Commit | Line | Data |
---|---|---|
f41d911f PM |
1 | /* |
2 | * Read-Copy Update mechanism for mutual exclusion (tree-based version) | |
3 | * Internal non-public definitions that provide either classic | |
6cc68793 | 4 | * or preemptible semantics. |
f41d911f PM |
5 | * |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License as published by | |
8 | * the Free Software Foundation; either version 2 of the License, or | |
9 | * (at your option) any later version. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | * GNU General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
87de1cfd PM |
17 | * along with this program; if not, you can access it online at |
18 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
f41d911f PM |
19 | * |
20 | * Copyright Red Hat, 2009 | |
21 | * Copyright IBM Corporation, 2009 | |
22 | * | |
23 | * Author: Ingo Molnar <mingo@elte.hu> | |
24 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> | |
25 | */ | |
26 | ||
d9a3da06 | 27 | #include <linux/delay.h> |
3fbfbf7a | 28 | #include <linux/gfp.h> |
b626c1b6 | 29 | #include <linux/oom.h> |
62ab7072 | 30 | #include <linux/smpboot.h> |
4102adab | 31 | #include "../time/tick-internal.h" |
f41d911f | 32 | |
5b61b0ba | 33 | #ifdef CONFIG_RCU_BOOST |
61cfd097 | 34 | |
abaa93d9 | 35 | #include "../locking/rtmutex_common.h" |
21871d7e | 36 | |
61cfd097 PM |
37 | /* |
38 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
39 | * handle all flavors of RCU. | |
40 | */ | |
41 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
42 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); | |
43 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); | |
44 | DEFINE_PER_CPU(char, rcu_cpu_has_work); | |
45 | ||
727b705b PM |
46 | #else /* #ifdef CONFIG_RCU_BOOST */ |
47 | ||
48 | /* | |
49 | * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST, | |
50 | * all uses are in dead code. Provide a definition to keep the compiler | |
51 | * happy, but add WARN_ON_ONCE() to complain if used in the wrong place. | |
52 | * This probably needs to be excluded from -rt builds. | |
53 | */ | |
54 | #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; }) | |
55 | ||
56 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ | |
5b61b0ba | 57 | |
3fbfbf7a PM |
58 | #ifdef CONFIG_RCU_NOCB_CPU |
59 | static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */ | |
60 | static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */ | |
1b0048a4 | 61 | static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */ |
3fbfbf7a PM |
62 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ |
63 | ||
26845c28 PM |
64 | /* |
65 | * Check the RCU kernel configuration parameters and print informative | |
699d4035 | 66 | * messages about anything out of the ordinary. |
26845c28 PM |
67 | */ |
68 | static void __init rcu_bootup_announce_oddness(void) | |
69 | { | |
ab6f5bd6 PM |
70 | if (IS_ENABLED(CONFIG_RCU_TRACE)) |
71 | pr_info("\tRCU debugfs-based tracing is enabled.\n"); | |
05c5df31 PM |
72 | if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) || |
73 | (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32)) | |
ab6f5bd6 | 74 | pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n", |
05c5df31 | 75 | RCU_FANOUT); |
7fa27001 | 76 | if (rcu_fanout_exact) |
ab6f5bd6 PM |
77 | pr_info("\tHierarchical RCU autobalancing is disabled.\n"); |
78 | if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ)) | |
79 | pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n"); | |
80 | if (IS_ENABLED(CONFIG_PROVE_RCU)) | |
81 | pr_info("\tRCU lockdep checking is enabled.\n"); | |
82 | if (IS_ENABLED(CONFIG_RCU_TORTURE_TEST_RUNNABLE)) | |
83 | pr_info("\tRCU torture testing starts during boot.\n"); | |
42621697 AG |
84 | if (RCU_NUM_LVLS >= 4) |
85 | pr_info("\tFour(or more)-level hierarchy is enabled.\n"); | |
47d631af | 86 | if (RCU_FANOUT_LEAF != 16) |
a3bd2c09 | 87 | pr_info("\tBuild-time adjustment of leaf fanout to %d.\n", |
47d631af PM |
88 | RCU_FANOUT_LEAF); |
89 | if (rcu_fanout_leaf != RCU_FANOUT_LEAF) | |
9a5739d7 | 90 | pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf); |
cca6f393 | 91 | if (nr_cpu_ids != NR_CPUS) |
efc151c3 | 92 | pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids); |
ab6f5bd6 PM |
93 | if (IS_ENABLED(CONFIG_RCU_BOOST)) |
94 | pr_info("\tRCU kthread priority: %d.\n", kthread_prio); | |
26845c28 PM |
95 | } |
96 | ||
28f6569a | 97 | #ifdef CONFIG_PREEMPT_RCU |
f41d911f | 98 | |
a41bfeb2 | 99 | RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu); |
b28a7c01 | 100 | static struct rcu_state *const rcu_state_p = &rcu_preempt_state; |
2927a689 | 101 | static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data; |
f41d911f | 102 | |
d19fb8d1 PM |
103 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
104 | bool wake); | |
d9a3da06 | 105 | |
f41d911f PM |
106 | /* |
107 | * Tell them what RCU they are running. | |
108 | */ | |
0e0fc1c2 | 109 | static void __init rcu_bootup_announce(void) |
f41d911f | 110 | { |
efc151c3 | 111 | pr_info("Preemptible hierarchical RCU implementation.\n"); |
26845c28 | 112 | rcu_bootup_announce_oddness(); |
f41d911f PM |
113 | } |
114 | ||
8203d6d0 PM |
115 | /* Flags for rcu_preempt_ctxt_queue() decision table. */ |
116 | #define RCU_GP_TASKS 0x8 | |
117 | #define RCU_EXP_TASKS 0x4 | |
118 | #define RCU_GP_BLKD 0x2 | |
119 | #define RCU_EXP_BLKD 0x1 | |
120 | ||
121 | /* | |
122 | * Queues a task preempted within an RCU-preempt read-side critical | |
123 | * section into the appropriate location within the ->blkd_tasks list, | |
124 | * depending on the states of any ongoing normal and expedited grace | |
125 | * periods. The ->gp_tasks pointer indicates which element the normal | |
126 | * grace period is waiting on (NULL if none), and the ->exp_tasks pointer | |
127 | * indicates which element the expedited grace period is waiting on (again, | |
128 | * NULL if none). If a grace period is waiting on a given element in the | |
129 | * ->blkd_tasks list, it also waits on all subsequent elements. Thus, | |
130 | * adding a task to the tail of the list blocks any grace period that is | |
131 | * already waiting on one of the elements. In contrast, adding a task | |
132 | * to the head of the list won't block any grace period that is already | |
133 | * waiting on one of the elements. | |
134 | * | |
135 | * This queuing is imprecise, and can sometimes make an ongoing grace | |
136 | * period wait for a task that is not strictly speaking blocking it. | |
137 | * Given the choice, we needlessly block a normal grace period rather than | |
138 | * blocking an expedited grace period. | |
139 | * | |
140 | * Note that an endless sequence of expedited grace periods still cannot | |
141 | * indefinitely postpone a normal grace period. Eventually, all of the | |
142 | * fixed number of preempted tasks blocking the normal grace period that are | |
143 | * not also blocking the expedited grace period will resume and complete | |
144 | * their RCU read-side critical sections. At that point, the ->gp_tasks | |
145 | * pointer will equal the ->exp_tasks pointer, at which point the end of | |
146 | * the corresponding expedited grace period will also be the end of the | |
147 | * normal grace period. | |
148 | */ | |
46a5d164 PM |
149 | static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp) |
150 | __releases(rnp->lock) /* But leaves rrupts disabled. */ | |
8203d6d0 PM |
151 | { |
152 | int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) + | |
153 | (rnp->exp_tasks ? RCU_EXP_TASKS : 0) + | |
154 | (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) + | |
155 | (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0); | |
156 | struct task_struct *t = current; | |
157 | ||
158 | /* | |
159 | * Decide where to queue the newly blocked task. In theory, | |
160 | * this could be an if-statement. In practice, when I tried | |
161 | * that, it was quite messy. | |
162 | */ | |
163 | switch (blkd_state) { | |
164 | case 0: | |
165 | case RCU_EXP_TASKS: | |
166 | case RCU_EXP_TASKS + RCU_GP_BLKD: | |
167 | case RCU_GP_TASKS: | |
168 | case RCU_GP_TASKS + RCU_EXP_TASKS: | |
169 | ||
170 | /* | |
171 | * Blocking neither GP, or first task blocking the normal | |
172 | * GP but not blocking the already-waiting expedited GP. | |
173 | * Queue at the head of the list to avoid unnecessarily | |
174 | * blocking the already-waiting GPs. | |
175 | */ | |
176 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); | |
177 | break; | |
178 | ||
179 | case RCU_EXP_BLKD: | |
180 | case RCU_GP_BLKD: | |
181 | case RCU_GP_BLKD + RCU_EXP_BLKD: | |
182 | case RCU_GP_TASKS + RCU_EXP_BLKD: | |
183 | case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
184 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
185 | ||
186 | /* | |
187 | * First task arriving that blocks either GP, or first task | |
188 | * arriving that blocks the expedited GP (with the normal | |
189 | * GP already waiting), or a task arriving that blocks | |
190 | * both GPs with both GPs already waiting. Queue at the | |
191 | * tail of the list to avoid any GP waiting on any of the | |
192 | * already queued tasks that are not blocking it. | |
193 | */ | |
194 | list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks); | |
195 | break; | |
196 | ||
197 | case RCU_EXP_TASKS + RCU_EXP_BLKD: | |
198 | case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
199 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD: | |
200 | ||
201 | /* | |
202 | * Second or subsequent task blocking the expedited GP. | |
203 | * The task either does not block the normal GP, or is the | |
204 | * first task blocking the normal GP. Queue just after | |
205 | * the first task blocking the expedited GP. | |
206 | */ | |
207 | list_add(&t->rcu_node_entry, rnp->exp_tasks); | |
208 | break; | |
209 | ||
210 | case RCU_GP_TASKS + RCU_GP_BLKD: | |
211 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD: | |
212 | ||
213 | /* | |
214 | * Second or subsequent task blocking the normal GP. | |
215 | * The task does not block the expedited GP. Queue just | |
216 | * after the first task blocking the normal GP. | |
217 | */ | |
218 | list_add(&t->rcu_node_entry, rnp->gp_tasks); | |
219 | break; | |
220 | ||
221 | default: | |
222 | ||
223 | /* Yet another exercise in excessive paranoia. */ | |
224 | WARN_ON_ONCE(1); | |
225 | break; | |
226 | } | |
227 | ||
228 | /* | |
229 | * We have now queued the task. If it was the first one to | |
230 | * block either grace period, update the ->gp_tasks and/or | |
231 | * ->exp_tasks pointers, respectively, to reference the newly | |
232 | * blocked tasks. | |
233 | */ | |
234 | if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) | |
235 | rnp->gp_tasks = &t->rcu_node_entry; | |
236 | if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD)) | |
237 | rnp->exp_tasks = &t->rcu_node_entry; | |
67c583a7 | 238 | raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */ |
8203d6d0 PM |
239 | |
240 | /* | |
241 | * Report the quiescent state for the expedited GP. This expedited | |
242 | * GP should not be able to end until we report, so there should be | |
243 | * no need to check for a subsequent expedited GP. (Though we are | |
244 | * still in a quiescent state in any case.) | |
245 | */ | |
246 | if (blkd_state & RCU_EXP_BLKD && | |
247 | t->rcu_read_unlock_special.b.exp_need_qs) { | |
248 | t->rcu_read_unlock_special.b.exp_need_qs = false; | |
249 | rcu_report_exp_rdp(rdp->rsp, rdp, true); | |
250 | } else { | |
251 | WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs); | |
252 | } | |
8203d6d0 PM |
253 | } |
254 | ||
f41d911f | 255 | /* |
6cc68793 | 256 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
f41d911f PM |
257 | * that this just means that the task currently running on the CPU is |
258 | * not in a quiescent state. There might be any number of tasks blocked | |
259 | * while in an RCU read-side critical section. | |
25502a6c | 260 | * |
1d082fd0 PM |
261 | * As with the other rcu_*_qs() functions, callers to this function |
262 | * must disable preemption. | |
f41d911f | 263 | */ |
284a8c93 | 264 | static void rcu_preempt_qs(void) |
f41d911f | 265 | { |
5b74c458 | 266 | if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) { |
284a8c93 | 267 | trace_rcu_grace_period(TPS("rcu_preempt"), |
2927a689 | 268 | __this_cpu_read(rcu_data_p->gpnum), |
284a8c93 | 269 | TPS("cpuqs")); |
5b74c458 | 270 | __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false); |
284a8c93 PM |
271 | barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */ |
272 | current->rcu_read_unlock_special.b.need_qs = false; | |
273 | } | |
f41d911f PM |
274 | } |
275 | ||
276 | /* | |
c3422bea PM |
277 | * We have entered the scheduler, and the current task might soon be |
278 | * context-switched away from. If this task is in an RCU read-side | |
279 | * critical section, we will no longer be able to rely on the CPU to | |
12f5f524 PM |
280 | * record that fact, so we enqueue the task on the blkd_tasks list. |
281 | * The task will dequeue itself when it exits the outermost enclosing | |
282 | * RCU read-side critical section. Therefore, the current grace period | |
283 | * cannot be permitted to complete until the blkd_tasks list entries | |
284 | * predating the current grace period drain, in other words, until | |
285 | * rnp->gp_tasks becomes NULL. | |
c3422bea | 286 | * |
46a5d164 | 287 | * Caller must disable interrupts. |
f41d911f | 288 | */ |
38200cf2 | 289 | static void rcu_preempt_note_context_switch(void) |
f41d911f PM |
290 | { |
291 | struct task_struct *t = current; | |
f41d911f PM |
292 | struct rcu_data *rdp; |
293 | struct rcu_node *rnp; | |
294 | ||
10f39bb1 | 295 | if (t->rcu_read_lock_nesting > 0 && |
1d082fd0 | 296 | !t->rcu_read_unlock_special.b.blocked) { |
f41d911f PM |
297 | |
298 | /* Possibly blocking in an RCU read-side critical section. */ | |
e63c887c | 299 | rdp = this_cpu_ptr(rcu_state_p->rda); |
f41d911f | 300 | rnp = rdp->mynode; |
46a5d164 | 301 | raw_spin_lock_rcu_node(rnp); |
1d082fd0 | 302 | t->rcu_read_unlock_special.b.blocked = true; |
86848966 | 303 | t->rcu_blocked_node = rnp; |
f41d911f PM |
304 | |
305 | /* | |
8203d6d0 PM |
306 | * Verify the CPU's sanity, trace the preemption, and |
307 | * then queue the task as required based on the states | |
308 | * of any ongoing and expedited grace periods. | |
f41d911f | 309 | */ |
0aa04b05 | 310 | WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0); |
e7d8842e | 311 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
d4c08f2a PM |
312 | trace_rcu_preempt_task(rdp->rsp->name, |
313 | t->pid, | |
314 | (rnp->qsmask & rdp->grpmask) | |
315 | ? rnp->gpnum | |
316 | : rnp->gpnum + 1); | |
46a5d164 | 317 | rcu_preempt_ctxt_queue(rnp, rdp); |
10f39bb1 | 318 | } else if (t->rcu_read_lock_nesting < 0 && |
1d082fd0 | 319 | t->rcu_read_unlock_special.s) { |
10f39bb1 PM |
320 | |
321 | /* | |
322 | * Complete exit from RCU read-side critical section on | |
323 | * behalf of preempted instance of __rcu_read_unlock(). | |
324 | */ | |
325 | rcu_read_unlock_special(t); | |
f41d911f PM |
326 | } |
327 | ||
328 | /* | |
329 | * Either we were not in an RCU read-side critical section to | |
330 | * begin with, or we have now recorded that critical section | |
331 | * globally. Either way, we can now note a quiescent state | |
332 | * for this CPU. Again, if we were in an RCU read-side critical | |
333 | * section, and if that critical section was blocking the current | |
334 | * grace period, then the fact that the task has been enqueued | |
335 | * means that we continue to block the current grace period. | |
336 | */ | |
284a8c93 | 337 | rcu_preempt_qs(); |
f41d911f PM |
338 | } |
339 | ||
fc2219d4 PM |
340 | /* |
341 | * Check for preempted RCU readers blocking the current grace period | |
342 | * for the specified rcu_node structure. If the caller needs a reliable | |
343 | * answer, it must hold the rcu_node's ->lock. | |
344 | */ | |
27f4d280 | 345 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 | 346 | { |
12f5f524 | 347 | return rnp->gp_tasks != NULL; |
fc2219d4 PM |
348 | } |
349 | ||
12f5f524 PM |
350 | /* |
351 | * Advance a ->blkd_tasks-list pointer to the next entry, instead | |
