]>
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 MG |
33 | #define RCU_KTHREAD_PRIO 1 |
34 | ||
35 | #ifdef CONFIG_RCU_BOOST | |
abaa93d9 | 36 | #include "../locking/rtmutex_common.h" |
5b61b0ba MG |
37 | #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO |
38 | #else | |
39 | #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO | |
40 | #endif | |
41 | ||
3fbfbf7a PM |
42 | #ifdef CONFIG_RCU_NOCB_CPU |
43 | static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */ | |
44 | static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */ | |
1b0048a4 | 45 | static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */ |
3fbfbf7a PM |
46 | static char __initdata nocb_buf[NR_CPUS * 5]; |
47 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ | |
48 | ||
26845c28 PM |
49 | /* |
50 | * Check the RCU kernel configuration parameters and print informative | |
51 | * messages about anything out of the ordinary. If you like #ifdef, you | |
52 | * will love this function. | |
53 | */ | |
54 | static void __init rcu_bootup_announce_oddness(void) | |
55 | { | |
56 | #ifdef CONFIG_RCU_TRACE | |
efc151c3 | 57 | pr_info("\tRCU debugfs-based tracing is enabled.\n"); |
26845c28 PM |
58 | #endif |
59 | #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32) | |
efc151c3 | 60 | pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n", |
26845c28 PM |
61 | CONFIG_RCU_FANOUT); |
62 | #endif | |
63 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
efc151c3 | 64 | pr_info("\tHierarchical RCU autobalancing is disabled.\n"); |
26845c28 PM |
65 | #endif |
66 | #ifdef CONFIG_RCU_FAST_NO_HZ | |
efc151c3 | 67 | pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n"); |
26845c28 PM |
68 | #endif |
69 | #ifdef CONFIG_PROVE_RCU | |
efc151c3 | 70 | pr_info("\tRCU lockdep checking is enabled.\n"); |
26845c28 PM |
71 | #endif |
72 | #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE | |
efc151c3 | 73 | pr_info("\tRCU torture testing starts during boot.\n"); |
26845c28 | 74 | #endif |
81a294c4 | 75 | #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE) |
efc151c3 | 76 | pr_info("\tDump stacks of tasks blocking RCU-preempt GP.\n"); |
a858af28 PM |
77 | #endif |
78 | #if defined(CONFIG_RCU_CPU_STALL_INFO) | |
efc151c3 | 79 | pr_info("\tAdditional per-CPU info printed with stalls.\n"); |
26845c28 PM |
80 | #endif |
81 | #if NUM_RCU_LVL_4 != 0 | |
efc151c3 | 82 | pr_info("\tFour-level hierarchy is enabled.\n"); |
26845c28 | 83 | #endif |
f885b7f2 | 84 | if (rcu_fanout_leaf != CONFIG_RCU_FANOUT_LEAF) |
9a5739d7 | 85 | pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf); |
cca6f393 | 86 | if (nr_cpu_ids != NR_CPUS) |
efc151c3 | 87 | pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids); |
3fbfbf7a | 88 | #ifdef CONFIG_RCU_NOCB_CPU |
911af505 PM |
89 | #ifndef CONFIG_RCU_NOCB_CPU_NONE |
90 | if (!have_rcu_nocb_mask) { | |
615ee544 | 91 | zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL); |
911af505 PM |
92 | have_rcu_nocb_mask = true; |
93 | } | |
94 | #ifdef CONFIG_RCU_NOCB_CPU_ZERO | |
9a5739d7 | 95 | pr_info("\tOffload RCU callbacks from CPU 0\n"); |
911af505 PM |
96 | cpumask_set_cpu(0, rcu_nocb_mask); |
97 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */ | |
98 | #ifdef CONFIG_RCU_NOCB_CPU_ALL | |
9a5739d7 | 99 | pr_info("\tOffload RCU callbacks from all CPUs\n"); |
5d5a0800 | 100 | cpumask_copy(rcu_nocb_mask, cpu_possible_mask); |
911af505 PM |
101 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */ |
102 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */ | |
3fbfbf7a | 103 | if (have_rcu_nocb_mask) { |
5d5a0800 KT |
104 | if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) { |
105 | pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n"); | |
106 | cpumask_and(rcu_nocb_mask, cpu_possible_mask, | |
107 | rcu_nocb_mask); | |
108 | } | |
3fbfbf7a | 109 | cpulist_scnprintf(nocb_buf, sizeof(nocb_buf), rcu_nocb_mask); |
9a5739d7 | 110 | pr_info("\tOffload RCU callbacks from CPUs: %s.\n", nocb_buf); |
3fbfbf7a | 111 | if (rcu_nocb_poll) |
9a5739d7 | 112 | pr_info("\tPoll for callbacks from no-CBs CPUs.\n"); |
3fbfbf7a PM |
113 | } |
114 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ | |
26845c28 PM |
115 | } |
116 | ||
f41d911f PM |
117 | #ifdef CONFIG_TREE_PREEMPT_RCU |
118 | ||
a41bfeb2 | 119 | RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu); |
e534165b | 120 | static struct rcu_state *rcu_state_p = &rcu_preempt_state; |
f41d911f | 121 | |
d9a3da06 PM |
122 | static int rcu_preempted_readers_exp(struct rcu_node *rnp); |
123 | ||
f41d911f PM |
124 | /* |
125 | * Tell them what RCU they are running. | |
126 | */ | |
0e0fc1c2 | 127 | static void __init rcu_bootup_announce(void) |
f41d911f | 128 | { |
efc151c3 | 129 | pr_info("Preemptible hierarchical RCU implementation.\n"); |
26845c28 | 130 | rcu_bootup_announce_oddness(); |
f41d911f PM |
131 | } |
132 | ||
133 | /* | |
134 | * Return the number of RCU-preempt batches processed thus far | |
135 | * for debug and statistics. | |
136 | */ | |
137 | long rcu_batches_completed_preempt(void) | |
138 | { | |
139 | return rcu_preempt_state.completed; | |
140 | } | |
141 | EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt); | |
142 | ||
143 | /* | |
144 | * Return the number of RCU batches processed thus far for debug & stats. | |
145 | */ | |
146 | long rcu_batches_completed(void) | |
147 | { | |
148 | return rcu_batches_completed_preempt(); | |
149 | } | |
150 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
151 | ||
152 | /* | |
6cc68793 | 153 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
f41d911f PM |
154 | * that this just means that the task currently running on the CPU is |
155 | * not in a quiescent state. There might be any number of tasks blocked | |
156 | * while in an RCU read-side critical section. | |
25502a6c PM |
157 | * |
158 | * Unlike the other rcu_*_qs() functions, callers to this function | |
159 | * must disable irqs in order to protect the assignment to | |
160 | * ->rcu_read_unlock_special. | |
f41d911f | 161 | */ |
c3422bea | 162 | static void rcu_preempt_qs(int cpu) |
f41d911f PM |
163 | { |
164 | struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu); | |
25502a6c | 165 | |
e4cc1f22 | 166 | if (rdp->passed_quiesce == 0) |
f7f7bac9 | 167 | trace_rcu_grace_period(TPS("rcu_preempt"), rdp->gpnum, TPS("cpuqs")); |
e4cc1f22 | 168 | rdp->passed_quiesce = 1; |
25502a6c | 169 | current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS; |
f41d911f PM |
170 | } |
171 | ||
172 | /* | |
c3422bea PM |
173 | * We have entered the scheduler, and the current task might soon be |
174 | * context-switched away from. If this task is in an RCU read-side | |
175 | * critical section, we will no longer be able to rely on the CPU to | |
12f5f524 PM |
176 | * record that fact, so we enqueue the task on the blkd_tasks list. |
177 | * The task will dequeue itself when it exits the outermost enclosing | |
178 | * RCU read-side critical section. Therefore, the current grace period | |
179 | * cannot be permitted to complete until the blkd_tasks list entries | |
180 | * predating the current grace period drain, in other words, until | |
181 | * rnp->gp_tasks becomes NULL. | |
c3422bea PM |
182 | * |
183 | * Caller must disable preemption. | |
f41d911f | 184 | */ |
cba6d0d6 | 185 | static void rcu_preempt_note_context_switch(int cpu) |
f41d911f PM |
186 | { |
187 | struct task_struct *t = current; | |
c3422bea | 188 | unsigned long flags; |
f41d911f PM |
189 | struct rcu_data *rdp; |
190 | struct rcu_node *rnp; | |
191 | ||
10f39bb1 | 192 | if (t->rcu_read_lock_nesting > 0 && |
f41d911f PM |
193 | (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) { |
194 | ||
195 | /* Possibly blocking in an RCU read-side critical section. */ | |
cba6d0d6 | 196 | rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu); |
f41d911f | 197 | rnp = rdp->mynode; |
1304afb2 | 198 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 199 | smp_mb__after_unlock_lock(); |
f41d911f | 200 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED; |
86848966 | 201 | t->rcu_blocked_node = rnp; |
f41d911f PM |
202 | |
203 | /* | |
204 | * If this CPU has already checked in, then this task | |
205 | * will hold up the next grace period rather than the | |
206 | * current grace period. Queue the task accordingly. | |
207 | * If the task is queued for the current grace period | |
208 | * (i.e., this CPU has not yet passed through a quiescent | |
209 | * state for the current grace period), then as long | |
210 | * as that task remains queued, the current grace period | |
12f5f524 PM |
211 | * cannot end. Note that there is some uncertainty as |
212 | * to exactly when the current grace period started. | |
213 | * We take a conservative approach, which can result | |
214 | * in unnecessarily waiting on tasks that started very | |
215 | * slightly after the current grace period began. C'est | |
216 | * la vie!!! | |
b0e165c0 PM |
217 | * |
218 | * But first, note that the current CPU must still be | |
219 | * on line! | |
f41d911f | 220 | */ |
b0e165c0 | 221 | WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0); |
e7d8842e | 222 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
12f5f524 PM |
223 | if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) { |
224 | list_add(&t->rcu_node_entry, rnp->gp_tasks->prev); | |
225 | rnp->gp_tasks = &t->rcu_node_entry; | |
27f4d280 PM |
226 | #ifdef CONFIG_RCU_BOOST |
227 | if (rnp->boost_tasks != NULL) | |
228 | rnp->boost_tasks = rnp->gp_tasks; | |
229 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
12f5f524 PM |
230 | } else { |
231 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); | |
232 | if (rnp->qsmask & rdp->grpmask) | |
233 | rnp->gp_tasks = &t->rcu_node_entry; | |
234 | } | |
d4c08f2a PM |
235 | trace_rcu_preempt_task(rdp->rsp->name, |
236 | t->pid, | |
237 | (rnp->qsmask & rdp->grpmask) | |
238 | ? rnp->gpnum | |
239 | : rnp->gpnum + 1); | |
1304afb2 | 240 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
10f39bb1 PM |
241 | } else if (t->rcu_read_lock_nesting < 0 && |
242 | t->rcu_read_unlock_special) { | |
243 | ||
244 | /* | |
245 | * Complete exit from RCU read-side critical section on | |
246 | * behalf of preempted instance of __rcu_read_unlock(). | |
247 | */ | |
248 | rcu_read_unlock_special(t); | |
f41d911f PM |
249 | } |
250 | ||
251 | /* | |
252 | * Either we were not in an RCU read-side critical section to | |
253 | * begin with, or we have now recorded that critical section | |
254 | * globally. Either way, we can now note a quiescent state | |
255 | * for this CPU. Again, if we were in an RCU read-side critical | |
256 | * section, and if that critical section was blocking the current | |
257 | * grace period, then the fact that the task has been enqueued | |
258 | * means that we continue to block the current grace period. | |
259 | */ | |
e7d8842e | 260 | local_irq_save(flags); |
cba6d0d6 | 261 | rcu_preempt_qs(cpu); |
e7d8842e | 262 | local_irq_restore(flags); |
f41d911f PM |
263 | } |
264 | ||
fc2219d4 PM |
265 | /* |
266 | * Check for preempted RCU readers blocking the current grace period | |
267 | * for the specified rcu_node structure. If the caller needs a reliable | |
268 | * answer, it must hold the rcu_node's ->lock. | |
269 | */ | |
27f4d280 | 270 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 | 271 | { |
12f5f524 | 272 | return rnp->gp_tasks != NULL; |
fc2219d4 PM |
273 | } |
274 | ||
b668c9cf PM |
275 | /* |
276 | * Record a quiescent state for all tasks that were previously queued | |
277 | * on the specified rcu_node structure and that were blocking the current | |
278 | * RCU grace period. The caller must hold the specified rnp->lock with | |
279 | * irqs disabled, and this lock is released upon return, but irqs remain | |
280 | * disabled. | |
281 | */ | |
d3f6bad3 | 282 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
b668c9cf PM |
283 | __releases(rnp->lock) |
284 | { | |
285 | unsigned long mask; | |
286 | struct rcu_node *rnp_p; | |
287 | ||
27f4d280 | 288 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
1304afb2 | 289 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf PM |
290 | return; /* Still need more quiescent states! */ |
291 | } | |
292 | ||
293 | rnp_p = rnp->parent; | |
294 | if (rnp_p == NULL) { | |
295 | /* | |
296 | * Either there is only one rcu_node in the tree, | |
297 | * or tasks were kicked up to root rcu_node due to | |
298 | * CPUs going offline. | |
299 | */ | |
d3f6bad3 | 300 | rcu_report_qs_rsp(&rcu_preempt_state, flags); |
b668c9cf PM |
301 | return; |
302 | } | |
303 | ||
304 | /* Report up the rest of the hierarchy. */ | |
305 | mask = rnp->grpmask; | |
1304afb2 PM |
306 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
307 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
6303b9c8 | 308 | smp_mb__after_unlock_lock(); |
d3f6bad3 | 309 | rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags); |
b668c9cf PM |
310 | } |
311 | ||
12f5f524 PM |
312 | /* |
313 | * Advance a ->blkd_tasks-list pointer to the next entry, instead | |
314 | * returning NULL if at the end of the list. | |
315 | */ | |
316 | static struct list_head *rcu_next_node_entry(struct task_struct *t, | |
317 | struct rcu_node *rnp) | |
318 | { | |
319 | struct list_head *np; | |
320 | ||
321 | np = t->rcu_node_entry.next; | |
322 | if (np == &rnp->blkd_tasks) | |
323 | np = NULL; | |
324 | return np; | |
325 | } | |
326 | ||
b668c9cf PM |
327 | /* |
328 | * Handle special cases during rcu_read_unlock(), such as needing to | |
329 | * notify RCU core processing or task having blocked during the RCU | |
330 | * read-side critical section. | |
331 | */ | |
2a3fa843 | 332 | void rcu_read_unlock_special(struct task_struct *t) |
f41d911f PM |
333 | { |
334 | int empty; | |
d9a3da06 | 335 | int empty_exp; |
389abd48 | 336 | int empty_exp_now; |
f41d911f | 337 | unsigned long flags; |
12f5f524 | 338 | struct list_head *np; |
82e78d80 | 339 | #ifdef CONFIG_RCU_BOOST |
abaa93d9 | 340 | bool drop_boost_mutex = false; |
82e78d80 | 341 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
f41d911f PM |
342 | struct rcu_node *rnp; |
343 | int special; | |
344 | ||
345 | /* NMI handlers cannot block and cannot safely manipulate state. */ | |
346 | if (in_nmi()) | |
347 | return; | |
348 | ||
349 | local_irq_save(flags); | |
350 | ||
351 | /* | |
352 | * If RCU core is waiting for this CPU to exit critical section, | |
353 | * let it know that we have done so. | |
354 | */ | |
355 | special = t->rcu_read_unlock_special; | |
356 | if (special & RCU_READ_UNLOCK_NEED_QS) { | |
c3422bea | 357 | rcu_preempt_qs(smp_processor_id()); |
79a62f95 LJ |
358 | if (!t->rcu_read_unlock_special) { |
359 | local_irq_restore(flags); | |
360 | return; | |
361 | } | |
f41d911f PM |
362 | } |
363 | ||
79a62f95 LJ |
364 | /* Hardware IRQ handlers cannot block, complain if they get here. */ |
365 | if (WARN_ON_ONCE(in_irq() || in_serving_softirq())) { | |
f41d911f PM |
366 | local_irq_restore(flags); |
367 | return; | |
368 | } | |
369 | ||
370 | /* Clean up if blocked during RCU read-side critical section. */ | |
371 | if (special & RCU_READ_UNLOCK_BLOCKED) { | |
372 | t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED; | |
373 | ||
dd5d19ba PM |
374 | /* |
375 | * Remove this task from the list it blocked on. The | |
376 | * task can migrate while we acquire the lock, but at | |
377 | * most one time. So at most two passes through loop. | |
378 | */ | |
379 | for (;;) { | |
86848966 | 380 | rnp = t->rcu_blocked_node; |
1304afb2 | 381 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
6303b9c8 | 382 | smp_mb__after_unlock_lock(); |
86848966 | 383 | if (rnp == t->rcu_blocked_node) |
dd5d19ba | 384 | break; |
