Loading...
1/*
2 * padata.c - generic interface to process data streams in parallel
3 *
4 * See Documentation/padata.txt for an api documentation.
5 *
6 * Copyright (C) 2008, 2009 secunet Security Networks AG
7 * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
8 *
9 * This program is free software; you can redistribute it and/or modify it
10 * under the terms and conditions of the GNU General Public License,
11 * version 2, as published by the Free Software Foundation.
12 *
13 * This program is distributed in the hope it will be useful, but WITHOUT
14 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
16 * more details.
17 *
18 * You should have received a copy of the GNU General Public License along with
19 * this program; if not, write to the Free Software Foundation, Inc.,
20 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
21 */
22
23#include <linux/export.h>
24#include <linux/cpumask.h>
25#include <linux/err.h>
26#include <linux/cpu.h>
27#include <linux/padata.h>
28#include <linux/mutex.h>
29#include <linux/sched.h>
30#include <linux/slab.h>
31#include <linux/sysfs.h>
32#include <linux/rcupdate.h>
33#include <linux/module.h>
34
35#define MAX_OBJ_NUM 1000
36
37static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
38{
39 int cpu, target_cpu;
40
41 target_cpu = cpumask_first(pd->cpumask.pcpu);
42 for (cpu = 0; cpu < cpu_index; cpu++)
43 target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
44
45 return target_cpu;
46}
47
48static int padata_cpu_hash(struct parallel_data *pd)
49{
50 unsigned int seq_nr;
51 int cpu_index;
52
53 /*
54 * Hash the sequence numbers to the cpus by taking
55 * seq_nr mod. number of cpus in use.
56 */
57
58 seq_nr = atomic_inc_return(&pd->seq_nr);
59 cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
60
61 return padata_index_to_cpu(pd, cpu_index);
62}
63
64static void padata_parallel_worker(struct work_struct *parallel_work)
65{
66 struct padata_parallel_queue *pqueue;
67 LIST_HEAD(local_list);
68
69 local_bh_disable();
70 pqueue = container_of(parallel_work,
71 struct padata_parallel_queue, work);
72
73 spin_lock(&pqueue->parallel.lock);
74 list_replace_init(&pqueue->parallel.list, &local_list);
75 spin_unlock(&pqueue->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 serial cpumask of padata(i.e. cpumask.cbcpu).
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_parallel_queue *queue;
108 struct parallel_data *pd;
109
110 rcu_read_lock_bh();
111
112 pd = rcu_dereference_bh(pinst->pd);
113
114 err = -EINVAL;
115 if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
116 goto out;
117
118 if (!cpumask_test_cpu(cb_cpu, pd->cpumask.cbcpu))
119 goto out;
120
121 err = -EBUSY;
122 if ((pinst->flags & PADATA_RESET))
123 goto out;
124
125 if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
126 goto out;
127
128 err = 0;
129 atomic_inc(&pd->refcnt);
130 padata->pd = pd;
131 padata->cb_cpu = cb_cpu;
132
133 target_cpu = padata_cpu_hash(pd);
134 queue = per_cpu_ptr(pd->pqueue, target_cpu);
135
136 spin_lock(&queue->parallel.lock);
137 list_add_tail(&padata->list, &queue->parallel.list);
138 spin_unlock(&queue->parallel.lock);
139
140 queue_work_on(target_cpu, pinst->wq, &queue->work);
141
142out:
143 rcu_read_unlock_bh();
144
145 return err;
146}
147EXPORT_SYMBOL(padata_do_parallel);
148
149/*
150 * padata_get_next - Get the next object that needs serialization.
151 *
152 * Return values are:
153 *
154 * A pointer to the control struct of the next object that needs
155 * serialization, if present in one of the percpu reorder queues.
156 *
157 * NULL, if all percpu reorder queues are empty.
158 *
159 * -EINPROGRESS, if the next object that needs serialization will
160 * be parallel processed by another cpu and is not yet present in
161 * the cpu's reorder queue.
162 *
163 * -ENODATA, if this cpu has to do the parallel processing for
164 * the next object.
165 */
166static struct padata_priv *padata_get_next(struct parallel_data *pd)
167{
168 int cpu, num_cpus;
169 unsigned int next_nr, next_index;
170 struct padata_parallel_queue *next_queue;
171 struct padata_priv *padata;
172 struct padata_list *reorder;
173
174 num_cpus = cpumask_weight(pd->cpumask.pcpu);
175
176 /*
177 * Calculate the percpu reorder queue and the sequence
178 * number of the next object.
179 */
180 next_nr = pd->processed;
181 next_index = next_nr % num_cpus;
182 cpu = padata_index_to_cpu(pd, next_index);
183 next_queue = per_cpu_ptr(pd->pqueue, cpu);
184
185 padata = NULL;
186
187 reorder = &next_queue->reorder;
188
189 spin_lock(&reorder->lock);
190 if (!list_empty(&reorder->list)) {
191 padata = list_entry(reorder->list.next,
192 struct padata_priv, list);
193
194 list_del_init(&padata->list);
195 atomic_dec(&pd->reorder_objects);
196
197 pd->processed++;
198
199 spin_unlock(&reorder->lock);
200 goto out;
201 }
202 spin_unlock(&reorder->lock);
203
204 if (__this_cpu_read(pd->pqueue->cpu_index) == next_queue->cpu_index) {
205 padata = ERR_PTR(-ENODATA);
206 goto out;
207 }
208
209 padata = ERR_PTR(-EINPROGRESS);
210out:
211 return padata;
212}
213
214static void padata_reorder(struct parallel_data *pd)
215{
216 int cb_cpu;
217 struct padata_priv *padata;
218 struct padata_serial_queue *squeue;
219 struct padata_instance *pinst = pd->pinst;
220
221 /*
222 * We need to ensure that only one cpu can work on dequeueing of
223 * the reorder queue the time. Calculating in which percpu reorder
224 * queue the next object will arrive takes some time. A spinlock
225 * would be highly contended. Also it is not clear in which order
226 * the objects arrive to the reorder queues. So a cpu could wait to
227 * get the lock just to notice that there is nothing to do at the
228 * moment. Therefore we use a trylock and let the holder of the lock
229 * care for all the objects enqueued during the holdtime of the lock.
230 */
231 if (!spin_trylock_bh(&pd->lock))
232 return;
233
234 while (1) {
235 padata = padata_get_next(pd);
236
237 /*
238 * All reorder queues are empty, or the next object that needs
239 * serialization is parallel processed by another cpu and is
240 * still on it's way to the cpu's reorder queue, nothing to
241 * do for now.
242 */
243 if (!padata || PTR_ERR(padata) == -EINPROGRESS)
244 break;
245
246 /*
247 * This cpu has to do the parallel processing of the next
248 * object. It's waiting in the cpu's parallelization queue,
249 * so exit immediately.
250 */
251 if (PTR_ERR(padata) == -ENODATA) {
252 del_timer(&pd->timer);
253 spin_unlock_bh(&pd->lock);
254 return;
255 }
256
257 cb_cpu = padata->cb_cpu;
258 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
259
260 spin_lock(&squeue->serial.lock);
261 list_add_tail(&padata->list, &squeue->serial.list);
262 spin_unlock(&squeue->serial.lock);
263
264 queue_work_on(cb_cpu, pinst->wq, &squeue->work);
265 }
266
267 spin_unlock_bh(&pd->lock);
268
269 /*
270 * The next object that needs serialization might have arrived to
271 * the reorder queues in the meantime, we will be called again
272 * from the timer function if no one else cares for it.
