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 | |
5b61b0ba MG |
30 | #define RCU_KTHREAD_PRIO 1 |
31 | ||
32 | #ifdef CONFIG_RCU_BOOST | |
33 | #define RCU_BOOST_PRIO CONFIG_RCU_BOOST_PRIO | |
34 | #else | |
35 | #define RCU_BOOST_PRIO RCU_KTHREAD_PRIO | |
36 | #endif | |
37 | ||
26845c28 PM |
38 | /* |
39 | * Check the RCU kernel configuration parameters and print informative | |
40 | * messages about anything out of the ordinary. If you like #ifdef, you | |
41 | * will love this function. | |
42 | */ | |
43 | static void __init rcu_bootup_announce_oddness(void) | |
44 | { | |
45 | #ifdef CONFIG_RCU_TRACE | |
46 | printk(KERN_INFO "\tRCU debugfs-based tracing is enabled.\n"); | |
47 | #endif | |
48 | #if (defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 64) || (!defined(CONFIG_64BIT) && CONFIG_RCU_FANOUT != 32) | |
49 | printk(KERN_INFO "\tCONFIG_RCU_FANOUT set to non-default value of %d\n", | |
50 | CONFIG_RCU_FANOUT); | |
51 | #endif | |
52 | #ifdef CONFIG_RCU_FANOUT_EXACT | |
53 | printk(KERN_INFO "\tHierarchical RCU autobalancing is disabled.\n"); | |
54 | #endif | |
55 | #ifdef CONFIG_RCU_FAST_NO_HZ | |
56 | printk(KERN_INFO | |
57 | "\tRCU dyntick-idle grace-period acceleration is enabled.\n"); | |
58 | #endif | |
59 | #ifdef CONFIG_PROVE_RCU | |
60 | printk(KERN_INFO "\tRCU lockdep checking is enabled.\n"); | |
61 | #endif | |
62 | #ifdef CONFIG_RCU_TORTURE_TEST_RUNNABLE | |
63 | printk(KERN_INFO "\tRCU torture testing starts during boot.\n"); | |
64 | #endif | |
81a294c4 | 65 | #if defined(CONFIG_TREE_PREEMPT_RCU) && !defined(CONFIG_RCU_CPU_STALL_VERBOSE) |
26845c28 PM |
66 | printk(KERN_INFO "\tVerbose stalled-CPUs detection is disabled.\n"); |
67 | #endif | |
68 | #if NUM_RCU_LVL_4 != 0 | |
69 | printk(KERN_INFO "\tExperimental four-level hierarchy is enabled.\n"); | |
70 | #endif | |
71 | } | |
72 | ||
f41d911f PM |
73 | #ifdef CONFIG_TREE_PREEMPT_RCU |
74 | ||
e99033c5 | 75 | struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt); |
f41d911f | 76 | DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data); |
27f4d280 | 77 | static struct rcu_state *rcu_state = &rcu_preempt_state; |
f41d911f | 78 | |
10f39bb1 | 79 | static void rcu_read_unlock_special(struct task_struct *t); |
d9a3da06 PM |
80 | static int rcu_preempted_readers_exp(struct rcu_node *rnp); |
81 | ||
f41d911f PM |
82 | /* |
83 | * Tell them what RCU they are running. | |
84 | */ | |
0e0fc1c2 | 85 | static void __init rcu_bootup_announce(void) |
f41d911f | 86 | { |
6cc68793 | 87 | printk(KERN_INFO "Preemptible hierarchical RCU implementation.\n"); |
26845c28 | 88 | rcu_bootup_announce_oddness(); |
f41d911f PM |
89 | } |
90 | ||
91 | /* | |
92 | * Return the number of RCU-preempt batches processed thus far | |
93 | * for debug and statistics. | |
94 | */ | |
95 | long rcu_batches_completed_preempt(void) | |
96 | { | |
97 | return rcu_preempt_state.completed; | |
98 | } | |
99 | EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt); | |
100 | ||
101 | /* | |
102 | * Return the number of RCU batches processed thus far for debug & stats. | |
103 | */ | |
104 | long rcu_batches_completed(void) | |
105 | { | |
106 | return rcu_batches_completed_preempt(); | |
107 | } | |
108 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
109 | ||
bf66f18e PM |
110 | /* |
111 | * Force a quiescent state for preemptible RCU. | |
112 | */ | |
113 | void rcu_force_quiescent_state(void) | |
114 | { | |
115 | force_quiescent_state(&rcu_preempt_state, 0); | |
116 | } | |
117 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
118 | ||
f41d911f | 119 | /* |
6cc68793 | 120 | * Record a preemptible-RCU quiescent state for the specified CPU. Note |
f41d911f PM |
121 | * that this just means that the task currently running on the CPU is |
122 | * not in a quiescent state. There might be any number of tasks blocked | |
123 | * while in an RCU read-side critical section. | |
25502a6c PM |
124 | * |
125 | * Unlike the other rcu_*_qs() functions, callers to this function | |
126 | * must disable irqs in order to protect the assignment to | |
127 | * ->rcu_read_unlock_special. | |
f41d911f | 128 | */ |
c3422bea | 129 | static void rcu_preempt_qs(int cpu) |
f41d911f PM |
130 | { |
131 | struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu); | |
25502a6c | 132 | |
e4cc1f22 | 133 | rdp->passed_quiesce_gpnum = rdp->gpnum; |
c3422bea | 134 | barrier(); |
e4cc1f22 | 135 | if (rdp->passed_quiesce == 0) |
d4c08f2a | 136 | trace_rcu_grace_period("rcu_preempt", rdp->gpnum, "cpuqs"); |
e4cc1f22 | 137 | rdp->passed_quiesce = 1; |
25502a6c | 138 | current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS; |
f41d911f PM |
139 | } |
140 | ||
141 | /* | |
c3422bea PM |
142 | * We have entered the scheduler, and the current task might soon be |
143 | * context-switched away from. If this task is in an RCU read-side | |
144 | * critical section, we will no longer be able to rely on the CPU to | |
12f5f524 PM |
145 | * record that fact, so we enqueue the task on the blkd_tasks list. |
146 | * The task will dequeue itself when it exits the outermost enclosing | |
147 | * RCU read-side critical section. Therefore, the current grace period | |
148 | * cannot be permitted to complete until the blkd_tasks list entries | |
149 | * predating the current grace period drain, in other words, until | |
150 | * rnp->gp_tasks becomes NULL. | |
c3422bea PM |
151 | * |
152 | * Caller must disable preemption. | |
f41d911f | 153 | */ |
c3422bea | 154 | static void rcu_preempt_note_context_switch(int cpu) |
f41d911f PM |
155 | { |
156 | struct task_struct *t = current; | |
c3422bea | 157 | unsigned long flags; |
f41d911f PM |
158 | struct rcu_data *rdp; |
159 | struct rcu_node *rnp; | |
160 | ||
10f39bb1 | 161 | if (t->rcu_read_lock_nesting > 0 && |
f41d911f PM |
162 | (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) { |
163 | ||
164 | /* Possibly blocking in an RCU read-side critical section. */ | |
394f99a9 | 165 | rdp = per_cpu_ptr(rcu_preempt_state.rda, cpu); |
f41d911f | 166 | rnp = rdp->mynode; |
1304afb2 | 167 | raw_spin_lock_irqsave(&rnp->lock, flags); |
f41d911f | 168 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED; |
86848966 | 169 | t->rcu_blocked_node = rnp; |
f41d911f PM |
170 | |
171 | /* | |
172 | * If this CPU has already checked in, then this task | |
173 | * will hold up the next grace period rather than the | |
174 | * current grace period. Queue the task accordingly. | |
175 | * If the task is queued for the current grace period | |
176 | * (i.e., this CPU has not yet passed through a quiescent | |
177 | * state for the current grace period), then as long | |
178 | * as that task remains queued, the current grace period | |
12f5f524 PM |
179 | * cannot end. Note that there is some uncertainty as |
180 | * to exactly when the current grace period started. | |
181 | * We take a conservative approach, which can result | |
182 | * in unnecessarily waiting on tasks that started very | |
183 | * slightly after the current grace period began. C'est | |
184 | * la vie!!! | |
b0e165c0 PM |
185 | * |
186 | * But first, note that the current CPU must still be | |
187 | * on line! | |
f41d911f | 188 | */ |
b0e165c0 | 189 | WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0); |
e7d8842e | 190 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); |
12f5f524 PM |
191 | if ((rnp->qsmask & rdp->grpmask) && rnp->gp_tasks != NULL) { |
192 | list_add(&t->rcu_node_entry, rnp->gp_tasks->prev); | |
193 | rnp->gp_tasks = &t->rcu_node_entry; | |
27f4d280 PM |
194 | #ifdef CONFIG_RCU_BOOST |
195 | if (rnp->boost_tasks != NULL) | |
196 | rnp->boost_tasks = rnp->gp_tasks; | |
197 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
12f5f524 PM |
198 | } else { |
199 | list_add(&t->rcu_node_entry, &rnp->blkd_tasks); | |
200 | if (rnp->qsmask & rdp->grpmask) | |
201 | rnp->gp_tasks = &t->rcu_node_entry; | |
202 | } | |
d4c08f2a PM |
203 | trace_rcu_preempt_task(rdp->rsp->name, |
204 | t->pid, | |
205 | (rnp->qsmask & rdp->grpmask) | |
206 | ? rnp->gpnum | |
207 | : rnp->gpnum + 1); | |
1304afb2 | 208 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
10f39bb1 PM |
209 | } else if (t->rcu_read_lock_nesting < 0 && |
210 | t->rcu_read_unlock_special) { | |
211 | ||
212 | /* | |
213 | * Complete exit from RCU read-side critical section on | |
214 | * behalf of preempted instance of __rcu_read_unlock(). | |
215 | */ | |
216 | rcu_read_unlock_special(t); | |
f41d911f PM |
217 | } |
218 | ||
219 | /* | |
220 | * Either we were not in an RCU read-side critical section to | |
221 | * begin with, or we have now recorded that critical section | |
222 | * globally. Either way, we can now note a quiescent state | |
223 | * for this CPU. Again, if we were in an RCU read-side critical | |
224 | * section, and if that critical section was blocking the current | |
225 | * grace period, then the fact that the task has been enqueued | |
226 | * means that we continue to block the current grace period. | |
227 | */ | |
e7d8842e | 228 | local_irq_save(flags); |
25502a6c | 229 | rcu_preempt_qs(cpu); |
e7d8842e | 230 | local_irq_restore(flags); |
f41d911f PM |
231 | } |
232 | ||
233 | /* | |
6cc68793 | 234 | * Tree-preemptible RCU implementation for rcu_read_lock(). |
f41d911f PM |
235 | * Just increment ->rcu_read_lock_nesting, shared state will be updated |
236 | * if we block. | |
237 | */ | |
238 | void __rcu_read_lock(void) | |
239 | { | |
80dcf60e | 240 | current->rcu_read_lock_nesting++; |
f41d911f PM |
241 | barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */ |
242 | } | |
243 | EXPORT_SYMBOL_GPL(__rcu_read_lock); | |
244 | ||
fc2219d4 PM |
245 | /* |
246 | * Check for preempted RCU readers blocking the current grace period | |
247 | * for the specified rcu_node structure. If the caller needs a reliable | |
248 | * answer, it must hold the rcu_node's ->lock. | |
249 | */ | |
27f4d280 | 250 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 | 251 | { |
12f5f524 | 252 | return rnp->gp_tasks != NULL; |
fc2219d4 PM |
253 | } |
254 | ||
b668c9cf PM |
255 | /* |
256 | * Record a quiescent state for all tasks that were previously queued | |
257 | * on the specified rcu_node structure and that were blocking the current | |
258 | * RCU grace period. The caller must hold the specified rnp->lock with | |
259 | * irqs disabled, and this lock is released upon return, but irqs remain | |
260 | * disabled. | |
261 | */ | |
d3f6bad3 | 262 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
b668c9cf PM |
263 | __releases(rnp->lock) |
264 | { | |
265 | unsigned long mask; | |
266 | struct rcu_node *rnp_p; | |
267 | ||
27f4d280 | 268 | if (rnp->qsmask != 0 || rcu_preempt_blocked_readers_cgp(rnp)) { |
1304afb2 | 269 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf PM |
270 | return; /* Still need more quiescent states! */ |
271 | } | |
272 | ||
273 | rnp_p = rnp->parent; | |
274 | if (rnp_p == NULL) { | |
275 | /* | |
276 | * Either there is only one rcu_node in the tree, | |
277 | * or tasks were kicked up to root rcu_node due to | |
278 | * CPUs going offline. | |
279 | */ | |
d3f6bad3 | 280 | rcu_report_qs_rsp(&rcu_preempt_state, flags); |
b668c9cf PM |
281 | return; |
282 | } | |
283 | ||
284 | /* Report up the rest of the hierarchy. */ | |
285 | mask = rnp->grpmask; | |
1304afb2 PM |
286 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
287 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
d3f6bad3 | 288 | rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags); |
b668c9cf PM |
289 | } |
290 | ||
12f5f524 PM |
291 | /* |
292 | * Advance a ->blkd_tasks-list pointer to the next entry, instead | |
293 | * returning NULL if at the end of the list. | |
294 | */ | |
295 | static struct list_head *rcu_next_node_entry(struct task_struct *t, | |
296 | struct rcu_node *rnp) | |
297 | { | |