1304afb2 | 385 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
dd5d19ba | 386 | } |
27f4d280 | 387 | empty = !rcu_preempt_blocked_readers_cgp(rnp); |
d9a3da06 PM |
388 | empty_exp = !rcu_preempted_readers_exp(rnp); |
389 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ | |
12f5f524 | 390 | np = rcu_next_node_entry(t, rnp); |
f41d911f | 391 | list_del_init(&t->rcu_node_entry); |
82e78d80 | 392 | t->rcu_blocked_node = NULL; |
f7f7bac9 | 393 | trace_rcu_unlock_preempted_task(TPS("rcu_preempt"), |
d4c08f2a | 394 | rnp->gpnum, t->pid); |
12f5f524 PM |
395 | if (&t->rcu_node_entry == rnp->gp_tasks) |
396 | rnp->gp_tasks = np; | |
397 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
398 | rnp->exp_tasks = np; | |
27f4d280 PM |
399 | #ifdef CONFIG_RCU_BOOST |
400 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
401 | rnp->boost_tasks = np; | |
abaa93d9 PM |
402 | /* Snapshot ->boost_mtx ownership with rcu_node lock held. */ |
403 | drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t; | |
27f4d280 | 404 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
f41d911f PM |
405 | |
406 | /* | |
407 | * If this was the last task on the current list, and if | |
408 | * we aren't waiting on any CPUs, report the quiescent state. | |
389abd48 PM |
409 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock, |
410 | * so we must take a snapshot of the expedited state. | |
f41d911f | 411 | */ |
389abd48 | 412 | empty_exp_now = !rcu_preempted_readers_exp(rnp); |
d4c08f2a | 413 | if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) { |
f7f7bac9 | 414 | trace_rcu_quiescent_state_report(TPS("preempt_rcu"), |
d4c08f2a PM |
415 | rnp->gpnum, |
416 | 0, rnp->qsmask, | |
417 | rnp->level, | |
418 | rnp->grplo, | |
419 | rnp->grphi, | |
420 | !!rnp->gp_tasks); | |
d3f6bad3 | 421 | rcu_report_unblock_qs_rnp(rnp, flags); |
c701d5d9 | 422 | } else { |
d4c08f2a | 423 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c701d5d9 | 424 | } |
d9a3da06 | 425 | |
27f4d280 PM |
426 | #ifdef CONFIG_RCU_BOOST |
427 | /* Unboost if we were boosted. */ | |
abaa93d9 PM |
428 | if (drop_boost_mutex) { |
429 | rt_mutex_unlock(&rnp->boost_mtx); | |
dfeb9765 PM |
430 | complete(&rnp->boost_completion); |
431 | } | |
27f4d280 PM |
432 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
433 | ||
d9a3da06 PM |
434 | /* |
435 | * If this was the last task on the expedited lists, | |
436 | * then we need to report up the rcu_node hierarchy. | |
437 | */ | |
389abd48 | 438 | if (!empty_exp && empty_exp_now) |
b40d293e | 439 | rcu_report_exp_rnp(&rcu_preempt_state, rnp, true); |
b668c9cf PM |
440 | } else { |
441 | local_irq_restore(flags); | |
f41d911f | 442 | } |
f41d911f PM |
443 | } |
444 | ||
1ed509a2 PM |
445 | #ifdef CONFIG_RCU_CPU_STALL_VERBOSE |
446 | ||
447 | /* | |
448 | * Dump detailed information for all tasks blocking the current RCU | |
449 | * grace period on the specified rcu_node structure. | |
450 | */ | |
451 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) | |
452 | { | |
453 | unsigned long flags; | |
1ed509a2 PM |
454 | struct task_struct *t; |
455 | ||
12f5f524 | 456 | raw_spin_lock_irqsave(&rnp->lock, flags); |
5fd4dc06 PM |
457 | if (!rcu_preempt_blocked_readers_cgp(rnp)) { |
458 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
459 | return; | |
460 | } | |
12f5f524 PM |
461 | t = list_entry(rnp->gp_tasks, |
462 | struct task_struct, rcu_node_entry); | |
463 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) | |
464 | sched_show_task(t); | |
465 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1ed509a2 PM |
466 | } |
467 | ||
468 | /* | |
469 | * Dump detailed information for all tasks blocking the current RCU | |
470 | * grace period. | |
471 | */ | |
472 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
473 | { | |
474 | struct rcu_node *rnp = rcu_get_root(rsp); | |
475 | ||
476 | rcu_print_detail_task_stall_rnp(rnp); | |
477 | rcu_for_each_leaf_node(rsp, rnp) | |
478 | rcu_print_detail_task_stall_rnp(rnp); | |
479 | } | |
480 | ||
481 | #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ | |
482 | ||
483 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
484 | { | |
485 | } | |
486 | ||
487 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ | |
488 | ||
a858af28 PM |
489 | #ifdef CONFIG_RCU_CPU_STALL_INFO |
490 | ||
491 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) | |
492 | { | |
efc151c3 | 493 | pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):", |
a858af28 PM |
494 | rnp->level, rnp->grplo, rnp->grphi); |
495 | } | |
496 | ||
497 | static void rcu_print_task_stall_end(void) | |
498 | { | |
efc151c3 | 499 | pr_cont("\n"); |
a858af28 PM |
500 | } |
501 | ||
502 | #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */ | |
503 | ||
504 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) | |
505 | { | |
506 | } | |
507 | ||
508 | static void rcu_print_task_stall_end(void) | |
509 | { | |
510 | } | |
511 | ||
512 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */ | |
513 | ||
f41d911f PM |
514 | /* |
515 | * Scan the current list of tasks blocked within RCU read-side critical | |
516 | * sections, printing out the tid of each. | |
517 | */ | |
9bc8b558 | 518 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 519 | { |
f41d911f | 520 | struct task_struct *t; |
9bc8b558 | 521 | int ndetected = 0; |
f41d911f | 522 | |
27f4d280 | 523 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
9bc8b558 | 524 | return 0; |
a858af28 | 525 | rcu_print_task_stall_begin(rnp); |
12f5f524 PM |
526 | t = list_entry(rnp->gp_tasks, |
527 | struct task_struct, rcu_node_entry); | |
9bc8b558 | 528 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { |
efc151c3 | 529 | pr_cont(" P%d", t->pid); |
9bc8b558 PM |
530 | ndetected++; |
531 | } | |
a858af28 | 532 | rcu_print_task_stall_end(); |
9bc8b558 | 533 | return ndetected; |
f41d911f PM |
534 | } |
535 | ||
b0e165c0 PM |
536 | /* |
537 | * Check that the list of blocked tasks for the newly completed grace | |
538 | * period is in fact empty. It is a serious bug to complete a grace | |
539 | * period that still has RCU readers blocked! This function must be | |
540 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock | |
541 | * must be held by the caller. | |
12f5f524 PM |
542 | * |
543 | * Also, if there are blocked tasks on the list, they automatically | |
544 | * block the newly created grace period, so set up ->gp_tasks accordingly. | |
b0e165c0 PM |
545 | */ |
546 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
547 | { | |
27f4d280 | 548 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
12f5f524 PM |
549 | if (!list_empty(&rnp->blkd_tasks)) |
550 | rnp->gp_tasks = rnp->blkd_tasks.next; | |
28ecd580 | 551 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
552 | } |
553 | ||
33f76148 PM |
554 | #ifdef CONFIG_HOTPLUG_CPU |
555 | ||
dd5d19ba PM |
556 | /* |
557 | * Handle tasklist migration for case in which all CPUs covered by the | |
558 | * specified rcu_node have gone offline. Move them up to the root | |
559 | * rcu_node. The reason for not just moving them to the immediate | |
560 | * parent is to remove the need for rcu_read_unlock_special() to | |
561 | * make more than two attempts to acquire the target rcu_node's lock. | |
b668c9cf PM |
562 | * Returns true if there were tasks blocking the current RCU grace |
563 | * period. | |
dd5d19ba | 564 | * |
237c80c5 PM |
565 | * Returns 1 if there was previously a task blocking the current grace |
566 | * period on the specified rcu_node structure. | |
567 | * | |
dd5d19ba PM |
568 | * The caller must hold rnp->lock with irqs disabled. |
569 | */ | |
237c80c5 PM |
570 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
571 | struct rcu_node *rnp, | |
572 | struct rcu_data *rdp) | |
dd5d19ba | 573 | { |
dd5d19ba PM |
574 | struct list_head *lp; |
575 | struct list_head *lp_root; | |
d9a3da06 | 576 | int retval = 0; |
dd5d19ba | 577 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
12f5f524 | 578 | struct task_struct *t; |
dd5d19ba | 579 | |
86848966 PM |
580 | if (rnp == rnp_root) { |
581 | WARN_ONCE(1, "Last CPU thought to be offlined?"); | |
237c80c5 | 582 | return 0; /* Shouldn't happen: at least one CPU online. */ |
86848966 | 583 | } |
12f5f524 PM |
584 | |
585 | /* If we are on an internal node, complain bitterly. */ | |
586 | WARN_ON_ONCE(rnp != rdp->mynode); | |
dd5d19ba PM |
587 | |
588 | /* | |
12f5f524 PM |
589 | * Move tasks up to root rcu_node. Don't try to get fancy for |
590 | * this corner-case operation -- just put this node's tasks | |
591 | * at the head of the root node's list, and update the root node's | |
592 | * ->gp_tasks and ->exp_tasks pointers to those of this node's, | |
593 | * if non-NULL. This might result in waiting for more tasks than | |
594 | * absolutely necessary, but this is a good performance/complexity | |
595 | * tradeoff. | |
dd5d19ba | 596 | */ |
2036d94a | 597 | if (rcu_preempt_blocked_readers_cgp(rnp) && rnp->qsmask == 0) |
d9a3da06 PM |
598 | retval |= RCU_OFL_TASKS_NORM_GP; |
599 | if (rcu_preempted_readers_exp(rnp)) | |
600 | retval |= RCU_OFL_TASKS_EXP_GP; | |
12f5f524 PM |
601 | lp = &rnp->blkd_tasks; |
602 | lp_root = &rnp_root->blkd_tasks; | |
603 | while (!list_empty(lp)) { | |
604 | t = list_entry(lp->next, typeof(*t), rcu_node_entry); | |
605 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ | |
6303b9c8 | 606 | smp_mb__after_unlock_lock(); |
12f5f524 PM |
607 | list_del(&t->rcu_node_entry); |
608 | t->rcu_blocked_node = rnp_root; | |
609 | list_add(&t->rcu_node_entry, lp_root); | |
610 | if (&t->rcu_node_entry == rnp->gp_tasks) | |
611 | rnp_root->gp_tasks = rnp->gp_tasks; | |
612 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
613 | rnp_root->exp_tasks = rnp->exp_tasks; | |
27f4d280 PM |
614 | #ifdef CONFIG_RCU_BOOST |
615 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
616 | rnp_root->boost_tasks = rnp->boost_tasks; | |
617 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
12f5f524 | 618 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ |
dd5d19ba | 619 | } |
27f4d280 | 620 | |
1e3fd2b3 PM |
621 | rnp->gp_tasks = NULL; |
622 | rnp->exp_tasks = NULL; | |
27f4d280 | 623 | #ifdef CONFIG_RCU_BOOST |
1e3fd2b3 | 624 | rnp->boost_tasks = NULL; |
5cc900cf PM |
625 | /* |
626 | * In case root is being boosted and leaf was not. Make sure | |
627 | * that we boost the tasks blocking the current grace period | |
628 | * in this case. | |
629 | */ | |
27f4d280 | 630 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ |
6303b9c8 | 631 | smp_mb__after_unlock_lock(); |
27f4d280 | 632 | if (rnp_root->boost_tasks != NULL && |
5cc900cf PM |
633 | rnp_root->boost_tasks != rnp_root->gp_tasks && |
634 | rnp_root->boost_tasks != rnp_root->exp_tasks) | |
27f4d280 PM |
635 | rnp_root->boost_tasks = rnp_root->gp_tasks; |
636 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ | |
637 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
638 | ||
237c80c5 | 639 | return retval; |
dd5d19ba PM |
640 | } |
641 | ||
e5601400 PM |
642 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
643 | ||
f41d911f PM |
644 | /* |
645 | * Check for a quiescent state from the current CPU. When a task blocks, | |
646 | * the task is recorded in the corresponding CPU's rcu_node structure, | |
647 | * which is checked elsewhere. | |
648 | * | |
649 | * Caller must disable hard irqs. | |
650 | */ | |
651 | static void rcu_preempt_check_callbacks(int cpu) | |
652 | { | |
653 | struct task_struct *t = current; | |
654 | ||
655 | if (t->rcu_read_lock_nesting == 0) { | |
c3422bea | 656 | rcu_preempt_qs(cpu); |
f41d911f PM |
657 | return; |
658 | } | |
10f39bb1 PM |
659 | if (t->rcu_read_lock_nesting > 0 && |
660 | per_cpu(rcu_preempt_data, cpu).qs_pending) | |
c3422bea | 661 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS; |
f41d911f PM |
662 | } |
663 | ||
a46e0899 PM |
664 | #ifdef CONFIG_RCU_BOOST |
665 | ||
09223371 SL |
666 | static void rcu_preempt_do_callbacks(void) |
667 | { | |
c9d4b0af | 668 | rcu_do_batch(&rcu_preempt_state, this_cpu_ptr(&rcu_preempt_data)); |
09223371 SL |
669 | } |
670 | ||
a46e0899 PM |
671 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
672 | ||
f41d911f | 673 | /* |
6cc68793 | 674 | * Queue a preemptible-RCU callback for invocation after a grace period. |
f41d911f PM |
675 | */ |
676 | void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
677 | { | |
3fbfbf7a | 678 | __call_rcu(head, func, &rcu_preempt_state, -1, 0); |
f41d911f PM |
679 | } |
680 | EXPORT_SYMBOL_GPL(call_rcu); | |
681 | ||
6ebb237b PM |
682 | /** |
683 | * synchronize_rcu - wait until a grace period has elapsed. | |
684 | * | |
685 | * Control will return to the caller some time after a full grace | |
686 | * period has elapsed, in other words after all currently executing RCU | |
77d8485a PM |
687 | * read-side critical sections have completed. Note, however, that |
688 | * upon return from synchronize_rcu(), the caller might well be executing | |
689 | * concurrently with new RCU read-side critical sections that began while | |
690 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
691 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
f0a0e6f2 PM |
692 | * |
693 | * See the description of synchronize_sched() for more detailed information | |
694 | * on memory ordering guarantees. | |
6ebb237b PM |
695 | */ |
696 | void synchronize_rcu(void) | |
697 | { | |
fe15d706 PM |
698 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
699 | !lock_is_held(&rcu_lock_map) && | |
700 | !lock_is_held(&rcu_sched_lock_map), | |
701 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
6ebb237b PM |
702 | if (!rcu_scheduler_active) |
703 | return; | |
3705b88d AM |
704 | if (rcu_expedited) |
705 | synchronize_rcu_expedited(); | |
706 | else | |
707 | wait_rcu_gp(call_rcu); | |
6ebb237b PM |
708 | } |
709 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
710 | ||
d9a3da06 | 711 | static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq); |
bcfa57ce | 712 | static unsigned long sync_rcu_preempt_exp_count; |
d9a3da06 PM |
713 | static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex); |
714 | ||
715 | /* | |
716 | * Return non-zero if there are any tasks in RCU read-side critical | |
717 | * sections blocking the current preemptible-RCU expedited grace period. | |
718 | * If there is no preemptible-RCU expedited grace period currently in | |
719 | * progress, returns zero unconditionally. | |
720 | */ | |
721 | static int rcu_preempted_readers_exp(struct rcu_node *rnp) | |
722 | { | |
12f5f524 | 723 | return rnp->exp_tasks != NULL; |
d9a3da06 PM |
724 | } |
725 | ||
726 | /* | |
727 | * return non-zero if there is no RCU expedited grace period in progress | |
728 | * for the specified rcu_node structure, in other words, if all CPUs and | |
729 | * tasks covered by the specified rcu_node structure have done their bit | |
730 | * for the current expedited grace period. Works only for preemptible | |
731 | * RCU -- other RCU implementation use other means. | |
732 | * | |
733 | * Caller must hold sync_rcu_preempt_exp_mutex. | |
734 | */ | |
735 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) | |
736 | { | |
737 | return !rcu_preempted_readers_exp(rnp) && | |
738 | ACCESS_ONCE(rnp->expmask) == 0; | |
739 | } | |
740 | ||
741 | /* | |
742 | * Report the exit from RCU read-side critical section for the last task | |
743 | * that queued itself during or before the current expedited preemptible-RCU | |
744 | * grace period. This event is reported either to the rcu_node structure on | |
745 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
746 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
747 | * iteratively!) | |
748 | * | |
b40d293e TG |
