | __call_srcu(sp, &rcu.head, wakeme_after_rcu, do_norm); |
dad81a20 | 853 | wait_for_completion(&rcu.completion); |
da915ad5 | 854 | destroy_rcu_head_on_stack(&rcu.head); |
dad81a20 PM |
855 | } |
856 | ||
857 | /** | |
858 | * synchronize_srcu_expedited - Brute-force SRCU grace period | |
859 | * @sp: srcu_struct with which to synchronize. | |
860 | * | |
861 | * Wait for an SRCU grace period to elapse, but be more aggressive about | |
862 | * spinning rather than blocking when waiting. | |
863 | * | |
864 | * Note that synchronize_srcu_expedited() has the same deadlock and | |
865 | * memory-ordering properties as does synchronize_srcu(). | |
866 | */ | |
867 | void synchronize_srcu_expedited(struct srcu_struct *sp) | |
868 | { | |
1e9a038b | 869 | __synchronize_srcu(sp, rcu_gp_is_normal()); |
dad81a20 PM |
870 | } |
871 | EXPORT_SYMBOL_GPL(synchronize_srcu_expedited); | |
872 | ||
873 | /** | |
874 | * synchronize_srcu - wait for prior SRCU read-side critical-section completion | |
875 | * @sp: srcu_struct with which to synchronize. | |
876 | * | |
877 | * Wait for the count to drain to zero of both indexes. To avoid the | |
878 | * possible starvation of synchronize_srcu(), it waits for the count of | |
da915ad5 PM |
879 | * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first, |
880 | * and then flip the srcu_idx and wait for the count of the other index. | |
dad81a20 PM |
881 | * |
882 | * Can block; must be called from process context. | |
883 | * | |
884 | * Note that it is illegal to call synchronize_srcu() from the corresponding | |
885 | * SRCU read-side critical section; doing so will result in deadlock. | |
886 | * However, it is perfectly legal to call synchronize_srcu() on one | |
887 | * srcu_struct from some other srcu_struct's read-side critical section, | |
888 | * as long as the resulting graph of srcu_structs is acyclic. | |
889 | * | |
890 | * There are memory-ordering constraints implied by synchronize_srcu(). | |
891 | * On systems with more than one CPU, when synchronize_srcu() returns, | |
892 | * each CPU is guaranteed to have executed a full memory barrier since | |
893 | * the end of its last corresponding SRCU-sched read-side critical section | |
894 | * whose beginning preceded the call to synchronize_srcu(). In addition, | |
895 | * each CPU having an SRCU read-side critical section that extends beyond | |
896 | * the return from synchronize_srcu() is guaranteed to have executed a | |
897 | * full memory barrier after the beginning of synchronize_srcu() and before | |
898 | * the beginning of that SRCU read-side critical section. Note that these | |
899 | * guarantees include CPUs that are offline, idle, or executing in user mode, | |
900 | * as well as CPUs that are executing in the kernel. | |
901 | * | |
902 | * Furthermore, if CPU A invoked synchronize_srcu(), which returned | |
903 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
904 | * to have executed a full memory barrier during the execution of | |
905 | * synchronize_srcu(). This guarantee applies even if CPU A and CPU B | |
906 | * are the same CPU, but again only if the system has more than one CPU. | |
907 | * | |
908 | * Of course, these memory-ordering guarantees apply only when | |
909 | * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are | |
910 | * passed the same srcu_struct structure. | |
2da4b2a7 PM |
911 | * |
912 | * If SRCU is likely idle, expedite the first request. This semantic | |
913 | * was provided by Classic SRCU, and is relied upon by its users, so TREE | |
914 | * SRCU must also provide it. Note that detecting idleness is heuristic | |
915 | * and subject to both false positives and negatives. | |
dad81a20 PM |
916 | */ |
