Commit | Line | Data |
---|---|---|
f41d911f PM |
1 | /* |
2 | * Read-Copy Update mechanism for mutual exclusion (tree-based version) | |
3 | * Internal non-public definitions that provide either classic | |
6cc68793 | 4 | * or preemptible semantics. |
f41d911f PM |
5 | * |
6 | * This program is free software; you can redistribute it and/or modify | |
7 | * it under the terms of the GNU General Public License as published by | |
8 | * the Free Software Foundation; either version 2 of the License, or | |
9 | * (at your option) any later version. | |
10 | * | |
11 | * This program is distributed in the hope that it will be useful, | |
12 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
13 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
14 | * GNU General Public License for more details. | |
15 | * | |
16 | * You should have received a copy of the GNU General Public License | |
87de1cfd PM |
17 | * along with this program; if not, you can access it online at |
18 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
f41d911f PM |
19 | * |
20 | * Copyright Red Hat, 2009 | |
21 | * Copyright IBM Corporation, 2009 | |
22 | * | |
23 | * Author: Ingo Molnar <mingo@elte.hu> | |
24 | * Paul E. McKenney <paulmck@linux.vnet.ibm.com> | |
25 | */ | |
26 | ||
d9a3da06 | 27 | #include <linux/delay.h> |
3fbfbf7a | 28 | #include <linux/gfp.h> |
b626c1b6 | 29 | #include <linux/oom.h> |
62ab7072 | 30 | #include <linux/smpboot.h> |
4102adab | 31 | #include "../time/tick-internal.h" |
f41d911f | 32 | |
5b61b0ba | 33 | #ifdef CONFIG_RCU_BOOST |
61cfd097 | 34 | |
abaa93d9 | 35 | #include "../locking/rtmutex_common.h" |
21871d7e | 36 | |
61cfd097 PM |
37 | /* |
38 | * Control variables for per-CPU and per-rcu_node kthreads. These | |
39 | * handle all flavors of RCU. | |
40 | */ | |
41 | static DEFINE_PER_CPU(struct task_struct *, rcu_cpu_kthread_task); | |
42 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_status); | |
43 | DEFINE_PER_CPU(unsigned int, rcu_cpu_kthread_loops); | |
44 | DEFINE_PER_CPU(char, rcu_cpu_has_work); | |
45 | ||
727b705b PM |
46 | #else /* #ifdef CONFIG_RCU_BOOST */ |
47 | ||
48 | /* | |
49 | * Some architectures do not define rt_mutexes, but if !CONFIG_RCU_BOOST, | |
50 | * all uses are in dead code. Provide a definition to keep the compiler | |
51 | * happy, but add WARN_ON_ONCE() to complain if used in the wrong place. | |
52 | * This probably needs to be excluded from -rt builds. | |
53 | */ | |
54 | #define rt_mutex_owner(a) ({ WARN_ON_ONCE(1); NULL; }) | |
55 | ||
56 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ | |
5b61b0ba | 57 | |
3fbfbf7a PM |
58 | #ifdef CONFIG_RCU_NOCB_CPU |
59 | static cpumask_var_t rcu_nocb_mask; /* CPUs to have callbacks offloaded. */ | |
60 | static bool have_rcu_nocb_mask; /* Was rcu_nocb_mask allocated? */ | |
1b0048a4 | 61 | static bool __read_mostly rcu_nocb_poll; /* Offload kthread are to poll. */ |
3fbfbf7a PM |
62 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU */ |
63 | ||
26845c28 PM |
64 | /* |
65 | * Check the RCU kernel configuration parameters and print informative | |
699d4035 | 66 | * messages about anything out of the ordinary. |
26845c28 PM |
67 | */ |
68 | static void __init rcu_bootup_announce_oddness(void) | |
69 | { | |
ab6f5bd6 PM |
70 | if (IS_ENABLED(CONFIG_RCU_TRACE)) |
71 | pr_info("\tRCU debugfs-based tracing is enabled.\n"); | |
05c5df31 PM |
72 | if ((IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 64) || |
73 | (!IS_ENABLED(CONFIG_64BIT) && RCU_FANOUT != 32)) | |
ab6f5bd6 | 74 | pr_info("\tCONFIG_RCU_FANOUT set to non-default value of %d\n", |
05c5df31 | 75 | RCU_FANOUT); |
7fa27001 | 76 | if (rcu_fanout_exact) |
ab6f5bd6 PM |
77 | pr_info("\tHierarchical RCU autobalancing is disabled.\n"); |
78 | if (IS_ENABLED(CONFIG_RCU_FAST_NO_HZ)) | |
79 | pr_info("\tRCU dyntick-idle grace-period acceleration is enabled.\n"); | |
80 | if (IS_ENABLED(CONFIG_PROVE_RCU)) | |
81 | pr_info("\tRCU lockdep checking is enabled.\n"); | |
42621697 AG |
82 | if (RCU_NUM_LVLS >= 4) |
83 | pr_info("\tFour(or more)-level hierarchy is enabled.\n"); | |
47d631af | 84 | if (RCU_FANOUT_LEAF != 16) |
a3bd2c09 | 85 | pr_info("\tBuild-time adjustment of leaf fanout to %d.\n", |
47d631af PM |
86 | RCU_FANOUT_LEAF); |
87 | if (rcu_fanout_leaf != RCU_FANOUT_LEAF) | |
9a5739d7 | 88 | pr_info("\tBoot-time adjustment of leaf fanout to %d.\n", rcu_fanout_leaf); |
cca6f393 | 89 | if (nr_cpu_ids != NR_CPUS) |
efc151c3 | 90 | pr_info("\tRCU restricting CPUs from NR_CPUS=%d to nr_cpu_ids=%d.\n", NR_CPUS, nr_cpu_ids); |
ab6f5bd6 PM |
91 | if (IS_ENABLED(CONFIG_RCU_BOOST)) |
92 | pr_info("\tRCU kthread priority: %d.\n", kthread_prio); | |
26845c28 PM |
93 | } |
94 | ||
28f6569a | 95 | #ifdef CONFIG_PREEMPT_RCU |
f41d911f | 96 | |
a41bfeb2 | 97 | RCU_STATE_INITIALIZER(rcu_preempt, 'p', call_rcu); |
b28a7c01 | 98 | static struct rcu_state *const rcu_state_p = &rcu_preempt_state; |
2927a689 | 99 | static struct rcu_data __percpu *const rcu_data_p = &rcu_preempt_data; |
f41d911f | 100 | |
d19fb8d1 PM |
101 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
102 | bool wake); | |
d9a3da06 | 103 | |
f41d911f PM |
104 | /* |
105 | * Tell them what RCU they are running. | |
106 | */ | |
0e0fc1c2 | 107 | static void __init rcu_bootup_announce(void) |
f41d911f | 108 | { |
efc151c3 | 109 | pr_info("Preemptible hierarchical RCU implementation.\n"); |
26845c28 | 110 | rcu_bootup_announce_oddness(); |
f41d911f PM |
111 | } |
112 | ||
8203d6d0 PM |
113 | /* Flags for rcu_preempt_ctxt_queue() decision table. */ |
114 | #define RCU_GP_TASKS 0x8 | |
115 | #define RCU_EXP_TASKS 0x4 | |
116 | #define RCU_GP_BLKD 0x2 | |
117 | #define RCU_EXP_BLKD 0x1 | |
118 | ||
119 | /* | |
120 | * Queues a task preempted within an RCU-preempt read-side critical | |
121 | * section into the appropriate location within the ->blkd_tasks list, | |
122 | * depending on the states of any ongoing normal and expedited grace | |
123 | * periods. The ->gp_tasks pointer indicates which element the normal | |
124 | * grace period is waiting on (NULL if none), and the ->exp_tasks pointer | |
125 | * indicates which element the expedited grace period is waiting on (again, | |
126 | * NULL if none). If a grace period is waiting on a given element in the | |
127 | * ->blkd_tasks list, it also waits on all subsequent elements. Thus, | |
128 | * adding a task to the tail of the list blocks any grace period that is | |
129 | * already waiting on one of the elements. In contrast, adding a task | |
130 | * to the head of the list won't block any grace period that is already | |
131 | * waiting on one of the elements. | |
132 | * | |
133 | * This queuing is imprecise, and can sometimes make an ongoing grace | |
134 | * period wait for a task that is not strictly speaking blocking it. | |
135 | * Given the choice, we needlessly block a normal grace period rather than | |
136 | * blocking an expedited grace period. | |
137 | * | |
138 | * Note that an endless sequence of expedited grace periods still cannot | |
139 | * indefinitely postpone a normal grace period. Eventually, all of the | |
140 | * fixed number of preempted tasks blocking the normal grace period that are | |
141 | * not also blocking the expedited grace period will resume and complete | |
142 | * their RCU read-side critical sections. At that point, the ->gp_tasks | |
143 | * pointer will equal the ->exp_tasks pointer, at which point the end of | |
144 | * the corresponding expedited grace period will also be the end of the | |
145 | * normal grace period. | |
146 | */ | |
46a5d164 PM |
147 | static void rcu_preempt_ctxt_queue(struct rcu_node *rnp, struct rcu_data *rdp) |
148 | __releases(rnp->lock) /* But leaves rrupts disabled. */ | |
8203d6d0 PM |
149 | { |
150 | int blkd_state = (rnp->gp_tasks ? RCU_GP_TASKS : 0) + | |
151 | (rnp->exp_tasks ? RCU_EXP_TASKS : 0) + | |
152 | (rnp->qsmask & rdp->grpmask ? RCU_GP_BLKD : 0) + | |
153 | (rnp->expmask & rdp->grpmask ? RCU_EXP_BLKD : 0); | |
154 | struct task_struct *t = current; | |
155 | ||
156 | /* | |
157 | * Decide where to queue the newly blocked task. In theory, | |
158 | * this could be an if-statement. In practice, when I tried | |
159 | * that, it was quite messy. | |
160 | */ | |
161 | switch (blkd_state) { | |
162 | case 0: | |
163 | case RCU_EXP_TASKS: | |
164 | case RCU_EXP_TASKS + RCU_GP_BLKD: | |
165 | case RCU_GP_TASKS: | |
166 | case RCU_GP_TASKS + RCU_EXP_TASKS: | |
167 | ||
168 | /* | |
169 | * Blocking neither GP, or first task blocking the normal | |
170 | * GP but not blocking the already-waiting expedited GP. | |
171 | * Queue at the head of the list to avoid unnecessarily | |
172 | * blocking the already-waiting GPs. | |
173 | */ | |
174 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); | |
175 | break; | |
176 | ||
177 | case RCU_EXP_BLKD: | |
178 | case RCU_GP_BLKD: | |
179 | case RCU_GP_BLKD + RCU_EXP_BLKD: | |
180 | case RCU_GP_TASKS + RCU_EXP_BLKD: | |
181 | case RCU_GP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
182 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
183 | ||
184 | /* | |
185 | * First task arriving that blocks either GP, or first task | |
186 | * arriving that blocks the expedited GP (with the normal | |
187 | * GP already waiting), or a task arriving that blocks | |
188 | * both GPs with both GPs already waiting. Queue at the | |
189 | * tail of the list to avoid any GP waiting on any of the | |
190 | * already queued tasks that are not blocking it. | |
191 | */ | |
192 | list_add_tail(&t->rcu_node_entry, &rnp->blkd_tasks); | |
193 | break; | |
194 | ||
195 | case RCU_EXP_TASKS + RCU_EXP_BLKD: | |
196 | case RCU_EXP_TASKS + RCU_GP_BLKD + RCU_EXP_BLKD: | |
197 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_EXP_BLKD: | |
198 | ||
199 | /* | |
200 | * Second or subsequent task blocking the expedited GP. | |
201 | * The task either does not block the normal GP, or is the | |
202 | * first task blocking the normal GP. Queue just after | |
203 | * the first task blocking the expedited GP. | |
204 | */ | |
205 | list_add(&t->rcu_node_entry, rnp->exp_tasks); | |
206 | break; | |
207 | ||
208 | case RCU_GP_TASKS + RCU_GP_BLKD: | |
209 | case RCU_GP_TASKS + RCU_EXP_TASKS + RCU_GP_BLKD: | |
210 | ||
211 | /* | |
212 | * Second or subsequent task blocking the normal GP. | |
213 | * The task does not block the expedited GP. Queue just | |
214 | * after the first task blocking the normal GP. | |
215 | */ | |
216 | list_add(&t->rcu_node_entry, rnp->gp_tasks); | |
217 | break; | |
218 | ||
219 | default: | |
220 | ||
221 | /* Yet another exercise in excessive paranoia. */ | |
222 | WARN_ON_ONCE(1); | |
223 | break; | |
224 | } | |
225 | ||
226 | /* | |
227 | * We have now queued the task. If it was the first one to | |
228 | * block either grace period, update the ->gp_tasks and/or | |
229 | * ->exp_tasks pointers, respectively, to reference the newly | |
230 | * blocked tasks. | |
231 | */ | |
232 | if (!rnp->gp_tasks && (blkd_state & RCU_GP_BLKD)) | |
233 | rnp->gp_tasks = &t->rcu_node_entry; | |
234 | if (!rnp->exp_tasks && (blkd_state & RCU_EXP_BLKD)) | |
235 | rnp->exp_tasks = &t->rcu_node_entry; | |
67c583a7 | 236 | raw_spin_unlock_rcu_node(rnp); /* interrupts remain disabled. */ |
8203d6d0 PM |
237 | |
238 | /* | |
239 | * Report the quiescent state for the expedited GP. This expedited | |
240 | * GP should not be able to end until we report, so there should be | |
241 | * no need to check for a subsequent expedited GP. (Though we are | |
242 | * still in a quiescent state in any case.) | |
243 | */ | |
244 | if (blkd_state & RCU_EXP_BLKD && | |
245 | t->rcu_read_unlock_special.b.exp_need_qs) { | |
246 | t->rcu_read_unlock_special.b.exp_need_qs = false; | |
247 | rcu_report_exp_rdp(rdp->rsp, rdp, true); | |
248 | } else { | |
249 | WARN_ON_ONCE(t->rcu_read_unlock_special.b.exp_need_qs); | |
250 | } | |
8203d6d0 PM |
251 | } |
252 | ||
f41d911f | 253 | /* |
6cc68793 | 254 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
f41d911f PM |
255 | * that this just means that the task currently running on the CPU is |
256 | * not in a quiescent state. There might be any number of tasks blocked | |
257 | * while in an RCU read-side critical section. | |
25502a6c | 258 | * |
1d082fd0 PM |
259 | * As with the other rcu_*_qs() functions, callers to this function |
260 | * must disable preemption. | |
f41d911f | 261 | */ |
284a8c93 | 262 | static void rcu_preempt_qs(void) |
f41d911f | 263 | { |
5b74c458 | 264 | if (__this_cpu_read(rcu_data_p->cpu_no_qs.s)) { |
284a8c93 | 265 | trace_rcu_grace_period(TPS("rcu_preempt"), |
2927a689 | 266 | __this_cpu_read(rcu_data_p->gpnum), |
284a8c93 | 267 | TPS("cpuqs")); |
5b74c458 | 268 | __this_cpu_write(rcu_data_p->cpu_no_qs.b.norm, false); |
284a8c93 PM |
269 | barrier(); /* Coordinate with rcu_preempt_check_callbacks(). */ |
270 | current->rcu_read_unlock_special.b.need_qs = false; | |
271 | } | |
f41d911f PM |
272 | } |
273 | ||
274 | /* | |
c3422bea PM |
275 | * We have entered the scheduler, and the current task might soon be |
276 | * context-switched away from. If this task is in an RCU read-side | |
277 | * critical section, we will no longer be able to rely on the CPU to | |
12f5f524 PM |
278 | * record that fact, so we enqueue the task on the blkd_tasks list. |
279 | * The task will dequeue itself when it exits the outermost enclosing | |
280 | * RCU read-side critical section. Therefore, the current grace period | |
281 | * cannot be permitted to complete until the blkd_tasks list entries | |
282 | * predating the current grace period drain, in other words, until | |
283 | * rnp->gp_tasks becomes NULL. | |
c3422bea | 284 | * |
46a5d164 | 285 | * Caller must disable interrupts. |
f41d911f | 286 | */ |
38200cf2 | 287 | static void rcu_preempt_note_context_switch(void) |
f41d911f PM |
288 | { |
289 | struct task_struct *t = current; | |
f41d911f PM |
290 | struct rcu_data *rdp; |
291 | struct rcu_node *rnp; | |
292 | ||
10f39bb1 | 293 | if (t->rcu_read_lock_nesting > 0 && |
1d082fd0 | 294 | !t->rcu_read_unlock_special.b.blocked) { |
f41d911f PM |
295 | |
296 | /* Possibly blocking in an RCU read-side critical section. */ | |
e63c887c | 297 | rdp = this_cpu_ptr(rcu_state_p->rda); |
f41d911f | 298 | rnp = rdp->mynode; |
46a5d164 | 299 | raw_spin_lock_rcu_node(rnp); |
1d082fd0 | 300 | t->rcu_read_unlock_special.b.blocked = true; |
86848966 | 301 | t->rcu_blocked_node = rnp; |
f41d911f PM |
302 | |
303 | /* | |
8203d6d0 PM |
304 | * Verify the CPU's sanity, trace the preemption, and |
305 | * then queue the task as required based on the states | |
306 | * of any ongoing and expedited grace periods. | |
f41d911f | 307 | */ |
0aa04b05 | 308 | WARN_ON_ONCE((rdp->grpmask & rcu_rnp_online_cpus(rnp)) == 0); |
e7d8842e | 309 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
d4c08f2a PM |
310 | trace_rcu_preempt_task(rdp->rsp->name, |
311 | t->pid, | |
312 | (rnp->qsmask & rdp->grpmask) | |
313 | ? rnp->gpnum | |
314 | : rnp->gpnum + 1); | |
46a5d164 | 315 | rcu_preempt_ctxt_queue(rnp, rdp); |
10f39bb1 | 316 | } else if (t->rcu_read_lock_nesting < 0 && |
1d082fd0 | 317 | t->rcu_read_unlock_special.s) { |
10f39bb1 PM |
318 | |
319 | /* | |
320 | * Complete exit from RCU read-side critical section on | |
321 | * behalf of preempted instance of __rcu_read_unlock(). | |
322 | */ | |
323 | rcu_read_unlock_special(t); | |
f41d911f PM |
324 | } |
325 | ||
326 | /* | |
327 | * Either we were not in an RCU read-side critical section to | |
328 | * begin with, or we have now recorded that critical section | |
329 | * globally. Either way, we can now note a quiescent state | |
330 | * for this CPU. Again, if we were in an RCU read-side critical | |
