Loading...
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 inline void padata_get_pd(struct parallel_data *pd)
51{
52 refcount_inc(&pd->refcnt);
53}
54
55static inline void padata_put_pd_cnt(struct parallel_data *pd, int cnt)
56{
57 if (refcount_sub_and_test(cnt, &pd->refcnt))
58 padata_free_pd(pd);
59}
60
61static inline void padata_put_pd(struct parallel_data *pd)
62{
63 padata_put_pd_cnt(pd, 1);
64}
65
66static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
67{
68 int cpu, target_cpu;
69
70 target_cpu = cpumask_first(pd->cpumask.pcpu);
71 for (cpu = 0; cpu < cpu_index; cpu++)
72 target_cpu = cpumask_next(target_cpu, pd->cpumask.pcpu);
73
74 return target_cpu;
75}
76
77static int padata_cpu_hash(struct parallel_data *pd, unsigned int seq_nr)
78{
79 /*
80 * Hash the sequence numbers to the cpus by taking
81 * seq_nr mod. number of cpus in use.
82 */
83 int cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
84
85 return padata_index_to_cpu(pd, cpu_index);
86}
87
88static struct padata_work *padata_work_alloc(void)
89{
90 struct padata_work *pw;
91
92 lockdep_assert_held(&padata_works_lock);
93
94 if (list_empty(&padata_free_works))
95 return NULL; /* No more work items allowed to be queued. */
96
97 pw = list_first_entry(&padata_free_works, struct padata_work, pw_list);
98 list_del(&pw->pw_list);
99 return pw;
100}
101
102/*
103 * This function is marked __ref because this function may be optimized in such
104 * a way that it directly refers to work_fn's address, which causes modpost to
105 * complain when work_fn is marked __init. This scenario was observed with clang
106 * LTO, where padata_work_init() was optimized to refer directly to
107 * padata_mt_helper() because the calls to padata_work_init() with other work_fn
108 * values were eliminated or inlined.
109 */
110static void __ref padata_work_init(struct padata_work *pw, work_func_t work_fn,
111 void *data, int flags)
112{
113 if (flags & PADATA_WORK_ONSTACK)
114 INIT_WORK_ONSTACK(&pw->pw_work, work_fn);
115 else
116 INIT_WORK(&pw->pw_work, work_fn);
117 pw->pw_data = data;
118}
119
120static int __init padata_work_alloc_mt(int nworks, void *data,
121 struct list_head *head)
122{
123 int i;
124
125 spin_lock_bh(&padata_works_lock);
126 /* Start at 1 because the current task participates in the job. */
127 for (i = 1; i < nworks; ++i) {
128 struct padata_work *pw = padata_work_alloc();
129
130 if (!pw)
131 break;
132 padata_work_init(pw, padata_mt_helper, data, 0);
133 list_add(&pw->pw_list, head);
134 }
135 spin_unlock_bh(&padata_works_lock);
136
137 return i;
138}
139
140static void padata_work_free(struct padata_work *pw)
141{
142 lockdep_assert_held(&padata_works_lock);
143 list_add(&pw->pw_list, &padata_free_works);
144}
145
146static void __init padata_works_free(struct list_head *works)
147{
148 struct padata_work *cur, *next;
149
150 if (list_empty(works))
151 return;
152
153 spin_lock_bh(&padata_works_lock);
154 list_for_each_entry_safe(cur, next, works, pw_list) {
155 list_del(&cur->pw_list);
156 padata_work_free(cur);
157 }
158 spin_unlock_bh(&padata_works_lock);
159}
160
161static void padata_parallel_worker(struct work_struct *parallel_work)
162{
163 struct padata_work *pw = container_of(parallel_work, struct padata_work,
164 pw_work);
165 struct padata_priv *padata = pw->pw_data;
166
167 local_bh_disable();
168 padata->parallel(padata);
169 spin_lock(&padata_works_lock);
170 padata_work_free(pw);
171 spin_unlock(&padata_works_lock);
172 local_bh_enable();
173}
174
175/**
176 * padata_do_parallel - padata parallelization function
177 *
178 * @ps: padatashell
179 * @padata: object to be parallelized
180 * @cb_cpu: pointer to the CPU that the serialization callback function should
181 * run on. If it's not in the serial cpumask of @pinst
182 * (i.e. cpumask.cbcpu), this function selects a fallback CPU and if
183 * none found, returns -EINVAL.
