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// SPDX-License-Identifier: GPL-2.0
2/*
3 * padata.c - generic interface to process data streams in parallel
4 *
5 * See Documentation/padata.txt for an api documentation.
6 *
7 * Copyright (C) 2008, 2009 secunet Security Networks AG
8 * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
9 *
10 * This program is free software; you can redistribute it and/or modify it
11 * under the terms and conditions of the GNU General Public License,
12 * version 2, as published by the Free Software Foundation.
13 *
14 * This program is distributed in the hope it will be useful, but WITHOUT
15 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
16 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
17 * more details.
18 *
19 * You should have received a copy of the GNU General Public License along with
20 * this program; if not, write to the Free Software Foundation, Inc.,
21 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
22 */
23
24#include <linux/export.h>
25#include <linux/cpumask.h>
26#include <linux/err.h>
27#include <linux/cpu.h>
28#include <linux/padata.h>
29#include <linux/mutex.h>
30#include <linux/sched.h>
31#include <linux/slab.h>
32#include <linux/sysfs.h>
33#include <linux/rcupdate.h>
34#include <linux/module.h>
35
36#define MAX_OBJ_NUM 1000
37
38static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
39{
40 int cpu, target_cpu;
41
42 target_cpu = cpumask_first(pd->cpumask.pcpu);
43 for (cpu = 0; cpu < cpu_index; cpu++)
44 target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
45
46 return target_cpu;
47}
48
49static int padata_cpu_hash(struct parallel_data *pd)
50{
51 unsigned int seq_nr;
52 int cpu_index;
53
54 /*
55 * Hash the sequence numbers to the cpus by taking
56 * seq_nr mod. number of cpus in use.
57 */
58
59 seq_nr = atomic_inc_return(&pd->seq_nr);
60 cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
61
62 return padata_index_to_cpu(pd, cpu_index);
63}
64
65static void padata_parallel_worker(struct work_struct *parallel_work)
66{
67 struct padata_parallel_queue *pqueue;
68 LIST_HEAD(local_list);
69
70 local_bh_disable();
71 pqueue = container_of(parallel_work,
72 struct padata_parallel_queue, work);
73
74 spin_lock(&pqueue->parallel.lock);
75 list_replace_init(&pqueue->parallel.list, &local_list);
76 spin_unlock(&pqueue->parallel.lock);
77
78 while (!list_empty(&local_list)) {
79 struct padata_priv *padata;
80
81 padata = list_entry(local_list.next,
82 struct padata_priv, list);
83
84 list_del_init(&padata->list);
85
86 padata->parallel(padata);
87 }
88
89 local_bh_enable();
90}
91
92/**
93 * padata_do_parallel - padata parallelization function
94 *
95 * @pinst: padata instance
96 * @padata: object to be parallelized
97 * @cb_cpu: cpu the serialization callback function will run on,
98 * must be in the serial cpumask of padata(i.e. cpumask.cbcpu).
99 *
100 * The parallelization callback function will run with BHs off.
101 * Note: Every object which is parallelized by padata_do_parallel
102 * must be seen by padata_do_serial.
103 */
104int padata_do_parallel(struct padata_instance *pinst,
105 struct padata_priv *padata, int cb_cpu)
106{
107 int target_cpu, err;
108 struct padata_parallel_queue *queue;
109 struct parallel_data *pd;
110
111 rcu_read_lock_bh();
112
113 pd = rcu_dereference_bh(pinst->pd);
114
115 err = -EINVAL;
116 if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
117 goto out;
118
119 if (!cpumask_test_cpu(cb_cpu, pd->cpumask.cbcpu))
120 goto out;
121
122 err = -EBUSY;
123 if ((pinst->flags & PADATA_RESET))
124 goto out;
125
126 if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
127 goto out;
128
129 err = 0;
130 atomic_inc(&pd->refcnt);
131 padata->pd = pd;
132 padata->cb_cpu = cb_cpu;
133
134 target_cpu = padata_cpu_hash(pd);
135 padata->cpu = target_cpu;
136 queue = per_cpu_ptr(pd->pqueue, target_cpu);
137
138 spin_lock(&queue->parallel.lock);
139 list_add_tail(&padata->list, &queue->parallel.list);
140 spin_unlock(&queue->parallel.lock);
141
142 queue_work_on(target_cpu, pinst->wq, &queue->work);
143
144out:
145 rcu_read_unlock_bh();
146
147 return err;
148}
149EXPORT_SYMBOL(padata_do_parallel);
150
151/*
152 * padata_get_next - Get the next object that needs serialization.