331 | * section, and if that critical section was blocking the current | |
332 | * grace period, then the fact that the task has been enqueued | |
333 | * means that we continue to block the current grace period. | |
334 | */ | |
284a8c93 | 335 | rcu_preempt_qs(); |
f41d911f PM |
336 | } |
337 | ||
fc2219d4 PM |
338 | /* |
339 | * Check for preempted RCU readers blocking the current grace period | |
340 | * for the specified rcu_node structure. If the caller needs a reliable | |
341 | * answer, it must hold the rcu_node's ->lock. | |
342 | */ | |
27f4d280 | 343 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 | 344 | { |
12f5f524 | 345 | return rnp->gp_tasks != NULL; |
fc2219d4 PM |
346 | } |
347 | ||
12f5f524 PM |
348 | /* |
349 | * Advance a ->blkd_tasks-list pointer to the next entry, instead | |
350 | * returning NULL if at the end of the list. | |
351 | */ | |
352 | static struct list_head *rcu_next_node_entry(struct task_struct *t, | |
353 | struct rcu_node *rnp) | |
354 | { | |
355 | struct list_head *np; | |
356 | ||
357 | np = t->rcu_node_entry.next; | |
358 | if (np == &rnp->blkd_tasks) | |
359 | np = NULL; | |
360 | return np; | |
361 | } | |
362 | ||
8af3a5e7 PM |
363 | /* |
364 | * Return true if the specified rcu_node structure has tasks that were | |
365 | * preempted within an RCU read-side critical section. | |
366 | */ | |
367 | static bool rcu_preempt_has_tasks(struct rcu_node *rnp) | |
368 | { | |
369 | return !list_empty(&rnp->blkd_tasks); | |
370 | } | |
371 | ||
b668c9cf PM |
372 | /* |
373 | * Handle special cases during rcu_read_unlock(), such as needing to | |
374 | * notify RCU core processing or task having blocked during the RCU | |
375 | * read-side critical section. | |
376 | */ | |
2a3fa843 | 377 | void rcu_read_unlock_special(struct task_struct *t) |
f41d911f | 378 | { |
b6a932d1 PM |
379 | bool empty_exp; |
380 | bool empty_norm; | |
381 | bool empty_exp_now; | |
f41d911f | 382 | unsigned long flags; |
12f5f524 | 383 | struct list_head *np; |
abaa93d9 | 384 | bool drop_boost_mutex = false; |
8203d6d0 | 385 | struct rcu_data *rdp; |
f41d911f | 386 | struct rcu_node *rnp; |
1d082fd0 | 387 | union rcu_special special; |
f41d911f PM |
388 | |
389 | /* NMI handlers cannot block and cannot safely manipulate state. */ | |
390 | if (in_nmi()) | |
391 | return; | |
392 | ||
393 | local_irq_save(flags); | |
394 | ||
395 | /* | |
8203d6d0 PM |
396 | * If RCU core is waiting for this CPU to exit its critical section, |
397 | * report the fact that it has exited. Because irqs are disabled, | |
1d082fd0 | 398 | * t->rcu_read_unlock_special cannot change. |
f41d911f PM |
399 | */ |
400 | special = t->rcu_read_unlock_special; | |
1d082fd0 | 401 | if (special.b.need_qs) { |
284a8c93 | 402 | rcu_preempt_qs(); |
c0135d07 | 403 | t->rcu_read_unlock_special.b.need_qs = false; |
1d082fd0 | 404 | if (!t->rcu_read_unlock_special.s) { |
79a62f95 LJ |
405 | local_irq_restore(flags); |
406 | return; | |
407 | } | |
f41d911f PM |
408 | } |
409 | ||
8203d6d0 PM |
410 | /* |
411 | * Respond to a request for an expedited grace period, but only if | |
412 | * we were not preempted, meaning that we were running on the same | |
413 | * CPU throughout. If we were preempted, the exp_need_qs flag | |
414 | * would have been cleared at the time of the first preemption, | |
415 | * and the quiescent state would be reported when we were dequeued. | |
416 | */ | |
417 | if (special.b.exp_need_qs) { | |
418 | WARN_ON_ONCE(special.b.blocked); | |
419 | t->rcu_read_unlock_special.b.exp_need_qs = false; | |
420 | rdp = this_cpu_ptr(rcu_state_p->rda); | |
421 | rcu_report_exp_rdp(rcu_state_p, rdp, true); | |
422 | if (!t->rcu_read_unlock_special.s) { | |
423 | local_irq_restore(flags); | |
424 | return; | |
425 | } | |
426 | } | |
427 | ||
79a62f95 | 428 | /* Hardware IRQ handlers cannot block, complain if they get here. */ |
d24209bb PM |
429 | if (in_irq() || in_serving_softirq()) { |
430 | lockdep_rcu_suspicious(__FILE__, __LINE__, | |
431 | "rcu_read_unlock() from irq or softirq with blocking in critical section!!!\n"); | |
8203d6d0 | 432 | pr_alert("->rcu_read_unlock_special: %#x (b: %d, enq: %d nq: %d)\n", |
d24209bb PM |
433 | t->rcu_read_unlock_special.s, |
434 | t->rcu_read_unlock_special.b.blocked, | |
8203d6d0 | 435 | t->rcu_read_unlock_special.b.exp_need_qs, |
d24209bb | 436 | t->rcu_read_unlock_special.b.need_qs); |
f41d911f PM |
437 | local_irq_restore(flags); |
438 | return; | |
439 | } | |
440 | ||
441 | /* Clean up if blocked during RCU read-side critical section. */ | |
1d082fd0 PM |
442 | if (special.b.blocked) { |
443 | t->rcu_read_unlock_special.b.blocked = false; | |
f41d911f | 444 | |
dd5d19ba | 445 | /* |
0a0ba1c9 | 446 | * Remove this task from the list it blocked on. The task |
8ba9153b PM |
447 | * now remains queued on the rcu_node corresponding to the |
448 | * CPU it first blocked on, so there is no longer any need | |
449 | * to loop. Retain a WARN_ON_ONCE() out of sheer paranoia. | |
dd5d19ba | 450 | */ |
8ba9153b PM |
451 | rnp = t->rcu_blocked_node; |
452 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ | |
453 | WARN_ON_ONCE(rnp != t->rcu_blocked_node); | |
74e871ac | 454 | empty_norm = !rcu_preempt_blocked_readers_cgp(rnp); |
8203d6d0 | 455 | empty_exp = sync_rcu_preempt_exp_done(rnp); |
d9a3da06 | 456 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ |
12f5f524 | 457 | np = rcu_next_node_entry(t, rnp); |
f41d911f | 458 | list_del_init(&t->rcu_node_entry); |
82e78d80 | 459 | t->rcu_blocked_node = NULL; |
f7f7bac9 | 460 | trace_rcu_unlock_preempted_task(TPS("rcu_preempt"), |
d4c08f2a | 461 | rnp->gpnum, t->pid); |
12f5f524 PM |
462 | if (&t->rcu_node_entry == rnp->gp_tasks) |
463 | rnp->gp_tasks = np; | |
464 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
465 | rnp->exp_tasks = np; | |
727b705b PM |
466 | if (IS_ENABLED(CONFIG_RCU_BOOST)) { |
467 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
468 | rnp->boost_tasks = np; | |
469 | /* Snapshot ->boost_mtx ownership w/rnp->lock held. */ | |
470 | drop_boost_mutex = rt_mutex_owner(&rnp->boost_mtx) == t; | |
471 | } | |
f41d911f PM |
472 | |
473 | /* | |
474 | * If this was the last task on the current list, and if | |
475 | * we aren't waiting on any CPUs, report the quiescent state. | |
389abd48 PM |
476 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock, |
477 | * so we must take a snapshot of the expedited state. | |
f41d911f | 478 | */ |
8203d6d0 | 479 | empty_exp_now = sync_rcu_preempt_exp_done(rnp); |
74e871ac | 480 | if (!empty_norm && !rcu_preempt_blocked_readers_cgp(rnp)) { |
f7f7bac9 | 481 | trace_rcu_quiescent_state_report(TPS("preempt_rcu"), |
d4c08f2a PM |
482 | rnp->gpnum, |
483 | 0, rnp->qsmask, | |
484 | rnp->level, | |
485 | rnp->grplo, | |
486 | rnp->grphi, | |
487 | !!rnp->gp_tasks); | |
e63c887c | 488 | rcu_report_unblock_qs_rnp(rcu_state_p, rnp, flags); |
c701d5d9 | 489 | } else { |
67c583a7 | 490 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
c701d5d9 | 491 | } |
d9a3da06 | 492 | |
27f4d280 | 493 | /* Unboost if we were boosted. */ |
727b705b | 494 | if (IS_ENABLED(CONFIG_RCU_BOOST) && drop_boost_mutex) |
abaa93d9 | 495 | rt_mutex_unlock(&rnp->boost_mtx); |
27f4d280 | 496 | |
d9a3da06 PM |
497 | /* |
498 | * If this was the last task on the expedited lists, | |
499 | * then we need to report up the rcu_node hierarchy. | |
500 | */ | |
389abd48 | 501 | if (!empty_exp && empty_exp_now) |
e63c887c | 502 | rcu_report_exp_rnp(rcu_state_p, rnp, true); |
b668c9cf PM |
503 | } else { |
504 | local_irq_restore(flags); | |
f41d911f | 505 | } |
f41d911f PM |
506 | } |
507 | ||
1ed509a2 PM |
508 | /* |
509 | * Dump detailed information for all tasks blocking the current RCU | |
510 | * grace period on the specified rcu_node structure. | |
511 | */ | |
512 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) | |
513 | { | |
514 | unsigned long flags; | |
1ed509a2 PM |
515 | struct task_struct *t; |
516 | ||
6cf10081 | 517 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
5fd4dc06 | 518 | if (!rcu_preempt_blocked_readers_cgp(rnp)) { |
67c583a7 | 519 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
5fd4dc06 PM |
520 | return; |
521 | } | |
82efed06 | 522 | t = list_entry(rnp->gp_tasks->prev, |
12f5f524 PM |
523 | struct task_struct, rcu_node_entry); |
524 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) | |
525 | sched_show_task(t); | |
67c583a7 | 526 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1ed509a2 PM |
527 | } |
528 | ||
529 | /* | |
530 | * Dump detailed information for all tasks blocking the current RCU | |
531 | * grace period. | |
532 | */ | |
533 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
534 | { | |
535 | struct rcu_node *rnp = rcu_get_root(rsp); | |
536 | ||
537 | rcu_print_detail_task_stall_rnp(rnp); | |
538 | rcu_for_each_leaf_node(rsp, rnp) | |
539 | rcu_print_detail_task_stall_rnp(rnp); | |
540 | } | |
541 | ||
a858af28 PM |
542 | static void rcu_print_task_stall_begin(struct rcu_node *rnp) |
543 | { | |
efc151c3 | 544 | pr_err("\tTasks blocked on level-%d rcu_node (CPUs %d-%d):", |
a858af28 PM |
545 | rnp->level, rnp->grplo, rnp->grphi); |
546 | } | |
547 | ||
548 | static void rcu_print_task_stall_end(void) | |
549 | { | |
efc151c3 | 550 | pr_cont("\n"); |
a858af28 PM |
551 | } |
552 | ||
f41d911f PM |
553 | /* |
554 | * Scan the current list of tasks blocked within RCU read-side critical | |
555 | * sections, printing out the tid of each. | |
556 | */ | |
9bc8b558 | 557 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 558 | { |
f41d911f | 559 | struct task_struct *t; |
9bc8b558 | 560 | int ndetected = 0; |
f41d911f | 561 | |
27f4d280 | 562 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
9bc8b558 | 563 | return 0; |
a858af28 | 564 | rcu_print_task_stall_begin(rnp); |
82efed06 | 565 | t = list_entry(rnp->gp_tasks->prev, |
12f5f524 | 566 | struct task_struct, rcu_node_entry); |
9bc8b558 | 567 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { |
efc151c3 | 568 | pr_cont(" P%d", t->pid); |
9bc8b558 PM |
569 | ndetected++; |
570 | } | |
a858af28 | 571 | rcu_print_task_stall_end(); |
9bc8b558 | 572 | return ndetected; |
f41d911f PM |
573 | } |
574 | ||
74611ecb PM |
575 | /* |
576 | * Scan the current list of tasks blocked within RCU read-side critical | |
577 | * sections, printing out the tid of each that is blocking the current | |
578 | * expedited grace period. | |
579 | */ | |
580 | static int rcu_print_task_exp_stall(struct rcu_node *rnp) | |
581 | { | |
582 | struct task_struct *t; | |
583 | int ndetected = 0; | |
584 | ||
585 | if (!rnp->exp_tasks) | |
586 | return 0; | |
587 | t = list_entry(rnp->exp_tasks->prev, | |
588 | struct task_struct, rcu_node_entry); | |
589 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { | |
590 | pr_cont(" P%d", t->pid); | |
591 | ndetected++; | |
592 | } | |
593 | return ndetected; | |
594 | } | |
595 | ||
b0e165c0 PM |
596 | /* |
597 | * Check that the list of blocked tasks for the newly completed grace | |
598 | * period is in fact empty. It is a serious bug to complete a grace | |
599 | * period that still has RCU readers blocked! This function must be | |
600 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock | |
601 | * must be held by the caller. | |
12f5f524 PM |
602 | * |
603 | * Also, if there are blocked tasks on the list, they automatically | |
604 | * block the newly created grace period, so set up ->gp_tasks accordingly. | |
b0e165c0 PM |
605 | */ |
606 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
607 | { | |
27f4d280 | 608 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
96e92021 | 609 | if (rcu_preempt_has_tasks(rnp)) |
12f5f524 | 610 | rnp->gp_tasks = rnp->blkd_tasks.next; |
28ecd580 | 611 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
612 | } |
613 | ||
f41d911f PM |
614 | /* |
615 | * Check for a quiescent state from the current CPU. When a task blocks, | |
616 | * the task is recorded in the corresponding CPU's rcu_node structure, | |
617 | * which is checked elsewhere. | |
618 | * | |
619 | * Caller must disable hard irqs. | |
620 | */ | |
86aea0e6 | 621 | static void rcu_preempt_check_callbacks(void) |
f41d911f PM |
622 | { |
623 | struct task_struct *t = current; | |
624 | ||
625 | if (t->rcu_read_lock_nesting == 0) { | |
284a8c93 | 626 | rcu_preempt_qs(); |
f41d911f PM |
627 | return; |
628 | } | |
10f39bb1 | 629 | if (t->rcu_read_lock_nesting > 0 && |
97c668b8 | 630 | __this_cpu_read(rcu_data_p->core_needs_qs) && |
5b74c458 | 631 | __this_cpu_read(rcu_data_p->cpu_no_qs.b.norm)) |
1d082fd0 | 632 | t->rcu_read_unlock_special.b.need_qs = true; |
f41d911f PM |
633 | } |
634 | ||
a46e0899 PM |
635 | #ifdef CONFIG_RCU_BOOST |
636 | ||
09223371 SL |
637 | static void rcu_preempt_do_callbacks(void) |
638 | { | |
2927a689 | 639 | rcu_do_batch(rcu_state_p, this_cpu_ptr(rcu_data_p)); |
09223371 SL |
640 | } |
641 | ||
a46e0899 PM |
642 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
643 | ||
f41d911f | 644 | /* |
6cc68793 | 645 | * Queue a preemptible-RCU callback for invocation after a grace period. |
f41d911f | 646 | */ |
b6a4ae76 | 647 | void call_rcu(struct rcu_head *head, rcu_callback_t func) |
f41d911f | 648 | { |
e63c887c | 649 | __call_rcu(head, func, rcu_state_p, -1, 0); |
f41d911f PM |
650 | } |
651 | EXPORT_SYMBOL_GPL(call_rcu); | |
652 | ||
6ebb237b PM |
653 | /** |
654 | * synchronize_rcu - wait until a grace period has elapsed. | |
655 | * | |
656 | * Control will return to the caller some time after a full grace | |
657 | * period has elapsed, in other words after all currently executing RCU | |
77d8485a PM |
658 | * read-side critical sections have completed. Note, however, that |
659 | * upon return from synchronize_rcu(), the caller might well be executing | |
660 | * concurrently with new RCU read-side critical sections that began while | |
661 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
662 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
f0a0e6f2 PM |
663 | * |
664 | * See the description of synchronize_sched() for more detailed information | |
665 | * on memory ordering guarantees. | |
6ebb237b PM |
666 | */ |
667 | void synchronize_rcu(void) | |
668 | { | |
f78f5b90 PM |
669 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
670 | lock_is_held(&rcu_lock_map) || | |
671 | lock_is_held(&rcu_sched_lock_map), | |
672 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
6ebb237b PM |
673 | if (!rcu_scheduler_active) |
674 | return; | |
5afff48b | 675 | if (rcu_gp_is_expedited()) |
3705b88d AM |
676 | synchronize_rcu_expedited(); |
677 | else | |
678 | wait_rcu_gp(call_rcu); | |
6ebb237b PM |
679 | } |
680 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
681 | ||
e74f4c45 PM |
682 | /** |
683 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
f0a0e6f2 PM |
684 | * |
685 | * Note that this primitive does not necessarily wait for an RCU grace period | |
686 | * to complete. For example, if there are no RCU callbacks queued anywhere | |
687 | * in the system, then rcu_barrier() is within its rights to return | |
688 | * immediately, without waiting for anything, much less an RCU grace period. | |
e74f4c45 PM |
689 | */ |
690 | void rcu_barrier(void) | |
691 | { | |
e63c887c | 692 | _rcu_barrier(rcu_state_p); |
e74f4c45 PM |
693 | } |
694 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
695 | ||
1eba8f84 | 696 | /* |
6cc68793 | 697 | * Initialize preemptible RCU's state structures. |
1eba8f84 PM |
698 | */ |
699 | static void __init __rcu_init_preempt(void) | |
700 | { | |
a87f203e | 701 | rcu_init_one(rcu_state_p); |
1eba8f84 PM |
702 | } |
703 | ||
2439b696 PM |
704 | /* |