352 | * returning NULL if at the end of the list. | |
353 | */ | |
354 | static struct list_head *rcu_next_node_entry(struct task_struct *t, | |
355 | struct rcu_node *rnp) | |
356 | { | |
357 | struct list_head *np; | |
358 | ||
359 | np = t->rcu_node_entry.next; | |
360 | if (np == &rnp->blkd_tasks) | |
361 | np = NULL; | |
362 | return np; | |
363 | } | |
364 | ||
8af3a5e7 PM |
365 | /* |
366 | * Return true if the specified rcu_node structure has tasks that were | |
367 | * preempted within an RCU read-side critical section. | |
368 | */ | |
369 | static bool rcu_preempt_has_tasks(struct rcu_node *rnp) | |
370 | { | |
371 | return !list_empty(&rnp->blkd_tasks); | |
372 | } | |
373 | ||
b668c9cf PM |
374 | /* |
375 | * Handle special cases during rcu_read_unlock(), such as needing to | |
376 | * notify RCU core processing or task having blocked during the RCU | |
377 | * read-side critical section. | |
378 | */ | |
2a3fa843 | 379 | void rcu_read_unlock_special(struct task_struct *t) |
f41d911f | 380 | { |
b6a932d1 PM |
381 | bool empty_exp; |
382 | bool empty_norm; | |
383 | bool empty_exp_now; | |
f41d911f | 384 | unsigned long flags; |
12f5f524 | 385 | struct list_head *np; |
abaa93d9 | 386 | bool drop_boost_mutex = false; |
8203d6d0 | 387 | struct rcu_data *rdp; |
f41d911f | 388 | struct rcu_node *rnp; |
1d082fd0 | 389 | union rcu_special special; |
f41d911f PM |
390 | |
391 | /* NMI handlers cannot block and cannot safely manipulate state. */ | |
392 | if (in_nmi()) | |
393 | return; | |
394 | ||
395 | local_irq_save(flags); | |
396 | ||
397 | /* | |
8203d6d0 PM |
398 | * If RCU core is waiting for this CPU to exit its critical section, |
399 | * report the fact that it has exited. Because irqs are disabled, | |
1d082fd0 | 400 | * t->rcu_read_unlock_special cannot change. |
f41d911f PM |
401 | */ |
402 | special = t->rcu_read_unlock_special; | |
1d082fd0 | 403 | if (special.b.need_qs) { |
284a8c93 | 404 | rcu_preempt_qs(); |
c0135d07 | 405 | t->rcu_read_unlock_special.b.need_qs = false; |
1d082fd0 | 406 | if (!t->rcu_read_unlock_special.s) { |
79a62f95 LJ |
407 | local_irq_restore(flags); |
408 | return; | |
409 | } | |
f41d911f PM |
410 | } |
411 | ||
8203d6d0 PM |
412 | /* |
413 | * Respond to a request for an expedited grace period, but only if | |
414 | * we were not preempted, meaning that we were running on the same | |
415 | * CPU throughout. If we were preempted, the exp_need_qs flag | |
416 | * would have been cleared at the time of the first preemption, | |
417 | * and the quiescent state would be reported when we were dequeued. | |
418 | */ | |
419 | if (special.b.exp_need_qs) { | |
420 | WARN_ON_ONCE(special.b.blocked); | |
421 | t->rcu_read_unlock_special.b.exp_need_qs = false; | |
422 | rdp = this_cpu_ptr(rcu_state_p->rda); | |
423 | rcu_report_exp_rdp(rcu_state_p, rdp, true); | |
424 | if (!t->rcu_read_unlock_special.s) { | |
425 | local_irq_restore(flags); | |
426 | return; | |
427 | } | |
428 | } | |
429 | ||
79a62f95 | 430 | /* Hardware IRQ handlers cannot block, complain if they get here. */ |
d24209bb PM |
431 | if (in_irq() || in_serving_softirq()) { |
432 | lockdep_rcu_suspicious(__FILE__, __LINE__, | |
433 | "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n"); | |
8203d6d0 | 434 | pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n", |
d24209bb PM |
435 | t->rcu_read_unlock_special.s, |
436 | t->rcu_read_unlock_special.b.blocked, | |
8203d6d0 | 437 | t->rcu_read_unlock_special.b.exp_need_qs, |
d24209bb | 438 | t->rcu_read_unlock_special.b.need_qs); |
f41d911f PM |
439 | local_irq_restore(flags); |
440 | return; | |
441 | } | |
442 | ||
443 | /* Clean up if blocked during RCU read-side critical section. */ | |
1d082fd0 PM |
444 | if (special.b.blocked) { |
445 | t->rcu_read_unlock_special.b.blocked = false; | |
f41d911f | 446 | |
dd5d19ba | 447 | /* |
0a0ba1c9 | 448 | * Remove this task from the list it blocked on. The task |
8ba9153b PM |
449 | * now remains queued on the rcu_node corresponding to the |
450 | * CPU it first blocked on, so there is no longer any need | |
451 | * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia. | |
dd5d19ba | 452 | */ |
8ba9153b PM |
453 | rnp = t->rcu_blocked_node; |
454 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ | |
455 | WARN_ON_ONCE(rnp != t->rcu_blocked_node); | |
74e871ac | 456 | empty_norm = !rcu_preempt_blocked_readers_cgp(rnp); |
8203d6d0 | 457 | empty_exp = sync_rcu_preempt_exp_done(rnp); |
d9a3da06 | 458 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ |
12f5f524 | 459 | np = rcu_next_node_entry(t, rnp); |
f41d911f | 460 | list_del_init(&t->rcu_node_entry); |
82e78d80 | 461 | t->rcu_blocked_node = NULL; |
f7f7bac9 | 462 | trace_rcu_unlock_preempted_task(TPS("rcu_preempt"), |
d4c08f2a | 463 | rnp->gpnum, t->pid); |
12f5f524 PM |
464 | if (&t->rcu_node_entry == rnp->gp_tasks) |
465 | rnp->gp_tasks = np; | |
466 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
467 | rnp->exp_tasks = np; | |
727b705b PM |
468 | if (IS_ENABLED(CONFIG_RCU_BOOST)) { |
469 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
470 | rnp->boost_tasks = np; | |
471 | /* Snapshot ->boost_mtx ownership w/rnp->lock held. */ | |
472 | drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t; | |
473 | } | |
f41d911f PM |
474 | |
475 | /* | |
476 | * If this was the last task on the current list, and if | |
477 | * we aren't waiting on any CPUs, report the quiescent state. | |
389abd48 PM |
478 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock, |
479 | * so we must take a snapshot of the expedited state. | |
f41d911f | 480 | */ |
8203d6d0 | 481 | empty_exp_now = sync_rcu_preempt_exp_done(rnp); |
74e871ac | 482 | if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) { |
f7f7bac9 | 483 | trace_rcu_quiescent_state_report(TPS("preempt_rcu"), |
d4c08f2a PM |
484 | rnp->gpnum, |
485 | 0, rnp->qsmask, | |
486 | rnp->level, | |
487 | rnp->grplo, | |
488 | rnp->grphi, | |
489 | !!rnp->gp_tasks); | |
e63c887c | 490 | rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags); |
c701d5d9 | 491 | } else { |
67c583a7 | 492 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
c701d5d9 | 493 | } |
d9a3da06 | 494 | |
27f4d280 | 495 | /* Unboost if we were boosted. */ |
727b705b | 496 | if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex) |
abaa93d9 | 497 | rt_mutex_unlock(&rnp->boost_mtx); |
27f4d280 | 498 | |
d9a3da06 PM |
499 | /* |
500 | * If this was the last task on the expedited lists, | |
501 | * then we need to report up the rcu_node hierarchy. | |
502 | */ | |
389abd48 | 503 | if (!empty_exp && empty_exp_now) |
e63c887c | 504 | rcu_report_exp_rnp(rcu_state_p, rnp, true); |
b668c9cf PM |
505 | } else { |
506 | local_irq_restore(flags); | |
f41d911f | 507 | } |
f41d911f PM |
508 | } |
509 | ||
1ed509a2 PM |
510 | /* |
511 | * Dump detailed information for all tasks blocking the current RCU | |
512 | * grace period on the specified rcu_node structure. | |
513 | */ | |
514 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) | |
515 | { | |
516 | unsigned long flags; | |
1ed509a2 PM |
517 | struct task_struct *t; |
518 | ||
6cf10081 | 519 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
5fd4dc06 | 520 | if (!rcu_preempt_blocked_readers_cgp(rnp)) { |
67c583a7 | 521 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
5fd4dc06 PM |
522 | return; |
523 | } | |
82efed06 | 524 | t = list_entry(rnp->gp_tasks->prev, |
12f5f524 PM |
525 | struct task_struct, rcu_node_entry); |
526 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) | |
527 | sched_show_task(t); | |
67c583a7 | 528 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1ed509a2 PM |
529 | } |
530 | ||
531 | /* | |
532 | * Dump detailed information for all tasks blocking the current RCU | |
533 | * grace period. | |
534 | */ | |
535 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
536 | { | |
537 | struct rcu_node *rnp = rcu_get_root(rsp); | |
538 | ||
539 | rcu_print_detail_task_stall_rnp(rnp); | |
540 | rcu_for_each_leaf_node(rsp, rnp) | |
541 | rcu_print_detail_task_stall_rnp(rnp); | |
542 | } | |
543 | ||
a858af28 PM |
544 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) |
545 | { | |
efc151c3 | 546 | pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):", |
a858af28 PM |
547 | rnp->level, rnp->grplo, rnp->grphi); |
548 | } | |
549 | ||
550 | static void rcu_print_task_stall_end(void) | |
551 | { | |
efc151c3 | 552 | pr_cont("\n"); |
a858af28 PM |
553 | } |
554 | ||
f41d911f PM |
555 | /* |
556 | * Scan the current list of tasks blocked within RCU read-side critical | |
557 | * sections, printing out the tid of each. | |
558 | */ | |
9bc8b558 | 559 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 560 | { |
f41d911f | 561 | struct task_struct *t; |
9bc8b558 | 562 | int ndetected = 0; |
f41d911f | 563 | |
27f4d280 | 564 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
9bc8b558 | 565 | return 0; |
a858af28 | 566 | rcu_print_task_stall_begin(rnp); |
82efed06 | 567 | t = list_entry(rnp->gp_tasks->prev, |
12f5f524 | 568 | struct task_struct, rcu_node_entry); |
9bc8b558 | 569 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { |
efc151c3 | 570 | pr_cont(" P%d", t->pid); |
9bc8b558 PM |
571 | ndetected++; |
572 | } | |
a858af28 | 573 | rcu_print_task_stall_end(); |
9bc8b558 | 574 | return ndetected; |
f41d911f PM |
575 | } |
576 | ||
74611ecb PM |
577 | /* |
578 | * Scan the current list of tasks blocked within RCU read-side critical | |
579 | * sections, printing out the tid of each that is blocking the current | |
580 | * expedited grace period. | |
581 | */ | |
582 | static int rcu_print_task_exp_stall(struct rcu_node *rnp) | |
583 | { | |
584 | struct task_struct *t; | |
585 | int ndetected = 0; | |
586 | ||
587 | if (!rnp->exp_tasks) | |
588 | return 0; | |
589 | t = list_entry(rnp->exp_tasks->prev, | |
590 | struct task_struct, rcu_node_entry); | |
591 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { | |
592 | pr_cont(" P%d", t->pid); | |
593 | ndetected++; | |
594 | } | |
595 | return ndetected; | |
596 | } | |
597 | ||
b0e165c0 PM |
598 | /* |
599 | * Check that the list of blocked tasks for the newly completed grace | |
600 | * period is in fact empty. It is a serious bug to complete a grace | |
601 | * period that still has RCU readers blocked! This function must be | |
602 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock | |
603 | * must be held by the caller. | |
12f5f524 PM |
604 | * |
605 | * Also, if there are blocked tasks on the list, they automatically | |
606 | * block the newly created grace period, so set up ->gp_tasks accordingly. | |
b0e165c0 PM |
607 | */ |
608 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
609 | { | |
27f4d280 | 610 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
96e92021 | 611 | if (rcu_preempt_has_tasks(rnp)) |
12f5f524 | 612 | rnp->gp_tasks = rnp->blkd_tasks.next; |
28ecd580 | 613 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
614 | } |
615 | ||
f41d911f PM |
616 | /* |
617 | * Check for a quiescent state from the current CPU. When a task blocks, | |
618 | * the task is recorded in the corresponding CPU's rcu_node structure, | |
619 | * which is checked elsewhere. | |
620 | * | |
621 | * Caller must disable hard irqs. | |
622 | */ | |
86aea0e6 | 623 | static void rcu_preempt_check_callbacks(void) |
f41d911f PM |
624 | { |
625 | struct task_struct *t = current; | |
626 | ||
627 | if (t->rcu_read_lock_nesting == 0) { | |
284a8c93 | 628 | rcu_preempt_qs(); |
f41d911f PM |
629 | return; |
630 | } | |
10f39bb1 | 631 | if (t->rcu_read_lock_nesting > 0 && |
97c668b8 | 632 | __this_cpu_read(rcu_data_p->core_needs_qs) && |
5b74c458 | 633 | __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm)) |
1d082fd0 | 634 | t->rcu_read_unlock_special.b.need_qs = true; |
f41d911f PM |
635 | } |
636 | ||
a46e0899 PM |
637 | #ifdef CONFIG_RCU_BOOST |
638 | ||
09223371 SL |
639 | static void rcu_preempt_do_callbacks(void) |
640 | { | |
2927a689 | 641 | rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p)); |
09223371 SL |
642 | } |
643 | ||
a46e0899 PM |
644 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
645 | ||
f41d911f | 646 | /* |
6cc68793 | 647 | * Queue a preemptible-RCU callback for invocation after a grace period. |
f41d911f | 648 | */ |
b6a4ae76 | 649 | void call_rcu(struct rcu_head *head, rcu_callback_t func) |
f41d911f | 650 | { |
e63c887c | 651 | __call_rcu(head, func, rcu_state_p, -1, 0); |
f41d911f PM |
652 | } |
653 | EXPORT_SYMBOL_GPL(call_rcu); | |
654 | ||
6ebb237b PM |
655 | /** |
656 | * synchronize_rcu - wait until a grace period has elapsed. | |
657 | * | |
658 | * Control will return to the caller some time after a full grace | |
659 | * period has elapsed, in other words after all currently executing RCU | |
77d8485a PM |
660 | * read-side critical sections have completed. Note, however, that |
661 | * upon return from synchronize_rcu(), the caller might well be executing | |
662 | * concurrently with new RCU read-side critical sections that began while | |
663 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
664 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
f0a0e6f2 PM |
665 | * |
666 | * See the description of synchronize_sched() for more detailed information | |
667 | * on memory ordering guarantees. | |
6ebb237b PM |
668 | */ |
669 | void synchronize_rcu(void) | |
670 | { | |
f78f5b90 PM |
671 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
672 | lock_is_held(&rcu_lock_map) || | |
673 | lock_is_held(&rcu_sched_lock_map), | |
674 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
6ebb237b PM |
675 | if (!rcu_scheduler_active) |
676 | return; | |
5afff48b | 677 | if (rcu_gp_is_expedited()) |
3705b88d AM |
678 | synchronize_rcu_expedited(); |
679 | else | |
680 | wait_rcu_gp(call_rcu); | |
6ebb237b PM |
681 | } |
682 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
683 | ||
d9a3da06 | 684 | /* |
8203d6d0 PM |
685 | * Remote handler for smp_call_function_single(). If there is an |
686 | * RCU read-side critical section in effect, request that the | |
687 | * next rcu_read_unlock() record the quiescent state up the | |
688 | * ->expmask fields in the rcu_node tree. Otherwise, immediately | |
689 | * report the quiescent state. | |
d9a3da06 | 690 | */ |
8203d6d0 | 691 | static void sync_rcu_exp_handler(void *info) |
d9a3da06 | 692 | { |
8203d6d0 PM |
693 | struct rcu_data *rdp; |
694 | struct rcu_state *rsp = info; | |
695 | struct task_struct *t = current; | |
8eb74b2b PM |
696 | |
697 | /* | |
8203d6d0 PM |
698 | * Within an RCU read-side critical section, request that the next |
699 | * rcu_read_unlock() report. Unless this RCU read-side critical | |
700 | * section has already blocked, in which case it is already set | |
701 | * up for the expedited grace period to wait on it. | |
8eb74b2b | 702 | */ |
8203d6d0 PM |
703 | if (t->rcu_read_lock_nesting > 0 && |
704 | !t->rcu_read_unlock_special.b.blocked) { | |
705 | t->rcu_read_unlock_special.b.exp_need_qs = true; | |
8eb74b2b | 706 | return; |
12f5f524 | 707 | } |
8eb74b2b | 708 | |
8203d6d0 PM |
709 | /* |
710 | * We are either exiting an RCU read-side critical section (negative | |
711 | * values of t->rcu_read_lock_nesting) or are not in one at all | |
712 | * (zero value of t->rcu_read_lock_nesting). Or we are in an RCU | |
713 | * read-side critical section that blocked before this expedited | |
714 | * grace period started. Either way, we can immediately report | |
715 | * the quiescent state. | |
716 | */ | |
717 | rdp = this_cpu_ptr(rsp->rda); | |
718 | rcu_report_exp_rdp(rsp, rdp, true); | |
d9a3da06 PM |
719 | } |
720 | ||
236fefaf PM |
721 | /** |
722 | * synchronize_rcu_expedited - Brute-force RCU grace period | |
723 | * | |
724 | * Wait for an RCU-preempt grace period, but expedite it. The basic | |
26ece8ef PM |
725 | * idea is to IPI all non-idle non-nohz online CPUs. The IPI handler |
726 | * checks whether the CPU is in an RCU-preempt critical section, and | |
727 | * if so, it sets a flag that causes the outermost rcu_read_unlock() | |
728 | * to report the quiescent state. On the other hand, if the CPU is | |
729 | * not in an RCU read-side critical section, the IPI handler reports | |