184 *
185 * The parallelization callback function will run with BHs off.
186 * Note: Every object which is parallelized by padata_do_parallel
187 * must be seen by padata_do_serial.
188 *
189 * Return: 0 on success or else negative error code.
190 */
191int padata_do_parallel(struct padata_shell *ps,
192 struct padata_priv *padata, int *cb_cpu)
193{
194 struct padata_instance *pinst = ps->pinst;
195 int i, cpu, cpu_index, err;
196 struct parallel_data *pd;
197 struct padata_work *pw;
198
199 rcu_read_lock_bh();
200
201 pd = rcu_dereference_bh(ps->pd);
202
203 err = -EINVAL;
204 if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
205 goto out;
206
207 if (!cpumask_test_cpu(*cb_cpu, pd->cpumask.cbcpu)) {
208 if (cpumask_empty(pd->cpumask.cbcpu))
209 goto out;
210
211 /* Select an alternate fallback CPU and notify the caller. */
212 cpu_index = *cb_cpu % cpumask_weight(pd->cpumask.cbcpu);
213
214 cpu = cpumask_first(pd->cpumask.cbcpu);
215 for (i = 0; i < cpu_index; i++)
216 cpu = cpumask_next(cpu, pd->cpumask.cbcpu);
217
218 *cb_cpu = cpu;
219 }
220
221 err = -EBUSY;
222 if ((pinst->flags & PADATA_RESET))
223 goto out;
224
225 padata_get_pd(pd);
226 padata->pd = pd;
227 padata->cb_cpu = *cb_cpu;
228
229 spin_lock(&padata_works_lock);
230 padata->seq_nr = ++pd->seq_nr;
231 pw = padata_work_alloc();
232 spin_unlock(&padata_works_lock);
233
234 if (!pw) {
235 /* Maximum works limit exceeded, run in the current task. */
236 padata->parallel(padata);
237 }
238
239 rcu_read_unlock_bh();
240
241 if (pw) {
242 padata_work_init(pw, padata_parallel_worker, padata, 0);
243 queue_work(pinst->parallel_wq, &pw->pw_work);
244 }
245
246 return 0;
247out:
248 rcu_read_unlock_bh();
249
250 return err;
251}
252EXPORT_SYMBOL(padata_do_parallel);
253
254/*
255 * padata_find_next - Find the next object that needs serialization.
256 *
257 * Return:
258 * * A pointer to the control struct of the next object that needs
259 * serialization, if present in one of the percpu reorder queues.
260 * * NULL, if the next object that needs serialization will
261 * be parallel processed by another cpu and is not yet present in
262 * the cpu's reorder queue.
263 */
264static struct padata_priv *padata_find_next(struct parallel_data *pd,
265 bool remove_object)
266{
267 struct padata_priv *padata;
268 struct padata_list *reorder;
269 int cpu = pd->cpu;
270
271 reorder = per_cpu_ptr(pd->reorder_list, cpu);
272
273 spin_lock(&reorder->lock);
274 if (list_empty(&reorder->list)) {
275 spin_unlock(&reorder->lock);
276 return NULL;
277 }
278
279 padata = list_entry(reorder->list.next, struct padata_priv, list);
280
281 /*
282 * Checks the rare case where two or more parallel jobs have hashed to
283 * the same CPU and one of the later ones finishes first.
