Commit | Line | Data |
---|---|---|
f41d911f PM |
1 | /* |
2 | * Read-Copy Update mechanism for mutual exclusion (tree-based version) | |
3 | * Internal non-public definitions that provide either classic | |
6cc68793 | 4 | * or preemptible semantics. |
f41d911f PM |
5 | * |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License as published by | |
8 | * the Free Software Foundation; either version 2 of the License, or | |
9 | * (at your option) any later version. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | * GNU General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
87de1cfd PM |
17 | * along with this program; if not, you can access it online at |
18 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
f41d911f PM |
19 | * |
20 | * Copyright Red Hat, 2009 | |
21 | * Copyright IBM Corporation, 2009 | |
22 | * | |
23 | * Author: Ingo Molnar <mingo@elte.hu> | |
24 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> | |
25 | */ | |
26 | ||
d9a3da06 | 27 | #include <linux/delay.h> |
3fbfbf7a | 28 | #include <linux/gfp.h> |
b626c1b6 | 29 | #include <linux/oom.h> |
62ab7072 | 30 | #include <linux/smpboot.h> |
4102adab | 31 | #include "../time/tick-internal.h" |
f41d911f | 32 | |
5b61b0ba | 33 | #ifdef CONFIG_RCU_BOOST |
61cfd097 | 34 | |
abaa93d9 | 35 | #include "../locking/rtmutex_common.h" |
21871d7e CW |
36 | |
37 | /* rcuc/rcub kthread realtime priority */ | |
38 | static int kthread_prio = CONFIG_RCU_KTHREAD_PRIO; | |
39 | module_param(kthread_prio, int, 0644); | |
61cfd097 PM |
40 | |
41 | /* | |
42 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
43 | * handle all flavors of RCU. | |
44 | */ | |
45 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
46 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); | |
47 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); | |
48 | DEFINE_PER_CPU(char, rcu_cpu_has_work); | |
49 | ||
21871d7e | 50 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
5b61b0ba | 51 | |
3fbfbf7a PM |
52 | #ifdef CONFIG_RCU_NOCB_CPU |
53 | static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */ | |
54 | static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */ | |
1b0048a4 | 55 | static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */ |
3fbfbf7a PM |
56 | static char __initdata nocb_buf[NR_CPUS * 5]; |
57 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ | |
58 | ||
26845c28 PM |
59 | /* |
60 | * Check the RCU kernel configuration parameters and print informative | |
61 | * messages about anything out of the ordinary. If you like #ifdef, you | |
62 | * will love this function. | |
63 | */ | |
64 | static void __init rcu_bootup_announce_oddness(void) | |
65 | { | |
66 | #ifdef CONFIG_RCU_TRACE | |
efc151c3 | 67 | pr_info("\tRCU debugfs-based tracing is enabled.\n"); |
26845c28 PM |
68 | #endif |
69 | #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32) | |
efc151c3 | 70 | pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n", |
26845c28 PM |
71 | CONFIG_RCU_FANOUT); |
72 | #endif | |
73 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
efc151c3 | 74 | pr_info("\tHierarchical RCU autobalancing is disabled.\n"); |
26845c28 PM |
75 | #endif |
76 | #ifdef CONFIG_RCU_FAST_NO_HZ | |
efc151c3 | 77 | pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n"); |
26845c28 PM |
78 | #endif |
79 | #ifdef CONFIG_PROVE_RCU | |
efc151c3 | 80 | pr_info("\tRCU lockdep checking is enabled.\n"); |
26845c28 PM |
81 | #endif |
82 | #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE | |
efc151c3 | 83 | pr_info("\tRCU torture testing starts during boot.\n"); |
26845c28 | 84 | #endif |
a858af28 | 85 | #if defined(CONFIG_RCU_CPU_STALL_INFO) |
efc151c3 | 86 | pr_info("\tAdditional per-CPU info printed with stalls.\n"); |
26845c28 PM |
87 | #endif |
88 | #if NUM_RCU_LVL_4 != 0 | |
efc151c3 | 89 | pr_info("\tFour-level hierarchy is enabled.\n"); |
26845c28 | 90 | #endif |
f885b7f2 | 91 | if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF) |
9a5739d7 | 92 | pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf); |
cca6f393 | 93 | if (nr_cpu_ids != NR_CPUS) |
efc151c3 | 94 | pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids); |
21871d7e CW |
95 | #ifdef CONFIG_RCU_BOOST |
96 | pr_info("\tRCU kthread priority: %d.\n", kthread_prio); | |
97 | #endif | |
26845c28 PM |
98 | } |
99 | ||
28f6569a | 100 | #ifdef CONFIG_PREEMPT_RCU |
f41d911f | 101 | |
a41bfeb2 | 102 | RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu); |
e534165b | 103 | static struct rcu_state *rcu_state_p = &rcu_preempt_state; |
f41d911f | 104 | |
d9a3da06 PM |
105 | static int rcu_preempted_readers_exp(struct rcu_node *rnp); |
106 | ||
f41d911f PM |
107 | /* |
108 | * Tell them what RCU they are running. | |
109 | */ | |
0e0fc1c2 | 110 | static void __init rcu_bootup_announce(void) |
f41d911f | 111 | { |
efc151c3 | 112 | pr_info("Preemptible hierarchical RCU implementation.\n"); |
26845c28 | 113 | rcu_bootup_announce_oddness(); |
f41d911f PM |
114 | } |
115 | ||
116 | /* | |
117 | * Return the number of RCU-preempt batches processed thus far | |
118 | * for debug and statistics. | |
119 | */ | |
bf33eb1a | 120 | static long rcu_batches_completed_preempt(void) |
f41d911f PM |
121 | { |
122 | return rcu_preempt_state.completed; | |
123 | } | |
124 | EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt); | |
125 | ||
126 | /* | |
127 | * Return the number of RCU batches processed thus far for debug & stats. | |
128 | */ | |
129 | long rcu_batches_completed(void) | |
130 | { | |
131 | return rcu_batches_completed_preempt(); | |
132 | } | |
133 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
134 | ||
135 | /* | |
6cc68793 | 136 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
f41d911f PM |
137 | * that this just means that the task currently running on the CPU is |
138 | * not in a quiescent state. There might be any number of tasks blocked | |
139 | * while in an RCU read-side critical section. | |
25502a6c | 140 | * |
1d082fd0 PM |
141 | * As with the other rcu_*_qs() functions, callers to this function |
142 | * must disable preemption. | |
f41d911f | 143 | */ |
284a8c93 | 144 | static void rcu_preempt_qs(void) |
f41d911f | 145 | { |
284a8c93 PM |
146 | if (!__this_cpu_read(rcu_preempt_data.passed_quiesce)) { |
147 | trace_rcu_grace_period(TPS("rcu_preempt"), | |
148 | __this_cpu_read(rcu_preempt_data.gpnum), | |
149 | TPS("cpuqs")); | |
150 | __this_cpu_write(rcu_preempt_data.passed_quiesce, 1); | |
151 | barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */ | |
152 | current->rcu_read_unlock_special.b.need_qs = false; | |
153 | } | |
f41d911f PM |
154 | } |
155 | ||
156 | /* | |
c3422bea PM |
157 | * We have entered the scheduler, and the current task might soon be |
158 | * context-switched away from. If this task is in an RCU read-side | |
159 | * critical section, we will no longer be able to rely on the CPU to | |
12f5f524 PM |
160 | * record that fact, so we enqueue the task on the blkd_tasks list. |
161 | * The task will dequeue itself when it exits the outermost enclosing | |
162 | * RCU read-side critical section. Therefore, the current grace period | |
163 | * cannot be permitted to complete until the blkd_tasks list entries | |
164 | * predating the current grace period drain, in other words, until | |
165 | * rnp->gp_tasks becomes NULL. | |
c3422bea PM |
166 | * |
167 | * Caller must disable preemption. | |
f41d911f | 168 | */ |
38200cf2 | 169 | static void rcu_preempt_note_context_switch(void) |
f41d911f PM |
170 | { |
171 | struct task_struct *t = current; | |
c3422bea | 172 | unsigned long flags; |
f41d911f PM |
173 | struct rcu_data *rdp; |
174 | struct rcu_node *rnp; | |
175 | ||
10f39bb1 | 176 | if (t->rcu_read_lock_nesting > 0 && |
1d082fd0 | 177 | !t->rcu_read_unlock_special.b.blocked) { |
f41d911f PM |
178 | |
179 | /* Possibly blocking in an RCU read-side critical section. */ | |
38200cf2 | 180 | rdp = this_cpu_ptr(rcu_preempt_state.rda); |
f41d911f | 181 | rnp = rdp->mynode; |
1304afb2 | 182 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 183 | smp_mb__after_unlock_lock(); |
1d082fd0 | 184 | t->rcu_read_unlock_special.b.blocked = true; |
86848966 | 185 | t->rcu_blocked_node = rnp; |
f41d911f PM |
186 | |
187 | /* | |
188 | * If this CPU has already checked in, then this task | |
189 | * will hold up the next grace period rather than the | |
190 | * current grace period. Queue the task accordingly. | |
191 | * If the task is queued for the current grace period | |
192 | * (i.e., this CPU has not yet passed through a quiescent | |
193 | * state for the current grace period), then as long | |
194 | * as that task remains queued, the current grace period | |
12f5f524 PM |
195 | * cannot end. Note that there is some uncertainty as |
196 | * to exactly when the current grace period started. | |
197 | * We take a conservative approach, which can result | |
198 | * in unnecessarily waiting on tasks that started very | |
199 | * slightly after the current grace period began. C'est | |
200 | * la vie!!! | |
b0e165c0 PM |
201 | * |
202 | * But first, note that the current CPU must still be | |
203 | * on line! | |
f41d911f | 204 | */ |
b0e165c0 | 205 | WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0); |
e7d8842e | 206 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
12f5f524 PM |
207 | if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) { |
208 | list_add(&t->rcu_node_entry, rnp->gp_tasks->prev); | |
209 | rnp->gp_tasks = &t->rcu_node_entry; | |
27f4d280 PM |
210 | #ifdef CONFIG_RCU_BOOST |
211 | if (rnp->boost_tasks != NULL) | |
212 | rnp->boost_tasks = rnp->gp_tasks; | |
213 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
12f5f524 PM |
214 | } else { |
215 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); | |
216 | if (rnp->qsmask & rdp->grpmask) | |
217 | rnp->gp_tasks = &t->rcu_node_entry; | |
218 | } | |
d4c08f2a PM |
219 | trace_rcu_preempt_task(rdp->rsp->name, |
220 | t->pid, | |
221 | (rnp->qsmask & rdp->grpmask) | |
222 | ? rnp->gpnum | |
223 | : rnp->gpnum + 1); | |
1304afb2 | 224 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
10f39bb1 | 225 | } else if (t->rcu_read_lock_nesting < 0 && |
1d082fd0 | 226 | t->rcu_read_unlock_special.s) { |
10f39bb1 PM |
227 | |
228 | /* | |
229 | * Complete exit from RCU read-side critical section on | |
230 | * behalf of preempted instance of __rcu_read_unlock(). | |
231 | */ | |
232 | rcu_read_unlock_special(t); | |
f41d911f PM |
233 | } |
234 | ||
235 | /* | |
236 | * Either we were not in an RCU read-side critical section to | |
237 | * begin with, or we have now recorded that critical section | |
238 | * globally. Either way, we can now note a quiescent state | |
239 | * for this CPU. Again, if we were in an RCU read-side critical | |
240 | * section, and if that critical section was blocking the current | |
241 | * grace period, then the fact that the task has been enqueued | |
242 | * means that we continue to block the current grace period. | |
243 | */ | |
284a8c93 | 244 | rcu_preempt_qs(); |
f41d911f PM |
245 | } |
246 | ||
fc2219d4 PM |
247 | /* |
248 | * Check for preempted RCU readers blocking the current grace period | |
249 | * for the specified rcu_node structure. If the caller needs a reliable | |
250 | * answer, it must hold the rcu_node's ->lock. | |
251 | */ | |
27f4d280 | 252 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 | 253 | { |
12f5f524 | 254 | return rnp->gp_tasks != NULL; |
fc2219d4 PM |
255 | } |
256 | ||
b668c9cf PM |
257 | /* |
258 | * Record a quiescent state for all tasks that were previously queued | |
259 | * on the specified rcu_node structure and that were blocking the current | |
260 | * RCU grace period. The caller must hold the specified rnp->lock with | |
261 | * irqs disabled, and this lock is released upon return, but irqs remain | |
262 | * disabled. | |
263 | */ | |
d3f6bad3 | 264 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
b668c9cf PM |
265 | __releases(rnp->lock) |
266 | { | |
267 | unsigned long mask; | |
268 | struct rcu_node *rnp_p; | |
269 | ||
27f4d280 | 270 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
1304afb2 | 271 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf PM |
272 | return; /* Still need more quiescent states! */ |
273 | } | |
274 | ||
275 | rnp_p = rnp->parent; | |
276 | if (rnp_p == NULL) { | |
277 | /* | |
278 | * Either there is only one rcu_node in the tree, | |
279 | * or tasks were kicked up to root rcu_node due to | |
280 | * CPUs going offline. | |
281 | */ | |
d3f6bad3 | 282 | rcu_report_qs_rsp(&rcu_preempt_state, flags); |
b668c9cf PM |
283 | return; |
284 | } | |
285 | ||
286 | /* Report up the rest of the hierarchy. */ | |
287 | mask = rnp->grpmask; | |
1304afb2 PM |
288 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
289 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
6303b9c8 | 290 | smp_mb__after_unlock_lock(); |
d3f6bad3 | 291 | rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags); |
b668c9cf PM |
292 | } |
293 | ||
12f5f524 PM |
294 | /* |
295 | * Advance a ->blkd_tasks-list pointer to the next entry, instead | |
296 | * returning NULL if at the end of the list. | |
297 | */ | |
298 | static struct list_head *rcu_next_node_entry(struct task_struct *t, | |
299 | struct rcu_node *rnp) | |
300 | { | |
301 | struct list_head *np; | |
302 | ||
303 | np = t->rcu_node_entry.next; | |
304 | if (np == &rnp->blkd_tasks) | |
305 | np = NULL; | |
306 | return np; | |
307 | } | |
308 | ||
b668c9cf PM |
309 | /* |
310 | * Handle special cases during rcu_read_unlock(), such as needing to | |
311 | * notify RCU core processing or task having blocked during the RCU | |
312 | * read-side critical section. | |
313 | */ | |
2a3fa843 | 314 | void rcu_read_unlock_special(struct task_struct *t) |
f41d911f PM |
315 | { |
316 | int empty; | |
d9a3da06 | 317 | int empty_exp; |
389abd48 | 318 | int empty_exp_now; |
f41d911f | 319 | unsigned long flags; |
12f5f524 | 320 | struct list_head *np; |
82e78d80 | 321 | #ifdef CONFIG_RCU_BOOST |
abaa93d9 | 322 | bool drop_boost_mutex = false; |
82e78d80 | 323 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
f41d911f | 324 | struct rcu_node *rnp; |
1d082fd0 | 325 | union rcu_special special; |
f41d911f PM |
326 | |
327 | /* NMI handlers cannot block and cannot safely manipulate state. */ | |
328 | if (in_nmi()) | |
329 | return; | |
330 | ||
331 | local_irq_save(flags); | |
332 | ||
333 | /* | |
334 | * If RCU core is waiting for this CPU to exit critical section, | |
1d082fd0 PM |
335 | * let it know that we have done so. Because irqs are disabled, |
336 | * t->rcu_read_unlock_special cannot change. | |
f41d911f PM |
337 | */ |
338 | special = t->rcu_read_unlock_special; | |
1d082fd0 | 339 | if (special.b.need_qs) { |
284a8c93 | 340 | rcu_preempt_qs(); |
1d082fd0 | 341 | if (!t->rcu_read_unlock_special.s) { |
79a62f95 LJ |
342 | local_irq_restore(flags); |
343 | return; | |
344 | } | |
f41d911f PM |
345 | } |
346 | ||
79a62f95 LJ |
347 | /* Hardware IRQ handlers cannot block, complain if they get here. */ |
348 | if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) { | |
f41d911f PM |
349 | local_irq_restore(flags); |
350 | return; | |
351 | } | |
352 | ||
353 | /* Clean up if blocked during RCU read-side critical section. */ | |
1d082fd0 PM |
354 | if (special.b.blocked) { |
355 | t->rcu_read_unlock_special.b.blocked = false; | |
f41d911f | 356 | |
dd5d19ba PM |
357 | /* |
358 | * Remove this task from the list it blocked on. The | |
359 | * task can migrate while we acquire the lock, but at | |
360 | * most one time. So at most two passes through loop. | |
361 | */ | |
362 | for (;;) { | |
86848966 | 363 | rnp = t->rcu_blocked_node; |
1304afb2 | 364 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
6303b9c8 | 365 | smp_mb__after_unlock_lock(); |
86848966 | 366 | if (rnp == t->rcu_blocked_node) |
dd5d19ba | 367 | break; |
1304afb2 | 368 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
dd5d19ba | 369 | } |
27f4d280 | 370 | empty = !rcu_preempt_blocked_readers_cgp(rnp); |
d9a3da06 PM |
371 | empty_exp = !rcu_preempted_readers_exp(rnp); |
372 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ | |
12f5f524 | 373 | np = rcu_next_node_entry(t, rnp); |
f41d911f | 374 | list_del_init(&t->rcu_node_entry); |
82e78d80 | 375 | t->rcu_blocked_node = NULL; |
f7f7bac9 | 376 | trace_rcu_unlock_preempted_task(TPS("rcu_preempt"), |
d4c08f2a | 377 | rnp->gpnum, t->pid); |
12f5f524 PM |
378 | if (&t->rcu_node_entry == rnp->gp_tasks) |
379 | rnp->gp_tasks = np; | |
380 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
381 | rnp->exp_tasks = np; | |
27f4d280 PM |
382 | #ifdef CONFIG_RCU_BOOST |
383 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
384 | rnp->boost_tasks = np; | |
abaa93d9 PM |
385 | /* Snapshot ->boost_mtx ownership with rcu_node lock held. */ |
386 | drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t; | |