730 | * the quiescent state immediately. | |
731 | * | |
732 | * Although this is a greate improvement over previous expedited | |
733 | * implementations, it is still unfriendly to real-time workloads, so is | |
734 | * thus not recommended for any sort of common-case code. In fact, if | |
735 | * you are using synchronize_rcu_expedited() in a loop, please restructure | |
736 | * your code to batch your updates, and then Use a single synchronize_rcu() | |
737 | * instead. | |
019129d5 PM |
738 | */ |
739 | void synchronize_rcu_expedited(void) | |
740 | { | |
d9a3da06 | 741 | struct rcu_node *rnp; |
29fd9309 | 742 | struct rcu_node *rnp_unlock; |
e63c887c | 743 | struct rcu_state *rsp = rcu_state_p; |
543c6158 | 744 | unsigned long s; |
d9a3da06 | 745 | |
5a9be7c6 PM |
746 | /* If expedited grace periods are prohibited, fall back to normal. */ |
747 | if (rcu_gp_is_normal()) { | |
748 | wait_rcu_gp(call_rcu); | |
749 | return; | |
750 | } | |
751 | ||
543c6158 | 752 | s = rcu_exp_gp_seq_snap(rsp); |
4f415302 | 753 | trace_rcu_exp_grace_period(rsp->name, s, TPS("snap")); |
d9a3da06 | 754 | |
29fd9309 PM |
755 | rnp_unlock = exp_funnel_lock(rsp, s); |
756 | if (rnp_unlock == NULL) | |
757 | return; /* Someone else did our work for us. */ | |
1943c89d | 758 | |
543c6158 | 759 | rcu_exp_gp_seq_start(rsp); |
4f415302 | 760 | trace_rcu_exp_grace_period(rsp->name, s, TPS("start")); |
d9a3da06 | 761 | |
b9585e94 | 762 | /* Initialize the rcu_node tree in preparation for the wait. */ |
dcdb8807 | 763 | sync_rcu_exp_select_cpus(rsp, sync_rcu_exp_handler); |
d9a3da06 | 764 | |
12f5f524 | 765 | /* Wait for snapshotted ->blkd_tasks lists to drain. */ |
d9a3da06 | 766 | rnp = rcu_get_root(rsp); |
b08517c7 | 767 | synchronize_sched_expedited_wait(rsp); |
d9a3da06 PM |
768 | |
769 | /* Clean up and exit. */ | |
543c6158 | 770 | rcu_exp_gp_seq_end(rsp); |
4f415302 | 771 | trace_rcu_exp_grace_period(rsp->name, s, TPS("end")); |
29fd9309 | 772 | mutex_unlock(&rnp_unlock->exp_funnel_mutex); |
bea2de44 PM |
773 | trace_rcu_exp_funnel_lock(rsp->name, rnp_unlock->level, |
774 | rnp_unlock->grplo, rnp_unlock->grphi, | |
775 | TPS("rel")); | |
019129d5 PM |
776 | } |
777 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
778 | ||
e74f4c45 PM |
779 | /** |
780 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
f0a0e6f2 PM |
781 | * |
782 | * Note that this primitive does not necessarily wait for an RCU grace period | |
783 | * to complete. For example, if there are no RCU callbacks queued anywhere | |
784 | * in the system, then rcu_barrier() is within its rights to return | |
785 | * immediately, without waiting for anything, much less an RCU grace period. | |
e74f4c45 PM |
786 | */ |
787 | void rcu_barrier(void) | |
788 | { | |
e63c887c | 789 | _rcu_barrier(rcu_state_p); |
e74f4c45 PM |
790 | } |
791 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
792 | ||
1eba8f84 | 793 | /* |
6cc68793 | 794 | * Initialize preemptible RCU's state structures. |
1eba8f84 PM |
795 | */ |
796 | static void __init __rcu_init_preempt(void) | |
797 | { | |
a87f203e | 798 | rcu_init_one(rcu_state_p); |
1eba8f84 PM |
799 | } |
800 | ||
2439b696 PM |
801 | /* |
802 | * Check for a task exiting while in a preemptible-RCU read-side | |
803 | * critical section, clean up if so. No need to issue warnings, | |
804 | * as debug_check_no_locks_held() already does this if lockdep | |
805 | * is enabled. | |
806 | */ | |
807 | void exit_rcu(void) | |
808 | { | |
809 | struct task_struct *t = current; | |
810 | ||
811 | if (likely(list_empty(¤t->rcu_node_entry))) | |
812 | return; | |
813 | t->rcu_read_lock_nesting = 1; | |
814 | barrier(); | |
1d082fd0 | 815 | t->rcu_read_unlock_special.b.blocked = true; |
2439b696 PM |
816 | __rcu_read_unlock(); |
817 | } | |
818 | ||
28f6569a | 819 | #else /* #ifdef CONFIG_PREEMPT_RCU */ |
f41d911f | 820 | |
b28a7c01 | 821 | static struct rcu_state *const rcu_state_p = &rcu_sched_state; |
27f4d280 | 822 | |
f41d911f PM |
823 | /* |
824 | * Tell them what RCU they are running. | |
825 | */ | |
0e0fc1c2 | 826 | static void __init rcu_bootup_announce(void) |
f41d911f | 827 | { |
efc151c3 | 828 | pr_info("Hierarchical RCU implementation.\n"); |
26845c28 | 829 | rcu_bootup_announce_oddness(); |
f41d911f PM |
830 | } |
831 | ||
cba6d0d6 PM |
832 | /* |
833 | * Because preemptible RCU does not exist, we never have to check for | |
834 | * CPUs being in quiescent states. | |
835 | */ | |
38200cf2 | 836 | static void rcu_preempt_note_context_switch(void) |
cba6d0d6 PM |
837 | { |
838 | } | |
839 | ||
fc2219d4 | 840 | /* |
6cc68793 | 841 | * Because preemptible RCU does not exist, there are never any preempted |
fc2219d4 PM |
842 | * RCU readers. |
843 | */ | |
27f4d280 | 844 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 PM |
845 | { |
846 | return 0; | |
847 | } | |
848 | ||
8af3a5e7 PM |
849 | /* |
850 | * Because there is no preemptible RCU, there can be no readers blocked. | |
851 | */ | |
852 | static bool rcu_preempt_has_tasks(struct rcu_node *rnp) | |
b668c9cf | 853 | { |
8af3a5e7 | 854 | return false; |
b668c9cf PM |
855 | } |
856 | ||
1ed509a2 | 857 | /* |
6cc68793 | 858 | * Because preemptible RCU does not exist, we never have to check for |
1ed509a2 PM |
859 | * tasks blocked within RCU read-side critical sections. |
860 | */ | |
861 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
862 | { | |
863 | } | |
864 | ||
f41d911f | 865 | /* |
6cc68793 | 866 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
867 | * tasks blocked within RCU read-side critical sections. |
868 | */ | |
9bc8b558 | 869 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 870 | { |
9bc8b558 | 871 | return 0; |
f41d911f PM |
872 | } |
873 | ||
74611ecb PM |
874 | /* |
875 | * Because preemptible RCU does not exist, we never have to check for | |
876 | * tasks blocked within RCU read-side critical sections that are | |
877 | * blocking the current expedited grace period. | |
878 | */ | |
879 | static int rcu_print_task_exp_stall(struct rcu_node *rnp) | |
880 | { | |
881 | return 0; | |
882 | } | |
883 | ||
b0e165c0 | 884 | /* |
6cc68793 | 885 | * Because there is no preemptible RCU, there can be no readers blocked, |
49e29126 PM |
886 | * so there is no need to check for blocked tasks. So check only for |
887 | * bogus qsmask values. | |
b0e165c0 PM |
888 | */ |
889 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
890 | { | |
49e29126 | 891 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
892 | } |
893 | ||
f41d911f | 894 | /* |
6cc68793 | 895 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
896 | * to check. |
897 | */ | |
86aea0e6 | 898 | static void rcu_preempt_check_callbacks(void) |
f41d911f PM |
899 | { |
900 | } | |
901 | ||
019129d5 PM |
902 | /* |
903 | * Wait for an rcu-preempt grace period, but make it happen quickly. | |
6cc68793 | 904 | * But because preemptible RCU does not exist, map to rcu-sched. |
019129d5 PM |
905 | */ |
906 | void synchronize_rcu_expedited(void) | |
907 | { | |
908 | synchronize_sched_expedited(); | |
909 | } | |
910 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
911 | ||
e74f4c45 | 912 | /* |
6cc68793 | 913 | * Because preemptible RCU does not exist, rcu_barrier() is just |
e74f4c45 PM |
914 | * another name for rcu_barrier_sched(). |
915 | */ | |
916 | void rcu_barrier(void) | |
917 | { | |
918 | rcu_barrier_sched(); | |
919 | } | |
920 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
921 | ||
1eba8f84 | 922 | /* |
6cc68793 | 923 | * Because preemptible RCU does not exist, it need not be initialized. |
1eba8f84 PM |
924 | */ |
925 | static void __init __rcu_init_preempt(void) | |
926 | { | |
927 | } | |
928 | ||
2439b696 PM |
929 | /* |
930 | * Because preemptible RCU does not exist, tasks cannot possibly exit | |
931 | * while in preemptible RCU read-side critical sections. | |
932 | */ | |
933 | void exit_rcu(void) | |
934 | { | |
935 | } | |
936 | ||
28f6569a | 937 | #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ |
8bd93a2c | 938 | |
27f4d280 PM |
939 | #ifdef CONFIG_RCU_BOOST |
940 | ||
1696a8be | 941 | #include "../locking/rtmutex_common.h" |
27f4d280 | 942 | |
0ea1f2eb PM |
943 | #ifdef CONFIG_RCU_TRACE |
944 | ||
945 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
946 | { | |
96e92021 | 947 | if (!rcu_preempt_has_tasks(rnp)) |
0ea1f2eb PM |
948 | rnp->n_balk_blkd_tasks++; |
949 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) | |
950 | rnp->n_balk_exp_gp_tasks++; | |
951 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) | |
952 | rnp->n_balk_boost_tasks++; | |
953 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) | |
954 | rnp->n_balk_notblocked++; | |
955 | else if (rnp->gp_tasks != NULL && | |
a9f4793d | 956 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
0ea1f2eb PM |
957 | rnp->n_balk_notyet++; |
958 | else | |
959 | rnp->n_balk_nos++; | |
960 | } | |
961 | ||
962 | #else /* #ifdef CONFIG_RCU_TRACE */ | |
963 | ||
964 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
965 | { | |
966 | } | |
967 | ||
968 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ | |
969 | ||
5d01bbd1 TG |
970 | static void rcu_wake_cond(struct task_struct *t, int status) |
971 | { | |
972 | /* | |
973 | * If the thread is yielding, only wake it when this | |
974 | * is invoked from idle | |
975 | */ | |
976 | if (status != RCU_KTHREAD_YIELDING || is_idle_task(current)) | |
977 | wake_up_process(t); | |
978 | } | |
979 | ||
27f4d280 PM |
980 | /* |
981 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks | |
982 | * or ->boost_tasks, advancing the pointer to the next task in the | |
983 | * ->blkd_tasks list. | |
984 | * | |
985 | * Note that irqs must be enabled: boosting the task can block. | |
986 | * Returns 1 if there are more tasks needing to be boosted. | |
987 | */ | |
988 | static int rcu_boost(struct rcu_node *rnp) | |
989 | { | |
990 | unsigned long flags; | |
27f4d280 PM |
991 | struct task_struct *t; |
992 | struct list_head *tb; | |
993 | ||
7d0ae808 PM |
994 | if (READ_ONCE(rnp->exp_tasks) == NULL && |
995 | READ_ONCE(rnp->boost_tasks) == NULL) | |
27f4d280 PM |
996 | return 0; /* Nothing left to boost. */ |
997 | ||
2a67e741 | 998 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
27f4d280 PM |
999 | |
1000 | /* | |
1001 | * Recheck under the lock: all tasks in need of boosting | |
1002 | * might exit their RCU read-side critical sections on their own. | |
1003 | */ | |
1004 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { | |
67c583a7 | 1005 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 PM |
1006 | return 0; |
1007 | } | |
1008 | ||
1009 | /* | |
1010 | * Preferentially boost tasks blocking expedited grace periods. | |
1011 | * This cannot starve the normal grace periods because a second | |
1012 | * expedited grace period must boost all blocked tasks, including | |
1013 | * those blocking the pre-existing normal grace period. | |
1014 | */ | |
0ea1f2eb | 1015 | if (rnp->exp_tasks != NULL) { |
27f4d280 | 1016 | tb = rnp->exp_tasks; |
0ea1f2eb PM |
1017 | rnp->n_exp_boosts++; |
1018 | } else { | |
27f4d280 | 1019 | tb = rnp->boost_tasks; |
0ea1f2eb PM |
1020 | rnp->n_normal_boosts++; |
1021 | } | |
1022 | rnp->n_tasks_boosted++; | |
27f4d280 PM |
1023 | |
1024 | /* | |
1025 | * We boost task t by manufacturing an rt_mutex that appears to | |
1026 | * be held by task t. We leave a pointer to that rt_mutex where | |
1027 | * task t can find it, and task t will release the mutex when it | |
1028 | * exits its outermost RCU read-side critical section. Then | |
1029 | * simply acquiring this artificial rt_mutex will boost task | |
1030 | * t's priority. (Thanks to tglx for suggesting this approach!) | |
1031 | * | |
1032 | * Note that task t must acquire rnp->lock to remove itself from | |
1033 | * the ->blkd_tasks list, which it will do from exit() if from | |
1034 | * nowhere else. We therefore are guaranteed that task t will | |
1035 | * stay around at least until we drop rnp->lock. Note that | |
1036 | * rnp->lock also resolves races between our priority boosting | |
1037 | * and task t's exiting its outermost RCU read-side critical | |
1038 | * section. | |
1039 | */ | |
1040 | t = container_of(tb, struct task_struct, rcu_node_entry); | |
abaa93d9 | 1041 | rt_mutex_init_proxy_locked(&rnp->boost_mtx, t); |
67c583a7 | 1042 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
abaa93d9 PM |
1043 | /* Lock only for side effect: boosts task t's priority. */ |
1044 | rt_mutex_lock(&rnp->boost_mtx); | |
1045 | rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */ | |
27f4d280 | 1046 | |
7d0ae808 PM |
1047 | return READ_ONCE(rnp->exp_tasks) != NULL || |
1048 | READ_ONCE(rnp->boost_tasks) != NULL; | |
27f4d280 PM |
1049 | } |
1050 | ||
27f4d280 | 1051 | /* |
bc17ea10 | 1052 | * Priority-boosting kthread, one per leaf rcu_node. |
27f4d280 PM |
1053 | */ |
1054 | static int rcu_boost_kthread(void *arg) | |
1055 | { | |
1056 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
1057 | int spincnt = 0; | |
1058 | int more2boost; | |
1059 | ||
f7f7bac9 | 1060 | trace_rcu_utilization(TPS("Start boost kthread@init")); |
27f4d280 | 1061 | for (;;) { |
d71df90e | 1062 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
f7f7bac9 | 1063 | trace_rcu_utilization(TPS("End boost kthread@rcu_wait")); |
08bca60a | 1064 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
f7f7bac9 | 1065 | trace_rcu_utilization(TPS("Start boost kthread@rcu_wait")); |
d71df90e | 1066 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
27f4d280 PM |
1067 | more2boost = rcu_boost(rnp); |
1068 | if (more2boost) | |
1069 | spincnt++; | |
1070 | else | |
1071 | spincnt = 0; | |
1072 | if (spincnt > 10) { | |
5d01bbd1 | 1073 | rnp->boost_kthread_status = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 1074 | trace_rcu_utilization(TPS("End boost kthread@rcu_yield")); |
5d01bbd1 | 1075 | schedule_timeout_interruptible(2); |
f7f7bac9 | 1076 | trace_rcu_utilization(TPS("Start boost kthread@rcu_yield")); |
27f4d280 PM |
1077 | spincnt = 0; |
1078 | } | |
1079 | } | |
1217ed1b | 1080 | /* NOTREACHED */ |
f7f7bac9 | 1081 | trace_rcu_utilization(TPS("End boost kthread@notreached")); |
27f4d280 PM |
1082 | return 0; |
1083 | } | |
1084 | ||
1085 | /* | |
1086 | * Check to see if it is time to start boosting RCU readers that are | |
1087 | * blocking the current grace period, and, if so, tell the per-rcu_node | |
1088 | * kthread to start boosting them. If there is an expedited grace | |
1089 | * period in progress, it is always time to boost. | |
1090 | * | |
b065a853 PM |
1091 | * The caller must hold rnp->lock, which this function releases. |
1092 | * The ->boost_kthread_task is immortal, so we don't need to worry | |
1093 | * about it going away. | |
27f4d280 | 1094 | */ |
1217ed1b | 1095 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 1096 | __releases(rnp->lock) |
27f4d280 PM |
1097 | { |
1098 | struct task_struct *t; | |
1099 | ||
0ea1f2eb PM |
1100 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
1101 | rnp->n_balk_exp_gp_tasks++; | |
67c583a7 | 1102 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 | 1103 | return; |
0ea1f2eb | 1104 | } |
27f4d280 PM |
1105 | if (rnp->exp_tasks != NULL || |
1106 | (rnp->gp_tasks != NULL && | |
1107 | rnp->boost_tasks == NULL && | |
1108 | rnp->qsmask == 0 && | |
1109 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { | |
1110 | if (rnp->exp_tasks == NULL) | |
1111 | rnp->boost_tasks = rnp->gp_tasks; | |
67c583a7 | 1112 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 | 1113 | t = rnp->boost_kthread_task; |
5d01bbd1 TG |
1114 | if (t) |
1115 | rcu_wake_cond(t, rnp->boost_kthread_status); | |
1217ed1b | 1116 | } else { |
0ea1f2eb | 1117 | rcu_initiate_boost_trace(rnp); |
67c583a7 | 1118 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1217ed1b | 1119 | } |
27f4d280 PM |
1120 | } |
1121 | ||
a46e0899 PM |
1122 | /* |
1123 | * Wake up the per-CPU kthread to invoke RCU callbacks. | |
1124 | */ | |
1125 | static void invoke_rcu_callbacks_kthread(void) | |
1126 | { | |
1127 | unsigned long flags; | |
1128 | ||
1129 | local_irq_save(flags); | |
1130 | __this_cpu_write(rcu_cpu_has_work, 1); | |
1eb52121 | 1131 | if (__this_cpu_read(rcu_cpu_kthread_task) != NULL && |