284 */
285 if (padata->seq_nr != pd->processed) {
286 spin_unlock(&reorder->lock);
287 return NULL;
288 }
289
290 if (remove_object) {
291 list_del_init(&padata->list);
292 ++pd->processed;
293 pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, -1, false);
294 }
295
296 spin_unlock(&reorder->lock);
297 return padata;
298}
299
300static void padata_reorder(struct parallel_data *pd)
301{
302 struct padata_instance *pinst = pd->ps->pinst;
303 int cb_cpu;
304 struct padata_priv *padata;
305 struct padata_serial_queue *squeue;
306 struct padata_list *reorder;
307
308 /*
309 * We need to ensure that only one cpu can work on dequeueing of
310 * the reorder queue the time. Calculating in which percpu reorder
311 * queue the next object will arrive takes some time. A spinlock
312 * would be highly contended. Also it is not clear in which order
313 * the objects arrive to the reorder queues. So a cpu could wait to
314 * get the lock just to notice that there is nothing to do at the
315 * moment. Therefore we use a trylock and let the holder of the lock
316 * care for all the objects enqueued during the holdtime of the lock.
317 */
318 if (!spin_trylock_bh(&pd->lock))
319 return;
320
321 while (1) {
322 padata = padata_find_next(pd, true);
323
324 /*
325 * If the next object that needs serialization is parallel
326 * processed by another cpu and is still on it's way to the
327 * cpu's reorder queue, nothing to do for now.
328 */
329 if (!padata)
330 break;
331
332 cb_cpu = padata->cb_cpu;
333 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
334
335 spin_lock(&squeue->serial.lock);
336 list_add_tail(&padata->list, &squeue->serial.list);
337 spin_unlock(&squeue->serial.lock);
338
339 queue_work_on(cb_cpu, pinst->serial_wq, &squeue->work);
340 }
341
342 spin_unlock_bh(&pd->lock);
343
344 /*
345 * The next object that needs serialization might have arrived to
346 * the reorder queues in the meantime.
347 *
348 * Ensure reorder queue is read after pd->lock is dropped so we see
349 * new objects from another task in padata_do_serial. Pairs with
350 * smp_mb in padata_do_serial.
351 */
352 smp_mb();
353
354 reorder = per_cpu_ptr(pd->reorder_list, pd->cpu);
355 if (!list_empty(&reorder->list) && padata_find_next(pd, false)) {
356 /*
357 * Other context(eg. the padata_serial_worker) can finish the request.
358 * To avoid UAF issue, add pd ref here, and put pd ref after reorder_work finish.
359 */
360 padata_get_pd(pd);
361 queue_work(pinst->serial_wq, &pd->reorder_work);
362 }
363}
364
365static void invoke_padata_reorder(struct work_struct *work)
366{
367 struct parallel_data *pd;
368
369 local_bh_disable();
370 pd = container_of(work, struct parallel_data, reorder_work);
371 padata_reorder(pd);
372 local_bh_enable();
373 /* Pairs with putting the reorder_work in the serial_wq */
374 padata_put_pd(pd);
375}
376
377static void padata_serial_worker(struct work_struct *serial_work)
378{
379 struct padata_serial_queue *squeue;
380 struct parallel_data *pd;
381 LIST_HEAD(local_list);
382 int cnt;
383
384 local_bh_disable();
385 squeue = container_of(serial_work, struct padata_serial_queue, work);
386 pd = squeue->pd;
387
388 spin_lock(&squeue->serial.lock);
389 list_replace_init(&squeue->serial.list, &local_list);
390 spin_unlock(&squeue->serial.lock);
391
392 cnt = 0;
393
394 while (!list_empty(&local_list)) {
395 struct padata_priv *padata;
396
397 padata = list_entry(local_list.next,
398 struct padata_priv, list);
399
400 list_del_init(&padata->list);
401
402 padata->serial(padata);
403 cnt++;
404 }
405 local_bh_enable();
406
407 padata_put_pd_cnt(pd, cnt);
408}
409
410/**
411 * padata_do_serial - padata serialization function
412 *
413 * @padata: object to be serialized.
