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 | |
17 | * along with this program; if not, write to the Free Software | |
18 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. | |
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> |
7b27d547 | 28 | #include <linux/stop_machine.h> |
f41d911f | 29 | |
26845c28 PM |
30 | /* |
31 | * Check the RCU kernel configuration parameters and print informative | |
32 | * messages about anything out of the ordinary. If you like #ifdef, you | |
33 | * will love this function. | |
34 | */ | |
35 | static void __init rcu_bootup_announce_oddness(void) | |
36 | { | |
37 | #ifdef CONFIG_RCU_TRACE | |
38 | printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n"); | |
39 | #endif | |
40 | #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32) | |
41 | printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n", | |
42 | CONFIG_RCU_FANOUT); | |
43 | #endif | |
44 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
45 | printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n"); | |
46 | #endif | |
47 | #ifdef CONFIG_RCU_FAST_NO_HZ | |
48 | printk(KERN_INFO | |
49 | "\tRCU dyntick-idle grace-period acceleration is enabled.\n"); | |
50 | #endif | |
51 | #ifdef CONFIG_PROVE_RCU | |
52 | printk(KERN_INFO "\tRCU lockdep checking is enabled.\n"); | |
53 | #endif | |
54 | #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE | |
55 | printk(KERN_INFO "\tRCU torture testing starts during boot.\n"); | |
56 | #endif | |
81a294c4 | 57 | #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE) |
26845c28 PM |
58 | printk(KERN_INFO "\tVerbose stalled-CPUs detection is disabled.\n"); |
59 | #endif | |
60 | #if NUM_RCU_LVL_4 != 0 | |
61 | printk(KERN_INFO "\tExperimental four-level hierarchy is enabled.\n"); | |
62 | #endif | |
63 | } | |
64 | ||
f41d911f PM |
65 | #ifdef CONFIG_TREE_PREEMPT_RCU |
66 | ||
67 | struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt_state); | |
68 | DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data); | |
27f4d280 | 69 | static struct rcu_state *rcu_state = &rcu_preempt_state; |
f41d911f | 70 | |
d9a3da06 PM |
71 | static int rcu_preempted_readers_exp(struct rcu_node *rnp); |
72 | ||
f41d911f PM |
73 | /* |
74 | * Tell them what RCU they are running. | |
75 | */ | |
0e0fc1c2 | 76 | static void __init rcu_bootup_announce(void) |
f41d911f | 77 | { |
6cc68793 | 78 | printk(KERN_INFO "Preemptible hierarchical RCU implementation.\n"); |
26845c28 | 79 | rcu_bootup_announce_oddness(); |
f41d911f PM |
80 | } |
81 | ||
82 | /* | |
83 | * Return the number of RCU-preempt batches processed thus far | |
84 | * for debug and statistics. | |
85 | */ | |
86 | long rcu_batches_completed_preempt(void) | |
87 | { | |
88 | return rcu_preempt_state.completed; | |
89 | } | |
90 | EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt); | |
91 | ||
92 | /* | |
93 | * Return the number of RCU batches processed thus far for debug & stats. | |
94 | */ | |
95 | long rcu_batches_completed(void) | |
96 | { | |
97 | return rcu_batches_completed_preempt(); | |
98 | } | |
99 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
100 | ||
bf66f18e PM |
101 | /* |
102 | * Force a quiescent state for preemptible RCU. | |
103 | */ | |
104 | void rcu_force_quiescent_state(void) | |
105 | { | |
106 | force_quiescent_state(&rcu_preempt_state, 0); | |
107 | } | |
108 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
109 | ||
f41d911f | 110 | /* |
6cc68793 | 111 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
f41d911f PM |
112 | * that this just means that the task currently running on the CPU is |
113 | * not in a quiescent state. There might be any number of tasks blocked | |
114 | * while in an RCU read-side critical section. | |
25502a6c PM |
115 | * |
116 | * Unlike the other rcu_*_qs() functions, callers to this function | |
117 | * must disable irqs in order to protect the assignment to | |
118 | * ->rcu_read_unlock_special. | |
f41d911f | 119 | */ |
c3422bea | 120 | static void rcu_preempt_qs(int cpu) |
f41d911f PM |
121 | { |
122 | struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu); | |
25502a6c | 123 | |
c64ac3ce | 124 | rdp->passed_quiesc_completed = rdp->gpnum - 1; |
c3422bea PM |
125 | barrier(); |
126 | rdp->passed_quiesc = 1; | |
25502a6c | 127 | current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS; |
f41d911f PM |
128 | } |
129 | ||
130 | /* | |
c3422bea PM |
131 | * We have entered the scheduler, and the current task might soon be |
132 | * context-switched away from. If this task is in an RCU read-side | |
133 | * critical section, we will no longer be able to rely on the CPU to | |
12f5f524 PM |
134 | * record that fact, so we enqueue the task on the blkd_tasks list. |
135 | * The task will dequeue itself when it exits the outermost enclosing | |
136 | * RCU read-side critical section. Therefore, the current grace period | |
137 | * cannot be permitted to complete until the blkd_tasks list entries | |
138 | * predating the current grace period drain, in other words, until | |
139 | * rnp->gp_tasks becomes NULL. | |
c3422bea PM |
140 | * |
141 | * Caller must disable preemption. | |
f41d911f | 142 | */ |
c3422bea | 143 | static void rcu_preempt_note_context_switch(int cpu) |
f41d911f PM |
144 | { |
145 | struct task_struct *t = current; | |
c3422bea | 146 | unsigned long flags; |
f41d911f PM |
147 | struct rcu_data *rdp; |
148 | struct rcu_node *rnp; | |
149 | ||
150 | if (t->rcu_read_lock_nesting && | |
151 | (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) { | |
152 | ||
153 | /* Possibly blocking in an RCU read-side critical section. */ | |
394f99a9 | 154 | rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu); |
f41d911f | 155 | rnp = rdp->mynode; |
1304afb2 | 156 | raw_spin_lock_irqsave(&rnp->lock, flags); |
f41d911f | 157 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED; |
86848966 | 158 | t->rcu_blocked_node = rnp; |
f41d911f PM |
159 | |
160 | /* | |
161 | * If this CPU has already checked in, then this task | |
162 | * will hold up the next grace period rather than the | |
163 | * current grace period. Queue the task accordingly. | |
164 | * If the task is queued for the current grace period | |
165 | * (i.e., this CPU has not yet passed through a quiescent | |
166 | * state for the current grace period), then as long | |
167 | * as that task remains queued, the current grace period | |
12f5f524 PM |
168 | * cannot end. Note that there is some uncertainty as |
169 | * to exactly when the current grace period started. | |
170 | * We take a conservative approach, which can result | |
171 | * in unnecessarily waiting on tasks that started very | |
172 | * slightly after the current grace period began. C'est | |
173 | * la vie!!! | |
b0e165c0 PM |
174 | * |
175 | * But first, note that the current CPU must still be | |
176 | * on line! | |
f41d911f | 177 | */ |
b0e165c0 | 178 | WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0); |
e7d8842e | 179 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
12f5f524 PM |
180 | if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) { |
181 | list_add(&t->rcu_node_entry, rnp->gp_tasks->prev); | |
182 | rnp->gp_tasks = &t->rcu_node_entry; | |
27f4d280 PM |
183 | #ifdef CONFIG_RCU_BOOST |
184 | if (rnp->boost_tasks != NULL) | |
185 | rnp->boost_tasks = rnp->gp_tasks; | |
186 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
12f5f524 PM |
187 | } else { |
188 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); | |
189 | if (rnp->qsmask & rdp->grpmask) | |
190 | rnp->gp_tasks = &t->rcu_node_entry; | |
191 | } | |
1304afb2 | 192 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
f41d911f PM |
193 | } |
194 | ||
195 | /* | |
196 | * Either we were not in an RCU read-side critical section to | |
197 | * begin with, or we have now recorded that critical section | |
198 | * globally. Either way, we can now note a quiescent state | |
199 | * for this CPU. Again, if we were in an RCU read-side critical | |
200 | * section, and if that critical section was blocking the current | |
201 | * grace period, then the fact that the task has been enqueued | |
202 | * means that we continue to block the current grace period. | |
203 | */ | |
e7d8842e | 204 | local_irq_save(flags); |
25502a6c | 205 | rcu_preempt_qs(cpu); |
e7d8842e | 206 | local_irq_restore(flags); |
f41d911f PM |
207 | } |
208 | ||
209 | /* | |
6cc68793 | 210 | * Tree-preemptible RCU implementation for rcu_read_lock(). |
f41d911f PM |
211 | * Just increment ->rcu_read_lock_nesting, shared state will be updated |
212 | * if we block. | |
213 | */ | |
214 | void __rcu_read_lock(void) | |
215 | { | |
80dcf60e | 216 | current->rcu_read_lock_nesting++; |
f41d911f PM |
217 | barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */ |
218 | } | |
219 | EXPORT_SYMBOL_GPL(__rcu_read_lock); | |
220 | ||
fc2219d4 PM |
221 | /* |
222 | * Check for preempted RCU readers blocking the current grace period | |
223 | * for the specified rcu_node structure. If the caller needs a reliable | |
224 | * answer, it must hold the rcu_node's ->lock. | |
225 | */ | |
27f4d280 | 226 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 | 227 | { |
12f5f524 | 228 | return rnp->gp_tasks != NULL; |
fc2219d4 PM |
229 | } |
230 | ||
b668c9cf PM |
231 | /* |
232 | * Record a quiescent state for all tasks that were previously queued | |
233 | * on the specified rcu_node structure and that were blocking the current | |
234 | * RCU grace period. The caller must hold the specified rnp->lock with | |
235 | * irqs disabled, and this lock is released upon return, but irqs remain | |
236 | * disabled. | |
237 | */ | |
d3f6bad3 | 238 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
b668c9cf PM |
239 | __releases(rnp->lock) |
240 | { | |
241 | unsigned long mask; | |
242 | struct rcu_node *rnp_p; | |
243 | ||
27f4d280 | 244 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
1304afb2 | 245 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf PM |