298 | struct list_head *np; | |
299 | ||
300 | np = t->rcu_node_entry.next; | |
301 | if (np == &rnp->blkd_tasks) | |
302 | np = NULL; | |
303 | return np; | |
304 | } | |
305 | ||
b668c9cf PM |
306 | /* |
307 | * Handle special cases during rcu_read_unlock(), such as needing to | |
308 | * notify RCU core processing or task having blocked during the RCU | |
309 | * read-side critical section. | |
310 | */ | |
be0e1e21 | 311 | static noinline void rcu_read_unlock_special(struct task_struct *t) |
f41d911f PM |
312 | { |
313 | int empty; | |
d9a3da06 | 314 | int empty_exp; |
389abd48 | 315 | int empty_exp_now; |
f41d911f | 316 | unsigned long flags; |
12f5f524 | 317 | struct list_head *np; |
82e78d80 PM |
318 | #ifdef CONFIG_RCU_BOOST |
319 | struct rt_mutex *rbmp = NULL; | |
320 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
f41d911f PM |
321 | struct rcu_node *rnp; |
322 | int special; | |
323 | ||
324 | /* NMI handlers cannot block and cannot safely manipulate state. */ | |
325 | if (in_nmi()) | |
326 | return; | |
327 | ||
328 | local_irq_save(flags); | |
329 | ||
330 | /* | |
331 | * If RCU core is waiting for this CPU to exit critical section, | |
332 | * let it know that we have done so. | |
333 | */ | |
334 | special = t->rcu_read_unlock_special; | |
335 | if (special & RCU_READ_UNLOCK_NEED_QS) { | |
c3422bea | 336 | rcu_preempt_qs(smp_processor_id()); |
f41d911f PM |
337 | } |
338 | ||
339 | /* Hardware IRQ handlers cannot block. */ | |
ec433f0c | 340 | if (in_irq() || in_serving_softirq()) { |
f41d911f PM |
341 | local_irq_restore(flags); |
342 | return; | |
343 | } | |
344 | ||
345 | /* Clean up if blocked during RCU read-side critical section. */ | |
346 | if (special & RCU_READ_UNLOCK_BLOCKED) { | |
347 | t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED; | |
348 | ||
dd5d19ba PM |
349 | /* |
350 | * Remove this task from the list it blocked on. The | |
351 | * task can migrate while we acquire the lock, but at | |
352 | * most one time. So at most two passes through loop. | |
353 | */ | |
354 | for (;;) { | |
86848966 | 355 | rnp = t->rcu_blocked_node; |
1304afb2 | 356 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
86848966 | 357 | if (rnp == t->rcu_blocked_node) |
dd5d19ba | 358 | break; |
1304afb2 | 359 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
dd5d19ba | 360 | } |
27f4d280 | 361 | empty = !rcu_preempt_blocked_readers_cgp(rnp); |
d9a3da06 PM |
362 | empty_exp = !rcu_preempted_readers_exp(rnp); |
363 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ | |
12f5f524 | 364 | np = rcu_next_node_entry(t, rnp); |
f41d911f | 365 | list_del_init(&t->rcu_node_entry); |
82e78d80 | 366 | t->rcu_blocked_node = NULL; |
d4c08f2a PM |
367 | trace_rcu_unlock_preempted_task("rcu_preempt", |
368 | rnp->gpnum, t->pid); | |
12f5f524 PM |
369 | if (&t->rcu_node_entry == rnp->gp_tasks) |
370 | rnp->gp_tasks = np; | |
371 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
372 | rnp->exp_tasks = np; | |
27f4d280 PM |
373 | #ifdef CONFIG_RCU_BOOST |
374 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
375 | rnp->boost_tasks = np; | |
82e78d80 PM |
376 | /* Snapshot/clear ->rcu_boost_mutex with rcu_node lock held. */ |
377 | if (t->rcu_boost_mutex) { | |
378 | rbmp = t->rcu_boost_mutex; | |
379 | t->rcu_boost_mutex = NULL; | |
7765be2f | 380 | } |
27f4d280 | 381 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
f41d911f PM |
382 | |
383 | /* | |
384 | * If this was the last task on the current list, and if | |
385 | * we aren't waiting on any CPUs, report the quiescent state. | |
389abd48 PM |
386 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock, |
387 | * so we must take a snapshot of the expedited state. | |
f41d911f | 388 | */ |
389abd48 | 389 | empty_exp_now = !rcu_preempted_readers_exp(rnp); |
d4c08f2a PM |
390 | if (!empty && !rcu_preempt_blocked_readers_cgp(rnp)) { |
391 | trace_rcu_quiescent_state_report("preempt_rcu", | |
392 | rnp->gpnum, | |
393 | 0, rnp->qsmask, | |
394 | rnp->level, | |
395 | rnp->grplo, | |
396 | rnp->grphi, | |
397 | !!rnp->gp_tasks); | |
d3f6bad3 | 398 | rcu_report_unblock_qs_rnp(rnp, flags); |
d4c08f2a PM |
399 | } else |
400 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
d9a3da06 | 401 | |
27f4d280 PM |
402 | #ifdef CONFIG_RCU_BOOST |
403 | /* Unboost if we were boosted. */ | |
82e78d80 PM |
404 | if (rbmp) |
405 | rt_mutex_unlock(rbmp); | |
27f4d280 PM |
406 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
407 | ||
d9a3da06 PM |
408 | /* |
409 | * If this was the last task on the expedited lists, | |
410 | * then we need to report up the rcu_node hierarchy. | |
411 | */ | |
389abd48 | 412 | if (!empty_exp && empty_exp_now) |
b40d293e | 413 | rcu_report_exp_rnp(&rcu_preempt_state, rnp, true); |
b668c9cf PM |
414 | } else { |
415 | local_irq_restore(flags); | |
f41d911f | 416 | } |
f41d911f PM |
417 | } |
418 | ||
419 | /* | |
6cc68793 | 420 | * Tree-preemptible RCU implementation for rcu_read_unlock(). |
f41d911f PM |
421 | * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost |
422 | * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then | |
423 | * invoke rcu_read_unlock_special() to clean up after a context switch | |
424 | * in an RCU read-side critical section and other special cases. | |
425 | */ | |
426 | void __rcu_read_unlock(void) | |
427 | { | |
428 | struct task_struct *t = current; | |
429 | ||
10f39bb1 PM |
430 | if (t->rcu_read_lock_nesting != 1) |
431 | --t->rcu_read_lock_nesting; | |
432 | else { | |
6206ab9b | 433 | barrier(); /* critical section before exit code. */ |
10f39bb1 PM |
434 | t->rcu_read_lock_nesting = INT_MIN; |
435 | barrier(); /* assign before ->rcu_read_unlock_special load */ | |
be0e1e21 PM |
436 | if (unlikely(ACCESS_ONCE(t->rcu_read_unlock_special))) |
437 | rcu_read_unlock_special(t); | |
10f39bb1 PM |
438 | barrier(); /* ->rcu_read_unlock_special load before assign */ |
439 | t->rcu_read_lock_nesting = 0; | |
be0e1e21 | 440 | } |
cba8244a | 441 | #ifdef CONFIG_PROVE_LOCKING |
10f39bb1 PM |
442 | { |
443 | int rrln = ACCESS_ONCE(t->rcu_read_lock_nesting); | |
444 | ||
445 | WARN_ON_ONCE(rrln < 0 && rrln > INT_MIN / 2); | |
446 | } | |
cba8244a | 447 | #endif /* #ifdef CONFIG_PROVE_LOCKING */ |
f41d911f PM |
448 | } |
449 | EXPORT_SYMBOL_GPL(__rcu_read_unlock); | |
450 | ||
1ed509a2 PM |
451 | #ifdef CONFIG_RCU_CPU_STALL_VERBOSE |
452 | ||
453 | /* | |
454 | * Dump detailed information for all tasks blocking the current RCU | |
455 | * grace period on the specified rcu_node structure. | |
456 | */ | |
457 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) | |
458 | { | |
459 | unsigned long flags; | |
1ed509a2 PM |
460 | struct task_struct *t; |
461 | ||
27f4d280 | 462 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
12f5f524 PM |
463 | return; |
464 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
465 | t = list_entry(rnp->gp_tasks, | |
466 | struct task_struct, rcu_node_entry); | |
467 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) | |
468 | sched_show_task(t); | |
469 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1ed509a2 PM |
470 | } |
471 | ||
472 | /* | |
473 | * Dump detailed information for all tasks blocking the current RCU | |
474 | * grace period. | |
475 | */ | |
476 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
477 | { | |
478 | struct rcu_node *rnp = rcu_get_root(rsp); | |
479 | ||
480 | rcu_print_detail_task_stall_rnp(rnp); | |
481 | rcu_for_each_leaf_node(rsp, rnp) | |
482 | rcu_print_detail_task_stall_rnp(rnp); | |
483 | } | |
484 | ||
485 | #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ | |
486 | ||
487 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
488 | { | |
489 | } | |
490 | ||
491 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ | |
492 | ||
f41d911f PM |
493 | /* |
494 | * Scan the current list of tasks blocked within RCU read-side critical | |
495 | * sections, printing out the tid of each. | |
496 | */ | |
9bc8b558 | 497 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 498 | { |
f41d911f | 499 | struct task_struct *t; |
9bc8b558 | 500 | int ndetected = 0; |
f41d911f | 501 | |
27f4d280 | 502 | if (!rcu_preempt_blocked_readers_cgp(rnp)) |
9bc8b558 | 503 | return 0; |
12f5f524 PM |
504 | t = list_entry(rnp->gp_tasks, |
505 | struct task_struct, rcu_node_entry); | |
9bc8b558 | 506 | list_for_each_entry_continue(t, &rnp->blkd_tasks, rcu_node_entry) { |
12f5f524 | 507 | printk(" P%d", t->pid); |
9bc8b558 PM |
508 | ndetected++; |
509 | } | |
510 | return ndetected; | |
f41d911f PM |
511 | } |
512 | ||
53d84e00 PM |
513 | /* |
514 | * Suppress preemptible RCU's CPU stall warnings by pushing the | |
515 | * time of the next stall-warning message comfortably far into the | |
516 | * future. | |
517 | */ | |
518 | static void rcu_preempt_stall_reset(void) | |
519 | { | |
520 | rcu_preempt_state.jiffies_stall = jiffies + ULONG_MAX / 2; | |
521 | } | |
522 | ||
b0e165c0 PM |
523 | /* |
524 | * Check that the list of blocked tasks for the newly completed grace | |
525 | * period is in fact empty. It is a serious bug to complete a grace | |
526 | * period that still has RCU readers blocked! This function must be | |
527 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock | |
528 | * must be held by the caller. | |
12f5f524 PM |
529 | * |
530 | * Also, if there are blocked tasks on the list, they automatically | |
531 | * block the newly created grace period, so set up ->gp_tasks accordingly. | |
b0e165c0 PM |
532 | */ |
533 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
534 | { | |
27f4d280 | 535 | WARN_ON_ONCE(rcu_preempt_blocked_readers_cgp(rnp)); |
12f5f524 PM |
536 | if (!list_empty(&rnp->blkd_tasks)) |
537 | rnp->gp_tasks = rnp->blkd_tasks.next; | |
28ecd580 | 538 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
539 | } |
540 | ||
33f76148 PM |
541 | #ifdef CONFIG_HOTPLUG_CPU |
542 | ||
dd5d19ba PM |
543 | /* |
544 | * Handle tasklist migration for case in which all CPUs covered by the | |
545 | * specified rcu_node have gone offline. Move them up to the root | |
546 | * rcu_node. The reason for not just moving them to the immediate | |
547 | * parent is to remove the need for rcu_read_unlock_special() to | |
548 | * make more than two attempts to acquire the target rcu_node's lock. | |
b668c9cf PM |
549 | * Returns true if there were tasks blocking the current RCU grace |
550 | * period. | |
dd5d19ba | 551 | * |
237c80c5 PM |
552 | * Returns 1 if there was previously a task blocking the current grace |
553 | * period on the specified rcu_node structure. | |
554 | * | |
dd5d19ba PM |
555 | * The caller must hold rnp->lock with irqs disabled. |
556 | */ | |
237c80c5 PM |
557 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
558 | struct rcu_node *rnp, | |
559 | struct rcu_data *rdp) | |
dd5d19ba | 560 | { |
dd5d19ba PM |
561 | struct list_head *lp; |
562 | struct list_head *lp_root; | |
d9a3da06 | 563 | int retval = 0; |
dd5d19ba | 564 | struct rcu_node *rnp_root = rcu_get_root(rsp); |
12f5f524 | 565 | struct task_struct *t; |
dd5d19ba | 566 | |
86848966 PM |
567 | if (rnp == rnp_root) { |
568 | WARN_ONCE(1, "Last CPU thought to be offlined?"); | |
237c80c5 | 569 | return 0; /* Shouldn't happen: at least one CPU online. */ |
86848966 | 570 | } |
12f5f524 PM |
571 | |
572 | /* If we are on an internal node, complain bitterly. */ | |
573 | WARN_ON_ONCE(rnp != rdp->mynode); | |
dd5d19ba PM |
574 | |
575 | /* | |
12f5f524 PM |
576 | * Move tasks up to root rcu_node. Don't try to get fancy for |
577 | * this corner-case operation -- just put this node's tasks | |
578 | * at the head of the root node's list, and update the root node's | |
579 | * ->gp_tasks and ->exp_tasks pointers to those of this node's, | |
580 | * if non-NULL. This might result in waiting for more tasks than | |
581 | * absolutely necessary, but this is a good performance/complexity | |