749 | * Most callers will set the "wake" flag, but the task initiating the |
750 | * expedited grace period need not wake itself. | |
751 | * | |
d9a3da06 PM |
752 | * Caller must hold sync_rcu_preempt_exp_mutex. |
753 | */ | |
b40d293e TG |
754 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
755 | bool wake) | |
d9a3da06 PM |
756 | { |
757 | unsigned long flags; | |
758 | unsigned long mask; | |
759 | ||
1304afb2 | 760 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 761 | smp_mb__after_unlock_lock(); |
d9a3da06 | 762 | for (;;) { |
131906b0 PM |
763 | if (!sync_rcu_preempt_exp_done(rnp)) { |
764 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
d9a3da06 | 765 | break; |
131906b0 | 766 | } |
d9a3da06 | 767 | if (rnp->parent == NULL) { |
131906b0 | 768 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
78e4bc34 PM |
769 | if (wake) { |
770 | smp_mb(); /* EGP done before wake_up(). */ | |
b40d293e | 771 | wake_up(&sync_rcu_preempt_exp_wq); |
78e4bc34 | 772 | } |
d9a3da06 PM |
773 | break; |
774 | } | |
775 | mask = rnp->grpmask; | |
1304afb2 | 776 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
d9a3da06 | 777 | rnp = rnp->parent; |
1304afb2 | 778 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
6303b9c8 | 779 | smp_mb__after_unlock_lock(); |
d9a3da06 PM |
780 | rnp->expmask &= ~mask; |
781 | } | |
d9a3da06 PM |
782 | } |
783 | ||
784 | /* | |
785 | * Snapshot the tasks blocking the newly started preemptible-RCU expedited | |
786 | * grace period for the specified rcu_node structure. If there are no such | |
787 | * tasks, report it up the rcu_node hierarchy. | |
788 | * | |
7b2e6011 PM |
789 | * Caller must hold sync_rcu_preempt_exp_mutex and must exclude |
790 | * CPU hotplug operations. | |
d9a3da06 PM |
791 | */ |
792 | static void | |
793 | sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp) | |
794 | { | |
1217ed1b | 795 | unsigned long flags; |
12f5f524 | 796 | int must_wait = 0; |
d9a3da06 | 797 | |
1217ed1b | 798 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 799 | smp_mb__after_unlock_lock(); |
c701d5d9 | 800 | if (list_empty(&rnp->blkd_tasks)) { |
1217ed1b | 801 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
c701d5d9 | 802 | } else { |
12f5f524 | 803 | rnp->exp_tasks = rnp->blkd_tasks.next; |
1217ed1b | 804 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
12f5f524 PM |
805 | must_wait = 1; |
806 | } | |
d9a3da06 | 807 | if (!must_wait) |
b40d293e | 808 | rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */ |
d9a3da06 PM |
809 | } |
810 | ||
236fefaf PM |
811 | /** |
812 | * synchronize_rcu_expedited - Brute-force RCU grace period | |
813 | * | |
814 | * Wait for an RCU-preempt grace period, but expedite it. The basic | |
815 | * idea is to invoke synchronize_sched_expedited() to push all the tasks to | |
816 | * the ->blkd_tasks lists and wait for this list to drain. This consumes | |
817 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
818 | * so is thus not recommended for any sort of common-case code. | |
819 | * In fact, if you are using synchronize_rcu_expedited() in a loop, | |
820 | * please restructure your code to batch your updates, and then Use a | |
821 | * single synchronize_rcu() instead. | |
822 | * | |
823 | * Note that it is illegal to call this function while holding any lock | |
824 | * that is acquired by a CPU-hotplug notifier. And yes, it is also illegal | |
825 | * to call this function from a CPU-hotplug notifier. Failing to observe | |
826 | * these restriction will result in deadlock. | |
019129d5 PM |
827 | */ |
828 | void synchronize_rcu_expedited(void) | |
829 | { | |
d9a3da06 PM |
830 | unsigned long flags; |
831 | struct rcu_node *rnp; | |
832 | struct rcu_state *rsp = &rcu_preempt_state; | |
bcfa57ce | 833 | unsigned long snap; |
d9a3da06 PM |
834 | int trycount = 0; |
835 | ||
836 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
837 | snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1; | |
838 | smp_mb(); /* Above access cannot bleed into critical section. */ | |
839 | ||
1943c89d PM |
840 | /* |
841 | * Block CPU-hotplug operations. This means that any CPU-hotplug | |
842 | * operation that finds an rcu_node structure with tasks in the | |
843 | * process of being boosted will know that all tasks blocking | |
844 | * this expedited grace period will already be in the process of | |
845 | * being boosted. This simplifies the process of moving tasks | |
846 | * from leaf to root rcu_node structures. | |
847 | */ | |
848 | get_online_cpus(); | |
849 | ||
d9a3da06 PM |
850 | /* |
851 | * Acquire lock, falling back to synchronize_rcu() if too many | |
852 | * lock-acquisition failures. Of course, if someone does the | |
853 | * expedited grace period for us, just leave. | |
854 | */ | |
855 | while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) { | |
1943c89d PM |
856 | if (ULONG_CMP_LT(snap, |
857 | ACCESS_ONCE(sync_rcu_preempt_exp_count))) { | |
858 | put_online_cpus(); | |
859 | goto mb_ret; /* Others did our work for us. */ | |
860 | } | |
c701d5d9 | 861 | if (trycount++ < 10) { |
d9a3da06 | 862 | udelay(trycount * num_online_cpus()); |
c701d5d9 | 863 | } else { |
1943c89d | 864 | put_online_cpus(); |
3705b88d | 865 | wait_rcu_gp(call_rcu); |
d9a3da06 PM |
866 | return; |
867 | } | |
d9a3da06 | 868 | } |
1943c89d PM |
869 | if (ULONG_CMP_LT(snap, ACCESS_ONCE(sync_rcu_preempt_exp_count))) { |
870 | put_online_cpus(); | |
d9a3da06 | 871 | goto unlock_mb_ret; /* Others did our work for us. */ |
1943c89d | 872 | } |
d9a3da06 | 873 | |
12f5f524 | 874 | /* force all RCU readers onto ->blkd_tasks lists. */ |
d9a3da06 PM |
875 | synchronize_sched_expedited(); |
876 | ||
d9a3da06 PM |
877 | /* Initialize ->expmask for all non-leaf rcu_node structures. */ |
878 | rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) { | |
1943c89d | 879 | raw_spin_lock_irqsave(&rnp->lock, flags); |
6303b9c8 | 880 | smp_mb__after_unlock_lock(); |
d9a3da06 | 881 | rnp->expmask = rnp->qsmaskinit; |
1943c89d | 882 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
d9a3da06 PM |
883 | } |
884 | ||
12f5f524 | 885 | /* Snapshot current state of ->blkd_tasks lists. */ |
d9a3da06 PM |
886 | rcu_for_each_leaf_node(rsp, rnp) |
887 | sync_rcu_preempt_exp_init(rsp, rnp); | |
888 | if (NUM_RCU_NODES > 1) | |
889 | sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp)); | |
890 | ||
1943c89d | 891 | put_online_cpus(); |
d9a3da06 | 892 | |
12f5f524 | 893 | /* Wait for snapshotted ->blkd_tasks lists to drain. */ |
d9a3da06 PM |
894 | rnp = rcu_get_root(rsp); |
895 | wait_event(sync_rcu_preempt_exp_wq, | |
896 | sync_rcu_preempt_exp_done(rnp)); | |
897 | ||
898 | /* Clean up and exit. */ | |
899 | smp_mb(); /* ensure expedited GP seen before counter increment. */ | |
900 | ACCESS_ONCE(sync_rcu_preempt_exp_count)++; | |
901 | unlock_mb_ret: | |
902 | mutex_unlock(&sync_rcu_preempt_exp_mutex); | |
903 | mb_ret: | |
904 | smp_mb(); /* ensure subsequent action seen after grace period. */ | |
019129d5 PM |
905 | } |
906 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
907 | ||
e74f4c45 PM |
908 | /** |
909 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
f0a0e6f2 PM |
910 | * |
911 | * Note that this primitive does not necessarily wait for an RCU grace period | |
912 | * to complete. For example, if there are no RCU callbacks queued anywhere | |
913 | * in the system, then rcu_barrier() is within its rights to return | |
914 | * immediately, without waiting for anything, much less an RCU grace period. | |
e74f4c45 PM |
915 | */ |
916 | void rcu_barrier(void) | |
917 | { | |
037b64ed | 918 | _rcu_barrier(&rcu_preempt_state); |
e74f4c45 PM |
919 | } |
920 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
921 | ||
1eba8f84 | 922 | /* |
6cc68793 | 923 | * Initialize preemptible RCU's state structures. |
1eba8f84 PM |
924 | */ |
925 | static void __init __rcu_init_preempt(void) | |
926 | { | |
394f99a9 | 927 | rcu_init_one(&rcu_preempt_state, &rcu_preempt_data); |
1eba8f84 PM |
928 | } |
929 | ||
2439b696 PM |
930 | /* |
931 | * Check for a task exiting while in a preemptible-RCU read-side | |
932 | * critical section, clean up if so. No need to issue warnings, | |
933 | * as debug_check_no_locks_held() already does this if lockdep | |
934 | * is enabled. | |
935 | */ | |
936 | void exit_rcu(void) | |
937 | { | |
938 | struct task_struct *t = current; | |
939 | ||
940 | if (likely(list_empty(¤t->rcu_node_entry))) | |
941 | return; | |
942 | t->rcu_read_lock_nesting = 1; | |
943 | barrier(); | |
944 | t->rcu_read_unlock_special = RCU_READ_UNLOCK_BLOCKED; | |
945 | __rcu_read_unlock(); | |
946 | } | |
947 | ||
f41d911f PM |
948 | #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */ |
949 | ||
e534165b | 950 | static struct rcu_state *rcu_state_p = &rcu_sched_state; |
27f4d280 | 951 | |
f41d911f PM |
952 | /* |
953 | * Tell them what RCU they are running. | |
954 | */ | |
0e0fc1c2 | 955 | static void __init rcu_bootup_announce(void) |
f41d911f | 956 | { |
efc151c3 | 957 | pr_info("Hierarchical RCU implementation.\n"); |
26845c28 | 958 | rcu_bootup_announce_oddness(); |
f41d911f PM |
959 | } |
960 | ||
961 | /* | |
962 | * Return the number of RCU batches processed thus far for debug & stats. | |
963 | */ | |
964 | long rcu_batches_completed(void) | |
965 | { | |
966 | return rcu_batches_completed_sched(); | |
967 | } | |
968 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
969 | ||
cba6d0d6 PM |
970 | /* |
971 | * Because preemptible RCU does not exist, we never have to check for | |
972 | * CPUs being in quiescent states. | |
973 | */ | |
974 | static void rcu_preempt_note_context_switch(int cpu) | |
975 | { | |
976 | } | |
977 | ||
fc2219d4 | 978 | /* |
6cc68793 | 979 | * Because preemptible RCU does not exist, there are never any preempted |
fc2219d4 PM |
980 | * RCU readers. |
981 | */ | |
27f4d280 | 982 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 PM |
983 | { |
984 | return 0; | |
985 | } | |
986 | ||
b668c9cf PM |
987 | #ifdef CONFIG_HOTPLUG_CPU |
988 | ||
989 | /* Because preemptible RCU does not exist, no quieting of tasks. */ | |
d3f6bad3 | 990 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
b41d1b92 | 991 | __releases(rnp->lock) |
b668c9cf | 992 | { |
1304afb2 | 993 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf PM |
994 | } |
995 | ||
996 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
997 | ||
1ed509a2 | 998 | /* |
6cc68793 | 999 | * Because preemptible RCU does not exist, we never have to check for |
1ed509a2 PM |
1000 | * tasks blocked within RCU read-side critical sections. |
1001 | */ | |
1002 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
1003 | { | |
1004 | } | |
1005 | ||
f41d911f | 1006 | /* |
6cc68793 | 1007 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
1008 | * tasks blocked within RCU read-side critical sections. |
1009 | */ | |
9bc8b558 | 1010 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 1011 | { |
9bc8b558 | 1012 | return 0; |
f41d911f PM |
1013 | } |
1014 | ||
b0e165c0 | 1015 | /* |
6cc68793 | 1016 | * Because there is no preemptible RCU, there can be no readers blocked, |
49e29126 PM |
1017 | * so there is no need to check for blocked tasks. So check only for |
1018 | * bogus qsmask values. | |
b0e165c0 PM |
1019 | */ |
1020 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
1021 | { | |
49e29126 | 1022 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
1023 | } |
1024 | ||
33f76148 PM |
1025 | #ifdef CONFIG_HOTPLUG_CPU |
1026 | ||
dd5d19ba | 1027 | /* |
6cc68793 | 1028 | * Because preemptible RCU does not exist, it never needs to migrate |
237c80c5 PM |
1029 | * tasks that were blocked within RCU read-side critical sections, and |
1030 | * such non-existent tasks cannot possibly have been blocking the current | |
1031 | * grace period. | |
dd5d19ba | 1032 | */ |
237c80c5 PM |
1033 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
1034 | struct rcu_node *rnp, | |
1035 | struct rcu_data *rdp) | |
dd5d19ba | 1036 | { |
237c80c5 | 1037 | return 0; |
dd5d19ba PM |
1038 | } |
1039 | ||
e5601400 PM |
1040 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
1041 | ||
f41d911f | 1042 | /* |
6cc68793 | 1043 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
1044 | * to check. |
1045 | */ | |
1eba8f84 | 1046 | static void rcu_preempt_check_callbacks(int cpu) |
f41d911f PM |
1047 | { |
1048 | } | |
1049 | ||
019129d5 PM |
1050 | /* |
1051 | * Wait for an rcu-preempt grace period, but make it happen quickly. | |
6cc68793 | 1052 | * But because preemptible RCU does not exist, map to rcu-sched. |
019129d5 PM |
1053 | */ |
1054 | void synchronize_rcu_expedited(void) | |
1055 | { | |
1056 | synchronize_sched_expedited(); | |
1057 | } | |
1058 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
1059 | ||
d9a3da06 PM |
1060 | #ifdef CONFIG_HOTPLUG_CPU |
1061 | ||
1062 | /* | |
6cc68793 | 1063 | * Because preemptible RCU does not exist, there is never any need to |
d9a3da06 PM |
1064 | * report on tasks preempted in RCU read-side critical sections during |
1065 | * expedited RCU grace periods. | |
1066 | */ | |
b40d293e TG |
1067 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
1068 | bool wake) | |
d9a3da06 | 1069 | { |
d9a3da06 PM |
1070 | } |
1071 | ||
1072 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1073 | ||
e74f4c45 | 1074 | /* |
6cc68793 | 1075 | * Because preemptible RCU does not exist, rcu_barrier() is just |
e74f4c45 PM |
1076 | * another name for rcu_barrier_sched(). |
1077 | */ | |
1078 | void rcu_barrier(void) | |
1079 | { | |
1080 | rcu_barrier_sched(); | |
1081 | } | |
1082 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
1083 | ||
1eba8f84 | 1084 | /* |
6cc68793 | 1085 | * Because preemptible RCU does not exist, it need not be initialized. |
1eba8f84 PM |
1086 | */ |
1087 | static void __init __rcu_init_preempt(void) | |
1088 | { | |
1089 | } | |
1090 | ||
2439b696 PM |
1091 | /* |
1092 | * Because preemptible RCU does not exist, tasks cannot possibly exit | |
1093 | * while in preemptible RCU read-side critical sections. | |
1094 | */ | |
1095 | void exit_rcu(void) | |
1096 | { | |
1097 | } | |
1098 | ||
f41d911f | 1099 | #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */ |
8bd93a2c | 1100 | |
27f4d280 PM |
1101 | #ifdef CONFIG_RCU_BOOST |
1102 | ||
1696a8be | 1103 | #include "../locking/rtmutex_common.h" |
27f4d280 | 1104 | |
0ea1f2eb PM |
1105 | #ifdef CONFIG_RCU_TRACE |
1106 | ||
1107 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
1108 | { | |
1109 | if (list_empty(&rnp->blkd_tasks)) | |
1110 | rnp->n_balk_blkd_tasks++; | |
1111 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) | |
1112 | rnp->n_balk_exp_gp_tasks++; | |
1113 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) | |
1114 | rnp->n_balk_boost_tasks++; | |
1115 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) | |
1116 | rnp->n_balk_notblocked++; | |
1117 | else if (rnp->gp_tasks != NULL && | |
a9f4793d | 1118 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
0ea1f2eb PM |
1119 | rnp->n_balk_notyet++; |
1120 | else | |
1121 | rnp->n_balk_nos++; | |
1122 | } | |
1123 | ||
1124 | #else /* #ifdef CONFIG_RCU_TRACE */ | |
1125 | ||
1126 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
1127 | { | |
1128 | } | |
1129 | ||
1130 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ | |
1131 | ||
5d01bbd1 TG |
1132 | static void rcu_wake_cond(struct task_struct *t, int status) |
1133 | { | |
1134 | /* | |
1135 | * If the thread is yielding, only wake it when this | |
1136 | * is invoked from idle | |
1137 | */ | |
1138 | if (status != RCU_KTHREAD_YIELDING || is_idle_task(current)) | |
1139 | wake_up_process(t); | |
1140 | } | |
1141 | ||
27f4d280 PM |
1142 | /* |
1143 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks | |
1144 | * or ->boost_tasks, advancing the pointer to the next task in the | |
1145 | * ->blkd_tasks list. | |
1146 | * | |
1147 | * Note that irqs must be enabled: boosting the task can block. | |
1148 | * Returns 1 if there are more tasks needing to be boosted. | |
1149 | */ | |
1150 | static int rcu_boost(struct rcu_node *rnp) | |
1151 | { | |
1152 | unsigned long flags; | |
27f4d280 PM |
1153 | struct task_struct *t; |
1154 | struct list_head *tb; | |
1155 | ||
1156 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) | |
1157 | return 0; /* Nothing left to boost. */ | |
1158 | ||
1159 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 1160 | smp_mb__after_unlock_lock(); |
27f4d280 PM |
1161 | |
1162 | /* | |
1163 | * Recheck under the lock: all tasks in need of boosting | |
1164 | * might exit their RCU read-side critical sections on their own. | |
1165 | */ | |
1166 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { | |
1167 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1168 | return 0; | |
1169 | } | |
1170 | ||
1171 | /* | |
1172 | * Preferentially boost tasks blocking expedited grace periods. | |
1173 | * This cannot starve the normal grace periods because a second | |
1174 | * expedited grace period must boost all blocked tasks, including | |
1175 | * those blocking the pre-existing normal grace period. | |
1176 | */ | |
0ea1f2eb | 1177 | if (rnp->exp_tasks != NULL) { |
27f4d280 | 1178 | tb = rnp->exp_tasks; |
0ea1f2eb PM |
1179 | rnp->n_exp_boosts++; |
1180 | } else { | |
27f4d280 | 1181 | tb = rnp->boost_tasks; |
0ea1f2eb PM |
1182 | rnp->n_normal_boosts++; |
1183 | } | |
1184 | rnp->n_tasks_boosted++; | |
27f4d280 PM |
1185 | |
1186 | /* | |
1187 | * We boost task t by manufacturing an rt_mutex that appears to | |
1188 | * be held by task t. We leave a pointer to that rt_mutex where | |
1189 | * task t can find it, and task t will release the mutex when it | |
1190 | * exits its outermost RCU read-side critical section. Then | |
1191 | * simply acquiring this artificial rt_mutex will boost task | |
1192 | * t's priority. (Thanks to tglx for suggesting this approach!) | |
1193 | * | |
1194 | * Note that task t must acquire rnp->lock to remove itself from | |
1195 | * the ->blkd_tasks list, which it will do from exit() if from | |
1196 | * nowhere else. We therefore are guaranteed that task t will | |
1197 | * stay around at least until we drop rnp->lock. Note that | |
1198 | * rnp->lock also resolves races between our priority boosting | |
1199 | * and task t's exiting its outermost RCU read-side critical | |
1200 | * section. | |
1201 | */ | |
1202 | t = container_of(tb, struct task_struct, rcu_node_entry); | |
abaa93d9 | 1203 | rt_mutex_init_proxy_locked(&rnp->boost_mtx, t); |
dfeb9765 | 1204 | init_completion(&rnp->boost_completion); |
27f4d280 | 1205 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
abaa93d9 PM |
1206 | /* Lock only for side effect: boosts task t's priority. */ |
1207 | rt_mutex_lock(&rnp->boost_mtx); | |
1208 | rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */ | |
27f4d280 | 1209 | |
abaa93d9 | 1210 | /* Wait for boostee to be done w/boost_mtx before reinitializing. */ |
dfeb9765 | 1211 | wait_for_completion(&rnp->boost_completion); |
27f4d280 | 1212 | |
4f89b336 PM |
1213 | return ACCESS_ONCE(rnp->exp_tasks) != NULL || |
1214 | ACCESS_ONCE(rnp->boost_tasks) != NULL; | |
27f4d280 PM |
1215 | } |
1216 | ||
27f4d280 PM |
1217 | /* |
1218 | * Priority-boosting kthread. One per leaf rcu_node and one for the | |
1219 | * root rcu_node. | |
1220 | */ | |
1221 | static int rcu_boost_kthread(void *arg) | |
1222 | { | |
1223 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
1224 | int spincnt = 0; | |
1225 | int more2boost; | |
1226 | ||
f7f7bac9 | 1227 | trace_rcu_utilization(TPS("Start boost kthread@init")); |
27f4d280 | 1228 | for (;;) { |
d71df90e | 1229 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
f7f7bac9 | 1230 | trace_rcu_utilization(TPS("End boost kthread@rcu_wait")); |
08bca60a | 1231 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
f7f7bac9 | 1232 | trace_rcu_utilization(TPS("Start boost kthread@rcu_wait")); |
d71df90e | 1233 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
27f4d280 PM |
1234 | more2boost = rcu_boost(rnp); |
1235 | if (more2boost) | |
1236 | spincnt++; | |
1237 | else | |
1238 | spincnt = 0; | |
1239 | if (spincnt > 10) { | |
5d01bbd1 | 1240 | rnp->boost_kthread_status = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 1241 | trace_rcu_utilization(TPS("End boost kthread@rcu_yield")); |
5d01bbd1 | 1242 | schedule_timeout_interruptible(2); |
f7f7bac9 | 1243 | trace_rcu_utilization(TPS("Start boost kthread@rcu_yield")); |
27f4d280 PM |
1244 | spincnt = 0; |
1245 | } | |
1246 | } | |
1217ed1b | 1247 | /* NOTREACHED */ |
f7f7bac9 | 1248 | trace_rcu_utilization(TPS("End boost kthread@notreached")); |
27f4d280 PM |
1249 | return 0; |
1250 | } | |
1251 | ||
1252 | /* | |
1253 | * Check to see if it is time to start boosting RCU readers that are | |
1254 | * blocking the current grace period, and, if so, tell the per-rcu_node | |
1255 | * kthread to start boosting them. If there is an expedited grace | |
1256 | * period in progress, it is always time to boost. | |
1257 | * | |
b065a853 PM |
1258 | * The caller must hold rnp->lock, which this function releases. |
1259 | * The ->boost_kthread_task is immortal, so we don't need to worry | |
1260 | * about it going away. | |
27f4d280 | 1261 | */ |
1217ed1b | 1262 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 1263 | __releases(rnp->lock) |
27f4d280 PM |
1264 | { |
1265 | struct task_struct *t; | |
1266 | ||
0ea1f2eb PM |
1267 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
1268 | rnp->n_balk_exp_gp_tasks++; | |
1217ed1b | 1269 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 | 1270 | return; |
0ea1f2eb | 1271 | } |
27f4d280 PM |
1272 | if (rnp->exp_tasks != NULL || |
1273 | (rnp->gp_tasks != NULL && | |
1274 | rnp->boost_tasks == NULL && | |
1275 | rnp->qsmask == 0 && | |
1276 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { | |
1277 | if (rnp->exp_tasks == NULL) | |
1278 | rnp->boost_tasks = rnp->gp_tasks; | |
1217ed1b | 1279 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 | 1280 | t = rnp->boost_kthread_task; |
5d01bbd1 TG |
1281 | if (t) |
1282 | rcu_wake_cond(t, rnp->boost_kthread_status); | |
1217ed1b | 1283 | } else { |
0ea1f2eb | 1284 | rcu_initiate_boost_trace(rnp); |
1217ed1b PM |
1285 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1286 | } | |
27f4d280 PM |
1287 | } |
1288 | ||
a46e0899 PM |
1289 | /* |
1290 | * Wake up the per-CPU kthread to invoke RCU callbacks. | |
1291 | */ | |
1292 | static void invoke_rcu_callbacks_kthread(void) | |
1293 | { | |
1294 | unsigned long flags; | |
1295 | ||
1296 | local_irq_save(flags); | |
1297 | __this_cpu_write(rcu_cpu_has_work, 1); | |
1eb52121 | 1298 | if (__this_cpu_read(rcu_cpu_kthread_task) != NULL && |
5d01bbd1 TG |
1299 | current != __this_cpu_read(rcu_cpu_kthread_task)) { |
1300 | rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task), | |
1301 | __this_cpu_read(rcu_cpu_kthread_status)); | |
1302 | } | |
a46e0899 PM |
1303 | local_irq_restore(flags); |
1304 | } | |
1305 | ||
dff1672d PM |
1306 | /* |
1307 | * Is the current CPU running the RCU-callbacks kthread? | |
1308 | * Caller must have preemption disabled. | |
1309 | */ | |
1310 | static bool rcu_is_callbacks_kthread(void) | |
1311 | { | |
c9d4b0af | 1312 | return __this_cpu_read(rcu_cpu_kthread_task) == current; |
dff1672d PM |
1313 | } |
1314 | ||
27f4d280 PM |
1315 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) |
1316 | ||
1317 | /* | |
1318 | * Do priority-boost accounting for the start of a new grace period. | |
1319 | */ | |
1320 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) | |
1321 | { | |
1322 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; | |
1323 | } | |
1324 | ||
27f4d280 PM |
1325 | /* |
1326 | * Create an RCU-boost kthread for the specified node if one does not | |
1327 | * already exist. We only create this kthread for preemptible RCU. | |
1328 | * Returns zero if all is well, a negated errno otherwise. | |
1329 | */ | |
49fb4c62 | 1330 | static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp, |
5d01bbd1 | 1331 | struct rcu_node *rnp) |
27f4d280 | 1332 | { |
5d01bbd1 | 1333 | int rnp_index = rnp - &rsp->node[0]; |
27f4d280 PM |
1334 | unsigned long flags; |
1335 | struct sched_param sp; | |
1336 | struct task_struct *t; | |
1337 | ||
1338 | if (&rcu_preempt_state != rsp) | |
1339 | return 0; | |
5d01bbd1 TG |
1340 | |
1341 | if (!rcu_scheduler_fully_active || rnp->qsmaskinit == 0) | |
1342 | return 0; | |
1343 | ||
a46e0899 | 1344 | rsp->boost = 1; |
27f4d280 PM |
1345 | if (rnp->boost_kthread_task != NULL) |
1346 | return 0; | |
1347 | t = kthread_create(rcu_boost_kthread, (void *)rnp, | |
5b61b0ba | 1348 | "rcub/%d", rnp_index); |
27f4d280 PM |
1349 | if (IS_ERR(t)) |
1350 | return PTR_ERR(t); | |
1351 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 1352 | smp_mb__after_unlock_lock(); |
27f4d280 PM |
1353 | rnp->boost_kthread_task = t; |
1354 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
5b61b0ba | 1355 | sp.sched_priority = RCU_BOOST_PRIO; |
27f4d280 | 1356 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); |
9a432736 | 1357 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
27f4d280 PM |
1358 | return 0; |
1359 | } | |
1360 | ||
f8b7fc6b PM |
1361 | static void rcu_kthread_do_work(void) |
1362 | { | |
c9d4b0af CL |
1363 | rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data)); |
1364 | rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data)); | |
f8b7fc6b PM |
1365 | rcu_preempt_do_callbacks(); |
1366 | } | |
1367 | ||
62ab7072 | 1368 | static void rcu_cpu_kthread_setup(unsigned int cpu) |
f8b7fc6b | 1369 | { |
f8b7fc6b | 1370 | struct sched_param sp; |
f8b7fc6b | 1371 | |
62ab7072 PM |
1372 | sp.sched_priority = RCU_KTHREAD_PRIO; |
1373 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); | |
f8b7fc6b PM |
1374 | } |
1375 | ||
62ab7072 | 1376 | static void rcu_cpu_kthread_park(unsigned int cpu) |
f8b7fc6b | 1377 | { |
62ab7072 | 1378 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; |
f8b7fc6b PM |
1379 | } |
1380 | ||
62ab7072 | 1381 | static int rcu_cpu_kthread_should_run(unsigned int cpu) |
f8b7fc6b | 1382 | { |
c9d4b0af | 1383 | return __this_cpu_read(rcu_cpu_has_work); |
f8b7fc6b PM |
1384 | } |
1385 | ||
1386 | /* | |
1387 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the | |
e0f23060 PM |
1388 | * RCU softirq used in flavors and configurations of RCU that do not |
1389 | * support RCU priority boosting. | |
f8b7fc6b | 1390 | */ |
62ab7072 | 1391 | static void rcu_cpu_kthread(unsigned int cpu) |
f8b7fc6b | 1392 | { |
c9d4b0af CL |
1393 | unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status); |
1394 | char work, *workp = this_cpu_ptr(&rcu_cpu_has_work); | |
62ab7072 | 1395 | int spincnt; |
f8b7fc6b | 1396 | |
62ab7072 | 1397 | for (spincnt = 0; spincnt < 10; spincnt++) { |
f7f7bac9 | 1398 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait")); |
f8b7fc6b | 1399 | local_bh_disable(); |
f8b7fc6b | 1400 | *statusp = RCU_KTHREAD_RUNNING; |
62ab7072 PM |
1401 | this_cpu_inc(rcu_cpu_kthread_loops); |
1402 | local_irq_disable(); | |
f8b7fc6b PM |
1403 | work = *workp; |
1404 | *workp = 0; | |
62ab7072 | 1405 | local_irq_enable(); |
f8b7fc6b PM |
1406 | if (work) |
1407 | rcu_kthread_do_work(); | |
1408 | local_bh_enable(); | |
62ab7072 | 1409 | if (*workp == 0) { |
f7f7bac9 | 1410 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); |
62ab7072 PM |
1411 | *statusp = RCU_KTHREAD_WAITING; |
1412 | return; | |
f8b7fc6b PM |
1413 | } |
1414 | } | |
62ab7072 | 1415 | *statusp = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 1416 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); |
62ab7072 | 1417 | schedule_timeout_interruptible(2); |
f7f7bac9 | 1418 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); |
62ab7072 | 1419 | *statusp = RCU_KTHREAD_WAITING; |
f8b7fc6b PM |
1420 | } |
1421 | ||
1422 | /* | |
1423 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are | |
1424 | * served by the rcu_node in question. The CPU hotplug lock is still | |
1425 | * held, so the value of rnp->qsmaskinit will be stable. | |
1426 | * | |
1427 | * We don't include outgoingcpu in the affinity set, use -1 if there is | |
1428 | * no outgoing CPU. If there are no CPUs left in the affinity set, | |
1429 | * this function allows the kthread to execute on any CPU. | |
1430 | */ | |
5d01bbd1 | 1431 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b | 1432 | { |
5d01bbd1 TG |
1433 | struct task_struct *t = rnp->boost_kthread_task; |
1434 | unsigned long mask = rnp->qsmaskinit; | |
f8b7fc6b PM |
1435 | cpumask_var_t cm; |
1436 | int cpu; | |
f8b7fc6b | 1437 | |
5d01bbd1 | 1438 | if (!t) |
f8b7fc6b | 1439 | return; |
5d01bbd1 | 1440 | if (!zalloc_cpumask_var(&cm, GFP_KERNEL)) |
f8b7fc6b | 1441 | return; |
f8b7fc6b PM |
1442 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) |
1443 | if ((mask & 0x1) && cpu != outgoingcpu) | |
1444 | cpumask_set_cpu(cpu, cm); | |
1445 | if (cpumask_weight(cm) == 0) { | |
1446 | cpumask_setall(cm); | |
1447 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) | |
1448 | cpumask_clear_cpu(cpu, cm); | |
1449 | WARN_ON_ONCE(cpumask_weight(cm) == 0); | |
1450 | } | |
5d01bbd1 | 1451 | set_cpus_allowed_ptr(t, cm); |
f8b7fc6b PM |
1452 | free_cpumask_var(cm); |
1453 | } | |
1454 | ||
62ab7072 PM |
1455 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { |
1456 | .store = &rcu_cpu_kthread_task, | |
1457 | .thread_should_run = rcu_cpu_kthread_should_run, | |
1458 | .thread_fn = rcu_cpu_kthread, | |
1459 | .thread_comm = "rcuc/%u", | |
1460 | .setup = rcu_cpu_kthread_setup, | |
1461 | .park = rcu_cpu_kthread_park, | |
1462 | }; | |
f8b7fc6b PM |
1463 | |
1464 | /* | |
1465 | * Spawn all kthreads -- called as soon as the scheduler is running. | |
1466 | */ | |
1467 | static int __init rcu_spawn_kthreads(void) | |
1468 | { | |
f8b7fc6b | 1469 | struct rcu_node *rnp; |
5d01bbd1 | 1470 | int cpu; |
f8b7fc6b | 1471 | |
b0d30417 | 1472 | rcu_scheduler_fully_active = 1; |
62ab7072 | 1473 | for_each_possible_cpu(cpu) |
f8b7fc6b | 1474 | per_cpu(rcu_cpu_has_work, cpu) = 0; |
62ab7072 | 1475 | BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec)); |
e534165b US |
1476 | rnp = rcu_get_root(rcu_state_p); |
1477 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); | |
f8b7fc6b | 1478 | if (NUM_RCU_NODES > 1) { |
e534165b US |
1479 | rcu_for_each_leaf_node(rcu_state_p, rnp) |
1480 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); | |
f8b7fc6b PM |
1481 | } |
1482 | return 0; | |
1483 | } | |
1484 | early_initcall(rcu_spawn_kthreads); | |
1485 | ||
49fb4c62 | 1486 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b | 1487 | { |
e534165b | 1488 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
f8b7fc6b PM |
1489 | struct rcu_node *rnp = rdp->mynode; |
1490 | ||
1491 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ | |
62ab7072 | 1492 | if (rcu_scheduler_fully_active) |
e534165b | 1493 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); |
f8b7fc6b PM |
1494 | } |
1495 | ||
27f4d280 PM |
1496 | #else /* #ifdef CONFIG_RCU_BOOST */ |
1497 | ||
1217ed1b | 1498 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 1499 | __releases(rnp->lock) |
27f4d280 | 1500 | { |
1217ed1b | 1501 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 PM |
1502 | } |
1503 | ||
a46e0899 | 1504 | static void invoke_rcu_callbacks_kthread(void) |
27f4d280 | 1505 | { |
a46e0899 | 1506 | WARN_ON_ONCE(1); |
27f4d280 PM |
1507 | } |
1508 | ||
dff1672d PM |
1509 | static bool rcu_is_callbacks_kthread(void) |