246 | return; /* Still need more quiescent states! */ |
247 | } | |
248 | ||
249 | rnp_p = rnp->parent; | |
250 | if (rnp_p == NULL) { | |
251 | /* | |
252 | * Either there is only one rcu_node in the tree, | |
253 | * or tasks were kicked up to root rcu_node due to | |
254 | * CPUs going offline. | |
255 | */ | |
d3f6bad3 | 256 | rcu_report_qs_rsp(&rcu_preempt_state, flags); |
b668c9cf PM |
257 | return; |
258 | } | |
259 | ||
260 | /* Report up the rest of the hierarchy. */ | |
261 | mask = rnp->grpmask; | |
1304afb2 PM |
262 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
263 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
d3f6bad3 | 264 | rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags); |
b668c9cf PM |
265 | } |
266 | ||
12f5f524 PM |
267 | /* |
268 | * Advance a ->blkd_tasks-list pointer to the next entry, instead | |
269 | * returning NULL if at the end of the list. | |
270 | */ | |
271 | static struct list_head *rcu_next_node_entry(struct task_struct *t, | |
272 | struct rcu_node *rnp) | |
273 | { | |
274 | struct list_head *np; | |
275 | ||
276 | np = t->rcu_node_entry.next; | |
277 | if (np == &rnp->blkd_tasks) | |
278 | np = NULL; | |
279 | return np; | |
280 | } | |
281 | ||
b668c9cf PM |
282 | /* |
283 | * Handle special cases during rcu_read_unlock(), such as needing to | |
284 | * notify RCU core processing or task having blocked during the RCU | |
285 | * read-side critical section. | |
286 | */ | |
f41d911f PM |
287 | static void rcu_read_unlock_special(struct task_struct *t) |
288 | { | |
289 | int empty; | |
d9a3da06 | 290 | int empty_exp; |
f41d911f | 291 | unsigned long flags; |
12f5f524 | 292 | struct list_head *np; |
f41d911f PM |
293 | struct rcu_node *rnp; |
294 | int special; | |
295 | ||
296 | /* NMI handlers cannot block and cannot safely manipulate state. */ | |
297 | if (in_nmi()) | |
298 | return; | |
299 | ||
300 | local_irq_save(flags); | |
301 | ||
302 | /* | |
303 | * If RCU core is waiting for this CPU to exit critical section, | |
304 | * let it know that we have done so. | |
305 | */ | |
306 | special = t->rcu_read_unlock_special; | |
307 | if (special & RCU_READ_UNLOCK_NEED_QS) { | |
c3422bea | 308 | rcu_preempt_qs(smp_processor_id()); |
f41d911f PM |
309 | } |
310 | ||
311 | /* Hardware IRQ handlers cannot block. */ | |
312 | if (in_irq()) { | |
313 | local_irq_restore(flags); | |
314 | return; | |
315 | } | |
316 | ||
317 | /* Clean up if blocked during RCU read-side critical section. */ | |
318 | if (special & RCU_READ_UNLOCK_BLOCKED) { | |
319 | t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED; | |
320 | ||
dd5d19ba PM |
321 | /* |
322 | * Remove this task from the list it blocked on. The | |
323 | * task can migrate while we acquire the lock, but at | |
324 | * most one time. So at most two passes through loop. | |
325 | */ | |
326 | for (;;) { | |
86848966 | 327 | rnp = t->rcu_blocked_node; |
1304afb2 | 328 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
86848966 | 329 | if (rnp == t->rcu_blocked_node) |
dd5d19ba | 330 | break; |
1304afb2 | 331 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
dd5d19ba | 332 | } |
27f4d280 | 333 | empty = !rcu_preempt_blocked_readers_cgp(rnp); |
d9a3da06 PM |
334 | empty_exp = !rcu_preempted_readers_exp(rnp); |
335 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ | |
12f5f524 | 336 | np = rcu_next_node_entry(t, rnp); |
f41d911f | 337 | list_del_init(&t->rcu_node_entry); |
12f5f524 PM |
338 | if (&t->rcu_node_entry == rnp->gp_tasks) |
339 | rnp->gp_tasks = np; | |
340 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
341 | rnp->exp_tasks = np; | |
27f4d280 PM |
342 | #ifdef CONFIG_RCU_BOOST |
343 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
344 | rnp->boost_tasks = np; | |
345 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
dd5d19ba | 346 | t->rcu_blocked_node = NULL; |
f41d911f PM |
347 | |
348 | /* | |
349 | * If this was the last task on the current list, and if | |
350 | * we aren't waiting on any CPUs, report the quiescent state. | |
d3f6bad3 | 351 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock. |
f41d911f | 352 | */ |
b668c9cf | 353 | if (empty) |
1304afb2 | 354 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf | 355 | else |
d3f6bad3 | 356 | rcu_report_unblock_qs_rnp(rnp, flags); |
d9a3da06 | 357 | |
27f4d280 PM |
358 | #ifdef CONFIG_RCU_BOOST |
359 | /* Unboost if we were boosted. */ | |
360 | if (special & RCU_READ_UNLOCK_BOOSTED) { | |
361 | t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BOOSTED; | |
362 | rt_mutex_unlock(t->rcu_boost_mutex); | |
363 | t->rcu_boost_mutex = NULL; | |
364 | } | |
365 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
366 | ||
d9a3da06 PM |
367 | /* |
368 | * If this was the last task on the expedited lists, | |
369 | * then we need to report up the rcu_node hierarchy. | |
370 | */ | |
371 | if (!empty_exp && !rcu_preempted_readers_exp(rnp)) | |
372 | rcu_report_exp_rnp(&rcu_preempt_state, rnp); | |
b668c9cf PM |
373 | } else { |
374 | local_irq_restore(flags); | |
f41d911f | 375 | } |
f41d911f PM |
376 | } |
377 | ||
378 | /* | |
6cc68793 | 379 | * Tree-preemptible RCU implementation for rcu_read_unlock(). |
f41d911f PM |
380 | * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost |
381 | * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then | |
382 | * invoke rcu_read_unlock_special() to clean up after a context switch | |
383 | * in an RCU read-side critical section and other special cases. | |
384 | */ | |
385 | void __rcu_read_unlock(void) | |
386 | { | |
387 | struct task_struct *t = current; | |
388 | ||
389 | barrier(); /* needed if we ever invoke rcu_read_unlock in rcutree.c */ | |
80dcf60e PM |
390 | --t->rcu_read_lock_nesting; |
391 | barrier(); /* decrement before load of ->rcu_read_unlock_special */ | |
392 | if (t->rcu_read_lock_nesting == 0 && | |
f41d911f PM |
393 | unlikely(ACCESS_ONCE(t->rcu_read_unlock_special))) |
394 | rcu_read_unlock_special(t); | |
cba8244a PM |
395 | #ifdef CONFIG_PROVE_LOCKING |
396 | WARN_ON_ONCE(ACCESS_ONCE(t->rcu_read_lock_nesting) < 0); | |
397 | #endif /* #ifdef CONFIG_PROVE_LOCKING */ | |
f41d911f PM |
398 | } |
399 | EXPORT_SYMBOL_GPL(__rcu_read_unlock); | |
400 | ||
1ed509a2 PM |
401 | #ifdef CONFIG_RCU_CPU_STALL_VERBOSE |
402 | ||
403 | /* | |
404 | * Dump detailed information for all tasks blocking the current RCU | |
405 | * grace period on the specified rcu_node structure. | |
406 | */ | |
407 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) | |
408 | { | |
409 | unsigned long flags; | |
1ed509a2 PM |
410 | struct task_struct *t; |
411 | ||
27f4d280 | 412 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
12f5f524 PM |
413 | return; |
414 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
415 | t = list_entry(rnp->gp_tasks, | |
416 | struct task_struct, rcu_node_entry); | |
417 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) | |
418 | sched_show_task(t); | |
419 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1ed509a2 PM |
420 | } |
421 | ||
422 | /* | |
423 | * Dump detailed information for all tasks blocking the current RCU | |
424 | * grace period. | |
425 | */ | |
426 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
427 | { | |
428 | struct rcu_node *rnp = rcu_get_root(rsp); | |
429 | ||
430 | rcu_print_detail_task_stall_rnp(rnp); | |
431 | rcu_for_each_leaf_node(rsp, rnp) | |
432 | rcu_print_detail_task_stall_rnp(rnp); | |
433 | } | |
434 | ||
435 | #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ | |
436 | ||
437 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
438 | { | |
439 | } | |
440 | ||
441 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ | |
442 | ||
f41d911f PM |
443 | /* |
444 | * Scan the current list of tasks blocked within RCU read-side critical | |
445 | * sections, printing out the tid of each. | |
446 | */ | |
447 | static void rcu_print_task_stall(struct rcu_node *rnp) | |
448 | { | |
f41d911f PM |
449 | struct task_struct *t; |
450 | ||
27f4d280 | 451 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
12f5f524 PM |
452 | return; |
453 | t = list_entry(rnp->gp_tasks, | |
454 | struct task_struct, rcu_node_entry); | |
455 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) | |
456 | printk(" P%d", t->pid); | |
f41d911f PM |
457 | } |
458 | ||
53d84e00 PM |
459 | /* |
460 | * Suppress preemptible RCU's CPU stall warnings by pushing the | |
461 | * time of the next stall-warning message comfortably far into the | |
462 | * future. | |
463 | */ | |
464 | static void rcu_preempt_stall_reset(void) | |
465 | { | |
466 | rcu_preempt_state.jiffies_stall = jiffies + ULONG_MAX / 2; | |
467 | } | |
468 | ||
b0e165c0 PM |
469 | /* |
470 | * Check that the list of blocked tasks for the newly completed grace | |
471 | * period is in fact empty. It is a serious bug to complete a grace | |
472 | * period that still has RCU readers blocked! This function must be | |
473 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock | |
474 | * must be held by the caller. | |
12f5f524 PM |
475 | * |
476 | * Also, if there are blocked tasks on the list, they automatically | |
477 | * block the newly created grace period, so set up ->gp_tasks accordingly. | |
b0e165c0 PM |
478 | */ |
479 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
480 | { | |
27f4d280 | 481 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
12f5f524 PM |
482 | if (!list_empty(&rnp->blkd_tasks)) |
483 | rnp->gp_tasks = rnp->blkd_tasks.next; | |
28ecd580 | 484 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
485 | } |
486 | ||
33f76148 PM |
487 | #ifdef CONFIG_HOTPLUG_CPU |
488 | ||
dd5d19ba PM |
489 | /* |
490 | * Handle tasklist migration for case in which all CPUs covered by the | |
491 | * specified rcu_node have gone offline. Move them up to the root | |
492 | * rcu_node. The reason for not just moving them to the immediate | |
493 | * parent is to remove the need for rcu_read_unlock_special() to | |