5d01bbd1 TG |
1132 | current != __this_cpu_read(rcu_cpu_kthread_task)) { |
1133 | rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task), | |
1134 | __this_cpu_read(rcu_cpu_kthread_status)); | |
1135 | } | |
a46e0899 PM |
1136 | local_irq_restore(flags); |
1137 | } | |
1138 | ||
dff1672d PM |
1139 | /* |
1140 | * Is the current CPU running the RCU-callbacks kthread? | |
1141 | * Caller must have preemption disabled. | |
1142 | */ | |
1143 | static bool rcu_is_callbacks_kthread(void) | |
1144 | { | |
c9d4b0af | 1145 | return __this_cpu_read(rcu_cpu_kthread_task) == current; |
dff1672d PM |
1146 | } |
1147 | ||
27f4d280 PM |
1148 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) |
1149 | ||
1150 | /* | |
1151 | * Do priority-boost accounting for the start of a new grace period. | |
1152 | */ | |
1153 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) | |
1154 | { | |
1155 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; | |
1156 | } | |
1157 | ||
27f4d280 PM |
1158 | /* |
1159 | * Create an RCU-boost kthread for the specified node if one does not | |
1160 | * already exist. We only create this kthread for preemptible RCU. | |
1161 | * Returns zero if all is well, a negated errno otherwise. | |
1162 | */ | |
49fb4c62 | 1163 | static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp, |
0aa04b05 | 1164 | struct rcu_node *rnp) |
27f4d280 | 1165 | { |
5d01bbd1 | 1166 | int rnp_index = rnp - &rsp->node[0]; |
27f4d280 PM |
1167 | unsigned long flags; |
1168 | struct sched_param sp; | |
1169 | struct task_struct *t; | |
1170 | ||
e63c887c | 1171 | if (rcu_state_p != rsp) |
27f4d280 | 1172 | return 0; |
5d01bbd1 | 1173 | |
0aa04b05 | 1174 | if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0) |
5d01bbd1 TG |
1175 | return 0; |
1176 | ||
a46e0899 | 1177 | rsp->boost = 1; |
27f4d280 PM |
1178 | if (rnp->boost_kthread_task != NULL) |
1179 | return 0; | |
1180 | t = kthread_create(rcu_boost_kthread, (void *)rnp, | |
5b61b0ba | 1181 | "rcub/%d", rnp_index); |
27f4d280 PM |
1182 | if (IS_ERR(t)) |
1183 | return PTR_ERR(t); | |
2a67e741 | 1184 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
27f4d280 | 1185 | rnp->boost_kthread_task = t; |
67c583a7 | 1186 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
21871d7e | 1187 | sp.sched_priority = kthread_prio; |
27f4d280 | 1188 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); |
9a432736 | 1189 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
27f4d280 PM |
1190 | return 0; |
1191 | } | |
1192 | ||
f8b7fc6b PM |
1193 | static void rcu_kthread_do_work(void) |
1194 | { | |
c9d4b0af CL |
1195 | rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data)); |
1196 | rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data)); | |
f8b7fc6b PM |
1197 | rcu_preempt_do_callbacks(); |
1198 | } | |
1199 | ||
62ab7072 | 1200 | static void rcu_cpu_kthread_setup(unsigned int cpu) |
f8b7fc6b | 1201 | { |
f8b7fc6b | 1202 | struct sched_param sp; |
f8b7fc6b | 1203 | |
21871d7e | 1204 | sp.sched_priority = kthread_prio; |
62ab7072 | 1205 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); |
f8b7fc6b PM |
1206 | } |
1207 | ||
62ab7072 | 1208 | static void rcu_cpu_kthread_park(unsigned int cpu) |
f8b7fc6b | 1209 | { |
62ab7072 | 1210 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; |
f8b7fc6b PM |
1211 | } |
1212 | ||
62ab7072 | 1213 | static int rcu_cpu_kthread_should_run(unsigned int cpu) |
f8b7fc6b | 1214 | { |
c9d4b0af | 1215 | return __this_cpu_read(rcu_cpu_has_work); |
f8b7fc6b PM |
1216 | } |
1217 | ||
1218 | /* | |
1219 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the | |
e0f23060 PM |
1220 | * RCU softirq used in flavors and configurations of RCU that do not |
1221 | * support RCU priority boosting. | |
f8b7fc6b | 1222 | */ |
62ab7072 | 1223 | static void rcu_cpu_kthread(unsigned int cpu) |
f8b7fc6b | 1224 | { |
c9d4b0af CL |
1225 | unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status); |
1226 | char work, *workp = this_cpu_ptr(&rcu_cpu_has_work); | |
62ab7072 | 1227 | int spincnt; |
f8b7fc6b | 1228 | |
62ab7072 | 1229 | for (spincnt = 0; spincnt < 10; spincnt++) { |
f7f7bac9 | 1230 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait")); |
f8b7fc6b | 1231 | local_bh_disable(); |
f8b7fc6b | 1232 | *statusp = RCU_KTHREAD_RUNNING; |
62ab7072 PM |
1233 | this_cpu_inc(rcu_cpu_kthread_loops); |
1234 | local_irq_disable(); | |
f8b7fc6b PM |
1235 | work = *workp; |
1236 | *workp = 0; | |
62ab7072 | 1237 | local_irq_enable(); |
f8b7fc6b PM |
1238 | if (work) |
1239 | rcu_kthread_do_work(); | |
1240 | local_bh_enable(); | |
62ab7072 | 1241 | if (*workp == 0) { |
f7f7bac9 | 1242 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); |
62ab7072 PM |
1243 | *statusp = RCU_KTHREAD_WAITING; |
1244 | return; | |
f8b7fc6b PM |
1245 | } |
1246 | } | |
62ab7072 | 1247 | *statusp = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 1248 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); |
62ab7072 | 1249 | schedule_timeout_interruptible(2); |
f7f7bac9 | 1250 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); |
62ab7072 | 1251 | *statusp = RCU_KTHREAD_WAITING; |
f8b7fc6b PM |
1252 | } |
1253 | ||
1254 | /* | |
1255 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are | |
1256 | * served by the rcu_node in question. The CPU hotplug lock is still | |
1257 | * held, so the value of rnp->qsmaskinit will be stable. | |
1258 | * | |
1259 | * We don't include outgoingcpu in the affinity set, use -1 if there is | |
1260 | * no outgoing CPU. If there are no CPUs left in the affinity set, | |
1261 | * this function allows the kthread to execute on any CPU. | |
1262 | */ | |
5d01bbd1 | 1263 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b | 1264 | { |
5d01bbd1 | 1265 | struct task_struct *t = rnp->boost_kthread_task; |
0aa04b05 | 1266 | unsigned long mask = rcu_rnp_online_cpus(rnp); |
f8b7fc6b PM |
1267 | cpumask_var_t cm; |
1268 | int cpu; | |
f8b7fc6b | 1269 | |
5d01bbd1 | 1270 | if (!t) |
f8b7fc6b | 1271 | return; |
5d01bbd1 | 1272 | if (!zalloc_cpumask_var(&cm, GFP_KERNEL)) |
f8b7fc6b | 1273 | return; |
f8b7fc6b PM |
1274 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) |
1275 | if ((mask & 0x1) && cpu != outgoingcpu) | |
1276 | cpumask_set_cpu(cpu, cm); | |
5d0b0249 | 1277 | if (cpumask_weight(cm) == 0) |
f8b7fc6b | 1278 | cpumask_setall(cm); |
5d01bbd1 | 1279 | set_cpus_allowed_ptr(t, cm); |
f8b7fc6b PM |
1280 | free_cpumask_var(cm); |
1281 | } | |
1282 | ||
62ab7072 PM |
1283 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { |
1284 | .store = &rcu_cpu_kthread_task, | |
1285 | .thread_should_run = rcu_cpu_kthread_should_run, | |
1286 | .thread_fn = rcu_cpu_kthread, | |
1287 | .thread_comm = "rcuc/%u", | |
1288 | .setup = rcu_cpu_kthread_setup, | |
1289 | .park = rcu_cpu_kthread_park, | |
1290 | }; | |
f8b7fc6b PM |
1291 | |
1292 | /* | |
9386c0b7 | 1293 | * Spawn boost kthreads -- called as soon as the scheduler is running. |
f8b7fc6b | 1294 | */ |
9386c0b7 | 1295 | static void __init rcu_spawn_boost_kthreads(void) |
f8b7fc6b | 1296 | { |
f8b7fc6b | 1297 | struct rcu_node *rnp; |
5d01bbd1 | 1298 | int cpu; |
f8b7fc6b | 1299 | |
62ab7072 | 1300 | for_each_possible_cpu(cpu) |
f8b7fc6b | 1301 | per_cpu(rcu_cpu_has_work, cpu) = 0; |
62ab7072 | 1302 | BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec)); |
3e9f5c70 PM |
1303 | rcu_for_each_leaf_node(rcu_state_p, rnp) |
1304 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); | |
f8b7fc6b | 1305 | } |
f8b7fc6b | 1306 | |
49fb4c62 | 1307 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b | 1308 | { |
e534165b | 1309 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
f8b7fc6b PM |
1310 | struct rcu_node *rnp = rdp->mynode; |
1311 | ||
1312 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ | |
62ab7072 | 1313 | if (rcu_scheduler_fully_active) |
e534165b | 1314 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); |
f8b7fc6b PM |
1315 | } |
1316 | ||
27f4d280 PM |
1317 | #else /* #ifdef CONFIG_RCU_BOOST */ |
1318 | ||
1217ed1b | 1319 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 1320 | __releases(rnp->lock) |
27f4d280 | 1321 | { |
67c583a7 | 1322 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 PM |
1323 | } |
1324 | ||
a46e0899 | 1325 | static void invoke_rcu_callbacks_kthread(void) |
27f4d280 | 1326 | { |
a46e0899 | 1327 | WARN_ON_ONCE(1); |
27f4d280 PM |
1328 | } |
1329 | ||
dff1672d PM |
1330 | static bool rcu_is_callbacks_kthread(void) |
1331 | { | |
1332 | return false; | |
1333 | } | |
1334 | ||
27f4d280 PM |
1335 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
1336 | { | |
1337 | } | |
1338 | ||
5d01bbd1 | 1339 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b PM |
1340 | { |
1341 | } | |
1342 | ||
9386c0b7 | 1343 | static void __init rcu_spawn_boost_kthreads(void) |
b0d30417 | 1344 | { |
b0d30417 | 1345 | } |
b0d30417 | 1346 | |
49fb4c62 | 1347 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b PM |
1348 | { |
1349 | } | |
1350 | ||
27f4d280 PM |
1351 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
1352 | ||
8bd93a2c PM |
1353 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
1354 | ||
1355 | /* | |
1356 | * Check to see if any future RCU-related work will need to be done | |
1357 | * by the current CPU, even if none need be done immediately, returning | |
1358 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
1359 | * an exported member of the RCU API. | |
1360 | * | |
7cb92499 PM |
1361 | * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs |
1362 | * any flavor of RCU. | |
8bd93a2c | 1363 | */ |
c1ad348b | 1364 | int rcu_needs_cpu(u64 basemono, u64 *nextevt) |
8bd93a2c | 1365 | { |
c1ad348b | 1366 | *nextevt = KTIME_MAX; |
3382adbc PM |
1367 | return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) |
1368 | ? 0 : rcu_cpu_has_callbacks(NULL); | |
7cb92499 PM |
1369 | } |
1370 | ||
1371 | /* | |
1372 | * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up | |
1373 | * after it. | |
1374 | */ | |
8fa7845d | 1375 | static void rcu_cleanup_after_idle(void) |
7cb92499 PM |
1376 | { |
1377 | } | |
1378 | ||
aea1b35e | 1379 | /* |
a858af28 | 1380 | * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n, |
aea1b35e PM |
1381 | * is nothing. |
1382 | */ | |
198bbf81 | 1383 | static void rcu_prepare_for_idle(void) |
aea1b35e PM |
1384 | { |
1385 | } | |
1386 | ||
c57afe80 PM |
1387 | /* |
1388 | * Don't bother keeping a running count of the number of RCU callbacks | |
1389 | * posted because CONFIG_RCU_FAST_NO_HZ=n. | |
1390 | */ | |
1391 | static void rcu_idle_count_callbacks_posted(void) | |
1392 | { | |
1393 | } | |
1394 | ||
8bd93a2c PM |
1395 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
1396 | ||
f23f7fa1 PM |
1397 | /* |
1398 | * This code is invoked when a CPU goes idle, at which point we want | |
1399 | * to have the CPU do everything required for RCU so that it can enter | |
1400 | * the energy-efficient dyntick-idle mode. This is handled by a | |
1401 | * state machine implemented by rcu_prepare_for_idle() below. | |
1402 | * | |
1403 | * The following three proprocessor symbols control this state machine: | |
1404 | * | |
f23f7fa1 PM |
1405 | * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted |
1406 | * to sleep in dyntick-idle mode with RCU callbacks pending. This | |
1407 | * is sized to be roughly one RCU grace period. Those energy-efficiency | |
1408 | * benchmarkers who might otherwise be tempted to set this to a large | |
1409 | * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your | |
1410 | * system. And if you are -that- concerned about energy efficiency, | |
1411 | * just power the system down and be done with it! | |
778d250a PM |
1412 | * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is |
1413 | * permitted to sleep in dyntick-idle mode with only lazy RCU | |
1414 | * callbacks pending. Setting this too high can OOM your system. | |
f23f7fa1 PM |
1415 | * |
1416 | * The values below work well in practice. If future workloads require | |
1417 | * adjustment, they can be converted into kernel config parameters, though | |
1418 | * making the state machine smarter might be a better option. | |
1419 | */ | |
e84c48ae | 1420 | #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */ |
778d250a | 1421 | #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */ |
f23f7fa1 | 1422 | |
5e44ce35 PM |
1423 | static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY; |
1424 | module_param(rcu_idle_gp_delay, int, 0644); | |
1425 | static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY; | |
1426 | module_param(rcu_idle_lazy_gp_delay, int, 0644); | |
486e2593 | 1427 | |
486e2593 | 1428 | /* |
c229828c PM |
1429 | * Try to advance callbacks for all flavors of RCU on the current CPU, but |
1430 | * only if it has been awhile since the last time we did so. Afterwards, | |
1431 | * if there are any callbacks ready for immediate invocation, return true. | |
486e2593 | 1432 | */ |
f1f399d1 | 1433 | static bool __maybe_unused rcu_try_advance_all_cbs(void) |
486e2593 | 1434 | { |
c0f4dfd4 PM |
1435 | bool cbs_ready = false; |
1436 | struct rcu_data *rdp; | |
c229828c | 1437 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1438 | struct rcu_node *rnp; |
1439 | struct rcu_state *rsp; | |
486e2593 | 1440 | |
c229828c PM |
1441 | /* Exit early if we advanced recently. */ |
1442 | if (jiffies == rdtp->last_advance_all) | |
d0bc90fd | 1443 | return false; |
c229828c PM |
1444 | rdtp->last_advance_all = jiffies; |
1445 | ||
c0f4dfd4 PM |
1446 | for_each_rcu_flavor(rsp) { |
1447 | rdp = this_cpu_ptr(rsp->rda); | |
1448 | rnp = rdp->mynode; | |
486e2593 | 1449 | |
c0f4dfd4 PM |
1450 | /* |
1451 | * Don't bother checking unless a grace period has | |
1452 | * completed since we last checked and there are | |
1453 | * callbacks not yet ready to invoke. | |
1454 | */ | |
e3663b10 | 1455 | if ((rdp->completed != rnp->completed || |
7d0ae808 | 1456 | unlikely(READ_ONCE(rdp->gpwrap))) && |
c0f4dfd4 | 1457 | rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL]) |
470716fc | 1458 | note_gp_changes(rsp, rdp); |
486e2593 | 1459 | |
c0f4dfd4 PM |
1460 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
1461 | cbs_ready = true; | |
1462 | } | |
1463 | return cbs_ready; | |
486e2593 PM |
1464 | } |
1465 | ||
aa9b1630 | 1466 | /* |
c0f4dfd4 PM |
1467 | * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready |
1468 | * to invoke. If the CPU has callbacks, try to advance them. Tell the | |
1469 | * caller to set the timeout based on whether or not there are non-lazy | |
1470 | * callbacks. | |
aa9b1630 | 1471 | * |
c0f4dfd4 | 1472 | * The caller must have disabled interrupts. |
aa9b1630 | 1473 | */ |
c1ad348b | 1474 | int rcu_needs_cpu(u64 basemono, u64 *nextevt) |
aa9b1630 | 1475 | { |
aa6da514 | 1476 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c1ad348b | 1477 | unsigned long dj; |
aa9b1630 | 1478 | |
3382adbc | 1479 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) { |
43224b96 | 1480 | *nextevt = KTIME_MAX; |
3382adbc PM |
1481 | return 0; |
1482 | } | |
1483 | ||
c0f4dfd4 PM |
1484 | /* Snapshot to detect later posting of non-lazy callback. */ |
1485 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
1486 | ||
aa9b1630 | 1487 | /* If no callbacks, RCU doesn't need the CPU. */ |
aa6da514 | 1488 | if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) { |
c1ad348b | 1489 | *nextevt = KTIME_MAX; |
aa9b1630 PM |
1490 | return 0; |
1491 | } | |
c0f4dfd4 PM |
1492 | |
1493 | /* Attempt to advance callbacks. */ | |
1494 | if (rcu_try_advance_all_cbs()) { | |
1495 | /* Some ready to invoke, so initiate later invocation. */ | |
1496 | invoke_rcu_core(); | |
aa9b1630 PM |
1497 | return 1; |
1498 | } | |
c0f4dfd4 PM |
1499 | rdtp->last_accelerate = jiffies; |
1500 | ||
1501 | /* Request timer delay depending on laziness, and round. */ | |
6faf7283 | 1502 | if (!rdtp->all_lazy) { |
c1ad348b | 1503 | dj = round_up(rcu_idle_gp_delay + jiffies, |
c0f4dfd4 | 1504 | rcu_idle_gp_delay) - jiffies; |
e84c48ae | 1505 | } else { |
c1ad348b | 1506 | dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies; |
e84c48ae | 1507 | } |
c1ad348b | 1508 | *nextevt = basemono + dj * TICK_NSEC; |
aa9b1630 PM |
1509 | return 0; |