582 | * tradeoff. | |
dd5d19ba | 583 | */ |
27f4d280 | 584 | if (rcu_preempt_blocked_readers_cgp(rnp)) |
d9a3da06 PM |
585 | retval |= RCU_OFL_TASKS_NORM_GP; |
586 | if (rcu_preempted_readers_exp(rnp)) | |
587 | retval |= RCU_OFL_TASKS_EXP_GP; | |
12f5f524 PM |
588 | lp = &rnp->blkd_tasks; |
589 | lp_root = &rnp_root->blkd_tasks; | |
590 | while (!list_empty(lp)) { | |
591 | t = list_entry(lp->next, typeof(*t), rcu_node_entry); | |
592 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ | |
593 | list_del(&t->rcu_node_entry); | |
594 | t->rcu_blocked_node = rnp_root; | |
595 | list_add(&t->rcu_node_entry, lp_root); | |
596 | if (&t->rcu_node_entry == rnp->gp_tasks) | |
597 | rnp_root->gp_tasks = rnp->gp_tasks; | |
598 | if (&t->rcu_node_entry == rnp->exp_tasks) | |
599 | rnp_root->exp_tasks = rnp->exp_tasks; | |
27f4d280 PM |
600 | #ifdef CONFIG_RCU_BOOST |
601 | if (&t->rcu_node_entry == rnp->boost_tasks) | |
602 | rnp_root->boost_tasks = rnp->boost_tasks; | |
603 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
12f5f524 | 604 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ |
dd5d19ba | 605 | } |
27f4d280 PM |
606 | |
607 | #ifdef CONFIG_RCU_BOOST | |
608 | /* In case root is being boosted and leaf is not. */ | |
609 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ | |
610 | if (rnp_root->boost_tasks != NULL && | |
611 | rnp_root->boost_tasks != rnp_root->gp_tasks) | |
612 | rnp_root->boost_tasks = rnp_root->gp_tasks; | |
613 | raw_spin_unlock(&rnp_root->lock); /* irqs still disabled */ | |
614 | #endif /* #ifdef CONFIG_RCU_BOOST */ | |
615 | ||
12f5f524 PM |
616 | rnp->gp_tasks = NULL; |
617 | rnp->exp_tasks = NULL; | |
237c80c5 | 618 | return retval; |
dd5d19ba PM |
619 | } |
620 | ||
e5601400 PM |
621 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
622 | ||
33f76148 | 623 | /* |
6cc68793 | 624 | * Do CPU-offline processing for preemptible RCU. |
33f76148 | 625 | */ |
e5601400 | 626 | static void rcu_preempt_cleanup_dead_cpu(int cpu) |
33f76148 | 627 | { |
e5601400 | 628 | rcu_cleanup_dead_cpu(cpu, &rcu_preempt_state); |
33f76148 PM |
629 | } |
630 | ||
f41d911f PM |
631 | /* |
632 | * Check for a quiescent state from the current CPU. When a task blocks, | |
633 | * the task is recorded in the corresponding CPU's rcu_node structure, | |
634 | * which is checked elsewhere. | |
635 | * | |
636 | * Caller must disable hard irqs. | |
637 | */ | |
638 | static void rcu_preempt_check_callbacks(int cpu) | |
639 | { | |
640 | struct task_struct *t = current; | |
641 | ||
642 | if (t->rcu_read_lock_nesting == 0) { | |
c3422bea | 643 | rcu_preempt_qs(cpu); |
f41d911f PM |
644 | return; |
645 | } | |
10f39bb1 PM |
646 | if (t->rcu_read_lock_nesting > 0 && |
647 | per_cpu(rcu_preempt_data, cpu).qs_pending) | |
c3422bea | 648 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS; |
f41d911f PM |
649 | } |
650 | ||
651 | /* | |
6cc68793 | 652 | * Process callbacks for preemptible RCU. |
f41d911f PM |
653 | */ |
654 | static void rcu_preempt_process_callbacks(void) | |
655 | { | |
656 | __rcu_process_callbacks(&rcu_preempt_state, | |
657 | &__get_cpu_var(rcu_preempt_data)); | |
658 | } | |
659 | ||
a46e0899 PM |
660 | #ifdef CONFIG_RCU_BOOST |
661 | ||
09223371 SL |
662 | static void rcu_preempt_do_callbacks(void) |
663 | { | |
664 | rcu_do_batch(&rcu_preempt_state, &__get_cpu_var(rcu_preempt_data)); | |
665 | } | |
666 | ||
a46e0899 PM |
667 | #endif /* #ifdef CONFIG_RCU_BOOST */ |
668 | ||
f41d911f | 669 | /* |
6cc68793 | 670 | * Queue a preemptible-RCU callback for invocation after a grace period. |
f41d911f PM |
671 | */ |
672 | void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
673 | { | |
486e2593 | 674 | __call_rcu(head, func, &rcu_preempt_state, 0); |
f41d911f PM |
675 | } |
676 | EXPORT_SYMBOL_GPL(call_rcu); | |
677 | ||
486e2593 PM |
678 | /* |
679 | * Queue an RCU callback for lazy invocation after a grace period. | |
680 | * This will likely be later named something like "call_rcu_lazy()", | |
681 | * but this change will require some way of tagging the lazy RCU | |
682 | * callbacks in the list of pending callbacks. Until then, this | |
683 | * function may only be called from __kfree_rcu(). | |
684 | */ | |
685 | void kfree_call_rcu(struct rcu_head *head, | |
686 | void (*func)(struct rcu_head *rcu)) | |
687 | { | |
688 | __call_rcu(head, func, &rcu_preempt_state, 1); | |
689 | } | |
690 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
691 | ||
6ebb237b PM |
692 | /** |
693 | * synchronize_rcu - wait until a grace period has elapsed. | |
694 | * | |
695 | * Control will return to the caller some time after a full grace | |
696 | * period has elapsed, in other words after all currently executing RCU | |
77d8485a PM |
697 | * read-side critical sections have completed. Note, however, that |
698 | * upon return from synchronize_rcu(), the caller might well be executing | |
699 | * concurrently with new RCU read-side critical sections that began while | |
700 | * synchronize_rcu() was waiting. RCU read-side critical sections are | |
701 | * delimited by rcu_read_lock() and rcu_read_unlock(), and may be nested. | |
6ebb237b PM |
702 | */ |
703 | void synchronize_rcu(void) | |
704 | { | |
fe15d706 PM |
705 | rcu_lockdep_assert(!lock_is_held(&rcu_bh_lock_map) && |
706 | !lock_is_held(&rcu_lock_map) && | |
707 | !lock_is_held(&rcu_sched_lock_map), | |
708 | "Illegal synchronize_rcu() in RCU read-side critical section"); | |
6ebb237b PM |
709 | if (!rcu_scheduler_active) |
710 | return; | |
2c42818e | 711 | wait_rcu_gp(call_rcu); |
6ebb237b PM |
712 | } |
713 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
714 | ||
d9a3da06 PM |
715 | static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq); |
716 | static long sync_rcu_preempt_exp_count; | |
717 | static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex); | |
718 | ||
719 | /* | |
720 | * Return non-zero if there are any tasks in RCU read-side critical | |
721 | * sections blocking the current preemptible-RCU expedited grace period. | |
722 | * If there is no preemptible-RCU expedited grace period currently in | |
723 | * progress, returns zero unconditionally. | |
724 | */ | |
725 | static int rcu_preempted_readers_exp(struct rcu_node *rnp) | |
726 | { | |
12f5f524 | 727 | return rnp->exp_tasks != NULL; |
d9a3da06 PM |
728 | } |
729 | ||
730 | /* | |
731 | * return non-zero if there is no RCU expedited grace period in progress | |
732 | * for the specified rcu_node structure, in other words, if all CPUs and | |
733 | * tasks covered by the specified rcu_node structure have done their bit | |
734 | * for the current expedited grace period. Works only for preemptible | |
735 | * RCU -- other RCU implementation use other means. | |
736 | * | |
737 | * Caller must hold sync_rcu_preempt_exp_mutex. | |
738 | */ | |
739 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) | |
740 | { | |
741 | return !rcu_preempted_readers_exp(rnp) && | |
742 | ACCESS_ONCE(rnp->expmask) == 0; | |
743 | } | |
744 | ||
745 | /* | |
746 | * Report the exit from RCU read-side critical section for the last task | |
747 | * that queued itself during or before the current expedited preemptible-RCU | |
748 | * grace period. This event is reported either to the rcu_node structure on | |
749 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
750 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
751 | * iteratively!) | |
752 | * | |
b40d293e TG |
753 | * Most callers will set the "wake" flag, but the task initiating the |
754 | * expedited grace period need not wake itself. | |
755 | * | |
d9a3da06 PM |
756 | * Caller must hold sync_rcu_preempt_exp_mutex. |
757 | */ | |
b40d293e TG |
758 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
759 | bool wake) | |
d9a3da06 PM |
760 | { |
761 | unsigned long flags; | |
762 | unsigned long mask; | |
763 | ||
1304afb2 | 764 | raw_spin_lock_irqsave(&rnp->lock, flags); |
d9a3da06 | 765 | for (;;) { |
131906b0 PM |
766 | if (!sync_rcu_preempt_exp_done(rnp)) { |
767 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
d9a3da06 | 768 | break; |
131906b0 | 769 | } |
d9a3da06 | 770 | if (rnp->parent == NULL) { |
131906b0 | 771 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b40d293e TG |
772 | if (wake) |
773 | wake_up(&sync_rcu_preempt_exp_wq); | |
d9a3da06 PM |
774 | break; |
775 | } | |
776 | mask = rnp->grpmask; | |
1304afb2 | 777 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ |
d9a3da06 | 778 | rnp = rnp->parent; |
1304afb2 | 779 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ |
d9a3da06 PM |
780 | rnp->expmask &= ~mask; |
781 | } | |
d9a3da06 PM |
782 | } |
783 | ||
784 | /* | |
785 | * Snapshot the tasks blocking the newly started preemptible-RCU expedited | |
786 | * grace period for the specified rcu_node structure. If there are no such | |
787 | * tasks, report it up the rcu_node hierarchy. | |
788 | * | |
789 | * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock. | |
790 | */ | |
791 | static void | |
792 | sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp) | |
793 | { | |
1217ed1b | 794 | unsigned long flags; |
12f5f524 | 795 | int must_wait = 0; |
d9a3da06 | 796 | |
1217ed1b PM |
797 | raw_spin_lock_irqsave(&rnp->lock, flags); |
798 | if (list_empty(&rnp->blkd_tasks)) | |
799 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
800 | else { | |
12f5f524 | 801 | rnp->exp_tasks = rnp->blkd_tasks.next; |
1217ed1b | 802 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock */ |
12f5f524 PM |
803 | must_wait = 1; |
804 | } | |
d9a3da06 | 805 | if (!must_wait) |
b40d293e | 806 | rcu_report_exp_rnp(rsp, rnp, false); /* Don't wake self. */ |
d9a3da06 PM |
807 | } |
808 | ||
019129d5 | 809 | /* |
d9a3da06 PM |
810 | * Wait for an rcu-preempt grace period, but expedite it. The basic idea |
811 | * is to invoke synchronize_sched_expedited() to push all the tasks to | |
12f5f524 | 812 | * the ->blkd_tasks lists and wait for this list to drain. |
019129d5 PM |
813 | */ |
814 | void synchronize_rcu_expedited(void) | |
815 | { | |
d9a3da06 PM |
816 | unsigned long flags; |
817 | struct rcu_node *rnp; | |
818 | struct rcu_state *rsp = &rcu_preempt_state; | |
819 | long snap; | |
820 | int trycount = 0; | |
821 | ||
822 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
823 | snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1; | |
824 | smp_mb(); /* Above access cannot bleed into critical section. */ | |
825 | ||
826 | /* | |
827 | * Acquire lock, falling back to synchronize_rcu() if too many | |
828 | * lock-acquisition failures. Of course, if someone does the | |
829 | * expedited grace period for us, just leave. | |
830 | */ | |
831 | while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) { | |
832 | if (trycount++ < 10) | |
833 | udelay(trycount * num_online_cpus()); | |
834 | else { | |
835 | synchronize_rcu(); | |
836 | return; | |
837 | } | |
838 | if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0) | |
839 | goto mb_ret; /* Others did our work for us. */ | |
840 | } | |
841 | if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0) | |
842 | goto unlock_mb_ret; /* Others did our work for us. */ | |
843 | ||
12f5f524 | 844 | /* force all RCU readers onto ->blkd_tasks lists. */ |
d9a3da06 PM |
845 | synchronize_sched_expedited(); |
846 | ||
1304afb2 | 847 | raw_spin_lock_irqsave(&rsp->onofflock, flags); |
d9a3da06 PM |
848 | |
849 | /* Initialize ->expmask for all non-leaf rcu_node structures. */ | |
850 | rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) { | |
1304afb2 | 851 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ |
d9a3da06 | 852 | rnp->expmask = rnp->qsmaskinit; |
1304afb2 | 853 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ |
d9a3da06 PM |
854 | } |
855 | ||
12f5f524 | 856 | /* Snapshot current state of ->blkd_tasks lists. */ |
d9a3da06 PM |