705 | * Check for a task exiting while in a preemptible-RCU read-side | |
706 | * critical section, clean up if so. No need to issue warnings, | |
707 | * as debug_check_no_locks_held() already does this if lockdep | |
708 | * is enabled. | |
709 | */ | |
710 | void exit_rcu(void) | |
711 | { | |
712 | struct task_struct *t = current; | |
713 | ||
714 | if (likely(list_empty(¤t->rcu_node_entry))) | |
715 | return; | |
716 | t->rcu_read_lock_nesting = 1; | |
717 | barrier(); | |
1d082fd0 | 718 | t->rcu_read_unlock_special.b.blocked = true; |
2439b696 PM |
719 | __rcu_read_unlock(); |
720 | } | |
721 | ||
28f6569a | 722 | #else /* #ifdef CONFIG_PREEMPT_RCU */ |
f41d911f | 723 | |
b28a7c01 | 724 | static struct rcu_state *const rcu_state_p = &rcu_sched_state; |
27f4d280 | 725 | |
f41d911f PM |
726 | /* |
727 | * Tell them what RCU they are running. | |
728 | */ | |
0e0fc1c2 | 729 | static void __init rcu_bootup_announce(void) |
f41d911f | 730 | { |
efc151c3 | 731 | pr_info("Hierarchical RCU implementation.\n"); |
26845c28 | 732 | rcu_bootup_announce_oddness(); |
f41d911f PM |
733 | } |
734 | ||
cba6d0d6 PM |
735 | /* |
736 | * Because preemptible RCU does not exist, we never have to check for | |
737 | * CPUs being in quiescent states. | |
738 | */ | |
38200cf2 | 739 | static void rcu_preempt_note_context_switch(void) |
cba6d0d6 PM |
740 | { |
741 | } | |
742 | ||
fc2219d4 | 743 | /* |
6cc68793 | 744 | * Because preemptible RCU does not exist, there are never any preempted |
fc2219d4 PM |
745 | * RCU readers. |
746 | */ | |
27f4d280 | 747 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 PM |
748 | { |
749 | return 0; | |
750 | } | |
751 | ||
8af3a5e7 PM |
752 | /* |
753 | * Because there is no preemptible RCU, there can be no readers blocked. | |
754 | */ | |
755 | static bool rcu_preempt_has_tasks(struct rcu_node *rnp) | |
b668c9cf | 756 | { |
8af3a5e7 | 757 | return false; |
b668c9cf PM |
758 | } |
759 | ||
1ed509a2 | 760 | /* |
6cc68793 | 761 | * Because preemptible RCU does not exist, we never have to check for |
1ed509a2 PM |
762 | * tasks blocked within RCU read-side critical sections. |
763 | */ | |
764 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
765 | { | |
766 | } | |
767 | ||
f41d911f | 768 | /* |
6cc68793 | 769 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
770 | * tasks blocked within RCU read-side critical sections. |
771 | */ | |
9bc8b558 | 772 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 773 | { |
9bc8b558 | 774 | return 0; |
f41d911f PM |
775 | } |
776 | ||
74611ecb PM |
777 | /* |
778 | * Because preemptible RCU does not exist, we never have to check for | |
779 | * tasks blocked within RCU read-side critical sections that are | |
780 | * blocking the current expedited grace period. | |
781 | */ | |
782 | static int rcu_print_task_exp_stall(struct rcu_node *rnp) | |
783 | { | |
784 | return 0; | |
785 | } | |
786 | ||
b0e165c0 | 787 | /* |
6cc68793 | 788 | * Because there is no preemptible RCU, there can be no readers blocked, |
49e29126 PM |
789 | * so there is no need to check for blocked tasks. So check only for |
790 | * bogus qsmask values. | |
b0e165c0 PM |
791 | */ |
792 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
793 | { | |
49e29126 | 794 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
795 | } |
796 | ||
f41d911f | 797 | /* |
6cc68793 | 798 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
799 | * to check. |
800 | */ | |
86aea0e6 | 801 | static void rcu_preempt_check_callbacks(void) |
f41d911f PM |
802 | { |
803 | } | |
804 | ||
e74f4c45 | 805 | /* |
6cc68793 | 806 | * Because preemptible RCU does not exist, rcu_barrier() is just |
e74f4c45 PM |
807 | * another name for rcu_barrier_sched(). |
808 | */ | |
809 | void rcu_barrier(void) | |
810 | { | |
811 | rcu_barrier_sched(); | |
812 | } | |
813 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
814 | ||
1eba8f84 | 815 | /* |
6cc68793 | 816 | * Because preemptible RCU does not exist, it need not be initialized. |
1eba8f84 PM |
817 | */ |
818 | static void __init __rcu_init_preempt(void) | |
819 | { | |
820 | } | |
821 | ||
2439b696 PM |
822 | /* |
823 | * Because preemptible RCU does not exist, tasks cannot possibly exit | |
824 | * while in preemptible RCU read-side critical sections. | |
825 | */ | |
826 | void exit_rcu(void) | |
827 | { | |
828 | } | |
829 | ||
28f6569a | 830 | #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ |
8bd93a2c | 831 | |
27f4d280 PM |
832 | #ifdef CONFIG_RCU_BOOST |
833 | ||
1696a8be | 834 | #include "../locking/rtmutex_common.h" |
27f4d280 | 835 | |
0ea1f2eb PM |
836 | #ifdef CONFIG_RCU_TRACE |
837 | ||
838 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
839 | { | |
96e92021 | 840 | if (!rcu_preempt_has_tasks(rnp)) |
0ea1f2eb PM |
841 | rnp->n_balk_blkd_tasks++; |
842 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) | |
843 | rnp->n_balk_exp_gp_tasks++; | |
844 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) | |
845 | rnp->n_balk_boost_tasks++; | |
846 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) | |
847 | rnp->n_balk_notblocked++; | |
848 | else if (rnp->gp_tasks != NULL && | |
a9f4793d | 849 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
0ea1f2eb PM |
850 | rnp->n_balk_notyet++; |
851 | else | |
852 | rnp->n_balk_nos++; | |
853 | } | |
854 | ||
855 | #else /* #ifdef CONFIG_RCU_TRACE */ | |
856 | ||
857 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
858 | { | |
859 | } | |
860 | ||
861 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ | |
862 | ||
5d01bbd1 TG |
863 | static void rcu_wake_cond(struct task_struct *t, int status) |
864 | { | |
865 | /* | |
866 | * If the thread is yielding, only wake it when this | |
867 | * is invoked from idle | |
868 | */ | |
869 | if (status != RCU_KTHREAD_YIELDING || is_idle_task(current)) | |
870 | wake_up_process(t); | |
871 | } | |
872 | ||
27f4d280 PM |
873 | /* |
874 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks | |
875 | * or ->boost_tasks, advancing the pointer to the next task in the | |
876 | * ->blkd_tasks list. | |
877 | * | |
878 | * Note that irqs must be enabled: boosting the task can block. | |
879 | * Returns 1 if there are more tasks needing to be boosted. | |
880 | */ | |
881 | static int rcu_boost(struct rcu_node *rnp) | |
882 | { | |
883 | unsigned long flags; | |
27f4d280 PM |
884 | struct task_struct *t; |
885 | struct list_head *tb; | |
886 | ||
7d0ae808 PM |
887 | if (READ_ONCE(rnp->exp_tasks) == NULL && |
888 | READ_ONCE(rnp->boost_tasks) == NULL) | |
27f4d280 PM |
889 | return 0; /* Nothing left to boost. */ |
890 | ||
2a67e741 | 891 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
27f4d280 PM |
892 | |
893 | /* | |
894 | * Recheck under the lock: all tasks in need of boosting | |
895 | * might exit their RCU read-side critical sections on their own. | |
896 | */ | |
897 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { | |
67c583a7 | 898 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 PM |
899 | return 0; |
900 | } | |
901 | ||
902 | /* | |
903 | * Preferentially boost tasks blocking expedited grace periods. | |
904 | * This cannot starve the normal grace periods because a second | |
905 | * expedited grace period must boost all blocked tasks, including | |
906 | * those blocking the pre-existing normal grace period. | |
907 | */ | |
0ea1f2eb | 908 | if (rnp->exp_tasks != NULL) { |
27f4d280 | 909 | tb = rnp->exp_tasks; |
0ea1f2eb PM |
910 | rnp->n_exp_boosts++; |
911 | } else { | |
27f4d280 | 912 | tb = rnp->boost_tasks; |
0ea1f2eb PM |
913 | rnp->n_normal_boosts++; |
914 | } | |
915 | rnp->n_tasks_boosted++; | |
27f4d280 PM |
916 | |
917 | /* | |
918 | * We boost task t by manufacturing an rt_mutex that appears to | |
919 | * be held by task t. We leave a pointer to that rt_mutex where | |
920 | * task t can find it, and task t will release the mutex when it | |
921 | * exits its outermost RCU read-side critical section. Then | |
922 | * simply acquiring this artificial rt_mutex will boost task | |
923 | * t's priority. (Thanks to tglx for suggesting this approach!) | |
924 | * | |
925 | * Note that task t must acquire rnp->lock to remove itself from | |
926 | * the ->blkd_tasks list, which it will do from exit() if from | |
927 | * nowhere else. We therefore are guaranteed that task t will | |
928 | * stay around at least until we drop rnp->lock. Note that | |
929 | * rnp->lock also resolves races between our priority boosting | |
930 | * and task t's exiting its outermost RCU read-side critical | |
931 | * section. | |
932 | */ | |
933 | t = container_of(tb, struct task_struct, rcu_node_entry); | |
abaa93d9 | 934 | rt_mutex_init_proxy_locked(&rnp->boost_mtx, t); |
67c583a7 | 935 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
abaa93d9 PM |
936 | /* Lock only for side effect: boosts task t's priority. */ |
937 | rt_mutex_lock(&rnp->boost_mtx); | |
938 | rt_mutex_unlock(&rnp->boost_mtx); /* Then keep lockdep happy. */ | |
27f4d280 | 939 | |
7d0ae808 PM |
940 | return READ_ONCE(rnp->exp_tasks) != NULL || |
941 | READ_ONCE(rnp->boost_tasks) != NULL; | |
27f4d280 PM |
942 | } |
943 | ||
27f4d280 | 944 | /* |
bc17ea10 | 945 | * Priority-boosting kthread, one per leaf rcu_node. |
27f4d280 PM |
946 | */ |
947 | static int rcu_boost_kthread(void *arg) | |
948 | { | |
949 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
950 | int spincnt = 0; | |
951 | int more2boost; | |
952 | ||
f7f7bac9 | 953 | trace_rcu_utilization(TPS("Start boost kthread@init")); |
27f4d280 | 954 | for (;;) { |
d71df90e | 955 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
f7f7bac9 | 956 | trace_rcu_utilization(TPS("End boost kthread@rcu_wait")); |
08bca60a | 957 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
f7f7bac9 | 958 | trace_rcu_utilization(TPS("Start boost kthread@rcu_wait")); |
d71df90e | 959 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
27f4d280 PM |
960 | more2boost = rcu_boost(rnp); |
961 | if (more2boost) | |
962 | spincnt++; | |
963 | else | |
964 | spincnt = 0; | |
965 | if (spincnt > 10) { | |
5d01bbd1 | 966 | rnp->boost_kthread_status = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 967 | trace_rcu_utilization(TPS("End boost kthread@rcu_yield")); |
5d01bbd1 | 968 | schedule_timeout_interruptible(2); |
f7f7bac9 | 969 | trace_rcu_utilization(TPS("Start boost kthread@rcu_yield")); |
27f4d280 PM |
970 | spincnt = 0; |
971 | } | |
972 | } | |
1217ed1b | 973 | /* NOTREACHED */ |
f7f7bac9 | 974 | trace_rcu_utilization(TPS("End boost kthread@notreached")); |
27f4d280 PM |
975 | return 0; |
976 | } | |
977 | ||
978 | /* | |
979 | * Check to see if it is time to start boosting RCU readers that are | |
980 | * blocking the current grace period, and, if so, tell the per-rcu_node | |
981 | * kthread to start boosting them. If there is an expedited grace | |
982 | * period in progress, it is always time to boost. | |
983 | * | |
b065a853 PM |
984 | * The caller must hold rnp->lock, which this function releases. |
985 | * The ->boost_kthread_task is immortal, so we don't need to worry | |
986 | * about it going away. | |
27f4d280 | 987 | */ |
1217ed1b | 988 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 989 | __releases(rnp->lock) |
27f4d280 PM |
990 | { |
991 | struct task_struct *t; | |
992 | ||
0ea1f2eb PM |
993 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
994 | rnp->n_balk_exp_gp_tasks++; | |
67c583a7 | 995 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 | 996 | return; |
0ea1f2eb | 997 | } |
27f4d280 PM |
998 | if (rnp->exp_tasks != NULL || |
999 | (rnp->gp_tasks != NULL && | |
1000 | rnp->boost_tasks == NULL && | |
1001 | rnp->qsmask == 0 && | |
1002 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { | |
1003 | if (rnp->exp_tasks == NULL) | |
1004 | rnp->boost_tasks = rnp->gp_tasks; | |
67c583a7 | 1005 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 | 1006 | t = rnp->boost_kthread_task; |
5d01bbd1 TG |
1007 | if (t) |
1008 | rcu_wake_cond(t, rnp->boost_kthread_status); | |
1217ed1b | 1009 | } else { |
0ea1f2eb | 1010 | rcu_initiate_boost_trace(rnp); |
67c583a7 | 1011 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1217ed1b | 1012 | } |
27f4d280 PM |
1013 | } |
1014 | ||
a46e0899 PM |
1015 | /* |
1016 | * Wake up the per-CPU kthread to invoke RCU callbacks. | |
1017 | */ | |
1018 | static void invoke_rcu_callbacks_kthread(void) | |
1019 | { | |
1020 | unsigned long flags; | |
1021 | ||
1022 | local_irq_save(flags); | |
1023 | __this_cpu_write(rcu_cpu_has_work, 1); | |
1eb52121 | 1024 | if (__this_cpu_read(rcu_cpu_kthread_task) != NULL && |
5d01bbd1 TG |
1025 | current != __this_cpu_read(rcu_cpu_kthread_task)) { |
1026 | rcu_wake_cond(__this_cpu_read(rcu_cpu_kthread_task), | |
1027 | __this_cpu_read(rcu_cpu_kthread_status)); | |
1028 | } | |
a46e0899 PM |
1029 | local_irq_restore(flags); |
1030 | } | |
1031 | ||
dff1672d PM |
1032 | /* |
1033 | * Is the current CPU running the RCU-callbacks kthread? | |
1034 | * Caller must have preemption disabled. | |
1035 | */ | |
1036 | static bool rcu_is_callbacks_kthread(void) | |
1037 | { | |
c9d4b0af | 1038 | return __this_cpu_read(rcu_cpu_kthread_task) == current; |
dff1672d PM |
1039 | } |
1040 | ||
27f4d280 PM |
1041 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) |
1042 | ||
1043 | /* | |
1044 | * Do priority-boost accounting for the start of a new grace period. | |
1045 | */ | |
1046 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) | |
1047 | { | |
1048 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; | |
1049 | } | |
1050 | ||
27f4d280 PM |
1051 | /* |
1052 | * Create an RCU-boost kthread for the specified node if one does not | |
1053 | * already exist. We only create this kthread for preemptible RCU. | |
1054 | * Returns zero if all is well, a negated errno otherwise. | |
1055 | */ | |
49fb4c62 | 1056 | static int rcu_spawn_one_boost_kthread(struct rcu_state *rsp, |
0aa04b05 | 1057 | struct rcu_node *rnp) |
27f4d280 | 1058 | { |
5d01bbd1 | 1059 | int rnp_index = rnp - &rsp->node[0]; |
27f4d280 PM |
1060 | unsigned long flags; |
1061 | struct sched_param sp; | |
1062 | struct task_struct *t; | |
1063 | ||
e63c887c | 1064 | if (rcu_state_p != rsp) |
27f4d280 | 1065 | return 0; |
5d01bbd1 | 1066 | |
0aa04b05 | 1067 | if (!rcu_scheduler_fully_active || rcu_rnp_online_cpus(rnp) == 0) |
5d01bbd1 TG |
1068 | return 0; |
1069 | ||
a46e0899 | 1070 | rsp->boost = 1; |
27f4d280 PM |
1071 | if (rnp->boost_kthread_task != NULL) |
1072 | return 0; | |
1073 | t = kthread_create(rcu_boost_kthread, (void *)rnp, | |
5b61b0ba | 1074 | "rcub/%d", rnp_index); |
27f4d280 PM |
1075 | if (IS_ERR(t)) |
1076 | return PTR_ERR(t); | |
2a67e741 | 1077 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
27f4d280 | 1078 | rnp->boost_kthread_task = t; |
67c583a7 | 1079 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
21871d7e | 1080 | sp.sched_priority = kthread_prio; |
27f4d280 | 1081 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); |
9a432736 | 1082 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
27f4d280 PM |
1083 | return 0; |
1084 | } | |
1085 | ||
f8b7fc6b PM |
1086 | static void rcu_kthread_do_work(void) |
1087 | { | |
c9d4b0af CL |
1088 | rcu_do_batch(&rcu_sched_state, this_cpu_ptr(&rcu_sched_data)); |
1089 | rcu_do_batch(&rcu_bh_state, this_cpu_ptr(&rcu_bh_data)); | |
f8b7fc6b PM |
1090 | rcu_preempt_do_callbacks(); |
1091 | } | |
1092 | ||
62ab7072 | 1093 | static void rcu_cpu_kthread_setup(unsigned int cpu) |
f8b7fc6b | 1094 | { |
f8b7fc6b | 1095 | struct sched_param sp; |
f8b7fc6b | 1096 | |
21871d7e | 1097 | sp.sched_priority = kthread_prio; |
62ab7072 | 1098 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); |
