rcu: Create rcutree plugins to handle hotplug CPU for multi-level trees
[safe/jmp/linux-2.6] / kernel / rcutree_plugin.h
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
28 #ifdef CONFIG_TREE_PREEMPT_RCU
29
30 struct rcu_state rcu_preempt_state = RCU_STATE_INITIALIZER(rcu_preempt_state);
31 DEFINE_PER_CPU(struct rcu_data, rcu_preempt_data);
32
33 /*
34  * Tell them what RCU they are running.
35  */
36 static inline void rcu_bootup_announce(void)
37 {
38         printk(KERN_INFO
39                "Experimental preemptable hierarchical RCU implementation.\n");
40 }
41
42 /*
43  * Return the number of RCU-preempt batches processed thus far
44  * for debug and statistics.
45  */
46 long rcu_batches_completed_preempt(void)
47 {
48         return rcu_preempt_state.completed;
49 }
50 EXPORT_SYMBOL_GPL(rcu_batches_completed_preempt);
51
52 /*
53  * Return the number of RCU batches processed thus far for debug & stats.
54  */
55 long rcu_batches_completed(void)
56 {
57         return rcu_batches_completed_preempt();
58 }
59 EXPORT_SYMBOL_GPL(rcu_batches_completed);
60
61 /*
62  * Record a preemptable-RCU quiescent state for the specified CPU.  Note
63  * that this just means that the task currently running on the CPU is
64  * not in a quiescent state.  There might be any number of tasks blocked
65  * while in an RCU read-side critical section.
66  */
67 static void rcu_preempt_qs_record(int cpu)
68 {
69         struct rcu_data *rdp = &per_cpu(rcu_preempt_data, cpu);
70         rdp->passed_quiesc = 1;
71         rdp->passed_quiesc_completed = rdp->completed;
72 }
73
74 /*
75  * We have entered the scheduler or are between softirqs in ksoftirqd.
76  * If we are in an RCU read-side critical section, we need to reflect
77  * that in the state of the rcu_node structure corresponding to this CPU.
78  * Caller must disable hardirqs.
79  */
80 static void rcu_preempt_qs(int cpu)
81 {
82         struct task_struct *t = current;
83         int phase;
84         struct rcu_data *rdp;
85         struct rcu_node *rnp;
86
87         if (t->rcu_read_lock_nesting &&
88             (t->rcu_read_unlock_special & RCU_READ_UNLOCK_BLOCKED) == 0) {
89
90                 /* Possibly blocking in an RCU read-side critical section. */
91                 rdp = rcu_preempt_state.rda[cpu];
92                 rnp = rdp->mynode;
93                 spin_lock(&rnp->lock);
94                 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_BLOCKED;
95                 t->rcu_blocked_node = (void *)rnp;
96
97                 /*
98                  * If this CPU has already checked in, then this task
99                  * will hold up the next grace period rather than the
100                  * current grace period.  Queue the task accordingly.
101                  * If the task is queued for the current grace period
102                  * (i.e., this CPU has not yet passed through a quiescent
103                  * state for the current grace period), then as long
104                  * as that task remains queued, the current grace period
105                  * cannot end.
106                  */
107                 phase = !(rnp->qsmask & rdp->grpmask) ^ (rnp->gpnum & 0x1);
108                 list_add(&t->rcu_node_entry, &rnp->blocked_tasks[phase]);
109                 smp_mb();  /* Ensure later ctxt swtch seen after above. */
110                 spin_unlock(&rnp->lock);
111         }
112
113         /*
114          * Either we were not in an RCU read-side critical section to
115          * begin with, or we have now recorded that critical section
116          * globally.  Either way, we can now note a quiescent state
117          * for this CPU.  Again, if we were in an RCU read-side critical
118          * section, and if that critical section was blocking the current
119          * grace period, then the fact that the task has been enqueued
120          * means that we continue to block the current grace period.
121          */
122         rcu_preempt_qs_record(cpu);
123         t->rcu_read_unlock_special &= ~(RCU_READ_UNLOCK_NEED_QS |
124                                         RCU_READ_UNLOCK_GOT_QS);
125 }
126
127 /*
128  * Tree-preemptable RCU implementation for rcu_read_lock().
129  * Just increment ->rcu_read_lock_nesting, shared state will be updated
130  * if we block.