857 | rcu_for_each_leaf_node(rsp, rnp) |
858 | sync_rcu_preempt_exp_init(rsp, rnp); | |
859 | if (NUM_RCU_NODES > 1) | |
860 | sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp)); | |
861 | ||
1304afb2 | 862 | raw_spin_unlock_irqrestore(&rsp->onofflock, flags); |
d9a3da06 | 863 | |
12f5f524 | 864 | /* Wait for snapshotted ->blkd_tasks lists to drain. */ |
d9a3da06 PM |
865 | rnp = rcu_get_root(rsp); |
866 | wait_event(sync_rcu_preempt_exp_wq, | |
867 | sync_rcu_preempt_exp_done(rnp)); | |
868 | ||
869 | /* Clean up and exit. */ | |
870 | smp_mb(); /* ensure expedited GP seen before counter increment. */ | |
871 | ACCESS_ONCE(sync_rcu_preempt_exp_count)++; | |
872 | unlock_mb_ret: | |
873 | mutex_unlock(&sync_rcu_preempt_exp_mutex); | |
874 | mb_ret: | |
875 | smp_mb(); /* ensure subsequent action seen after grace period. */ | |
019129d5 PM |
876 | } |
877 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
878 | ||
f41d911f | 879 | /* |
6cc68793 | 880 | * Check to see if there is any immediate preemptible-RCU-related work |
f41d911f PM |
881 | * to be done. |
882 | */ | |
883 | static int rcu_preempt_pending(int cpu) | |
884 | { | |
885 | return __rcu_pending(&rcu_preempt_state, | |
886 | &per_cpu(rcu_preempt_data, cpu)); | |
887 | } | |
888 | ||
889 | /* | |
6cc68793 | 890 | * Does preemptible RCU need the CPU to stay out of dynticks mode? |
f41d911f PM |
891 | */ |
892 | static int rcu_preempt_needs_cpu(int cpu) | |
893 | { | |
894 | return !!per_cpu(rcu_preempt_data, cpu).nxtlist; | |
895 | } | |
896 | ||
e74f4c45 PM |
897 | /** |
898 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
899 | */ | |
900 | void rcu_barrier(void) | |
901 | { | |
902 | _rcu_barrier(&rcu_preempt_state, call_rcu); | |
903 | } | |
904 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
905 | ||
f41d911f | 906 | /* |
6cc68793 | 907 | * Initialize preemptible RCU's per-CPU data. |
f41d911f PM |
908 | */ |
909 | static void __cpuinit rcu_preempt_init_percpu_data(int cpu) | |
910 | { | |
911 | rcu_init_percpu_data(cpu, &rcu_preempt_state, 1); | |
912 | } | |
913 | ||
e74f4c45 | 914 | /* |
e5601400 PM |
915 | * Move preemptible RCU's callbacks from dying CPU to other online CPU |
916 | * and record a quiescent state. | |
e74f4c45 | 917 | */ |
e5601400 | 918 | static void rcu_preempt_cleanup_dying_cpu(void) |
e74f4c45 | 919 | { |
e5601400 | 920 | rcu_cleanup_dying_cpu(&rcu_preempt_state); |
e74f4c45 PM |
921 | } |
922 | ||
1eba8f84 | 923 | /* |
6cc68793 | 924 | * Initialize preemptible RCU's state structures. |
1eba8f84 PM |
925 | */ |
926 | static void __init __rcu_init_preempt(void) | |
927 | { | |
394f99a9 | 928 | rcu_init_one(&rcu_preempt_state, &rcu_preempt_data); |
1eba8f84 PM |
929 | } |
930 | ||
f41d911f | 931 | /* |
6cc68793 | 932 | * Check for a task exiting while in a preemptible-RCU read-side |
f41d911f PM |
933 | * critical section, clean up if so. No need to issue warnings, |
934 | * as debug_check_no_locks_held() already does this if lockdep | |
935 | * is enabled. | |
936 | */ | |
937 | void exit_rcu(void) | |
938 | { | |
939 | struct task_struct *t = current; | |
940 | ||
941 | if (t->rcu_read_lock_nesting == 0) | |
942 | return; | |
943 | t->rcu_read_lock_nesting = 1; | |
13491a0e | 944 | __rcu_read_unlock(); |
f41d911f PM |
945 | } |
946 | ||
947 | #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */ | |
948 | ||
27f4d280 PM |
949 | static struct rcu_state *rcu_state = &rcu_sched_state; |
950 | ||
f41d911f PM |
951 | /* |
952 | * Tell them what RCU they are running. | |
953 | */ | |
0e0fc1c2 | 954 | static void __init rcu_bootup_announce(void) |
f41d911f PM |
955 | { |
956 | printk(KERN_INFO "Hierarchical RCU implementation.\n"); | |
26845c28 | 957 | rcu_bootup_announce_oddness(); |
f41d911f PM |
958 | } |
959 | ||
960 | /* | |
961 | * Return the number of RCU batches processed thus far for debug & stats. | |
962 | */ | |
963 | long rcu_batches_completed(void) | |
964 | { | |
965 | return rcu_batches_completed_sched(); | |
966 | } | |
967 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
968 | ||
bf66f18e PM |
969 | /* |
970 | * Force a quiescent state for RCU, which, because there is no preemptible | |
971 | * RCU, becomes the same as rcu-sched. | |
972 | */ | |
973 | void rcu_force_quiescent_state(void) | |
974 | { | |
975 | rcu_sched_force_quiescent_state(); | |
976 | } | |
977 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
978 | ||
f41d911f | 979 | /* |
6cc68793 | 980 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
981 | * CPUs being in quiescent states. |
982 | */ | |
c3422bea | 983 | static void rcu_preempt_note_context_switch(int cpu) |
f41d911f PM |
984 | { |
985 | } | |
986 | ||
fc2219d4 | 987 | /* |
6cc68793 | 988 | * Because preemptible RCU does not exist, there are never any preempted |
fc2219d4 PM |
989 | * RCU readers. |
990 | */ | |
27f4d280 | 991 | static int rcu_preempt_blocked_readers_cgp(struct rcu_node *rnp) |
fc2219d4 PM |
992 | { |
993 | return 0; | |
994 | } | |
995 | ||
b668c9cf PM |
996 | #ifdef CONFIG_HOTPLUG_CPU |
997 | ||
998 | /* Because preemptible RCU does not exist, no quieting of tasks. */ | |
d3f6bad3 | 999 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) |
b668c9cf | 1000 | { |
1304afb2 | 1001 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
b668c9cf PM |
1002 | } |
1003 | ||
1004 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1005 | ||
1ed509a2 | 1006 | /* |
6cc68793 | 1007 | * Because preemptible RCU does not exist, we never have to check for |
1ed509a2 PM |
1008 | * tasks blocked within RCU read-side critical sections. |
1009 | */ | |
1010 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
1011 | { | |
1012 | } | |
1013 | ||
f41d911f | 1014 | /* |
6cc68793 | 1015 | * Because preemptible RCU does not exist, we never have to check for |
f41d911f PM |
1016 | * tasks blocked within RCU read-side critical sections. |
1017 | */ | |
9bc8b558 | 1018 | static int rcu_print_task_stall(struct rcu_node *rnp) |
f41d911f | 1019 | { |
9bc8b558 | 1020 | return 0; |
f41d911f PM |
1021 | } |
1022 | ||
53d84e00 PM |
1023 | /* |
1024 | * Because preemptible RCU does not exist, there is no need to suppress | |
1025 | * its CPU stall warnings. | |
1026 | */ | |
1027 | static void rcu_preempt_stall_reset(void) | |
1028 | { | |
1029 | } | |
1030 | ||
b0e165c0 | 1031 | /* |
6cc68793 | 1032 | * Because there is no preemptible RCU, there can be no readers blocked, |
49e29126 PM |
1033 | * so there is no need to check for blocked tasks. So check only for |
1034 | * bogus qsmask values. | |
b0e165c0 PM |
1035 | */ |
1036 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
1037 | { | |
49e29126 | 1038 | WARN_ON_ONCE(rnp->qsmask); |
b0e165c0 PM |
1039 | } |
1040 | ||
33f76148 PM |
1041 | #ifdef CONFIG_HOTPLUG_CPU |
1042 | ||
dd5d19ba | 1043 | /* |
6cc68793 | 1044 | * Because preemptible RCU does not exist, it never needs to migrate |
237c80c5 PM |
1045 | * tasks that were blocked within RCU read-side critical sections, and |
1046 | * such non-existent tasks cannot possibly have been blocking the current | |
1047 | * grace period. | |
dd5d19ba | 1048 | */ |
237c80c5 PM |
1049 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, |
1050 | struct rcu_node *rnp, | |
1051 | struct rcu_data *rdp) | |
dd5d19ba | 1052 | { |
237c80c5 | 1053 | return 0; |
dd5d19ba PM |
1054 | } |
1055 | ||
e5601400 PM |
1056 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ |
1057 | ||
33f76148 | 1058 | /* |
6cc68793 | 1059 | * Because preemptible RCU does not exist, it never needs CPU-offline |
33f76148 PM |
1060 | * processing. |
1061 | */ | |
e5601400 | 1062 | static void rcu_preempt_cleanup_dead_cpu(int cpu) |
33f76148 PM |
1063 | { |
1064 | } | |
1065 | ||
f41d911f | 1066 | /* |
6cc68793 | 1067 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
1068 | * to check. |
1069 | */ | |
1eba8f84 | 1070 | static void rcu_preempt_check_callbacks(int cpu) |
f41d911f PM |
1071 | { |
1072 | } | |
1073 | ||
1074 | /* | |
6cc68793 | 1075 | * Because preemptible RCU does not exist, it never has any callbacks |
f41d911f PM |
1076 | * to process. |
1077 | */ | |
1eba8f84 | 1078 | static void rcu_preempt_process_callbacks(void) |
f41d911f PM |
1079 | { |
1080 | } | |
1081 | ||
486e2593 PM |
1082 | /* |
1083 | * Queue an RCU callback for lazy invocation after a grace period. | |
1084 | * This will likely be later named something like "call_rcu_lazy()", | |
1085 | * but this change will require some way of tagging the lazy RCU | |
1086 | * callbacks in the list of pending callbacks. Until then, this | |
1087 | * function may only be called from __kfree_rcu(). | |
1088 | * | |
1089 | * Because there is no preemptible RCU, we use RCU-sched instead. | |
1090 | */ | |
1091 | void kfree_call_rcu(struct rcu_head *head, | |
1092 | void (*func)(struct rcu_head *rcu)) | |
1093 | { | |
1094 | __call_rcu(head, func, &rcu_sched_state, 1); | |
1095 | } | |
1096 | EXPORT_SYMBOL_GPL(kfree_call_rcu); | |
1097 | ||
019129d5 PM |
1098 | /* |
1099 | * Wait for an rcu-preempt grace period, but make it happen quickly. | |
6cc68793 | 1100 | * But because preemptible RCU does not exist, map to rcu-sched. |
019129d5 PM |
1101 | */ |
1102 | void synchronize_rcu_expedited(void) | |
1103 | { | |
1104 | synchronize_sched_expedited(); | |
1105 | } | |
1106 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
1107 | ||
d9a3da06 PM |
1108 | #ifdef CONFIG_HOTPLUG_CPU |
1109 | ||
1110 | /* | |
6cc68793 | 1111 | * Because preemptible RCU does not exist, there is never any need to |
d9a3da06 PM |
1112 | * report on tasks preempted in RCU read-side critical sections during |
1113 | * expedited RCU grace periods. | |
1114 | */ | |
b40d293e TG |
1115 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp, |
1116 | bool wake) | |
d9a3da06 | 1117 | { |
d9a3da06 PM |
1118 | } |
1119 | ||
1120 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1121 | ||
f41d911f | 1122 | /* |
6cc68793 | 1123 | * Because preemptible RCU does not exist, it never has any work to do. |
f41d911f PM |
1124 | */ |
1125 | static int rcu_preempt_pending(int cpu) | |
1126 | { | |
1127 | return 0; | |
1128 | } | |
1129 | ||
1130 | /* | |
6cc68793 | 1131 | * Because preemptible RCU does not exist, it never needs any CPU. |
f41d911f PM |
1132 | */ |
1133 | static int rcu_preempt_needs_cpu(int cpu) | |
1134 | { | |
1135 | return 0; | |
1136 | } | |
1137 | ||
e74f4c45 | 1138 | /* |
6cc68793 | 1139 | * Because preemptible RCU does not exist, rcu_barrier() is just |
e74f4c45 PM |
1140 | * another name for rcu_barrier_sched(). |
1141 | */ | |
1142 | void rcu_barrier(void) | |
1143 | { | |
1144 | rcu_barrier_sched(); | |
1145 | } | |
1146 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
1147 | ||
f41d911f | 1148 | /* |
6cc68793 | 1149 | * Because preemptible RCU does not exist, there is no per-CPU |
f41d911f PM |
1150 | * data to initialize. |
1151 | */ | |
1152 | static void __cpuinit rcu_preempt_init_percpu_data(int cpu) | |
1153 | { | |
1154 | } | |
1155 | ||
e74f4c45 | 1156 | /* |
e5601400 | 1157 | * Because there is no preemptible RCU, there is no cleanup to do. |
e74f4c45 | 1158 | */ |
e5601400 | 1159 | static void rcu_preempt_cleanup_dying_cpu(void) |
e74f4c45 PM |
1160 | { |
1161 | } | |
1162 | ||
1eba8f84 | 1163 | /* |
6cc68793 | 1164 | * Because preemptible RCU does not exist, it need not be initialized. |
1eba8f84 PM |
1165 | */ |
1166 | static void __init __rcu_init_preempt(void) | |
1167 | { | |
1168 | } | |
1169 | ||
f41d911f | 1170 | #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */ |
8bd93a2c | 1171 | |
27f4d280 PM |
1172 | #ifdef CONFIG_RCU_BOOST |
1173 | ||
1174 | #include "rtmutex_common.h" | |
1175 | ||
0ea1f2eb PM |
1176 | #ifdef CONFIG_RCU_TRACE |
1177 | ||
1178 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
1179 | { | |
1180 | if (list_empty(&rnp->blkd_tasks)) | |