494 | * make more than two attempts to acquire the target rcu_node's lock. | |
b668c9cf PM |
495 | * Returns true if there were tasks blocking the current RCU grace |
496 | * period. | |
dd5d19ba | 497 | * |
237c80c5 PM |
498 | * Returns 1 if there was previously a task blocking the current grace |
499 | * period on the specified rcu_node structure. | |
500 | * | |
dd5d19ba PM |
501 | * The caller must hold rnp->lock with irqs disabled. |
502 | */ | |
237c80c5 PM |
503 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
504 | struct rcu_node *rnp, | |
505 | struct rcu_data *rdp) | |
dd5d19ba | 506 | { |
dd5d19ba PM |
507 | struct list_head *lp; |
508 | struct list_head *lp_root; | |
d9a3da06 | 509 | int retval = 0; |
dd5d19ba | 510 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
12f5f524 | 511 | struct task_struct *t; |
dd5d19ba | 512 | |
86848966 PM |
513 | if (rnp == rnp_root) { |
514 | WARN_ONCE(1, "Last CPU thought to be offlined?"); | |
237c80c5 | 515 | return 0; /* Shouldn't happen: at least one CPU online. */ |
86848966 | 516 | } |
12f5f524 PM |
517 | |
518 | /* If we are on an internal node, complain bitterly. */ | |
519 | WARN_ON_ONCE(rnp != rdp->mynode); | |
dd5d19ba PM |
520 | |
521 | /* | |
12f5f524 PM |
522 | * Move tasks up to root rcu_node. Don't try to get fancy for |
523 | * this corner-case operation -- just put this node's tasks | |
524 | * at the head of the root node's list, and update the root node's | |
525 | * ->gp_tasks and ->exp_tasks pointers to those of this node's, | |
526 | * if non-NULL. This might result in waiting for more tasks than | |
527 | * absolutely necessary, but this is a good performance/complexity | |
528 | * tradeoff. | |
dd5d19ba | 529 | */ |
27f4d280 | 530 | if (rcu_preempt_blocked_readers_cgp(rnp)) |
d9a3da06 PM |
531 | retval |= RCU_OFL_TASKS_NORM_GP; |
532 | if (rcu_preempted_readers_exp(rnp)) | |
533 | retval |= RCU_OFL_TASKS_EXP_GP; | |
12f5f524 PM |
534 | lp = &rnp->blkd_tasks; |
535 | lp_root = &rnp_root->blkd_tasks; | |
536 | while (!list_empty(lp)) { | |
537 | t = list_entry(lp->next, typeof(*t), rcu_node_entry); | |
538 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ | |
539 | list_del(&t->rcu_node_entry); | |
540 | t->rcu_blocked_node = rnp_root; | |
541 | list_add(&t->rcu_node_entry, lp_root); | |
542 | if (&t->rcu_node_entry == rnp->gp_tasks) | |
543 | rnp_root->gp_tasks = rnp->gp_tasks; | |
544 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
545 | rnp_root->exp_tasks = rnp->exp_tasks; | |
27f4d280 PM |
546 | #ifdef CONFIG_RCU_BOOST |
547 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
548 | rnp_root->boost_tasks = rnp->boost_tasks; | |
549 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
12f5f524 | 550 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ |
dd5d19ba | 551 | } |
27f4d280 PM |
552 | |
553 | #ifdef CONFIG_RCU_BOOST | |
554 | /* In case root is being boosted and leaf is not. */ | |
555 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ | |
556 | if (rnp_root->boost_tasks != NULL && | |
557 | rnp_root->boost_tasks != rnp_root->gp_tasks) | |
558 | rnp_root->boost_tasks = rnp_root->gp_tasks; | |
559 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ | |
560 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
561 | ||
12f5f524 PM |
562 | rnp->gp_tasks = NULL; |
563 | rnp->exp_tasks = NULL; | |
237c80c5 | 564 | return retval; |
dd5d19ba PM |
565 | } |
566 | ||
33f76148 | 567 | /* |
6cc68793 | 568 | * Do CPU-offline processing for preemptible RCU. |
33f76148 PM |
569 | */ |
570 | static void rcu_preempt_offline_cpu(int cpu) | |
571 | { | |
572 | __rcu_offline_cpu(cpu, &rcu_preempt_state); | |
573 | } | |
574 | ||
575 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
576 | ||
f41d911f PM |
577 | /* |
578 | * Check for a quiescent state from the current CPU. When a task blocks, | |
579 | * the task is recorded in the corresponding CPU's rcu_node structure, | |
580 | * which is checked elsewhere. | |
581 | * | |
582 | * Caller must disable hard irqs. | |
583 | */ | |
584 | static void rcu_preempt_check_callbacks(int cpu) | |
585 | { | |
586 | struct task_struct *t = current; | |
587 | ||
588 | if (t->rcu_read_lock_nesting == 0) { | |
c3422bea | 589 | rcu_preempt_qs(cpu); |
f41d911f PM |
590 | return; |
591 | } | |
a71fca58 | 592 | if (per_cpu(rcu_preempt_data, cpu).qs_pending) |
c3422bea | 593 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS; |
f41d911f PM |
594 | } |
595 | ||
596 | /* | |
6cc68793 | 597 | * Process callbacks for preemptible RCU. |
f41d911f PM |
598 | */ |
599 | static void rcu_preempt_process_callbacks(void) | |
600 | { | |
601 | __rcu_process_callbacks(&rcu_preempt_state, | |
602 | &__get_cpu_var(rcu_preempt_data)); | |
603 | } | |
604 | ||
a46e0899 PM |
605 | #ifdef CONFIG_RCU_BOOST |
606 | ||
09223371 SL |
607 | static void rcu_preempt_do_callbacks(void) |
608 | { | |
609 | rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data)); | |
610 | } | |
611 | ||
a46e0899 PM |
612 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
613 | ||
f41d911f | 614 | /* |
6cc68793 | 615 | * Queue a preemptible-RCU callback for invocation after a grace period. |
f41d911f PM |
616 | */ |
617 | void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
618 | { | |
619 | __call_rcu(head, func, &rcu_preempt_state); | |
620 | } | |
621 | EXPORT_SYMBOL_GPL(call_rcu); | |
622 | ||
6ebb237b PM |
623 | /** |
624 | * synchronize_rcu - wait until a grace period has elapsed. | |
625 | * | |
626 | * Control will return to the caller some time after a full grace | |
627 | * period has elapsed, in other words after all currently executing RCU | |
77d8485a PM |
628 | * read-side critical sections have completed. Note, however, that |
629 | * upon return from synchronize_rcu(), the caller might well be executing | |
630 | * concurrently with new RCU read-side critical sections that began while | |
631 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
632 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
6ebb237b PM |
633 | */ |
634 | void synchronize_rcu(void) | |
635 | { | |
636 | struct rcu_synchronize rcu; | |
637 | ||
638 | if (!rcu_scheduler_active) | |
639 | return; | |
640 | ||
72d5a9f7 | 641 | init_rcu_head_on_stack(&rcu.head); |
6ebb237b PM |
642 | init_completion(&rcu.completion); |
643 | /* Will wake me after RCU finished. */ | |
644 | call_rcu(&rcu.head, wakeme_after_rcu); | |
645 | /* Wait for it. */ | |
646 | wait_for_completion(&rcu.completion); | |
72d5a9f7 | 647 | destroy_rcu_head_on_stack(&rcu.head); |
6ebb237b PM |
648 | } |
649 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
650 | ||
d9a3da06 PM |
651 | static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq); |
652 | static long sync_rcu_preempt_exp_count; | |
653 | static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex); | |
654 | ||
655 | /* | |
656 | * Return non-zero if there are any tasks in RCU read-side critical | |
657 | * sections blocking the current preemptible-RCU expedited grace period. | |
658 | * If there is no preemptible-RCU expedited grace period currently in | |
659 | * progress, returns zero unconditionally. | |
660 | */ | |
661 | static int rcu_preempted_readers_exp(struct rcu_node *rnp) | |
662 | { | |
12f5f524 | 663 | return rnp->exp_tasks != NULL; |
d9a3da06 PM |
664 | } |
665 | ||
666 | /* | |
667 | * return non-zero if there is no RCU expedited grace period in progress | |
668 | * for the specified rcu_node structure, in other words, if all CPUs and | |
669 | * tasks covered by the specified rcu_node structure have done their bit | |
670 | * for the current expedited grace period. Works only for preemptible | |
671 | * RCU -- other RCU implementation use other means. | |
672 | * | |
673 | * Caller must hold sync_rcu_preempt_exp_mutex. | |
674 | */ | |
675 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) | |
676 | { | |
677 | return !rcu_preempted_readers_exp(rnp) && | |
678 | ACCESS_ONCE(rnp->expmask) == 0; | |
679 | } | |
680 | ||
681 | /* | |
682 | * Report the exit from RCU read-side critical section for the last task | |
683 | * that queued itself during or before the current expedited preemptible-RCU | |
684 | * grace period. This event is reported either to the rcu_node structure on | |
685 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
686 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
687 | * iteratively!) | |
688 | * | |
689 | * Caller must hold sync_rcu_preempt_exp_mutex. | |
690 | */ | |
691 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp) | |
692 | { | |
693 | unsigned long flags; | |
694 | unsigned long mask; | |
695 | ||
1304afb2 | 696 | raw_spin_lock_irqsave(&rnp->lock, flags); |
d9a3da06 | 697 | for (;;) { |
131906b0 PM |
698 | if (!sync_rcu_preempt_exp_done(rnp)) { |
699 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
d9a3da06 | 700 | break; |
131906b0 | 701 | } |
d9a3da06 | 702 | if (rnp->parent == NULL) { |
131906b0 | 703 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
d9a3da06 PM |
704 | wake_up(&sync_rcu_preempt_exp_wq); |
705 | break; | |
706 | } | |
707 | mask = rnp->grpmask; | |
1304afb2 | 708 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
d9a3da06 | 709 | rnp = rnp->parent; |
1304afb2 | 710 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
d9a3da06 PM |
711 | rnp->expmask &= ~mask; |
712 | } | |
d9a3da06 PM |
713 | } |
714 | ||
715 | /* | |
716 | * Snapshot the tasks blocking the newly started preemptible-RCU expedited | |
717 | * grace period for the specified rcu_node structure. If there are no such | |
718 | * tasks, report it up the rcu_node hierarchy. | |
719 | * | |
720 | * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock. | |
721 | */ | |
722 | static void | |
723 | sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp) | |
724 | { | |
1217ed1b | 725 | unsigned long flags; |
12f5f524 | 726 | int must_wait = 0; |
d9a3da06 | 727 | |
1217ed1b PM |
