Commit | Line | Data |
---|---|---|
3549c2bc PM |
1 | /* |
2 | * RCU expedited grace periods | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public License | |
15 | * along with this program; if not, you can access it online at | |
16 | * http://www.gnu.org/licenses/gpl-2.0.html. | |
17 | * | |
18 | * Copyright IBM Corporation, 2016 | |
19 | * | |
20 | * Authors: Paul E. McKenney <paulmck@linux.vnet.ibm.com> | |
21 | */ | |
22 | ||
55ebfce0 BF |
23 | #include <linux/lockdep.h> |
24 | ||
09e2db37 PM |
25 | /* |
26 | * Record the start of an expedited grace period. | |
27 | */ | |
3549c2bc PM |
28 | static void rcu_exp_gp_seq_start(struct rcu_state *rsp) |
29 | { | |
30 | rcu_seq_start(&rsp->expedited_sequence); | |
31 | } | |
09e2db37 | 32 | |
9a414201 PM |
33 | /* |
34 | * Return then value that expedited-grace-period counter will have | |
35 | * at the end of the current grace period. | |
36 | */ | |
7f5d42d0 | 37 | static __maybe_unused unsigned long rcu_exp_gp_seq_endval(struct rcu_state *rsp) |
9a414201 PM |
38 | { |
39 | return rcu_seq_endval(&rsp->expedited_sequence); | |
40 | } | |
41 | ||
09e2db37 PM |
42 | /* |
43 | * Record the end of an expedited grace period. | |
44 | */ | |
3549c2bc PM |
45 | static void rcu_exp_gp_seq_end(struct rcu_state *rsp) |
46 | { | |
47 | rcu_seq_end(&rsp->expedited_sequence); | |
48 | smp_mb(); /* Ensure that consecutive grace periods serialize. */ | |
49 | } | |
09e2db37 PM |
50 | |
51 | /* | |
52 | * Take a snapshot of the expedited-grace-period counter. | |
53 | */ | |
3549c2bc PM |
54 | static unsigned long rcu_exp_gp_seq_snap(struct rcu_state *rsp) |
55 | { | |
56 | unsigned long s; | |
57 | ||
58 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
59 | s = rcu_seq_snap(&rsp->expedited_sequence); | |
60 | trace_rcu_exp_grace_period(rsp->name, s, TPS("snap")); | |
61 | return s; | |
62 | } | |
09e2db37 PM |
63 | |
64 | /* | |
65 | * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true | |
66 | * if a full expedited grace period has elapsed since that snapshot | |
67 | * was taken. | |
68 | */ | |
3549c2bc PM |
69 | static bool rcu_exp_gp_seq_done(struct rcu_state *rsp, unsigned long s) |
70 | { | |
71 | return rcu_seq_done(&rsp->expedited_sequence, s); | |
72 | } | |
73 | ||
74 | /* | |
75 | * Reset the ->expmaskinit values in the rcu_node tree to reflect any | |
76 | * recent CPU-online activity. Note that these masks are not cleared | |
77 | * when CPUs go offline, so they reflect the union of all CPUs that have | |
78 | * ever been online. This means that this function normally takes its | |
79 | * no-work-to-do fastpath. | |
80 | */ | |
81 | static void sync_exp_reset_tree_hotplug(struct rcu_state *rsp) | |
82 | { | |
83 | bool done; | |
84 | unsigned long flags; | |
85 | unsigned long mask; | |
86 | unsigned long oldmask; | |
313517fc | 87 | int ncpus = smp_load_acquire(&rsp->ncpus); /* Order against locking. */ |
3549c2bc PM |
88 | struct rcu_node *rnp; |
89 | struct rcu_node *rnp_up; | |
90 | ||
91 | /* If no new CPUs onlined since last time, nothing to do. */ | |
92 | if (likely(ncpus == rsp->ncpus_snap)) | |
93 | return; | |
94 | rsp->ncpus_snap = ncpus; | |
95 | ||
96 | /* | |
97 | * Each pass through the following loop propagates newly onlined | |
98 | * CPUs for the current rcu_node structure up the rcu_node tree. | |
99 | */ | |
100 | rcu_for_each_leaf_node(rsp, rnp) { | |
101 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
102 | if (rnp->expmaskinit == rnp->expmaskinitnext) { | |
103 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
104 | continue; /* No new CPUs, nothing to do. */ | |
105 | } | |
106 | ||
107 | /* Update this node's mask, track old value for propagation. */ | |
108 | oldmask = rnp->expmaskinit; | |
109 | rnp->expmaskinit = rnp->expmaskinitnext; | |
110 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
111 | ||
112 | /* If was already nonzero, nothing to propagate. */ | |
113 | if (oldmask) | |
114 | continue; | |
115 | ||
116 | /* Propagate the new CPU up the tree. */ | |
