]>
Commit | Line | Data |
---|---|---|
1 | /* | |
2 | * Read-Copy Update mechanism for mutual exclusion (tree-based version) | |
3 | * Internal non-public definitions that provide either classic | |
4 | * or preemptable semantics. | |
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 | ||
27 | #include <linux/delay.h> | |
28 | ||
29 | #ifdef CONFIG_TREE_PREEMPT_RCU | |
30 | ||
31 | struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt_state); | |
32 | DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data); | |
33 | ||
34 | static int rcu_preempted_readers_exp(struct rcu_node *rnp); | |
35 | ||
36 | /* | |
37 | * Tell them what RCU they are running. | |
38 | */ | |
39 | static void __init rcu_bootup_announce(void) | |
40 | { | |
41 | printk(KERN_INFO | |
42 | "Experimental preemptable hierarchical RCU implementation.\n"); | |
43 | } | |
44 | ||
45 | /* | |
46 | * Return the number of RCU-preempt batches processed thus far | |
47 | * for debug and statistics. | |
48 | */ | |
49 | long rcu_batches_completed_preempt(void) | |
50 | { | |
51 | return rcu_preempt_state.completed; | |
52 | } | |
53 | EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt); | |
54 | ||
55 | /* | |
56 | * Return the number of RCU batches processed thus far for debug & stats. | |
57 | */ | |
58 | long rcu_batches_completed(void) | |
59 | { | |
60 | return rcu_batches_completed_preempt(); | |
61 | } | |
62 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
63 | ||
64 | /* | |
65 | * Force a quiescent state for preemptible RCU. | |
66 | */ | |
67 | void rcu_force_quiescent_state(void) | |
68 | { | |
69 | force_quiescent_state(&rcu_preempt_state, 0); | |
70 | } | |
71 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
72 | ||
73 | /* | |
74 | * Record a preemptable-RCU quiescent state for the specified CPU. Note | |
75 | * that this just means that the task currently running on the CPU is | |
76 | * not in a quiescent state. There might be any number of tasks blocked | |
77 | * while in an RCU read-side critical section. | |
78 | * | |
79 | * Unlike the other rcu_*_qs() functions, callers to this function | |
80 | * must disable irqs in order to protect the assignment to | |
81 | * ->rcu_read_unlock_special. | |
82 | */ | |
83 | static void rcu_preempt_qs(int cpu) | |
84 | { | |
85 | struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu); | |
86 | ||
87 | rdp->passed_quiesc_completed = rdp->gpnum - 1; | |
88 | barrier(); | |
89 | rdp->passed_quiesc = 1; | |
90 | current->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS; | |
91 | } | |
92 | ||
93 | /* | |
94 | * We have entered the scheduler, and the current task might soon be | |
95 | * context-switched away from. If this task is in an RCU read-side | |
96 | * critical section, we will no longer be able to rely on the CPU to | |
97 | * record that fact, so we enqueue the task on the appropriate entry | |
98 | * of the blocked_tasks[] array. The task will dequeue itself when | |
99 | * it exits the outermost enclosing RCU read-side critical section. | |
100 | * Therefore, the current grace period cannot be permitted to complete | |
101 | * until the blocked_tasks[] entry indexed by the low-order bit of | |
102 | * rnp->gpnum empties. | |
103 | * | |
104 | * Caller must disable preemption. | |
105 | */ | |
106 | static void rcu_preempt_note_context_switch(int cpu) | |
107 | { | |
108 | struct task_struct *t = current; | |
109 | unsigned long flags; | |
110 | int phase; | |
111 | struct rcu_data *rdp; | |
112 | struct rcu_node *rnp; | |
113 | ||
114 | if (t->rcu_read_lock_nesting && | |
115 | (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) { | |
116 | ||
117 | /* Possibly blocking in an RCU read-side critical section. */ | |
118 | rdp = rcu_preempt_state.rda[cpu]; | |
119 | rnp = rdp->mynode; | |
120 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
121 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED; | |
122 | t->rcu_blocked_node = rnp; | |
123 | ||
124 | /* | |
125 | * If this CPU has already checked in, then this task | |
126 | * will hold up the next grace period rather than the | |
127 | * current grace period. Queue the task accordingly. | |
128 | * If the task is queued for the current grace period | |
129 | * (i.e., this CPU has not yet passed through a quiescent | |
130 | * state for the current grace period), then as long | |
131 | * as that task remains queued, the current grace period | |
132 | * cannot end. | |
133 | * | |
134 | * But first, note that the current CPU must still be | |
135 | * on line! | |
136 | */ | |
137 | WARN_ON_ONCE((rdp->grpmask & rnp->qsmaskinit) == 0); | |
138 | WARN_ON_ONCE(!list_empty(&t->rcu_node_entry)); | |
139 | phase = (rnp->gpnum + !(rnp->qsmask & rdp->grpmask)) & 0x1; | |
140 | list_add(&t->rcu_node_entry, &rnp->blocked_tasks[phase]); | |
141 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
142 | } | |
143 | ||
