projects/linux·blob v2.6.35 · refs

774 lines · 17.1 KiB

1/*
2 * padata.c - generic interface to process data streams in parallel
3 *
4 * Copyright (C) 2008, 2009 secunet Security Networks AG
5 * Copyright (C) 2008, 2009 Steffen Klassert <[email protected]>
6 *
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms and conditions of the GNU General Public License,
9 * version 2, as published by the Free Software Foundation.
10 *
11 * This program is distributed in the hope it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
14 * more details.
15 *
16 * You should have received a copy of the GNU General Public License along with
17 * this program; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
19 */
20
21#include <linux/module.h>
22#include <linux/cpumask.h>
23#include <linux/err.h>
24#include <linux/cpu.h>
25#include <linux/padata.h>
26#include <linux/mutex.h>
27#include <linux/sched.h>
28#include <linux/slab.h>
29#include <linux/rcupdate.h>
30
31#define MAX_SEQ_NR INT_MAX - NR_CPUS
32#define MAX_OBJ_NUM 1000
33
34static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
35{
36 int cpu, target_cpu;
37
38 target_cpu = cpumask_first(pd->cpumask);
39 for (cpu = 0; cpu < cpu_index; cpu++)
40 target_cpu = cpumask_next(target_cpu, pd->cpumask);
41
42 return target_cpu;
43}
44
45static int padata_cpu_hash(struct padata_priv *padata)
46{
47 int cpu_index;
48 struct parallel_data *pd;
49
50 pd = padata->pd;
51
52 /*
53 * Hash the sequence numbers to the cpus by taking
54 * seq_nr mod. number of cpus in use.
55 */
56 cpu_index = padata->seq_nr % cpumask_weight(pd->cpumask);
57
58 return padata_index_to_cpu(pd, cpu_index);
59}
60
61static void padata_parallel_worker(struct work_struct *work)
62{
63 struct padata_queue *queue;
64 struct parallel_data *pd;
65 struct padata_instance *pinst;
66 LIST_HEAD(local_list);
67
68 local_bh_disable();
69 queue = container_of(work, struct padata_queue, pwork);
70 pd = queue->pd;
71 pinst = pd->pinst;
72
73 spin_lock(&queue->parallel.lock);
74 list_replace_init(&queue->parallel.list, &local_list);
75 spin_unlock(&queue->parallel.lock);
76
77 while (!list_empty(&local_list)) {
78 struct padata_priv *padata;
79
80 padata = list_entry(local_list.next,
81 struct padata_priv, list);
82
83 list_del_init(&padata->list);
84
85 padata->parallel(padata);
86 }
87
88 local_bh_enable();
89}
90
91/**
92 * padata_do_parallel - padata parallelization function
93 *
94 * @pinst: padata instance
95 * @padata: object to be parallelized
96 * @cb_cpu: cpu the serialization callback function will run on,
97 * must be in the cpumask of padata.
98 *
99 * The parallelization callback function will run with BHs off.
100 * Note: Every object which is parallelized by padata_do_parallel
101 * must be seen by padata_do_serial.
102 */
103int padata_do_parallel(struct padata_instance *pinst,
104 struct padata_priv *padata, int cb_cpu)
105{
106 int target_cpu, err;
107 struct padata_queue *queue;
108 struct parallel_data *pd;
109
110 rcu_read_lock_bh();
111
112 pd = rcu_dereference(pinst->pd);
113
114 err = 0;
115 if (!(pinst->flags & PADATA_INIT))
116 goto out;
117
118 err = -EBUSY;
119 if ((pinst->flags & PADATA_RESET))
120 goto out;
121
122 if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
123 goto out;
124
125 err = -EINVAL;
126 if (!cpumask_test_cpu(cb_cpu, pd->cpumask))
127 goto out;
128
129 err = -EINPROGRESS;
130 atomic_inc(&pd->refcnt);
131 padata->pd = pd;
132 padata->cb_cpu = cb_cpu;
133
134 if (unlikely(atomic_read(&pd->seq_nr) == pd->max_seq_nr))
135 atomic_set(&pd->seq_nr, -1);
136
137 padata->seq_nr = atomic_inc_return(&pd->seq_nr);
138
139 target_cpu = padata_cpu_hash(padata);
140 queue = per_cpu_ptr(pd->queue, target_cpu);
141
142 spin_lock(&queue->parallel.lock);
143 list_add_tail(&padata->list, &queue->parallel.list);
144 spin_unlock(&queue->parallel.lock);
145
146 queue_work_on(target_cpu, pinst->wq, &queue->pwork);
147
148out:
149 rcu_read_unlock_bh();
150
151 return err;
152}
153EXPORT_SYMBOL(padata_do_parallel);
154
155/*
156 * padata_get_next - Get the next object that needs serialization.