728 | raw_spin_lock_irqsave(&rnp->lock, flags); |
729 | if (list_empty(&rnp->blkd_tasks)) | |
730 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
731 | else { | |
12f5f524 | 732 | rnp->exp_tasks = rnp->blkd_tasks.next; |
1217ed1b | 733 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
12f5f524 PM |
734 | must_wait = 1; |
735 | } | |
d9a3da06 PM |
736 | if (!must_wait) |
737 | rcu_report_exp_rnp(rsp, rnp); | |
738 | } | |
739 | ||
019129d5 | 740 | /* |
d9a3da06 PM |
741 | * Wait for an rcu-preempt grace period, but expedite it. The basic idea |
742 | * is to invoke synchronize_sched_expedited() to push all the tasks to | |
12f5f524 | 743 | * the ->blkd_tasks lists and wait for this list to drain. |
019129d5 PM |
744 | */ |
745 | void synchronize_rcu_expedited(void) | |
746 | { | |
d9a3da06 PM |
747 | unsigned long flags; |
748 | struct rcu_node *rnp; | |
749 | struct rcu_state *rsp = &rcu_preempt_state; | |
750 | long snap; | |
751 | int trycount = 0; | |
752 | ||
753 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
754 | snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1; | |
755 | smp_mb(); /* Above access cannot bleed into critical section. */ | |
756 | ||
757 | /* | |
758 | * Acquire lock, falling back to synchronize_rcu() if too many | |
759 | * lock-acquisition failures. Of course, if someone does the | |
760 | * expedited grace period for us, just leave. | |
761 | */ | |
762 | while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) { | |
763 | if (trycount++ < 10) | |
764 | udelay(trycount * num_online_cpus()); | |
765 | else { | |
766 | synchronize_rcu(); | |
767 | return; | |
768 | } | |
769 | if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0) | |
770 | goto mb_ret; /* Others did our work for us. */ | |
771 | } | |
772 | if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0) | |
773 | goto unlock_mb_ret; /* Others did our work for us. */ | |
774 | ||
12f5f524 | 775 | /* force all RCU readers onto ->blkd_tasks lists. */ |
d9a3da06 PM |
776 | synchronize_sched_expedited(); |
777 | ||
1304afb2 | 778 | raw_spin_lock_irqsave(&rsp->onofflock, flags); |
d9a3da06 PM |
779 | |
780 | /* Initialize ->expmask for all non-leaf rcu_node structures. */ | |
781 | rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) { | |
1304afb2 | 782 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
d9a3da06 | 783 | rnp->expmask = rnp->qsmaskinit; |
1304afb2 | 784 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
d9a3da06 PM |
785 | } |
786 | ||
12f5f524 | 787 | /* Snapshot current state of ->blkd_tasks lists. */ |
d9a3da06 PM |
788 | rcu_for_each_leaf_node(rsp, rnp) |
789 | sync_rcu_preempt_exp_init(rsp, rnp); | |
790 | if (NUM_RCU_NODES > 1) | |
791 | sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp)); | |
792 | ||
1304afb2 | 793 | raw_spin_unlock_irqrestore(&rsp->onofflock, flags); |
d9a3da06 | 794 | |
12f5f524 | 795 | /* Wait for snapshotted ->blkd_tasks lists to drain. */ |
d9a3da06 PM |
796 | rnp = rcu_get_root(rsp); |
797 | wait_event(sync_rcu_preempt_exp_wq, | |
798 | sync_rcu_preempt_exp_done(rnp)); | |
799 | ||
800 | /* Clean up and exit. */ | |
801 | smp_mb(); /* ensure expedited GP seen before counter increment. */ | |
802 | ACCESS_ONCE(sync_rcu_preempt_exp_count)++; | |
803 | unlock_mb_ret: | |
804 | mutex_unlock(&sync_rcu_preempt_exp_mutex); | |
805 | mb_ret: | |
806 | smp_mb(); /* ensure subsequent action seen after grace period. */ | |
019129d5 PM |
807 | } |
808 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
809 | ||
f41d911f | 810 | /* |
6cc68793 | 811 | * Check to see if there is any immediate preemptible-RCU-related work |
f41d911f PM |
812 | * to be done. |
813 | */ | |
814 | static int rcu_preempt_pending(int cpu) | |
815 | { | |
816 | return __rcu_pending(&rcu_preempt_state, | |
817 | &per_cpu(rcu_preempt_data, cpu)); | |
818 | } | |
819 | ||
820 | /* | |
6cc68793 | 821 | * Does preemptible RCU need the CPU to stay out of dynticks mode? |
f41d911f PM |
822 | */ |
823 | static int rcu_preempt_needs_cpu(int cpu) | |
824 | { | |
825 | return !!per_cpu(rcu_preempt_data, cpu).nxtlist; | |
826 | } | |
827 | ||
e74f4c45 PM |
828 | /** |
829 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
830 | */ | |
831 | void rcu_barrier(void) | |
832 | { | |
833 | _rcu_barrier(&rcu_preempt_state, call_rcu); | |
834 | } | |
835 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
836 | ||
f41d911f | 837 | /* |
6cc68793 | 838 | * Initialize preemptible RCU's per-CPU data. |
f41d911f PM |
839 | */ |
840 | static void __cpuinit rcu_preempt_init_percpu_data(int cpu) | |
841 | { | |
842 | rcu_init_percpu_data(cpu, &rcu_preempt_state, 1); | |
843 | } | |
844 | ||
e74f4c45 | 845 | /* |
6cc68793 | 846 | * Move preemptible RCU's callbacks from dying CPU to other online CPU. |
e74f4c45 | 847 | */ |
29494be7 | 848 | static void rcu_preempt_send_cbs_to_online(void) |
e74f4c45 | 849 | { |
29494be7 | 850 | rcu_send_cbs_to_online(&rcu_preempt_state); |
e74f4c45 PM |
851 | } |
852 | ||
1eba8f84 | 853 | /* |
6cc68793 | 854 | * Initialize preemptible RCU's state structures. |
1eba8f84 PM |
855 | */ |
856 | static void __init __rcu_init_preempt(void) | |
857 | { | |
394f99a9 | 858 | rcu_init_one(&rcu_preempt_state, &rcu_preempt_data); |
1eba8f84 PM |
859 | } |
860 | ||
f41d911f | 861 | /* |
6cc68793 | 862 | * Check for a task exiting while in a preemptible-RCU read-side |
f41d911f PM |
863 | * critical section, clean up if so. No need to issue warnings, |
864 | * as debug_check_no_locks_held() already does this if lockdep | |
865 | * is enabled. | |
866 | */ | |
867 | void exit_rcu(void) | |
868 | { | |
869 | struct task_struct *t = current; | |
870 | ||
871 | if (t->rcu_read_lock_nesting == 0) | |
872 | return; | |
873 | t->rcu_read_lock_nesting = 1; | |
13491a0e | 874 | __rcu_read_unlock(); |
f41d911f PM |
875 | } |
876 | ||
877 | #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */ | |
878 | ||
27f4d280 PM |
879 | static struct rcu_state *rcu_state = &rcu_sched_state; |
880 | ||
f41d911f PM |
881 | /* |
882 | * Tell them what RCU they are running. | |
883 | */ | |
0e0fc1c2 | 884 | static void __init rcu_bootup_announce(void) |
f41d911f PM |
885 | { |
886 | printk(KERN_INFO "Hierarchical RCU implementation.\n"); | |
26845c28 | 887 | rcu_bootup_announce_oddness(); |
f41d911f PM |
888 | } |
889 | ||
890 | /* | |
891 | * Return the number of RCU batches processed thus far for debug & stats. | |
892 | */ | |
893 | long rcu_batches_completed(void) | |
894 | { | |
895 | return rcu_batches_completed_sched(); | |
896 | } | |
897 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
898 | ||
bf66f18e PM |
899 | /* |
900 | * Force a quiescent state for RCU, which, because there is no preemptible | |
901 | * RCU, becomes the same as rcu-sched. | |
902 | */ | |
903 | void rcu_force_quiescent_state(void) | |
904 | { | |
905 | rcu_sched_force_quiescent_state(); | |
906 | } | |
907 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
908 | ||
f41d911f | 909 | /* |
6cc68793 | 910 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
911 | * CPUs being in quiescent states. |
912 | */ | |
c3422bea | 913 | static void rcu_preempt_note_context_switch(int cpu) |
f41d911f PM |
914 | { |
915 | } | |
916 | ||
fc2219d4 | 917 | /* |
6cc68793 | 918 | * Because preemptible RCU does not exist, there are never any preempted |
fc2219d4 PM |
919 | * RCU readers. |
920 | */ | |
27f4d280 | 921 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 PM |
922 | { |
923 | return 0; | |
924 | } | |
925 | ||
b668c9cf PM |
926 | #ifdef CONFIG_HOTPLUG_CPU |
927 | ||
928 | /* Because preemptible RCU does not exist, no quieting of tasks. */ | |
d3f6bad3 | 929 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
b668c9cf | 930 | { |
1304afb2 | 931 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf PM |
932 | } |
933 | ||
934 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
935 | ||
1ed509a2 | 936 | /* |
6cc68793 | 937 | * Because preemptible RCU does not exist, we never have to check for |
1ed509a2 PM |
938 | * tasks blocked within RCU read-side critical sections. |
939 | */ | |
940 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
941 | { | |
942 | } | |
943 | ||
f41d911f | 944 | /* |
6cc68793 | 945 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
946 | * tasks blocked within RCU read-side critical sections. |
947 | */ | |
948 | static void rcu_print_task_stall(struct rcu_node *rnp) | |
949 | { | |
950 | } | |
951 | ||
53d84e00 PM |
952 | /* |
953 | * Because preemptible RCU does not exist, there is no need to suppress | |
954 | * its CPU stall warnings. | |
955 | */ | |
956 | static void rcu_preempt_stall_reset(void) | |
957 | { | |
958 | } | |
959 | ||
b0e165c0 | 960 | /* |
6cc68793 | 961 | * Because there is no preemptible RCU, there can be no readers blocked, |
49e29126 PM |
962 | * so there is no need to check for blocked tasks. So check only for |
963 | * bogus qsmask values. | |
b0e165c0 PM |
964 | */ |
965 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
966 | { | |
49e29126 | 967 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
968 | } |
969 | ||
33f76148 PM |
970 | #ifdef CONFIG_HOTPLUG_CPU |
971 | ||
dd5d19ba | 972 | /* |
6cc68793 | 973 | * Because preemptible RCU does not exist, it never needs to migrate |
237c80c5 PM |
974 | * tasks that were blocked within RCU read-side critical sections, and |
975 | * such non-existent tasks cannot possibly have been blocking the current | |
976 | * grace period. | |
dd5d19ba | 977 | */ |
237c80c5 PM |
978 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
979 | struct rcu_node *rnp, | |
980 | struct rcu_data *rdp) | |
dd5d19ba | 981 | { |
237c80c5 | 982 | return 0; |
dd5d19ba PM |
983 | } |
984 | ||
33f76148 | 985 | /* |
6cc68793 | 986 | * Because preemptible RCU does not exist, it never needs CPU-offline |
33f76148 PM |
987 | * processing. |
988 | */ | |
989 | static void rcu_preempt_offline_cpu(int cpu) | |
990 | { | |
991 | } | |
992 | ||