273 */
274 if (atomic_read(&pd->reorder_objects)
275 && !(pinst->flags & PADATA_RESET))
276 mod_timer(&pd->timer, jiffies + HZ);
277 else
278 del_timer(&pd->timer);
279
280 return;
281}
282
283static void padata_reorder_timer(unsigned long arg)
284{
285 struct parallel_data *pd = (struct parallel_data *)arg;
286
287 padata_reorder(pd);
288}
289
290static void padata_serial_worker(struct work_struct *serial_work)
291{
292 struct padata_serial_queue *squeue;
293 struct parallel_data *pd;
294 LIST_HEAD(local_list);
295
296 local_bh_disable();
297 squeue = container_of(serial_work, struct padata_serial_queue, work);
298 pd = squeue->pd;
299
300 spin_lock(&squeue->serial.lock);
301 list_replace_init(&squeue->serial.list, &local_list);
302 spin_unlock(&squeue->serial.lock);
303
304 while (!list_empty(&local_list)) {
305 struct padata_priv *padata;
306
307 padata = list_entry(local_list.next,
308 struct padata_priv, list);
309
310 list_del_init(&padata->list);
311
312 padata->serial(padata);
313 atomic_dec(&pd->refcnt);
314 }
315 local_bh_enable();
316}
317
318/**
319 * padata_do_serial - padata serialization function
320 *
321 * @padata: object to be serialized.
322 *
323 * padata_do_serial must be called for every parallelized object.
324 * The serialization callback function will run with BHs off.
325 */
326void padata_do_serial(struct padata_priv *padata)
327{
328 int cpu;
329 struct padata_parallel_queue *pqueue;
330 struct parallel_data *pd;
331
332 pd = padata->pd;
333
334 cpu = get_cpu();
335 pqueue = per_cpu_ptr(pd->pqueue, cpu);
336
337 spin_lock(&pqueue->reorder.lock);
338 atomic_inc(&pd->reorder_objects);
339 list_add_tail(&padata->list, &pqueue->reorder.list);
340 spin_unlock(&pqueue->reorder.lock);
341
342 put_cpu();
343
344 padata_reorder(pd);
345}
346EXPORT_SYMBOL(padata_do_serial);
347
348static int padata_setup_cpumasks(struct parallel_data *pd,
349 const struct cpumask *pcpumask,
350 const struct cpumask *cbcpumask)
351{
352 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
353 return -ENOMEM;
354
355 cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_online_mask);
356 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL)) {
357 free_cpumask_var(pd->cpumask.cbcpu);
358 return -ENOMEM;
359 }
360
361 cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_online_mask);
362 return 0;
363}
364
365static void __padata_list_init(struct padata_list *pd_list)
366{
367 INIT_LIST_HEAD(&pd_list->list);
368 spin_lock_init(&pd_list->lock);
369}
370
371/* Initialize all percpu queues used by serial workers */
372static void padata_init_squeues(struct parallel_data *pd)
373{
374 int cpu;
375 struct padata_serial_queue *squeue;
376
377 for_each_cpu(cpu, pd->cpumask.cbcpu) {
378 squeue = per_cpu_ptr(pd->squeue, cpu);
379 squeue->pd = pd;
380 __padata_list_init(&squeue->serial);
381 INIT_WORK(&squeue->work, padata_serial_worker);
382 }
383}
384
385/* Initialize all percpu queues used by parallel workers */
386static void padata_init_pqueues(struct parallel_data *pd)
387{
388 int cpu_index, cpu;
389 struct padata_parallel_queue *pqueue;
390
391 cpu_index = 0;
392 for_each_cpu(cpu, pd->cpumask.pcpu) {
393 pqueue = per_cpu_ptr(pd->pqueue, cpu);
394 pqueue->pd = pd;
395 pqueue->cpu_index = cpu_index;
396 cpu_index++;
397
398 __padata_list_init(&pqueue->reorder);
399 __padata_list_init(&pqueue->parallel);
400 INIT_WORK(&pqueue->work, padata_parallel_worker);
401 atomic_set(&pqueue->num_obj, 0);
402 }
403}
404
405/* Allocate and initialize the internal cpumask dependend resources. */
406static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
407 const struct cpumask *pcpumask,
408 const struct cpumask *cbcpumask)
409{
410 struct parallel_data *pd;
411
412 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
413 if (!pd)
414 goto err;
415
416 pd->pqueue = alloc_percpu(struct padata_parallel_queue);
417 if (!pd->pqueue)
418 goto err_free_pd;
419
420 pd->squeue = alloc_percpu(struct padata_serial_queue);
421 if (!pd->squeue)
422 goto err_free_pqueue;
423 if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
424 goto err_free_squeue;
425
426 padata_init_pqueues(pd);
427 padata_init_squeues(pd);
428 setup_timer(&pd->timer, padata_reorder_timer, (unsigned long)pd);
429 atomic_set(&pd->seq_nr, -1);
430 atomic_set(&pd->reorder_objects, 0);
431 atomic_set(&pd->refcnt, 0);
432 pd->pinst = pinst;
433 spin_lock_init(&pd->lock);
434
435 return pd;
436
437err_free_squeue:
438 free_percpu(pd->squeue);
439err_free_pqueue:
440 free_percpu(pd->pqueue);
441err_free_pd:
442 kfree(pd);
443err:
444 return NULL;
445}
446
447static void padata_free_pd(struct parallel_data *pd)
448{
449 free_cpumask_var(pd->cpumask.pcpu);
450 free_cpumask_var(pd->cpumask.cbcpu);
451 free_percpu(pd->pqueue);
452 free_percpu(pd->squeue);
453 kfree(pd);
454}
455
456/* Flush all objects out of the padata queues. */
457static void padata_flush_queues(struct parallel_data *pd)
458{
459 int cpu;
460 struct padata_parallel_queue *pqueue;
461 struct padata_serial_queue *squeue;
462
463 for_each_cpu(cpu, pd->cpumask.pcpu) {
464 pqueue = per_cpu_ptr(pd->pqueue, cpu);
465 flush_work(&pqueue->work);
466 }
467
468 del_timer_sync(&pd->timer);
469
470 if (atomic_read(&pd->reorder_objects))
471 padata_reorder(pd);
472
473 for_each_cpu(cpu, pd->cpumask.cbcpu) {
474 squeue = per_cpu_ptr(pd->squeue, cpu);
475 flush_work(&squeue->work);
476 }
477
478 BUG_ON(atomic_read(&pd->refcnt) != 0);
479}
480
481static void __padata_start(struct padata_instance *pinst)
482{
483 pinst->flags |= PADATA_INIT;
484}
485
486static void __padata_stop(struct padata_instance *pinst)
487{
488 if (!(pinst->flags & PADATA_INIT))
489 return;
490
491 pinst->flags &= ~PADATA_INIT;
492
493 synchronize_rcu();
494
495 get_online_cpus();
496 padata_flush_queues(pinst->pd);
497 put_online_cpus();
498}
499
500/* Replace the internal control structure with a new one. */
501static void padata_replace(struct padata_instance *pinst,
502 struct parallel_data *pd_new)
503{
504 struct parallel_data *pd_old = pinst->pd;
505 int notification_mask = 0;
506
507 pinst->flags |= PADATA_RESET;
508
509 rcu_assign_pointer(pinst->pd, pd_new);
510
511 synchronize_rcu();
512
513 if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
514 notification_mask |= PADATA_CPU_PARALLEL;
515 if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
516 notification_mask |= PADATA_CPU_SERIAL;
517
518 padata_flush_queues(pd_old);
519 padata_free_pd(pd_old);
520
521 if (notification_mask)
522 blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
523 notification_mask,
524 &pd_new->cpumask);
525
526 pinst->flags &= ~PADATA_RESET;
527}
528
529/**
530 * padata_register_cpumask_notifier - Registers a notifier that will be called
531 * if either pcpu or cbcpu or both cpumasks change.
532 *
533 * @pinst: A poineter to padata instance
534 * @nblock: A pointer to notifier block.