117 | mask = rnp->grpmask; | |
118 | rnp_up = rnp->parent; | |
119 | done = false; | |
120 | while (rnp_up) { | |
121 | raw_spin_lock_irqsave_rcu_node(rnp_up, flags); | |
122 | if (rnp_up->expmaskinit) | |
123 | done = true; | |
124 | rnp_up->expmaskinit |= mask; | |
125 | raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags); | |
126 | if (done) | |
127 | break; | |
128 | mask = rnp_up->grpmask; | |
129 | rnp_up = rnp_up->parent; | |
130 | } | |
131 | } | |
132 | } | |
133 | ||
134 | /* | |
135 | * Reset the ->expmask values in the rcu_node tree in preparation for | |
136 | * a new expedited grace period. | |
137 | */ | |
138 | static void __maybe_unused sync_exp_reset_tree(struct rcu_state *rsp) | |
139 | { | |
140 | unsigned long flags; | |
141 | struct rcu_node *rnp; | |
142 | ||
143 | sync_exp_reset_tree_hotplug(rsp); | |
144 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
145 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
146 | WARN_ON_ONCE(rnp->expmask); | |
147 | rnp->expmask = rnp->expmaskinit; | |
148 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
149 | } | |
150 | } | |
151 | ||
152 | /* | |
153 | * Return non-zero if there is no RCU expedited grace period in progress | |
154 | * for the specified rcu_node structure, in other words, if all CPUs and | |
155 | * tasks covered by the specified rcu_node structure have done their bit | |
156 | * for the current expedited grace period. Works only for preemptible | |
157 | * RCU -- other RCU implementation use other means. | |
158 | * | |
7be8c56f | 159 | * Caller must hold the specificed rcu_node structure's ->lock |
3549c2bc | 160 | */ |
dcfc315b | 161 | static bool sync_rcu_preempt_exp_done(struct rcu_node *rnp) |
3549c2bc | 162 | { |
55ebfce0 BF |
163 | raw_lockdep_assert_held_rcu_node(rnp); |
164 | ||
3549c2bc PM |
165 | return rnp->exp_tasks == NULL && |
166 | READ_ONCE(rnp->expmask) == 0; | |
167 | } | |
168 | ||
55ebfce0 BF |
169 | /* |
170 | * Like sync_rcu_preempt_exp_done(), but this function assumes the caller | |
171 | * doesn't hold the rcu_node's ->lock, and will acquire and release the lock | |
172 | * itself | |
173 | */ | |
174 | static bool sync_rcu_preempt_exp_done_unlocked(struct rcu_node *rnp) | |
175 | { | |
176 | unsigned long flags; | |
177 | bool ret; | |
178 | ||
179 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
180 | ret = sync_rcu_preempt_exp_done(rnp); | |
181 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
182 | ||
183 | return ret; | |
184 | } | |
185 | ||
186 | ||
3549c2bc PM |
187 | /* |
188 | * Report the exit from RCU read-side critical section for the last task | |
189 | * that queued itself during or before the current expedited preemptible-RCU | |
190 | * grace period. This event is reported either to the rcu_node structure on | |
191 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
192 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
193 | * iteratively!) | |
194 | * | |
7be8c56f | 195 | * Caller must hold the specified rcu_node structure's ->lock. |
3549c2bc PM |
196 | */ |
197 | static void __rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, | |
198 | bool wake, unsigned long flags) | |
199 | __releases(rnp->lock) | |
200 | { | |
201 | unsigned long mask; | |
202 | ||
203 | for (;;) { | |
204 | if (!sync_rcu_preempt_exp_done(rnp)) { | |
205 | if (!rnp->expmask) | |
206 | rcu_initiate_boost(rnp, flags); | |
207 | else | |
208 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
209 | break; | |
210 | } | |
211 | if (rnp->parent == NULL) { | |
212 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
213 | if (wake) { | |
214 | smp_mb(); /* EGP done before wake_up(). */ | |
215 | swake_up(&rsp->expedited_wq); | |
216 | } | |
217 | break; | |
218 | } | |
219 | mask = rnp->grpmask; | |
220 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */ | |
221 | rnp = rnp->parent; | |
222 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled */ | |
223 | WARN_ON_ONCE(!(rnp->expmask & mask)); | |