1181 | rnp->n_balk_blkd_tasks++; | |
1182 | else if (rnp->exp_tasks == NULL && rnp->gp_tasks == NULL) | |
1183 | rnp->n_balk_exp_gp_tasks++; | |
1184 | else if (rnp->gp_tasks != NULL && rnp->boost_tasks != NULL) | |
1185 | rnp->n_balk_boost_tasks++; | |
1186 | else if (rnp->gp_tasks != NULL && rnp->qsmask != 0) | |
1187 | rnp->n_balk_notblocked++; | |
1188 | else if (rnp->gp_tasks != NULL && | |
a9f4793d | 1189 | ULONG_CMP_LT(jiffies, rnp->boost_time)) |
0ea1f2eb PM |
1190 | rnp->n_balk_notyet++; |
1191 | else | |
1192 | rnp->n_balk_nos++; | |
1193 | } | |
1194 | ||
1195 | #else /* #ifdef CONFIG_RCU_TRACE */ | |
1196 | ||
1197 | static void rcu_initiate_boost_trace(struct rcu_node *rnp) | |
1198 | { | |
1199 | } | |
1200 | ||
1201 | #endif /* #else #ifdef CONFIG_RCU_TRACE */ | |
1202 | ||
27f4d280 PM |
1203 | /* |
1204 | * Carry out RCU priority boosting on the task indicated by ->exp_tasks | |
1205 | * or ->boost_tasks, advancing the pointer to the next task in the | |
1206 | * ->blkd_tasks list. | |
1207 | * | |
1208 | * Note that irqs must be enabled: boosting the task can block. | |
1209 | * Returns 1 if there are more tasks needing to be boosted. | |
1210 | */ | |
1211 | static int rcu_boost(struct rcu_node *rnp) | |
1212 | { | |
1213 | unsigned long flags; | |
1214 | struct rt_mutex mtx; | |
1215 | struct task_struct *t; | |
1216 | struct list_head *tb; | |
1217 | ||
1218 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) | |
1219 | return 0; /* Nothing left to boost. */ | |
1220 | ||
1221 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1222 | ||
1223 | /* | |
1224 | * Recheck under the lock: all tasks in need of boosting | |
1225 | * might exit their RCU read-side critical sections on their own. | |
1226 | */ | |
1227 | if (rnp->exp_tasks == NULL && rnp->boost_tasks == NULL) { | |
1228 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1229 | return 0; | |
1230 | } | |
1231 | ||
1232 | /* | |
1233 | * Preferentially boost tasks blocking expedited grace periods. | |
1234 | * This cannot starve the normal grace periods because a second | |
1235 | * expedited grace period must boost all blocked tasks, including | |
1236 | * those blocking the pre-existing normal grace period. | |
1237 | */ | |
0ea1f2eb | 1238 | if (rnp->exp_tasks != NULL) { |
27f4d280 | 1239 | tb = rnp->exp_tasks; |
0ea1f2eb PM |
1240 | rnp->n_exp_boosts++; |
1241 | } else { | |
27f4d280 | 1242 | tb = rnp->boost_tasks; |
0ea1f2eb PM |
1243 | rnp->n_normal_boosts++; |
1244 | } | |
1245 | rnp->n_tasks_boosted++; | |
27f4d280 PM |
1246 | |
1247 | /* | |
1248 | * We boost task t by manufacturing an rt_mutex that appears to | |
1249 | * be held by task t. We leave a pointer to that rt_mutex where | |
1250 | * task t can find it, and task t will release the mutex when it | |
1251 | * exits its outermost RCU read-side critical section. Then | |
1252 | * simply acquiring this artificial rt_mutex will boost task | |
1253 | * t's priority. (Thanks to tglx for suggesting this approach!) | |
1254 | * | |
1255 | * Note that task t must acquire rnp->lock to remove itself from | |
1256 | * the ->blkd_tasks list, which it will do from exit() if from | |
1257 | * nowhere else. We therefore are guaranteed that task t will | |
1258 | * stay around at least until we drop rnp->lock. Note that | |
1259 | * rnp->lock also resolves races between our priority boosting | |
1260 | * and task t's exiting its outermost RCU read-side critical | |
1261 | * section. | |
1262 | */ | |
1263 | t = container_of(tb, struct task_struct, rcu_node_entry); | |
1264 | rt_mutex_init_proxy_locked(&mtx, t); | |
1265 | t->rcu_boost_mutex = &mtx; | |
27f4d280 PM |
1266 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1267 | rt_mutex_lock(&mtx); /* Side effect: boosts task t's priority. */ | |
1268 | rt_mutex_unlock(&mtx); /* Keep lockdep happy. */ | |
1269 | ||
4f89b336 PM |
1270 | return ACCESS_ONCE(rnp->exp_tasks) != NULL || |
1271 | ACCESS_ONCE(rnp->boost_tasks) != NULL; | |
27f4d280 PM |
1272 | } |
1273 | ||
1274 | /* | |
1275 | * Timer handler to initiate waking up of boost kthreads that | |
1276 | * have yielded the CPU due to excessive numbers of tasks to | |
1277 | * boost. We wake up the per-rcu_node kthread, which in turn | |
1278 | * will wake up the booster kthread. | |
1279 | */ | |
1280 | static void rcu_boost_kthread_timer(unsigned long arg) | |
1281 | { | |
1217ed1b | 1282 | invoke_rcu_node_kthread((struct rcu_node *)arg); |
27f4d280 PM |
1283 | } |
1284 | ||
1285 | /* | |
1286 | * Priority-boosting kthread. One per leaf rcu_node and one for the | |
1287 | * root rcu_node. | |
1288 | */ | |
1289 | static int rcu_boost_kthread(void *arg) | |
1290 | { | |
1291 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
1292 | int spincnt = 0; | |
1293 | int more2boost; | |
1294 | ||
385680a9 | 1295 | trace_rcu_utilization("Start boost kthread@init"); |
27f4d280 | 1296 | for (;;) { |
d71df90e | 1297 | rnp->boost_kthread_status = RCU_KTHREAD_WAITING; |
385680a9 | 1298 | trace_rcu_utilization("End boost kthread@rcu_wait"); |
08bca60a | 1299 | rcu_wait(rnp->boost_tasks || rnp->exp_tasks); |
385680a9 | 1300 | trace_rcu_utilization("Start boost kthread@rcu_wait"); |
d71df90e | 1301 | rnp->boost_kthread_status = RCU_KTHREAD_RUNNING; |
27f4d280 PM |
1302 | more2boost = rcu_boost(rnp); |
1303 | if (more2boost) | |
1304 | spincnt++; | |
1305 | else | |
1306 | spincnt = 0; | |
1307 | if (spincnt > 10) { | |
385680a9 | 1308 | trace_rcu_utilization("End boost kthread@rcu_yield"); |
27f4d280 | 1309 | rcu_yield(rcu_boost_kthread_timer, (unsigned long)rnp); |
385680a9 | 1310 | trace_rcu_utilization("Start boost kthread@rcu_yield"); |
27f4d280 PM |
1311 | spincnt = 0; |
1312 | } | |
1313 | } | |
1217ed1b | 1314 | /* NOTREACHED */ |
385680a9 | 1315 | trace_rcu_utilization("End boost kthread@notreached"); |
27f4d280 PM |
1316 | return 0; |
1317 | } | |
1318 | ||
1319 | /* | |
1320 | * Check to see if it is time to start boosting RCU readers that are | |
1321 | * blocking the current grace period, and, if so, tell the per-rcu_node | |
1322 | * kthread to start boosting them. If there is an expedited grace | |
1323 | * period in progress, it is always time to boost. | |
1324 | * | |
1217ed1b PM |
1325 | * The caller must hold rnp->lock, which this function releases, |
1326 | * but irqs remain disabled. The ->boost_kthread_task is immortal, | |
1327 | * so we don't need to worry about it going away. | |
27f4d280 | 1328 | */ |
1217ed1b | 1329 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
27f4d280 PM |
1330 | { |
1331 | struct task_struct *t; | |
1332 | ||
0ea1f2eb PM |
1333 | if (!rcu_preempt_blocked_readers_cgp(rnp) && rnp->exp_tasks == NULL) { |
1334 | rnp->n_balk_exp_gp_tasks++; | |
1217ed1b | 1335 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 | 1336 | return; |
0ea1f2eb | 1337 | } |
27f4d280 PM |
1338 | if (rnp->exp_tasks != NULL || |
1339 | (rnp->gp_tasks != NULL && | |
1340 | rnp->boost_tasks == NULL && | |
1341 | rnp->qsmask == 0 && | |
1342 | ULONG_CMP_GE(jiffies, rnp->boost_time))) { | |
1343 | if (rnp->exp_tasks == NULL) | |
1344 | rnp->boost_tasks = rnp->gp_tasks; | |
1217ed1b | 1345 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 PM |
1346 | t = rnp->boost_kthread_task; |
1347 | if (t != NULL) | |
1348 | wake_up_process(t); | |
1217ed1b | 1349 | } else { |
0ea1f2eb | 1350 | rcu_initiate_boost_trace(rnp); |
1217ed1b PM |
1351 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
1352 | } | |
27f4d280 PM |
1353 | } |
1354 | ||
a46e0899 PM |
1355 | /* |
1356 | * Wake up the per-CPU kthread to invoke RCU callbacks. | |
1357 | */ | |
1358 | static void invoke_rcu_callbacks_kthread(void) | |
1359 | { | |
1360 | unsigned long flags; | |
1361 | ||
1362 | local_irq_save(flags); | |
1363 | __this_cpu_write(rcu_cpu_has_work, 1); | |
1eb52121 SL |
1364 | if (__this_cpu_read(rcu_cpu_kthread_task) != NULL && |
1365 | current != __this_cpu_read(rcu_cpu_kthread_task)) | |
1366 | wake_up_process(__this_cpu_read(rcu_cpu_kthread_task)); | |
a46e0899 PM |
1367 | local_irq_restore(flags); |
1368 | } | |
1369 | ||
dff1672d PM |
1370 | /* |
1371 | * Is the current CPU running the RCU-callbacks kthread? | |
1372 | * Caller must have preemption disabled. | |
1373 | */ | |
1374 | static bool rcu_is_callbacks_kthread(void) | |
1375 | { | |
1376 | return __get_cpu_var(rcu_cpu_kthread_task) == current; | |
1377 | } | |
1378 | ||
0f962a5e PM |
1379 | /* |
1380 | * Set the affinity of the boost kthread. The CPU-hotplug locks are | |
1381 | * held, so no one should be messing with the existence of the boost | |
1382 | * kthread. | |
1383 | */ | |
27f4d280 PM |
1384 | static void rcu_boost_kthread_setaffinity(struct rcu_node *rnp, |
1385 | cpumask_var_t cm) | |
1386 | { | |
27f4d280 PM |
1387 | struct task_struct *t; |
1388 | ||
27f4d280 PM |
1389 | t = rnp->boost_kthread_task; |
1390 | if (t != NULL) | |
1391 | set_cpus_allowed_ptr(rnp->boost_kthread_task, cm); | |
27f4d280 PM |
1392 | } |
1393 | ||
1394 | #define RCU_BOOST_DELAY_JIFFIES DIV_ROUND_UP(CONFIG_RCU_BOOST_DELAY * HZ, 1000) | |
1395 | ||
1396 | /* | |
1397 | * Do priority-boost accounting for the start of a new grace period. | |
1398 | */ | |
1399 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) | |
1400 | { | |
1401 | rnp->boost_time = jiffies + RCU_BOOST_DELAY_JIFFIES; | |
1402 | } | |
1403 | ||
27f4d280 PM |
1404 | /* |
1405 | * Create an RCU-boost kthread for the specified node if one does not | |
1406 | * already exist. We only create this kthread for preemptible RCU. | |
1407 | * Returns zero if all is well, a negated errno otherwise. | |
1408 | */ | |
1409 | static int __cpuinit rcu_spawn_one_boost_kthread(struct rcu_state *rsp, | |
1410 | struct rcu_node *rnp, | |
1411 | int rnp_index) | |
1412 | { | |
1413 | unsigned long flags; | |
1414 | struct sched_param sp; | |
1415 | struct task_struct *t; | |
1416 | ||
1417 | if (&rcu_preempt_state != rsp) | |
1418 | return 0; | |
a46e0899 | 1419 | rsp->boost = 1; |
27f4d280 PM |
1420 | if (rnp->boost_kthread_task != NULL) |
1421 | return 0; | |
1422 | t = kthread_create(rcu_boost_kthread, (void *)rnp, | |
5b61b0ba | 1423 | "rcub/%d", rnp_index); |
27f4d280 PM |
1424 | if (IS_ERR(t)) |
1425 | return PTR_ERR(t); | |
1426 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1427 | rnp->boost_kthread_task = t; | |
1428 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
5b61b0ba | 1429 | sp.sched_priority = RCU_BOOST_PRIO; |
27f4d280 | 1430 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); |
9a432736 | 1431 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ |
27f4d280 PM |
1432 | return 0; |
1433 | } | |
1434 | ||
f8b7fc6b PM |
1435 | #ifdef CONFIG_HOTPLUG_CPU |
1436 | ||
1437 | /* | |
1438 | * Stop the RCU's per-CPU kthread when its CPU goes offline,. | |
1439 | */ | |
1440 | static void rcu_stop_cpu_kthread(int cpu) | |
1441 | { | |
1442 | struct task_struct *t; | |
1443 | ||
1444 | /* Stop the CPU's kthread. */ | |
1445 | t = per_cpu(rcu_cpu_kthread_task, cpu); | |
1446 | if (t != NULL) { | |
1447 | per_cpu(rcu_cpu_kthread_task, cpu) = NULL; | |
1448 | kthread_stop(t); | |
1449 | } | |
1450 | } | |
1451 | ||
1452 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1453 | ||
1454 | static void rcu_kthread_do_work(void) | |
1455 | { | |
1456 | rcu_do_batch(&rcu_sched_state, &__get_cpu_var(rcu_sched_data)); | |
1457 | rcu_do_batch(&rcu_bh_state, &__get_cpu_var(rcu_bh_data)); | |
1458 | rcu_preempt_do_callbacks(); | |
1459 | } | |
1460 | ||
1461 | /* | |
1462 | * Wake up the specified per-rcu_node-structure kthread. | |
1463 | * Because the per-rcu_node kthreads are immortal, we don't need | |
1464 | * to do anything to keep them alive. | |
1465 | */ | |
1466 | static void invoke_rcu_node_kthread(struct rcu_node *rnp) | |
1467 | { | |
1468 | struct task_struct *t; | |
1469 | ||
1470 | t = rnp->node_kthread_task; | |