f8b7fc6b PM |
1099 | } |
1100 | ||
62ab7072 | 1101 | static void rcu_cpu_kthread_park(unsigned int cpu) |
f8b7fc6b | 1102 | { |
62ab7072 | 1103 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; |
f8b7fc6b PM |
1104 | } |
1105 | ||
62ab7072 | 1106 | static int rcu_cpu_kthread_should_run(unsigned int cpu) |
f8b7fc6b | 1107 | { |
c9d4b0af | 1108 | return __this_cpu_read(rcu_cpu_has_work); |
f8b7fc6b PM |
1109 | } |
1110 | ||
1111 | /* | |
1112 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the | |
e0f23060 PM |
1113 | * RCU softirq used in flavors and configurations of RCU that do not |
1114 | * support RCU priority boosting. | |
f8b7fc6b | 1115 | */ |
62ab7072 | 1116 | static void rcu_cpu_kthread(unsigned int cpu) |
f8b7fc6b | 1117 | { |
c9d4b0af CL |
1118 | unsigned int *statusp = this_cpu_ptr(&rcu_cpu_kthread_status); |
1119 | char work, *workp = this_cpu_ptr(&rcu_cpu_has_work); | |
62ab7072 | 1120 | int spincnt; |
f8b7fc6b | 1121 | |
62ab7072 | 1122 | for (spincnt = 0; spincnt < 10; spincnt++) { |
f7f7bac9 | 1123 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_wait")); |
f8b7fc6b | 1124 | local_bh_disable(); |
f8b7fc6b | 1125 | *statusp = RCU_KTHREAD_RUNNING; |
62ab7072 PM |
1126 | this_cpu_inc(rcu_cpu_kthread_loops); |
1127 | local_irq_disable(); | |
f8b7fc6b PM |
1128 | work = *workp; |
1129 | *workp = 0; | |
62ab7072 | 1130 | local_irq_enable(); |
f8b7fc6b PM |
1131 | if (work) |
1132 | rcu_kthread_do_work(); | |
1133 | local_bh_enable(); | |
62ab7072 | 1134 | if (*workp == 0) { |
f7f7bac9 | 1135 | trace_rcu_utilization(TPS("End CPU kthread@rcu_wait")); |
62ab7072 PM |
1136 | *statusp = RCU_KTHREAD_WAITING; |
1137 | return; | |
f8b7fc6b PM |
1138 | } |
1139 | } | |
62ab7072 | 1140 | *statusp = RCU_KTHREAD_YIELDING; |
f7f7bac9 | 1141 | trace_rcu_utilization(TPS("Start CPU kthread@rcu_yield")); |
62ab7072 | 1142 | schedule_timeout_interruptible(2); |
f7f7bac9 | 1143 | trace_rcu_utilization(TPS("End CPU kthread@rcu_yield")); |
62ab7072 | 1144 | *statusp = RCU_KTHREAD_WAITING; |
f8b7fc6b PM |
1145 | } |
1146 | ||
1147 | /* | |
1148 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are | |
1149 | * served by the rcu_node in question. The CPU hotplug lock is still | |
1150 | * held, so the value of rnp->qsmaskinit will be stable. | |
1151 | * | |
1152 | * We don't include outgoingcpu in the affinity set, use -1 if there is | |
1153 | * no outgoing CPU. If there are no CPUs left in the affinity set, | |
1154 | * this function allows the kthread to execute on any CPU. | |
1155 | */ | |
5d01bbd1 | 1156 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b | 1157 | { |
5d01bbd1 | 1158 | struct task_struct *t = rnp->boost_kthread_task; |
0aa04b05 | 1159 | unsigned long mask = rcu_rnp_online_cpus(rnp); |
f8b7fc6b PM |
1160 | cpumask_var_t cm; |
1161 | int cpu; | |
f8b7fc6b | 1162 | |
5d01bbd1 | 1163 | if (!t) |
f8b7fc6b | 1164 | return; |
5d01bbd1 | 1165 | if (!zalloc_cpumask_var(&cm, GFP_KERNEL)) |
f8b7fc6b | 1166 | return; |
bc75e999 MR |
1167 | for_each_leaf_node_possible_cpu(rnp, cpu) |
1168 | if ((mask & leaf_node_cpu_bit(rnp, cpu)) && | |
1169 | cpu != outgoingcpu) | |
f8b7fc6b | 1170 | cpumask_set_cpu(cpu, cm); |
5d0b0249 | 1171 | if (cpumask_weight(cm) == 0) |
f8b7fc6b | 1172 | cpumask_setall(cm); |
5d01bbd1 | 1173 | set_cpus_allowed_ptr(t, cm); |
f8b7fc6b PM |
1174 | free_cpumask_var(cm); |
1175 | } | |
1176 | ||
62ab7072 PM |
1177 | static struct smp_hotplug_thread rcu_cpu_thread_spec = { |
1178 | .store = &rcu_cpu_kthread_task, | |
1179 | .thread_should_run = rcu_cpu_kthread_should_run, | |
1180 | .thread_fn = rcu_cpu_kthread, | |
1181 | .thread_comm = "rcuc/%u", | |
1182 | .setup = rcu_cpu_kthread_setup, | |
1183 | .park = rcu_cpu_kthread_park, | |
1184 | }; | |
f8b7fc6b PM |
1185 | |
1186 | /* | |
9386c0b7 | 1187 | * Spawn boost kthreads -- called as soon as the scheduler is running. |
f8b7fc6b | 1188 | */ |
9386c0b7 | 1189 | static void __init rcu_spawn_boost_kthreads(void) |
f8b7fc6b | 1190 | { |
f8b7fc6b | 1191 | struct rcu_node *rnp; |
5d01bbd1 | 1192 | int cpu; |
f8b7fc6b | 1193 | |
62ab7072 | 1194 | for_each_possible_cpu(cpu) |
f8b7fc6b | 1195 | per_cpu(rcu_cpu_has_work, cpu) = 0; |
62ab7072 | 1196 | BUG_ON(smpboot_register_percpu_thread(&rcu_cpu_thread_spec)); |
3e9f5c70 PM |
1197 | rcu_for_each_leaf_node(rcu_state_p, rnp) |
1198 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); | |
f8b7fc6b | 1199 | } |
f8b7fc6b | 1200 | |
49fb4c62 | 1201 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b | 1202 | { |
e534165b | 1203 | struct rcu_data *rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
f8b7fc6b PM |
1204 | struct rcu_node *rnp = rdp->mynode; |
1205 | ||
1206 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ | |
62ab7072 | 1207 | if (rcu_scheduler_fully_active) |
e534165b | 1208 | (void)rcu_spawn_one_boost_kthread(rcu_state_p, rnp); |
f8b7fc6b PM |
1209 | } |
1210 | ||
27f4d280 PM |
1211 | #else /* #ifdef CONFIG_RCU_BOOST */ |
1212 | ||
1217ed1b | 1213 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
615e41c6 | 1214 | __releases(rnp->lock) |
27f4d280 | 1215 | { |
67c583a7 | 1216 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
27f4d280 PM |
1217 | } |
1218 | ||
a46e0899 | 1219 | static void invoke_rcu_callbacks_kthread(void) |
27f4d280 | 1220 | { |
a46e0899 | 1221 | WARN_ON_ONCE(1); |
27f4d280 PM |
1222 | } |
1223 | ||
dff1672d PM |
1224 | static bool rcu_is_callbacks_kthread(void) |
1225 | { | |
1226 | return false; | |
1227 | } | |
1228 | ||
27f4d280 PM |
1229 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
1230 | { | |
1231 | } | |
1232 | ||
5d01bbd1 | 1233 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) |
f8b7fc6b PM |
1234 | { |
1235 | } | |
1236 | ||
9386c0b7 | 1237 | static void __init rcu_spawn_boost_kthreads(void) |
b0d30417 | 1238 | { |
b0d30417 | 1239 | } |
b0d30417 | 1240 | |
49fb4c62 | 1241 | static void rcu_prepare_kthreads(int cpu) |
f8b7fc6b PM |
1242 | { |
1243 | } | |
1244 | ||
27f4d280 PM |
1245 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
1246 | ||
8bd93a2c PM |
1247 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
1248 | ||
1249 | /* | |
1250 | * Check to see if any future RCU-related work will need to be done | |
1251 | * by the current CPU, even if none need be done immediately, returning | |
1252 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
1253 | * an exported member of the RCU API. | |
1254 | * | |
7cb92499 PM |
1255 | * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs |
1256 | * any flavor of RCU. | |
8bd93a2c | 1257 | */ |
c1ad348b | 1258 | int rcu_needs_cpu(u64 basemono, u64 *nextevt) |
8bd93a2c | 1259 | { |
c1ad348b | 1260 | *nextevt = KTIME_MAX; |
3382adbc PM |
1261 | return IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) |
1262 | ? 0 : rcu_cpu_has_callbacks(NULL); | |
7cb92499 PM |
1263 | } |
1264 | ||
1265 | /* | |
1266 | * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up | |
1267 | * after it. | |
1268 | */ | |
8fa7845d | 1269 | static void rcu_cleanup_after_idle(void) |
7cb92499 PM |
1270 | { |
1271 | } | |
1272 | ||
aea1b35e | 1273 | /* |
a858af28 | 1274 | * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=n, |
aea1b35e PM |
1275 | * is nothing. |
1276 | */ | |
198bbf81 | 1277 | static void rcu_prepare_for_idle(void) |
aea1b35e PM |
1278 | { |
1279 | } | |
1280 | ||
c57afe80 PM |
1281 | /* |
1282 | * Don't bother keeping a running count of the number of RCU callbacks | |
1283 | * posted because CONFIG_RCU_FAST_NO_HZ=n. | |
1284 | */ | |
1285 | static void rcu_idle_count_callbacks_posted(void) | |
1286 | { | |
1287 | } | |
1288 | ||
8bd93a2c PM |
1289 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
1290 | ||
f23f7fa1 PM |
1291 | /* |
1292 | * This code is invoked when a CPU goes idle, at which point we want | |
1293 | * to have the CPU do everything required for RCU so that it can enter | |
1294 | * the energy-efficient dyntick-idle mode. This is handled by a | |
1295 | * state machine implemented by rcu_prepare_for_idle() below. | |
1296 | * | |
1297 | * The following three proprocessor symbols control this state machine: | |
1298 | * | |
f23f7fa1 PM |
1299 | * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted |
1300 | * to sleep in dyntick-idle mode with RCU callbacks pending. This | |
1301 | * is sized to be roughly one RCU grace period. Those energy-efficiency | |
1302 | * benchmarkers who might otherwise be tempted to set this to a large | |
1303 | * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your | |
1304 | * system. And if you are -that- concerned about energy efficiency, | |
1305 | * just power the system down and be done with it! | |
778d250a PM |
1306 | * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is |
1307 | * permitted to sleep in dyntick-idle mode with only lazy RCU | |
1308 | * callbacks pending. Setting this too high can OOM your system. | |
f23f7fa1 PM |
1309 | * |
1310 | * The values below work well in practice. If future workloads require | |
1311 | * adjustment, they can be converted into kernel config parameters, though | |
1312 | * making the state machine smarter might be a better option. | |
1313 | */ | |
e84c48ae | 1314 | #define RCU_IDLE_GP_DELAY 4 /* Roughly one grace period. */ |
778d250a | 1315 | #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */ |
f23f7fa1 | 1316 | |
5e44ce35 PM |
1317 | static int rcu_idle_gp_delay = RCU_IDLE_GP_DELAY; |
1318 | module_param(rcu_idle_gp_delay, int, 0644); | |
1319 | static int rcu_idle_lazy_gp_delay = RCU_IDLE_LAZY_GP_DELAY; | |
1320 | module_param(rcu_idle_lazy_gp_delay, int, 0644); | |
486e2593 | 1321 | |
486e2593 | 1322 | /* |
c229828c PM |
1323 | * Try to advance callbacks for all flavors of RCU on the current CPU, but |
1324 | * only if it has been awhile since the last time we did so. Afterwards, | |
1325 | * if there are any callbacks ready for immediate invocation, return true. | |
486e2593 | 1326 | */ |
f1f399d1 | 1327 | static bool __maybe_unused rcu_try_advance_all_cbs(void) |
486e2593 | 1328 | { |
c0f4dfd4 PM |
1329 | bool cbs_ready = false; |
1330 | struct rcu_data *rdp; | |
c229828c | 1331 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1332 | struct rcu_node *rnp; |
1333 | struct rcu_state *rsp; | |
486e2593 | 1334 | |
c229828c PM |
1335 | /* Exit early if we advanced recently. */ |
1336 | if (jiffies == rdtp->last_advance_all) | |
d0bc90fd | 1337 | return false; |
c229828c PM |
1338 | rdtp->last_advance_all = jiffies; |
1339 | ||
c0f4dfd4 PM |
1340 | for_each_rcu_flavor(rsp) { |
1341 | rdp = this_cpu_ptr(rsp->rda); | |
1342 | rnp = rdp->mynode; | |
486e2593 | 1343 | |
c0f4dfd4 PM |
1344 | /* |
1345 | * Don't bother checking unless a grace period has | |
1346 | * completed since we last checked and there are | |
1347 | * callbacks not yet ready to invoke. | |
1348 | */ | |
e3663b10 | 1349 | if ((rdp->completed != rnp->completed || |
7d0ae808 | 1350 | unlikely(READ_ONCE(rdp->gpwrap))) && |
c0f4dfd4 | 1351 | rdp->nxttail[RCU_DONE_TAIL] != rdp->nxttail[RCU_NEXT_TAIL]) |
470716fc | 1352 | note_gp_changes(rsp, rdp); |
486e2593 | 1353 | |
c0f4dfd4 PM |
1354 | if (cpu_has_callbacks_ready_to_invoke(rdp)) |
1355 | cbs_ready = true; | |
1356 | } | |
1357 | return cbs_ready; | |
486e2593 PM |
1358 | } |
1359 | ||
aa9b1630 | 1360 | /* |
c0f4dfd4 PM |
1361 | * Allow the CPU to enter dyntick-idle mode unless it has callbacks ready |
1362 | * to invoke. If the CPU has callbacks, try to advance them. Tell the | |
1363 | * caller to set the timeout based on whether or not there are non-lazy | |
1364 | * callbacks. | |
aa9b1630 | 1365 | * |
c0f4dfd4 | 1366 | * The caller must have disabled interrupts. |
aa9b1630 | 1367 | */ |
c1ad348b | 1368 | int rcu_needs_cpu(u64 basemono, u64 *nextevt) |
aa9b1630 | 1369 | { |
aa6da514 | 1370 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c1ad348b | 1371 | unsigned long dj; |
aa9b1630 | 1372 | |
3382adbc | 1373 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL)) { |
43224b96 | 1374 | *nextevt = KTIME_MAX; |
3382adbc PM |
1375 | return 0; |
1376 | } | |
1377 | ||
c0f4dfd4 PM |
1378 | /* Snapshot to detect later posting of non-lazy callback. */ |
1379 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
1380 | ||
aa9b1630 | 1381 | /* If no callbacks, RCU doesn't need the CPU. */ |
aa6da514 | 1382 | if (!rcu_cpu_has_callbacks(&rdtp->all_lazy)) { |
c1ad348b | 1383 | *nextevt = KTIME_MAX; |
aa9b1630 PM |
1384 | return 0; |
1385 | } | |
c0f4dfd4 PM |
1386 | |
1387 | /* Attempt to advance callbacks. */ | |
1388 | if (rcu_try_advance_all_cbs()) { | |
1389 | /* Some ready to invoke, so initiate later invocation. */ | |
1390 | invoke_rcu_core(); | |
aa9b1630 PM |
1391 | return 1; |
1392 | } | |
c0f4dfd4 PM |
1393 | rdtp->last_accelerate = jiffies; |
1394 | ||
1395 | /* Request timer delay depending on laziness, and round. */ | |
6faf7283 | 1396 | if (!rdtp->all_lazy) { |
c1ad348b | 1397 | dj = round_up(rcu_idle_gp_delay + jiffies, |
c0f4dfd4 | 1398 | rcu_idle_gp_delay) - jiffies; |
e84c48ae | 1399 | } else { |
c1ad348b | 1400 | dj = round_jiffies(rcu_idle_lazy_gp_delay + jiffies) - jiffies; |
e84c48ae | 1401 | } |
c1ad348b | 1402 | *nextevt = basemono + dj * TICK_NSEC; |
aa9b1630 PM |
1403 | return 0; |
1404 | } | |
1405 | ||
21e52e15 | 1406 | /* |
c0f4dfd4 PM |
1407 | * Prepare a CPU for idle from an RCU perspective. The first major task |
1408 | * is to sense whether nohz mode has been enabled or disabled via sysfs. | |
1409 | * The second major task is to check to see if a non-lazy callback has | |
1410 | * arrived at a CPU that previously had only lazy callbacks. The third | |
1411 | * major task is to accelerate (that is, assign grace-period numbers to) | |
1412 | * any recently arrived callbacks. | |
aea1b35e PM |
1413 | * |
1414 | * The caller must have disabled interrupts. | |
8bd93a2c | 1415 | */ |
198bbf81 | 1416 | static void rcu_prepare_for_idle(void) |
8bd93a2c | 1417 | { |
48a7639c | 1418 | bool needwake; |
c0f4dfd4 | 1419 | struct rcu_data *rdp; |
198bbf81 | 1420 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
c0f4dfd4 PM |
1421 | struct rcu_node *rnp; |
1422 | struct rcu_state *rsp; | |
9d2ad243 PM |
1423 | int tne; |
1424 | ||
f0f2e7d3 PM |
1425 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) || |
1426 | rcu_is_nocb_cpu(smp_processor_id())) | |
3382adbc PM |
1427 | return; |
1428 | ||
9d2ad243 | 1429 | /* Handle nohz enablement switches conservatively. */ |
7d0ae808 | 1430 | tne = READ_ONCE(tick_nohz_active); |
9d2ad243 | 1431 | if (tne != rdtp->tick_nohz_enabled_snap) { |
aa6da514 | 1432 | if (rcu_cpu_has_callbacks(NULL)) |
9d2ad243 PM |
1433 | invoke_rcu_core(); /* force nohz to see update. */ |
1434 | rdtp->tick_nohz_enabled_snap = tne; | |
1435 | return; | |
1436 | } | |
1437 | if (!tne) | |
1438 | return; | |
f511fc62 | 1439 | |
c57afe80 | 1440 | /* |
c0f4dfd4 PM |
1441 | * If a non-lazy callback arrived at a CPU having only lazy |
1442 | * callbacks, invoke RCU core for the side-effect of recalculating | |
1443 | * idle duration on re-entry to idle. | |
c57afe80 | 1444 | */ |
c0f4dfd4 PM |
1445 | if (rdtp->all_lazy && |
1446 | rdtp->nonlazy_posted != rdtp->nonlazy_posted_snap) { | |
c337f8f5 PM |
1447 | rdtp->all_lazy = false; |
1448 | rdtp->nonlazy_posted_snap = rdtp->nonlazy_posted; | |
c0f4dfd4 | 1449 | invoke_rcu_core(); |
c57afe80 PM |
1450 | return; |
1451 | } | |
c57afe80 | 1452 | |
3084f2f8 | 1453 | /* |
c0f4dfd4 PM |
1454 | * If we have not yet accelerated this jiffy, accelerate all |
1455 | * callbacks on this CPU. | |
3084f2f8 | 1456 | */ |
c0f4dfd4 | 1457 | if (rdtp->last_accelerate == jiffies) |
aea1b35e | 1458 | return; |
c0f4dfd4 PM |
1459 | rdtp->last_accelerate = jiffies; |
