]>
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> |
b17b0153 | 30 | #include <linux/sched/debug.h> |
62ab7072 | 31 | #include <linux/smpboot.h> |
ae7e81c0 | 32 | #include <uapi/linux/sched/types.h> |
4102adab | 33 | #include "../time/tick-internal.h" |
f41d911f | 34 | |
5b61b0ba | 35 | #ifdef CONFIG_RCU_BOOST |
61cfd097 | 36 | |
abaa93d9 | 37 | #include "../locking/rtmutex_common.h" |
21871d7e | 38 | |
61cfd097 PM |
39 | /* |
40 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
41 | * handle all flavors of RCU. | |
42 | */ | |
43 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
44 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); | |
45 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); | |
46 | DEFINE_PER_CPU(char, rcu_cpu_has_work); | |
47 | ||
727b705b PM |
48 | #else /* #ifdef CONFIG_RCU_BOOST */ |
49 | ||
50 | /* | |
51 | * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST, | |
52 | * all uses are in dead code. Provide a definition to keep the compiler | |
53 | * happy, but add WARN_ON_ONCE() to complain if used in the wrong place. | |
54 | * This probably needs to be excluded from -rt builds. | |
55 | */ | |
56 | #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; }) | |
57 | ||
58 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ | |
5b61b0ba | 59 | |
3fbfbf7a PM |
60 | #ifdef CONFIG_RCU_NOCB_CPU |
61 | static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */ | |
62 | static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */ | |
1b0048a4 | 63 | static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */ |
3fbfbf7a PM |
64 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ |
65 | ||
26845c28 PM |
66 | /* |
67 | * Check the RCU kernel configuration parameters and print informative | |
699d4035 | 68 | * messages about anything out of the ordinary. |
26845c28 PM |
69 | */ |
70 | static void __init rcu_bootup_announce_oddness(void) | |
71 | { | |
ab6f5bd6 PM |
72 | if (IS_ENABLED(CONFIG_RCU_TRACE)) |
73 | pr_info("\tRCU debugfs-based tracing is enabled.\n"); | |
05c5df31 PM |
74 | if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) || |
75 | (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32)) | |
ab6f5bd6 | 76 | pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n", |
05c5df31 | 77 | RCU_FANOUT); |
7fa27001 | 78 | if (rcu_fanout_exact) |
ab6f5bd6 PM |
79 | pr_info("\tHierarchical RCU autobalancing is disabled.\n"); |
80 | if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ)) | |
81 | pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n"); | |
82 | if (IS_ENABLED(CONFIG_PROVE_RCU)) | |
83 | pr_info("\tRCU lockdep checking is enabled.\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); | |
59d80fd8 | 95 | rcupdate_announce_bootup_oddness(); |
26845c28 PM |
96 | } |
97 | ||
28f6569a | 98 | #ifdef CONFIG_PREEMPT_RCU |
f41d911f | 99 | |
a41bfeb2 | 100 | RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu); |
b28a7c01 | 101 | static struct rcu_state *const rcu_state_p = &rcu_preempt_state; |
2927a689 | 102 | static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data; |
f41d911f | 103 | |
d19fb8d1 PM |
104 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
105 | bool wake); | |
d9a3da06 | 106 | |
f41d911f PM |
107 | /* |
108 | * Tell them what RCU they are running. | |
109 | */ | |
0e0fc1c2 | 110 | static void __init rcu_bootup_announce(void) |
f41d911f | 111 | { |
efc151c3 | 112 | pr_info("Preemptible hierarchical RCU implementation.\n"); |
26845c28 | 113 | rcu_bootup_announce_oddness(); |
f41d911f PM |
114 | } |
115 | ||
8203d6d0 PM |
116 | /* Flags for rcu_preempt_ctxt_queue() decision table. */ |
117 | #define RCU_GP_TASKS 0x8 | |
118 | #define RCU_EXP_TASKS 0x4 | |
119 | #define RCU_GP_BLKD 0x2 | |
120 | #define RCU_EXP_BLKD 0x1 | |
121 | ||
122 | /* | |
123 | * Queues a task preempted within an RCU-preempt read-side critical | |
124 | * section into the appropriate location within the ->blkd_tasks list, | |
125 | * depending on the states of any ongoing normal and expedited grace | |
126 | * periods. The ->gp_tasks pointer indicates which element the normal | |
127 | * grace period is waiting on (NULL if none), and the ->exp_tasks pointer | |
128 | * indicates which element the expedited grace period is waiting on (again, | |
129 | * NULL if none). If a grace period is waiting on a given element in the | |
130 | * ->blkd_tasks list, it also waits on all subsequent elements. Thus, | |
131 | * adding a task to the tail of the list blocks any grace period that is | |
132 | * already waiting on one of the elements. In contrast, adding a task | |
133 | * to the head of the list won't block any grace period that is already | |
134 | * waiting on one of the elements. | |
135 | * | |
136 | * This queuing is imprecise, and can sometimes make an ongoing grace | |
137 | * period wait for a task that is not strictly speaking blocking it. | |
138 | * Given the choice, we needlessly block a normal grace period rather than | |
139 | * blocking an expedited grace period. | |
140 | * | |
141 | * Note that an endless sequence of expedited grace periods still cannot | |
142 | * indefinitely postpone a normal grace period. Eventually, all of the | |
143 | * fixed number of preempted tasks blocking the normal grace period that are | |
144 | * not also blocking the expedited grace period will resume and complete | |
145 | * their RCU read-side critical sections. At that point, the ->gp_tasks | |
146 | * pointer will equal the ->exp_tasks pointer, at which point the end of | |
147 | * the corresponding expedited grace period will also be the end of the | |
148 | * normal grace period. | |
149 | */ | |
46a5d164 PM |
150 | static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp) |
151 | __releases(rnp->lock) /* But leaves rrupts disabled. */ | |
8203d6d0 PM |
152 | { |
153 | int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) + | |
154 | (rnp->exp_tasks ? RCU_EXP_TASKS : 0) + | |
155 | (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) + | |
156 | (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0); | |
157 | struct task_struct *t = current; | |
158 | ||
159 | /* | |
160 | * Decide where to queue the newly blocked task. In theory, | |
161 | * this could be an if-statement. In practice, when I tried | |
162 | * that, it was quite messy. | |
163 | */ | |
164 | switch (blkd_state) { | |
165 | case 0: | |
166 | case RCU_EXP_TASKS: | |
167 | case RCU_EXP_TASKS + RCU_GP_BLKD: | |
168 | case RCU_GP_TASKS: | |
169 | case RCU_GP_TASKS + RCU_EXP_TASKS: | |
170 | ||
171 | /* | |
172 | * Blocking neither GP, or first task blocking the normal | |
173 | * GP but not blocking the already-waiting expedited GP. | |
174 | * Queue at the head of the list to avoid unnecessarily | |
175 | * blocking the already-waiting GPs. | |
176 | */ | |
177 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); | |
178 | break; | |
179 | ||
180 | case RCU_EXP_BLKD: | |
181 | case RCU_GP_BLKD: | |
182 | case RCU_GP_BLKD + RCU_EXP_BLKD: | |
183 | case RCU_GP_TASKS + RCU_EXP_BLKD: | |
184 | case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
185 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
186 | ||
187 | /* | |
188 | * First task arriving that blocks either GP, or first task | |
189 | * arriving that blocks the expedited GP (with the normal | |
190 | * GP already waiting), or a task arriving that blocks | |
191 | * both GPs with both GPs already waiting. Queue at the | |
192 | * tail of the list to avoid any GP waiting on any of the | |
193 | * already queued tasks that are not blocking it. | |
194 | */ | |
195 | list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks); | |
196 | break; | |
197 | ||
198 | case RCU_EXP_TASKS + RCU_EXP_BLKD: | |
199 | case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
200 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD: | |
201 | ||
202 | /* | |
203 | * Second or subsequent task blocking the expedited GP. | |
204 | * The task either does not block the normal GP, or is the | |
205 | * first task blocking the normal GP. Queue just after | |
206 | * the first task blocking the expedited GP. | |
207 | */ | |
208 | list_add(&t->rcu_node_entry, rnp->exp_tasks); | |
209 | break; | |
210 | ||
211 | case RCU_GP_TASKS + RCU_GP_BLKD: | |
212 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD: | |
213 | ||
214 | /* | |
215 | * Second or subsequent task blocking the normal GP. | |
216 | * The task does not block the expedited GP. Queue just | |
217 | * after the first task blocking the normal GP. | |
218 | */ | |
219 | list_add(&t->rcu_node_entry, rnp->gp_tasks); | |
220 | break; | |
221 | ||
222 | default: | |
223 | ||
224 | /* Yet another exercise in excessive paranoia. */ | |
225 | WARN_ON_ONCE(1); | |
226 | break; | |
227 | } | |
228 | ||
229 | /* | |
230 | * We have now queued the task. If it was the first one to | |
231 | * block either grace period, update the ->gp_tasks and/or | |
232 | * ->exp_tasks pointers, respectively, to reference the newly | |
233 | * blocked tasks. | |
234 | */ | |
235 | if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) | |
236 | rnp->gp_tasks = &t->rcu_node_entry; | |
237 | if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD)) | |
238 | rnp->exp_tasks = &t->rcu_node_entry; | |
67c583a7 | 239 | raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */ |
8203d6d0 PM |
240 | |
241 | /* | |
242 | * Report the quiescent state for the expedited GP. This expedited | |
243 | * GP should not be able to end until we report, so there should be | |
244 | * no need to check for a subsequent expedited GP. (Though we are | |
245 | * still in a quiescent state in any case.) | |
246 | */ | |
247 | if (blkd_state & RCU_EXP_BLKD && | |
248 | t->rcu_read_unlock_special.b.exp_need_qs) { | |
249 | t->rcu_read_unlock_special.b.exp_need_qs = false; | |
250 | rcu_report_exp_rdp(rdp->rsp, rdp, true); | |
251 | } else { | |
252 | WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs); | |
253 | } | |
8203d6d0 PM |
254 | } |
255 | ||
f41d911f | 256 | /* |
6cc68793 | 257 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
f41d911f PM |
258 | * that this just means that the task currently running on the CPU is |
259 | * not in a quiescent state. There might be any number of tasks blocked | |
260 | * while in an RCU read-side critical section. | |
25502a6c | 261 | * |
1d082fd0 PM |
262 | * As with the other rcu_*_qs() functions, callers to this function |
263 | * must disable preemption. | |
f41d911f | 264 | */ |
284a8c93 | 265 | static void rcu_preempt_qs(void) |
f41d911f | 266 | { |
5b74c458 | 267 | if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) { |
284a8c93 | 268 | trace_rcu_grace_period(TPS("rcu_preempt"), |
2927a689 | 269 | __this_cpu_read(rcu_data_p->gpnum), |
284a8c93 | 270 | TPS("cpuqs")); |
5b74c458 | 271 | __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false); |
284a8c93 PM |
272 | barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */ |
273 | current->rcu_read_unlock_special.b.need_qs = false; | |
274 | } | |
f41d911f PM |
275 | } |
276 | ||
277 | /* | |
c3422bea PM |
278 | * We have entered the scheduler, and the current task might soon be |
279 | * context-switched away from. If this task is in an RCU read-side | |
280 | * critical section, we will no longer be able to rely on the CPU to | |
12f5f524 PM |
281 | * record that fact, so we enqueue the task on the blkd_tasks list. |
282 | * The task will dequeue itself when it exits the outermost enclosing | |
283 | * RCU read-side critical section. Therefore, the current grace period | |
284 | * cannot be permitted to complete until the blkd_tasks list entries | |
285 | * predating the current grace period drain, in other words, until | |
286 | * rnp->gp_tasks becomes NULL. | |
c3422bea | 287 | * |
46a5d164 | 288 | * Caller must disable interrupts. |
f41d911f | 289 | */ |
5b72f964 | 290 | static void rcu_preempt_note_context_switch(bool preempt) |
f41d911f PM |
291 | { |
292 | struct task_struct *t = current; | |
f41d911f PM |
293 | struct rcu_data *rdp; |
294 | struct rcu_node *rnp; | |
295 | ||
5b72f964 | 296 | WARN_ON_ONCE(!preempt && t->rcu_read_lock_nesting > 0); |
10f39bb1 | 297 | if (t->rcu_read_lock_nesting > 0 && |
1d082fd0 | 298 | !t->rcu_read_unlock_special.b.blocked) { |
f41d911f PM |
299 | |
300 | /* Possibly blocking in an RCU read-side critical section. */ | |
e63c887c | 301 | rdp = this_cpu_ptr(rcu_state_p->rda); |
f41d911f | 302 | rnp = rdp->mynode; |
46a5d164 | 303 | raw_spin_lock_rcu_node(rnp); |
1d082fd0 | 304 | t->rcu_read_unlock_special.b.blocked = true; |
86848966 | 305 | t->rcu_blocked_node = rnp; |
f41d911f PM |
306 | |
307 | /* | |
8203d6d0 PM |
308 | * Verify the CPU's sanity, trace the preemption, and |
309 | * then queue the task as required based on the states | |
310 | * of any ongoing and expedited grace periods. | |
f41d911f | 311 | */ |
0aa04b05 | 312 | WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0); |
e7d8842e | 313 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
d4c08f2a PM |
314 | trace_rcu_preempt_task(rdp->rsp->name, |
315 | t->pid, | |
316 | (rnp->qsmask & rdp->grpmask) | |
317 | ? rnp->gpnum | |
318 | : rnp->gpnum + 1); | |
46a5d164 | 319 | rcu_preempt_ctxt_queue(rnp, rdp); |
10f39bb1 | 320 | } else if (t->rcu_read_lock_nesting < 0 && |
1d082fd0 | 321 | t->rcu_read_unlock_special.s) { |
10f39bb1 PM |
322 | |
323 | /* | |
324 | * Complete exit from RCU read-side critical section on | |
325 | * behalf of preempted instance of __rcu_read_unlock(). | |
326 | */ | |
327 | rcu_read_unlock_special(t); | |
f41d911f PM |
328 | } |
329 | ||
330 | /* | |
331 | * Either we were not in an RCU read-side critical section to | |
332 | * begin with, or we have now recorded that critical section | |
333 | * globally. Either way, we can now note a quiescent state | |
334 | * for this CPU. Again, if we were in an RCU read-side critical | |
335 | * section, and if that critical section was blocking the current | |
336 | * grace period, then the fact that the task has been enqueued | |
337 | * means that we continue to block the current grace period. | |
338 | */ | |
284a8c93 | 339 | rcu_preempt_qs(); |
f41d911f PM |
340 | } |
341 | ||
fc2219d4 PM |
342 | /* |
343 | * Check for preempted RCU readers blocking the current grace period | |
344 | * for the specified rcu_node structure. If the caller needs a reliable | |
345 | * answer, it must hold the rcu_node's ->lock. | |
346 | */ | |
27f4d280 | 347 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 | 348 | { |
12f5f524 | 349 | return rnp->gp_tasks != NULL; |
fc2219d4 PM |
350 | } |
351 | ||
12f5f524 PM |
352 | /* |
353 | * Advance a ->blkd_tasks-list pointer to the next entry, instead | |
354 | * returning NULL if at the end of the list. | |
355 | */ | |
356 | static struct list_head *rcu_next_node_entry(struct task_struct *t, | |
357 | struct rcu_node *rnp) | |
358 | { | |
359 | struct list_head *np; | |
360 | ||
361 | np = t->rcu_node_entry.next; | |
362 | if (np == &rnp->blkd_tasks) | |
363 | np = NULL; | |
364 | return np; | |
365 | } | |
366 | ||
8af3a5e7 PM |
367 | /* |
368 | * Return true if the specified rcu_node structure has tasks that were | |
369 | * preempted within an RCU read-side critical section. | |
370 | */ | |
371 | static bool rcu_preempt_has_tasks(struct rcu_node *rnp) | |
372 | { | |
373 | return !list_empty(&rnp->blkd_tasks); | |
374 | } | |
375 | ||
b668c9cf PM |
376 | /* |
377 | * Handle special cases during rcu_read_unlock(), such as needing to | |
