Commit | Line | Data |
---|---|---|
22e40925 | 1 | /* SPDX-License-Identifier: GPL-2.0+ */ |
3549c2bc PM |
2 | /* |
3 | * RCU expedited grace periods | |
4 | * | |
3549c2bc PM |
5 | * Copyright IBM Corporation, 2016 |
6 | * | |
22e40925 | 7 | * Authors: Paul E. McKenney <paulmck@linux.ibm.com> |
3549c2bc PM |
8 | */ |
9 | ||
55ebfce0 BF |
10 | #include <linux/lockdep.h> |
11 | ||
142d106d | 12 | static void rcu_exp_handler(void *unused); |
d87cda50 | 13 | static int rcu_print_task_exp_stall(struct rcu_node *rnp); |
142d106d | 14 | |
09e2db37 PM |
15 | /* |
16 | * Record the start of an expedited grace period. | |
17 | */ | |
63d4c8c9 | 18 | static void rcu_exp_gp_seq_start(void) |
3549c2bc | 19 | { |
63d4c8c9 | 20 | rcu_seq_start(&rcu_state.expedited_sequence); |
3549c2bc | 21 | } |
09e2db37 | 22 | |
9a414201 | 23 | /* |
de8cd0a5 | 24 | * Return the value that the expedited-grace-period counter will have |
9a414201 PM |
25 | * at the end of the current grace period. |
26 | */ | |
63d4c8c9 | 27 | static __maybe_unused unsigned long rcu_exp_gp_seq_endval(void) |
9a414201 | 28 | { |
63d4c8c9 | 29 | return rcu_seq_endval(&rcu_state.expedited_sequence); |
9a414201 PM |
30 | } |
31 | ||
09e2db37 PM |
32 | /* |
33 | * Record the end of an expedited grace period. | |
34 | */ | |
63d4c8c9 | 35 | static void rcu_exp_gp_seq_end(void) |
3549c2bc | 36 | { |
63d4c8c9 | 37 | rcu_seq_end(&rcu_state.expedited_sequence); |
3549c2bc PM |
38 | smp_mb(); /* Ensure that consecutive grace periods serialize. */ |
39 | } | |
09e2db37 PM |
40 | |
41 | /* | |
de8cd0a5 PM |
42 | * Take a snapshot of the expedited-grace-period counter, which is the |
43 | * earliest value that will indicate that a full grace period has | |
44 | * elapsed since the current time. | |
09e2db37 | 45 | */ |
63d4c8c9 | 46 | static unsigned long rcu_exp_gp_seq_snap(void) |
3549c2bc PM |
47 | { |
48 | unsigned long s; | |
49 | ||
50 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
63d4c8c9 PM |
51 | s = rcu_seq_snap(&rcu_state.expedited_sequence); |
52 | trace_rcu_exp_grace_period(rcu_state.name, s, TPS("snap")); | |
3549c2bc PM |
53 | return s; |
54 | } | |
09e2db37 PM |
55 | |
56 | /* | |
57 | * Given a counter snapshot from rcu_exp_gp_seq_snap(), return true | |
58 | * if a full expedited grace period has elapsed since that snapshot | |
59 | * was taken. | |
60 | */ | |
63d4c8c9 | 61 | static bool rcu_exp_gp_seq_done(unsigned long s) |
3549c2bc | 62 | { |
63d4c8c9 | 63 | return rcu_seq_done(&rcu_state.expedited_sequence, s); |
3549c2bc PM |
64 | } |
65 | ||
66 | /* | |
67 | * Reset the ->expmaskinit values in the rcu_node tree to reflect any | |
68 | * recent CPU-online activity. Note that these masks are not cleared | |
69 | * when CPUs go offline, so they reflect the union of all CPUs that have | |
70 | * ever been online. This means that this function normally takes its | |
71 | * no-work-to-do fastpath. | |
72 | */ | |
63d4c8c9 | 73 | static void sync_exp_reset_tree_hotplug(void) |
3549c2bc PM |
74 | { |
75 | bool done; | |
76 | unsigned long flags; | |
77 | unsigned long mask; | |
78 | unsigned long oldmask; | |
63d4c8c9 | 79 | int ncpus = smp_load_acquire(&rcu_state.ncpus); /* Order vs. locking. */ |
3549c2bc PM |
80 | struct rcu_node *rnp; |
81 | struct rcu_node *rnp_up; | |
82 | ||
83 | /* If no new CPUs onlined since last time, nothing to do. */ | |
63d4c8c9 | 84 | if (likely(ncpus == rcu_state.ncpus_snap)) |
3549c2bc | 85 | return; |
63d4c8c9 | 86 | rcu_state.ncpus_snap = ncpus; |
3549c2bc PM |
87 | |
88 | /* | |
89 | * Each pass through the following loop propagates newly onlined | |
90 | * CPUs for the current rcu_node structure up the rcu_node tree. | |
91 | */ | |
aedf4ba9 | 92 | rcu_for_each_leaf_node(rnp) { |
3549c2bc PM |
93 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
94 | if (rnp->expmaskinit == rnp->expmaskinitnext) { | |
95 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
96 | continue; /* No new CPUs, nothing to do. */ | |
97 | } | |
98 | ||
99 | /* Update this node's mask, track old value for propagation. */ | |
100 | oldmask = rnp->expmaskinit; | |
101 | rnp->expmaskinit = rnp->expmaskinitnext; | |
102 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
103 | ||
104 | /* If was already nonzero, nothing to propagate. */ | |
105 | if (oldmask) | |
106 | continue; | |
107 | ||
108 | /* Propagate the new CPU up the tree. */ | |
109 | mask = rnp->grpmask; | |
110 | rnp_up = rnp->parent; | |
111 | done = false; | |
112 | while (rnp_up) { | |
113 | raw_spin_lock_irqsave_rcu_node(rnp_up, flags); | |
114 | if (rnp_up->expmaskinit) | |
115 | done = true; | |