153 *
154 * Return values are:
155 *
156 * A pointer to the control struct of the next object that needs
157 * serialization, if present in one of the percpu reorder queues.
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 reorder = &next_queue->reorder;
186
187 spin_lock(&reorder->lock);
188 if (!list_empty(&reorder->list)) {
189 padata = list_entry(reorder->list.next,
190 struct padata_priv, list);
191
192 list_del_init(&padata->list);
193 atomic_dec(&pd->reorder_objects);
194
195 pd->processed++;
196
197 spin_unlock(&reorder->lock);
198 goto out;
199 }
200 spin_unlock(&reorder->lock);
201
202 if (__this_cpu_read(pd->pqueue->cpu_index) == next_queue->cpu_index) {
203 padata = ERR_PTR(-ENODATA);
204 goto out;
205 }
206
207 padata = ERR_PTR(-EINPROGRESS);
208out:
209 return padata;
210}
211
212static void padata_reorder(struct parallel_data *pd)
213{
214 int cb_cpu;
215 struct padata_priv *padata;
216 struct padata_serial_queue *squeue;
217 struct padata_instance *pinst = pd->pinst;
218
219 /*
220 * We need to ensure that only one cpu can work on dequeueing of
221 * the reorder queue the time. Calculating in which percpu reorder
222 * queue the next object will arrive takes some time. A spinlock
223 * would be highly contended. Also it is not clear in which order
224 * the objects arrive to the reorder queues. So a cpu could wait to
225 * get the lock just to notice that there is nothing to do at the
226 * moment. Therefore we use a trylock and let the holder of the lock
227 * care for all the objects enqueued during the holdtime of the lock.
228 */
229 if (!spin_trylock_bh(&pd->lock))
230 return;
231
232 while (1) {
233 padata = padata_get_next(pd);
234
235 /*
236 * If the next object that needs serialization is parallel
237 * processed by another cpu and is still on it's way to the
238 * cpu's reorder queue, nothing to do for now.
239 */
240 if (PTR_ERR(padata) == -EINPROGRESS)
241 break;
242
243 /*
244 * This cpu has to do the parallel processing of the next
245 * object. It's waiting in the cpu's parallelization queue,
246 * so exit immediately.
247 */
248 if (PTR_ERR(padata) == -ENODATA) {
249 del_timer(&pd->timer);
250 spin_unlock_bh(&pd->lock);
251 return;
252 }
253
254 cb_cpu = padata->cb_cpu;
255 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
256
257 spin_lock(&squeue->serial.lock);
258 list_add_tail(&padata->list, &squeue->serial.list);
259 spin_unlock(&squeue->serial.lock);
260
261 queue_work_on(cb_cpu, pinst->wq, &squeue->work);
262 }
263
264 spin_unlock_bh(&pd->lock);
265
266 /*
267 * The next object that needs serialization might have arrived to
268 * the reorder queues in the meantime, we will be called again
269 * from the timer function if no one else cares for it.