378 | * notify RCU core processing or task having blocked during the RCU | |
379 | * read-side critical section. | |
380 | */ | |
2a3fa843 | 381 | void rcu_read_unlock_special(struct task_struct *t) |
f41d911f | 382 | { |
b6a932d1 PM |
383 | bool empty_exp; |
384 | bool empty_norm; | |
385 | bool empty_exp_now; | |
f41d911f | 386 | unsigned long flags; |
12f5f524 | 387 | struct list_head *np; |
abaa93d9 | 388 | bool drop_boost_mutex = false; |
8203d6d0 | 389 | struct rcu_data *rdp; |
f41d911f | 390 | struct rcu_node *rnp; |
1d082fd0 | 391 | union rcu_special special; |
f41d911f PM |
392 | |
393 | /* NMI handlers cannot block and cannot safely manipulate state. */ | |
394 | if (in_nmi()) | |
395 | return; | |
396 | ||
397 | local_irq_save(flags); | |
398 | ||
399 | /* | |
8203d6d0 PM |
400 | * If RCU core is waiting for this CPU to exit its critical section, |
401 | * report the fact that it has exited. Because irqs are disabled, | |
1d082fd0 | 402 | * t->rcu_read_unlock_special cannot change. |
f41d911f PM |
403 | */ |
404 | special = t->rcu_read_unlock_special; | |
1d082fd0 | 405 | if (special.b.need_qs) { |
284a8c93 | 406 | rcu_preempt_qs(); |
c0135d07 | 407 | t->rcu_read_unlock_special.b.need_qs = false; |
1d082fd0 | 408 | if (!t->rcu_read_unlock_special.s) { |
79a62f95 LJ |
409 | local_irq_restore(flags); |
410 | return; | |
411 | } | |
f41d911f PM |
412 | } |
413 | ||
8203d6d0 PM |
414 | /* |
415 | * Respond to a request for an expedited grace period, but only if | |
416 | * we were not preempted, meaning that we were running on the same | |
417 | * CPU throughout. If we were preempted, the exp_need_qs flag | |
418 | * would have been cleared at the time of the first preemption, | |
419 | * and the quiescent state would be reported when we were dequeued. | |
420 | */ | |
421 | if (special.b.exp_need_qs) { | |
422 | WARN_ON_ONCE(special.b.blocked); | |
423 | t->rcu_read_unlock_special.b.exp_need_qs = false; | |
424 | rdp = this_cpu_ptr(rcu_state_p->rda); | |
425 | rcu_report_exp_rdp(rcu_state_p, rdp, true); | |
426 | if (!t->rcu_read_unlock_special.s) { | |
427 | local_irq_restore(flags); | |
428 | return; | |
429 | } | |
430 | } | |
431 | ||
79a62f95 | 432 | /* Hardware IRQ handlers cannot block, complain if they get here. */ |
d24209bb PM |
433 | if (in_irq() || in_serving_softirq()) { |
434 | lockdep_rcu_suspicious(__FILE__, __LINE__, | |
435 | "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n"); | |
8203d6d0 | 436 | pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n", |
d24209bb PM |
437 | t->rcu_read_unlock_special.s, |
438 | t->rcu_read_unlock_special.b.blocked, | |
8203d6d0 | 439 | t->rcu_read_unlock_special.b.exp_need_qs, |
d24209bb | 440 | t->rcu_read_unlock_special.b.need_qs); |
f41d911f PM |
441 | local_irq_restore(flags); |
442 | return; | |
443 | } | |
444 | ||
445 | /* Clean up if blocked during RCU read-side critical section. */ | |
1d082fd0 PM |
446 | if (special.b.blocked) { |
447 | t->rcu_read_unlock_special.b.blocked = false; | |
f41d911f | 448 | |
dd5d19ba | 449 | /* |
0a0ba1c9 | 450 | * Remove this task from the list it blocked on. The task |
8ba9153b PM |
451 | * now remains queued on the rcu_node corresponding to the |
452 | * CPU it first blocked on, so there is no longer any need | |
453 | * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia. | |
dd5d19ba | 454 | */ |
8ba9153b PM |
455 | rnp = t->rcu_blocked_node; |
456 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ | |
457 | WARN_ON_ONCE(rnp != t->rcu_blocked_node); | |
74e871ac | 458 | empty_norm = !rcu_preempt_blocked_readers_cgp(rnp); |
8203d6d0 | 459 | empty_exp = sync_rcu_preempt_exp_done(rnp); |
d9a3da06 | 460 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ |
12f5f524 | 461 | np = rcu_next_node_entry(t, rnp); |
f41d911f | 462 | list_del_init(&t->rcu_node_entry); |
82e78d80 | 463 | t->rcu_blocked_node = NULL; |
f7f7bac9 | 464 | trace_rcu_unlock_preempted_task(TPS("rcu_preempt"), |
d4c08f2a | 465 | rnp->gpnum, t->pid); |
12f5f524 PM |
466 | if (&t->rcu_node_entry == rnp->gp_tasks) |
467 | rnp->gp_tasks = np; | |
468 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
469 | rnp->exp_tasks = np; | |
727b705b PM |
470 | if (IS_ENABLED(CONFIG_RCU_BOOST)) { |
471 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
472 | rnp->boost_tasks = np; | |
473 | /* Snapshot ->boost_mtx ownership w/rnp->lock held. */ | |
474 | drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t; | |
475 | } | |
f41d911f PM |
476 | |
477 | /* | |
478 | * If this was the last task on the current list, and if | |
479 | * we aren't waiting on any CPUs, report the quiescent state. | |
389abd48 PM |
480 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock, |
481 | * so we must take a snapshot of the expedited state. | |
f41d911f | 482 | */ |
8203d6d0 | 483 | empty_exp_now = sync_rcu_preempt_exp_done(rnp); |
74e871ac | 484 | if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) { |
f7f7bac9 | 485 | trace_rcu_quiescent_state_report(TPS("preempt_rcu"), |
d4c08f2a PM |
486 | rnp->gpnum, |
487 | 0, rnp->qsmask, | |
488 | rnp->level, | |
489 | rnp->grplo, | |
490 | rnp->grphi, | |
491 | !!rnp->gp_tasks); | |
e63c887c | 492 | rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags); |
c701d5d9 | 493 | } else { |
67c583a7 | 494 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
c701d5d9 | 495 | } |
d9a3da06 | 496 | |
27f4d280 | 497 | /* Unboost if we were boosted. */ |
727b705b | 498 | if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex) |
abaa93d9 | 499 | rt_mutex_unlock(&rnp->boost_mtx); |
27f4d280 | 500 | |
d9a3da06 PM |
501 | /* |
502 | * If this was the last task on the expedited lists, | |
503 | * then we need to report up the rcu_node hierarchy. | |
504 | */ | |
389abd48 | 505 | if (!empty_exp && empty_exp_now) |
e63c887c | 506 | rcu_report_exp_rnp(rcu_state_p, rnp, true); |
b668c9cf PM |
507 | } else { |
508 | local_irq_restore(flags); | |
f41d911f | 509 | } |
f41d911f PM |
510 | } |
511 | ||
1ed509a2 PM |
512 | /* |
513 | * Dump detailed information for all tasks blocking the current RCU | |
514 | * grace period on the specified rcu_node structure. | |
515 | */ | |
516 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) | |
517 | { | |
518 | unsigned long flags; | |
1ed509a2 PM |
519 | struct task_struct *t; |
520 | ||
6cf10081 | 521 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
5fd4dc06 | 522 | if (!rcu_preempt_blocked_readers_cgp(rnp)) { |
67c583a7 | 523 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
5fd4dc06 PM |
524 | return; |
525 | } | |
82efed06 | 526 | t = list_entry(rnp->gp_tasks->prev, |
12f5f524 PM |
527 | struct task_struct, rcu_node_entry); |
528 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) | |
529 | sched_show_task(t); | |
67c583a7 | 530 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1ed509a2 PM |
531 | } |
532 | ||
533 | /* | |
534 | * Dump detailed information for all tasks blocking the current RCU | |
535 | * grace period. | |
536 | */ | |
537 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
538 | { | |
539 | struct rcu_node *rnp = rcu_get_root(rsp); | |
540 | ||
541 | rcu_print_detail_task_stall_rnp(rnp); | |
542 | rcu_for_each_leaf_node(rsp, rnp) | |
543 | rcu_print_detail_task_stall_rnp(rnp); | |
544 | } | |
545 | ||
a858af28 PM |
546 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) |
547 | { | |
efc151c3 | 548 | pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):", |
a858af28 PM |
549 | rnp->level, rnp->grplo, rnp->grphi); |
550 | } | |
551 | ||
552 | static void rcu_print_task_stall_end(void) | |
553 | { | |
efc151c3 | 554 | pr_cont("\n"); |
a858af28 PM |
555 | } |
556 | ||
f41d911f PM |
557 | /* |
558 | * Scan the current list of tasks blocked within RCU read-side critical | |
559 | * sections, printing out the tid of each. | |
560 | */ | |
9bc8b558 | 561 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 562 | { |
f41d911f | 563 | struct task_struct *t; |
9bc8b558 | 564 | int ndetected = 0; |
f41d911f | 565 | |
27f4d280 | 566 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
9bc8b558 | 567 | return 0; |
a858af28 | 568 | rcu_print_task_stall_begin(rnp); |
82efed06 | 569 | t = list_entry(rnp->gp_tasks->prev, |
12f5f524 | 570 | struct task_struct, rcu_node_entry); |
9bc8b558 | 571 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { |
efc151c3 | 572 | pr_cont(" P%d", t->pid); |
9bc8b558 PM |
573 | ndetected++; |
574 | } | |
a858af28 | 575 | rcu_print_task_stall_end(); |
9bc8b558 | 576 | return ndetected; |
f41d911f PM |
577 | } |
578 | ||
74611ecb PM |
579 | /* |
580 | * Scan the current list of tasks blocked within RCU read-side critical | |
581 | * sections, printing out the tid of each that is blocking the current | |
582 | * expedited grace period. | |
583 | */ | |
584 | static int rcu_print_task_exp_stall(struct rcu_node *rnp) | |
585 | { | |
586 | struct task_struct *t; | |
587 | int ndetected = 0; | |
588 | ||
589 | if (!rnp->exp_tasks) | |
590 | return 0; | |
591 | t = list_entry(rnp->exp_tasks->prev, | |
592 | struct task_struct, rcu_node_entry); | |
593 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { | |
594 | pr_cont(" P%d", t->pid); | |
595 | ndetected++; | |
596 | } | |
597 | return ndetected; | |
598 | } | |
599 | ||
b0e165c0 PM |
600 | /* |
601 | * Check that the list of blocked tasks for the newly completed grace | |
602 | * period is in fact empty. It is a serious bug to complete a grace | |
603 | * period that still has RCU readers blocked! This function must be | |
604 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock | |
605 | * must be held by the caller. | |
12f5f524 PM |
606 | * |
607 | * Also, if there are blocked tasks on the list, they automatically | |
608 | * block the newly created grace period, so set up ->gp_tasks accordingly. | |
b0e165c0 PM |
609 | */ |
610 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
611 | { | |
27f4d280 | 612 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
96e92021 | 613 | if (rcu_preempt_has_tasks(rnp)) |
12f5f524 | 614 | rnp->gp_tasks = rnp->blkd_tasks.next; |
28ecd580 | 615 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
616 | } |
617 | ||
f41d911f PM |
618 | /* |
619 | * Check for a quiescent state from the current CPU. When a task blocks, | |
620 | * the task is recorded in the corresponding CPU's rcu_node structure, | |
621 | * which is checked elsewhere. | |
622 | * | |
623 | * Caller must disable hard irqs. | |
624 | */ | |
86aea0e6 | 625 | static void rcu_preempt_check_callbacks(void) |
f41d911f PM |
626 | { |
627 | struct task_struct *t = current; | |
628 | ||
629 | if (t->rcu_read_lock_nesting == 0) { | |
284a8c93 | 630 | rcu_preempt_qs(); |
f41d911f PM |
631 | return; |
632 | } | |
10f39bb1 | 633 | if (t->rcu_read_lock_nesting > 0 && |
97c668b8 | 634 | __this_cpu_read(rcu_data_p->core_needs_qs) && |
5b74c458 | 635 | __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm)) |
1d082fd0 | 636 | t->rcu_read_unlock_special.b.need_qs = true; |
f41d911f PM |
637 | } |
638 | ||
a46e0899 PM |
639 | #ifdef CONFIG_RCU_BOOST |
640 | ||
09223371 SL |
641 | static void rcu_preempt_do_callbacks(void) |
642 | { | |
2927a689 | 643 | rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p)); |
09223371 SL |
644 | } |
645 | ||
a46e0899 PM |
646 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
647 | ||
f41d911f | 648 | /* |
6cc68793 | 649 | * Queue a preemptible-RCU callback for invocation after a grace period. |
f41d911f | 650 | */ |
b6a4ae76 | 651 | void call_rcu(struct rcu_head *head, rcu_callback_t func) |
f41d911f | 652 | { |
e63c887c | 653 | __call_rcu(head, func, rcu_state_p, -1, 0); |
f41d911f PM |
654 | } |
655 | EXPORT_SYMBOL_GPL(call_rcu); | |
656 | ||
6ebb237b PM |
657 | /** |
658 | * synchronize_rcu - wait until a grace period has elapsed. | |
659 | * | |
660 | * Control will return to the caller some time after a full grace | |
661 | * period has elapsed, in other words after all currently executing RCU | |
77d8485a PM |
662 | * read-side critical sections have completed. Note, however, that |
663 | * upon return from synchronize_rcu(), the caller might well be executing | |
664 | * concurrently with new RCU read-side critical sections that began while | |
665 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
666 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
f0a0e6f2 | 667 | * |
e28371c8 PM |
668 | * See the description of synchronize_sched() for more detailed |
669 | * information on memory-ordering guarantees. However, please note | |
670 | * that -only- the memory-ordering guarantees apply. For example, | |
671 | * synchronize_rcu() is -not- guaranteed to wait on things like code | |
672 | * protected by preempt_disable(), instead, synchronize_rcu() is -only- | |
673 | * guaranteed to wait on RCU read-side critical sections, that is, sections | |
674 | * of code protected by rcu_read_lock(). | |
6ebb237b PM |
675 | */ |
676 | void synchronize_rcu(void) | |
677 | { | |
f78f5b90 PM |
678 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
679 | lock_is_held(&rcu_lock_map) || | |
680 | lock_is_held(&rcu_sched_lock_map), | |
681 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
52d7e48b | 682 | if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE) |
6ebb237b | 683 | return; |
5afff48b | 684 | if (rcu_gp_is_expedited()) |
3705b88d AM |
685 | synchronize_rcu_expedited(); |
686 | else | |
687 | wait_rcu_gp(call_rcu); | |
6ebb237b PM |
688 | } |
689 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
690 | ||
e74f4c45 PM |
691 | /** |
692 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
f0a0e6f2 PM |
693 | * |
694 | * Note that this primitive does not necessarily wait for an RCU grace period | |
695 | * to complete. For example, if there are no RCU callbacks queued anywhere | |
696 | * in the system, then rcu_barrier() is within its rights to return | |
697 | * immediately, without waiting for anything, much less an RCU grace period. | |
e74f4c45 PM |
698 | */ |
699 | void rcu_barrier(void) | |
700 | { | |
e63c887c | 701 | _rcu_barrier(rcu_state_p); |
e74f4c45 PM |
702 | } |
703 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
704 | ||
1eba8f84 | 705 | /* |
6cc68793 | 706 | * Initialize preemptible RCU's state structures. |
1eba8f84 PM |
707 | */ |
708 | static void __init __rcu_init_preempt(void) | |
709 | { | |
a87f203e | 710 | rcu_init_one(rcu_state_p); |
1eba8f84 PM |
711 | } |
712 | ||
2439b696 PM |
713 | /* |
714 | * Check for a task exiting while in a preemptible-RCU read-side | |
715 | * critical section, clean up if so. No need to issue warnings, | |
716 | * as debug_check_no_locks_held() already does this if lockdep | |
717 | * is enabled. | |
718 | */ | |
719 | void exit_rcu(void) | |
720 | { | |
721 | struct task_struct *t = current; | |
722 | ||
723 | if (likely(list_empty(¤t->rcu_node_entry))) | |
724 | return; | |
725 | t->rcu_read_lock_nesting = 1; | |
726 | barrier(); | |
1d082fd0 | 727 | t->rcu_read_unlock_special.b.blocked = true; |
2439b696 PM |
728 | __rcu_read_unlock(); |
729 | } | |
730 | ||
28f6569a | 731 | #else /* #ifdef CONFIG_PREEMPT_RCU */ |
f41d911f | 732 | |
b28a7c01 | 733 | static struct rcu_state *const rcu_state_p = &rcu_sched_state; |
27f4d280 | 734 | |
f41d911f PM |
735 | /* |
736 | * Tell them what RCU they are running. | |
737 | */ | |
0e0fc1c2 | 738 | static void __init rcu_bootup_announce(void) |
f41d911f | 739 | { |
efc151c3 | 740 | pr_info("Hierarchical RCU implementation.\n"); |
26845c28 | 741 | rcu_bootup_announce_oddness(); |
f41d911f PM |
742 | } |
743 | ||
cba6d0d6 PM |
744 | /* |
745 | * Because preemptible RCU does not exist, we never have to check for | |
746 | * CPUs being in quiescent states. | |
747 | */ | |
5b72f964 | 748 | static void rcu_preempt_note_context_switch(bool preempt) |
cba6d0d6 PM |
749 | { |
750 | } | |
751 | ||
fc2219d4 | 752 | /* |
6cc68793 | 753 | * Because preemptible RCU does not exist, there are never any preempted |
fc2219d4 PM |
754 | * RCU readers. |
755 | */ | |
27f4d280 | 756 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 PM |
757 | { |
758 | return 0; | |
759 | } | |
760 | ||
8af3a5e7 PM |
761 | /* |
762 | * Because there is no preemptible RCU, there can be no readers blocked. | |
763 | */ | |
764 | static bool rcu_preempt_has_tasks(struct rcu_node *rnp) | |
b668c9cf | 765 | { |
8af3a5e7 | 766 | return false; |
b668c9cf PM |
767 | } |
768 | ||
1ed509a2 | 769 | /* |
6cc68793 | 770 | * Because preemptible RCU does not exist, we never have to check for |
1ed509a2 PM |
771 | * tasks blocked within RCU read-side critical sections. |
772 | */ | |
773 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
774 | { | |
775 | } | |
776 | ||
f41d911f | 777 | /* |
6cc68793 | 778 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
779 | * tasks blocked within RCU read-side critical sections. |
780 | */ | |
9bc8b558 | 781 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 782 | { |
9bc8b558 | 783 | return 0; |
f41d911f PM |
784 | } |
785 | ||
74611ecb PM |
786 | /* |
787 | * Because preemptible RCU does not exist, we never have to check for | |
788 | * tasks blocked within RCU read-side critical sections that are | |
789 | * blocking the current expedited grace period. | |
790 | */ | |
791 | static int rcu_print_task_exp_stall(struct rcu_node *rnp) | |
792 | { | |
793 | return 0; | |
794 | } | |
795 | ||
b0e165c0 | 796 | /* |
6cc68793 | 797 | * Because there is no preemptible RCU, there can be no readers blocked, |
49e29126 PM |
798 | * so there is no need to check for blocked tasks. So check only for |
799 | * bogus qsmask values. | |
b0e165c0 PM |
800 | */ |
801 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
802 | { | |
49e29126 | 803 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
804 | } |
805 | ||
f41d911f | 806 | /* |
6cc68793 | 807 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
808 | * to check. |
809 | */ | |
86aea0e6 | 810 | static void rcu_preempt_check_callbacks(void) |
f41d911f PM |
811 | { |
812 | } | |
813 | ||
e74f4c45 | 814 | /* |
6cc68793 | 815 | * Because preemptible RCU does not exist, rcu_barrier() is just |
e74f4c45 PM |
816 | * another name for rcu_barrier_sched(). |
817 | */ | |
818 | void rcu_barrier(void) | |
819 | { | |
820 | rcu_barrier_sched(); | |
821 | } | |
822 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
823 | ||
1eba8f84 | 824 | /* |
6cc68793 | 825 | * Because preemptible RCU does not exist, it need not be initialized. |
1eba8f84 PM |
826 | */ |
827 | static void __init __rcu_init_preempt(void) | |
828 | { | |
829 | } | |
830 | ||
2439b696 PM |
831 | /* |
832 | * Because preemptible RCU does not exist, tasks cannot possibly exit | |
833 | * while in preemptible RCU read-side critical sections. | |
834 | */ | |
835 | void exit_rcu(void) | |
836 | { | |
837 | } | |
838 | ||
28f6569a | 839 | #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ |
8bd93a2c | 840 | |
27f4d280 PM |
841 | #ifdef CONFIG_RCU_BOOST |
842 | ||
1696a8be | 843 | #include "../locking/rtmutex_common.h" |
27f4d280 | 844 | |
0ea1f2eb PM |
845 | #ifdef CONFIG_RCU_TRACE |
846 | ||
847 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
848 | { | |
96e92021 | 849 | if (!rcu_preempt_has_tasks(rnp)) |
0ea1f2eb PM |
850 | rnp->n_balk_blkd_tasks++; |
851 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) | |
852 | rnp->n_balk_exp_gp_tasks++; | |
853 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) | |
854 | rnp->n_balk_boost_tasks++; | |
855 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) | |
856 | rnp->n_balk_notblocked++; | |
857 | else if (rnp->gp_tasks != NULL && | |
a9f4793d | 858 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
0ea1f2eb PM |
859 | rnp->n_balk_notyet++; |
860 | else | |
861 | rnp->n_balk_nos++; | |
862 | } | |
863 | ||
864 | #else /* #ifdef CONFIG_RCU_TRACE */ | |
865 | ||
866 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
867 | { | |
868 | } | |
869 | ||
870 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ | |
871 | ||
5d01bbd1 TG |
872 | static void rcu_wake_cond(struct task_struct *t, int status) |
873 | { | |
874 | /* | |
875 | * If the thread is yielding, only wake it when this | |
876 | * is invoked from idle | |
877 | */ | |
878 | if (status != RCU_KTHREAD_YIELDING || is_idle_task(current)) | |
879 | wake_up_process(t); | |
880 | } | |
881 | ||
27f4d280 PM |
882 | /* |
883 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks | |
884 | * or ->boost_tasks, advancing the pointer to the next task in the | |
885 | * ->blkd_tasks list. | |
886 | * | |
887 | * Note that irqs must be enabled: boosting the task can block. | |
888 | * Returns 1 if there are more tasks needing to be boosted. | |
889 | */ | |
890 | static int rcu_boost(struct rcu_node *rnp) | |
891 | { | |
892 | unsigned long flags; | |
27f4d280 PM |
893 | struct task_struct *t; |
894 | struct list_head *tb; | |
895 | ||
7d0ae808 PM |
896 | if (READ_ONCE(rnp->exp_tasks) == NULL && |
897 | READ_ONCE(rnp->boost_tasks) == NULL) | |
27f4d280 PM |
898 | return 0; /* Nothing left to boost. */ |
899 | ||
2a67e741 | 900 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
27f4d280 PM |
901 | |
902 | /* | |
903 | * Recheck under the lock: all tasks in need of boosting | |
904 | * might exit their RCU read-side critical sections on their own. | |
905 | */ | |
906 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { | |
67c583a7 | 907 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 PM |
908 | return 0; |
909 | } | |
910 | ||
911 | /* | |
912 | * Preferentially boost tasks blocking expedited grace periods. | |
913 | * This cannot starve the normal grace periods because a second | |
914 | * expedited grace period must boost all blocked tasks, including | |
915 | * those blocking the pre-existing normal grace period. | |
916 | */ | |
0ea1f2eb | 917 | if (rnp->exp_tasks != NULL) { |
27f4d280 | 918 | tb = rnp->exp_tasks; |
0ea1f2eb PM |
919 | rnp->n_exp_boosts++; |
920 | } else { | |
27f4d280 | 921 | tb = rnp->boost_tasks; |
0ea1f2eb PM |
922 | rnp->n_normal_boosts++; |
923 | } | |
924 | rnp->n_tasks_boosted++; | |
27f4d280 PM |
925 | |
926 | /* | |
927 | * We boost task t by manufacturing an rt_mutex that appears to | |
928 | * be held by task t. We leave a pointer to that rt_mutex where | |
929 | * task t can find it, and task t will release the mutex when it | |
930 | * exits its outermost RCU read-side critical section. Then | |
931 | * simply acquiring this artificial rt_mutex will boost task | |
932 | * t's priority. (Thanks to tglx for suggesting this approach!) | |
933 | * | |
934 | * Note that task t must acquire rnp->lock to remove itself from | |
935 | * the ->blkd_tasks list, which it will do from exit() if from | |
936 | * nowhere else. We therefore are guaranteed that task t will | |
937 | * stay around at least until we drop rnp->lock. Note that | |
938 | * rnp->lock also resolves races between our priority boosting | |
939 | * and task t's exiting its outermost RCU read-side critical | |
940 | * section. | |
941 | */ | |
942 | t = container_of(tb, struct task_struct, rcu_node_entry); | |
abaa93d9 | 943 | rt_mutex_init_proxy_locked(&rnp->boost_mtx, t); |
67c583a7 | 944 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
abaa93d9 PM |
945 | /* Lock only for side effect: boosts task t's priority. */ |
946 | rt_mutex_lock(&rnp->boost_mtx); | |
947 | rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */ | |
27f4d280 | 948 | |
7d0ae808 PM |
949 | return READ_ONCE(rnp->exp_tasks) != NULL || |
950 | READ_ONCE(rnp->boost_tasks) != NULL; | |
27f4d280 PM |
951 | } |
952 | ||
27f4d280 | 953 | /* |
bc17ea10 | 954 | * Priority-boosting kthread, one per leaf rcu_node. |
27f4d280 PM |
955 | */ |
956 | static int rcu_boost_kthread(void *arg) | |
957 | { | |
958 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
959 | int spincnt = 0; | |
960 | int more2boost; | |
961 | ||
f7f7bac9 | 962 | trace_rcu_utilization(TPS("Start boost kthread@init")); |
27f4d280 | 963 | for (;;) { |
d71df90e | 964 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
f7f7bac9 | 965 | trace_rcu_utilization(TPS("End boost kthread@rcu_wait")); |
08bca60a | 966 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
f7f7bac9 | 967 | trace_rcu_utilization(TPS("Start boost kthread@rcu_wait")); |
d71df90e | 968 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
27f4d280 PM |
969 | more2boost = rcu_boost(rnp); |
970 | if (more2boost) | |
971 | spincnt++; | |
972 | else | |
973 | spincnt = 0; | |
974 | if (spincnt > 10) { | |
5d01bbd1 | 975 | rnp->boost_kthread_status = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 976 | trace_rcu_utilization(TPS("End boost kthread@rcu_yield")); |
5d01bbd1 | 977 | schedule_timeout_interruptible(2); |
f7f7bac9 | 978 | trace_rcu_utilization(TPS("Start boost kthread@rcu_yield")); |
27f4d280 PM |
979 | spincnt = 0; |
980 | } | |
981 | } | |
1217ed1b | 982 | /* NOTREACHED */ |
f7f7bac9 | 983 | trace_rcu_utilization(TPS("End boost kthread@notreached")); |
27f4d280 PM |
984 | return 0; |
985 | } | |
986 | ||
987 | /* | |
988 | * Check to see if it is time to start boosting RCU readers that are | |
989 | * blocking the current grace period, and, if so, tell the per-rcu_node | |
990 | * kthread to start boosting them. If there is an expedited grace | |
991 | * period in progress, it is always time to boost. | |
992 | * | |
b065a853 PM |
993 | * The caller must hold rnp->lock, which this function releases. |
994 | * The ->boost_kthread_task is immortal, so we don't need to worry | |
995 | * about it going away. | |
27f4d280 | 996 | */ |
1217ed1b | 997 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 998 | __releases(rnp->lock) |
27f4d280 PM |
999 | { |
1000 | struct task_struct *t; | |
1001 | ||
0ea1f2eb PM |
1002 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
1003 | rnp->n_balk_exp_gp_tasks++; | |
67c583a7 | 1004 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 | 1005 | return; |
0ea1f2eb | 1006 | } |
27f4d280 PM |
1007 | if (rnp->exp_tasks != NULL || |
1008 | (rnp->gp_tasks != NULL && | |
1009 | rnp->boost_tasks == NULL && | |
1010 | rnp->qsmask == 0 && | |
1011 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { | |
1012 | if (rnp->exp_tasks == NULL) | |
1013 | rnp->boost_tasks = rnp->gp_tasks; | |
67c583a7 | 1014 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 | 1015 | t = rnp->boost_kthread_task; |
5d01bbd1 TG |
1016 | if (t) |
1017 | rcu_wake_cond(t, rnp->boost_kthread_status); | |
1217ed1b | 1018 | } else { |
0ea1f2eb | 1019 | rcu_initiate_boost_trace(rnp); |
67c583a7 | 1020 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1217ed1b | 1021 | } |
27f4d280 PM |
1022 | } |
1023 | ||
a46e0899 PM |
1024 | /* |
1025 | * Wake up the per-CPU kthread to invoke RCU callbacks. | |
1026 | */ | |
1027 | static void invoke_rcu_callbacks_kthread(void) | |
1028 | { | |
1029 | unsigned long flags; | |
1030 | ||
1031 | local_irq_save(flags); | |
1032 | __this_cpu_write(rcu_cpu_has_work, 1); | |
1eb52121 | 1033 | if (__this_cpu_read(rcu_cpu_kthread_task) != NULL && |
5d01bbd1 TG |
1034 | current != __this_cpu_read(rcu_cpu_kthread_task)) { |
1035 | rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task), | |
1036 | __this_cpu_read(rcu_cpu_kthread_status)); | |
1037 | } | |
a46e0899 PM |
1038 | local_irq_restore(flags); |
1039 | } | |
1040 | ||
dff1672d PM |
1041 | /* |
1042 | * Is the current CPU running the RCU-callbacks kthread? | |
1043 | * Caller must have preemption disabled. | |
1044 | */ | |
1045 | static bool rcu_is_callbacks_kthread(void) | |
1046 | { | |
c9d4b0af | 1047 | return __this_cpu_read(rcu_cpu_kthread_task) == current; |
dff1672d PM |
1048 | } |
1049 | ||
27f4d280 PM |
1050 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) |
1051 | ||
1052 | /* | |
1053 | * Do priority-boost accounting for the start of a new grace period. | |
1054 | */ | |
1055 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) | |
1056 | { | |
1057 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; | |
1058 | } | |
1059 | ||
27f4d280 PM |
1060 | /* |
1061 | * Create an RCU-boost kthread for the specified node if one does not | |
1062 | * already exist. We only create this kthread for preemptible RCU. | |
1063 | * Returns zero if all is well, a negated errno otherwise. | |
1064 | */ | |
49fb4c62 | 1065 | static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp, |
0aa04b05 | 1066 | struct rcu_node *rnp) |
27f4d280 | 1067 | { |
5d01bbd1 | 1068 | int rnp_index = rnp - &rsp->node[0]; |
27f4d280 PM |
1069 | unsigned long flags; |
1070 | struct sched_param sp; | |
1071 | struct task_struct *t; | |
1072 | ||
e63c887c | 1073 | if (rcu_state_p != rsp) |
27f4d280 | 1074 | return 0; |
5d01bbd1 | 1075 | |
0aa04b05 | 1076 | if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0) |
5d01bbd1 TG |
1077 | return 0; |
1078 | ||
a46e0899 | 1079 | rsp->boost = 1; |
27f4d280 PM |
1080 | if (rnp->boost_kthread_task != NULL) |
1081 | return 0; | |
1082 | t = kthread_create(rcu_boost_kthread, (void *)rnp, | |
5b61b0ba | 1083 | "rcub/%d", rnp_index); |
27f4d280 PM |
1084 | if (IS_ERR(t)) |
1085 | return PTR_ERR(t); | |
2a67e741 | 1086 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
27f4d280 | 1087 | rnp->boost_kthread_task = t; |
67c583a7 | 1088 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
21871d7e | 1089 | sp.sched_priority = kthread_prio; |
27f4d280 | 1090 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); |
9a432736 | 1091 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
27f4d280 PM |
1092 | return 0; |
1093 | } | |
1094 | ||
f8b7fc6b PM |
1095 | static void rcu_kthread_do_work(void) |
1096 | { | |
c9d4b0af CL |
1097 | rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data)); |
1098 | rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data)); | |
f8b7fc6b PM |
1099 | rcu_preempt_do_callbacks(); |
1100 | } | |
1101 | ||
62ab7072 | 1102 | static void rcu_cpu_kthread_setup(unsigned int cpu) |
f8b7fc6b | 1103 | { |
f8b7fc6b | 1104 | struct sched_param sp; |
f8b7fc6b | 1105 | |
21871d7e | 1106 | sp.sched_priority = kthread_prio; |
62ab7072 | 1107 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); |
f8b7fc6b PM |
1108 | } |
1109 | ||
62ab7072 | 1110 | static void rcu_cpu_kthread_park(unsigned int cpu) |
f8b7fc6b | 1111 | { |
62ab7072 | 1112 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; |
f8b7fc6b PM |
1113 | } |
1114 | ||
62ab7072 | 1115 | static int rcu_cpu_kthread_should_run(unsigned int cpu) |
f8b7fc6b | 1116 | { |
c9d4b0af | 1117 | return __this_cpu_read(rcu_cpu_has_work); |
f8b7fc6b PM |
1118 | } |
1119 | ||
1120 | /* | |
1121 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the | |
e0f23060 PM |
1122 | * RCU softirq used in flavors and configurations of RCU that do not |
1123 | * support RCU priority boosting. | |
f8b7fc6b | 1124 | */ |
62ab7072 | 1125 | static void rcu_cpu_kthread(unsigned int cpu) |
f8b7fc6b | 1126 | { |
c9d4b0af CL |
1127 | unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status); |
1128 | char work, *workp = this_cpu_ptr(&rcu_cpu_has_work); | |
62ab7072 | 1129 | int spincnt; |
f8b7fc6b | 1130 | |
62ab7072 | 1131 | for (spincnt = 0; spincnt < 10; spincnt++) { |
f7f7bac9 | 1132 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait")); |
f8b7fc6b | 1133 | local_bh_disable(); |
f8b7fc6b | 1134 | *statusp = RCU_KTHREAD_RUNNING; |
62ab7072 PM |
1135 | this_cpu_inc(rcu_cpu_kthread_loops); |
1136 | local_irq_disable(); | |
f8b7fc6b PM |
1137 | work = *workp; |
1138 | *workp = 0; | |
62ab7072 | 1139 | local_irq_enable(); |
f8b7fc6b PM |
1140 | if (work) |
1141 | rcu_kthread_do_work(); | |
1142 | local_bh_enable(); | |
62ab7072 | 1143 | if (*workp == 0) { |
f7f7bac9 | 1144 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); |
62ab7072 PM |
1145 | *statusp = RCU_KTHREAD_WAITING; |
1146 | return; | |
f8b7fc6b PM |
1147 | } |
1148 | } | |
62ab7072 | 1149 | *statusp = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 1150 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); |
62ab7072 | 1151 | schedule_timeout_interruptible(2); |
f7f7bac9 | 1152 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); |
62ab7072 | 1153 | *statusp = RCU_KTHREAD_WAITING; |
f8b7fc6b PM |
1154 | } |
1155 | ||
1156 | /* | |
1157 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are | |
1158 | * served by the rcu_node in question. The CPU hotplug lock is still | |
1159 | * held, so the value of rnp->qsmaskinit will be stable. | |
1160 | * | |
1161 | * We don't include outgoingcpu in the affinity set, use -1 if there is | |
1162 | * no outgoing CPU. If there are no CPUs left in the affinity set, | |
1163 | * this function allows the kthread to execute on any CPU. | |
1164 | */ | |
5d01bbd1 | 1165 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b | 1166 | { |
5d01bbd1 | 1167 | struct task_struct *t = rnp->boost_kthread_task; |
0aa04b05 | 1168 | unsigned long mask = rcu_rnp_online_cpus(rnp); |
f8b7fc6b PM |
1169 | cpumask_var_t cm; |
1170 | int cpu; | |
f8b7fc6b | 1171 | |
5d01bbd1 | 1172 | if (!t) |
f8b7fc6b | 1173 | return; |
5d01bbd1 | 1174 | if (!zalloc_cpumask_var(&cm, GFP_KERNEL)) |
f8b7fc6b | 1175 | return; |
bc75e999 MR |
1176 | for_each_leaf_node_possible_cpu(rnp, cpu) |
1177 | if ((mask & leaf_node_cpu_bit(rnp, cpu)) && | |
1178 | cpu != outgoingcpu) | |
f8b7fc6b | 1179 | cpumask_set_cpu(cpu, cm); |
5d0b0249 | 1180 | if (cpumask_weight(cm) == 0) |
f8b7fc6b | 1181 | cpumask_setall(cm); |
5d01bbd1 | 1182 | set_cpus_allowed_ptr(t, cm); |
f8b7fc6b PM |
1183 | free_cpumask_var(cm); |
1184 | } | |
1185 | ||
62ab7072 PM |
1186 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { |
1187 | .store = &rcu_cpu_kthread_task, | |
1188 | .thread_should_run = rcu_cpu_kthread_should_run, | |
1189 | .thread_fn = rcu_cpu_kthread, | |
1190 | .thread_comm = "rcuc/%u", | |
1191 | .setup = rcu_cpu_kthread_setup, | |
1192 | .park = rcu_cpu_kthread_park, | |
1193 | }; | |
f8b7fc6b PM |
1194 | |
1195 | /* | |
9386c0b7 | 1196 | * Spawn boost kthreads -- called as soon as the scheduler is running. |
f8b7fc6b | 1197 | */ |
9386c0b7 | 1198 | static void __init rcu_spawn_boost_kthreads(void) |
f8b7fc6b | 1199 | { |
f8b7fc6b | 1200 | struct rcu_node *rnp; |
5d01bbd1 | 1201 | int cpu; |
f8b7fc6b | 1202 | |
62ab7072 | 1203 | for_each_possible_cpu(cpu) |
f8b7fc6b | 1204 | per_cpu(rcu_cpu_has_work, cpu) = 0; |
62ab7072 | 1205 | BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec)); |
3e9f5c70 PM |
1206 | rcu_for_each_leaf_node(rcu_state_p, rnp) |
1207 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); | |
f8b7fc6b | 1208 | } |
f8b7fc6b | 1209 | |
49fb4c62 | 1210 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b | 1211 | { |
e534165b | 1212 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
f8b7fc6b PM |
1213 | struct rcu_node *rnp = rdp->mynode; |
1214 | ||
1215 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ | |
62ab7072 | 1216 | if (rcu_scheduler_fully_active) |
e534165b | 1217 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); |
f8b7fc6b PM |
1218 | } |
1219 | ||
27f4d280 PM |
1220 | #else /* #ifdef CONFIG_RCU_BOOST */ |
1221 | ||
1217ed1b | 1222 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 1223 | __releases(rnp->lock) |
27f4d280 | 1224 | { |
67c583a7 | 1225 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 PM |
1226 | } |
1227 | ||
a46e0899 | 1228 | static void invoke_rcu_callbacks_kthread(void) |
27f4d280 | 1229 | { |
a46e0899 | 1230 | WARN_ON_ONCE(1); |
27f4d280 PM |
1231 | } |
1232 | ||
dff1672d PM |
1233 | static bool rcu_is_callbacks_kthread(void) |
1234 | { | |
1235 | return false; | |
1236 | } | |
1237 | ||
27f4d280 PM |
1238 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
1239 | { | |
1240 | } | |
1241 | ||
5d01bbd1 | 1242 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b PM |
1243 | { |
1244 | } | |
1245 | ||
9386c0b7 | 1246 | static void __init rcu_spawn_boost_kthreads(void) |
b0d30417 | 1247 | { |
b0d30417 | 1248 | } |
b0d30417 | 1249 | |
49fb4c62 | 1250 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b PM |
1251 | { |
1252 | } | |
1253 | ||
27f4d280 PM |
1254 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
1255 | ||
8bd93a2c PM |
1256 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
1257 | ||
1258 | /* | |
1259 | * Check to see if any future RCU-related work will need to be done | |
1260 | * by the current CPU, even if none need be done immediately, returning | |
1261 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
1262 | * an exported member of the RCU API. | |
1263 | * | |
7cb92499 PM |
1264 | * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs |
1265 | * any flavor of RCU. | |
8bd93a2c | 1266 | */ |
c1ad348b | 1267 | int rcu_needs_cpu(u64 basemono, u64 *nextevt) |
8bd93a2c | 1268 | { |
c1ad348b | 1269 | *nextevt = KTIME_MAX; |
3382adbc PM |
1270 | return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) |
1271 | ? 0 : rcu_cpu_has_callbacks(NULL); | |
7cb92499 PM |
1272 | } |
1273 | ||
1274 | /* | |
1275 | * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up | |
1276 | * after it. | |
1277 | */ | |
8fa7845d | 1278 | static void rcu_cleanup_after_idle(void) |
7cb92499 PM |
1279 | { |
1280 | } | |
1281 | ||
aea1b35e | 1282 | /* |
a858af28 | 1283 | * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n, |
aea1b35e PM |
1284 | * is nothing. |
1285 | */ | |
198bbf81 | 1286 | static void rcu_prepare_for_idle(void) |
aea1b35e PM |
1287 | { |
1288 | } | |
1289 | ||
c57afe80 PM |
1290 | /* |
1291 | * Don't bother keeping a running count of the number of RCU callbacks | |
1292 | * posted because CONFIG_RCU_FAST_NO_HZ=n. | |
1293 | */ | |
1294 | static void rcu_idle_count_callbacks_posted(void) | |
1295 | { | |
1296 | } | |
1297 | ||
8bd93a2c PM |
1298 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
1299 | ||
f23f7fa1 PM |
1300 | /* |
1301 | * This code is invoked when a CPU goes idle, at which point we want | |
1302 | * to have the CPU do everything required for RCU so that it can enter | |
1303 | * the energy-efficient dyntick-idle mode. This is handled by a | |
1304 | * state machine implemented by rcu_prepare_for_idle() below. | |
1305 | * | |
1306 | * The following three proprocessor symbols control this state machine: | |
1307 | * | |
f23f7fa1 PM |
1308 | * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted |
1309 | * to sleep in dyntick-idle mode with RCU callbacks pending. This | |
1310 | * is sized to be roughly one RCU grace period. Those energy-efficiency | |
1311 | * benchmarkers who might otherwise be tempted to set this to a large | |
1312 | * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your | |
1313 | * system. And if you are -that- concerned about energy efficiency, | |
1314 | * just power the system down and be done with it! | |
778d250a PM |
1315 | * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is |
1316 | * permitted to sleep in dyntick-idle mode with only lazy RCU | |
1317 | * callbacks pending. Setting this too high can OOM your system. | |
f23f7fa1 PM |
1318 | * |
1319 | * The values below work well in practice. If future workloads require | |
1320 | * adjustment, they can be converted into kernel config parameters, though | |
1321 | * making the state machine smarter might be a better option. | |
1322 | */ | |
e84c48ae | 1323 | #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */ |
778d250a | 1324 | #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */ |
f23f7fa1 | 1325 | |
5e44ce35 PM |
1326 | static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY; |
1327 | module_param(rcu_idle_gp_delay, int, 0644); | |
1328 | static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY; | |
1329 | module_param(rcu_idle_lazy_gp_delay, int, 0644); | |
486e2593 | 1330 | |
486e2593 | 1331 | /* |
c229828c PM |
1332 | * Try to advance callbacks for all flavors of RCU on the current CPU, but |
1333 | * only if it has been awhile since the last time we did so. Afterwards, | |
1334 | * if there are any callbacks ready for immediate invocation, return true. | |
486e2593 | 1335 | */ |
f1f399d1 | 1336 | static bool __maybe_unused rcu_try_advance_all_cbs(void) |
486e2593 | 1337 | { |
c0f4dfd4 PM |
1338 | bool cbs_ready = false; |
1339 | struct rcu_data *rdp; | |
c229828c | 1340 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1341 | struct rcu_node *rnp; |
1342 | struct rcu_state *rsp; | |
486e2593 | 1343 | |
c229828c PM |
1344 | /* Exit early if we advanced recently. */ |
1345 | if (jiffies == rdtp->last_advance_all) | |
d0bc90fd | 1346 | return false; |
c229828c PM |
1347 | rdtp->last_advance_all = jiffies; |
1348 | ||
c0f4dfd4 PM |
1349 | for_each_rcu_flavor(rsp) { |
1350 | rdp = this_cpu_ptr(rsp->rda); | |
1351 | rnp = rdp->mynode; | |
486e2593 | 1352 | |
c0f4dfd4 PM |
1353 | /* |
1354 | * Don't bother checking unless a grace period has | |
1355 | * completed since we last checked and there are | |
1356 | * callbacks not yet ready to invoke. | |
1357 | */ | |
e3663b10 | 1358 | if ((rdp->completed != rnp->completed || |
7d0ae808 | 1359 | unlikely(READ_ONCE(rdp->gpwrap))) && |
15fecf89 | 1360 | rcu_segcblist_pend_cbs(&rdp->cblist)) |
470716fc | 1361 | note_gp_changes(rsp, rdp); |
486e2593 | 1362 | |
15fecf89 | 1363 | if (rcu_segcblist_ready_cbs(&rdp->cblist)) |
c0f4dfd4 PM |
1364 | cbs_ready = true; |
1365 | } | |
1366 | return cbs_ready; | |
486e2593 PM |
1367 | } |
1368 | ||
aa9b1630 | 1369 | /* |
c0f4dfd4 PM |
1370 | * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready |
1371 | * to invoke. If the CPU has callbacks, try to advance them. Tell the | |
1372 | * caller to set the timeout based on whether or not there are non-lazy | |
1373 | * callbacks. | |
aa9b1630 | 1374 | * |
c0f4dfd4 | 1375 | * The caller must have disabled interrupts. |
aa9b1630 | 1376 | */ |
c1ad348b | 1377 | int rcu_needs_cpu(u64 basemono, u64 *nextevt) |
aa9b1630 | 1378 | { |
aa6da514 | 1379 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c1ad348b | 1380 | unsigned long dj; |
aa9b1630 | 1381 | |
3382adbc | 1382 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) { |
43224b96 | 1383 | *nextevt = KTIME_MAX; |
3382adbc PM |
1384 | return 0; |
1385 | } | |
1386 | ||
c0f4dfd4 PM |
1387 | /* Snapshot to detect later posting of non-lazy callback. */ |
1388 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
1389 | ||
aa9b1630 | 1390 | /* If no callbacks, RCU doesn't need the CPU. */ |
aa6da514 | 1391 | if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) { |
c1ad348b | 1392 | *nextevt = KTIME_MAX; |
aa9b1630 PM |
1393 | return 0; |
1394 | } | |
c0f4dfd4 PM |
1395 | |
1396 | /* Attempt to advance callbacks. */ | |
1397 | if (rcu_try_advance_all_cbs()) { | |
1398 | /* Some ready to invoke, so initiate later invocation. */ | |
1399 | invoke_rcu_core(); | |
aa9b1630 PM |
1400 | return 1; |
1401 | } | |
c0f4dfd4 PM |
1402 | rdtp->last_accelerate = jiffies; |
1403 | ||
1404 | /* Request timer delay depending on laziness, and round. */ | |
6faf7283 | 1405 | if (!rdtp->all_lazy) { |
c1ad348b | 1406 | dj = round_up(rcu_idle_gp_delay + jiffies, |
c0f4dfd4 | 1407 | rcu_idle_gp_delay) - jiffies; |
e84c48ae | 1408 | } else { |
c1ad348b | 1409 | dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies; |
e84c48ae | 1410 | } |
c1ad348b | 1411 | *nextevt = basemono + dj * TICK_NSEC; |
aa9b1630 PM |
1412 | return 0; |
1413 | } | |
1414 | ||
21e52e15 | 1415 | /* |
c0f4dfd4 PM |
1416 | * Prepare a CPU for idle from an RCU perspective. The first major task |
1417 | * is to sense whether nohz mode has been enabled or disabled via sysfs. | |
1418 | * The second major task is to check to see if a non-lazy callback has | |
1419 | * arrived at a CPU that previously had only lazy callbacks. The third | |
1420 | * major task is to accelerate (that is, assign grace-period numbers to) | |
1421 | * any recently arrived callbacks. | |
aea1b35e PM |
1422 | * |
1423 | * The caller must have disabled interrupts. | |
8bd93a2c | 1424 | */ |
198bbf81 | 1425 | static void rcu_prepare_for_idle(void) |
8bd93a2c | 1426 | { |
48a7639c | 1427 | bool needwake; |
c0f4dfd4 | 1428 | struct rcu_data *rdp; |
198bbf81 | 1429 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1430 | struct rcu_node *rnp; |
1431 | struct rcu_state *rsp; | |
9d2ad243 PM |
1432 | int tne; |
1433 | ||
f0f2e7d3 PM |
1434 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) || |
1435 | rcu_is_nocb_cpu(smp_processor_id())) | |
3382adbc PM |
1436 | return; |
1437 | ||
9d2ad243 | 1438 | /* Handle nohz enablement switches conservatively. */ |
7d0ae808 | 1439 | tne = READ_ONCE(tick_nohz_active); |
9d2ad243 | 1440 | if (tne != rdtp->tick_nohz_enabled_snap) { |
aa6da514 | 1441 | if (rcu_cpu_has_callbacks(NULL)) |
9d2ad243 PM |
1442 | invoke_rcu_core(); /* force nohz to see update. */ |
1443 | rdtp->tick_nohz_enabled_snap = tne; | |
1444 | return; | |
1445 | } | |
1446 | if (!tne) | |
1447 | return; | |
f511fc62 | 1448 | |
c57afe80 | 1449 | /* |
c0f4dfd4 PM |
1450 | * If a non-lazy callback arrived at a CPU having only lazy |
1451 | * callbacks, invoke RCU core for the side-effect of recalculating | |
1452 | * idle duration on re-entry to idle. | |
c57afe80 | 1453 | */ |
c0f4dfd4 PM |
1454 | if (rdtp->all_lazy && |
1455 | rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) { | |
c337f8f5 PM |
1456 | rdtp->all_lazy = false; |
1457 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
c0f4dfd4 | 1458 | invoke_rcu_core(); |
c57afe80 PM |
1459 | return; |
1460 | } | |
c57afe80 | 1461 | |
3084f2f8 | 1462 | /* |
c0f4dfd4 PM |
1463 | * If we have not yet accelerated this jiffy, accelerate all |
1464 | * callbacks on this CPU. | |
3084f2f8 | 1465 | */ |
c0f4dfd4 | 1466 | if (rdtp->last_accelerate == jiffies) |
aea1b35e | 1467 | return; |
c0f4dfd4 PM |
1468 | rdtp->last_accelerate = jiffies; |
1469 | for_each_rcu_flavor(rsp) { | |
198bbf81 | 1470 | rdp = this_cpu_ptr(rsp->rda); |
15fecf89 | 1471 | if (rcu_segcblist_pend_cbs(&rdp->cblist)) |
c0f4dfd4 PM |
1472 | continue; |
1473 | rnp = rdp->mynode; | |
2a67e741 | 1474 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
48a7639c | 1475 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
67c583a7 | 1476 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
48a7639c PM |
1477 | if (needwake) |
1478 | rcu_gp_kthread_wake(rsp); | |
77e38ed3 | 1479 | } |
c0f4dfd4 | 1480 | } |
3084f2f8 | 1481 | |
c0f4dfd4 PM |
1482 | /* |
1483 | * Clean up for exit from idle. Attempt to advance callbacks based on | |
1484 | * any grace periods that elapsed while the CPU was idle, and if any | |
1485 | * callbacks are now ready to invoke, initiate invocation. | |
1486 | */ | |
8fa7845d | 1487 | static void rcu_cleanup_after_idle(void) |