535 */
536int padata_register_cpumask_notifier(struct padata_instance *pinst,
537 struct notifier_block *nblock)
538{
539 return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
540 nblock);
541}
542EXPORT_SYMBOL(padata_register_cpumask_notifier);
543
544/**
545 * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
546 * registered earlier using padata_register_cpumask_notifier
547 *
548 * @pinst: A pointer to data instance.
549 * @nlock: A pointer to notifier block.
550 */
551int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
552 struct notifier_block *nblock)
553{
554 return blocking_notifier_chain_unregister(
555 &pinst->cpumask_change_notifier,
556 nblock);
557}
558EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
559
560
561/* If cpumask contains no active cpu, we mark the instance as invalid. */
562static bool padata_validate_cpumask(struct padata_instance *pinst,
563 const struct cpumask *cpumask)
564{
565 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
566 pinst->flags |= PADATA_INVALID;
567 return false;
568 }
569
570 pinst->flags &= ~PADATA_INVALID;
571 return true;
572}
573
574static int __padata_set_cpumasks(struct padata_instance *pinst,
575 cpumask_var_t pcpumask,
576 cpumask_var_t cbcpumask)
577{
578 int valid;
579 struct parallel_data *pd;
580
581 valid = padata_validate_cpumask(pinst, pcpumask);
582 if (!valid) {
583 __padata_stop(pinst);
584 goto out_replace;
585 }
586
587 valid = padata_validate_cpumask(pinst, cbcpumask);
588 if (!valid)
589 __padata_stop(pinst);
590
591out_replace:
592 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
593 if (!pd)
594 return -ENOMEM;
595
596 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
597 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
598
599 padata_replace(pinst, pd);
600
601 if (valid)
602 __padata_start(pinst);
603
604 return 0;
605}
606
607/**
608 * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
609 * equivalent to @cpumask.
610 *
611 * @pinst: padata instance
612 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
613 * to parallel and serial cpumasks respectively.
614 * @cpumask: the cpumask to use
615 */
616int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
617 cpumask_var_t cpumask)
618{
619 struct cpumask *serial_mask, *parallel_mask;
620 int err = -EINVAL;
621
622 mutex_lock(&pinst->lock);
623 get_online_cpus();
624
625 switch (cpumask_type) {
626 case PADATA_CPU_PARALLEL:
627 serial_mask = pinst->cpumask.cbcpu;
628 parallel_mask = cpumask;
629 break;
630 case PADATA_CPU_SERIAL:
631 parallel_mask = pinst->cpumask.pcpu;
632 serial_mask = cpumask;
633 break;
634 default:
635 goto out;
636 }
637
638 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
639
640out:
641 put_online_cpus();
642 mutex_unlock(&pinst->lock);
643
644 return err;
645}
646EXPORT_SYMBOL(padata_set_cpumask);
647
648/**
649 * padata_start - start the parallel processing
650 *
651 * @pinst: padata instance to start
652 */
653int padata_start(struct padata_instance *pinst)
654{
655 int err = 0;
656
657 mutex_lock(&pinst->lock);
658
659 if (pinst->flags & PADATA_INVALID)
660 err = -EINVAL;
661
662 __padata_start(pinst);
663
664 mutex_unlock(&pinst->lock);
665
666 return err;
667}
668EXPORT_SYMBOL(padata_start);
669
670/**
671 * padata_stop - stop the parallel processing
672 *
673 * @pinst: padata instance to stop
674 */
675void padata_stop(struct padata_instance *pinst)
676{
677 mutex_lock(&pinst->lock);
678 __padata_stop(pinst);
679 mutex_unlock(&pinst->lock);
680}
681EXPORT_SYMBOL(padata_stop);
682
683#ifdef CONFIG_HOTPLUG_CPU
684
685static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
686{
687 struct parallel_data *pd;
688
689 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
690 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
691 pinst->cpumask.cbcpu);
692 if (!pd)
693 return -ENOMEM;
694
695 padata_replace(pinst, pd);
696
697 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
698 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
699 __padata_start(pinst);
700 }
701
702 return 0;
703}
704
705static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
706{
707 struct parallel_data *pd = NULL;
708
709 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
710
711 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
712 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
713 __padata_stop(pinst);
714
715 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
716 pinst->cpumask.cbcpu);
717 if (!pd)
718 return -ENOMEM;
719
720 padata_replace(pinst, pd);
721
722 cpumask_clear_cpu(cpu, pd->cpumask.cbcpu);
723 cpumask_clear_cpu(cpu, pd->cpumask.pcpu);
724 }
725
726 return 0;
727}
728
729 /**
730 * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
731 * padata cpumasks.
732 *
733 * @pinst: padata instance
734 * @cpu: cpu to remove
735 * @mask: bitmask specifying from which cpumask @cpu should be removed
736 * The @mask may be any combination of the following flags:
737 * PADATA_CPU_SERIAL - serial cpumask
738 * PADATA_CPU_PARALLEL - parallel cpumask
739 */
740int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
741{
742 int err;
743
744 if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
745 return -EINVAL;
746
747 mutex_lock(&pinst->lock);
748
749 get_online_cpus();
750 if (mask & PADATA_CPU_SERIAL)
751 cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
752 if (mask & PADATA_CPU_PARALLEL)
753 cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
754
755 err = __padata_remove_cpu(pinst, cpu);
756 put_online_cpus();
757
758 mutex_unlock(&pinst->lock);
759
760 return err;
761}
762EXPORT_SYMBOL(padata_remove_cpu);
763
764static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
765{
766 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
767 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
768}
769
770static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
771{
772 struct padata_instance *pinst;
773 int ret;
774
775 pinst = hlist_entry_safe(node, struct padata_instance, node);
776 if (!pinst_has_cpu(pinst, cpu))
777 return 0;
778
779 mutex_lock(&pinst->lock);
780 ret = __padata_add_cpu(pinst, cpu);
781 mutex_unlock(&pinst->lock);
782 return ret;
783}
784
785static int padata_cpu_prep_down(unsigned int cpu, struct hlist_node *node)
786{
787 struct padata_instance *pinst;
788 int ret;
789
790 pinst = hlist_entry_safe(node, struct padata_instance, node);
791 if (!pinst_has_cpu(pinst, cpu))
792 return 0;
793
794 mutex_lock(&pinst->lock);
795 ret = __padata_remove_cpu(pinst, cpu);
796 mutex_unlock(&pinst->lock);
797 return ret;
798}
799
800static enum cpuhp_state hp_online;
801#endif
802
803static void __padata_free(struct padata_instance *pinst)
804{
805#ifdef CONFIG_HOTPLUG_CPU
806 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->node);
807#endif
808
809 padata_stop(pinst);
810 padata_free_pd(pinst->pd);
811 free_cpumask_var(pinst->cpumask.pcpu);
812 free_cpumask_var(pinst->cpumask.cbcpu);
813 kfree(pinst);
814}
815
816#define kobj2pinst(_kobj) \
817 container_of(_kobj, struct padata_instance, kobj)
818#define attr2pentry(_attr) \
819 container_of(_attr, struct padata_sysfs_entry, attr)
820
821static void padata_sysfs_release(struct kobject *kobj)
822{
823 struct padata_instance *pinst = kobj2pinst(kobj);
824 __padata_free(pinst);
825}
826
827struct padata_sysfs_entry {
828 struct attribute attr;
829 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
830 ssize_t (*store)(struct padata_instance *, struct attribute *,
831 const char *, size_t);
832};
833
834static ssize_t show_cpumask(struct padata_instance *pinst,
835 struct attribute *attr, char *buf)
836{
837 struct cpumask *cpumask;
838 ssize_t len;
839
840 mutex_lock(&pinst->lock);
841 if (!strcmp(attr->name, "serial_cpumask"))
842 cpumask = pinst->cpumask.cbcpu;
843 else
844 cpumask = pinst->cpumask.pcpu;
845
846 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
847 nr_cpu_ids, cpumask_bits(cpumask));
848 mutex_unlock(&pinst->lock);
849 return len < PAGE_SIZE ? len : -EINVAL;
850}
851
852static ssize_t store_cpumask(struct padata_instance *pinst,
853 struct attribute *attr,
854 const char *buf, size_t count)
855{
856 cpumask_var_t new_cpumask;
857 ssize_t ret;
858 int mask_type;
859
860 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
861 return -ENOMEM;
862
863 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
864 nr_cpumask_bits);
865 if (ret < 0)
866 goto out;
867
868 mask_type = !strcmp(attr->name, "serial_cpumask") ?