131  */
132 void __rcu_read_lock(void)
133 {
134         ACCESS_ONCE(current->rcu_read_lock_nesting)++;
135         barrier();  /* needed if we ever invoke rcu_read_lock in rcutree.c */
136 }
137 EXPORT_SYMBOL_GPL(__rcu_read_lock);
138
139 static void rcu_read_unlock_special(struct task_struct *t)
140 {
141         int empty;
142         unsigned long flags;
143         unsigned long mask;
144         struct rcu_node *rnp;
145         int special;
146
147         /* NMI handlers cannot block and cannot safely manipulate state. */
148         if (in_nmi())
149                 return;
150
151         local_irq_save(flags);
152
153         /*
154          * If RCU core is waiting for this CPU to exit critical section,
155          * let it know that we have done so.
156          */
157         special = t->rcu_read_unlock_special;
158         if (special & RCU_READ_UNLOCK_NEED_QS) {
159                 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_NEED_QS;
160                 t->rcu_read_unlock_special |= RCU_READ_UNLOCK_GOT_QS;
161         }
162
163         /* Hardware IRQ handlers cannot block. */
164         if (in_irq()) {
165                 local_irq_restore(flags);
166                 return;
167         }
168
169         /* Clean up if blocked during RCU read-side critical section. */
170         if (special & RCU_READ_UNLOCK_BLOCKED) {
171                 t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_BLOCKED;
172
173                 /*
174                  * Remove this task from the list it blocked on.  The
175                  * task can migrate while we acquire the lock, but at
176                  * most one time.  So at most two passes through loop.
177                  */
178                 for (;;) {
179                         rnp = (struct rcu_node *)t->rcu_blocked_node;
180                         spin_lock(&rnp->lock);
181                         if (rnp == (struct rcu_node *)t->rcu_blocked_node)
182                                 break;
183                         spin_unlock(&rnp->lock);
184                 }
185                 empty = list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1]);
186                 list_del_init(&t->rcu_node_entry);
187                 t->rcu_blocked_node = NULL;
188
189                 /*
190                  * If this was the last task on the current list, and if
191                  * we aren't waiting on any CPUs, report the quiescent state.
192                  * Note that both cpu_quiet_msk_finish() and cpu_quiet_msk()
193                  * drop rnp->lock and restore irq.
194                  */
195                 if (!empty && rnp->qsmask == 0 &&
196                     list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1])) {
197                         t->rcu_read_unlock_special &=
198                                 ~(RCU_READ_UNLOCK_NEED_QS |
199                                   RCU_READ_UNLOCK_GOT_QS);
200                         if (rnp->parent == NULL) {
201                                 /* Only one rcu_node in the tree. */
202                                 cpu_quiet_msk_finish(&rcu_preempt_state, flags);
203                                 return;
204                         }
205                         /* Report up the rest of the hierarchy. */
206                         mask = rnp->grpmask;
207                         spin_unlock_irqrestore(&rnp->lock, flags);
208                         rnp = rnp->parent;
209                         spin_lock_irqsave(&rnp->lock, flags);
210                         cpu_quiet_msk(mask, &rcu_preempt_state, rnp, flags);
211                         return;
212                 }
213                 spin_unlock(&rnp->lock);
214         }
215         local_irq_restore(flags);
216 }
217
218 /*
219  * Tree-preemptable RCU implementation for rcu_read_unlock().
220  * Decrement ->rcu_read_lock_nesting.  If the result is zero (outermost
221  * rcu_read_unlock()) and ->rcu_read_unlock_special is non-zero, then
222  * invoke rcu_read_unlock_special() to clean up after a context switch
223  * in an RCU read-side critical section and other special cases.
224  */
225 void __rcu_read_unlock(void)
226 {
227         struct task_struct *t = current;
228
229         barrier();  /* needed if we ever invoke rcu_read_unlock in rcutree.c */
230         if (--ACCESS_ONCE(t->rcu_read_lock_nesting) == 0 &&
231             unlikely(ACCESS_ONCE(t->rcu_read_unlock_special)))
232                 rcu_read_unlock_special(t);
233 }
234 EXPORT_SYMBOL_GPL(__rcu_read_unlock);
235
236 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
237
238 /*
239  * Scan the current list of tasks blocked within RCU read-side critical
240  * sections, printing out the tid of each.
241  */
242 static void rcu_print_task_stall(struct rcu_node *rnp)
243 {
244         unsigned long flags;
245         struct list_head *lp;
246         int phase = rnp->gpnum & 0x1;
247         struct task_struct *t;
248
249         if (!list_empty(&rnp->blocked_tasks[phase])) {
250                 spin_lock_irqsave(&rnp->lock, flags);
251                 phase = rnp->gpnum & 0x1; /* re-read under lock. */
252                 lp = &rnp->blocked_tasks[phase];
253                 list_for_each_entry(t, lp, rcu_node_entry)
254                         printk(" P%d", t->pid);
255                 spin_unlock_irqrestore(&rnp->lock, flags);
256         }
257 }
258
259 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
260
261 /*
262  * Check for preempted RCU readers for the specified rcu_node structure.
263  * If the caller needs a reliable answer, it must hold the rcu_node's
264  * >lock.
265  */
266 static int rcu_preempted_readers(struct rcu_node *rnp)
267 {
268         return !list_empty(&rnp->blocked_tasks[rnp->gpnum & 0x1]);
269 }
270
271 #ifdef CONFIG_HOTPLUG_CPU
272
273 /*
274  * Handle tasklist migration for case in which all CPUs covered by the
275  * specified rcu_node have gone offline.  Move them up to the root
276  * rcu_node.  The reason for not just moving them to the immediate
277  * parent is to remove the need for rcu_read_unlock_special() to
278  * make more than two attempts to acquire the target rcu_node's lock.
279  *
280  * The caller must hold rnp->lock with irqs disabled.
281  */
282 static void rcu_preempt_offline_tasks(struct rcu_state *rsp,
283                                       struct rcu_node *rnp)
284 {
285         int i;
286         struct list_head *lp;
287         struct list_head *lp_root;
288         struct rcu_node *rnp_root = rcu_get_root(rsp);
289         struct task_struct *tp;
290
291         if (rnp == rnp_root)
292                 return;  /* Shouldn't happen: at least one CPU online. */
293
294         /*
295          * Move tasks up to root rcu_node.  Rely on the fact that the
296          * root rcu_node can be at most one ahead of the rest of the
297          * rcu_nodes in terms of gp_num value.  This fact allows us to
298          * move the blocked_tasks[] array directly, element by element.
299          */
300         for (i = 0; i < 2; i++) {
301                 lp = &rnp->blocked_tasks[i];
302                 lp_root = &rnp_root->blocked_tasks[i];
303                 while (!list_empty(lp)) {
304                         tp = list_entry(lp->next, typeof(*tp), rcu_node_entry);
305                         spin_lock(&rnp_root->lock); /* irqs already disabled */
306                         list_del(&tp->rcu_node_entry);
307                         tp->rcu_blocked_node = rnp_root;
308                         list_add(&tp->rcu_node_entry, lp_root);
309                         spin_unlock(&rnp_root->lock); /* irqs remain disabled */
310                 }
311         }
312 }
313
314 /*
315  * Do CPU-offline processing for preemptable RCU.
316  */
317 static void rcu_preempt_offline_cpu(int cpu)
318 {
319         __rcu_offline_cpu(cpu, &rcu_preempt_state);
320 }
321
322 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
323
324 /*
325  * Check for a quiescent state from the current CPU.  When a task blocks,
326  * the task is recorded in the corresponding CPU's rcu_node structure,
327  * which is checked elsewhere.
328  *
329  * Caller must disable hard irqs.
330  */
331 static void rcu_preempt_check_callbacks(int cpu)
332 {
333         struct task_struct *t = current;
334
335         if (t->rcu_read_lock_nesting == 0) {
336                 t->rcu_read_unlock_special &=
337                         ~(RCU_READ_UNLOCK_NEED_QS | RCU_READ_UNLOCK_GOT_QS);
338                 rcu_preempt_qs_record(cpu);
339                 return;
340         }
341         if (per_cpu(rcu_preempt_data, cpu).qs_pending) {
342                 if (t->rcu_read_unlock_special & RCU_READ_UNLOCK_GOT_QS) {
343                         rcu_preempt_qs_record(cpu);
344                         t->rcu_read_unlock_special &= ~RCU_READ_UNLOCK_GOT_QS;
345                 } else if (!(t->rcu_read_unlock_special &
346                              RCU_READ_UNLOCK_NEED_QS)) {
347                         t->rcu_read_unlock_special |= RCU_READ_UNLOCK_NEED_QS;
348                 }
349         }
350 }
351
352 /*
353  * Process callbacks for preemptable RCU.
354  */
355 static void rcu_preempt_process_callbacks(void)
356 {
357         __rcu_process_callbacks(&rcu_preempt_state,
358                                 &__get_cpu_var(rcu_preempt_data));
359 }
360
361 /*
362  * Queue a preemptable-RCU callback for invocation after a grace period.
363  */
364 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
365 {
366         __call_rcu(head, func, &rcu_preempt_state);
367 }
368 EXPORT_SYMBOL_GPL(call_rcu);
369
370 /*
371  * Check to see if there is any immediate preemptable-RCU-related work
372  * to be done.
373  */
374 static int rcu_preempt_pending(int cpu)
375 {
376         return __rcu_pending(&rcu_preempt_state,
377                              &per_cpu(rcu_preempt_data, cpu));
378 }
379
380 /*
381  * Does preemptable RCU need the CPU to stay out of dynticks mode?
382  */
383 static int rcu_preempt_needs_cpu(int cpu)
384 {
385         return !!per_cpu(rcu_preempt_data, cpu).nxtlist;
386 }
387
388 /*
389  * Initialize preemptable RCU's per-CPU data.
390  */
391 static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
392 {
393         rcu_init_percpu_data(cpu, &rcu_preempt_state, 1);
394 }
395
396 /*
397  * Check for a task exiting while in a preemptable-RCU read-side
398  * critical section, clean up if so.  No need to issue warnings,
399  * as debug_check_no_locks_held() already does this if lockdep
400  * is enabled.
401  */
402 void exit_rcu(void)
403 {
404         struct task_struct *t = current;
405
406         if (t->rcu_read_lock_nesting == 0)
407                 return;
408         t->rcu_read_lock_nesting = 1;
409         rcu_read_unlock();
410 }
411
412 #else /* #ifdef CONFIG_TREE_PREEMPT_RCU */
413
414 /*
415  * Tell them what RCU they are running.
416  */
417 static inline void rcu_bootup_announce(void)
418 {
419         printk(KERN_INFO "Hierarchical RCU implementation.\n");
420 }
421
422 /*
423  * Return the number of RCU batches processed thus far for debug & stats.
424  */
425 long rcu_batches_completed(void)
426 {
427         return rcu_batches_completed_sched();
428 }
429 EXPORT_SYMBOL_GPL(rcu_batches_completed);
430
431 /*
432  * Because preemptable RCU does not exist, we never have to check for
433  * CPUs being in quiescent states.
434  */
435 static void rcu_preempt_qs(int cpu)
436 {
437 }
438
439 #ifdef CONFIG_RCU_CPU_STALL_DETECTOR
440
441 /*
442  * Because preemptable RCU does not exist, we never have to check for
443  * tasks blocked within RCU read-side critical sections.
444  */
445 static void rcu_print_task_stall(struct rcu_node *rnp)
446 {
447 }
448
449 #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */
450
451 /*
452  * Because preemptable RCU does not exist, there are never any preempted
453  * RCU readers.
454  */
455 static int rcu_preempted_readers(struct rcu_node *rnp)
456 {
457         return 0;
458 }
459
460 #ifdef CONFIG_HOTPLUG_CPU
461
462 /*
463  * Because preemptable RCU does not exist, it never needs to migrate
464  * tasks that were blocked within RCU read-side critical sections.
465  */
466 static void rcu_preempt_offline_tasks(struct rcu_state *rsp,
467                                       struct rcu_node *rnp)
468 {
469 }
470
471 /*
472  * Because preemptable RCU does not exist, it never needs CPU-offline
473  * processing.
474  */
475 static void rcu_preempt_offline_cpu(int cpu)
476 {
477 }
478
479 #endif /* #ifdef CONFIG_HOTPLUG_CPU */
480
481 /*
482  * Because preemptable RCU does not exist, it never has any callbacks
483  * to check.
484  */
485 void rcu_preempt_check_callbacks(int cpu)
486 {
487 }
488
489 /*
490  * Because preemptable RCU does not exist, it never has any callbacks
491  * to process.
492  */
493 void rcu_preempt_process_callbacks(void)
494 {
495 }
496
497 /*
498  * In classic RCU, call_rcu() is just call_rcu_sched().
499  */
500 void call_rcu(struct rcu_head *head, void (*func)(struct rcu_head *rcu))
501 {
502         call_rcu_sched(head, func);
503 }
504 EXPORT_SYMBOL_GPL(call_rcu);
505
506 /*
507  * Because preemptable RCU does not exist, it never has any work to do.
508  */
509 static int rcu_preempt_pending(int cpu)
510 {
511         return 0;
512 }
513
514 /*
515  * Because preemptable RCU does not exist, it never needs any CPU.
516  */
517 static int rcu_preempt_needs_cpu(int cpu)
518 {
519         return 0;
520 }
521
522 /*
523  * Because preemptable RCU does not exist, there is no per-CPU
524  * data to initialize.
525  */
526 static void __cpuinit rcu_preempt_init_percpu_data(int cpu)
527 {
528 }
529
530 #endif /* #else #ifdef CONFIG_TREE_PREEMPT_RCU */