Commit | Line | Data |
---|---|---|
e7ee1501 | 1 | // SPDX-License-Identifier: GPL-2.0+ |
dad81a20 PM |
2 | /* |
3 | * Sleepable Read-Copy Update mechanism for mutual exclusion. | |
4 | * | |
dad81a20 PM |
5 | * Copyright (C) IBM Corporation, 2006 |
6 | * Copyright (C) Fujitsu, 2012 | |
7 | * | |
e7ee1501 | 8 | * Author: Paul McKenney <paulmck@linux.ibm.com> |
dad81a20 PM |
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 | ||
a7538352 JP |
16 | #define pr_fmt(fmt) "rcu: " fmt |
17 | ||
dad81a20 PM |
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> | |
22607d66 | 26 | #include <linux/module.h> |
dad81a20 PM |
27 | #include <linux/srcu.h> |
28 | ||
dad81a20 | 29 | #include "rcu.h" |
45753c5f | 30 | #include "rcu_segcblist.h" |
dad81a20 | 31 | |
0c8e0e3c PM |
32 | /* Holdoff in nanoseconds for auto-expediting. */ |
33 | #define DEFAULT_SRCU_EXP_HOLDOFF (25 * 1000) | |
34 | static ulong exp_holdoff = DEFAULT_SRCU_EXP_HOLDOFF; | |
22607d66 PM |
35 | module_param(exp_holdoff, ulong, 0444); |
36 | ||
c350c008 PM |
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 | ||
e0fcba9a PM |
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 | ||
da915ad5 | 45 | static void srcu_invoke_callbacks(struct work_struct *work); |
aacb5d91 | 46 | static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay); |
0d8a1e83 | 47 | static void process_srcu(struct work_struct *work); |
e81baf4c | 48 | static void srcu_delay_timer(struct timer_list *t); |
da915ad5 | 49 | |
d6331980 PM |
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 | ||
da915ad5 PM |
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 | */ | |
aacb5d91 | 83 | static void init_srcu_struct_nodes(struct srcu_struct *ssp, bool is_static) |
dad81a20 | 84 | { |
da915ad5 PM |
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. */ | |
aacb5d91 | 94 | ssp->level[0] = &ssp->node[0]; |
da915ad5 | 95 | for (i = 1; i < rcu_num_lvls; i++) |
aacb5d91 | 96 | ssp->level[i] = ssp->level[i - 1] + num_rcu_lvl[i - 1]; |
da915ad5 PM |
97 | rcu_init_levelspread(levelspread, num_rcu_lvl); |
98 | ||
99 | /* Each pass through this loop initializes one srcu_node structure. */ | |
aacb5d91 | 100 | srcu_for_each_node_breadth_first(ssp, snp) { |
d6331980 | 101 | spin_lock_init(&ACCESS_PRIVATE(snp, lock)); |
c7e88067 PM |
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++) { | |
da915ad5 | 105 | snp->srcu_have_cbs[i] = 0; |
c7e88067 PM |
106 | snp->srcu_data_have_cbs[i] = 0; |
107 | } | |
1e9a038b | 108 | snp->srcu_gp_seq_needed_exp = 0; |
da915ad5 PM |
109 | snp->grplo = -1; |
110 | snp->grphi = -1; | |
aacb5d91 | 111 | if (snp == &ssp->node[0]) { |
da915ad5 PM |
112 | /* Root node, special case. */ |
113 | snp->srcu_parent = NULL; | |
114 | continue; | |
115 | } | |
116 | ||
117 | /* Non-root node. */ | |
aacb5d91 | 118 | if (snp == ssp->level[level + 1]) |
da915ad5 | 119 | level++; |
aacb5d91 PM |
120 | snp->srcu_parent = ssp->level[level - 1] + |
121 | (snp - ssp->level[level]) / | |
da915ad5 PM |
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; | |
aacb5d91 | 132 | snp_first = ssp->level[level]; |
da915ad5 | 133 | for_each_possible_cpu(cpu) { |
aacb5d91 | 134 | sdp = per_cpu_ptr(ssp->sda, cpu); |
d6331980 | 135 | spin_lock_init(&ACCESS_PRIVATE(sdp, lock)); |
da915ad5 PM |
136 | rcu_segcblist_init(&sdp->srcu_cblist); |
137 | sdp->srcu_cblist_invoking = false; | |
aacb5d91 PM |
138 | sdp->srcu_gp_seq_needed = ssp->srcu_gp_seq; |
139 | sdp->srcu_gp_seq_needed_exp = ssp->srcu_gp_seq; | |
da915ad5 PM |
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; | |
e81baf4c SAS |
147 | INIT_WORK(&sdp->work, srcu_invoke_callbacks); |
148 | timer_setup(&sdp->delay_work, srcu_delay_timer, 0); | |
aacb5d91 | 149 | sdp->ssp = ssp; |
c7e88067 | 150 | sdp->grpmask = 1 << (cpu - sdp->mynode->grplo); |
da915ad5 PM |
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 | */ | |
aacb5d91 | 168 | static int init_srcu_struct_fields(struct srcu_struct *ssp, bool is_static) |
da915ad5 | 169 | { |
aacb5d91 PM |
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); | |
da915ad5 | 178 | if (!is_static) |
aacb5d91 PM |
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; | |
dad81a20 PM |
185 | } |
186 | ||
187 | #ifdef CONFIG_DEBUG_LOCK_ALLOC | |
188 | ||
aacb5d91 | 189 | int __init_srcu_struct(struct srcu_struct *ssp, const char *name, |
dad81a20 PM |
190 | struct lock_class_key *key) |
191 | { | |
192 | /* Don't re-initialize a lock while it is held. */ | |
aacb5d91 PM |
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); | |
dad81a20 PM |
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 | |
aacb5d91 | 204 | * @ssp: structure to initialize. |
dad81a20 PM |
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 | */ | |
aacb5d91 | 210 | int init_srcu_struct(struct srcu_struct *ssp) |
dad81a20 | 211 | { |
aacb5d91 PM |
212 | spin_lock_init(&ACCESS_PRIVATE(ssp, lock)); |
213 | return init_srcu_struct_fields(ssp, false); | |
dad81a20 PM |
214 | } |
215 | EXPORT_SYMBOL_GPL(init_srcu_struct); | |
216 | ||
217 | #endif /* #else #ifdef CONFIG_DEBUG_LOCK_ALLOC */ | |
218 | ||
219 | /* | |
da915ad5 PM |
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 | |
aacb5d91 | 223 | * to each update-side SRCU primitive. Use ssp->lock, which -is- |
da915ad5 PM |
224 | * compile-time initialized, to resolve races involving multiple |
225 | * CPUs trying to garner first-use privileges. | |
226 | */ | |
aacb5d91 | 227 | static void check_init_srcu_struct(struct srcu_struct *ssp) |
da915ad5 PM |
228 | { |
229 | unsigned long flags; | |
230 | ||
da915ad5 | 231 | /* The smp_load_acquire() pairs with the smp_store_release(). */ |
aacb5d91 | 232 | if (!rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq_needed))) /*^^^*/ |
da915ad5 | 233 | return; /* Already initialized. */ |
aacb5d91 PM |
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); | |
da915ad5 PM |
237 | return; |
238 | } | |
aacb5d91 PM |
239 | init_srcu_struct_fields(ssp, true); |
240 | spin_unlock_irqrestore_rcu_node(ssp, flags); | |
da915ad5 PM |
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. | |
dad81a20 | 246 | */ |
aacb5d91 | 247 | static unsigned long srcu_readers_lock_idx(struct srcu_struct *ssp, int idx) |
dad81a20 PM |
248 | { |
249 | int cpu; | |