224 | rnp->expmask &= ~mask; | |
225 | } | |
226 | } | |
227 | ||
228 | /* | |
229 | * Report expedited quiescent state for specified node. This is a | |
230 | * lock-acquisition wrapper function for __rcu_report_exp_rnp(). | |
3549c2bc PM |
231 | */ |
232 | static void __maybe_unused rcu_report_exp_rnp(struct rcu_state *rsp, | |
233 | struct rcu_node *rnp, bool wake) | |
234 | { | |
235 | unsigned long flags; | |
236 | ||
237 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
238 | __rcu_report_exp_rnp(rsp, rnp, wake, flags); | |
239 | } | |
240 | ||
241 | /* | |
242 | * Report expedited quiescent state for multiple CPUs, all covered by the | |
7be8c56f | 243 | * specified leaf rcu_node structure. |
3549c2bc PM |
244 | */ |
245 | static void rcu_report_exp_cpu_mult(struct rcu_state *rsp, struct rcu_node *rnp, | |
246 | unsigned long mask, bool wake) | |
247 | { | |
248 | unsigned long flags; | |
249 | ||
250 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
251 | if (!(rnp->expmask & mask)) { | |
252 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
253 | return; | |
254 | } | |
255 | rnp->expmask &= ~mask; | |
256 | __rcu_report_exp_rnp(rsp, rnp, wake, flags); /* Releases rnp->lock. */ | |
257 | } | |
258 | ||
259 | /* | |
260 | * Report expedited quiescent state for specified rcu_data (CPU). | |
261 | */ | |
262 | static void rcu_report_exp_rdp(struct rcu_state *rsp, struct rcu_data *rdp, | |
263 | bool wake) | |
264 | { | |
265 | rcu_report_exp_cpu_mult(rsp, rdp->mynode, rdp->grpmask, wake); | |
266 | } | |
267 | ||
268 | /* Common code for synchronize_{rcu,sched}_expedited() work-done checking. */ | |
6fba2b37 | 269 | static bool sync_exp_work_done(struct rcu_state *rsp, unsigned long s) |
3549c2bc PM |
270 | { |
271 | if (rcu_exp_gp_seq_done(rsp, s)) { | |
272 | trace_rcu_exp_grace_period(rsp->name, s, TPS("done")); | |
273 | /* Ensure test happens before caller kfree(). */ | |
274 | smp_mb__before_atomic(); /* ^^^ */ | |
3549c2bc PM |
275 | return true; |
276 | } | |
277 | return false; | |
278 | } | |
279 | ||
280 | /* | |
281 | * Funnel-lock acquisition for expedited grace periods. Returns true | |
282 | * if some other task completed an expedited grace period that this task | |
283 | * can piggy-back on, and with no mutex held. Otherwise, returns false | |
284 | * with the mutex held, indicating that the caller must actually do the | |
285 | * expedited grace period. | |
286 | */ | |
287 | static bool exp_funnel_lock(struct rcu_state *rsp, unsigned long s) | |
288 | { | |
289 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id()); | |
290 | struct rcu_node *rnp = rdp->mynode; | |
291 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
292 | ||
293 | /* Low-contention fastpath. */ | |
294 | if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) && | |
295 | (rnp == rnp_root || | |
296 | ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) && | |
3549c2bc PM |
297 | mutex_trylock(&rsp->exp_mutex)) |
298 | goto fastpath; | |
299 | ||
300 | /* | |
301 | * Each pass through the following loop works its way up | |
302 | * the rcu_node tree, returning if others have done the work or | |
303 | * otherwise falls through to acquire rsp->exp_mutex. The mapping | |
304 | * from CPU to rcu_node structure can be inexact, as it is just | |
305 | * promoting locality and is not strictly needed for correctness. | |
306 | */ | |
307 | for (; rnp != NULL; rnp = rnp->parent) { | |
6fba2b37 | 308 | if (sync_exp_work_done(rsp, s)) |
3549c2bc PM |
309 | return true; |
310 | ||
311 | /* Work not done, either wait here or go up. */ | |
312 | spin_lock(&rnp->exp_lock); | |
313 | if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) { | |
314 | ||
315 | /* Someone else doing GP, so wait for them. */ | |
316 | spin_unlock(&rnp->exp_lock); | |
317 | trace_rcu_exp_funnel_lock(rsp->name, rnp->level, | |
318 | rnp->grplo, rnp->grphi, | |
319 | TPS("wait")); | |
031aeee0 | 320 | wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3], |