157 *
158 * Return values are:
159 *
160 * A pointer to the control struct of the next object that needs
161 * serialization, if present in one of the percpu reorder queues.
162 *
163 * NULL, if all percpu reorder queues are empty.
164 *
165 * -EINPROGRESS, if the next object that needs serialization will
166 * be parallel processed by another cpu and is not yet present in
167 * the cpu's reorder queue.
168 *
169 * -ENODATA, if this cpu has to do the parallel processing for
170 * the next object.
171 */
172static struct padata_priv *padata_get_next(struct parallel_data *pd)
173{
174 int cpu, num_cpus, empty, calc_seq_nr;
175 int seq_nr, next_nr, overrun, next_overrun;
176 struct padata_queue *queue, *next_queue;
177 struct padata_priv *padata;
178 struct padata_list *reorder;
179
180 empty = 0;
181 next_nr = -1;
182 next_overrun = 0;
183 next_queue = NULL;
184
185 num_cpus = cpumask_weight(pd->cpumask);
186
187 for_each_cpu(cpu, pd->cpumask) {
188 queue = per_cpu_ptr(pd->queue, cpu);
189 reorder = &queue->reorder;
190
191 /*
192 * Calculate the seq_nr of the object that should be
193 * next in this reorder queue.
194 */
195 overrun = 0;
196 calc_seq_nr = (atomic_read(&queue->num_obj) * num_cpus)
197 + queue->cpu_index;
198
199 if (unlikely(calc_seq_nr > pd->max_seq_nr)) {
200 calc_seq_nr = calc_seq_nr - pd->max_seq_nr - 1;
201 overrun = 1;
202 }
203
204 if (!list_empty(&reorder->list)) {
205 padata = list_entry(reorder->list.next,
206 struct padata_priv, list);
207
208 seq_nr = padata->seq_nr;
209 BUG_ON(calc_seq_nr != seq_nr);
210 } else {
211 seq_nr = calc_seq_nr;
212 empty++;
213 }
214
215 if (next_nr < 0 || seq_nr < next_nr
216 || (next_overrun && !overrun)) {
217 next_nr = seq_nr;
218 next_overrun = overrun;
219 next_queue = queue;
220 }
221 }
222
223 padata = NULL;
224
225 if (empty == num_cpus)
226 goto out;
227
228 reorder = &next_queue->reorder;
229
230 if (!list_empty(&reorder->list)) {
231 padata = list_entry(reorder->list.next,
232 struct padata_priv, list);
233
234 if (unlikely(next_overrun)) {
235 for_each_cpu(cpu, pd->cpumask) {
236 queue = per_cpu_ptr(pd->queue, cpu);
237 atomic_set(&queue->num_obj, 0);
238 }
239 }
240
241 spin_lock(&reorder->lock);
242 list_del_init(&padata->list);
243 atomic_dec(&pd->reorder_objects);
244 spin_unlock(&reorder->lock);
245
246 atomic_inc(&next_queue->num_obj);
247
248 goto out;
249 }
250
251 queue = per_cpu_ptr(pd->queue, smp_processor_id());
252 if (queue->cpu_index == next_queue->cpu_index) {
253 padata = ERR_PTR(-ENODATA);
254 goto out;
255 }
256
257 padata = ERR_PTR(-EINPROGRESS);
258out:
259 return padata;
260}
261
262static void padata_reorder(struct parallel_data *pd)
263{
264 struct padata_priv *padata;
265 struct padata_queue *queue;
266 struct padata_instance *pinst = pd->pinst;
267
268 /*
269 * We need to ensure that only one cpu can work on dequeueing of
270 * the reorder queue the time. Calculating in which percpu reorder
271 * queue the next object will arrive takes some time. A spinlock
272 * would be highly contended. Also it is not clear in which order
273 * the objects arrive to the reorder queues. So a cpu could wait to
274 * get the lock just to notice that there is nothing to do at the
275 * moment. Therefore we use a trylock and let the holder of the lock
276 * care for all the objects enqueued during the holdtime of the lock.