144 | /* | |
145 | * Either we were not in an RCU read-side critical section to | |
146 | * begin with, or we have now recorded that critical section | |
147 | * globally. Either way, we can now note a quiescent state | |
148 | * for this CPU. Again, if we were in an RCU read-side critical | |
149 | * section, and if that critical section was blocking the current | |
150 | * grace period, then the fact that the task has been enqueued | |
151 | * means that we continue to block the current grace period. | |
152 | */ | |
153 | local_irq_save(flags); | |
154 | rcu_preempt_qs(cpu); | |
155 | local_irq_restore(flags); | |
156 | } | |
157 | ||
158 | /* | |
159 | * Tree-preemptable RCU implementation for rcu_read_lock(). | |
160 | * Just increment ->rcu_read_lock_nesting, shared state will be updated | |
161 | * if we block. | |
162 | */ | |
163 | void __rcu_read_lock(void) | |
164 | { | |
165 | ACCESS_ONCE(current->rcu_read_lock_nesting)++; | |
166 | barrier(); /* needed if we ever invoke rcu_read_lock in rcutree.c */ | |
167 | } | |
168 | EXPORT_SYMBOL_GPL(__rcu_read_lock); | |
169 | ||
170 | /* | |
171 | * Check for preempted RCU readers blocking the current grace period | |
172 | * for the specified rcu_node structure. If the caller needs a reliable | |
173 | * answer, it must hold the rcu_node's ->lock. | |
174 | */ | |
175 | static int rcu_preempted_readers(struct rcu_node *rnp) | |
176 | { | |
177 | int phase = rnp->gpnum & 0x1; | |
178 | ||
179 | return !list_empty(&rnp->blocked_tasks[phase]) || | |
180 | !list_empty(&rnp->blocked_tasks[phase + 2]); | |
181 | } | |
182 | ||
183 | /* | |
184 | * Record a quiescent state for all tasks that were previously queued | |
185 | * on the specified rcu_node structure and that were blocking the current | |
186 | * RCU grace period. The caller must hold the specified rnp->lock with | |
187 | * irqs disabled, and this lock is released upon return, but irqs remain | |
188 | * disabled. | |
189 | */ | |
190 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) | |
191 | __releases(rnp->lock) | |
192 | { | |
193 | unsigned long mask; | |
194 | struct rcu_node *rnp_p; | |
195 | ||
196 | if (rnp->qsmask != 0 || rcu_preempted_readers(rnp)) { | |
197 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
198 | return; /* Still need more quiescent states! */ | |
199 | } | |
200 | ||
201 | rnp_p = rnp->parent; | |
202 | if (rnp_p == NULL) { | |
203 | /* | |
204 | * Either there is only one rcu_node in the tree, | |
205 | * or tasks were kicked up to root rcu_node due to | |
206 | * CPUs going offline. | |
207 | */ | |
208 | rcu_report_qs_rsp(&rcu_preempt_state, flags); | |
209 | return; | |
210 | } | |
211 | ||
212 | /* Report up the rest of the hierarchy. */ | |
213 | mask = rnp->grpmask; | |
214 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
215 | raw_spin_lock(&rnp_p->lock); /* irqs already disabled. */ | |
216 | rcu_report_qs_rnp(mask, &rcu_preempt_state, rnp_p, flags); | |
217 | } | |
218 | ||
219 | /* | |
220 | * Handle special cases during rcu_read_unlock(), such as needing to | |
221 | * notify RCU core processing or task having blocked during the RCU | |
222 | * read-side critical section. | |
223 | */ | |
224 | static void rcu_read_unlock_special(struct task_struct *t) | |
225 | { | |
226 | int empty; | |
227 | int empty_exp; | |
228 | unsigned long flags; | |
229 | struct rcu_node *rnp; | |
230 | int special; | |
231 | ||
232 | /* NMI handlers cannot block and cannot safely manipulate state. */ | |
233 | if (in_nmi()) | |
234 | return; | |
235 | ||
236 | local_irq_save(flags); | |
237 | ||
238 | /* | |
239 | * If RCU core is waiting for this CPU to exit critical section, | |
240 | * let it know that we have done so. | |
241 | */ | |
242 | special = t->rcu_read_unlock_special; | |
243 | if (special & RCU_READ_UNLOCK_NEED_QS) { | |
244 | rcu_preempt_qs(smp_processor_id()); | |
245 | } | |
246 | ||
247 | /* Hardware IRQ handlers cannot block. */ | |
248 | if (in_irq()) { | |
249 | local_irq_restore(flags); | |
250 | return; | |
251 | } | |
252 | ||
253 | /* Clean up if blocked during RCU read-side critical section. */ | |
254 | if (special & RCU_READ_UNLOCK_BLOCKED) { | |
255 | t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED; | |
256 | ||
257 | /* | |
258 | * Remove this task from the list it blocked on. The | |
259 | * task can migrate while we acquire the lock, but at | |
260 | * most one time. So at most two passes through loop. | |
261 | */ | |
262 | for (;;) { | |
263 | rnp = t->rcu_blocked_node; | |
264 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
265 | if (rnp == t->rcu_blocked_node) | |
266 | break; | |
267 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
268 | } | |
269 | empty = !rcu_preempted_readers(rnp); | |
270 | empty_exp = !rcu_preempted_readers_exp(rnp); | |
271 | smp_mb(); /* ensure expedited fastpath sees end of RCU c-s. */ | |