6fba2b37 | 321 | sync_exp_work_done(rsp, s)); |
3549c2bc PM |
322 | return true; |
323 | } | |
324 | rnp->exp_seq_rq = s; /* Followers can wait on us. */ | |
325 | spin_unlock(&rnp->exp_lock); | |
326 | trace_rcu_exp_funnel_lock(rsp->name, rnp->level, rnp->grplo, | |
327 | rnp->grphi, TPS("nxtlvl")); | |
328 | } | |
329 | mutex_lock(&rsp->exp_mutex); | |
330 | fastpath: | |
6fba2b37 | 331 | if (sync_exp_work_done(rsp, s)) { |
3549c2bc PM |
332 | mutex_unlock(&rsp->exp_mutex); |
333 | return true; | |
334 | } | |
335 | rcu_exp_gp_seq_start(rsp); | |
336 | trace_rcu_exp_grace_period(rsp->name, s, TPS("start")); | |
337 | return false; | |
338 | } | |
339 | ||
340 | /* Invoked on each online non-idle CPU for expedited quiescent state. */ | |
341 | static void sync_sched_exp_handler(void *data) | |
342 | { | |
343 | struct rcu_data *rdp; | |
344 | struct rcu_node *rnp; | |
345 | struct rcu_state *rsp = data; | |
346 | ||
347 | rdp = this_cpu_ptr(rsp->rda); | |
348 | rnp = rdp->mynode; | |
349 | if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) || | |
350 | __this_cpu_read(rcu_sched_data.cpu_no_qs.b.exp)) | |
351 | return; | |
352 | if (rcu_is_cpu_rrupt_from_idle()) { | |
353 | rcu_report_exp_rdp(&rcu_sched_state, | |
354 | this_cpu_ptr(&rcu_sched_data), true); | |
355 | return; | |
356 | } | |
357 | __this_cpu_write(rcu_sched_data.cpu_no_qs.b.exp, true); | |
9226b10d PM |
358 | /* Store .exp before .rcu_urgent_qs. */ |
359 | smp_store_release(this_cpu_ptr(&rcu_dynticks.rcu_urgent_qs), true); | |
3549c2bc PM |
360 | resched_cpu(smp_processor_id()); |
361 | } | |
362 | ||
363 | /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */ | |
364 | static void sync_sched_exp_online_cleanup(int cpu) | |
365 | { | |
366 | struct rcu_data *rdp; | |
367 | int ret; | |
368 | struct rcu_node *rnp; | |
369 | struct rcu_state *rsp = &rcu_sched_state; | |
370 | ||
371 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
372 | rnp = rdp->mynode; | |
373 | if (!(READ_ONCE(rnp->expmask) & rdp->grpmask)) | |
374 | return; | |
375 | ret = smp_call_function_single(cpu, sync_sched_exp_handler, rsp, 0); | |
376 | WARN_ON_ONCE(ret); | |
377 | } | |
378 | ||
379 | /* | |
25f3d7ef PM |
380 | * Select the CPUs within the specified rcu_node that the upcoming |
381 | * expedited grace period needs to wait for. | |
3549c2bc | 382 | */ |
25f3d7ef | 383 | static void sync_rcu_exp_select_node_cpus(struct work_struct *wp) |
3549c2bc PM |
384 | { |
385 | int cpu; | |
386 | unsigned long flags; | |
25f3d7ef | 387 | smp_call_func_t func; |
3549c2bc PM |
388 | unsigned long mask_ofl_test; |
389 | unsigned long mask_ofl_ipi; | |
390 | int ret; | |
25f3d7ef PM |
391 | struct rcu_exp_work *rewp = |
392 | container_of(wp, struct rcu_exp_work, rew_work); | |
393 | struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew); | |
394 | struct rcu_state *rsp = rewp->rew_rsp; | |
3549c2bc | 395 | |
25f3d7ef PM |
396 | func = rewp->rew_func; |
397 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
3549c2bc | 398 | |
25f3d7ef PM |
399 | /* Each pass checks a CPU for identity, offline, and idle. */ |
400 | mask_ofl_test = 0; | |
401 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) { | |
402 | unsigned long mask = leaf_node_cpu_bit(rnp, cpu); | |
403 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
404 | struct rcu_dynticks *rdtp = per_cpu_ptr(&rcu_dynticks, cpu); | |
405 | int snap; | |
406 | ||
407 | if (raw_smp_processor_id() == cpu || | |
408 | !(rnp->qsmaskinitnext & mask)) { | |
409 | mask_ofl_test |= mask; | |
410 | } else { | |
411 | snap = rcu_dynticks_snap(rdtp); | |
412 | if (rcu_dynticks_in_eqs(snap)) | |
65963d24 | 413 | mask_ofl_test |= mask; |
25f3d7ef PM |
414 | else |
415 | rdp->exp_dynticks_snap = snap; | |
3549c2bc | 416 | } |
25f3d7ef PM |
417 | } |
418 | mask_ofl_ipi = rnp->expmask & ~mask_ofl_test; | |
3549c2bc | 419 | |
25f3d7ef PM |
420 | /* |
421 | * Need to wait for any blocked tasks as well. Note that | |
422 | * additional blocking tasks will also block the expedited GP | |