993 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
994 | ||
f41d911f | 995 | /* |
6cc68793 | 996 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
997 | * to check. |
998 | */ | |
1eba8f84 | 999 | static void rcu_preempt_check_callbacks(int cpu) |
f41d911f PM |
1000 | { |
1001 | } | |
1002 | ||
1003 | /* | |
6cc68793 | 1004 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
1005 | * to process. |
1006 | */ | |
1eba8f84 | 1007 | static void rcu_preempt_process_callbacks(void) |
f41d911f PM |
1008 | { |
1009 | } | |
1010 | ||
019129d5 PM |
1011 | /* |
1012 | * Wait for an rcu-preempt grace period, but make it happen quickly. | |
6cc68793 | 1013 | * But because preemptible RCU does not exist, map to rcu-sched. |
019129d5 PM |
1014 | */ |
1015 | void synchronize_rcu_expedited(void) | |
1016 | { | |
1017 | synchronize_sched_expedited(); | |
1018 | } | |
1019 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
1020 | ||
d9a3da06 PM |
1021 | #ifdef CONFIG_HOTPLUG_CPU |
1022 | ||
1023 | /* | |
6cc68793 | 1024 | * Because preemptible RCU does not exist, there is never any need to |
d9a3da06 PM |
1025 | * report on tasks preempted in RCU read-side critical sections during |
1026 | * expedited RCU grace periods. | |
1027 | */ | |
1028 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp) | |
1029 | { | |
1030 | return; | |
1031 | } | |
1032 | ||
1033 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1034 | ||
f41d911f | 1035 | /* |
6cc68793 | 1036 | * Because preemptible RCU does not exist, it never has any work to do. |
f41d911f PM |
1037 | */ |
1038 | static int rcu_preempt_pending(int cpu) | |
1039 | { | |
1040 | return 0; | |
1041 | } | |
1042 | ||
1043 | /* | |
6cc68793 | 1044 | * Because preemptible RCU does not exist, it never needs any CPU. |
f41d911f PM |
1045 | */ |
1046 | static int rcu_preempt_needs_cpu(int cpu) | |
1047 | { | |
1048 | return 0; | |
1049 | } | |
1050 | ||
e74f4c45 | 1051 | /* |
6cc68793 | 1052 | * Because preemptible RCU does not exist, rcu_barrier() is just |
e74f4c45 PM |
1053 | * another name for rcu_barrier_sched(). |
1054 | */ | |
1055 | void rcu_barrier(void) | |
1056 | { | |
1057 | rcu_barrier_sched(); | |
1058 | } | |
1059 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
1060 | ||
f41d911f | 1061 | /* |
6cc68793 | 1062 | * Because preemptible RCU does not exist, there is no per-CPU |
f41d911f PM |
1063 | * data to initialize. |
1064 | */ | |
1065 | static void __cpuinit rcu_preempt_init_percpu_data(int cpu) | |
1066 | { | |
1067 | } | |
1068 | ||
e74f4c45 | 1069 | /* |
6cc68793 | 1070 | * Because there is no preemptible RCU, there are no callbacks to move. |
e74f4c45 | 1071 | */ |
29494be7 | 1072 | static void rcu_preempt_send_cbs_to_online(void) |
e74f4c45 PM |
1073 | { |
1074 | } | |
1075 | ||
1eba8f84 | 1076 | /* |
6cc68793 | 1077 | * Because preemptible RCU does not exist, it need not be initialized. |
1eba8f84 PM |
1078 | */ |
1079 | static void __init __rcu_init_preempt(void) | |
1080 | { | |
1081 | } | |
1082 | ||
f41d911f | 1083 | #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */ |
8bd93a2c | 1084 | |
27f4d280 PM |
1085 | #ifdef CONFIG_RCU_BOOST |
1086 | ||
1087 | #include "rtmutex_common.h" | |
1088 | ||
0ea1f2eb PM |
1089 | #ifdef CONFIG_RCU_TRACE |
1090 | ||
1091 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
1092 | { | |
1093 | if (list_empty(&rnp->blkd_tasks)) | |
1094 | rnp->n_balk_blkd_tasks++; | |
1095 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) | |
1096 | rnp->n_balk_exp_gp_tasks++; | |
1097 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) | |
1098 | rnp->n_balk_boost_tasks++; | |
1099 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) | |
1100 | rnp->n_balk_notblocked++; | |
1101 | else if (rnp->gp_tasks != NULL && | |
a9f4793d | 1102 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
0ea1f2eb PM |
1103 | rnp->n_balk_notyet++; |
1104 | else | |
1105 | rnp->n_balk_nos++; | |
1106 | } | |
1107 | ||
1108 | #else /* #ifdef CONFIG_RCU_TRACE */ | |
1109 | ||
1110 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
1111 | { | |
1112 | } | |
1113 | ||
1114 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ | |
1115 | ||
27f4d280 PM |
1116 | /* |
1117 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks | |
1118 | * or ->boost_tasks, advancing the pointer to the next task in the | |
1119 | * ->blkd_tasks list. | |
1120 | * | |
1121 | * Note that irqs must be enabled: boosting the task can block. | |
1122 | * Returns 1 if there are more tasks needing to be boosted. | |
1123 | */ | |
1124 | static int rcu_boost(struct rcu_node *rnp) | |
1125 | { | |
1126 | unsigned long flags; | |
1127 | struct rt_mutex mtx; | |
1128 | struct task_struct *t; | |
1129 | struct list_head *tb; | |
1130 | ||
1131 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) | |
1132 | return 0; /* Nothing left to boost. */ | |
1133 | ||
1134 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1135 | ||
1136 | /* | |
1137 | * Recheck under the lock: all tasks in need of boosting | |
1138 | * might exit their RCU read-side critical sections on their own. | |
1139 | */ | |
1140 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { | |
1141 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1142 | return 0; | |
1143 | } | |
1144 | ||
1145 | /* | |
1146 | * Preferentially boost tasks blocking expedited grace periods. | |
1147 | * This cannot starve the normal grace periods because a second | |
1148 | * expedited grace period must boost all blocked tasks, including | |
1149 | * those blocking the pre-existing normal grace period. | |
1150 | */ | |
0ea1f2eb | 1151 | if (rnp->exp_tasks != NULL) { |
27f4d280 | 1152 | tb = rnp->exp_tasks; |
0ea1f2eb PM |
1153 | rnp->n_exp_boosts++; |
1154 | } else { | |
27f4d280 | 1155 | tb = rnp->boost_tasks; |
0ea1f2eb PM |
1156 | rnp->n_normal_boosts++; |
1157 | } | |
1158 | rnp->n_tasks_boosted++; | |
27f4d280 PM |
1159 | |
1160 | /* | |
1161 | * We boost task t by manufacturing an rt_mutex that appears to | |
1162 | * be held by task t. We leave a pointer to that rt_mutex where | |
1163 | * task t can find it, and task t will release the mutex when it | |
1164 | * exits its outermost RCU read-side critical section. Then | |
1165 | * simply acquiring this artificial rt_mutex will boost task | |
1166 | * t's priority. (Thanks to tglx for suggesting this approach!) | |
1167 | * | |
1168 | * Note that task t must acquire rnp->lock to remove itself from | |
1169 | * the ->blkd_tasks list, which it will do from exit() if from | |
1170 | * nowhere else. We therefore are guaranteed that task t will | |
1171 | * stay around at least until we drop rnp->lock. Note that | |
1172 | * rnp->lock also resolves races between our priority boosting | |
1173 | * and task t's exiting its outermost RCU read-side critical | |
1174 | * section. | |
1175 | */ | |
1176 | t = container_of(tb, struct task_struct, rcu_node_entry); | |
1177 | rt_mutex_init_proxy_locked(&mtx, t); | |
1178 | t->rcu_boost_mutex = &mtx; | |
1179 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BOOSTED; | |
1180 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1181 | rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */ | |
1182 | rt_mutex_unlock(&mtx); /* Keep lockdep happy. */ | |
1183 | ||
1184 | return rnp->exp_tasks != NULL || rnp->boost_tasks != NULL; | |
1185 | } | |
1186 | ||
1187 | /* | |
1188 | * Timer handler to initiate waking up of boost kthreads that | |
1189 | * have yielded the CPU due to excessive numbers of tasks to | |
1190 | * boost. We wake up the per-rcu_node kthread, which in turn | |
1191 | * will wake up the booster kthread. | |
1192 | */ | |
1193 | static void rcu_boost_kthread_timer(unsigned long arg) | |
1194 | { | |
1217ed1b | 1195 | invoke_rcu_node_kthread((struct rcu_node *)arg); |
27f4d280 PM |
1196 | } |
1197 | ||
1198 | /* | |
1199 | * Priority-boosting kthread. One per leaf rcu_node and one for the | |
1200 | * root rcu_node. | |
1201 | */ | |
1202 | static int rcu_boost_kthread(void *arg) | |
1203 | { | |
1204 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
1205 | int spincnt = 0; | |
1206 | int more2boost; | |
1207 | ||
1208 | for (;;) { | |
d71df90e | 1209 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
08bca60a | 1210 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
d71df90e | 1211 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
27f4d280 PM |
1212 | more2boost = rcu_boost(rnp); |
1213 | if (more2boost) | |
1214 | spincnt++; | |
1215 | else | |
1216 | spincnt = 0; | |
1217 | if (spincnt > 10) { | |
1218 | rcu_yield(rcu_boost_kthread_timer, (unsigned long)rnp); | |
1219 | spincnt = 0; | |
1220 | } | |
1221 | } | |
1217ed1b | 1222 | /* NOTREACHED */ |
27f4d280 PM |
1223 | return 0; |
1224 | } | |
1225 | ||
1226 | /* | |
1227 | * Check to see if it is time to start boosting RCU readers that are | |
1228 | * blocking the current grace period, and, if so, tell the per-rcu_node | |
1229 | * kthread to start boosting them. If there is an expedited grace | |
1230 | * period in progress, it is always time to boost. | |
1231 | * | |
1217ed1b PM |
1232 | * The caller must hold rnp->lock, which this function releases, |
1233 | * but irqs remain disabled. The ->boost_kthread_task is immortal, | |
1234 | * so we don't need to worry about it going away. | |
27f4d280 | 1235 | */ |
1217ed1b | 1236 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
27f4d280 PM |
1237 | { |
1238 | struct task_struct *t; | |
1239 | ||
0ea1f2eb PM |
1240 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
1241 | rnp->n_balk_exp_gp_tasks++; | |
1217ed1b | 1242 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 | 1243 | return; |
0ea1f2eb | 1244 | } |
27f4d280 PM |
1245 | if (rnp->exp_tasks != NULL || |
1246 | (rnp->gp_tasks != NULL && | |
1247 | rnp->boost_tasks == NULL && | |
1248 | rnp->qsmask == 0 && | |
1249 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { | |
1250 | if (rnp->exp_tasks == NULL) | |