272 | list_del_init(&t->rcu_node_entry); | |
273 | t->rcu_blocked_node = NULL; | |
274 | ||
275 | /* | |
276 | * If this was the last task on the current list, and if | |
277 | * we aren't waiting on any CPUs, report the quiescent state. | |
278 | * Note that rcu_report_unblock_qs_rnp() releases rnp->lock. | |
279 | */ | |
280 | if (empty) | |
281 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
282 | else | |
283 | rcu_report_unblock_qs_rnp(rnp, flags); | |
284 | ||
285 | /* | |
286 | * If this was the last task on the expedited lists, | |
287 | * then we need to report up the rcu_node hierarchy. | |
288 | */ | |
289 | if (!empty_exp && !rcu_preempted_readers_exp(rnp)) | |
290 | rcu_report_exp_rnp(&rcu_preempt_state, rnp); | |
291 | } else { | |
292 | local_irq_restore(flags); | |
293 | } | |
294 | } | |
295 | ||
296 | /* | |
297 | * Tree-preemptable RCU implementation for rcu_read_unlock(). | |
298 | * Decrement ->rcu_read_lock_nesting. If the result is zero (outermost | |
299 | * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then | |
300 | * invoke rcu_read_unlock_special() to clean up after a context switch | |
301 | * in an RCU read-side critical section and other special cases. | |
302 | */ | |
303 | void __rcu_read_unlock(void) | |
304 | { | |
305 | struct task_struct *t = current; | |
306 | ||
307 | barrier(); /* needed if we ever invoke rcu_read_unlock in rcutree.c */ | |
308 | if (--ACCESS_ONCE(t->rcu_read_lock_nesting) == 0 && | |
309 | unlikely(ACCESS_ONCE(t->rcu_read_unlock_special))) | |
310 | rcu_read_unlock_special(t); | |
311 | #ifdef CONFIG_PROVE_LOCKING | |
312 | WARN_ON_ONCE(ACCESS_ONCE(t->rcu_read_lock_nesting) < 0); | |
313 | #endif /* #ifdef CONFIG_PROVE_LOCKING */ | |
314 | } | |
315 | EXPORT_SYMBOL_GPL(__rcu_read_unlock); | |
316 | ||
317 | #ifdef CONFIG_RCU_CPU_STALL_DETECTOR | |
318 | ||
319 | #ifdef CONFIG_RCU_CPU_STALL_VERBOSE | |
320 | ||
321 | /* | |
322 | * Dump detailed information for all tasks blocking the current RCU | |
323 | * grace period on the specified rcu_node structure. | |
324 | */ | |
325 | static void rcu_print_detail_task_stall_rnp(struct rcu_node *rnp) | |
326 | { | |
327 | unsigned long flags; | |
328 | struct list_head *lp; | |
329 | int phase; | |
330 | struct task_struct *t; | |
331 | ||
332 | if (rcu_preempted_readers(rnp)) { | |
333 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
334 | phase = rnp->gpnum & 0x1; | |
335 | lp = &rnp->blocked_tasks[phase]; | |
336 | list_for_each_entry(t, lp, rcu_node_entry) | |
337 | sched_show_task(t); | |
338 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
339 | } | |
340 | } | |
341 | ||
342 | /* | |
343 | * Dump detailed information for all tasks blocking the current RCU | |
344 | * grace period. | |
345 | */ | |
346 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
347 | { | |
348 | struct rcu_node *rnp = rcu_get_root(rsp); | |
349 | ||
350 | rcu_print_detail_task_stall_rnp(rnp); | |
351 | rcu_for_each_leaf_node(rsp, rnp) | |
352 | rcu_print_detail_task_stall_rnp(rnp); | |
353 | } | |
354 | ||
355 | #else /* #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ | |
356 | ||
357 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
358 | { | |
359 | } | |
360 | ||
361 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_VERBOSE */ | |
362 | ||
363 | /* | |
364 | * Scan the current list of tasks blocked within RCU read-side critical | |
365 | * sections, printing out the tid of each. | |
366 | */ | |
367 | static void rcu_print_task_stall(struct rcu_node *rnp) | |
368 | { | |
369 | struct list_head *lp; | |
370 | int phase; | |
371 | struct task_struct *t; | |
372 | ||
373 | if (rcu_preempted_readers(rnp)) { | |
374 | phase = rnp->gpnum & 0x1; | |
375 | lp = &rnp->blocked_tasks[phase]; | |
376 | list_for_each_entry(t, lp, rcu_node_entry) | |
377 | printk(" P%d", t->pid); | |
378 | } | |
379 | } | |
380 | ||
381 | #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */ | |
382 | ||
383 | /* | |
384 | * Check that the list of blocked tasks for the newly completed grace | |
385 | * period is in fact empty. It is a serious bug to complete a grace | |
386 | * period that still has RCU readers blocked! This function must be | |
387 | * invoked -before- updating this rnp's ->gpnum, and the rnp's ->lock | |
388 | * must be held by the caller. | |
389 | */ | |
390 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
391 | { | |
392 | WARN_ON_ONCE(rcu_preempted_readers(rnp)); | |
393 | WARN_ON_ONCE(rnp->qsmask); | |
394 | } | |
395 | ||
396 | #ifdef CONFIG_HOTPLUG_CPU | |
397 | ||
398 | /* | |
399 | * Handle tasklist migration for case in which all CPUs covered by the | |
400 | * specified rcu_node have gone offline. Move them up to the root | |
401 | * rcu_node. The reason for not just moving them to the immediate | |
402 | * parent is to remove the need for rcu_read_unlock_special() to | |