c0f4dfd4 | 1488 | { |
3382adbc PM |
1489 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) || |
1490 | rcu_is_nocb_cpu(smp_processor_id())) | |
aea1b35e | 1491 | return; |
7a497c96 PM |
1492 | if (rcu_try_advance_all_cbs()) |
1493 | invoke_rcu_core(); | |
8bd93a2c PM |
1494 | } |
1495 | ||
c57afe80 | 1496 | /* |
98248a0e PM |
1497 | * Keep a running count of the number of non-lazy callbacks posted |
1498 | * on this CPU. This running counter (which is never decremented) allows | |
1499 | * rcu_prepare_for_idle() to detect when something out of the idle loop | |
1500 | * posts a callback, even if an equal number of callbacks are invoked. | |
1501 | * Of course, callbacks should only be posted from within a trace event | |
1502 | * designed to be called from idle or from within RCU_NONIDLE(). | |
c57afe80 PM |
1503 | */ |
1504 | static void rcu_idle_count_callbacks_posted(void) | |
1505 | { | |
5955f7ee | 1506 | __this_cpu_add(rcu_dynticks.nonlazy_posted, 1); |
c57afe80 PM |
1507 | } |
1508 | ||
b626c1b6 PM |
1509 | /* |
1510 | * Data for flushing lazy RCU callbacks at OOM time. | |
1511 | */ | |
1512 | static atomic_t oom_callback_count; | |
1513 | static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq); | |
1514 | ||
1515 | /* | |
1516 | * RCU OOM callback -- decrement the outstanding count and deliver the | |
1517 | * wake-up if we are the last one. | |
1518 | */ | |
1519 | static void rcu_oom_callback(struct rcu_head *rhp) | |
1520 | { | |
1521 | if (atomic_dec_and_test(&oom_callback_count)) | |
1522 | wake_up(&oom_callback_wq); | |
1523 | } | |
1524 | ||
1525 | /* | |
1526 | * Post an rcu_oom_notify callback on the current CPU if it has at | |
1527 | * least one lazy callback. This will unnecessarily post callbacks | |
1528 | * to CPUs that already have a non-lazy callback at the end of their | |
1529 | * callback list, but this is an infrequent operation, so accept some | |
1530 | * extra overhead to keep things simple. | |
1531 | */ | |
1532 | static void rcu_oom_notify_cpu(void *unused) | |
1533 | { | |
1534 | struct rcu_state *rsp; | |
1535 | struct rcu_data *rdp; | |
1536 | ||
1537 | for_each_rcu_flavor(rsp) { | |
fa07a58f | 1538 | rdp = raw_cpu_ptr(rsp->rda); |
15fecf89 | 1539 | if (rcu_segcblist_n_lazy_cbs(&rdp->cblist)) { |
b626c1b6 PM |
1540 | atomic_inc(&oom_callback_count); |
1541 | rsp->call(&rdp->oom_head, rcu_oom_callback); | |
1542 | } | |
1543 | } | |
1544 | } | |
1545 | ||
1546 | /* | |
1547 | * If low on memory, ensure that each CPU has a non-lazy callback. | |
1548 | * This will wake up CPUs that have only lazy callbacks, in turn | |
1549 | * ensuring that they free up the corresponding memory in a timely manner. | |
1550 | * Because an uncertain amount of memory will be freed in some uncertain | |
1551 | * timeframe, we do not claim to have freed anything. | |
1552 | */ | |
1553 | static int rcu_oom_notify(struct notifier_block *self, | |
1554 | unsigned long notused, void *nfreed) | |
1555 | { | |
1556 | int cpu; | |
1557 | ||
1558 | /* Wait for callbacks from earlier instance to complete. */ | |
1559 | wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0); | |
78e4bc34 | 1560 | smp_mb(); /* Ensure callback reuse happens after callback invocation. */ |
b626c1b6 PM |
1561 | |
1562 | /* | |
1563 | * Prevent premature wakeup: ensure that all increments happen | |
1564 | * before there is a chance of the counter reaching zero. | |
1565 | */ | |
1566 | atomic_set(&oom_callback_count, 1); | |
1567 | ||
b626c1b6 PM |
1568 | for_each_online_cpu(cpu) { |
1569 | smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1); | |
bde6c3aa | 1570 | cond_resched_rcu_qs(); |
b626c1b6 | 1571 | } |
b626c1b6 PM |
1572 | |
1573 | /* Unconditionally decrement: no need to wake ourselves up. */ | |
1574 | atomic_dec(&oom_callback_count); | |
1575 | ||
1576 | return NOTIFY_OK; | |
1577 | } | |
1578 | ||
1579 | static struct notifier_block rcu_oom_nb = { | |
1580 | .notifier_call = rcu_oom_notify | |
1581 | }; | |
1582 | ||
1583 | static int __init rcu_register_oom_notifier(void) | |
1584 | { | |
1585 | register_oom_notifier(&rcu_oom_nb); | |
1586 | return 0; | |
1587 | } | |
1588 | early_initcall(rcu_register_oom_notifier); | |
1589 | ||
8bd93a2c | 1590 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
a858af28 | 1591 | |
a858af28 PM |
1592 | #ifdef CONFIG_RCU_FAST_NO_HZ |
1593 | ||
1594 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1595 | { | |
5955f7ee | 1596 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
c0f4dfd4 | 1597 | unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap; |
a858af28 | 1598 | |
c0f4dfd4 PM |
1599 | sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c", |
1600 | rdtp->last_accelerate & 0xffff, jiffies & 0xffff, | |
1601 | ulong2long(nlpd), | |
1602 | rdtp->all_lazy ? 'L' : '.', | |
1603 | rdtp->tick_nohz_enabled_snap ? '.' : 'D'); | |
a858af28 PM |
1604 | } |
1605 | ||
1606 | #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1607 | ||
1608 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1609 | { | |
1c17e4d4 | 1610 | *cp = '\0'; |
a858af28 PM |
1611 | } |
1612 | ||
1613 | #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1614 | ||
1615 | /* Initiate the stall-info list. */ | |
1616 | static void print_cpu_stall_info_begin(void) | |
1617 | { | |
efc151c3 | 1618 | pr_cont("\n"); |
a858af28 PM |
1619 | } |
1620 | ||
1621 | /* | |
1622 | * Print out diagnostic information for the specified stalled CPU. | |
1623 | * | |
1624 | * If the specified CPU is aware of the current RCU grace period | |
1625 | * (flavor specified by rsp), then print the number of scheduling | |
1626 | * clock interrupts the CPU has taken during the time that it has | |
1627 | * been aware. Otherwise, print the number of RCU grace periods | |
1628 | * that this CPU is ignorant of, for example, "1" if the CPU was | |
1629 | * aware of the previous grace period. | |
1630 | * | |
1631 | * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info. | |
1632 | */ | |
1633 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) | |
1634 | { | |
1635 | char fast_no_hz[72]; | |
1636 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
1637 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
1638 | char *ticks_title; | |
1639 | unsigned long ticks_value; | |
1640 | ||
1641 | if (rsp->gpnum == rdp->gpnum) { | |
1642 | ticks_title = "ticks this GP"; | |
1643 | ticks_value = rdp->ticks_this_gp; | |
1644 | } else { | |
1645 | ticks_title = "GPs behind"; | |
1646 | ticks_value = rsp->gpnum - rdp->gpnum; | |
1647 | } | |
1648 | print_cpu_stall_fast_no_hz(fast_no_hz, cpu); | |
7f21aeef PM |
1649 | pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n", |
1650 | cpu, | |
1651 | "O."[!!cpu_online(cpu)], | |
1652 | "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)], | |
1653 | "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)], | |
1654 | ticks_value, ticks_title, | |
02a5c550 | 1655 | rcu_dynticks_snap(rdtp) & 0xfff, |
a858af28 | 1656 | rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting, |
6231069b | 1657 | rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu), |
7d0ae808 | 1658 | READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart, |
a858af28 PM |
1659 | fast_no_hz); |
1660 | } | |
1661 | ||
1662 | /* Terminate the stall-info list. */ | |
1663 | static void print_cpu_stall_info_end(void) | |
1664 | { | |
efc151c3 | 1665 | pr_err("\t"); |
a858af28 PM |
1666 | } |
1667 | ||
1668 | /* Zero ->ticks_this_gp for all flavors of RCU. */ | |
1669 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) | |
1670 | { | |
1671 | rdp->ticks_this_gp = 0; | |
6231069b | 1672 | rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id()); |
a858af28 PM |
1673 | } |
1674 | ||
1675 | /* Increment ->ticks_this_gp for all flavors of RCU. */ | |
1676 | static void increment_cpu_stall_ticks(void) | |
1677 | { | |
115f7a7c PM |
1678 | struct rcu_state *rsp; |
1679 | ||
1680 | for_each_rcu_flavor(rsp) | |
fa07a58f | 1681 | raw_cpu_inc(rsp->rda->ticks_this_gp); |
a858af28 PM |
1682 | } |
1683 | ||
3fbfbf7a PM |
1684 | #ifdef CONFIG_RCU_NOCB_CPU |
1685 | ||
1686 | /* | |
1687 | * Offload callback processing from the boot-time-specified set of CPUs | |
1688 | * specified by rcu_nocb_mask. For each CPU in the set, there is a | |
1689 | * kthread created that pulls the callbacks from the corresponding CPU, | |
1690 | * waits for a grace period to elapse, and invokes the callbacks. | |
1691 | * The no-CBs CPUs do a wake_up() on their kthread when they insert | |
1692 | * a callback into any empty list, unless the rcu_nocb_poll boot parameter | |
1693 | * has been specified, in which case each kthread actively polls its | |
1694 | * CPU. (Which isn't so great for energy efficiency, but which does | |
1695 | * reduce RCU's overhead on that CPU.) | |
1696 | * | |
1697 | * This is intended to be used in conjunction with Frederic Weisbecker's | |
1698 | * adaptive-idle work, which would seriously reduce OS jitter on CPUs | |
1699 | * running CPU-bound user-mode computations. | |
1700 | * | |
1701 | * Offloading of callback processing could also in theory be used as | |
1702 | * an energy-efficiency measure because CPUs with no RCU callbacks | |
1703 | * queued are more aggressive about entering dyntick-idle mode. | |
1704 | */ | |
1705 | ||
1706 | ||
1707 | /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */ | |
1708 | static int __init rcu_nocb_setup(char *str) | |
1709 | { | |
1710 | alloc_bootmem_cpumask_var(&rcu_nocb_mask); | |
1711 | have_rcu_nocb_mask = true; | |
1712 | cpulist_parse(str, rcu_nocb_mask); | |
1713 | return 1; | |
1714 | } | |
1715 | __setup("rcu_nocbs=", rcu_nocb_setup); | |
1716 | ||
1b0048a4 PG |
1717 | static int __init parse_rcu_nocb_poll(char *arg) |
1718 | { | |
5455a7f6 | 1719 | rcu_nocb_poll = true; |
1b0048a4 PG |
1720 | return 0; |
1721 | } | |
1722 | early_param("rcu_nocb_poll", parse_rcu_nocb_poll); | |
1723 | ||
dae6e64d | 1724 | /* |
0446be48 PM |
1725 | * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended |
1726 | * grace period. | |
dae6e64d | 1727 | */ |
abedf8e2 | 1728 | static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq) |
dae6e64d | 1729 | { |
abedf8e2 | 1730 | swake_up_all(sq); |
dae6e64d PM |
1731 | } |
1732 | ||
1733 | /* | |
8b425aa8 | 1734 | * Set the root rcu_node structure's ->need_future_gp field |
dae6e64d PM |
1735 | * based on the sum of those of all rcu_node structures. This does |
1736 | * double-count the root rcu_node structure's requests, but this | |
1737 | * is necessary to handle the possibility of a rcu_nocb_kthread() | |
1738 | * having awakened during the time that the rcu_node structures | |
1739 | * were being updated for the end of the previous grace period. | |
34ed6246 | 1740 | */ |
dae6e64d PM |
1741 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
1742 | { | |
8b425aa8 | 1743 | rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq; |
dae6e64d PM |
1744 | } |
1745 | ||
abedf8e2 | 1746 | static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp) |
065bb78c DW |
1747 | { |
1748 | return &rnp->nocb_gp_wq[rnp->completed & 0x1]; | |
1749 | } | |
1750 | ||
dae6e64d | 1751 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
34ed6246 | 1752 | { |
abedf8e2 PG |
1753 | init_swait_queue_head(&rnp->nocb_gp_wq[0]); |
1754 | init_swait_queue_head(&rnp->nocb_gp_wq[1]); | |
34ed6246 PM |
1755 | } |
1756 | ||
2f33b512 | 1757 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
24342c96 | 1758 | /* Is the specified CPU a no-CBs CPU? */ |
d1e43fa5 | 1759 | bool rcu_is_nocb_cpu(int cpu) |
3fbfbf7a PM |
1760 | { |
1761 | if (have_rcu_nocb_mask) | |
1762 | return cpumask_test_cpu(cpu, rcu_nocb_mask); | |
1763 | return false; | |
1764 | } | |
2f33b512 | 1765 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
3fbfbf7a | 1766 | |
fbce7497 PM |
1767 | /* |
1768 | * Kick the leader kthread for this NOCB group. | |
1769 | */ | |
1770 | static void wake_nocb_leader(struct rcu_data *rdp, bool force) | |
1771 | { | |
1772 | struct rcu_data *rdp_leader = rdp->nocb_leader; | |
1773 | ||
7d0ae808 | 1774 | if (!READ_ONCE(rdp_leader->nocb_kthread)) |
fbce7497 | 1775 | return; |
7d0ae808 | 1776 | if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) { |
39953dfd | 1777 | /* Prior smp_mb__after_atomic() orders against prior enqueue. */ |
7d0ae808 | 1778 | WRITE_ONCE(rdp_leader->nocb_leader_sleep, false); |
abedf8e2 | 1779 | swake_up(&rdp_leader->nocb_wq); |
fbce7497 PM |
1780 | } |
1781 | } | |
1782 | ||
d7e29933 PM |
1783 | /* |
1784 | * Does the specified CPU need an RCU callback for the specified flavor | |
1785 | * of rcu_barrier()? | |
1786 | */ | |
1787 | static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu) | |
1788 | { | |
1789 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
41050a00 PM |
1790 | unsigned long ret; |
1791 | #ifdef CONFIG_PROVE_RCU | |
d7e29933 | 1792 | struct rcu_head *rhp; |
41050a00 | 1793 | #endif /* #ifdef CONFIG_PROVE_RCU */ |
d7e29933 | 1794 | |
41050a00 PM |
1795 | /* |
1796 | * Check count of all no-CBs callbacks awaiting invocation. | |
1797 | * There needs to be a barrier before this function is called, | |
1798 | * but associated with a prior determination that no more | |
1799 | * callbacks would be posted. In the worst case, the first | |
1800 | * barrier in _rcu_barrier() suffices (but the caller cannot | |
1801 | * necessarily rely on this, not a substitute for the caller | |
1802 | * getting the concurrency design right!). There must also be | |
1803 | * a barrier between the following load an posting of a callback | |
1804 | * (if a callback is in fact needed). This is associated with an | |
1805 | * atomic_inc() in the caller. | |
1806 | */ | |
1807 | ret = atomic_long_read(&rdp->nocb_q_count); | |
d7e29933 | 1808 | |
41050a00 | 1809 | #ifdef CONFIG_PROVE_RCU |
7d0ae808 | 1810 | rhp = READ_ONCE(rdp->nocb_head); |
d7e29933 | 1811 | if (!rhp) |
7d0ae808 | 1812 | rhp = READ_ONCE(rdp->nocb_gp_head); |
d7e29933 | 1813 | if (!rhp) |
7d0ae808 | 1814 | rhp = READ_ONCE(rdp->nocb_follower_head); |
d7e29933 PM |
1815 | |
1816 | /* Having no rcuo kthread but CBs after scheduler starts is bad! */ | |
7d0ae808 | 1817 | if (!READ_ONCE(rdp->nocb_kthread) && rhp && |
59f792d1 | 1818 | rcu_scheduler_fully_active) { |
d7e29933 PM |
1819 | /* RCU callback enqueued before CPU first came online??? */ |
1820 | pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n", | |
1821 | cpu, rhp->func); | |
1822 | WARN_ON_ONCE(1); | |
1823 | } | |
41050a00 | 1824 | #endif /* #ifdef CONFIG_PROVE_RCU */ |
d7e29933 | 1825 | |
41050a00 | 1826 | return !!ret; |
d7e29933 PM |
1827 | } |
1828 | ||
3fbfbf7a PM |
1829 | /* |
1830 | * Enqueue the specified string of rcu_head structures onto the specified | |
1831 | * CPU's no-CBs lists. The CPU is specified by rdp, the head of the | |
1832 | * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy | |
1833 | * counts are supplied by rhcount and rhcount_lazy. | |
1834 | * | |
1835 | * If warranted, also wake up the kthread servicing this CPUs queues. | |
1836 | */ | |
1837 | static void __call_rcu_nocb_enqueue(struct rcu_data *rdp, | |
1838 | struct rcu_head *rhp, | |
1839 | struct rcu_head **rhtp, | |
96d3fd0d PM |
1840 | int rhcount, int rhcount_lazy, |
1841 | unsigned long flags) | |
3fbfbf7a PM |
1842 | { |
1843 | int len; | |
1844 | struct rcu_head **old_rhpp; | |
1845 | struct task_struct *t; | |
1846 | ||
1847 | /* Enqueue the callback on the nocb list and update counts. */ | |
41050a00 PM |