1510 | } | |
1511 | ||
21e52e15 | 1512 | /* |
c0f4dfd4 PM |
1513 | * Prepare a CPU for idle from an RCU perspective. The first major task |
1514 | * is to sense whether nohz mode has been enabled or disabled via sysfs. | |
1515 | * The second major task is to check to see if a non-lazy callback has | |
1516 | * arrived at a CPU that previously had only lazy callbacks. The third | |
1517 | * major task is to accelerate (that is, assign grace-period numbers to) | |
1518 | * any recently arrived callbacks. | |
aea1b35e PM |
1519 | * |
1520 | * The caller must have disabled interrupts. | |
8bd93a2c | 1521 | */ |
198bbf81 | 1522 | static void rcu_prepare_for_idle(void) |
8bd93a2c | 1523 | { |
48a7639c | 1524 | bool needwake; |
c0f4dfd4 | 1525 | struct rcu_data *rdp; |
198bbf81 | 1526 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1527 | struct rcu_node *rnp; |
1528 | struct rcu_state *rsp; | |
9d2ad243 PM |
1529 | int tne; |
1530 | ||
f0f2e7d3 PM |
1531 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) || |
1532 | rcu_is_nocb_cpu(smp_processor_id())) | |
3382adbc PM |
1533 | return; |
1534 | ||
9d2ad243 | 1535 | /* Handle nohz enablement switches conservatively. */ |
7d0ae808 | 1536 | tne = READ_ONCE(tick_nohz_active); |
9d2ad243 | 1537 | if (tne != rdtp->tick_nohz_enabled_snap) { |
aa6da514 | 1538 | if (rcu_cpu_has_callbacks(NULL)) |
9d2ad243 PM |
1539 | invoke_rcu_core(); /* force nohz to see update. */ |
1540 | rdtp->tick_nohz_enabled_snap = tne; | |
1541 | return; | |
1542 | } | |
1543 | if (!tne) | |
1544 | return; | |
f511fc62 | 1545 | |
c57afe80 | 1546 | /* |
c0f4dfd4 PM |
1547 | * If a non-lazy callback arrived at a CPU having only lazy |
1548 | * callbacks, invoke RCU core for the side-effect of recalculating | |
1549 | * idle duration on re-entry to idle. | |
c57afe80 | 1550 | */ |
c0f4dfd4 PM |
1551 | if (rdtp->all_lazy && |
1552 | rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) { | |
c337f8f5 PM |
1553 | rdtp->all_lazy = false; |
1554 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
c0f4dfd4 | 1555 | invoke_rcu_core(); |
c57afe80 PM |
1556 | return; |
1557 | } | |
c57afe80 | 1558 | |
3084f2f8 | 1559 | /* |
c0f4dfd4 PM |
1560 | * If we have not yet accelerated this jiffy, accelerate all |
1561 | * callbacks on this CPU. | |
3084f2f8 | 1562 | */ |
c0f4dfd4 | 1563 | if (rdtp->last_accelerate == jiffies) |
aea1b35e | 1564 | return; |
c0f4dfd4 PM |
1565 | rdtp->last_accelerate = jiffies; |
1566 | for_each_rcu_flavor(rsp) { | |
198bbf81 | 1567 | rdp = this_cpu_ptr(rsp->rda); |
c0f4dfd4 PM |
1568 | if (!*rdp->nxttail[RCU_DONE_TAIL]) |
1569 | continue; | |
1570 | rnp = rdp->mynode; | |
2a67e741 | 1571 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
48a7639c | 1572 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
67c583a7 | 1573 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
48a7639c PM |
1574 | if (needwake) |
1575 | rcu_gp_kthread_wake(rsp); | |
77e38ed3 | 1576 | } |
c0f4dfd4 | 1577 | } |
3084f2f8 | 1578 | |
c0f4dfd4 PM |
1579 | /* |
1580 | * Clean up for exit from idle. Attempt to advance callbacks based on | |
1581 | * any grace periods that elapsed while the CPU was idle, and if any | |
1582 | * callbacks are now ready to invoke, initiate invocation. | |
1583 | */ | |
8fa7845d | 1584 | static void rcu_cleanup_after_idle(void) |
c0f4dfd4 | 1585 | { |
3382adbc PM |
1586 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) || |
1587 | rcu_is_nocb_cpu(smp_processor_id())) | |
aea1b35e | 1588 | return; |
7a497c96 PM |
1589 | if (rcu_try_advance_all_cbs()) |
1590 | invoke_rcu_core(); | |
8bd93a2c PM |
1591 | } |
1592 | ||
c57afe80 | 1593 | /* |
98248a0e PM |
1594 | * Keep a running count of the number of non-lazy callbacks posted |
1595 | * on this CPU. This running counter (which is never decremented) allows | |
1596 | * rcu_prepare_for_idle() to detect when something out of the idle loop | |
1597 | * posts a callback, even if an equal number of callbacks are invoked. | |
1598 | * Of course, callbacks should only be posted from within a trace event | |
1599 | * designed to be called from idle or from within RCU_NONIDLE(). | |
c57afe80 PM |
1600 | */ |
1601 | static void rcu_idle_count_callbacks_posted(void) | |
1602 | { | |
5955f7ee | 1603 | __this_cpu_add(rcu_dynticks.nonlazy_posted, 1); |
c57afe80 PM |
1604 | } |
1605 | ||
b626c1b6 PM |
1606 | /* |
1607 | * Data for flushing lazy RCU callbacks at OOM time. | |
1608 | */ | |
1609 | static atomic_t oom_callback_count; | |
1610 | static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq); | |
1611 | ||
1612 | /* | |
1613 | * RCU OOM callback -- decrement the outstanding count and deliver the | |
1614 | * wake-up if we are the last one. | |
1615 | */ | |
1616 | static void rcu_oom_callback(struct rcu_head *rhp) | |
1617 | { | |
1618 | if (atomic_dec_and_test(&oom_callback_count)) | |
1619 | wake_up(&oom_callback_wq); | |
1620 | } | |
1621 | ||
1622 | /* | |
1623 | * Post an rcu_oom_notify callback on the current CPU if it has at | |
1624 | * least one lazy callback. This will unnecessarily post callbacks | |
1625 | * to CPUs that already have a non-lazy callback at the end of their | |
1626 | * callback list, but this is an infrequent operation, so accept some | |
1627 | * extra overhead to keep things simple. | |
1628 | */ | |
1629 | static void rcu_oom_notify_cpu(void *unused) | |
1630 | { | |
1631 | struct rcu_state *rsp; | |
1632 | struct rcu_data *rdp; | |
1633 | ||
1634 | for_each_rcu_flavor(rsp) { | |
fa07a58f | 1635 | rdp = raw_cpu_ptr(rsp->rda); |
b626c1b6 PM |
1636 | if (rdp->qlen_lazy != 0) { |
1637 | atomic_inc(&oom_callback_count); | |
1638 | rsp->call(&rdp->oom_head, rcu_oom_callback); | |
1639 | } | |
1640 | } | |
1641 | } | |
1642 | ||
1643 | /* | |
1644 | * If low on memory, ensure that each CPU has a non-lazy callback. | |
1645 | * This will wake up CPUs that have only lazy callbacks, in turn | |
1646 | * ensuring that they free up the corresponding memory in a timely manner. | |
1647 | * Because an uncertain amount of memory will be freed in some uncertain | |
1648 | * timeframe, we do not claim to have freed anything. | |
1649 | */ | |
1650 | static int rcu_oom_notify(struct notifier_block *self, | |
1651 | unsigned long notused, void *nfreed) | |
1652 | { | |
1653 | int cpu; | |
1654 | ||
1655 | /* Wait for callbacks from earlier instance to complete. */ | |
1656 | wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0); | |
78e4bc34 | 1657 | smp_mb(); /* Ensure callback reuse happens after callback invocation. */ |
b626c1b6 PM |
1658 | |
1659 | /* | |
1660 | * Prevent premature wakeup: ensure that all increments happen | |
1661 | * before there is a chance of the counter reaching zero. | |
1662 | */ | |
1663 | atomic_set(&oom_callback_count, 1); | |
1664 | ||
b626c1b6 PM |
1665 | for_each_online_cpu(cpu) { |
1666 | smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1); | |
bde6c3aa | 1667 | cond_resched_rcu_qs(); |
b626c1b6 | 1668 | } |
b626c1b6 PM |
1669 | |
1670 | /* Unconditionally decrement: no need to wake ourselves up. */ | |
1671 | atomic_dec(&oom_callback_count); | |
1672 | ||
1673 | return NOTIFY_OK; | |
1674 | } | |
1675 | ||
1676 | static struct notifier_block rcu_oom_nb = { | |
1677 | .notifier_call = rcu_oom_notify | |
1678 | }; | |
1679 | ||
1680 | static int __init rcu_register_oom_notifier(void) | |
1681 | { | |
1682 | register_oom_notifier(&rcu_oom_nb); | |
1683 | return 0; | |
1684 | } | |
1685 | early_initcall(rcu_register_oom_notifier); | |
1686 | ||
8bd93a2c | 1687 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
a858af28 | 1688 | |
a858af28 PM |
1689 | #ifdef CONFIG_RCU_FAST_NO_HZ |
1690 | ||
1691 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1692 | { | |
5955f7ee | 1693 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
c0f4dfd4 | 1694 | unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap; |
a858af28 | 1695 | |
c0f4dfd4 PM |
1696 | sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c", |
1697 | rdtp->last_accelerate & 0xffff, jiffies & 0xffff, | |
1698 | ulong2long(nlpd), | |
1699 | rdtp->all_lazy ? 'L' : '.', | |
1700 | rdtp->tick_nohz_enabled_snap ? '.' : 'D'); | |
a858af28 PM |
1701 | } |
1702 | ||
1703 | #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1704 | ||
1705 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1706 | { | |
1c17e4d4 | 1707 | *cp = '\0'; |
a858af28 PM |
1708 | } |
1709 | ||
1710 | #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1711 | ||
1712 | /* Initiate the stall-info list. */ | |
1713 | static void print_cpu_stall_info_begin(void) | |
1714 | { | |
efc151c3 | 1715 | pr_cont("\n"); |
a858af28 PM |
1716 | } |
1717 | ||
1718 | /* | |
1719 | * Print out diagnostic information for the specified stalled CPU. | |
1720 | * | |
1721 | * If the specified CPU is aware of the current RCU grace period | |
1722 | * (flavor specified by rsp), then print the number of scheduling | |
1723 | * clock interrupts the CPU has taken during the time that it has | |
1724 | * been aware. Otherwise, print the number of RCU grace periods | |
1725 | * that this CPU is ignorant of, for example, "1" if the CPU was | |
1726 | * aware of the previous grace period. | |
1727 | * | |
1728 | * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info. | |
1729 | */ | |
1730 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) | |
1731 | { | |
1732 | char fast_no_hz[72]; | |
1733 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
1734 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
1735 | char *ticks_title; | |
1736 | unsigned long ticks_value; | |
1737 | ||
1738 | if (rsp->gpnum == rdp->gpnum) { | |
1739 | ticks_title = "ticks this GP"; | |
1740 | ticks_value = rdp->ticks_this_gp; | |
1741 | } else { | |
1742 | ticks_title = "GPs behind"; | |
1743 | ticks_value = rsp->gpnum - rdp->gpnum; | |
1744 | } | |
1745 | print_cpu_stall_fast_no_hz(fast_no_hz, cpu); | |
7f21aeef PM |
1746 | pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n", |
1747 | cpu, | |
1748 | "O."[!!cpu_online(cpu)], | |
1749 | "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)], | |
1750 | "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)], | |
1751 | ticks_value, ticks_title, | |
a858af28 PM |
1752 | atomic_read(&rdtp->dynticks) & 0xfff, |
1753 | rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting, | |
6231069b | 1754 | rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu), |
7d0ae808 | 1755 | READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart, |
a858af28 PM |
1756 | fast_no_hz); |
1757 | } | |
1758 | ||
1759 | /* Terminate the stall-info list. */ | |
1760 | static void print_cpu_stall_info_end(void) | |
1761 | { | |
efc151c3 | 1762 | pr_err("\t"); |
a858af28 PM |
1763 | } |
1764 | ||
1765 | /* Zero ->ticks_this_gp for all flavors of RCU. */ | |
1766 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) | |
1767 | { | |
1768 | rdp->ticks_this_gp = 0; | |
6231069b | 1769 | rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id()); |
a858af28 PM |
1770 | } |
1771 | ||
1772 | /* Increment ->ticks_this_gp for all flavors of RCU. */ | |
1773 | static void increment_cpu_stall_ticks(void) | |
1774 | { | |
115f7a7c PM |
1775 | struct rcu_state *rsp; |
1776 | ||
1777 | for_each_rcu_flavor(rsp) | |
fa07a58f | 1778 | raw_cpu_inc(rsp->rda->ticks_this_gp); |
a858af28 PM |
1779 | } |
1780 | ||
3fbfbf7a PM |
1781 | #ifdef CONFIG_RCU_NOCB_CPU |
1782 | ||
1783 | /* | |
1784 | * Offload callback processing from the boot-time-specified set of CPUs | |
1785 | * specified by rcu_nocb_mask. For each CPU in the set, there is a | |
1786 | * kthread created that pulls the callbacks from the corresponding CPU, | |
1787 | * waits for a grace period to elapse, and invokes the callbacks. | |
1788 | * The no-CBs CPUs do a wake_up() on their kthread when they insert | |
1789 | * a callback into any empty list, unless the rcu_nocb_poll boot parameter | |
1790 | * has been specified, in which case each kthread actively polls its | |
1791 | * CPU. (Which isn't so great for energy efficiency, but which does | |
1792 | * reduce RCU's overhead on that CPU.) | |
1793 | * | |
1794 | * This is intended to be used in conjunction with Frederic Weisbecker's | |
1795 | * adaptive-idle work, which would seriously reduce OS jitter on CPUs | |
1796 | * running CPU-bound user-mode computations. | |
1797 | * | |
1798 | * Offloading of callback processing could also in theory be used as | |
1799 | * an energy-efficiency measure because CPUs with no RCU callbacks | |
1800 | * queued are more aggressive about entering dyntick-idle mode. | |
1801 | */ | |
1802 | ||
1803 | ||
1804 | /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */ | |
1805 | static int __init rcu_nocb_setup(char *str) | |
1806 | { | |
1807 | alloc_bootmem_cpumask_var(&rcu_nocb_mask); | |
1808 | have_rcu_nocb_mask = true; | |
1809 | cpulist_parse(str, rcu_nocb_mask); | |
1810 | return 1; | |
1811 | } | |
1812 | __setup("rcu_nocbs=", rcu_nocb_setup); | |
1813 | ||
1b0048a4 PG |
1814 | static int __init parse_rcu_nocb_poll(char *arg) |
1815 | { | |
1816 | rcu_nocb_poll = 1; | |
1817 | return 0; | |
1818 | } | |
1819 | early_param("rcu_nocb_poll", parse_rcu_nocb_poll); | |
1820 | ||
dae6e64d | 1821 | /* |
0446be48 PM |
1822 | * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended |
1823 | * grace period. | |
dae6e64d | 1824 | */ |
abedf8e2 | 1825 | static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq) |
dae6e64d | 1826 | { |
abedf8e2 | 1827 | swake_up_all(sq); |
dae6e64d PM |
1828 | } |
1829 | ||
1830 | /* | |
8b425aa8 | 1831 | * Set the root rcu_node structure's ->need_future_gp field |
dae6e64d PM |
1832 | * based on the sum of those of all rcu_node structures. This does |
1833 | * double-count the root rcu_node structure's requests, but this | |
1834 | * is necessary to handle the possibility of a rcu_nocb_kthread() | |
1835 | * having awakened during the time that the rcu_node structures | |
1836 | * were being updated for the end of the previous grace period. | |
34ed6246 | 1837 | */ |
dae6e64d PM |
1838 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
1839 | { | |
8b425aa8 | 1840 | rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq; |
dae6e64d PM |
1841 | } |
1842 | ||
abedf8e2 | 1843 | static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp) |
065bb78c DW |
1844 | { |
1845 | return &rnp->nocb_gp_wq[rnp->completed & 0x1]; | |
1846 | } | |
1847 | ||
dae6e64d | 1848 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
34ed6246 | 1849 | { |
abedf8e2 PG |
1850 | init_swait_queue_head(&rnp->nocb_gp_wq[0]); |
1851 | init_swait_queue_head(&rnp->nocb_gp_wq[1]); | |
34ed6246 PM |
1852 | } |
1853 | ||
2f33b512 | 1854 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
24342c96 | 1855 | /* Is the specified CPU a no-CBs CPU? */ |
d1e43fa5 | 1856 | bool rcu_is_nocb_cpu(int cpu) |
3fbfbf7a PM |
1857 | { |
1858 | if (have_rcu_nocb_mask) | |
1859 | return cpumask_test_cpu(cpu, rcu_nocb_mask); | |
1860 | return false; | |
1861 | } | |
2f33b512 | 1862 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
3fbfbf7a | 1863 | |
fbce7497 PM |
1864 | /* |
1865 | * Kick the leader kthread for this NOCB group. | |
1866 | */ | |
1867 | static void wake_nocb_leader(struct rcu_data *rdp, bool force) | |
1868 | { | |
1869 | struct rcu_data *rdp_leader = rdp->nocb_leader; | |
1870 | ||
7d0ae808 | 1871 | if (!READ_ONCE(rdp_leader->nocb_kthread)) |
fbce7497 | 1872 | return; |
7d0ae808 | 1873 | if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) { |
39953dfd | 1874 | /* Prior smp_mb__after_atomic() orders against prior enqueue. */ |
7d0ae808 | 1875 | WRITE_ONCE(rdp_leader->nocb_leader_sleep, false); |
abedf8e2 | 1876 | swake_up(&rdp_leader->nocb_wq); |
fbce7497 PM |
1877 | } |
1878 | } | |
1879 | ||
d7e29933 PM |
1880 | /* |
1881 | * Does the specified CPU need an RCU callback for the specified flavor | |
1882 | * of rcu_barrier()? | |
1883 | */ | |
1884 | static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu) | |
1885 | { | |
1886 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
41050a00 PM |
1887 | unsigned long ret; |
1888 | #ifdef CONFIG_PROVE_RCU | |
d7e29933 | 1889 | struct rcu_head *rhp; |
41050a00 | 1890 | #endif /* #ifdef CONFIG_PROVE_RCU */ |