250 | unsigned long sum = 0; | |
251 | ||
252 | for_each_possible_cpu(cpu) { | |
aacb5d91 | 253 | struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu); |
dad81a20 | 254 | |
da915ad5 | 255 | sum += READ_ONCE(cpuc->srcu_lock_count[idx]); |
dad81a20 PM |
256 | } |
257 | return sum; | |
258 | } | |
259 | ||
260 | /* | |
da915ad5 PM |
261 | * Returns approximate total of the readers' ->srcu_unlock_count[] values |
262 | * for the rank of per-CPU counters specified by idx. | |
dad81a20 | 263 | */ |
aacb5d91 | 264 | static unsigned long srcu_readers_unlock_idx(struct srcu_struct *ssp, int idx) |
dad81a20 PM |
265 | { |
266 | int cpu; | |
267 | unsigned long sum = 0; | |
268 | ||
269 | for_each_possible_cpu(cpu) { | |
aacb5d91 | 270 | struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu); |
dad81a20 | 271 | |
da915ad5 | 272 | sum += READ_ONCE(cpuc->srcu_unlock_count[idx]); |
dad81a20 PM |
273 | } |
274 | return sum; | |
275 | } | |
276 | ||
277 | /* | |
278 | * Return true if the number of pre-existing readers is determined to | |
279 | * be zero. | |
280 | */ | |
aacb5d91 | 281 | static bool srcu_readers_active_idx_check(struct srcu_struct *ssp, int idx) |
dad81a20 PM |
282 | { |
283 | unsigned long unlocks; | |
284 | ||
aacb5d91 | 285 | unlocks = srcu_readers_unlock_idx(ssp, idx); |
dad81a20 PM |
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 | * | |
881ec9d2 PM |
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. | |
dad81a20 | 320 | */ |
aacb5d91 | 321 | return srcu_readers_lock_idx(ssp, idx) == unlocks; |
dad81a20 PM |
322 | } |
323 | ||
324 | /** | |
325 | * srcu_readers_active - returns true if there are readers. and false | |
326 | * otherwise | |
aacb5d91 | 327 | * @ssp: which srcu_struct to count active readers (holding srcu_read_lock). |
dad81a20 PM |
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 | */ | |
aacb5d91 | 333 | static bool srcu_readers_active(struct srcu_struct *ssp) |
dad81a20 PM |
334 | { |
335 | int cpu; | |
336 | unsigned long sum = 0; | |
337 | ||
338 | for_each_possible_cpu(cpu) { | |
aacb5d91 | 339 | struct srcu_data *cpuc = per_cpu_ptr(ssp->sda, cpu); |
dad81a20 | 340 | |
da915ad5 PM |
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]); | |
dad81a20 PM |
345 | } |
346 | return sum; | |
347 | } | |
348 | ||
349 | #define SRCU_INTERVAL 1 | |
350 | ||
1e9a038b PM |
351 | /* |
352 | * Return grace-period delay, zero if there are expedited grace | |
353 | * periods pending, SRCU_INTERVAL otherwise. | |
354 | */ | |
aacb5d91 | 355 | static unsigned long srcu_get_delay(struct srcu_struct *ssp) |
1e9a038b | 356 | { |
aacb5d91 PM |
357 | if (ULONG_CMP_LT(READ_ONCE(ssp->srcu_gp_seq), |
358 | READ_ONCE(ssp->srcu_gp_seq_needed_exp))) | |
1e9a038b PM |
359 | return 0; |
360 | return SRCU_INTERVAL; | |
361 | } | |
362 | ||
f5ad3991 PM |
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) | |
dad81a20 | 371 | { |
da915ad5 PM |
372 | int cpu; |
373 | ||
aacb5d91 | 374 | if (WARN_ON(!srcu_get_delay(ssp))) |
f7194ac3 | 375 | return; /* Just leak it! */ |
aacb5d91 | 376 | if (WARN_ON(srcu_readers_active(ssp))) |
f7194ac3 | 377 | return; /* Just leak it! */ |
f5ad3991 | 378 | flush_delayed_work(&ssp->work); |
e81baf4c SAS |
379 | for_each_possible_cpu(cpu) { |
380 | struct srcu_data *sdp = per_cpu_ptr(ssp->sda, cpu); | |
381 | ||
f5ad3991 PM |
382 | del_timer_sync(&sdp->delay_work); |
383 | flush_work(&sdp->work); | |
5cdfd174 PM |
384 | if (WARN_ON(rcu_segcblist_n_cbs(&sdp->srcu_cblist))) |
385 | return; /* Forgot srcu_barrier(), so just leak it! */ | |
e81baf4c | 386 | } |
aacb5d91 PM |
387 | if (WARN_ON(rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) != SRCU_STATE_IDLE) || |
388 | WARN_ON(srcu_readers_active(ssp))) { | |
a7538352 | 389 | pr_info("%s: Active srcu_struct %p state: %d\n", |
aacb5d91 | 390 | __func__, ssp, rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq))); |
dad81a20 PM |
391 | return; /* Caller forgot to stop doing call_srcu()? */ |
392 | } | |
aacb5d91 PM |
393 | free_percpu(ssp->sda); |
394 | ssp->sda = NULL; | |
dad81a20 | 395 | } |
f5ad3991 | 396 | EXPORT_SYMBOL_GPL(cleanup_srcu_struct); |
dad81a20 PM |
397 | |
398 | /* | |
399 | * Counts the new reader in the appropriate per-CPU element of the | |
cdf7abc4 | 400 | * srcu_struct. |
dad81a20 PM |
401 | * Returns an index that must be passed to the matching srcu_read_unlock(). |
402 | */ | |
aacb5d91 | 403 | int __srcu_read_lock(struct srcu_struct *ssp) |
dad81a20 PM |
404 | { |
405 | int idx; | |
406 | ||
aacb5d91 PM |
407 | idx = READ_ONCE(ssp->srcu_idx) & 0x1; |
408 | this_cpu_inc(ssp->sda->srcu_lock_count[idx]); | |
dad81a20 PM |
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(). | |
dad81a20 | 418 | */ |
aacb5d91 | 419 | void __srcu_read_unlock(struct srcu_struct *ssp, int idx) |
dad81a20 PM |
420 | { |
421 | smp_mb(); /* C */ /* Avoid leaking the critical section. */ | |
aacb5d91 | 422 | this_cpu_inc(ssp->sda->srcu_unlock_count[idx]); |
dad81a20 PM |
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 | */ | |
aacb5d91 | 438 | static void srcu_gp_start(struct srcu_struct *ssp) |
dad81a20 | 439 | { |
aacb5d91 | 440 | struct srcu_data *sdp = this_cpu_ptr(ssp->sda); |
dad81a20 PM |
441 | int state; |
442 | ||
aacb5d91 PM |
443 | lockdep_assert_held(&ACCESS_PRIVATE(ssp, lock)); |
444 | WARN_ON_ONCE(ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)); | |
eb4c2382 | 445 | spin_lock_rcu_node(sdp); /* Interrupts already disabled. */ |
da915ad5 | 446 | rcu_segcblist_advance(&sdp->srcu_cblist, |
aacb5d91 | 447 | rcu_seq_current(&ssp->srcu_gp_seq)); |
da915ad5 | 448 | (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, |
aacb5d91 | 449 | rcu_seq_snap(&ssp->srcu_gp_seq)); |
eb4c2382 | 450 | spin_unlock_rcu_node(sdp); /* Interrupts remain disabled. */ |
2da4b2a7 | 451 | smp_mb(); /* Order prior store to ->srcu_gp_seq_needed vs. GP start. */ |
aacb5d91 PM |
452 | rcu_seq_start(&ssp->srcu_gp_seq); |
453 | state = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)); | |
dad81a20 PM |
454 | WARN_ON_ONCE(state != SRCU_STATE_SCAN1); |
455 | } | |
456 | ||
da915ad5 | 457 | |
e81baf4c | 458 | static void srcu_delay_timer(struct timer_list *t) |
da915ad5 | 459 | { |
e81baf4c | 460 | struct srcu_data *sdp = container_of(t, struct srcu_data, delay_work); |