116 | rnp_up->expmaskinit |= mask; | |
117 | raw_spin_unlock_irqrestore_rcu_node(rnp_up, flags); | |
118 | if (done) | |
119 | break; | |
120 | mask = rnp_up->grpmask; | |
121 | rnp_up = rnp_up->parent; | |
122 | } | |
123 | } | |
124 | } | |
125 | ||
126 | /* | |
127 | * Reset the ->expmask values in the rcu_node tree in preparation for | |
128 | * a new expedited grace period. | |
129 | */ | |
63d4c8c9 | 130 | static void __maybe_unused sync_exp_reset_tree(void) |
3549c2bc PM |
131 | { |
132 | unsigned long flags; | |
133 | struct rcu_node *rnp; | |
134 | ||
63d4c8c9 | 135 | sync_exp_reset_tree_hotplug(); |
aedf4ba9 | 136 | rcu_for_each_node_breadth_first(rnp) { |
3549c2bc PM |
137 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
138 | WARN_ON_ONCE(rnp->expmask); | |
15c7c972 | 139 | WRITE_ONCE(rnp->expmask, rnp->expmaskinit); |
3549c2bc PM |
140 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
141 | } | |
142 | } | |
143 | ||
144 | /* | |
145 | * Return non-zero if there is no RCU expedited grace period in progress | |
146 | * for the specified rcu_node structure, in other words, if all CPUs and | |
147 | * tasks covered by the specified rcu_node structure have done their bit | |
de8cd0a5 | 148 | * for the current expedited grace period. |
3549c2bc | 149 | */ |
6c7d7dbf | 150 | static bool sync_rcu_exp_done(struct rcu_node *rnp) |
3549c2bc | 151 | { |
55ebfce0 | 152 | raw_lockdep_assert_held_rcu_node(rnp); |
314eeb43 | 153 | return READ_ONCE(rnp->exp_tasks) == NULL && |
3549c2bc PM |
154 | READ_ONCE(rnp->expmask) == 0; |
155 | } | |
156 | ||
55ebfce0 | 157 | /* |
de8cd0a5 PM |
158 | * Like sync_rcu_exp_done(), but where the caller does not hold the |
159 | * rcu_node's ->lock. | |
55ebfce0 | 160 | */ |
6c7d7dbf | 161 | static bool sync_rcu_exp_done_unlocked(struct rcu_node *rnp) |
55ebfce0 BF |
162 | { |
163 | unsigned long flags; | |
164 | bool ret; | |
165 | ||
166 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
6c7d7dbf | 167 | ret = sync_rcu_exp_done(rnp); |
55ebfce0 BF |
168 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
169 | ||
170 | return ret; | |
171 | } | |
172 | ||
173 | ||
3549c2bc PM |
174 | /* |
175 | * Report the exit from RCU read-side critical section for the last task | |
176 | * that queued itself during or before the current expedited preemptible-RCU | |
177 | * grace period. This event is reported either to the rcu_node structure on | |
178 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
179 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
180 | * iteratively!) | |
3549c2bc | 181 | */ |
63d4c8c9 | 182 | static void __rcu_report_exp_rnp(struct rcu_node *rnp, |
3549c2bc PM |
183 | bool wake, unsigned long flags) |
184 | __releases(rnp->lock) | |
185 | { | |
186 | unsigned long mask; | |
187 | ||
de8cd0a5 | 188 | raw_lockdep_assert_held_rcu_node(rnp); |
3549c2bc | 189 | for (;;) { |
6c7d7dbf | 190 | if (!sync_rcu_exp_done(rnp)) { |
3549c2bc PM |
191 | if (!rnp->expmask) |
192 | rcu_initiate_boost(rnp, flags); | |
193 | else | |
194 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
195 | break; | |
196 | } | |
197 | if (rnp->parent == NULL) { | |
198 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
199 | if (wake) { | |
200 | smp_mb(); /* EGP done before wake_up(). */ | |
63d4c8c9 | 201 | swake_up_one(&rcu_state.expedited_wq); |
3549c2bc PM |
202 | } |
203 | break; | |
204 | } | |
205 | mask = rnp->grpmask; | |
206 | raw_spin_unlock_rcu_node(rnp); /* irqs remain disabled */ | |
207 | rnp = rnp->parent; | |
208 | raw_spin_lock_rcu_node(rnp); /* irqs already disabled */ | |
209 | WARN_ON_ONCE(!(rnp->expmask & mask)); | |
15c7c972 | 210 | WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask); |
3549c2bc PM |
211 | } |
212 | } | |
213 | ||
214 | /* | |
215 | * Report expedited quiescent state for specified node. This is a | |
216 | * lock-acquisition wrapper function for __rcu_report_exp_rnp(). | |
3549c2bc | 217 | */ |
63d4c8c9 | 218 | static void __maybe_unused rcu_report_exp_rnp(struct rcu_node *rnp, bool wake) |
3549c2bc PM |
219 | { |
220 | unsigned long flags; | |
221 | ||
222 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
63d4c8c9 | 223 | __rcu_report_exp_rnp(rnp, wake, flags); |
3549c2bc PM |
224 | } |
225 | ||
226 | /* | |
227 | * Report expedited quiescent state for multiple CPUs, all covered by the | |
7be8c56f | 228 | * specified leaf rcu_node structure. |
3549c2bc | 229 | */ |
63d4c8c9 | 230 | static void rcu_report_exp_cpu_mult(struct rcu_node *rnp, |
3549c2bc PM |
231 | unsigned long mask, bool wake) |
232 | { | |
df1e849a | 233 | int cpu; |