d7e29933 | 1891 | |
41050a00 PM |
1892 | /* |
1893 | * Check count of all no-CBs callbacks awaiting invocation. | |
1894 | * There needs to be a barrier before this function is called, | |
1895 | * but associated with a prior determination that no more | |
1896 | * callbacks would be posted. In the worst case, the first | |
1897 | * barrier in _rcu_barrier() suffices (but the caller cannot | |
1898 | * necessarily rely on this, not a substitute for the caller | |
1899 | * getting the concurrency design right!). There must also be | |
1900 | * a barrier between the following load an posting of a callback | |
1901 | * (if a callback is in fact needed). This is associated with an | |
1902 | * atomic_inc() in the caller. | |
1903 | */ | |
1904 | ret = atomic_long_read(&rdp->nocb_q_count); | |
d7e29933 | 1905 | |
41050a00 | 1906 | #ifdef CONFIG_PROVE_RCU |
7d0ae808 | 1907 | rhp = READ_ONCE(rdp->nocb_head); |
d7e29933 | 1908 | if (!rhp) |
7d0ae808 | 1909 | rhp = READ_ONCE(rdp->nocb_gp_head); |
d7e29933 | 1910 | if (!rhp) |
7d0ae808 | 1911 | rhp = READ_ONCE(rdp->nocb_follower_head); |
d7e29933 PM |
1912 | |
1913 | /* Having no rcuo kthread but CBs after scheduler starts is bad! */ | |
7d0ae808 | 1914 | if (!READ_ONCE(rdp->nocb_kthread) && rhp && |
59f792d1 | 1915 | rcu_scheduler_fully_active) { |
d7e29933 PM |
1916 | /* RCU callback enqueued before CPU first came online??? */ |
1917 | pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n", | |
1918 | cpu, rhp->func); | |
1919 | WARN_ON_ONCE(1); | |
1920 | } | |
41050a00 | 1921 | #endif /* #ifdef CONFIG_PROVE_RCU */ |
d7e29933 | 1922 | |
41050a00 | 1923 | return !!ret; |
d7e29933 PM |
1924 | } |
1925 | ||
3fbfbf7a PM |
1926 | /* |
1927 | * Enqueue the specified string of rcu_head structures onto the specified | |
1928 | * CPU's no-CBs lists. The CPU is specified by rdp, the head of the | |
1929 | * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy | |
1930 | * counts are supplied by rhcount and rhcount_lazy. | |
1931 | * | |
1932 | * If warranted, also wake up the kthread servicing this CPUs queues. | |
1933 | */ | |
1934 | static void __call_rcu_nocb_enqueue(struct rcu_data *rdp, | |
1935 | struct rcu_head *rhp, | |
1936 | struct rcu_head **rhtp, | |
96d3fd0d PM |
1937 | int rhcount, int rhcount_lazy, |
1938 | unsigned long flags) | |
3fbfbf7a PM |
1939 | { |
1940 | int len; | |
1941 | struct rcu_head **old_rhpp; | |
1942 | struct task_struct *t; | |
1943 | ||
1944 | /* Enqueue the callback on the nocb list and update counts. */ | |
41050a00 PM |
1945 | atomic_long_add(rhcount, &rdp->nocb_q_count); |
1946 | /* rcu_barrier() relies on ->nocb_q_count add before xchg. */ | |
3fbfbf7a | 1947 | old_rhpp = xchg(&rdp->nocb_tail, rhtp); |
7d0ae808 | 1948 | WRITE_ONCE(*old_rhpp, rhp); |
3fbfbf7a | 1949 | atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy); |
39953dfd | 1950 | smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */ |
3fbfbf7a PM |
1951 | |
1952 | /* If we are not being polled and there is a kthread, awaken it ... */ | |
7d0ae808 | 1953 | t = READ_ONCE(rdp->nocb_kthread); |
25e03a74 | 1954 | if (rcu_nocb_poll || !t) { |
9261dd0d PM |
1955 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1956 | TPS("WakeNotPoll")); | |
3fbfbf7a | 1957 | return; |
9261dd0d | 1958 | } |
3fbfbf7a PM |
1959 | len = atomic_long_read(&rdp->nocb_q_count); |
1960 | if (old_rhpp == &rdp->nocb_head) { | |
96d3fd0d | 1961 | if (!irqs_disabled_flags(flags)) { |
fbce7497 PM |
1962 | /* ... if queue was empty ... */ |
1963 | wake_nocb_leader(rdp, false); | |
96d3fd0d PM |
1964 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1965 | TPS("WakeEmpty")); | |
1966 | } else { | |
9fdd3bc9 | 1967 | rdp->nocb_defer_wakeup = RCU_NOGP_WAKE; |
96d3fd0d PM |
1968 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1969 | TPS("WakeEmptyIsDeferred")); | |
1970 | } | |
3fbfbf7a PM |
1971 | rdp->qlen_last_fqs_check = 0; |
1972 | } else if (len > rdp->qlen_last_fqs_check + qhimark) { | |
fbce7497 | 1973 | /* ... or if many callbacks queued. */ |
9fdd3bc9 PM |
1974 | if (!irqs_disabled_flags(flags)) { |
1975 | wake_nocb_leader(rdp, true); | |
1976 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
1977 | TPS("WakeOvf")); | |
1978 | } else { | |
1979 | rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE; | |
1980 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
1981 | TPS("WakeOvfIsDeferred")); | |
1982 | } | |
3fbfbf7a | 1983 | rdp->qlen_last_fqs_check = LONG_MAX / 2; |
9261dd0d PM |
1984 | } else { |
1985 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot")); | |
3fbfbf7a PM |
1986 | } |
1987 | return; | |
1988 | } | |
1989 | ||
1990 | /* | |
1991 | * This is a helper for __call_rcu(), which invokes this when the normal | |
1992 | * callback queue is inoperable. If this is not a no-CBs CPU, this | |
1993 | * function returns failure back to __call_rcu(), which can complain | |
1994 | * appropriately. | |
1995 | * | |
1996 | * Otherwise, this function queues the callback where the corresponding | |
1997 | * "rcuo" kthread can find it. | |
1998 | */ | |
1999 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, | |
96d3fd0d | 2000 | bool lazy, unsigned long flags) |
3fbfbf7a PM |
2001 | { |
2002 | ||
d1e43fa5 | 2003 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
c271d3a9 | 2004 | return false; |
96d3fd0d | 2005 | __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags); |
21e7a608 PM |
2006 | if (__is_kfree_rcu_offset((unsigned long)rhp->func)) |
2007 | trace_rcu_kfree_callback(rdp->rsp->name, rhp, | |
2008 | (unsigned long)rhp->func, | |
756cbf6b PM |
2009 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
2010 | -atomic_long_read(&rdp->nocb_q_count)); | |
21e7a608 PM |
2011 | else |
2012 | trace_rcu_callback(rdp->rsp->name, rhp, | |
756cbf6b PM |
2013 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
2014 | -atomic_long_read(&rdp->nocb_q_count)); | |
1772947b PM |
2015 | |
2016 | /* | |
2017 | * If called from an extended quiescent state with interrupts | |
2018 | * disabled, invoke the RCU core in order to allow the idle-entry | |
2019 | * deferred-wakeup check to function. | |
2020 | */ | |
2021 | if (irqs_disabled_flags(flags) && | |
2022 | !rcu_is_watching() && | |
2023 | cpu_online(smp_processor_id())) | |
2024 | invoke_rcu_core(); | |
2025 | ||
c271d3a9 | 2026 | return true; |
3fbfbf7a PM |
2027 | } |
2028 | ||
2029 | /* | |
2030 | * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is | |
2031 | * not a no-CBs CPU. | |
2032 | */ | |
2033 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
2034 | struct rcu_data *rdp, |
2035 | unsigned long flags) | |
3fbfbf7a PM |
2036 | { |
2037 | long ql = rsp->qlen; | |
2038 | long qll = rsp->qlen_lazy; | |
2039 | ||
2040 | /* If this is not a no-CBs CPU, tell the caller to do it the old way. */ | |
d1e43fa5 | 2041 | if (!rcu_is_nocb_cpu(smp_processor_id())) |
0a9e1e11 | 2042 | return false; |
3fbfbf7a PM |
2043 | rsp->qlen = 0; |
2044 | rsp->qlen_lazy = 0; | |
2045 | ||
2046 | /* First, enqueue the donelist, if any. This preserves CB ordering. */ | |
2047 | if (rsp->orphan_donelist != NULL) { | |
2048 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist, | |
96d3fd0d | 2049 | rsp->orphan_donetail, ql, qll, flags); |
3fbfbf7a PM |
2050 | ql = qll = 0; |
2051 | rsp->orphan_donelist = NULL; | |
2052 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
2053 | } | |
2054 | if (rsp->orphan_nxtlist != NULL) { | |
2055 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist, | |
96d3fd0d | 2056 | rsp->orphan_nxttail, ql, qll, flags); |
3fbfbf7a PM |
2057 | ql = qll = 0; |
2058 | rsp->orphan_nxtlist = NULL; | |
2059 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
2060 | } | |
0a9e1e11 | 2061 | return true; |
3fbfbf7a PM |
2062 | } |
2063 | ||
2064 | /* | |
34ed6246 PM |
2065 | * If necessary, kick off a new grace period, and either way wait |
2066 | * for a subsequent grace period to complete. | |
3fbfbf7a | 2067 | */ |
34ed6246 | 2068 | static void rcu_nocb_wait_gp(struct rcu_data *rdp) |
3fbfbf7a | 2069 | { |
34ed6246 | 2070 | unsigned long c; |
dae6e64d | 2071 | bool d; |
34ed6246 | 2072 | unsigned long flags; |
48a7639c | 2073 | bool needwake; |
34ed6246 PM |
2074 | struct rcu_node *rnp = rdp->mynode; |
2075 | ||
2a67e741 | 2076 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
48a7639c | 2077 | needwake = rcu_start_future_gp(rnp, rdp, &c); |
67c583a7 | 2078 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
48a7639c PM |
2079 | if (needwake) |
2080 | rcu_gp_kthread_wake(rdp->rsp); | |
3fbfbf7a PM |
2081 | |
2082 | /* | |
34ed6246 PM |
2083 | * Wait for the grace period. Do so interruptibly to avoid messing |
2084 | * up the load average. | |
3fbfbf7a | 2085 | */ |
f7f7bac9 | 2086 | trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait")); |
34ed6246 | 2087 | for (;;) { |
abedf8e2 | 2088 | swait_event_interruptible( |
dae6e64d | 2089 | rnp->nocb_gp_wq[c & 0x1], |
7d0ae808 | 2090 | (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c))); |
dae6e64d | 2091 | if (likely(d)) |
34ed6246 | 2092 | break; |
73a860cd | 2093 | WARN_ON(signal_pending(current)); |
f7f7bac9 | 2094 | trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait")); |
34ed6246 | 2095 | } |
f7f7bac9 | 2096 | trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait")); |
34ed6246 | 2097 | smp_mb(); /* Ensure that CB invocation happens after GP end. */ |
3fbfbf7a PM |
2098 | } |
2099 | ||
fbce7497 PM |
2100 | /* |
2101 | * Leaders come here to wait for additional callbacks to show up. | |
2102 | * This function does not return until callbacks appear. | |
2103 | */ | |
2104 | static void nocb_leader_wait(struct rcu_data *my_rdp) | |
2105 | { | |
2106 | bool firsttime = true; | |
2107 | bool gotcbs; | |
2108 | struct rcu_data *rdp; | |
2109 | struct rcu_head **tail; | |
2110 | ||
2111 | wait_again: | |
2112 | ||
2113 | /* Wait for callbacks to appear. */ | |
2114 | if (!rcu_nocb_poll) { | |
2115 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep"); | |
abedf8e2 | 2116 | swait_event_interruptible(my_rdp->nocb_wq, |
7d0ae808 | 2117 | !READ_ONCE(my_rdp->nocb_leader_sleep)); |
fbce7497 PM |
2118 | /* Memory barrier handled by smp_mb() calls below and repoll. */ |
2119 | } else if (firsttime) { | |
2120 | firsttime = false; /* Don't drown trace log with "Poll"! */ | |
2121 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll"); | |
2122 | } | |
2123 | ||
2124 | /* | |
2125 | * Each pass through the following loop checks a follower for CBs. | |
2126 | * We are our own first follower. Any CBs found are moved to | |
2127 | * nocb_gp_head, where they await a grace period. | |
2128 | */ | |
2129 | gotcbs = false; | |
2130 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
7d0ae808 | 2131 | rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head); |
fbce7497 PM |
2132 | if (!rdp->nocb_gp_head) |
2133 | continue; /* No CBs here, try next follower. */ | |
2134 | ||
2135 | /* Move callbacks to wait-for-GP list, which is empty. */ | |
7d0ae808 | 2136 | WRITE_ONCE(rdp->nocb_head, NULL); |
fbce7497 | 2137 | rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head); |
fbce7497 PM |
2138 | gotcbs = true; |
2139 | } | |
2140 | ||
2141 | /* | |
2142 | * If there were no callbacks, sleep a bit, rescan after a | |
2143 | * memory barrier, and go retry. | |
2144 | */ | |
2145 | if (unlikely(!gotcbs)) { | |
2146 | if (!rcu_nocb_poll) | |
2147 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, | |
2148 | "WokeEmpty"); | |
73a860cd | 2149 | WARN_ON(signal_pending(current)); |
fbce7497 PM |
2150 | schedule_timeout_interruptible(1); |
2151 | ||
2152 | /* Rescan in case we were a victim of memory ordering. */ | |
11ed7f93 PK |
2153 | my_rdp->nocb_leader_sleep = true; |
2154 | smp_mb(); /* Ensure _sleep true before scan. */ | |
fbce7497 | 2155 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) |
7d0ae808 | 2156 | if (READ_ONCE(rdp->nocb_head)) { |
fbce7497 | 2157 | /* Found CB, so short-circuit next wait. */ |
11ed7f93 | 2158 | my_rdp->nocb_leader_sleep = false; |
fbce7497 PM |
2159 | break; |
2160 | } | |
2161 | goto wait_again; | |
2162 | } | |
2163 | ||
2164 | /* Wait for one grace period. */ | |
2165 | rcu_nocb_wait_gp(my_rdp); | |
2166 | ||
2167 | /* | |
11ed7f93 PK |
2168 | * We left ->nocb_leader_sleep unset to reduce cache thrashing. |
2169 | * We set it now, but recheck for new callbacks while | |
fbce7497 PM |
2170 | * traversing our follower list. |
2171 | */ | |
11ed7f93 PK |
2172 | my_rdp->nocb_leader_sleep = true; |
2173 | smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */ | |
fbce7497 PM |
2174 | |
2175 | /* Each pass through the following loop wakes a follower, if needed. */ | |
2176 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
7d0ae808 | 2177 | if (READ_ONCE(rdp->nocb_head)) |
11ed7f93 | 2178 | my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/ |
fbce7497 PM |
2179 | if (!rdp->nocb_gp_head) |
2180 | continue; /* No CBs, so no need to wake follower. */ | |
2181 | ||
2182 | /* Append callbacks to follower's "done" list. */ | |
2183 | tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail); | |
2184 | *tail = rdp->nocb_gp_head; | |
c847f142 | 2185 | smp_mb__after_atomic(); /* Store *tail before wakeup. */ |
fbce7497 PM |
2186 | if (rdp != my_rdp && tail == &rdp->nocb_follower_head) { |
2187 | /* | |
2188 | * List was empty, wake up the follower. | |
2189 | * Memory barriers supplied by atomic_long_add(). | |
2190 | */ | |
abedf8e2 | 2191 | swake_up(&rdp->nocb_wq); |
fbce7497 PM |
2192 | } |
2193 | } | |
2194 | ||
2195 | /* If we (the leader) don't have CBs, go wait some more. */ | |
2196 | if (!my_rdp->nocb_follower_head) | |
2197 | goto wait_again; | |
2198 | } | |
2199 | ||
2200 | /* | |
2201 | * Followers come here to wait for additional callbacks to show up. | |
2202 | * This function does not return until callbacks appear. | |
2203 | */ | |
2204 | static void nocb_follower_wait(struct rcu_data *rdp) | |
2205 | { | |
2206 | bool firsttime = true; | |
2207 | ||
2208 | for (;;) { | |
2209 | if (!rcu_nocb_poll) { | |
2210 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2211 | "FollowerSleep"); | |
abedf8e2 | 2212 | swait_event_interruptible(rdp->nocb_wq, |
7d0ae808 | 2213 | READ_ONCE(rdp->nocb_follower_head)); |
fbce7497 PM |
2214 | } else if (firsttime) { |
2215 | /* Don't drown trace log with "Poll"! */ | |
2216 | firsttime = false; | |
2217 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll"); | |
2218 | } | |
2219 | if (smp_load_acquire(&rdp->nocb_follower_head)) { | |
2220 | /* ^^^ Ensure CB invocation follows _head test. */ | |
2221 | return; | |
2222 | } | |
2223 | if (!rcu_nocb_poll) | |
2224 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2225 | "WokeEmpty"); | |
73a860cd | 2226 | WARN_ON(signal_pending(current)); |
fbce7497 PM |
2227 | schedule_timeout_interruptible(1); |
2228 | } | |
2229 | } | |
2230 | ||
3fbfbf7a PM |
2231 | /* |
2232 | * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes | |
fbce7497 PM |
2233 | * callbacks queued by the corresponding no-CBs CPU, however, there is |
2234 | * an optional leader-follower relationship so that the grace-period | |
2235 | * kthreads don't have to do quite so many wakeups. | |
3fbfbf7a PM |
2236 | */ |
2237 | static int rcu_nocb_kthread(void *arg) | |
2238 | { | |
2239 | int c, cl; | |
2240 | struct rcu_head *list; | |
2241 | struct rcu_head *next; | |
2242 | struct rcu_head **tail; | |
2243 | struct rcu_data *rdp = arg; | |
2244 | ||
2245 | /* Each pass through this loop invokes one batch of callbacks */ | |
2246 | for (;;) { | |
fbce7497 PM |
2247 | /* Wait for callbacks. */ |
2248 | if (rdp->nocb_leader == rdp) | |
2249 | nocb_leader_wait(rdp); | |
2250 | else | |
2251 | nocb_follower_wait(rdp); | |
2252 | ||
2253 | /* Pull the ready-to-invoke callbacks onto local list. */ | |
7d0ae808 | 2254 | list = READ_ONCE(rdp->nocb_follower_head); |
fbce7497 PM |
2255 | BUG_ON(!list); |
2256 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty"); | |
7d0ae808 | 2257 | WRITE_ONCE(rdp->nocb_follower_head, NULL); |
fbce7497 | 2258 | tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head); |
3fbfbf7a PM |
2259 | |
2260 | /* Each pass through the following loop invokes a callback. */ | |