1460 | for_each_rcu_flavor(rsp) { | |
198bbf81 | 1461 | rdp = this_cpu_ptr(rsp->rda); |
c0f4dfd4 PM |
1462 | if (!*rdp->nxttail[RCU_DONE_TAIL]) |
1463 | continue; | |
1464 | rnp = rdp->mynode; | |
2a67e741 | 1465 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled. */ |
48a7639c | 1466 | needwake = rcu_accelerate_cbs(rsp, rnp, rdp); |
67c583a7 | 1467 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled. */ |
48a7639c PM |
1468 | if (needwake) |
1469 | rcu_gp_kthread_wake(rsp); | |
77e38ed3 | 1470 | } |
c0f4dfd4 | 1471 | } |
3084f2f8 | 1472 | |
c0f4dfd4 PM |
1473 | /* |
1474 | * Clean up for exit from idle. Attempt to advance callbacks based on | |
1475 | * any grace periods that elapsed while the CPU was idle, and if any | |
1476 | * callbacks are now ready to invoke, initiate invocation. | |
1477 | */ | |
8fa7845d | 1478 | static void rcu_cleanup_after_idle(void) |
c0f4dfd4 | 1479 | { |
3382adbc PM |
1480 | if (IS_ENABLED(CONFIG_RCU_NOCB_CPU_ALL) || |
1481 | rcu_is_nocb_cpu(smp_processor_id())) | |
aea1b35e | 1482 | return; |
7a497c96 PM |
1483 | if (rcu_try_advance_all_cbs()) |
1484 | invoke_rcu_core(); | |
8bd93a2c PM |
1485 | } |
1486 | ||
c57afe80 | 1487 | /* |
98248a0e PM |
1488 | * Keep a running count of the number of non-lazy callbacks posted |
1489 | * on this CPU. This running counter (which is never decremented) allows | |
1490 | * rcu_prepare_for_idle() to detect when something out of the idle loop | |
1491 | * posts a callback, even if an equal number of callbacks are invoked. | |
1492 | * Of course, callbacks should only be posted from within a trace event | |
1493 | * designed to be called from idle or from within RCU_NONIDLE(). | |
c57afe80 PM |
1494 | */ |
1495 | static void rcu_idle_count_callbacks_posted(void) | |
1496 | { | |
5955f7ee | 1497 | __this_cpu_add(rcu_dynticks.nonlazy_posted, 1); |
c57afe80 PM |
1498 | } |
1499 | ||
b626c1b6 PM |
1500 | /* |
1501 | * Data for flushing lazy RCU callbacks at OOM time. | |
1502 | */ | |
1503 | static atomic_t oom_callback_count; | |
1504 | static DECLARE_WAIT_QUEUE_HEAD(oom_callback_wq); | |
1505 | ||
1506 | /* | |
1507 | * RCU OOM callback -- decrement the outstanding count and deliver the | |
1508 | * wake-up if we are the last one. | |
1509 | */ | |
1510 | static void rcu_oom_callback(struct rcu_head *rhp) | |
1511 | { | |
1512 | if (atomic_dec_and_test(&oom_callback_count)) | |
1513 | wake_up(&oom_callback_wq); | |
1514 | } | |
1515 | ||
1516 | /* | |
1517 | * Post an rcu_oom_notify callback on the current CPU if it has at | |
1518 | * least one lazy callback. This will unnecessarily post callbacks | |
1519 | * to CPUs that already have a non-lazy callback at the end of their | |
1520 | * callback list, but this is an infrequent operation, so accept some | |
1521 | * extra overhead to keep things simple. | |
1522 | */ | |
1523 | static void rcu_oom_notify_cpu(void *unused) | |
1524 | { | |
1525 | struct rcu_state *rsp; | |
1526 | struct rcu_data *rdp; | |
1527 | ||
1528 | for_each_rcu_flavor(rsp) { | |
fa07a58f | 1529 | rdp = raw_cpu_ptr(rsp->rda); |
b626c1b6 PM |
1530 | if (rdp->qlen_lazy != 0) { |
1531 | atomic_inc(&oom_callback_count); | |
1532 | rsp->call(&rdp->oom_head, rcu_oom_callback); | |
1533 | } | |
1534 | } | |
1535 | } | |
1536 | ||
1537 | /* | |
1538 | * If low on memory, ensure that each CPU has a non-lazy callback. | |
1539 | * This will wake up CPUs that have only lazy callbacks, in turn | |
1540 | * ensuring that they free up the corresponding memory in a timely manner. | |
1541 | * Because an uncertain amount of memory will be freed in some uncertain | |
1542 | * timeframe, we do not claim to have freed anything. | |
1543 | */ | |
1544 | static int rcu_oom_notify(struct notifier_block *self, | |
1545 | unsigned long notused, void *nfreed) | |
1546 | { | |
1547 | int cpu; | |
1548 | ||
1549 | /* Wait for callbacks from earlier instance to complete. */ | |
1550 | wait_event(oom_callback_wq, atomic_read(&oom_callback_count) == 0); | |
78e4bc34 | 1551 | smp_mb(); /* Ensure callback reuse happens after callback invocation. */ |
b626c1b6 PM |
1552 | |
1553 | /* | |
1554 | * Prevent premature wakeup: ensure that all increments happen | |
1555 | * before there is a chance of the counter reaching zero. | |
1556 | */ | |
1557 | atomic_set(&oom_callback_count, 1); | |
1558 | ||
b626c1b6 PM |
1559 | for_each_online_cpu(cpu) { |
1560 | smp_call_function_single(cpu, rcu_oom_notify_cpu, NULL, 1); | |
bde6c3aa | 1561 | cond_resched_rcu_qs(); |
b626c1b6 | 1562 | } |
b626c1b6 PM |
1563 | |
1564 | /* Unconditionally decrement: no need to wake ourselves up. */ | |
1565 | atomic_dec(&oom_callback_count); | |
1566 | ||
1567 | return NOTIFY_OK; | |
1568 | } | |
1569 | ||
1570 | static struct notifier_block rcu_oom_nb = { | |
1571 | .notifier_call = rcu_oom_notify | |
1572 | }; | |
1573 | ||
1574 | static int __init rcu_register_oom_notifier(void) | |
1575 | { | |
1576 | register_oom_notifier(&rcu_oom_nb); | |
1577 | return 0; | |
1578 | } | |
1579 | early_initcall(rcu_register_oom_notifier); | |
1580 | ||
8bd93a2c | 1581 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
a858af28 | 1582 | |
a858af28 PM |
1583 | #ifdef CONFIG_RCU_FAST_NO_HZ |
1584 | ||
1585 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1586 | { | |
5955f7ee | 1587 | struct rcu_dynticks *rdtp = &per_cpu(rcu_dynticks, cpu); |
c0f4dfd4 | 1588 | unsigned long nlpd = rdtp->nonlazy_posted - rdtp->nonlazy_posted_snap; |
a858af28 | 1589 | |
c0f4dfd4 PM |
1590 | sprintf(cp, "last_accelerate: %04lx/%04lx, nonlazy_posted: %ld, %c%c", |
1591 | rdtp->last_accelerate & 0xffff, jiffies & 0xffff, | |
1592 | ulong2long(nlpd), | |
1593 | rdtp->all_lazy ? 'L' : '.', | |
1594 | rdtp->tick_nohz_enabled_snap ? '.' : 'D'); | |
a858af28 PM |
1595 | } |
1596 | ||
1597 | #else /* #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1598 | ||
1599 | static void print_cpu_stall_fast_no_hz(char *cp, int cpu) | |
1600 | { | |
1c17e4d4 | 1601 | *cp = '\0'; |
a858af28 PM |
1602 | } |
1603 | ||
1604 | #endif /* #else #ifdef CONFIG_RCU_FAST_NO_HZ */ | |
1605 | ||
1606 | /* Initiate the stall-info list. */ | |
1607 | static void print_cpu_stall_info_begin(void) | |
1608 | { | |
efc151c3 | 1609 | pr_cont("\n"); |
a858af28 PM |
1610 | } |
1611 | ||
1612 | /* | |
1613 | * Print out diagnostic information for the specified stalled CPU. | |
1614 | * | |
1615 | * If the specified CPU is aware of the current RCU grace period | |
1616 | * (flavor specified by rsp), then print the number of scheduling | |
1617 | * clock interrupts the CPU has taken during the time that it has | |
1618 | * been aware. Otherwise, print the number of RCU grace periods | |
1619 | * that this CPU is ignorant of, for example, "1" if the CPU was | |
1620 | * aware of the previous grace period. | |
1621 | * | |
1622 | * Also print out idle and (if CONFIG_RCU_FAST_NO_HZ) idle-entry info. | |
1623 | */ | |
1624 | static void print_cpu_stall_info(struct rcu_state *rsp, int cpu) | |
1625 | { | |
1626 | char fast_no_hz[72]; | |
1627 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
1628 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
1629 | char *ticks_title; | |
1630 | unsigned long ticks_value; | |
1631 | ||
1632 | if (rsp->gpnum == rdp->gpnum) { | |
1633 | ticks_title = "ticks this GP"; | |
1634 | ticks_value = rdp->ticks_this_gp; | |
1635 | } else { | |
1636 | ticks_title = "GPs behind"; | |
1637 | ticks_value = rsp->gpnum - rdp->gpnum; | |
1638 | } | |
1639 | print_cpu_stall_fast_no_hz(fast_no_hz, cpu); | |
7f21aeef PM |
1640 | pr_err("\t%d-%c%c%c: (%lu %s) idle=%03x/%llx/%d softirq=%u/%u fqs=%ld %s\n", |
1641 | cpu, | |
1642 | "O."[!!cpu_online(cpu)], | |
1643 | "o."[!!(rdp->grpmask & rdp->mynode->qsmaskinit)], | |
1644 | "N."[!!(rdp->grpmask & rdp->mynode->qsmaskinitnext)], | |
1645 | ticks_value, ticks_title, | |
a858af28 PM |
1646 | atomic_read(&rdtp->dynticks) & 0xfff, |
1647 | rdtp->dynticks_nesting, rdtp->dynticks_nmi_nesting, | |
6231069b | 1648 | rdp->softirq_snap, kstat_softirqs_cpu(RCU_SOFTIRQ, cpu), |
7d0ae808 | 1649 | READ_ONCE(rsp->n_force_qs) - rsp->n_force_qs_gpstart, |
a858af28 PM |
1650 | fast_no_hz); |
1651 | } | |
1652 | ||
1653 | /* Terminate the stall-info list. */ | |
1654 | static void print_cpu_stall_info_end(void) | |
1655 | { | |
efc151c3 | 1656 | pr_err("\t"); |
a858af28 PM |
1657 | } |
1658 | ||
1659 | /* Zero ->ticks_this_gp for all flavors of RCU. */ | |
1660 | static void zero_cpu_stall_ticks(struct rcu_data *rdp) | |
1661 | { | |
1662 | rdp->ticks_this_gp = 0; | |
6231069b | 1663 | rdp->softirq_snap = kstat_softirqs_cpu(RCU_SOFTIRQ, smp_processor_id()); |
a858af28 PM |
1664 | } |
1665 | ||
1666 | /* Increment ->ticks_this_gp for all flavors of RCU. */ | |
1667 | static void increment_cpu_stall_ticks(void) | |
1668 | { | |
115f7a7c PM |
1669 | struct rcu_state *rsp; |
1670 | ||
1671 | for_each_rcu_flavor(rsp) | |
fa07a58f | 1672 | raw_cpu_inc(rsp->rda->ticks_this_gp); |
a858af28 PM |
1673 | } |
1674 | ||
3fbfbf7a PM |
1675 | #ifdef CONFIG_RCU_NOCB_CPU |
1676 | ||
1677 | /* | |
1678 | * Offload callback processing from the boot-time-specified set of CPUs | |
1679 | * specified by rcu_nocb_mask. For each CPU in the set, there is a | |
1680 | * kthread created that pulls the callbacks from the corresponding CPU, | |
1681 | * waits for a grace period to elapse, and invokes the callbacks. | |
1682 | * The no-CBs CPUs do a wake_up() on their kthread when they insert | |
1683 | * a callback into any empty list, unless the rcu_nocb_poll boot parameter | |
1684 | * has been specified, in which case each kthread actively polls its | |
1685 | * CPU. (Which isn't so great for energy efficiency, but which does | |
1686 | * reduce RCU's overhead on that CPU.) | |
1687 | * | |
1688 | * This is intended to be used in conjunction with Frederic Weisbecker's | |
1689 | * adaptive-idle work, which would seriously reduce OS jitter on CPUs | |
1690 | * running CPU-bound user-mode computations. | |
1691 | * | |
1692 | * Offloading of callback processing could also in theory be used as | |
1693 | * an energy-efficiency measure because CPUs with no RCU callbacks | |
1694 | * queued are more aggressive about entering dyntick-idle mode. | |
1695 | */ | |
1696 | ||
1697 | ||
1698 | /* Parse the boot-time rcu_nocb_mask CPU list from the kernel parameters. */ | |
1699 | static int __init rcu_nocb_setup(char *str) | |
1700 | { | |
1701 | alloc_bootmem_cpumask_var(&rcu_nocb_mask); | |
1702 | have_rcu_nocb_mask = true; | |
1703 | cpulist_parse(str, rcu_nocb_mask); | |
1704 | return 1; | |
1705 | } | |
1706 | __setup("rcu_nocbs=", rcu_nocb_setup); | |
1707 | ||
1b0048a4 PG |
1708 | static int __init parse_rcu_nocb_poll(char *arg) |
1709 | { | |
1710 | rcu_nocb_poll = 1; | |
1711 | return 0; | |
1712 | } | |
1713 | early_param("rcu_nocb_poll", parse_rcu_nocb_poll); | |
1714 | ||
dae6e64d | 1715 | /* |
0446be48 PM |
1716 | * Wake up any no-CBs CPUs' kthreads that were waiting on the just-ended |
1717 | * grace period. | |
dae6e64d | 1718 | */ |
abedf8e2 | 1719 | static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq) |
dae6e64d | 1720 | { |
abedf8e2 | 1721 | swake_up_all(sq); |
dae6e64d PM |
1722 | } |
1723 | ||
1724 | /* | |
8b425aa8 | 1725 | * Set the root rcu_node structure's ->need_future_gp field |
dae6e64d PM |
1726 | * based on the sum of those of all rcu_node structures. This does |
1727 | * double-count the root rcu_node structure's requests, but this | |
1728 | * is necessary to handle the possibility of a rcu_nocb_kthread() | |
1729 | * having awakened during the time that the rcu_node structures | |
1730 | * were being updated for the end of the previous grace period. | |
34ed6246 | 1731 | */ |
dae6e64d PM |
1732 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
1733 | { | |
8b425aa8 | 1734 | rnp->need_future_gp[(rnp->completed + 1) & 0x1] += nrq; |
dae6e64d PM |
1735 | } |
1736 | ||
abedf8e2 | 1737 | static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp) |
065bb78c DW |
1738 | { |
1739 | return &rnp->nocb_gp_wq[rnp->completed & 0x1]; | |
1740 | } | |
1741 | ||
dae6e64d | 1742 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
34ed6246 | 1743 | { |
abedf8e2 PG |
1744 | init_swait_queue_head(&rnp->nocb_gp_wq[0]); |
1745 | init_swait_queue_head(&rnp->nocb_gp_wq[1]); | |
34ed6246 PM |
1746 | } |
1747 | ||
2f33b512 | 1748 | #ifndef CONFIG_RCU_NOCB_CPU_ALL |
24342c96 | 1749 | /* Is the specified CPU a no-CBs CPU? */ |
d1e43fa5 | 1750 | bool rcu_is_nocb_cpu(int cpu) |
3fbfbf7a PM |
1751 | { |
1752 | if (have_rcu_nocb_mask) | |
1753 | return cpumask_test_cpu(cpu, rcu_nocb_mask); | |
1754 | return false; | |
1755 | } | |
2f33b512 | 1756 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_ALL */ |
3fbfbf7a | 1757 | |
fbce7497 PM |
1758 | /* |
1759 | * Kick the leader kthread for this NOCB group. | |
1760 | */ | |
1761 | static void wake_nocb_leader(struct rcu_data *rdp, bool force) | |
1762 | { | |
1763 | struct rcu_data *rdp_leader = rdp->nocb_leader; | |
1764 | ||
7d0ae808 | 1765 | if (!READ_ONCE(rdp_leader->nocb_kthread)) |
fbce7497 | 1766 | return; |
7d0ae808 | 1767 | if (READ_ONCE(rdp_leader->nocb_leader_sleep) || force) { |
39953dfd | 1768 | /* Prior smp_mb__after_atomic() orders against prior enqueue. */ |
7d0ae808 | 1769 | WRITE_ONCE(rdp_leader->nocb_leader_sleep, false); |
abedf8e2 | 1770 | swake_up(&rdp_leader->nocb_wq); |
fbce7497 PM |
1771 | } |
1772 | } | |
1773 | ||
d7e29933 PM |
1774 | /* |
1775 | * Does the specified CPU need an RCU callback for the specified flavor | |
1776 | * of rcu_barrier()? | |
1777 | */ | |
1778 | static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu) | |
1779 | { | |
1780 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
41050a00 PM |
1781 | unsigned long ret; |
1782 | #ifdef CONFIG_PROVE_RCU | |
d7e29933 | 1783 | struct rcu_head *rhp; |
41050a00 | 1784 | #endif /* #ifdef CONFIG_PROVE_RCU */ |
d7e29933 | 1785 | |
41050a00 PM |
1786 | /* |
1787 | * Check count of all no-CBs callbacks awaiting invocation. | |
1788 | * There needs to be a barrier before this function is called, | |
1789 | * but associated with a prior determination that no more | |
1790 | * callbacks would be posted. In the worst case, the first | |
1791 | * barrier in _rcu_barrier() suffices (but the caller cannot | |
1792 | * necessarily rely on this, not a substitute for the caller | |
1793 | * getting the concurrency design right!). There must also be | |
1794 | * a barrier between the following load an posting of a callback | |
1795 | * (if a callback is in fact needed). This is associated with an | |
1796 | * atomic_inc() in the caller. | |
1797 | */ | |
1798 | ret = atomic_long_read(&rdp->nocb_q_count); | |
d7e29933 | 1799 | |
41050a00 | 1800 | #ifdef CONFIG_PROVE_RCU |
7d0ae808 | 1801 | rhp = READ_ONCE(rdp->nocb_head); |
d7e29933 | 1802 | if (!rhp) |
7d0ae808 | 1803 | rhp = READ_ONCE(rdp->nocb_gp_head); |
d7e29933 | 1804 | if (!rhp) |
7d0ae808 | 1805 | rhp = READ_ONCE(rdp->nocb_follower_head); |
d7e29933 PM |
1806 | |
1807 | /* Having no rcuo kthread but CBs after scheduler starts is bad! */ | |
7d0ae808 | 1808 | if (!READ_ONCE(rdp->nocb_kthread) && rhp && |
59f792d1 | 1809 | rcu_scheduler_fully_active) { |
d7e29933 PM |
1810 | /* RCU callback enqueued before CPU first came online??? */ |
1811 | pr_err("RCU: Never-onlined no-CBs CPU %d has CB %p\n", | |
1812 | cpu, rhp->func); | |
1813 | WARN_ON_ONCE(1); | |
1814 | } | |
41050a00 | 1815 | #endif /* #ifdef CONFIG_PROVE_RCU */ |
d7e29933 | 1816 | |
41050a00 | 1817 | return !!ret; |
d7e29933 PM |
1818 | } |
1819 | ||
3fbfbf7a PM |
1820 | /* |
1821 | * Enqueue the specified string of rcu_head structures onto the specified | |