423 | * until such time as the ->expmask bits are cleared. | |
424 | */ | |
425 | if (rcu_preempt_has_tasks(rnp)) | |
426 | rnp->exp_tasks = rnp->blkd_tasks.next; | |
427 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
428 | ||
429 | /* IPI the remaining CPUs for expedited quiescent state. */ | |
430 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) { | |
431 | unsigned long mask = leaf_node_cpu_bit(rnp, cpu); | |
432 | struct rcu_data *rdp = per_cpu_ptr(rsp->rda, cpu); | |
0742ac3e | 433 | |
25f3d7ef PM |
434 | if (!(mask_ofl_ipi & mask)) |
435 | continue; | |
3549c2bc | 436 | retry_ipi: |
25f3d7ef PM |
437 | if (rcu_dynticks_in_eqs_since(rdp->dynticks, |
438 | rdp->exp_dynticks_snap)) { | |
439 | mask_ofl_test |= mask; | |
440 | continue; | |
441 | } | |
442 | ret = smp_call_function_single(cpu, func, rsp, 0); | |
443 | if (!ret) { | |
444 | mask_ofl_ipi &= ~mask; | |
445 | continue; | |
446 | } | |
447 | /* Failed, raced with CPU hotplug operation. */ | |
448 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
449 | if ((rnp->qsmaskinitnext & mask) && | |
450 | (rnp->expmask & mask)) { | |
451 | /* Online, so delay for a bit and try again. */ | |
3549c2bc | 452 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
25f3d7ef PM |
453 | trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("selectofl")); |
454 | schedule_timeout_uninterruptible(1); | |
455 | goto retry_ipi; | |
456 | } | |
457 | /* CPU really is offline, so we can ignore it. */ | |
458 | if (!(rnp->expmask & mask)) | |
459 | mask_ofl_ipi &= ~mask; | |
460 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
461 | } | |
462 | /* Report quiescent states for those that went offline. */ | |
463 | mask_ofl_test |= mask_ofl_ipi; | |
464 | if (mask_ofl_test) | |
465 | rcu_report_exp_cpu_mult(rsp, rnp, mask_ofl_test, false); | |
466 | } | |
467 | ||
468 | /* | |
469 | * Select the nodes that the upcoming expedited grace period needs | |
470 | * to wait for. | |
471 | */ | |
472 | static void sync_rcu_exp_select_cpus(struct rcu_state *rsp, | |
473 | smp_call_func_t func) | |
474 | { | |
475 | struct rcu_node *rnp; | |
476 | ||
477 | trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("reset")); | |
478 | sync_exp_reset_tree(rsp); | |
479 | trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("select")); | |
480 | ||
481 | /* Schedule work for each leaf rcu_node structure. */ | |
482 | rcu_for_each_leaf_node(rsp, rnp) { | |
483 | rnp->exp_need_flush = false; | |
484 | if (!READ_ONCE(rnp->expmask)) | |
485 | continue; /* Avoid early boot non-existent wq. */ | |
486 | rnp->rew.rew_func = func; | |
487 | rnp->rew.rew_rsp = rsp; | |
488 | if (!READ_ONCE(rcu_par_gp_wq) || | |
489 | rcu_scheduler_active != RCU_SCHEDULER_RUNNING) { | |
490 | /* No workqueues yet. */ | |
491 | sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work); | |
492 | continue; | |
3549c2bc | 493 | } |
25f3d7ef PM |
494 | INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus); |
495 | queue_work_on(rnp->grplo, rcu_par_gp_wq, &rnp->rew.rew_work); | |
496 | rnp->exp_need_flush = true; | |
3549c2bc | 497 | } |
25f3d7ef PM |
498 | |
499 | /* Wait for workqueue jobs (if any) to complete. */ | |
500 | rcu_for_each_leaf_node(rsp, rnp) | |
501 | if (rnp->exp_need_flush) | |
502 | flush_work(&rnp->rew.rew_work); | |
3549c2bc PM |
503 | } |
504 | ||
505 | static void synchronize_sched_expedited_wait(struct rcu_state *rsp) | |
506 | { | |
507 | int cpu; | |
508 | unsigned long jiffies_stall; | |
509 | unsigned long jiffies_start; | |
510 | unsigned long mask; | |
511 | int ndetected; | |
512 | struct rcu_node *rnp; | |
513 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
514 | int ret; | |
515 | ||
9a414201 | 516 | trace_rcu_exp_grace_period(rsp->name, rcu_exp_gp_seq_endval(rsp), TPS("startwait")); |
3549c2bc PM |
517 | jiffies_stall = rcu_jiffies_till_stall_check(); |
518 | jiffies_start = jiffies; | |
519 | ||
520 | for (;;) { | |
521 | ret = swait_event_timeout( | |
522 | rsp->expedited_wq, | |
