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