277 */
278 if (!spin_trylock_bh(&pd->lock))
279 return;
280
281 while (1) {
282 padata = padata_get_next(pd);
283
284 /*
285 * All reorder queues are empty, or the next object that needs
286 * serialization is parallel processed by another cpu and is
287 * still on it's way to the cpu's reorder queue, nothing to
288 * do for now.
289 */
290 if (!padata || PTR_ERR(padata) == -EINPROGRESS)
291 break;
292
293 /*
294 * This cpu has to do the parallel processing of the next
295 * object. It's waiting in the cpu's parallelization queue,
296 * so exit imediately.
297 */
298 if (PTR_ERR(padata) == -ENODATA) {
299 del_timer(&pd->timer);
300 spin_unlock_bh(&pd->lock);
301 return;
302 }
303
304 queue = per_cpu_ptr(pd->queue, padata->cb_cpu);
305
306 spin_lock(&queue->serial.lock);
307 list_add_tail(&padata->list, &queue->serial.list);
308 spin_unlock(&queue->serial.lock);
309
310 queue_work_on(padata->cb_cpu, pinst->wq, &queue->swork);
311 }
312
313 spin_unlock_bh(&pd->lock);
314
315 /*
316 * The next object that needs serialization might have arrived to
317 * the reorder queues in the meantime, we will be called again
318 * from the timer function if noone else cares for it.
319 */
320 if (atomic_read(&pd->reorder_objects)
321 && !(pinst->flags & PADATA_RESET))
322 mod_timer(&pd->timer, jiffies + HZ);
323 else
324 del_timer(&pd->timer);
325
326 return;
327}
328
329static void padata_reorder_timer(unsigned long arg)
330{
331 struct parallel_data *pd = (struct parallel_data *)arg;
332
333 padata_reorder(pd);
334}
335
336static void padata_serial_worker(struct work_struct *work)
337{
338 struct padata_queue *queue;
339 struct parallel_data *pd;
340 LIST_HEAD(local_list);
341
342 local_bh_disable();
343 queue = container_of(work, struct padata_queue, swork);
344 pd = queue->pd;
345
346 spin_lock(&queue->serial.lock);
347 list_replace_init(&queue->serial.list, &local_list);
348 spin_unlock(&queue->serial.lock);
349
350 while (!list_empty(&local_list)) {
351 struct padata_priv *padata;
352
353 padata = list_entry(local_list.next,
354 struct padata_priv, list);
355
356 list_del_init(&padata->list);
357
358 padata->serial(padata);
359 atomic_dec(&pd->refcnt);
360 }
361 local_bh_enable();
362}
363
364/**
365 * padata_do_serial - padata serialization function
366 *
367 * @padata: object to be serialized.
368 *
369 * padata_do_serial must be called for every parallelized object.
370 * The serialization callback function will run with BHs off.
