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, unsigned int seq_nr)
50{
51 /*
52 * Hash the sequence numbers to the cpus by taking
53 * seq_nr mod. number of cpus in use.
54 */
55 int cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
56
57 return padata_index_to_cpu(pd, cpu_index);
58}
59
60static void padata_parallel_worker(struct work_struct *parallel_work)
61{
62 struct padata_parallel_queue *pqueue;
63 LIST_HEAD(local_list);
64
65 local_bh_disable();
66 pqueue = container_of(parallel_work,
67 struct padata_parallel_queue, work);
68
69 spin_lock(&pqueue->parallel.lock);
70 list_replace_init(&pqueue->parallel.list, &local_list);
71 spin_unlock(&pqueue->parallel.lock);
72
73 while (!list_empty(&local_list)) {
74 struct padata_priv *padata;
75
76 padata = list_entry(local_list.next,
77 struct padata_priv, list);
78
79 list_del_init(&padata->list);
80
81 padata->parallel(padata);
82 }
83
84 local_bh_enable();
85}
86
87/**
88 * padata_do_parallel - padata parallelization function
89 *
90 * @pinst: padata instance
91 * @padata: object to be parallelized
92 * @cb_cpu: pointer to the CPU that the serialization callback function should
93 * run on. If it's not in the serial cpumask of @pinst
94 * (i.e. cpumask.cbcpu), this function selects a fallback CPU and if
95 * none found, returns -EINVAL.
96 *
97 * The parallelization callback function will run with BHs off.
98 * Note: Every object which is parallelized by padata_do_parallel
99 * must be seen by padata_do_serial.
100 */
101int padata_do_parallel(struct padata_instance *pinst,
102 struct padata_priv *padata, int *cb_cpu)
103{
104 int i, cpu, cpu_index, target_cpu, err;
105 struct padata_parallel_queue *queue;
106 struct parallel_data *pd;
107
108 rcu_read_lock_bh();
109
110 pd = rcu_dereference_bh(pinst->pd);
111
112 err = -EINVAL;
113 if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
114 goto out;
115
116 if (!cpumask_test_cpu(*cb_cpu, pd->cpumask.cbcpu)) {
117 if (!cpumask_weight(pd->cpumask.cbcpu))
118 goto out;
119
120 /* Select an alternate fallback CPU and notify the caller. */
121 cpu_index = *cb_cpu % cpumask_weight(pd->cpumask.cbcpu);
122
123 cpu = cpumask_first(pd->cpumask.cbcpu);
124 for (i = 0; i < cpu_index; i++)
125 cpu = cpumask_next(cpu, pd->cpumask.cbcpu);
126
127 *cb_cpu = cpu;
128 }
129
130 err = -EBUSY;
131 if ((pinst->flags & PADATA_RESET))
132 goto out;
133
134 if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
135 goto out;
136
137 err = 0;
138 atomic_inc(&pd->refcnt);
139 padata->pd = pd;
140 padata->cb_cpu = *cb_cpu;
141
142 padata->seq_nr = atomic_inc_return(&pd->seq_nr);
143 target_cpu = padata_cpu_hash(pd, padata->seq_nr);
144 padata->cpu = target_cpu;
145 queue = per_cpu_ptr(pd->pqueue, target_cpu);
146
147 spin_lock(&queue->parallel.lock);
148 list_add_tail(&padata->list, &queue->parallel.list);
149 spin_unlock(&queue->parallel.lock);
150
151 queue_work(pinst->parallel_wq, &queue->work);
152
153out:
154 rcu_read_unlock_bh();
155
156 return err;
157}
158EXPORT_SYMBOL(padata_do_parallel);
159
160/*
161 * padata_find_next - Find the next object that needs serialization.
162 *
163 * Return values are:
164 *
165 * A pointer to the control struct of the next object that needs
166 * serialization, if present in one of the percpu reorder queues.
167 *
168 * NULL, if the next object that needs serialization will
169 * be parallel processed by another cpu and is not yet present in
170 * the cpu's reorder queue.