414 *
415 * padata_do_serial must be called for every parallelized object.
416 * The serialization callback function will run with BHs off.
417 */
418void padata_do_serial(struct padata_priv *padata)
419{
420 struct parallel_data *pd = padata->pd;
421 int hashed_cpu = padata_cpu_hash(pd, padata->seq_nr);
422 struct padata_list *reorder = per_cpu_ptr(pd->reorder_list, hashed_cpu);
423 struct padata_priv *cur;
424 struct list_head *pos;
425
426 spin_lock(&reorder->lock);
427 /* Sort in ascending order of sequence number. */
428 list_for_each_prev(pos, &reorder->list) {
429 cur = list_entry(pos, struct padata_priv, list);
430 /* Compare by difference to consider integer wrap around */
431 if ((signed int)(cur->seq_nr - padata->seq_nr) < 0)
432 break;
433 }
434 list_add(&padata->list, pos);
435 spin_unlock(&reorder->lock);
436
437 /*
438 * Ensure the addition to the reorder list is ordered correctly
439 * with the trylock of pd->lock in padata_reorder. Pairs with smp_mb
440 * in padata_reorder.
441 */
442 smp_mb();
443
444 padata_reorder(pd);
445}
446EXPORT_SYMBOL(padata_do_serial);
447
448static int padata_setup_cpumasks(struct padata_instance *pinst)
449{
450 struct workqueue_attrs *attrs;
451 int err;
452
453 attrs = alloc_workqueue_attrs();
454 if (!attrs)
455 return -ENOMEM;
456
457 /* Restrict parallel_wq workers to pd->cpumask.pcpu. */
458 cpumask_copy(attrs->cpumask, pinst->cpumask.pcpu);
459 err = apply_workqueue_attrs(pinst->parallel_wq, attrs);
460 free_workqueue_attrs(attrs);
461
462 return err;
463}
464
465static void __init padata_mt_helper(struct work_struct *w)
466{
467 struct padata_work *pw = container_of(w, struct padata_work, pw_work);
468 struct padata_mt_job_state *ps = pw->pw_data;
469 struct padata_mt_job *job = ps->job;
470 bool done;
471
472 spin_lock(&ps->lock);
473
474 while (job->size > 0) {
475 unsigned long start, size, end;
476
477 start = job->start;
478 /* So end is chunk size aligned if enough work remains. */
479 size = roundup(start + 1, ps->chunk_size) - start;
480 size = min(size, job->size);
481 end = start + size;
482
483 job->start = end;
484 job->size -= size;
485
486 spin_unlock(&ps->lock);
487 job->thread_fn(start, end, job->fn_arg);
488 spin_lock(&ps->lock);
489 }
490
491 ++ps->nworks_fini;
492 done = (ps->nworks_fini == ps->nworks);
493 spin_unlock(&ps->lock);
494
495 if (done)
496 complete(&ps->completion);
497}
498
499/**
500 * padata_do_multithreaded - run a multithreaded job
501 * @job: Description of the job.
502 *
503 * See the definition of struct padata_mt_job for more details.
504 */
505void __init padata_do_multithreaded(struct padata_mt_job *job)
506{
507 /* In case threads finish at different times. */
508 static const unsigned long load_balance_factor = 4;
509 struct padata_work my_work, *pw;
510 struct padata_mt_job_state ps;
511 LIST_HEAD(works);
512 int nworks, nid;
513 static atomic_t last_used_nid __initdata;
514
515 if (job->size == 0)
516 return;
517
518 /* Ensure at least one thread when size < min_chunk. */
519 nworks = max(job->size / max(job->min_chunk, job->align), 1ul);
520 nworks = min(nworks, job->max_threads);
521
522 if (nworks == 1) {
523 /* Single thread, no coordination needed, cut to the chase. */
524 job->thread_fn(job->start, job->start + job->size, job->fn_arg);
525 return;
526 }
527
528 spin_lock_init(&ps.lock);
529 init_completion(&ps.completion);
530 ps.job = job;
531 ps.nworks = padata_work_alloc_mt(nworks, &ps, &works);
532 ps.nworks_fini = 0;
533
534 /*
535 * Chunk size is the amount of work a helper does per call to the
536 * thread function. Load balance large jobs between threads by
537 * increasing the number of chunks, guarantee at least the minimum
538 * chunk size from the caller, and honor the caller's alignment.
539 * Ensure chunk_size is at least 1 to prevent divide-by-0
540 * panic in padata_mt_helper().
541 */
542 ps.chunk_size = job->size / (ps.nworks * load_balance_factor);
543 ps.chunk_size = max(ps.chunk_size, job->min_chunk);
544 ps.chunk_size = max(ps.chunk_size, 1ul);
545 ps.chunk_size = roundup(ps.chunk_size, job->align);
546
547 list_for_each_entry(pw, &works, pw_list)
548 if (job->numa_aware) {
549 int old_node = atomic_read(&last_used_nid);
550
551 do {
552 nid = next_node_in(old_node, node_states[N_CPU]);
553 } while (!atomic_try_cmpxchg(&last_used_nid, &old_node, nid));
554 queue_work_node(nid, system_unbound_wq, &pw->pw_work);
555 } else {
556 queue_work(system_unbound_wq, &pw->pw_work);
557 }
558
559 /* Use the current thread, which saves starting a workqueue worker. */
560 padata_work_init(&my_work, padata_mt_helper, &ps, PADATA_WORK_ONSTACK);
561 padata_mt_helper(&my_work.pw_work);
562
563 /* Wait for all the helpers to finish. */
564 wait_for_completion(&ps.completion);
565
566 destroy_work_on_stack(&my_work.pw_work);
567 padata_works_free(&works);
568}
569
570static void __padata_list_init(struct padata_list *pd_list)
571{
572 INIT_LIST_HEAD(&pd_list->list);
573 spin_lock_init(&pd_list->lock);
574}
575
576/* Initialize all percpu queues used by serial workers */
577static void padata_init_squeues(struct parallel_data *pd)
578{
579 int cpu;
580 struct padata_serial_queue *squeue;
581
582 for_each_cpu(cpu, pd->cpumask.cbcpu) {
583 squeue = per_cpu_ptr(pd->squeue, cpu);
584 squeue->pd = pd;
585 __padata_list_init(&squeue->serial);
586 INIT_WORK(&squeue->work, padata_serial_worker);
587 }
588}
589
590/* Initialize per-CPU reorder lists */
591static void padata_init_reorder_list(struct parallel_data *pd)
592{
593 int cpu;
594 struct padata_list *list;
595
596 for_each_cpu(cpu, pd->cpumask.pcpu) {
597 list = per_cpu_ptr(pd->reorder_list, cpu);
598 __padata_list_init(list);
599 }
600}
601
602/* Allocate and initialize the internal cpumask dependend resources. */
603static struct parallel_data *padata_alloc_pd(struct padata_shell *ps)
604{
605 struct padata_instance *pinst = ps->pinst;
606 struct parallel_data *pd;
607
608 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
609 if (!pd)
610 goto err;
611
612 pd->reorder_list = alloc_percpu(struct padata_list);
613 if (!pd->reorder_list)
614 goto err_free_pd;
615
616 pd->squeue = alloc_percpu(struct padata_serial_queue);
617 if (!pd->squeue)
618 goto err_free_reorder_list;
619
620 pd->ps = ps;
621
622 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
623 goto err_free_squeue;
624 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
625 goto err_free_pcpu;
626
627 cpumask_and(pd->cpumask.pcpu, pinst->cpumask.pcpu, cpu_online_mask);
628 cpumask_and(pd->cpumask.cbcpu, pinst->cpumask.cbcpu, cpu_online_mask);
629
630 padata_init_reorder_list(pd);
631 padata_init_squeues(pd);
632 pd->seq_nr = -1;
633 refcount_set(&pd->refcnt, 1);
634 spin_lock_init(&pd->lock);
635 pd->cpu = cpumask_first(pd->cpumask.pcpu);
636 INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
637
638 return pd;
639
640err_free_pcpu:
641 free_cpumask_var(pd->cpumask.pcpu);
642err_free_squeue:
643 free_percpu(pd->squeue);
644err_free_reorder_list:
645 free_percpu(pd->reorder_list);
646err_free_pd:
647 kfree(pd);
648err:
649 return NULL;
650}
651
652static void padata_free_pd(struct parallel_data *pd)
653{
654 free_cpumask_var(pd->cpumask.pcpu);
655 free_cpumask_var(pd->cpumask.cbcpu);
656 free_percpu(pd->reorder_list);
657 free_percpu(pd->squeue);
658 kfree(pd);
659}
660
661static void __padata_start(struct padata_instance *pinst)
662{
663 pinst->flags |= PADATA_INIT;
664}
665
666static void __padata_stop(struct padata_instance *pinst)
667{
668 if (!(pinst->flags & PADATA_INIT))
669 return;
670
671 pinst->flags &= ~PADATA_INIT;
672
673 synchronize_rcu();
674}
675
676/* Replace the internal control structure with a new one. */
677static int padata_replace_one(struct padata_shell *ps)
678{
679 struct parallel_data *pd_new;
680
681 pd_new = padata_alloc_pd(ps);
682 if (!pd_new)
683 return -ENOMEM;
684
685 ps->opd = rcu_dereference_protected(ps->pd, 1);
686 rcu_assign_pointer(ps->pd, pd_new);
687
688 return 0;
689}
690
691static int padata_replace(struct padata_instance *pinst)
692{
693 struct padata_shell *ps;
694 int err = 0;
695
696 pinst->flags |= PADATA_RESET;
697
698 list_for_each_entry(ps, &pinst->pslist, list) {
699 err = padata_replace_one(ps);
700 if (err)
701 break;
702 }
703
704 synchronize_rcu();
705
706 list_for_each_entry_continue_reverse(ps, &pinst->pslist, list)
707 padata_put_pd(ps->opd);
708
709 pinst->flags &= ~PADATA_RESET;
710
711 return err;
712}
713
714/* If cpumask contains no active cpu, we mark the instance as invalid. */
715static bool padata_validate_cpumask(struct padata_instance *pinst,
716 const struct cpumask *cpumask)
717{
718 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
719 pinst->flags |= PADATA_INVALID;
720 return false;
721 }
722
723 pinst->flags &= ~PADATA_INVALID;
724 return true;
725}
726
727static int __padata_set_cpumasks(struct padata_instance *pinst,
728 cpumask_var_t pcpumask,
729 cpumask_var_t cbcpumask)
730{
731 int valid;
732 int err;
733
734 valid = padata_validate_cpumask(pinst, pcpumask);
735 if (!valid) {
736 __padata_stop(pinst);
737 goto out_replace;
738 }
739
740 valid = padata_validate_cpumask(pinst, cbcpumask);
741 if (!valid)
742 __padata_stop(pinst);
743
744out_replace:
745 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
746 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
747
748 err = padata_setup_cpumasks(pinst) ?: padata_replace(pinst);
749
750 if (valid)
751 __padata_start(pinst);
752
753 return err;
754}
755
756/**
757 * padata_set_cpumask - Sets specified by @cpumask_type cpumask to the value
758 * equivalent to @cpumask.
759 * @pinst: padata instance
760 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
761 * to parallel and serial cpumasks respectively.
762 * @cpumask: the cpumask to use
763 *
764 * Return: 0 on success or negative error code
765 */
766int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
767 cpumask_var_t cpumask)
768{
769 struct cpumask *serial_mask, *parallel_mask;
770 int err = -EINVAL;
771
772 cpus_read_lock();
773 mutex_lock(&pinst->lock);
774
775 switch (cpumask_type) {
776 case PADATA_CPU_PARALLEL:
777 serial_mask = pinst->cpumask.cbcpu;
778 parallel_mask = cpumask;
779 break;
780 case PADATA_CPU_SERIAL:
781 parallel_mask = pinst->cpumask.pcpu;
782 serial_mask = cpumask;
783 break;
784 default:
785 goto out;
786 }
787
788 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
789
790out:
791 mutex_unlock(&pinst->lock);
792 cpus_read_unlock();
793
794 return err;
795}
796EXPORT_SYMBOL(padata_set_cpumask);
797
798#ifdef CONFIG_HOTPLUG_CPU
799
800static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
801{
802 int err = 0;
803
804 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
805 err = padata_replace(pinst);
806
807 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
808 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
809 __padata_start(pinst);
810 }
811
812 return err;
813}
814
815static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
816{
817 int err = 0;
818
819 if (!cpumask_test_cpu(cpu, cpu_online_mask)) {
820 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
821 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
822 __padata_stop(pinst);
823
824 err = padata_replace(pinst);
825 }
826
827 return err;
828}
829
830static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
831{
832 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
833 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
834}
835
836static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
837{
838 struct padata_instance *pinst;
839 int ret;
840
841 pinst = hlist_entry_safe(node, struct padata_instance, cpu_online_node);
842 if (!pinst_has_cpu(pinst, cpu))
843 return 0;
844
845 mutex_lock(&pinst->lock);
846 ret = __padata_add_cpu(pinst, cpu);
847 mutex_unlock(&pinst->lock);
848 return ret;
849}
850
851static int padata_cpu_dead(unsigned int cpu, struct hlist_node *node)
852{
853 struct padata_instance *pinst;
854 int ret;
855
856 pinst = hlist_entry_safe(node, struct padata_instance, cpu_dead_node);
857 if (!pinst_has_cpu(pinst, cpu))
858 return 0;
859
860 mutex_lock(&pinst->lock);
861 ret = __padata_remove_cpu(pinst, cpu);
862 mutex_unlock(&pinst->lock);
863 return ret;
864}
865
866static enum cpuhp_state hp_online;
867#endif
868
869static void __padata_free(struct padata_instance *pinst)
870{
871#ifdef CONFIG_HOTPLUG_CPU
872 cpuhp_state_remove_instance_nocalls(CPUHP_PADATA_DEAD,
873 &pinst->cpu_dead_node);
874 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->cpu_online_node);
875#endif
876
877 WARN_ON(!list_empty(&pinst->pslist));
878
879 free_cpumask_var(pinst->cpumask.pcpu);
880 free_cpumask_var(pinst->cpumask.cbcpu);
881 destroy_workqueue(pinst->serial_wq);
882 destroy_workqueue(pinst->parallel_wq);
883 kfree(pinst);
884}
885
886#define kobj2pinst(_kobj) \
887 container_of(_kobj, struct padata_instance, kobj)
888#define attr2pentry(_attr) \
889 container_of(_attr, struct padata_sysfs_entry, attr)
890
891static void padata_sysfs_release(struct kobject *kobj)
892{
893 struct padata_instance *pinst = kobj2pinst(kobj);
894 __padata_free(pinst);
895}
896
897struct padata_sysfs_entry {
898 struct attribute attr;
899 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
900 ssize_t (*store)(struct padata_instance *, struct attribute *,
901 const char *, size_t);
902};
903
904static ssize_t show_cpumask(struct padata_instance *pinst,
905 struct attribute *attr, char *buf)
906{
907 struct cpumask *cpumask;
908 ssize_t len;
909
910 mutex_lock(&pinst->lock);
911 if (!strcmp(attr->name, "serial_cpumask"))
912 cpumask = pinst->cpumask.cbcpu;
913 else
914 cpumask = pinst->cpumask.pcpu;
915
916 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
917 nr_cpu_ids, cpumask_bits(cpumask));
918 mutex_unlock(&pinst->lock);
919 return len < PAGE_SIZE ? len : -EINVAL;
920}
921
922static ssize_t store_cpumask(struct padata_instance *pinst,
923 struct attribute *attr,
924 const char *buf, size_t count)
925{
926 cpumask_var_t new_cpumask;
927 ssize_t ret;
928 int mask_type;
929
930 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
931 return -ENOMEM;
932
933 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
934 nr_cpumask_bits);
935 if (ret < 0)
936 goto out;
937
938 mask_type = !strcmp(attr->name, "serial_cpumask") ?