3549c2bc | 234 | unsigned long flags; |
df1e849a | 235 | struct rcu_data *rdp; |
3549c2bc PM |
236 | |
237 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
238 | if (!(rnp->expmask & mask)) { | |
239 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); | |
240 | return; | |
241 | } | |
15c7c972 | 242 | WRITE_ONCE(rnp->expmask, rnp->expmask & ~mask); |
df1e849a PM |
243 | for_each_leaf_node_cpu_mask(rnp, cpu, mask) { |
244 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
245 | if (!IS_ENABLED(CONFIG_NO_HZ_FULL) || !rdp->rcu_forced_tick_exp) | |
246 | continue; | |
247 | rdp->rcu_forced_tick_exp = false; | |
248 | tick_dep_clear_cpu(cpu, TICK_DEP_BIT_RCU_EXP); | |
249 | } | |
63d4c8c9 | 250 | __rcu_report_exp_rnp(rnp, wake, flags); /* Releases rnp->lock. */ |
3549c2bc PM |
251 | } |
252 | ||
253 | /* | |
254 | * Report expedited quiescent state for specified rcu_data (CPU). | |
255 | */ | |
63d4c8c9 | 256 | static void rcu_report_exp_rdp(struct rcu_data *rdp) |
3549c2bc | 257 | { |
6e16b0f7 | 258 | WRITE_ONCE(rdp->cpu_no_qs.b.exp, false); |
63d4c8c9 | 259 | rcu_report_exp_cpu_mult(rdp->mynode, rdp->grpmask, true); |
3549c2bc PM |
260 | } |
261 | ||
45975c7d | 262 | /* Common code for work-done checking. */ |
63d4c8c9 | 263 | static bool sync_exp_work_done(unsigned long s) |
3549c2bc | 264 | { |
63d4c8c9 PM |
265 | if (rcu_exp_gp_seq_done(s)) { |
266 | trace_rcu_exp_grace_period(rcu_state.name, s, TPS("done")); | |
96050c68 | 267 | smp_mb(); /* Ensure test happens before caller kfree(). */ |
3549c2bc PM |
268 | return true; |
269 | } | |
270 | return false; | |
271 | } | |
272 | ||
273 | /* | |
274 | * Funnel-lock acquisition for expedited grace periods. Returns true | |
275 | * if some other task completed an expedited grace period that this task | |
276 | * can piggy-back on, and with no mutex held. Otherwise, returns false | |
277 | * with the mutex held, indicating that the caller must actually do the | |
278 | * expedited grace period. | |
279 | */ | |
63d4c8c9 | 280 | static bool exp_funnel_lock(unsigned long s) |
3549c2bc | 281 | { |
da1df50d | 282 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, raw_smp_processor_id()); |
3549c2bc | 283 | struct rcu_node *rnp = rdp->mynode; |
336a4f6c | 284 | struct rcu_node *rnp_root = rcu_get_root(); |
3549c2bc PM |
285 | |
286 | /* Low-contention fastpath. */ | |
287 | if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s) && | |
288 | (rnp == rnp_root || | |
289 | ULONG_CMP_LT(READ_ONCE(rnp_root->exp_seq_rq), s)) && | |
63d4c8c9 | 290 | mutex_trylock(&rcu_state.exp_mutex)) |
3549c2bc PM |
291 | goto fastpath; |
292 | ||
293 | /* | |
294 | * Each pass through the following loop works its way up | |
295 | * the rcu_node tree, returning if others have done the work or | |
63d4c8c9 | 296 | * otherwise falls through to acquire ->exp_mutex. The mapping |
3549c2bc PM |
297 | * from CPU to rcu_node structure can be inexact, as it is just |
298 | * promoting locality and is not strictly needed for correctness. | |
299 | */ | |
300 | for (; rnp != NULL; rnp = rnp->parent) { | |
63d4c8c9 | 301 | if (sync_exp_work_done(s)) |
3549c2bc PM |
302 | return true; |
303 | ||
304 | /* Work not done, either wait here or go up. */ | |
305 | spin_lock(&rnp->exp_lock); | |
306 | if (ULONG_CMP_GE(rnp->exp_seq_rq, s)) { | |
307 | ||
308 | /* Someone else doing GP, so wait for them. */ | |
309 | spin_unlock(&rnp->exp_lock); | |
63d4c8c9 | 310 | trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level, |
3549c2bc PM |
311 | rnp->grplo, rnp->grphi, |
312 | TPS("wait")); | |
031aeee0 | 313 | wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3], |
63d4c8c9 | 314 | sync_exp_work_done(s)); |
3549c2bc PM |
315 | return true; |
316 | } | |
b0c18c87 | 317 | WRITE_ONCE(rnp->exp_seq_rq, s); /* Followers can wait on us. */ |
3549c2bc | 318 | spin_unlock(&rnp->exp_lock); |
63d4c8c9 PM |
319 | trace_rcu_exp_funnel_lock(rcu_state.name, rnp->level, |
320 | rnp->grplo, rnp->grphi, TPS("nxtlvl")); | |
3549c2bc | 321 | } |
63d4c8c9 | 322 | mutex_lock(&rcu_state.exp_mutex); |
3549c2bc | 323 | fastpath: |
63d4c8c9 PM |
324 | if (sync_exp_work_done(s)) { |
325 | mutex_unlock(&rcu_state.exp_mutex); | |
3549c2bc PM |
326 | return true; |
327 | } | |
63d4c8c9 PM |
328 | rcu_exp_gp_seq_start(); |
329 | trace_rcu_exp_grace_period(rcu_state.name, s, TPS("start")); | |
3549c2bc PM |
330 | return false; |
331 | } | |
332 | ||
3549c2bc | 333 | /* |
25f3d7ef PM |
334 | * Select the CPUs within the specified rcu_node that the upcoming |
335 | * expedited grace period needs to wait for. | |
3549c2bc | 336 | */ |