270 */
271 if (atomic_read(&pd->reorder_objects)
272 && !(pinst->flags & PADATA_RESET))
273 mod_timer(&pd->timer, jiffies + HZ);
274 else
275 del_timer(&pd->timer);
276
277 return;
278}
279
280static void invoke_padata_reorder(struct work_struct *work)
281{
282 struct padata_parallel_queue *pqueue;
283 struct parallel_data *pd;
284
285 local_bh_disable();
286 pqueue = container_of(work, struct padata_parallel_queue, reorder_work);
287 pd = pqueue->pd;
288 padata_reorder(pd);
289 local_bh_enable();
290}
291
292static void padata_reorder_timer(struct timer_list *t)
293{
294 struct parallel_data *pd = from_timer(pd, t, timer);
295 unsigned int weight;
296 int target_cpu, cpu;
297
298 cpu = get_cpu();
299
300 /* We don't lock pd here to not interfere with parallel processing
301 * padata_reorder() calls on other CPUs. We just need any CPU out of
302 * the cpumask.pcpu set. It would be nice if it's the right one but
303 * it doesn't matter if we're off to the next one by using an outdated
304 * pd->processed value.
305 */
306 weight = cpumask_weight(pd->cpumask.pcpu);
307 target_cpu = padata_index_to_cpu(pd, pd->processed % weight);
308
309 /* ensure to call the reorder callback on the correct CPU */
310 if (cpu != target_cpu) {
311 struct padata_parallel_queue *pqueue;
312 struct padata_instance *pinst;
313
314 /* The timer function is serialized wrt itself -- no locking
315 * needed.
316 */
317 pinst = pd->pinst;
318 pqueue = per_cpu_ptr(pd->pqueue, target_cpu);
319 queue_work_on(target_cpu, pinst->wq, &pqueue->reorder_work);
320 } else {
321 padata_reorder(pd);
322 }
323
324 put_cpu();
325}
326
327static void padata_serial_worker(struct work_struct *serial_work)
328{
329 struct padata_serial_queue *squeue;
330 struct parallel_data *pd;
331 LIST_HEAD(local_list);
332
333 local_bh_disable();
334 squeue = container_of(serial_work, struct padata_serial_queue, work);
335 pd = squeue->pd;
336
337 spin_lock(&squeue->serial.lock);
338 list_replace_init(&squeue->serial.list, &local_list);
339 spin_unlock(&squeue->serial.lock);
340
341 while (!list_empty(&local_list)) {
342 struct padata_priv *padata;
343
344 padata = list_entry(local_list.next,
345 struct padata_priv, list);
346
347 list_del_init(&padata->list);
348
349 padata->serial(padata);
350 atomic_dec(&pd->refcnt);
351 }
352 local_bh_enable();
353}
354
355/**
356 * padata_do_serial - padata serialization function
357 *
358 * @padata: object to be serialized.
359 *
360 * padata_do_serial must be called for every parallelized object.
361 * The serialization callback function will run with BHs off.
362 */
363void padata_do_serial(struct padata_priv *padata)
364{
365 int cpu;
366 struct padata_parallel_queue *pqueue;
367 struct parallel_data *pd;
368 int reorder_via_wq = 0;
369
370 pd = padata->pd;
371
372 cpu = get_cpu();
373
374 /* We need to run on the same CPU padata_do_parallel(.., padata, ..)
375 * was called on -- or, at least, enqueue the padata object into the
376 * correct per-cpu queue.
377 */
378 if (cpu != padata->cpu) {
379 reorder_via_wq = 1;
380 cpu = padata->cpu;
381 }
382
383 pqueue = per_cpu_ptr(pd->pqueue, cpu);
384
385 spin_lock(&pqueue->reorder.lock);
386 atomic_inc(&pd->reorder_objects);
387 list_add_tail(&padata->list, &pqueue->reorder.list);
388 spin_unlock(&pqueue->reorder.lock);
389
390 put_cpu();
391
392 /* If we're running on the wrong CPU, call padata_reorder() via a
393 * kernel worker.