da915ad5 | 461 | |
e81baf4c | 462 | queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work); |
da915ad5 PM |
463 | } |
464 | ||
e81baf4c | 465 | static void srcu_queue_delayed_work_on(struct srcu_data *sdp, |
da915ad5 PM |
466 | unsigned long delay) |
467 | { | |
e81baf4c SAS |
468 | if (!delay) { |
469 | queue_work_on(sdp->cpu, rcu_gp_wq, &sdp->work); | |
470 | return; | |
471 | } | |
da915ad5 | 472 | |
e81baf4c | 473 | timer_reduce(&sdp->delay_work, jiffies + delay); |
da915ad5 PM |
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 | { | |
e81baf4c | 482 | srcu_queue_delayed_work_on(sdp, delay); |
da915ad5 PM |
483 | } |
484 | ||
485 | /* | |
486 | * Schedule callback invocation for all srcu_data structures associated | |
c7e88067 PM |
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. | |
da915ad5 | 490 | */ |
aacb5d91 | 491 | static void srcu_schedule_cbs_snp(struct srcu_struct *ssp, struct srcu_node *snp, |
1e9a038b | 492 | unsigned long mask, unsigned long delay) |
da915ad5 PM |
493 | { |
494 | int cpu; | |
495 | ||
c7e88067 PM |
496 | for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) { |
497 | if (!(mask & (1 << (cpu - snp->grplo)))) | |
498 | continue; | |
aacb5d91 | 499 | srcu_schedule_cbs_sdp(per_cpu_ptr(ssp->sda, cpu), delay); |
c7e88067 | 500 | } |
da915ad5 PM |
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 | */ | |
aacb5d91 | 512 | static void srcu_gp_end(struct srcu_struct *ssp) |
da915ad5 | 513 | { |
1e9a038b | 514 | unsigned long cbdelay; |
da915ad5 | 515 | bool cbs; |
8ddbd883 | 516 | bool last_lvl; |
c350c008 PM |
517 | int cpu; |
518 | unsigned long flags; | |
da915ad5 PM |
519 | unsigned long gpseq; |
520 | int idx; | |
c7e88067 | 521 | unsigned long mask; |
c350c008 | 522 | struct srcu_data *sdp; |
da915ad5 PM |
523 | struct srcu_node *snp; |
524 | ||
525 | /* Prevent more than one additional grace period. */ | |
aacb5d91 | 526 | mutex_lock(&ssp->srcu_cb_mutex); |
da915ad5 PM |
527 | |
528 | /* End the current grace period. */ | |
aacb5d91 PM |
529 | spin_lock_irq_rcu_node(ssp); |
530 | idx = rcu_seq_state(ssp->srcu_gp_seq); | |
da915ad5 | 531 | WARN_ON_ONCE(idx != SRCU_STATE_SCAN2); |
aacb5d91 PM |
532 | cbdelay = srcu_get_delay(ssp); |
533 | 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); | |
da915ad5 PM |
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); | |
aacb5d91 | 544 | srcu_for_each_node_breadth_first(ssp, snp) { |
d6331980 | 545 | spin_lock_irq_rcu_node(snp); |
da915ad5 | 546 | cbs = false; |
aacb5d91 | 547 | last_lvl = snp >= ssp->level[rcu_num_lvls - 1]; |
8ddbd883 | 548 | if (last_lvl) |
da915ad5 PM |
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); | |
1e9a038b PM |
552 | if (ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, gpseq)) |
553 | snp->srcu_gp_seq_needed_exp = gpseq; | |
c7e88067 PM |
554 | mask = snp->srcu_data_have_cbs[idx]; |
555 | snp->srcu_data_have_cbs[idx] = 0; | |
d6331980 | 556 | spin_unlock_irq_rcu_node(snp); |
a3883df3 | 557 | if (cbs) |
aacb5d91 | 558 | srcu_schedule_cbs_snp(ssp, snp, mask, cbdelay); |
c350c008 PM |
559 | |
560 | /* Occasionally prevent srcu_data counter wrap. */ | |
8ddbd883 | 561 | if (!(gpseq & counter_wrap_check) && last_lvl) |
c350c008 | 562 | for (cpu = snp->grplo; cpu <= snp->grphi; cpu++) { |
aacb5d91 | 563 | sdp = per_cpu_ptr(ssp->sda, cpu); |
d6331980 | 564 | spin_lock_irqsave_rcu_node(sdp, flags); |
c350c008 PM |
565 | if (ULONG_CMP_GE(gpseq, |
566 | sdp->srcu_gp_seq_needed + 100)) | |
567 | sdp->srcu_gp_seq_needed = gpseq; | |
a35d13ec II |
568 | if (ULONG_CMP_GE(gpseq, |
569 | sdp->srcu_gp_seq_needed_exp + 100)) | |
570 | sdp->srcu_gp_seq_needed_exp = gpseq; | |
d6331980 | 571 | spin_unlock_irqrestore_rcu_node(sdp, flags); |
c350c008 | 572 | } |
da915ad5 PM |
573 | } |
574 | ||
575 | /* Callback initiation done, allow grace periods after next. */ | |
aacb5d91 | 576 | mutex_unlock(&ssp->srcu_cb_mutex); |
da915ad5 PM |
577 | |
578 | /* Start a new grace period if needed. */ | |
aacb5d91 PM |
579 | spin_lock_irq_rcu_node(ssp); |
580 | gpseq = rcu_seq_current(&ssp->srcu_gp_seq); | |
da915ad5 | 581 | if (!rcu_seq_state(gpseq) && |
aacb5d91 PM |
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); | |
da915ad5 | 586 | } else { |
aacb5d91 | 587 | spin_unlock_irq_rcu_node(ssp); |
da915ad5 PM |
588 | } |
589 | } | |
590 | ||
1e9a038b PM |
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 | */ | |
aacb5d91 | 598 | static void srcu_funnel_exp_start(struct srcu_struct *ssp, struct srcu_node *snp, |
1e9a038b PM |
599 | unsigned long s) |
600 | { | |
601 | unsigned long flags; | |
602 | ||
603 | for (; snp != NULL; snp = snp->srcu_parent) { | |
aacb5d91 | 604 | if (rcu_seq_done(&ssp->srcu_gp_seq, s) || |
1e9a038b PM |
605 | ULONG_CMP_GE(READ_ONCE(snp->srcu_gp_seq_needed_exp), s)) |
606 | return; | |
d6331980 | 607 | spin_lock_irqsave_rcu_node(snp, flags); |
1e9a038b | 608 | if (ULONG_CMP_GE(snp->srcu_gp_seq_needed_exp, s)) { |
d6331980 | 609 | spin_unlock_irqrestore_rcu_node(snp, flags); |
1e9a038b PM |
610 | return; |
611 | } | |
612 | WRITE_ONCE(snp->srcu_gp_seq_needed_exp, s); | |
d6331980 | 613 | spin_unlock_irqrestore_rcu_node(snp, flags); |
1e9a038b | 614 | } |
aacb5d91 PM |
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); | |
1e9a038b PM |
619 | } |
620 | ||
da915ad5 PM |
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. | |
17294ce6 PM |
627 | * |
628 | * Note that this function also does the work of srcu_funnel_exp_start(), | |
629 | * in some cases by directly invoking it. | |
da915ad5 | 630 | */ |
aacb5d91 | 631 | static void srcu_funnel_gp_start(struct srcu_struct *ssp, struct srcu_data *sdp, |
1e9a038b | 632 | unsigned long s, bool do_norm) |
da915ad5 PM |
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) { | |
aacb5d91 | 641 | if (rcu_seq_done(&ssp->srcu_gp_seq, s) && snp != sdp->mynode) |
da915ad5 | 642 | return; /* GP already done and CBs recorded. */ |
d6331980 | 643 | spin_lock_irqsave_rcu_node(snp, flags); |
da915ad5 PM |
644 | if (ULONG_CMP_GE(snp->srcu_have_cbs[idx], s)) { |
645 | snp_seq = snp->srcu_have_cbs[idx]; | |
c7e88067 PM |
646 | if (snp == sdp->mynode && snp_seq == s) |
647 | snp->srcu_data_have_cbs[idx] |= sdp->grpmask; | |
d6331980 | 648 | spin_unlock_irqrestore_rcu_node(snp, flags); |