869 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
870 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
871 if (!ret)
872 ret = count;
873
874out:
875 free_cpumask_var(new_cpumask);
876 return ret;
877}
878
879#define PADATA_ATTR_RW(_name, _show_name, _store_name) \
880 static struct padata_sysfs_entry _name##_attr = \
881 __ATTR(_name, 0644, _show_name, _store_name)
882#define PADATA_ATTR_RO(_name, _show_name) \
883 static struct padata_sysfs_entry _name##_attr = \
884 __ATTR(_name, 0400, _show_name, NULL)
885
886PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
887PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
888
889/*
890 * Padata sysfs provides the following objects:
891 * serial_cpumask [RW] - cpumask for serial workers
892 * parallel_cpumask [RW] - cpumask for parallel workers
893 */
894static struct attribute *padata_default_attrs[] = {
895 &serial_cpumask_attr.attr,
896 ¶llel_cpumask_attr.attr,
897 NULL,
898};
899
900static ssize_t padata_sysfs_show(struct kobject *kobj,
901 struct attribute *attr, char *buf)
902{
903 struct padata_instance *pinst;
904 struct padata_sysfs_entry *pentry;
905 ssize_t ret = -EIO;
906
907 pinst = kobj2pinst(kobj);
908 pentry = attr2pentry(attr);
909 if (pentry->show)
910 ret = pentry->show(pinst, attr, buf);
911
912 return ret;
913}
914
915static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
916 const char *buf, size_t count)
917{
918 struct padata_instance *pinst;
919 struct padata_sysfs_entry *pentry;
920 ssize_t ret = -EIO;
921
922 pinst = kobj2pinst(kobj);
923 pentry = attr2pentry(attr);
924 if (pentry->show)
925 ret = pentry->store(pinst, attr, buf, count);
926
927 return ret;
928}
929
930static const struct sysfs_ops padata_sysfs_ops = {
931 .show = padata_sysfs_show,
932 .store = padata_sysfs_store,
933};
934
935static struct kobj_type padata_attr_type = {
936 .sysfs_ops = &padata_sysfs_ops,
937 .default_attrs = padata_default_attrs,
938 .release = padata_sysfs_release,
939};
940
941/**
942 * padata_alloc_possible - Allocate and initialize padata instance.
943 * Use the cpu_possible_mask for serial and
944 * parallel workers.
945 *
946 * @wq: workqueue to use for the allocated padata instance
947 */
948struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq)
949{
950 return padata_alloc(wq, cpu_possible_mask, cpu_possible_mask);
951}
952EXPORT_SYMBOL(padata_alloc_possible);
953
954/**
955 * padata_alloc - allocate and initialize a padata instance and specify
956 * cpumasks for serial and parallel workers.
957 *
958 * @wq: workqueue to use for the allocated padata instance
959 * @pcpumask: cpumask that will be used for padata parallelization
960 * @cbcpumask: cpumask that will be used for padata serialization
961 */
962struct padata_instance *padata_alloc(struct workqueue_struct *wq,
963 const struct cpumask *pcpumask,
964 const struct cpumask *cbcpumask)
965{
966 struct padata_instance *pinst;
967 struct parallel_data *pd = NULL;
968
969 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
970 if (!pinst)
971 goto err;
972
973 get_online_cpus();
974 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
975 goto err_free_inst;
976 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
977 free_cpumask_var(pinst->cpumask.pcpu);
978 goto err_free_inst;
979 }
980 if (!padata_validate_cpumask(pinst, pcpumask) ||
981 !padata_validate_cpumask(pinst, cbcpumask))
982 goto err_free_masks;
983
984 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
985 if (!pd)
986 goto err_free_masks;
987
988 rcu_assign_pointer(pinst->pd, pd);
989
990 pinst->wq = wq;
991
992 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
993 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
994
995 pinst->flags = 0;
996
997 put_online_cpus();
998
999 BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
1000 kobject_init(&pinst->kobj, &padata_attr_type);
1001 mutex_init(&pinst->lock);
1002
1003#ifdef CONFIG_HOTPLUG_CPU
1004 cpuhp_state_add_instance_nocalls(hp_online, &pinst->node);
1005#endif
1006 return pinst;
1007
1008err_free_masks:
1009 free_cpumask_var(pinst->cpumask.pcpu);
1010 free_cpumask_var(pinst->cpumask.cbcpu);
1011err_free_inst:
1012 kfree(pinst);
1013 put_online_cpus();
1014err:
1015 return NULL;
1016}
1017
1018/**
1019 * padata_free - free a padata instance
1020 *
1021 * @padata_inst: padata instance to free
1022 */
1023void padata_free(struct padata_instance *pinst)
1024{
1025 kobject_put(&pinst->kobj);
1026}
1027EXPORT_SYMBOL(padata_free);
1028
1029#ifdef CONFIG_HOTPLUG_CPU
1030
1031static __init int padata_driver_init(void)
1032{
1033 int ret;
1034
1035 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1036 padata_cpu_online,
1037 padata_cpu_prep_down);
1038 if (ret < 0)
1039 return ret;
1040 hp_online = ret;
1041 return 0;
1042}
1043module_init(padata_driver_init);
1044
1045static __exit void padata_driver_exit(void)
1046{
1047 cpuhp_remove_multi_state(hp_online);
1048}
1049module_exit(padata_driver_exit);
1050#endif
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 <steffen.klassert@secunet.com>
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/sysfs.h>
30#include <linux/rcupdate.h>
31
32#define MAX_SEQ_NR (INT_MAX - NR_CPUS)
33#define MAX_OBJ_NUM 1000
34
35static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
36{
37 int cpu, target_cpu;
38
39 target_cpu = cpumask_first(pd->cpumask.pcpu);
40 for (cpu = 0; cpu < cpu_index; cpu++)
41 target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
42
43 return target_cpu;
44}
45
46static int padata_cpu_hash(struct padata_priv *padata)
47{
48 int cpu_index;
49 struct parallel_data *pd;
50
51 pd = padata->pd;
52
53 /*
54 * Hash the sequence numbers to the cpus by taking
55 * seq_nr mod. number of cpus in use.
56 */
57 cpu_index = padata->seq_nr % cpumask_weight(pd->cpumask.pcpu);
58
59 return padata_index_to_cpu(pd, cpu_index);
60}
61
62static void padata_parallel_worker(struct work_struct *parallel_work)
63{
64 struct padata_parallel_queue *pqueue;
65 struct parallel_data *pd;
66 struct padata_instance *pinst;
67 LIST_HEAD(local_list);
68
69 local_bh_disable();
70 pqueue = container_of(parallel_work,
71 struct padata_parallel_queue, work);
72 pd = pqueue->pd;
73 pinst = pd->pinst;
74
75 spin_lock(&pqueue->parallel.lock);
76 list_replace_init(&pqueue->parallel.list, &local_list);
77 spin_unlock(&pqueue->parallel.lock);
78
79 while (!list_empty(&local_list)) {
80 struct padata_priv *padata;
81
82 padata = list_entry(local_list.next,
83 struct padata_priv, list);
84
85 list_del_init(&padata->list);
86
87 padata->parallel(padata);
88 }
89
90 local_bh_enable();
91}
92
93/**
94 * padata_do_parallel - padata parallelization function
95 *
96 * @pinst: padata instance
97 * @padata: object to be parallelized
98 * @cb_cpu: cpu the serialization callback function will run on,
99 * must be in the serial cpumask of padata(i.e. cpumask.cbcpu).