1251 | rnp->boost_tasks = rnp->gp_tasks; | |
1217ed1b | 1252 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 PM |
1253 | t = rnp->boost_kthread_task; |
1254 | if (t != NULL) | |
1255 | wake_up_process(t); | |
1217ed1b | 1256 | } else { |
0ea1f2eb | 1257 | rcu_initiate_boost_trace(rnp); |
1217ed1b PM |
1258 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1259 | } | |
27f4d280 PM |
1260 | } |
1261 | ||
a46e0899 PM |
1262 | /* |
1263 | * Wake up the per-CPU kthread to invoke RCU callbacks. | |
1264 | */ | |
1265 | static void invoke_rcu_callbacks_kthread(void) | |
1266 | { | |
1267 | unsigned long flags; | |
1268 | ||
1269 | local_irq_save(flags); | |
1270 | __this_cpu_write(rcu_cpu_has_work, 1); | |
1271 | if (__this_cpu_read(rcu_cpu_kthread_task) == NULL) { | |
1272 | local_irq_restore(flags); | |
1273 | return; | |
1274 | } | |
1275 | wake_up_process(__this_cpu_read(rcu_cpu_kthread_task)); | |
1276 | local_irq_restore(flags); | |
1277 | } | |
1278 | ||
0f962a5e PM |
1279 | /* |
1280 | * Set the affinity of the boost kthread. The CPU-hotplug locks are | |
1281 | * held, so no one should be messing with the existence of the boost | |
1282 | * kthread. | |
1283 | */ | |
27f4d280 PM |
1284 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, |
1285 | cpumask_var_t cm) | |
1286 | { | |
27f4d280 PM |
1287 | struct task_struct *t; |
1288 | ||
27f4d280 PM |
1289 | t = rnp->boost_kthread_task; |
1290 | if (t != NULL) | |
1291 | set_cpus_allowed_ptr(rnp->boost_kthread_task, cm); | |
27f4d280 PM |
1292 | } |
1293 | ||
1294 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) | |
1295 | ||
1296 | /* | |
1297 | * Do priority-boost accounting for the start of a new grace period. | |
1298 | */ | |
1299 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) | |
1300 | { | |
1301 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; | |
1302 | } | |
1303 | ||
27f4d280 PM |
1304 | /* |
1305 | * Create an RCU-boost kthread for the specified node if one does not | |
1306 | * already exist. We only create this kthread for preemptible RCU. | |
1307 | * Returns zero if all is well, a negated errno otherwise. | |
1308 | */ | |
1309 | static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp, | |
1310 | struct rcu_node *rnp, | |
1311 | int rnp_index) | |
1312 | { | |
1313 | unsigned long flags; | |
1314 | struct sched_param sp; | |
1315 | struct task_struct *t; | |
1316 | ||
1317 | if (&rcu_preempt_state != rsp) | |
1318 | return 0; | |
a46e0899 | 1319 | rsp->boost = 1; |
27f4d280 PM |
1320 | if (rnp->boost_kthread_task != NULL) |
1321 | return 0; | |
1322 | t = kthread_create(rcu_boost_kthread, (void *)rnp, | |
1323 | "rcub%d", rnp_index); | |
1324 | if (IS_ERR(t)) | |
1325 | return PTR_ERR(t); | |
1326 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1327 | rnp->boost_kthread_task = t; | |
1328 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
27f4d280 PM |
1329 | sp.sched_priority = RCU_KTHREAD_PRIO; |
1330 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
9a432736 | 1331 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
27f4d280 PM |
1332 | return 0; |
1333 | } | |
1334 | ||
f8b7fc6b PM |
1335 | #ifdef CONFIG_HOTPLUG_CPU |
1336 | ||
1337 | /* | |
1338 | * Stop the RCU's per-CPU kthread when its CPU goes offline,. | |
1339 | */ | |
1340 | static void rcu_stop_cpu_kthread(int cpu) | |
1341 | { | |
1342 | struct task_struct *t; | |
1343 | ||
1344 | /* Stop the CPU's kthread. */ | |
1345 | t = per_cpu(rcu_cpu_kthread_task, cpu); | |
1346 | if (t != NULL) { | |
1347 | per_cpu(rcu_cpu_kthread_task, cpu) = NULL; | |
1348 | kthread_stop(t); | |
1349 | } | |
1350 | } | |
1351 | ||
1352 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1353 | ||
1354 | static void rcu_kthread_do_work(void) | |
1355 | { | |
1356 | rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data)); | |
1357 | rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data)); | |
1358 | rcu_preempt_do_callbacks(); | |
1359 | } | |
1360 | ||
1361 | /* | |
1362 | * Wake up the specified per-rcu_node-structure kthread. | |
1363 | * Because the per-rcu_node kthreads are immortal, we don't need | |
1364 | * to do anything to keep them alive. | |
1365 | */ | |
1366 | static void invoke_rcu_node_kthread(struct rcu_node *rnp) | |
1367 | { | |
1368 | struct task_struct *t; | |
1369 | ||
1370 | t = rnp->node_kthread_task; | |
1371 | if (t != NULL) | |
1372 | wake_up_process(t); | |
1373 | } | |
1374 | ||
1375 | /* | |
1376 | * Set the specified CPU's kthread to run RT or not, as specified by | |
1377 | * the to_rt argument. The CPU-hotplug locks are held, so the task | |
1378 | * is not going away. | |
1379 | */ | |
1380 | static void rcu_cpu_kthread_setrt(int cpu, int to_rt) | |
1381 | { | |
1382 | int policy; | |
1383 | struct sched_param sp; | |
1384 | struct task_struct *t; | |
1385 | ||
1386 | t = per_cpu(rcu_cpu_kthread_task, cpu); | |
1387 | if (t == NULL) | |
1388 | return; | |
1389 | if (to_rt) { | |
1390 | policy = SCHED_FIFO; | |
1391 | sp.sched_priority = RCU_KTHREAD_PRIO; | |
1392 | } else { | |
1393 | policy = SCHED_NORMAL; | |
1394 | sp.sched_priority = 0; | |
1395 | } | |
1396 | sched_setscheduler_nocheck(t, policy, &sp); | |
1397 | } | |
1398 | ||
1399 | /* | |
1400 | * Timer handler to initiate the waking up of per-CPU kthreads that | |
1401 | * have yielded the CPU due to excess numbers of RCU callbacks. | |
1402 | * We wake up the per-rcu_node kthread, which in turn will wake up | |
1403 | * the booster kthread. | |
1404 | */ | |
1405 | static void rcu_cpu_kthread_timer(unsigned long arg) | |
1406 | { | |
1407 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg); | |
1408 | struct rcu_node *rnp = rdp->mynode; | |
1409 | ||
1410 | atomic_or(rdp->grpmask, &rnp->wakemask); | |
1411 | invoke_rcu_node_kthread(rnp); | |
1412 | } | |
1413 | ||
1414 | /* | |
1415 | * Drop to non-real-time priority and yield, but only after posting a | |
1416 | * timer that will cause us to regain our real-time priority if we | |
1417 | * remain preempted. Either way, we restore our real-time priority | |
1418 | * before returning. | |
1419 | */ | |
1420 | static void rcu_yield(void (*f)(unsigned long), unsigned long arg) | |
1421 | { | |
1422 | struct sched_param sp; | |
1423 | struct timer_list yield_timer; | |
1424 | ||
1425 | setup_timer_on_stack(&yield_timer, f, arg); | |
1426 | mod_timer(&yield_timer, jiffies + 2); | |
1427 | sp.sched_priority = 0; | |
1428 | sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp); | |
1429 | set_user_nice(current, 19); | |
1430 | schedule(); | |
1431 | sp.sched_priority = RCU_KTHREAD_PRIO; | |
1432 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); | |
1433 | del_timer(&yield_timer); | |
1434 | } | |
1435 | ||
1436 | /* | |
1437 | * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU. | |
1438 | * This can happen while the corresponding CPU is either coming online | |
1439 | * or going offline. We cannot wait until the CPU is fully online | |
1440 | * before starting the kthread, because the various notifier functions | |
1441 | * can wait for RCU grace periods. So we park rcu_cpu_kthread() until | |
1442 | * the corresponding CPU is online. | |
1443 | * | |
1444 | * Return 1 if the kthread needs to stop, 0 otherwise. | |
1445 | * | |
1446 | * Caller must disable bh. This function can momentarily enable it. | |
1447 | */ | |
1448 | static int rcu_cpu_kthread_should_stop(int cpu) | |
1449 | { | |
1450 | while (cpu_is_offline(cpu) || | |
1451 | !cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)) || | |
1452 | smp_processor_id() != cpu) { | |
1453 | if (kthread_should_stop()) | |
1454 | return 1; | |
1455 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; | |
1456 | per_cpu(rcu_cpu_kthread_cpu, cpu) = raw_smp_processor_id(); | |
1457 | local_bh_enable(); | |
1458 | schedule_timeout_uninterruptible(1); | |
1459 | if (!cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu))) | |
1460 | set_cpus_allowed_ptr(current, cpumask_of(cpu)); | |
1461 | local_bh_disable(); | |
1462 | } | |
1463 | per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu; | |
1464 | return 0; | |
1465 | } | |
1466 | ||
1467 | /* | |
1468 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the | |
1469 | * earlier RCU softirq. | |
1470 | */ | |
1471 | static int rcu_cpu_kthread(void *arg) | |
1472 | { | |
1473 | int cpu = (int)(long)arg; | |
1474 | unsigned long flags; | |
1475 | int spincnt = 0; | |
1476 | unsigned int *statusp = &per_cpu(rcu_cpu_kthread_status, cpu); | |
1477 | char work; | |
1478 | char *workp = &per_cpu(rcu_cpu_has_work, cpu); | |
1479 | ||
1480 | for (;;) { | |
1481 | *statusp = RCU_KTHREAD_WAITING; | |
1482 | rcu_wait(*workp != 0 || kthread_should_stop()); | |
1483 | local_bh_disable(); | |
1484 | if (rcu_cpu_kthread_should_stop(cpu)) { | |
1485 | local_bh_enable(); | |
1486 | break; | |
1487 | } | |
1488 | *statusp = RCU_KTHREAD_RUNNING; | |
1489 | per_cpu(rcu_cpu_kthread_loops, cpu)++; | |
1490 | local_irq_save(flags); | |
1491 | work = *workp; | |
1492 | *workp = 0; | |
1493 | local_irq_restore(flags); | |
1494 | if (work) | |
1495 | rcu_kthread_do_work(); | |
1496 | local_bh_enable(); | |
1497 | if (*workp != 0) | |
1498 | spincnt++; | |
1499 | else | |
1500 | spincnt = 0; | |
1501 | if (spincnt > 10) { | |
1502 | *statusp = RCU_KTHREAD_YIELDING; | |
1503 | rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu); | |
1504 | spincnt = 0; | |
1505 | } | |
1506 | } | |
1507 | *statusp = RCU_KTHREAD_STOPPED; | |
1508 | return 0; | |
1509 | } | |
1510 | ||
1511 | /* | |
1512 | * Spawn a per-CPU kthread, setting up affinity and priority. | |
1513 | * Because the CPU hotplug lock is held, no other CPU will be attempting | |
1514 | * to manipulate rcu_cpu_kthread_task. There might be another CPU | |
1515 | * attempting to access it during boot, but the locking in kthread_bind() | |
1516 | * will enforce sufficient ordering. | |
1517 | * | |
1518 | * Please note that we cannot simply refuse to wake up the per-CPU | |