394 */
395 if (reorder_via_wq)
396 queue_work_on(cpu, pd->pinst->wq, &pqueue->reorder_work);
397 else
398 padata_reorder(pd);
399}
400EXPORT_SYMBOL(padata_do_serial);
401
402static int padata_setup_cpumasks(struct parallel_data *pd,
403 const struct cpumask *pcpumask,
404 const struct cpumask *cbcpumask)
405{
406 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
407 return -ENOMEM;
408
409 cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_online_mask);
410 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL)) {
411 free_cpumask_var(pd->cpumask.pcpu);
412 return -ENOMEM;
413 }
414
415 cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_online_mask);
416 return 0;
417}
418
419static void __padata_list_init(struct padata_list *pd_list)
420{
421 INIT_LIST_HEAD(&pd_list->list);
422 spin_lock_init(&pd_list->lock);
423}
424
425/* Initialize all percpu queues used by serial workers */
426static void padata_init_squeues(struct parallel_data *pd)
427{
428 int cpu;
429 struct padata_serial_queue *squeue;
430
431 for_each_cpu(cpu, pd->cpumask.cbcpu) {
432 squeue = per_cpu_ptr(pd->squeue, cpu);
433 squeue->pd = pd;
434 __padata_list_init(&squeue->serial);
435 INIT_WORK(&squeue->work, padata_serial_worker);
436 }
437}
438
439/* Initialize all percpu queues used by parallel workers */
440static void padata_init_pqueues(struct parallel_data *pd)
441{
442 int cpu_index, cpu;
443 struct padata_parallel_queue *pqueue;
444
445 cpu_index = 0;
446 for_each_possible_cpu(cpu) {
447 pqueue = per_cpu_ptr(pd->pqueue, cpu);
448
449 if (!cpumask_test_cpu(cpu, pd->cpumask.pcpu)) {
450 pqueue->cpu_index = -1;
451 continue;
452 }
453
454 pqueue->pd = pd;
455 pqueue->cpu_index = cpu_index;
456 cpu_index++;
457
458 __padata_list_init(&pqueue->reorder);
459 __padata_list_init(&pqueue->parallel);
460 INIT_WORK(&pqueue->work, padata_parallel_worker);
461 INIT_WORK(&pqueue->reorder_work, invoke_padata_reorder);
462 atomic_set(&pqueue->num_obj, 0);
463 }
464}
465
466/* Allocate and initialize the internal cpumask dependend resources. */
467static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
468 const struct cpumask *pcpumask,
469 const struct cpumask *cbcpumask)
470{
471 struct parallel_data *pd;
472
473 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
474 if (!pd)
475 goto err;
476
477 pd->pqueue = alloc_percpu(struct padata_parallel_queue);
478 if (!pd->pqueue)
479 goto err_free_pd;
480
481 pd->squeue = alloc_percpu(struct padata_serial_queue);
482 if (!pd->squeue)
483 goto err_free_pqueue;
484 if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
485 goto err_free_squeue;
486
487 padata_init_pqueues(pd);
488 padata_init_squeues(pd);
489 timer_setup(&pd->timer, padata_reorder_timer, 0);
490 atomic_set(&pd->seq_nr, -1);
491 atomic_set(&pd->reorder_objects, 0);
492 atomic_set(&pd->refcnt, 0);
493 pd->pinst = pinst;
494 spin_lock_init(&pd->lock);
495
496 return pd;
497
498err_free_squeue:
499 free_percpu(pd->squeue);
500err_free_pqueue:
501 free_percpu(pd->pqueue);
502err_free_pd:
503 kfree(pd);
504err:
505 return NULL;
506}
507
508static void padata_free_pd(struct parallel_data *pd)
509{
510 free_cpumask_var(pd->cpumask.pcpu);
511 free_cpumask_var(pd->cpumask.cbcpu);
512 free_percpu(pd->pqueue);
513 free_percpu(pd->squeue);
514 kfree(pd);
515}
516
517/* Flush all objects out of the padata queues. */
518static void padata_flush_queues(struct parallel_data *pd)
519{
520 int cpu;
521 struct padata_parallel_queue *pqueue;
522 struct padata_serial_queue *squeue;
523
524 for_each_cpu(cpu, pd->cpumask.pcpu) {
525 pqueue = per_cpu_ptr(pd->pqueue, cpu);
526 flush_work(&pqueue->work);
527 }
528
529 del_timer_sync(&pd->timer);
530
531 if (atomic_read(&pd->reorder_objects))
532 padata_reorder(pd);
533
534 for_each_cpu(cpu, pd->cpumask.cbcpu) {
535 squeue = per_cpu_ptr(pd->squeue, cpu);
536 flush_work(&squeue->work);
537 }
538
539 BUG_ON(atomic_read(&pd->refcnt) != 0);
540}
541
542static void __padata_start(struct padata_instance *pinst)
543{
544 pinst->flags |= PADATA_INIT;
545}
546
547static void __padata_stop(struct padata_instance *pinst)
548{
549 if (!(pinst->flags & PADATA_INIT))
550 return;
551
552 pinst->flags &= ~PADATA_INIT;
553
554 synchronize_rcu();
555
556 get_online_cpus();
557 padata_flush_queues(pinst->pd);
558 put_online_cpus();
559}
560
561/* Replace the internal control structure with a new one. */
562static void padata_replace(struct padata_instance *pinst,
563 struct parallel_data *pd_new)
564{
565 struct parallel_data *pd_old = pinst->pd;
566 int notification_mask = 0;
567
568 pinst->flags |= PADATA_RESET;
569
570 rcu_assign_pointer(pinst->pd, pd_new);
571
572 synchronize_rcu();
573
574 if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
575 notification_mask |= PADATA_CPU_PARALLEL;
576 if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
577 notification_mask |= PADATA_CPU_SERIAL;
578
579 padata_flush_queues(pd_old);
580 padata_free_pd(pd_old);
581
582 if (notification_mask)
583 blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
584 notification_mask,
585 &pd_new->cpumask);
586
587 pinst->flags &= ~PADATA_RESET;
588}
589
590/**
591 * padata_register_cpumask_notifier - Registers a notifier that will be called
592 * if either pcpu or cbcpu or both cpumasks change.
593 *
594 * @pinst: A poineter to padata instance
595 * @nblock: A pointer to notifier block.
596 */
597int padata_register_cpumask_notifier(struct padata_instance *pinst,
598 struct notifier_block *nblock)
599{
600 return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
601 nblock);
602}
603EXPORT_SYMBOL(padata_register_cpumask_notifier);
604
605/**
606 * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
607 * registered earlier using padata_register_cpumask_notifier
608 *
609 * @pinst: A pointer to data instance.
610 * @nlock: A pointer to notifier block.
611 */
612int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
613 struct notifier_block *nblock)
614{
615 return blocking_notifier_chain_unregister(
616 &pinst->cpumask_change_notifier,
617 nblock);
618}
619EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
620
621
622/* If cpumask contains no active cpu, we mark the instance as invalid. */
623static bool padata_validate_cpumask(struct padata_instance *pinst,
624 const struct cpumask *cpumask)
625{
626 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
627 pinst->flags |= PADATA_INVALID;
628 return false;
629 }
630
631 pinst->flags &= ~PADATA_INVALID;
632 return true;
633}
634
635static int __padata_set_cpumasks(struct padata_instance *pinst,
636 cpumask_var_t pcpumask,
637 cpumask_var_t cbcpumask)
638{
639 int valid;
640 struct parallel_data *pd;
641
642 valid = padata_validate_cpumask(pinst, pcpumask);
643 if (!valid) {
644 __padata_stop(pinst);
645 goto out_replace;
646 }
647
648 valid = padata_validate_cpumask(pinst, cbcpumask);
649 if (!valid)
650 __padata_stop(pinst);
651
652out_replace:
653 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
654 if (!pd)
655 return -ENOMEM;
656
657 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
658 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
659
660 padata_replace(pinst, pd);
661
662 if (valid)
663 __padata_start(pinst);
664
665 return 0;
666}
667
668/**
669 * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
670 * equivalent to @cpumask.
671 *
672 * @pinst: padata instance
673 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
674 * to parallel and serial cpumasks respectively.
675 * @cpumask: the cpumask to use
676 */
677int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
678 cpumask_var_t cpumask)
679{
680 struct cpumask *serial_mask, *parallel_mask;
681 int err = -EINVAL;
682
683 mutex_lock(&pinst->lock);
684 get_online_cpus();
685
686 switch (cpumask_type) {
687 case PADATA_CPU_PARALLEL:
688 serial_mask = pinst->cpumask.cbcpu;
689 parallel_mask = cpumask;
690 break;
691 case PADATA_CPU_SERIAL:
692 parallel_mask = pinst->cpumask.pcpu;
693 serial_mask = cpumask;
694 break;
695 default:
696 goto out;
697 }
698
699 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
700
701out:
702 put_online_cpus();
703 mutex_unlock(&pinst->lock);
704
705 return err;
706}
707EXPORT_SYMBOL(padata_set_cpumask);
708
709/**
710 * padata_start - start the parallel processing
711 *
712 * @pinst: padata instance to start
713 */
714int padata_start(struct padata_instance *pinst)
715{
716 int err = 0;
717
718 mutex_lock(&pinst->lock);
719
720 if (pinst->flags & PADATA_INVALID)
721 err = -EINVAL;
722
723 __padata_start(pinst);
724
725 mutex_unlock(&pinst->lock);
726
727 return err;
728}
729EXPORT_SYMBOL(padata_start);
730
731/**
732 * padata_stop - stop the parallel processing
733 *
734 * @pinst: padata instance to stop
735 */
736void padata_stop(struct padata_instance *pinst)
737{
738 mutex_lock(&pinst->lock);
739 __padata_stop(pinst);
740 mutex_unlock(&pinst->lock);
741}
742EXPORT_SYMBOL(padata_stop);
743
744#ifdef CONFIG_HOTPLUG_CPU
745
746static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
747{
748 struct parallel_data *pd;
749
750 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
751 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
752 pinst->cpumask.cbcpu);
753 if (!pd)
754 return -ENOMEM;
755
756 padata_replace(pinst, pd);
757
758 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
759 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
760 __padata_start(pinst);
761 }
762
763 return 0;
764}
765
766static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
767{
768 struct parallel_data *pd = NULL;
769
770 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
771
772 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
773 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
774 __padata_stop(pinst);
775
776 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
777 pinst->cpumask.cbcpu);
778 if (!pd)
779 return -ENOMEM;
780
781 padata_replace(pinst, pd);
782
783 cpumask_clear_cpu(cpu, pd->cpumask.cbcpu);
784 cpumask_clear_cpu(cpu, pd->cpumask.pcpu);
785 }
786
787 return 0;
788}
789
790 /**
791 * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
792 * padata cpumasks.
793 *
794 * @pinst: padata instance
795 * @cpu: cpu to remove
796 * @mask: bitmask specifying from which cpumask @cpu should be removed
797 * The @mask may be any combination of the following flags:
798 * PADATA_CPU_SERIAL - serial cpumask
799 * PADATA_CPU_PARALLEL - parallel cpumask
800 */
801int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
802{
803 int err;
804
805 if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
806 return -EINVAL;
807
808 mutex_lock(&pinst->lock);
809
810 get_online_cpus();
811 if (mask & PADATA_CPU_SERIAL)
812 cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
813 if (mask & PADATA_CPU_PARALLEL)
814 cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
815
816 err = __padata_remove_cpu(pinst, cpu);
817 put_online_cpus();
818
819 mutex_unlock(&pinst->lock);
820
821 return err;
822}
823EXPORT_SYMBOL(padata_remove_cpu);
824
825static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
826{
827 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
828 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
829}
830
831static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
832{
833 struct padata_instance *pinst;
834 int ret;
835
836 pinst = hlist_entry_safe(node, struct padata_instance, node);
837 if (!pinst_has_cpu(pinst, cpu))
838 return 0;
839
840 mutex_lock(&pinst->lock);
841 ret = __padata_add_cpu(pinst, cpu);
842 mutex_unlock(&pinst->lock);
843 return ret;
844}
845
846static int padata_cpu_prep_down(unsigned int cpu, struct hlist_node *node)
847{
848 struct padata_instance *pinst;
849 int ret;
850
851 pinst = hlist_entry_safe(node, struct padata_instance, node);
852 if (!pinst_has_cpu(pinst, cpu))
853 return 0;
854
855 mutex_lock(&pinst->lock);
856 ret = __padata_remove_cpu(pinst, cpu);
857 mutex_unlock(&pinst->lock);
858 return ret;
859}
860
861static enum cpuhp_state hp_online;
862#endif
863
864static void __padata_free(struct padata_instance *pinst)
865{
866#ifdef CONFIG_HOTPLUG_CPU
867 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->node);
868#endif
869
870 padata_stop(pinst);
871 padata_free_pd(pinst->pd);
872 free_cpumask_var(pinst->cpumask.pcpu);
873 free_cpumask_var(pinst->cpumask.cbcpu);
874 kfree(pinst);
875}
876
877#define kobj2pinst(_kobj) \
878 container_of(_kobj, struct padata_instance, kobj)
879#define attr2pentry(_attr) \
880 container_of(_attr, struct padata_sysfs_entry, attr)
881
882static void padata_sysfs_release(struct kobject *kobj)
883{
884 struct padata_instance *pinst = kobj2pinst(kobj);
885 __padata_free(pinst);
886}
887
888struct padata_sysfs_entry {
889 struct attribute attr;
890 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
891 ssize_t (*store)(struct padata_instance *, struct attribute *,
892 const char *, size_t);
893};
894
895static ssize_t show_cpumask(struct padata_instance *pinst,
896 struct attribute *attr, char *buf)
897{
898 struct cpumask *cpumask;
899 ssize_t len;
900
901 mutex_lock(&pinst->lock);
902 if (!strcmp(attr->name, "serial_cpumask"))
903 cpumask = pinst->cpumask.cbcpu;
904 else
905 cpumask = pinst->cpumask.pcpu;
906
907 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
908 nr_cpu_ids, cpumask_bits(cpumask));
909 mutex_unlock(&pinst->lock);
910 return len < PAGE_SIZE ? len : -EINVAL;
911}
912
913static ssize_t store_cpumask(struct padata_instance *pinst,
914 struct attribute *attr,
915 const char *buf, size_t count)
916{
917 cpumask_var_t new_cpumask;
918 ssize_t ret;
919 int mask_type;
920
921 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
922 return -ENOMEM;
923
924 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
925 nr_cpumask_bits);
926 if (ret < 0)
927 goto out;
928
929 mask_type = !strcmp(attr->name, "serial_cpumask") ?
930 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
931 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
932 if (!ret)
933 ret = count;
934
935out:
936 free_cpumask_var(new_cpumask);
937 return ret;
938}
939
940#define PADATA_ATTR_RW(_name, _show_name, _store_name) \
941 static struct padata_sysfs_entry _name##_attr = \
942 __ATTR(_name, 0644, _show_name, _store_name)
943#define PADATA_ATTR_RO(_name, _show_name) \
944 static struct padata_sysfs_entry _name##_attr = \
945 __ATTR(_name, 0400, _show_name, NULL)
946
947PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
948PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
949
950/*
951 * Padata sysfs provides the following objects:
952 * serial_cpumask [RW] - cpumask for serial workers
953 * parallel_cpumask [RW] - cpumask for parallel workers
954 */
955static struct attribute *padata_default_attrs[] = {
956 &serial_cpumask_attr.attr,
957 ¶llel_cpumask_attr.attr,
958 NULL,
959};
960
961static ssize_t padata_sysfs_show(struct kobject *kobj,
962 struct attribute *attr, char *buf)
963{
964 struct padata_instance *pinst;
965 struct padata_sysfs_entry *pentry;
966 ssize_t ret = -EIO;
967
968 pinst = kobj2pinst(kobj);
969 pentry = attr2pentry(attr);
970 if (pentry->show)
971 ret = pentry->show(pinst, attr, buf);
972
973 return ret;
974}
975
976static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
977 const char *buf, size_t count)
978{
979 struct padata_instance *pinst;
980 struct padata_sysfs_entry *pentry;
981 ssize_t ret = -EIO;
982
983 pinst = kobj2pinst(kobj);
984 pentry = attr2pentry(attr);
985 if (pentry->show)
986 ret = pentry->store(pinst, attr, buf, count);
987
988 return ret;
989}
990
991static const struct sysfs_ops padata_sysfs_ops = {
992 .show = padata_sysfs_show,
993 .store = padata_sysfs_store,
994};
995
996static struct kobj_type padata_attr_type = {
997 .sysfs_ops = &padata_sysfs_ops,
998 .default_attrs = padata_default_attrs,
999 .release = padata_sysfs_release,
1000};
1001
1002/**
1003 * padata_alloc - allocate and initialize a padata instance and specify
1004 * cpumasks for serial and parallel workers.
1005 *
1006 * @wq: workqueue to use for the allocated padata instance
1007 * @pcpumask: cpumask that will be used for padata parallelization
1008 * @cbcpumask: cpumask that will be used for padata serialization
1009 *
1010 * Must be called from a cpus_read_lock() protected region
1011 */
1012static struct padata_instance *padata_alloc(struct workqueue_struct *wq,
1013 const struct cpumask *pcpumask,
1014 const struct cpumask *cbcpumask)
1015{
1016 struct padata_instance *pinst;
1017 struct parallel_data *pd = NULL;
1018
1019 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
1020 if (!pinst)
1021 goto err;
1022
1023 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
1024 goto err_free_inst;
1025 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
1026 free_cpumask_var(pinst->cpumask.pcpu);
1027 goto err_free_inst;
1028 }
1029 if (!padata_validate_cpumask(pinst, pcpumask) ||
1030 !padata_validate_cpumask(pinst, cbcpumask))
1031 goto err_free_masks;
1032
1033 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
1034 if (!pd)
1035 goto err_free_masks;
1036
1037 rcu_assign_pointer(pinst->pd, pd);
1038
1039 pinst->wq = wq;
1040
1041 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
1042 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
1043
1044 pinst->flags = 0;
1045
1046 BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
1047 kobject_init(&pinst->kobj, &padata_attr_type);
1048 mutex_init(&pinst->lock);
1049
1050#ifdef CONFIG_HOTPLUG_CPU
1051 cpuhp_state_add_instance_nocalls_cpuslocked(hp_online, &pinst->node);
1052#endif
1053 return pinst;
1054
1055err_free_masks:
1056 free_cpumask_var(pinst->cpumask.pcpu);
1057 free_cpumask_var(pinst->cpumask.cbcpu);
1058err_free_inst:
1059 kfree(pinst);
1060err:
1061 return NULL;
1062}
1063
1064/**
1065 * padata_alloc_possible - Allocate and initialize padata instance.
1066 * Use the cpu_possible_mask for serial and
1067 * parallel workers.
1068 *
1069 * @wq: workqueue to use for the allocated padata instance
1070 *
1071 * Must be called from a cpus_read_lock() protected region
1072 */
1073struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq)
1074{
1075 lockdep_assert_cpus_held();
1076 return padata_alloc(wq, cpu_possible_mask, cpu_possible_mask);
1077}
1078EXPORT_SYMBOL(padata_alloc_possible);
1079
1080/**
1081 * padata_free - free a padata instance
1082 *
1083 * @padata_inst: padata instance to free
1084 */
1085void padata_free(struct padata_instance *pinst)
1086{
1087 kobject_put(&pinst->kobj);
1088}
1089EXPORT_SYMBOL(padata_free);
1090
1091#ifdef CONFIG_HOTPLUG_CPU
1092
1093static __init int padata_driver_init(void)
1094{
1095 int ret;
1096
1097 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1098 padata_cpu_online,
1099 padata_cpu_prep_down);
1100 if (ret < 0)
1101 return ret;
1102 hp_online = ret;
1103 return 0;
1104}
1105module_init(padata_driver_init);
1106
1107static __exit void padata_driver_exit(void)
1108{
1109 cpuhp_remove_multi_state(hp_online);
1110}
1111module_exit(padata_driver_exit);
1112#endif