1471 | if (t != NULL) | |
1472 | wake_up_process(t); | |
1473 | } | |
1474 | ||
1475 | /* | |
1476 | * Set the specified CPU's kthread to run RT or not, as specified by | |
1477 | * the to_rt argument. The CPU-hotplug locks are held, so the task | |
1478 | * is not going away. | |
1479 | */ | |
1480 | static void rcu_cpu_kthread_setrt(int cpu, int to_rt) | |
1481 | { | |
1482 | int policy; | |
1483 | struct sched_param sp; | |
1484 | struct task_struct *t; | |
1485 | ||
1486 | t = per_cpu(rcu_cpu_kthread_task, cpu); | |
1487 | if (t == NULL) | |
1488 | return; | |
1489 | if (to_rt) { | |
1490 | policy = SCHED_FIFO; | |
1491 | sp.sched_priority = RCU_KTHREAD_PRIO; | |
1492 | } else { | |
1493 | policy = SCHED_NORMAL; | |
1494 | sp.sched_priority = 0; | |
1495 | } | |
1496 | sched_setscheduler_nocheck(t, policy, &sp); | |
1497 | } | |
1498 | ||
1499 | /* | |
1500 | * Timer handler to initiate the waking up of per-CPU kthreads that | |
1501 | * have yielded the CPU due to excess numbers of RCU callbacks. | |
1502 | * We wake up the per-rcu_node kthread, which in turn will wake up | |
1503 | * the booster kthread. | |
1504 | */ | |
1505 | static void rcu_cpu_kthread_timer(unsigned long arg) | |
1506 | { | |
1507 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, arg); | |
1508 | struct rcu_node *rnp = rdp->mynode; | |
1509 | ||
1510 | atomic_or(rdp->grpmask, &rnp->wakemask); | |
1511 | invoke_rcu_node_kthread(rnp); | |
1512 | } | |
1513 | ||
1514 | /* | |
1515 | * Drop to non-real-time priority and yield, but only after posting a | |
1516 | * timer that will cause us to regain our real-time priority if we | |
1517 | * remain preempted. Either way, we restore our real-time priority | |
1518 | * before returning. | |
1519 | */ | |
1520 | static void rcu_yield(void (*f)(unsigned long), unsigned long arg) | |
1521 | { | |
1522 | struct sched_param sp; | |
1523 | struct timer_list yield_timer; | |
5b61b0ba | 1524 | int prio = current->rt_priority; |
f8b7fc6b PM |
1525 | |
1526 | setup_timer_on_stack(&yield_timer, f, arg); | |
1527 | mod_timer(&yield_timer, jiffies + 2); | |
1528 | sp.sched_priority = 0; | |
1529 | sched_setscheduler_nocheck(current, SCHED_NORMAL, &sp); | |
1530 | set_user_nice(current, 19); | |
1531 | schedule(); | |
5b61b0ba MG |
1532 | set_user_nice(current, 0); |
1533 | sp.sched_priority = prio; | |
f8b7fc6b PM |
1534 | sched_setscheduler_nocheck(current, SCHED_FIFO, &sp); |
1535 | del_timer(&yield_timer); | |
1536 | } | |
1537 | ||
1538 | /* | |
1539 | * Handle cases where the rcu_cpu_kthread() ends up on the wrong CPU. | |
1540 | * This can happen while the corresponding CPU is either coming online | |
1541 | * or going offline. We cannot wait until the CPU is fully online | |
1542 | * before starting the kthread, because the various notifier functions | |
1543 | * can wait for RCU grace periods. So we park rcu_cpu_kthread() until | |
1544 | * the corresponding CPU is online. | |
1545 | * | |
1546 | * Return 1 if the kthread needs to stop, 0 otherwise. | |
1547 | * | |
1548 | * Caller must disable bh. This function can momentarily enable it. | |
1549 | */ | |
1550 | static int rcu_cpu_kthread_should_stop(int cpu) | |
1551 | { | |
1552 | while (cpu_is_offline(cpu) || | |
1553 | !cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu)) || | |
1554 | smp_processor_id() != cpu) { | |
1555 | if (kthread_should_stop()) | |
1556 | return 1; | |
1557 | per_cpu(rcu_cpu_kthread_status, cpu) = RCU_KTHREAD_OFFCPU; | |
1558 | per_cpu(rcu_cpu_kthread_cpu, cpu) = raw_smp_processor_id(); | |
1559 | local_bh_enable(); | |
1560 | schedule_timeout_uninterruptible(1); | |
1561 | if (!cpumask_equal(¤t->cpus_allowed, cpumask_of(cpu))) | |
1562 | set_cpus_allowed_ptr(current, cpumask_of(cpu)); | |
1563 | local_bh_disable(); | |
1564 | } | |
1565 | per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu; | |
1566 | return 0; | |
1567 | } | |
1568 | ||
1569 | /* | |
1570 | * Per-CPU kernel thread that invokes RCU callbacks. This replaces the | |
e0f23060 PM |
1571 | * RCU softirq used in flavors and configurations of RCU that do not |
1572 | * support RCU priority boosting. | |
f8b7fc6b PM |
1573 | */ |
1574 | static int rcu_cpu_kthread(void *arg) | |
1575 | { | |
1576 | int cpu = (int)(long)arg; | |
1577 | unsigned long flags; | |
1578 | int spincnt = 0; | |
1579 | unsigned int *statusp = &per_cpu(rcu_cpu_kthread_status, cpu); | |
1580 | char work; | |
1581 | char *workp = &per_cpu(rcu_cpu_has_work, cpu); | |
1582 | ||
385680a9 | 1583 | trace_rcu_utilization("Start CPU kthread@init"); |
f8b7fc6b PM |
1584 | for (;;) { |
1585 | *statusp = RCU_KTHREAD_WAITING; | |
385680a9 | 1586 | trace_rcu_utilization("End CPU kthread@rcu_wait"); |
f8b7fc6b | 1587 | rcu_wait(*workp != 0 || kthread_should_stop()); |
385680a9 | 1588 | trace_rcu_utilization("Start CPU kthread@rcu_wait"); |
f8b7fc6b PM |
1589 | local_bh_disable(); |
1590 | if (rcu_cpu_kthread_should_stop(cpu)) { | |
1591 | local_bh_enable(); | |
1592 | break; | |
1593 | } | |
1594 | *statusp = RCU_KTHREAD_RUNNING; | |
1595 | per_cpu(rcu_cpu_kthread_loops, cpu)++; | |
1596 | local_irq_save(flags); | |
1597 | work = *workp; | |
1598 | *workp = 0; | |
1599 | local_irq_restore(flags); | |
1600 | if (work) | |
1601 | rcu_kthread_do_work(); | |
1602 | local_bh_enable(); | |
1603 | if (*workp != 0) | |
1604 | spincnt++; | |
1605 | else | |
1606 | spincnt = 0; | |
1607 | if (spincnt > 10) { | |
1608 | *statusp = RCU_KTHREAD_YIELDING; | |
385680a9 | 1609 | trace_rcu_utilization("End CPU kthread@rcu_yield"); |
f8b7fc6b | 1610 | rcu_yield(rcu_cpu_kthread_timer, (unsigned long)cpu); |
385680a9 | 1611 | trace_rcu_utilization("Start CPU kthread@rcu_yield"); |
f8b7fc6b PM |
1612 | spincnt = 0; |
1613 | } | |
1614 | } | |
1615 | *statusp = RCU_KTHREAD_STOPPED; | |
385680a9 | 1616 | trace_rcu_utilization("End CPU kthread@term"); |
f8b7fc6b PM |
1617 | return 0; |
1618 | } | |
1619 | ||
1620 | /* | |
1621 | * Spawn a per-CPU kthread, setting up affinity and priority. | |
1622 | * Because the CPU hotplug lock is held, no other CPU will be attempting | |
1623 | * to manipulate rcu_cpu_kthread_task. There might be another CPU | |
1624 | * attempting to access it during boot, but the locking in kthread_bind() | |
1625 | * will enforce sufficient ordering. | |
1626 | * | |
1627 | * Please note that we cannot simply refuse to wake up the per-CPU | |
1628 | * kthread because kthreads are created in TASK_UNINTERRUPTIBLE state, | |
1629 | * which can result in softlockup complaints if the task ends up being | |
1630 | * idle for more than a couple of minutes. | |
1631 | * | |
1632 | * However, please note also that we cannot bind the per-CPU kthread to its | |
1633 | * CPU until that CPU is fully online. We also cannot wait until the | |
1634 | * CPU is fully online before we create its per-CPU kthread, as this would | |
1635 | * deadlock the system when CPU notifiers tried waiting for grace | |
1636 | * periods. So we bind the per-CPU kthread to its CPU only if the CPU | |
1637 | * is online. If its CPU is not yet fully online, then the code in | |
1638 | * rcu_cpu_kthread() will wait until it is fully online, and then do | |
1639 | * the binding. | |
1640 | */ | |
1641 | static int __cpuinit rcu_spawn_one_cpu_kthread(int cpu) | |
1642 | { | |
1643 | struct sched_param sp; | |
1644 | struct task_struct *t; | |
1645 | ||
b0d30417 | 1646 | if (!rcu_scheduler_fully_active || |
f8b7fc6b PM |
1647 | per_cpu(rcu_cpu_kthread_task, cpu) != NULL) |
1648 | return 0; | |
1f288094 ED |
1649 | t = kthread_create_on_node(rcu_cpu_kthread, |
1650 | (void *)(long)cpu, | |
1651 | cpu_to_node(cpu), | |
5b61b0ba | 1652 | "rcuc/%d", cpu); |
f8b7fc6b PM |
1653 | if (IS_ERR(t)) |
1654 | return PTR_ERR(t); | |
1655 | if (cpu_online(cpu)) | |
1656 | kthread_bind(t, cpu); | |
1657 | per_cpu(rcu_cpu_kthread_cpu, cpu) = cpu; | |
1658 | WARN_ON_ONCE(per_cpu(rcu_cpu_kthread_task, cpu) != NULL); | |
1659 | sp.sched_priority = RCU_KTHREAD_PRIO; | |
1660 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
1661 | per_cpu(rcu_cpu_kthread_task, cpu) = t; | |
1662 | wake_up_process(t); /* Get to TASK_INTERRUPTIBLE quickly. */ | |
1663 | return 0; | |
1664 | } | |
1665 | ||
1666 | /* | |
1667 | * Per-rcu_node kthread, which is in charge of waking up the per-CPU | |
1668 | * kthreads when needed. We ignore requests to wake up kthreads | |
1669 | * for offline CPUs, which is OK because force_quiescent_state() | |
1670 | * takes care of this case. | |
1671 | */ | |
1672 | static int rcu_node_kthread(void *arg) | |
1673 | { | |
1674 | int cpu; | |
1675 | unsigned long flags; | |
1676 | unsigned long mask; | |
1677 | struct rcu_node *rnp = (struct rcu_node *)arg; | |
1678 | struct sched_param sp; | |
1679 | struct task_struct *t; | |
1680 | ||
1681 | for (;;) { | |
1682 | rnp->node_kthread_status = RCU_KTHREAD_WAITING; | |
1683 | rcu_wait(atomic_read(&rnp->wakemask) != 0); | |
1684 | rnp->node_kthread_status = RCU_KTHREAD_RUNNING; | |
1685 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1686 | mask = atomic_xchg(&rnp->wakemask, 0); | |
1687 | rcu_initiate_boost(rnp, flags); /* releases rnp->lock. */ | |
1688 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) { | |
1689 | if ((mask & 0x1) == 0) | |
1690 | continue; | |
1691 | preempt_disable(); | |
1692 | t = per_cpu(rcu_cpu_kthread_task, cpu); | |
1693 | if (!cpu_online(cpu) || t == NULL) { | |
1694 | preempt_enable(); | |
1695 | continue; | |
1696 | } | |
1697 | per_cpu(rcu_cpu_has_work, cpu) = 1; | |
1698 | sp.sched_priority = RCU_KTHREAD_PRIO; | |
1699 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
1700 | preempt_enable(); | |
1701 | } | |
1702 | } | |
1703 | /* NOTREACHED */ | |
1704 | rnp->node_kthread_status = RCU_KTHREAD_STOPPED; | |
1705 | return 0; | |
1706 | } | |
1707 | ||
1708 | /* | |
1709 | * Set the per-rcu_node kthread's affinity to cover all CPUs that are | |
1710 | * served by the rcu_node in question. The CPU hotplug lock is still | |
1711 | * held, so the value of rnp->qsmaskinit will be stable. | |
1712 | * | |
1713 | * We don't include outgoingcpu in the affinity set, use -1 if there is | |
1714 | * no outgoing CPU. If there are no CPUs left in the affinity set, | |
1715 | * this function allows the kthread to execute on any CPU. | |
1716 | */ | |
1717 | static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) | |
1718 | { | |
1719 | cpumask_var_t cm; | |
1720 | int cpu; | |
1721 | unsigned long mask = rnp->qsmaskinit; | |
1722 | ||
1723 | if (rnp->node_kthread_task == NULL) | |
1724 | return; | |
1725 | if (!alloc_cpumask_var(&cm, GFP_KERNEL)) | |
1726 | return; | |
1727 | cpumask_clear(cm); | |
1728 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++, mask >>= 1) | |
1729 | if ((mask & 0x1) && cpu != outgoingcpu) | |
1730 | cpumask_set_cpu(cpu, cm); | |
1731 | if (cpumask_weight(cm) == 0) { | |
1732 | cpumask_setall(cm); | |
1733 | for (cpu = rnp->grplo; cpu <= rnp->grphi; cpu++) | |
1734 | cpumask_clear_cpu(cpu, cm); | |
1735 | WARN_ON_ONCE(cpumask_weight(cm) == 0); | |
1736 | } | |
1737 | set_cpus_allowed_ptr(rnp->node_kthread_task, cm); | |
1738 | rcu_boost_kthread_setaffinity(rnp, cm); | |
1739 | free_cpumask_var(cm); | |
1740 | } | |
1741 | ||
1742 | /* | |
1743 | * Spawn a per-rcu_node kthread, setting priority and affinity. | |
1744 | * Called during boot before online/offline can happen, or, if | |
1745 | * during runtime, with the main CPU-hotplug locks held. So only | |
1746 | * one of these can be executing at a time. | |
1747 | */ | |
1748 | static int __cpuinit rcu_spawn_one_node_kthread(struct rcu_state *rsp, | |
1749 | struct rcu_node *rnp) | |
1750 | { | |
1751 | unsigned long flags; | |
1752 | int rnp_index = rnp - &rsp->node[0]; | |
1753 | struct sched_param sp; | |
1754 | struct task_struct *t; | |
1755 | ||
b0d30417 | 1756 | if (!rcu_scheduler_fully_active || |