27f4d280 | 387 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
f41d911f PM |
388 | |
389 | /* | |
390 | * If this was the last task on the current list, and if | |
391 | * we aren't waiting on any CPUs, report the quiescent state. | |
389abd48 PM |
392 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock, |
393 | * so we must take a snapshot of the expedited state. | |
f41d911f | 394 | */ |
389abd48 | 395 | empty_exp_now = !rcu_preempted_readers_exp(rnp); |
d4c08f2a | 396 | if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) { |
f7f7bac9 | 397 | trace_rcu_quiescent_state_report(TPS("preempt_rcu"), |
d4c08f2a PM |
398 | rnp->gpnum, |
399 | 0, rnp->qsmask, | |
400 | rnp->level, | |
401 | rnp->grplo, | |
402 | rnp->grphi, | |
403 | !!rnp->gp_tasks); | |
d3f6bad3 | 404 | rcu_report_unblock_qs_rnp(rnp, flags); |
c701d5d9 | 405 | } else { |
d4c08f2a | 406 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c701d5d9 | 407 | } |
d9a3da06 | 408 | |
27f4d280 PM |
409 | #ifdef CONFIG_RCU_BOOST |
410 | /* Unboost if we were boosted. */ | |
abaa93d9 PM |
411 | if (drop_boost_mutex) { |
412 | rt_mutex_unlock(&rnp->boost_mtx); | |
dfeb9765 PM |
413 | complete(&rnp->boost_completion); |
414 | } | |
27f4d280 PM |
415 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
416 | ||
d9a3da06 PM |
417 | /* |
418 | * If this was the last task on the expedited lists, | |
419 | * then we need to report up the rcu_node hierarchy. | |
420 | */ | |
389abd48 | 421 | if (!empty_exp && empty_exp_now) |
b40d293e | 422 | rcu_report_exp_rnp(&rcu_preempt_state, rnp, true); |
b668c9cf PM |
423 | } else { |
424 | local_irq_restore(flags); | |
f41d911f | 425 | } |
f41d911f PM |
426 | } |
427 | ||
1ed509a2 PM |
428 | /* |
429 | * Dump detailed information for all tasks blocking the current RCU | |
430 | * grace period on the specified rcu_node structure. | |
431 | */ | |
432 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) | |
433 | { | |
434 | unsigned long flags; | |
1ed509a2 PM |
435 | struct task_struct *t; |
436 | ||
12f5f524 | 437 | raw_spin_lock_irqsave(&rnp->lock, flags); |
5fd4dc06 PM |
438 | if (!rcu_preempt_blocked_readers_cgp(rnp)) { |
439 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
440 | return; | |
441 | } | |
12f5f524 PM |
442 | t = list_entry(rnp->gp_tasks, |
443 | struct task_struct, rcu_node_entry); | |
444 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) | |
445 | sched_show_task(t); | |
446 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1ed509a2 PM |
447 | } |
448 | ||
449 | /* | |
450 | * Dump detailed information for all tasks blocking the current RCU | |
451 | * grace period. | |
452 | */ | |
453 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
454 | { | |
455 | struct rcu_node *rnp = rcu_get_root(rsp); | |
456 | ||
457 | rcu_print_detail_task_stall_rnp(rnp); | |
458 | rcu_for_each_leaf_node(rsp, rnp) | |
459 | rcu_print_detail_task_stall_rnp(rnp); | |
460 | } | |
461 | ||
a858af28 PM |
462 | #ifdef CONFIG_RCU_CPU_STALL_INFO |
463 | ||
464 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) | |
465 | { | |
efc151c3 | 466 | pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):", |
a858af28 PM |
467 | rnp->level, rnp->grplo, rnp->grphi); |
468 | } | |
469 | ||
470 | static void rcu_print_task_stall_end(void) | |
471 | { | |
efc151c3 | 472 | pr_cont("\n"); |
a858af28 PM |
473 | } |
474 | ||
475 | #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */ | |
476 | ||
477 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) | |
478 | { | |
479 | } | |
480 | ||
481 | static void rcu_print_task_stall_end(void) | |
482 | { | |
483 | } | |
484 | ||
485 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */ | |
486 | ||
f41d911f PM |
487 | /* |
488 | * Scan the current list of tasks blocked within RCU read-side critical | |
489 | * sections, printing out the tid of each. | |
490 | */ | |
9bc8b558 | 491 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 492 | { |
f41d911f | 493 | struct task_struct *t; |
9bc8b558 | 494 | int ndetected = 0; |
f41d911f | 495 | |
27f4d280 | 496 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
9bc8b558 | 497 | return 0; |
a858af28 | 498 | rcu_print_task_stall_begin(rnp); |
12f5f524 PM |
499 | t = list_entry(rnp->gp_tasks, |
500 | struct task_struct, rcu_node_entry); | |
9bc8b558 | 501 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { |
efc151c3 | 502 | pr_cont(" P%d", t->pid); |
9bc8b558 PM |
503 | ndetected++; |
504 | } | |
a858af28 | 505 | rcu_print_task_stall_end(); |
9bc8b558 | 506 | return ndetected; |
f41d911f PM |
507 | } |
508 | ||
b0e165c0 PM |
509 | /* |
510 | * Check that the list of blocked tasks for the newly completed grace | |
511 | * period is in fact empty. It is a serious bug to complete a grace | |
512 | * period that still has RCU readers blocked! This function must be | |
513 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock | |
514 | * must be held by the caller. | |
12f5f524 PM |
515 | * |
516 | * Also, if there are blocked tasks on the list, they automatically | |
517 | * block the newly created grace period, so set up ->gp_tasks accordingly. | |
b0e165c0 PM |
518 | */ |
519 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
520 | { | |
27f4d280 | 521 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
12f5f524 PM |
522 | if (!list_empty(&rnp->blkd_tasks)) |
523 | rnp->gp_tasks = rnp->blkd_tasks.next; | |
28ecd580 | 524 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
525 | } |
526 | ||
33f76148 PM |
527 | #ifdef CONFIG_HOTPLUG_CPU |
528 | ||
dd5d19ba PM |
529 | /* |
530 | * Handle tasklist migration for case in which all CPUs covered by the | |
531 | * specified rcu_node have gone offline. Move them up to the root | |
532 | * rcu_node. The reason for not just moving them to the immediate | |
533 | * parent is to remove the need for rcu_read_unlock_special() to | |
534 | * make more than two attempts to acquire the target rcu_node's lock. | |
b668c9cf PM |
535 | * Returns true if there were tasks blocking the current RCU grace |
536 | * period. | |
dd5d19ba | 537 | * |
237c80c5 PM |
538 | * Returns 1 if there was previously a task blocking the current grace |
539 | * period on the specified rcu_node structure. | |
540 | * | |
dd5d19ba PM |
541 | * The caller must hold rnp->lock with irqs disabled. |
542 | */ | |
237c80c5 PM |
543 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
544 | struct rcu_node *rnp, | |
545 | struct rcu_data *rdp) | |
dd5d19ba | 546 | { |
dd5d19ba PM |
547 | struct list_head *lp; |
548 | struct list_head *lp_root; | |
d9a3da06 | 549 | int retval = 0; |
dd5d19ba | 550 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
12f5f524 | 551 | struct task_struct *t; |
dd5d19ba | 552 | |
86848966 PM |
553 | if (rnp == rnp_root) { |
554 | WARN_ONCE(1, "Last CPU thought to be offlined?"); | |
237c80c5 | 555 | return 0; /* Shouldn't happen: at least one CPU online. */ |
86848966 | 556 | } |
12f5f524 PM |
557 | |
558 | /* If we are on an internal node, complain bitterly. */ | |
559 | WARN_ON_ONCE(rnp != rdp->mynode); | |
dd5d19ba PM |
560 | |
561 | /* | |
12f5f524 PM |
562 | * Move tasks up to root rcu_node. Don't try to get fancy for |
563 | * this corner-case operation -- just put this node's tasks | |
564 | * at the head of the root node's list, and update the root node's | |
565 | * ->gp_tasks and ->exp_tasks pointers to those of this node's, | |
566 | * if non-NULL. This might result in waiting for more tasks than | |
567 | * absolutely necessary, but this is a good performance/complexity | |
568 | * tradeoff. | |
dd5d19ba | 569 | */ |
2036d94a | 570 | if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0) |
d9a3da06 PM |
571 | retval |= RCU_OFL_TASKS_NORM_GP; |
572 | if (rcu_preempted_readers_exp(rnp)) | |
573 | retval |= RCU_OFL_TASKS_EXP_GP; | |
12f5f524 PM |
574 | lp = &rnp->blkd_tasks; |
575 | lp_root = &rnp_root->blkd_tasks; | |
576 | while (!list_empty(lp)) { | |
577 | t = list_entry(lp->next, typeof(*t), rcu_node_entry); | |
578 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ | |
6303b9c8 | 579 | smp_mb__after_unlock_lock(); |
12f5f524 PM |
580 | list_del(&t->rcu_node_entry); |
581 | t->rcu_blocked_node = rnp_root; | |
582 | list_add(&t->rcu_node_entry, lp_root); | |
583 | if (&t->rcu_node_entry == rnp->gp_tasks) | |
584 | rnp_root->gp_tasks = rnp->gp_tasks; | |
585 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
586 | rnp_root->exp_tasks = rnp->exp_tasks; | |
27f4d280 PM |
587 | #ifdef CONFIG_RCU_BOOST |
588 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
589 | rnp_root->boost_tasks = rnp->boost_tasks; | |
590 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
12f5f524 | 591 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ |
dd5d19ba | 592 | } |
27f4d280 | 593 | |
1e3fd2b3 PM |
594 | rnp->gp_tasks = NULL; |
595 | rnp->exp_tasks = NULL; | |
27f4d280 | 596 | #ifdef CONFIG_RCU_BOOST |
1e3fd2b3 | 597 | rnp->boost_tasks = NULL; |
5cc900cf PM |
598 | /* |
599 | * In case root is being boosted and leaf was not. Make sure | |
600 | * that we boost the tasks blocking the current grace period | |
601 | * in this case. | |
602 | */ | |
27f4d280 | 603 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ |
6303b9c8 | 604 | smp_mb__after_unlock_lock(); |
27f4d280 | 605 | if (rnp_root->boost_tasks != NULL && |
5cc900cf PM |
606 | rnp_root->boost_tasks != rnp_root->gp_tasks && |
607 | rnp_root->boost_tasks != rnp_root->exp_tasks) | |
27f4d280 PM |
608 | rnp_root->boost_tasks = rnp_root->gp_tasks; |
609 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ | |
610 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
611 | ||
237c80c5 | 612 | return retval; |
dd5d19ba PM |
613 | } |
614 | ||
e5601400 PM |
615 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
616 | ||
f41d911f PM |
617 | /* |
618 | * Check for a quiescent state from the current CPU. When a task blocks, | |
619 | * the task is recorded in the corresponding CPU's rcu_node structure, | |
620 | * which is checked elsewhere. | |
621 | * | |
622 | * Caller must disable hard irqs. | |
623 | */ | |
86aea0e6 | 624 | static void rcu_preempt_check_callbacks(void) |
f41d911f PM |
625 | { |
626 | struct task_struct *t = current; | |
627 | ||
628 | if (t->rcu_read_lock_nesting == 0) { | |
284a8c93 | 629 | rcu_preempt_qs(); |
f41d911f PM |
630 | return; |
631 | } | |
10f39bb1 | 632 | if (t->rcu_read_lock_nesting > 0 && |
86aea0e6 PM |
633 | __this_cpu_read(rcu_preempt_data.qs_pending) && |
634 | !__this_cpu_read(rcu_preempt_data.passed_quiesce)) | |
1d082fd0 | 635 | t->rcu_read_unlock_special.b.need_qs = true; |
f41d911f PM |
636 | } |
637 | ||
a46e0899 PM |
638 | #ifdef CONFIG_RCU_BOOST |
639 | ||
09223371 SL |
640 | static void rcu_preempt_do_callbacks(void) |
641 | { | |
c9d4b0af | 642 | rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data)); |
09223371 SL |
643 | } |
644 | ||
a46e0899 PM |
645 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
646 | ||
f41d911f | 647 | /* |
6cc68793 | 648 | * Queue a preemptible-RCU callback for invocation after a grace period. |
f41d911f PM |
649 | */ |
650 | void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
651 | { | |
3fbfbf7a | 652 | __call_rcu(head, func, &rcu_preempt_state, -1, 0); |
f41d911f PM |
653 | } |
654 | EXPORT_SYMBOL_GPL(call_rcu); | |
655 | ||
6ebb237b PM |
656 | /** |
657 | * synchronize_rcu - wait until a grace period has elapsed. | |
658 | * | |
659 | * Control will return to the caller some time after a full grace | |
660 | * period has elapsed, in other words after all currently executing RCU | |
77d8485a PM |
661 | * read-side critical sections have completed. Note, however, that |
662 | * upon return from synchronize_rcu(), the caller might well be executing | |
663 | * concurrently with new RCU read-side critical sections that began while | |
664 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
665 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
f0a0e6f2 PM |
666 | * |
667 | * See the description of synchronize_sched() for more detailed information | |
668 | * on memory ordering guarantees. | |
6ebb237b PM |
669 | */ |
670 | void synchronize_rcu(void) | |
671 | { | |
fe15d706 PM |
672 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
673 | !lock_is_held(&rcu_lock_map) && | |
674 | !lock_is_held(&rcu_sched_lock_map), | |
675 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
6ebb237b PM |
676 | if (!rcu_scheduler_active) |
677 | return; | |
3705b88d AM |
678 | if (rcu_expedited) |
679 | synchronize_rcu_expedited(); | |
680 | else | |
681 | wait_rcu_gp(call_rcu); | |
6ebb237b PM |
682 | } |
683 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
684 | ||
d9a3da06 | 685 | static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq); |
bcfa57ce | 686 | static unsigned long sync_rcu_preempt_exp_count; |
d9a3da06 PM |
687 | static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex); |
688 | ||
689 | /* | |
690 | * Return non-zero if there are any tasks in RCU read-side critical | |
691 | * sections blocking the current preemptible-RCU expedited grace period. | |
692 | * If there is no preemptible-RCU expedited grace period currently in | |
693 | * progress, returns zero unconditionally. | |
694 | */ | |
695 | static int rcu_preempted_readers_exp(struct rcu_node *rnp) | |
696 | { | |
12f5f524 | 697 | return rnp->exp_tasks != NULL; |
d9a3da06 PM |
698 | } |
699 | ||
700 | /* | |
701 | * return non-zero if there is no RCU expedited grace period in progress | |
702 | * for the specified rcu_node structure, in other words, if all CPUs and | |
703 | * tasks covered by the specified rcu_node structure have done their bit | |
704 | * for the current expedited grace period. Works only for preemptible | |
705 | * RCU -- other RCU implementation use other means. | |
706 | * | |
707 | * Caller must hold sync_rcu_preempt_exp_mutex. | |
708 | */ | |
709 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) | |
710 | { | |
711 | return !rcu_preempted_readers_exp(rnp) && | |
712 | ACCESS_ONCE(rnp->expmask) == 0; | |
713 | } | |
714 | ||
715 | /* | |
716 | * Report the exit from RCU read-side critical section for the last task | |
717 | * that queued itself during or before the current expedited preemptible-RCU | |
718 | * grace period. This event is reported either to the rcu_node structure on | |
719 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
720 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
721 | * iteratively!) | |
722 | * | |
b40d293e TG |
723 | * Most callers will set the "wake" flag, but the task initiating the |
724 | * expedited grace period need not wake itself. | |
725 | * | |
d9a3da06 PM |
726 | * Caller must hold sync_rcu_preempt_exp_mutex. |
727 | */ | |
b40d293e TG |
728 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
729 | bool wake) | |
d9a3da06 PM |
730 | { |
731 | unsigned long flags; | |
732 | unsigned long mask; | |
733 | ||
1304afb2 | 734 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 735 | smp_mb__after_unlock_lock(); |
d9a3da06 | 736 | for (;;) { |
131906b0 PM |
737 | if (!sync_rcu_preempt_exp_done(rnp)) { |
738 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
d9a3da06 | 739 | break; |
131906b0 | 740 | } |
d9a3da06 | 741 | if (rnp->parent == NULL) { |
131906b0 | 742 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
78e4bc34 PM |
743 | if (wake) { |
744 | smp_mb(); /* EGP done before wake_up(). */ | |
b40d293e | 745 | wake_up(&sync_rcu_preempt_exp_wq); |
78e4bc34 | 746 | } |
d9a3da06 PM |
747 | break; |
748 | } | |
749 | mask = rnp->grpmask; | |
1304afb2 | 750 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
d9a3da06 | 751 | rnp = rnp->parent; |
1304afb2 | 752 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
6303b9c8 | 753 | smp_mb__after_unlock_lock(); |
d9a3da06 PM |
754 | rnp->expmask &= ~mask; |
755 | } | |
d9a3da06 PM |
756 | } |
757 | ||
758 | /* | |
759 | * Snapshot the tasks blocking the newly started preemptible-RCU expedited | |
760 | * grace period for the specified rcu_node structure. If there are no such | |
761 | * tasks, report it up the rcu_node hierarchy. | |
762 | * | |
7b2e6011 PM |
763 | * Caller must hold sync_rcu_preempt_exp_mutex and must exclude |
764 | * CPU hotplug operations. | |
d9a3da06 PM |
765 | */ |
766 | static void | |
767 | sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp) | |
768 | { | |
1217ed1b | 769 | unsigned long flags; |
12f5f524 | 770 | int must_wait = 0; |
d9a3da06 | 771 | |
1217ed1b | 772 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 773 | smp_mb__after_unlock_lock(); |
c701d5d9 | 774 | if (list_empty(&rnp->blkd_tasks)) { |
1217ed1b | 775 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c701d5d9 | 776 | } else { |
12f5f524 | 777 | rnp->exp_tasks = rnp->blkd_tasks.next; |
1217ed1b | 778 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
12f5f524 PM |
779 | must_wait = 1; |