1822 | * CPU's no-CBs lists. The CPU is specified by rdp, the head of the | |
1823 | * string by rhp, and the tail of the string by rhtp. The non-lazy/lazy | |
1824 | * counts are supplied by rhcount and rhcount_lazy. | |
1825 | * | |
1826 | * If warranted, also wake up the kthread servicing this CPUs queues. | |
1827 | */ | |
1828 | static void __call_rcu_nocb_enqueue(struct rcu_data *rdp, | |
1829 | struct rcu_head *rhp, | |
1830 | struct rcu_head **rhtp, | |
96d3fd0d PM |
1831 | int rhcount, int rhcount_lazy, |
1832 | unsigned long flags) | |
3fbfbf7a PM |
1833 | { |
1834 | int len; | |
1835 | struct rcu_head **old_rhpp; | |
1836 | struct task_struct *t; | |
1837 | ||
1838 | /* Enqueue the callback on the nocb list and update counts. */ | |
41050a00 PM |
1839 | atomic_long_add(rhcount, &rdp->nocb_q_count); |
1840 | /* rcu_barrier() relies on ->nocb_q_count add before xchg. */ | |
3fbfbf7a | 1841 | old_rhpp = xchg(&rdp->nocb_tail, rhtp); |
7d0ae808 | 1842 | WRITE_ONCE(*old_rhpp, rhp); |
3fbfbf7a | 1843 | atomic_long_add(rhcount_lazy, &rdp->nocb_q_count_lazy); |
39953dfd | 1844 | smp_mb__after_atomic(); /* Store *old_rhpp before _wake test. */ |
3fbfbf7a PM |
1845 | |
1846 | /* If we are not being polled and there is a kthread, awaken it ... */ | |
7d0ae808 | 1847 | t = READ_ONCE(rdp->nocb_kthread); |
25e03a74 | 1848 | if (rcu_nocb_poll || !t) { |
9261dd0d PM |
1849 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1850 | TPS("WakeNotPoll")); | |
3fbfbf7a | 1851 | return; |
9261dd0d | 1852 | } |
3fbfbf7a PM |
1853 | len = atomic_long_read(&rdp->nocb_q_count); |
1854 | if (old_rhpp == &rdp->nocb_head) { | |
96d3fd0d | 1855 | if (!irqs_disabled_flags(flags)) { |
fbce7497 PM |
1856 | /* ... if queue was empty ... */ |
1857 | wake_nocb_leader(rdp, false); | |
96d3fd0d PM |
1858 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1859 | TPS("WakeEmpty")); | |
1860 | } else { | |
9fdd3bc9 | 1861 | rdp->nocb_defer_wakeup = RCU_NOGP_WAKE; |
96d3fd0d PM |
1862 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
1863 | TPS("WakeEmptyIsDeferred")); | |
1864 | } | |
3fbfbf7a PM |
1865 | rdp->qlen_last_fqs_check = 0; |
1866 | } else if (len > rdp->qlen_last_fqs_check + qhimark) { | |
fbce7497 | 1867 | /* ... or if many callbacks queued. */ |
9fdd3bc9 PM |
1868 | if (!irqs_disabled_flags(flags)) { |
1869 | wake_nocb_leader(rdp, true); | |
1870 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
1871 | TPS("WakeOvf")); | |
1872 | } else { | |
1873 | rdp->nocb_defer_wakeup = RCU_NOGP_WAKE_FORCE; | |
1874 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
1875 | TPS("WakeOvfIsDeferred")); | |
1876 | } | |
3fbfbf7a | 1877 | rdp->qlen_last_fqs_check = LONG_MAX / 2; |
9261dd0d PM |
1878 | } else { |
1879 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("WakeNot")); | |
3fbfbf7a PM |
1880 | } |
1881 | return; | |
1882 | } | |
1883 | ||
1884 | /* | |
1885 | * This is a helper for __call_rcu(), which invokes this when the normal | |
1886 | * callback queue is inoperable. If this is not a no-CBs CPU, this | |
1887 | * function returns failure back to __call_rcu(), which can complain | |
1888 | * appropriately. | |
1889 | * | |
1890 | * Otherwise, this function queues the callback where the corresponding | |
1891 | * "rcuo" kthread can find it. | |
1892 | */ | |
1893 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, | |
96d3fd0d | 1894 | bool lazy, unsigned long flags) |
3fbfbf7a PM |
1895 | { |
1896 | ||
d1e43fa5 | 1897 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
c271d3a9 | 1898 | return false; |
96d3fd0d | 1899 | __call_rcu_nocb_enqueue(rdp, rhp, &rhp->next, 1, lazy, flags); |
21e7a608 PM |
1900 | if (__is_kfree_rcu_offset((unsigned long)rhp->func)) |
1901 | trace_rcu_kfree_callback(rdp->rsp->name, rhp, | |
1902 | (unsigned long)rhp->func, | |
756cbf6b PM |
1903 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
1904 | -atomic_long_read(&rdp->nocb_q_count)); | |
21e7a608 PM |
1905 | else |
1906 | trace_rcu_callback(rdp->rsp->name, rhp, | |
756cbf6b PM |
1907 | -atomic_long_read(&rdp->nocb_q_count_lazy), |
1908 | -atomic_long_read(&rdp->nocb_q_count)); | |
1772947b PM |
1909 | |
1910 | /* | |
1911 | * If called from an extended quiescent state with interrupts | |
1912 | * disabled, invoke the RCU core in order to allow the idle-entry | |
1913 | * deferred-wakeup check to function. | |
1914 | */ | |
1915 | if (irqs_disabled_flags(flags) && | |
1916 | !rcu_is_watching() && | |
1917 | cpu_online(smp_processor_id())) | |
1918 | invoke_rcu_core(); | |
1919 | ||
c271d3a9 | 1920 | return true; |
3fbfbf7a PM |
1921 | } |
1922 | ||
1923 | /* | |
1924 | * Adopt orphaned callbacks on a no-CBs CPU, or return 0 if this is | |
1925 | * not a no-CBs CPU. | |
1926 | */ | |
1927 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
1928 | struct rcu_data *rdp, |
1929 | unsigned long flags) | |
3fbfbf7a PM |
1930 | { |
1931 | long ql = rsp->qlen; | |
1932 | long qll = rsp->qlen_lazy; | |
1933 | ||
1934 | /* If this is not a no-CBs CPU, tell the caller to do it the old way. */ | |
d1e43fa5 | 1935 | if (!rcu_is_nocb_cpu(smp_processor_id())) |
0a9e1e11 | 1936 | return false; |
3fbfbf7a PM |
1937 | rsp->qlen = 0; |
1938 | rsp->qlen_lazy = 0; | |
1939 | ||
1940 | /* First, enqueue the donelist, if any. This preserves CB ordering. */ | |
1941 | if (rsp->orphan_donelist != NULL) { | |
1942 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_donelist, | |
96d3fd0d | 1943 | rsp->orphan_donetail, ql, qll, flags); |
3fbfbf7a PM |
1944 | ql = qll = 0; |
1945 | rsp->orphan_donelist = NULL; | |
1946 | rsp->orphan_donetail = &rsp->orphan_donelist; | |
1947 | } | |
1948 | if (rsp->orphan_nxtlist != NULL) { | |
1949 | __call_rcu_nocb_enqueue(rdp, rsp->orphan_nxtlist, | |
96d3fd0d | 1950 | rsp->orphan_nxttail, ql, qll, flags); |
3fbfbf7a PM |
1951 | ql = qll = 0; |
1952 | rsp->orphan_nxtlist = NULL; | |
1953 | rsp->orphan_nxttail = &rsp->orphan_nxtlist; | |
1954 | } | |
0a9e1e11 | 1955 | return true; |
3fbfbf7a PM |
1956 | } |
1957 | ||
1958 | /* | |
34ed6246 PM |
1959 | * If necessary, kick off a new grace period, and either way wait |
1960 | * for a subsequent grace period to complete. | |
3fbfbf7a | 1961 | */ |
34ed6246 | 1962 | static void rcu_nocb_wait_gp(struct rcu_data *rdp) |
3fbfbf7a | 1963 | { |
34ed6246 | 1964 | unsigned long c; |
dae6e64d | 1965 | bool d; |
34ed6246 | 1966 | unsigned long flags; |
48a7639c | 1967 | bool needwake; |
34ed6246 PM |
1968 | struct rcu_node *rnp = rdp->mynode; |
1969 | ||
2a67e741 | 1970 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
48a7639c | 1971 | needwake = rcu_start_future_gp(rnp, rdp, &c); |
67c583a7 | 1972 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
48a7639c PM |
1973 | if (needwake) |
1974 | rcu_gp_kthread_wake(rdp->rsp); | |
3fbfbf7a PM |
1975 | |
1976 | /* | |
34ed6246 PM |
1977 | * Wait for the grace period. Do so interruptibly to avoid messing |
1978 | * up the load average. | |
3fbfbf7a | 1979 | */ |
f7f7bac9 | 1980 | trace_rcu_future_gp(rnp, rdp, c, TPS("StartWait")); |
34ed6246 | 1981 | for (;;) { |
abedf8e2 | 1982 | swait_event_interruptible( |
dae6e64d | 1983 | rnp->nocb_gp_wq[c & 0x1], |
7d0ae808 | 1984 | (d = ULONG_CMP_GE(READ_ONCE(rnp->completed), c))); |
dae6e64d | 1985 | if (likely(d)) |
34ed6246 | 1986 | break; |
73a860cd | 1987 | WARN_ON(signal_pending(current)); |
f7f7bac9 | 1988 | trace_rcu_future_gp(rnp, rdp, c, TPS("ResumeWait")); |
34ed6246 | 1989 | } |
f7f7bac9 | 1990 | trace_rcu_future_gp(rnp, rdp, c, TPS("EndWait")); |
34ed6246 | 1991 | smp_mb(); /* Ensure that CB invocation happens after GP end. */ |
3fbfbf7a PM |
1992 | } |
1993 | ||
fbce7497 PM |
1994 | /* |
1995 | * Leaders come here to wait for additional callbacks to show up. | |
1996 | * This function does not return until callbacks appear. | |
1997 | */ | |
1998 | static void nocb_leader_wait(struct rcu_data *my_rdp) | |
1999 | { | |
2000 | bool firsttime = true; | |
2001 | bool gotcbs; | |
2002 | struct rcu_data *rdp; | |
2003 | struct rcu_head **tail; | |
2004 | ||
2005 | wait_again: | |
2006 | ||
2007 | /* Wait for callbacks to appear. */ | |
2008 | if (!rcu_nocb_poll) { | |
2009 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Sleep"); | |
abedf8e2 | 2010 | swait_event_interruptible(my_rdp->nocb_wq, |
7d0ae808 | 2011 | !READ_ONCE(my_rdp->nocb_leader_sleep)); |
fbce7497 PM |
2012 | /* Memory barrier handled by smp_mb() calls below and repoll. */ |
2013 | } else if (firsttime) { | |
2014 | firsttime = false; /* Don't drown trace log with "Poll"! */ | |
2015 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, "Poll"); | |
2016 | } | |
2017 | ||
2018 | /* | |
2019 | * Each pass through the following loop checks a follower for CBs. | |
2020 | * We are our own first follower. Any CBs found are moved to | |
2021 | * nocb_gp_head, where they await a grace period. | |
2022 | */ | |
2023 | gotcbs = false; | |
2024 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
7d0ae808 | 2025 | rdp->nocb_gp_head = READ_ONCE(rdp->nocb_head); |
fbce7497 PM |
2026 | if (!rdp->nocb_gp_head) |
2027 | continue; /* No CBs here, try next follower. */ | |
2028 | ||
2029 | /* Move callbacks to wait-for-GP list, which is empty. */ | |
7d0ae808 | 2030 | WRITE_ONCE(rdp->nocb_head, NULL); |
fbce7497 | 2031 | rdp->nocb_gp_tail = xchg(&rdp->nocb_tail, &rdp->nocb_head); |
fbce7497 PM |
2032 | gotcbs = true; |
2033 | } | |
2034 | ||
2035 | /* | |
2036 | * If there were no callbacks, sleep a bit, rescan after a | |
2037 | * memory barrier, and go retry. | |
2038 | */ | |
2039 | if (unlikely(!gotcbs)) { | |
2040 | if (!rcu_nocb_poll) | |
2041 | trace_rcu_nocb_wake(my_rdp->rsp->name, my_rdp->cpu, | |
2042 | "WokeEmpty"); | |
73a860cd | 2043 | WARN_ON(signal_pending(current)); |
fbce7497 PM |
2044 | schedule_timeout_interruptible(1); |
2045 | ||
2046 | /* Rescan in case we were a victim of memory ordering. */ | |
11ed7f93 PK |
2047 | my_rdp->nocb_leader_sleep = true; |
2048 | smp_mb(); /* Ensure _sleep true before scan. */ | |
fbce7497 | 2049 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) |
7d0ae808 | 2050 | if (READ_ONCE(rdp->nocb_head)) { |
fbce7497 | 2051 | /* Found CB, so short-circuit next wait. */ |
11ed7f93 | 2052 | my_rdp->nocb_leader_sleep = false; |
fbce7497 PM |
2053 | break; |
2054 | } | |
2055 | goto wait_again; | |
2056 | } | |
2057 | ||
2058 | /* Wait for one grace period. */ | |
2059 | rcu_nocb_wait_gp(my_rdp); | |
2060 | ||
2061 | /* | |
11ed7f93 PK |
2062 | * We left ->nocb_leader_sleep unset to reduce cache thrashing. |
2063 | * We set it now, but recheck for new callbacks while | |
fbce7497 PM |
2064 | * traversing our follower list. |
2065 | */ | |
11ed7f93 PK |
2066 | my_rdp->nocb_leader_sleep = true; |
2067 | smp_mb(); /* Ensure _sleep true before scan of ->nocb_head. */ | |
fbce7497 PM |
2068 | |
2069 | /* Each pass through the following loop wakes a follower, if needed. */ | |
2070 | for (rdp = my_rdp; rdp; rdp = rdp->nocb_next_follower) { | |
7d0ae808 | 2071 | if (READ_ONCE(rdp->nocb_head)) |
11ed7f93 | 2072 | my_rdp->nocb_leader_sleep = false;/* No need to sleep.*/ |
fbce7497 PM |
2073 | if (!rdp->nocb_gp_head) |
2074 | continue; /* No CBs, so no need to wake follower. */ | |
2075 | ||
2076 | /* Append callbacks to follower's "done" list. */ | |
2077 | tail = xchg(&rdp->nocb_follower_tail, rdp->nocb_gp_tail); | |
2078 | *tail = rdp->nocb_gp_head; | |
c847f142 | 2079 | smp_mb__after_atomic(); /* Store *tail before wakeup. */ |
fbce7497 PM |
2080 | if (rdp != my_rdp && tail == &rdp->nocb_follower_head) { |
2081 | /* | |
2082 | * List was empty, wake up the follower. | |
2083 | * Memory barriers supplied by atomic_long_add(). | |
2084 | */ | |
abedf8e2 | 2085 | swake_up(&rdp->nocb_wq); |
fbce7497 PM |
2086 | } |
2087 | } | |
2088 | ||
2089 | /* If we (the leader) don't have CBs, go wait some more. */ | |
2090 | if (!my_rdp->nocb_follower_head) | |
2091 | goto wait_again; | |
2092 | } | |
2093 | ||
2094 | /* | |
2095 | * Followers come here to wait for additional callbacks to show up. | |
2096 | * This function does not return until callbacks appear. | |
2097 | */ | |
2098 | static void nocb_follower_wait(struct rcu_data *rdp) | |
2099 | { | |
2100 | bool firsttime = true; | |
2101 | ||
2102 | for (;;) { | |
2103 | if (!rcu_nocb_poll) { | |
2104 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2105 | "FollowerSleep"); | |
abedf8e2 | 2106 | swait_event_interruptible(rdp->nocb_wq, |
7d0ae808 | 2107 | READ_ONCE(rdp->nocb_follower_head)); |
fbce7497 PM |
2108 | } else if (firsttime) { |
2109 | /* Don't drown trace log with "Poll"! */ | |
2110 | firsttime = false; | |
2111 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "Poll"); | |
2112 | } | |
2113 | if (smp_load_acquire(&rdp->nocb_follower_head)) { | |
2114 | /* ^^^ Ensure CB invocation follows _head test. */ | |
2115 | return; | |
2116 | } | |
2117 | if (!rcu_nocb_poll) | |
2118 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, | |
2119 | "WokeEmpty"); | |
73a860cd | 2120 | WARN_ON(signal_pending(current)); |
fbce7497 PM |
2121 | schedule_timeout_interruptible(1); |
2122 | } | |
2123 | } | |
2124 | ||
3fbfbf7a PM |
2125 | /* |
2126 | * Per-rcu_data kthread, but only for no-CBs CPUs. Each kthread invokes | |
fbce7497 PM |
2127 | * callbacks queued by the corresponding no-CBs CPU, however, there is |
2128 | * an optional leader-follower relationship so that the grace-period | |
2129 | * kthreads don't have to do quite so many wakeups. | |
3fbfbf7a PM |
2130 | */ |
2131 | static int rcu_nocb_kthread(void *arg) | |
2132 | { | |
2133 | int c, cl; | |
2134 | struct rcu_head *list; | |
2135 | struct rcu_head *next; | |
2136 | struct rcu_head **tail; | |
2137 | struct rcu_data *rdp = arg; | |
2138 | ||
2139 | /* Each pass through this loop invokes one batch of callbacks */ | |
2140 | for (;;) { | |
fbce7497 PM |
2141 | /* Wait for callbacks. */ |
2142 | if (rdp->nocb_leader == rdp) | |
2143 | nocb_leader_wait(rdp); | |
2144 | else | |
2145 | nocb_follower_wait(rdp); | |
2146 | ||
2147 | /* Pull the ready-to-invoke callbacks onto local list. */ | |
7d0ae808 | 2148 | list = READ_ONCE(rdp->nocb_follower_head); |
fbce7497 PM |
2149 | BUG_ON(!list); |
2150 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, "WokeNonEmpty"); | |
7d0ae808 | 2151 | WRITE_ONCE(rdp->nocb_follower_head, NULL); |
fbce7497 | 2152 | tail = xchg(&rdp->nocb_follower_tail, &rdp->nocb_follower_head); |
3fbfbf7a PM |
2153 | |
2154 | /* Each pass through the following loop invokes a callback. */ | |
41050a00 PM |
2155 | trace_rcu_batch_start(rdp->rsp->name, |
2156 | atomic_long_read(&rdp->nocb_q_count_lazy), | |
2157 | atomic_long_read(&rdp->nocb_q_count), -1); | |
3fbfbf7a PM |
2158 | c = cl = 0; |
2159 | while (list) { | |
2160 | next = list->next; | |
2161 | /* Wait for enqueuing to complete, if needed. */ | |
2162 | while (next == NULL && &list->next != tail) { | |
69a79bb1 PM |
2163 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2164 | TPS("WaitQueue")); | |
3fbfbf7a | 2165 | schedule_timeout_interruptible(1); |
69a79bb1 PM |
2166 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, |
2167 | TPS("WokeQueue")); | |
3fbfbf7a PM |
2168 | next = list->next; |
2169 | } | |
2170 | debug_rcu_head_unqueue(list); | |
2171 | local_bh_disable(); | |
2172 | if (__rcu_reclaim(rdp->rsp->name, list)) | |
2173 | cl++; | |
2174 | c++; | |
2175 | local_bh_enable(); | |
2176 | list = next; | |
2177 | } | |
2178 | trace_rcu_batch_end(rdp->rsp->name, c, !!list, 0, 0, 1); | |
41050a00 PM |
2179 | smp_mb__before_atomic(); /* _add after CB invocation. */ |
2180 | atomic_long_add(-c, &rdp->nocb_q_count); | |
2181 | atomic_long_add(-cl, &rdp->nocb_q_count_lazy); | |
c635a4e1 | 2182 | rdp->n_nocbs_invoked += c; |