371 */
372void padata_do_serial(struct padata_priv *padata)
373{
374 int cpu;
375 struct padata_queue *queue;
376 struct parallel_data *pd;
377
378 pd = padata->pd;
379
380 cpu = get_cpu();
381 queue = per_cpu_ptr(pd->queue, cpu);
382
383 spin_lock(&queue->reorder.lock);
384 atomic_inc(&pd->reorder_objects);
385 list_add_tail(&padata->list, &queue->reorder.list);
386 spin_unlock(&queue->reorder.lock);
387
388 put_cpu();
389
390 padata_reorder(pd);
391}
392EXPORT_SYMBOL(padata_do_serial);
393
394/* Allocate and initialize the internal cpumask dependend resources. */
395static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
396 const struct cpumask *cpumask)
397{
398 int cpu, cpu_index, num_cpus;
399 struct padata_queue *queue;
400 struct parallel_data *pd;
401
402 cpu_index = 0;
403
404 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
405 if (!pd)
406 goto err;
407
408 pd->queue = alloc_percpu(struct padata_queue);
409 if (!pd->queue)
410 goto err_free_pd;
411
412 if (!alloc_cpumask_var(&pd->cpumask, GFP_KERNEL))
413 goto err_free_queue;
414
415 cpumask_and(pd->cpumask, cpumask, cpu_active_mask);
416
417 for_each_cpu(cpu, pd->cpumask) {
418 queue = per_cpu_ptr(pd->queue, cpu);
419
420 queue->pd = pd;
421
422 queue->cpu_index = cpu_index;
423 cpu_index++;
424
425 INIT_LIST_HEAD(&queue->reorder.list);
426 INIT_LIST_HEAD(&queue->parallel.list);
427 INIT_LIST_HEAD(&queue->serial.list);
428 spin_lock_init(&queue->reorder.lock);
429 spin_lock_init(&queue->parallel.lock);
430 spin_lock_init(&queue->serial.lock);
431
432 INIT_WORK(&queue->pwork, padata_parallel_worker);
433 INIT_WORK(&queue->swork, padata_serial_worker);
434 atomic_set(&queue->num_obj, 0);
435 }
436
437 num_cpus = cpumask_weight(pd->cpumask);
438 pd->max_seq_nr = (MAX_SEQ_NR / num_cpus) * num_cpus - 1;
439
440 setup_timer(&pd->timer, padata_reorder_timer, (unsigned long)pd);
441 atomic_set(&pd->seq_nr, -1);
442 atomic_set(&pd->reorder_objects, 0);
443 atomic_set(&pd->refcnt, 0);
444 pd->pinst = pinst;
445 spin_lock_init(&pd->lock);
446
447 return pd;
448
449err_free_queue:
450 free_percpu(pd->queue);
451err_free_pd:
452 kfree(pd);
453err:
454 return NULL;
455}
456
457static void padata_free_pd(struct parallel_data *pd)
458{
459 free_cpumask_var(pd->cpumask);
460 free_percpu(pd->queue);
461 kfree(pd);
462}
463
464/* Flush all objects out of the padata queues. */
465static void padata_flush_queues(struct parallel_data *pd)
466{
467 int cpu;
468 struct padata_queue *queue;
469
470 for_each_cpu(cpu, pd->cpumask) {
471 queue = per_cpu_ptr(pd->queue, cpu);
472 flush_work(&queue->pwork);
473 }
474
475 del_timer_sync(&pd->timer);
476
477 if (atomic_read(&pd->reorder_objects))
478 padata_reorder(pd);
479
480 for_each_cpu(cpu, pd->cpumask) {
481 queue = per_cpu_ptr(pd->queue, cpu);
482 flush_work(&queue->swork);
483 }
484
485 BUG_ON(atomic_read(&pd->refcnt) != 0);
486}
487
488/* Replace the internal control stucture with a new one. */
489static void padata_replace(struct padata_instance *pinst,
490 struct parallel_data *pd_new)
491{
492 struct parallel_data *pd_old = pinst->pd;
493