780 | } | |
d9a3da06 | 781 | if (!must_wait) |
b40d293e | 782 | rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */ |
d9a3da06 PM |
783 | } |
784 | ||
236fefaf PM |
785 | /** |
786 | * synchronize_rcu_expedited - Brute-force RCU grace period | |
787 | * | |
788 | * Wait for an RCU-preempt grace period, but expedite it. The basic | |
789 | * idea is to invoke synchronize_sched_expedited() to push all the tasks to | |
790 | * the ->blkd_tasks lists and wait for this list to drain. This consumes | |
791 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
792 | * so is thus not recommended for any sort of common-case code. | |
793 | * In fact, if you are using synchronize_rcu_expedited() in a loop, | |
794 | * please restructure your code to batch your updates, and then Use a | |
795 | * single synchronize_rcu() instead. | |
019129d5 PM |
796 | */ |
797 | void synchronize_rcu_expedited(void) | |
798 | { | |
d9a3da06 PM |
799 | unsigned long flags; |
800 | struct rcu_node *rnp; | |
801 | struct rcu_state *rsp = &rcu_preempt_state; | |
bcfa57ce | 802 | unsigned long snap; |
d9a3da06 PM |
803 | int trycount = 0; |
804 | ||
805 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
806 | snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1; | |
807 | smp_mb(); /* Above access cannot bleed into critical section. */ | |
808 | ||
1943c89d PM |
809 | /* |
810 | * Block CPU-hotplug operations. This means that any CPU-hotplug | |
811 | * operation that finds an rcu_node structure with tasks in the | |
812 | * process of being boosted will know that all tasks blocking | |
813 | * this expedited grace period will already be in the process of | |
814 | * being boosted. This simplifies the process of moving tasks | |
815 | * from leaf to root rcu_node structures. | |
816 | */ | |
dd56af42 PM |
817 | if (!try_get_online_cpus()) { |
818 | /* CPU-hotplug operation in flight, fall back to normal GP. */ | |
819 | wait_rcu_gp(call_rcu); | |
820 | return; | |
821 | } | |
1943c89d | 822 | |
d9a3da06 PM |
823 | /* |
824 | * Acquire lock, falling back to synchronize_rcu() if too many | |
825 | * lock-acquisition failures. Of course, if someone does the | |
826 | * expedited grace period for us, just leave. | |
827 | */ | |
828 | while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) { | |
1943c89d PM |
829 | if (ULONG_CMP_LT(snap, |
830 | ACCESS_ONCE(sync_rcu_preempt_exp_count))) { | |
831 | put_online_cpus(); | |
832 | goto mb_ret; /* Others did our work for us. */ | |
833 | } | |
c701d5d9 | 834 | if (trycount++ < 10) { |
d9a3da06 | 835 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 836 | } else { |
1943c89d | 837 | put_online_cpus(); |
3705b88d | 838 | wait_rcu_gp(call_rcu); |
d9a3da06 PM |
839 | return; |
840 | } | |
d9a3da06 | 841 | } |
1943c89d PM |
842 | if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) { |
843 | put_online_cpus(); | |
d9a3da06 | 844 | goto unlock_mb_ret; /* Others did our work for us. */ |
1943c89d | 845 | } |
d9a3da06 | 846 | |
12f5f524 | 847 | /* force all RCU readers onto ->blkd_tasks lists. */ |
d9a3da06 PM |
848 | synchronize_sched_expedited(); |
849 | ||
d9a3da06 PM |
850 | /* Initialize ->expmask for all non-leaf rcu_node structures. */ |
851 | rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) { | |
1943c89d | 852 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 853 | smp_mb__after_unlock_lock(); |
d9a3da06 | 854 | rnp->expmask = rnp->qsmaskinit; |
1943c89d | 855 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
d9a3da06 PM |
856 | } |
857 | ||
12f5f524 | 858 | /* Snapshot current state of ->blkd_tasks lists. */ |
d9a3da06 PM |
859 | rcu_for_each_leaf_node(rsp, rnp) |
860 | sync_rcu_preempt_exp_init(rsp, rnp); | |
861 | if (NUM_RCU_NODES > 1) | |
862 | sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp)); | |
863 | ||
1943c89d | 864 | put_online_cpus(); |
d9a3da06 | 865 | |
12f5f524 | 866 | /* Wait for snapshotted ->blkd_tasks lists to drain. */ |
d9a3da06 PM |
867 | rnp = rcu_get_root(rsp); |
868 | wait_event(sync_rcu_preempt_exp_wq, | |
869 | sync_rcu_preempt_exp_done(rnp)); | |
870 | ||
871 | /* Clean up and exit. */ | |
872 | smp_mb(); /* ensure expedited GP seen before counter increment. */ | |
4de376a1 PK |
873 | ACCESS_ONCE(sync_rcu_preempt_exp_count) = |
874 | sync_rcu_preempt_exp_count + 1; | |
d9a3da06 PM |
875 | unlock_mb_ret: |
876 | mutex_unlock(&sync_rcu_preempt_exp_mutex); | |
877 | mb_ret: | |
878 | smp_mb(); /* ensure subsequent action seen after grace period. */ | |
019129d5 PM |
879 | } |
880 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
881 | ||
e74f4c45 PM |
882 | /** |
883 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
f0a0e6f2 PM |
884 | * |
885 | * Note that this primitive does not necessarily wait for an RCU grace period | |
886 | * to complete. For example, if there are no RCU callbacks queued anywhere | |
887 | * in the system, then rcu_barrier() is within its rights to return | |
888 | * immediately, without waiting for anything, much less an RCU grace period. | |
e74f4c45 PM |
889 | */ |
890 | void rcu_barrier(void) | |
891 | { | |
037b64ed | 892 | _rcu_barrier(&rcu_preempt_state); |
e74f4c45 PM |
893 | } |
894 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
895 | ||
1eba8f84 | 896 | /* |
6cc68793 | 897 | * Initialize preemptible RCU's state structures. |
1eba8f84 PM |
898 | */ |
899 | static void __init __rcu_init_preempt(void) | |
900 | { | |
394f99a9 | 901 | rcu_init_one(&rcu_preempt_state, &rcu_preempt_data); |
1eba8f84 PM |
902 | } |
903 | ||
2439b696 PM |
904 | /* |
905 | * Check for a task exiting while in a preemptible-RCU read-side | |
906 | * critical section, clean up if so. No need to issue warnings, | |
907 | * as debug_check_no_locks_held() already does this if lockdep | |
908 | * is enabled. | |
909 | */ | |
910 | void exit_rcu(void) | |
911 | { | |
912 | struct task_struct *t = current; | |
913 | ||
914 | if (likely(list_empty(¤t->rcu_node_entry))) | |
915 | return; | |
916 | t->rcu_read_lock_nesting = 1; | |
917 | barrier(); | |
1d082fd0 | 918 | t->rcu_read_unlock_special.b.blocked = true; |
2439b696 PM |
919 | __rcu_read_unlock(); |
920 | } | |
921 | ||
28f6569a | 922 | #else /* #ifdef CONFIG_PREEMPT_RCU */ |
f41d911f | 923 | |
e534165b | 924 | static struct rcu_state *rcu_state_p = &rcu_sched_state; |
27f4d280 | 925 | |
f41d911f PM |
926 | /* |
927 | * Tell them what RCU they are running. | |
928 | */ | |
0e0fc1c2 | 929 | static void __init rcu_bootup_announce(void) |
f41d911f | 930 | { |
efc151c3 | 931 | pr_info("Hierarchical RCU implementation.\n"); |
26845c28 | 932 | rcu_bootup_announce_oddness(); |
f41d911f PM |
933 | } |
934 | ||
935 | /* | |
936 | * Return the number of RCU batches processed thus far for debug & stats. | |
937 | */ | |
938 | long rcu_batches_completed(void) | |
939 | { | |
940 | return rcu_batches_completed_sched(); | |
941 | } | |
942 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
943 | ||
cba6d0d6 PM |
944 | /* |
945 | * Because preemptible RCU does not exist, we never have to check for | |
946 | * CPUs being in quiescent states. | |
947 | */ | |
38200cf2 | 948 | static void rcu_preempt_note_context_switch(void) |
cba6d0d6 PM |
949 | { |
950 | } | |
951 | ||
fc2219d4 | 952 | /* |
6cc68793 | 953 | * Because preemptible RCU does not exist, there are never any preempted |
fc2219d4 PM |
954 | * RCU readers. |
955 | */ | |
27f4d280 | 956 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 PM |
957 | { |
958 | return 0; | |
959 | } | |
960 | ||
b668c9cf PM |
961 | #ifdef CONFIG_HOTPLUG_CPU |
962 | ||
963 | /* Because preemptible RCU does not exist, no quieting of tasks. */ | |
d3f6bad3 | 964 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
b41d1b92 | 965 | __releases(rnp->lock) |
b668c9cf | 966 | { |
1304afb2 | 967 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf PM |
968 | } |
969 | ||
970 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
971 | ||
1ed509a2 | 972 | /* |
6cc68793 | 973 | * Because preemptible RCU does not exist, we never have to check for |
1ed509a2 PM |
974 | * tasks blocked within RCU read-side critical sections. |
975 | */ | |
976 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
977 | { | |
978 | } | |
979 | ||
f41d911f | 980 | /* |
6cc68793 | 981 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
982 | * tasks blocked within RCU read-side critical sections. |
983 | */ | |
9bc8b558 | 984 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 985 | { |
9bc8b558 | 986 | return 0; |
f41d911f PM |
987 | } |
988 | ||
b0e165c0 | 989 | /* |
6cc68793 | 990 | * Because there is no preemptible RCU, there can be no readers blocked, |
49e29126 PM |
991 | * so there is no need to check for blocked tasks. So check only for |
992 | * bogus qsmask values. | |
b0e165c0 PM |
993 | */ |
994 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
995 | { | |
49e29126 | 996 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
997 | } |
998 | ||
33f76148 PM |
999 | #ifdef CONFIG_HOTPLUG_CPU |
1000 | ||
dd5d19ba | 1001 | /* |
6cc68793 | 1002 | * Because preemptible RCU does not exist, it never needs to migrate |
237c80c5 PM |
1003 | * tasks that were blocked within RCU read-side critical sections, and |
1004 | * such non-existent tasks cannot possibly have been blocking the current | |
1005 | * grace period. | |
dd5d19ba | 1006 | */ |
237c80c5 PM |
1007 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
1008 | struct rcu_node *rnp, | |
1009 | struct rcu_data *rdp) | |
dd5d19ba | 1010 | { |
237c80c5 | 1011 | return 0; |
dd5d19ba PM |
1012 | } |
1013 | ||
e5601400 PM |
1014 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
1015 | ||
f41d911f | 1016 | /* |
6cc68793 | 1017 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
1018 | * to check. |
1019 | */ | |
86aea0e6 | 1020 | static void rcu_preempt_check_callbacks(void) |
f41d911f PM |
1021 | { |
1022 | } | |
1023 | ||
019129d5 PM |
1024 | /* |
1025 | * Wait for an rcu-preempt grace period, but make it happen quickly. | |
6cc68793 | 1026 | * But because preemptible RCU does not exist, map to rcu-sched. |
019129d5 PM |
1027 | */ |
1028 | void synchronize_rcu_expedited(void) | |
1029 | { | |
1030 | synchronize_sched_expedited(); | |
1031 | } | |
1032 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
1033 | ||
d9a3da06 PM |
1034 | #ifdef CONFIG_HOTPLUG_CPU |
1035 | ||
1036 | /* | |
6cc68793 | 1037 | * Because preemptible RCU does not exist, there is never any need to |
d9a3da06 PM |
1038 | * report on tasks preempted in RCU read-side critical sections during |
1039 | * expedited RCU grace periods. | |
1040 | */ | |
b40d293e TG |
1041 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
1042 | bool wake) | |
d9a3da06 | 1043 | { |
d9a3da06 PM |
1044 | } |
1045 | ||
1046 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1047 | ||
e74f4c45 | 1048 | /* |
6cc68793 | 1049 | * Because preemptible RCU does not exist, rcu_barrier() is just |
e74f4c45 PM |
1050 | * another name for rcu_barrier_sched(). |
1051 | */ | |
1052 | void rcu_barrier(void) | |
1053 | { | |
1054 | rcu_barrier_sched(); | |
1055 | } | |
1056 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
1057 | ||
1eba8f84 | 1058 | /* |
6cc68793 | 1059 | * Because preemptible RCU does not exist, it need not be initialized. |
1eba8f84 PM |
1060 | */ |
1061 | static void __init __rcu_init_preempt(void) | |
1062 | { | |
1063 | } | |
1064 | ||
2439b696 PM |
1065 | /* |
1066 | * Because preemptible RCU does not exist, tasks cannot possibly exit | |
1067 | * while in preemptible RCU read-side critical sections. | |
1068 | */ | |
1069 | void exit_rcu(void) | |
1070 | { | |
1071 | } | |
1072 | ||
28f6569a | 1073 | #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ |
8bd93a2c | 1074 | |
27f4d280 PM |
1075 | #ifdef CONFIG_RCU_BOOST |
1076 | ||
1696a8be | 1077 | #include "../locking/rtmutex_common.h" |
27f4d280 | 1078 | |
0ea1f2eb PM |
1079 | #ifdef CONFIG_RCU_TRACE |
1080 | ||
1081 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
1082 | { | |
1083 | if (list_empty(&rnp->blkd_tasks)) | |
1084 | rnp->n_balk_blkd_tasks++; | |
1085 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) | |
1086 | rnp->n_balk_exp_gp_tasks++; | |
1087 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) | |
1088 | rnp->n_balk_boost_tasks++; | |
1089 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) | |
1090 | rnp->n_balk_notblocked++; | |
1091 | else if (rnp->gp_tasks != NULL && | |
a9f4793d | 1092 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
0ea1f2eb PM |
1093 | rnp->n_balk_notyet++; |
1094 | else | |
1095 | rnp->n_balk_nos++; | |
1096 | } | |
1097 | ||
1098 | #else /* #ifdef CONFIG_RCU_TRACE */ | |
1099 | ||
1100 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
1101 | { | |
1102 | } | |
1103 | ||
1104 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ | |
1105 | ||
5d01bbd1 TG |
1106 | static void rcu_wake_cond(struct task_struct *t, int status) |
1107 | { | |
1108 | /* | |
1109 | * If the thread is yielding, only wake it when this | |
1110 | * is invoked from idle | |
1111 | */ | |
1112 | if (status != RCU_KTHREAD_YIELDING || is_idle_task(current)) | |
1113 | wake_up_process(t); | |
1114 | } | |
1115 | ||
27f4d280 PM |
1116 | /* |
1117 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks | |
1118 | * or ->boost_tasks, advancing the pointer to the next task in the | |
1119 | * ->blkd_tasks list. | |
1120 | * | |
1121 | * Note that irqs must be enabled: boosting the task can block. | |
1122 | * Returns 1 if there are more tasks needing to be boosted. | |
1123 | */ | |
1124 | static int rcu_boost(struct rcu_node *rnp) | |
1125 | { | |
1126 | unsigned long flags; | |
27f4d280 PM |
1127 | struct task_struct *t; |
1128 | struct list_head *tb; | |
1129 | ||
1130 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) | |
1131 | return 0; /* Nothing left to boost. */ | |
1132 | ||
1133 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 1134 | smp_mb__after_unlock_lock(); |
27f4d280 PM |
1135 | |
1136 | /* | |
1137 | * Recheck under the lock: all tasks in need of boosting | |
1138 | * might exit their RCU read-side critical sections on their own. | |
1139 | */ | |
1140 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { | |
1141 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1142 | return 0; | |
1143 | } | |
1144 | ||
1145 | /* | |
1146 | * Preferentially boost tasks blocking expedited grace periods. | |
1147 | * This cannot starve the normal grace periods because a second | |
1148 | * expedited grace period must boost all blocked tasks, including | |
1149 | * those blocking the pre-existing normal grace period. | |
1150 | */ | |
0ea1f2eb | 1151 | if (rnp->exp_tasks != NULL) { |
27f4d280 | 1152 | tb = rnp->exp_tasks; |
0ea1f2eb PM |
1153 | rnp->n_exp_boosts++; |
1154 | } else { | |
27f4d280 | 1155 | tb = rnp->boost_tasks; |
0ea1f2eb PM |
1156 | rnp->n_normal_boosts++; |
1157 | } | |
1158 | rnp->n_tasks_boosted++; | |
27f4d280 PM |
1159 | |
1160 | /* | |
1161 | * We boost task t by manufacturing an rt_mutex that appears to | |
1162 | * be held by task t. We leave a pointer to that rt_mutex where | |
1163 | * task t can find it, and task t will release the mutex when it | |
1164 | * exits its outermost RCU read-side critical section. Then | |
1165 | * simply acquiring this artificial rt_mutex will boost task | |
1166 | * t's priority. (Thanks to tglx for suggesting this approach!) | |
1167 | * | |
1168 | * Note that task t must acquire rnp->lock to remove itself from | |
1169 | * the ->blkd_tasks list, which it will do from exit() if from | |
1170 | * nowhere else. We therefore are guaranteed that task t will | |
1171 | * stay around at least until we drop rnp->lock. Note that | |
1172 | * rnp->lock also resolves races between our priority boosting | |
1173 | * and task t's exiting its outermost RCU read-side critical | |
1174 | * section. | |
1175 | */ | |
1176 | t = container_of(tb, struct task_struct, rcu_node_entry); | |
abaa93d9 | 1177 | rt_mutex_init_proxy_locked(&rnp->boost_mtx, t); |
dfeb9765 | 1178 | init_completion(&rnp->boost_completion); |
27f4d280 | 1179 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
abaa93d9 PM |
1180 | /* Lock only for side effect: boosts task t's priority. */ |
1181 | rt_mutex_lock(&rnp->boost_mtx); | |
1182 | rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */ | |
27f4d280 | 1183 | |
abaa93d9 | 1184 | /* Wait for boostee to be done w/boost_mtx before reinitializing. */ |
dfeb9765 | 1185 | wait_for_completion(&rnp->boost_completion); |
27f4d280 | 1186 | |
4f89b336 PM |
1187 | return ACCESS_ONCE(rnp->exp_tasks) != NULL || |
1188 | ACCESS_ONCE(rnp->boost_tasks) != NULL; | |
27f4d280 PM |
1189 | } |
1190 | ||
27f4d280 PM |
1191 | /* |
1192 | * Priority-boosting kthread. One per leaf rcu_node and one for the | |
1193 | * root rcu_node. | |
1194 | */ | |
1195 | static int rcu_boost_kthread(void *arg) | |
1196 | { | |
1197 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
1198 | int spincnt = 0; | |
1199 | int more2boost; | |
1200 | ||
f7f7bac9 | 1201 | trace_rcu_utilization(TPS("Start boost kthread@init")); |
27f4d280 | 1202 | for (;;) { |
d71df90e | 1203 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
f7f7bac9 | 1204 | trace_rcu_utilization(TPS("End boost kthread@rcu_wait")); |
08bca60a | 1205 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
f7f7bac9 | 1206 | trace_rcu_utilization(TPS("Start boost kthread@rcu_wait")); |
d71df90e | 1207 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
27f4d280 PM |
1208 | more2boost = rcu_boost(rnp); |
1209 | if (more2boost) | |
1210 | spincnt++; | |
1211 | else | |
1212 | spincnt = 0; | |
1213 | if (spincnt > 10) { | |
5d01bbd1 | 1214 | rnp->boost_kthread_status = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 1215 | trace_rcu_utilization(TPS("End boost kthread@rcu_yield")); |
5d01bbd1 | 1216 | schedule_timeout_interruptible(2); |
f7f7bac9 | 1217 | trace_rcu_utilization(TPS("Start boost kthread@rcu_yield")); |
27f4d280 PM |
1218 | spincnt = 0; |
1219 | } | |
1220 | } | |
1217ed1b | 1221 | /* NOTREACHED */ |
f7f7bac9 | 1222 | trace_rcu_utilization(TPS("End boost kthread@notreached")); |
27f4d280 PM |
1223 | return 0; |
1224 | } | |
1225 | ||
1226 | /* | |
1227 | * Check to see if it is time to start boosting RCU readers that are | |
1228 | * blocking the current grace period, and, if so, tell the per-rcu_node | |
1229 | * kthread to start boosting them. If there is an expedited grace | |
1230 | * period in progress, it is always time to boost. | |
1231 | * | |
b065a853 PM |
1232 | * The caller must hold rnp->lock, which this function releases. |
1233 | * The ->boost_kthread_task is immortal, so we don't need to worry | |
1234 | * about it going away. | |
27f4d280 | 1235 | */ |
1217ed1b | 1236 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 1237 | __releases(rnp->lock) |
27f4d280 PM |
1238 | { |
1239 | struct task_struct *t; | |
1240 | ||
0ea1f2eb PM |
1241 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
1242 | rnp->n_balk_exp_gp_tasks++; | |
1217ed1b | 1243 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 | 1244 | return; |
0ea1f2eb | 1245 | } |
27f4d280 PM |
1246 | if (rnp->exp_tasks != NULL || |
1247 | (rnp->gp_tasks != NULL && | |
1248 | rnp->boost_tasks == NULL && | |
1249 | rnp->qsmask == 0 && | |
1250 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { | |
1251 | if (rnp->exp_tasks == NULL) | |
1252 | rnp->boost_tasks = rnp->gp_tasks; | |
1217ed1b | 1253 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 | 1254 | t = rnp->boost_kthread_task; |
5d01bbd1 TG |
1255 | if (t) |
1256 | rcu_wake_cond(t, rnp->boost_kthread_status); | |
1217ed1b | 1257 | } else { |
0ea1f2eb | 1258 | rcu_initiate_boost_trace(rnp); |
1217ed1b PM |
1259 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1260 | } | |
27f4d280 PM |
1261 | } |
1262 | ||
a46e0899 PM |
1263 | /* |
1264 | * Wake up the per-CPU kthread to invoke RCU callbacks. | |
1265 | */ | |
1266 | static void invoke_rcu_callbacks_kthread(void) | |
1267 | { | |
1268 | unsigned long flags; | |
1269 | ||
1270 | local_irq_save(flags); | |
1271 | __this_cpu_write(rcu_cpu_has_work, 1); | |
1eb52121 | 1272 | if (__this_cpu_read(rcu_cpu_kthread_task) != NULL && |
5d01bbd1 TG |
1273 | current != __this_cpu_read(rcu_cpu_kthread_task)) { |
1274 | rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task), | |
1275 | __this_cpu_read(rcu_cpu_kthread_status)); | |
1276 | } | |
a46e0899 PM |
1277 | local_irq_restore(flags); |
1278 | } | |
1279 | ||
dff1672d PM |
1280 | /* |
1281 | * Is the current CPU running the RCU-callbacks kthread? | |
1282 | * Caller must have preemption disabled. | |
1283 | */ | |
1284 | static bool rcu_is_callbacks_kthread(void) | |
1285 | { | |
c9d4b0af | 1286 | return __this_cpu_read(rcu_cpu_kthread_task) == current; |
dff1672d PM |
1287 | } |
1288 | ||
27f4d280 PM |
1289 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) |
1290 | ||
1291 | /* | |
1292 | * Do priority-boost accounting for the start of a new grace period. | |
1293 | */ | |
1294 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) | |
1295 | { | |
1296 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; | |
1297 | } | |
1298 | ||
27f4d280 PM |
1299 | /* |
1300 | * Create an RCU-boost kthread for the specified node if one does not | |
1301 | * already exist. We only create this kthread for preemptible RCU. | |
1302 | * Returns zero if all is well, a negated errno otherwise. | |
1303 | */ | |
49fb4c62 | 1304 | static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp, |
5d01bbd1 | 1305 | struct rcu_node *rnp) |
27f4d280 | 1306 | { |
5d01bbd1 | 1307 | int rnp_index = rnp - &rsp->node[0]; |
27f4d280 PM |
1308 | unsigned long flags; |
1309 | struct sched_param sp; | |
1310 | struct task_struct *t; | |
1311 | ||
1312 | if (&rcu_preempt_state != rsp) | |
1313 | return 0; | |
5d01bbd1 TG |
1314 | |
1315 | if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0) | |
1316 | return 0; | |
1317 | ||
a46e0899 | 1318 | rsp->boost = 1; |
27f4d280 PM |
1319 | if (rnp->boost_kthread_task != NULL) |
1320 | return 0; | |
1321 | t = kthread_create(rcu_boost_kthread, (void *)rnp, | |
5b61b0ba | 1322 | "rcub/%d", rnp_index); |
27f4d280 PM |
1323 | if (IS_ERR(t)) |
1324 | return PTR_ERR(t); | |
1325 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 1326 | smp_mb__after_unlock_lock(); |
27f4d280 PM |
1327 | rnp->boost_kthread_task = t; |
1328 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
21871d7e | 1329 | sp.sched_priority = kthread_prio; |
27f4d280 | 1330 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); |
9a432736 | 1331 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
27f4d280 PM |
1332 | return 0; |
1333 | } | |
1334 | ||
f8b7fc6b PM |
1335 | static void rcu_kthread_do_work(void) |
1336 | { | |
c9d4b0af CL |
1337 | rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data)); |
1338 | rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data)); | |
f8b7fc6b PM |
1339 | rcu_preempt_do_callbacks(); |
1340 | } | |
1341 | ||
62ab7072 | 1342 | static void rcu_cpu_kthread_setup(unsigned int cpu) |
f8b7fc6b | 1343 | { |
f8b7fc6b | 1344 | struct sched_param sp; |
f8b7fc6b | 1345 | |
21871d7e | 1346 | sp.sched_priority = kthread_prio; |
62ab7072 | 1347 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); |
f8b7fc6b PM |
1348 | } |
1349 | ||
62ab7072 | 1350 | static void rcu_cpu_kthread_park(unsigned int cpu) |
f8b7fc6b | 1351 | { |
62ab7072 | 1352 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; |
f8b7fc6b PM |
1353 | } |
1354 | ||
62ab7072 | 1355 | static int rcu_cpu_kthread_should_run(unsigned int cpu) |
f8b7fc6b | 1356 | { |
c9d4b0af | 1357 | return __this_cpu_read(rcu_cpu_has_work); |
f8b7fc6b PM |
1358 | } |
1359 | ||
1360 | /* | |
1361 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the | |
e0f23060 PM |
1362 | * RCU softirq used in flavors and configurations of RCU that do not |
1363 | * support RCU priority boosting. | |
f8b7fc6b | 1364 | */ |
62ab7072 | 1365 | static void rcu_cpu_kthread(unsigned int cpu) |
f8b7fc6b | 1366 | { |
c9d4b0af CL |
1367 | unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status); |
1368 | char work, *workp = this_cpu_ptr(&rcu_cpu_has_work); | |
62ab7072 | 1369 | int spincnt; |
f8b7fc6b | 1370 | |
62ab7072 | 1371 | for (spincnt = 0; spincnt < 10; spincnt++) { |
f7f7bac9 | 1372 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait")); |
f8b7fc6b | 1373 | local_bh_disable(); |
f8b7fc6b | 1374 | *statusp = RCU_KTHREAD_RUNNING; |
62ab7072 PM |
1375 | this_cpu_inc(rcu_cpu_kthread_loops); |
1376 | local_irq_disable(); | |
f8b7fc6b PM |
1377 | work = *workp; |
1378 | *workp = 0; | |
62ab7072 | 1379 | local_irq_enable(); |
f8b7fc6b PM |
1380 | if (work) |
1381 | rcu_kthread_do_work(); | |
1382 | local_bh_enable(); | |
62ab7072 | 1383 | if (*workp == 0) { |
f7f7bac9 | 1384 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); |
62ab7072 PM |
1385 | *statusp = RCU_KTHREAD_WAITING; |
1386 | return; | |
f8b7fc6b PM |
1387 | } |
1388 | } | |
62ab7072 | 1389 | *statusp = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 1390 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); |
62ab7072 | 1391 | schedule_timeout_interruptible(2); |
f7f7bac9 | 1392 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); |
62ab7072 | 1393 | *statusp = RCU_KTHREAD_WAITING; |
f8b7fc6b PM |
1394 | } |
1395 | ||
1396 | /* | |
1397 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are | |
1398 | * served by the rcu_node in question. The CPU hotplug lock is still | |
1399 | * held, so the value of rnp->qsmaskinit will be stable. | |
1400 | * | |
1401 | * We don't include outgoingcpu in the affinity set, use -1 if there is | |
1402 | * no outgoing CPU. If there are no CPUs left in the affinity set, | |
1403 | * this function allows the kthread to execute on any CPU. | |
1404 | */ | |
5d01bbd1 | 1405 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b | 1406 | { |
5d01bbd1 TG |
1407 | struct task_struct *t = rnp->boost_kthread_task; |
1408 | unsigned long mask = rnp->qsmaskinit; | |
f8b7fc6b PM |
1409 | cpumask_var_t cm; |
1410 | int cpu; | |
f8b7fc6b | 1411 | |
5d01bbd1 | 1412 | if (!t) |
f8b7fc6b | 1413 | return; |
5d01bbd1 | 1414 | if (!zalloc_cpumask_var(&cm, GFP_KERNEL)) |
f8b7fc6b | 1415 | return; |
f8b7fc6b PM |
1416 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) |
1417 | if ((mask & 0x1) && cpu != outgoingcpu) | |
1418 | cpumask_set_cpu(cpu, cm); | |
1419 | if (cpumask_weight(cm) == 0) { | |
1420 | cpumask_setall(cm); | |
1421 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) | |
1422 | cpumask_clear_cpu(cpu, cm); | |
1423 | WARN_ON_ONCE(cpumask_weight(cm) == 0); | |
1424 | } | |
5d01bbd1 | 1425 | set_cpus_allowed_ptr(t, cm); |
f8b7fc6b PM |
1426 | free_cpumask_var(cm); |
1427 | } | |
1428 | ||
62ab7072 PM |
1429 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { |
1430 | .store = &rcu_cpu_kthread_task, | |
1431 | .thread_should_run = rcu_cpu_kthread_should_run, | |
1432 | .thread_fn = rcu_cpu_kthread, | |
1433 | .thread_comm = "rcuc/%u", | |
1434 | .setup = rcu_cpu_kthread_setup, | |
1435 | .park = rcu_cpu_kthread_park, | |
1436 | }; | |
f8b7fc6b PM |
1437 | |
1438 | /* | |
9386c0b7 | 1439 | * Spawn boost kthreads -- called as soon as the scheduler is running. |
f8b7fc6b | 1440 | */ |
9386c0b7 | 1441 | static void __init rcu_spawn_boost_kthreads(void) |
f8b7fc6b | 1442 | { |
f8b7fc6b | 1443 | struct rcu_node *rnp; |
5d01bbd1 | 1444 | int cpu; |
f8b7fc6b | 1445 | |
62ab7072 | 1446 | for_each_possible_cpu(cpu) |
f8b7fc6b | 1447 | per_cpu(rcu_cpu_has_work, cpu) = 0; |
62ab7072 | 1448 | BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec)); |
e534165b US |
1449 | rnp = rcu_get_root(rcu_state_p); |
1450 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); | |
f8b7fc6b | 1451 | if (NUM_RCU_NODES > 1) { |
e534165b US |
1452 | rcu_for_each_leaf_node(rcu_state_p, rnp) |
1453 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); | |
f8b7fc6b | 1454 | } |
f8b7fc6b | 1455 | } |
f8b7fc6b | 1456 | |
49fb4c62 | 1457 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b | 1458 | { |
e534165b | 1459 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
f8b7fc6b PM |
1460 | struct rcu_node *rnp = rdp->mynode; |
1461 | ||
1462 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ | |
62ab7072 | 1463 | if (rcu_scheduler_fully_active) |
e534165b | 1464 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); |
f8b7fc6b PM |
1465 | } |
1466 | ||
27f4d280 PM |
1467 | #else /* #ifdef CONFIG_RCU_BOOST */ |
1468 | ||
1217ed1b | 1469 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 1470 | __releases(rnp->lock) |
27f4d280 | 1471 | { |
1217ed1b | 1472 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 PM |
1473 | } |
1474 | ||
a46e0899 | 1475 | static void invoke_rcu_callbacks_kthread(void) |
27f4d280 | 1476 | { |
a46e0899 | 1477 | WARN_ON_ONCE(1); |
27f4d280 PM |
1478 | } |
1479 | ||
dff1672d PM |
1480 | static bool rcu_is_callbacks_kthread(void) |
1481 | { | |
1482 | return false; | |
1483 | } | |
1484 | ||
27f4d280 PM |
1485 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
1486 | { | |
1487 | } | |
1488 | ||
5d01bbd1 | 1489 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b PM |
1490 | { |
1491 | } | |
1492 | ||
9386c0b7 | 1493 | static void __init rcu_spawn_boost_kthreads(void) |
b0d30417 | 1494 | { |
b0d30417 | 1495 | } |
b0d30417 | 1496 | |
49fb4c62 | 1497 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b PM |
1498 | { |
1499 | } | |
1500 | ||
27f4d280 PM |
1501 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
1502 | ||
8bd93a2c PM |
1503 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
1504 | ||
1505 | /* | |
1506 | * Check to see if any future RCU-related work will need to be done | |
1507 | * by the current CPU, even if none need be done immediately, returning | |
1508 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
1509 | * an exported member of the RCU API. | |
1510 | * | |
7cb92499 PM |
1511 | * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs |
1512 | * any flavor of RCU. | |
8bd93a2c | 1513 | */ |
ffa83fb5 | 1514 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
aa6da514 | 1515 | int rcu_needs_cpu(unsigned long *delta_jiffies) |
8bd93a2c | 1516 | { |
aa9b1630 | 1517 | *delta_jiffies = ULONG_MAX; |
aa6da514 | 1518 | return rcu_cpu_has_callbacks(NULL); |
7cb92499 | 1519 | } |
ffa83fb5 | 1520 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
7cb92499 PM |
1521 | |
1522 | /* | |
1523 | * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up | |
1524 | * after it. | |
1525 | */ | |
8fa7845d | 1526 | static void rcu_cleanup_after_idle(void) |
7cb92499 PM |
1527 | { |
1528 | } | |
1529 | ||
aea1b35e | 1530 | /* |
a858af28 | 1531 | * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n, |
aea1b35e PM |
1532 | * is nothing. |
1533 | */ | |
198bbf81 | 1534 | static void rcu_prepare_for_idle(void) |
aea1b35e PM |
1535 | { |
1536 | } | |
1537 | ||
c57afe80 PM |
1538 | /* |
1539 | * Don't bother keeping a running count of the number of RCU callbacks | |
1540 | * posted because CONFIG_RCU_FAST_NO_HZ=n. | |
1541 | */ | |
1542 | static void rcu_idle_count_callbacks_posted(void) | |
1543 | { | |
1544 | } | |
1545 | ||
8bd93a2c PM |
1546 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
1547 | ||
f23f7fa1 PM |
1548 | /* |
1549 | * This code is invoked when a CPU goes idle, at which point we want | |
1550 | * to have the CPU do everything required for RCU so that it can enter | |
1551 | * the energy-efficient dyntick-idle mode. This is handled by a | |
1552 | * state machine implemented by rcu_prepare_for_idle() below. | |
1553 | * | |
1554 | * The following three proprocessor symbols control this state machine: | |
1555 | * | |
f23f7fa1 PM |
1556 | * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted |
1557 | * to sleep in dyntick-idle mode with RCU callbacks pending. This | |
1558 | * is sized to be roughly one RCU grace period. Those energy-efficiency | |
1559 | * benchmarkers who might otherwise be tempted to set this to a large | |
1560 | * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your | |
1561 | * system. And if you are -that- concerned about energy efficiency, | |
1562 | * just power the system down and be done with it! | |
778d250a PM |
1563 | * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is |
1564 | * permitted to sleep in dyntick-idle mode with only lazy RCU | |
1565 | * callbacks pending. Setting this too high can OOM your system. | |
f23f7fa1 PM |
1566 | * |
1567 | * The values below work well in practice. If future workloads require | |
1568 | * adjustment, they can be converted into kernel config parameters, though | |
1569 | * making the state machine smarter might be a better option. | |
1570 | */ | |
e84c48ae | 1571 | #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */ |
778d250a | 1572 | #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */ |
f23f7fa1 | 1573 | |
5e44ce35 PM |
1574 | static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY; |
1575 | module_param(rcu_idle_gp_delay, int, 0644); | |
1576 | static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY; | |
1577 | module_param(rcu_idle_lazy_gp_delay, int, 0644); | |
486e2593 | 1578 | |
d689fe22 | 1579 | extern int tick_nohz_active; |
486e2593 PM |
1580 | |
1581 | /* | |
c229828c PM |
1582 | * Try to advance callbacks for all flavors of RCU on the current CPU, but |
1583 | * only if it has been awhile since the last time we did so. Afterwards, | |
1584 | * if there are any callbacks ready for immediate invocation, return true. | |
486e2593 | 1585 | */ |
f1f399d1 | 1586 | static bool __maybe_unused rcu_try_advance_all_cbs(void) |
486e2593 | 1587 | { |
c0f4dfd4 PM |
1588 | bool cbs_ready = false; |
1589 | struct rcu_data *rdp; | |
c229828c | 1590 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1591 | struct rcu_node *rnp; |
1592 | struct rcu_state *rsp; | |
486e2593 | 1593 | |
c229828c PM |
1594 | /* Exit early if we advanced recently. */ |
1595 | if (jiffies == rdtp->last_advance_all) | |
d0bc90fd | 1596 | return false; |
c229828c PM |
1597 | rdtp->last_advance_all = jiffies; |
1598 | ||
c0f4dfd4 PM |
1599 | for_each_rcu_flavor(rsp) { |
1600 | rdp = this_cpu_ptr(rsp->rda); | |
1601 | rnp = rdp->mynode; | |
486e2593 | 1602 | |
c0f4dfd4 PM |
1603 | /* |
1604 | * Don't bother checking unless a grace period has | |
1605 | * completed since we last checked and there are | |
1606 | * callbacks not yet ready to invoke. | |
1607 | */ | |
1608 | if (rdp->completed != rnp->completed && | |
1609 | rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL]) | |
470716fc | 1610 | note_gp_changes(rsp, rdp); |
486e2593 | 1611 | |
c0f4dfd4 PM |
1612 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
1613 | cbs_ready = true; | |
1614 | } | |
1615 | return cbs_ready; | |
486e2593 PM |
1616 | } |
1617 | ||
aa9b1630 | 1618 | /* |
c0f4dfd4 PM |
1619 | * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready |
1620 | * to invoke. If the CPU has callbacks, try to advance them. Tell the | |
1621 | * caller to set the timeout based on whether or not there are non-lazy | |
1622 | * callbacks. | |
aa9b1630 | 1623 | * |
c0f4dfd4 | 1624 | * The caller must have disabled interrupts. |
aa9b1630 | 1625 | */ |
ffa83fb5 | 1626 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
aa6da514 | 1627 | int rcu_needs_cpu(unsigned long *dj) |
aa9b1630 | 1628 | { |
aa6da514 | 1629 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
aa9b1630 | 1630 | |
c0f4dfd4 PM |
1631 | /* Snapshot to detect later posting of non-lazy callback. */ |
1632 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
1633 | ||
aa9b1630 | 1634 | /* If no callbacks, RCU doesn't need the CPU. */ |
aa6da514 | 1635 | if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) { |
c0f4dfd4 | 1636 | *dj = ULONG_MAX; |
aa9b1630 PM |
1637 | return 0; |
1638 | } | |
c0f4dfd4 PM |
1639 | |
1640 | /* Attempt to advance callbacks. */ | |
1641 | if (rcu_try_advance_all_cbs()) { | |
1642 | /* Some ready to invoke, so initiate later invocation. */ | |
1643 | invoke_rcu_core(); | |
aa9b1630 PM |
1644 | return 1; |
1645 | } | |
c0f4dfd4 PM |
1646 | rdtp->last_accelerate = jiffies; |
1647 | ||
1648 | /* Request timer delay depending on laziness, and round. */ | |
6faf7283 | 1649 | if (!rdtp->all_lazy) { |
c0f4dfd4 PM |
1650 | *dj = round_up(rcu_idle_gp_delay + jiffies, |
1651 | rcu_idle_gp_delay) - jiffies; | |
e84c48ae | 1652 | } else { |
c0f4dfd4 | 1653 | *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies; |
e84c48ae | 1654 | } |
aa9b1630 PM |
1655 | return 0; |
1656 | } | |
ffa83fb5 | 1657 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
aa9b1630 | 1658 | |
21e52e15 | 1659 | /* |
c0f4dfd4 PM |
1660 | * Prepare a CPU for idle from an RCU perspective. The first major task |
1661 | * is to sense whether nohz mode has been enabled or disabled via sysfs. | |
1662 | * The second major task is to check to see if a non-lazy callback has | |
1663 | * arrived at a CPU that previously had only lazy callbacks. The third | |
1664 | * major task is to accelerate (that is, assign grace-period numbers to) | |
1665 | * any recently arrived callbacks. | |
aea1b35e PM |
1666 | * |
1667 | * The caller must have disabled interrupts. | |
8bd93a2c | 1668 | */ |
198bbf81 | 1669 | static void rcu_prepare_for_idle(void) |
8bd93a2c | 1670 | { |
f1f399d1 | 1671 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
48a7639c | 1672 | bool needwake; |
c0f4dfd4 | 1673 | struct rcu_data *rdp; |
198bbf81 | 1674 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1675 | struct rcu_node *rnp; |
1676 | struct rcu_state *rsp; | |
9d2ad243 PM |
1677 | int tne; |
1678 | ||
1679 | /* Handle nohz enablement switches conservatively. */ | |
d689fe22 | 1680 | tne = ACCESS_ONCE(tick_nohz_active); |
9d2ad243 | 1681 | if (tne != rdtp->tick_nohz_enabled_snap) { |
aa6da514 | 1682 | if (rcu_cpu_has_callbacks(NULL)) |
9d2ad243 PM |
1683 | invoke_rcu_core(); /* force nohz to see update. */ |
1684 | rdtp->tick_nohz_enabled_snap = tne; | |
1685 | return; | |
1686 | } | |
1687 | if (!tne) | |
1688 | return; | |
f511fc62 | 1689 | |
c0f4dfd4 | 1690 | /* If this is a no-CBs CPU, no callbacks, just return. */ |
198bbf81 | 1691 | if (rcu_is_nocb_cpu(smp_processor_id())) |
9a0c6fef | 1692 | return; |
9a0c6fef | 1693 | |
c57afe80 | 1694 | /* |
c0f4dfd4 PM |
1695 | * If a non-lazy callback arrived at a CPU having only lazy |
1696 | * callbacks, invoke RCU core for the side-effect of recalculating | |
1697 | * idle duration on re-entry to idle. | |
c57afe80 | 1698 | */ |
c0f4dfd4 PM |
1699 | if (rdtp->all_lazy && |
1700 | rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) { | |
c337f8f5 PM |
1701 | rdtp->all_lazy = false; |
1702 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
c0f4dfd4 | 1703 | invoke_rcu_core(); |
c57afe80 PM |
1704 | return; |
1705 | } | |
c57afe80 | 1706 | |
3084f2f8 | 1707 | /* |
c0f4dfd4 PM |
1708 | * If we have not yet accelerated this jiffy, accelerate all |
1709 | * callbacks on this CPU. | |
3084f2f8 | 1710 | */ |
c0f4dfd4 | 1711 | if (rdtp->last_accelerate == jiffies) |
aea1b35e | 1712 | return; |
c0f4dfd4 PM |
1713 | rdtp->last_accelerate = jiffies; |
1714 | for_each_rcu_flavor(rsp) { | |
198bbf81 | 1715 | rdp = this_cpu_ptr(rsp->rda); |
c0f4dfd4 PM |
1716 | if (!*rdp->nxttail[RCU_DONE_TAIL]) |
1717 | continue; | |
1718 | rnp = rdp->mynode; | |
1719 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
6303b9c8 | 1720 | smp_mb__after_unlock_lock(); |
48a7639c | 1721 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
c0f4dfd4 | 1722 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
48a7639c PM |
1723 | if (needwake) |
1724 | rcu_gp_kthread_wake(rsp); | |
77e38ed3 | 1725 | } |
f1f399d1 | 1726 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
c0f4dfd4 | 1727 | } |
3084f2f8 | 1728 | |
c0f4dfd4 PM |
1729 | /* |
1730 | * Clean up for exit from idle. Attempt to advance callbacks based on | |
1731 | * any grace periods that elapsed while the CPU was idle, and if any | |
1732 | * callbacks are now ready to invoke, initiate invocation. | |
1733 | */ | |
8fa7845d | 1734 | static void rcu_cleanup_after_idle(void) |
c0f4dfd4 | 1735 | { |
f1f399d1 | 1736 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
8fa7845d | 1737 | if (rcu_is_nocb_cpu(smp_processor_id())) |
aea1b35e | 1738 | return; |
7a497c96 PM |
1739 | if (rcu_try_advance_all_cbs()) |
1740 | invoke_rcu_core(); | |
f1f399d1 | 1741 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
8bd93a2c PM |
1742 | } |
1743 | ||
c57afe80 | 1744 | /* |
98248a0e PM |
1745 | * Keep a running count of the number of non-lazy callbacks posted |
1746 | * on this CPU. This running counter (which is never decremented) allows | |
1747 | * rcu_prepare_for_idle() to detect when something out of the idle loop | |
1748 | * posts a callback, even if an equal number of callbacks are invoked. | |
1749 | * Of course, callbacks should only be posted from within a trace event | |
1750 | * designed to be called from idle or from within RCU_NONIDLE(). | |
c57afe80 PM |
1751 | */ |
1752 | static void rcu_idle_count_callbacks_posted(void) | |
1753 | { | |
5955f7ee | 1754 | __this_cpu_add(rcu_dynticks.nonlazy_posted, 1); |
c57afe80 PM |
1755 | } |
1756 | ||
b626c1b6 PM |
1757 | /* |
1758 | * Data for flushing lazy RCU callbacks at OOM time. | |
1759 | */ | |
1760 | static atomic_t oom_callback_count; | |
1761 | static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq); | |
1762 | ||
1763 | /* | |
1764 | * RCU OOM callback -- decrement the outstanding count and deliver the | |
1765 | * wake-up if we are the last one. | |
1766 | */ | |
1767 | static void rcu_oom_callback(struct rcu_head *rhp) | |
1768 | { | |
1769 | if (atomic_dec_and_test(&oom_callback_count)) | |
1770 | wake_up(&oom_callback_wq); | |
1771 | } | |
1772 | ||
1773 | /* | |
1774 | * Post an rcu_oom_notify callback on the current CPU if it has at | |
1775 | * least one lazy callback. This will unnecessarily post callbacks | |
1776 | * to CPUs that already have a non-lazy callback at the end of their | |
1777 | * callback list, but this is an infrequent operation, so accept some | |
1778 | * extra overhead to keep things simple. | |
1779 | */ | |
1780 | static void rcu_oom_notify_cpu(void *unused) | |
1781 | { | |
1782 | struct rcu_state *rsp; | |
1783 | struct rcu_data *rdp; | |
1784 | ||
1785 | for_each_rcu_flavor(rsp) { | |
fa07a58f | 1786 | rdp = raw_cpu_ptr(rsp->rda); |
b626c1b6 PM |
1787 | if (rdp->qlen_lazy != 0) { |
1788 | atomic_inc(&oom_callback_count); | |
1789 | rsp->call(&rdp->oom_head, rcu_oom_callback); | |
1790 | } | |
1791 | } | |
1792 | } | |
1793 | ||
1794 | /* | |
1795 | * If low on memory, ensure that each CPU has a non-lazy callback. | |
1796 | * This will wake up CPUs that have only lazy callbacks, in turn | |
1797 | * ensuring that they free up the corresponding memory in a timely manner. | |
1798 | * Because an uncertain amount of memory will be freed in some uncertain | |
1799 | * timeframe, we do not claim to have freed anything. | |
1800 | */ | |
1801 | static int rcu_oom_notify(struct notifier_block *self, | |
1802 | unsigned long notused, void *nfreed) | |
1803 | { | |
1804 | int cpu; | |
1805 | ||
1806 | /* Wait for callbacks from earlier instance to complete. */ | |
1807 | wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0); | |
78e4bc34 | 1808 | smp_mb(); /* Ensure callback reuse happens after callback invocation. */ |
b626c1b6 PM |
1809 | |
1810 | /* | |
1811 | * Prevent premature wakeup: ensure that all increments happen | |
1812 | * before there is a chance of the counter reaching zero. | |
1813 | */ | |
1814 | atomic_set(&oom_callback_count, 1); | |
1815 | ||
1816 | get_online_cpus(); | |
1817 | for_each_online_cpu(cpu) { | |
1818 | smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1); | |
bde6c3aa | 1819 | cond_resched_rcu_qs(); |
b626c1b6 PM |
1820 | } |
1821 | put_online_cpus(); | |
1822 | ||
1823 | /* Unconditionally decrement: no need to wake ourselves up. */ | |
1824 | atomic_dec(&oom_callback_count); | |
1825 | ||
1826 | return NOTIFY_OK; | |
1827 | } | |
1828 | ||
1829 | static struct notifier_block rcu_oom_nb = { | |
1830 | .notifier_call = rcu_oom_notify | |
1831 | }; | |
1832 | ||
1833 | static int __init rcu_register_oom_notifier(void) | |
1834 | { | |
1835 | register_oom_notifier(&rcu_oom_nb); | |
1836 | return 0; | |
1837 | } | |
1838 | early_initcall(rcu_register_oom_notifier); | |
1839 | ||
8bd93a2c | 1840 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
a858af28 PM |
1841 | |
1842 | #ifdef CONFIG_RCU_CPU_STALL_INFO | |
1843 | ||
1844 | #ifdef CONFIG_RCU_FAST_NO_HZ | |
1845 | ||
1846 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1847 | { | |
5955f7ee | 1848 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
c0f4dfd4 | 1849 | unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap; |
a858af28 | 1850 | |
c0f4dfd4 PM |
1851 | sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c", |
1852 | rdtp->last_accelerate & 0xffff, jiffies & 0xffff, | |
1853 | ulong2long(nlpd), | |
1854 | rdtp->all_lazy ? 'L' : '.', | |
1855 | rdtp->tick_nohz_enabled_snap ? '.' : 'D'); | |
a858af28 PM |
1856 | } |
1857 | ||
1858 | #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1859 | ||
1860 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1861 | { | |
1c17e4d4 | 1862 | *cp = '\0'; |
a858af28 PM |
1863 | } |
1864 | ||
1865 | #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1866 | ||
1867 | /* Initiate the stall-info list. */ | |
1868 | static void print_cpu_stall_info_begin(void) | |
1869 | { | |
efc151c3 | 1870 | pr_cont("\n"); |
a858af28 PM |
1871 | } |
1872 | ||
1873 | /* | |
1874 | * Print out diagnostic information for the specified stalled CPU. | |
1875 | * | |
1876 | * If the specified CPU is aware of the current RCU grace period | |
1877 | * (flavor specified by rsp), then print the number of scheduling | |
1878 | * clock interrupts the CPU has taken during the time that it has | |
1879 | * been aware. Otherwise, print the number of RCU grace periods | |
1880 | * that this CPU is ignorant of, for example, "1" if the CPU was | |
1881 | * aware of the previous grace period. | |
1882 | * | |
1883 | * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info. | |
1884 | */ | |
1885 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) | |
1886 | { | |
1887 | char fast_no_hz[72]; | |
1888 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
1889 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
1890 | char *ticks_title; | |
1891 | unsigned long ticks_value; | |
1892 | ||
1893 | if (rsp->gpnum == rdp->gpnum) { | |
1894 | ticks_title = "ticks this GP"; | |
1895 | ticks_value = rdp->ticks_this_gp; | |
1896 | } else { | |
1897 | ticks_title = "GPs behind"; | |
1898 | ticks_value = rsp->gpnum - rdp->gpnum; | |
1899 | } | |
1900 | print_cpu_stall_fast_no_hz(fast_no_hz, cpu); | |
efc151c3 | 1901 | pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n", |
a858af28 PM |
1902 | cpu, ticks_value, ticks_title, |
1903 | atomic_read(&rdtp->dynticks) & 0xfff, | |
1904 | rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting, | |
6231069b | 1905 | rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu), |
a858af28 PM |
1906 | fast_no_hz); |
1907 | } | |
1908 | ||
1909 | /* Terminate the stall-info list. */ | |
1910 | static void print_cpu_stall_info_end(void) | |
1911 | { | |
efc151c3 | 1912 | pr_err("\t"); |
a858af28 PM |
1913 | } |
1914 | ||
1915 | /* Zero ->ticks_this_gp for all flavors of RCU. */ | |
1916 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) | |
1917 | { | |
1918 | rdp->ticks_this_gp = 0; | |
6231069b | 1919 | rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id()); |
a858af28 PM |
1920 | } |
1921 | ||
1922 | /* Increment ->ticks_this_gp for all flavors of RCU. */ | |
1923 | static void increment_cpu_stall_ticks(void) | |
1924 | { | |
115f7a7c PM |
1925 | struct rcu_state *rsp; |
1926 | ||
1927 | for_each_rcu_flavor(rsp) | |
fa07a58f | 1928 | raw_cpu_inc(rsp->rda->ticks_this_gp); |
a858af28 PM |
1929 | } |
1930 | ||
1931 | #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */ | |
1932 | ||
1933 | static void print_cpu_stall_info_begin(void) | |
1934 | { | |
efc151c3 | 1935 | pr_cont(" {"); |
a858af28 PM |
1936 | } |
1937 | ||
1938 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) | |
1939 | { | |
efc151c3 | 1940 | pr_cont(" %d", cpu); |
a858af28 PM |
1941 | } |
1942 | ||
1943 | static void print_cpu_stall_info_end(void) | |
1944 | { | |
efc151c3 | 1945 | pr_cont("} "); |
a858af28 PM |
1946 | } |
1947 | ||
1948 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) | |
1949 | { | |
1950 | } | |
1951 | ||
1952 | static void increment_cpu_stall_ticks(void) | |
1953 | { | |
1954 | } | |
1955 | ||
1956 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */ | |
3fbfbf7a PM |
1957 | |
1958 | #ifdef CONFIG_RCU_NOCB_CPU | |
1959 | ||
1960 | /* | |
1961 | * Offload callback processing from the boot-time-specified set of CPUs | |
1962 | * specified by rcu_nocb_mask. For each CPU in the set, there is a | |
1963 | * kthread created that pulls the callbacks from the corresponding CPU, | |
1964 | * waits for a grace period to elapse, and invokes the callbacks. | |
1965 | * The no-CBs CPUs do a wake_up() on their kthread when they insert | |
1966 | * a callback into any empty list, unless the rcu_nocb_poll boot parameter | |
1967 | * has been specified, in which case each kthread actively polls its | |
1968 | * CPU. (Which isn't so great for energy efficiency, but which does | |
1969 | * reduce RCU's overhead on that CPU.) | |
1970 | * | |
1971 | * This is intended to be used in conjunction with Frederic Weisbecker's | |
1972 | * adaptive-idle work, which would seriously reduce OS jitter on CPUs | |
1973 | * running CPU-bound user-mode computations. | |
1974 | * | |
1975 | * Offloading of callback processing could also in theory be used as | |
1976 | * an energy-efficiency measure because CPUs with no RCU callbacks | |
1977 | * queued are more aggressive about entering dyntick-idle mode. | |
1978 | */ | |
1979 | ||
1980 | ||
1981 | /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */ | |
1982 | static int __init rcu_nocb_setup(char *str) | |
1983 | { | |
1984 | alloc_bootmem_cpumask_var(&rcu_nocb_mask); | |
1985 | have_rcu_nocb_mask = true; | |
1986 | cpulist_parse(str, rcu_nocb_mask); | |
1987 | return 1; | |
1988 | } | |
1989 | __setup("rcu_nocbs=", rcu_nocb_setup); | |
1990 | ||
1b0048a4 PG |
1991 | static int __init parse_rcu_nocb_poll(char *arg) |
1992 | { | |
1993 | rcu_nocb_poll = 1; | |
1994 | return 0; | |
1995 | } | |
1996 | early_param("rcu_nocb_poll", parse_rcu_nocb_poll); | |
1997 | ||
dae6e64d | 1998 | /* |
0446be48 PM |
1999 | * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended |
2000 | * grace period. | |
dae6e64d | 2001 | */ |
0446be48 | 2002 | static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
dae6e64d | 2003 | { |
0446be48 | 2004 | wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]); |
dae6e64d PM |
2005 | } |
2006 | ||
2007 | /* | |
8b425aa8 | 2008 | * Set the root rcu_node structure's ->need_future_gp field |
dae6e64d PM |
2009 | * based on the sum of those of all rcu_node structures. This does |
2010 | * double-count the root rcu_node structure's requests, but this | |
2011 | * is necessary to handle the possibility of a rcu_nocb_kthread() | |
2012 | * having awakened during the time that the rcu_node structures | |
2013 | * were being updated for the end of the previous grace period. | |
34ed6246 | 2014 | */ |
dae6e64d PM |
2015 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
2016 | { | |
8b425aa8 | 2017 | rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq; |
dae6e64d PM |
2018 | } |
2019 | ||
2020 | static void rcu_init_one_nocb(struct rcu_node *rnp) | |
34ed6246 | 2021 | { |
dae6e64d PM |
2022 | init_waitqueue_head(&rnp->nocb_gp_wq[0]); |
2023 | init_waitqueue_head(&rnp->nocb_gp_wq[1]); | |
34ed6246 PM |
2024 | } |
2025 | ||
2f33b512 | 2026 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
24342c96 | 2027 | /* Is the specified CPU a no-CBs CPU? */ |
d1e43fa5 | 2028 | bool rcu_is_nocb_cpu(int cpu) |
3fbfbf7a PM |
2029 | { |
2030 | if (have_rcu_nocb_mask) | |
2031 | return cpumask_test_cpu(cpu, rcu_nocb_mask); | |
2032 | return false; | |
2033 | } | |
2f33b512 | 2034 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
3fbfbf7a | 2035 | |
fbce7497 PM |
2036 | /* |
2037 | * Kick the leader kthread for this NOCB group. | |
2038 | */ | |
2039 | static void wake_nocb_leader(struct rcu_data *rdp, bool force) | |
2040 | { | |
2041 | struct rcu_data *rdp_leader = rdp->nocb_leader; | |
2042 | ||
2043 | if (!ACCESS_ONCE(rdp_leader->nocb_kthread)) | |
2044 | return; | |
11ed7f93 | 2045 | if (ACCESS_ONCE(rdp_leader->nocb_leader_sleep) || force) { |
39953dfd | 2046 | /* Prior smp_mb__after_atomic() orders against prior enqueue. */ |
11ed7f93 | 2047 | ACCESS_ONCE(rdp_leader->nocb_leader_sleep) = false; |
fbce7497 PM |
2048 | wake_up(&rdp_leader->nocb_wq); |
2049 | } | |
2050 | } | |
2051 | ||
d7e29933 PM |
2052 | /* |
2053 | * Does the specified CPU need an RCU callback for the specified flavor | |
2054 | * of rcu_barrier()? | |
2055 | */ | |
2056 | static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu) | |
2057 | { | |
2058 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
41050a00 PM |
2059 | unsigned long ret; |
2060 | #ifdef CONFIG_PROVE_RCU | |
d7e29933 | 2061 | struct rcu_head *rhp; |
41050a00 PM |
2062 | #endif /* #ifdef CONFIG_PROVE_RCU */ |
2063 | ||
2064 | /* | |
2065 | * Check count of all no-CBs callbacks awaiting invocation. | |
2066 | * There needs to be a barrier before this function is called, | |
2067 | * but associated with a prior determination that no more | |
2068 | * callbacks would be posted. In the worst case, the first | |
2069 | * barrier in _rcu_barrier() suffices (but the caller cannot | |
2070 | * necessarily rely on this, not a substitute for the caller | |
2071 | * getting the concurrency design right!). There must also be | |
2072 | * a barrier between the following load an posting of a callback | |
2073 | * (if a callback is in fact needed). This is associated with an | |
2074 | * atomic_inc() in the caller. | |
2075 | */ | |
2076 | ret = atomic_long_read(&rdp->nocb_q_count); | |
d7e29933 | 2077 | |
41050a00 | 2078 | #ifdef CONFIG_PROVE_RCU |
d7e29933 PM |
2079 | rhp = ACCESS_ONCE(rdp->nocb_head); |
2080 | if (!rhp) | |
2081 | rhp = ACCESS_ONCE(rdp->nocb_gp_head); | |
2082 | if (!rhp) | |
2083 | rhp = ACCESS_ONCE(rdp->nocb_follower_head); | |
2084 | ||
2085 | /* Having no rcuo kthread but CBs after scheduler starts is bad! */ | |
2086 | if (!ACCESS_ONCE(rdp->nocb_kthread) && rhp) { | |
2087 | /* RCU callback enqueued before CPU first came online??? */ | |
2088 | pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n", | |
2089 | cpu, rhp->func); | |
2090 | WARN_ON_ONCE(1); | |
2091 | } | |
41050a00 | 2092 | #endif /* #ifdef CONFIG_PROVE_RCU */ |
d7e29933 | 2093 | |
41050a00 | 2094 | return !!ret; |
d7e29933 PM |
2095 | } |
2096 | ||
3fbfbf7a PM |
2097 | /* |
2098 | * Enqueue the specified string of rcu_head structures onto the specified | |
2099 | * CPU's no-CBs lists. The CPU is specified by rdp, the head of the | |
2100 | * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy | |
2101 | * counts are supplied by rhcount and rhcount_lazy. | |
2102 | * | |
2103 | * If warranted, also wake up the kthread servicing this CPUs queues. | |
2104 | */ | |
2105 | static void __call_rcu_nocb_enqueue(struct rcu_data *rdp, | |
2106 | struct rcu_head *rhp, | |
2107 | struct rcu_head **rhtp, | |
96d3fd0d PM |
2108 | int rhcount, int rhcount_lazy, |
2109 | unsigned long flags) | |
3fbfbf7a PM |
2110 | { |
2111 | int len; | |
2112 | struct rcu_head **old_rhpp; | |
2113 | struct task_struct *t; | |
2114 | ||
2115 | /* Enqueue the callback on the nocb list and update counts. */ | |
41050a00 PM |
2116 | atomic_long_add(rhcount, &rdp->nocb_q_count); |
2117 | /* rcu_barrier() relies on ->nocb_q_count add before xchg. */ | |
3fbfbf7a PM |
2118 | old_rhpp = xchg(&rdp->nocb_tail, rhtp); |
2119 | ACCESS_ONCE(*old_rhpp) = rhp; | |
3fbfbf7a | 2120 | atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy); |
39953dfd | 2121 | smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */ |
3fbfbf7a PM |
2122 | |
2123 | /* If we are not being polled and there is a kthread, awaken it ... */ | |
2124 | t = ACCESS_ONCE(rdp->nocb_kthread); | |
25e03a74 | 2125 | if (rcu_nocb_poll || !t) { |
9261dd0d PM |
2126 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2127 | TPS("WakeNotPoll")); | |
3fbfbf7a | 2128 | return; |
9261dd0d | 2129 | } |
3fbfbf7a PM |
2130 | len = atomic_long_read(&rdp->nocb_q_count); |
2131 | if (old_rhpp == &rdp->nocb_head) { | |
96d3fd0d | 2132 | if (!irqs_disabled_flags(flags)) { |
fbce7497 PM |
2133 | /* ... if queue was empty ... */ |
2134 | wake_nocb_leader(rdp, false); | |
96d3fd0d PM |
2135 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2136 | TPS("WakeEmpty")); | |
2137 | } else { | |
9fdd3bc9 | 2138 | rdp->nocb_defer_wakeup = RCU_NOGP_WAKE; |
96d3fd0d PM |
2139 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2140 | TPS("WakeEmptyIsDeferred")); | |
2141 | } | |
3fbfbf7a PM |
2142 | rdp->qlen_last_fqs_check = 0; |
2143 | } else if (len > rdp->qlen_last_fqs_check + qhimark) { | |
fbce7497 | 2144 | /* ... or if many callbacks queued. */ |
9fdd3bc9 PM |
2145 | if (!irqs_disabled_flags(flags)) { |
2146 | wake_nocb_leader(rdp, true); | |
2147 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2148 | TPS("WakeOvf")); | |
2149 | } else { | |
2150 | rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE; | |
2151 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2152 | TPS("WakeOvfIsDeferred")); | |
2153 | } | |
3fbfbf7a | 2154 | rdp->qlen_last_fqs_check = LONG_MAX / 2; |
9261dd0d PM |
2155 | } else { |
2156 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot")); | |
3fbfbf7a PM |
2157 | } |
2158 | return; | |
2159 | } | |
2160 | ||
2161 | /* | |
2162 | * This is a helper for __call_rcu(), which invokes this when the normal | |
2163 | * callback queue is inoperable. If this is not a no-CBs CPU, this | |
2164 | * function returns failure back to __call_rcu(), which can complain | |
2165 | * appropriately. | |
2166 | * | |
2167 | * Otherwise, this function queues the callback where the corresponding | |
2168 | * "rcuo" kthread can find it. | |
2169 | */ | |
2170 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, | |
96d3fd0d | 2171 | bool lazy, unsigned long flags) |
3fbfbf7a PM |
2172 | { |
2173 | ||
d1e43fa5 | 2174 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
c271d3a9 | 2175 | return false; |
96d3fd0d | 2176 | __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags); |
21e7a608 PM |
2177 | if (__is_kfree_rcu_offset((unsigned long)rhp->func)) |
2178 | trace_rcu_kfree_callback(rdp->rsp->name, rhp, | |
2179 | (unsigned long)rhp->func, | |
756cbf6b PM |
2180 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
2181 | -atomic_long_read(&rdp->nocb_q_count)); | |
21e7a608 PM |
2182 | else |
2183 | trace_rcu_callback(rdp->rsp->name, rhp, | |
756cbf6b PM |
2184 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
2185 | -atomic_long_read(&rdp->nocb_q_count)); | |
1772947b PM |
2186 | |
2187 | /* | |
2188 | * If called from an extended quiescent state with interrupts | |
2189 | * disabled, invoke the RCU core in order to allow the idle-entry | |
2190 | * deferred-wakeup check to function. | |
2191 | */ | |
2192 | if (irqs_disabled_flags(flags) && | |
2193 | !rcu_is_watching() && | |
2194 | cpu_online(smp_processor_id())) | |
2195 | invoke_rcu_core(); | |
2196 | ||
c271d3a9 | 2197 | return true; |
3fbfbf7a PM |
2198 | } |
2199 | ||
2200 | /* | |
2201 | * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is | |
2202 | * not a no-CBs CPU. | |
2203 | */ | |
2204 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
2205 | struct rcu_data *rdp, |
2206 | unsigned long flags) | |
3fbfbf7a PM |
2207 | { |
2208 | long ql = rsp->qlen; | |
2209 | long qll = rsp->qlen_lazy; | |
2210 | ||
2211 | /* If this is not a no-CBs CPU, tell the caller to do it the old way. */ | |
d1e43fa5 | 2212 | if (!rcu_is_nocb_cpu(smp_processor_id())) |
0a9e1e11 | 2213 | return false; |
3fbfbf7a PM |
2214 | rsp->qlen = 0; |
2215 | rsp->qlen_lazy = 0; | |
2216 | ||
2217 | /* First, enqueue the donelist, if any. This preserves CB ordering. */ | |
2218 | if (rsp->orphan_donelist != NULL) { | |
2219 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist, | |
96d3fd0d | 2220 | rsp->orphan_donetail, ql, qll, flags); |
3fbfbf7a PM |
2221 | ql = qll = 0; |
2222 | rsp->orphan_donelist = NULL; | |
2223 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2224 | } | |
2225 | if (rsp->orphan_nxtlist != NULL) { | |
2226 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist, | |
96d3fd0d | 2227 | rsp->orphan_nxttail, ql, qll, flags); |
3fbfbf7a PM |
2228 | ql = qll = 0; |
2229 | rsp->orphan_nxtlist = NULL; | |
2230 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2231 | } | |
0a9e1e11 | 2232 | return true; |
3fbfbf7a PM |
2233 | } |
2234 | ||
2235 | /* | |
34ed6246 PM |
2236 | * If necessary, kick off a new grace period, and either way wait |
2237 | * for a subsequent grace period to complete. | |
3fbfbf7a | 2238 | */ |
34ed6246 | 2239 | static void rcu_nocb_wait_gp(struct rcu_data *rdp) |
3fbfbf7a | 2240 | { |
34ed6246 | 2241 | unsigned long c; |
dae6e64d | 2242 | bool d; |
34ed6246 | 2243 | unsigned long flags; |
48a7639c | 2244 | bool needwake; |
34ed6246 PM |
2245 | struct rcu_node *rnp = rdp->mynode; |
2246 | ||
2247 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 2248 | smp_mb__after_unlock_lock(); |
48a7639c | 2249 | needwake = rcu_start_future_gp(rnp, rdp, &c); |
0446be48 | 2250 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
48a7639c PM |
2251 | if (needwake) |
2252 | rcu_gp_kthread_wake(rdp->rsp); | |
3fbfbf7a PM |
2253 | |
2254 | /* | |
34ed6246 PM |
2255 | * Wait for the grace period. Do so interruptibly to avoid messing |
2256 | * up the load average. | |
3fbfbf7a | 2257 | */ |
f7f7bac9 | 2258 | trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait")); |
34ed6246 | 2259 | for (;;) { |
dae6e64d PM |
2260 | wait_event_interruptible( |
2261 | rnp->nocb_gp_wq[c & 0x1], | |
2262 | (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c))); | |
2263 | if (likely(d)) | |
34ed6246 | 2264 | break; |
73a860cd | 2265 | WARN_ON(signal_pending(current)); |
f7f7bac9 | 2266 | trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait")); |
34ed6246 | 2267 | } |
f7f7bac9 | 2268 | trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait")); |
34ed6246 | 2269 | smp_mb(); /* Ensure that CB invocation happens after GP end. */ |
3fbfbf7a PM |
2270 | } |
2271 | ||
fbce7497 PM |
2272 | /* |
2273 | * Leaders come here to wait for additional callbacks to show up. | |
2274 | * This function does not return until callbacks appear. | |
2275 | */ | |
2276 | static void nocb_leader_wait(struct rcu_data *my_rdp) | |
2277 | { | |
2278 | bool firsttime = true; | |
2279 | bool gotcbs; | |
2280 | struct rcu_data *rdp; | |
2281 | struct rcu_head **tail; | |
2282 | ||
2283 | wait_again: | |
2284 | ||
2285 | /* Wait for callbacks to appear. */ | |
2286 | if (!rcu_nocb_poll) { | |
2287 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep"); | |
2288 | wait_event_interruptible(my_rdp->nocb_wq, | |
11ed7f93 | 2289 | !ACCESS_ONCE(my_rdp->nocb_leader_sleep)); |
fbce7497 PM |
2290 | /* Memory barrier handled by smp_mb() calls below and repoll. */ |
2291 | } else if (firsttime) { | |
2292 | firsttime = false; /* Don't drown trace log with "Poll"! */ | |
2293 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll"); | |
2294 | } | |
2295 | ||
2296 | /* | |
2297 | * Each pass through the following loop checks a follower for CBs. | |
2298 | * We are our own first follower. Any CBs found are moved to | |
2299 | * nocb_gp_head, where they await a grace period. | |
2300 | */ | |
2301 | gotcbs = false; | |
2302 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
2303 | rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head); | |
2304 | if (!rdp->nocb_gp_head) | |
2305 | continue; /* No CBs here, try next follower. */ | |
2306 | ||
2307 | /* Move callbacks to wait-for-GP list, which is empty. */ | |
2308 | ACCESS_ONCE(rdp->nocb_head) = NULL; | |
2309 | rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head); | |
fbce7497 PM |
2310 | gotcbs = true; |
2311 | } | |
2312 | ||
2313 | /* | |
2314 | * If there were no callbacks, sleep a bit, rescan after a | |
2315 | * memory barrier, and go retry. | |
2316 | */ | |
2317 | if (unlikely(!gotcbs)) { | |
2318 | if (!rcu_nocb_poll) | |
2319 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, | |
2320 | "WokeEmpty"); | |
73a860cd | 2321 | WARN_ON(signal_pending(current)); |
fbce7497 PM |
2322 | schedule_timeout_interruptible(1); |
2323 | ||
2324 | /* Rescan in case we were a victim of memory ordering. */ | |
11ed7f93 PK |
2325 | my_rdp->nocb_leader_sleep = true; |
2326 | smp_mb(); /* Ensure _sleep true before scan. */ | |
fbce7497 PM |
2327 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) |
2328 | if (ACCESS_ONCE(rdp->nocb_head)) { | |
2329 | /* Found CB, so short-circuit next wait. */ | |
11ed7f93 | 2330 | my_rdp->nocb_leader_sleep = false; |
fbce7497 PM |
2331 | break; |
2332 | } | |
2333 | goto wait_again; | |
2334 | } | |
2335 | ||
2336 | /* Wait for one grace period. */ | |
2337 | rcu_nocb_wait_gp(my_rdp); | |
2338 | ||
2339 | /* | |
11ed7f93 PK |
2340 | * We left ->nocb_leader_sleep unset to reduce cache thrashing. |
2341 | * We set it now, but recheck for new callbacks while | |
fbce7497 PM |
2342 | * traversing our follower list. |
2343 | */ | |
11ed7f93 PK |
2344 | my_rdp->nocb_leader_sleep = true; |
2345 | smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */ | |
fbce7497 PM |
2346 | |
2347 | /* Each pass through the following loop wakes a follower, if needed. */ | |
2348 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
2349 | if (ACCESS_ONCE(rdp->nocb_head)) | |
11ed7f93 | 2350 | my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/ |
fbce7497 PM |
2351 | if (!rdp->nocb_gp_head) |
2352 | continue; /* No CBs, so no need to wake follower. */ | |
2353 | ||
2354 | /* Append callbacks to follower's "done" list. */ | |
2355 | tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail); | |
2356 | *tail = rdp->nocb_gp_head; | |
c847f142 | 2357 | smp_mb__after_atomic(); /* Store *tail before wakeup. */ |
fbce7497 PM |
2358 | if (rdp != my_rdp && tail == &rdp->nocb_follower_head) { |
2359 | /* | |
2360 | * List was empty, wake up the follower. | |
2361 | * Memory barriers supplied by atomic_long_add(). | |
2362 | */ | |
2363 | wake_up(&rdp->nocb_wq); | |
2364 | } | |
2365 | } | |
2366 | ||
2367 | /* If we (the leader) don't have CBs, go wait some more. */ | |
2368 | if (!my_rdp->nocb_follower_head) | |
2369 | goto wait_again; | |
2370 | } | |
2371 | ||
2372 | /* | |
2373 | * Followers come here to wait for additional callbacks to show up. | |
2374 | * This function does not return until callbacks appear. | |
2375 | */ | |
2376 | static void nocb_follower_wait(struct rcu_data *rdp) | |
2377 | { | |
2378 | bool firsttime = true; | |
2379 | ||
2380 | for (;;) { | |
2381 | if (!rcu_nocb_poll) { | |
2382 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2383 | "FollowerSleep"); | |
2384 | wait_event_interruptible(rdp->nocb_wq, | |
2385 | ACCESS_ONCE(rdp->nocb_follower_head)); | |
2386 | } else if (firsttime) { | |
2387 | /* Don't drown trace log with "Poll"! */ | |
2388 | firsttime = false; | |
2389 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll"); | |
2390 | } | |
2391 | if (smp_load_acquire(&rdp->nocb_follower_head)) { | |
2392 | /* ^^^ Ensure CB invocation follows _head test. */ | |
2393 | return; | |
2394 | } | |
2395 | if (!rcu_nocb_poll) | |
2396 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2397 | "WokeEmpty"); | |
73a860cd | 2398 | WARN_ON(signal_pending(current)); |
fbce7497 PM |
2399 | schedule_timeout_interruptible(1); |
2400 | } | |
2401 | } | |
2402 | ||
3fbfbf7a PM |
2403 | /* |
2404 | * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes | |
fbce7497 PM |
2405 | * callbacks queued by the corresponding no-CBs CPU, however, there is |
2406 | * an optional leader-follower relationship so that the grace-period | |
2407 | * kthreads don't have to do quite so many wakeups. | |
3fbfbf7a PM |
2408 | */ |
2409 | static int rcu_nocb_kthread(void *arg) | |
2410 | { | |
2411 | int c, cl; | |
2412 | struct rcu_head *list; | |
2413 | struct rcu_head *next; | |
2414 | struct rcu_head **tail; | |
2415 | struct rcu_data *rdp = arg; | |
2416 | ||
2417 | /* Each pass through this loop invokes one batch of callbacks */ | |
2418 | for (;;) { | |
fbce7497 PM |
2419 | /* Wait for callbacks. */ |
2420 | if (rdp->nocb_leader == rdp) | |
2421 | nocb_leader_wait(rdp); | |
2422 | else | |
2423 | nocb_follower_wait(rdp); | |
2424 | ||
2425 | /* Pull the ready-to-invoke callbacks onto local list. */ | |
2426 | list = ACCESS_ONCE(rdp->nocb_follower_head); | |
2427 | BUG_ON(!list); | |
2428 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty"); | |
2429 | ACCESS_ONCE(rdp->nocb_follower_head) = NULL; | |
2430 | tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head); | |
3fbfbf7a PM |
2431 | |
2432 | /* Each pass through the following loop invokes a callback. */ | |
41050a00 PM |
2433 | trace_rcu_batch_start(rdp->rsp->name, |
2434 | atomic_long_read(&rdp->nocb_q_count_lazy), | |
2435 | atomic_long_read(&rdp->nocb_q_count), -1); | |
3fbfbf7a PM |
2436 | c = cl = 0; |
2437 | while (list) { | |
2438 | next = list->next; | |
2439 | /* Wait for enqueuing to complete, if needed. */ | |
2440 | while (next == NULL && &list->next != tail) { | |
69a79bb1 PM |
2441 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2442 | TPS("WaitQueue")); | |
3fbfbf7a | 2443 | schedule_timeout_interruptible(1); |
69a79bb1 PM |
2444 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2445 | TPS("WokeQueue")); | |
3fbfbf7a PM |
2446 | next = list->next; |
2447 | } | |
2448 | debug_rcu_head_unqueue(list); | |
2449 | local_bh_disable(); | |
2450 | if (__rcu_reclaim(rdp->rsp->name, list)) | |
2451 | cl++; | |
2452 | c++; | |
2453 | local_bh_enable(); | |
2454 | list = next; | |
2455 | } | |
2456 | trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); | |
41050a00 PM |
2457 | smp_mb__before_atomic(); /* _add after CB invocation. */ |
2458 | atomic_long_add(-c, &rdp->nocb_q_count); | |
2459 | atomic_long_add(-cl, &rdp->nocb_q_count_lazy); | |
c635a4e1 | 2460 | rdp->n_nocbs_invoked += c; |
3fbfbf7a PM |
2461 | } |
2462 | return 0; | |
2463 | } | |
2464 | ||
96d3fd0d | 2465 | /* Is a deferred wakeup of rcu_nocb_kthread() required? */ |
9fdd3bc9 | 2466 | static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
96d3fd0d PM |
2467 | { |
2468 | return ACCESS_ONCE(rdp->nocb_defer_wakeup); | |
2469 | } | |
2470 | ||
2471 | /* Do a deferred wakeup of rcu_nocb_kthread(). */ | |
2472 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2473 | { | |
9fdd3bc9 PM |
2474 | int ndw; |
2475 | ||
96d3fd0d PM |
2476 | if (!rcu_nocb_need_deferred_wakeup(rdp)) |
2477 | return; | |
9fdd3bc9 PM |
2478 | ndw = ACCESS_ONCE(rdp->nocb_defer_wakeup); |
2479 | ACCESS_ONCE(rdp->nocb_defer_wakeup) = RCU_NOGP_WAKE_NOT; | |
2480 | wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE); | |
2481 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake")); | |
96d3fd0d PM |
2482 | } |
2483 | ||
f4579fc5 PM |
2484 | void __init rcu_init_nohz(void) |
2485 | { | |
2486 | int cpu; | |
2487 | bool need_rcu_nocb_mask = true; | |
2488 | struct rcu_state *rsp; | |
2489 | ||
2490 | #ifdef CONFIG_RCU_NOCB_CPU_NONE | |
2491 | need_rcu_nocb_mask = false; | |
2492 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */ | |
2493 | ||
2494 | #if defined(CONFIG_NO_HZ_FULL) | |
2495 | if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask)) | |
2496 | need_rcu_nocb_mask = true; | |
2497 | #endif /* #if defined(CONFIG_NO_HZ_FULL) */ | |
2498 | ||
2499 | if (!have_rcu_nocb_mask && need_rcu_nocb_mask) { | |
949cccdb PK |
2500 | if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) { |
2501 | pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n"); | |
2502 | return; | |
2503 | } | |
f4579fc5 PM |
2504 | have_rcu_nocb_mask = true; |
2505 | } | |
2506 | if (!have_rcu_nocb_mask) | |
2507 | return; | |
2508 | ||
2509 | #ifdef CONFIG_RCU_NOCB_CPU_ZERO | |
2510 | pr_info("\tOffload RCU callbacks from CPU 0\n"); | |
2511 | cpumask_set_cpu(0, rcu_nocb_mask); | |
2512 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */ | |
2513 | #ifdef CONFIG_RCU_NOCB_CPU_ALL | |
2514 | pr_info("\tOffload RCU callbacks from all CPUs\n"); | |
2515 | cpumask_copy(rcu_nocb_mask, cpu_possible_mask); | |
2516 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */ | |
2517 | #if defined(CONFIG_NO_HZ_FULL) | |
2518 | if (tick_nohz_full_running) | |
2519 | cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask); | |
2520 | #endif /* #if defined(CONFIG_NO_HZ_FULL) */ | |
2521 | ||
2522 | if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) { | |
2523 | pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n"); | |
2524 | cpumask_and(rcu_nocb_mask, cpu_possible_mask, | |
2525 | rcu_nocb_mask); | |
2526 | } | |
2527 | cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask); | |
2528 | pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf); | |
2529 | if (rcu_nocb_poll) | |
2530 | pr_info("\tPoll for callbacks from no-CBs CPUs.\n"); | |
2531 | ||
2532 | for_each_rcu_flavor(rsp) { | |
2533 | for_each_cpu(cpu, rcu_nocb_mask) { | |
2534 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
2535 | ||
2536 | /* | |
2537 | * If there are early callbacks, they will need | |
2538 | * to be moved to the nocb lists. | |
2539 | */ | |
2540 | WARN_ON_ONCE(rdp->nxttail[RCU_NEXT_TAIL] != | |
2541 | &rdp->nxtlist && | |
2542 | rdp->nxttail[RCU_NEXT_TAIL] != NULL); | |
2543 | init_nocb_callback_list(rdp); | |
2544 | } | |
35ce7f29 | 2545 | rcu_organize_nocb_kthreads(rsp); |
f4579fc5 | 2546 | } |
96d3fd0d PM |
2547 | } |
2548 | ||
3fbfbf7a PM |
2549 | /* Initialize per-rcu_data variables for no-CBs CPUs. */ |
2550 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) | |
2551 | { | |
2552 | rdp->nocb_tail = &rdp->nocb_head; | |
2553 | init_waitqueue_head(&rdp->nocb_wq); | |
fbce7497 | 2554 | rdp->nocb_follower_tail = &rdp->nocb_follower_head; |
3fbfbf7a PM |
2555 | } |
2556 | ||
35ce7f29 PM |
2557 | /* |
2558 | * If the specified CPU is a no-CBs CPU that does not already have its | |
2559 | * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are | |
2560 | * brought online out of order, this can require re-organizing the | |
2561 | * leader-follower relationships. | |
2562 | */ | |
2563 | static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu) | |
2564 | { | |
2565 | struct rcu_data *rdp; | |
2566 | struct rcu_data *rdp_last; | |
2567 | struct rcu_data *rdp_old_leader; | |
2568 | struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu); | |
2569 | struct task_struct *t; | |
2570 | ||
2571 | /* | |
2572 | * If this isn't a no-CBs CPU or if it already has an rcuo kthread, | |
2573 | * then nothing to do. | |
2574 | */ | |
2575 | if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread) | |
2576 | return; | |
2577 | ||
2578 | /* If we didn't spawn the leader first, reorganize! */ | |
2579 | rdp_old_leader = rdp_spawn->nocb_leader; | |
2580 | if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) { | |
2581 | rdp_last = NULL; | |
2582 | rdp = rdp_old_leader; | |
2583 | do { | |
2584 | rdp->nocb_leader = rdp_spawn; | |
2585 | if (rdp_last && rdp != rdp_spawn) | |
2586 | rdp_last->nocb_next_follower = rdp; | |
bbe5d7a9 PM |
2587 | if (rdp == rdp_spawn) { |
2588 | rdp = rdp->nocb_next_follower; | |
2589 | } else { | |
2590 | rdp_last = rdp; | |
2591 | rdp = rdp->nocb_next_follower; | |
2592 | rdp_last->nocb_next_follower = NULL; | |
2593 | } | |
35ce7f29 PM |
2594 | } while (rdp); |
2595 | rdp_spawn->nocb_next_follower = rdp_old_leader; | |
2596 | } | |
2597 | ||
2598 | /* Spawn the kthread for this CPU and RCU flavor. */ | |
2599 | t = kthread_run(rcu_nocb_kthread, rdp_spawn, | |
2600 | "rcuo%c/%d", rsp->abbr, cpu); | |
2601 | BUG_ON(IS_ERR(t)); | |
2602 | ACCESS_ONCE(rdp_spawn->nocb_kthread) = t; | |
2603 | } | |
2604 | ||
2605 | /* | |
2606 | * If the specified CPU is a no-CBs CPU that does not already have its | |
2607 | * rcuo kthreads, spawn them. | |
2608 | */ | |
2609 | static void rcu_spawn_all_nocb_kthreads(int cpu) | |
2610 | { | |
2611 | struct rcu_state *rsp; | |
2612 | ||
2613 | if (rcu_scheduler_fully_active) | |
2614 | for_each_rcu_flavor(rsp) | |
2615 | rcu_spawn_one_nocb_kthread(rsp, cpu); | |
2616 | } | |
2617 | ||
2618 | /* | |
2619 | * Once the scheduler is running, spawn rcuo kthreads for all online | |
2620 | * no-CBs CPUs. This assumes that the early_initcall()s happen before | |
2621 | * non-boot CPUs come online -- if this changes, we will need to add | |
2622 | * some mutual exclusion. | |
2623 | */ | |
2624 | static void __init rcu_spawn_nocb_kthreads(void) | |
2625 | { | |
2626 | int cpu; | |
2627 | ||
2628 | for_each_online_cpu(cpu) | |
2629 | rcu_spawn_all_nocb_kthreads(cpu); | |
2630 | } | |
2631 | ||
fbce7497 PM |
2632 | /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */ |
2633 | static int rcu_nocb_leader_stride = -1; | |
2634 | module_param(rcu_nocb_leader_stride, int, 0444); | |
2635 | ||
2636 | /* | |
35ce7f29 | 2637 | * Initialize leader-follower relationships for all no-CBs CPU. |
fbce7497 | 2638 | */ |
35ce7f29 | 2639 | static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp) |
3fbfbf7a PM |
2640 | { |
2641 | int cpu; | |
fbce7497 PM |
2642 | int ls = rcu_nocb_leader_stride; |
2643 | int nl = 0; /* Next leader. */ | |
3fbfbf7a | 2644 | struct rcu_data *rdp; |
fbce7497 PM |
2645 | struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */ |
2646 | struct rcu_data *rdp_prev = NULL; | |
3fbfbf7a | 2647 | |
22c2f669 | 2648 | if (!have_rcu_nocb_mask) |
3fbfbf7a | 2649 | return; |
fbce7497 PM |
2650 | if (ls == -1) { |
2651 | ls = int_sqrt(nr_cpu_ids); | |
2652 | rcu_nocb_leader_stride = ls; | |
2653 | } | |
2654 | ||
2655 | /* | |
2656 | * Each pass through this loop sets up one rcu_data structure and | |
2657 | * spawns one rcu_nocb_kthread(). | |
2658 | */ | |
3fbfbf7a PM |
2659 | for_each_cpu(cpu, rcu_nocb_mask) { |
2660 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
fbce7497 PM |
2661 | if (rdp->cpu >= nl) { |
2662 | /* New leader, set up for followers & next leader. */ | |
2663 | nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls; | |
2664 | rdp->nocb_leader = rdp; | |
2665 | rdp_leader = rdp; | |
2666 | } else { | |
2667 | /* Another follower, link to previous leader. */ | |
2668 | rdp->nocb_leader = rdp_leader; | |
2669 | rdp_prev->nocb_next_follower = rdp; | |
2670 | } | |
2671 | rdp_prev = rdp; | |
3fbfbf7a PM |
2672 | } |
2673 | } | |
2674 | ||
2675 | /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */ | |
34ed6246 | 2676 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2677 | { |
22c2f669 | 2678 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
34ed6246 | 2679 | return false; |
22c2f669 | 2680 | |
3fbfbf7a | 2681 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
34ed6246 | 2682 | return true; |
3fbfbf7a PM |
2683 | } |
2684 | ||
34ed6246 PM |
2685 | #else /* #ifdef CONFIG_RCU_NOCB_CPU */ |
2686 | ||
d7e29933 PM |
2687 | static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu) |
2688 | { | |
2689 | WARN_ON_ONCE(1); /* Should be dead code. */ | |
2690 | return false; | |
2691 | } | |
2692 | ||
0446be48 | 2693 | static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
3fbfbf7a | 2694 | { |
3fbfbf7a PM |
2695 | } |
2696 | ||
dae6e64d PM |
2697 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
2698 | { | |
2699 | } | |
2700 | ||
2701 | static void rcu_init_one_nocb(struct rcu_node *rnp) | |
2702 | { | |
2703 | } | |
3fbfbf7a | 2704 | |
3fbfbf7a | 2705 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
96d3fd0d | 2706 | bool lazy, unsigned long flags) |
3fbfbf7a | 2707 | { |
4afc7e26 | 2708 | return false; |
3fbfbf7a PM |
2709 | } |
2710 | ||
2711 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
2712 | struct rcu_data *rdp, |
2713 | unsigned long flags) | |
3fbfbf7a | 2714 | { |
f4aa84ba | 2715 | return false; |
3fbfbf7a PM |
2716 | } |
2717 | ||
3fbfbf7a PM |
2718 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
2719 | { | |
2720 | } | |
2721 | ||
9fdd3bc9 | 2722 | static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
96d3fd0d PM |
2723 | { |
2724 | return false; | |
2725 | } | |
2726 | ||
2727 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2728 | { | |
2729 | } | |
2730 | ||
35ce7f29 PM |
2731 | static void rcu_spawn_all_nocb_kthreads(int cpu) |
2732 | { | |
2733 | } | |
2734 | ||
2735 | static void __init rcu_spawn_nocb_kthreads(void) | |
3fbfbf7a PM |
2736 | { |
2737 | } | |
2738 | ||
34ed6246 | 2739 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2740 | { |
34ed6246 | 2741 | return false; |
3fbfbf7a PM |
2742 | } |
2743 | ||
2744 | #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */ | |
65d798f0 PM |
2745 | |
2746 | /* | |
2747 | * An adaptive-ticks CPU can potentially execute in kernel mode for an | |
2748 | * arbitrarily long period of time with the scheduling-clock tick turned | |
2749 | * off. RCU will be paying attention to this CPU because it is in the | |
2750 | * kernel, but the CPU cannot be guaranteed to be executing the RCU state | |
2751 | * machine because the scheduling-clock tick has been disabled. Therefore, | |
2752 | * if an adaptive-ticks CPU is failing to respond to the current grace | |
2753 | * period and has not be idle from an RCU perspective, kick it. | |
2754 | */ | |
4a81e832 | 2755 | static void __maybe_unused rcu_kick_nohz_cpu(int cpu) |
65d798f0 PM |
2756 | { |
2757 | #ifdef CONFIG_NO_HZ_FULL | |
2758 | if (tick_nohz_full_cpu(cpu)) | |
2759 | smp_send_reschedule(cpu); | |
2760 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ | |
2761 | } | |
2333210b PM |
2762 | |
2763 | ||
2764 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
2765 | ||
0edd1b17 | 2766 | static int full_sysidle_state; /* Current system-idle state. */ |
d4bd54fb PM |
2767 | #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */ |
2768 | #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */ | |
2769 | #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */ | |
2770 | #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */ | |
2771 | #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */ | |
2772 | ||
eb348b89 PM |
2773 | /* |
2774 | * Invoked to note exit from irq or task transition to idle. Note that | |
2775 | * usermode execution does -not- count as idle here! After all, we want | |
2776 | * to detect full-system idle states, not RCU quiescent states and grace | |
2777 | * periods. The caller must have disabled interrupts. | |
2778 | */ | |
28ced795 | 2779 | static void rcu_sysidle_enter(int irq) |
eb348b89 PM |
2780 | { |
2781 | unsigned long j; | |
28ced795 | 2782 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
eb348b89 | 2783 | |
663e1310 PM |
2784 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2785 | if (!tick_nohz_full_enabled()) | |
2786 | return; | |
2787 | ||
eb348b89 PM |
2788 | /* Adjust nesting, check for fully idle. */ |
2789 | if (irq) { | |
2790 | rdtp->dynticks_idle_nesting--; | |
2791 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2792 | if (rdtp->dynticks_idle_nesting != 0) | |
2793 | return; /* Still not fully idle. */ | |
2794 | } else { | |
2795 | if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) == | |
2796 | DYNTICK_TASK_NEST_VALUE) { | |
2797 | rdtp->dynticks_idle_nesting = 0; | |
2798 | } else { | |
2799 | rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE; | |
2800 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2801 | return; /* Still not fully idle. */ | |
2802 | } | |
2803 | } | |
2804 | ||
2805 | /* Record start of fully idle period. */ | |
2806 | j = jiffies; | |
2807 | ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j; | |
4e857c58 | 2808 | smp_mb__before_atomic(); |
eb348b89 | 2809 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2810 | smp_mb__after_atomic(); |
eb348b89 PM |
2811 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1); |
2812 | } | |
2813 | ||
0edd1b17 PM |
2814 | /* |
2815 | * Unconditionally force exit from full system-idle state. This is | |
2816 | * invoked when a normal CPU exits idle, but must be called separately | |
2817 | * for the timekeeping CPU (tick_do_timer_cpu). The reason for this | |
2818 | * is that the timekeeping CPU is permitted to take scheduling-clock | |
2819 | * interrupts while the system is in system-idle state, and of course | |
2820 | * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock | |
2821 | * interrupt from any other type of interrupt. | |
2822 | */ | |
2823 | void rcu_sysidle_force_exit(void) | |
2824 | { | |
2825 | int oldstate = ACCESS_ONCE(full_sysidle_state); | |
2826 | int newoldstate; | |
2827 | ||
2828 | /* | |
2829 | * Each pass through the following loop attempts to exit full | |
2830 | * system-idle state. If contention proves to be a problem, | |
2831 | * a trylock-based contention tree could be used here. | |
2832 | */ | |
2833 | while (oldstate > RCU_SYSIDLE_SHORT) { | |
2834 | newoldstate = cmpxchg(&full_sysidle_state, | |
2835 | oldstate, RCU_SYSIDLE_NOT); | |
2836 | if (oldstate == newoldstate && | |
2837 | oldstate == RCU_SYSIDLE_FULL_NOTED) { | |
2838 | rcu_kick_nohz_cpu(tick_do_timer_cpu); | |
2839 | return; /* We cleared it, done! */ | |
2840 | } | |
2841 | oldstate = newoldstate; | |
2842 | } | |
2843 | smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */ | |
2844 | } | |
2845 | ||
eb348b89 PM |
2846 | /* |
2847 | * Invoked to note entry to irq or task transition from idle. Note that | |
2848 | * usermode execution does -not- count as idle here! The caller must | |
2849 | * have disabled interrupts. | |
2850 | */ | |
28ced795 | 2851 | static void rcu_sysidle_exit(int irq) |
eb348b89 | 2852 | { |
28ced795 CL |
2853 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
2854 | ||
663e1310 PM |
2855 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2856 | if (!tick_nohz_full_enabled()) | |
2857 | return; | |
2858 | ||
eb348b89 PM |
2859 | /* Adjust nesting, check for already non-idle. */ |
2860 | if (irq) { | |
2861 | rdtp->dynticks_idle_nesting++; | |
2862 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2863 | if (rdtp->dynticks_idle_nesting != 1) | |
2864 | return; /* Already non-idle. */ | |
2865 | } else { | |
2866 | /* | |
2867 | * Allow for irq misnesting. Yes, it really is possible | |
2868 | * to enter an irq handler then never leave it, and maybe | |
2869 | * also vice versa. Handle both possibilities. | |
2870 | */ | |
2871 | if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) { | |
2872 | rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE; | |
2873 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2874 | return; /* Already non-idle. */ | |
2875 | } else { | |
2876 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE; | |
2877 | } | |
2878 | } | |
2879 | ||
2880 | /* Record end of idle period. */ | |
4e857c58 | 2881 | smp_mb__before_atomic(); |
eb348b89 | 2882 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2883 | smp_mb__after_atomic(); |
eb348b89 | 2884 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1)); |
0edd1b17 PM |
2885 | |
2886 | /* | |
2887 | * If we are the timekeeping CPU, we are permitted to be non-idle | |
2888 | * during a system-idle state. This must be the case, because | |
2889 | * the timekeeping CPU has to take scheduling-clock interrupts | |
2890 | * during the time that the system is transitioning to full | |
2891 | * system-idle state. This means that the timekeeping CPU must | |
2892 | * invoke rcu_sysidle_force_exit() directly if it does anything | |
2893 | * more than take a scheduling-clock interrupt. | |
2894 | */ | |
2895 | if (smp_processor_id() == tick_do_timer_cpu) | |
2896 | return; | |
2897 | ||
2898 | /* Update system-idle state: We are clearly no longer fully idle! */ | |
2899 | rcu_sysidle_force_exit(); | |
2900 | } | |
2901 | ||
2902 | /* | |
2903 | * Check to see if the current CPU is idle. Note that usermode execution | |
2904 | * does not count as idle. The caller must have disabled interrupts. | |
2905 | */ | |
2906 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, | |
2907 | unsigned long *maxj) | |
2908 | { | |
2909 | int cur; | |
2910 | unsigned long j; | |
2911 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
2912 | ||
663e1310 PM |
2913 | /* If there are no nohz_full= CPUs, don't check system-wide idleness. */ |
2914 | if (!tick_nohz_full_enabled()) | |
2915 | return; | |
2916 | ||
0edd1b17 PM |
2917 | /* |
2918 | * If some other CPU has already reported non-idle, if this is | |
2919 | * not the flavor of RCU that tracks sysidle state, or if this | |
2920 | * is an offline or the timekeeping CPU, nothing to do. | |
2921 | */ | |
417e8d26 | 2922 | if (!*isidle || rdp->rsp != rcu_state_p || |
0edd1b17 PM |
2923 | cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu) |
2924 | return; | |
eb75767b PM |
2925 | if (rcu_gp_in_progress(rdp->rsp)) |
2926 | WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu); | |
0edd1b17 PM |
2927 | |
2928 | /* Pick up current idle and NMI-nesting counter and check. */ | |
2929 | cur = atomic_read(&rdtp->dynticks_idle); | |
2930 | if (cur & 0x1) { | |
2931 | *isidle = false; /* We are not idle! */ | |
2932 | return; | |
2933 | } | |
2934 | smp_mb(); /* Read counters before timestamps. */ | |
2935 | ||
2936 | /* Pick up timestamps. */ | |
2937 | j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies); | |
2938 | /* If this CPU entered idle more recently, update maxj timestamp. */ | |
2939 | if (ULONG_CMP_LT(*maxj, j)) | |
2940 | *maxj = j; | |
2941 | } | |
2942 | ||
2943 | /* | |
2944 | * Is this the flavor of RCU that is handling full-system idle? | |
2945 | */ | |
2946 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2947 | { | |
417e8d26 | 2948 | return rsp == rcu_state_p; |
0edd1b17 PM |
2949 | } |
2950 | ||
2951 | /* | |
2952 | * Return a delay in jiffies based on the number of CPUs, rcu_node | |
2953 | * leaf fanout, and jiffies tick rate. The idea is to allow larger | |
2954 | * systems more time to transition to full-idle state in order to | |
2955 | * avoid the cache thrashing that otherwise occur on the state variable. | |
2956 | * Really small systems (less than a couple of tens of CPUs) should | |
2957 | * instead use a single global atomically incremented counter, and later | |
2958 | * versions of this will automatically reconfigure themselves accordingly. | |
2959 | */ | |
2960 | static unsigned long rcu_sysidle_delay(void) | |
2961 | { | |
2962 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2963 | return 0; | |
2964 | return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000); | |
2965 | } | |
2966 | ||
2967 | /* | |
2968 | * Advance the full-system-idle state. This is invoked when all of | |
2969 | * the non-timekeeping CPUs are idle. | |
2970 | */ | |
2971 | static void rcu_sysidle(unsigned long j) | |
2972 | { | |
2973 | /* Check the current state. */ | |
2974 | switch (ACCESS_ONCE(full_sysidle_state)) { | |
2975 | case RCU_SYSIDLE_NOT: | |
2976 | ||
2977 | /* First time all are idle, so note a short idle period. */ | |
2978 | ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT; | |
2979 | break; | |
2980 | ||
2981 | case RCU_SYSIDLE_SHORT: | |
2982 | ||
2983 | /* | |
2984 | * Idle for a bit, time to advance to next state? | |
2985 | * cmpxchg failure means race with non-idle, let them win. | |
2986 | */ | |
2987 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2988 | (void)cmpxchg(&full_sysidle_state, | |
2989 | RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG); | |
2990 | break; | |
2991 | ||
2992 | case RCU_SYSIDLE_LONG: | |
2993 | ||
2994 | /* | |
2995 | * Do an additional check pass before advancing to full. | |
2996 | * cmpxchg failure means race with non-idle, let them win. | |
2997 | */ | |
2998 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2999 | (void)cmpxchg(&full_sysidle_state, | |
3000 | RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL); | |
3001 | break; | |
3002 | ||
3003 | default: | |
3004 | break; | |
3005 | } | |
3006 | } | |
3007 | ||
3008 | /* | |
3009 | * Found a non-idle non-timekeeping CPU, so kick the system-idle state | |
3010 | * back to the beginning. | |
3011 | */ | |
3012 | static void rcu_sysidle_cancel(void) | |
3013 | { | |
3014 | smp_mb(); | |
becb41bf PM |
3015 | if (full_sysidle_state > RCU_SYSIDLE_SHORT) |
3016 | ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT; | |
0edd1b17 PM |
3017 | } |
3018 | ||
3019 | /* | |
3020 | * Update the sysidle state based on the results of a force-quiescent-state | |
3021 | * scan of the CPUs' dyntick-idle state. | |
3022 | */ | |
3023 | static void rcu_sysidle_report(struct rcu_state *rsp, int isidle, | |
3024 | unsigned long maxj, bool gpkt) | |
3025 | { | |
417e8d26 | 3026 | if (rsp != rcu_state_p) |
0edd1b17 PM |
3027 | return; /* Wrong flavor, ignore. */ |
3028 | if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
3029 | return; /* Running state machine from timekeeping CPU. */ | |
3030 | if (isidle) | |
3031 | rcu_sysidle(maxj); /* More idle! */ | |
3032 | else | |
3033 | rcu_sysidle_cancel(); /* Idle is over. */ | |
3034 | } | |
3035 | ||
3036 | /* | |
3037 | * Wrapper for rcu_sysidle_report() when called from the grace-period | |
3038 | * kthread's context. | |
3039 | */ | |
3040 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
3041 | unsigned long maxj) | |
3042 | { | |
663e1310 PM |
3043 | /* If there are no nohz_full= CPUs, no need to track this. */ |
3044 | if (!tick_nohz_full_enabled()) | |
3045 | return; | |
3046 | ||
0edd1b17 PM |
3047 | rcu_sysidle_report(rsp, isidle, maxj, true); |
3048 | } | |
3049 | ||
3050 | /* Callback and function for forcing an RCU grace period. */ | |
3051 | struct rcu_sysidle_head { | |
3052 | struct rcu_head rh; | |
3053 | int inuse; | |
3054 | }; | |
3055 | ||
3056 | static void rcu_sysidle_cb(struct rcu_head *rhp) | |
3057 | { | |
3058 | struct rcu_sysidle_head *rshp; | |
3059 | ||
3060 | /* | |
3061 | * The following memory barrier is needed to replace the | |
3062 | * memory barriers that would normally be in the memory | |
3063 | * allocator. | |
3064 | */ | |
3065 | smp_mb(); /* grace period precedes setting inuse. */ | |
3066 | ||
3067 | rshp = container_of(rhp, struct rcu_sysidle_head, rh); | |
3068 | ACCESS_ONCE(rshp->inuse) = 0; | |
3069 | } | |
3070 | ||
3071 | /* | |
3072 | * Check to see if the system is fully idle, other than the timekeeping CPU. | |
663e1310 PM |
3073 | * The caller must have disabled interrupts. This is not intended to be |
3074 | * called unless tick_nohz_full_enabled(). | |
0edd1b17 PM |
3075 | */ |
3076 | bool rcu_sys_is_idle(void) | |
3077 | { | |
3078 | static struct rcu_sysidle_head rsh; | |
3079 | int rss = ACCESS_ONCE(full_sysidle_state); | |
3080 | ||
3081 | if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu)) | |
3082 | return false; | |
3083 | ||
3084 | /* Handle small-system case by doing a full scan of CPUs. */ | |
3085 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) { | |
3086 | int oldrss = rss - 1; | |
3087 | ||
3088 | /* | |
3089 | * One pass to advance to each state up to _FULL. | |
3090 | * Give up if any pass fails to advance the state. | |
3091 | */ | |
3092 | while (rss < RCU_SYSIDLE_FULL && oldrss < rss) { | |
3093 | int cpu; | |
3094 | bool isidle = true; | |
3095 | unsigned long maxj = jiffies - ULONG_MAX / 4; | |
3096 | struct rcu_data *rdp; | |
3097 | ||
3098 | /* Scan all the CPUs looking for nonidle CPUs. */ | |
3099 | for_each_possible_cpu(cpu) { | |
417e8d26 | 3100 | rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
0edd1b17 PM |
3101 | rcu_sysidle_check_cpu(rdp, &isidle, &maxj); |
3102 | if (!isidle) | |
3103 | break; | |
3104 | } | |
417e8d26 | 3105 | rcu_sysidle_report(rcu_state_p, isidle, maxj, false); |
0edd1b17 PM |
3106 | oldrss = rss; |
3107 | rss = ACCESS_ONCE(full_sysidle_state); | |
3108 | } | |
3109 | } | |
3110 | ||
3111 | /* If this is the first observation of an idle period, record it. */ | |
3112 | if (rss == RCU_SYSIDLE_FULL) { | |
3113 | rss = cmpxchg(&full_sysidle_state, | |
3114 | RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED); | |
3115 | return rss == RCU_SYSIDLE_FULL; | |
3116 | } | |
3117 | ||
3118 | smp_mb(); /* ensure rss load happens before later caller actions. */ | |
3119 | ||
3120 | /* If already fully idle, tell the caller (in case of races). */ | |
3121 | if (rss == RCU_SYSIDLE_FULL_NOTED) | |
3122 | return true; | |
3123 | ||
3124 | /* | |
3125 | * If we aren't there yet, and a grace period is not in flight, | |
3126 | * initiate a grace period. Either way, tell the caller that | |
3127 | * we are not there yet. We use an xchg() rather than an assignment | |
3128 | * to make up for the memory barriers that would otherwise be | |
3129 | * provided by the memory allocator. | |
3130 | */ | |
3131 | if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL && | |
417e8d26 | 3132 | !rcu_gp_in_progress(rcu_state_p) && |
0edd1b17 PM |
3133 | !rsh.inuse && xchg(&rsh.inuse, 1) == 0) |
3134 | call_rcu(&rsh.rh, rcu_sysidle_cb); | |
3135 | return false; | |
eb348b89 PM |
3136 | } |
3137 | ||
2333210b PM |
3138 | /* |
3139 | * Initialize dynticks sysidle state for CPUs coming online. | |
3140 | */ | |
3141 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) | |
3142 | { | |
3143 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE; | |
3144 | } | |
3145 | ||
3146 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
3147 | ||
28ced795 | 3148 | static void rcu_sysidle_enter(int irq) |
eb348b89 PM |
3149 | { |
3150 | } | |
3151 | ||
28ced795 | 3152 | static void rcu_sysidle_exit(int irq) |
eb348b89 PM |
3153 | { |
3154 | } | |
3155 | ||
0edd1b17 PM |
3156 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, |
3157 | unsigned long *maxj) | |
3158 | { | |
3159 | } | |
3160 | ||
3161 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
3162 | { | |
3163 | return false; | |
3164 | } | |
3165 | ||
3166 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
3167 | unsigned long maxj) | |
3168 | { | |
3169 | } | |
3170 | ||
2333210b PM |
3171 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) |
3172 | { | |
3173 | } | |
3174 | ||
3175 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
a096932f PM |
3176 | |
3177 | /* | |
3178 | * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the | |
3179 | * grace-period kthread will do force_quiescent_state() processing? | |
3180 | * The idea is to avoid waking up RCU core processing on such a | |
3181 | * CPU unless the grace period has extended for too long. | |
3182 | * | |
3183 | * This code relies on the fact that all NO_HZ_FULL CPUs are also | |
52e2bb95 | 3184 | * CONFIG_RCU_NOCB_CPU CPUs. |
a096932f PM |
3185 | */ |
3186 | static bool rcu_nohz_full_cpu(struct rcu_state *rsp) | |
3187 | { | |
3188 | #ifdef CONFIG_NO_HZ_FULL | |
3189 | if (tick_nohz_full_cpu(smp_processor_id()) && | |
3190 | (!rcu_gp_in_progress(rsp) || | |
3191 | ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ))) | |
3192 | return 1; | |
3193 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ | |
3194 | return 0; | |
3195 | } | |
5057f55e PM |
3196 | |
3197 | /* | |
3198 | * Bind the grace-period kthread for the sysidle flavor of RCU to the | |
3199 | * timekeeping CPU. | |
3200 | */ | |
3201 | static void rcu_bind_gp_kthread(void) | |
3202 | { | |
c0f489d2 | 3203 | int __maybe_unused cpu; |
5057f55e | 3204 | |
c0f489d2 | 3205 | if (!tick_nohz_full_enabled()) |
5057f55e | 3206 | return; |
c0f489d2 PM |
3207 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
3208 | cpu = tick_do_timer_cpu; | |
3209 | if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu) | |
5057f55e | 3210 | set_cpus_allowed_ptr(current, cpumask_of(cpu)); |
c0f489d2 PM |
3211 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
3212 | if (!is_housekeeping_cpu(raw_smp_processor_id())) | |
3213 | housekeeping_affine(current); | |
3214 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
5057f55e | 3215 | } |
176f8f7a PM |
3216 | |
3217 | /* Record the current task on dyntick-idle entry. */ | |
3218 | static void rcu_dynticks_task_enter(void) | |
3219 | { | |
3220 | #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) | |
3221 | ACCESS_ONCE(current->rcu_tasks_idle_cpu) = smp_processor_id(); | |
3222 | #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ | |
3223 | } | |
3224 | ||
3225 | /* Record no current task on dyntick-idle exit. */ | |
3226 | static void rcu_dynticks_task_exit(void) | |
3227 | { | |
3228 | #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) | |
3229 | ACCESS_ONCE(current->rcu_tasks_idle_cpu) = -1; | |
3230 | #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ | |
3231 | } |