da915ad5 | 649 | if (snp == sdp->mynode && snp_seq != s) { |
1e9a038b PM |
650 | srcu_schedule_cbs_sdp(sdp, do_norm |
651 | ? SRCU_INTERVAL | |
652 | : 0); | |
653 | return; | |
da915ad5 | 654 | } |
1e9a038b | 655 | if (!do_norm) |
aacb5d91 | 656 | srcu_funnel_exp_start(ssp, snp, s); |
da915ad5 PM |
657 | return; |
658 | } | |
659 | snp->srcu_have_cbs[idx] = s; | |
c7e88067 PM |
660 | if (snp == sdp->mynode) |
661 | snp->srcu_data_have_cbs[idx] |= sdp->grpmask; | |
1e9a038b PM |
662 | if (!do_norm && ULONG_CMP_LT(snp->srcu_gp_seq_needed_exp, s)) |
663 | snp->srcu_gp_seq_needed_exp = s; | |
d6331980 | 664 | spin_unlock_irqrestore_rcu_node(snp, flags); |
da915ad5 PM |
665 | } |
666 | ||
667 | /* Top of tree, must ensure the grace period will be started. */ | |
aacb5d91 PM |
668 | spin_lock_irqsave_rcu_node(ssp, flags); |
669 | if (ULONG_CMP_LT(ssp->srcu_gp_seq_needed, s)) { | |
da915ad5 PM |
670 | /* |
671 | * Record need for grace period s. Pair with load | |
672 | * acquire setting up for initialization. | |
673 | */ | |
aacb5d91 | 674 | smp_store_release(&ssp->srcu_gp_seq_needed, s); /*^^^*/ |
da915ad5 | 675 | } |
aacb5d91 PM |
676 | if (!do_norm && ULONG_CMP_LT(ssp->srcu_gp_seq_needed_exp, s)) |
677 | ssp->srcu_gp_seq_needed_exp = s; | |
da915ad5 PM |
678 | |
679 | /* If grace period not already done and none in progress, start it. */ | |
aacb5d91 PM |
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); | |
e0fcba9a | 684 | if (likely(srcu_init_done)) |
aacb5d91 PM |
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); | |
da915ad5 | 689 | } |
aacb5d91 | 690 | spin_unlock_irqrestore_rcu_node(ssp, flags); |
da915ad5 PM |
691 | } |
692 | ||
dad81a20 PM |
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. | |
da915ad5 | 696 | * The caller must ensure that ->srcu_idx is not changed while checking. |
dad81a20 | 697 | */ |
aacb5d91 | 698 | static bool try_check_zero(struct srcu_struct *ssp, int idx, int trycount) |
dad81a20 PM |
699 | { |
700 | for (;;) { | |
aacb5d91 | 701 | if (srcu_readers_active_idx_check(ssp, idx)) |
dad81a20 | 702 | return true; |
aacb5d91 | 703 | if (--trycount + !srcu_get_delay(ssp) <= 0) |
dad81a20 PM |
704 | return false; |
705 | udelay(SRCU_RETRY_CHECK_DELAY); | |
706 | } | |
707 | } | |
708 | ||
709 | /* | |
da915ad5 PM |
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 | |
dad81a20 PM |
712 | * us to wait for pre-existing readers in a starvation-free manner. |
713 | */ | |
aacb5d91 | 714 | static void srcu_flip(struct srcu_struct *ssp) |
dad81a20 | 715 | { |
881ec9d2 PM |
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 | ||
aacb5d91 | 726 | WRITE_ONCE(ssp->srcu_idx, ssp->srcu_idx + 1); |
dad81a20 PM |
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 | ||
2da4b2a7 PM |
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 | */ | |
aacb5d91 | 759 | static bool srcu_might_be_idle(struct srcu_struct *ssp) |
2da4b2a7 | 760 | { |
22607d66 | 761 | unsigned long curseq; |
2da4b2a7 PM |
762 | unsigned long flags; |
763 | struct srcu_data *sdp; | |
22607d66 | 764 | unsigned long t; |
2da4b2a7 PM |
765 | |
766 | /* If the local srcu_data structure has callbacks, not idle. */ | |
767 | local_irq_save(flags); | |
aacb5d91 | 768 | sdp = this_cpu_ptr(ssp->sda); |
2da4b2a7 PM |
769 | if (rcu_segcblist_pend_cbs(&sdp->srcu_cblist)) { |
770 | local_irq_restore(flags); | |
771 | return false; /* Callbacks already present, so not idle. */ | |
772 | } | |
773 | local_irq_restore(flags); | |
774 | ||
775 | /* | |
776 | * No local callbacks, so probabalistically probe global state. | |
777 | * Exact information would require acquiring locks, which would | |
778 | * kill scalability, hence the probabalistic nature of the probe. | |
779 | */ | |
22607d66 PM |
780 | |
781 | /* First, see if enough time has passed since the last GP. */ | |
782 | t = ktime_get_mono_fast_ns(); | |
783 | if (exp_holdoff == 0 || | |
aacb5d91 PM |
784 | time_in_range_open(t, ssp->srcu_last_gp_end, |
785 | ssp->srcu_last_gp_end + exp_holdoff)) | |
22607d66 PM |
786 | return false; /* Too soon after last GP. */ |
787 | ||
788 | /* Next, check for probable idleness. */ | |
aacb5d91 | 789 | curseq = rcu_seq_current(&ssp->srcu_gp_seq); |
2da4b2a7 | 790 | smp_mb(); /* Order ->srcu_gp_seq with ->srcu_gp_seq_needed. */ |
aacb5d91 | 791 | if (ULONG_CMP_LT(curseq, READ_ONCE(ssp->srcu_gp_seq_needed))) |
2da4b2a7 PM |
792 | return false; /* Grace period in progress, so not idle. */ |
793 | smp_mb(); /* Order ->srcu_gp_seq with prior access. */ | |
aacb5d91 | 794 | if (curseq != rcu_seq_current(&ssp->srcu_gp_seq)) |
2da4b2a7 PM |
795 | return false; /* GP # changed, so not idle. */ |
796 | return true; /* With reasonable probability, idle! */ | |
797 | } | |
798 | ||
a602538e PM |
799 | /* |
800 | * SRCU callback function to leak a callback. | |
801 | */ | |
802 | static void srcu_leak_callback(struct rcu_head *rhp) | |
803 | { | |
804 | } | |
805 | ||
dad81a20 | 806 | /* |
da915ad5 PM |
807 | * Enqueue an SRCU callback on the srcu_data structure associated with |
808 | * the current CPU and the specified srcu_struct structure, initiating | |
809 | * grace-period processing if it is not already running. | |
dad81a20 PM |
810 | * |
811 | * Note that all CPUs must agree that the grace period extended beyond | |
812 | * all pre-existing SRCU read-side critical section. On systems with | |
813 | * more than one CPU, this means that when "func()" is invoked, each CPU | |
814 | * is guaranteed to have executed a full memory barrier since the end of | |
815 | * its last corresponding SRCU read-side critical section whose beginning | |
5ef98a63 | 816 | * preceded the call to call_srcu(). It also means that each CPU executing |
dad81a20 | 817 | * an SRCU read-side critical section that continues beyond the start of |
5ef98a63 | 818 | * "func()" must have executed a memory barrier after the call_srcu() |
dad81a20 PM |
819 | * but before the beginning of that SRCU read-side critical section. |
820 | * Note that these guarantees include CPUs that are offline, idle, or | |
821 | * executing in user mode, as well as CPUs that are executing in the kernel. | |
822 | * | |
5ef98a63 | 823 | * Furthermore, if CPU A invoked call_srcu() and CPU B invoked the |
dad81a20 PM |