917 | void synchronize_srcu(struct srcu_struct *sp) | |
918 | { | |
2da4b2a7 | 919 | if (srcu_might_be_idle(sp) || rcu_gp_is_expedited()) |
dad81a20 PM |
920 | synchronize_srcu_expedited(sp); |
921 | else | |
1e9a038b | 922 | __synchronize_srcu(sp, true); |
dad81a20 PM |
923 | } |
924 | EXPORT_SYMBOL_GPL(synchronize_srcu); | |
925 | ||
da915ad5 PM |
926 | /* |
927 | * Callback function for srcu_barrier() use. | |
928 | */ | |
929 | static void srcu_barrier_cb(struct rcu_head *rhp) | |
930 | { | |
931 | struct srcu_data *sdp; | |
932 | struct srcu_struct *sp; | |
933 | ||
934 | sdp = container_of(rhp, struct srcu_data, srcu_barrier_head); | |
935 | sp = sdp->sp; | |
936 | if (atomic_dec_and_test(&sp->srcu_barrier_cpu_cnt)) | |
937 | complete(&sp->srcu_barrier_completion); | |
938 | } | |
939 | ||
dad81a20 PM |
940 | /** |
941 | * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete. | |
942 | * @sp: srcu_struct on which to wait for in-flight callbacks. | |
943 | */ | |
944 | void srcu_barrier(struct srcu_struct *sp) | |
945 | { | |
da915ad5 PM |
946 | int cpu; |
947 | struct srcu_data *sdp; | |
948 | unsigned long s = rcu_seq_snap(&sp->srcu_barrier_seq); | |
949 | ||
950 | check_init_srcu_struct(sp); | |
951 | mutex_lock(&sp->srcu_barrier_mutex); | |
952 | if (rcu_seq_done(&sp->srcu_barrier_seq, s)) { | |
953 | smp_mb(); /* Force ordering following return. */ | |
954 | mutex_unlock(&sp->srcu_barrier_mutex); | |
955 | return; /* Someone else did our work for us. */ | |
956 | } | |
957 | rcu_seq_start(&sp->srcu_barrier_seq); | |
958 | init_completion(&sp->srcu_barrier_completion); | |
959 | ||
960 | /* Initial count prevents reaching zero until all CBs are posted. */ | |
961 | atomic_set(&sp->srcu_barrier_cpu_cnt, 1); | |
962 | ||
963 | /* | |
964 | * Each pass through this loop enqueues a callback, but only | |
965 | * on CPUs already having callbacks enqueued. Note that if | |
966 | * a CPU already has callbacks enqueue, it must have already | |
967 | * registered the need for a future grace period, so all we | |
968 | * need do is enqueue a callback that will use the same | |
969 | * grace period as the last callback already in the queue. | |
970 | */ | |
971 | for_each_possible_cpu(cpu) { | |
972 | sdp = per_cpu_ptr(sp->sda, cpu); | |
973 | spin_lock_irq(&sdp->lock); | |
974 | atomic_inc(&sp->srcu_barrier_cpu_cnt); | |
975 | sdp->srcu_barrier_head.func = srcu_barrier_cb; | |
976 | if (!rcu_segcblist_entrain(&sdp->srcu_cblist, | |
977 | &sdp->srcu_barrier_head, 0)) | |
978 | atomic_dec(&sp->srcu_barrier_cpu_cnt); | |
979 | spin_unlock_irq(&sdp->lock); | |
980 | } | |
981 | ||
982 | /* Remove the initial count, at which point reaching zero can happen. */ | |
983 | if (atomic_dec_and_test(&sp->srcu_barrier_cpu_cnt)) | |
984 | complete(&sp->srcu_barrier_completion); | |
985 | wait_for_completion(&sp->srcu_barrier_completion); | |
986 | ||
987 | rcu_seq_end(&sp->srcu_barrier_seq); | |
988 | mutex_unlock(&sp->srcu_barrier_mutex); | |
dad81a20 PM |
989 | } |
990 | EXPORT_SYMBOL_GPL(srcu_barrier); | |
991 | ||
992 | /** | |
993 | * srcu_batches_completed - return batches completed. | |
994 | * @sp: srcu_struct on which to report batch completion. | |
995 | * | |
996 | * Report the number of batches, correlated with, but not necessarily | |
997 | * precisely the same as, the number of grace periods that have elapsed. | |
998 | */ | |
999 | unsigned long srcu_batches_completed(struct srcu_struct *sp) | |
1000 | { | |
da915ad5 | 1001 | return sp->srcu_idx; |
dad81a20 PM |
1002 | } |
1003 | EXPORT_SYMBOL_GPL(srcu_batches_completed); | |
1004 | ||
1005 | /* | |
da915ad5 PM |
1006 | * Core SRCU state machine. Push state bits of ->srcu_gp_seq |
1007 | * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has | |
1008 | * completed in that state. | |
dad81a20 | 1009 | */ |
da915ad5 | 1010 | static void srcu_advance_state(struct srcu_struct *sp) |
dad81a20 PM |
1011 | { |
1012 | int idx; | |
1013 | ||
da915ad5 PM |
1014 | mutex_lock(&sp->srcu_gp_mutex); |
1015 | ||
dad81a20 PM |
1016 | /* |
1017 | * Because readers might be delayed for an extended period after | |
da915ad5 | 1018 | * fetching ->srcu_idx for their index, at any point in time there |
dad81a20 PM |
1019 | * might well be readers using both idx=0 and idx=1. We therefore |
1020 | * need to wait for readers to clear from both index values before | |
1021 | * invoking a callback. | |
1022 | * | |
1023 | * The load-acquire ensures that we see the accesses performed | |
1024 | * by the prior grace period. | |
1025 | */ | |
1026 | idx = rcu_seq_state(smp_load_acquire(&sp->srcu_gp_seq)); /* ^^^ */ | |
1027 | if (idx == SRCU_STATE_IDLE) { | |
da915ad5 PM |
1028 | spin_lock_irq(&sp->gp_lock); |
1029 | if (ULONG_CMP_GE(sp->srcu_gp_seq, sp->srcu_gp_seq_needed)) { | |
1030 | WARN_ON_ONCE(rcu_seq_state(sp->srcu_gp_seq)); | |
1031 | spin_unlock_irq(&sp->gp_lock); | |
1032 | mutex_unlock(&sp->srcu_gp_mutex); | |
dad81a20 PM |
1033 | return; |
1034 | } | |
1035 | idx = rcu_seq_state(READ_ONCE(sp->srcu_gp_seq)); | |
1036 | if (idx == SRCU_STATE_IDLE) | |
1037 | srcu_gp_start(sp); | |
da915ad5 PM |
1038 | spin_unlock_irq(&sp->gp_lock); |
1039 | if (idx != SRCU_STATE_IDLE) { | |
1040 | mutex_unlock(&sp->srcu_gp_mutex); | |
dad81a20 | 1041 | return; /* Someone else started the grace period. */ |
da915ad5 | 1042 | } |
dad81a20 PM |
1043 | } |
1044 | ||
1045 | if (rcu_seq_state(READ_ONCE(sp->srcu_gp_seq)) == SRCU_STATE_SCAN1) { | |
da915ad5 PM |
1046 | idx = 1 ^ (sp->srcu_idx & 1); |
1047 | if (!try_check_zero(sp, idx, 1)) { | |
1048 | mutex_unlock(&sp->srcu_gp_mutex); | |
dad81a20 | 1049 | return; /* readers present, retry later. */ |
da915ad5 | 1050 | } |
dad81a20 PM |
1051 | srcu_flip(sp); |
1052 | rcu_seq_set_state(&sp->srcu_gp_seq, SRCU_STATE_SCAN2); | |
1053 | } | |
1054 | ||
1055 | if (rcu_seq_state(READ_ONCE(sp->srcu_gp_seq)) == SRCU_STATE_SCAN2) { | |
1056 | ||
1057 | /* | |
1058 | * SRCU read-side critical sections are normally short, | |
1059 | * so check at least twice in quick succession after a flip. | |
1060 | */ | |
da915ad5 PM |
1061 | idx = 1 ^ (sp->srcu_idx & 1); |
1062 | if (!try_check_zero(sp, idx, 2)) { | |
1063 | mutex_unlock(&sp->srcu_gp_mutex); | |
1064 | return; /* readers present, retry later. */ | |
1065 | } | |
1066 | srcu_gp_end(sp); /* Releases ->srcu_gp_mutex. */ | |
dad81a20 PM |
1067 | } |
1068 | } | |
1069 | ||
1070 | /* | |
1071 | * Invoke a limited number of SRCU callbacks that have passed through | |
1072 | * their grace period. If there are more to do, SRCU will reschedule | |
1073 | * the workqueue. Note that needed memory barriers have been executed | |
1074 | * in this task's context by srcu_readers_active_idx_check(). | |
1075 | */ | |
da915ad5 | 1076 | static void srcu_invoke_callbacks(struct work_struct *work) |
dad81a20 | 1077 | { |
da915ad5 | 1078 | bool more; |
dad81a20 PM |
1079 | struct rcu_cblist ready_cbs; |
1080 | struct rcu_head *rhp; | |
da915ad5 PM |
1081 | struct srcu_data *sdp; |
1082 | struct srcu_struct *sp; | |
dad81a20 | 1083 | |
da915ad5 PM |
1084 | sdp = container_of(work, struct srcu_data, work.work); |
1085 | sp = sdp->sp; | |
dad81a20 | 1086 | rcu_cblist_init(&ready_cbs); |
da915ad5 PM |
1087 | spin_lock_irq(&sdp->lock); |
1088 | smp_mb(); /* Old grace periods before callback invocation! */ | |
1089 | rcu_segcblist_advance(&sdp->srcu_cblist, | |
1090 | rcu_seq_current(&sp->srcu_gp_seq)); | |
1091 | if (sdp->srcu_cblist_invoking || | |
1092 | !rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) { | |
1093 | spin_unlock_irq(&sdp->lock); | |
1094 | return; /* Someone else on the job or nothing to do. */ | |
1095 | } | |
1096 | ||
1097 | /* We are on the job! Extract and invoke ready callbacks. */ | |
1098 | sdp->srcu_cblist_invoking = true; | |
1099 | rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs); | |
1100 | spin_unlock_irq(&sdp->lock); | |
dad81a20 PM |
1101 | rhp = rcu_cblist_dequeue(&ready_cbs); |
1102 | for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) { | |
1103 | local_bh_disable(); | |
1104 | rhp->func(rhp); | |
1105 | local_bh_enable(); | |
1106 | } | |
da915ad5 PM |
1107 | |
1108 | /* | |
1109 | * Update counts, accelerate new callbacks, and if needed, | |
1110 | * schedule another round of callback invocation. | |
1111 | */ | |
1112 | spin_lock_irq(&sdp->lock); | |
1113 | rcu_segcblist_insert_count(&sdp->srcu_cblist, &ready_cbs); | |
1114 | (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, | |
1115 | rcu_seq_snap(&sp->srcu_gp_seq)); | |
1116 | sdp->srcu_cblist_invoking = false; | |
1117 | more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist); | |
1118 | spin_unlock_irq(&sdp->lock); | |
1119 | if (more) | |
1120 | srcu_schedule_cbs_sdp(sdp, 0); | |
dad81a20 PM |
1121 | } |
1122 | ||
1123 | /* | |
1124 | * Finished one round of SRCU grace period. Start another if there are | |
1125 | * more SRCU callbacks queued, otherwise put SRCU into not-running state. | |
1126 | */ | |
1127 | static void srcu_reschedule(struct srcu_struct *sp, unsigned long delay) | |
1128 | { | |
da915ad5 | 1129 | bool pushgp = true; |
dad81a20 | 1130 | |
da915ad5 PM |
1131 | spin_lock_irq(&sp->gp_lock); |
1132 | if (ULONG_CMP_GE(sp->srcu_gp_seq, sp->srcu_gp_seq_needed)) { | |
1133 | if (!WARN_ON_ONCE(rcu_seq_state(sp->srcu_gp_seq))) { | |
1134 | /* All requests fulfilled, time to go idle. */ | |
1135 | pushgp = false; | |
1136 | } | |
1137 | } else if (!rcu_seq_state(sp->srcu_gp_seq)) { | |
1138 | /* Outstanding request and no GP. Start one. */ | |
1139 | srcu_gp_start(sp); | |
dad81a20 | 1140 | } |
da915ad5 | 1141 | spin_unlock_irq(&sp->gp_lock); |
dad81a20 | 1142 | |
da915ad5 | 1143 | if (pushgp) |
dad81a20 PM |
1144 | queue_delayed_work(system_power_efficient_wq, &sp->work, delay); |
1145 | } | |
1146 | ||
1147 | /* | |
1148 | * This is the work-queue function that handles SRCU grace periods. | |
1149 | */ | |
1150 | void process_srcu(struct work_struct *work) | |
1151 | { | |
1152 | struct srcu_struct *sp; | |
1153 | ||
1154 | sp = container_of(work, struct srcu_struct, work.work); | |
1155 | ||
da915ad5 | 1156 | srcu_advance_state(sp); |
1e9a038b | 1157 | srcu_reschedule(sp, srcu_get_delay(sp)); |
dad81a20 PM |
1158 | } |
1159 | EXPORT_SYMBOL_GPL(process_srcu); | |
7f6733c3 PM |
1160 | |
1161 | void srcutorture_get_gp_data(enum rcutorture_type test_type, | |
1e9a038b PM |
1162 | struct srcu_struct *sp, int *flags, |
1163 | unsigned long *gpnum, unsigned long *completed) | |
7f6733c3 PM |
1164 | { |
1165 | if (test_type != SRCU_FLAVOR) | |
1166 | return; | |
1167 | *flags = 0; | |
1168 | *completed = rcu_seq_ctr(sp->srcu_gp_seq); | |
1169 | *gpnum = rcu_seq_ctr(sp->srcu_gp_seq_needed); | |
1170 | } | |
1171 | EXPORT_SYMBOL_GPL(srcutorture_get_gp_data); | |
1f4f6da1 PM |
1172 | |
1173 | static int __init srcu_bootup_announce(void) | |
1174 | { | |
1175 | pr_info("Hierarchical SRCU implementation.\n"); | |
0c8e0e3c PM |
1176 | if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF) |
1177 | pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff); | |
1f4f6da1 PM |
1178 | return 0; |
1179 | } | |
1180 | early_initcall(srcu_bootup_announce); |