1510 | { | |
1511 | return false; | |
1512 | } | |
1513 | ||
27f4d280 PM |
1514 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
1515 | { | |
1516 | } | |
1517 | ||
5d01bbd1 | 1518 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b PM |
1519 | { |
1520 | } | |
1521 | ||
b0d30417 PM |
1522 | static int __init rcu_scheduler_really_started(void) |
1523 | { | |
1524 | rcu_scheduler_fully_active = 1; | |
1525 | return 0; | |
1526 | } | |
1527 | early_initcall(rcu_scheduler_really_started); | |
1528 | ||
49fb4c62 | 1529 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b PM |
1530 | { |
1531 | } | |
1532 | ||
27f4d280 PM |
1533 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
1534 | ||
8bd93a2c PM |
1535 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
1536 | ||
1537 | /* | |
1538 | * Check to see if any future RCU-related work will need to be done | |
1539 | * by the current CPU, even if none need be done immediately, returning | |
1540 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
1541 | * an exported member of the RCU API. | |
1542 | * | |
7cb92499 PM |
1543 | * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs |
1544 | * any flavor of RCU. | |
8bd93a2c | 1545 | */ |
ffa83fb5 | 1546 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
aa9b1630 | 1547 | int rcu_needs_cpu(int cpu, unsigned long *delta_jiffies) |
8bd93a2c | 1548 | { |
aa9b1630 | 1549 | *delta_jiffies = ULONG_MAX; |
c0f4dfd4 | 1550 | return rcu_cpu_has_callbacks(cpu, NULL); |
7cb92499 | 1551 | } |
ffa83fb5 | 1552 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
7cb92499 PM |
1553 | |
1554 | /* | |
1555 | * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up | |
1556 | * after it. | |
1557 | */ | |
1558 | static void rcu_cleanup_after_idle(int cpu) | |
1559 | { | |
1560 | } | |
1561 | ||
aea1b35e | 1562 | /* |
a858af28 | 1563 | * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n, |
aea1b35e PM |
1564 | * is nothing. |
1565 | */ | |
1566 | static void rcu_prepare_for_idle(int cpu) | |
1567 | { | |
1568 | } | |
1569 | ||
c57afe80 PM |
1570 | /* |
1571 | * Don't bother keeping a running count of the number of RCU callbacks | |
1572 | * posted because CONFIG_RCU_FAST_NO_HZ=n. | |
1573 | */ | |
1574 | static void rcu_idle_count_callbacks_posted(void) | |
1575 | { | |
1576 | } | |
1577 | ||
8bd93a2c PM |
1578 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
1579 | ||
f23f7fa1 PM |
1580 | /* |
1581 | * This code is invoked when a CPU goes idle, at which point we want | |
1582 | * to have the CPU do everything required for RCU so that it can enter | |
1583 | * the energy-efficient dyntick-idle mode. This is handled by a | |
1584 | * state machine implemented by rcu_prepare_for_idle() below. | |
1585 | * | |
1586 | * The following three proprocessor symbols control this state machine: | |
1587 | * | |
f23f7fa1 PM |
1588 | * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted |
1589 | * to sleep in dyntick-idle mode with RCU callbacks pending. This | |
1590 | * is sized to be roughly one RCU grace period. Those energy-efficiency | |
1591 | * benchmarkers who might otherwise be tempted to set this to a large | |
1592 | * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your | |
1593 | * system. And if you are -that- concerned about energy efficiency, | |
1594 | * just power the system down and be done with it! | |
778d250a PM |
1595 | * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is |
1596 | * permitted to sleep in dyntick-idle mode with only lazy RCU | |
1597 | * callbacks pending. Setting this too high can OOM your system. | |
f23f7fa1 PM |
1598 | * |
1599 | * The values below work well in practice. If future workloads require | |
1600 | * adjustment, they can be converted into kernel config parameters, though | |
1601 | * making the state machine smarter might be a better option. | |
1602 | */ | |
e84c48ae | 1603 | #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */ |
778d250a | 1604 | #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */ |
f23f7fa1 | 1605 | |
5e44ce35 PM |
1606 | static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY; |
1607 | module_param(rcu_idle_gp_delay, int, 0644); | |
1608 | static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY; | |
1609 | module_param(rcu_idle_lazy_gp_delay, int, 0644); | |
486e2593 | 1610 | |
d689fe22 | 1611 | extern int tick_nohz_active; |
486e2593 PM |
1612 | |
1613 | /* | |
c229828c PM |
1614 | * Try to advance callbacks for all flavors of RCU on the current CPU, but |
1615 | * only if it has been awhile since the last time we did so. Afterwards, | |
1616 | * if there are any callbacks ready for immediate invocation, return true. | |
486e2593 | 1617 | */ |
f1f399d1 | 1618 | static bool __maybe_unused rcu_try_advance_all_cbs(void) |
486e2593 | 1619 | { |
c0f4dfd4 PM |
1620 | bool cbs_ready = false; |
1621 | struct rcu_data *rdp; | |
c229828c | 1622 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1623 | struct rcu_node *rnp; |
1624 | struct rcu_state *rsp; | |
486e2593 | 1625 | |
c229828c PM |
1626 | /* Exit early if we advanced recently. */ |
1627 | if (jiffies == rdtp->last_advance_all) | |
1628 | return 0; | |
1629 | rdtp->last_advance_all = jiffies; | |
1630 | ||
c0f4dfd4 PM |
1631 | for_each_rcu_flavor(rsp) { |
1632 | rdp = this_cpu_ptr(rsp->rda); | |
1633 | rnp = rdp->mynode; | |
486e2593 | 1634 | |
c0f4dfd4 PM |
1635 | /* |
1636 | * Don't bother checking unless a grace period has | |
1637 | * completed since we last checked and there are | |
1638 | * callbacks not yet ready to invoke. | |
1639 | */ | |
1640 | if (rdp->completed != rnp->completed && | |
1641 | rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL]) | |
470716fc | 1642 | note_gp_changes(rsp, rdp); |
486e2593 | 1643 | |
c0f4dfd4 PM |
1644 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
1645 | cbs_ready = true; | |
1646 | } | |
1647 | return cbs_ready; | |
486e2593 PM |
1648 | } |
1649 | ||
aa9b1630 | 1650 | /* |
c0f4dfd4 PM |
1651 | * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready |
1652 | * to invoke. If the CPU has callbacks, try to advance them. Tell the | |
1653 | * caller to set the timeout based on whether or not there are non-lazy | |
1654 | * callbacks. | |
aa9b1630 | 1655 | * |
c0f4dfd4 | 1656 | * The caller must have disabled interrupts. |
aa9b1630 | 1657 | */ |
ffa83fb5 | 1658 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
c0f4dfd4 | 1659 | int rcu_needs_cpu(int cpu, unsigned long *dj) |
aa9b1630 PM |
1660 | { |
1661 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); | |
1662 | ||
c0f4dfd4 PM |
1663 | /* Snapshot to detect later posting of non-lazy callback. */ |
1664 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
1665 | ||
aa9b1630 | 1666 | /* If no callbacks, RCU doesn't need the CPU. */ |
c0f4dfd4 PM |
1667 | if (!rcu_cpu_has_callbacks(cpu, &rdtp->all_lazy)) { |
1668 | *dj = ULONG_MAX; | |
aa9b1630 PM |
1669 | return 0; |
1670 | } | |
c0f4dfd4 PM |
1671 | |
1672 | /* Attempt to advance callbacks. */ | |
1673 | if (rcu_try_advance_all_cbs()) { | |
1674 | /* Some ready to invoke, so initiate later invocation. */ | |
1675 | invoke_rcu_core(); | |
aa9b1630 PM |
1676 | return 1; |
1677 | } | |
c0f4dfd4 PM |
1678 | rdtp->last_accelerate = jiffies; |
1679 | ||
1680 | /* Request timer delay depending on laziness, and round. */ | |
6faf7283 | 1681 | if (!rdtp->all_lazy) { |
c0f4dfd4 PM |
1682 | *dj = round_up(rcu_idle_gp_delay + jiffies, |
1683 | rcu_idle_gp_delay) - jiffies; | |
e84c48ae | 1684 | } else { |
c0f4dfd4 | 1685 | *dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies; |
e84c48ae | 1686 | } |
aa9b1630 PM |
1687 | return 0; |
1688 | } | |
ffa83fb5 | 1689 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
aa9b1630 | 1690 | |
21e52e15 | 1691 | /* |
c0f4dfd4 PM |
1692 | * Prepare a CPU for idle from an RCU perspective. The first major task |
1693 | * is to sense whether nohz mode has been enabled or disabled via sysfs. | |
1694 | * The second major task is to check to see if a non-lazy callback has | |
1695 | * arrived at a CPU that previously had only lazy callbacks. The third | |
1696 | * major task is to accelerate (that is, assign grace-period numbers to) | |
1697 | * any recently arrived callbacks. | |
aea1b35e PM |
1698 | * |
1699 | * The caller must have disabled interrupts. | |
8bd93a2c | 1700 | */ |
aea1b35e | 1701 | static void rcu_prepare_for_idle(int cpu) |
8bd93a2c | 1702 | { |
f1f399d1 | 1703 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
48a7639c | 1704 | bool needwake; |
c0f4dfd4 | 1705 | struct rcu_data *rdp; |
5955f7ee | 1706 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
c0f4dfd4 PM |
1707 | struct rcu_node *rnp; |
1708 | struct rcu_state *rsp; | |
9d2ad243 PM |
1709 | int tne; |
1710 | ||
1711 | /* Handle nohz enablement switches conservatively. */ | |
d689fe22 | 1712 | tne = ACCESS_ONCE(tick_nohz_active); |
9d2ad243 | 1713 | if (tne != rdtp->tick_nohz_enabled_snap) { |
c0f4dfd4 | 1714 | if (rcu_cpu_has_callbacks(cpu, NULL)) |
9d2ad243 PM |
1715 | invoke_rcu_core(); /* force nohz to see update. */ |
1716 | rdtp->tick_nohz_enabled_snap = tne; | |
1717 | return; | |
1718 | } | |
1719 | if (!tne) | |
1720 | return; | |
f511fc62 | 1721 | |
c0f4dfd4 | 1722 | /* If this is a no-CBs CPU, no callbacks, just return. */ |
534c97b0 | 1723 | if (rcu_is_nocb_cpu(cpu)) |
9a0c6fef | 1724 | return; |
9a0c6fef | 1725 | |
c57afe80 | 1726 | /* |
c0f4dfd4 PM |
1727 | * If a non-lazy callback arrived at a CPU having only lazy |
1728 | * callbacks, invoke RCU core for the side-effect of recalculating | |
1729 | * idle duration on re-entry to idle. | |
c57afe80 | 1730 | */ |
c0f4dfd4 PM |
1731 | if (rdtp->all_lazy && |
1732 | rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) { | |
c337f8f5 PM |
1733 | rdtp->all_lazy = false; |
1734 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
c0f4dfd4 | 1735 | invoke_rcu_core(); |
c57afe80 PM |
1736 | return; |
1737 | } | |
c57afe80 | 1738 | |
3084f2f8 | 1739 | /* |
c0f4dfd4 PM |
1740 | * If we have not yet accelerated this jiffy, accelerate all |
1741 | * callbacks on this CPU. | |
3084f2f8 | 1742 | */ |
c0f4dfd4 | 1743 | if (rdtp->last_accelerate == jiffies) |
aea1b35e | 1744 | return; |
c0f4dfd4 PM |
1745 | rdtp->last_accelerate = jiffies; |
1746 | for_each_rcu_flavor(rsp) { | |
1747 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
1748 | if (!*rdp->nxttail[RCU_DONE_TAIL]) | |
1749 | continue; | |
1750 | rnp = rdp->mynode; | |
1751 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
6303b9c8 | 1752 | smp_mb__after_unlock_lock(); |
48a7639c | 1753 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
c0f4dfd4 | 1754 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
48a7639c PM |
1755 | if (needwake) |
1756 | rcu_gp_kthread_wake(rsp); | |
77e38ed3 | 1757 | } |
f1f399d1 | 1758 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
c0f4dfd4 | 1759 | } |
3084f2f8 | 1760 | |
c0f4dfd4 PM |
1761 | /* |
1762 | * Clean up for exit from idle. Attempt to advance callbacks based on | |
1763 | * any grace periods that elapsed while the CPU was idle, and if any | |
1764 | * callbacks are now ready to invoke, initiate invocation. | |
1765 | */ | |
1766 | static void rcu_cleanup_after_idle(int cpu) | |
1767 | { | |
f1f399d1 | 1768 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
534c97b0 | 1769 | if (rcu_is_nocb_cpu(cpu)) |
aea1b35e | 1770 | return; |
7a497c96 PM |
1771 | if (rcu_try_advance_all_cbs()) |
1772 | invoke_rcu_core(); | |
f1f399d1 | 1773 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
8bd93a2c PM |
1774 | } |
1775 | ||
c57afe80 | 1776 | /* |
98248a0e PM |
1777 | * Keep a running count of the number of non-lazy callbacks posted |
1778 | * on this CPU. This running counter (which is never decremented) allows | |
1779 | * rcu_prepare_for_idle() to detect when something out of the idle loop | |
1780 | * posts a callback, even if an equal number of callbacks are invoked. | |
1781 | * Of course, callbacks should only be posted from within a trace event | |
1782 | * designed to be called from idle or from within RCU_NONIDLE(). | |
c57afe80 PM |
1783 | */ |
1784 | static void rcu_idle_count_callbacks_posted(void) | |
1785 | { | |
5955f7ee | 1786 | __this_cpu_add(rcu_dynticks.nonlazy_posted, 1); |
c57afe80 PM |
1787 | } |
1788 | ||
b626c1b6 PM |
1789 | /* |
1790 | * Data for flushing lazy RCU callbacks at OOM time. | |
1791 | */ | |
1792 | static atomic_t oom_callback_count; | |
1793 | static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq); | |
1794 | ||
1795 | /* | |
1796 | * RCU OOM callback -- decrement the outstanding count and deliver the | |
1797 | * wake-up if we are the last one. | |
1798 | */ | |
1799 | static void rcu_oom_callback(struct rcu_head *rhp) | |
1800 | { | |
1801 | if (atomic_dec_and_test(&oom_callback_count)) | |
1802 | wake_up(&oom_callback_wq); | |
1803 | } | |
1804 | ||
1805 | /* | |
1806 | * Post an rcu_oom_notify callback on the current CPU if it has at | |
1807 | * least one lazy callback. This will unnecessarily post callbacks | |
1808 | * to CPUs that already have a non-lazy callback at the end of their | |
1809 | * callback list, but this is an infrequent operation, so accept some | |
1810 | * extra overhead to keep things simple. | |
1811 | */ | |
1812 | static void rcu_oom_notify_cpu(void *unused) | |
1813 | { | |
1814 | struct rcu_state *rsp; | |
1815 | struct rcu_data *rdp; | |
1816 | ||
1817 | for_each_rcu_flavor(rsp) { | |
fa07a58f | 1818 | rdp = raw_cpu_ptr(rsp->rda); |
b626c1b6 PM |
1819 | if (rdp->qlen_lazy != 0) { |
1820 | atomic_inc(&oom_callback_count); | |
1821 | rsp->call(&rdp->oom_head, rcu_oom_callback); | |
1822 | } | |
1823 | } | |
1824 | } | |
1825 | ||
1826 | /* | |
1827 | * If low on memory, ensure that each CPU has a non-lazy callback. | |
1828 | * This will wake up CPUs that have only lazy callbacks, in turn | |
1829 | * ensuring that they free up the corresponding memory in a timely manner. | |
1830 | * Because an uncertain amount of memory will be freed in some uncertain | |
1831 | * timeframe, we do not claim to have freed anything. | |
1832 | */ | |
1833 | static int rcu_oom_notify(struct notifier_block *self, | |
1834 | unsigned long notused, void *nfreed) | |
1835 | { | |
1836 | int cpu; | |
1837 | ||
1838 | /* Wait for callbacks from earlier instance to complete. */ | |
1839 | wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0); | |
78e4bc34 | 1840 | smp_mb(); /* Ensure callback reuse happens after callback invocation. */ |
b626c1b6 PM |
1841 | |
1842 | /* | |
1843 | * Prevent premature wakeup: ensure that all increments happen | |
1844 | * before there is a chance of the counter reaching zero. | |
1845 | */ | |
1846 | atomic_set(&oom_callback_count, 1); | |
1847 | ||
1848 | get_online_cpus(); | |
1849 | for_each_online_cpu(cpu) { | |
1850 | smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1); | |
1851 | cond_resched(); | |
1852 | } | |
1853 | put_online_cpus(); | |
1854 | ||
1855 | /* Unconditionally decrement: no need to wake ourselves up. */ | |
1856 | atomic_dec(&oom_callback_count); | |
1857 | ||
1858 | return NOTIFY_OK; | |
1859 | } | |
1860 | ||
1861 | static struct notifier_block rcu_oom_nb = { | |
1862 | .notifier_call = rcu_oom_notify | |
1863 | }; | |
1864 | ||
1865 | static int __init rcu_register_oom_notifier(void) | |
1866 | { | |
1867 | register_oom_notifier(&rcu_oom_nb); | |
1868 | return 0; | |
1869 | } | |
1870 | early_initcall(rcu_register_oom_notifier); | |
1871 | ||
8bd93a2c | 1872 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
a858af28 PM |
1873 | |
1874 | #ifdef CONFIG_RCU_CPU_STALL_INFO | |
1875 | ||
1876 | #ifdef CONFIG_RCU_FAST_NO_HZ | |
1877 | ||
1878 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1879 | { | |