1519 | * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state, | |
1520 | * which can result in softlockup complaints if the task ends up being | |
1521 | * idle for more than a couple of minutes. | |
1522 | * | |
1523 | * However, please note also that we cannot bind the per-CPU kthread to its | |
1524 | * CPU until that CPU is fully online. We also cannot wait until the | |
1525 | * CPU is fully online before we create its per-CPU kthread, as this would | |
1526 | * deadlock the system when CPU notifiers tried waiting for grace | |
1527 | * periods. So we bind the per-CPU kthread to its CPU only if the CPU | |
1528 | * is online. If its CPU is not yet fully online, then the code in | |
1529 | * rcu_cpu_kthread() will wait until it is fully online, and then do | |
1530 | * the binding. | |
1531 | */ | |
1532 | static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu) | |
1533 | { | |
1534 | struct sched_param sp; | |
1535 | struct task_struct *t; | |
1536 | ||
b0d30417 | 1537 | if (!rcu_scheduler_fully_active || |
f8b7fc6b PM |
1538 | per_cpu(rcu_cpu_kthread_task, cpu) != NULL) |
1539 | return 0; | |
1540 | t = kthread_create(rcu_cpu_kthread, (void *)(long)cpu, "rcuc%d", cpu); | |
1541 | if (IS_ERR(t)) | |
1542 | return PTR_ERR(t); | |
1543 | if (cpu_online(cpu)) | |
1544 | kthread_bind(t, cpu); | |
1545 | per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu; | |
1546 | WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL); | |
1547 | sp.sched_priority = RCU_KTHREAD_PRIO; | |
1548 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
1549 | per_cpu(rcu_cpu_kthread_task, cpu) = t; | |
1550 | wake_up_process(t); /* Get to TASK_INTERRUPTIBLE quickly. */ | |
1551 | return 0; | |
1552 | } | |
1553 | ||
1554 | /* | |
1555 | * Per-rcu_node kthread, which is in charge of waking up the per-CPU | |
1556 | * kthreads when needed. We ignore requests to wake up kthreads | |
1557 | * for offline CPUs, which is OK because force_quiescent_state() | |
1558 | * takes care of this case. | |
1559 | */ | |
1560 | static int rcu_node_kthread(void *arg) | |
1561 | { | |
1562 | int cpu; | |
1563 | unsigned long flags; | |
1564 | unsigned long mask; | |
1565 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
1566 | struct sched_param sp; | |
1567 | struct task_struct *t; | |
1568 | ||
1569 | for (;;) { | |
1570 | rnp->node_kthread_status = RCU_KTHREAD_WAITING; | |
1571 | rcu_wait(atomic_read(&rnp->wakemask) != 0); | |
1572 | rnp->node_kthread_status = RCU_KTHREAD_RUNNING; | |
1573 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1574 | mask = atomic_xchg(&rnp->wakemask, 0); | |
1575 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ | |
1576 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) { | |
1577 | if ((mask & 0x1) == 0) | |
1578 | continue; | |
1579 | preempt_disable(); | |
1580 | t = per_cpu(rcu_cpu_kthread_task, cpu); | |
1581 | if (!cpu_online(cpu) || t == NULL) { | |
1582 | preempt_enable(); | |
1583 | continue; | |
1584 | } | |
1585 | per_cpu(rcu_cpu_has_work, cpu) = 1; | |
1586 | sp.sched_priority = RCU_KTHREAD_PRIO; | |
1587 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
1588 | preempt_enable(); | |
1589 | } | |
1590 | } | |
1591 | /* NOTREACHED */ | |
1592 | rnp->node_kthread_status = RCU_KTHREAD_STOPPED; | |
1593 | return 0; | |
1594 | } | |
1595 | ||
1596 | /* | |
1597 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are | |
1598 | * served by the rcu_node in question. The CPU hotplug lock is still | |
1599 | * held, so the value of rnp->qsmaskinit will be stable. | |
1600 | * | |
1601 | * We don't include outgoingcpu in the affinity set, use -1 if there is | |
1602 | * no outgoing CPU. If there are no CPUs left in the affinity set, | |
1603 | * this function allows the kthread to execute on any CPU. | |
1604 | */ | |
1605 | static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) | |
1606 | { | |
1607 | cpumask_var_t cm; | |
1608 | int cpu; | |
1609 | unsigned long mask = rnp->qsmaskinit; | |
1610 | ||
1611 | if (rnp->node_kthread_task == NULL) | |
1612 | return; | |
1613 | if (!alloc_cpumask_var(&cm, GFP_KERNEL)) | |
1614 | return; | |
1615 | cpumask_clear(cm); | |
1616 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) | |
1617 | if ((mask & 0x1) && cpu != outgoingcpu) | |
1618 | cpumask_set_cpu(cpu, cm); | |
1619 | if (cpumask_weight(cm) == 0) { | |
1620 | cpumask_setall(cm); | |
1621 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) | |
1622 | cpumask_clear_cpu(cpu, cm); | |
1623 | WARN_ON_ONCE(cpumask_weight(cm) == 0); | |
1624 | } | |
1625 | set_cpus_allowed_ptr(rnp->node_kthread_task, cm); | |
1626 | rcu_boost_kthread_setaffinity(rnp, cm); | |
1627 | free_cpumask_var(cm); | |
1628 | } | |
1629 | ||
1630 | /* | |
1631 | * Spawn a per-rcu_node kthread, setting priority and affinity. | |
1632 | * Called during boot before online/offline can happen, or, if | |
1633 | * during runtime, with the main CPU-hotplug locks held. So only | |
1634 | * one of these can be executing at a time. | |
1635 | */ | |
1636 | static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp, | |
1637 | struct rcu_node *rnp) | |
1638 | { | |
1639 | unsigned long flags; | |
1640 | int rnp_index = rnp - &rsp->node[0]; | |
1641 | struct sched_param sp; | |
1642 | struct task_struct *t; | |
1643 | ||
b0d30417 | 1644 | if (!rcu_scheduler_fully_active || |
f8b7fc6b PM |
1645 | rnp->qsmaskinit == 0) |
1646 | return 0; | |
1647 | if (rnp->node_kthread_task == NULL) { | |
1648 | t = kthread_create(rcu_node_kthread, (void *)rnp, | |
1649 | "rcun%d", rnp_index); | |
1650 | if (IS_ERR(t)) | |
1651 | return PTR_ERR(t); | |
1652 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1653 | rnp->node_kthread_task = t; | |
1654 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1655 | sp.sched_priority = 99; | |
1656 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
1657 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ | |
1658 | } | |
1659 | return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index); | |
1660 | } | |
1661 | ||
1662 | /* | |
1663 | * Spawn all kthreads -- called as soon as the scheduler is running. | |
1664 | */ | |
1665 | static int __init rcu_spawn_kthreads(void) | |
1666 | { | |
1667 | int cpu; | |
1668 | struct rcu_node *rnp; | |
1669 | ||
b0d30417 | 1670 | rcu_scheduler_fully_active = 1; |
f8b7fc6b PM |
1671 | for_each_possible_cpu(cpu) { |
1672 | per_cpu(rcu_cpu_has_work, cpu) = 0; | |
1673 | if (cpu_online(cpu)) | |
1674 | (void)rcu_spawn_one_cpu_kthread(cpu); | |
1675 | } | |
1676 | rnp = rcu_get_root(rcu_state); | |
1677 | (void)rcu_spawn_one_node_kthread(rcu_state, rnp); | |
1678 | if (NUM_RCU_NODES > 1) { | |
1679 | rcu_for_each_leaf_node(rcu_state, rnp) | |
1680 | (void)rcu_spawn_one_node_kthread(rcu_state, rnp); | |
1681 | } | |
1682 | return 0; | |
1683 | } | |
1684 | early_initcall(rcu_spawn_kthreads); | |
1685 | ||
1686 | static void __cpuinit rcu_prepare_kthreads(int cpu) | |
1687 | { | |
1688 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); | |
1689 | struct rcu_node *rnp = rdp->mynode; | |
1690 | ||
1691 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ | |
b0d30417 | 1692 | if (rcu_scheduler_fully_active) { |
f8b7fc6b PM |
1693 | (void)rcu_spawn_one_cpu_kthread(cpu); |
1694 | if (rnp->node_kthread_task == NULL) | |
1695 | (void)rcu_spawn_one_node_kthread(rcu_state, rnp); | |
1696 | } | |
1697 | } | |
1698 | ||
27f4d280 PM |
1699 | #else /* #ifdef CONFIG_RCU_BOOST */ |
1700 | ||
1217ed1b | 1701 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
27f4d280 | 1702 | { |
1217ed1b | 1703 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 PM |
1704 | } |
1705 | ||
a46e0899 | 1706 | static void invoke_rcu_callbacks_kthread(void) |
27f4d280 | 1707 | { |
a46e0899 | 1708 | WARN_ON_ONCE(1); |
27f4d280 PM |
1709 | } |
1710 | ||
1711 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) | |
1712 | { | |
1713 | } | |
1714 | ||
f8b7fc6b PM |
1715 | #ifdef CONFIG_HOTPLUG_CPU |
1716 | ||
1717 | static void rcu_stop_cpu_kthread(int cpu) | |
1718 | { | |
1719 | } | |
1720 | ||
1721 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1722 | ||
1723 | static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) | |
1724 | { | |
1725 | } | |
1726 | ||
1727 | static void rcu_cpu_kthread_setrt(int cpu, int to_rt) | |
1728 | { | |
1729 | } | |
1730 | ||
b0d30417 PM |
1731 | static int __init rcu_scheduler_really_started(void) |
1732 | { | |
1733 | rcu_scheduler_fully_active = 1; | |
1734 | return 0; | |
1735 | } | |
1736 | early_initcall(rcu_scheduler_really_started); | |
1737 | ||
f8b7fc6b PM |
1738 | static void __cpuinit rcu_prepare_kthreads(int cpu) |
1739 | { | |
1740 | } | |
1741 | ||
27f4d280 PM |
1742 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
1743 | ||
7b27d547 LJ |
1744 | #ifndef CONFIG_SMP |
1745 | ||
1746 | void synchronize_sched_expedited(void) | |
1747 | { | |
1748 | cond_resched(); | |
1749 | } | |
1750 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
1751 | ||
1752 | #else /* #ifndef CONFIG_SMP */ | |
1753 | ||
e27fc964 TH |
1754 | static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0); |
1755 | static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0); | |
7b27d547 LJ |
1756 | |
1757 | static int synchronize_sched_expedited_cpu_stop(void *data) | |
1758 | { | |
1759 | /* | |
1760 | * There must be a full memory barrier on each affected CPU | |
1761 | * between the time that try_stop_cpus() is called and the | |
1762 | * time that it returns. | |
1763 | * | |
1764 | * In the current initial implementation of cpu_stop, the | |
1765 | * above condition is already met when the control reaches | |
1766 | * this point and the following smp_mb() is not strictly | |
1767 | * necessary. Do smp_mb() anyway for documentation and | |
1768 | * robustness against future implementation changes. | |
1769 | */ | |
1770 | smp_mb(); /* See above comment block. */ | |
1771 | return 0; | |
1772 | } | |
1773 | ||
1774 | /* | |