403 | * make more than two attempts to acquire the target rcu_node's lock. | |
404 | * Returns true if there were tasks blocking the current RCU grace | |
405 | * period. | |
406 | * | |
407 | * Returns 1 if there was previously a task blocking the current grace | |
408 | * period on the specified rcu_node structure. | |
409 | * | |
410 | * The caller must hold rnp->lock with irqs disabled. | |
411 | */ | |
412 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, | |
413 | struct rcu_node *rnp, | |
414 | struct rcu_data *rdp) | |
415 | { | |
416 | int i; | |
417 | struct list_head *lp; | |
418 | struct list_head *lp_root; | |
419 | int retval = 0; | |
420 | struct rcu_node *rnp_root = rcu_get_root(rsp); | |
421 | struct task_struct *tp; | |
422 | ||
423 | if (rnp == rnp_root) { | |
424 | WARN_ONCE(1, "Last CPU thought to be offlined?"); | |
425 | return 0; /* Shouldn't happen: at least one CPU online. */ | |
426 | } | |
427 | WARN_ON_ONCE(rnp != rdp->mynode && | |
428 | (!list_empty(&rnp->blocked_tasks[0]) || | |
429 | !list_empty(&rnp->blocked_tasks[1]) || | |
430 | !list_empty(&rnp->blocked_tasks[2]) || | |
431 | !list_empty(&rnp->blocked_tasks[3]))); | |
432 | ||
433 | /* | |
434 | * Move tasks up to root rcu_node. Rely on the fact that the | |
435 | * root rcu_node can be at most one ahead of the rest of the | |
436 | * rcu_nodes in terms of gp_num value. This fact allows us to | |
437 | * move the blocked_tasks[] array directly, element by element. | |
438 | */ | |
439 | if (rcu_preempted_readers(rnp)) | |
440 | retval |= RCU_OFL_TASKS_NORM_GP; | |
441 | if (rcu_preempted_readers_exp(rnp)) | |
442 | retval |= RCU_OFL_TASKS_EXP_GP; | |
443 | for (i = 0; i < 4; i++) { | |
444 | lp = &rnp->blocked_tasks[i]; | |
445 | lp_root = &rnp_root->blocked_tasks[i]; | |
446 | while (!list_empty(lp)) { | |
447 | tp = list_entry(lp->next, typeof(*tp), rcu_node_entry); | |
448 | raw_spin_lock(&rnp_root->lock); /* irqs already disabled */ | |
449 | list_del(&tp->rcu_node_entry); | |
450 | tp->rcu_blocked_node = rnp_root; | |
451 | list_add(&tp->rcu_node_entry, lp_root); | |
452 | raw_spin_unlock(&rnp_root->lock); /* irqs remain disabled */ | |
453 | } | |
454 | } | |
455 | return retval; | |
456 | } | |
457 | ||
458 | /* | |
459 | * Do CPU-offline processing for preemptable RCU. | |
460 | */ | |
461 | static void rcu_preempt_offline_cpu(int cpu) | |
462 | { | |
463 | __rcu_offline_cpu(cpu, &rcu_preempt_state); | |
464 | } | |
465 | ||
466 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
467 | ||
468 | /* | |
469 | * Check for a quiescent state from the current CPU. When a task blocks, | |
470 | * the task is recorded in the corresponding CPU's rcu_node structure, | |
471 | * which is checked elsewhere. | |
472 | * | |
473 | * Caller must disable hard irqs. | |
474 | */ | |
475 | static void rcu_preempt_check_callbacks(int cpu) | |
476 | { | |
477 | struct task_struct *t = current; | |
478 | ||
479 | if (t->rcu_read_lock_nesting == 0) { | |
480 | rcu_preempt_qs(cpu); | |
481 | return; | |
482 | } | |
483 | if (per_cpu(rcu_preempt_data, cpu).qs_pending) | |
484 | t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS; | |
485 | } | |
486 | ||
487 | /* | |
488 | * Process callbacks for preemptable RCU. | |
489 | */ | |
490 | static void rcu_preempt_process_callbacks(void) | |
491 | { | |
492 | __rcu_process_callbacks(&rcu_preempt_state, | |
493 | &__get_cpu_var(rcu_preempt_data)); | |
494 | } | |
495 | ||
496 | /* | |
497 | * Queue a preemptable-RCU callback for invocation after a grace period. | |
498 | */ | |
499 | void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
500 | { | |
501 | __call_rcu(head, func, &rcu_preempt_state); | |
502 | } | |
503 | EXPORT_SYMBOL_GPL(call_rcu); | |
504 | ||
505 | /** | |
506 | * synchronize_rcu - wait until a grace period has elapsed. | |
507 | * | |
508 | * Control will return to the caller some time after a full grace | |
509 | * period has elapsed, in other words after all currently executing RCU | |
510 | * read-side critical sections have completed. RCU read-side critical | |
511 | * sections are delimited by rcu_read_lock() and rcu_read_unlock(), | |
512 | * and may be nested. | |
513 | */ | |
514 | void synchronize_rcu(void) | |
515 | { | |
516 | struct rcu_synchronize rcu; | |
517 | ||
518 | if (!rcu_scheduler_active) | |
519 | return; | |
520 | ||
521 | init_completion(&rcu.completion); | |
522 | /* Will wake me after RCU finished. */ | |
523 | call_rcu(&rcu.head, wakeme_after_rcu); | |
524 | /* Wait for it. */ | |
525 | wait_for_completion(&rcu.completion); | |
526 | } | |
527 | EXPORT_SYMBOL_GPL(synchronize_rcu); | |
528 | ||
529 | static DECLARE_WAIT_QUEUE_HEAD(sync_rcu_preempt_exp_wq); | |
530 | static long sync_rcu_preempt_exp_count; | |
531 | static DEFINE_MUTEX(sync_rcu_preempt_exp_mutex); | |
532 | ||
533 | /* | |
534 | * Return non-zero if there are any tasks in RCU read-side critical | |
535 | * sections blocking the current preemptible-RCU expedited grace period. | |
536 | * If there is no preemptible-RCU expedited grace period currently in | |
537 | * progress, returns zero unconditionally. | |
538 | */ | |
539 | static int rcu_preempted_readers_exp(struct rcu_node *rnp) | |
540 | { | |
541 | return !list_empty(&rnp->blocked_tasks[2]) || | |
542 | !list_empty(&rnp->blocked_tasks[3]); | |
543 | } | |
544 | ||
545 | /* | |
546 | * return non-zero if there is no RCU expedited grace period in progress | |
547 | * for the specified rcu_node structure, in other words, if all CPUs and | |
548 | * tasks covered by the specified rcu_node structure have done their bit | |
549 | * for the current expedited grace period. Works only for preemptible | |
550 | * RCU -- other RCU implementation use other means. | |
551 | * | |
552 | * Caller must hold sync_rcu_preempt_exp_mutex. | |
553 | */ | |
554 | static int sync_rcu_preempt_exp_done(struct rcu_node *rnp) | |
555 | { | |
556 | return !rcu_preempted_readers_exp(rnp) && | |
557 | ACCESS_ONCE(rnp->expmask) == 0; | |
558 | } | |
559 | ||
560 | /* | |
561 | * Report the exit from RCU read-side critical section for the last task | |
562 | * that queued itself during or before the current expedited preemptible-RCU | |
563 | * grace period. This event is reported either to the rcu_node structure on | |
564 | * which the task was queued or to one of that rcu_node structure's ancestors, | |
565 | * recursively up the tree. (Calm down, calm down, we do the recursion | |
566 | * iteratively!) | |
567 | * | |
568 | * Caller must hold sync_rcu_preempt_exp_mutex. | |
569 | */ | |
570 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp) | |
571 | { | |
572 | unsigned long flags; | |
573 | unsigned long mask; | |
574 | ||
575 | raw_spin_lock_irqsave(&rnp->lock, flags); | |
576 | for (;;) { | |
577 | if (!sync_rcu_preempt_exp_done(rnp)) | |
578 | break; | |
579 | if (rnp->parent == NULL) { | |
580 | wake_up(&sync_rcu_preempt_exp_wq); | |
581 | break; | |
582 | } | |
583 | mask = rnp->grpmask; | |
584 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ | |
585 | rnp = rnp->parent; | |
586 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ | |
587 | rnp->expmask &= ~mask; | |
588 | } | |
589 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
590 | } | |
591 | ||
592 | /* | |
593 | * Snapshot the tasks blocking the newly started preemptible-RCU expedited | |
594 | * grace period for the specified rcu_node structure. If there are no such | |
595 | * tasks, report it up the rcu_node hierarchy. | |
596 | * | |
597 | * Caller must hold sync_rcu_preempt_exp_mutex and rsp->onofflock. | |
598 | */ | |
599 | static void | |
600 | sync_rcu_preempt_exp_init(struct rcu_state *rsp, struct rcu_node *rnp) | |
601 | { | |
602 | int must_wait; | |
603 | ||
604 | raw_spin_lock(&rnp->lock); /* irqs already disabled */ | |
605 | list_splice_init(&rnp->blocked_tasks[0], &rnp->blocked_tasks[2]); | |
606 | list_splice_init(&rnp->blocked_tasks[1], &rnp->blocked_tasks[3]); | |
607 | must_wait = rcu_preempted_readers_exp(rnp); | |
608 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled */ | |
609 | if (!must_wait) | |
610 | rcu_report_exp_rnp(rsp, rnp); | |
611 | } | |
612 | ||
613 | /* | |
614 | * Wait for an rcu-preempt grace period, but expedite it. The basic idea | |
615 | * is to invoke synchronize_sched_expedited() to push all the tasks to | |
616 | * the ->blocked_tasks[] lists, move all entries from the first set of | |
617 | * ->blocked_tasks[] lists to the second set, and finally wait for this | |
618 | * second set to drain. | |
619 | */ | |
620 | void synchronize_rcu_expedited(void) | |
621 | { | |
622 | unsigned long flags; | |
623 | struct rcu_node *rnp; | |
624 | struct rcu_state *rsp = &rcu_preempt_state; | |
625 | long snap; | |
626 | int trycount = 0; | |
627 | ||
628 | smp_mb(); /* Caller's modifications seen first by other CPUs. */ | |
629 | snap = ACCESS_ONCE(sync_rcu_preempt_exp_count) + 1; | |
630 | smp_mb(); /* Above access cannot bleed into critical section. */ | |
631 | ||
632 | /* | |
633 | * Acquire lock, falling back to synchronize_rcu() if too many | |
634 | * lock-acquisition failures. Of course, if someone does the | |
635 | * expedited grace period for us, just leave. | |
636 | */ | |
637 | while (!mutex_trylock(&sync_rcu_preempt_exp_mutex)) { | |
638 | if (trycount++ < 10) | |
639 | udelay(trycount * num_online_cpus()); | |
640 | else { | |
641 | synchronize_rcu(); | |
642 | return; | |
643 | } | |
644 | if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0) | |
645 | goto mb_ret; /* Others did our work for us. */ | |
646 | } | |
647 | if ((ACCESS_ONCE(sync_rcu_preempt_exp_count) - snap) > 0) | |
648 | goto unlock_mb_ret; /* Others did our work for us. */ | |
649 | ||
650 | /* force all RCU readers onto blocked_tasks[]. */ | |
651 | synchronize_sched_expedited(); | |
652 | ||