55ebfce0 | 523 | sync_rcu_preempt_exp_done_unlocked(rnp_root), |
3549c2bc | 524 | jiffies_stall); |
55ebfce0 | 525 | if (ret > 0 || sync_rcu_preempt_exp_done_unlocked(rnp_root)) |
3549c2bc | 526 | return; |
908d2c1f | 527 | WARN_ON(ret < 0); /* workqueues should not be signaled. */ |
24a6cff2 PM |
528 | if (rcu_cpu_stall_suppress) |
529 | continue; | |
530 | panic_on_rcu_stall(); | |
3549c2bc PM |
531 | pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {", |
532 | rsp->name); | |
533 | ndetected = 0; | |
534 | rcu_for_each_leaf_node(rsp, rnp) { | |
535 | ndetected += rcu_print_task_exp_stall(rnp); | |
bc75e999 | 536 | for_each_leaf_node_possible_cpu(rnp, cpu) { |
3549c2bc PM |
537 | struct rcu_data *rdp; |
538 | ||
bc75e999 | 539 | mask = leaf_node_cpu_bit(rnp, cpu); |
3549c2bc PM |
540 | if (!(rnp->expmask & mask)) |
541 | continue; | |
542 | ndetected++; | |
543 | rdp = per_cpu_ptr(rsp->rda, cpu); | |
544 | pr_cont(" %d-%c%c%c", cpu, | |
545 | "O."[!!cpu_online(cpu)], | |
546 | "o."[!!(rdp->grpmask & rnp->expmaskinit)], | |
547 | "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]); | |
548 | } | |
3549c2bc PM |
549 | } |
550 | pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n", | |
551 | jiffies - jiffies_start, rsp->expedited_sequence, | |
552 | rnp_root->expmask, ".T"[!!rnp_root->exp_tasks]); | |
553 | if (ndetected) { | |
554 | pr_err("blocking rcu_node structures:"); | |
555 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
556 | if (rnp == rnp_root) | |
557 | continue; /* printed unconditionally */ | |
55ebfce0 | 558 | if (sync_rcu_preempt_exp_done_unlocked(rnp)) |
3549c2bc PM |
559 | continue; |
560 | pr_cont(" l=%u:%d-%d:%#lx/%c", | |
561 | rnp->level, rnp->grplo, rnp->grphi, | |
562 | rnp->expmask, | |
563 | ".T"[!!rnp->exp_tasks]); | |
564 | } | |
565 | pr_cont("\n"); | |
566 | } | |
567 | rcu_for_each_leaf_node(rsp, rnp) { | |
bc75e999 MR |
568 | for_each_leaf_node_possible_cpu(rnp, cpu) { |
569 | mask = leaf_node_cpu_bit(rnp, cpu); | |
3549c2bc PM |
570 | if (!(rnp->expmask & mask)) |
571 | continue; | |
572 | dump_cpu_task(cpu); | |
573 | } | |
574 | } | |
575 | jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3; | |
576 | } | |
577 | } | |
578 | ||
579 | /* | |
580 | * Wait for the current expedited grace period to complete, and then | |
581 | * wake up everyone who piggybacked on the just-completed expedited | |
582 | * grace period. Also update all the ->exp_seq_rq counters as needed | |
583 | * in order to avoid counter-wrap problems. | |
584 | */ | |
585 | static void rcu_exp_wait_wake(struct rcu_state *rsp, unsigned long s) | |
586 | { | |
587 | struct rcu_node *rnp; | |
588 | ||
589 | synchronize_sched_expedited_wait(rsp); | |
590 | rcu_exp_gp_seq_end(rsp); | |
591 | trace_rcu_exp_grace_period(rsp->name, s, TPS("end")); | |
592 | ||
593 | /* | |
594 | * Switch over to wakeup mode, allowing the next GP, but -only- the | |
595 | * next GP, to proceed. | |
596 | */ | |
597 | mutex_lock(&rsp->exp_wake_mutex); | |
3549c2bc PM |
598 | |
599 | rcu_for_each_node_breadth_first(rsp, rnp) { | |
600 | if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) { | |
601 | spin_lock(&rnp->exp_lock); | |
602 | /* Recheck, avoid hang in case someone just arrived. */ | |
603 | if (ULONG_CMP_LT(rnp->exp_seq_rq, s)) | |
604 | rnp->exp_seq_rq = s; | |
605 | spin_unlock(&rnp->exp_lock); | |
606 | } | |
3c345825 | 607 | smp_mb(); /* All above changes before wakeup. */ |
031aeee0 | 608 | wake_up_all(&rnp->exp_wq[rcu_seq_ctr(rsp->expedited_sequence) & 0x3]); |
3549c2bc PM |
609 | } |
610 | trace_rcu_exp_grace_period(rsp->name, s, TPS("endwake")); | |
611 | mutex_unlock(&rsp->exp_wake_mutex); | |
612 | } | |
613 | ||
52d7e48b PM |
614 | /* |
615 | * Common code to drive an expedited grace period forward, used by | |
616 | * workqueues and mid-boot-time tasks. | |
617 | */ | |
618 | static void rcu_exp_sel_wait_wake(struct rcu_state *rsp, | |
619 | smp_call_func_t func, unsigned long s) | |
620 | { | |
621 | /* Initialize the rcu_node tree in preparation for the wait. */ | |
