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