25f3d7ef | 337 | static void sync_rcu_exp_select_node_cpus(struct work_struct *wp) |
3549c2bc PM |
338 | { |
339 | int cpu; | |
340 | unsigned long flags; | |
3549c2bc PM |
341 | unsigned long mask_ofl_test; |
342 | unsigned long mask_ofl_ipi; | |
343 | int ret; | |
25f3d7ef PM |
344 | struct rcu_exp_work *rewp = |
345 | container_of(wp, struct rcu_exp_work, rew_work); | |
346 | struct rcu_node *rnp = container_of(rewp, struct rcu_node, rew); | |
3549c2bc | 347 | |
25f3d7ef | 348 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
3549c2bc | 349 | |
25f3d7ef PM |
350 | /* Each pass checks a CPU for identity, offline, and idle. */ |
351 | mask_ofl_test = 0; | |
352 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) { | |
da1df50d | 353 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
aca2991a | 354 | unsigned long mask = rdp->grpmask; |
25f3d7ef PM |
355 | int snap; |
356 | ||
357 | if (raw_smp_processor_id() == cpu || | |
358 | !(rnp->qsmaskinitnext & mask)) { | |
359 | mask_ofl_test |= mask; | |
360 | } else { | |
dc5a4f29 | 361 | snap = rcu_dynticks_snap(rdp); |
25f3d7ef | 362 | if (rcu_dynticks_in_eqs(snap)) |
65963d24 | 363 | mask_ofl_test |= mask; |
25f3d7ef PM |
364 | else |
365 | rdp->exp_dynticks_snap = snap; | |
3549c2bc | 366 | } |
25f3d7ef PM |
367 | } |
368 | mask_ofl_ipi = rnp->expmask & ~mask_ofl_test; | |
3549c2bc | 369 | |
25f3d7ef PM |
370 | /* |
371 | * Need to wait for any blocked tasks as well. Note that | |
372 | * additional blocking tasks will also block the expedited GP | |
373 | * until such time as the ->expmask bits are cleared. | |
374 | */ | |
375 | if (rcu_preempt_has_tasks(rnp)) | |
314eeb43 | 376 | WRITE_ONCE(rnp->exp_tasks, rnp->blkd_tasks.next); |
25f3d7ef PM |
377 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
378 | ||
379 | /* IPI the remaining CPUs for expedited quiescent state. */ | |
15c7c972 | 380 | for_each_leaf_node_cpu_mask(rnp, cpu, mask_ofl_ipi) { |
da1df50d | 381 | struct rcu_data *rdp = per_cpu_ptr(&rcu_data, cpu); |
aca2991a | 382 | unsigned long mask = rdp->grpmask; |
0742ac3e | 383 | |
3549c2bc | 384 | retry_ipi: |
dc5a4f29 | 385 | if (rcu_dynticks_in_eqs_since(rdp, rdp->exp_dynticks_snap)) { |
25f3d7ef PM |
386 | mask_ofl_test |= mask; |
387 | continue; | |
388 | } | |
b9ad4d6e | 389 | if (get_cpu() == cpu) { |
81f6d49c | 390 | mask_ofl_test |= mask; |
b9ad4d6e PM |
391 | put_cpu(); |
392 | continue; | |
393 | } | |
142d106d | 394 | ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0); |
b9ad4d6e | 395 | put_cpu(); |
9f08cf08 BF |
396 | /* The CPU will report the QS in response to the IPI. */ |
397 | if (!ret) | |
25f3d7ef | 398 | continue; |
9f08cf08 | 399 | |
25f3d7ef PM |
400 | /* Failed, raced with CPU hotplug operation. */ |
401 | raw_spin_lock_irqsave_rcu_node(rnp, flags); | |
402 | if ((rnp->qsmaskinitnext & mask) && | |
403 | (rnp->expmask & mask)) { | |
404 | /* Online, so delay for a bit and try again. */ | |
3549c2bc | 405 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
63d4c8c9 | 406 | trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("selectofl")); |
68c2f27e | 407 | schedule_timeout_idle(1); |
25f3d7ef PM |
408 | goto retry_ipi; |
409 | } | |
9f08cf08 BF |
410 | /* CPU really is offline, so we must report its QS. */ |
411 | if (rnp->expmask & mask) | |
412 | mask_ofl_test |= mask; | |
25f3d7ef PM |
413 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
414 | } | |
415 | /* Report quiescent states for those that went offline. */ | |
25f3d7ef | 416 | if (mask_ofl_test) |
63d4c8c9 | 417 | rcu_report_exp_cpu_mult(rnp, mask_ofl_test, false); |
25f3d7ef PM |
418 | } |
419 | ||
420 | /* | |
421 | * Select the nodes that the upcoming expedited grace period needs | |
422 | * to wait for. | |
423 | */ | |
142d106d | 424 | static void sync_rcu_exp_select_cpus(void) |
25f3d7ef | 425 | { |
fcc63543 | 426 | int cpu; |
25f3d7ef PM |
427 | struct rcu_node *rnp; |
428 | ||
63d4c8c9 PM |
429 | trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("reset")); |
430 | sync_exp_reset_tree(); | |
431 | trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("select")); | |
25f3d7ef PM |
432 | |
433 | /* Schedule work for each leaf rcu_node structure. */ | |
aedf4ba9 | 434 | rcu_for_each_leaf_node(rnp) { |
25f3d7ef PM |
435 | rnp->exp_need_flush = false; |
436 | if (!READ_ONCE(rnp->expmask)) | |
437 | continue; /* Avoid early boot non-existent wq. */ | |
25f3d7ef | 438 | if (!READ_ONCE(rcu_par_gp_wq) || |
5257514d | 439 | rcu_scheduler_active != RCU_SCHEDULER_RUNNING || |
aedf4ba9 | 440 | rcu_is_last_leaf_node(rnp)) { |
5257514d | 441 | /* No workqueues yet or last leaf, do direct call. */ |