100 *
101 * The parallelization callback function will run with BHs off.
102 * Note: Every object which is parallelized by padata_do_parallel
103 * must be seen by padata_do_serial.
104 */
105int padata_do_parallel(struct padata_instance *pinst,
106 struct padata_priv *padata, int cb_cpu)
107{
108 int target_cpu, err;
109 struct padata_parallel_queue *queue;
110 struct parallel_data *pd;
111
112 rcu_read_lock_bh();
113
114 pd = rcu_dereference(pinst->pd);
115
116 err = -EINVAL;
117 if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
118 goto out;
119
120 if (!cpumask_test_cpu(cb_cpu, pd->cpumask.cbcpu))
121 goto out;
122
123 err = -EBUSY;
124 if ((pinst->flags & PADATA_RESET))
125 goto out;
126
127 if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
128 goto out;
129
130 err = 0;
131 atomic_inc(&pd->refcnt);
132 padata->pd = pd;
133 padata->cb_cpu = cb_cpu;
134
135 if (unlikely(atomic_read(&pd->seq_nr) == pd->max_seq_nr))
136 atomic_set(&pd->seq_nr, -1);
137
138 padata->seq_nr = atomic_inc_return(&pd->seq_nr);
139
140 target_cpu = padata_cpu_hash(padata);
141 queue = per_cpu_ptr(pd->pqueue, target_cpu);
142
143 spin_lock(&queue->parallel.lock);
144 list_add_tail(&padata->list, &queue->parallel.list);
145 spin_unlock(&queue->parallel.lock);
146
147 queue_work_on(target_cpu, pinst->wq, &queue->work);
148
149out:
150 rcu_read_unlock_bh();
151
152 return err;
153}
154EXPORT_SYMBOL(padata_do_parallel);
155
156/*
157 * padata_get_next - Get the next object that needs serialization.
158 *
159 * Return values are:
160 *
161 * A pointer to the control struct of the next object that needs
162 * serialization, if present in one of the percpu reorder queues.
163 *
164 * NULL, if all percpu reorder queues are empty.
165 *
166 * -EINPROGRESS, if the next object that needs serialization will
167 * be parallel processed by another cpu and is not yet present in
168 * the cpu's reorder queue.
169 *
170 * -ENODATA, if this cpu has to do the parallel processing for
171 * the next object.
172 */
173static struct padata_priv *padata_get_next(struct parallel_data *pd)
174{
175 int cpu, num_cpus;
176 int next_nr, next_index;
177 struct padata_parallel_queue *queue, *next_queue;
178 struct padata_priv *padata;
179 struct padata_list *reorder;
180
181 num_cpus = cpumask_weight(pd->cpumask.pcpu);
182
183 /*
184 * Calculate the percpu reorder queue and the sequence
185 * number of the next object.
186 */
187 next_nr = pd->processed;
188 next_index = next_nr % num_cpus;
189 cpu = padata_index_to_cpu(pd, next_index);
190 next_queue = per_cpu_ptr(pd->pqueue, cpu);
191
192 if (unlikely(next_nr > pd->max_seq_nr)) {
193 next_nr = next_nr - pd->max_seq_nr - 1;
194 next_index = next_nr % num_cpus;
195 cpu = padata_index_to_cpu(pd, next_index);
196 next_queue = per_cpu_ptr(pd->pqueue, cpu);
197 pd->processed = 0;
198 }
199
200 padata = NULL;
201
202 reorder = &next_queue->reorder;
203
204 if (!list_empty(&reorder->list)) {
205 padata = list_entry(reorder->list.next,
206 struct padata_priv, list);
207
208 BUG_ON(next_nr != padata->seq_nr);
209
210 spin_lock(&reorder->lock);
211 list_del_init(&padata->list);
212 atomic_dec(&pd->reorder_objects);
213 spin_unlock(&reorder->lock);
214
215 pd->processed++;
216
217 goto out;
218 }
219
220 queue = per_cpu_ptr(pd->pqueue, smp_processor_id());
221 if (queue->cpu_index == next_queue->cpu_index) {
222 padata = ERR_PTR(-ENODATA);
223 goto out;
224 }
225
226 padata = ERR_PTR(-EINPROGRESS);
227out:
228 return padata;
229}
230
231static void padata_reorder(struct parallel_data *pd)
232{
233 struct padata_priv *padata;
234 struct padata_serial_queue *squeue;
235 struct padata_instance *pinst = pd->pinst;
236
237 /*
238 * We need to ensure that only one cpu can work on dequeueing of
239 * the reorder queue the time. Calculating in which percpu reorder
240 * queue the next object will arrive takes some time. A spinlock
241 * would be highly contended. Also it is not clear in which order
242 * the objects arrive to the reorder queues. So a cpu could wait to
243 * get the lock just to notice that there is nothing to do at the
244 * moment. Therefore we use a trylock and let the holder of the lock
245 * care for all the objects enqueued during the holdtime of the lock.
246 */
247 if (!spin_trylock_bh(&pd->lock))
248 return;
249
250 while (1) {
251 padata = padata_get_next(pd);
252
253 /*
254 * All reorder queues are empty, or the next object that needs
255 * serialization is parallel processed by another cpu and is
256 * still on it's way to the cpu's reorder queue, nothing to
257 * do for now.
258 */
259 if (!padata || PTR_ERR(padata) == -EINPROGRESS)
260 break;
261
262 /*
263 * This cpu has to do the parallel processing of the next
264 * object. It's waiting in the cpu's parallelization queue,
265 * so exit immediately.
266 */
267 if (PTR_ERR(padata) == -ENODATA) {
268 del_timer(&pd->timer);
269 spin_unlock_bh(&pd->lock);
270 return;
271 }
272
273 squeue = per_cpu_ptr(pd->squeue, padata->cb_cpu);
274
275 spin_lock(&squeue->serial.lock);
276 list_add_tail(&padata->list, &squeue->serial.list);
277 spin_unlock(&squeue->serial.lock);
278
279 queue_work_on(padata->cb_cpu, pinst->wq, &squeue->work);
280 }
281
282 spin_unlock_bh(&pd->lock);
283
284 /*
285 * The next object that needs serialization might have arrived to
286 * the reorder queues in the meantime, we will be called again
287 * from the timer function if no one else cares for it.
288 */
289 if (atomic_read(&pd->reorder_objects)
290 && !(pinst->flags & PADATA_RESET))
291 mod_timer(&pd->timer, jiffies + HZ);
292 else
293 del_timer(&pd->timer);
294
295 return;
296}
297
298static void padata_reorder_timer(unsigned long arg)
299{
300 struct parallel_data *pd = (struct parallel_data *)arg;
301
302 padata_reorder(pd);
303}
304
305static void padata_serial_worker(struct work_struct *serial_work)
306{
307 struct padata_serial_queue *squeue;
308 struct parallel_data *pd;
309 LIST_HEAD(local_list);
310
311 local_bh_disable();
312 squeue = container_of(serial_work, struct padata_serial_queue, work);
313 pd = squeue->pd;
314
315 spin_lock(&squeue->serial.lock);
316 list_replace_init(&squeue->serial.list, &local_list);
317 spin_unlock(&squeue->serial.lock);
318
319 while (!list_empty(&local_list)) {
320 struct padata_priv *padata;
321
322 padata = list_entry(local_list.next,
323 struct padata_priv, list);
324
325 list_del_init(&padata->list);
326
327 padata->serial(padata);
328 atomic_dec(&pd->refcnt);
329 }
330 local_bh_enable();
331}
332
333/**
334 * padata_do_serial - padata serialization function
335 *
336 * @padata: object to be serialized.
337 *
338 * padata_do_serial must be called for every parallelized object.
339 * The serialization callback function will run with BHs off.
340 */
341void padata_do_serial(struct padata_priv *padata)
342{
343 int cpu;
344 struct padata_parallel_queue *pqueue;
345 struct parallel_data *pd;
346
347 pd = padata->pd;
348
349 cpu = get_cpu();
350 pqueue = per_cpu_ptr(pd->pqueue, cpu);
351
352 spin_lock(&pqueue->reorder.lock);
353 atomic_inc(&pd->reorder_objects);
354 list_add_tail(&padata->list, &pqueue->reorder.list);
355 spin_unlock(&pqueue->reorder.lock);
356
357 put_cpu();
358
359 padata_reorder(pd);
360}
361EXPORT_SYMBOL(padata_do_serial);
362
363static int padata_setup_cpumasks(struct parallel_data *pd,
364 const struct cpumask *pcpumask,
365 const struct cpumask *cbcpumask)
366{
367 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
368 return -ENOMEM;
369
370 cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_active_mask);
371 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL)) {
372 free_cpumask_var(pd->cpumask.cbcpu);
373 return -ENOMEM;
374 }
375
376 cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_active_mask);
377 return 0;
378}
379
380static void __padata_list_init(struct padata_list *pd_list)
381{
382 INIT_LIST_HEAD(&pd_list->list);
383 spin_lock_init(&pd_list->lock);
384}
385
386/* Initialize all percpu queues used by serial workers */
387static void padata_init_squeues(struct parallel_data *pd)
388{
389 int cpu;
390 struct padata_serial_queue *squeue;
391
392 for_each_cpu(cpu, pd->cpumask.cbcpu) {
393 squeue = per_cpu_ptr(pd->squeue, cpu);
394 squeue->pd = pd;
395 __padata_list_init(&squeue->serial);
396 INIT_WORK(&squeue->work, padata_serial_worker);
397 }
398}
399
400/* Initialize all percpu queues used by parallel workers */
401static void padata_init_pqueues(struct parallel_data *pd)
402{
403 int cpu_index, num_cpus, cpu;
404 struct padata_parallel_queue *pqueue;
405
406 cpu_index = 0;
407 for_each_cpu(cpu, pd->cpumask.pcpu) {
408 pqueue = per_cpu_ptr(pd->pqueue, cpu);
409 pqueue->pd = pd;
410 pqueue->cpu_index = cpu_index;
411 cpu_index++;
412
413 __padata_list_init(&pqueue->reorder);
414 __padata_list_init(&pqueue->parallel);
415 INIT_WORK(&pqueue->work, padata_parallel_worker);
416 atomic_set(&pqueue->num_obj, 0);
417 }
418
419 num_cpus = cpumask_weight(pd->cpumask.pcpu);
420 pd->max_seq_nr = num_cpus ? (MAX_SEQ_NR / num_cpus) * num_cpus - 1 : 0;
421}
422
423/* Allocate and initialize the internal cpumask dependend resources. */
424static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
425 const struct cpumask *pcpumask,
426 const struct cpumask *cbcpumask)
427{
428 struct parallel_data *pd;
429
430 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
431 if (!pd)
432 goto err;
433
434 pd->pqueue = alloc_percpu(struct padata_parallel_queue);
435 if (!pd->pqueue)
436 goto err_free_pd;
437
438 pd->squeue = alloc_percpu(struct padata_serial_queue);
439 if (!pd->squeue)
440 goto err_free_pqueue;
441 if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
442 goto err_free_squeue;
443
444 padata_init_pqueues(pd);
445 padata_init_squeues(pd);
446 setup_timer(&pd->timer, padata_reorder_timer, (unsigned long)pd);
447 atomic_set(&pd->seq_nr, -1);
448 atomic_set(&pd->reorder_objects, 0);
449 atomic_set(&pd->refcnt, 0);
450 pd->pinst = pinst;
451 spin_lock_init(&pd->lock);
452
453 return pd;
454
455err_free_squeue:
456 free_percpu(pd->squeue);
457err_free_pqueue:
458 free_percpu(pd->pqueue);
459err_free_pd:
460 kfree(pd);
461err:
462 return NULL;
463}
464
465static void padata_free_pd(struct parallel_data *pd)
466{
467 free_cpumask_var(pd->cpumask.pcpu);
468 free_cpumask_var(pd->cpumask.cbcpu);
469 free_percpu(pd->pqueue);
470 free_percpu(pd->squeue);
471 kfree(pd);
472}
473
474/* Flush all objects out of the padata queues. */
475static void padata_flush_queues(struct parallel_data *pd)
476{
477 int cpu;
478 struct padata_parallel_queue *pqueue;
479 struct padata_serial_queue *squeue;
480
481 for_each_cpu(cpu, pd->cpumask.pcpu) {
482 pqueue = per_cpu_ptr(pd->pqueue, cpu);
483 flush_work(&pqueue->work);
484 }
485
486 del_timer_sync(&pd->timer);
487
488 if (atomic_read(&pd->reorder_objects))
489 padata_reorder(pd);
490
491 for_each_cpu(cpu, pd->cpumask.cbcpu) {
492 squeue = per_cpu_ptr(pd->squeue, cpu);
493 flush_work(&squeue->work);
494 }
495
496 BUG_ON(atomic_read(&pd->refcnt) != 0);
497}
498
499static void __padata_start(struct padata_instance *pinst)
500{
501 pinst->flags |= PADATA_INIT;
502}
503
504static void __padata_stop(struct padata_instance *pinst)
505{
506 if (!(pinst->flags & PADATA_INIT))
507 return;
508
509 pinst->flags &= ~PADATA_INIT;
510
511 synchronize_rcu();
512
513 get_online_cpus();
514 padata_flush_queues(pinst->pd);
515 put_online_cpus();
516}
517
518/* Replace the internal control structure with a new one. */
519static void padata_replace(struct padata_instance *pinst,
520 struct parallel_data *pd_new)
521{
522 struct parallel_data *pd_old = pinst->pd;
523 int notification_mask = 0;
524
525 pinst->flags |= PADATA_RESET;
526
527 rcu_assign_pointer(pinst->pd, pd_new);
528
529 synchronize_rcu();
530
531 if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
532 notification_mask |= PADATA_CPU_PARALLEL;
533 if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
534 notification_mask |= PADATA_CPU_SERIAL;
535
536 padata_flush_queues(pd_old);
537 padata_free_pd(pd_old);
538
539 if (notification_mask)
540 blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
541 notification_mask,
542 &pd_new->cpumask);
543
544 pinst->flags &= ~PADATA_RESET;
545}
546
547/**
548 * padata_register_cpumask_notifier - Registers a notifier that will be called
549 * if either pcpu or cbcpu or both cpumasks change.
550 *
551 * @pinst: A poineter to padata instance
552 * @nblock: A pointer to notifier block.
553 */
554int padata_register_cpumask_notifier(struct padata_instance *pinst,
555 struct notifier_block *nblock)
556{
557 return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
558 nblock);
559}
560EXPORT_SYMBOL(padata_register_cpumask_notifier);
561
562/**
563 * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
564 * registered earlier using padata_register_cpumask_notifier
565 *
566 * @pinst: A pointer to data instance.
567 * @nlock: A pointer to notifier block.
568 */
569int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
570 struct notifier_block *nblock)
571{
572 return blocking_notifier_chain_unregister(
573 &pinst->cpumask_change_notifier,
574 nblock);
575}
576EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
577
578
579/* If cpumask contains no active cpu, we mark the instance as invalid. */
580static bool padata_validate_cpumask(struct padata_instance *pinst,
581 const struct cpumask *cpumask)
582{
583 if (!cpumask_intersects(cpumask, cpu_active_mask)) {
584 pinst->flags |= PADATA_INVALID;
585 return false;
586 }
587
588 pinst->flags &= ~PADATA_INVALID;
589 return true;
590}
591
592static int __padata_set_cpumasks(struct padata_instance *pinst,
593 cpumask_var_t pcpumask,
594 cpumask_var_t cbcpumask)
595{
596 int valid;
597 struct parallel_data *pd;
598
599 valid = padata_validate_cpumask(pinst, pcpumask);
600 if (!valid) {
601 __padata_stop(pinst);
602 goto out_replace;
603 }
604
605 valid = padata_validate_cpumask(pinst, cbcpumask);
606 if (!valid)
607 __padata_stop(pinst);
608
609out_replace:
610 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
611 if (!pd)
612 return -ENOMEM;
613
614 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
615 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
616
617 padata_replace(pinst, pd);
618
619 if (valid)
620 __padata_start(pinst);
621
622 return 0;
623}
624
625/**
626 * padata_set_cpumasks - Set both parallel and serial cpumasks. The first
627 * one is used by parallel workers and the second one
628 * by the wokers doing serialization.
629 *
630 * @pinst: padata instance
631 * @pcpumask: the cpumask to use for parallel workers
632 * @cbcpumask: the cpumsak to use for serial workers
633 */
634int padata_set_cpumasks(struct padata_instance *pinst, cpumask_var_t pcpumask,
635 cpumask_var_t cbcpumask)
636{
637 int err;
638
639 mutex_lock(&pinst->lock);
640 get_online_cpus();
641
642 err = __padata_set_cpumasks(pinst, pcpumask, cbcpumask);
643
644 put_online_cpus();
645 mutex_unlock(&pinst->lock);
646
647 return err;
648
649}
650EXPORT_SYMBOL(padata_set_cpumasks);
651
652/**
653 * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
654 * equivalent to @cpumask.
655 *
656 * @pinst: padata instance
657 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
658 * to parallel and serial cpumasks respectively.
659 * @cpumask: the cpumask to use
660 */
661int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
662 cpumask_var_t cpumask)
663{
664 struct cpumask *serial_mask, *parallel_mask;
665 int err = -EINVAL;
666
667 mutex_lock(&pinst->lock);
668 get_online_cpus();
669
670 switch (cpumask_type) {
671 case PADATA_CPU_PARALLEL:
672 serial_mask = pinst->cpumask.cbcpu;
673 parallel_mask = cpumask;
674 break;
675 case PADATA_CPU_SERIAL:
676 parallel_mask = pinst->cpumask.pcpu;
677 serial_mask = cpumask;
678 break;
679 default:
680 goto out;
681 }
682
683 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
684
685out:
686 put_online_cpus();
687 mutex_unlock(&pinst->lock);
688
689 return err;
690}
691EXPORT_SYMBOL(padata_set_cpumask);
692
693static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
694{
695 struct parallel_data *pd;
696
697 if (cpumask_test_cpu(cpu, cpu_active_mask)) {
698 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
699 pinst->cpumask.cbcpu);
700 if (!pd)
701 return -ENOMEM;
702
703 padata_replace(pinst, pd);
704
705 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
706 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
707 __padata_start(pinst);
708 }
709
710 return 0;
711}
712
713 /**
714 * padata_add_cpu - add a cpu to one or both(parallel and serial)
715 * padata cpumasks.
716 *
717 * @pinst: padata instance
718 * @cpu: cpu to add
719 * @mask: bitmask of flags specifying to which cpumask @cpu shuld be added.
720 * The @mask may be any combination of the following flags:
721 * PADATA_CPU_SERIAL - serial cpumask
722 * PADATA_CPU_PARALLEL - parallel cpumask
723 */
724
725int padata_add_cpu(struct padata_instance *pinst, int cpu, int mask)
726{
727 int err;
728
729 if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
730 return -EINVAL;
731
732 mutex_lock(&pinst->lock);
733
734 get_online_cpus();
735 if (mask & PADATA_CPU_SERIAL)
736 cpumask_set_cpu(cpu, pinst->cpumask.cbcpu);
737 if (mask & PADATA_CPU_PARALLEL)
738 cpumask_set_cpu(cpu, pinst->cpumask.pcpu);
739
740 err = __padata_add_cpu(pinst, cpu);
741 put_online_cpus();
742
743 mutex_unlock(&pinst->lock);
744
745 return err;
746}
747EXPORT_SYMBOL(padata_add_cpu);
748
749static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
750{
751 struct parallel_data *pd = NULL;
752
753 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
754
755 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
756 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
757 __padata_stop(pinst);
758
759 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
760 pinst->cpumask.cbcpu);
761 if (!pd)
762 return -ENOMEM;
763
764 padata_replace(pinst, pd);
765 }
766
767 return 0;
768}
769
770 /**
771 * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
772 * padata cpumasks.
773 *
774 * @pinst: padata instance
775 * @cpu: cpu to remove
776 * @mask: bitmask specifying from which cpumask @cpu should be removed
777 * The @mask may be any combination of the following flags:
778 * PADATA_CPU_SERIAL - serial cpumask
779 * PADATA_CPU_PARALLEL - parallel cpumask
780 */
781int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
782{
783 int err;
784
785 if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
786 return -EINVAL;
787
788 mutex_lock(&pinst->lock);
789
790 get_online_cpus();
791 if (mask & PADATA_CPU_SERIAL)
792 cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
793 if (mask & PADATA_CPU_PARALLEL)
794 cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
795
796 err = __padata_remove_cpu(pinst, cpu);
797 put_online_cpus();
798
799 mutex_unlock(&pinst->lock);
800
801 return err;
802}
803EXPORT_SYMBOL(padata_remove_cpu);
804
805/**
806 * padata_start - start the parallel processing
807 *
808 * @pinst: padata instance to start
809 */
810int padata_start(struct padata_instance *pinst)
811{
812 int err = 0;
813
814 mutex_lock(&pinst->lock);
815
816 if (pinst->flags & PADATA_INVALID)
817 err =-EINVAL;
818
819 __padata_start(pinst);
820
821 mutex_unlock(&pinst->lock);
822
823 return err;
824}
825EXPORT_SYMBOL(padata_start);
826
827/**
828 * padata_stop - stop the parallel processing
829 *
830 * @pinst: padata instance to stop
831 */
832void padata_stop(struct padata_instance *pinst)
833{
834 mutex_lock(&pinst->lock);
835 __padata_stop(pinst);
836 mutex_unlock(&pinst->lock);
837}
838EXPORT_SYMBOL(padata_stop);
839
840#ifdef CONFIG_HOTPLUG_CPU
841
842static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
843{
844 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
845 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
846}
847
848
849static int padata_cpu_callback(struct notifier_block *nfb,
850 unsigned long action, void *hcpu)
851{
852 int err;
853 struct padata_instance *pinst;
854 int cpu = (unsigned long)hcpu;
855
856 pinst = container_of(nfb, struct padata_instance, cpu_notifier);
857
858 switch (action) {
859 case CPU_ONLINE:
860 case CPU_ONLINE_FROZEN:
861 if (!pinst_has_cpu(pinst, cpu))
862 break;
863 mutex_lock(&pinst->lock);
864 err = __padata_add_cpu(pinst, cpu);
865 mutex_unlock(&pinst->lock);
866 if (err)
867 return notifier_from_errno(err);
868 break;
869
870 case CPU_DOWN_PREPARE:
871 case CPU_DOWN_PREPARE_FROZEN:
872 if (!pinst_has_cpu(pinst, cpu))
873 break;
874 mutex_lock(&pinst->lock);
875 err = __padata_remove_cpu(pinst, cpu);
876 mutex_unlock(&pinst->lock);
877 if (err)
878 return notifier_from_errno(err);
879 break;
880
881 case CPU_UP_CANCELED:
882 case CPU_UP_CANCELED_FROZEN:
883 if (!pinst_has_cpu(pinst, cpu))
884 break;
885 mutex_lock(&pinst->lock);
886 __padata_remove_cpu(pinst, cpu);
887 mutex_unlock(&pinst->lock);
888
889 case CPU_DOWN_FAILED:
890 case CPU_DOWN_FAILED_FROZEN:
891 if (!pinst_has_cpu(pinst, cpu))
892 break;
893 mutex_lock(&pinst->lock);
894 __padata_add_cpu(pinst, cpu);
895 mutex_unlock(&pinst->lock);
896 }
897
898 return NOTIFY_OK;
899}
900#endif
901
902static void __padata_free(struct padata_instance *pinst)
903{
904#ifdef CONFIG_HOTPLUG_CPU
905 unregister_hotcpu_notifier(&pinst->cpu_notifier);
906#endif
907
908 padata_stop(pinst);
909 padata_free_pd(pinst->pd);
910 free_cpumask_var(pinst->cpumask.pcpu);
911 free_cpumask_var(pinst->cpumask.cbcpu);
912 kfree(pinst);
913}
914
915#define kobj2pinst(_kobj) \
916 container_of(_kobj, struct padata_instance, kobj)
917#define attr2pentry(_attr) \
918 container_of(_attr, struct padata_sysfs_entry, attr)
919
920static void padata_sysfs_release(struct kobject *kobj)
921{
922 struct padata_instance *pinst = kobj2pinst(kobj);
923 __padata_free(pinst);
924}
925
926struct padata_sysfs_entry {
927 struct attribute attr;
928 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
929 ssize_t (*store)(struct padata_instance *, struct attribute *,
930 const char *, size_t);
931};
932
933static ssize_t show_cpumask(struct padata_instance *pinst,
934 struct attribute *attr, char *buf)
935{
936 struct cpumask *cpumask;
937 ssize_t len;
938
939 mutex_lock(&pinst->lock);
940 if (!strcmp(attr->name, "serial_cpumask"))
941 cpumask = pinst->cpumask.cbcpu;
942 else
943 cpumask = pinst->cpumask.pcpu;
944
945 len = bitmap_scnprintf(buf, PAGE_SIZE, cpumask_bits(cpumask),
946 nr_cpu_ids);
947 if (PAGE_SIZE - len < 2)
948 len = -EINVAL;
949 else
950 len += sprintf(buf + len, "\n");
951
952 mutex_unlock(&pinst->lock);
953 return len;
954}
955
956static ssize_t store_cpumask(struct padata_instance *pinst,
957 struct attribute *attr,
958 const char *buf, size_t count)
959{
960 cpumask_var_t new_cpumask;
961 ssize_t ret;
962 int mask_type;
963
964 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
965 return -ENOMEM;
966
967 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
968 nr_cpumask_bits);
969 if (ret < 0)
970 goto out;
971
972 mask_type = !strcmp(attr->name, "serial_cpumask") ?
973 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
974 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
975 if (!ret)
976 ret = count;
977
978out:
979 free_cpumask_var(new_cpumask);
980 return ret;
981}
982
983#define PADATA_ATTR_RW(_name, _show_name, _store_name) \
984 static struct padata_sysfs_entry _name##_attr = \
985 __ATTR(_name, 0644, _show_name, _store_name)
986#define PADATA_ATTR_RO(_name, _show_name) \
987 static struct padata_sysfs_entry _name##_attr = \
988 __ATTR(_name, 0400, _show_name, NULL)
989
990PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
991PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
992
993/*
994 * Padata sysfs provides the following objects:
995 * serial_cpumask [RW] - cpumask for serial workers
996 * parallel_cpumask [RW] - cpumask for parallel workers
997 */
998static struct attribute *padata_default_attrs[] = {
999 &serial_cpumask_attr.attr,
1000 ¶llel_cpumask_attr.attr,
1001 NULL,
1002};
1003
1004static ssize_t padata_sysfs_show(struct kobject *kobj,
1005 struct attribute *attr, char *buf)
1006{
1007 struct padata_instance *pinst;
1008 struct padata_sysfs_entry *pentry;
1009 ssize_t ret = -EIO;
1010
1011 pinst = kobj2pinst(kobj);
1012 pentry = attr2pentry(attr);
1013 if (pentry->show)
1014 ret = pentry->show(pinst, attr, buf);
1015
1016 return ret;
1017}
1018
1019static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
1020 const char *buf, size_t count)
1021{
1022 struct padata_instance *pinst;
1023 struct padata_sysfs_entry *pentry;
1024 ssize_t ret = -EIO;
1025
1026 pinst = kobj2pinst(kobj);
1027 pentry = attr2pentry(attr);
1028 if (pentry->show)
1029 ret = pentry->store(pinst, attr, buf, count);
1030
1031 return ret;
1032}
1033
1034static const struct sysfs_ops padata_sysfs_ops = {
1035 .show = padata_sysfs_show,
1036 .store = padata_sysfs_store,
1037};
1038
1039static struct kobj_type padata_attr_type = {
1040 .sysfs_ops = &padata_sysfs_ops,
1041 .default_attrs = padata_default_attrs,
1042 .release = padata_sysfs_release,
1043};
1044
1045/**
1046 * padata_alloc_possible - Allocate and initialize padata instance.
1047 * Use the cpu_possible_mask for serial and
1048 * parallel workers.
1049 *
1050 * @wq: workqueue to use for the allocated padata instance
1051 */
1052struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq)
1053{
1054 return padata_alloc(wq, cpu_possible_mask, cpu_possible_mask);
1055}
1056EXPORT_SYMBOL(padata_alloc_possible);
1057
1058/**
1059 * padata_alloc - allocate and initialize a padata instance and specify
1060 * cpumasks for serial and parallel workers.
1061 *
1062 * @wq: workqueue to use for the allocated padata instance
1063 * @pcpumask: cpumask that will be used for padata parallelization
1064 * @cbcpumask: cpumask that will be used for padata serialization
1065 */
1066struct padata_instance *padata_alloc(struct workqueue_struct *wq,
1067 const struct cpumask *pcpumask,
1068 const struct cpumask *cbcpumask)
1069{
1070 struct padata_instance *pinst;
1071 struct parallel_data *pd = NULL;
1072
1073 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
1074 if (!pinst)
1075 goto err;
1076
1077 get_online_cpus();
1078 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
1079 goto err_free_inst;
1080 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
1081 free_cpumask_var(pinst->cpumask.pcpu);
1082 goto err_free_inst;
1083 }
1084 if (!padata_validate_cpumask(pinst, pcpumask) ||
1085 !padata_validate_cpumask(pinst, cbcpumask))
1086 goto err_free_masks;
1087
1088 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
1089 if (!pd)
1090 goto err_free_masks;
1091
1092 rcu_assign_pointer(pinst->pd, pd);
1093
1094 pinst->wq = wq;
1095
1096 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
1097 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
1098
1099 pinst->flags = 0;
1100
1101#ifdef CONFIG_HOTPLUG_CPU
1102 pinst->cpu_notifier.notifier_call = padata_cpu_callback;
1103 pinst->cpu_notifier.priority = 0;
1104 register_hotcpu_notifier(&pinst->cpu_notifier);
1105#endif
1106
1107 put_online_cpus();
1108
1109 BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
1110 kobject_init(&pinst->kobj, &padata_attr_type);
1111 mutex_init(&pinst->lock);
1112
1113 return pinst;
1114
1115err_free_masks:
1116 free_cpumask_var(pinst->cpumask.pcpu);
1117 free_cpumask_var(pinst->cpumask.cbcpu);
1118err_free_inst:
1119 kfree(pinst);
1120 put_online_cpus();
1121err:
1122 return NULL;
1123}
1124EXPORT_SYMBOL(padata_alloc);
1125
1126/**
1127 * padata_free - free a padata instance
1128 *
1129 * @padata_inst: padata instance to free
1130 */
1131void padata_free(struct padata_instance *pinst)
1132{
1133 kobject_put(&pinst->kobj);
1134}
1135EXPORT_SYMBOL(padata_free);