494 pinst->flags |= PADATA_RESET;
495
496 rcu_assign_pointer(pinst->pd, pd_new);
497
498 synchronize_rcu();
499
500 padata_flush_queues(pd_old);
501 padata_free_pd(pd_old);
502
503 pinst->flags &= ~PADATA_RESET;
504}
505
506/**
507 * padata_set_cpumask - set the cpumask that padata should use
508 *
509 * @pinst: padata instance
510 * @cpumask: the cpumask to use
511 */
512int padata_set_cpumask(struct padata_instance *pinst,
513 cpumask_var_t cpumask)
514{
515 struct parallel_data *pd;
516 int err = 0;
517
518 mutex_lock(&pinst->lock);
519
520 get_online_cpus();
521
522 pd = padata_alloc_pd(pinst, cpumask);
523 if (!pd) {
524 err = -ENOMEM;
525 goto out;
526 }
527
528 cpumask_copy(pinst->cpumask, cpumask);
529
530 padata_replace(pinst, pd);
531
532out:
533 put_online_cpus();
534
535 mutex_unlock(&pinst->lock);
536
537 return err;
538}
539EXPORT_SYMBOL(padata_set_cpumask);
540
541static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
542{
543 struct parallel_data *pd;
544
545 if (cpumask_test_cpu(cpu, cpu_active_mask)) {
546 pd = padata_alloc_pd(pinst, pinst->cpumask);
547 if (!pd)
548 return -ENOMEM;
549
550 padata_replace(pinst, pd);
551 }
552
553 return 0;
554}
555
556/**
557 * padata_add_cpu - add a cpu to the padata cpumask
558 *
559 * @pinst: padata instance
560 * @cpu: cpu to add
561 */
562int padata_add_cpu(struct padata_instance *pinst, int cpu)
563{
564 int err;
565
566 mutex_lock(&pinst->lock);
567
568 get_online_cpus();
569 cpumask_set_cpu(cpu, pinst->cpumask);
570 err = __padata_add_cpu(pinst, cpu);
571 put_online_cpus();
572
573 mutex_unlock(&pinst->lock);
574
575 return err;
576}
577EXPORT_SYMBOL(padata_add_cpu);
578
579static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
580{
581 struct parallel_data *pd;
582
583 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
584 pd = padata_alloc_pd(pinst, pinst->cpumask);
585 if (!pd)
586 return -ENOMEM;
587
588 padata_replace(pinst, pd);
589 }
590
591 return 0;
592}
593
594/**
595 * padata_remove_cpu - remove a cpu from the padata cpumask
596 *
597 * @pinst: padata instance
598 * @cpu: cpu to remove
599 */
600int padata_remove_cpu(struct padata_instance *pinst, int cpu)
601{
602 int err;
603
604 mutex_lock(&pinst->lock);
605
606 get_online_cpus();
607 cpumask_clear_cpu(cpu, pinst->cpumask);
608 err = __padata_remove_cpu(pinst, cpu);
609 put_online_cpus();
610
611 mutex_unlock(&pinst->lock);
612
613 return err;
614}
615EXPORT_SYMBOL(padata_remove_cpu);
616
617/**
618 * padata_start - start the parallel processing
619 *
620 * @pinst: padata instance to start
621 */
622void padata_start(struct padata_instance *pinst)
623{
624 mutex_lock(&pinst->lock);
625 pinst->flags |= PADATA_INIT;
626 mutex_unlock(&pinst->lock);
627}
628EXPORT_SYMBOL(padata_start);
629
630/**
631 * padata_stop - stop the parallel processing
632 *
633 * @pinst: padata instance to stop
634 */
635void padata_stop(struct padata_instance *pinst)
636{
637 mutex_lock(&pinst->lock);
638 pinst->flags &= ~PADATA_INIT;
639 mutex_unlock(&pinst->lock);
640}
641EXPORT_SYMBOL(padata_stop);
642
643#ifdef CONFIG_HOTPLUG_CPU