1775 | * Wait for an rcu-sched grace period to elapse, but use "big hammer" | |
1776 | * approach to force grace period to end quickly. This consumes | |
1777 | * significant time on all CPUs, and is thus not recommended for | |
1778 | * any sort of common-case code. | |
1779 | * | |
1780 | * Note that it is illegal to call this function while holding any | |
1781 | * lock that is acquired by a CPU-hotplug notifier. Failing to | |
1782 | * observe this restriction will result in deadlock. | |
db3a8920 | 1783 | * |
e27fc964 TH |
1784 | * This implementation can be thought of as an application of ticket |
1785 | * locking to RCU, with sync_sched_expedited_started and | |
1786 | * sync_sched_expedited_done taking on the roles of the halves | |
1787 | * of the ticket-lock word. Each task atomically increments | |
1788 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
1789 | * then attempts to stop all the CPUs. If this succeeds, then each | |
1790 | * CPU will have executed a context switch, resulting in an RCU-sched | |
1791 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
1792 | * update sync_sched_expedited_done to match our snapshot -- but | |
1793 | * only if someone else has not already advanced past our snapshot. | |
1794 | * | |
1795 | * On the other hand, if try_stop_cpus() fails, we check the value | |
1796 | * of sync_sched_expedited_done. If it has advanced past our | |
1797 | * initial snapshot, then someone else must have forced a grace period | |
1798 | * some time after we took our snapshot. In this case, our work is | |
1799 | * done for us, and we can simply return. Otherwise, we try again, | |
1800 | * but keep our initial snapshot for purposes of checking for someone | |
1801 | * doing our work for us. | |
1802 | * | |
1803 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
7b27d547 LJ |
1804 | */ |
1805 | void synchronize_sched_expedited(void) | |
1806 | { | |
e27fc964 | 1807 | int firstsnap, s, snap, trycount = 0; |
7b27d547 | 1808 | |
e27fc964 TH |
1809 | /* Note that atomic_inc_return() implies full memory barrier. */ |
1810 | firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started); | |
7b27d547 | 1811 | get_online_cpus(); |
e27fc964 TH |
1812 | |
1813 | /* | |
1814 | * Each pass through the following loop attempts to force a | |
1815 | * context switch on each CPU. | |
1816 | */ | |
7b27d547 LJ |
1817 | while (try_stop_cpus(cpu_online_mask, |
1818 | synchronize_sched_expedited_cpu_stop, | |
1819 | NULL) == -EAGAIN) { | |
1820 | put_online_cpus(); | |
e27fc964 TH |
1821 | |
1822 | /* No joy, try again later. Or just synchronize_sched(). */ | |
7b27d547 LJ |
1823 | if (trycount++ < 10) |
1824 | udelay(trycount * num_online_cpus()); | |
1825 | else { | |
1826 | synchronize_sched(); | |
1827 | return; | |
1828 | } | |
e27fc964 TH |
1829 | |
1830 | /* Check to see if someone else did our work for us. */ | |
1831 | s = atomic_read(&sync_sched_expedited_done); | |
1832 | if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) { | |
7b27d547 LJ |
1833 | smp_mb(); /* ensure test happens before caller kfree */ |
1834 | return; | |
1835 | } | |
e27fc964 TH |
1836 | |
1837 | /* | |
1838 | * Refetching sync_sched_expedited_started allows later | |
1839 | * callers to piggyback on our grace period. We subtract | |
1840 | * 1 to get the same token that the last incrementer got. | |
1841 | * We retry after they started, so our grace period works | |
1842 | * for them, and they started after our first try, so their | |
1843 | * grace period works for us. | |
1844 | */ | |
7b27d547 | 1845 | get_online_cpus(); |
e27fc964 TH |
1846 | snap = atomic_read(&sync_sched_expedited_started) - 1; |
1847 | smp_mb(); /* ensure read is before try_stop_cpus(). */ | |
7b27d547 | 1848 | } |
e27fc964 TH |
1849 | |
1850 | /* | |
1851 | * Everyone up to our most recent fetch is covered by our grace | |
1852 | * period. Update the counter, but only if our work is still | |
1853 | * relevant -- which it won't be if someone who started later | |
1854 | * than we did beat us to the punch. | |
1855 | */ | |
1856 | do { | |
1857 | s = atomic_read(&sync_sched_expedited_done); | |
1858 | if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) { | |
1859 | smp_mb(); /* ensure test happens before caller kfree */ | |
1860 | break; | |
1861 | } | |
1862 | } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s); | |
1863 | ||
7b27d547 LJ |
1864 | put_online_cpus(); |
1865 | } | |
1866 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
1867 | ||
1868 | #endif /* #else #ifndef CONFIG_SMP */ | |
1869 | ||
8bd93a2c PM |
1870 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
1871 | ||
1872 | /* | |
1873 | * Check to see if any future RCU-related work will need to be done | |
1874 | * by the current CPU, even if none need be done immediately, returning | |
1875 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
1876 | * an exported member of the RCU API. | |
1877 | * | |
1878 | * Because we have preemptible RCU, just check whether this CPU needs | |
1879 | * any flavor of RCU. Do not chew up lots of CPU cycles with preemption | |
1880 | * disabled in a most-likely vain attempt to cause RCU not to need this CPU. | |
1881 | */ | |
1882 | int rcu_needs_cpu(int cpu) | |
1883 | { | |
1884 | return rcu_needs_cpu_quick_check(cpu); | |
1885 | } | |
1886 | ||
a47cd880 PM |
1887 | /* |
1888 | * Check to see if we need to continue a callback-flush operations to | |
1889 | * allow the last CPU to enter dyntick-idle mode. But fast dyntick-idle | |
1890 | * entry is not configured, so we never do need to. | |
1891 | */ | |
1892 | static void rcu_needs_cpu_flush(void) | |
1893 | { | |
1894 | } | |
1895 | ||
8bd93a2c PM |
1896 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
1897 | ||
1898 | #define RCU_NEEDS_CPU_FLUSHES 5 | |
a47cd880 | 1899 | static DEFINE_PER_CPU(int, rcu_dyntick_drain); |
71da8132 | 1900 | static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff); |
8bd93a2c PM |
1901 | |
1902 | /* | |
1903 | * Check to see if any future RCU-related work will need to be done | |
1904 | * by the current CPU, even if none need be done immediately, returning | |
1905 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
1906 | * an exported member of the RCU API. | |
1907 | * | |
1908 | * Because we are not supporting preemptible RCU, attempt to accelerate | |
1909 | * any current grace periods so that RCU no longer needs this CPU, but | |
1910 | * only if all other CPUs are already in dynticks-idle mode. This will | |
1911 | * allow the CPU cores to be powered down immediately, as opposed to after | |
1912 | * waiting many milliseconds for grace periods to elapse. | |
a47cd880 PM |
1913 | * |
1914 | * Because it is not legal to invoke rcu_process_callbacks() with irqs | |
1915 | * disabled, we do one pass of force_quiescent_state(), then do a | |
a46e0899 | 1916 | * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked |
27f4d280 | 1917 | * later. The per-cpu rcu_dyntick_drain variable controls the sequencing. |
8bd93a2c PM |
1918 | */ |
1919 | int rcu_needs_cpu(int cpu) | |
1920 | { | |
a47cd880 | 1921 | int c = 0; |
77e38ed3 | 1922 | int snap; |
8bd93a2c PM |
1923 | int thatcpu; |
1924 | ||
622ea685 PM |
1925 | /* Check for being in the holdoff period. */ |
1926 | if (per_cpu(rcu_dyntick_holdoff, cpu) == jiffies) | |
1927 | return rcu_needs_cpu_quick_check(cpu); | |
1928 | ||
8bd93a2c | 1929 | /* Don't bother unless we are the last non-dyntick-idle CPU. */ |
77e38ed3 PM |
1930 | for_each_online_cpu(thatcpu) { |
1931 | if (thatcpu == cpu) | |
1932 | continue; | |
23b5c8fa PM |
1933 | snap = atomic_add_return(0, &per_cpu(rcu_dynticks, |
1934 | thatcpu).dynticks); | |
77e38ed3 | 1935 | smp_mb(); /* Order sampling of snap with end of grace period. */ |
23b5c8fa | 1936 | if ((snap & 0x1) != 0) { |
a47cd880 | 1937 | per_cpu(rcu_dyntick_drain, cpu) = 0; |
71da8132 | 1938 | per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1; |
8bd93a2c | 1939 | return rcu_needs_cpu_quick_check(cpu); |
8bd93a2c | 1940 | } |
77e38ed3 | 1941 | } |
a47cd880 PM |
1942 | |
1943 | /* Check and update the rcu_dyntick_drain sequencing. */ | |
1944 | if (per_cpu(rcu_dyntick_drain, cpu) <= 0) { | |
1945 | /* First time through, initialize the counter. */ | |
1946 | per_cpu(rcu_dyntick_drain, cpu) = RCU_NEEDS_CPU_FLUSHES; | |
1947 | } else if (--per_cpu(rcu_dyntick_drain, cpu) <= 0) { | |
1948 | /* We have hit the limit, so time to give up. */ | |
71da8132 | 1949 | per_cpu(rcu_dyntick_holdoff, cpu) = jiffies; |
a47cd880 PM |
1950 | return rcu_needs_cpu_quick_check(cpu); |
1951 | } | |
1952 | ||
1953 | /* Do one step pushing remaining RCU callbacks through. */ | |
1954 | if (per_cpu(rcu_sched_data, cpu).nxtlist) { | |
1955 | rcu_sched_qs(cpu); | |
1956 | force_quiescent_state(&rcu_sched_state, 0); | |
1957 | c = c || per_cpu(rcu_sched_data, cpu).nxtlist; | |
1958 | } | |
1959 | if (per_cpu(rcu_bh_data, cpu).nxtlist) { | |
1960 | rcu_bh_qs(cpu); | |
1961 | force_quiescent_state(&rcu_bh_state, 0); | |
1962 | c = c || per_cpu(rcu_bh_data, cpu).nxtlist; | |
8bd93a2c PM |
1963 | } |
1964 | ||
1965 | /* If RCU callbacks are still pending, RCU still needs this CPU. */ | |
622ea685 | 1966 | if (c) |
a46e0899 | 1967 | invoke_rcu_core(); |
8bd93a2c PM |
1968 | return c; |
1969 | } | |
1970 | ||
a47cd880 PM |
1971 | /* |
1972 | * Check to see if we need to continue a callback-flush operations to | |
1973 | * allow the last CPU to enter dyntick-idle mode. | |
1974 | */ | |
1975 | static void rcu_needs_cpu_flush(void) | |
1976 | { | |
1977 | int cpu = smp_processor_id(); | |
71da8132 | 1978 | unsigned long flags; |
a47cd880 PM |
1979 | |
1980 | if (per_cpu(rcu_dyntick_drain, cpu) <= 0) | |
1981 | return; | |
71da8132 | 1982 | local_irq_save(flags); |
a47cd880 | 1983 | (void)rcu_needs_cpu(cpu); |
71da8132 | 1984 | local_irq_restore(flags); |
a47cd880 PM |
1985 | } |
1986 | ||
8bd93a2c | 1987 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |