Loading...
1/*
2 * padata.c - generic interface to process data streams in parallel
3 *
4 * See Documentation/padata.txt for an api documentation.
5 *
6 * Copyright (C) 2008, 2009 secunet Security Networks AG
7 * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
8 *
9 * This program is free software; you can redistribute it and/or modify it
10 * under the terms and conditions of the GNU General Public License,
11 * version 2, as published by the Free Software Foundation.
12 *
13 * This program is distributed in the hope it will be useful, but WITHOUT
14 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
15 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
16 * more details.
17 *
18 * You should have received a copy of the GNU General Public License along with
19 * this program; if not, write to the Free Software Foundation, Inc.,
20 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
21 */
22
23#include <linux/export.h>
24#include <linux/cpumask.h>
25#include <linux/err.h>
26#include <linux/cpu.h>
27#include <linux/padata.h>
28#include <linux/mutex.h>
29#include <linux/sched.h>
30#include <linux/slab.h>
31#include <linux/sysfs.h>
32#include <linux/rcupdate.h>
33#include <linux/module.h>
34
35#define MAX_OBJ_NUM 1000
36
37static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
38{
39 int cpu, target_cpu;
40
41 target_cpu = cpumask_first(pd->cpumask.pcpu);
42 for (cpu = 0; cpu < cpu_index; cpu++)
43 target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
44
45 return target_cpu;
46}
47
48static int padata_cpu_hash(struct parallel_data *pd)
49{
50 unsigned int seq_nr;
51 int cpu_index;
52
53 /*
54 * Hash the sequence numbers to the cpus by taking
55 * seq_nr mod. number of cpus in use.
56 */
57
58 seq_nr = atomic_inc_return(&pd->seq_nr);
59 cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
60
61 return padata_index_to_cpu(pd, cpu_index);
62}
63
64static void padata_parallel_worker(struct work_struct *parallel_work)
65{
66 struct padata_parallel_queue *pqueue;
67 LIST_HEAD(local_list);
68
69 local_bh_disable();
70 pqueue = container_of(parallel_work,
71 struct padata_parallel_queue, work);
72
73 spin_lock(&pqueue->parallel.lock);
74 list_replace_init(&pqueue->parallel.list, &local_list);
75 spin_unlock(&pqueue->parallel.lock);
76
77 while (!list_empty(&local_list)) {
78 struct padata_priv *padata;
79
80 padata = list_entry(local_list.next,
81 struct padata_priv, list);
82
83 list_del_init(&padata->list);
84
85 padata->parallel(padata);
86 }
87
88 local_bh_enable();
89}
90
91/**
92 * padata_do_parallel - padata parallelization function
93 *
94 * @pinst: padata instance
95 * @padata: object to be parallelized
96 * @cb_cpu: cpu the serialization callback function will run on,
97 * must be in the serial cpumask of padata(i.e. cpumask.cbcpu).
98 *
99 * The parallelization callback function will run with BHs off.
100 * Note: Every object which is parallelized by padata_do_parallel
101 * must be seen by padata_do_serial.
102 */
103int padata_do_parallel(struct padata_instance *pinst,
104 struct padata_priv *padata, int cb_cpu)
105{
106 int target_cpu, err;
107 struct padata_parallel_queue *queue;
108 struct parallel_data *pd;
109
110 rcu_read_lock_bh();
111
112 pd = rcu_dereference_bh(pinst->pd);
113
114 err = -EINVAL;
115 if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
116 goto out;
117
118 if (!cpumask_test_cpu(cb_cpu, pd->cpumask.cbcpu))
119 goto out;
120
121 err = -EBUSY;
122 if ((pinst->flags & PADATA_RESET))
123 goto out;
124
125 if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
126 goto out;
127
128 err = 0;
129 atomic_inc(&pd->refcnt);
130 padata->pd = pd;
131 padata->cb_cpu = cb_cpu;
132
133 target_cpu = padata_cpu_hash(pd);
134 queue = per_cpu_ptr(pd->pqueue, target_cpu);
135
136 spin_lock(&queue->parallel.lock);
137 list_add_tail(&padata->list, &queue->parallel.list);
138 spin_unlock(&queue->parallel.lock);
139
140 queue_work_on(target_cpu, pinst->wq, &queue->work);
141
142out:
143 rcu_read_unlock_bh();
144
145 return err;
146}
147EXPORT_SYMBOL(padata_do_parallel);
148
149/*
150 * padata_get_next - Get the next object that needs serialization.
151 *
152 * Return values are:
153 *
154 * A pointer to the control struct of the next object that needs
155 * serialization, if present in one of the percpu reorder queues.
156 *
157 * NULL, if all percpu reorder queues are empty.
158 *
159 * -EINPROGRESS, if the next object that needs serialization will
160 * be parallel processed by another cpu and is not yet present in
161 * the cpu's reorder queue.
162 *
163 * -ENODATA, if this cpu has to do the parallel processing for
164 * the next object.
165 */
166static struct padata_priv *padata_get_next(struct parallel_data *pd)
167{
168 int cpu, num_cpus;
169 unsigned int next_nr, next_index;
170 struct padata_parallel_queue *next_queue;
171 struct padata_priv *padata;
172 struct padata_list *reorder;
173
174 num_cpus = cpumask_weight(pd->cpumask.pcpu);
175
176 /*
177 * Calculate the percpu reorder queue and the sequence
178 * number of the next object.
179 */
180 next_nr = pd->processed;
181 next_index = next_nr % num_cpus;
182 cpu = padata_index_to_cpu(pd, next_index);
183 next_queue = per_cpu_ptr(pd->pqueue, cpu);
184
185 padata = NULL;
186
187 reorder = &next_queue->reorder;
188
189 spin_lock(&reorder->lock);
190 if (!list_empty(&reorder->list)) {
191 padata = list_entry(reorder->list.next,
192 struct padata_priv, list);
193
194 list_del_init(&padata->list);
195 atomic_dec(&pd->reorder_objects);
196
197 pd->processed++;
198
199 spin_unlock(&reorder->lock);
200 goto out;
201 }
202 spin_unlock(&reorder->lock);
203
204 if (__this_cpu_read(pd->pqueue->cpu_index) == next_queue->cpu_index) {
205 padata = ERR_PTR(-ENODATA);
206 goto out;
207 }
208
209 padata = ERR_PTR(-EINPROGRESS);
210out:
211 return padata;
212}
213
214static void padata_reorder(struct parallel_data *pd)
215{
216 int cb_cpu;
217 struct padata_priv *padata;
218 struct padata_serial_queue *squeue;
219 struct padata_instance *pinst = pd->pinst;
220
221 /*
222 * We need to ensure that only one cpu can work on dequeueing of
223 * the reorder queue the time. Calculating in which percpu reorder
224 * queue the next object will arrive takes some time. A spinlock
225 * would be highly contended. Also it is not clear in which order
226 * the objects arrive to the reorder queues. So a cpu could wait to
227 * get the lock just to notice that there is nothing to do at the
228 * moment. Therefore we use a trylock and let the holder of the lock
229 * care for all the objects enqueued during the holdtime of the lock.
230 */
231 if (!spin_trylock_bh(&pd->lock))
232 return;
233
234 while (1) {
235 padata = padata_get_next(pd);
236
237 /*
238 * All reorder queues are empty, or the next object that needs
239 * serialization is parallel processed by another cpu and is
240 * still on it's way to the cpu's reorder queue, nothing to
241 * do for now.
242 */
243 if (!padata || PTR_ERR(padata) == -EINPROGRESS)
244 break;
245
246 /*
247 * This cpu has to do the parallel processing of the next
248 * object. It's waiting in the cpu's parallelization queue,
249 * so exit immediately.
250 */
251 if (PTR_ERR(padata) == -ENODATA) {
252 del_timer(&pd->timer);
253 spin_unlock_bh(&pd->lock);
254 return;
255 }
256
257 cb_cpu = padata->cb_cpu;
258 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
259
260 spin_lock(&squeue->serial.lock);
261 list_add_tail(&padata->list, &squeue->serial.list);
262 spin_unlock(&squeue->serial.lock);
263
264 queue_work_on(cb_cpu, pinst->wq, &squeue->work);
265 }
266
267 spin_unlock_bh(&pd->lock);
268
269 /*
270 * The next object that needs serialization might have arrived to
271 * the reorder queues in the meantime, we will be called again
272 * from the timer function if no one else cares for it.
273 */
274 if (atomic_read(&pd->reorder_objects)
275 && !(pinst->flags & PADATA_RESET))
276 mod_timer(&pd->timer, jiffies + HZ);
277 else
278 del_timer(&pd->timer);
279
280 return;
281}
282
283static void padata_reorder_timer(unsigned long arg)
284{
285 struct parallel_data *pd = (struct parallel_data *)arg;
286
287 padata_reorder(pd);
288}
289
290static void padata_serial_worker(struct work_struct *serial_work)
291{
292 struct padata_serial_queue *squeue;
293 struct parallel_data *pd;
294 LIST_HEAD(local_list);
295
296 local_bh_disable();
297 squeue = container_of(serial_work, struct padata_serial_queue, work);
298 pd = squeue->pd;
299
300 spin_lock(&squeue->serial.lock);
301 list_replace_init(&squeue->serial.list, &local_list);
302 spin_unlock(&squeue->serial.lock);
303
304 while (!list_empty(&local_list)) {
305 struct padata_priv *padata;
306
307 padata = list_entry(local_list.next,
308 struct padata_priv, list);
309
310 list_del_init(&padata->list);
311
312 padata->serial(padata);
313 atomic_dec(&pd->refcnt);
314 }
315 local_bh_enable();
316}
317
318/**
319 * padata_do_serial - padata serialization function
320 *
321 * @padata: object to be serialized.
322 *
323 * padata_do_serial must be called for every parallelized object.
324 * The serialization callback function will run with BHs off.
325 */
326void padata_do_serial(struct padata_priv *padata)
327{
328 int cpu;
329 struct padata_parallel_queue *pqueue;
330 struct parallel_data *pd;
331
332 pd = padata->pd;
333
334 cpu = get_cpu();
335 pqueue = per_cpu_ptr(pd->pqueue, cpu);
336
337 spin_lock(&pqueue->reorder.lock);
338 atomic_inc(&pd->reorder_objects);
339 list_add_tail(&padata->list, &pqueue->reorder.list);
340 spin_unlock(&pqueue->reorder.lock);
341
342 put_cpu();
343
344 padata_reorder(pd);
345}
346EXPORT_SYMBOL(padata_do_serial);
347
348static int padata_setup_cpumasks(struct parallel_data *pd,
349 const struct cpumask *pcpumask,
350 const struct cpumask *cbcpumask)
351{
352 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
353 return -ENOMEM;
354
355 cpumask_and(pd->cpumask.pcpu, pcpumask, cpu_online_mask);
356 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL)) {
357 free_cpumask_var(pd->cpumask.cbcpu);
358 return -ENOMEM;
359 }
360
361 cpumask_and(pd->cpumask.cbcpu, cbcpumask, cpu_online_mask);
362 return 0;
363}
364
365static void __padata_list_init(struct padata_list *pd_list)
366{
367 INIT_LIST_HEAD(&pd_list->list);
368 spin_lock_init(&pd_list->lock);
369}
370
371/* Initialize all percpu queues used by serial workers */
372static void padata_init_squeues(struct parallel_data *pd)
373{
374 int cpu;
375 struct padata_serial_queue *squeue;
376
377 for_each_cpu(cpu, pd->cpumask.cbcpu) {
378 squeue = per_cpu_ptr(pd->squeue, cpu);
379 squeue->pd = pd;
380 __padata_list_init(&squeue->serial);
381 INIT_WORK(&squeue->work, padata_serial_worker);
382 }
383}
384
385/* Initialize all percpu queues used by parallel workers */
386static void padata_init_pqueues(struct parallel_data *pd)
387{
388 int cpu_index, cpu;
389 struct padata_parallel_queue *pqueue;
390
391 cpu_index = 0;
392 for_each_cpu(cpu, pd->cpumask.pcpu) {
393 pqueue = per_cpu_ptr(pd->pqueue, cpu);
394 pqueue->pd = pd;
395 pqueue->cpu_index = cpu_index;
396 cpu_index++;
397
398 __padata_list_init(&pqueue->reorder);
399 __padata_list_init(&pqueue->parallel);
400 INIT_WORK(&pqueue->work, padata_parallel_worker);
401 atomic_set(&pqueue->num_obj, 0);
402 }
403}
404
405/* Allocate and initialize the internal cpumask dependend resources. */
406static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
407 const struct cpumask *pcpumask,
408 const struct cpumask *cbcpumask)
409{
410 struct parallel_data *pd;
411
412 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
413 if (!pd)
414 goto err;
415
416 pd->pqueue = alloc_percpu(struct padata_parallel_queue);
417 if (!pd->pqueue)
418 goto err_free_pd;
419
420 pd->squeue = alloc_percpu(struct padata_serial_queue);
421 if (!pd->squeue)
422 goto err_free_pqueue;
423 if (padata_setup_cpumasks(pd, pcpumask, cbcpumask) < 0)
424 goto err_free_squeue;
425
426 padata_init_pqueues(pd);
427 padata_init_squeues(pd);
428 setup_timer(&pd->timer, padata_reorder_timer, (unsigned long)pd);
429 atomic_set(&pd->seq_nr, -1);
430 atomic_set(&pd->reorder_objects, 0);
431 atomic_set(&pd->refcnt, 0);
432 pd->pinst = pinst;
433 spin_lock_init(&pd->lock);
434
435 return pd;
436
437err_free_squeue:
438 free_percpu(pd->squeue);
439err_free_pqueue:
440 free_percpu(pd->pqueue);
441err_free_pd:
442 kfree(pd);
443err:
444 return NULL;
445}
446
447static void padata_free_pd(struct parallel_data *pd)
448{
449 free_cpumask_var(pd->cpumask.pcpu);
450 free_cpumask_var(pd->cpumask.cbcpu);
451 free_percpu(pd->pqueue);
452 free_percpu(pd->squeue);
453 kfree(pd);
454}
455
456/* Flush all objects out of the padata queues. */
457static void padata_flush_queues(struct parallel_data *pd)
458{
459 int cpu;
460 struct padata_parallel_queue *pqueue;
461 struct padata_serial_queue *squeue;
462
463 for_each_cpu(cpu, pd->cpumask.pcpu) {
464 pqueue = per_cpu_ptr(pd->pqueue, cpu);
465 flush_work(&pqueue->work);
466 }
467
468 del_timer_sync(&pd->timer);
469
470 if (atomic_read(&pd->reorder_objects))
471 padata_reorder(pd);
472
473 for_each_cpu(cpu, pd->cpumask.cbcpu) {
474 squeue = per_cpu_ptr(pd->squeue, cpu);
475 flush_work(&squeue->work);
476 }
477
478 BUG_ON(atomic_read(&pd->refcnt) != 0);
479}
480
481static void __padata_start(struct padata_instance *pinst)
482{
483 pinst->flags |= PADATA_INIT;
484}
485
486static void __padata_stop(struct padata_instance *pinst)
487{
488 if (!(pinst->flags & PADATA_INIT))
489 return;
490
491 pinst->flags &= ~PADATA_INIT;
492
493 synchronize_rcu();
494
495 get_online_cpus();
496 padata_flush_queues(pinst->pd);
497 put_online_cpus();
498}
499
500/* Replace the internal control structure with a new one. */
501static void padata_replace(struct padata_instance *pinst,
502 struct parallel_data *pd_new)
503{
504 struct parallel_data *pd_old = pinst->pd;
505 int notification_mask = 0;
506
507 pinst->flags |= PADATA_RESET;
508
509 rcu_assign_pointer(pinst->pd, pd_new);
510
511 synchronize_rcu();
512
513 if (!cpumask_equal(pd_old->cpumask.pcpu, pd_new->cpumask.pcpu))
514 notification_mask |= PADATA_CPU_PARALLEL;
515 if (!cpumask_equal(pd_old->cpumask.cbcpu, pd_new->cpumask.cbcpu))
516 notification_mask |= PADATA_CPU_SERIAL;
517
518 padata_flush_queues(pd_old);
519 padata_free_pd(pd_old);
520
521 if (notification_mask)
522 blocking_notifier_call_chain(&pinst->cpumask_change_notifier,
523 notification_mask,
524 &pd_new->cpumask);
525
526 pinst->flags &= ~PADATA_RESET;
527}
528
529/**
530 * padata_register_cpumask_notifier - Registers a notifier that will be called
531 * if either pcpu or cbcpu or both cpumasks change.
532 *
533 * @pinst: A poineter to padata instance
534 * @nblock: A pointer to notifier block.
535 */
536int padata_register_cpumask_notifier(struct padata_instance *pinst,
537 struct notifier_block *nblock)
538{
539 return blocking_notifier_chain_register(&pinst->cpumask_change_notifier,
540 nblock);
541}
542EXPORT_SYMBOL(padata_register_cpumask_notifier);
543
544/**
545 * padata_unregister_cpumask_notifier - Unregisters cpumask notifier
546 * registered earlier using padata_register_cpumask_notifier
547 *
548 * @pinst: A pointer to data instance.
549 * @nlock: A pointer to notifier block.
550 */
551int padata_unregister_cpumask_notifier(struct padata_instance *pinst,
552 struct notifier_block *nblock)
553{
554 return blocking_notifier_chain_unregister(
555 &pinst->cpumask_change_notifier,
556 nblock);
557}
558EXPORT_SYMBOL(padata_unregister_cpumask_notifier);
559
560
561/* If cpumask contains no active cpu, we mark the instance as invalid. */
562static bool padata_validate_cpumask(struct padata_instance *pinst,
563 const struct cpumask *cpumask)
564{
565 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
566 pinst->flags |= PADATA_INVALID;
567 return false;
568 }
569
570 pinst->flags &= ~PADATA_INVALID;
571 return true;
572}
573
574static int __padata_set_cpumasks(struct padata_instance *pinst,
575 cpumask_var_t pcpumask,
576 cpumask_var_t cbcpumask)
577{
578 int valid;
579 struct parallel_data *pd;
580
581 valid = padata_validate_cpumask(pinst, pcpumask);
582 if (!valid) {
583 __padata_stop(pinst);
584 goto out_replace;
585 }
586
587 valid = padata_validate_cpumask(pinst, cbcpumask);
588 if (!valid)
589 __padata_stop(pinst);
590
591out_replace:
592 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
593 if (!pd)
594 return -ENOMEM;
595
596 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
597 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
598
599 padata_replace(pinst, pd);
600
601 if (valid)
602 __padata_start(pinst);
603
604 return 0;
605}
606
607/**
608 * padata_set_cpumask: Sets specified by @cpumask_type cpumask to the value
609 * equivalent to @cpumask.
610 *
611 * @pinst: padata instance
612 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
613 * to parallel and serial cpumasks respectively.
614 * @cpumask: the cpumask to use
615 */
616int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
617 cpumask_var_t cpumask)
618{
619 struct cpumask *serial_mask, *parallel_mask;
620 int err = -EINVAL;
621
622 mutex_lock(&pinst->lock);
623 get_online_cpus();
624
625 switch (cpumask_type) {
626 case PADATA_CPU_PARALLEL:
627 serial_mask = pinst->cpumask.cbcpu;
628 parallel_mask = cpumask;
629 break;
630 case PADATA_CPU_SERIAL:
631 parallel_mask = pinst->cpumask.pcpu;
632 serial_mask = cpumask;
633 break;
634 default:
635 goto out;
636 }
637
638 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
639
640out:
641 put_online_cpus();
642 mutex_unlock(&pinst->lock);
643
644 return err;
645}
646EXPORT_SYMBOL(padata_set_cpumask);
647
648/**
649 * padata_start - start the parallel processing
650 *
651 * @pinst: padata instance to start
652 */
653int padata_start(struct padata_instance *pinst)
654{
655 int err = 0;
656
657 mutex_lock(&pinst->lock);
658
659 if (pinst->flags & PADATA_INVALID)
660 err = -EINVAL;
661
662 __padata_start(pinst);
663
664 mutex_unlock(&pinst->lock);
665
666 return err;
667}
668EXPORT_SYMBOL(padata_start);
669
670/**
671 * padata_stop - stop the parallel processing
672 *
673 * @pinst: padata instance to stop
674 */
675void padata_stop(struct padata_instance *pinst)
676{
677 mutex_lock(&pinst->lock);
678 __padata_stop(pinst);
679 mutex_unlock(&pinst->lock);
680}
681EXPORT_SYMBOL(padata_stop);
682
683#ifdef CONFIG_HOTPLUG_CPU
684
685static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
686{
687 struct parallel_data *pd;
688
689 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
690 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
691 pinst->cpumask.cbcpu);
692 if (!pd)
693 return -ENOMEM;
694
695 padata_replace(pinst, pd);
696
697 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
698 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
699 __padata_start(pinst);
700 }
701
702 return 0;
703}
704
705static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
706{
707 struct parallel_data *pd = NULL;
708
709 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
710
711 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
712 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
713 __padata_stop(pinst);
714
715 pd = padata_alloc_pd(pinst, pinst->cpumask.pcpu,
716 pinst->cpumask.cbcpu);
717 if (!pd)
718 return -ENOMEM;
719
720 padata_replace(pinst, pd);
721
722 cpumask_clear_cpu(cpu, pd->cpumask.cbcpu);
723 cpumask_clear_cpu(cpu, pd->cpumask.pcpu);
724 }
725
726 return 0;
727}
728
729 /**
730 * padata_remove_cpu - remove a cpu from the one or both(serial and parallel)
731 * padata cpumasks.
732 *
733 * @pinst: padata instance
734 * @cpu: cpu to remove
735 * @mask: bitmask specifying from which cpumask @cpu should be removed
736 * The @mask may be any combination of the following flags:
737 * PADATA_CPU_SERIAL - serial cpumask
738 * PADATA_CPU_PARALLEL - parallel cpumask
739 */
740int padata_remove_cpu(struct padata_instance *pinst, int cpu, int mask)
741{
742 int err;
743
744 if (!(mask & (PADATA_CPU_SERIAL | PADATA_CPU_PARALLEL)))
745 return -EINVAL;
746
747 mutex_lock(&pinst->lock);
748
749 get_online_cpus();
750 if (mask & PADATA_CPU_SERIAL)
751 cpumask_clear_cpu(cpu, pinst->cpumask.cbcpu);
752 if (mask & PADATA_CPU_PARALLEL)
753 cpumask_clear_cpu(cpu, pinst->cpumask.pcpu);
754
755 err = __padata_remove_cpu(pinst, cpu);
756 put_online_cpus();
757
758 mutex_unlock(&pinst->lock);
759
760 return err;
761}
762EXPORT_SYMBOL(padata_remove_cpu);
763
764static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
765{
766 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
767 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
768}
769
770static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
771{
772 struct padata_instance *pinst;
773 int ret;
774
775 pinst = hlist_entry_safe(node, struct padata_instance, node);
776 if (!pinst_has_cpu(pinst, cpu))
777 return 0;
778
779 mutex_lock(&pinst->lock);
780 ret = __padata_add_cpu(pinst, cpu);
781 mutex_unlock(&pinst->lock);
782 return ret;
783}
784
785static int padata_cpu_prep_down(unsigned int cpu, struct hlist_node *node)
786{
787 struct padata_instance *pinst;
788 int ret;
789
790 pinst = hlist_entry_safe(node, struct padata_instance, node);
791 if (!pinst_has_cpu(pinst, cpu))
792 return 0;
793
794 mutex_lock(&pinst->lock);
795 ret = __padata_remove_cpu(pinst, cpu);
796 mutex_unlock(&pinst->lock);
797 return ret;
798}
799
800static enum cpuhp_state hp_online;
801#endif
802
803static void __padata_free(struct padata_instance *pinst)
804{
805#ifdef CONFIG_HOTPLUG_CPU
806 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->node);
807#endif
808
809 padata_stop(pinst);
810 padata_free_pd(pinst->pd);
811 free_cpumask_var(pinst->cpumask.pcpu);
812 free_cpumask_var(pinst->cpumask.cbcpu);
813 kfree(pinst);
814}
815
816#define kobj2pinst(_kobj) \
817 container_of(_kobj, struct padata_instance, kobj)
818#define attr2pentry(_attr) \
819 container_of(_attr, struct padata_sysfs_entry, attr)
820
821static void padata_sysfs_release(struct kobject *kobj)
822{
823 struct padata_instance *pinst = kobj2pinst(kobj);
824 __padata_free(pinst);
825}
826
827struct padata_sysfs_entry {
828 struct attribute attr;
829 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
830 ssize_t (*store)(struct padata_instance *, struct attribute *,
831 const char *, size_t);
832};
833
834static ssize_t show_cpumask(struct padata_instance *pinst,
835 struct attribute *attr, char *buf)
836{
837 struct cpumask *cpumask;
838 ssize_t len;
839
840 mutex_lock(&pinst->lock);
841 if (!strcmp(attr->name, "serial_cpumask"))
842 cpumask = pinst->cpumask.cbcpu;
843 else
844 cpumask = pinst->cpumask.pcpu;
845
846 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
847 nr_cpu_ids, cpumask_bits(cpumask));
848 mutex_unlock(&pinst->lock);
849 return len < PAGE_SIZE ? len : -EINVAL;
850}
851
852static ssize_t store_cpumask(struct padata_instance *pinst,
853 struct attribute *attr,
854 const char *buf, size_t count)
855{
856 cpumask_var_t new_cpumask;
857 ssize_t ret;
858 int mask_type;
859
860 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
861 return -ENOMEM;
862
863 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
864 nr_cpumask_bits);
865 if (ret < 0)
866 goto out;
867
868 mask_type = !strcmp(attr->name, "serial_cpumask") ?
869 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
870 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
871 if (!ret)
872 ret = count;
873
874out:
875 free_cpumask_var(new_cpumask);
876 return ret;
877}
878
879#define PADATA_ATTR_RW(_name, _show_name, _store_name) \
880 static struct padata_sysfs_entry _name##_attr = \
881 __ATTR(_name, 0644, _show_name, _store_name)
882#define PADATA_ATTR_RO(_name, _show_name) \
883 static struct padata_sysfs_entry _name##_attr = \
884 __ATTR(_name, 0400, _show_name, NULL)
885
886PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
887PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
888
889/*
890 * Padata sysfs provides the following objects:
891 * serial_cpumask [RW] - cpumask for serial workers
892 * parallel_cpumask [RW] - cpumask for parallel workers
893 */
894static struct attribute *padata_default_attrs[] = {
895 &serial_cpumask_attr.attr,
896 ¶llel_cpumask_attr.attr,
897 NULL,
898};
899
900static ssize_t padata_sysfs_show(struct kobject *kobj,
901 struct attribute *attr, char *buf)
902{
903 struct padata_instance *pinst;
904 struct padata_sysfs_entry *pentry;
905 ssize_t ret = -EIO;
906
907 pinst = kobj2pinst(kobj);
908 pentry = attr2pentry(attr);
909 if (pentry->show)
910 ret = pentry->show(pinst, attr, buf);
911
912 return ret;
913}
914
915static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
916 const char *buf, size_t count)
917{
918 struct padata_instance *pinst;
919 struct padata_sysfs_entry *pentry;
920 ssize_t ret = -EIO;
921
922 pinst = kobj2pinst(kobj);
923 pentry = attr2pentry(attr);
924 if (pentry->show)
925 ret = pentry->store(pinst, attr, buf, count);
926
927 return ret;
928}
929
930static const struct sysfs_ops padata_sysfs_ops = {
931 .show = padata_sysfs_show,
932 .store = padata_sysfs_store,
933};
934
935static struct kobj_type padata_attr_type = {
936 .sysfs_ops = &padata_sysfs_ops,
937 .default_attrs = padata_default_attrs,
938 .release = padata_sysfs_release,
939};
940
941/**
942 * padata_alloc_possible - Allocate and initialize padata instance.
943 * Use the cpu_possible_mask for serial and
944 * parallel workers.
945 *
946 * @wq: workqueue to use for the allocated padata instance
947 */
948struct padata_instance *padata_alloc_possible(struct workqueue_struct *wq)
949{
950 return padata_alloc(wq, cpu_possible_mask, cpu_possible_mask);
951}
952EXPORT_SYMBOL(padata_alloc_possible);
953
954/**
955 * padata_alloc - allocate and initialize a padata instance and specify
956 * cpumasks for serial and parallel workers.
957 *
958 * @wq: workqueue to use for the allocated padata instance
959 * @pcpumask: cpumask that will be used for padata parallelization
960 * @cbcpumask: cpumask that will be used for padata serialization
961 */
962struct padata_instance *padata_alloc(struct workqueue_struct *wq,
963 const struct cpumask *pcpumask,
964 const struct cpumask *cbcpumask)
965{
966 struct padata_instance *pinst;
967 struct parallel_data *pd = NULL;
968
969 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
970 if (!pinst)
971 goto err;
972
973 get_online_cpus();
974 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
975 goto err_free_inst;
976 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
977 free_cpumask_var(pinst->cpumask.pcpu);
978 goto err_free_inst;
979 }
980 if (!padata_validate_cpumask(pinst, pcpumask) ||
981 !padata_validate_cpumask(pinst, cbcpumask))
982 goto err_free_masks;
983
984 pd = padata_alloc_pd(pinst, pcpumask, cbcpumask);
985 if (!pd)
986 goto err_free_masks;
987
988 rcu_assign_pointer(pinst->pd, pd);
989
990 pinst->wq = wq;
991
992 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
993 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
994
995 pinst->flags = 0;
996
997 put_online_cpus();
998
999 BLOCKING_INIT_NOTIFIER_HEAD(&pinst->cpumask_change_notifier);
1000 kobject_init(&pinst->kobj, &padata_attr_type);
1001 mutex_init(&pinst->lock);
1002
1003#ifdef CONFIG_HOTPLUG_CPU
1004 cpuhp_state_add_instance_nocalls(hp_online, &pinst->node);
1005#endif
1006 return pinst;
1007
1008err_free_masks:
1009 free_cpumask_var(pinst->cpumask.pcpu);
1010 free_cpumask_var(pinst->cpumask.cbcpu);
1011err_free_inst:
1012 kfree(pinst);
1013 put_online_cpus();
1014err:
1015 return NULL;
1016}
1017
1018/**
1019 * padata_free - free a padata instance
1020 *
1021 * @padata_inst: padata instance to free
1022 */
1023void padata_free(struct padata_instance *pinst)
1024{
1025 kobject_put(&pinst->kobj);
1026}
1027EXPORT_SYMBOL(padata_free);
1028
1029#ifdef CONFIG_HOTPLUG_CPU
1030
1031static __init int padata_driver_init(void)
1032{
1033 int ret;
1034
1035 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1036 padata_cpu_online,
1037 padata_cpu_prep_down);
1038 if (ret < 0)
1039 return ret;
1040 hp_online = ret;
1041 return 0;
1042}
1043module_init(padata_driver_init);
1044
1045static __exit void padata_driver_exit(void)
1046{
1047 cpuhp_remove_multi_state(hp_online);
1048}
1049module_exit(padata_driver_exit);
1050#endif
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * padata.c - generic interface to process data streams in parallel
4 *
5 * See Documentation/core-api/padata.rst for more information.
6 *
7 * Copyright (C) 2008, 2009 secunet Security Networks AG
8 * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
9 *
10 * Copyright (c) 2020 Oracle and/or its affiliates.
11 * Author: Daniel Jordan <daniel.m.jordan@oracle.com>
12 */
13
14#include <linux/completion.h>
15#include <linux/export.h>
16#include <linux/cpumask.h>
17#include <linux/err.h>
18#include <linux/cpu.h>
19#include <linux/padata.h>
20#include <linux/mutex.h>
21#include <linux/sched.h>
22#include <linux/slab.h>
23#include <linux/sysfs.h>
24#include <linux/rcupdate.h>
25
26#define PADATA_WORK_ONSTACK 1 /* Work's memory is on stack */
27
28struct padata_work {
29 struct work_struct pw_work;
30 struct list_head pw_list; /* padata_free_works linkage */
31 void *pw_data;
32};
33
34static DEFINE_SPINLOCK(padata_works_lock);
35static struct padata_work *padata_works;
36static LIST_HEAD(padata_free_works);
37
38struct padata_mt_job_state {
39 spinlock_t lock;
40 struct completion completion;
41 struct padata_mt_job *job;
42 int nworks;
43 int nworks_fini;
44 unsigned long chunk_size;
45};
46
47static void padata_free_pd(struct parallel_data *pd);
48static void __init padata_mt_helper(struct work_struct *work);
49
50static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
51{
52 int cpu, target_cpu;
53
54 target_cpu = cpumask_first(pd->cpumask.pcpu);
55 for (cpu = 0; cpu < cpu_index; cpu++)
56 target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
57
58 return target_cpu;
59}
60
61static int padata_cpu_hash(struct parallel_data *pd, unsigned int seq_nr)
62{
63 /*
64 * Hash the sequence numbers to the cpus by taking
65 * seq_nr mod. number of cpus in use.
66 */
67 int cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
68
69 return padata_index_to_cpu(pd, cpu_index);
70}
71
72static struct padata_work *padata_work_alloc(void)
73{
74 struct padata_work *pw;
75
76 lockdep_assert_held(&padata_works_lock);
77
78 if (list_empty(&padata_free_works))
79 return NULL; /* No more work items allowed to be queued. */
80
81 pw = list_first_entry(&padata_free_works, struct padata_work, pw_list);
82 list_del(&pw->pw_list);
83 return pw;
84}
85
86/*
87 * This function is marked __ref because this function may be optimized in such
88 * a way that it directly refers to work_fn's address, which causes modpost to
89 * complain when work_fn is marked __init. This scenario was observed with clang
90 * LTO, where padata_work_init() was optimized to refer directly to
91 * padata_mt_helper() because the calls to padata_work_init() with other work_fn
92 * values were eliminated or inlined.
93 */
94static void __ref padata_work_init(struct padata_work *pw, work_func_t work_fn,
95 void *data, int flags)
96{
97 if (flags & PADATA_WORK_ONSTACK)
98 INIT_WORK_ONSTACK(&pw->pw_work, work_fn);
99 else
100 INIT_WORK(&pw->pw_work, work_fn);
101 pw->pw_data = data;
102}
103
104static int __init padata_work_alloc_mt(int nworks, void *data,
105 struct list_head *head)
106{
107 int i;
108
109 spin_lock(&padata_works_lock);
110 /* Start at 1 because the current task participates in the job. */
111 for (i = 1; i < nworks; ++i) {
112 struct padata_work *pw = padata_work_alloc();
113
114 if (!pw)
115 break;
116 padata_work_init(pw, padata_mt_helper, data, 0);
117 list_add(&pw->pw_list, head);
118 }
119 spin_unlock(&padata_works_lock);
120
121 return i;
122}
123
124static void padata_work_free(struct padata_work *pw)
125{
126 lockdep_assert_held(&padata_works_lock);
127 list_add(&pw->pw_list, &padata_free_works);
128}
129
130static void __init padata_works_free(struct list_head *works)
131{
132 struct padata_work *cur, *next;
133
134 if (list_empty(works))
135 return;
136
137 spin_lock(&padata_works_lock);
138 list_for_each_entry_safe(cur, next, works, pw_list) {
139 list_del(&cur->pw_list);
140 padata_work_free(cur);
141 }
142 spin_unlock(&padata_works_lock);
143}
144
145static void padata_parallel_worker(struct work_struct *parallel_work)
146{
147 struct padata_work *pw = container_of(parallel_work, struct padata_work,
148 pw_work);
149 struct padata_priv *padata = pw->pw_data;
150
151 local_bh_disable();
152 padata->parallel(padata);
153 spin_lock(&padata_works_lock);
154 padata_work_free(pw);
155 spin_unlock(&padata_works_lock);
156 local_bh_enable();
157}
158
159/**
160 * padata_do_parallel - padata parallelization function
161 *
162 * @ps: padatashell
163 * @padata: object to be parallelized
164 * @cb_cpu: pointer to the CPU that the serialization callback function should
165 * run on. If it's not in the serial cpumask of @pinst
166 * (i.e. cpumask.cbcpu), this function selects a fallback CPU and if
167 * none found, returns -EINVAL.
168 *
169 * The parallelization callback function will run with BHs off.
170 * Note: Every object which is parallelized by padata_do_parallel
171 * must be seen by padata_do_serial.
172 *
173 * Return: 0 on success or else negative error code.
174 */
175int padata_do_parallel(struct padata_shell *ps,
176 struct padata_priv *padata, int *cb_cpu)
177{
178 struct padata_instance *pinst = ps->pinst;
179 int i, cpu, cpu_index, err;
180 struct parallel_data *pd;
181 struct padata_work *pw;
182
183 rcu_read_lock_bh();
184
185 pd = rcu_dereference_bh(ps->pd);
186
187 err = -EINVAL;
188 if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
189 goto out;
190
191 if (!cpumask_test_cpu(*cb_cpu, pd->cpumask.cbcpu)) {
192 if (cpumask_empty(pd->cpumask.cbcpu))
193 goto out;
194
195 /* Select an alternate fallback CPU and notify the caller. */
196 cpu_index = *cb_cpu % cpumask_weight(pd->cpumask.cbcpu);
197
198 cpu = cpumask_first(pd->cpumask.cbcpu);
199 for (i = 0; i < cpu_index; i++)
200 cpu = cpumask_next(cpu, pd->cpumask.cbcpu);
201
202 *cb_cpu = cpu;
203 }
204
205 err = -EBUSY;
206 if ((pinst->flags & PADATA_RESET))
207 goto out;
208
209 refcount_inc(&pd->refcnt);
210 padata->pd = pd;
211 padata->cb_cpu = *cb_cpu;
212
213 spin_lock(&padata_works_lock);
214 padata->seq_nr = ++pd->seq_nr;
215 pw = padata_work_alloc();
216 spin_unlock(&padata_works_lock);
217
218 if (!pw) {
219 /* Maximum works limit exceeded, run in the current task. */
220 padata->parallel(padata);
221 }
222
223 rcu_read_unlock_bh();
224
225 if (pw) {
226 padata_work_init(pw, padata_parallel_worker, padata, 0);
227 queue_work(pinst->parallel_wq, &pw->pw_work);
228 }
229
230 return 0;
231out:
232 rcu_read_unlock_bh();
233
234 return err;
235}
236EXPORT_SYMBOL(padata_do_parallel);
237
238/*
239 * padata_find_next - Find the next object that needs serialization.
240 *
241 * Return:
242 * * A pointer to the control struct of the next object that needs
243 * serialization, if present in one of the percpu reorder queues.
244 * * NULL, if the next object that needs serialization will
245 * be parallel processed by another cpu and is not yet present in
246 * the cpu's reorder queue.
247 */
248static struct padata_priv *padata_find_next(struct parallel_data *pd,
249 bool remove_object)
250{
251 struct padata_priv *padata;
252 struct padata_list *reorder;
253 int cpu = pd->cpu;
254
255 reorder = per_cpu_ptr(pd->reorder_list, cpu);
256
257 spin_lock(&reorder->lock);
258 if (list_empty(&reorder->list)) {
259 spin_unlock(&reorder->lock);
260 return NULL;
261 }
262
263 padata = list_entry(reorder->list.next, struct padata_priv, list);
264
265 /*
266 * Checks the rare case where two or more parallel jobs have hashed to
267 * the same CPU and one of the later ones finishes first.
268 */
269 if (padata->seq_nr != pd->processed) {
270 spin_unlock(&reorder->lock);
271 return NULL;
272 }
273
274 if (remove_object) {
275 list_del_init(&padata->list);
276 ++pd->processed;
277 pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, -1, false);
278 }
279
280 spin_unlock(&reorder->lock);
281 return padata;
282}
283
284static void padata_reorder(struct parallel_data *pd)
285{
286 struct padata_instance *pinst = pd->ps->pinst;
287 int cb_cpu;
288 struct padata_priv *padata;
289 struct padata_serial_queue *squeue;
290 struct padata_list *reorder;
291
292 /*
293 * We need to ensure that only one cpu can work on dequeueing of
294 * the reorder queue the time. Calculating in which percpu reorder
295 * queue the next object will arrive takes some time. A spinlock
296 * would be highly contended. Also it is not clear in which order
297 * the objects arrive to the reorder queues. So a cpu could wait to
298 * get the lock just to notice that there is nothing to do at the
299 * moment. Therefore we use a trylock and let the holder of the lock
300 * care for all the objects enqueued during the holdtime of the lock.
301 */
302 if (!spin_trylock_bh(&pd->lock))
303 return;
304
305 while (1) {
306 padata = padata_find_next(pd, true);
307
308 /*
309 * If the next object that needs serialization is parallel
310 * processed by another cpu and is still on it's way to the
311 * cpu's reorder queue, nothing to do for now.
312 */
313 if (!padata)
314 break;
315
316 cb_cpu = padata->cb_cpu;
317 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
318
319 spin_lock(&squeue->serial.lock);
320 list_add_tail(&padata->list, &squeue->serial.list);
321 spin_unlock(&squeue->serial.lock);
322
323 queue_work_on(cb_cpu, pinst->serial_wq, &squeue->work);
324 }
325
326 spin_unlock_bh(&pd->lock);
327
328 /*
329 * The next object that needs serialization might have arrived to
330 * the reorder queues in the meantime.
331 *
332 * Ensure reorder queue is read after pd->lock is dropped so we see
333 * new objects from another task in padata_do_serial. Pairs with
334 * smp_mb in padata_do_serial.