25f3d7ef PM |
442 | sync_rcu_exp_select_node_cpus(&rnp->rew.rew_work); |
443 | continue; | |
3549c2bc | 444 | } |
25f3d7ef | 445 | INIT_WORK(&rnp->rew.rew_work, sync_rcu_exp_select_node_cpus); |
9cac83a5 | 446 | cpu = find_next_bit(&rnp->ffmask, BITS_PER_LONG, -1); |
fcc63543 | 447 | /* If all offline, queue the work on an unbound CPU. */ |
9cac83a5 | 448 | if (unlikely(cpu > rnp->grphi - rnp->grplo)) |
fcc63543 | 449 | cpu = WORK_CPU_UNBOUND; |
9cac83a5 PM |
450 | else |
451 | cpu += rnp->grplo; | |
fcc63543 | 452 | queue_work_on(cpu, rcu_par_gp_wq, &rnp->rew.rew_work); |
25f3d7ef | 453 | rnp->exp_need_flush = true; |
3549c2bc | 454 | } |
25f3d7ef PM |
455 | |
456 | /* Wait for workqueue jobs (if any) to complete. */ | |
aedf4ba9 | 457 | rcu_for_each_leaf_node(rnp) |
25f3d7ef PM |
458 | if (rnp->exp_need_flush) |
459 | flush_work(&rnp->rew.rew_work); | |
3549c2bc PM |
460 | } |
461 | ||
df1e849a PM |
462 | /* |
463 | * Wait for the expedited grace period to elapse, within time limit. | |
464 | * If the time limit is exceeded without the grace period elapsing, | |
465 | * return false, otherwise return true. | |
466 | */ | |
467 | static bool synchronize_rcu_expedited_wait_once(long tlimit) | |
468 | { | |
469 | int t; | |
470 | struct rcu_node *rnp_root = rcu_get_root(); | |
471 | ||
472 | t = swait_event_timeout_exclusive(rcu_state.expedited_wq, | |
473 | sync_rcu_exp_done_unlocked(rnp_root), | |
474 | tlimit); | |
475 | // Workqueues should not be signaled. | |
476 | if (t > 0 || sync_rcu_exp_done_unlocked(rnp_root)) | |
477 | return true; | |
478 | WARN_ON(t < 0); /* workqueues should not be signaled. */ | |
479 | return false; | |
480 | } | |
481 | ||
de8cd0a5 PM |
482 | /* |
483 | * Wait for the expedited grace period to elapse, issuing any needed | |
484 | * RCU CPU stall warnings along the way. | |
485 | */ | |
28f0361f | 486 | static void synchronize_rcu_expedited_wait(void) |
3549c2bc PM |
487 | { |
488 | int cpu; | |
59ee0326 | 489 | unsigned long j; |
3549c2bc PM |
490 | unsigned long jiffies_stall; |
491 | unsigned long jiffies_start; | |
492 | unsigned long mask; | |
493 | int ndetected; | |
df1e849a | 494 | struct rcu_data *rdp; |
3549c2bc | 495 | struct rcu_node *rnp; |
336a4f6c | 496 | struct rcu_node *rnp_root = rcu_get_root(); |
3549c2bc | 497 | |
63d4c8c9 | 498 | trace_rcu_exp_grace_period(rcu_state.name, rcu_exp_gp_seq_endval(), TPS("startwait")); |
3549c2bc PM |
499 | jiffies_stall = rcu_jiffies_till_stall_check(); |
500 | jiffies_start = jiffies; | |
59ee0326 | 501 | if (tick_nohz_full_enabled() && rcu_inkernel_boot_has_ended()) { |
df1e849a PM |
502 | if (synchronize_rcu_expedited_wait_once(1)) |
503 | return; | |
504 | rcu_for_each_leaf_node(rnp) { | |
505 | for_each_leaf_node_cpu_mask(rnp, cpu, rnp->expmask) { | |
506 | rdp = per_cpu_ptr(&rcu_data, cpu); | |
507 | if (rdp->rcu_forced_tick_exp) | |
508 | continue; | |
509 | rdp->rcu_forced_tick_exp = true; | |
147f04b1 PM |
510 | preempt_disable(); |
511 | if (cpu_online(cpu)) | |
512 | tick_dep_set_cpu(cpu, TICK_DEP_BIT_RCU_EXP); | |
513 | preempt_enable(); | |
df1e849a PM |
514 | } |
515 | } | |
59ee0326 PM |
516 | j = READ_ONCE(jiffies_till_first_fqs); |
517 | if (synchronize_rcu_expedited_wait_once(j + HZ)) | |
518 | return; | |
df1e849a | 519 | } |
3549c2bc PM |
520 | |
521 | for (;;) { | |
df1e849a | 522 | if (synchronize_rcu_expedited_wait_once(jiffies_stall)) |
3549c2bc | 523 | return; |
58c53360 | 524 | if (rcu_stall_is_suppressed()) |
24a6cff2 PM |
525 | continue; |
526 | panic_on_rcu_stall(); | |
565cfb9e | 527 | trace_rcu_stall_warning(rcu_state.name, TPS("ExpeditedStall")); |
3549c2bc | 528 | pr_err("INFO: %s detected expedited stalls on CPUs/tasks: {", |
63d4c8c9 | 529 | rcu_state.name); |
3549c2bc | 530 | ndetected = 0; |
aedf4ba9 | 531 | rcu_for_each_leaf_node(rnp) { |
3549c2bc | 532 | ndetected += rcu_print_task_exp_stall(rnp); |
bc75e999 | 533 | for_each_leaf_node_possible_cpu(rnp, cpu) { |
3549c2bc PM |
534 | struct rcu_data *rdp; |
535 | ||
bc75e999 | 536 | mask = leaf_node_cpu_bit(rnp, cpu); |
15c7c972 | 537 | if (!(READ_ONCE(rnp->expmask) & mask)) |
3549c2bc PM |
538 | continue; |
539 | ndetected++; | |
da1df50d | 540 | rdp = per_cpu_ptr(&rcu_data, cpu); |
3549c2bc PM |
541 | pr_cont(" %d-%c%c%c", cpu, |
542 | "O."[!!cpu_online(cpu)], | |
543 | "o."[!!(rdp->grpmask & rnp->expmaskinit)], | |
544 | "N."[!!(rdp->grpmask & rnp->expmaskinitnext)]); | |
545 | } | |
3549c2bc PM |
546 | } |
547 | pr_cont(" } %lu jiffies s: %lu root: %#lx/%c\n", | |
63d4c8c9 | 548 | jiffies - jiffies_start, rcu_state.expedited_sequence, |