f8b7fc6b PM |
1757 | rnp->qsmaskinit == 0) |
1758 | return 0; | |
1759 | if (rnp->node_kthread_task == NULL) { | |
1760 | t = kthread_create(rcu_node_kthread, (void *)rnp, | |
5b61b0ba | 1761 | "rcun/%d", rnp_index); |
f8b7fc6b PM |
1762 | if (IS_ERR(t)) |
1763 | return PTR_ERR(t); | |
1764 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
1765 | rnp->node_kthread_task = t; | |
1766 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
1767 | sp.sched_priority = 99; | |
1768 | sched_setscheduler_nocheck(t, SCHED_FIFO, &sp); | |
1769 | wake_up_process(t); /* get to TASK_INTERRUPTIBLE quickly. */ | |
1770 | } | |
1771 | return rcu_spawn_one_boost_kthread(rsp, rnp, rnp_index); | |
1772 | } | |
1773 | ||
1774 | /* | |
1775 | * Spawn all kthreads -- called as soon as the scheduler is running. | |
1776 | */ | |
1777 | static int __init rcu_spawn_kthreads(void) | |
1778 | { | |
1779 | int cpu; | |
1780 | struct rcu_node *rnp; | |
1781 | ||
b0d30417 | 1782 | rcu_scheduler_fully_active = 1; |
f8b7fc6b PM |
1783 | for_each_possible_cpu(cpu) { |
1784 | per_cpu(rcu_cpu_has_work, cpu) = 0; | |
1785 | if (cpu_online(cpu)) | |
1786 | (void)rcu_spawn_one_cpu_kthread(cpu); | |
1787 | } | |
1788 | rnp = rcu_get_root(rcu_state); | |
1789 | (void)rcu_spawn_one_node_kthread(rcu_state, rnp); | |
1790 | if (NUM_RCU_NODES > 1) { | |
1791 | rcu_for_each_leaf_node(rcu_state, rnp) | |
1792 | (void)rcu_spawn_one_node_kthread(rcu_state, rnp); | |
1793 | } | |
1794 | return 0; | |
1795 | } | |
1796 | early_initcall(rcu_spawn_kthreads); | |
1797 | ||
1798 | static void __cpuinit rcu_prepare_kthreads(int cpu) | |
1799 | { | |
1800 | struct rcu_data *rdp = per_cpu_ptr(rcu_state->rda, cpu); | |
1801 | struct rcu_node *rnp = rdp->mynode; | |
1802 | ||
1803 | /* Fire up the incoming CPU's kthread and leaf rcu_node kthread. */ | |
b0d30417 | 1804 | if (rcu_scheduler_fully_active) { |
f8b7fc6b PM |
1805 | (void)rcu_spawn_one_cpu_kthread(cpu); |
1806 | if (rnp->node_kthread_task == NULL) | |
1807 | (void)rcu_spawn_one_node_kthread(rcu_state, rnp); | |
1808 | } | |
1809 | } | |
1810 | ||
27f4d280 PM |
1811 | #else /* #ifdef CONFIG_RCU_BOOST */ |
1812 | ||
1217ed1b | 1813 | static void rcu_initiate_boost(struct rcu_node *rnp, unsigned long flags) |
27f4d280 | 1814 | { |
1217ed1b | 1815 | raw_spin_unlock_irqrestore(&rnp->lock, flags); |
27f4d280 PM |
1816 | } |
1817 | ||
a46e0899 | 1818 | static void invoke_rcu_callbacks_kthread(void) |
27f4d280 | 1819 | { |
a46e0899 | 1820 | WARN_ON_ONCE(1); |
27f4d280 PM |
1821 | } |
1822 | ||
dff1672d PM |
1823 | static bool rcu_is_callbacks_kthread(void) |
1824 | { | |
1825 | return false; | |
1826 | } | |
1827 | ||
27f4d280 PM |
1828 | static void rcu_preempt_boost_start_gp(struct rcu_node *rnp) |
1829 | { | |
1830 | } | |
1831 | ||
f8b7fc6b PM |
1832 | #ifdef CONFIG_HOTPLUG_CPU |
1833 | ||
1834 | static void rcu_stop_cpu_kthread(int cpu) | |
1835 | { | |
1836 | } | |
1837 | ||
1838 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
1839 | ||
1840 | static void rcu_node_kthread_setaffinity(struct rcu_node *rnp, int outgoingcpu) | |
1841 | { | |
1842 | } | |
1843 | ||
1844 | static void rcu_cpu_kthread_setrt(int cpu, int to_rt) | |
1845 | { | |
1846 | } | |
1847 | ||
b0d30417 PM |
1848 | static int __init rcu_scheduler_really_started(void) |
1849 | { | |
1850 | rcu_scheduler_fully_active = 1; | |
1851 | return 0; | |
1852 | } | |
1853 | early_initcall(rcu_scheduler_really_started); | |
1854 | ||
f8b7fc6b PM |
1855 | static void __cpuinit rcu_prepare_kthreads(int cpu) |
1856 | { | |
1857 | } | |
1858 | ||
27f4d280 PM |
1859 | #endif /* #else #ifdef CONFIG_RCU_BOOST */ |
1860 | ||
7b27d547 LJ |
1861 | #ifndef CONFIG_SMP |
1862 | ||
1863 | void synchronize_sched_expedited(void) | |
1864 | { | |
1865 | cond_resched(); | |
1866 | } | |
1867 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
1868 | ||
1869 | #else /* #ifndef CONFIG_SMP */ | |
1870 | ||
e27fc964 TH |
1871 | static atomic_t sync_sched_expedited_started = ATOMIC_INIT(0); |
1872 | static atomic_t sync_sched_expedited_done = ATOMIC_INIT(0); | |
7b27d547 LJ |
1873 | |
1874 | static int synchronize_sched_expedited_cpu_stop(void *data) | |
1875 | { | |
1876 | /* | |
1877 | * There must be a full memory barrier on each affected CPU | |
1878 | * between the time that try_stop_cpus() is called and the | |
1879 | * time that it returns. | |
1880 | * | |
1881 | * In the current initial implementation of cpu_stop, the | |
1882 | * above condition is already met when the control reaches | |
1883 | * this point and the following smp_mb() is not strictly | |
1884 | * necessary. Do smp_mb() anyway for documentation and | |
1885 | * robustness against future implementation changes. | |
1886 | */ | |
1887 | smp_mb(); /* See above comment block. */ | |
1888 | return 0; | |
1889 | } | |
1890 | ||
1891 | /* | |
1892 | * Wait for an rcu-sched grace period to elapse, but use "big hammer" | |
1893 | * approach to force grace period to end quickly. This consumes | |
1894 | * significant time on all CPUs, and is thus not recommended for | |
1895 | * any sort of common-case code. | |
1896 | * | |
1897 | * Note that it is illegal to call this function while holding any | |
1898 | * lock that is acquired by a CPU-hotplug notifier. Failing to | |
1899 | * observe this restriction will result in deadlock. | |
db3a8920 | 1900 | * |
e27fc964 TH |
1901 | * This implementation can be thought of as an application of ticket |
1902 | * locking to RCU, with sync_sched_expedited_started and | |
1903 | * sync_sched_expedited_done taking on the roles of the halves | |
1904 | * of the ticket-lock word. Each task atomically increments | |
1905 | * sync_sched_expedited_started upon entry, snapshotting the old value, | |
1906 | * then attempts to stop all the CPUs. If this succeeds, then each | |
1907 | * CPU will have executed a context switch, resulting in an RCU-sched | |
1908 | * grace period. We are then done, so we use atomic_cmpxchg() to | |
1909 | * update sync_sched_expedited_done to match our snapshot -- but | |
1910 | * only if someone else has not already advanced past our snapshot. | |
1911 | * | |
1912 | * On the other hand, if try_stop_cpus() fails, we check the value | |
1913 | * of sync_sched_expedited_done. If it has advanced past our | |
1914 | * initial snapshot, then someone else must have forced a grace period | |
1915 | * some time after we took our snapshot. In this case, our work is | |
1916 | * done for us, and we can simply return. Otherwise, we try again, | |
1917 | * but keep our initial snapshot for purposes of checking for someone | |
1918 | * doing our work for us. | |
1919 | * | |
1920 | * If we fail too many times in a row, we fall back to synchronize_sched(). | |
7b27d547 LJ |
1921 | */ |
1922 | void synchronize_sched_expedited(void) | |
1923 | { | |
e27fc964 | 1924 | int firstsnap, s, snap, trycount = 0; |
7b27d547 | 1925 | |
e27fc964 TH |
1926 | /* Note that atomic_inc_return() implies full memory barrier. */ |
1927 | firstsnap = snap = atomic_inc_return(&sync_sched_expedited_started); | |
7b27d547 | 1928 | get_online_cpus(); |
e27fc964 TH |
1929 | |
1930 | /* | |
1931 | * Each pass through the following loop attempts to force a | |
1932 | * context switch on each CPU. | |
1933 | */ | |
7b27d547 LJ |
1934 | while (try_stop_cpus(cpu_online_mask, |
1935 | synchronize_sched_expedited_cpu_stop, | |
1936 | NULL) == -EAGAIN) { | |
1937 | put_online_cpus(); | |
e27fc964 TH |
1938 | |
1939 | /* No joy, try again later. Or just synchronize_sched(). */ | |
7b27d547 LJ |
1940 | if (trycount++ < 10) |
1941 | udelay(trycount * num_online_cpus()); | |
1942 | else { | |
1943 | synchronize_sched(); | |
1944 | return; | |
1945 | } | |
e27fc964 TH |
1946 | |
1947 | /* Check to see if someone else did our work for us. */ | |
1948 | s = atomic_read(&sync_sched_expedited_done); | |
1949 | if (UINT_CMP_GE((unsigned)s, (unsigned)firstsnap)) { | |
7b27d547 LJ |
1950 | smp_mb(); /* ensure test happens before caller kfree */ |
1951 | return; | |
1952 | } | |
e27fc964 TH |
1953 | |
1954 | /* | |
1955 | * Refetching sync_sched_expedited_started allows later | |
1956 | * callers to piggyback on our grace period. We subtract | |
1957 | * 1 to get the same token that the last incrementer got. | |
1958 | * We retry after they started, so our grace period works | |
1959 | * for them, and they started after our first try, so their | |
1960 | * grace period works for us. | |
1961 | */ | |
7b27d547 | 1962 | get_online_cpus(); |
7077714e | 1963 | snap = atomic_read(&sync_sched_expedited_started); |
e27fc964 | 1964 | smp_mb(); /* ensure read is before try_stop_cpus(). */ |
7b27d547 | 1965 | } |
e27fc964 TH |
1966 | |
1967 | /* | |
1968 | * Everyone up to our most recent fetch is covered by our grace | |
1969 | * period. Update the counter, but only if our work is still | |
1970 | * relevant -- which it won't be if someone who started later | |
1971 | * than we did beat us to the punch. | |
1972 | */ | |
1973 | do { | |
1974 | s = atomic_read(&sync_sched_expedited_done); | |
1975 | if (UINT_CMP_GE((unsigned)s, (unsigned)snap)) { | |
1976 | smp_mb(); /* ensure test happens before caller kfree */ | |
1977 | break; | |
1978 | } | |
1979 | } while (atomic_cmpxchg(&sync_sched_expedited_done, s, snap) != s); | |
1980 | ||
7b27d547 LJ |
1981 | put_online_cpus(); |
1982 | } | |
1983 | EXPORT_SYMBOL_GPL(synchronize_sched_expedited); | |
1984 | ||
1985 | #endif /* #else #ifndef CONFIG_SMP */ | |
1986 | ||
8bd93a2c PM |
1987 | #if !defined(CONFIG_RCU_FAST_NO_HZ) |
1988 | ||
1989 | /* | |
1990 | * Check to see if any future RCU-related work will need to be done | |
1991 | * by the current CPU, even if none need be done immediately, returning | |
1992 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
1993 | * an exported member of the RCU API. | |
1994 | * | |
7cb92499 PM |
1995 | * Because we not have RCU_FAST_NO_HZ, just check whether this CPU needs |
1996 | * any flavor of RCU. | |
8bd93a2c PM |
1997 | */ |
1998 | int rcu_needs_cpu(int cpu) | |
1999 | { | |
aea1b35e PM |
2000 | return rcu_cpu_has_callbacks(cpu); |
2001 | } | |
2002 | ||
7cb92499 PM |
2003 | /* |
2004 | * Because we do not have RCU_FAST_NO_HZ, don't bother initializing for it. | |
2005 | */ | |
2006 | static void rcu_prepare_for_idle_init(int cpu) | |
2007 | { | |
2008 | } | |
2009 | ||
2010 | /* | |
2011 | * Because we do not have RCU_FAST_NO_HZ, don't bother cleaning up | |
2012 | * after it. | |
2013 | */ | |
2014 | static void rcu_cleanup_after_idle(int cpu) | |
2015 | { | |
2016 | } | |
2017 | ||
aea1b35e PM |
2018 | /* |
2019 | * Do the idle-entry grace-period work, which, because CONFIG_RCU_FAST_NO_HZ=y, | |
2020 | * is nothing. | |
2021 | */ | |
2022 | static void rcu_prepare_for_idle(int cpu) | |
2023 | { | |
2024 | } | |
2025 | ||
8bd93a2c PM |
2026 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |
2027 | ||
f23f7fa1 PM |
2028 | /* |
2029 | * This code is invoked when a CPU goes idle, at which point we want | |
2030 | * to have the CPU do everything required for RCU so that it can enter | |
2031 | * the energy-efficient dyntick-idle mode. This is handled by a | |
2032 | * state machine implemented by rcu_prepare_for_idle() below. | |
2033 | * | |
2034 | * The following three proprocessor symbols control this state machine: | |
2035 | * | |
2036 | * RCU_IDLE_FLUSHES gives the maximum number of times that we will attempt | |
2037 | * to satisfy RCU. Beyond this point, it is better to incur a periodic | |
2038 | * scheduling-clock interrupt than to loop through the state machine | |
2039 | * at full power. | |
2040 | * RCU_IDLE_OPT_FLUSHES gives the number of RCU_IDLE_FLUSHES that are | |
2041 | * optional if RCU does not need anything immediately from this | |
2042 | * CPU, even if this CPU still has RCU callbacks queued. The first | |