644static int padata_cpu_callback(struct notifier_block *nfb,
645 unsigned long action, void *hcpu)
646{
647 int err;
648 struct padata_instance *pinst;
649 int cpu = (unsigned long)hcpu;
650
651 pinst = container_of(nfb, struct padata_instance, cpu_notifier);
652
653 switch (action) {
654 case CPU_ONLINE:
655 case CPU_ONLINE_FROZEN:
656 if (!cpumask_test_cpu(cpu, pinst->cpumask))
657 break;
658 mutex_lock(&pinst->lock);
659 err = __padata_add_cpu(pinst, cpu);
660 mutex_unlock(&pinst->lock);
661 if (err)
662 return notifier_from_errno(err);
663 break;
664
665 case CPU_DOWN_PREPARE:
666 case CPU_DOWN_PREPARE_FROZEN:
667 if (!cpumask_test_cpu(cpu, pinst->cpumask))
668 break;
669 mutex_lock(&pinst->lock);
670 err = __padata_remove_cpu(pinst, cpu);
671 mutex_unlock(&pinst->lock);
672 if (err)
673 return notifier_from_errno(err);
674 break;
675
676 case CPU_UP_CANCELED:
677 case CPU_UP_CANCELED_FROZEN:
678 if (!cpumask_test_cpu(cpu, pinst->cpumask))
679 break;
680 mutex_lock(&pinst->lock);
681 __padata_remove_cpu(pinst, cpu);
682 mutex_unlock(&pinst->lock);
683
684 case CPU_DOWN_FAILED:
685 case CPU_DOWN_FAILED_FROZEN:
686 if (!cpumask_test_cpu(cpu, pinst->cpumask))
687 break;
688 mutex_lock(&pinst->lock);
689 __padata_add_cpu(pinst, cpu);
690 mutex_unlock(&pinst->lock);
691 }
692
693 return NOTIFY_OK;
694}
695#endif
696
697/**
698 * padata_alloc - allocate and initialize a padata instance
699 *
700 * @cpumask: cpumask that padata uses for parallelization
701 * @wq: workqueue to use for the allocated padata instance
702 */
703struct padata_instance *padata_alloc(const struct cpumask *cpumask,
704 struct workqueue_struct *wq)
705{
706 struct padata_instance *pinst;
707 struct parallel_data *pd;
708
709 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
710 if (!pinst)
711 goto err;
712
713 get_online_cpus();
714
715 pd = padata_alloc_pd(pinst, cpumask);
716 if (!pd)
717 goto err_free_inst;
718
719 if (!alloc_cpumask_var(&pinst->cpumask, GFP_KERNEL))
720 goto err_free_pd;
721
722 rcu_assign_pointer(pinst->pd, pd);
723
724 pinst->wq = wq;
725
726 cpumask_copy(pinst->cpumask, cpumask);
727
728 pinst->flags = 0;
729
730#ifdef CONFIG_HOTPLUG_CPU
731 pinst->cpu_notifier.notifier_call = padata_cpu_callback;
732 pinst->cpu_notifier.priority = 0;
733 register_hotcpu_notifier(&pinst->cpu_notifier);
734#endif
735
736 put_online_cpus();
737
738 mutex_init(&pinst->lock);
739
740 return pinst;
741
742err_free_pd:
743 padata_free_pd(pd);
744err_free_inst:
745 kfree(pinst);
746 put_online_cpus();
747err:
748 return NULL;
749}
750EXPORT_SYMBOL(padata_alloc);
751
752/**
753 * padata_free - free a padata instance
754 *
755 * @padata_inst: padata instance to free
756 */
757void padata_free(struct padata_instance *pinst)
758{
759 padata_stop(pinst);
760
761 synchronize_rcu();
762
763#ifdef CONFIG_HOTPLUG_CPU
764 unregister_hotcpu_notifier(&pinst->cpu_notifier);
765#endif
766 get_online_cpus();
767 padata_flush_queues(pinst->pd);
768 put_online_cpus();
769
770 padata_free_pd(pinst->pd);
771 free_cpumask_var(pinst->cpumask);
772 kfree(pinst);
773}
774EXPORT_SYMBOL(padata_free);