824 | * resulting SRCU callback function "func()", then both CPU A and CPU |
825 | * B are guaranteed to execute a full memory barrier during the time | |
5ef98a63 | 826 | * interval between the call to call_srcu() and the invocation of "func()". |
dad81a20 PM |
827 | * This guarantee applies even if CPU A and CPU B are the same CPU (but |
828 | * again only if the system has more than one CPU). | |
829 | * | |
830 | * Of course, these guarantees apply only for invocations of call_srcu(), | |
831 | * srcu_read_lock(), and srcu_read_unlock() that are all passed the same | |
832 | * srcu_struct structure. | |
833 | */ | |
11b00045 JB |
834 | static void __call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp, |
835 | rcu_callback_t func, bool do_norm) | |
dad81a20 PM |
836 | { |
837 | unsigned long flags; | |
0607ba84 | 838 | int idx; |
1e9a038b | 839 | bool needexp = false; |
da915ad5 PM |
840 | bool needgp = false; |
841 | unsigned long s; | |
842 | struct srcu_data *sdp; | |
843 | ||
aacb5d91 | 844 | check_init_srcu_struct(ssp); |
a602538e PM |
845 | if (debug_rcu_head_queue(rhp)) { |
846 | /* Probable double call_srcu(), so leak the callback. */ | |
847 | WRITE_ONCE(rhp->func, srcu_leak_callback); | |
848 | WARN_ONCE(1, "call_srcu(): Leaked duplicate callback\n"); | |
849 | return; | |
850 | } | |
da915ad5 | 851 | rhp->func = func; |
aacb5d91 | 852 | idx = srcu_read_lock(ssp); |
da915ad5 | 853 | local_irq_save(flags); |
aacb5d91 | 854 | sdp = this_cpu_ptr(ssp->sda); |
d6331980 | 855 | spin_lock_rcu_node(sdp); |
da915ad5 PM |
856 | rcu_segcblist_enqueue(&sdp->srcu_cblist, rhp, false); |
857 | rcu_segcblist_advance(&sdp->srcu_cblist, | |
aacb5d91 PM |
858 | rcu_seq_current(&ssp->srcu_gp_seq)); |
859 | s = rcu_seq_snap(&ssp->srcu_gp_seq); | |
da915ad5 PM |
860 | (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, s); |
861 | if (ULONG_CMP_LT(sdp->srcu_gp_seq_needed, s)) { | |
862 | sdp->srcu_gp_seq_needed = s; | |
863 | needgp = true; | |
dad81a20 | 864 | } |
1e9a038b PM |
865 | if (!do_norm && ULONG_CMP_LT(sdp->srcu_gp_seq_needed_exp, s)) { |
866 | sdp->srcu_gp_seq_needed_exp = s; | |
867 | needexp = true; | |
868 | } | |
d6331980 | 869 | spin_unlock_irqrestore_rcu_node(sdp, flags); |
da915ad5 | 870 | if (needgp) |
aacb5d91 | 871 | srcu_funnel_gp_start(ssp, sdp, s, do_norm); |
1e9a038b | 872 | else if (needexp) |
aacb5d91 PM |
873 | srcu_funnel_exp_start(ssp, sdp->mynode, s); |
874 | srcu_read_unlock(ssp, idx); | |
1e9a038b PM |
875 | } |
876 | ||
5a0465e1 PM |
877 | /** |
878 | * call_srcu() - Queue a callback for invocation after an SRCU grace period | |
aacb5d91 | 879 | * @ssp: srcu_struct in queue the callback |
27fdb35f | 880 | * @rhp: structure to be used for queueing the SRCU callback. |
5a0465e1 PM |
881 | * @func: function to be invoked after the SRCU grace period |
882 | * | |
883 | * The callback function will be invoked some time after a full SRCU | |
884 | * grace period elapses, in other words after all pre-existing SRCU | |
885 | * read-side critical sections have completed. However, the callback | |
886 | * function might well execute concurrently with other SRCU read-side | |
887 | * critical sections that started after call_srcu() was invoked. SRCU | |
888 | * read-side critical sections are delimited by srcu_read_lock() and | |
889 | * srcu_read_unlock(), and may be nested. | |
890 | * | |
891 | * The callback will be invoked from process context, but must nevertheless | |
892 | * be fast and must not block. | |
893 | */ | |
aacb5d91 | 894 | void call_srcu(struct srcu_struct *ssp, struct rcu_head *rhp, |
1e9a038b PM |
895 | rcu_callback_t func) |
896 | { | |
aacb5d91 | 897 | __call_srcu(ssp, rhp, func, true); |
dad81a20 PM |
898 | } |
899 | EXPORT_SYMBOL_GPL(call_srcu); | |
900 | ||
dad81a20 PM |
901 | /* |
902 | * Helper function for synchronize_srcu() and synchronize_srcu_expedited(). | |
903 | */ | |
aacb5d91 | 904 | static void __synchronize_srcu(struct srcu_struct *ssp, bool do_norm) |
dad81a20 PM |
905 | { |
906 | struct rcu_synchronize rcu; | |
dad81a20 | 907 | |
aacb5d91 | 908 | RCU_LOCKDEP_WARN(lock_is_held(&ssp->dep_map) || |
dad81a20 PM |
909 | lock_is_held(&rcu_bh_lock_map) || |
910 | lock_is_held(&rcu_lock_map) || | |
911 | lock_is_held(&rcu_sched_lock_map), | |
912 | "Illegal synchronize_srcu() in same-type SRCU (or in RCU) read-side critical section"); | |
913 | ||
914 | if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE) | |
915 | return; | |
916 | might_sleep(); | |
aacb5d91 | 917 | check_init_srcu_struct(ssp); |
dad81a20 | 918 | init_completion(&rcu.completion); |
da915ad5 | 919 | init_rcu_head_on_stack(&rcu.head); |
aacb5d91 | 920 | __call_srcu(ssp, &rcu.head, wakeme_after_rcu, do_norm); |
dad81a20 | 921 | wait_for_completion(&rcu.completion); |
da915ad5 | 922 | destroy_rcu_head_on_stack(&rcu.head); |
35732cf9 PM |
923 | |
924 | /* | |
925 | * Make sure that later code is ordered after the SRCU grace | |
d6331980 | 926 | * period. This pairs with the spin_lock_irq_rcu_node() |
35732cf9 PM |
927 | * in srcu_invoke_callbacks(). Unlike Tree RCU, this is needed |
928 | * because the current CPU might have been totally uninvolved with | |
929 | * (and thus unordered against) that grace period. | |
930 | */ | |
931 | smp_mb(); | |
dad81a20 PM |
932 | } |
933 | ||
934 | /** | |
935 | * synchronize_srcu_expedited - Brute-force SRCU grace period | |
aacb5d91 | 936 | * @ssp: srcu_struct with which to synchronize. |
dad81a20 PM |
937 | * |
938 | * Wait for an SRCU grace period to elapse, but be more aggressive about | |
939 | * spinning rather than blocking when waiting. | |
940 | * | |
941 | * Note that synchronize_srcu_expedited() has the same deadlock and | |
942 | * memory-ordering properties as does synchronize_srcu(). | |
943 | */ | |
aacb5d91 | 944 | void synchronize_srcu_expedited(struct srcu_struct *ssp) |
dad81a20 | 945 | { |
aacb5d91 | 946 | __synchronize_srcu(ssp, rcu_gp_is_normal()); |
dad81a20 PM |
947 | } |
948 | EXPORT_SYMBOL_GPL(synchronize_srcu_expedited); | |
949 | ||
950 | /** | |
951 | * synchronize_srcu - wait for prior SRCU read-side critical-section completion | |
aacb5d91 | 952 | * @ssp: srcu_struct with which to synchronize. |
dad81a20 PM |
953 | * |