171 */
172static struct padata_priv *padata_find_next(struct parallel_data *pd,
173 bool remove_object)
174{
175 struct padata_parallel_queue *next_queue;
176 struct padata_priv *padata;
177 struct padata_list *reorder;
178 int cpu = pd->cpu;
179
180 next_queue = per_cpu_ptr(pd->pqueue, cpu);
181 reorder = &next_queue->reorder;
182
183 spin_lock(&reorder->lock);
184 if (list_empty(&reorder->list)) {
185 spin_unlock(&reorder->lock);
186 return NULL;
187 }
188
189 padata = list_entry(reorder->list.next, struct padata_priv, list);
190
191 /*
192 * Checks the rare case where two or more parallel jobs have hashed to
193 * the same CPU and one of the later ones finishes first.
194 */
195 if (padata->seq_nr != pd->processed) {
196 spin_unlock(&reorder->lock);
197 return NULL;
198 }
199
200 if (remove_object) {
201 list_del_init(&padata->list);
202 atomic_dec(&pd->reorder_objects);
203 ++pd->processed;
204 pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, -1, false);
205 }
206
207 spin_unlock(&reorder->lock);
208 return padata;
209}
210
211static void padata_reorder(struct parallel_data *pd)
212{
213 int cb_cpu;
214 struct padata_priv *padata;
215 struct padata_serial_queue *squeue;
216 struct padata_instance *pinst = pd->pinst;
217 struct padata_parallel_queue *next_queue;
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_find_next(pd, true);
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 (!padata)
241 break;
242
243 cb_cpu = padata->cb_cpu;
244 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
245
246 spin_lock(&squeue->serial.lock);
247 list_add_tail(&padata->list, &squeue->serial.list);
248 spin_unlock(&squeue->serial.lock);
249
250 queue_work_on(cb_cpu, pinst->serial_wq, &squeue->work);
251 }
252
253 spin_unlock_bh(&pd->lock);
254
255 /*
256 * The next object that needs serialization might have arrived to
257 * the reorder queues in the meantime.
258 *
259 * Ensure reorder queue is read after pd->lock is dropped so we see
260 * new objects from another task in padata_do_serial. Pairs with
261 * smp_mb__after_atomic in padata_do_serial.
262 */
263 smp_mb();
264
265 next_queue = per_cpu_ptr(pd->pqueue, pd->cpu);
266 if (!list_empty(&next_queue->reorder.list) &&
267 padata_find_next(pd, false))
268 queue_work(pinst->serial_wq, &pd->reorder_work);
269}
270
271static void invoke_padata_reorder(struct work_struct *work)
272{
273 struct parallel_data *pd;
274
275 local_bh_disable();
276 pd = container_of(work, struct parallel_data, reorder_work);
277 padata_reorder(pd);
278 local_bh_enable();
279}
280
281static void padata_serial_worker(struct work_struct *serial_work)
282{
283 struct padata_serial_queue *squeue;
284 struct parallel_data *pd;
285 LIST_HEAD(local_list);
286
287 local_bh_disable();
288 squeue = container_of(serial_work, struct padata_serial_queue, work);
289 pd = squeue->pd;
290
291 spin_lock(&squeue->serial.lock);
292 list_replace_init(&squeue->serial.list, &local_list);
293 spin_unlock(&squeue->serial.lock);
294
295 while (!list_empty(&local_list)) {
296 struct padata_priv *padata;
297
298 padata = list_entry(local_list.next,
299 struct padata_priv, list);
300
301 list_del_init(&padata->list);
302
303 padata->serial(padata);
304 atomic_dec(&pd->refcnt);
305 }
306 local_bh_enable();
307}
308
309/**
310 * padata_do_serial - padata serialization function
311 *
312 * @padata: object to be serialized.
313 *
314 * padata_do_serial must be called for every parallelized object.
315 * The serialization callback function will run with BHs off.
316 */
317void padata_do_serial(struct padata_priv *padata)
318{
319 struct parallel_data *pd = padata->pd;
320 struct padata_parallel_queue *pqueue = per_cpu_ptr(pd->pqueue,
321 padata->cpu);
322 struct padata_priv *cur;
323
324 spin_lock(&pqueue->reorder.lock);
325 /* Sort in ascending order of sequence number. */
326 list_for_each_entry_reverse(cur, &pqueue->reorder.list, list)
327 if (cur->seq_nr < padata->seq_nr)
328 break;
329 list_add(&padata->list, &cur->list);
330 atomic_inc(&pd->reorder_objects);
331 spin_unlock(&pqueue->reorder.lock);
332
333 /*
334 * Ensure the addition to the reorder list is ordered correctly
335 * with the trylock of pd->lock in padata_reorder. Pairs with smp_mb
336 * in padata_reorder.
337 */
338 smp_mb__after_atomic();
339
340 padata_reorder(pd);
341}
342EXPORT_SYMBOL(padata_do_serial);
343
344static int padata_setup_cpumasks(struct parallel_data *pd,
345 const struct cpumask *pcpumask,
346 const struct cpumask *cbcpumask)
347{
348 struct workqueue_attrs *attrs;
349 int err = -ENOMEM;
350
351 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
352 goto out;
353 cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_online_mask);
354
355 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
356 goto free_pcpu_mask;
357 cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_online_mask);
358
359 attrs = alloc_workqueue_attrs();
360 if (!attrs)
361 goto free_cbcpu_mask;
362
363 /* Restrict parallel_wq workers to pd->cpumask.pcpu. */
364 cpumask_copy(attrs->cpumask, pd->cpumask.pcpu);
365 err = apply_workqueue_attrs(pd->pinst->parallel_wq, attrs);
366 free_workqueue_attrs(attrs);
367 if (err < 0)
368 goto free_cbcpu_mask;
369
370 return 0;
371
372free_cbcpu_mask:
373 free_cpumask_var(pd->cpumask.cbcpu);
374free_pcpu_mask:
375 free_cpumask_var(pd->cpumask.pcpu);
376out:
377 return err;
378}
379
380static void __padata_list_init(struct padata_list *pd_list)
381{
382 INIT_LIST_HEAD(&pd_list->list);
383 spin_lock_init(&pd_list->lock);
384}
385
386/* Initialize all percpu queues used by serial workers */
387static void padata_init_squeues(struct parallel_data *pd)
388{
389 int cpu;
390 struct padata_serial_queue *squeue;
391
392 for_each_cpu(cpu, pd->cpumask.cbcpu) {
393 squeue = per_cpu_ptr(pd->squeue, cpu);
394 squeue->pd = pd;
395 __padata_list_init(&squeue->serial);
396 INIT_WORK(&squeue->work, padata_serial_worker);
397 }
398}
399
400/* Initialize all percpu queues used by parallel workers */
401static void padata_init_pqueues(struct parallel_data *pd)
402{
403 int cpu;
404 struct padata_parallel_queue *pqueue;
405
406 for_each_cpu(cpu, pd->cpumask.pcpu) {
407 pqueue = per_cpu_ptr(pd->pqueue, cpu);
408
409 __padata_list_init(&pqueue->reorder);
410 __padata_list_init(&pqueue->parallel);
411 INIT_WORK(&pqueue->work, padata_parallel_worker);
412 atomic_set(&pqueue->num_obj, 0);
413 }
414}
415
416/* Allocate and initialize the internal cpumask dependend resources. */
417static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
418 const struct cpumask *pcpumask,
419 const struct cpumask *cbcpumask)
420{
421 struct parallel_data *pd;
422
423 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
424 if (!pd)
425 goto err;
426
427 pd->pqueue = alloc_percpu(struct padata_parallel_queue);
428 if (!pd->pqueue)
429 goto err_free_pd;
430
431 pd->squeue = alloc_percpu(struct padata_serial_queue);
432 if (!pd->squeue)
433 goto err_free_pqueue;
434
435 pd->pinst = pinst;
436 if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
437 goto err_free_squeue;
438
439 padata_init_pqueues(pd);
440 padata_init_squeues(pd);
441 atomic_set(&pd->seq_nr, -1);
442 atomic_set(&pd->reorder_objects, 0);
443 atomic_set(&pd->refcnt, 0);
444 spin_lock_init(&pd->lock);
445 pd->cpu = cpumask_first(pd->cpumask.pcpu);
446 INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
447
448 return pd;
449
450err_free_squeue:
451 free_percpu(pd->squeue);
452err_free_pqueue:
453 free_percpu(pd->pqueue);
454err_free_pd:
455 kfree(pd);
456err:
457 return NULL;
458}
459
460static void padata_free_pd(struct parallel_data *pd)
461{
462 free_cpumask_var(pd->cpumask.pcpu);
463 free_cpumask_var(pd->cpumask.cbcpu);
464 free_percpu(pd->pqueue);
465 free_percpu(pd->squeue);
466 kfree(pd);
467}
468
469/* Flush all objects out of the padata queues. */
470static void padata_flush_queues(struct parallel_data *pd)
471{
472 int cpu;
473 struct padata_parallel_queue *pqueue;
474 struct padata_serial_queue *squeue;
475
476 for_each_cpu(cpu, pd->cpumask.pcpu) {
477 pqueue = per_cpu_ptr(pd->pqueue, cpu);
478 flush_work(&pqueue->work);
479 }
480
481 if (atomic_read(&pd->reorder_objects))
482 padata_reorder(pd);
483
484 for_each_cpu(cpu, pd->cpumask.cbcpu) {
485 squeue = per_cpu_ptr(pd->squeue, cpu);
486 flush_work(&squeue->work);
487 }
488
489 BUG_ON(atomic_read(&pd->refcnt) != 0);
490}
491
492static void __padata_start(struct padata_instance *pinst)
493{
494 pinst->flags |= PADATA_INIT;
495}
496
497static void __padata_stop(struct padata_instance *pinst)
498{
499 if (!(pinst->flags & PADATA_INIT))
500 return;
501
502 pinst->flags &= ~PADATA_INIT;
503
504 synchronize_rcu();
505
506 get_online_cpus();
507 padata_flush_queues(pinst->pd);
508 put_online_cpus();
509}
510
511/* Replace the internal control structure with a new one. */
512static void padata_replace(struct padata_instance *pinst,
513 struct parallel_data *pd_new)
514{
515 struct parallel_data *pd_old = pinst->pd;
516 int notification_mask = 0;
517
518 pinst->flags |= PADATA_RESET;
519
520 rcu_assign_pointer(pinst->pd, pd_new);
521
522 synchronize_rcu();
523
524 if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
525 notification_mask |= PADATA_CPU_PARALLEL;
526 if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
527 notification_mask |= PADATA_CPU_SERIAL;
528
529 padata_flush_queues(pd_old);
530 padata_free_pd(pd_old);
531
532 if (notification_mask)
533 blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
534 notification_mask,
535 &pd_new->cpumask);
536
537 pinst->flags &= ~PADATA_RESET;
538}
539
540/**
541 * padata_register_cpumask_notifier - Registers a notifier that will be called
542 * if either pcpu or cbcpu or both cpumasks change.
543 *
544 * @pinst: A poineter to padata instance
545 * @nblock: A pointer to notifier block.
546 */
547int padata_register_cpumask_notifier(struct padata_instance *pinst,
548 struct notifier_block *nblock)
549{
550 return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
551 nblock);
552}
553EXPORT_SYMBOL(padata_register_cpumask_notifier);
554
555/**
556 * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
557 * registered earlier using padata_register_cpumask_notifier
558 *
559 * @pinst: A pointer to data instance.
560 * @nlock: A pointer to notifier block.
561 */
562int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
563 struct notifier_block *nblock)
564{
565 return blocking_notifier_chain_unregister(
566 &pinst->cpumask_change_notifier,
567 nblock);
568}
569EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
570
571
572/* If cpumask contains no active cpu, we mark the instance as invalid. */
573static bool padata_validate_cpumask(struct padata_instance *pinst,
574 const struct cpumask *cpumask)
575{
576 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
577 pinst->flags |= PADATA_INVALID;
578 return false;
579 }
580
581 pinst->flags &= ~PADATA_INVALID;
582 return true;
583}
584
585static int __padata_set_cpumasks(struct padata_instance *pinst,
586 cpumask_var_t pcpumask,
587 cpumask_var_t cbcpumask)
588{
589 int valid;
590 struct parallel_data *pd;
591
592 valid = padata_validate_cpumask(pinst, pcpumask);
593 if (!valid) {
594 __padata_stop(pinst);
595 goto out_replace;
596 }
597
598 valid = padata_validate_cpumask(pinst, cbcpumask);
599 if (!valid)
600 __padata_stop(pinst);
601
602out_replace:
603 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
604 if (!pd)
605 return -ENOMEM;
606
607 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
608 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
609
610 padata_replace(pinst, pd);
611
612 if (valid)
613 __padata_start(pinst);
614
615 return 0;
616}
617
618/**
619 * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
620 * equivalent to @cpumask.
621 *
622 * @pinst: padata instance
623 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
624 * to parallel and serial cpumasks respectively.
625 * @cpumask: the cpumask to use
626 */
627int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
628 cpumask_var_t cpumask)
629{
630 struct cpumask *serial_mask, *parallel_mask;
631 int err = -EINVAL;
632
633 mutex_lock(&pinst->lock);
634 get_online_cpus();
635
636 switch (cpumask_type) {
637 case PADATA_CPU_PARALLEL:
638 serial_mask = pinst->cpumask.cbcpu;
639 parallel_mask = cpumask;
640 break;
641 case PADATA_CPU_SERIAL:
642 parallel_mask = pinst->cpumask.pcpu;
643 serial_mask = cpumask;
644 break;
645 default:
646 goto out;
647 }
648
649 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
650
651out:
652 put_online_cpus();
653 mutex_unlock(&pinst->lock);
654
655 return err;
656}
657EXPORT_SYMBOL(padata_set_cpumask);
658
659/**
660 * padata_start - start the parallel processing
661 *
662 * @pinst: padata instance to start
663 */
664int padata_start(struct padata_instance *pinst)
665{
666 int err = 0;
667
668 mutex_lock(&pinst->lock);
669
670 if (pinst->flags & PADATA_INVALID)
671 err = -EINVAL;
672
673 __padata_start(pinst);
674
675 mutex_unlock(&pinst->lock);
676
677 return err;
678}
679EXPORT_SYMBOL(padata_start);
680
681/**
682 * padata_stop - stop the parallel processing
683 *
684 * @pinst: padata instance to stop
685 */
686void padata_stop(struct padata_instance *pinst)
687{
688 mutex_lock(&pinst->lock);
689 __padata_stop(pinst);
690 mutex_unlock(&pinst->lock);
691}
692EXPORT_SYMBOL(padata_stop);
693
694#ifdef CONFIG_HOTPLUG_CPU
695
696static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
697{
698 struct parallel_data *pd;
699
700 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
701 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
702 pinst->cpumask.cbcpu);
703 if (!pd)
704 return -ENOMEM;
705
706 padata_replace(pinst, pd);
707
708 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
709 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
710 __padata_start(pinst);
711 }
712
713 return 0;
714}
715
716static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
717{
718 struct parallel_data *pd = NULL;
719
720 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
721
722 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
723 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
724 __padata_stop(pinst);
725
726 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
727 pinst->cpumask.cbcpu);
728 if (!pd)
729 return -ENOMEM;
730
731 padata_replace(pinst, pd);
732
733 cpumask_clear_cpu(cpu, pd->cpumask.cbcpu);
734 cpumask_clear_cpu(cpu, pd->cpumask.pcpu);
735 }
736
737 return 0;
738}
739
740 /**
741 * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
742 * padata cpumasks.
743 *
744 * @pinst: padata instance
745 * @cpu: cpu to remove
746 * @mask: bitmask specifying from which cpumask @cpu should be removed
747 * The @mask may be any combination of the following flags:
748 * PADATA_CPU_SERIAL - serial cpumask
749 * PADATA_CPU_PARALLEL - parallel cpumask
750 */
751int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
752{
753 int err;
754
755 if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
756 return -EINVAL;
757
758 mutex_lock(&pinst->lock);
759
760 get_online_cpus();
761 if (mask & PADATA_CPU_SERIAL)
762 cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
763 if (mask & PADATA_CPU_PARALLEL)
764 cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
765
766 err = __padata_remove_cpu(pinst, cpu);
767 put_online_cpus();
768
769 mutex_unlock(&pinst->lock);
770
771 return err;
772}
773EXPORT_SYMBOL(padata_remove_cpu);
774
775static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
776{
777 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
778 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
779}
780
781static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
782{
783 struct padata_instance *pinst;
784 int ret;
785
786 pinst = hlist_entry_safe(node, struct padata_instance, node);
787 if (!pinst_has_cpu(pinst, cpu))
788 return 0;
789
790 mutex_lock(&pinst->lock);
791 ret = __padata_add_cpu(pinst, cpu);
792 mutex_unlock(&pinst->lock);
793 return ret;
794}
795
796static int padata_cpu_prep_down(unsigned int cpu, struct hlist_node *node)
797{
798 struct padata_instance *pinst;
799 int ret;
800
801 pinst = hlist_entry_safe(node, struct padata_instance, node);
802 if (!pinst_has_cpu(pinst, cpu))
803 return 0;
804
805 mutex_lock(&pinst->lock);
806 ret = __padata_remove_cpu(pinst, cpu);
807 mutex_unlock(&pinst->lock);
808 return ret;
809}
810
811static enum cpuhp_state hp_online;
812#endif
813
814static void __padata_free(struct padata_instance *pinst)
815{
816#ifdef CONFIG_HOTPLUG_CPU
817 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->node);
818#endif
819
820 padata_stop(pinst);
821 padata_free_pd(pinst->pd);
822 free_cpumask_var(pinst->cpumask.pcpu);
823 free_cpumask_var(pinst->cpumask.cbcpu);
824 destroy_workqueue(pinst->serial_wq);
825 destroy_workqueue(pinst->parallel_wq);
826 kfree(pinst);
827}
828
829#define kobj2pinst(_kobj) \
830 container_of(_kobj, struct padata_instance, kobj)
831#define attr2pentry(_attr) \
832 container_of(_attr, struct padata_sysfs_entry, attr)
833
834static void padata_sysfs_release(struct kobject *kobj)
835{
836 struct padata_instance *pinst = kobj2pinst(kobj);
837 __padata_free(pinst);
838}
839
840struct padata_sysfs_entry {
841 struct attribute attr;
842 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
843 ssize_t (*store)(struct padata_instance *, struct attribute *,
844 const char *, size_t);
845};
846
847static ssize_t show_cpumask(struct padata_instance *pinst,
848 struct attribute *attr, char *buf)
849{
850 struct cpumask *cpumask;
851 ssize_t len;
852
853 mutex_lock(&pinst->lock);
854 if (!strcmp(attr->name, "serial_cpumask"))
855 cpumask = pinst->cpumask.cbcpu;
856 else
857 cpumask = pinst->cpumask.pcpu;
858
859 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
860 nr_cpu_ids, cpumask_bits(cpumask));
861 mutex_unlock(&pinst->lock);
862 return len < PAGE_SIZE ? len : -EINVAL;
863}
864
865static ssize_t store_cpumask(struct padata_instance *pinst,
866 struct attribute *attr,
867 const char *buf, size_t count)
868{
869 cpumask_var_t new_cpumask;
870 ssize_t ret;
871 int mask_type;
872
873 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
874 return -ENOMEM;
875
876 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
877 nr_cpumask_bits);
878 if (ret < 0)
879 goto out;
880
881 mask_type = !strcmp(attr->name, "serial_cpumask") ?
882 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
883 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
884 if (!ret)
885 ret = count;
886
887out:
888 free_cpumask_var(new_cpumask);
889 return ret;
890}
891
892#define PADATA_ATTR_RW(_name, _show_name, _store_name) \
893 static struct padata_sysfs_entry _name##_attr = \
894 __ATTR(_name, 0644, _show_name, _store_name)
895#define PADATA_ATTR_RO(_name, _show_name) \
896 static struct padata_sysfs_entry _name##_attr = \
897 __ATTR(_name, 0400, _show_name, NULL)
898
899PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
900PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
901
902/*
903 * Padata sysfs provides the following objects:
904 * serial_cpumask [RW] - cpumask for serial workers
905 * parallel_cpumask [RW] - cpumask for parallel workers
906 */
907static struct attribute *padata_default_attrs[] = {
908 &serial_cpumask_attr.attr,
909 ¶llel_cpumask_attr.attr,
910 NULL,
911};
912ATTRIBUTE_GROUPS(padata_default);
913
914static ssize_t padata_sysfs_show(struct kobject *kobj,
915 struct attribute *attr, char *buf)
916{
917 struct padata_instance *pinst;
918 struct padata_sysfs_entry *pentry;
919 ssize_t ret = -EIO;
920
921 pinst = kobj2pinst(kobj);
922 pentry = attr2pentry(attr);
923 if (pentry->show)
924 ret = pentry->show(pinst, attr, buf);
925
926 return ret;
927}
928
929static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
930 const char *buf, size_t count)
931{
932 struct padata_instance *pinst;
933 struct padata_sysfs_entry *pentry;
934 ssize_t ret = -EIO;
935
936 pinst = kobj2pinst(kobj);
937 pentry = attr2pentry(attr);
938 if (pentry->show)
939 ret = pentry->store(pinst, attr, buf, count);
940
941 return ret;
942}
943
944static const struct sysfs_ops padata_sysfs_ops = {
945 .show = padata_sysfs_show,
946 .store = padata_sysfs_store,
947};
948
949static struct kobj_type padata_attr_type = {
950 .sysfs_ops = &padata_sysfs_ops,
951 .default_groups = padata_default_groups,
952 .release = padata_sysfs_release,
953};
954
955/**
956 * padata_alloc - allocate and initialize a padata instance and specify
957 * cpumasks for serial and parallel workers.
958 *
959 * @name: used to identify the instance
960 * @pcpumask: cpumask that will be used for padata parallelization
961 * @cbcpumask: cpumask that will be used for padata serialization
962 */
963static struct padata_instance *padata_alloc(const char *name,
964 const struct cpumask *pcpumask,
965 const struct cpumask *cbcpumask)
966{
967 struct padata_instance *pinst;
968 struct parallel_data *pd = NULL;
969
970 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
971 if (!pinst)
972 goto err;
973
974 pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
975 name);
976 if (!pinst->parallel_wq)
977 goto err_free_inst;
978
979 get_online_cpus();
980
981 pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
982 WQ_CPU_INTENSIVE, 1, name);
983 if (!pinst->serial_wq)
984 goto err_put_cpus;
985
986 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
987 goto err_free_serial_wq;
988 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
989 free_cpumask_var(pinst->cpumask.pcpu);
990 goto err_free_serial_wq;
991 }
992 if (!padata_validate_cpumask(pinst, pcpumask) ||
993 !padata_validate_cpumask(pinst, cbcpumask))
994 goto err_free_masks;
995
996 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
997 if (!pd)
998 goto err_free_masks;
999
1000 rcu_assign_pointer(pinst->pd, pd);
1001
1002 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
1003 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
1004
1005 pinst->flags = 0;
1006
1007 BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
1008 kobject_init(&pinst->kobj, &padata_attr_type);
1009 mutex_init(&pinst->lock);
1010
1011#ifdef CONFIG_HOTPLUG_CPU
1012 cpuhp_state_add_instance_nocalls_cpuslocked(hp_online, &pinst->node);
1013#endif
1014
1015 put_online_cpus();
1016
1017 return pinst;
1018
1019err_free_masks:
1020 free_cpumask_var(pinst->cpumask.pcpu);
1021 free_cpumask_var(pinst->cpumask.cbcpu);
1022err_free_serial_wq:
1023 destroy_workqueue(pinst->serial_wq);
1024err_put_cpus:
1025 put_online_cpus();
1026 destroy_workqueue(pinst->parallel_wq);
1027err_free_inst:
1028 kfree(pinst);
1029err:
1030 return NULL;
1031}
1032
1033/**
1034 * padata_alloc_possible - Allocate and initialize padata instance.
1035 * Use the cpu_possible_mask for serial and
1036 * parallel workers.
1037 *
1038 * @name: used to identify the instance
1039 */
1040struct padata_instance *padata_alloc_possible(const char *name)
1041{
1042 return padata_alloc(name, cpu_possible_mask, cpu_possible_mask);
1043}
1044EXPORT_SYMBOL(padata_alloc_possible);
1045
1046/**
1047 * padata_free - free a padata instance
1048 *
1049 * @padata_inst: padata instance to free
1050 */
1051void padata_free(struct padata_instance *pinst)
1052{
1053 kobject_put(&pinst->kobj);
1054}
1055EXPORT_SYMBOL(padata_free);
1056
1057#ifdef CONFIG_HOTPLUG_CPU
1058
1059static __init int padata_driver_init(void)
1060{
1061 int ret;
1062
1063 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1064 padata_cpu_online,
1065 padata_cpu_prep_down);
1066 if (ret < 0)
1067 return ret;
1068 hp_online = ret;
1069 return 0;
1070}
1071module_init(padata_driver_init);
1072
1073static __exit void padata_driver_exit(void)
1074{
1075 cpuhp_remove_multi_state(hp_online);
1076}
1077module_exit(padata_driver_exit);
1078#endif