5955f7ee | 1880 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
c0f4dfd4 | 1881 | unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap; |
a858af28 | 1882 | |
c0f4dfd4 PM |
1883 | sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c", |
1884 | rdtp->last_accelerate & 0xffff, jiffies & 0xffff, | |
1885 | ulong2long(nlpd), | |
1886 | rdtp->all_lazy ? 'L' : '.', | |
1887 | rdtp->tick_nohz_enabled_snap ? '.' : 'D'); | |
a858af28 PM |
1888 | } |
1889 | ||
1890 | #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1891 | ||
1892 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1893 | { | |
1c17e4d4 | 1894 | *cp = '\0'; |
a858af28 PM |
1895 | } |
1896 | ||
1897 | #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1898 | ||
1899 | /* Initiate the stall-info list. */ | |
1900 | static void print_cpu_stall_info_begin(void) | |
1901 | { | |
efc151c3 | 1902 | pr_cont("\n"); |
a858af28 PM |
1903 | } |
1904 | ||
1905 | /* | |
1906 | * Print out diagnostic information for the specified stalled CPU. | |
1907 | * | |
1908 | * If the specified CPU is aware of the current RCU grace period | |
1909 | * (flavor specified by rsp), then print the number of scheduling | |
1910 | * clock interrupts the CPU has taken during the time that it has | |
1911 | * been aware. Otherwise, print the number of RCU grace periods | |
1912 | * that this CPU is ignorant of, for example, "1" if the CPU was | |
1913 | * aware of the previous grace period. | |
1914 | * | |
1915 | * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info. | |
1916 | */ | |
1917 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) | |
1918 | { | |
1919 | char fast_no_hz[72]; | |
1920 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
1921 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
1922 | char *ticks_title; | |
1923 | unsigned long ticks_value; | |
1924 | ||
1925 | if (rsp->gpnum == rdp->gpnum) { | |
1926 | ticks_title = "ticks this GP"; | |
1927 | ticks_value = rdp->ticks_this_gp; | |
1928 | } else { | |
1929 | ticks_title = "GPs behind"; | |
1930 | ticks_value = rsp->gpnum - rdp->gpnum; | |
1931 | } | |
1932 | print_cpu_stall_fast_no_hz(fast_no_hz, cpu); | |
efc151c3 | 1933 | pr_err("\t%d: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u %s\n", |
a858af28 PM |
1934 | cpu, ticks_value, ticks_title, |
1935 | atomic_read(&rdtp->dynticks) & 0xfff, | |
1936 | rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting, | |
6231069b | 1937 | rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu), |
a858af28 PM |
1938 | fast_no_hz); |
1939 | } | |
1940 | ||
1941 | /* Terminate the stall-info list. */ | |
1942 | static void print_cpu_stall_info_end(void) | |
1943 | { | |
efc151c3 | 1944 | pr_err("\t"); |
a858af28 PM |
1945 | } |
1946 | ||
1947 | /* Zero ->ticks_this_gp for all flavors of RCU. */ | |
1948 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) | |
1949 | { | |
1950 | rdp->ticks_this_gp = 0; | |
6231069b | 1951 | rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id()); |
a858af28 PM |
1952 | } |
1953 | ||
1954 | /* Increment ->ticks_this_gp for all flavors of RCU. */ | |
1955 | static void increment_cpu_stall_ticks(void) | |
1956 | { | |
115f7a7c PM |
1957 | struct rcu_state *rsp; |
1958 | ||
1959 | for_each_rcu_flavor(rsp) | |
fa07a58f | 1960 | raw_cpu_inc(rsp->rda->ticks_this_gp); |
a858af28 PM |
1961 | } |
1962 | ||
1963 | #else /* #ifdef CONFIG_RCU_CPU_STALL_INFO */ | |
1964 | ||
1965 | static void print_cpu_stall_info_begin(void) | |
1966 | { | |
efc151c3 | 1967 | pr_cont(" {"); |
a858af28 PM |
1968 | } |
1969 | ||
1970 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) | |
1971 | { | |
efc151c3 | 1972 | pr_cont(" %d", cpu); |
a858af28 PM |
1973 | } |
1974 | ||
1975 | static void print_cpu_stall_info_end(void) | |
1976 | { | |
efc151c3 | 1977 | pr_cont("} "); |
a858af28 PM |
1978 | } |
1979 | ||
1980 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) | |
1981 | { | |
1982 | } | |
1983 | ||
1984 | static void increment_cpu_stall_ticks(void) | |
1985 | { | |
1986 | } | |
1987 | ||
1988 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_INFO */ | |
3fbfbf7a PM |
1989 | |
1990 | #ifdef CONFIG_RCU_NOCB_CPU | |
1991 | ||
1992 | /* | |
1993 | * Offload callback processing from the boot-time-specified set of CPUs | |
1994 | * specified by rcu_nocb_mask. For each CPU in the set, there is a | |
1995 | * kthread created that pulls the callbacks from the corresponding CPU, | |
1996 | * waits for a grace period to elapse, and invokes the callbacks. | |
1997 | * The no-CBs CPUs do a wake_up() on their kthread when they insert | |
1998 | * a callback into any empty list, unless the rcu_nocb_poll boot parameter | |
1999 | * has been specified, in which case each kthread actively polls its | |
2000 | * CPU. (Which isn't so great for energy efficiency, but which does | |
2001 | * reduce RCU's overhead on that CPU.) | |
2002 | * | |
2003 | * This is intended to be used in conjunction with Frederic Weisbecker's | |
2004 | * adaptive-idle work, which would seriously reduce OS jitter on CPUs | |
2005 | * running CPU-bound user-mode computations. | |
2006 | * | |
2007 | * Offloading of callback processing could also in theory be used as | |
2008 | * an energy-efficiency measure because CPUs with no RCU callbacks | |
2009 | * queued are more aggressive about entering dyntick-idle mode. | |
2010 | */ | |
2011 | ||
2012 | ||
2013 | /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */ | |
2014 | static int __init rcu_nocb_setup(char *str) | |
2015 | { | |
2016 | alloc_bootmem_cpumask_var(&rcu_nocb_mask); | |
2017 | have_rcu_nocb_mask = true; | |
2018 | cpulist_parse(str, rcu_nocb_mask); | |
2019 | return 1; | |
2020 | } | |
2021 | __setup("rcu_nocbs=", rcu_nocb_setup); | |
2022 | ||
1b0048a4 PG |
2023 | static int __init parse_rcu_nocb_poll(char *arg) |
2024 | { | |
2025 | rcu_nocb_poll = 1; | |
2026 | return 0; | |
2027 | } | |
2028 | early_param("rcu_nocb_poll", parse_rcu_nocb_poll); | |
2029 | ||
dae6e64d | 2030 | /* |
0446be48 PM |
2031 | * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended |
2032 | * grace period. | |
dae6e64d | 2033 | */ |
0446be48 | 2034 | static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
dae6e64d | 2035 | { |
0446be48 | 2036 | wake_up_all(&rnp->nocb_gp_wq[rnp->completed & 0x1]); |
dae6e64d PM |
2037 | } |
2038 | ||
2039 | /* | |
8b425aa8 | 2040 | * Set the root rcu_node structure's ->need_future_gp field |
dae6e64d PM |
2041 | * based on the sum of those of all rcu_node structures. This does |
2042 | * double-count the root rcu_node structure's requests, but this | |
2043 | * is necessary to handle the possibility of a rcu_nocb_kthread() | |
2044 | * having awakened during the time that the rcu_node structures | |
2045 | * were being updated for the end of the previous grace period. | |
34ed6246 | 2046 | */ |
dae6e64d PM |
2047 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
2048 | { | |
8b425aa8 | 2049 | rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq; |
dae6e64d PM |
2050 | } |
2051 | ||
2052 | static void rcu_init_one_nocb(struct rcu_node *rnp) | |
34ed6246 | 2053 | { |
dae6e64d PM |
2054 | init_waitqueue_head(&rnp->nocb_gp_wq[0]); |
2055 | init_waitqueue_head(&rnp->nocb_gp_wq[1]); | |
34ed6246 PM |
2056 | } |
2057 | ||
2f33b512 | 2058 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
24342c96 | 2059 | /* Is the specified CPU a no-CBs CPU? */ |
d1e43fa5 | 2060 | bool rcu_is_nocb_cpu(int cpu) |
3fbfbf7a PM |
2061 | { |
2062 | if (have_rcu_nocb_mask) | |
2063 | return cpumask_test_cpu(cpu, rcu_nocb_mask); | |
2064 | return false; | |
2065 | } | |
2f33b512 | 2066 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
3fbfbf7a | 2067 | |
fbce7497 PM |
2068 | /* |
2069 | * Kick the leader kthread for this NOCB group. | |
2070 | */ | |
2071 | static void wake_nocb_leader(struct rcu_data *rdp, bool force) | |
2072 | { | |
2073 | struct rcu_data *rdp_leader = rdp->nocb_leader; | |
2074 | ||
2075 | if (!ACCESS_ONCE(rdp_leader->nocb_kthread)) | |
2076 | return; | |
11ed7f93 | 2077 | if (ACCESS_ONCE(rdp_leader->nocb_leader_sleep) || force) { |
fbce7497 | 2078 | /* Prior xchg orders against prior callback enqueue. */ |
11ed7f93 | 2079 | ACCESS_ONCE(rdp_leader->nocb_leader_sleep) = false; |
fbce7497 PM |
2080 | wake_up(&rdp_leader->nocb_wq); |
2081 | } | |
2082 | } | |
2083 | ||
3fbfbf7a PM |
2084 | /* |
2085 | * Enqueue the specified string of rcu_head structures onto the specified | |
2086 | * CPU's no-CBs lists. The CPU is specified by rdp, the head of the | |
2087 | * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy | |
2088 | * counts are supplied by rhcount and rhcount_lazy. | |
2089 | * | |
2090 | * If warranted, also wake up the kthread servicing this CPUs queues. | |
2091 | */ | |
2092 | static void __call_rcu_nocb_enqueue(struct rcu_data *rdp, | |
2093 | struct rcu_head *rhp, | |
2094 | struct rcu_head **rhtp, | |
96d3fd0d PM |
2095 | int rhcount, int rhcount_lazy, |
2096 | unsigned long flags) | |
3fbfbf7a PM |
2097 | { |
2098 | int len; | |
2099 | struct rcu_head **old_rhpp; | |
2100 | struct task_struct *t; | |
2101 | ||
2102 | /* Enqueue the callback on the nocb list and update counts. */ | |
2103 | old_rhpp = xchg(&rdp->nocb_tail, rhtp); | |
2104 | ACCESS_ONCE(*old_rhpp) = rhp; | |
2105 | atomic_long_add(rhcount, &rdp->nocb_q_count); | |
2106 | atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy); | |
2107 | ||
2108 | /* If we are not being polled and there is a kthread, awaken it ... */ | |
2109 | t = ACCESS_ONCE(rdp->nocb_kthread); | |
25e03a74 | 2110 | if (rcu_nocb_poll || !t) { |
9261dd0d PM |
2111 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2112 | TPS("WakeNotPoll")); | |
3fbfbf7a | 2113 | return; |
9261dd0d | 2114 | } |
3fbfbf7a PM |
2115 | len = atomic_long_read(&rdp->nocb_q_count); |
2116 | if (old_rhpp == &rdp->nocb_head) { | |
96d3fd0d | 2117 | if (!irqs_disabled_flags(flags)) { |
fbce7497 PM |
2118 | /* ... if queue was empty ... */ |
2119 | wake_nocb_leader(rdp, false); | |
96d3fd0d PM |
2120 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2121 | TPS("WakeEmpty")); | |
2122 | } else { | |
2123 | rdp->nocb_defer_wakeup = true; | |
2124 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2125 | TPS("WakeEmptyIsDeferred")); | |
2126 | } | |
3fbfbf7a PM |
2127 | rdp->qlen_last_fqs_check = 0; |
2128 | } else if (len > rdp->qlen_last_fqs_check + qhimark) { | |
fbce7497 PM |
2129 | /* ... or if many callbacks queued. */ |
2130 | wake_nocb_leader(rdp, true); | |
3fbfbf7a | 2131 | rdp->qlen_last_fqs_check = LONG_MAX / 2; |
9261dd0d PM |
2132 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeOvf")); |
2133 | } else { | |
2134 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot")); | |
3fbfbf7a PM |
2135 | } |
2136 | return; | |
2137 | } | |
2138 | ||
2139 | /* | |
2140 | * This is a helper for __call_rcu(), which invokes this when the normal | |
2141 | * callback queue is inoperable. If this is not a no-CBs CPU, this | |
2142 | * function returns failure back to __call_rcu(), which can complain | |
2143 | * appropriately. | |
2144 | * | |
2145 | * Otherwise, this function queues the callback where the corresponding | |
2146 | * "rcuo" kthread can find it. | |
2147 | */ | |
2148 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, | |
96d3fd0d | 2149 | bool lazy, unsigned long flags) |
3fbfbf7a PM |
2150 | { |
2151 | ||
d1e43fa5 | 2152 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
3fbfbf7a | 2153 | return 0; |
96d3fd0d | 2154 | __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags); |
21e7a608 PM |
2155 | if (__is_kfree_rcu_offset((unsigned long)rhp->func)) |
2156 | trace_rcu_kfree_callback(rdp->rsp->name, rhp, | |
2157 | (unsigned long)rhp->func, | |
756cbf6b PM |
2158 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
2159 | -atomic_long_read(&rdp->nocb_q_count)); | |
21e7a608 PM |
2160 | else |
2161 | trace_rcu_callback(rdp->rsp->name, rhp, | |
756cbf6b PM |
2162 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
2163 | -atomic_long_read(&rdp->nocb_q_count)); | |
3fbfbf7a PM |
2164 | return 1; |
2165 | } | |
2166 | ||
2167 | /* | |
2168 | * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is | |
2169 | * not a no-CBs CPU. | |
2170 | */ | |
2171 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
2172 | struct rcu_data *rdp, |
2173 | unsigned long flags) | |
3fbfbf7a PM |
2174 | { |
2175 | long ql = rsp->qlen; | |
2176 | long qll = rsp->qlen_lazy; | |
2177 | ||
2178 | /* If this is not a no-CBs CPU, tell the caller to do it the old way. */ | |
d1e43fa5 | 2179 | if (!rcu_is_nocb_cpu(smp_processor_id())) |
3fbfbf7a PM |
2180 | return 0; |
2181 | rsp->qlen = 0; | |
2182 | rsp->qlen_lazy = 0; | |
2183 | ||
2184 | /* First, enqueue the donelist, if any. This preserves CB ordering. */ | |
2185 | if (rsp->orphan_donelist != NULL) { | |
2186 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist, | |
96d3fd0d | 2187 | rsp->orphan_donetail, ql, qll, flags); |
3fbfbf7a PM |
2188 | ql = qll = 0; |
2189 | rsp->orphan_donelist = NULL; | |
2190 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2191 | } | |
2192 | if (rsp->orphan_nxtlist != NULL) { | |
2193 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist, | |
96d3fd0d | 2194 | rsp->orphan_nxttail, ql, qll, flags); |
3fbfbf7a PM |
2195 | ql = qll = 0; |
2196 | rsp->orphan_nxtlist = NULL; | |
2197 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2198 | } | |
2199 | return 1; | |
2200 | } | |
2201 | ||
2202 | /* | |
34ed6246 PM |
2203 | * If necessary, kick off a new grace period, and either way wait |
2204 | * for a subsequent grace period to complete. | |
3fbfbf7a | 2205 | */ |
34ed6246 | 2206 | static void rcu_nocb_wait_gp(struct rcu_data *rdp) |
3fbfbf7a | 2207 | { |
34ed6246 | 2208 | unsigned long c; |
dae6e64d | 2209 | bool d; |
34ed6246 | 2210 | unsigned long flags; |
48a7639c | 2211 | bool needwake; |
34ed6246 PM |
2212 | struct rcu_node *rnp = rdp->mynode; |
2213 | ||
2214 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
6303b9c8 | 2215 | smp_mb__after_unlock_lock(); |
48a7639c | 2216 | needwake = rcu_start_future_gp(rnp, rdp, &c); |
0446be48 | 2217 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
48a7639c PM |
2218 | if (needwake) |
2219 | rcu_gp_kthread_wake(rdp->rsp); | |
3fbfbf7a PM |
2220 | |
2221 | /* | |
34ed6246 PM |
2222 | * Wait for the grace period. Do so interruptibly to avoid messing |
2223 | * up the load average. | |
3fbfbf7a | 2224 | */ |
f7f7bac9 | 2225 | trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait")); |
34ed6246 | 2226 | for (;;) { |
dae6e64d PM |
2227 | wait_event_interruptible( |
2228 | rnp->nocb_gp_wq[c & 0x1], | |
2229 | (d = ULONG_CMP_GE(ACCESS_ONCE(rnp->completed), c))); | |
2230 | if (likely(d)) | |
34ed6246 | 2231 | break; |
dae6e64d | 2232 | flush_signals(current); |
f7f7bac9 | 2233 | trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait")); |
34ed6246 | 2234 | } |
f7f7bac9 | 2235 | trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait")); |
34ed6246 | 2236 | smp_mb(); /* Ensure that CB invocation happens after GP end. */ |
3fbfbf7a PM |
2237 | } |
2238 | ||
fbce7497 PM |
2239 | /* |
2240 | * Leaders come here to wait for additional callbacks to show up. | |
2241 | * This function does not return until callbacks appear. | |
2242 | */ | |
2243 | static void nocb_leader_wait(struct rcu_data *my_rdp) | |
2244 | { | |
2245 | bool firsttime = true; | |
2246 | bool gotcbs; | |
2247 | struct rcu_data *rdp; | |
2248 | struct rcu_head **tail; | |
2249 | ||
2250 | wait_again: | |
2251 | ||
2252 | /* Wait for callbacks to appear. */ | |
2253 | if (!rcu_nocb_poll) { | |
2254 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep"); | |