653 | raw_spin_lock_irqsave(&rsp->onofflock, flags); | |
654 | ||
655 | /* Initialize ->expmask for all non-leaf rcu_node structures. */ | |
656 | rcu_for_each_nonleaf_node_breadth_first(rsp, rnp) { | |
657 | raw_spin_lock(&rnp->lock); /* irqs already disabled. */ | |
658 | rnp->expmask = rnp->qsmaskinit; | |
659 | raw_spin_unlock(&rnp->lock); /* irqs remain disabled. */ | |
660 | } | |
661 | ||
662 | /* Snapshot current state of ->blocked_tasks[] lists. */ | |
663 | rcu_for_each_leaf_node(rsp, rnp) | |
664 | sync_rcu_preempt_exp_init(rsp, rnp); | |
665 | if (NUM_RCU_NODES > 1) | |
666 | sync_rcu_preempt_exp_init(rsp, rcu_get_root(rsp)); | |
667 | ||
668 | raw_spin_unlock_irqrestore(&rsp->onofflock, flags); | |
669 | ||
670 | /* Wait for snapshotted ->blocked_tasks[] lists to drain. */ | |
671 | rnp = rcu_get_root(rsp); | |
672 | wait_event(sync_rcu_preempt_exp_wq, | |
673 | sync_rcu_preempt_exp_done(rnp)); | |
674 | ||
675 | /* Clean up and exit. */ | |
676 | smp_mb(); /* ensure expedited GP seen before counter increment. */ | |
677 | ACCESS_ONCE(sync_rcu_preempt_exp_count)++; | |
678 | unlock_mb_ret: | |
679 | mutex_unlock(&sync_rcu_preempt_exp_mutex); | |
680 | mb_ret: | |
681 | smp_mb(); /* ensure subsequent action seen after grace period. */ | |
682 | } | |
683 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
684 | ||
685 | /* | |
686 | * Check to see if there is any immediate preemptable-RCU-related work | |
687 | * to be done. | |
688 | */ | |
689 | static int rcu_preempt_pending(int cpu) | |
690 | { | |
691 | return __rcu_pending(&rcu_preempt_state, | |
692 | &per_cpu(rcu_preempt_data, cpu)); | |
693 | } | |
694 | ||
695 | /* | |
696 | * Does preemptable RCU need the CPU to stay out of dynticks mode? | |
697 | */ | |
698 | static int rcu_preempt_needs_cpu(int cpu) | |
699 | { | |
700 | return !!per_cpu(rcu_preempt_data, cpu).nxtlist; | |
701 | } | |
702 | ||
703 | /** | |
704 | * rcu_barrier - Wait until all in-flight call_rcu() callbacks complete. | |
705 | */ | |
706 | void rcu_barrier(void) | |
707 | { | |
708 | _rcu_barrier(&rcu_preempt_state, call_rcu); | |
709 | } | |
710 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
711 | ||
712 | /* | |
713 | * Initialize preemptable RCU's per-CPU data. | |
714 | */ | |
715 | static void __cpuinit rcu_preempt_init_percpu_data(int cpu) | |
716 | { | |
717 | rcu_init_percpu_data(cpu, &rcu_preempt_state, 1); | |
718 | } | |
719 | ||
720 | /* | |
721 | * Move preemptable RCU's callbacks to ->orphan_cbs_list. | |
722 | */ | |
723 | static void rcu_preempt_send_cbs_to_orphanage(void) | |
724 | { | |
725 | rcu_send_cbs_to_orphanage(&rcu_preempt_state); | |
726 | } | |
727 | ||
728 | /* | |
729 | * Initialize preemptable RCU's state structures. | |
730 | */ | |
731 | static void __init __rcu_init_preempt(void) | |
732 | { | |
733 | RCU_INIT_FLAVOR(&rcu_preempt_state, rcu_preempt_data); | |
734 | } | |
735 | ||
736 | /* | |
737 | * Check for a task exiting while in a preemptable-RCU read-side | |
738 | * critical section, clean up if so. No need to issue warnings, | |
739 | * as debug_check_no_locks_held() already does this if lockdep | |
740 | * is enabled. | |
741 | */ | |
742 | void exit_rcu(void) | |
743 | { | |
744 | struct task_struct *t = current; | |
745 | ||
746 | if (t->rcu_read_lock_nesting == 0) | |
747 | return; | |
748 | t->rcu_read_lock_nesting = 1; | |
749 | rcu_read_unlock(); | |
750 | } | |
751 | ||
752 | #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */ | |
753 | ||
754 | /* | |
755 | * Tell them what RCU they are running. | |
756 | */ | |
757 | static void __init rcu_bootup_announce(void) | |
758 | { | |
759 | printk(KERN_INFO "Hierarchical RCU implementation.\n"); | |
760 | } | |
761 | ||
762 | /* | |
763 | * Return the number of RCU batches processed thus far for debug & stats. | |
764 | */ | |
765 | long rcu_batches_completed(void) | |
766 | { | |
767 | return rcu_batches_completed_sched(); | |
768 | } | |
769 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
770 | ||
771 | /* | |
772 | * Force a quiescent state for RCU, which, because there is no preemptible | |
773 | * RCU, becomes the same as rcu-sched. | |
774 | */ | |
775 | void rcu_force_quiescent_state(void) | |
776 | { | |
777 | rcu_sched_force_quiescent_state(); | |
778 | } | |
779 | EXPORT_SYMBOL_GPL(rcu_force_quiescent_state); | |
780 | ||
781 | /* | |
782 | * Because preemptable RCU does not exist, we never have to check for | |
783 | * CPUs being in quiescent states. | |
784 | */ | |
785 | static void rcu_preempt_note_context_switch(int cpu) | |
786 | { | |
787 | } | |
788 | ||
789 | /* | |
790 | * Because preemptable RCU does not exist, there are never any preempted | |
791 | * RCU readers. | |
792 | */ | |
793 | static int rcu_preempted_readers(struct rcu_node *rnp) | |
794 | { | |
795 | return 0; | |
796 | } | |
797 | ||
798 | #ifdef CONFIG_HOTPLUG_CPU | |
799 | ||