622 | sync_rcu_exp_select_cpus(rsp, func); | |
623 | ||
624 | /* Wait and clean up, including waking everyone. */ | |
625 | rcu_exp_wait_wake(rsp, s); | |
626 | } | |
627 | ||
8b355e3b PM |
628 | /* |
629 | * Work-queue handler to drive an expedited grace period forward. | |
630 | */ | |
631 | static void wait_rcu_exp_gp(struct work_struct *wp) | |
632 | { | |
633 | struct rcu_exp_work *rewp; | |
634 | ||
8b355e3b | 635 | rewp = container_of(wp, struct rcu_exp_work, rew_work); |
52d7e48b | 636 | rcu_exp_sel_wait_wake(rewp->rew_rsp, rewp->rew_func, rewp->rew_s); |
8b355e3b PM |
637 | } |
638 | ||
f7b8eb84 PM |
639 | /* |
640 | * Given an rcu_state pointer and a smp_call_function() handler, kick | |
641 | * off the specified flavor of expedited grace period. | |
642 | */ | |
643 | static void _synchronize_rcu_expedited(struct rcu_state *rsp, | |
644 | smp_call_func_t func) | |
645 | { | |
8b355e3b PM |
646 | struct rcu_data *rdp; |
647 | struct rcu_exp_work rew; | |
648 | struct rcu_node *rnp; | |
f7b8eb84 PM |
649 | unsigned long s; |
650 | ||
651 | /* If expedited grace periods are prohibited, fall back to normal. */ | |
652 | if (rcu_gp_is_normal()) { | |
653 | wait_rcu_gp(rsp->call); | |
654 | return; | |
655 | } | |
656 | ||
657 | /* Take a snapshot of the sequence number. */ | |
658 | s = rcu_exp_gp_seq_snap(rsp); | |
659 | if (exp_funnel_lock(rsp, s)) | |
660 | return; /* Someone else did our work for us. */ | |
661 | ||
52d7e48b PM |
662 | /* Ensure that load happens before action based on it. */ |
663 | if (unlikely(rcu_scheduler_active == RCU_SCHEDULER_INIT)) { | |
664 | /* Direct call during scheduler init and early_initcalls(). */ | |
665 | rcu_exp_sel_wait_wake(rsp, func, s); | |
666 | } else { | |
667 | /* Marshall arguments & schedule the expedited grace period. */ | |
668 | rew.rew_func = func; | |
669 | rew.rew_rsp = rsp; | |
670 | rew.rew_s = s; | |
671 | INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp); | |
ad7c946b | 672 | queue_work(rcu_gp_wq, &rew.rew_work); |
52d7e48b | 673 | } |
8b355e3b PM |
674 | |
675 | /* Wait for expedited grace period to complete. */ | |
676 | rdp = per_cpu_ptr(rsp->rda, raw_smp_processor_id()); | |
677 | rnp = rcu_get_root(rsp); | |
031aeee0 | 678 | wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3], |
6fba2b37 | 679 | sync_exp_work_done(rsp, s)); |
3c345825 | 680 | smp_mb(); /* Workqueue actions happen before return. */ |
8b355e3b PM |
681 | |
682 | /* Let the next expedited grace period start. */ | |
683 | mutex_unlock(&rsp->exp_mutex); | |
f7b8eb84 PM |
684 | } |
685 | ||
3549c2bc PM |
686 | /** |
687 | * synchronize_sched_expedited - Brute-force RCU-sched grace period | |
688 | * | |
689 | * Wait for an RCU-sched grace period to elapse, but use a "big hammer" | |
690 | * approach to force the grace period to end quickly. This consumes | |
691 | * significant time on all CPUs and is unfriendly to real-time workloads, | |
692 | * so is thus not recommended for any sort of common-case code. In fact, | |
693 | * if you are using synchronize_sched_expedited() in a loop, please | |
694 | * restructure your code to batch your updates, and then use a single | |
695 | * synchronize_sched() instead. | |
696 | * | |
697 | * This implementation can be thought of as an application of sequence | |
698 | * locking to expedited grace periods, but using the sequence counter to | |
699 | * determine when someone else has already done the work instead of for | |
700 | * retrying readers. | |
701 | */ | |
702 | void synchronize_sched_expedited(void) | |
703 | { | |
3549c2bc PM |
704 | struct rcu_state *rsp = &rcu_sched_state; |
705 | ||
8dc79888 PM |
706 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
707 | lock_is_held(&rcu_lock_map) || | |
708 | lock_is_held(&rcu_sched_lock_map), | |
709 | "Illegal synchronize_sched_expedited() in RCU read-side critical section"); | |
710 | ||
3549c2bc PM |
711 | /* If only one CPU, this is automatically a grace period. */ |