41050a00 PM |
2261 | trace_rcu_batch_start(rdp->rsp->name, |
2262 | atomic_long_read(&rdp->nocb_q_count_lazy), | |
2263 | atomic_long_read(&rdp->nocb_q_count), -1); | |
3fbfbf7a PM |
2264 | c = cl = 0; |
2265 | while (list) { | |
2266 | next = list->next; | |
2267 | /* Wait for enqueuing to complete, if needed. */ | |
2268 | while (next == NULL && &list->next != tail) { | |
69a79bb1 PM |
2269 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2270 | TPS("WaitQueue")); | |
3fbfbf7a | 2271 | schedule_timeout_interruptible(1); |
69a79bb1 PM |
2272 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2273 | TPS("WokeQueue")); | |
3fbfbf7a PM |
2274 | next = list->next; |
2275 | } | |
2276 | debug_rcu_head_unqueue(list); | |
2277 | local_bh_disable(); | |
2278 | if (__rcu_reclaim(rdp->rsp->name, list)) | |
2279 | cl++; | |
2280 | c++; | |
2281 | local_bh_enable(); | |
2282 | list = next; | |
2283 | } | |
2284 | trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); | |
41050a00 PM |
2285 | smp_mb__before_atomic(); /* _add after CB invocation. */ |
2286 | atomic_long_add(-c, &rdp->nocb_q_count); | |
2287 | atomic_long_add(-cl, &rdp->nocb_q_count_lazy); | |
c635a4e1 | 2288 | rdp->n_nocbs_invoked += c; |
3fbfbf7a PM |
2289 | } |
2290 | return 0; | |
2291 | } | |
2292 | ||
96d3fd0d | 2293 | /* Is a deferred wakeup of rcu_nocb_kthread() required? */ |
9fdd3bc9 | 2294 | static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
96d3fd0d | 2295 | { |
7d0ae808 | 2296 | return READ_ONCE(rdp->nocb_defer_wakeup); |
96d3fd0d PM |
2297 | } |
2298 | ||
2299 | /* Do a deferred wakeup of rcu_nocb_kthread(). */ | |
2300 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2301 | { | |
9fdd3bc9 PM |
2302 | int ndw; |
2303 | ||
96d3fd0d PM |
2304 | if (!rcu_nocb_need_deferred_wakeup(rdp)) |
2305 | return; | |
7d0ae808 PM |
2306 | ndw = READ_ONCE(rdp->nocb_defer_wakeup); |
2307 | WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT); | |
9fdd3bc9 PM |
2308 | wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE); |
2309 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake")); | |
96d3fd0d PM |
2310 | } |
2311 | ||
f4579fc5 PM |
2312 | void __init rcu_init_nohz(void) |
2313 | { | |
2314 | int cpu; | |
2315 | bool need_rcu_nocb_mask = true; | |
2316 | struct rcu_state *rsp; | |
2317 | ||
2318 | #ifdef CONFIG_RCU_NOCB_CPU_NONE | |
2319 | need_rcu_nocb_mask = false; | |
2320 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */ | |
2321 | ||
2322 | #if defined(CONFIG_NO_HZ_FULL) | |
2323 | if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask)) | |
2324 | need_rcu_nocb_mask = true; | |
2325 | #endif /* #if defined(CONFIG_NO_HZ_FULL) */ | |
2326 | ||
2327 | if (!have_rcu_nocb_mask && need_rcu_nocb_mask) { | |
949cccdb PK |
2328 | if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) { |
2329 | pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n"); | |
2330 | return; | |
2331 | } | |
f4579fc5 PM |
2332 | have_rcu_nocb_mask = true; |
2333 | } | |
2334 | if (!have_rcu_nocb_mask) | |
2335 | return; | |
2336 | ||
2337 | #ifdef CONFIG_RCU_NOCB_CPU_ZERO | |
2338 | pr_info("\tOffload RCU callbacks from CPU 0\n"); | |
2339 | cpumask_set_cpu(0, rcu_nocb_mask); | |
2340 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */ | |
2341 | #ifdef CONFIG_RCU_NOCB_CPU_ALL | |
2342 | pr_info("\tOffload RCU callbacks from all CPUs\n"); | |
2343 | cpumask_copy(rcu_nocb_mask, cpu_possible_mask); | |
2344 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */ | |
2345 | #if defined(CONFIG_NO_HZ_FULL) | |
2346 | if (tick_nohz_full_running) | |
2347 | cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask); | |
2348 | #endif /* #if defined(CONFIG_NO_HZ_FULL) */ | |
2349 | ||
2350 | if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) { | |
2351 | pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n"); | |
2352 | cpumask_and(rcu_nocb_mask, cpu_possible_mask, | |
2353 | rcu_nocb_mask); | |
2354 | } | |
ad853b48 TH |
2355 | pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n", |
2356 | cpumask_pr_args(rcu_nocb_mask)); | |
f4579fc5 PM |
2357 | if (rcu_nocb_poll) |
2358 | pr_info("\tPoll for callbacks from no-CBs CPUs.\n"); | |
2359 | ||
2360 | for_each_rcu_flavor(rsp) { | |
34404ca8 PM |
2361 | for_each_cpu(cpu, rcu_nocb_mask) |
2362 | init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu)); | |
35ce7f29 | 2363 | rcu_organize_nocb_kthreads(rsp); |
f4579fc5 | 2364 | } |
96d3fd0d PM |
2365 | } |
2366 | ||
3fbfbf7a PM |
2367 | /* Initialize per-rcu_data variables for no-CBs CPUs. */ |
2368 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) | |
2369 | { | |
2370 | rdp->nocb_tail = &rdp->nocb_head; | |
abedf8e2 | 2371 | init_swait_queue_head(&rdp->nocb_wq); |
fbce7497 | 2372 | rdp->nocb_follower_tail = &rdp->nocb_follower_head; |
3fbfbf7a PM |
2373 | } |
2374 | ||
35ce7f29 PM |
2375 | /* |
2376 | * If the specified CPU is a no-CBs CPU that does not already have its | |
2377 | * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are | |
2378 | * brought online out of order, this can require re-organizing the | |
2379 | * leader-follower relationships. | |
2380 | */ | |
2381 | static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu) | |
2382 | { | |
2383 | struct rcu_data *rdp; | |
2384 | struct rcu_data *rdp_last; | |
2385 | struct rcu_data *rdp_old_leader; | |
2386 | struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu); | |
2387 | struct task_struct *t; | |
2388 | ||
2389 | /* | |
2390 | * If this isn't a no-CBs CPU or if it already has an rcuo kthread, | |
2391 | * then nothing to do. | |
2392 | */ | |
2393 | if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread) | |
2394 | return; | |
2395 | ||
2396 | /* If we didn't spawn the leader first, reorganize! */ | |
2397 | rdp_old_leader = rdp_spawn->nocb_leader; | |
2398 | if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) { | |
2399 | rdp_last = NULL; | |
2400 | rdp = rdp_old_leader; | |
2401 | do { | |
2402 | rdp->nocb_leader = rdp_spawn; | |
2403 | if (rdp_last && rdp != rdp_spawn) | |
2404 | rdp_last->nocb_next_follower = rdp; | |
bbe5d7a9 PM |
2405 | if (rdp == rdp_spawn) { |
2406 | rdp = rdp->nocb_next_follower; | |
2407 | } else { | |
2408 | rdp_last = rdp; | |
2409 | rdp = rdp->nocb_next_follower; | |
2410 | rdp_last->nocb_next_follower = NULL; | |
2411 | } | |
35ce7f29 PM |
2412 | } while (rdp); |
2413 | rdp_spawn->nocb_next_follower = rdp_old_leader; | |
2414 | } | |
2415 | ||
2416 | /* Spawn the kthread for this CPU and RCU flavor. */ | |
2417 | t = kthread_run(rcu_nocb_kthread, rdp_spawn, | |
2418 | "rcuo%c/%d", rsp->abbr, cpu); | |
2419 | BUG_ON(IS_ERR(t)); | |
7d0ae808 | 2420 | WRITE_ONCE(rdp_spawn->nocb_kthread, t); |
35ce7f29 PM |
2421 | } |
2422 | ||
2423 | /* | |
2424 | * If the specified CPU is a no-CBs CPU that does not already have its | |
2425 | * rcuo kthreads, spawn them. | |
2426 | */ | |
2427 | static void rcu_spawn_all_nocb_kthreads(int cpu) | |
2428 | { | |
2429 | struct rcu_state *rsp; | |
2430 | ||
2431 | if (rcu_scheduler_fully_active) | |
2432 | for_each_rcu_flavor(rsp) | |
2433 | rcu_spawn_one_nocb_kthread(rsp, cpu); | |
2434 | } | |
2435 | ||
2436 | /* | |
2437 | * Once the scheduler is running, spawn rcuo kthreads for all online | |
2438 | * no-CBs CPUs. This assumes that the early_initcall()s happen before | |
2439 | * non-boot CPUs come online -- if this changes, we will need to add | |
2440 | * some mutual exclusion. | |
2441 | */ | |
2442 | static void __init rcu_spawn_nocb_kthreads(void) | |
2443 | { | |
2444 | int cpu; | |
2445 | ||
2446 | for_each_online_cpu(cpu) | |
2447 | rcu_spawn_all_nocb_kthreads(cpu); | |
2448 | } | |
2449 | ||
fbce7497 PM |
2450 | /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */ |
2451 | static int rcu_nocb_leader_stride = -1; | |
2452 | module_param(rcu_nocb_leader_stride, int, 0444); | |
2453 | ||
2454 | /* | |
35ce7f29 | 2455 | * Initialize leader-follower relationships for all no-CBs CPU. |
fbce7497 | 2456 | */ |
35ce7f29 | 2457 | static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp) |
3fbfbf7a PM |
2458 | { |
2459 | int cpu; | |
fbce7497 PM |
2460 | int ls = rcu_nocb_leader_stride; |
2461 | int nl = 0; /* Next leader. */ | |
3fbfbf7a | 2462 | struct rcu_data *rdp; |
fbce7497 PM |
2463 | struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */ |
2464 | struct rcu_data *rdp_prev = NULL; | |
3fbfbf7a | 2465 | |
22c2f669 | 2466 | if (!have_rcu_nocb_mask) |
3fbfbf7a | 2467 | return; |
fbce7497 PM |
2468 | if (ls == -1) { |
2469 | ls = int_sqrt(nr_cpu_ids); | |
2470 | rcu_nocb_leader_stride = ls; | |
2471 | } | |
2472 | ||
2473 | /* | |
2474 | * Each pass through this loop sets up one rcu_data structure and | |
2475 | * spawns one rcu_nocb_kthread(). | |
2476 | */ | |
3fbfbf7a PM |
2477 | for_each_cpu(cpu, rcu_nocb_mask) { |
2478 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
fbce7497 PM |
2479 | if (rdp->cpu >= nl) { |
2480 | /* New leader, set up for followers & next leader. */ | |
2481 | nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls; | |
2482 | rdp->nocb_leader = rdp; | |
2483 | rdp_leader = rdp; | |
2484 | } else { | |
2485 | /* Another follower, link to previous leader. */ | |
2486 | rdp->nocb_leader = rdp_leader; | |
2487 | rdp_prev->nocb_next_follower = rdp; | |
2488 | } | |
2489 | rdp_prev = rdp; | |
3fbfbf7a PM |
2490 | } |
2491 | } | |
2492 | ||
2493 | /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */ | |
34ed6246 | 2494 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2495 | { |
22c2f669 | 2496 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
34ed6246 | 2497 | return false; |
22c2f669 | 2498 | |
34404ca8 PM |
2499 | /* If there are early-boot callbacks, move them to nocb lists. */ |
2500 | if (rdp->nxtlist) { | |
2501 | rdp->nocb_head = rdp->nxtlist; | |
2502 | rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2503 | atomic_long_set(&rdp->nocb_q_count, rdp->qlen); | |
2504 | atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy); | |
2505 | rdp->nxtlist = NULL; | |
2506 | rdp->qlen = 0; | |
2507 | rdp->qlen_lazy = 0; | |
2508 | } | |
3fbfbf7a | 2509 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
34ed6246 | 2510 | return true; |
3fbfbf7a PM |
2511 | } |
2512 | ||
34ed6246 PM |
2513 | #else /* #ifdef CONFIG_RCU_NOCB_CPU */ |
2514 | ||
d7e29933 PM |
2515 | static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu) |
2516 | { | |
2517 | WARN_ON_ONCE(1); /* Should be dead code. */ | |
2518 | return false; | |
2519 | } | |
2520 | ||
abedf8e2 | 2521 | static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq) |
3fbfbf7a | 2522 | { |
3fbfbf7a PM |
2523 | } |
2524 | ||
dae6e64d PM |
2525 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
2526 | { | |
2527 | } | |
2528 | ||
abedf8e2 | 2529 | static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp) |
065bb78c DW |
2530 | { |
2531 | return NULL; | |
2532 | } | |
2533 | ||
dae6e64d PM |
2534 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
2535 | { | |
2536 | } | |
3fbfbf7a | 2537 | |
3fbfbf7a | 2538 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
96d3fd0d | 2539 | bool lazy, unsigned long flags) |
3fbfbf7a | 2540 | { |
4afc7e26 | 2541 | return false; |
3fbfbf7a PM |
2542 | } |
2543 | ||
2544 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
2545 | struct rcu_data *rdp, |
2546 | unsigned long flags) | |
3fbfbf7a | 2547 | { |
f4aa84ba | 2548 | return false; |
3fbfbf7a PM |
2549 | } |
2550 | ||
3fbfbf7a PM |
2551 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
2552 | { | |
2553 | } | |
2554 | ||
9fdd3bc9 | 2555 | static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
96d3fd0d PM |
2556 | { |
2557 | return false; | |
2558 | } | |
2559 | ||
2560 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2561 | { | |
2562 | } | |
2563 | ||
35ce7f29 PM |
2564 | static void rcu_spawn_all_nocb_kthreads(int cpu) |
2565 | { | |
2566 | } | |
2567 | ||
2568 | static void __init rcu_spawn_nocb_kthreads(void) | |
3fbfbf7a PM |
2569 | { |
2570 | } | |
2571 | ||
34ed6246 | 2572 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2573 | { |
34ed6246 | 2574 | return false; |
3fbfbf7a PM |
2575 | } |
2576 | ||
2577 | #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */ | |
65d798f0 PM |
2578 | |
2579 | /* | |
2580 | * An adaptive-ticks CPU can potentially execute in kernel mode for an | |
2581 | * arbitrarily long period of time with the scheduling-clock tick turned | |
2582 | * off. RCU will be paying attention to this CPU because it is in the | |
2583 | * kernel, but the CPU cannot be guaranteed to be executing the RCU state | |
2584 | * machine because the scheduling-clock tick has been disabled. Therefore, | |
2585 | * if an adaptive-ticks CPU is failing to respond to the current grace | |
2586 | * period and has not be idle from an RCU perspective, kick it. | |
2587 | */ | |
4a81e832 | 2588 | static void __maybe_unused rcu_kick_nohz_cpu(int cpu) |
65d798f0 PM |
2589 | { |
2590 | #ifdef CONFIG_NO_HZ_FULL | |
2591 | if (tick_nohz_full_cpu(cpu)) | |
2592 | smp_send_reschedule(cpu); | |
2593 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ | |
2594 | } | |
2333210b PM |
2595 | |
2596 | ||
2597 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
2598 | ||
0edd1b17 | 2599 | static int full_sysidle_state; /* Current system-idle state. */ |
d4bd54fb PM |
2600 | #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */ |
2601 | #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */ | |
2602 | #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */ | |
2603 | #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */ | |
2604 | #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */ | |
2605 | ||
eb348b89 PM |
2606 | /* |
2607 | * Invoked to note exit from irq or task transition to idle. Note that | |
2608 | * usermode execution does -not- count as idle here! After all, we want | |
2609 | * to detect full-system idle states, not RCU quiescent states and grace | |
2610 | * periods. The caller must have disabled interrupts. | |
2611 | */ | |
28ced795 | 2612 | static void rcu_sysidle_enter(int irq) |
eb348b89 PM |
2613 | { |
2614 | unsigned long j; | |
28ced795 | 2615 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
eb348b89 | 2616 | |
663e1310 PM |
2617 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2618 | if (!tick_nohz_full_enabled()) | |
2619 | return; | |
2620 | ||
eb348b89 PM |
2621 | /* Adjust nesting, check for fully idle. */ |
2622 | if (irq) { | |
2623 | rdtp->dynticks_idle_nesting--; | |
2624 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2625 | if (rdtp->dynticks_idle_nesting != 0) | |
2626 | return; /* Still not fully idle. */ | |
2627 | } else { | |
2628 | if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) == | |
2629 | DYNTICK_TASK_NEST_VALUE) { | |
2630 | rdtp->dynticks_idle_nesting = 0; | |
2631 | } else { | |
2632 | rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE; | |
2633 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2634 | return; /* Still not fully idle. */ | |
2635 | } | |
2636 | } | |
2637 | ||
2638 | /* Record start of fully idle period. */ | |
2639 | j = jiffies; | |
7d0ae808 | 2640 | WRITE_ONCE(rdtp->dynticks_idle_jiffies, j); |
4e857c58 | 2641 | smp_mb__before_atomic(); |
eb348b89 | 2642 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2643 | smp_mb__after_atomic(); |
eb348b89 PM |
2644 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1); |
2645 | } | |
2646 | ||
0edd1b17 PM |
2647 | /* |
2648 | * Unconditionally force exit from full system-idle state. This is | |
2649 | * invoked when a normal CPU exits idle, but must be called separately | |
2650 | * for the timekeeping CPU (tick_do_timer_cpu). The reason for this | |
2651 | * is that the timekeeping CPU is permitted to take scheduling-clock | |
2652 | * interrupts while the system is in system-idle state, and of course | |
2653 | * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock | |
2654 | * interrupt from any other type of interrupt. | |
2655 | */ | |
2656 | void rcu_sysidle_force_exit(void) | |
2657 | { | |
7d0ae808 | 2658 | int oldstate = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2659 | int newoldstate; |
2660 | ||
2661 | /* | |