335 */
336 smp_mb();
337
338 reorder = per_cpu_ptr(pd->reorder_list, pd->cpu);
339 if (!list_empty(&reorder->list) && padata_find_next(pd, false))
340 queue_work(pinst->serial_wq, &pd->reorder_work);
341}
342
343static void invoke_padata_reorder(struct work_struct *work)
344{
345 struct parallel_data *pd;
346
347 local_bh_disable();
348 pd = container_of(work, struct parallel_data, reorder_work);
349 padata_reorder(pd);
350 local_bh_enable();
351}
352
353static void padata_serial_worker(struct work_struct *serial_work)
354{
355 struct padata_serial_queue *squeue;
356 struct parallel_data *pd;
357 LIST_HEAD(local_list);
358 int cnt;
359
360 local_bh_disable();
361 squeue = container_of(serial_work, struct padata_serial_queue, work);
362 pd = squeue->pd;
363
364 spin_lock(&squeue->serial.lock);
365 list_replace_init(&squeue->serial.list, &local_list);
366 spin_unlock(&squeue->serial.lock);
367
368 cnt = 0;
369
370 while (!list_empty(&local_list)) {
371 struct padata_priv *padata;
372
373 padata = list_entry(local_list.next,
374 struct padata_priv, list);
375
376 list_del_init(&padata->list);
377
378 padata->serial(padata);
379 cnt++;
380 }
381 local_bh_enable();
382
383 if (refcount_sub_and_test(cnt, &pd->refcnt))
384 padata_free_pd(pd);
385}
386
387/**
388 * padata_do_serial - padata serialization function
389 *
390 * @padata: object to be serialized.
391 *
392 * padata_do_serial must be called for every parallelized object.
393 * The serialization callback function will run with BHs off.
394 */
395void padata_do_serial(struct padata_priv *padata)
396{
397 struct parallel_data *pd = padata->pd;
398 int hashed_cpu = padata_cpu_hash(pd, padata->seq_nr);
399 struct padata_list *reorder = per_cpu_ptr(pd->reorder_list, hashed_cpu);
400 struct padata_priv *cur;
401 struct list_head *pos;
402
403 spin_lock(&reorder->lock);
404 /* Sort in ascending order of sequence number. */
405 list_for_each_prev(pos, &reorder->list) {
406 cur = list_entry(pos, struct padata_priv, list);
407 if (cur->seq_nr < padata->seq_nr)
408 break;
409 }
410 list_add(&padata->list, pos);
411 spin_unlock(&reorder->lock);
412
413 /*
414 * Ensure the addition to the reorder list is ordered correctly
415 * with the trylock of pd->lock in padata_reorder. Pairs with smp_mb
416 * in padata_reorder.
417 */
418 smp_mb();
419
420 padata_reorder(pd);
421}
422EXPORT_SYMBOL(padata_do_serial);
423
424static int padata_setup_cpumasks(struct padata_instance *pinst)
425{
426 struct workqueue_attrs *attrs;
427 int err;
428
429 attrs = alloc_workqueue_attrs();
430 if (!attrs)
431 return -ENOMEM;
432
433 /* Restrict parallel_wq workers to pd->cpumask.pcpu. */
434 cpumask_copy(attrs->cpumask, pinst->cpumask.pcpu);
435 err = apply_workqueue_attrs(pinst->parallel_wq, attrs);
436 free_workqueue_attrs(attrs);
437
438 return err;
439}
440
441static void __init padata_mt_helper(struct work_struct *w)
442{
443 struct padata_work *pw = container_of(w, struct padata_work, pw_work);
444 struct padata_mt_job_state *ps = pw->pw_data;
445 struct padata_mt_job *job = ps->job;
446 bool done;
447
448 spin_lock(&ps->lock);
449
450 while (job->size > 0) {
451 unsigned long start, size, end;
452
453 start = job->start;
454 /* So end is chunk size aligned if enough work remains. */
455 size = roundup(start + 1, ps->chunk_size) - start;
456 size = min(size, job->size);
457 end = start + size;
458
459 job->start = end;
460 job->size -= size;
461
462 spin_unlock(&ps->lock);
463 job->thread_fn(start, end, job->fn_arg);
464 spin_lock(&ps->lock);
465 }
466
467 ++ps->nworks_fini;
468 done = (ps->nworks_fini == ps->nworks);
469 spin_unlock(&ps->lock);
470
471 if (done)
472 complete(&ps->completion);
473}
474
475/**
476 * padata_do_multithreaded - run a multithreaded job
477 * @job: Description of the job.
478 *
479 * See the definition of struct padata_mt_job for more details.
480 */
481void __init padata_do_multithreaded(struct padata_mt_job *job)
482{
483 /* In case threads finish at different times. */
484 static const unsigned long load_balance_factor = 4;
485 struct padata_work my_work, *pw;
486 struct padata_mt_job_state ps;
487 LIST_HEAD(works);
488 int nworks, nid;
489 static atomic_t last_used_nid __initdata;
490
491 if (job->size == 0)
492 return;
493
494 /* Ensure at least one thread when size < min_chunk. */
495 nworks = max(job->size / max(job->min_chunk, job->align), 1ul);
496 nworks = min(nworks, job->max_threads);
497
498 if (nworks == 1) {
499 /* Single thread, no coordination needed, cut to the chase. */
500 job->thread_fn(job->start, job->start + job->size, job->fn_arg);
501 return;
502 }
503
504 spin_lock_init(&ps.lock);
505 init_completion(&ps.completion);
506 ps.job = job;
507 ps.nworks = padata_work_alloc_mt(nworks, &ps, &works);
508 ps.nworks_fini = 0;
509
510 /*
511 * Chunk size is the amount of work a helper does per call to the
512 * thread function. Load balance large jobs between threads by
513 * increasing the number of chunks, guarantee at least the minimum
514 * chunk size from the caller, and honor the caller's alignment.
515 */
516 ps.chunk_size = job->size / (ps.nworks * load_balance_factor);
517 ps.chunk_size = max(ps.chunk_size, job->min_chunk);
518 ps.chunk_size = roundup(ps.chunk_size, job->align);
519
520 list_for_each_entry(pw, &works, pw_list)
521 if (job->numa_aware) {
522 int old_node = atomic_read(&last_used_nid);
523
524 do {
525 nid = next_node_in(old_node, node_states[N_CPU]);
526 } while (!atomic_try_cmpxchg(&last_used_nid, &old_node, nid));
527 queue_work_node(nid, system_unbound_wq, &pw->pw_work);
528 } else {
529 queue_work(system_unbound_wq, &pw->pw_work);
530 }
531
532 /* Use the current thread, which saves starting a workqueue worker. */
533 padata_work_init(&my_work, padata_mt_helper, &ps, PADATA_WORK_ONSTACK);
534 padata_mt_helper(&my_work.pw_work);
535
536 /* Wait for all the helpers to finish. */
537 wait_for_completion(&ps.completion);
538
539 destroy_work_on_stack(&my_work.pw_work);
540 padata_works_free(&works);
541}
542
543static void __padata_list_init(struct padata_list *pd_list)
544{
545 INIT_LIST_HEAD(&pd_list->list);
546 spin_lock_init(&pd_list->lock);
547}
548
549/* Initialize all percpu queues used by serial workers */
550static void padata_init_squeues(struct parallel_data *pd)
551{
552 int cpu;
553 struct padata_serial_queue *squeue;
554
555 for_each_cpu(cpu, pd->cpumask.cbcpu) {
556 squeue = per_cpu_ptr(pd->squeue, cpu);
557 squeue->pd = pd;
558 __padata_list_init(&squeue->serial);
559 INIT_WORK(&squeue->work, padata_serial_worker);
560 }
561}
562
563/* Initialize per-CPU reorder lists */
564static void padata_init_reorder_list(struct parallel_data *pd)
565{
566 int cpu;
567 struct padata_list *list;
568
569 for_each_cpu(cpu, pd->cpumask.pcpu) {
570 list = per_cpu_ptr(pd->reorder_list, cpu);
571 __padata_list_init(list);
572 }
573}
574
575/* Allocate and initialize the internal cpumask dependend resources. */
576static struct parallel_data *padata_alloc_pd(struct padata_shell *ps)
577{
578 struct padata_instance *pinst = ps->pinst;
579 struct parallel_data *pd;
580
581 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
582 if (!pd)
583 goto err;
584
585 pd->reorder_list = alloc_percpu(struct padata_list);
586 if (!pd->reorder_list)
587 goto err_free_pd;
588
589 pd->squeue = alloc_percpu(struct padata_serial_queue);
590 if (!pd->squeue)
591 goto err_free_reorder_list;
592
593 pd->ps = ps;
594
595 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
596 goto err_free_squeue;
597 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
598 goto err_free_pcpu;
599
600 cpumask_and(pd->cpumask.pcpu, pinst->cpumask.pcpu, cpu_online_mask);
601 cpumask_and(pd->cpumask.cbcpu, pinst->cpumask.cbcpu, cpu_online_mask);
602
603 padata_init_reorder_list(pd);