3fbfbf7a PM |
2183 | } |
2184 | return 0; | |
2185 | } | |
2186 | ||
96d3fd0d | 2187 | /* Is a deferred wakeup of rcu_nocb_kthread() required? */ |
9fdd3bc9 | 2188 | static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
96d3fd0d | 2189 | { |
7d0ae808 | 2190 | return READ_ONCE(rdp->nocb_defer_wakeup); |
96d3fd0d PM |
2191 | } |
2192 | ||
2193 | /* Do a deferred wakeup of rcu_nocb_kthread(). */ | |
2194 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2195 | { | |
9fdd3bc9 PM |
2196 | int ndw; |
2197 | ||
96d3fd0d PM |
2198 | if (!rcu_nocb_need_deferred_wakeup(rdp)) |
2199 | return; | |
7d0ae808 PM |
2200 | ndw = READ_ONCE(rdp->nocb_defer_wakeup); |
2201 | WRITE_ONCE(rdp->nocb_defer_wakeup, RCU_NOGP_WAKE_NOT); | |
9fdd3bc9 PM |
2202 | wake_nocb_leader(rdp, ndw == RCU_NOGP_WAKE_FORCE); |
2203 | trace_rcu_nocb_wake(rdp->rsp->name, rdp->cpu, TPS("DeferredWake")); | |
96d3fd0d PM |
2204 | } |
2205 | ||
f4579fc5 PM |
2206 | void __init rcu_init_nohz(void) |
2207 | { | |
2208 | int cpu; | |
2209 | bool need_rcu_nocb_mask = true; | |
2210 | struct rcu_state *rsp; | |
2211 | ||
2212 | #ifdef CONFIG_RCU_NOCB_CPU_NONE | |
2213 | need_rcu_nocb_mask = false; | |
2214 | #endif /* #ifndef CONFIG_RCU_NOCB_CPU_NONE */ | |
2215 | ||
2216 | #if defined(CONFIG_NO_HZ_FULL) | |
2217 | if (tick_nohz_full_running && cpumask_weight(tick_nohz_full_mask)) | |
2218 | need_rcu_nocb_mask = true; | |
2219 | #endif /* #if defined(CONFIG_NO_HZ_FULL) */ | |
2220 | ||
2221 | if (!have_rcu_nocb_mask && need_rcu_nocb_mask) { | |
949cccdb PK |
2222 | if (!zalloc_cpumask_var(&rcu_nocb_mask, GFP_KERNEL)) { |
2223 | pr_info("rcu_nocb_mask allocation failed, callback offloading disabled.\n"); | |
2224 | return; | |
2225 | } | |
f4579fc5 PM |
2226 | have_rcu_nocb_mask = true; |
2227 | } | |
2228 | if (!have_rcu_nocb_mask) | |
2229 | return; | |
2230 | ||
2231 | #ifdef CONFIG_RCU_NOCB_CPU_ZERO | |
2232 | pr_info("\tOffload RCU callbacks from CPU 0\n"); | |
2233 | cpumask_set_cpu(0, rcu_nocb_mask); | |
2234 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ZERO */ | |
2235 | #ifdef CONFIG_RCU_NOCB_CPU_ALL | |
2236 | pr_info("\tOffload RCU callbacks from all CPUs\n"); | |
2237 | cpumask_copy(rcu_nocb_mask, cpu_possible_mask); | |
2238 | #endif /* #ifdef CONFIG_RCU_NOCB_CPU_ALL */ | |
2239 | #if defined(CONFIG_NO_HZ_FULL) | |
2240 | if (tick_nohz_full_running) | |
2241 | cpumask_or(rcu_nocb_mask, rcu_nocb_mask, tick_nohz_full_mask); | |
2242 | #endif /* #if defined(CONFIG_NO_HZ_FULL) */ | |
2243 | ||
2244 | if (!cpumask_subset(rcu_nocb_mask, cpu_possible_mask)) { | |
2245 | pr_info("\tNote: kernel parameter 'rcu_nocbs=' contains nonexistent CPUs.\n"); | |
2246 | cpumask_and(rcu_nocb_mask, cpu_possible_mask, | |
2247 | rcu_nocb_mask); | |
2248 | } | |
ad853b48 TH |
2249 | pr_info("\tOffload RCU callbacks from CPUs: %*pbl.\n", |
2250 | cpumask_pr_args(rcu_nocb_mask)); | |
f4579fc5 PM |
2251 | if (rcu_nocb_poll) |
2252 | pr_info("\tPoll for callbacks from no-CBs CPUs.\n"); | |
2253 | ||
2254 | for_each_rcu_flavor(rsp) { | |
34404ca8 PM |
2255 | for_each_cpu(cpu, rcu_nocb_mask) |
2256 | init_nocb_callback_list(per_cpu_ptr(rsp->rda, cpu)); | |
35ce7f29 | 2257 | rcu_organize_nocb_kthreads(rsp); |
f4579fc5 | 2258 | } |
96d3fd0d PM |
2259 | } |
2260 | ||
3fbfbf7a PM |
2261 | /* Initialize per-rcu_data variables for no-CBs CPUs. */ |
2262 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) | |
2263 | { | |
2264 | rdp->nocb_tail = &rdp->nocb_head; | |
abedf8e2 | 2265 | init_swait_queue_head(&rdp->nocb_wq); |
fbce7497 | 2266 | rdp->nocb_follower_tail = &rdp->nocb_follower_head; |
3fbfbf7a PM |
2267 | } |
2268 | ||
35ce7f29 PM |
2269 | /* |
2270 | * If the specified CPU is a no-CBs CPU that does not already have its | |
2271 | * rcuo kthread for the specified RCU flavor, spawn it. If the CPUs are | |
2272 | * brought online out of order, this can require re-organizing the | |
2273 | * leader-follower relationships. | |
2274 | */ | |
2275 | static void rcu_spawn_one_nocb_kthread(struct rcu_state *rsp, int cpu) | |
2276 | { | |
2277 | struct rcu_data *rdp; | |
2278 | struct rcu_data *rdp_last; | |
2279 | struct rcu_data *rdp_old_leader; | |
2280 | struct rcu_data *rdp_spawn = per_cpu_ptr(rsp->rda, cpu); | |
2281 | struct task_struct *t; | |
2282 | ||
2283 | /* | |
2284 | * If this isn't a no-CBs CPU or if it already has an rcuo kthread, | |
2285 | * then nothing to do. | |
2286 | */ | |
2287 | if (!rcu_is_nocb_cpu(cpu) || rdp_spawn->nocb_kthread) | |
2288 | return; | |
2289 | ||
2290 | /* If we didn't spawn the leader first, reorganize! */ | |
2291 | rdp_old_leader = rdp_spawn->nocb_leader; | |
2292 | if (rdp_old_leader != rdp_spawn && !rdp_old_leader->nocb_kthread) { | |
2293 | rdp_last = NULL; | |
2294 | rdp = rdp_old_leader; | |
2295 | do { | |
2296 | rdp->nocb_leader = rdp_spawn; | |
2297 | if (rdp_last && rdp != rdp_spawn) | |
2298 | rdp_last->nocb_next_follower = rdp; | |
bbe5d7a9 PM |
2299 | if (rdp == rdp_spawn) { |
2300 | rdp = rdp->nocb_next_follower; | |
2301 | } else { | |
2302 | rdp_last = rdp; | |
2303 | rdp = rdp->nocb_next_follower; | |
2304 | rdp_last->nocb_next_follower = NULL; | |
2305 | } | |
35ce7f29 PM |
2306 | } while (rdp); |
2307 | rdp_spawn->nocb_next_follower = rdp_old_leader; | |
2308 | } | |
2309 | ||
2310 | /* Spawn the kthread for this CPU and RCU flavor. */ | |
2311 | t = kthread_run(rcu_nocb_kthread, rdp_spawn, | |
2312 | "rcuo%c/%d", rsp->abbr, cpu); | |
2313 | BUG_ON(IS_ERR(t)); | |
7d0ae808 | 2314 | WRITE_ONCE(rdp_spawn->nocb_kthread, t); |
35ce7f29 PM |
2315 | } |
2316 | ||
2317 | /* | |
2318 | * If the specified CPU is a no-CBs CPU that does not already have its | |
2319 | * rcuo kthreads, spawn them. | |
2320 | */ | |
2321 | static void rcu_spawn_all_nocb_kthreads(int cpu) | |
2322 | { | |
2323 | struct rcu_state *rsp; | |
2324 | ||
2325 | if (rcu_scheduler_fully_active) | |
2326 | for_each_rcu_flavor(rsp) | |
2327 | rcu_spawn_one_nocb_kthread(rsp, cpu); | |
2328 | } | |
2329 | ||
2330 | /* | |
2331 | * Once the scheduler is running, spawn rcuo kthreads for all online | |
2332 | * no-CBs CPUs. This assumes that the early_initcall()s happen before | |
2333 | * non-boot CPUs come online -- if this changes, we will need to add | |
2334 | * some mutual exclusion. | |
2335 | */ | |
2336 | static void __init rcu_spawn_nocb_kthreads(void) | |
2337 | { | |
2338 | int cpu; | |
2339 | ||
2340 | for_each_online_cpu(cpu) | |
2341 | rcu_spawn_all_nocb_kthreads(cpu); | |
2342 | } | |
2343 | ||
fbce7497 PM |
2344 | /* How many follower CPU IDs per leader? Default of -1 for sqrt(nr_cpu_ids). */ |
2345 | static int rcu_nocb_leader_stride = -1; | |
2346 | module_param(rcu_nocb_leader_stride, int, 0444); | |
2347 | ||
2348 | /* | |
35ce7f29 | 2349 | * Initialize leader-follower relationships for all no-CBs CPU. |
fbce7497 | 2350 | */ |
35ce7f29 | 2351 | static void __init rcu_organize_nocb_kthreads(struct rcu_state *rsp) |
3fbfbf7a PM |
2352 | { |
2353 | int cpu; | |
fbce7497 PM |
2354 | int ls = rcu_nocb_leader_stride; |
2355 | int nl = 0; /* Next leader. */ | |
3fbfbf7a | 2356 | struct rcu_data *rdp; |
fbce7497 PM |
2357 | struct rcu_data *rdp_leader = NULL; /* Suppress misguided gcc warn. */ |
2358 | struct rcu_data *rdp_prev = NULL; | |
3fbfbf7a | 2359 | |
22c2f669 | 2360 | if (!have_rcu_nocb_mask) |
3fbfbf7a | 2361 | return; |
fbce7497 PM |
2362 | if (ls == -1) { |
2363 | ls = int_sqrt(nr_cpu_ids); | |
2364 | rcu_nocb_leader_stride = ls; | |
2365 | } | |
2366 | ||
2367 | /* | |
2368 | * Each pass through this loop sets up one rcu_data structure and | |
2369 | * spawns one rcu_nocb_kthread(). | |
2370 | */ | |
3fbfbf7a PM |
2371 | for_each_cpu(cpu, rcu_nocb_mask) { |
2372 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
fbce7497 PM |
2373 | if (rdp->cpu >= nl) { |
2374 | /* New leader, set up for followers & next leader. */ | |
2375 | nl = DIV_ROUND_UP(rdp->cpu + 1, ls) * ls; | |
2376 | rdp->nocb_leader = rdp; | |
2377 | rdp_leader = rdp; | |
2378 | } else { | |
2379 | /* Another follower, link to previous leader. */ | |
2380 | rdp->nocb_leader = rdp_leader; | |
2381 | rdp_prev->nocb_next_follower = rdp; | |
2382 | } | |
2383 | rdp_prev = rdp; | |
3fbfbf7a PM |
2384 | } |
2385 | } | |
2386 | ||
2387 | /* Prevent __call_rcu() from enqueuing callbacks on no-CBs CPUs */ | |
34ed6246 | 2388 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2389 | { |
22c2f669 | 2390 | if (!rcu_is_nocb_cpu(rdp->cpu)) |
34ed6246 | 2391 | return false; |
22c2f669 | 2392 | |
34404ca8 PM |
2393 | /* If there are early-boot callbacks, move them to nocb lists. */ |
2394 | if (rdp->nxtlist) { | |
2395 | rdp->nocb_head = rdp->nxtlist; | |
2396 | rdp->nocb_tail = rdp->nxttail[RCU_NEXT_TAIL]; | |
2397 | atomic_long_set(&rdp->nocb_q_count, rdp->qlen); | |
2398 | atomic_long_set(&rdp->nocb_q_count_lazy, rdp->qlen_lazy); | |
2399 | rdp->nxtlist = NULL; | |
2400 | rdp->qlen = 0; | |
2401 | rdp->qlen_lazy = 0; | |
2402 | } | |
3fbfbf7a | 2403 | rdp->nxttail[RCU_NEXT_TAIL] = NULL; |
34ed6246 | 2404 | return true; |
3fbfbf7a PM |
2405 | } |
2406 | ||
34ed6246 PM |
2407 | #else /* #ifdef CONFIG_RCU_NOCB_CPU */ |
2408 | ||
d7e29933 PM |
2409 | static bool rcu_nocb_cpu_needs_barrier(struct rcu_state *rsp, int cpu) |
2410 | { | |
2411 | WARN_ON_ONCE(1); /* Should be dead code. */ | |
2412 | return false; | |
2413 | } | |
2414 | ||
abedf8e2 | 2415 | static void rcu_nocb_gp_cleanup(struct swait_queue_head *sq) |
3fbfbf7a | 2416 | { |
3fbfbf7a PM |
2417 | } |
2418 | ||
dae6e64d PM |
2419 | static void rcu_nocb_gp_set(struct rcu_node *rnp, int nrq) |
2420 | { | |
2421 | } | |
2422 | ||
abedf8e2 | 2423 | static struct swait_queue_head *rcu_nocb_gp_get(struct rcu_node *rnp) |
065bb78c DW |
2424 | { |
2425 | return NULL; | |
2426 | } | |
2427 | ||
dae6e64d PM |
2428 | static void rcu_init_one_nocb(struct rcu_node *rnp) |
2429 | { | |
2430 | } | |
3fbfbf7a | 2431 | |
3fbfbf7a | 2432 | static bool __call_rcu_nocb(struct rcu_data *rdp, struct rcu_head *rhp, |
96d3fd0d | 2433 | bool lazy, unsigned long flags) |
3fbfbf7a | 2434 | { |
4afc7e26 | 2435 | return false; |
3fbfbf7a PM |
2436 | } |
2437 | ||
2438 | static bool __maybe_unused rcu_nocb_adopt_orphan_cbs(struct rcu_state *rsp, | |
96d3fd0d PM |
2439 | struct rcu_data *rdp, |
2440 | unsigned long flags) | |
3fbfbf7a | 2441 | { |
f4aa84ba | 2442 | return false; |
3fbfbf7a PM |
2443 | } |
2444 | ||
3fbfbf7a PM |
2445 | static void __init rcu_boot_init_nocb_percpu_data(struct rcu_data *rdp) |
2446 | { | |
2447 | } | |
2448 | ||
9fdd3bc9 | 2449 | static int rcu_nocb_need_deferred_wakeup(struct rcu_data *rdp) |
96d3fd0d PM |
2450 | { |
2451 | return false; | |
2452 | } | |
2453 | ||
2454 | static void do_nocb_deferred_wakeup(struct rcu_data *rdp) | |
2455 | { | |
2456 | } | |
2457 | ||
35ce7f29 PM |
2458 | static void rcu_spawn_all_nocb_kthreads(int cpu) |
2459 | { | |
2460 | } | |
2461 | ||
2462 | static void __init rcu_spawn_nocb_kthreads(void) | |
3fbfbf7a PM |
2463 | { |
2464 | } | |
2465 | ||
34ed6246 | 2466 | static bool init_nocb_callback_list(struct rcu_data *rdp) |
3fbfbf7a | 2467 | { |
34ed6246 | 2468 | return false; |
3fbfbf7a PM |
2469 | } |
2470 | ||
2471 | #endif /* #else #ifdef CONFIG_RCU_NOCB_CPU */ | |
65d798f0 PM |
2472 | |
2473 | /* | |
2474 | * An adaptive-ticks CPU can potentially execute in kernel mode for an | |
2475 | * arbitrarily long period of time with the scheduling-clock tick turned | |
2476 | * off. RCU will be paying attention to this CPU because it is in the | |
2477 | * kernel, but the CPU cannot be guaranteed to be executing the RCU state | |
2478 | * machine because the scheduling-clock tick has been disabled. Therefore, | |
2479 | * if an adaptive-ticks CPU is failing to respond to the current grace | |
2480 | * period and has not be idle from an RCU perspective, kick it. | |
2481 | */ | |
4a81e832 | 2482 | static void __maybe_unused rcu_kick_nohz_cpu(int cpu) |
65d798f0 PM |
2483 | { |
2484 | #ifdef CONFIG_NO_HZ_FULL | |
2485 | if (tick_nohz_full_cpu(cpu)) | |
2486 | smp_send_reschedule(cpu); | |
2487 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ | |
2488 | } | |
2333210b PM |
2489 | |
2490 | ||
2491 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE | |
2492 | ||
0edd1b17 | 2493 | static int full_sysidle_state; /* Current system-idle state. */ |
d4bd54fb PM |
2494 | #define RCU_SYSIDLE_NOT 0 /* Some CPU is not idle. */ |
2495 | #define RCU_SYSIDLE_SHORT 1 /* All CPUs idle for brief period. */ | |
2496 | #define RCU_SYSIDLE_LONG 2 /* All CPUs idle for long enough. */ | |
2497 | #define RCU_SYSIDLE_FULL 3 /* All CPUs idle, ready for sysidle. */ | |
2498 | #define RCU_SYSIDLE_FULL_NOTED 4 /* Actually entered sysidle state. */ | |
2499 | ||
eb348b89 PM |
2500 | /* |
2501 | * Invoked to note exit from irq or task transition to idle. Note that | |
2502 | * usermode execution does -not- count as idle here! After all, we want | |
2503 | * to detect full-system idle states, not RCU quiescent states and grace | |
2504 | * periods. The caller must have disabled interrupts. | |
2505 | */ | |
28ced795 | 2506 | static void rcu_sysidle_enter(int irq) |
eb348b89 PM |
2507 | { |
2508 | unsigned long j; | |
28ced795 | 2509 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
eb348b89 | 2510 | |
663e1310 PM |
2511 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2512 | if (!tick_nohz_full_enabled()) | |
2513 | return; | |
2514 | ||
eb348b89 PM |
2515 | /* Adjust nesting, check for fully idle. */ |
2516 | if (irq) { | |
2517 | rdtp->dynticks_idle_nesting--; | |
2518 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2519 | if (rdtp->dynticks_idle_nesting != 0) | |
2520 | return; /* Still not fully idle. */ | |
2521 | } else { | |
2522 | if ((rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) == | |
2523 | DYNTICK_TASK_NEST_VALUE) { | |
2524 | rdtp->dynticks_idle_nesting = 0; | |
2525 | } else { | |
2526 | rdtp->dynticks_idle_nesting -= DYNTICK_TASK_NEST_VALUE; | |
2527 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting < 0); | |
2528 | return; /* Still not fully idle. */ | |
2529 | } | |
2530 | } | |
2531 | ||
2532 | /* Record start of fully idle period. */ | |
2533 | j = jiffies; | |
7d0ae808 | 2534 | WRITE_ONCE(rdtp->dynticks_idle_jiffies, j); |
4e857c58 | 2535 | smp_mb__before_atomic(); |
eb348b89 | 2536 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2537 | smp_mb__after_atomic(); |
eb348b89 PM |
2538 | WARN_ON_ONCE(atomic_read(&rdtp->dynticks_idle) & 0x1); |
2539 | } | |
2540 | ||
0edd1b17 PM |
2541 | /* |
2542 | * Unconditionally force exit from full system-idle state. This is | |
2543 | * invoked when a normal CPU exits idle, but must be called separately | |
2544 | * for the timekeeping CPU (tick_do_timer_cpu). The reason for this | |
2545 | * is that the timekeeping CPU is permitted to take scheduling-clock | |
2546 | * interrupts while the system is in system-idle state, and of course | |
2547 | * rcu_sysidle_exit() has no way of distinguishing a scheduling-clock | |
2548 | * interrupt from any other type of interrupt. | |
2549 | */ | |
2550 | void rcu_sysidle_force_exit(void) | |
2551 | { | |
7d0ae808 | 2552 | int oldstate = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2553 | int newoldstate; |
2554 | ||
2555 | /* | |
2556 | * Each pass through the following loop attempts to exit full | |
2557 | * system-idle state. If contention proves to be a problem, | |
2558 | * a trylock-based contention tree could be used here. | |
2559 | */ | |
2560 | while (oldstate > RCU_SYSIDLE_SHORT) { | |
2561 | newoldstate = cmpxchg(&full_sysidle_state, | |
2562 | oldstate, RCU_SYSIDLE_NOT); | |
2563 | if (oldstate == newoldstate && | |
2564 | oldstate == RCU_SYSIDLE_FULL_NOTED) { | |