1848 | atomic_long_add(rhcount, &rdp->nocb_q_count); |
1849 | /* rcu_barrier() relies on ->nocb_q_count add before xchg. */ | |
3fbfbf7a | 1850 | old_rhpp = xchg(&rdp->nocb_tail, rhtp); |
7d0ae808 | 1851 | WRITE_ONCE(*old_rhpp, rhp); |
3fbfbf7a | 1852 | atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy); |
39953dfd | 1853 | smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */ |
3fbfbf7a PM |
1854 | |
1855 | /* If we are not being polled and there is a kthread, awaken it ... */ | |
7d0ae808 | 1856 | t = READ_ONCE(rdp->nocb_kthread); |
25e03a74 | 1857 | if (rcu_nocb_poll || !t) { |
9261dd0d PM |
1858 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1859 | TPS("WakeNotPoll")); | |
3fbfbf7a | 1860 | return; |
9261dd0d | 1861 | } |
3fbfbf7a PM |
1862 | len = atomic_long_read(&rdp->nocb_q_count); |
1863 | if (old_rhpp == &rdp->nocb_head) { | |
96d3fd0d | 1864 | if (!irqs_disabled_flags(flags)) { |
fbce7497 PM |
1865 | /* ... if queue was empty ... */ |
1866 | wake_nocb_leader(rdp, false); | |
96d3fd0d PM |
1867 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1868 | TPS("WakeEmpty")); | |
1869 | } else { | |
9226b10d PM |
1870 | WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE); |
1871 | /* Store ->nocb_defer_wakeup before ->rcu_urgent_qs. */ | |
1872 | smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true); | |
96d3fd0d PM |
1873 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1874 | TPS("WakeEmptyIsDeferred")); | |
1875 | } | |
3fbfbf7a PM |
1876 | rdp->qlen_last_fqs_check = 0; |
1877 | } else if (len > rdp->qlen_last_fqs_check + qhimark) { | |
fbce7497 | 1878 | /* ... or if many callbacks queued. */ |
9fdd3bc9 PM |
1879 | if (!irqs_disabled_flags(flags)) { |
1880 | wake_nocb_leader(rdp, true); | |
1881 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
1882 | TPS("WakeOvf")); | |
1883 | } else { | |
9226b10d PM |
1884 | WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_FORCE); |
1885 | /* Store ->nocb_defer_wakeup before ->rcu_urgent_qs. */ | |
1886 | smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true); | |
9fdd3bc9 PM |
1887 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1888 | TPS("WakeOvfIsDeferred")); | |
1889 | } | |
3fbfbf7a | 1890 | rdp->qlen_last_fqs_check = LONG_MAX / 2; |
9261dd0d PM |
1891 | } else { |
1892 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot")); | |
3fbfbf7a PM |
1893 | } |
1894 | return; | |
1895 | } | |
1896 | ||
1897 | /* | |
1898 | * This is a helper for __call_rcu(), which invokes this when the normal | |
1899 | * callback queue is inoperable. If this is not a no-CBs CPU, this | |
1900 | * function returns failure back to __call_rcu(), which can complain | |
1901 | * appropriately. | |
1902 | * | |
1903 | * Otherwise, this function queues the callback where the corresponding | |
1904 | * "rcuo" kthread can find it. | |
1905 | */ | |
1906 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, | |
96d3fd0d | 1907 | bool lazy, unsigned long flags) |
3fbfbf7a PM |
1908 | { |
1909 | ||
d1e43fa5 | 1910 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
c271d3a9 | 1911 | return false; |
96d3fd0d | 1912 | __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags); |
21e7a608 PM |
1913 | if (__is_kfree_rcu_offset((unsigned long)rhp->func)) |
1914 | trace_rcu_kfree_callback(rdp->rsp->name, rhp, | |
1915 | (unsigned long)rhp->func, | |
756cbf6b PM |
1916 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
1917 | -atomic_long_read(&rdp->nocb_q_count)); | |
21e7a608 PM |
1918 | else |
1919 | trace_rcu_callback(rdp->rsp->name, rhp, | |
756cbf6b PM |
1920 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
1921 | -atomic_long_read(&rdp->nocb_q_count)); | |
1772947b PM |
1922 | |
1923 | /* | |
1924 | * If called from an extended quiescent state with interrupts | |
1925 | * disabled, invoke the RCU core in order to allow the idle-entry | |
1926 | * deferred-wakeup check to function. | |
1927 | */ | |
1928 | if (irqs_disabled_flags(flags) && | |
1929 | !rcu_is_watching() && | |
1930 | cpu_online(smp_processor_id())) | |
1931 | invoke_rcu_core(); | |
1932 | ||
c271d3a9 | 1933 | return true; |
3fbfbf7a PM |
1934 | } |
1935 | ||
1936 | /* | |
1937 | * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is | |
1938 | * not a no-CBs CPU. | |
1939 | */ | |
1940 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
1941 | struct rcu_data *rdp, |
1942 | unsigned long flags) | |
3fbfbf7a | 1943 | { |
4b27f20b | 1944 | long ql = rsp->orphan_done.len; |
933dfbd7 | 1945 | long qll = rsp->orphan_done.len_lazy; |
3fbfbf7a PM |
1946 | |
1947 | /* If this is not a no-CBs CPU, tell the caller to do it the old way. */ | |
d1e43fa5 | 1948 | if (!rcu_is_nocb_cpu(smp_processor_id())) |
0a9e1e11 | 1949 | return false; |
3fbfbf7a PM |
1950 | |
1951 | /* First, enqueue the donelist, if any. This preserves CB ordering. */ | |
8ef0f37e | 1952 | if (rsp->orphan_done.head) { |
15fecf89 PM |
1953 | __call_rcu_nocb_enqueue(rdp, rcu_cblist_head(&rsp->orphan_done), |
1954 | rcu_cblist_tail(&rsp->orphan_done), | |
1955 | ql, qll, flags); | |
3fbfbf7a | 1956 | } |
8ef0f37e | 1957 | if (rsp->orphan_pend.head) { |
15fecf89 PM |
1958 | __call_rcu_nocb_enqueue(rdp, rcu_cblist_head(&rsp->orphan_pend), |
1959 | rcu_cblist_tail(&rsp->orphan_pend), | |
1960 | ql, qll, flags); | |
3fbfbf7a | 1961 | } |
15fecf89 PM |
1962 | rcu_cblist_init(&rsp->orphan_done); |
1963 | rcu_cblist_init(&rsp->orphan_pend); | |
0a9e1e11 | 1964 | return true; |
3fbfbf7a PM |
1965 | } |
1966 | ||
1967 | /* | |
34ed6246 PM |
1968 | * If necessary, kick off a new grace period, and either way wait |
1969 | * for a subsequent grace period to complete. | |
3fbfbf7a | 1970 | */ |
34ed6246 | 1971 | static void rcu_nocb_wait_gp(struct rcu_data *rdp) |
3fbfbf7a | 1972 | { |
34ed6246 | 1973 | unsigned long c; |
dae6e64d | 1974 | bool d; |
34ed6246 | 1975 | unsigned long flags; |
48a7639c | 1976 | bool needwake; |
34ed6246 PM |
1977 | struct rcu_node *rnp = rdp->mynode; |
1978 | ||
2a67e741 | 1979 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
48a7639c | 1980 | needwake = rcu_start_future_gp(rnp, rdp, &c); |
67c583a7 | 1981 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
48a7639c PM |
1982 | if (needwake) |
1983 | rcu_gp_kthread_wake(rdp->rsp); | |
3fbfbf7a PM |
1984 | |
1985 | /* | |
34ed6246 PM |
1986 | * Wait for the grace period. Do so interruptibly to avoid messing |
1987 | * up the load average. | |
3fbfbf7a | 1988 | */ |
f7f7bac9 | 1989 | trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait")); |
34ed6246 | 1990 | for (;;) { |
abedf8e2 | 1991 | swait_event_interruptible( |
dae6e64d | 1992 | rnp->nocb_gp_wq[c & 0x1], |
7d0ae808 | 1993 | (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c))); |
dae6e64d | 1994 | if (likely(d)) |
34ed6246 | 1995 | break; |
73a860cd | 1996 | WARN_ON(signal_pending(current)); |
f7f7bac9 | 1997 | trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait")); |
34ed6246 | 1998 | } |
f7f7bac9 | 1999 | trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait")); |
34ed6246 | 2000 | smp_mb(); /* Ensure that CB invocation happens after GP end. */ |
3fbfbf7a PM |
2001 | } |
2002 | ||
fbce7497 PM |
2003 | /* |
2004 | * Leaders come here to wait for additional callbacks to show up. | |
2005 | * This function does not return until callbacks appear. | |
2006 | */ | |
2007 | static void nocb_leader_wait(struct rcu_data *my_rdp) | |
2008 | { | |
2009 | bool firsttime = true; | |
2010 | bool gotcbs; | |
2011 | struct rcu_data *rdp; | |
2012 | struct rcu_head **tail; | |
2013 | ||
2014 | wait_again: | |
2015 | ||
2016 | /* Wait for callbacks to appear. */ | |
2017 | if (!rcu_nocb_poll) { | |
2018 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep"); | |
abedf8e2 | 2019 | swait_event_interruptible(my_rdp->nocb_wq, |
7d0ae808 | 2020 | !READ_ONCE(my_rdp->nocb_leader_sleep)); |
fbce7497 PM |
2021 | /* Memory barrier handled by smp_mb() calls below and repoll. */ |
2022 | } else if (firsttime) { | |
2023 | firsttime = false; /* Don't drown trace log with "Poll"! */ | |
2024 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll"); | |
2025 | } | |
2026 | ||
2027 | /* | |
2028 | * Each pass through the following loop checks a follower for CBs. | |
2029 | * We are our own first follower. Any CBs found are moved to | |
2030 | * nocb_gp_head, where they await a grace period. | |
2031 | */ | |
2032 | gotcbs = false; | |
2033 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
7d0ae808 | 2034 | rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head); |
fbce7497 PM |
2035 | if (!rdp->nocb_gp_head) |
2036 | continue; /* No CBs here, try next follower. */ | |
2037 | ||
2038 | /* Move callbacks to wait-for-GP list, which is empty. */ | |
7d0ae808 | 2039 | WRITE_ONCE(rdp->nocb_head, NULL); |
fbce7497 | 2040 | rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head); |
fbce7497 PM |
2041 | gotcbs = true; |
2042 | } | |
2043 | ||
2044 | /* | |
2045 | * If there were no callbacks, sleep a bit, rescan after a | |
2046 | * memory barrier, and go retry. | |
2047 | */ | |
2048 | if (unlikely(!gotcbs)) { | |
2049 | if (!rcu_nocb_poll) | |
2050 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, | |
2051 | "WokeEmpty"); | |
73a860cd | 2052 | WARN_ON(signal_pending(current)); |
fbce7497 PM |
2053 | schedule_timeout_interruptible(1); |
2054 | ||
2055 | /* Rescan in case we were a victim of memory ordering. */ | |
11ed7f93 PK |
2056 | my_rdp->nocb_leader_sleep = true; |
2057 | smp_mb(); /* Ensure _sleep true before scan. */ | |
fbce7497 | 2058 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) |
7d0ae808 | 2059 | if (READ_ONCE(rdp->nocb_head)) { |
fbce7497 | 2060 | /* Found CB, so short-circuit next wait. */ |
11ed7f93 | 2061 | my_rdp->nocb_leader_sleep = false; |
fbce7497 PM |
2062 | break; |
2063 | } | |
2064 | goto wait_again; | |
2065 | } | |
2066 | ||
2067 | /* Wait for one grace period. */ | |
2068 | rcu_nocb_wait_gp(my_rdp); | |
2069 | ||
2070 | /* | |
11ed7f93 PK |
2071 | * We left ->nocb_leader_sleep unset to reduce cache thrashing. |
2072 | * We set it now, but recheck for new callbacks while | |
fbce7497 PM |
2073 | * traversing our follower list. |
2074 | */ | |
11ed7f93 PK |
2075 | my_rdp->nocb_leader_sleep = true; |
2076 | smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */ | |
fbce7497 PM |
2077 | |
2078 | /* Each pass through the following loop wakes a follower, if needed. */ | |
2079 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
7d0ae808 | 2080 | if (READ_ONCE(rdp->nocb_head)) |
11ed7f93 | 2081 | my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/ |
fbce7497 PM |
2082 | if (!rdp->nocb_gp_head) |
2083 | continue; /* No CBs, so no need to wake follower. */ | |
2084 | ||
2085 | /* Append callbacks to follower's "done" list. */ | |
2086 | tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail); | |
2087 | *tail = rdp->nocb_gp_head; | |
c847f142 | 2088 | smp_mb__after_atomic(); /* Store *tail before wakeup. */ |
fbce7497 PM |
2089 | if (rdp != my_rdp && tail == &rdp->nocb_follower_head) { |
2090 | /* | |
2091 | * List was empty, wake up the follower. | |
2092 | * Memory barriers supplied by atomic_long_add(). | |
2093 | */ | |
abedf8e2 | 2094 | swake_up(&rdp->nocb_wq); |
fbce7497 PM |
2095 | } |
2096 | } | |
2097 | ||
2098 | /* If we (the leader) don't have CBs, go wait some more. */ | |
2099 | if (!my_rdp->nocb_follower_head) | |
2100 | goto wait_again; | |
2101 | } | |
2102 | ||
2103 | /* | |
2104 | * Followers come here to wait for additional callbacks to show up. | |
2105 | * This function does not return until callbacks appear. | |
2106 | */ | |
2107 | static void nocb_follower_wait(struct rcu_data *rdp) | |
2108 | { | |
2109 | bool firsttime = true; | |
2110 | ||
2111 | for (;;) { | |
2112 | if (!rcu_nocb_poll) { | |
2113 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2114 | "FollowerSleep"); | |
abedf8e2 | 2115 | swait_event_interruptible(rdp->nocb_wq, |
7d0ae808 | 2116 | READ_ONCE(rdp->nocb_follower_head)); |
fbce7497 PM |
2117 | } else if (firsttime) { |
2118 | /* Don't drown trace log with "Poll"! */ | |
2119 | firsttime = false; | |
2120 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll"); | |
2121 | } | |
2122 | if (smp_load_acquire(&rdp->nocb_follower_head)) { | |
2123 | /* ^^^ Ensure CB invocation follows _head test. */ | |
2124 | return; | |
2125 | } | |
2126 | if (!rcu_nocb_poll) | |
2127 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2128 | "WokeEmpty"); | |
73a860cd | 2129 | WARN_ON(signal_pending(current)); |
fbce7497 PM |
2130 | schedule_timeout_interruptible(1); |
2131 | } | |
2132 | } | |
2133 | ||
3fbfbf7a PM |
2134 | /* |
2135 | * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes | |
fbce7497 PM |
2136 | * callbacks queued by the corresponding no-CBs CPU, however, there is |
2137 | * an optional leader-follower relationship so that the grace-period | |
2138 | * kthreads don't have to do quite so many wakeups. | |
3fbfbf7a PM |
2139 | */ |
2140 | static int rcu_nocb_kthread(void *arg) | |
2141 | { | |
2142 | int c, cl; | |
2143 | struct rcu_head *list; | |
2144 | struct rcu_head *next; | |
2145 | struct rcu_head **tail; | |
2146 | struct rcu_data *rdp = arg; | |
2147 | ||
2148 | /* Each pass through this loop invokes one batch of callbacks */ | |
2149 | for (;;) { | |
fbce7497 PM |
2150 | /* Wait for callbacks. */ |
2151 | if (rdp->nocb_leader == rdp) | |
2152 | nocb_leader_wait(rdp); | |
2153 | else | |
2154 | nocb_follower_wait(rdp); | |
2155 | ||
2156 | /* Pull the ready-to-invoke callbacks onto local list. */ | |
7d0ae808 | 2157 | list = READ_ONCE(rdp->nocb_follower_head); |
fbce7497 PM |
2158 | BUG_ON(!list); |
2159 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty"); | |
7d0ae808 | 2160 | WRITE_ONCE(rdp->nocb_follower_head, NULL); |
fbce7497 | 2161 | tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head); |
3fbfbf7a PM |
2162 | |
2163 | /* Each pass through the following loop invokes a callback. */ | |
41050a00 PM |
2164 | trace_rcu_batch_start(rdp->rsp->name, |
2165 | atomic_long_read(&rdp->nocb_q_count_lazy), | |
2166 | atomic_long_read(&rdp->nocb_q_count), -1); | |
3fbfbf7a PM |
2167 | c = cl = 0; |
2168 | while (list) { | |
2169 | next = list->next; | |
2170 | /* Wait for enqueuing to complete, if needed. */ | |
2171 | while (next == NULL && &list->next != tail) { | |
69a79bb1 PM |
2172 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2173 | TPS("WaitQueue")); | |
3fbfbf7a | 2174 | schedule_timeout_interruptible(1); |
69a79bb1 PM |
2175 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2176 | TPS("WokeQueue")); | |
3fbfbf7a PM |
2177 | next = list->next; |
2178 | } | |
2179 | debug_rcu_head_unqueue(list); | |
2180 | local_bh_disable(); | |
2181 | if (__rcu_reclaim(rdp->rsp->name, list)) | |
2182 | cl++; | |
2183 | c++; | |
2184 | local_bh_enable(); | |
bedc1969 | 2185 | cond_resched_rcu_qs(); |
3fbfbf7a PM |
2186 | list = next; |
2187 | } | |
2188 | trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); | |
41050a00 PM |
2189 | smp_mb__before_atomic(); /* _add after CB invocation. */ |
2190 | atomic_long_add(-c, &rdp->nocb_q_count); | |
2191 | atomic_long_add(-cl, &rdp->nocb_q_count_lazy); | |
c635a4e1 | 2192 | rdp->n_nocbs_invoked += c; |
3fbfbf7a PM |
2193 | } |
2194 | return 0; | |
2195 | } | |
2196 | ||
96d3fd0d | 2197 | /* Is a deferred wakeup of rcu_nocb_kthread() required? */ |
9fdd3bc9 | 2198 | static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
96d3fd0d | 2199 | { |
7d0ae808 | 2200 | return READ_ONCE(rdp->nocb_defer_wakeup); |
96d3fd0d PM |
2201 | } |
2202 | ||
2203 | /* Do a deferred wakeup of rcu_nocb_kthread(). */ | |
2204 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2205 | { | |
9fdd3bc9 PM |
2206 | int ndw; |
2207 | ||
96d3fd0d PM |
2208 | if (!rcu_nocb_need_deferred_wakeup(rdp)) |
2209 | return; | |
7d0ae808 PM |
2210 | ndw = READ_ONCE(rdp->nocb_defer_wakeup); |
2211 | WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT); | |
9fdd3bc9 PM |
2212 | wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE); |
2213 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake")); | |
96d3fd0d PM |
2214 | } |
2215 | ||
f4579fc5 PM |
2216 | void __init rcu_init_nohz(void) |
2217 | { | |
2218 | int cpu; | |
2219 | bool need_rcu_nocb_mask = true; | |
2220 | struct rcu_state *rsp; | |
2221 | ||
2222 | #ifdef CONFIG_RCU_NOCB_CPU_NONE | |
2223 | need_rcu_nocb_mask = false; | |
2224 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */ | |
2225 | ||
2226 | #if defined(CONFIG_NO_HZ_FULL) | |
2227 | if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask)) | |
2228 | need_rcu_nocb_mask = true; | |
2229 | #endif /* #if defined(CONFIG_NO_HZ_FULL) */ | |
2230 | ||
2231 | if (!have_rcu_nocb_mask && need_rcu_nocb_mask) { | |
949cccdb PK |
2232 | if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) { |
2233 | pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n"); | |
2234 | return; | |
2235 | } | |
f4579fc5 PM |
2236 | have_rcu_nocb_mask = true; |
2237 | } | |
2238 | if (!have_rcu_nocb_mask) | |
2239 | return; | |
2240 | ||
2241 | #ifdef CONFIG_RCU_NOCB_CPU_ZERO | |
2242 | pr_info("\tOffload RCU callbacks from CPU 0\n"); | |
2243 | cpumask_set_cpu(0, rcu_nocb_mask); | |
2244 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */ | |
2245 | #ifdef CONFIG_RCU_NOCB_CPU_ALL | |
2246 | pr_info("\tOffload RCU callbacks from all CPUs\n"); | |
2247 | cpumask_copy(rcu_nocb_mask, cpu_possible_mask); | |
2248 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */ | |
2249 | #if defined(CONFIG_NO_HZ_FULL) | |
2250 | if (tick_nohz_full_running) | |
2251 | cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask); | |
2252 | #endif /* #if defined(CONFIG_NO_HZ_FULL) */ | |
2253 | ||
2254 | if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) { | |
2255 | pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n"); | |
2256 | cpumask_and(rcu_nocb_mask, cpu_possible_mask, | |
2257 | rcu_nocb_mask); | |
2258 | } | |
ad853b48 TH |
2259 | pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n", |
2260 | cpumask_pr_args(rcu_nocb_mask)); | |
f4579fc5 PM |
2261 | if (rcu_nocb_poll) |
2262 | pr_info("\tPoll for callbacks from no-CBs CPUs.\n"); | |
2263 | ||
2264 | for_each_rcu_flavor(rsp) { | |
34404ca8 PM |
2265 | for_each_cpu(cpu, rcu_nocb_mask) |
2266 | init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu)); | |
35ce7f29 | 2267 | rcu_organize_nocb_kthreads(rsp); |
f4579fc5 | 2268 | } |
96d3fd0d PM |
2269 | } |
2270 | ||
3fbfbf7a PM |
2271 | /* Initialize per-rcu_data variables for no-CBs CPUs. */ |
2272 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) | |
2273 | { | |
2274 | rdp->nocb_tail = &rdp->nocb_head; | |
abedf8e2 | 2275 | init_swait_queue_head(&rdp->nocb_wq); |
fbce7497 | 2276 | rdp->nocb_follower_tail = &rdp->nocb_follower_head; |
3fbfbf7a PM |
2277 | } |
2278 | ||
35ce7f29 PM |
2279 | /* |
2280 | * If the specified CPU is a no-CBs CPU that does not already have its | |
2281 | * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are | |
2282 | * brought online out of order, this can require re-organizing the | |
2283 | * leader-follower relationships. | |
2284 | */ | |
2285 | static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu) | |
2286 | { | |
2287 | struct rcu_data *rdp; | |
2288 | struct rcu_data *rdp_last; | |
2289 | struct rcu_data *rdp_old_leader; | |
2290 | struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu); | |
2291 | struct task_struct *t; | |
2292 | ||
2293 | /* | |
2294 | * If this isn't a no-CBs CPU or if it already has an rcuo kthread, | |
2295 | * then nothing to do. | |
2296 | */ | |
2297 | if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread) | |
2298 | return; | |
2299 | ||
2300 | /* If we didn't spawn the leader first, reorganize! */ | |
2301 | rdp_old_leader = rdp_spawn->nocb_leader; | |
2302 | if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) { | |
2303 | rdp_last = NULL; | |
2304 | rdp = rdp_old_leader; | |
2305 | do { | |
2306 | rdp->nocb_leader = rdp_spawn; | |
2307 | if (rdp_last && rdp != rdp_spawn) | |
2308 | rdp_last->nocb_next_follower = rdp; | |
bbe5d7a9 PM |
2309 | if (rdp == rdp_spawn) { |
2310 | rdp = rdp->nocb_next_follower; | |
2311 | } else { | |
2312 | rdp_last = rdp; | |
2313 | rdp = rdp->nocb_next_follower; | |
2314 | rdp_last->nocb_next_follower = NULL; | |
2315 | } | |
35ce7f29 PM |
2316 | } while (rdp); |
2317 | rdp_spawn->nocb_next_follower = rdp_old_leader; | |
2318 | } | |
2319 | ||
2320 | /* Spawn the kthread for this CPU and RCU flavor. */ | |
2321 | t = kthread_run(rcu_nocb_kthread, rdp_spawn, | |
2322 | "rcuo%c/%d", rsp->abbr, cpu); | |
2323 | BUG_ON(IS_ERR(t)); | |
7d0ae808 | 2324 | WRITE_ONCE(rdp_spawn->nocb_kthread, t); |
35ce7f29 PM |
2325 | } |
2326 | ||
2327 | /* | |
2328 | * If the specified CPU is a no-CBs CPU that does not already have its | |
2329 | * rcuo kthreads, spawn them. | |
2330 | */ | |
2331 | static void rcu_spawn_all_nocb_kthreads(int cpu) | |
2332 | { | |
2333 | struct rcu_state *rsp; | |
2334 | ||
2335 | if (rcu_scheduler_fully_active) | |
2336 | for_each_rcu_flavor(rsp) | |
2337 | rcu_spawn_one_nocb_kthread(rsp, cpu); | |
2338 | } | |
2339 | ||
2340 | /* | |
2341 | * Once the scheduler is running, spawn rcuo kthreads for all online | |
2342 | * no-CBs CPUs. This assumes that the early_initcall()s happen before | |
2343 | * non-boot CPUs come online -- if this changes, we will need to add | |
2344 | * some mutual exclusion. | |
2345 | */ | |
2346 | static void __init rcu_spawn_nocb_kthreads(void) | |
2347 | { | |
2348 | int cpu; | |
2349 | ||
2350 | for_each_online_cpu(cpu) | |
2351 | rcu_spawn_all_nocb_kthreads(cpu); | |
2352 | } | |
2353 | ||
fbce7497 PM |
2354 | /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */ |
2355 | static int rcu_nocb_leader_stride = -1; | |
2356 | module_param(rcu_nocb_leader_stride, int, 0444); | |
2357 | ||
2358 | /* | |
35ce7f29 | 2359 | * Initialize leader-follower relationships for all no-CBs CPU. |
fbce7497 | 2360 | */ |
35ce7f29 | 2361 | static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp) |
3fbfbf7a PM |
2362 | { |
2363 | int cpu; | |
fbce7497 PM |
2364 | int ls = rcu_nocb_leader_stride; |
2365 | int nl = 0; /* Next leader. */ | |
3fbfbf7a | 2366 | struct rcu_data *rdp; |
fbce7497 PM |
2367 | struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */ |
2368 | struct rcu_data *rdp_prev = NULL; | |
3fbfbf7a | 2369 | |
22c2f669 | 2370 | if (!have_rcu_nocb_mask) |
3fbfbf7a | 2371 | return; |
fbce7497 PM |
2372 | if (ls == -1) { |
2373 | ls = int_sqrt(nr_cpu_ids); | |
2374 | rcu_nocb_leader_stride = ls; | |
2375 | } | |
2376 | ||
2377 | /* | |
9831ce3b PM |
2378 | * Each pass through this loop sets up one rcu_data structure. |
2379 | * Should the corresponding CPU come online in the future, then | |
2380 | * we will spawn the needed set of rcu_nocb_kthread() kthreads. | |
fbce7497 | 2381 | */ |
3fbfbf7a PM |
2382 | for_each_cpu(cpu, rcu_nocb_mask) { |
2383 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
fbce7497 PM |
2384 | if (rdp->cpu >= nl) { |
2385 | /* New leader, set up for followers & next leader. */ | |
2386 | nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls; | |
2387 | rdp->nocb_leader = rdp; | |
2388 | rdp_leader = rdp; | |
2389 | } else { | |
2390 | /* Another follower, link to previous leader. */ | |
2391 | rdp->nocb_leader = rdp_leader; | |
2392 | rdp_prev->nocb_next_follower = rdp; | |
2393 | } | |
2394 | rdp_prev = rdp; | |
3fbfbf7a PM |
2395 | } |
2396 | } | |
2397 | ||
2398 | /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */ | |
34ed6246 | 2399 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2400 | { |
22c2f669 | 2401 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
34ed6246 | 2402 | return false; |
22c2f669 | 2403 | |
34404ca8 | 2404 | /* If there are early-boot callbacks, move them to nocb lists. */ |
15fecf89 PM |
2405 | if (!rcu_segcblist_empty(&rdp->cblist)) { |
2406 | rdp->nocb_head = rcu_segcblist_head(&rdp->cblist); | |
2407 | rdp->nocb_tail = rcu_segcblist_tail(&rdp->cblist); | |
2408 | atomic_long_set(&rdp->nocb_q_count, | |
2409 | rcu_segcblist_n_cbs(&rdp->cblist)); | |
2410 | atomic_long_set(&rdp->nocb_q_count_lazy, | |
2411 | rcu_segcblist_n_lazy_cbs(&rdp->cblist)); | |
2412 | rcu_segcblist_init(&rdp->cblist); | |
34404ca8 | 2413 | } |
15fecf89 | 2414 | rcu_segcblist_disable(&rdp->cblist); |
34ed6246 | 2415 | return true; |
3fbfbf7a PM |
2416 | } |
2417 | ||
34ed6246 PM |
2418 | #else /* #ifdef CONFIG_RCU_NOCB_CPU */ |
2419 | ||
d7e29933 PM |
2420 | static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu) |
2421 | { | |
2422 | WARN_ON_ONCE(1); /* Should be dead code. */ | |
2423 | return false; | |
2424 | } | |
2425 | ||
abedf8e2 | 2426 | static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq) |
3fbfbf7a | 2427 | { |
3fbfbf7a PM |
2428 | } |
2429 | ||
dae6e64d PM |
2430 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
2431 | { | |
2432 | } | |
2433 | ||
abedf8e2 | 2434 | static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp) |
065bb78c DW |
2435 | { |
2436 | return NULL; | |
2437 | } | |
2438 | ||
dae6e64d PM |
2439 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
2440 | { | |
2441 | } | |
3fbfbf7a | 2442 | |
3fbfbf7a | 2443 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
96d3fd0d | 2444 | bool lazy, unsigned long flags) |
3fbfbf7a | 2445 | { |
4afc7e26 | 2446 | return false; |
3fbfbf7a PM |
2447 | } |
2448 | ||
2449 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
2450 | struct rcu_data *rdp, |
2451 | unsigned long flags) | |
3fbfbf7a | 2452 | { |
f4aa84ba | 2453 | return false; |
3fbfbf7a PM |
2454 | } |
2455 | ||
3fbfbf7a PM |
2456 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
2457 | { | |
2458 | } | |
2459 | ||
9fdd3bc9 | 2460 | static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
96d3fd0d PM |
2461 | { |
2462 | return false; | |
2463 | } | |
2464 | ||
2465 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2466 | { | |
2467 | } | |
2468 | ||
35ce7f29 PM |
2469 | static void rcu_spawn_all_nocb_kthreads(int cpu) |
2470 | { | |
2471 | } | |
2472 | ||
2473 | static void __init rcu_spawn_nocb_kthreads(void) | |
3fbfbf7a PM |
2474 | { |
2475 | } | |
2476 | ||
34ed6246 | 2477 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2478 | { |
34ed6246 | 2479 | return false; |
3fbfbf7a PM |
2480 | } |
2481 | ||
2482 | #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */ | |
65d798f0 PM |
2483 | |
2484 | /* | |
2485 | * An adaptive-ticks CPU can potentially execute in kernel mode for an | |
2486 | * arbitrarily long period of time with the scheduling-clock tick turned | |
2487 | * off. RCU will be paying attention to this CPU because it is in the | |
2488 | * kernel, but the CPU cannot be guaranteed to be executing the RCU state | |
2489 | * machine because the scheduling-clock tick has been disabled. Therefore, | |
2490 | * if an adaptive-ticks CPU is failing to respond to the current grace | |
2491 | * period and has not be idle from an RCU perspective, kick it. | |
2492 | */ | |
4a81e832 | 2493 | static void __maybe_unused rcu_kick_nohz_cpu(int cpu) |
65d798f0 PM |
2494 | { |
2495 | #ifdef CONFIG_NO_HZ_FULL | |
2496 | if (tick_nohz_full_cpu(cpu)) | |
2497 | smp_send_reschedule(cpu); | |
2498 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ | |
2499 | } | |
2333210b PM |
2500 | |
2501 | ||
2502 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
2503 | ||
0edd1b17 | 2504 | static int full_sysidle_state; /* Current system-idle state. */ |
d4bd54fb PM |
2505 | #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */ |
2506 | #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */ | |
2507 | #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */ | |
2508 | #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */ | |
2509 | #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */ | |
2510 | ||
eb348b89 PM |
2511 | /* |
2512 | * Invoked to note exit from irq or task transition to idle. Note that | |
2513 | * usermode execution does -not- count as idle here! After all, we want | |
2514 | * to detect full-system idle states, not RCU quiescent states and grace | |
2515 | * periods. The caller must have disabled interrupts. | |
2516 | */ | |
28ced795 | 2517 | static void rcu_sysidle_enter(int irq) |
eb348b89 PM |
2518 | { |
2519 | unsigned long j; | |
28ced795 | 2520 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
eb348b89 | 2521 | |
663e1310 PM |
2522 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2523 | if (!tick_nohz_full_enabled()) | |
2524 | return; | |
2525 | ||
eb348b89 PM |
2526 | /* Adjust nesting, check for fully idle. */ |
2527 | if (irq) { | |
2528 | rdtp->dynticks_idle_nesting--; | |
2529 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2530 | if (rdtp->dynticks_idle_nesting != 0) | |
2531 | return; /* Still not fully idle. */ | |
2532 | } else { | |
2533 | if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) == | |
2534 | DYNTICK_TASK_NEST_VALUE) { | |
2535 | rdtp->dynticks_idle_nesting = 0; | |
2536 | } else { | |
2537 | rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE; | |
2538 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2539 | return; /* Still not fully idle. */ | |
2540 | } | |
2541 | } | |
2542 | ||
2543 | /* Record start of fully idle period. */ | |
2544 | j = jiffies; | |
7d0ae808 | 2545 | WRITE_ONCE(rdtp->dynticks_idle_jiffies, j); |
4e857c58 | 2546 | smp_mb__before_atomic(); |
eb348b89 | 2547 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2548 | smp_mb__after_atomic(); |
eb348b89 PM |
2549 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1); |
2550 | } | |
2551 | ||
0edd1b17 PM |
2552 | /* |
2553 | * Unconditionally force exit from full system-idle state. This is | |
2554 | * invoked when a normal CPU exits idle, but must be called separately | |
2555 | * for the timekeeping CPU (tick_do_timer_cpu). The reason for this | |
2556 | * is that the timekeeping CPU is permitted to take scheduling-clock | |
2557 | * interrupts while the system is in system-idle state, and of course | |
2558 | * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock | |
2559 | * interrupt from any other type of interrupt. | |
2560 | */ | |
2561 | void rcu_sysidle_force_exit(void) | |
2562 | { | |
7d0ae808 | 2563 | int oldstate = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2564 | int newoldstate; |
2565 | ||
2566 | /* | |
2567 | * Each pass through the following loop attempts to exit full | |
2568 | * system-idle state. If contention proves to be a problem, | |
2569 | * a trylock-based contention tree could be used here. | |
2570 | */ | |
2571 | while (oldstate > RCU_SYSIDLE_SHORT) { | |
2572 | newoldstate = cmpxchg(&full_sysidle_state, | |
2573 | oldstate, RCU_SYSIDLE_NOT); | |
2574 | if (oldstate == newoldstate && | |
2575 | oldstate == RCU_SYSIDLE_FULL_NOTED) { | |
2576 | rcu_kick_nohz_cpu(tick_do_timer_cpu); | |
2577 | return; /* We cleared it, done! */ | |
2578 | } | |
2579 | oldstate = newoldstate; | |
2580 | } | |
2581 | smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */ | |
2582 | } | |
2583 | ||
eb348b89 PM |
2584 | /* |
2585 | * Invoked to note entry to irq or task transition from idle. Note that | |
2586 | * usermode execution does -not- count as idle here! The caller must | |
2587 | * have disabled interrupts. | |
2588 | */ | |
28ced795 | 2589 | static void rcu_sysidle_exit(int irq) |
eb348b89 | 2590 | { |
28ced795 CL |
2591 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
2592 | ||
663e1310 PM |
2593 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2594 | if (!tick_nohz_full_enabled()) | |
2595 | return; | |
2596 | ||
eb348b89 PM |
2597 | /* Adjust nesting, check for already non-idle. */ |
2598 | if (irq) { | |
2599 | rdtp->dynticks_idle_nesting++; | |
2600 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2601 | if (rdtp->dynticks_idle_nesting != 1) | |
2602 | return; /* Already non-idle. */ | |
2603 | } else { | |
2604 | /* | |
2605 | * Allow for irq misnesting. Yes, it really is possible | |
2606 | * to enter an irq handler then never leave it, and maybe | |
2607 | * also vice versa. Handle both possibilities. | |
2608 | */ | |
2609 | if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) { | |
2610 | rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE; | |
2611 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2612 | return; /* Already non-idle. */ | |
2613 | } else { | |
2614 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE; | |
2615 | } | |
2616 | } | |
2617 | ||
2618 | /* Record end of idle period. */ | |
4e857c58 | 2619 | smp_mb__before_atomic(); |
eb348b89 | 2620 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2621 | smp_mb__after_atomic(); |
eb348b89 | 2622 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1)); |
0edd1b17 PM |
2623 | |
2624 | /* | |
2625 | * If we are the timekeeping CPU, we are permitted to be non-idle | |
2626 | * during a system-idle state. This must be the case, because | |
2627 | * the timekeeping CPU has to take scheduling-clock interrupts | |
2628 | * during the time that the system is transitioning to full | |
2629 | * system-idle state. This means that the timekeeping CPU must | |
2630 | * invoke rcu_sysidle_force_exit() directly if it does anything | |
2631 | * more than take a scheduling-clock interrupt. | |
2632 | */ | |
2633 | if (smp_processor_id() == tick_do_timer_cpu) | |
2634 | return; | |
2635 | ||
2636 | /* Update system-idle state: We are clearly no longer fully idle! */ | |
2637 | rcu_sysidle_force_exit(); | |
2638 | } | |
2639 | ||
2640 | /* | |
2641 | * Check to see if the current CPU is idle. Note that usermode execution | |
5871968d PM |
2642 | * does not count as idle. The caller must have disabled interrupts, |
2643 | * and must be running on tick_do_timer_cpu. | |
0edd1b17 PM |
2644 | */ |
2645 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, | |
2646 | unsigned long *maxj) | |
2647 | { | |
2648 | int cur; | |
2649 | unsigned long j; | |
2650 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
2651 | ||
663e1310 PM |
2652 | /* If there are no nohz_full= CPUs, don't check system-wide idleness. */ |
2653 | if (!tick_nohz_full_enabled()) | |
2654 | return; | |
2655 | ||
0edd1b17 PM |
2656 | /* |
2657 | * If some other CPU has already reported non-idle, if this is | |
2658 | * not the flavor of RCU that tracks sysidle state, or if this | |
2659 | * is an offline or the timekeeping CPU, nothing to do. | |
2660 | */ | |
417e8d26 | 2661 | if (!*isidle || rdp->rsp != rcu_state_p || |
0edd1b17 PM |
2662 | cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu) |
2663 | return; | |
5871968d PM |
2664 | /* Verify affinity of current kthread. */ |
2665 | WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu); | |
0edd1b17 PM |
2666 | |
2667 | /* Pick up current idle and NMI-nesting counter and check. */ | |
2668 | cur = atomic_read(&rdtp->dynticks_idle); | |
2669 | if (cur & 0x1) { | |
2670 | *isidle = false; /* We are not idle! */ | |
2671 | return; | |
2672 | } | |
2673 | smp_mb(); /* Read counters before timestamps. */ | |
2674 | ||
2675 | /* Pick up timestamps. */ | |
7d0ae808 | 2676 | j = READ_ONCE(rdtp->dynticks_idle_jiffies); |
0edd1b17 PM |
2677 | /* If this CPU entered idle more recently, update maxj timestamp. */ |
2678 | if (ULONG_CMP_LT(*maxj, j)) | |
2679 | *maxj = j; | |
2680 | } | |
2681 | ||
2682 | /* | |
2683 | * Is this the flavor of RCU that is handling full-system idle? | |
2684 | */ | |
2685 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2686 | { | |
417e8d26 | 2687 | return rsp == rcu_state_p; |
0edd1b17 PM |
2688 | } |
2689 | ||
2690 | /* | |
2691 | * Return a delay in jiffies based on the number of CPUs, rcu_node | |
2692 | * leaf fanout, and jiffies tick rate. The idea is to allow larger | |
2693 | * systems more time to transition to full-idle state in order to | |
2694 | * avoid the cache thrashing that otherwise occur on the state variable. | |
2695 | * Really small systems (less than a couple of tens of CPUs) should | |
2696 | * instead use a single global atomically incremented counter, and later | |
2697 | * versions of this will automatically reconfigure themselves accordingly. | |
2698 | */ | |
2699 | static unsigned long rcu_sysidle_delay(void) | |
2700 | { | |
2701 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2702 | return 0; | |
2703 | return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000); | |
2704 | } | |
2705 | ||
2706 | /* | |
2707 | * Advance the full-system-idle state. This is invoked when all of | |
2708 | * the non-timekeeping CPUs are idle. | |
2709 | */ | |
2710 | static void rcu_sysidle(unsigned long j) | |
2711 | { | |
2712 | /* Check the current state. */ | |
7d0ae808 | 2713 | switch (READ_ONCE(full_sysidle_state)) { |
0edd1b17 PM |
2714 | case RCU_SYSIDLE_NOT: |
2715 | ||
2716 | /* First time all are idle, so note a short idle period. */ | |
7d0ae808 | 2717 | WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT); |
0edd1b17 PM |
2718 | break; |
2719 | ||
2720 | case RCU_SYSIDLE_SHORT: | |
2721 | ||
2722 | /* | |
2723 | * Idle for a bit, time to advance to next state? | |
2724 | * cmpxchg failure means race with non-idle, let them win. | |
2725 | */ | |
2726 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2727 | (void)cmpxchg(&full_sysidle_state, | |
2728 | RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG); | |
2729 | break; | |
2730 | ||
2731 | case RCU_SYSIDLE_LONG: | |
2732 | ||
2733 | /* | |
2734 | * Do an additional check pass before advancing to full. | |
2735 | * cmpxchg failure means race with non-idle, let them win. | |
2736 | */ | |
2737 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2738 | (void)cmpxchg(&full_sysidle_state, | |
2739 | RCU_SYSIDLE_LONG, RCU_SYSIDLE_FULL); | |
2740 | break; | |
2741 | ||
2742 | default: | |
2743 | break; | |
2744 | } | |
2745 | } | |
2746 | ||
2747 | /* | |
2748 | * Found a non-idle non-timekeeping CPU, so kick the system-idle state | |
2749 | * back to the beginning. | |
2750 | */ | |
2751 | static void rcu_sysidle_cancel(void) | |
2752 | { | |
2753 | smp_mb(); | |
becb41bf | 2754 | if (full_sysidle_state > RCU_SYSIDLE_SHORT) |
7d0ae808 | 2755 | WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT); |
0edd1b17 PM |
2756 | } |
2757 | ||
2758 | /* | |
2759 | * Update the sysidle state based on the results of a force-quiescent-state | |
2760 | * scan of the CPUs' dyntick-idle state. | |
2761 | */ | |
2762 | static void rcu_sysidle_report(struct rcu_state *rsp, int isidle, | |
2763 | unsigned long maxj, bool gpkt) | |
2764 | { | |
417e8d26 | 2765 | if (rsp != rcu_state_p) |
0edd1b17 PM |
2766 | return; /* Wrong flavor, ignore. */ |
2767 | if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2768 | return; /* Running state machine from timekeeping CPU. */ | |
2769 | if (isidle) | |
2770 | rcu_sysidle(maxj); /* More idle! */ | |
2771 | else | |
2772 | rcu_sysidle_cancel(); /* Idle is over. */ | |
2773 | } | |
2774 | ||
2775 | /* | |
2776 | * Wrapper for rcu_sysidle_report() when called from the grace-period | |
2777 | * kthread's context. | |
2778 | */ | |
2779 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
2780 | unsigned long maxj) | |
2781 | { | |
663e1310 PM |
2782 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2783 | if (!tick_nohz_full_enabled()) | |
2784 | return; | |
2785 | ||
0edd1b17 PM |
2786 | rcu_sysidle_report(rsp, isidle, maxj, true); |
2787 | } | |
2788 | ||
2789 | /* Callback and function for forcing an RCU grace period. */ | |
2790 | struct rcu_sysidle_head { | |
2791 | struct rcu_head rh; | |
2792 | int inuse; | |
2793 | }; | |
2794 | ||
2795 | static void rcu_sysidle_cb(struct rcu_head *rhp) | |
2796 | { | |
2797 | struct rcu_sysidle_head *rshp; | |
2798 | ||
2799 | /* | |
2800 | * The following memory barrier is needed to replace the | |
2801 | * memory barriers that would normally be in the memory | |
2802 | * allocator. | |
2803 | */ | |
2804 | smp_mb(); /* grace period precedes setting inuse. */ | |
2805 | ||
2806 | rshp = container_of(rhp, struct rcu_sysidle_head, rh); | |
7d0ae808 | 2807 | WRITE_ONCE(rshp->inuse, 0); |
0edd1b17 PM |
2808 | } |
2809 | ||
2810 | /* | |
2811 | * Check to see if the system is fully idle, other than the timekeeping CPU. | |
663e1310 PM |
2812 | * The caller must have disabled interrupts. This is not intended to be |
2813 | * called unless tick_nohz_full_enabled(). | |
0edd1b17 PM |
2814 | */ |
2815 | bool rcu_sys_is_idle(void) | |
2816 | { | |
2817 | static struct rcu_sysidle_head rsh; | |
7d0ae808 | 2818 | int rss = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2819 | |
2820 | if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu)) | |
2821 | return false; | |
2822 | ||
2823 | /* Handle small-system case by doing a full scan of CPUs. */ | |
2824 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) { | |
2825 | int oldrss = rss - 1; | |
2826 | ||
2827 | /* | |
2828 | * One pass to advance to each state up to _FULL. | |
2829 | * Give up if any pass fails to advance the state. | |
2830 | */ | |
2831 | while (rss < RCU_SYSIDLE_FULL && oldrss < rss) { | |
2832 | int cpu; | |
2833 | bool isidle = true; | |
2834 | unsigned long maxj = jiffies - ULONG_MAX / 4; | |
2835 | struct rcu_data *rdp; | |
2836 | ||
2837 | /* Scan all the CPUs looking for nonidle CPUs. */ | |
2838 | for_each_possible_cpu(cpu) { | |
417e8d26 | 2839 | rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
0edd1b17 PM |
2840 | rcu_sysidle_check_cpu(rdp, &isidle, &maxj); |
2841 | if (!isidle) | |
2842 | break; | |
2843 | } | |
417e8d26 | 2844 | rcu_sysidle_report(rcu_state_p, isidle, maxj, false); |
0edd1b17 | 2845 | oldrss = rss; |
7d0ae808 | 2846 | rss = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2847 | } |
2848 | } | |
2849 | ||
2850 | /* If this is the first observation of an idle period, record it. */ | |
2851 | if (rss == RCU_SYSIDLE_FULL) { | |
2852 | rss = cmpxchg(&full_sysidle_state, | |
2853 | RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED); | |
2854 | return rss == RCU_SYSIDLE_FULL; | |
2855 | } | |
2856 | ||
2857 | smp_mb(); /* ensure rss load happens before later caller actions. */ | |
2858 | ||
2859 | /* If already fully idle, tell the caller (in case of races). */ | |
2860 | if (rss == RCU_SYSIDLE_FULL_NOTED) | |
2861 | return true; | |
2862 | ||
2863 | /* | |
2864 | * If we aren't there yet, and a grace period is not in flight, | |
2865 | * initiate a grace period. Either way, tell the caller that | |
2866 | * we are not there yet. We use an xchg() rather than an assignment | |
2867 | * to make up for the memory barriers that would otherwise be | |
2868 | * provided by the memory allocator. | |
2869 | */ | |
2870 | if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL && | |
417e8d26 | 2871 | !rcu_gp_in_progress(rcu_state_p) && |
0edd1b17 PM |
2872 | !rsh.inuse && xchg(&rsh.inuse, 1) == 0) |
2873 | call_rcu(&rsh.rh, rcu_sysidle_cb); | |
2874 | return false; | |
eb348b89 PM |
2875 | } |
2876 | ||
2333210b PM |
2877 | /* |
2878 | * Initialize dynticks sysidle state for CPUs coming online. | |
2879 | */ | |
2880 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) | |
2881 | { | |
2882 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE; | |
2883 | } | |
2884 | ||
2885 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
2886 | ||
28ced795 | 2887 | static void rcu_sysidle_enter(int irq) |
eb348b89 PM |
2888 | { |
2889 | } | |
2890 | ||
28ced795 | 2891 | static void rcu_sysidle_exit(int irq) |
eb348b89 PM |
2892 | { |
2893 | } | |
2894 | ||
0edd1b17 PM |
2895 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, |
2896 | unsigned long *maxj) | |
2897 | { | |
2898 | } | |
2899 | ||
2900 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2901 | { | |
2902 | return false; | |
2903 | } | |
2904 | ||
2905 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
2906 | unsigned long maxj) | |
2907 | { | |
2908 | } | |
2909 | ||
2333210b PM |
2910 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) |
2911 | { | |
2912 | } | |
2913 | ||
2914 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
a096932f PM |
2915 | |
2916 | /* | |
2917 | * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the | |
2918 | * grace-period kthread will do force_quiescent_state() processing? | |
2919 | * The idea is to avoid waking up RCU core processing on such a | |
2920 | * CPU unless the grace period has extended for too long. | |
2921 | * | |
2922 | * This code relies on the fact that all NO_HZ_FULL CPUs are also | |
52e2bb95 | 2923 | * CONFIG_RCU_NOCB_CPU CPUs. |
a096932f PM |
2924 | */ |
2925 | static bool rcu_nohz_full_cpu(struct rcu_state *rsp) | |
2926 | { | |
2927 | #ifdef CONFIG_NO_HZ_FULL | |
2928 | if (tick_nohz_full_cpu(smp_processor_id()) && | |
2929 | (!rcu_gp_in_progress(rsp) || | |
7d0ae808 | 2930 | ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ))) |
5ce035fb | 2931 | return true; |
a096932f | 2932 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ |
5ce035fb | 2933 | return false; |
a096932f | 2934 | } |
5057f55e PM |
2935 | |
2936 | /* | |
2937 | * Bind the grace-period kthread for the sysidle flavor of RCU to the | |
2938 | * timekeeping CPU. | |
2939 | */ | |
2940 | static void rcu_bind_gp_kthread(void) | |
2941 | { | |
c0f489d2 | 2942 | int __maybe_unused cpu; |
5057f55e | 2943 | |
c0f489d2 | 2944 | if (!tick_nohz_full_enabled()) |
5057f55e | 2945 | return; |
c0f489d2 PM |
2946 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
2947 | cpu = tick_do_timer_cpu; | |
5871968d | 2948 | if (cpu >= 0 && cpu < nr_cpu_ids) |
5057f55e | 2949 | set_cpus_allowed_ptr(current, cpumask_of(cpu)); |
c0f489d2 | 2950 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
5871968d | 2951 | housekeeping_affine(current); |
c0f489d2 | 2952 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
5057f55e | 2953 | } |
176f8f7a PM |
2954 | |
2955 | /* Record the current task on dyntick-idle entry. */ | |
2956 | static void rcu_dynticks_task_enter(void) | |
2957 | { | |
2958 | #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) | |
7d0ae808 | 2959 | WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id()); |
176f8f7a PM |
2960 | #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ |
2961 | } | |
2962 | ||
2963 | /* Record no current task on dyntick-idle exit. */ | |
2964 | static void rcu_dynticks_task_exit(void) | |
2965 | { | |
2966 | #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) | |
7d0ae808 | 2967 | WRITE_ONCE(current->rcu_tasks_idle_cpu, -1); |
176f8f7a PM |
2968 | #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ |
2969 | } |