954 | * Wait for the count to drain to zero of both indexes. To avoid the | |
955 | * possible starvation of synchronize_srcu(), it waits for the count of | |
da915ad5 PM |
956 | * the index=((->srcu_idx & 1) ^ 1) to drain to zero at first, |
957 | * and then flip the srcu_idx and wait for the count of the other index. | |
dad81a20 PM |
958 | * |
959 | * Can block; must be called from process context. | |
960 | * | |
961 | * Note that it is illegal to call synchronize_srcu() from the corresponding | |
962 | * SRCU read-side critical section; doing so will result in deadlock. | |
963 | * However, it is perfectly legal to call synchronize_srcu() on one | |
964 | * srcu_struct from some other srcu_struct's read-side critical section, | |
965 | * as long as the resulting graph of srcu_structs is acyclic. | |
966 | * | |
967 | * There are memory-ordering constraints implied by synchronize_srcu(). | |
968 | * On systems with more than one CPU, when synchronize_srcu() returns, | |
969 | * each CPU is guaranteed to have executed a full memory barrier since | |
6eb95cc4 | 970 | * the end of its last corresponding SRCU read-side critical section |
dad81a20 PM |
971 | * whose beginning preceded the call to synchronize_srcu(). In addition, |
972 | * each CPU having an SRCU read-side critical section that extends beyond | |
973 | * the return from synchronize_srcu() is guaranteed to have executed a | |
974 | * full memory barrier after the beginning of synchronize_srcu() and before | |
975 | * the beginning of that SRCU read-side critical section. Note that these | |
976 | * guarantees include CPUs that are offline, idle, or executing in user mode, | |
977 | * as well as CPUs that are executing in the kernel. | |
978 | * | |
979 | * Furthermore, if CPU A invoked synchronize_srcu(), which returned | |
980 | * to its caller on CPU B, then both CPU A and CPU B are guaranteed | |
981 | * to have executed a full memory barrier during the execution of | |
982 | * synchronize_srcu(). This guarantee applies even if CPU A and CPU B | |
983 | * are the same CPU, but again only if the system has more than one CPU. | |
984 | * | |
985 | * Of course, these memory-ordering guarantees apply only when | |
986 | * synchronize_srcu(), srcu_read_lock(), and srcu_read_unlock() are | |
987 | * passed the same srcu_struct structure. | |
2da4b2a7 PM |
988 | * |
989 | * If SRCU is likely idle, expedite the first request. This semantic | |
990 | * was provided by Classic SRCU, and is relied upon by its users, so TREE | |
991 | * SRCU must also provide it. Note that detecting idleness is heuristic | |
992 | * and subject to both false positives and negatives. | |
dad81a20 | 993 | */ |
aacb5d91 | 994 | void synchronize_srcu(struct srcu_struct *ssp) |
dad81a20 | 995 | { |
aacb5d91 PM |
996 | if (srcu_might_be_idle(ssp) || rcu_gp_is_expedited()) |
997 | synchronize_srcu_expedited(ssp); | |
dad81a20 | 998 | else |
aacb5d91 | 999 | __synchronize_srcu(ssp, true); |
dad81a20 PM |
1000 | } |
1001 | EXPORT_SYMBOL_GPL(synchronize_srcu); | |
1002 | ||
da915ad5 PM |
1003 | /* |
1004 | * Callback function for srcu_barrier() use. | |
1005 | */ | |
1006 | static void srcu_barrier_cb(struct rcu_head *rhp) | |
1007 | { | |
1008 | struct srcu_data *sdp; | |
aacb5d91 | 1009 | struct srcu_struct *ssp; |
da915ad5 PM |
1010 | |
1011 | sdp = container_of(rhp, struct srcu_data, srcu_barrier_head); | |
aacb5d91 PM |
1012 | ssp = sdp->ssp; |
1013 | if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt)) | |
1014 | complete(&ssp->srcu_barrier_completion); | |
da915ad5 PM |
1015 | } |
1016 | ||
dad81a20 PM |
1017 | /** |
1018 | * srcu_barrier - Wait until all in-flight call_srcu() callbacks complete. | |
aacb5d91 | 1019 | * @ssp: srcu_struct on which to wait for in-flight callbacks. |
dad81a20 | 1020 | */ |
aacb5d91 | 1021 | void srcu_barrier(struct srcu_struct *ssp) |
dad81a20 | 1022 | { |
da915ad5 PM |
1023 | int cpu; |
1024 | struct srcu_data *sdp; | |
aacb5d91 | 1025 | unsigned long s = rcu_seq_snap(&ssp->srcu_barrier_seq); |
da915ad5 | 1026 | |
aacb5d91 PM |
1027 | check_init_srcu_struct(ssp); |
1028 | mutex_lock(&ssp->srcu_barrier_mutex); | |
1029 | if (rcu_seq_done(&ssp->srcu_barrier_seq, s)) { | |
da915ad5 | 1030 | smp_mb(); /* Force ordering following return. */ |
aacb5d91 | 1031 | mutex_unlock(&ssp->srcu_barrier_mutex); |
da915ad5 PM |
1032 | return; /* Someone else did our work for us. */ |
1033 | } | |
aacb5d91 PM |
1034 | rcu_seq_start(&ssp->srcu_barrier_seq); |
1035 | init_completion(&ssp->srcu_barrier_completion); | |
da915ad5 PM |
1036 | |
1037 | /* Initial count prevents reaching zero until all CBs are posted. */ | |
aacb5d91 | 1038 | atomic_set(&ssp->srcu_barrier_cpu_cnt, 1); |
da915ad5 PM |
1039 | |
1040 | /* | |
1041 | * Each pass through this loop enqueues a callback, but only | |
1042 | * on CPUs already having callbacks enqueued. Note that if | |
1043 | * a CPU already has callbacks enqueue, it must have already | |
1044 | * registered the need for a future grace period, so all we | |
1045 | * need do is enqueue a callback that will use the same | |
1046 | * grace period as the last callback already in the queue. | |
1047 | */ | |
1048 | for_each_possible_cpu(cpu) { | |
aacb5d91 | 1049 | sdp = per_cpu_ptr(ssp->sda, cpu); |
d6331980 | 1050 | spin_lock_irq_rcu_node(sdp); |
aacb5d91 | 1051 | atomic_inc(&ssp->srcu_barrier_cpu_cnt); |
da915ad5 | 1052 | sdp->srcu_barrier_head.func = srcu_barrier_cb; |
a602538e | 1053 | debug_rcu_head_queue(&sdp->srcu_barrier_head); |
da915ad5 | 1054 | if (!rcu_segcblist_entrain(&sdp->srcu_cblist, |
a602538e PM |
1055 | &sdp->srcu_barrier_head, 0)) { |
1056 | debug_rcu_head_unqueue(&sdp->srcu_barrier_head); | |
aacb5d91 | 1057 | atomic_dec(&ssp->srcu_barrier_cpu_cnt); |
a602538e | 1058 | } |
d6331980 | 1059 | spin_unlock_irq_rcu_node(sdp); |
da915ad5 PM |
1060 | } |
1061 | ||
1062 | /* Remove the initial count, at which point reaching zero can happen. */ | |
aacb5d91 PM |
1063 | if (atomic_dec_and_test(&ssp->srcu_barrier_cpu_cnt)) |
1064 | complete(&ssp->srcu_barrier_completion); | |
1065 | wait_for_completion(&ssp->srcu_barrier_completion); | |
da915ad5 | 1066 | |
aacb5d91 PM |
1067 | rcu_seq_end(&ssp->srcu_barrier_seq); |
1068 | mutex_unlock(&ssp->srcu_barrier_mutex); | |
dad81a20 PM |
1069 | } |
1070 | EXPORT_SYMBOL_GPL(srcu_barrier); | |
1071 | ||
1072 | /** | |
1073 | * srcu_batches_completed - return batches completed. | |
aacb5d91 | 1074 | * @ssp: srcu_struct on which to report batch completion. |
dad81a20 PM |
1075 | * |
1076 | * Report the number of batches, correlated with, but not necessarily | |
1077 | * precisely the same as, the number of grace periods that have elapsed. | |
1078 | */ | |
aacb5d91 | 1079 | unsigned long srcu_batches_completed(struct srcu_struct *ssp) |
dad81a20 | 1080 | { |
aacb5d91 | 1081 | return ssp->srcu_idx; |
dad81a20 PM |
1082 | } |
1083 | EXPORT_SYMBOL_GPL(srcu_batches_completed); | |
1084 | ||
1085 | /* | |
da915ad5 PM |
1086 | * Core SRCU state machine. Push state bits of ->srcu_gp_seq |
1087 | * to SRCU_STATE_SCAN2, and invoke srcu_gp_end() when scan has | |
1088 | * completed in that state. | |
dad81a20 | 1089 | */ |
aacb5d91 | 1090 | static void srcu_advance_state(struct srcu_struct *ssp) |
dad81a20 PM |
1091 | { |
1092 | int idx; | |
1093 | ||
aacb5d91 | 1094 | mutex_lock(&ssp->srcu_gp_mutex); |
da915ad5 | 1095 | |
dad81a20 PM |
1096 | /* |
1097 | * Because readers might be delayed for an extended period after | |
da915ad5 | 1098 | * fetching ->srcu_idx for their index, at any point in time there |
dad81a20 PM |
1099 | * might well be readers using both idx=0 and idx=1. We therefore |
1100 | * need to wait for readers to clear from both index values before | |
1101 | * invoking a callback. | |
1102 | * | |
1103 | * The load-acquire ensures that we see the accesses performed | |
1104 | * by the prior grace period. | |
1105 | */ | |
aacb5d91 | 1106 | idx = rcu_seq_state(smp_load_acquire(&ssp->srcu_gp_seq)); /* ^^^ */ |
dad81a20 | 1107 | if (idx == SRCU_STATE_IDLE) { |
aacb5d91 PM |
1108 | spin_lock_irq_rcu_node(ssp); |
1109 | if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) { | |
1110 | WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq)); | |
1111 | spin_unlock_irq_rcu_node(ssp); | |
1112 | mutex_unlock(&ssp->srcu_gp_mutex); | |
dad81a20 PM |
1113 | return; |
1114 | } | |
aacb5d91 | 1115 | idx = rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)); |
dad81a20 | 1116 | if (idx == SRCU_STATE_IDLE) |
aacb5d91 PM |
1117 | srcu_gp_start(ssp); |
1118 | spin_unlock_irq_rcu_node(ssp); | |
da915ad5 | 1119 | if (idx != SRCU_STATE_IDLE) { |
aacb5d91 | 1120 | mutex_unlock(&ssp->srcu_gp_mutex); |
dad81a20 | 1121 | return; /* Someone else started the grace period. */ |
da915ad5 | 1122 | } |
dad81a20 PM |
1123 | } |
1124 | ||
aacb5d91 PM |
1125 | if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN1) { |
1126 | idx = 1 ^ (ssp->srcu_idx & 1); | |
1127 | if (!try_check_zero(ssp, idx, 1)) { | |
1128 | mutex_unlock(&ssp->srcu_gp_mutex); | |
dad81a20 | 1129 | return; /* readers present, retry later. */ |
da915ad5 | 1130 | } |
aacb5d91 PM |
1131 | srcu_flip(ssp); |
1132 | rcu_seq_set_state(&ssp->srcu_gp_seq, SRCU_STATE_SCAN2); | |
dad81a20 PM |
1133 | } |
1134 | ||
aacb5d91 | 1135 | if (rcu_seq_state(READ_ONCE(ssp->srcu_gp_seq)) == SRCU_STATE_SCAN2) { |
dad81a20 PM |
1136 | |
1137 | /* | |
1138 | * SRCU read-side critical sections are normally short, | |
1139 | * so check at least twice in quick succession after a flip. | |
1140 | */ | |
aacb5d91 PM |
1141 | idx = 1 ^ (ssp->srcu_idx & 1); |
1142 | if (!try_check_zero(ssp, idx, 2)) { | |
1143 | mutex_unlock(&ssp->srcu_gp_mutex); | |
da915ad5 PM |
1144 | return; /* readers present, retry later. */ |
1145 | } | |
aacb5d91 | 1146 | srcu_gp_end(ssp); /* Releases ->srcu_gp_mutex. */ |
dad81a20 PM |
1147 | } |
1148 | } | |
1149 | ||
1150 | /* | |
1151 | * Invoke a limited number of SRCU callbacks that have passed through | |
1152 | * their grace period. If there are more to do, SRCU will reschedule | |
1153 | * the workqueue. Note that needed memory barriers have been executed | |
1154 | * in this task's context by srcu_readers_active_idx_check(). | |
1155 | */ | |
da915ad5 | 1156 | static void srcu_invoke_callbacks(struct work_struct *work) |
dad81a20 | 1157 | { |
da915ad5 | 1158 | bool more; |
dad81a20 PM |
1159 | struct rcu_cblist ready_cbs; |
1160 | struct rcu_head *rhp; | |
da915ad5 | 1161 | struct srcu_data *sdp; |
aacb5d91 | 1162 | struct srcu_struct *ssp; |
dad81a20 | 1163 | |
e81baf4c SAS |
1164 | sdp = container_of(work, struct srcu_data, work); |
1165 | ||
aacb5d91 | 1166 | ssp = sdp->ssp; |
dad81a20 | 1167 | rcu_cblist_init(&ready_cbs); |
d6331980 | 1168 | spin_lock_irq_rcu_node(sdp); |
da915ad5 | 1169 | rcu_segcblist_advance(&sdp->srcu_cblist, |
aacb5d91 | 1170 | rcu_seq_current(&ssp->srcu_gp_seq)); |
da915ad5 PM |
1171 | if (sdp->srcu_cblist_invoking || |
1172 | !rcu_segcblist_ready_cbs(&sdp->srcu_cblist)) { | |
d6331980 | 1173 | spin_unlock_irq_rcu_node(sdp); |
da915ad5 PM |
1174 | return; /* Someone else on the job or nothing to do. */ |
1175 | } | |
1176 | ||
1177 | /* We are on the job! Extract and invoke ready callbacks. */ | |
1178 | sdp->srcu_cblist_invoking = true; | |
1179 | rcu_segcblist_extract_done_cbs(&sdp->srcu_cblist, &ready_cbs); | |
d6331980 | 1180 | spin_unlock_irq_rcu_node(sdp); |
dad81a20 PM |
1181 | rhp = rcu_cblist_dequeue(&ready_cbs); |
1182 | for (; rhp != NULL; rhp = rcu_cblist_dequeue(&ready_cbs)) { | |
a602538e | 1183 | debug_rcu_head_unqueue(rhp); |
dad81a20 PM |
1184 | local_bh_disable(); |
1185 | rhp->func(rhp); | |
1186 | local_bh_enable(); | |
1187 | } | |
da915ad5 PM |
1188 | |
1189 | /* | |
1190 | * Update counts, accelerate new callbacks, and if needed, | |
1191 | * schedule another round of callback invocation. | |
1192 | */ | |
d6331980 | 1193 | spin_lock_irq_rcu_node(sdp); |
da915ad5 PM |
1194 | rcu_segcblist_insert_count(&sdp->srcu_cblist, &ready_cbs); |
1195 | (void)rcu_segcblist_accelerate(&sdp->srcu_cblist, | |
aacb5d91 | 1196 | rcu_seq_snap(&ssp->srcu_gp_seq)); |
da915ad5 PM |
1197 | sdp->srcu_cblist_invoking = false; |
1198 | more = rcu_segcblist_ready_cbs(&sdp->srcu_cblist); | |
d6331980 | 1199 | spin_unlock_irq_rcu_node(sdp); |
da915ad5 PM |
1200 | if (more) |
1201 | srcu_schedule_cbs_sdp(sdp, 0); | |
dad81a20 PM |
1202 | } |
1203 | ||
1204 | /* | |
1205 | * Finished one round of SRCU grace period. Start another if there are | |
1206 | * more SRCU callbacks queued, otherwise put SRCU into not-running state. | |
1207 | */ | |
aacb5d91 | 1208 | static void srcu_reschedule(struct srcu_struct *ssp, unsigned long delay) |
dad81a20 | 1209 | { |
da915ad5 | 1210 | bool pushgp = true; |
dad81a20 | 1211 | |
aacb5d91 PM |
1212 | spin_lock_irq_rcu_node(ssp); |
1213 | if (ULONG_CMP_GE(ssp->srcu_gp_seq, ssp->srcu_gp_seq_needed)) { | |
1214 | if (!WARN_ON_ONCE(rcu_seq_state(ssp->srcu_gp_seq))) { | |
da915ad5 PM |
1215 | /* All requests fulfilled, time to go idle. */ |
1216 | pushgp = false; | |
1217 | } | |
aacb5d91 | 1218 | } else if (!rcu_seq_state(ssp->srcu_gp_seq)) { |
da915ad5 | 1219 | /* Outstanding request and no GP. Start one. */ |
aacb5d91 | 1220 | srcu_gp_start(ssp); |
dad81a20 | 1221 | } |
aacb5d91 | 1222 | spin_unlock_irq_rcu_node(ssp); |
dad81a20 | 1223 | |
da915ad5 | 1224 | if (pushgp) |
aacb5d91 | 1225 | queue_delayed_work(rcu_gp_wq, &ssp->work, delay); |
dad81a20 PM |
1226 | } |
1227 | ||
1228 | /* | |
1229 | * This is the work-queue function that handles SRCU grace periods. | |
1230 | */ | |
0d8a1e83 | 1231 | static void process_srcu(struct work_struct *work) |
dad81a20 | 1232 | { |
aacb5d91 | 1233 | struct srcu_struct *ssp; |
dad81a20 | 1234 | |
aacb5d91 | 1235 | ssp = container_of(work, struct srcu_struct, work.work); |
dad81a20 | 1236 | |
aacb5d91 PM |
1237 | srcu_advance_state(ssp); |
1238 | srcu_reschedule(ssp, srcu_get_delay(ssp)); | |
dad81a20 | 1239 | } |
7f6733c3 PM |
1240 | |
1241 | void srcutorture_get_gp_data(enum rcutorture_type test_type, | |
aacb5d91 | 1242 | struct srcu_struct *ssp, int *flags, |
aebc8264 | 1243 | unsigned long *gp_seq) |
7f6733c3 PM |
1244 | { |
1245 | if (test_type != SRCU_FLAVOR) | |
1246 | return; | |
1247 | *flags = 0; | |
aacb5d91 | 1248 | *gp_seq = rcu_seq_current(&ssp->srcu_gp_seq); |
7f6733c3 PM |
1249 | } |
1250 | EXPORT_SYMBOL_GPL(srcutorture_get_gp_data); | |
1f4f6da1 | 1251 | |
aacb5d91 | 1252 | void srcu_torture_stats_print(struct srcu_struct *ssp, char *tt, char *tf) |
115a1a52 PM |
1253 | { |
1254 | int cpu; | |
1255 | int idx; | |
ac3748c6 | 1256 | unsigned long s0 = 0, s1 = 0; |
115a1a52 | 1257 | |
aacb5d91 | 1258 | idx = ssp->srcu_idx & 0x1; |
52e17ba1 | 1259 | pr_alert("%s%s Tree SRCU g%ld per-CPU(idx=%d):", |
aacb5d91 | 1260 | tt, tf, rcu_seq_current(&ssp->srcu_gp_seq), idx); |
115a1a52 PM |
1261 | for_each_possible_cpu(cpu) { |
1262 | unsigned long l0, l1; | |
1263 | unsigned long u0, u1; | |
1264 | long c0, c1; | |
5ab07a8d | 1265 | struct srcu_data *sdp; |
115a1a52 | 1266 | |
aacb5d91 | 1267 | sdp = per_cpu_ptr(ssp->sda, cpu); |
5ab07a8d PM |
1268 | u0 = sdp->srcu_unlock_count[!idx]; |
1269 | u1 = sdp->srcu_unlock_count[idx]; | |
115a1a52 PM |
1270 | |
1271 | /* | |
1272 | * Make sure that a lock is always counted if the corresponding | |
1273 | * unlock is counted. | |
1274 | */ | |
1275 | smp_rmb(); | |
1276 | ||
5ab07a8d PM |
1277 | l0 = sdp->srcu_lock_count[!idx]; |
1278 | l1 = sdp->srcu_lock_count[idx]; | |
115a1a52 PM |
1279 | |
1280 | c0 = l0 - u0; | |
1281 | c1 = l1 - u1; | |
5ab07a8d PM |
1282 | pr_cont(" %d(%ld,%ld %1p)", |
1283 | cpu, c0, c1, rcu_segcblist_head(&sdp->srcu_cblist)); | |
ac3748c6 PM |
1284 | s0 += c0; |
1285 | s1 += c1; | |
115a1a52 | 1286 | } |
ac3748c6 | 1287 | pr_cont(" T(%ld,%ld)\n", s0, s1); |
115a1a52 PM |
1288 | } |
1289 | EXPORT_SYMBOL_GPL(srcu_torture_stats_print); | |
1290 | ||
1f4f6da1 PM |
1291 | static int __init srcu_bootup_announce(void) |
1292 | { | |
1293 | pr_info("Hierarchical SRCU implementation.\n"); | |
0c8e0e3c PM |
1294 | if (exp_holdoff != DEFAULT_SRCU_EXP_HOLDOFF) |
1295 | pr_info("\tNon-default auto-expedite holdoff of %lu ns.\n", exp_holdoff); | |
1f4f6da1 PM |
1296 | return 0; |
1297 | } | |
1298 | early_initcall(srcu_bootup_announce); | |
e0fcba9a PM |
1299 | |
1300 | void __init srcu_init(void) | |
1301 | { | |
aacb5d91 | 1302 | struct srcu_struct *ssp; |
e0fcba9a PM |
1303 | |
1304 | srcu_init_done = true; | |
1305 | while (!list_empty(&srcu_boot_list)) { | |
aacb5d91 | 1306 | ssp = list_first_entry(&srcu_boot_list, struct srcu_struct, |
4e6ea4ef | 1307 | work.work.entry); |
aacb5d91 PM |
1308 | check_init_srcu_struct(ssp); |
1309 | list_del_init(&ssp->work.work.entry); | |
1310 | queue_work(rcu_gp_wq, &ssp->work.work); | |
e0fcba9a PM |
1311 | } |
1312 | } | |
fe15b50c PM |
1313 | |
1314 | #ifdef CONFIG_MODULES | |
1315 | ||
1316 | /* Initialize any global-scope srcu_struct structures used by this module. */ | |
1317 | static int srcu_module_coming(struct module *mod) | |
1318 | { | |
1319 | int i; | |
1320 | struct srcu_struct **sspp = mod->srcu_struct_ptrs; | |
1321 | int ret; | |
1322 | ||
1323 | for (i = 0; i < mod->num_srcu_structs; i++) { | |
1324 | ret = init_srcu_struct(*(sspp++)); | |
1325 | if (WARN_ON_ONCE(ret)) | |
1326 | return ret; | |
1327 | } | |
1328 | return 0; | |
1329 | } | |
1330 | ||
1331 | /* Clean up any global-scope srcu_struct structures used by this module. */ | |
1332 | static void srcu_module_going(struct module *mod) | |
1333 | { | |
1334 | int i; | |
1335 | struct srcu_struct **sspp = mod->srcu_struct_ptrs; | |
1336 | ||
1337 | for (i = 0; i < mod->num_srcu_structs; i++) | |
1338 | cleanup_srcu_struct(*(sspp++)); | |
1339 | } | |
1340 | ||
1341 | /* Handle one module, either coming or going. */ | |
1342 | static int srcu_module_notify(struct notifier_block *self, | |
1343 | unsigned long val, void *data) | |
1344 | { | |
1345 | struct module *mod = data; | |
1346 | int ret = 0; | |
1347 | ||
1348 | switch (val) { | |
1349 | case MODULE_STATE_COMING: | |
1350 | ret = srcu_module_coming(mod); | |
1351 | break; | |
1352 | case MODULE_STATE_GOING: | |
1353 | srcu_module_going(mod); | |
1354 | break; | |
1355 | default: | |
1356 | break; | |
1357 | } | |
1358 | return ret; | |
1359 | } | |
1360 | ||
1361 | static struct notifier_block srcu_module_nb = { | |
1362 | .notifier_call = srcu_module_notify, | |
1363 | .priority = 0, | |
1364 | }; | |
1365 | ||
1366 | static __init int init_srcu_module_notifier(void) | |
1367 | { | |
1368 | int ret; | |
1369 | ||
1370 | ret = register_module_notifier(&srcu_module_nb); | |
1371 | if (ret) | |
1372 | pr_warn("Failed to register srcu module notifier\n"); | |
1373 | return ret; | |
1374 | } | |
1375 | late_initcall(init_srcu_module_notifier); | |
1376 | ||
1377 | #endif /* #ifdef CONFIG_MODULES */ |