2565 | rcu_kick_nohz_cpu(tick_do_timer_cpu); | |
2566 | return; /* We cleared it, done! */ | |
2567 | } | |
2568 | oldstate = newoldstate; | |
2569 | } | |
2570 | smp_mb(); /* Order initial oldstate fetch vs. later non-idle work. */ | |
2571 | } | |
2572 | ||
eb348b89 PM |
2573 | /* |
2574 | * Invoked to note entry to irq or task transition from idle. Note that | |
2575 | * usermode execution does -not- count as idle here! The caller must | |
2576 | * have disabled interrupts. | |
2577 | */ | |
28ced795 | 2578 | static void rcu_sysidle_exit(int irq) |
eb348b89 | 2579 | { |
28ced795 CL |
2580 | struct rcu_dynticks *rdtp = this_cpu_ptr(&rcu_dynticks); |
2581 | ||
663e1310 PM |
2582 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2583 | if (!tick_nohz_full_enabled()) | |
2584 | return; | |
2585 | ||
eb348b89 PM |
2586 | /* Adjust nesting, check for already non-idle. */ |
2587 | if (irq) { | |
2588 | rdtp->dynticks_idle_nesting++; | |
2589 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2590 | if (rdtp->dynticks_idle_nesting != 1) | |
2591 | return; /* Already non-idle. */ | |
2592 | } else { | |
2593 | /* | |
2594 | * Allow for irq misnesting. Yes, it really is possible | |
2595 | * to enter an irq handler then never leave it, and maybe | |
2596 | * also vice versa. Handle both possibilities. | |
2597 | */ | |
2598 | if (rdtp->dynticks_idle_nesting & DYNTICK_TASK_NEST_MASK) { | |
2599 | rdtp->dynticks_idle_nesting += DYNTICK_TASK_NEST_VALUE; | |
2600 | WARN_ON_ONCE(rdtp->dynticks_idle_nesting <= 0); | |
2601 | return; /* Already non-idle. */ | |
2602 | } else { | |
2603 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_EXIT_IDLE; | |
2604 | } | |
2605 | } | |
2606 | ||
2607 | /* Record end of idle period. */ | |
4e857c58 | 2608 | smp_mb__before_atomic(); |
eb348b89 | 2609 | atomic_inc(&rdtp->dynticks_idle); |
4e857c58 | 2610 | smp_mb__after_atomic(); |
eb348b89 | 2611 | WARN_ON_ONCE(!(atomic_read(&rdtp->dynticks_idle) & 0x1)); |
0edd1b17 PM |
2612 | |
2613 | /* | |
2614 | * If we are the timekeeping CPU, we are permitted to be non-idle | |
2615 | * during a system-idle state. This must be the case, because | |
2616 | * the timekeeping CPU has to take scheduling-clock interrupts | |
2617 | * during the time that the system is transitioning to full | |
2618 | * system-idle state. This means that the timekeeping CPU must | |
2619 | * invoke rcu_sysidle_force_exit() directly if it does anything | |
2620 | * more than take a scheduling-clock interrupt. | |
2621 | */ | |
2622 | if (smp_processor_id() == tick_do_timer_cpu) | |
2623 | return; | |
2624 | ||
2625 | /* Update system-idle state: We are clearly no longer fully idle! */ | |
2626 | rcu_sysidle_force_exit(); | |
2627 | } | |
2628 | ||
2629 | /* | |
2630 | * Check to see if the current CPU is idle. Note that usermode execution | |
5871968d PM |
2631 | * does not count as idle. The caller must have disabled interrupts, |
2632 | * and must be running on tick_do_timer_cpu. | |
0edd1b17 PM |
2633 | */ |
2634 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, | |
2635 | unsigned long *maxj) | |
2636 | { | |
2637 | int cur; | |
2638 | unsigned long j; | |
2639 | struct rcu_dynticks *rdtp = rdp->dynticks; | |
2640 | ||
663e1310 PM |
2641 | /* If there are no nohz_full= CPUs, don't check system-wide idleness. */ |
2642 | if (!tick_nohz_full_enabled()) | |
2643 | return; | |
2644 | ||
0edd1b17 PM |
2645 | /* |
2646 | * If some other CPU has already reported non-idle, if this is | |
2647 | * not the flavor of RCU that tracks sysidle state, or if this | |
2648 | * is an offline or the timekeeping CPU, nothing to do. | |
2649 | */ | |
417e8d26 | 2650 | if (!*isidle || rdp->rsp != rcu_state_p || |
0edd1b17 PM |
2651 | cpu_is_offline(rdp->cpu) || rdp->cpu == tick_do_timer_cpu) |
2652 | return; | |
5871968d PM |
2653 | /* Verify affinity of current kthread. */ |
2654 | WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu); | |
0edd1b17 PM |
2655 | |
2656 | /* Pick up current idle and NMI-nesting counter and check. */ | |
2657 | cur = atomic_read(&rdtp->dynticks_idle); | |
2658 | if (cur & 0x1) { | |
2659 | *isidle = false; /* We are not idle! */ | |
2660 | return; | |
2661 | } | |
2662 | smp_mb(); /* Read counters before timestamps. */ | |
2663 | ||
2664 | /* Pick up timestamps. */ | |
7d0ae808 | 2665 | j = READ_ONCE(rdtp->dynticks_idle_jiffies); |
0edd1b17 PM |
2666 | /* If this CPU entered idle more recently, update maxj timestamp. */ |
2667 | if (ULONG_CMP_LT(*maxj, j)) | |
2668 | *maxj = j; | |
2669 | } | |
2670 | ||
2671 | /* | |
2672 | * Is this the flavor of RCU that is handling full-system idle? | |
2673 | */ | |
2674 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2675 | { | |
417e8d26 | 2676 | return rsp == rcu_state_p; |
0edd1b17 PM |
2677 | } |
2678 | ||
2679 | /* | |
2680 | * Return a delay in jiffies based on the number of CPUs, rcu_node | |
2681 | * leaf fanout, and jiffies tick rate. The idea is to allow larger | |
2682 | * systems more time to transition to full-idle state in order to | |
2683 | * avoid the cache thrashing that otherwise occur on the state variable. | |
2684 | * Really small systems (less than a couple of tens of CPUs) should | |
2685 | * instead use a single global atomically incremented counter, and later | |
2686 | * versions of this will automatically reconfigure themselves accordingly. | |
2687 | */ | |
2688 | static unsigned long rcu_sysidle_delay(void) | |
2689 | { | |
2690 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2691 | return 0; | |
2692 | return DIV_ROUND_UP(nr_cpu_ids * HZ, rcu_fanout_leaf * 1000); | |
2693 | } | |
2694 | ||
2695 | /* | |
2696 | * Advance the full-system-idle state. This is invoked when all of | |
2697 | * the non-timekeeping CPUs are idle. | |
2698 | */ | |
2699 | static void rcu_sysidle(unsigned long j) | |
2700 | { | |
2701 | /* Check the current state. */ | |
7d0ae808 | 2702 | switch (READ_ONCE(full_sysidle_state)) { |
0edd1b17 PM |
2703 | case RCU_SYSIDLE_NOT: |
2704 | ||
2705 | /* First time all are idle, so note a short idle period. */ | |
7d0ae808 | 2706 | WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_SHORT); |
0edd1b17 PM |
2707 | break; |
2708 | ||
2709 | case RCU_SYSIDLE_SHORT: | |
2710 | ||
2711 | /* | |
2712 | * Idle for a bit, time to advance to next state? | |
2713 | * cmpxchg failure means race with non-idle, let them win. | |
2714 | */ | |
2715 | if (ULONG_CMP_GE(jiffies, j + rcu_sysidle_delay())) | |
2716 | (void)cmpxchg(&full_sysidle_state, | |
2717 | RCU_SYSIDLE_SHORT, RCU_SYSIDLE_LONG); | |
2718 | break; | |
2719 | ||
2720 | case RCU_SYSIDLE_LONG: | |
2721 | ||
2722 | /* | |
2723 | * Do an additional check pass before advancing to full. | |
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_LONG, RCU_SYSIDLE_FULL); | |
2729 | break; | |
2730 | ||
2731 | default: | |
2732 | break; | |
2733 | } | |
2734 | } | |
2735 | ||
2736 | /* | |
2737 | * Found a non-idle non-timekeeping CPU, so kick the system-idle state | |
2738 | * back to the beginning. | |
2739 | */ | |
2740 | static void rcu_sysidle_cancel(void) | |
2741 | { | |
2742 | smp_mb(); | |
becb41bf | 2743 | if (full_sysidle_state > RCU_SYSIDLE_SHORT) |
7d0ae808 | 2744 | WRITE_ONCE(full_sysidle_state, RCU_SYSIDLE_NOT); |
0edd1b17 PM |
2745 | } |
2746 | ||
2747 | /* | |
2748 | * Update the sysidle state based on the results of a force-quiescent-state | |
2749 | * scan of the CPUs' dyntick-idle state. | |
2750 | */ | |
2751 | static void rcu_sysidle_report(struct rcu_state *rsp, int isidle, | |
2752 | unsigned long maxj, bool gpkt) | |
2753 | { | |
417e8d26 | 2754 | if (rsp != rcu_state_p) |
0edd1b17 PM |
2755 | return; /* Wrong flavor, ignore. */ |
2756 | if (gpkt && nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) | |
2757 | return; /* Running state machine from timekeeping CPU. */ | |
2758 | if (isidle) | |
2759 | rcu_sysidle(maxj); /* More idle! */ | |
2760 | else | |
2761 | rcu_sysidle_cancel(); /* Idle is over. */ | |
2762 | } | |
2763 | ||
2764 | /* | |
2765 | * Wrapper for rcu_sysidle_report() when called from the grace-period | |
2766 | * kthread's context. | |
2767 | */ | |
2768 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
2769 | unsigned long maxj) | |
2770 | { | |
663e1310 PM |
2771 | /* If there are no nohz_full= CPUs, no need to track this. */ |
2772 | if (!tick_nohz_full_enabled()) | |
2773 | return; | |
2774 | ||
0edd1b17 PM |
2775 | rcu_sysidle_report(rsp, isidle, maxj, true); |
2776 | } | |
2777 | ||
2778 | /* Callback and function for forcing an RCU grace period. */ | |
2779 | struct rcu_sysidle_head { | |
2780 | struct rcu_head rh; | |
2781 | int inuse; | |
2782 | }; | |
2783 | ||
2784 | static void rcu_sysidle_cb(struct rcu_head *rhp) | |
2785 | { | |
2786 | struct rcu_sysidle_head *rshp; | |
2787 | ||
2788 | /* | |
2789 | * The following memory barrier is needed to replace the | |
2790 | * memory barriers that would normally be in the memory | |
2791 | * allocator. | |
2792 | */ | |
2793 | smp_mb(); /* grace period precedes setting inuse. */ | |
2794 | ||
2795 | rshp = container_of(rhp, struct rcu_sysidle_head, rh); | |
7d0ae808 | 2796 | WRITE_ONCE(rshp->inuse, 0); |
0edd1b17 PM |
2797 | } |
2798 | ||
2799 | /* | |
2800 | * Check to see if the system is fully idle, other than the timekeeping CPU. | |
663e1310 PM |
2801 | * The caller must have disabled interrupts. This is not intended to be |
2802 | * called unless tick_nohz_full_enabled(). | |
0edd1b17 PM |
2803 | */ |
2804 | bool rcu_sys_is_idle(void) | |
2805 | { | |
2806 | static struct rcu_sysidle_head rsh; | |
7d0ae808 | 2807 | int rss = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2808 | |
2809 | if (WARN_ON_ONCE(smp_processor_id() != tick_do_timer_cpu)) | |
2810 | return false; | |
2811 | ||
2812 | /* Handle small-system case by doing a full scan of CPUs. */ | |
2813 | if (nr_cpu_ids <= CONFIG_NO_HZ_FULL_SYSIDLE_SMALL) { | |
2814 | int oldrss = rss - 1; | |
2815 | ||
2816 | /* | |
2817 | * One pass to advance to each state up to _FULL. | |
2818 | * Give up if any pass fails to advance the state. | |
2819 | */ | |
2820 | while (rss < RCU_SYSIDLE_FULL && oldrss < rss) { | |
2821 | int cpu; | |
2822 | bool isidle = true; | |
2823 | unsigned long maxj = jiffies - ULONG_MAX / 4; | |
2824 | struct rcu_data *rdp; | |
2825 | ||
2826 | /* Scan all the CPUs looking for nonidle CPUs. */ | |
2827 | for_each_possible_cpu(cpu) { | |
417e8d26 | 2828 | rdp = per_cpu_ptr(rcu_state_p->rda, cpu); |
0edd1b17 PM |
2829 | rcu_sysidle_check_cpu(rdp, &isidle, &maxj); |
2830 | if (!isidle) | |
2831 | break; | |
2832 | } | |
417e8d26 | 2833 | rcu_sysidle_report(rcu_state_p, isidle, maxj, false); |
0edd1b17 | 2834 | oldrss = rss; |
7d0ae808 | 2835 | rss = READ_ONCE(full_sysidle_state); |
0edd1b17 PM |
2836 | } |
2837 | } | |
2838 | ||
2839 | /* If this is the first observation of an idle period, record it. */ | |
2840 | if (rss == RCU_SYSIDLE_FULL) { | |
2841 | rss = cmpxchg(&full_sysidle_state, | |
2842 | RCU_SYSIDLE_FULL, RCU_SYSIDLE_FULL_NOTED); | |
2843 | return rss == RCU_SYSIDLE_FULL; | |
2844 | } | |
2845 | ||
2846 | smp_mb(); /* ensure rss load happens before later caller actions. */ | |
2847 | ||
2848 | /* If already fully idle, tell the caller (in case of races). */ | |
2849 | if (rss == RCU_SYSIDLE_FULL_NOTED) | |
2850 | return true; | |
2851 | ||
2852 | /* | |
2853 | * If we aren't there yet, and a grace period is not in flight, | |
2854 | * initiate a grace period. Either way, tell the caller that | |
2855 | * we are not there yet. We use an xchg() rather than an assignment | |
2856 | * to make up for the memory barriers that would otherwise be | |
2857 | * provided by the memory allocator. | |
2858 | */ | |
2859 | if (nr_cpu_ids > CONFIG_NO_HZ_FULL_SYSIDLE_SMALL && | |
417e8d26 | 2860 | !rcu_gp_in_progress(rcu_state_p) && |
0edd1b17 PM |
2861 | !rsh.inuse && xchg(&rsh.inuse, 1) == 0) |
2862 | call_rcu(&rsh.rh, rcu_sysidle_cb); | |
2863 | return false; | |
eb348b89 PM |
2864 | } |
2865 | ||
2333210b PM |
2866 | /* |
2867 | * Initialize dynticks sysidle state for CPUs coming online. | |
2868 | */ | |
2869 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) | |
2870 | { | |
2871 | rdtp->dynticks_idle_nesting = DYNTICK_TASK_NEST_VALUE; | |
2872 | } | |
2873 | ||
2874 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
2875 | ||
28ced795 | 2876 | static void rcu_sysidle_enter(int irq) |
eb348b89 PM |
2877 | { |
2878 | } | |
2879 | ||
28ced795 | 2880 | static void rcu_sysidle_exit(int irq) |
eb348b89 PM |
2881 | { |
2882 | } | |
2883 | ||
0edd1b17 PM |
2884 | static void rcu_sysidle_check_cpu(struct rcu_data *rdp, bool *isidle, |
2885 | unsigned long *maxj) | |
2886 | { | |
2887 | } | |
2888 | ||
2889 | static bool is_sysidle_rcu_state(struct rcu_state *rsp) | |
2890 | { | |
2891 | return false; | |
2892 | } | |
2893 | ||
2894 | static void rcu_sysidle_report_gp(struct rcu_state *rsp, int isidle, | |
2895 | unsigned long maxj) | |
2896 | { | |
2897 | } | |
2898 | ||
2333210b PM |
2899 | static void rcu_sysidle_init_percpu_data(struct rcu_dynticks *rdtp) |
2900 | { | |
2901 | } | |
2902 | ||
2903 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ | |
a096932f PM |
2904 | |
2905 | /* | |
2906 | * Is this CPU a NO_HZ_FULL CPU that should ignore RCU so that the | |
2907 | * grace-period kthread will do force_quiescent_state() processing? | |
2908 | * The idea is to avoid waking up RCU core processing on such a | |
2909 | * CPU unless the grace period has extended for too long. | |
2910 | * | |
2911 | * This code relies on the fact that all NO_HZ_FULL CPUs are also | |
52e2bb95 | 2912 | * CONFIG_RCU_NOCB_CPU CPUs. |
a096932f PM |
2913 | */ |
2914 | static bool rcu_nohz_full_cpu(struct rcu_state *rsp) | |
2915 | { | |
2916 | #ifdef CONFIG_NO_HZ_FULL | |
2917 | if (tick_nohz_full_cpu(smp_processor_id()) && | |
2918 | (!rcu_gp_in_progress(rsp) || | |
7d0ae808 | 2919 | ULONG_CMP_LT(jiffies, READ_ONCE(rsp->gp_start) + HZ))) |
5ce035fb | 2920 | return true; |
a096932f | 2921 | #endif /* #ifdef CONFIG_NO_HZ_FULL */ |
5ce035fb | 2922 | return false; |
a096932f | 2923 | } |
5057f55e PM |
2924 | |
2925 | /* | |
2926 | * Bind the grace-period kthread for the sysidle flavor of RCU to the | |
2927 | * timekeeping CPU. | |
2928 | */ | |
2929 | static void rcu_bind_gp_kthread(void) | |
2930 | { | |
c0f489d2 | 2931 | int __maybe_unused cpu; |
5057f55e | 2932 | |
c0f489d2 | 2933 | if (!tick_nohz_full_enabled()) |
5057f55e | 2934 | return; |
c0f489d2 PM |
2935 | #ifdef CONFIG_NO_HZ_FULL_SYSIDLE |
2936 | cpu = tick_do_timer_cpu; | |
5871968d | 2937 | if (cpu >= 0 && cpu < nr_cpu_ids) |
5057f55e | 2938 | set_cpus_allowed_ptr(current, cpumask_of(cpu)); |
c0f489d2 | 2939 | #else /* #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
5871968d | 2940 | housekeeping_affine(current); |
c0f489d2 | 2941 | #endif /* #else #ifdef CONFIG_NO_HZ_FULL_SYSIDLE */ |
5057f55e | 2942 | } |
176f8f7a PM |
2943 | |
2944 | /* Record the current task on dyntick-idle entry. */ | |
2945 | static void rcu_dynticks_task_enter(void) | |
2946 | { | |
2947 | #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) | |
7d0ae808 | 2948 | WRITE_ONCE(current->rcu_tasks_idle_cpu, smp_processor_id()); |
176f8f7a PM |
2949 | #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ |
2950 | } | |
2951 | ||
2952 | /* Record no current task on dyntick-idle exit. */ | |
2953 | static void rcu_dynticks_task_exit(void) | |
2954 | { | |
2955 | #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) | |
7d0ae808 | 2956 | WRITE_ONCE(current->rcu_tasks_idle_cpu, -1); |
176f8f7a PM |
2957 | #endif /* #if defined(CONFIG_TASKS_RCU) && defined(CONFIG_NO_HZ_FULL) */ |
2958 | } |