2255 | wait_event_interruptible(my_rdp->nocb_wq, | |
11ed7f93 | 2256 | !ACCESS_ONCE(my_rdp->nocb_leader_sleep)); |
fbce7497 PM |
2257 | /* Memory barrier handled by smp_mb() calls below and repoll. */ |
2258 | } else if (firsttime) { | |
2259 | firsttime = false; /* Don't drown trace log with "Poll"! */ | |
2260 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll"); | |
2261 | } | |
2262 | ||
2263 | /* | |
2264 | * Each pass through the following loop checks a follower for CBs. | |
2265 | * We are our own first follower. Any CBs found are moved to | |
2266 | * nocb_gp_head, where they await a grace period. | |
2267 | */ | |
2268 | gotcbs = false; | |
2269 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
2270 | rdp->nocb_gp_head = ACCESS_ONCE(rdp->nocb_head); | |
2271 | if (!rdp->nocb_gp_head) | |
2272 | continue; /* No CBs here, try next follower. */ | |
2273 | ||
2274 | /* Move callbacks to wait-for-GP list, which is empty. */ | |
2275 | ACCESS_ONCE(rdp->nocb_head) = NULL; | |
2276 | rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head); | |
2277 | rdp->nocb_gp_count = atomic_long_xchg(&rdp->nocb_q_count, 0); | |
2278 | rdp->nocb_gp_count_lazy = | |
2279 | atomic_long_xchg(&rdp->nocb_q_count_lazy, 0); | |
2280 | gotcbs = true; | |
2281 | } | |
2282 | ||
2283 | /* | |
2284 | * If there were no callbacks, sleep a bit, rescan after a | |
2285 | * memory barrier, and go retry. | |
2286 | */ | |
2287 | if (unlikely(!gotcbs)) { | |
2288 | if (!rcu_nocb_poll) | |
2289 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, | |
2290 | "WokeEmpty"); | |
2291 | flush_signals(current); | |
2292 | schedule_timeout_interruptible(1); | |
2293 | ||
2294 | /* Rescan in case we were a victim of memory ordering. */ | |
11ed7f93 PK |
2295 | my_rdp->nocb_leader_sleep = true; |
2296 | smp_mb(); /* Ensure _sleep true before scan. */ | |
fbce7497 PM |
2297 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) |
2298 | if (ACCESS_ONCE(rdp->nocb_head)) { | |
2299 | /* Found CB, so short-circuit next wait. */ | |
11ed7f93 | 2300 | my_rdp->nocb_leader_sleep = false; |
fbce7497 PM |
2301 | break; |
2302 | } | |
2303 | goto wait_again; | |
2304 | } | |
2305 | ||
2306 | /* Wait for one grace period. */ | |
2307 | rcu_nocb_wait_gp(my_rdp); | |
2308 | ||
2309 | /* | |
11ed7f93 PK |
2310 | * We left ->nocb_leader_sleep unset to reduce cache thrashing. |
2311 | * We set it now, but recheck for new callbacks while | |
fbce7497 PM |
2312 | * traversing our follower list. |
2313 | */ | |
11ed7f93 PK |
2314 | my_rdp->nocb_leader_sleep = true; |
2315 | smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */ | |
fbce7497 PM |
2316 | |
2317 | /* Each pass through the following loop wakes a follower, if needed. */ | |
2318 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
2319 | if (ACCESS_ONCE(rdp->nocb_head)) | |
11ed7f93 | 2320 | my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/ |
fbce7497 PM |
2321 | if (!rdp->nocb_gp_head) |
2322 | continue; /* No CBs, so no need to wake follower. */ | |
2323 | ||
2324 | /* Append callbacks to follower's "done" list. */ | |
2325 | tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail); | |
2326 | *tail = rdp->nocb_gp_head; | |
2327 | atomic_long_add(rdp->nocb_gp_count, &rdp->nocb_follower_count); | |
2328 | atomic_long_add(rdp->nocb_gp_count_lazy, | |
2329 | &rdp->nocb_follower_count_lazy); | |
2330 | if (rdp != my_rdp && tail == &rdp->nocb_follower_head) { | |
2331 | /* | |
2332 | * List was empty, wake up the follower. | |
2333 | * Memory barriers supplied by atomic_long_add(). | |
2334 | */ | |
2335 | wake_up(&rdp->nocb_wq); | |
2336 | } | |
2337 | } | |
2338 | ||
2339 | /* If we (the leader) don't have CBs, go wait some more. */ | |
2340 | if (!my_rdp->nocb_follower_head) | |
2341 | goto wait_again; | |
2342 | } | |
2343 | ||
2344 | /* | |
2345 | * Followers come here to wait for additional callbacks to show up. | |
2346 | * This function does not return until callbacks appear. | |
2347 | */ | |
2348 | static void nocb_follower_wait(struct rcu_data *rdp) | |
2349 | { | |
2350 | bool firsttime = true; | |
2351 | ||
2352 | for (;;) { | |
2353 | if (!rcu_nocb_poll) { | |
2354 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2355 | "FollowerSleep"); | |
2356 | wait_event_interruptible(rdp->nocb_wq, | |
2357 | ACCESS_ONCE(rdp->nocb_follower_head)); | |
2358 | } else if (firsttime) { | |
2359 | /* Don't drown trace log with "Poll"! */ | |
2360 | firsttime = false; | |
2361 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll"); | |
2362 | } | |
2363 | if (smp_load_acquire(&rdp->nocb_follower_head)) { | |
2364 | /* ^^^ Ensure CB invocation follows _head test. */ | |
2365 | return; | |
2366 | } | |
2367 | if (!rcu_nocb_poll) | |
2368 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2369 | "WokeEmpty"); | |
2370 | flush_signals(current); | |
2371 | schedule_timeout_interruptible(1); | |
2372 | } | |
2373 | } | |
2374 | ||
3fbfbf7a PM |
2375 | /* |
2376 | * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes | |
fbce7497 PM |
2377 | * callbacks queued by the corresponding no-CBs CPU, however, there is |
2378 | * an optional leader-follower relationship so that the grace-period | |
2379 | * kthreads don't have to do quite so many wakeups. | |
3fbfbf7a PM |
2380 | */ |
2381 | static int rcu_nocb_kthread(void *arg) | |
2382 | { | |
2383 | int c, cl; | |
2384 | struct rcu_head *list; | |
2385 | struct rcu_head *next; | |
2386 | struct rcu_head **tail; | |
2387 | struct rcu_data *rdp = arg; | |
2388 | ||
2389 | /* Each pass through this loop invokes one batch of callbacks */ | |
2390 | for (;;) { | |
fbce7497 PM |
2391 | /* Wait for callbacks. */ |
2392 | if (rdp->nocb_leader == rdp) | |
2393 | nocb_leader_wait(rdp); | |
2394 | else | |
2395 | nocb_follower_wait(rdp); | |
2396 | ||
2397 | /* Pull the ready-to-invoke callbacks onto local list. */ | |
2398 | list = ACCESS_ONCE(rdp->nocb_follower_head); | |
2399 | BUG_ON(!list); | |
2400 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty"); | |
2401 | ACCESS_ONCE(rdp->nocb_follower_head) = NULL; | |
2402 | tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head); | |
2403 | c = atomic_long_xchg(&rdp->nocb_follower_count, 0); | |
2404 | cl = atomic_long_xchg(&rdp->nocb_follower_count_lazy, 0); | |
2405 | rdp->nocb_p_count += c; | |
2406 | rdp->nocb_p_count_lazy += cl; | |
3fbfbf7a PM |
2407 | |
2408 | /* Each pass through the following loop invokes a callback. */ | |
2409 | trace_rcu_batch_start(rdp->rsp->name, cl, c, -1); | |
2410 | c = cl = 0; | |
2411 | while (list) { | |
2412 | next = list->next; | |
2413 | /* Wait for enqueuing to complete, if needed. */ | |
2414 | while (next == NULL && &list->next != tail) { | |
69a79bb1 PM |
2415 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2416 | TPS("WaitQueue")); | |
3fbfbf7a | 2417 | schedule_timeout_interruptible(1); |
69a79bb1 PM |
2418 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2419 | TPS("WokeQueue")); | |
3fbfbf7a PM |
2420 | next = list->next; |
2421 | } | |
2422 | debug_rcu_head_unqueue(list); | |
2423 | local_bh_disable(); | |
2424 | if (__rcu_reclaim(rdp->rsp->name, list)) | |
2425 | cl++; | |
2426 | c++; | |
2427 | local_bh_enable(); | |
2428 | list = next; | |
2429 | } | |
2430 | trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); | |
2431 | ACCESS_ONCE(rdp->nocb_p_count) -= c; | |
2432 | ACCESS_ONCE(rdp->nocb_p_count_lazy) -= cl; | |
c635a4e1 | 2433 | rdp->n_nocbs_invoked += c; |
3fbfbf7a PM |
2434 | } |
2435 | return 0; | |
2436 | } | |
2437 | ||
96d3fd0d PM |
2438 | /* Is a deferred wakeup of rcu_nocb_kthread() required? */ |
2439 | static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) | |
2440 | { | |
2441 | return ACCESS_ONCE(rdp->nocb_defer_wakeup); | |
2442 | } | |
2443 | ||
2444 | /* Do a deferred wakeup of rcu_nocb_kthread(). */ | |
2445 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2446 | { | |
2447 | if (!rcu_nocb_need_deferred_wakeup(rdp)) | |
2448 | return; | |
2449 | ACCESS_ONCE(rdp->nocb_defer_wakeup) = false; | |
fbce7497 | 2450 | wake_nocb_leader(rdp, false); |
96d3fd0d PM |
2451 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWakeEmpty")); |
2452 | } | |
2453 | ||
3fbfbf7a PM |
2454 | /* Initialize per-rcu_data variables for no-CBs CPUs. */ |
2455 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) | |
2456 | { | |
2457 | rdp->nocb_tail = &rdp->nocb_head; | |
2458 | init_waitqueue_head(&rdp->nocb_wq); | |
fbce7497 | 2459 | rdp->nocb_follower_tail = &rdp->nocb_follower_head; |
3fbfbf7a PM |
2460 | } |
2461 | ||
fbce7497 PM |
2462 | /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */ |
2463 | static int rcu_nocb_leader_stride = -1; | |
2464 | module_param(rcu_nocb_leader_stride, int, 0444); | |
2465 | ||
2466 | /* | |
2467 | * Create a kthread for each RCU flavor for each no-CBs CPU. | |
2468 | * Also initialize leader-follower relationships. | |
2469 | */ | |
3fbfbf7a PM |
2470 | static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp) |
2471 | { | |
2472 | int cpu; | |
fbce7497 PM |
2473 | int ls = rcu_nocb_leader_stride; |
2474 | int nl = 0; /* Next leader. */ | |
3fbfbf7a | 2475 | struct rcu_data *rdp; |
fbce7497 PM |
2476 | struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */ |
2477 | struct rcu_data *rdp_prev = NULL; | |
3fbfbf7a PM |
2478 | struct task_struct *t; |
2479 | ||
2480 | if (rcu_nocb_mask == NULL) | |
2481 | return; | |
187497fa PM |
2482 | #if defined(CONFIG_NO_HZ_FULL) && !defined(CONFIG_NO_HZ_FULL_ALL) |
2483 | if (tick_nohz_full_running) | |
2484 | cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask); | |
2485 | #endif /* #if defined(CONFIG_NO_HZ_FULL) && !defined(CONFIG_NO_HZ_FULL_ALL) */ | |
fbce7497 PM |
2486 | if (ls == -1) { |
2487 | ls = int_sqrt(nr_cpu_ids); | |
2488 | rcu_nocb_leader_stride = ls; | |
2489 | } | |
2490 | ||
2491 | /* | |
2492 | * Each pass through this loop sets up one rcu_data structure and | |
2493 | * spawns one rcu_nocb_kthread(). | |
2494 | */ | |
3fbfbf7a PM |
2495 | for_each_cpu(cpu, rcu_nocb_mask) { |
2496 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
fbce7497 PM |
2497 | if (rdp->cpu >= nl) { |
2498 | /* New leader, set up for followers & next leader. */ | |
2499 | nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls; | |
2500 | rdp->nocb_leader = rdp; | |
2501 | rdp_leader = rdp; | |
2502 | } else { | |
2503 | /* Another follower, link to previous leader. */ | |
2504 | rdp->nocb_leader = rdp_leader; | |
2505 | rdp_prev->nocb_next_follower = rdp; | |
2506 | } | |
2507 | rdp_prev = rdp; | |
2508 | ||
2509 | /* Spawn the kthread for this CPU. */ | |
a4889858 PM |
2510 | t = kthread_run(rcu_nocb_kthread, rdp, |
2511 | "rcuo%c/%d", rsp->abbr, cpu); | |
3fbfbf7a PM |
2512 | BUG_ON(IS_ERR(t)); |
2513 | ACCESS_ONCE(rdp->nocb_kthread) = t; | |
2514 | } | |
2515 | } | |
2516 | ||
2517 | /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */ | |
34ed6246 | 2518 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a PM |
2519 | { |
2520 | if (rcu_nocb_mask == NULL || | |
2521 | !cpumask_test_cpu(rdp->cpu, rcu_nocb_mask)) | |
34ed6246 | 2522 | return false; |
3fbfbf7a | 2523 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
34ed6246 | 2524 | return true; |
3fbfbf7a PM |
2525 | } |
2526 | ||
34ed6246 PM |
2527 | #else /* #ifdef CONFIG_RCU_NOCB_CPU */ |
2528 | ||
0446be48 | 2529 | static void rcu_nocb_gp_cleanup(struct rcu_state *rsp, struct rcu_node *rnp) |
3fbfbf7a | 2530 | { |
3fbfbf7a PM |
2531 | } |
2532 | ||
dae6e64d PM |
2533 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
2534 | { | |
2535 | } | |
2536 | ||
2537 | static void rcu_init_one_nocb(struct rcu_node *rnp) | |
2538 | { | |
2539 | } | |
3fbfbf7a | 2540 | |
3fbfbf7a | 2541 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
96d3fd0d | 2542 | bool lazy, unsigned long flags) |
3fbfbf7a PM |
2543 | { |
2544 | return 0; | |
2545 | } | |
2546 | ||
2547 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
2548 | struct rcu_data *rdp, |
2549 | unsigned long flags) | |
3fbfbf7a PM |
2550 | { |
2551 | return 0; | |
2552 | } | |
2553 | ||
3fbfbf7a PM |
2554 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
2555 | { | |
2556 | } | |
2557 | ||
96d3fd0d PM |
2558 | static bool rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
2559 | { | |
2560 | return false; | |
2561 | } | |
2562 | ||
2563 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2564 | { | |
2565 | } | |
2566 | ||
3fbfbf7a PM |
2567 | static void __init rcu_spawn_nocb_kthreads(struct rcu_state *rsp) |
2568 | { | |
2569 | } | |
2570 | ||
34ed6246 | 2571 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2572 | { |
34ed6246 | 2573 | return false; |
3fbfbf7a PM |
2574 | } |
2575 | ||
2576 | #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */ | |
65d798f0 PM |
2577 | |
2578 | /* | |
2579 | * An adaptive-ticks CPU can potentially execute in kernel mode for an | |
2580 | * arbitrarily long period of time with the scheduling-clock tick turned | |
2581 | * off. RCU will be paying attention to this CPU because it is in the | |
2582 | * kernel, but the CPU cannot be guaranteed to be executing the RCU state | |
2583 | * machine because the scheduling-clock tick has been disabled. Therefore, | |
2584 | * if an adaptive-ticks CPU is failing to respond to the current grace | |
2585 | * period and has not be idle from an RCU perspective, kick it. | |
2586 | */ | |
4a81e832 | 2587 | static void __maybe_unused rcu_kick_nohz_cpu(int cpu) |
65d798f0 PM |
2588 | { |
2589 | #ifdef CONFIG_NO_HZ_FULL | |
2590 | if (tick_nohz_full_cpu(cpu)) | |
2591 | smp_send_reschedule(cpu); | |
2592 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ | |
2593 | } | |
2333210b PM |
2594 | |
2595 | ||
2596 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
2597 | ||
d4bd54fb PM |
2598 | /* |
2599 | * Define RCU flavor that holds sysidle state. This needs to be the | |
2600 | * most active flavor of RCU. | |
2601 | */ | |
2602 | #ifdef CONFIG_PREEMPT_RCU | |
0edd1b17 | 2603 | static struct rcu_state *rcu_sysidle_state = &rcu_preempt_state; |
d4bd54fb | 2604 | #else /* #ifdef CONFIG_PREEMPT_RCU */ |
0edd1b17 | 2605 | static struct rcu_state *rcu_sysidle_state = &rcu_sched_state; |
d4bd54fb PM |
2606 | #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ |
2607 | ||
0edd1b17 | 2608 | static int full_sysidle_state; /* Current system-idle state. */ |
d4bd54fb PM |
2609 | #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */ |
2610 | #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */ | |
2611 | #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */ | |
2612 | #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */ | |
2613 | #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */ | |
2614 | ||
eb348b89 PM |
2615 | /* |
2616 | * Invoked to note exit from irq or task transition to idle. Note that | |
2617 | * usermode execution does -not- count as idle here! After all, we want | |
2618 | * to detect full-system idle states, not RCU quiescent states and grace | |
2619 | * periods. The caller must have disabled interrupts. | |
2620 | */ | |
2621 | static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq) | |
2622 | { | |
2623 | unsigned long j; | |
2624 | ||
2625 | /* Adjust nesting, check for fully idle. */ | |
2626 | if (irq) { | |
2627 | rdtp->dynticks_idle_nesting--; | |
2628 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2629 | if (rdtp->dynticks_idle_nesting != 0) | |
2630 | return; /* Still not fully idle. */ | |
2631 | } else { | |
2632 | if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) == | |
2633 | DYNTICK_TASK_NEST_VALUE) { | |
2634 | rdtp->dynticks_idle_nesting = 0; | |
2635 | } else { | |
2636 | rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE; | |