604 padata_init_squeues(pd);
605 pd->seq_nr = -1;
606 refcount_set(&pd->refcnt, 1);
607 spin_lock_init(&pd->lock);
608 pd->cpu = cpumask_first(pd->cpumask.pcpu);
609 INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
610
611 return pd;
612
613err_free_pcpu:
614 free_cpumask_var(pd->cpumask.pcpu);
615err_free_squeue:
616 free_percpu(pd->squeue);
617err_free_reorder_list:
618 free_percpu(pd->reorder_list);
619err_free_pd:
620 kfree(pd);
621err:
622 return NULL;
623}
624
625static void padata_free_pd(struct parallel_data *pd)
626{
627 free_cpumask_var(pd->cpumask.pcpu);
628 free_cpumask_var(pd->cpumask.cbcpu);
629 free_percpu(pd->reorder_list);
630 free_percpu(pd->squeue);
631 kfree(pd);
632}
633
634static void __padata_start(struct padata_instance *pinst)
635{
636 pinst->flags |= PADATA_INIT;
637}
638
639static void __padata_stop(struct padata_instance *pinst)
640{
641 if (!(pinst->flags & PADATA_INIT))
642 return;
643
644 pinst->flags &= ~PADATA_INIT;
645
646 synchronize_rcu();
647}
648
649/* Replace the internal control structure with a new one. */
650static int padata_replace_one(struct padata_shell *ps)
651{
652 struct parallel_data *pd_new;
653
654 pd_new = padata_alloc_pd(ps);
655 if (!pd_new)
656 return -ENOMEM;
657
658 ps->opd = rcu_dereference_protected(ps->pd, 1);
659 rcu_assign_pointer(ps->pd, pd_new);
660
661 return 0;
662}
663
664static int padata_replace(struct padata_instance *pinst)
665{
666 struct padata_shell *ps;
667 int err = 0;
668
669 pinst->flags |= PADATA_RESET;
670
671 list_for_each_entry(ps, &pinst->pslist, list) {
672 err = padata_replace_one(ps);
673 if (err)
674 break;
675 }
676
677 synchronize_rcu();
678
679 list_for_each_entry_continue_reverse(ps, &pinst->pslist, list)
680 if (refcount_dec_and_test(&ps->opd->refcnt))
681 padata_free_pd(ps->opd);
682
683 pinst->flags &= ~PADATA_RESET;
684
685 return err;
686}
687
688/* If cpumask contains no active cpu, we mark the instance as invalid. */
689static bool padata_validate_cpumask(struct padata_instance *pinst,
690 const struct cpumask *cpumask)
691{
692 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
693 pinst->flags |= PADATA_INVALID;
694 return false;
695 }
696
697 pinst->flags &= ~PADATA_INVALID;
698 return true;
699}
700
701static int __padata_set_cpumasks(struct padata_instance *pinst,
702 cpumask_var_t pcpumask,
703 cpumask_var_t cbcpumask)
704{
705 int valid;
706 int err;
707
708 valid = padata_validate_cpumask(pinst, pcpumask);
709 if (!valid) {
710 __padata_stop(pinst);
711 goto out_replace;
712 }
713
714 valid = padata_validate_cpumask(pinst, cbcpumask);
715 if (!valid)
716 __padata_stop(pinst);
717
718out_replace:
719 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
720 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
721
722 err = padata_setup_cpumasks(pinst) ?: padata_replace(pinst);
723
724 if (valid)
725 __padata_start(pinst);
726
727 return err;
728}
729
730/**
731 * padata_set_cpumask - Sets specified by @cpumask_type cpumask to the value
732 * equivalent to @cpumask.
733 * @pinst: padata instance
734 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
735 * to parallel and serial cpumasks respectively.
736 * @cpumask: the cpumask to use
737 *
738 * Return: 0 on success or negative error code
739 */
740int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
741 cpumask_var_t cpumask)
742{
743 struct cpumask *serial_mask, *parallel_mask;
744 int err = -EINVAL;
745
746 cpus_read_lock();
747 mutex_lock(&pinst->lock);
748
749 switch (cpumask_type) {
750 case PADATA_CPU_PARALLEL:
751 serial_mask = pinst->cpumask.cbcpu;
752 parallel_mask = cpumask;
753 break;
754 case PADATA_CPU_SERIAL:
755 parallel_mask = pinst->cpumask.pcpu;
756 serial_mask = cpumask;
757 break;
758 default:
759 goto out;
760 }
761
762 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
763
764out:
765 mutex_unlock(&pinst->lock);
766 cpus_read_unlock();
767
768 return err;
769}
770EXPORT_SYMBOL(padata_set_cpumask);
771
772#ifdef CONFIG_HOTPLUG_CPU
773
774static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
775{
776 int err = 0;
777
778 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
779 err = padata_replace(pinst);
780
781 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
782 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
783 __padata_start(pinst);
784 }
785
786 return err;
787}
788
789static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
790{
791 int err = 0;
792
793 if (!cpumask_test_cpu(cpu, cpu_online_mask)) {
794 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
795 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
796 __padata_stop(pinst);
797
798 err = padata_replace(pinst);
799 }
800
801 return err;
802}
803
804static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
805{
806 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
807 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
808}
809
810static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
811{
812 struct padata_instance *pinst;
813 int ret;
814
815 pinst = hlist_entry_safe(node, struct padata_instance, cpu_online_node);
816 if (!pinst_has_cpu(pinst, cpu))
817 return 0;
818
819 mutex_lock(&pinst->lock);
820 ret = __padata_add_cpu(pinst, cpu);
821 mutex_unlock(&pinst->lock);
822 return ret;
823}
824
825static int padata_cpu_dead(unsigned int cpu, struct hlist_node *node)
826{
827 struct padata_instance *pinst;
828 int ret;
829
830 pinst = hlist_entry_safe(node, struct padata_instance, cpu_dead_node);
831 if (!pinst_has_cpu(pinst, cpu))
832 return 0;
833
834 mutex_lock(&pinst->lock);
835 ret = __padata_remove_cpu(pinst, cpu);
836 mutex_unlock(&pinst->lock);
837 return ret;
838}
839
840static enum cpuhp_state hp_online;
841#endif
842
843static void __padata_free(struct padata_instance *pinst)
844{
845#ifdef CONFIG_HOTPLUG_CPU
846 cpuhp_state_remove_instance_nocalls(CPUHP_PADATA_DEAD,
847 &pinst->cpu_dead_node);
848 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->cpu_online_node);
849#endif
850
851 WARN_ON(!list_empty(&pinst->pslist));
852
853 free_cpumask_var(pinst->cpumask.pcpu);
854 free_cpumask_var(pinst->cpumask.cbcpu);
855 destroy_workqueue(pinst->serial_wq);
856 destroy_workqueue(pinst->parallel_wq);
857 kfree(pinst);
858}
859
860#define kobj2pinst(_kobj) \
861 container_of(_kobj, struct padata_instance, kobj)
862#define attr2pentry(_attr) \
863 container_of(_attr, struct padata_sysfs_entry, attr)
864
865static void padata_sysfs_release(struct kobject *kobj)
866{
867 struct padata_instance *pinst = kobj2pinst(kobj);
868 __padata_free(pinst);
869}
870
871struct padata_sysfs_entry {
872 struct attribute attr;
873 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
874 ssize_t (*store)(struct padata_instance *, struct attribute *,
875 const char *, size_t);
876};
877
878static ssize_t show_cpumask(struct padata_instance *pinst,
879 struct attribute *attr, char *buf)
880{
881 struct cpumask *cpumask;
882 ssize_t len;
883
884 mutex_lock(&pinst->lock);
885 if (!strcmp(attr->name, "serial_cpumask"))
886 cpumask = pinst->cpumask.cbcpu;
887 else
888 cpumask = pinst->cpumask.pcpu;
889
890 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
891 nr_cpu_ids, cpumask_bits(cpumask));
892 mutex_unlock(&pinst->lock);
893 return len < PAGE_SIZE ? len : -EINVAL;
894}
895
896static ssize_t store_cpumask(struct padata_instance *pinst,
897 struct attribute *attr,
898 const char *buf, size_t count)
899{
900 cpumask_var_t new_cpumask;
901 ssize_t ret;
902 int mask_type;
903
904 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
905 return -ENOMEM;
906
907 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
908 nr_cpumask_bits);
909 if (ret < 0)
910 goto out;
911
912 mask_type = !strcmp(attr->name, "serial_cpumask") ?
913 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
914 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
915 if (!ret)
916 ret = count;
917
918out:
919 free_cpumask_var(new_cpumask);
920 return ret;
921}
922
923#define PADATA_ATTR_RW(_name, _show_name, _store_name) \
924 static struct padata_sysfs_entry _name##_attr = \
925 __ATTR(_name, 0644, _show_name, _store_name)
926#define PADATA_ATTR_RO(_name, _show_name) \
927 static struct padata_sysfs_entry _name##_attr = \
928 __ATTR(_name, 0400, _show_name, NULL)
929
930PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
931PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
932
933/*
934 * Padata sysfs provides the following objects:
935 * serial_cpumask [RW] - cpumask for serial workers
936 * parallel_cpumask [RW] - cpumask for parallel workers
937 */
938static struct attribute *padata_default_attrs[] = {
939 &serial_cpumask_attr.attr,
940 ¶llel_cpumask_attr.attr,
941 NULL,
942};
943ATTRIBUTE_GROUPS(padata_default);
944
945static ssize_t padata_sysfs_show(struct kobject *kobj,
946 struct attribute *attr, char *buf)
947{
948 struct padata_instance *pinst;
949 struct padata_sysfs_entry *pentry;
950 ssize_t ret = -EIO;
951
952 pinst = kobj2pinst(kobj);
953 pentry = attr2pentry(attr);
954 if (pentry->show)
955 ret = pentry->show(pinst, attr, buf);
956
957 return ret;
958}
959
960static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
961 const char *buf, size_t count)
962{
963 struct padata_instance *pinst;
964 struct padata_sysfs_entry *pentry;
965 ssize_t ret = -EIO;
966
967 pinst = kobj2pinst(kobj);
968 pentry = attr2pentry(attr);
969 if (pentry->show)
970 ret = pentry->store(pinst, attr, buf, count);
971
972 return ret;
973}
974
975static const struct sysfs_ops padata_sysfs_ops = {
976 .show = padata_sysfs_show,
977 .store = padata_sysfs_store,
978};
979
980static const struct kobj_type padata_attr_type = {
981 .sysfs_ops = &padata_sysfs_ops,
982 .default_groups = padata_default_groups,
983 .release = padata_sysfs_release,
984};
985
986/**
987 * padata_alloc - allocate and initialize a padata instance
988 * @name: used to identify the instance
989 *
990 * Return: new instance on success, NULL on error
991 */
992struct padata_instance *padata_alloc(const char *name)
993{
994 struct padata_instance *pinst;
995
996 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
997 if (!pinst)
998 goto err;
999
1000 pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
1001 name);
1002 if (!pinst->parallel_wq)
1003 goto err_free_inst;
1004
1005 cpus_read_lock();
1006
1007 pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
1008 WQ_CPU_INTENSIVE, 1, name);
1009 if (!pinst->serial_wq)
1010 goto err_put_cpus;
1011
1012 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
1013 goto err_free_serial_wq;
1014 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
1015 free_cpumask_var(pinst->cpumask.pcpu);
1016 goto err_free_serial_wq;
1017 }
1018
1019 INIT_LIST_HEAD(&pinst->pslist);
1020
1021 cpumask_copy(pinst->cpumask.pcpu, cpu_possible_mask);
1022 cpumask_copy(pinst->cpumask.cbcpu, cpu_possible_mask);
1023
1024 if (padata_setup_cpumasks(pinst))
1025 goto err_free_masks;
1026
1027 __padata_start(pinst);
1028
1029 kobject_init(&pinst->kobj, &padata_attr_type);
1030 mutex_init(&pinst->lock);
1031
1032#ifdef CONFIG_HOTPLUG_CPU
1033 cpuhp_state_add_instance_nocalls_cpuslocked(hp_online,
1034 &pinst->cpu_online_node);
1035 cpuhp_state_add_instance_nocalls_cpuslocked(CPUHP_PADATA_DEAD,
1036 &pinst->cpu_dead_node);
1037#endif
1038
1039 cpus_read_unlock();
1040
1041 return pinst;
1042
1043err_free_masks:
1044 free_cpumask_var(pinst->cpumask.pcpu);
1045 free_cpumask_var(pinst->cpumask.cbcpu);
1046err_free_serial_wq:
1047 destroy_workqueue(pinst->serial_wq);
1048err_put_cpus:
1049 cpus_read_unlock();
1050 destroy_workqueue(pinst->parallel_wq);
1051err_free_inst:
1052 kfree(pinst);
1053err:
1054 return NULL;
1055}
1056EXPORT_SYMBOL(padata_alloc);
1057
1058/**
1059 * padata_free - free a padata instance
1060 *
1061 * @pinst: padata instance to free
1062 */
1063void padata_free(struct padata_instance *pinst)
1064{
1065 kobject_put(&pinst->kobj);
1066}
1067EXPORT_SYMBOL(padata_free);
1068
1069/**
1070 * padata_alloc_shell - Allocate and initialize padata shell.
1071 *
1072 * @pinst: Parent padata_instance object.
1073 *
1074 * Return: new shell on success, NULL on error
1075 */
1076struct padata_shell *padata_alloc_shell(struct padata_instance *pinst)
1077{
1078 struct parallel_data *pd;
1079 struct padata_shell *ps;
1080
1081 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
1082 if (!ps)
1083 goto out;
1084
1085 ps->pinst = pinst;
1086
1087 cpus_read_lock();
1088 pd = padata_alloc_pd(ps);
1089 cpus_read_unlock();
1090
1091 if (!pd)
1092 goto out_free_ps;
1093
1094 mutex_lock(&pinst->lock);
1095 RCU_INIT_POINTER(ps->pd, pd);
1096 list_add(&ps->list, &pinst->pslist);
1097 mutex_unlock(&pinst->lock);
1098
1099 return ps;
1100
1101out_free_ps:
1102 kfree(ps);
1103out:
1104 return NULL;
1105}
1106EXPORT_SYMBOL(padata_alloc_shell);
1107
1108/**
1109 * padata_free_shell - free a padata shell
1110 *
1111 * @ps: padata shell to free
1112 */
1113void padata_free_shell(struct padata_shell *ps)
1114{
1115 struct parallel_data *pd;
1116
1117 if (!ps)
1118 return;
1119
1120 mutex_lock(&ps->pinst->lock);
1121 list_del(&ps->list);
1122 pd = rcu_dereference_protected(ps->pd, 1);
1123 if (refcount_dec_and_test(&pd->refcnt))
1124 padata_free_pd(pd);
1125 mutex_unlock(&ps->pinst->lock);
1126
1127 kfree(ps);
1128}
1129EXPORT_SYMBOL(padata_free_shell);
1130
1131void __init padata_init(void)
1132{
1133 unsigned int i, possible_cpus;
1134#ifdef CONFIG_HOTPLUG_CPU
1135 int ret;
1136
1137 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1138 padata_cpu_online, NULL);
1139 if (ret < 0)
1140 goto err;
1141 hp_online = ret;
1142
1143 ret = cpuhp_setup_state_multi(CPUHP_PADATA_DEAD, "padata:dead",
1144 NULL, padata_cpu_dead);
1145 if (ret < 0)
1146 goto remove_online_state;
1147#endif
1148
1149 possible_cpus = num_possible_cpus();
1150 padata_works = kmalloc_array(possible_cpus, sizeof(struct padata_work),
1151 GFP_KERNEL);
1152 if (!padata_works)
1153 goto remove_dead_state;
1154
1155 for (i = 0; i < possible_cpus; ++i)
1156 list_add(&padata_works[i].pw_list, &padata_free_works);
1157
1158 return;
1159
1160remove_dead_state:
1161#ifdef CONFIG_HOTPLUG_CPU
1162 cpuhp_remove_multi_state(CPUHP_PADATA_DEAD);
1163remove_online_state:
1164 cpuhp_remove_multi_state(hp_online);
1165err:
1166#endif
1167 pr_warn("padata: initialization failed\n");
1168}