2662 | * Each pass through the following loop attempts to exit full | |
2663 | * system-idle state. If contention proves to be a problem, | |
2664 | * a trylock-based contention tree could be used here. | |
2665 | */ | |
2666 | while (oldstate > RCU_SYSIDLE_SHORT) { | |
2667 | newoldstate = cmpxchg(&full_sysidle_state, | |
2668 | oldstate, RCU_SYSIDLE_NOT); | |
2669 | if (oldstate == newoldstate && | |
2670 | oldstate == RCU_SYSIDLE_FULL_NOTED) { | |
2671 | rcu_kick_nohz_cpu(tick_do_timer_cpu); | |
2672 | return; /* We cleared it, done! */ | |
2673 | } | |
2674 | oldstate = newoldstate; | |
2675 | } | |
2676 | smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */ | |
2677 | } | |
2678 | ||
eb348b89 PM |
2679 | /* |
2680 | * Invoked to note entry to irq or task transition from idle. Note that | |
2681 | * usermode execution does -not- count as idle here! The caller must | |
2682 | * have disabled interrupts. | |
2683 | */ | |
28ced795 | 2684 | static void rcu_sysidle_exit(int irq) |
eb348b89 | 2685 | { |
28ced795 CL |
2686 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
2687 | ||
663e1310 PM |
2688 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2689 | if (!tick_nohz_full_enabled()) | |
2690 | return; | |
2691 | ||
eb348b89 PM |
2692 | /* Adjust nesting, check for already non-idle. */ |
2693 | if (irq) { | |
2694 | rdtp->dynticks_idle_nesting++; | |
2695 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2696 | if (rdtp->dynticks_idle_nesting != 1) | |
2697 | return; /* Already non-idle. */ | |
2698 | } else { | |
2699 | /* | |
2700 | * Allow for irq misnesting. Yes, it really is possible | |
2701 | * to enter an irq handler then never leave it, and maybe | |
2702 | * also vice versa. Handle both possibilities. | |
2703 | */ | |
2704 | if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) { | |
2705 | rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE; | |
2706 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2707 | return; /* Already non-idle. */ | |
2708 | } else { | |
2709 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE; | |
2710 | } | |
2711 | } | |
2712 | ||
2713 | /* Record end of idle period. */ | |
4e857c58 | 2714 | smp_mb__before_atomic(); |
eb348b89 | 2715 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2716 | smp_mb__after_atomic(); |
eb348b89 | 2717 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1)); |
0edd1b17 PM |
2718 | |
2719 | /* | |
2720 | * If we are the timekeeping CPU, we are permitted to be non-idle | |
2721 | * during a system-idle state. This must be the case, because | |
2722 | * the timekeeping CPU has to take scheduling-clock interrupts | |
2723 | * during the time that the system is transitioning to full | |
2724 | * system-idle state. This means that the timekeeping CPU must | |
2725 | * invoke rcu_sysidle_force_exit() directly if it does anything | |
2726 | * more than take a scheduling-clock interrupt. | |
2727 | */ | |
2728 | if (smp_processor_id() == tick_do_timer_cpu) | |
2729 | return; | |
2730 | ||
2731 | /* Update system-idle state: We are clearly no longer fully idle! */ | |
2732 | rcu_sysidle_force_exit(); | |
2733 | } | |
2734 | ||
2735 | /* | |
2736 | * Check to see if the current CPU is idle. Note that usermode execution | |
5871968d PM |
2737 | * does not count as idle. The caller must have disabled interrupts, |
2738 | * and must be running on tick_do_timer_cpu. | |
0edd1b17 PM |
2739 | */ |
2740 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, | |
2741 | unsigned long *maxj) | |
2742 | { | |
2743 | int cur; | |
2744 | unsigned long j; | |
2745 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
2746 | ||
663e1310 PM |
2747 | /* If there are no nohz_full= CPUs, don't check system-wide idleness. */ |
2748 | if (!tick_nohz_full_enabled()) | |
2749 | return; | |
2750 | ||
0edd1b17 PM |
2751 | /* |
2752 | * If some other CPU has already reported non-idle, if this is | |
2753 | * not the flavor of RCU that tracks sysidle state, or if this | |
2754 | * is an offline or the timekeeping CPU, nothing to do. | |
2755 | */ | |
417e8d26 | 2756 | if (!*isidle || rdp->rsp != rcu_state_p || |
0edd1b17 PM |
2757 | cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu) |
2758 | return; | |
5871968d PM |
2759 | /* Verify affinity of current kthread. */ |
2760 | WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu); | |
0edd1b17 PM |
2761 | |
2762 | /* Pick up current idle and NMI-nesting counter and check. */ | |
2763 | cur = atomic_read(&rdtp->dynticks_idle); | |
2764 | if (cur & 0x1) { | |
2765 | *isidle = false; /* We are not idle! */ | |
2766 | return; | |
2767 | } | |
2768 | smp_mb(); /* Read counters before timestamps. */ | |
2769 | ||
2770 | /* Pick up timestamps. */ | |
7d0ae808 | 2771 | j = READ_ONCE(rdtp->dynticks_idle_jiffies); |
0edd1b17 PM |
2772 | /* If this CPU entered idle more recently, update maxj timestamp. */ |
2773 | if (ULONG_CMP_LT(*maxj, j)) | |
2774 | *maxj = j; | |
2775 | } | |
2776 | ||
2777 | /* | |
2778 | * Is this the flavor of RCU that is handling full-system idle? | |
2779 | */ | |
2780 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2781 | { | |
417e8d26 | 2782 | return rsp == rcu_state_p; |
0edd1b17 PM |
2783 | } |
2784 | ||
2785 | /* | |
2786 | * Return a delay in jiffies based on the number of CPUs, rcu_node | |
2787 | * leaf fanout, and jiffies tick rate. The idea is to allow larger | |
2788 | * systems more time to transition to full-idle state in order to | |
2789 | * avoid the cache thrashing that otherwise occur on the state variable. | |
2790 | * Really small systems (less than a couple of tens of CPUs) should | |
2791 | * instead use a single global atomically incremented counter, and later | |
2792 | * versions of this will automatically reconfigure themselves accordingly. | |
2793 | */ | |
2794 | static unsigned long rcu_sysidle_delay(void) | |
2795 | { | |
2796 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2797 | return 0; | |
2798 | return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000); | |
2799 | } | |
2800 | ||
2801 | /* | |
2802 | * Advance the full-system-idle state. This is invoked when all of | |
2803 | * the non-timekeeping CPUs are idle. | |
2804 | */ | |
2805 | static void rcu_sysidle(unsigned long j) | |
2806 | { | |
2807 | /* Check the current state. */ | |
7d0ae808 | 2808 | switch (READ_ONCE(full_sysidle_state)) { |
0edd1b17 PM |
2809 | case RCU_SYSIDLE_NOT: |
2810 | ||
2811 | /* First time all are idle, so note a short idle period. */ | |
7d0ae808 | 2812 | WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT); |
0edd1b17 PM |
2813 | break; |
2814 | ||
2815 | case RCU_SYSIDLE_SHORT: | |
2816 | ||
2817 | /* | |
2818 | * Idle for a bit, time to advance to next state? | |
2819 | * cmpxchg failure means race with non-idle, let them win. | |
2820 | */ | |
2821 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2822 | (void)cmpxchg(&full_sysidle_state, | |
2823 | RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG); | |
2824 | break; | |
2825 | ||
2826 | case RCU_SYSIDLE_LONG: | |
2827 | ||
2828 | /* | |
2829 | * Do an additional check pass before advancing to full. | |
2830 | * cmpxchg failure means race with non-idle, let them win. | |
2831 | */ | |
2832 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2833 | (void)cmpxchg(&full_sysidle_state, | |
2834 | RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL); | |
2835 | break; | |
2836 | ||
2837 | default: | |
2838 | break; | |
2839 | } | |
2840 | } | |
2841 | ||
2842 | /* | |
2843 | * Found a non-idle non-timekeeping CPU, so kick the system-idle state | |
2844 | * back to the beginning. | |
2845 | */ | |
2846 | static void rcu_sysidle_cancel(void) | |
2847 | { | |
2848 | smp_mb(); | |
becb41bf | 2849 | if (full_sysidle_state > RCU_SYSIDLE_SHORT) |
7d0ae808 | 2850 | WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT); |
0edd1b17 PM |
2851 | } |
2852 | ||
2853 | /* | |
2854 | * Update the sysidle state based on the results of a force-quiescent-state | |
2855 | * scan of the CPUs' dyntick-idle state. | |
2856 | */ | |
2857 | static void rcu_sysidle_report(struct rcu_state *rsp, int isidle, | |
2858 | unsigned long maxj, bool gpkt) | |
2859 | { | |
417e8d26 | 2860 | if (rsp != rcu_state_p) |
0edd1b17 PM |
2861 | return; /* Wrong flavor, ignore. */ |
2862 | if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2863 | return; /* Running state machine from timekeeping CPU. */ | |
2864 | if (isidle) | |
2865 | rcu_sysidle(maxj); /* More idle! */ | |
2866 | else | |
2867 | rcu_sysidle_cancel(); /* Idle is over. */ | |
2868 | } | |
2869 | ||
2870 | /* | |
2871 | * Wrapper for rcu_sysidle_report() when called from the grace-period | |
2872 | * kthread's context. | |
2873 | */ | |
2874 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
2875 | unsigned long maxj) | |
2876 | { | |
663e1310 PM |
2877 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2878 | if (!tick_nohz_full_enabled()) | |
2879 | return; | |
2880 | ||
0edd1b17 PM |
2881 | rcu_sysidle_report(rsp, isidle, maxj, true); |
2882 | } | |
2883 | ||
2884 | /* Callback and function for forcing an RCU grace period. */ | |
2885 | struct rcu_sysidle_head { | |
2886 | struct rcu_head rh; | |
2887 | int inuse; | |
2888 | }; | |
2889 | ||
2890 | static void rcu_sysidle_cb(struct rcu_head *rhp) | |
2891 | { | |
2892 | struct rcu_sysidle_head *rshp; | |
2893 | ||
2894 | /* | |
2895 | * The following memory barrier is needed to replace the | |
2896 | * memory barriers that would normally be in the memory | |
2897 | * allocator. | |
2898 | */ | |
2899 | smp_mb(); /* grace period precedes setting inuse. */ | |
2900 | ||
2901 | rshp = container_of(rhp, struct rcu_sysidle_head, rh); | |
7d0ae808 | 2902 | WRITE_ONCE(rshp->inuse, 0); |
0edd1b17 PM |
2903 | } |
2904 | ||
2905 | /* | |
2906 | * Check to see if the system is fully idle, other than the timekeeping CPU. | |
663e1310 PM |
2907 | * The caller must have disabled interrupts. This is not intended to be |
2908 | * called unless tick_nohz_full_enabled(). | |
0edd1b17 PM |
2909 | */ |
2910 | bool rcu_sys_is_idle(void) | |
2911 | { | |
2912 | static struct rcu_sysidle_head rsh; | |
7d0ae808 | 2913 | int rss = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2914 | |
2915 | if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu)) | |
2916 | return false; | |
2917 | ||
2918 | /* Handle small-system case by doing a full scan of CPUs. */ | |
2919 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) { | |
2920 | int oldrss = rss - 1; | |
2921 | ||
2922 | /* | |
2923 | * One pass to advance to each state up to _FULL. | |
2924 | * Give up if any pass fails to advance the state. | |
2925 | */ | |
2926 | while (rss < RCU_SYSIDLE_FULL && oldrss < rss) { | |
2927 | int cpu; | |
2928 | bool isidle = true; | |
2929 | unsigned long maxj = jiffies - ULONG_MAX / 4; | |
2930 | struct rcu_data *rdp; | |
2931 | ||
2932 | /* Scan all the CPUs looking for nonidle CPUs. */ | |
2933 | for_each_possible_cpu(cpu) { | |
417e8d26 | 2934 | rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
0edd1b17 PM |
2935 | rcu_sysidle_check_cpu(rdp, &isidle, &maxj); |
2936 | if (!isidle) | |
2937 | break; | |
2938 | } | |
417e8d26 | 2939 | rcu_sysidle_report(rcu_state_p, isidle, maxj, false); |
0edd1b17 | 2940 | oldrss = rss; |
7d0ae808 | 2941 | rss = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2942 | } |
2943 | } | |
2944 | ||
2945 | /* If this is the first observation of an idle period, record it. */ | |
2946 | if (rss == RCU_SYSIDLE_FULL) { | |
2947 | rss = cmpxchg(&full_sysidle_state, | |
2948 | RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED); | |
2949 | return rss == RCU_SYSIDLE_FULL; | |
2950 | } | |
2951 | ||
2952 | smp_mb(); /* ensure rss load happens before later caller actions. */ | |
2953 | ||
2954 | /* If already fully idle, tell the caller (in case of races). */ | |
2955 | if (rss == RCU_SYSIDLE_FULL_NOTED) | |
2956 | return true; | |
2957 | ||
2958 | /* | |
2959 | * If we aren't there yet, and a grace period is not in flight, | |
2960 | * initiate a grace period. Either way, tell the caller that | |
2961 | * we are not there yet. We use an xchg() rather than an assignment | |
2962 | * to make up for the memory barriers that would otherwise be | |
2963 | * provided by the memory allocator. | |
2964 | */ | |
2965 | if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL && | |
417e8d26 | 2966 | !rcu_gp_in_progress(rcu_state_p) && |
0edd1b17 PM |
2967 | !rsh.inuse && xchg(&rsh.inuse, 1) == 0) |
2968 | call_rcu(&rsh.rh, rcu_sysidle_cb); | |
2969 | return false; | |
eb348b89 PM |
2970 | } |
2971 | ||
2333210b PM |
2972 | /* |
2973 | * Initialize dynticks sysidle state for CPUs coming online. | |
2974 | */ | |
2975 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) | |
2976 | { | |
2977 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE; | |
2978 | } | |
2979 | ||
2980 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
2981 | ||
28ced795 | 2982 | static void rcu_sysidle_enter(int irq) |
eb348b89 PM |
2983 | { |
2984 | } | |
2985 | ||
28ced795 | 2986 | static void rcu_sysidle_exit(int irq) |
eb348b89 PM |
2987 | { |
2988 | } | |
2989 | ||
0edd1b17 PM |
2990 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, |
2991 | unsigned long *maxj) | |
2992 | { | |
2993 | } | |
2994 | ||
2995 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2996 | { | |
2997 | return false; | |
2998 | } | |
2999 | ||
3000 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
3001 | unsigned long maxj) | |
3002 | { | |
3003 | } | |
3004 | ||
2333210b PM |
3005 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) |
3006 | { | |
3007 | } | |
3008 | ||
3009 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
a096932f PM |
3010 | |
3011 | /* | |
3012 | * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the | |
3013 | * grace-period kthread will do force_quiescent_state() processing? | |
3014 | * The idea is to avoid waking up RCU core processing on such a | |
3015 | * CPU unless the grace period has extended for too long. | |
3016 | * | |
3017 | * This code relies on the fact that all NO_HZ_FULL CPUs are also | |
52e2bb95 | 3018 | * CONFIG_RCU_NOCB_CPU CPUs. |
a096932f PM |
3019 | */ |
3020 | static bool rcu_nohz_full_cpu(struct rcu_state *rsp) | |
3021 | { | |
3022 | #ifdef CONFIG_NO_HZ_FULL | |
3023 | if (tick_nohz_full_cpu(smp_processor_id()) && | |
3024 | (!rcu_gp_in_progress(rsp) || | |
7d0ae808 | 3025 | ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ))) |
5ce035fb | 3026 | return true; |
a096932f | 3027 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ |
5ce035fb | 3028 | return false; |
a096932f | 3029 | } |
5057f55e PM |
3030 | |
3031 | /* | |
3032 | * Bind the grace-period kthread for the sysidle flavor of RCU to the | |
3033 | * timekeeping CPU. | |
3034 | */ | |
3035 | static void rcu_bind_gp_kthread(void) | |
3036 | { | |
c0f489d2 | 3037 | int __maybe_unused cpu; |
5057f55e | 3038 | |
c0f489d2 | 3039 | if (!tick_nohz_full_enabled()) |
5057f55e | 3040 | return; |
c0f489d2 PM |
3041 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
3042 | cpu = tick_do_timer_cpu; | |
5871968d | 3043 | if (cpu >= 0 && cpu < nr_cpu_ids) |
5057f55e | 3044 | set_cpus_allowed_ptr(current, cpumask_of(cpu)); |
c0f489d2 | 3045 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
5871968d | 3046 | housekeeping_affine(current); |
c0f489d2 | 3047 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
5057f55e | 3048 | } |
176f8f7a PM |
3049 | |
3050 | /* Record the current task on dyntick-idle entry. */ | |
3051 | static void rcu_dynticks_task_enter(void) | |
3052 | { | |
3053 | #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) | |
7d0ae808 | 3054 | WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id()); |
176f8f7a PM |
3055 | #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ |
3056 | } | |
3057 | ||
3058 | /* Record no current task on dyntick-idle exit. */ | |
3059 | static void rcu_dynticks_task_exit(void) | |
3060 | { | |
3061 | #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) | |
7d0ae808 | 3062 | WRITE_ONCE(current->rcu_tasks_idle_cpu, -1); |
176f8f7a PM |
3063 | #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ |
3064 | } |