800 | /* Because preemptible RCU does not exist, no quieting of tasks. */ | |
801 | static void rcu_report_unblock_qs_rnp(struct rcu_node *rnp, unsigned long flags) | |
802 | { | |
803 | raw_spin_unlock_irqrestore(&rnp->lock, flags); | |
804 | } | |
805 | ||
806 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
807 | ||
808 | #ifdef CONFIG_RCU_CPU_STALL_DETECTOR | |
809 | ||
810 | /* | |
811 | * Because preemptable RCU does not exist, we never have to check for | |
812 | * tasks blocked within RCU read-side critical sections. | |
813 | */ | |
814 | static void rcu_print_detail_task_stall(struct rcu_state *rsp) | |
815 | { | |
816 | } | |
817 | ||
818 | /* | |
819 | * Because preemptable RCU does not exist, we never have to check for | |
820 | * tasks blocked within RCU read-side critical sections. | |
821 | */ | |
822 | static void rcu_print_task_stall(struct rcu_node *rnp) | |
823 | { | |
824 | } | |
825 | ||
826 | #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */ | |
827 | ||
828 | /* | |
829 | * Because there is no preemptable RCU, there can be no readers blocked, | |
830 | * so there is no need to check for blocked tasks. So check only for | |
831 | * bogus qsmask values. | |
832 | */ | |
833 | static void rcu_preempt_check_blocked_tasks(struct rcu_node *rnp) | |
834 | { | |
835 | WARN_ON_ONCE(rnp->qsmask); | |
836 | } | |
837 | ||
838 | #ifdef CONFIG_HOTPLUG_CPU | |
839 | ||
840 | /* | |
841 | * Because preemptable RCU does not exist, it never needs to migrate | |
842 | * tasks that were blocked within RCU read-side critical sections, and | |
843 | * such non-existent tasks cannot possibly have been blocking the current | |
844 | * grace period. | |
845 | */ | |
846 | static int rcu_preempt_offline_tasks(struct rcu_state *rsp, | |
847 | struct rcu_node *rnp, | |
848 | struct rcu_data *rdp) | |
849 | { | |
850 | return 0; | |
851 | } | |
852 | ||
853 | /* | |
854 | * Because preemptable RCU does not exist, it never needs CPU-offline | |
855 | * processing. | |
856 | */ | |
857 | static void rcu_preempt_offline_cpu(int cpu) | |
858 | { | |
859 | } | |
860 | ||
861 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
862 | ||
863 | /* | |
864 | * Because preemptable RCU does not exist, it never has any callbacks | |
865 | * to check. | |
866 | */ | |
867 | static void rcu_preempt_check_callbacks(int cpu) | |
868 | { | |
869 | } | |
870 | ||
871 | /* | |
872 | * Because preemptable RCU does not exist, it never has any callbacks | |
873 | * to process. | |
874 | */ | |
875 | static void rcu_preempt_process_callbacks(void) | |
876 | { | |
877 | } | |
878 | ||
879 | /* | |
880 | * In classic RCU, call_rcu() is just call_rcu_sched(). | |
881 | */ | |
882 | void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu)) | |
883 | { | |
884 | call_rcu_sched(head, func); | |
885 | } | |
886 | EXPORT_SYMBOL_GPL(call_rcu); | |
887 | ||
888 | /* | |
889 | * Wait for an rcu-preempt grace period, but make it happen quickly. | |
890 | * But because preemptable RCU does not exist, map to rcu-sched. | |
891 | */ | |
892 | void synchronize_rcu_expedited(void) | |
893 | { | |
894 | synchronize_sched_expedited(); | |
895 | } | |
896 | EXPORT_SYMBOL_GPL(synchronize_rcu_expedited); | |
897 | ||
898 | #ifdef CONFIG_HOTPLUG_CPU | |
899 | ||
900 | /* | |
901 | * Because preemptable RCU does not exist, there is never any need to | |
902 | * report on tasks preempted in RCU read-side critical sections during | |
903 | * expedited RCU grace periods. | |
904 | */ | |
905 | static void rcu_report_exp_rnp(struct rcu_state *rsp, struct rcu_node *rnp) | |
906 | { | |
907 | return; | |
908 | } | |
909 | ||
910 | #endif /* #ifdef CONFIG_HOTPLUG_CPU */ | |
911 | ||
912 | /* | |
913 | * Because preemptable RCU does not exist, it never has any work to do. | |
914 | */ | |
915 | static int rcu_preempt_pending(int cpu) | |
916 | { | |
917 | return 0; | |
918 | } | |
919 | ||
920 | /* | |
921 | * Because preemptable RCU does not exist, it never needs any CPU. | |
922 | */ | |
923 | static int rcu_preempt_needs_cpu(int cpu) | |
924 | { | |
925 | return 0; | |
926 | } | |
927 | ||
928 | /* | |
929 | * Because preemptable RCU does not exist, rcu_barrier() is just | |
930 | * another name for rcu_barrier_sched(). | |
931 | */ | |
932 | void rcu_barrier(void) | |
933 | { | |
934 | rcu_barrier_sched(); | |
935 | } | |
936 | EXPORT_SYMBOL_GPL(rcu_barrier); | |
937 | ||
938 | /* | |
939 | * Because preemptable RCU does not exist, there is no per-CPU | |
940 | * data to initialize. | |
941 | */ | |
942 | static void __cpuinit rcu_preempt_init_percpu_data(int cpu) | |
943 | { | |
944 | } | |
945 | ||
946 | /* | |
947 | * Because there is no preemptable RCU, there are no callbacks to move. | |
948 | */ | |
949 | static void rcu_preempt_send_cbs_to_orphanage(void) | |
950 | { | |
951 | } | |
952 | ||
953 | /* | |
954 | * Because preemptable RCU does not exist, it need not be initialized. | |
955 | */ | |
956 | static void __init __rcu_init_preempt(void) | |
957 | { | |
958 | } | |
959 | ||
960 | #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */ | |
961 | ||
962 | #if !defined(CONFIG_RCU_FAST_NO_HZ) | |
963 | ||
964 | /* | |
965 | * Check to see if any future RCU-related work will need to be done | |
966 | * by the current CPU, even if none need be done immediately, returning | |
967 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
968 | * an exported member of the RCU API. | |
969 | * | |
970 | * Because we have preemptible RCU, just check whether this CPU needs | |
971 | * any flavor of RCU. Do not chew up lots of CPU cycles with preemption | |
972 | * disabled in a most-likely vain attempt to cause RCU not to need this CPU. | |
973 | */ | |
974 | int rcu_needs_cpu(int cpu) | |
975 | { | |
976 | return rcu_needs_cpu_quick_check(cpu); | |
977 | } | |
978 | ||
979 | /* | |
980 | * Check to see if we need to continue a callback-flush operations to | |
981 | * allow the last CPU to enter dyntick-idle mode. But fast dyntick-idle | |
982 | * entry is not configured, so we never do need to. | |
983 | */ | |
984 | static void rcu_needs_cpu_flush(void) | |
985 | { | |
986 | } | |
987 | ||
988 | #else /* #if !defined(CONFIG_RCU_FAST_NO_HZ) */ | |
989 | ||
990 | #define RCU_NEEDS_CPU_FLUSHES 5 | |
991 | static DEFINE_PER_CPU(int, rcu_dyntick_drain); | |
992 | static DEFINE_PER_CPU(unsigned long, rcu_dyntick_holdoff); | |
993 | ||
994 | /* | |
995 | * Check to see if any future RCU-related work will need to be done | |
996 | * by the current CPU, even if none need be done immediately, returning | |
997 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
998 | * an exported member of the RCU API. | |
999 | * | |
1000 | * Because we are not supporting preemptible RCU, attempt to accelerate | |
1001 | * any current grace periods so that RCU no longer needs this CPU, but | |
1002 | * only if all other CPUs are already in dynticks-idle mode. This will | |
1003 | * allow the CPU cores to be powered down immediately, as opposed to after | |
1004 | * waiting many milliseconds for grace periods to elapse. | |
1005 | * | |
1006 | * Because it is not legal to invoke rcu_process_callbacks() with irqs | |
1007 | * disabled, we do one pass of force_quiescent_state(), then do a | |
1008 | * raise_softirq() to cause rcu_process_callbacks() to be invoked later. | |
1009 | * The per-cpu rcu_dyntick_drain variable controls the sequencing. | |
1010 | */ | |
1011 | int rcu_needs_cpu(int cpu) | |
1012 | { | |
1013 | int c = 0; | |
1014 | int thatcpu; | |
1015 | ||
1016 | /* Check for being in the holdoff period. */ | |
1017 | if (per_cpu(rcu_dyntick_holdoff, cpu) == jiffies) | |
1018 | return rcu_needs_cpu_quick_check(cpu); | |
1019 | ||
1020 | /* Don't bother unless we are the last non-dyntick-idle CPU. */ | |
1021 | for_each_cpu_not(thatcpu, nohz_cpu_mask) | |
1022 | if (cpu_online(thatcpu) && thatcpu != cpu) { | |
1023 | per_cpu(rcu_dyntick_drain, cpu) = 0; | |
1024 | per_cpu(rcu_dyntick_holdoff, cpu) = jiffies - 1; | |
1025 | return rcu_needs_cpu_quick_check(cpu); | |
1026 | } | |
1027 | ||
1028 | /* Check and update the rcu_dyntick_drain sequencing. */ | |
1029 | if (per_cpu(rcu_dyntick_drain, cpu) <= 0) { | |
1030 | /* First time through, initialize the counter. */ | |
1031 | per_cpu(rcu_dyntick_drain, cpu) = RCU_NEEDS_CPU_FLUSHES; | |
1032 | } else if (--per_cpu(rcu_dyntick_drain, cpu) <= 0) { | |
1033 | /* We have hit the limit, so time to give up. */ | |
1034 | per_cpu(rcu_dyntick_holdoff, cpu) = jiffies; | |
1035 | return rcu_needs_cpu_quick_check(cpu); | |
1036 | } | |
1037 | ||
1038 | /* Do one step pushing remaining RCU callbacks through. */ | |
1039 | if (per_cpu(rcu_sched_data, cpu).nxtlist) { | |
1040 | rcu_sched_qs(cpu); | |
1041 | force_quiescent_state(&rcu_sched_state, 0); | |
1042 | c = c || per_cpu(rcu_sched_data, cpu).nxtlist; | |
1043 | } | |
1044 | if (per_cpu(rcu_bh_data, cpu).nxtlist) { | |
1045 | rcu_bh_qs(cpu); | |
1046 | force_quiescent_state(&rcu_bh_state, 0); | |
1047 | c = c || per_cpu(rcu_bh_data, cpu).nxtlist; | |
1048 | } | |
1049 | ||
1050 | /* If RCU callbacks are still pending, RCU still needs this CPU. */ | |
1051 | if (c) | |
1052 | raise_softirq(RCU_SOFTIRQ); | |
1053 | return c; | |
1054 | } | |
1055 | ||
1056 | /* | |
1057 | * Check to see if we need to continue a callback-flush operations to | |
1058 | * allow the last CPU to enter dyntick-idle mode. | |
1059 | */ | |
1060 | static void rcu_needs_cpu_flush(void) | |
1061 | { | |
1062 | int cpu = smp_processor_id(); | |
1063 | unsigned long flags; | |
1064 | ||
1065 | if (per_cpu(rcu_dyntick_drain, cpu) <= 0) | |
1066 | return; | |
1067 | local_irq_save(flags); | |
1068 | (void)rcu_needs_cpu(cpu); | |
1069 | local_irq_restore(flags); | |
1070 | } | |
1071 | ||
1072 | #endif /* #else #if !defined(CONFIG_RCU_FAST_NO_HZ) */ |