712 | if (rcu_blocking_is_gp()) | |
713 | return; | |
714 | ||
f7b8eb84 | 715 | _synchronize_rcu_expedited(rsp, sync_sched_exp_handler); |
3549c2bc PM |
716 | } |
717 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
40e0a6cf PM |
718 | |
719 | #ifdef CONFIG_PREEMPT_RCU | |
720 | ||
721 | /* | |
722 | * Remote handler for smp_call_function_single(). If there is an | |
723 | * RCU read-side critical section in effect, request that the | |
724 | * next rcu_read_unlock() record the quiescent state up the | |
725 | * ->expmask fields in the rcu_node tree. Otherwise, immediately | |
726 | * report the quiescent state. | |
727 | */ | |
728 | static void sync_rcu_exp_handler(void *info) | |
729 | { | |
730 | struct rcu_data *rdp; | |
731 | struct rcu_state *rsp = info; | |
732 | struct task_struct *t = current; | |
733 | ||
734 | /* | |
735 | * Within an RCU read-side critical section, request that the next | |
736 | * rcu_read_unlock() report. Unless this RCU read-side critical | |
737 | * section has already blocked, in which case it is already set | |
738 | * up for the expedited grace period to wait on it. | |
739 | */ | |
740 | if (t->rcu_read_lock_nesting > 0 && | |
741 | !t->rcu_read_unlock_special.b.blocked) { | |
742 | t->rcu_read_unlock_special.b.exp_need_qs = true; | |
743 | return; | |
744 | } | |
745 | ||
746 | /* | |
747 | * We are either exiting an RCU read-side critical section (negative | |
748 | * values of t->rcu_read_lock_nesting) or are not in one at all | |
749 | * (zero value of t->rcu_read_lock_nesting). Or we are in an RCU | |
750 | * read-side critical section that blocked before this expedited | |
751 | * grace period started. Either way, we can immediately report | |
752 | * the quiescent state. | |
753 | */ | |
754 | rdp = this_cpu_ptr(rsp->rda); | |
755 | rcu_report_exp_rdp(rsp, rdp, true); | |
756 | } | |
757 | ||
758 | /** | |
759 | * synchronize_rcu_expedited - Brute-force RCU grace period | |
760 | * | |
761 | * Wait for an RCU-preempt grace period, but expedite it. The basic | |
762 | * idea is to IPI all non-idle non-nohz online CPUs. The IPI handler | |
763 | * checks whether the CPU is in an RCU-preempt critical section, and | |
764 | * if so, it sets a flag that causes the outermost rcu_read_unlock() | |
765 | * to report the quiescent state. On the other hand, if the CPU is | |
766 | * not in an RCU read-side critical section, the IPI handler reports | |
767 | * the quiescent state immediately. | |
768 | * | |
769 | * Although this is a greate improvement over previous expedited | |
770 | * implementations, it is still unfriendly to real-time workloads, so is | |
771 | * thus not recommended for any sort of common-case code. In fact, if | |
772 | * you are using synchronize_rcu_expedited() in a loop, please restructure | |
773 | * your code to batch your updates, and then Use a single synchronize_rcu() | |
774 | * instead. | |
775 | */ | |
776 | void synchronize_rcu_expedited(void) | |
777 | { | |
778 | struct rcu_state *rsp = rcu_state_p; | |
40e0a6cf | 779 | |
8dc79888 PM |
780 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
781 | lock_is_held(&rcu_lock_map) || | |
782 | lock_is_held(&rcu_sched_lock_map), | |
783 | "Illegal synchronize_rcu_expedited() in RCU read-side critical section"); | |
784 | ||
52d7e48b PM |
785 | if (rcu_scheduler_active == RCU_SCHEDULER_INACTIVE) |
786 | return; | |
f7b8eb84 | 787 | _synchronize_rcu_expedited(rsp, sync_rcu_exp_handler); |
40e0a6cf PM |
788 | } |
789 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
790 | ||
791 | #else /* #ifdef CONFIG_PREEMPT_RCU */ | |
792 | ||
793 | /* | |
794 | * Wait for an rcu-preempt grace period, but make it happen quickly. | |
795 | * But because preemptible RCU does not exist, map to rcu-sched. | |
796 | */ | |
797 | void synchronize_rcu_expedited(void) | |
798 | { | |
799 | synchronize_sched_expedited(); | |
800 | } | |
801 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
802 | ||
803 | #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ |