Loading...
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
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