2637 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2638 | return; /* Still not fully idle. */ | |
2639 | } | |
2640 | } | |
2641 | ||
2642 | /* Record start of fully idle period. */ | |
2643 | j = jiffies; | |
2644 | ACCESS_ONCE(rdtp->dynticks_idle_jiffies) = j; | |
4e857c58 | 2645 | smp_mb__before_atomic(); |
eb348b89 | 2646 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2647 | smp_mb__after_atomic(); |
eb348b89 PM |
2648 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1); |
2649 | } | |
2650 | ||
0edd1b17 PM |
2651 | /* |
2652 | * Unconditionally force exit from full system-idle state. This is | |
2653 | * invoked when a normal CPU exits idle, but must be called separately | |
2654 | * for the timekeeping CPU (tick_do_timer_cpu). The reason for this | |
2655 | * is that the timekeeping CPU is permitted to take scheduling-clock | |
2656 | * interrupts while the system is in system-idle state, and of course | |
2657 | * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock | |
2658 | * interrupt from any other type of interrupt. | |
2659 | */ | |
2660 | void rcu_sysidle_force_exit(void) | |
2661 | { | |
2662 | int oldstate = ACCESS_ONCE(full_sysidle_state); | |
2663 | int newoldstate; | |
2664 | ||
2665 | /* | |
2666 | * Each pass through the following loop attempts to exit full | |
2667 | * system-idle state. If contention proves to be a problem, | |
2668 | * a trylock-based contention tree could be used here. | |
2669 | */ | |
2670 | while (oldstate > RCU_SYSIDLE_SHORT) { | |
2671 | newoldstate = cmpxchg(&full_sysidle_state, | |
2672 | oldstate, RCU_SYSIDLE_NOT); | |
2673 | if (oldstate == newoldstate && | |
2674 | oldstate == RCU_SYSIDLE_FULL_NOTED) { | |
2675 | rcu_kick_nohz_cpu(tick_do_timer_cpu); | |
2676 | return; /* We cleared it, done! */ | |
2677 | } | |
2678 | oldstate = newoldstate; | |
2679 | } | |
2680 | smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */ | |
2681 | } | |
2682 | ||
eb348b89 PM |
2683 | /* |
2684 | * Invoked to note entry to irq or task transition from idle. Note that | |
2685 | * usermode execution does -not- count as idle here! The caller must | |
2686 | * have disabled interrupts. | |
2687 | */ | |
2688 | static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq) | |
2689 | { | |
2690 | /* Adjust nesting, check for already non-idle. */ | |
2691 | if (irq) { | |
2692 | rdtp->dynticks_idle_nesting++; | |
2693 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2694 | if (rdtp->dynticks_idle_nesting != 1) | |
2695 | return; /* Already non-idle. */ | |
2696 | } else { | |
2697 | /* | |
2698 | * Allow for irq misnesting. Yes, it really is possible | |
2699 | * to enter an irq handler then never leave it, and maybe | |
2700 | * also vice versa. Handle both possibilities. | |
2701 | */ | |
2702 | if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) { | |
2703 | rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE; | |
2704 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2705 | return; /* Already non-idle. */ | |
2706 | } else { | |
2707 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE; | |
2708 | } | |
2709 | } | |
2710 | ||
2711 | /* Record end of idle period. */ | |
4e857c58 | 2712 | smp_mb__before_atomic(); |
eb348b89 | 2713 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2714 | smp_mb__after_atomic(); |
eb348b89 | 2715 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1)); |
0edd1b17 PM |
2716 | |
2717 | /* | |
2718 | * If we are the timekeeping CPU, we are permitted to be non-idle | |
2719 | * during a system-idle state. This must be the case, because | |
2720 | * the timekeeping CPU has to take scheduling-clock interrupts | |
2721 | * during the time that the system is transitioning to full | |
2722 | * system-idle state. This means that the timekeeping CPU must | |
2723 | * invoke rcu_sysidle_force_exit() directly if it does anything | |
2724 | * more than take a scheduling-clock interrupt. | |
2725 | */ | |
2726 | if (smp_processor_id() == tick_do_timer_cpu) | |
2727 | return; | |
2728 | ||
2729 | /* Update system-idle state: We are clearly no longer fully idle! */ | |
2730 | rcu_sysidle_force_exit(); | |
2731 | } | |
2732 | ||
2733 | /* | |
2734 | * Check to see if the current CPU is idle. Note that usermode execution | |
2735 | * does not count as idle. The caller must have disabled interrupts. | |
2736 | */ | |
2737 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, | |
2738 | unsigned long *maxj) | |
2739 | { | |
2740 | int cur; | |
2741 | unsigned long j; | |
2742 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
2743 | ||
2744 | /* | |
2745 | * If some other CPU has already reported non-idle, if this is | |
2746 | * not the flavor of RCU that tracks sysidle state, or if this | |
2747 | * is an offline or the timekeeping CPU, nothing to do. | |
2748 | */ | |
2749 | if (!*isidle || rdp->rsp != rcu_sysidle_state || | |
2750 | cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu) | |
2751 | return; | |
eb75767b PM |
2752 | if (rcu_gp_in_progress(rdp->rsp)) |
2753 | WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu); | |
0edd1b17 PM |
2754 | |
2755 | /* Pick up current idle and NMI-nesting counter and check. */ | |
2756 | cur = atomic_read(&rdtp->dynticks_idle); | |
2757 | if (cur & 0x1) { | |
2758 | *isidle = false; /* We are not idle! */ | |
2759 | return; | |
2760 | } | |
2761 | smp_mb(); /* Read counters before timestamps. */ | |
2762 | ||
2763 | /* Pick up timestamps. */ | |
2764 | j = ACCESS_ONCE(rdtp->dynticks_idle_jiffies); | |
2765 | /* If this CPU entered idle more recently, update maxj timestamp. */ | |
2766 | if (ULONG_CMP_LT(*maxj, j)) | |
2767 | *maxj = j; | |
2768 | } | |
2769 | ||
2770 | /* | |
2771 | * Is this the flavor of RCU that is handling full-system idle? | |
2772 | */ | |
2773 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2774 | { | |
2775 | return rsp == rcu_sysidle_state; | |
2776 | } | |
2777 | ||
2778 | /* | |
2779 | * Return a delay in jiffies based on the number of CPUs, rcu_node | |
2780 | * leaf fanout, and jiffies tick rate. The idea is to allow larger | |
2781 | * systems more time to transition to full-idle state in order to | |
2782 | * avoid the cache thrashing that otherwise occur on the state variable. | |
2783 | * Really small systems (less than a couple of tens of CPUs) should | |
2784 | * instead use a single global atomically incremented counter, and later | |
2785 | * versions of this will automatically reconfigure themselves accordingly. | |
2786 | */ | |
2787 | static unsigned long rcu_sysidle_delay(void) | |
2788 | { | |
2789 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2790 | return 0; | |
2791 | return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000); | |
2792 | } | |
2793 | ||
2794 | /* | |
2795 | * Advance the full-system-idle state. This is invoked when all of | |
2796 | * the non-timekeeping CPUs are idle. | |
2797 | */ | |
2798 | static void rcu_sysidle(unsigned long j) | |
2799 | { | |
2800 | /* Check the current state. */ | |
2801 | switch (ACCESS_ONCE(full_sysidle_state)) { | |
2802 | case RCU_SYSIDLE_NOT: | |
2803 | ||
2804 | /* First time all are idle, so note a short idle period. */ | |
2805 | ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_SHORT; | |
2806 | break; | |
2807 | ||
2808 | case RCU_SYSIDLE_SHORT: | |
2809 | ||
2810 | /* | |
2811 | * Idle for a bit, time to advance to next state? | |
2812 | * cmpxchg failure means race with non-idle, let them win. | |
2813 | */ | |
2814 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2815 | (void)cmpxchg(&full_sysidle_state, | |
2816 | RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG); | |
2817 | break; | |
2818 | ||
2819 | case RCU_SYSIDLE_LONG: | |
2820 | ||
2821 | /* | |
2822 | * Do an additional check pass before advancing to full. | |
2823 | * cmpxchg failure means race with non-idle, let them win. | |
2824 | */ | |
2825 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2826 | (void)cmpxchg(&full_sysidle_state, | |
2827 | RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL); | |
2828 | break; | |
2829 | ||
2830 | default: | |
2831 | break; | |
2832 | } | |
2833 | } | |
2834 | ||
2835 | /* | |
2836 | * Found a non-idle non-timekeeping CPU, so kick the system-idle state | |
2837 | * back to the beginning. | |
2838 | */ | |
2839 | static void rcu_sysidle_cancel(void) | |
2840 | { | |
2841 | smp_mb(); | |
becb41bf PM |
2842 | if (full_sysidle_state > RCU_SYSIDLE_SHORT) |
2843 | ACCESS_ONCE(full_sysidle_state) = RCU_SYSIDLE_NOT; | |
0edd1b17 PM |
2844 | } |
2845 | ||
2846 | /* | |
2847 | * Update the sysidle state based on the results of a force-quiescent-state | |
2848 | * scan of the CPUs' dyntick-idle state. | |
2849 | */ | |
2850 | static void rcu_sysidle_report(struct rcu_state *rsp, int isidle, | |
2851 | unsigned long maxj, bool gpkt) | |
2852 | { | |
2853 | if (rsp != rcu_sysidle_state) | |
2854 | return; /* Wrong flavor, ignore. */ | |
2855 | if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2856 | return; /* Running state machine from timekeeping CPU. */ | |
2857 | if (isidle) | |
2858 | rcu_sysidle(maxj); /* More idle! */ | |
2859 | else | |
2860 | rcu_sysidle_cancel(); /* Idle is over. */ | |
2861 | } | |
2862 | ||
2863 | /* | |
2864 | * Wrapper for rcu_sysidle_report() when called from the grace-period | |
2865 | * kthread's context. | |
2866 | */ | |
2867 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
2868 | unsigned long maxj) | |
2869 | { | |
2870 | rcu_sysidle_report(rsp, isidle, maxj, true); | |
2871 | } | |
2872 | ||
2873 | /* Callback and function for forcing an RCU grace period. */ | |
2874 | struct rcu_sysidle_head { | |
2875 | struct rcu_head rh; | |
2876 | int inuse; | |
2877 | }; | |
2878 | ||
2879 | static void rcu_sysidle_cb(struct rcu_head *rhp) | |
2880 | { | |
2881 | struct rcu_sysidle_head *rshp; | |
2882 | ||
2883 | /* | |
2884 | * The following memory barrier is needed to replace the | |
2885 | * memory barriers that would normally be in the memory | |
2886 | * allocator. | |
2887 | */ | |
2888 | smp_mb(); /* grace period precedes setting inuse. */ | |
2889 | ||
2890 | rshp = container_of(rhp, struct rcu_sysidle_head, rh); | |
2891 | ACCESS_ONCE(rshp->inuse) = 0; | |
2892 | } | |
2893 | ||
2894 | /* | |
2895 | * Check to see if the system is fully idle, other than the timekeeping CPU. | |
2896 | * The caller must have disabled interrupts. | |
2897 | */ | |
2898 | bool rcu_sys_is_idle(void) | |
2899 | { | |
2900 | static struct rcu_sysidle_head rsh; | |
2901 | int rss = ACCESS_ONCE(full_sysidle_state); | |
2902 | ||
2903 | if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu)) | |
2904 | return false; | |
2905 | ||
2906 | /* Handle small-system case by doing a full scan of CPUs. */ | |
2907 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) { | |
2908 | int oldrss = rss - 1; | |
2909 | ||
2910 | /* | |
2911 | * One pass to advance to each state up to _FULL. | |
2912 | * Give up if any pass fails to advance the state. | |
2913 | */ | |
2914 | while (rss < RCU_SYSIDLE_FULL && oldrss < rss) { | |
2915 | int cpu; | |
2916 | bool isidle = true; | |
2917 | unsigned long maxj = jiffies - ULONG_MAX / 4; | |
2918 | struct rcu_data *rdp; | |
2919 | ||
2920 | /* Scan all the CPUs looking for nonidle CPUs. */ | |
2921 | for_each_possible_cpu(cpu) { | |
2922 | rdp = per_cpu_ptr(rcu_sysidle_state->rda, cpu); | |
2923 | rcu_sysidle_check_cpu(rdp, &isidle, &maxj); | |
2924 | if (!isidle) | |
2925 | break; | |
2926 | } | |
2927 | rcu_sysidle_report(rcu_sysidle_state, | |
2928 | isidle, maxj, false); | |
2929 | oldrss = rss; | |
2930 | rss = ACCESS_ONCE(full_sysidle_state); | |
2931 | } | |
2932 | } | |
2933 | ||
2934 | /* If this is the first observation of an idle period, record it. */ | |
2935 | if (rss == RCU_SYSIDLE_FULL) { | |
2936 | rss = cmpxchg(&full_sysidle_state, | |
2937 | RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED); | |
2938 | return rss == RCU_SYSIDLE_FULL; | |
2939 | } | |
2940 | ||
2941 | smp_mb(); /* ensure rss load happens before later caller actions. */ | |
2942 | ||
2943 | /* If already fully idle, tell the caller (in case of races). */ | |
2944 | if (rss == RCU_SYSIDLE_FULL_NOTED) | |
2945 | return true; | |
2946 | ||
2947 | /* | |
2948 | * If we aren't there yet, and a grace period is not in flight, | |
2949 | * initiate a grace period. Either way, tell the caller that | |
2950 | * we are not there yet. We use an xchg() rather than an assignment | |
2951 | * to make up for the memory barriers that would otherwise be | |
2952 | * provided by the memory allocator. | |
2953 | */ | |
2954 | if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL && | |
2955 | !rcu_gp_in_progress(rcu_sysidle_state) && | |
2956 | !rsh.inuse && xchg(&rsh.inuse, 1) == 0) | |
2957 | call_rcu(&rsh.rh, rcu_sysidle_cb); | |
2958 | return false; | |
eb348b89 PM |
2959 | } |
2960 | ||
2333210b PM |
2961 | /* |
2962 | * Initialize dynticks sysidle state for CPUs coming online. | |
2963 | */ | |
2964 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) | |
2965 | { | |
2966 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE; | |
2967 | } | |
2968 | ||
2969 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
2970 | ||
eb348b89 PM |
2971 | static void rcu_sysidle_enter(struct rcu_dynticks *rdtp, int irq) |
2972 | { | |
2973 | } | |
2974 | ||
2975 | static void rcu_sysidle_exit(struct rcu_dynticks *rdtp, int irq) | |
2976 | { | |
2977 | } | |
2978 | ||
0edd1b17 PM |
2979 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, |
2980 | unsigned long *maxj) | |
2981 | { | |
2982 | } | |
2983 | ||
2984 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2985 | { | |
2986 | return false; | |
2987 | } | |
2988 | ||
2989 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
2990 | unsigned long maxj) | |
2991 | { | |
2992 | } | |
2993 | ||
2333210b PM |
2994 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) |
2995 | { | |
2996 | } | |
2997 | ||
2998 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
a096932f PM |
2999 | |
3000 | /* | |
3001 | * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the | |
3002 | * grace-period kthread will do force_quiescent_state() processing? | |
3003 | * The idea is to avoid waking up RCU core processing on such a | |
3004 | * CPU unless the grace period has extended for too long. | |
3005 | * | |
3006 | * This code relies on the fact that all NO_HZ_FULL CPUs are also | |
52e2bb95 | 3007 | * CONFIG_RCU_NOCB_CPU CPUs. |
a096932f PM |
3008 | */ |
3009 | static bool rcu_nohz_full_cpu(struct rcu_state *rsp) | |
3010 | { | |
3011 | #ifdef CONFIG_NO_HZ_FULL | |
3012 | if (tick_nohz_full_cpu(smp_processor_id()) && | |
3013 | (!rcu_gp_in_progress(rsp) || | |
3014 | ULONG_CMP_LT(jiffies, ACCESS_ONCE(rsp->gp_start) + HZ))) | |
3015 | return 1; | |
3016 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ | |
3017 | return 0; | |
3018 | } | |
5057f55e PM |
3019 | |
3020 | /* | |
3021 | * Bind the grace-period kthread for the sysidle flavor of RCU to the | |
3022 | * timekeeping CPU. | |
3023 | */ | |
3024 | static void rcu_bind_gp_kthread(void) | |
3025 | { | |
c0f489d2 | 3026 | int __maybe_unused cpu; |
5057f55e | 3027 | |
c0f489d2 | 3028 | if (!tick_nohz_full_enabled()) |
5057f55e | 3029 | return; |
c0f489d2 PM |
3030 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
3031 | cpu = tick_do_timer_cpu; | |
3032 | if (cpu >= 0 && cpu < nr_cpu_ids && raw_smp_processor_id() != cpu) | |
5057f55e | 3033 | set_cpus_allowed_ptr(current, cpumask_of(cpu)); |
c0f489d2 PM |
3034 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
3035 | if (!is_housekeeping_cpu(raw_smp_processor_id())) | |
3036 | housekeeping_affine(current); | |
3037 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
5057f55e | 3038 | } |