2043 | * times through the state machine are mandatory: we need to give | |
2044 | * the state machine a chance to communicate a quiescent state | |
2045 | * to the RCU core. | |
2046 | * RCU_IDLE_GP_DELAY gives the number of jiffies that a CPU is permitted | |
2047 | * to sleep in dyntick-idle mode with RCU callbacks pending. This | |
2048 | * is sized to be roughly one RCU grace period. Those energy-efficiency | |
2049 | * benchmarkers who might otherwise be tempted to set this to a large | |
2050 | * number, be warned: Setting RCU_IDLE_GP_DELAY too high can hang your | |
2051 | * system. And if you are -that- concerned about energy efficiency, | |
2052 | * just power the system down and be done with it! | |
778d250a PM |
2053 | * RCU_IDLE_LAZY_GP_DELAY gives the number of jiffies that a CPU is |
2054 | * permitted to sleep in dyntick-idle mode with only lazy RCU | |
2055 | * callbacks pending. Setting this too high can OOM your system. | |
f23f7fa1 PM |
2056 | * |
2057 | * The values below work well in practice. If future workloads require | |
2058 | * adjustment, they can be converted into kernel config parameters, though | |
2059 | * making the state machine smarter might be a better option. | |
2060 | */ | |
2061 | #define RCU_IDLE_FLUSHES 5 /* Number of dyntick-idle tries. */ | |
2062 | #define RCU_IDLE_OPT_FLUSHES 3 /* Optional dyntick-idle tries. */ | |
7cb92499 | 2063 | #define RCU_IDLE_GP_DELAY 6 /* Roughly one grace period. */ |
778d250a | 2064 | #define RCU_IDLE_LAZY_GP_DELAY (6 * HZ) /* Roughly six seconds. */ |
f23f7fa1 | 2065 | |
a47cd880 | 2066 | static DEFINE_PER_CPU(int, rcu_dyntick_drain); |
71da8132 | 2067 | static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff); |
7cb92499 | 2068 | static DEFINE_PER_CPU(struct hrtimer, rcu_idle_gp_timer); |
778d250a PM |
2069 | static ktime_t rcu_idle_gp_wait; /* If some non-lazy callbacks. */ |
2070 | static ktime_t rcu_idle_lazy_gp_wait; /* If only lazy callbacks. */ | |
8bd93a2c PM |
2071 | |
2072 | /* | |
aea1b35e PM |
2073 | * Allow the CPU to enter dyntick-idle mode if either: (1) There are no |
2074 | * callbacks on this CPU, (2) this CPU has not yet attempted to enter | |
2075 | * dyntick-idle mode, or (3) this CPU is in the process of attempting to | |
2076 | * enter dyntick-idle mode. Otherwise, if we have recently tried and failed | |
2077 | * to enter dyntick-idle mode, we refuse to try to enter it. After all, | |
2078 | * it is better to incur scheduling-clock interrupts than to spin | |
2079 | * continuously for the same time duration! | |
2080 | */ | |
2081 | int rcu_needs_cpu(int cpu) | |
2082 | { | |
2083 | /* If no callbacks, RCU doesn't need the CPU. */ | |
2084 | if (!rcu_cpu_has_callbacks(cpu)) | |
2085 | return 0; | |
2086 | /* Otherwise, RCU needs the CPU only if it recently tried and failed. */ | |
2087 | return per_cpu(rcu_dyntick_holdoff, cpu) == jiffies; | |
2088 | } | |
2089 | ||
486e2593 PM |
2090 | /* |
2091 | * Does the specified flavor of RCU have non-lazy callbacks pending on | |
2092 | * the specified CPU? Both RCU flavor and CPU are specified by the | |
2093 | * rcu_data structure. | |
2094 | */ | |
2095 | static bool __rcu_cpu_has_nonlazy_callbacks(struct rcu_data *rdp) | |
2096 | { | |
2097 | return rdp->qlen != rdp->qlen_lazy; | |
2098 | } | |
2099 | ||
2100 | #ifdef CONFIG_TREE_PREEMPT_RCU | |
2101 | ||
2102 | /* | |
2103 | * Are there non-lazy RCU-preempt callbacks? (There cannot be if there | |
2104 | * is no RCU-preempt in the kernel.) | |
2105 | */ | |
2106 | static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu) | |
2107 | { | |
2108 | struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu); | |
2109 | ||
2110 | return __rcu_cpu_has_nonlazy_callbacks(rdp); | |
2111 | } | |
2112 | ||
2113 | #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */ | |
2114 | ||
2115 | static bool rcu_preempt_cpu_has_nonlazy_callbacks(int cpu) | |
2116 | { | |
2117 | return 0; | |
2118 | } | |
2119 | ||
2120 | #endif /* else #ifdef CONFIG_TREE_PREEMPT_RCU */ | |
2121 | ||
2122 | /* | |
2123 | * Does any flavor of RCU have non-lazy callbacks on the specified CPU? | |
2124 | */ | |
2125 | static bool rcu_cpu_has_nonlazy_callbacks(int cpu) | |
2126 | { | |
2127 | return __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_sched_data, cpu)) || | |
2128 | __rcu_cpu_has_nonlazy_callbacks(&per_cpu(rcu_bh_data, cpu)) || | |
2129 | rcu_preempt_cpu_has_nonlazy_callbacks(cpu); | |
2130 | } | |
2131 | ||
7cb92499 PM |
2132 | /* |
2133 | * Timer handler used to force CPU to start pushing its remaining RCU | |
2134 | * callbacks in the case where it entered dyntick-idle mode with callbacks | |
2135 | * pending. The hander doesn't really need to do anything because the | |
2136 | * real work is done upon re-entry to idle, or by the next scheduling-clock | |
2137 | * interrupt should idle not be re-entered. | |
2138 | */ | |
2139 | static enum hrtimer_restart rcu_idle_gp_timer_func(struct hrtimer *hrtp) | |
2140 | { | |
2141 | trace_rcu_prep_idle("Timer"); | |
2142 | return HRTIMER_NORESTART; | |
2143 | } | |
2144 | ||
2145 | /* | |
2146 | * Initialize the timer used to pull CPUs out of dyntick-idle mode. | |
2147 | */ | |
2148 | static void rcu_prepare_for_idle_init(int cpu) | |
2149 | { | |
2150 | static int firsttime = 1; | |
2151 | struct hrtimer *hrtp = &per_cpu(rcu_idle_gp_timer, cpu); | |
2152 | ||
2153 | hrtimer_init(hrtp, CLOCK_MONOTONIC, HRTIMER_MODE_REL); | |
2154 | hrtp->function = rcu_idle_gp_timer_func; | |
2155 | if (firsttime) { | |
2156 | unsigned int upj = jiffies_to_usecs(RCU_IDLE_GP_DELAY); | |
2157 | ||
2158 | rcu_idle_gp_wait = ns_to_ktime(upj * (u64)1000); | |
778d250a PM |
2159 | upj = jiffies_to_usecs(RCU_IDLE_LAZY_GP_DELAY); |
2160 | rcu_idle_lazy_gp_wait = ns_to_ktime(upj * (u64)1000); | |
7cb92499 PM |
2161 | firsttime = 0; |
2162 | } | |
2163 | } | |
2164 | ||
2165 | /* | |
2166 | * Clean up for exit from idle. Because we are exiting from idle, there | |
2167 | * is no longer any point to rcu_idle_gp_timer, so cancel it. This will | |
2168 | * do nothing if this timer is not active, so just cancel it unconditionally. | |
2169 | */ | |
2170 | static void rcu_cleanup_after_idle(int cpu) | |
2171 | { | |
2172 | hrtimer_cancel(&per_cpu(rcu_idle_gp_timer, cpu)); | |
2173 | } | |
2174 | ||
aea1b35e PM |
2175 | /* |
2176 | * Check to see if any RCU-related work can be done by the current CPU, | |
2177 | * and if so, schedule a softirq to get it done. This function is part | |
2178 | * of the RCU implementation; it is -not- an exported member of the RCU API. | |
8bd93a2c | 2179 | * |
aea1b35e PM |
2180 | * The idea is for the current CPU to clear out all work required by the |
2181 | * RCU core for the current grace period, so that this CPU can be permitted | |
2182 | * to enter dyntick-idle mode. In some cases, it will need to be awakened | |
2183 | * at the end of the grace period by whatever CPU ends the grace period. | |
2184 | * This allows CPUs to go dyntick-idle more quickly, and to reduce the | |
2185 | * number of wakeups by a modest integer factor. | |
a47cd880 PM |
2186 | * |
2187 | * Because it is not legal to invoke rcu_process_callbacks() with irqs | |
2188 | * disabled, we do one pass of force_quiescent_state(), then do a | |
a46e0899 | 2189 | * invoke_rcu_core() to cause rcu_process_callbacks() to be invoked |
27f4d280 | 2190 | * later. The per-cpu rcu_dyntick_drain variable controls the sequencing. |
aea1b35e PM |
2191 | * |
2192 | * The caller must have disabled interrupts. | |
8bd93a2c | 2193 | */ |
aea1b35e | 2194 | static void rcu_prepare_for_idle(int cpu) |
8bd93a2c | 2195 | { |
84ad00cb PM |
2196 | unsigned long flags; |
2197 | ||
2198 | local_irq_save(flags); | |
8bd93a2c | 2199 | |
3084f2f8 | 2200 | /* |
f535a607 PM |
2201 | * If there are no callbacks on this CPU, enter dyntick-idle mode. |
2202 | * Also reset state to avoid prejudicing later attempts. | |
3084f2f8 | 2203 | */ |
aea1b35e PM |
2204 | if (!rcu_cpu_has_callbacks(cpu)) { |
2205 | per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1; | |
3084f2f8 | 2206 | per_cpu(rcu_dyntick_drain, cpu) = 0; |
84ad00cb | 2207 | local_irq_restore(flags); |
433cdddc | 2208 | trace_rcu_prep_idle("No callbacks"); |
aea1b35e | 2209 | return; |
77e38ed3 | 2210 | } |
3084f2f8 PM |
2211 | |
2212 | /* | |
2213 | * If in holdoff mode, just return. We will presumably have | |
2214 | * refrained from disabling the scheduling-clock tick. | |
2215 | */ | |
433cdddc | 2216 | if (per_cpu(rcu_dyntick_holdoff, cpu) == jiffies) { |
84ad00cb | 2217 | local_irq_restore(flags); |
433cdddc | 2218 | trace_rcu_prep_idle("In holdoff"); |
aea1b35e | 2219 | return; |
433cdddc | 2220 | } |
a47cd880 PM |
2221 | |
2222 | /* Check and update the rcu_dyntick_drain sequencing. */ | |
2223 | if (per_cpu(rcu_dyntick_drain, cpu) <= 0) { | |
2224 | /* First time through, initialize the counter. */ | |
f23f7fa1 PM |
2225 | per_cpu(rcu_dyntick_drain, cpu) = RCU_IDLE_FLUSHES; |
2226 | } else if (per_cpu(rcu_dyntick_drain, cpu) <= RCU_IDLE_OPT_FLUSHES && | |
2227 | !rcu_pending(cpu)) { | |
7cb92499 | 2228 | /* Can we go dyntick-idle despite still having callbacks? */ |
f23f7fa1 PM |
2229 | trace_rcu_prep_idle("Dyntick with callbacks"); |
2230 | per_cpu(rcu_dyntick_drain, cpu) = 0; | |
2231 | per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1; | |
486e2593 PM |
2232 | if (rcu_cpu_has_nonlazy_callbacks(cpu)) |
2233 | hrtimer_start(&per_cpu(rcu_idle_gp_timer, cpu), | |
2234 | rcu_idle_gp_wait, HRTIMER_MODE_REL); | |
778d250a PM |
2235 | else |
2236 | hrtimer_start(&per_cpu(rcu_idle_gp_timer, cpu), | |
2237 | rcu_idle_lazy_gp_wait, HRTIMER_MODE_REL); | |
f23f7fa1 PM |
2238 | return; /* Nothing more to do immediately. */ |
2239 | } else if (--per_cpu(rcu_dyntick_drain, cpu) <= 0) { | |
a47cd880 | 2240 | /* We have hit the limit, so time to give up. */ |
71da8132 | 2241 | per_cpu(rcu_dyntick_holdoff, cpu) = jiffies; |
84ad00cb | 2242 | local_irq_restore(flags); |
433cdddc | 2243 | trace_rcu_prep_idle("Begin holdoff"); |
aea1b35e PM |
2244 | invoke_rcu_core(); /* Force the CPU out of dyntick-idle. */ |
2245 | return; | |
a47cd880 PM |
2246 | } |
2247 | ||
aea1b35e PM |
2248 | /* |
2249 | * Do one step of pushing the remaining RCU callbacks through | |
2250 | * the RCU core state machine. | |
2251 | */ | |
2252 | #ifdef CONFIG_TREE_PREEMPT_RCU | |
2253 | if (per_cpu(rcu_preempt_data, cpu).nxtlist) { | |
3ad0decf | 2254 | local_irq_restore(flags); |
aea1b35e PM |
2255 | rcu_preempt_qs(cpu); |
2256 | force_quiescent_state(&rcu_preempt_state, 0); | |
3ad0decf | 2257 | local_irq_save(flags); |
aea1b35e PM |
2258 | } |
2259 | #endif /* #ifdef CONFIG_TREE_PREEMPT_RCU */ | |
a47cd880 | 2260 | if (per_cpu(rcu_sched_data, cpu).nxtlist) { |
3ad0decf | 2261 | local_irq_restore(flags); |
a47cd880 PM |
2262 | rcu_sched_qs(cpu); |
2263 | force_quiescent_state(&rcu_sched_state, 0); | |
3ad0decf | 2264 | local_irq_save(flags); |
a47cd880 PM |
2265 | } |
2266 | if (per_cpu(rcu_bh_data, cpu).nxtlist) { | |
3ad0decf | 2267 | local_irq_restore(flags); |
a47cd880 PM |
2268 | rcu_bh_qs(cpu); |
2269 | force_quiescent_state(&rcu_bh_state, 0); | |
3ad0decf | 2270 | local_irq_save(flags); |
8bd93a2c PM |
2271 | } |
2272 | ||
433cdddc PM |
2273 | /* |
2274 | * If RCU callbacks are still pending, RCU still needs this CPU. | |
2275 | * So try forcing the callbacks through the grace period. | |
2276 | */ | |
3ad0decf | 2277 | if (rcu_cpu_has_callbacks(cpu)) { |
84ad00cb | 2278 | local_irq_restore(flags); |
433cdddc | 2279 | trace_rcu_prep_idle("More callbacks"); |
a46e0899 | 2280 | invoke_rcu_core(); |
84ad00cb PM |
2281 | } else { |
2282 | local_irq_restore(flags); | |
433cdddc | 2283 | trace_rcu_prep_idle("Callbacks drained"); |
84ad00cb | 2284 | } |
8bd93a2c PM |
2285 | } |
2286 | ||
2287 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |