Commit | Line | Data |
---|---|---|
01c1c660 PM |
1 | /* |
2 | * Read-Copy Update mechanism for mutual exclusion | |
3 | * | |
4 | * This program is free software; you can redistribute it and/or modify | |
5 | * it under the terms of the GNU General Public License as published by | |
6 | * the Free Software Foundation; either version 2 of the License, or | |
7 | * (at your option) any later version. | |
8 | * | |
9 | * This program is distributed in the hope that it will be useful, | |
10 | * but WITHOUT ANY WARRANTY; without even the implied warranty of | |
11 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the | |
12 | * GNU General Public License for more details. | |
13 | * | |
14 | * You should have received a copy of the GNU General Public License | |
15 | * along with this program; if not, write to the Free Software | |
16 | * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111-1307, USA. | |
17 | * | |
18 | * Copyright IBM Corporation, 2001 | |
19 | * | |
20 | * Authors: Dipankar Sarma <dipankar@in.ibm.com> | |
21 | * Manfred Spraul <manfred@colorfullife.com> | |
22 | * | |
23 | * Based on the original work by Paul McKenney <paulmck@us.ibm.com> | |
24 | * and inputs from Rusty Russell, Andrea Arcangeli and Andi Kleen. | |
25 | * Papers: | |
26 | * http://www.rdrop.com/users/paulmck/paper/rclockpdcsproof.pdf | |
27 | * http://lse.sourceforge.net/locking/rclock_OLS.2001.05.01c.sc.pdf (OLS2001) | |
28 | * | |
29 | * For detailed explanation of Read-Copy Update mechanism see - | |
30 | * Documentation/RCU | |
31 | * | |
32 | */ | |
33 | #include <linux/types.h> | |
34 | #include <linux/kernel.h> | |
35 | #include <linux/init.h> | |
36 | #include <linux/spinlock.h> | |
37 | #include <linux/smp.h> | |
38 | #include <linux/rcupdate.h> | |
39 | #include <linux/interrupt.h> | |
40 | #include <linux/sched.h> | |
41 | #include <asm/atomic.h> | |
42 | #include <linux/bitops.h> | |
43 | #include <linux/module.h> | |
44 | #include <linux/completion.h> | |
45 | #include <linux/moduleparam.h> | |
46 | #include <linux/percpu.h> | |
47 | #include <linux/notifier.h> | |
01c1c660 PM |
48 | #include <linux/cpu.h> |
49 | #include <linux/mutex.h> | |
67182ae1 | 50 | #include <linux/time.h> |
01c1c660 PM |
51 | |
52 | #ifdef CONFIG_DEBUG_LOCK_ALLOC | |
53 | static struct lock_class_key rcu_lock_key; | |
54 | struct lockdep_map rcu_lock_map = | |
55 | STATIC_LOCKDEP_MAP_INIT("rcu_read_lock", &rcu_lock_key); | |
56 | EXPORT_SYMBOL_GPL(rcu_lock_map); | |
57 | #endif | |
58 | ||
59 | ||
60 | /* Definition for rcupdate control block. */ | |
61 | static struct rcu_ctrlblk rcu_ctrlblk = { | |
62 | .cur = -300, | |
63 | .completed = -300, | |
3cac97cb | 64 | .pending = -300, |
01c1c660 | 65 | .lock = __SPIN_LOCK_UNLOCKED(&rcu_ctrlblk.lock), |
bd232f97 | 66 | .cpumask = CPU_BITS_NONE, |
01c1c660 PM |
67 | }; |
68 | static struct rcu_ctrlblk rcu_bh_ctrlblk = { | |
69 | .cur = -300, | |
70 | .completed = -300, | |
3cac97cb | 71 | .pending = -300, |
01c1c660 | 72 | .lock = __SPIN_LOCK_UNLOCKED(&rcu_bh_ctrlblk.lock), |
bd232f97 | 73 | .cpumask = CPU_BITS_NONE, |
01c1c660 PM |
74 | }; |
75 | ||
76 | DEFINE_PER_CPU(struct rcu_data, rcu_data) = { 0L }; | |
77 | DEFINE_PER_CPU(struct rcu_data, rcu_bh_data) = { 0L }; | |
78 | ||
79 | static int blimit = 10; | |
80 | static int qhimark = 10000; | |
81 | static int qlowmark = 100; | |
82 | ||
83 | #ifdef CONFIG_SMP | |
84 | static void force_quiescent_state(struct rcu_data *rdp, | |
85 | struct rcu_ctrlblk *rcp) | |
86 | { | |
87 | int cpu; | |
eff9b713 PM |
88 | unsigned long flags; |
89 | ||
01c1c660 | 90 | set_need_resched(); |
eff9b713 | 91 | spin_lock_irqsave(&rcp->lock, flags); |
01c1c660 PM |
92 | if (unlikely(!rcp->signaled)) { |
93 | rcp->signaled = 1; | |
94 | /* | |
95 | * Don't send IPI to itself. With irqs disabled, | |
96 | * rdp->cpu is the current cpu. | |
8558f8f8 | 97 | * |
bd232f97 | 98 | * cpu_online_mask is updated by the _cpu_down() |
9b1a4d38 RR |
99 | * using __stop_machine(). Since we're in irqs disabled |
100 | * section, __stop_machine() is not exectuting, hence | |
bd232f97 | 101 | * the cpu_online_mask is stable. |
8558f8f8 GS |
102 | * |
103 | * However, a cpu might have been offlined _just_ before | |
104 | * we disabled irqs while entering here. | |
105 | * And rcu subsystem might not yet have handled the CPU_DEAD | |
106 | * notification, leading to the offlined cpu's bit | |
107 | * being set in the rcp->cpumask. | |
108 | * | |
bd232f97 | 109 | * Hence cpumask = (rcp->cpumask & cpu_online_mask) to prevent |
8558f8f8 | 110 | * sending smp_reschedule() to an offlined CPU. |
01c1c660 | 111 | */ |
bd232f97 RR |
112 | for_each_cpu_and(cpu, |
113 | to_cpumask(rcp->cpumask), cpu_online_mask) { | |
114 | if (cpu != rdp->cpu) | |
115 | smp_send_reschedule(cpu); | |
116 | } | |
01c1c660 | 117 | } |
eff9b713 | 118 | spin_unlock_irqrestore(&rcp->lock, flags); |
01c1c660 PM |
119 | } |
120 | #else | |
121 | static inline void force_quiescent_state(struct rcu_data *rdp, | |
122 | struct rcu_ctrlblk *rcp) | |
123 | { | |
124 | set_need_resched(); | |
125 | } | |
126 | #endif | |
127 | ||
5127bed5 LJ |
128 | static void __call_rcu(struct rcu_head *head, struct rcu_ctrlblk *rcp, |
129 | struct rcu_data *rdp) | |
130 | { | |
131 | long batch; | |
1f7b94cd PM |
132 | |
133 | head->next = NULL; | |
134 | smp_mb(); /* Read of rcu->cur must happen after any change by caller. */ | |
5127bed5 LJ |
135 | |
136 | /* | |
137 | * Determine the batch number of this callback. | |
138 | * | |
139 | * Using ACCESS_ONCE to avoid the following error when gcc eliminates | |
140 | * local variable "batch" and emits codes like this: | |
141 | * 1) rdp->batch = rcp->cur + 1 # gets old value | |
142 | * ...... | |
143 | * 2)rcu_batch_after(rcp->cur + 1, rdp->batch) # gets new value | |
144 | * then [*nxttail[0], *nxttail[1]) may contain callbacks | |
145 | * that batch# = rdp->batch, see the comment of struct rcu_data. | |
146 | */ | |
147 | batch = ACCESS_ONCE(rcp->cur) + 1; | |
148 | ||
149 | if (rdp->nxtlist && rcu_batch_after(batch, rdp->batch)) { | |
150 | /* process callbacks */ | |
151 | rdp->nxttail[0] = rdp->nxttail[1]; | |
152 | rdp->nxttail[1] = rdp->nxttail[2]; | |
153 | if (rcu_batch_after(batch - 1, rdp->batch)) | |
154 | rdp->nxttail[0] = rdp->nxttail[2]; | |
155 | } | |
156 | ||
157 | rdp->batch = batch; | |
158 | *rdp->nxttail[2] = head; | |
159 | rdp->nxttail[2] = &head->next; | |
160 | ||
161 | if (unlikely(++rdp->qlen > qhimark)) { | |
162 | rdp->blimit = INT_MAX; | |
163 | force_quiescent_state(rdp, &rcu_ctrlblk); | |
164 | } | |
165 | } | |
166 | ||
2133b5d7 PM |
167 | #ifdef CONFIG_RCU_CPU_STALL_DETECTOR |
168 | ||
169 | static void record_gp_stall_check_time(struct rcu_ctrlblk *rcp) | |
170 | { | |
171 | rcp->gp_start = jiffies; | |
172 | rcp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_CHECK; | |
173 | } | |
174 | ||
175 | static void print_other_cpu_stall(struct rcu_ctrlblk *rcp) | |
176 | { | |
177 | int cpu; | |
178 | long delta; | |
179 | unsigned long flags; | |
180 | ||
181 | /* Only let one CPU complain about others per time interval. */ | |
182 | ||
183 | spin_lock_irqsave(&rcp->lock, flags); | |
184 | delta = jiffies - rcp->jiffies_stall; | |
185 | if (delta < 2 || rcp->cur != rcp->completed) { | |
186 | spin_unlock_irqrestore(&rcp->lock, flags); | |
187 | return; | |
188 | } | |
189 | rcp->jiffies_stall = jiffies + RCU_SECONDS_TILL_STALL_RECHECK; | |
190 | spin_unlock_irqrestore(&rcp->lock, flags); | |
191 | ||
192 | /* OK, time to rat on our buddy... */ | |
193 | ||
be19ef82 | 194 | printk(KERN_ERR "INFO: RCU detected CPU stalls:"); |
2133b5d7 | 195 | for_each_possible_cpu(cpu) { |
bd232f97 | 196 | if (cpumask_test_cpu(cpu, to_cpumask(rcp->cpumask))) |
2133b5d7 PM |
197 | printk(" %d", cpu); |
198 | } | |
199 | printk(" (detected by %d, t=%ld jiffies)\n", | |
200 | smp_processor_id(), (long)(jiffies - rcp->gp_start)); | |
201 | } | |
202 | ||
203 | static void print_cpu_stall(struct rcu_ctrlblk *rcp) | |
204 | { | |
205 | unsigned long flags; | |
206 | ||
be19ef82 | 207 | printk(KERN_ERR "INFO: RCU detected CPU %d stall (t=%lu/%lu jiffies)\n", |
2133b5d7 PM |
208 | smp_processor_id(), jiffies, |
209 | jiffies - rcp->gp_start); | |
210 | dump_stack(); | |
211 | spin_lock_irqsave(&rcp->lock, flags); | |
212 | if ((long)(jiffies - rcp->jiffies_stall) >= 0) | |
213 | rcp->jiffies_stall = | |
214 | jiffies + RCU_SECONDS_TILL_STALL_RECHECK; | |
215 | spin_unlock_irqrestore(&rcp->lock, flags); | |
216 | set_need_resched(); /* kick ourselves to get things going. */ | |
217 | } | |
218 | ||
219 | static void check_cpu_stall(struct rcu_ctrlblk *rcp) | |
220 | { | |
221 | long delta; | |
222 | ||
223 | delta = jiffies - rcp->jiffies_stall; | |
bd232f97 RR |
224 | if (cpumask_test_cpu(smp_processor_id(), to_cpumask(rcp->cpumask)) && |
225 | delta >= 0) { | |
2133b5d7 PM |
226 | |
227 | /* We haven't checked in, so go dump stack. */ | |
228 | print_cpu_stall(rcp); | |
229 | ||
230 | } else if (rcp->cur != rcp->completed && delta >= 2) { | |
231 | ||
232 | /* They had two seconds to dump stack, so complain. */ | |
233 | print_other_cpu_stall(rcp); | |
234 | } | |
235 | } | |
236 | ||
237 | #else /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */ | |
238 | ||
239 | static void record_gp_stall_check_time(struct rcu_ctrlblk *rcp) | |
240 | { | |
241 | } | |
242 | ||
2ec2b482 | 243 | static inline void check_cpu_stall(struct rcu_ctrlblk *rcp) |
2133b5d7 PM |
244 | { |
245 | } | |
246 | ||
247 | #endif /* #else #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */ | |
248 | ||
01c1c660 PM |
249 | /** |
250 | * call_rcu - Queue an RCU callback for invocation after a grace period. | |
251 | * @head: structure to be used for queueing the RCU updates. | |
252 | * @func: actual update function to be invoked after the grace period | |
253 | * | |
254 | * The update function will be invoked some time after a full grace | |
255 | * period elapses, in other words after all currently executing RCU | |
256 | * read-side critical sections have completed. RCU read-side critical | |
257 | * sections are delimited by rcu_read_lock() and rcu_read_unlock(), | |
258 | * and may be nested. | |
259 | */ | |
260 | void call_rcu(struct rcu_head *head, | |
261 | void (*func)(struct rcu_head *rcu)) | |
262 | { | |
263 | unsigned long flags; | |
01c1c660 PM |
264 | |
265 | head->func = func; | |
01c1c660 | 266 | local_irq_save(flags); |
5127bed5 | 267 | __call_rcu(head, &rcu_ctrlblk, &__get_cpu_var(rcu_data)); |
01c1c660 PM |
268 | local_irq_restore(flags); |
269 | } | |
270 | EXPORT_SYMBOL_GPL(call_rcu); | |
271 | ||
272 | /** | |
273 | * call_rcu_bh - Queue an RCU for invocation after a quicker grace period. | |
274 | * @head: structure to be used for queueing the RCU updates. | |
275 | * @func: actual update function to be invoked after the grace period | |
276 | * | |
277 | * The update function will be invoked some time after a full grace | |
278 | * period elapses, in other words after all currently executing RCU | |
279 | * read-side critical sections have completed. call_rcu_bh() assumes | |
280 | * that the read-side critical sections end on completion of a softirq | |
281 | * handler. This means that read-side critical sections in process | |
282 | * context must not be interrupted by softirqs. This interface is to be | |
283 | * used when most of the read-side critical sections are in softirq context. | |
284 | * RCU read-side critical sections are delimited by rcu_read_lock() and | |
285 | * rcu_read_unlock(), * if in interrupt context or rcu_read_lock_bh() | |
286 | * and rcu_read_unlock_bh(), if in process context. These may be nested. | |
287 | */ | |
288 | void call_rcu_bh(struct rcu_head *head, | |
289 | void (*func)(struct rcu_head *rcu)) | |
290 | { | |
291 | unsigned long flags; | |
01c1c660 PM |
292 | |
293 | head->func = func; | |
01c1c660 | 294 | local_irq_save(flags); |
5127bed5 | 295 | __call_rcu(head, &rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data)); |
01c1c660 PM |
296 | local_irq_restore(flags); |
297 | } | |
298 | EXPORT_SYMBOL_GPL(call_rcu_bh); | |
299 | ||
300 | /* | |
301 | * Return the number of RCU batches processed thus far. Useful | |
302 | * for debug and statistics. | |
303 | */ | |
304 | long rcu_batches_completed(void) | |
305 | { | |
306 | return rcu_ctrlblk.completed; | |
307 | } | |
308 | EXPORT_SYMBOL_GPL(rcu_batches_completed); | |
309 | ||
310 | /* | |
311 | * Return the number of RCU batches processed thus far. Useful | |
312 | * for debug and statistics. | |
313 | */ | |
314 | long rcu_batches_completed_bh(void) | |
315 | { | |
316 | return rcu_bh_ctrlblk.completed; | |
317 | } | |
318 | EXPORT_SYMBOL_GPL(rcu_batches_completed_bh); | |
319 | ||
320 | /* Raises the softirq for processing rcu_callbacks. */ | |
321 | static inline void raise_rcu_softirq(void) | |
322 | { | |
323 | raise_softirq(RCU_SOFTIRQ); | |
01c1c660 PM |
324 | } |
325 | ||
326 | /* | |
327 | * Invoke the completed RCU callbacks. They are expected to be in | |
328 | * a per-cpu list. | |
329 | */ | |
330 | static void rcu_do_batch(struct rcu_data *rdp) | |
331 | { | |
275a89bd | 332 | unsigned long flags; |
01c1c660 PM |
333 | struct rcu_head *next, *list; |
334 | int count = 0; | |
335 | ||
336 | list = rdp->donelist; | |
337 | while (list) { | |
338 | next = list->next; | |
339 | prefetch(next); | |
340 | list->func(list); | |
341 | list = next; | |
342 | if (++count >= rdp->blimit) | |
343 | break; | |
344 | } | |
345 | rdp->donelist = list; | |
346 | ||
275a89bd | 347 | local_irq_save(flags); |
01c1c660 | 348 | rdp->qlen -= count; |
275a89bd | 349 | local_irq_restore(flags); |
01c1c660 PM |
350 | if (rdp->blimit == INT_MAX && rdp->qlen <= qlowmark) |
351 | rdp->blimit = blimit; | |
352 | ||
353 | if (!rdp->donelist) | |
354 | rdp->donetail = &rdp->donelist; | |
355 | else | |
356 | raise_rcu_softirq(); | |
357 | } | |
358 | ||
359 | /* | |
360 | * Grace period handling: | |
361 | * The grace period handling consists out of two steps: | |
362 | * - A new grace period is started. | |
363 | * This is done by rcu_start_batch. The start is not broadcasted to | |
364 | * all cpus, they must pick this up by comparing rcp->cur with | |
365 | * rdp->quiescbatch. All cpus are recorded in the | |
366 | * rcu_ctrlblk.cpumask bitmap. | |
367 | * - All cpus must go through a quiescent state. | |
368 | * Since the start of the grace period is not broadcasted, at least two | |
369 | * calls to rcu_check_quiescent_state are required: | |
370 | * The first call just notices that a new grace period is running. The | |
371 | * following calls check if there was a quiescent state since the beginning | |
372 | * of the grace period. If so, it updates rcu_ctrlblk.cpumask. If | |
373 | * the bitmap is empty, then the grace period is completed. | |
374 | * rcu_check_quiescent_state calls rcu_start_batch(0) to start the next grace | |
375 | * period (if necessary). | |
376 | */ | |
67182ae1 | 377 | |
01c1c660 PM |
378 | /* |
379 | * Register a new batch of callbacks, and start it up if there is currently no | |
380 | * active batch and the batch to be registered has not already occurred. | |
381 | * Caller must hold rcu_ctrlblk.lock. | |
382 | */ | |
383 | static void rcu_start_batch(struct rcu_ctrlblk *rcp) | |
384 | { | |
3cac97cb | 385 | if (rcp->cur != rcp->pending && |
01c1c660 | 386 | rcp->completed == rcp->cur) { |
01c1c660 | 387 | rcp->cur++; |
2133b5d7 | 388 | record_gp_stall_check_time(rcp); |
01c1c660 PM |
389 | |
390 | /* | |
391 | * Accessing nohz_cpu_mask before incrementing rcp->cur needs a | |
392 | * Barrier Otherwise it can cause tickless idle CPUs to be | |
393 | * included in rcp->cpumask, which will extend graceperiods | |
394 | * unnecessarily. | |
395 | */ | |
396 | smp_mb(); | |
6a7b3dc3 | 397 | cpumask_andnot(&rcp->cpumask, cpu_online_mask, nohz_cpu_mask); |
01c1c660 PM |
398 | |
399 | rcp->signaled = 0; | |
400 | } | |
401 | } | |
402 | ||
403 | /* | |
404 | * cpu went through a quiescent state since the beginning of the grace period. | |
405 | * Clear it from the cpu mask and complete the grace period if it was the last | |
406 | * cpu. Start another grace period if someone has further entries pending | |
407 | */ | |
408 | static void cpu_quiet(int cpu, struct rcu_ctrlblk *rcp) | |
409 | { | |
bd232f97 RR |
410 | cpumask_clear_cpu(cpu, to_cpumask(rcp->cpumask)); |
411 | if (cpumask_empty(to_cpumask(rcp->cpumask))) { | |
01c1c660 PM |
412 | /* batch completed ! */ |
413 | rcp->completed = rcp->cur; | |
414 | rcu_start_batch(rcp); | |
415 | } | |
416 | } | |
417 | ||
418 | /* | |
419 | * Check if the cpu has gone through a quiescent state (say context | |
420 | * switch). If so and if it already hasn't done so in this RCU | |
421 | * quiescent cycle, then indicate that it has done so. | |
422 | */ | |
423 | static void rcu_check_quiescent_state(struct rcu_ctrlblk *rcp, | |
424 | struct rcu_data *rdp) | |
425 | { | |
eff9b713 PM |
426 | unsigned long flags; |
427 | ||
01c1c660 PM |
428 | if (rdp->quiescbatch != rcp->cur) { |
429 | /* start new grace period: */ | |
430 | rdp->qs_pending = 1; | |
431 | rdp->passed_quiesc = 0; | |
432 | rdp->quiescbatch = rcp->cur; | |
433 | return; | |
434 | } | |
435 | ||
436 | /* Grace period already completed for this cpu? | |
437 | * qs_pending is checked instead of the actual bitmap to avoid | |
438 | * cacheline trashing. | |
439 | */ | |
440 | if (!rdp->qs_pending) | |
441 | return; | |
442 | ||
443 | /* | |
444 | * Was there a quiescent state since the beginning of the grace | |
445 | * period? If no, then exit and wait for the next call. | |
446 | */ | |
447 | if (!rdp->passed_quiesc) | |
448 | return; | |
449 | rdp->qs_pending = 0; | |
450 | ||
eff9b713 | 451 | spin_lock_irqsave(&rcp->lock, flags); |
01c1c660 PM |
452 | /* |
453 | * rdp->quiescbatch/rcp->cur and the cpu bitmap can come out of sync | |
454 | * during cpu startup. Ignore the quiescent state. | |
455 | */ | |
456 | if (likely(rdp->quiescbatch == rcp->cur)) | |
457 | cpu_quiet(rdp->cpu, rcp); | |
458 | ||
eff9b713 | 459 | spin_unlock_irqrestore(&rcp->lock, flags); |
01c1c660 PM |
460 | } |
461 | ||
462 | ||
463 | #ifdef CONFIG_HOTPLUG_CPU | |
464 | ||
465 | /* warning! helper for rcu_offline_cpu. do not use elsewhere without reviewing | |
466 | * locking requirements, the list it's pulling from has to belong to a cpu | |
467 | * which is dead and hence not processing interrupts. | |
468 | */ | |
469 | static void rcu_move_batch(struct rcu_data *this_rdp, struct rcu_head *list, | |
5127bed5 | 470 | struct rcu_head **tail, long batch) |
01c1c660 | 471 | { |
275a89bd PM |
472 | unsigned long flags; |
473 | ||
5127bed5 | 474 | if (list) { |
275a89bd | 475 | local_irq_save(flags); |
5127bed5 LJ |
476 | this_rdp->batch = batch; |
477 | *this_rdp->nxttail[2] = list; | |
478 | this_rdp->nxttail[2] = tail; | |
275a89bd | 479 | local_irq_restore(flags); |
5127bed5 | 480 | } |
01c1c660 PM |
481 | } |
482 | ||
483 | static void __rcu_offline_cpu(struct rcu_data *this_rdp, | |
484 | struct rcu_ctrlblk *rcp, struct rcu_data *rdp) | |
485 | { | |
eff9b713 PM |
486 | unsigned long flags; |
487 | ||
1f7b94cd PM |
488 | /* |
489 | * if the cpu going offline owns the grace period | |
01c1c660 PM |
490 | * we can block indefinitely waiting for it, so flush |
491 | * it here | |
492 | */ | |
eff9b713 | 493 | spin_lock_irqsave(&rcp->lock, flags); |
01c1c660 PM |
494 | if (rcp->cur != rcp->completed) |
495 | cpu_quiet(rdp->cpu, rcp); | |
5127bed5 LJ |
496 | rcu_move_batch(this_rdp, rdp->donelist, rdp->donetail, rcp->cur + 1); |
497 | rcu_move_batch(this_rdp, rdp->nxtlist, rdp->nxttail[2], rcp->cur + 1); | |
eff9b713 | 498 | spin_unlock(&rcp->lock); |
199a9528 | 499 | |
199a9528 | 500 | this_rdp->qlen += rdp->qlen; |
eff9b713 | 501 | local_irq_restore(flags); |
01c1c660 PM |
502 | } |
503 | ||
504 | static void rcu_offline_cpu(int cpu) | |
505 | { | |
506 | struct rcu_data *this_rdp = &get_cpu_var(rcu_data); | |
507 | struct rcu_data *this_bh_rdp = &get_cpu_var(rcu_bh_data); | |
508 | ||
509 | __rcu_offline_cpu(this_rdp, &rcu_ctrlblk, | |
510 | &per_cpu(rcu_data, cpu)); | |
511 | __rcu_offline_cpu(this_bh_rdp, &rcu_bh_ctrlblk, | |
512 | &per_cpu(rcu_bh_data, cpu)); | |
513 | put_cpu_var(rcu_data); | |
514 | put_cpu_var(rcu_bh_data); | |
515 | } | |
516 | ||
517 | #else | |
518 | ||
519 | static void rcu_offline_cpu(int cpu) | |
520 | { | |
521 | } | |
522 | ||
523 | #endif | |
524 | ||
525 | /* | |
526 | * This does the RCU processing work from softirq context. | |
527 | */ | |
528 | static void __rcu_process_callbacks(struct rcu_ctrlblk *rcp, | |
529 | struct rcu_data *rdp) | |
530 | { | |
275a89bd | 531 | unsigned long flags; |
1f7b94cd PM |
532 | long completed_snap; |
533 | ||
5127bed5 | 534 | if (rdp->nxtlist) { |
275a89bd | 535 | local_irq_save(flags); |
1f7b94cd | 536 | completed_snap = ACCESS_ONCE(rcp->completed); |
01c1c660 PM |
537 | |
538 | /* | |
5127bed5 LJ |
539 | * move the other grace-period-completed entries to |
540 | * [rdp->nxtlist, *rdp->nxttail[0]) temporarily | |
541 | */ | |
1f7b94cd | 542 | if (!rcu_batch_before(completed_snap, rdp->batch)) |
5127bed5 | 543 | rdp->nxttail[0] = rdp->nxttail[1] = rdp->nxttail[2]; |
1f7b94cd | 544 | else if (!rcu_batch_before(completed_snap, rdp->batch - 1)) |
5127bed5 LJ |
545 | rdp->nxttail[0] = rdp->nxttail[1]; |
546 | ||
547 | /* | |
548 | * the grace period for entries in | |
549 | * [rdp->nxtlist, *rdp->nxttail[0]) has completed and | |
550 | * move these entries to donelist | |
01c1c660 | 551 | */ |
5127bed5 LJ |
552 | if (rdp->nxttail[0] != &rdp->nxtlist) { |
553 | *rdp->donetail = rdp->nxtlist; | |
554 | rdp->donetail = rdp->nxttail[0]; | |
555 | rdp->nxtlist = *rdp->nxttail[0]; | |
556 | *rdp->donetail = NULL; | |
557 | ||
558 | if (rdp->nxttail[1] == rdp->nxttail[0]) | |
559 | rdp->nxttail[1] = &rdp->nxtlist; | |
560 | if (rdp->nxttail[2] == rdp->nxttail[0]) | |
561 | rdp->nxttail[2] = &rdp->nxtlist; | |
562 | rdp->nxttail[0] = &rdp->nxtlist; | |
563 | } | |
01c1c660 | 564 | |
275a89bd | 565 | local_irq_restore(flags); |
01c1c660 | 566 | |
3cac97cb | 567 | if (rcu_batch_after(rdp->batch, rcp->pending)) { |
3a72dc8e | 568 | unsigned long flags2; |
0c925d79 | 569 | |
01c1c660 | 570 | /* and start it/schedule start if it's a new batch */ |
3a72dc8e | 571 | spin_lock_irqsave(&rcp->lock, flags2); |
3cac97cb LJ |
572 | if (rcu_batch_after(rdp->batch, rcp->pending)) { |
573 | rcp->pending = rdp->batch; | |
574 | rcu_start_batch(rcp); | |
575 | } | |
3a72dc8e | 576 | spin_unlock_irqrestore(&rcp->lock, flags2); |
01c1c660 PM |
577 | } |
578 | } | |
579 | ||
580 | rcu_check_quiescent_state(rcp, rdp); | |
581 | if (rdp->donelist) | |
582 | rcu_do_batch(rdp); | |
583 | } | |
584 | ||
585 | static void rcu_process_callbacks(struct softirq_action *unused) | |
586 | { | |
1f7b94cd PM |
587 | /* |
588 | * Memory references from any prior RCU read-side critical sections | |
589 | * executed by the interrupted code must be see before any RCU | |
590 | * grace-period manupulations below. | |
591 | */ | |
592 | ||
593 | smp_mb(); /* See above block comment. */ | |
594 | ||
01c1c660 PM |
595 | __rcu_process_callbacks(&rcu_ctrlblk, &__get_cpu_var(rcu_data)); |
596 | __rcu_process_callbacks(&rcu_bh_ctrlblk, &__get_cpu_var(rcu_bh_data)); | |
1f7b94cd PM |
597 | |
598 | /* | |
599 | * Memory references from any later RCU read-side critical sections | |
600 | * executed by the interrupted code must be see after any RCU | |
601 | * grace-period manupulations above. | |
602 | */ | |
603 | ||
604 | smp_mb(); /* See above block comment. */ | |
01c1c660 PM |
605 | } |
606 | ||
607 | static int __rcu_pending(struct rcu_ctrlblk *rcp, struct rcu_data *rdp) | |
608 | { | |
67182ae1 | 609 | /* Check for CPU stalls, if enabled. */ |
2133b5d7 | 610 | check_cpu_stall(rcp); |
67182ae1 | 611 | |
5127bed5 | 612 | if (rdp->nxtlist) { |
1f7b94cd PM |
613 | long completed_snap = ACCESS_ONCE(rcp->completed); |
614 | ||
5127bed5 LJ |
615 | /* |
616 | * This cpu has pending rcu entries and the grace period | |
617 | * for them has completed. | |
618 | */ | |
1f7b94cd | 619 | if (!rcu_batch_before(completed_snap, rdp->batch)) |
5127bed5 | 620 | return 1; |
1f7b94cd | 621 | if (!rcu_batch_before(completed_snap, rdp->batch - 1) && |
5127bed5 LJ |
622 | rdp->nxttail[0] != rdp->nxttail[1]) |
623 | return 1; | |
624 | if (rdp->nxttail[0] != &rdp->nxtlist) | |
625 | return 1; | |
01c1c660 | 626 | |
5127bed5 LJ |
627 | /* |
628 | * This cpu has pending rcu entries and the new batch | |
629 | * for then hasn't been started nor scheduled start | |
630 | */ | |
631 | if (rcu_batch_after(rdp->batch, rcp->pending)) | |
632 | return 1; | |
633 | } | |
01c1c660 PM |
634 | |
635 | /* This cpu has finished callbacks to invoke */ | |
636 | if (rdp->donelist) | |
637 | return 1; | |
638 | ||
639 | /* The rcu core waits for a quiescent state from the cpu */ | |
640 | if (rdp->quiescbatch != rcp->cur || rdp->qs_pending) | |
641 | return 1; | |
642 | ||
643 | /* nothing to do */ | |
644 | return 0; | |
645 | } | |
646 | ||
647 | /* | |
648 | * Check to see if there is any immediate RCU-related work to be done | |
649 | * by the current CPU, returning 1 if so. This function is part of the | |
650 | * RCU implementation; it is -not- an exported member of the RCU API. | |
651 | */ | |
652 | int rcu_pending(int cpu) | |
653 | { | |
654 | return __rcu_pending(&rcu_ctrlblk, &per_cpu(rcu_data, cpu)) || | |
655 | __rcu_pending(&rcu_bh_ctrlblk, &per_cpu(rcu_bh_data, cpu)); | |
656 | } | |
657 | ||
658 | /* | |
659 | * Check to see if any future RCU-related work will need to be done | |
660 | * by the current CPU, even if none need be done immediately, returning | |
661 | * 1 if so. This function is part of the RCU implementation; it is -not- | |
662 | * an exported member of the RCU API. | |
663 | */ | |
664 | int rcu_needs_cpu(int cpu) | |
665 | { | |
666 | struct rcu_data *rdp = &per_cpu(rcu_data, cpu); | |
667 | struct rcu_data *rdp_bh = &per_cpu(rcu_bh_data, cpu); | |
668 | ||
5127bed5 | 669 | return !!rdp->nxtlist || !!rdp_bh->nxtlist || rcu_pending(cpu); |
01c1c660 PM |
670 | } |
671 | ||
1f7b94cd PM |
672 | /* |
673 | * Top-level function driving RCU grace-period detection, normally | |
674 | * invoked from the scheduler-clock interrupt. This function simply | |
675 | * increments counters that are read only from softirq by this same | |
676 | * CPU, so there are no memory barriers required. | |
677 | */ | |
01c1c660 PM |
678 | void rcu_check_callbacks(int cpu, int user) |
679 | { | |
680 | if (user || | |
681 | (idle_cpu(cpu) && !in_softirq() && | |
682 | hardirq_count() <= (1 << HARDIRQ_SHIFT))) { | |
8db559b8 PM |
683 | |
684 | /* | |
685 | * Get here if this CPU took its interrupt from user | |
686 | * mode or from the idle loop, and if this is not a | |
687 | * nested interrupt. In this case, the CPU is in | |
688 | * a quiescent state, so count it. | |
689 | * | |
690 | * Also do a memory barrier. This is needed to handle | |
691 | * the case where writes from a preempt-disable section | |
692 | * of code get reordered into schedule() by this CPU's | |
693 | * write buffer. The memory barrier makes sure that | |
694 | * the rcu_qsctr_inc() and rcu_bh_qsctr_inc() are see | |
695 | * by other CPUs to happen after any such write. | |
696 | */ | |
697 | ||
698 | smp_mb(); /* See above block comment. */ | |
01c1c660 PM |
699 | rcu_qsctr_inc(cpu); |
700 | rcu_bh_qsctr_inc(cpu); | |
8db559b8 PM |
701 | |
702 | } else if (!in_softirq()) { | |
703 | ||
704 | /* | |
705 | * Get here if this CPU did not take its interrupt from | |
706 | * softirq, in other words, if it is not interrupting | |
707 | * a rcu_bh read-side critical section. This is an _bh | |
708 | * critical section, so count it. The memory barrier | |
709 | * is needed for the same reason as is the above one. | |
710 | */ | |
711 | ||
712 | smp_mb(); /* See above block comment. */ | |
01c1c660 | 713 | rcu_bh_qsctr_inc(cpu); |
8db559b8 | 714 | } |
01c1c660 PM |
715 | raise_rcu_softirq(); |
716 | } | |
717 | ||
718 | static void rcu_init_percpu_data(int cpu, struct rcu_ctrlblk *rcp, | |
719 | struct rcu_data *rdp) | |
720 | { | |
0cd418dd | 721 | unsigned long flags; |
cd958517 PM |
722 | |
723 | spin_lock_irqsave(&rcp->lock, flags); | |
01c1c660 | 724 | memset(rdp, 0, sizeof(*rdp)); |
5127bed5 | 725 | rdp->nxttail[0] = rdp->nxttail[1] = rdp->nxttail[2] = &rdp->nxtlist; |
01c1c660 PM |
726 | rdp->donetail = &rdp->donelist; |
727 | rdp->quiescbatch = rcp->completed; | |
728 | rdp->qs_pending = 0; | |
729 | rdp->cpu = cpu; | |
730 | rdp->blimit = blimit; | |
cd958517 | 731 | spin_unlock_irqrestore(&rcp->lock, flags); |
01c1c660 PM |
732 | } |
733 | ||
734 | static void __cpuinit rcu_online_cpu(int cpu) | |
735 | { | |
736 | struct rcu_data *rdp = &per_cpu(rcu_data, cpu); | |
737 | struct rcu_data *bh_rdp = &per_cpu(rcu_bh_data, cpu); | |
738 | ||
739 | rcu_init_percpu_data(cpu, &rcu_ctrlblk, rdp); | |
740 | rcu_init_percpu_data(cpu, &rcu_bh_ctrlblk, bh_rdp); | |
962cf36c | 741 | open_softirq(RCU_SOFTIRQ, rcu_process_callbacks); |
01c1c660 PM |
742 | } |
743 | ||
744 | static int __cpuinit rcu_cpu_notify(struct notifier_block *self, | |
745 | unsigned long action, void *hcpu) | |
746 | { | |
747 | long cpu = (long)hcpu; | |
748 | ||
749 | switch (action) { | |
750 | case CPU_UP_PREPARE: | |
751 | case CPU_UP_PREPARE_FROZEN: | |
752 | rcu_online_cpu(cpu); | |
753 | break; | |
754 | case CPU_DEAD: | |
755 | case CPU_DEAD_FROZEN: | |
756 | rcu_offline_cpu(cpu); | |
757 | break; | |
758 | default: | |
759 | break; | |
760 | } | |
761 | return NOTIFY_OK; | |
762 | } | |
763 | ||
764 | static struct notifier_block __cpuinitdata rcu_nb = { | |
765 | .notifier_call = rcu_cpu_notify, | |
766 | }; | |
767 | ||
768 | /* | |
769 | * Initializes rcu mechanism. Assumed to be called early. | |
770 | * That is before local timer(SMP) or jiffie timer (uniproc) is setup. | |
771 | * Note that rcu_qsctr and friends are implicitly | |
772 | * initialized due to the choice of ``0'' for RCU_CTR_INVALID. | |
773 | */ | |
774 | void __init __rcu_init(void) | |
775 | { | |
2133b5d7 PM |
776 | #ifdef CONFIG_RCU_CPU_STALL_DETECTOR |
777 | printk(KERN_INFO "RCU-based detection of stalled CPUs is enabled.\n"); | |
778 | #endif /* #ifdef CONFIG_RCU_CPU_STALL_DETECTOR */ | |
01c1c660 PM |
779 | rcu_cpu_notify(&rcu_nb, CPU_UP_PREPARE, |
780 | (void *)(long)smp_processor_id()); | |
781 | /* Register notifier for non-boot CPUs */ | |
782 | register_cpu_notifier(&rcu_nb); | |
783 | } | |
784 | ||
785 | module_param(blimit, int, 0); | |
786 | module_param(qhimark, int, 0); | |
787 | module_param(qlowmark, int, 0); |