314eeb43 PM |
549 | data_race(rnp_root->expmask), |
550 | ".T"[!!data_race(rnp_root->exp_tasks)]); | |
3549c2bc | 551 | if (ndetected) { |
b08ea1de | 552 | pr_err("blocking rcu_node structures (internal RCU debug):"); |
aedf4ba9 | 553 | rcu_for_each_node_breadth_first(rnp) { |
3549c2bc PM |
554 | if (rnp == rnp_root) |
555 | continue; /* printed unconditionally */ | |
6c7d7dbf | 556 | if (sync_rcu_exp_done_unlocked(rnp)) |
3549c2bc PM |
557 | continue; |
558 | pr_cont(" l=%u:%d-%d:%#lx/%c", | |
559 | rnp->level, rnp->grplo, rnp->grphi, | |
314eeb43 PM |
560 | data_race(rnp->expmask), |
561 | ".T"[!!data_race(rnp->exp_tasks)]); | |
3549c2bc PM |
562 | } |
563 | pr_cont("\n"); | |
564 | } | |
aedf4ba9 | 565 | rcu_for_each_leaf_node(rnp) { |
bc75e999 MR |
566 | for_each_leaf_node_possible_cpu(rnp, cpu) { |
567 | mask = leaf_node_cpu_bit(rnp, cpu); | |
15c7c972 | 568 | if (!(READ_ONCE(rnp->expmask) & mask)) |
3549c2bc PM |
569 | continue; |
570 | dump_cpu_task(cpu); | |
571 | } | |
572 | } | |
573 | jiffies_stall = 3 * rcu_jiffies_till_stall_check() + 3; | |
574 | } | |
575 | } | |
576 | ||
577 | /* | |
578 | * Wait for the current expedited grace period to complete, and then | |
579 | * wake up everyone who piggybacked on the just-completed expedited | |
580 | * grace period. Also update all the ->exp_seq_rq counters as needed | |
581 | * in order to avoid counter-wrap problems. | |
582 | */ | |
63d4c8c9 | 583 | static void rcu_exp_wait_wake(unsigned long s) |
3549c2bc PM |
584 | { |
585 | struct rcu_node *rnp; | |
586 | ||
28f0361f | 587 | synchronize_rcu_expedited_wait(); |
3549c2bc | 588 | |
4bc6b745 NU |
589 | // Switch over to wakeup mode, allowing the next GP to proceed. |
590 | // End the previous grace period only after acquiring the mutex | |
591 | // to ensure that only one GP runs concurrently with wakeups. | |
63d4c8c9 | 592 | mutex_lock(&rcu_state.exp_wake_mutex); |
4bc6b745 NU |
593 | rcu_exp_gp_seq_end(); |
594 | trace_rcu_exp_grace_period(rcu_state.name, s, TPS("end")); | |
3549c2bc | 595 | |
aedf4ba9 | 596 | rcu_for_each_node_breadth_first(rnp) { |
3549c2bc PM |
597 | if (ULONG_CMP_LT(READ_ONCE(rnp->exp_seq_rq), s)) { |
598 | spin_lock(&rnp->exp_lock); | |
599 | /* Recheck, avoid hang in case someone just arrived. */ | |
600 | if (ULONG_CMP_LT(rnp->exp_seq_rq, s)) | |
24bb9ecc | 601 | WRITE_ONCE(rnp->exp_seq_rq, s); |
3549c2bc PM |
602 | spin_unlock(&rnp->exp_lock); |
603 | } | |
3c345825 | 604 | smp_mb(); /* All above changes before wakeup. */ |
fd6bc19d | 605 | wake_up_all(&rnp->exp_wq[rcu_seq_ctr(s) & 0x3]); |
3549c2bc | 606 | } |
63d4c8c9 PM |
607 | trace_rcu_exp_grace_period(rcu_state.name, s, TPS("endwake")); |
608 | mutex_unlock(&rcu_state.exp_wake_mutex); | |
3549c2bc PM |
609 | } |
610 | ||
52d7e48b PM |
611 | /* |
612 | * Common code to drive an expedited grace period forward, used by | |
613 | * workqueues and mid-boot-time tasks. | |
614 | */ | |
142d106d | 615 | static void rcu_exp_sel_wait_wake(unsigned long s) |
52d7e48b PM |
616 | { |
617 | /* Initialize the rcu_node tree in preparation for the wait. */ | |
142d106d | 618 | sync_rcu_exp_select_cpus(); |
52d7e48b PM |
619 | |
620 | /* Wait and clean up, including waking everyone. */ | |
63d4c8c9 | 621 | rcu_exp_wait_wake(s); |
52d7e48b PM |
622 | } |
623 | ||
8b355e3b PM |
624 | /* |
625 | * Work-queue handler to drive an expedited grace period forward. | |
626 | */ | |
627 | static void wait_rcu_exp_gp(struct work_struct *wp) | |
628 | { | |
629 | struct rcu_exp_work *rewp; | |
630 | ||
8b355e3b | 631 | rewp = container_of(wp, struct rcu_exp_work, rew_work); |
142d106d | 632 | rcu_exp_sel_wait_wake(rewp->rew_s); |
8b355e3b PM |
633 | } |
634 | ||
40e0a6cf PM |
635 | #ifdef CONFIG_PREEMPT_RCU |
636 | ||
637 | /* | |
638 | * Remote handler for smp_call_function_single(). If there is an | |
639 | * RCU read-side critical section in effect, request that the | |
640 | * next rcu_read_unlock() record the quiescent state up the | |
641 | * ->expmask fields in the rcu_node tree. Otherwise, immediately | |
642 | * report the quiescent state. | |
643 | */ | |
142d106d | 644 | static void rcu_exp_handler(void *unused) |
40e0a6cf | 645 | { |
5f5fa7ea | 646 | int depth = rcu_preempt_depth(); |
3e310098 | 647 | unsigned long flags; |
da1df50d | 648 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
3e310098 | 649 | struct rcu_node *rnp = rdp->mynode; |
40e0a6cf PM |
650 | struct task_struct *t = current; |
651 | ||
652 | /* | |
3e310098 PM |
653 | * First, the common case of not being in an RCU read-side |
654 | * critical section. If also enabled or idle, immediately | |
655 | * report the quiescent state, otherwise defer. | |
40e0a6cf | 656 | */ |
5f5fa7ea | 657 | if (!depth) { |
3e310098 PM |
658 | if (!(preempt_count() & (PREEMPT_MASK | SOFTIRQ_MASK)) || |
659 | rcu_dynticks_curr_cpu_in_eqs()) { | |
63d4c8c9 | 660 | rcu_report_exp_rdp(rdp); |
3e310098 | 661 | } else { |
6120b72e | 662 | WRITE_ONCE(rdp->cpu_no_qs.b.exp, true); |
fced9c8c PM |
663 | set_tsk_need_resched(t); |
664 | set_preempt_need_resched(); | |
3e310098 | 665 | } |
40e0a6cf PM |
666 | return; |
667 | } | |
668 | ||
669 | /* | |
3e310098 PM |
670 | * Second, the less-common case of being in an RCU read-side |
671 | * critical section. In this case we can count on a future | |
672 | * rcu_read_unlock(). However, this rcu_read_unlock() might | |
673 | * execute on some other CPU, but in that case there will be | |
674 | * a future context switch. Either way, if the expedited | |
675 | * grace period is still waiting on this CPU, set ->deferred_qs | |
676 | * so that the eventual quiescent state will be reported. | |
677 | * Note that there is a large group of race conditions that | |
678 | * can have caused this quiescent state to already have been | |
679 | * reported, so we really do need to check ->expmask. | |
40e0a6cf | 680 | */ |
5f5fa7ea | 681 | if (depth > 0) { |
3e310098 | 682 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
05f41571 | 683 | if (rnp->expmask & rdp->grpmask) { |
6120b72e | 684 | WRITE_ONCE(rdp->cpu_no_qs.b.exp, true); |
add0d37b | 685 | t->rcu_read_unlock_special.b.exp_hint = true; |
05f41571 | 686 | } |
3e310098 | 687 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
1de462ed | 688 | return; |
3e310098 PM |
689 | } |
690 | ||
5f5fa7ea LJ |
691 | // Finally, negative nesting depth should not happen. |
692 | WARN_ON_ONCE(1); | |
40e0a6cf PM |
693 | } |
694 | ||
90326f05 | 695 | /* PREEMPTION=y, so no PREEMPTION=n expedited grace period to clean up after. */ |
45975c7d PM |
696 | static void sync_sched_exp_online_cleanup(int cpu) |
697 | { | |
698 | } | |
699 | ||
d87cda50 PM |
700 | /* |
701 | * Scan the current list of tasks blocked within RCU read-side critical | |
702 | * sections, printing out the tid of each that is blocking the current | |
703 | * expedited grace period. | |
704 | */ | |
705 | static int rcu_print_task_exp_stall(struct rcu_node *rnp) | |
706 | { | |
314eeb43 | 707 | unsigned long flags; |
d87cda50 | 708 | int ndetected = 0; |
314eeb43 | 709 | struct task_struct *t; |
d87cda50 | 710 | |
314eeb43 | 711 | if (!READ_ONCE(rnp->exp_tasks)) |
d87cda50 | 712 | return 0; |
314eeb43 | 713 | raw_spin_lock_irqsave_rcu_node(rnp, flags); |
d87cda50 PM |
714 | t = list_entry(rnp->exp_tasks->prev, |
715 | struct task_struct, rcu_node_entry); | |
716 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { | |
717 | pr_cont(" P%d", t->pid); | |
718 | ndetected++; | |
719 | } | |
314eeb43 | 720 | raw_spin_unlock_irqrestore_rcu_node(rnp, flags); |
d87cda50 PM |
721 | return ndetected; |
722 | } | |
723 | ||
40e0a6cf PM |
724 | #else /* #ifdef CONFIG_PREEMPT_RCU */ |
725 | ||
e015a341 PM |
726 | /* Request an expedited quiescent state. */ |
727 | static void rcu_exp_need_qs(void) | |
728 | { | |
729 | __this_cpu_write(rcu_data.cpu_no_qs.b.exp, true); | |
730 | /* Store .exp before .rcu_urgent_qs. */ | |
731 | smp_store_release(this_cpu_ptr(&rcu_data.rcu_urgent_qs), true); | |
732 | set_tsk_need_resched(current); | |
733 | set_preempt_need_resched(); | |
734 | } | |
735 | ||
45975c7d | 736 | /* Invoked on each online non-idle CPU for expedited quiescent state. */ |
142d106d | 737 | static void rcu_exp_handler(void *unused) |
45975c7d | 738 | { |
7487ea07 PM |
739 | struct rcu_data *rdp = this_cpu_ptr(&rcu_data); |
740 | struct rcu_node *rnp = rdp->mynode; | |
45975c7d | 741 | |
45975c7d PM |
742 | if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) || |
743 | __this_cpu_read(rcu_data.cpu_no_qs.b.exp)) | |
744 | return; | |
745 | if (rcu_is_cpu_rrupt_from_idle()) { | |
63d4c8c9 | 746 | rcu_report_exp_rdp(this_cpu_ptr(&rcu_data)); |
45975c7d PM |
747 | return; |
748 | } | |
e015a341 | 749 | rcu_exp_need_qs(); |
45975c7d PM |
750 | } |
751 | ||
752 | /* Send IPI for expedited cleanup if needed at end of CPU-hotplug operation. */ | |
753 | static void sync_sched_exp_online_cleanup(int cpu) | |
754 | { | |
e015a341 PM |
755 | unsigned long flags; |
756 | int my_cpu; | |
45975c7d PM |
757 | struct rcu_data *rdp; |
758 | int ret; | |
759 | struct rcu_node *rnp; | |
45975c7d | 760 | |
da1df50d | 761 | rdp = per_cpu_ptr(&rcu_data, cpu); |
45975c7d | 762 | rnp = rdp->mynode; |
e015a341 PM |
763 | my_cpu = get_cpu(); |
764 | /* Quiescent state either not needed or already requested, leave. */ | |
765 | if (!(READ_ONCE(rnp->expmask) & rdp->grpmask) || | |
5401cc52 | 766 | READ_ONCE(rdp->cpu_no_qs.b.exp)) { |
e015a341 | 767 | put_cpu(); |
45975c7d | 768 | return; |
e015a341 PM |
769 | } |
770 | /* Quiescent state needed on current CPU, so set it up locally. */ | |
771 | if (my_cpu == cpu) { | |
772 | local_irq_save(flags); | |
773 | rcu_exp_need_qs(); | |
774 | local_irq_restore(flags); | |
775 | put_cpu(); | |
776 | return; | |
777 | } | |
778 | /* Quiescent state needed on some other CPU, send IPI. */ | |
142d106d | 779 | ret = smp_call_function_single(cpu, rcu_exp_handler, NULL, 0); |
e015a341 | 780 | put_cpu(); |
45975c7d PM |
781 | WARN_ON_ONCE(ret); |
782 | } | |
783 | ||
d87cda50 PM |
784 | /* |
785 | * Because preemptible RCU does not exist, we never have to check for | |
786 | * tasks blocked within RCU read-side critical sections that are | |
787 | * blocking the current expedited grace period. | |
788 | */ | |
789 | static int rcu_print_task_exp_stall(struct rcu_node *rnp) | |
790 | { | |
791 | return 0; | |
792 | } | |
793 | ||
3cd4ca47 | 794 | #endif /* #else #ifdef CONFIG_PREEMPT_RCU */ |
45975c7d | 795 | |
3cd4ca47 PM |
796 | /** |
797 | * synchronize_rcu_expedited - Brute-force RCU grace period | |
798 | * | |
799 | * Wait for an RCU grace period, but expedite it. The basic idea is to | |
800 | * IPI all non-idle non-nohz online CPUs. The IPI handler checks whether | |
801 | * the CPU is in an RCU critical section, and if so, it sets a flag that | |
802 | * causes the outermost rcu_read_unlock() to report the quiescent state | |
803 | * for RCU-preempt or asks the scheduler for help for RCU-sched. On the | |
804 | * other hand, if the CPU is not in an RCU read-side critical section, | |
805 | * the IPI handler reports the quiescent state immediately. | |
806 | * | |
511b44f7 | 807 | * Although this is a great improvement over previous expedited |
3cd4ca47 PM |
808 | * implementations, it is still unfriendly to real-time workloads, so is |
809 | * thus not recommended for any sort of common-case code. In fact, if | |
810 | * you are using synchronize_rcu_expedited() in a loop, please restructure | |
de8cd0a5 | 811 | * your code to batch your updates, and then use a single synchronize_rcu() |
3cd4ca47 PM |
812 | * instead. |
813 | * | |
814 | * This has the same semantics as (but is more brutal than) synchronize_rcu(). | |
815 | */ | |
40e0a6cf PM |
816 | void synchronize_rcu_expedited(void) |
817 | { | |
fbad01af | 818 | bool boottime = (rcu_scheduler_active == RCU_SCHEDULER_INIT); |
89230726 PM |
819 | struct rcu_exp_work rew; |
820 | struct rcu_node *rnp; | |
821 | unsigned long s; | |
822 | ||
45975c7d PM |
823 | RCU_LOCKDEP_WARN(lock_is_held(&rcu_bh_lock_map) || |
824 | lock_is_held(&rcu_lock_map) || | |
825 | lock_is_held(&rcu_sched_lock_map), | |
8fa946d4 | 826 | "Illegal synchronize_rcu_expedited() in RCU read-side critical section"); |
45975c7d | 827 | |
3cd4ca47 | 828 | /* Is the state is such that the call is a grace period? */ |
45975c7d PM |
829 | if (rcu_blocking_is_gp()) |
830 | return; | |
831 | ||
89230726 PM |
832 | /* If expedited grace periods are prohibited, fall back to normal. */ |
833 | if (rcu_gp_is_normal()) { | |
834 | wait_rcu_gp(call_rcu); | |
835 | return; | |
836 | } | |
837 | ||
838 | /* Take a snapshot of the sequence number. */ | |
839 | s = rcu_exp_gp_seq_snap(); | |
840 | if (exp_funnel_lock(s)) | |
841 | return; /* Someone else did our work for us. */ | |
842 | ||
843 | /* Ensure that load happens before action based on it. */ | |
fbad01af | 844 | if (unlikely(boottime)) { |
89230726 PM |
845 | /* Direct call during scheduler init and early_initcalls(). */ |
846 | rcu_exp_sel_wait_wake(s); | |
847 | } else { | |
848 | /* Marshall arguments & schedule the expedited grace period. */ | |
849 | rew.rew_s = s; | |
850 | INIT_WORK_ONSTACK(&rew.rew_work, wait_rcu_exp_gp); | |
851 | queue_work(rcu_gp_wq, &rew.rew_work); | |
852 | } | |
853 | ||
854 | /* Wait for expedited grace period to complete. */ | |
89230726 PM |
855 | rnp = rcu_get_root(); |
856 | wait_event(rnp->exp_wq[rcu_seq_ctr(s) & 0x3], | |
857 | sync_exp_work_done(s)); | |
858 | smp_mb(); /* Workqueue actions happen before return. */ | |
859 | ||
860 | /* Let the next expedited grace period start. */ | |
861 | mutex_unlock(&rcu_state.exp_mutex); | |
fbad01af PM |
862 | |
863 | if (likely(!boottime)) | |
864 | destroy_work_on_stack(&rew.rew_work); | |
40e0a6cf PM |
865 | } |
866 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); |