939 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
940 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
941 if (!ret)
942 ret = count;
943
944out:
945 free_cpumask_var(new_cpumask);
946 return ret;
947}
948
949#define PADATA_ATTR_RW(_name, _show_name, _store_name) \
950 static struct padata_sysfs_entry _name##_attr = \
951 __ATTR(_name, 0644, _show_name, _store_name)
952#define PADATA_ATTR_RO(_name, _show_name) \
953 static struct padata_sysfs_entry _name##_attr = \
954 __ATTR(_name, 0400, _show_name, NULL)
955
956PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
957PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
958
959/*
960 * Padata sysfs provides the following objects:
961 * serial_cpumask [RW] - cpumask for serial workers
962 * parallel_cpumask [RW] - cpumask for parallel workers
963 */
964static struct attribute *padata_default_attrs[] = {
965 &serial_cpumask_attr.attr,
966 ¶llel_cpumask_attr.attr,
967 NULL,
968};
969ATTRIBUTE_GROUPS(padata_default);
970
971static ssize_t padata_sysfs_show(struct kobject *kobj,
972 struct attribute *attr, char *buf)
973{
974 struct padata_instance *pinst;
975 struct padata_sysfs_entry *pentry;
976 ssize_t ret = -EIO;
977
978 pinst = kobj2pinst(kobj);
979 pentry = attr2pentry(attr);
980 if (pentry->show)
981 ret = pentry->show(pinst, attr, buf);
982
983 return ret;
984}
985
986static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
987 const char *buf, size_t count)
988{
989 struct padata_instance *pinst;
990 struct padata_sysfs_entry *pentry;
991 ssize_t ret = -EIO;
992
993 pinst = kobj2pinst(kobj);
994 pentry = attr2pentry(attr);
995 if (pentry->store)
996 ret = pentry->store(pinst, attr, buf, count);
997
998 return ret;
999}
1000
1001static const struct sysfs_ops padata_sysfs_ops = {
1002 .show = padata_sysfs_show,
1003 .store = padata_sysfs_store,
1004};
1005
1006static const struct kobj_type padata_attr_type = {
1007 .sysfs_ops = &padata_sysfs_ops,
1008 .default_groups = padata_default_groups,
1009 .release = padata_sysfs_release,
1010};
1011
1012/**
1013 * padata_alloc - allocate and initialize a padata instance
1014 * @name: used to identify the instance
1015 *
1016 * Return: new instance on success, NULL on error
1017 */
1018struct padata_instance *padata_alloc(const char *name)
1019{
1020 struct padata_instance *pinst;
1021
1022 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
1023 if (!pinst)
1024 goto err;
1025
1026 pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
1027 name);
1028 if (!pinst->parallel_wq)
1029 goto err_free_inst;
1030
1031 cpus_read_lock();
1032
1033 pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
1034 WQ_CPU_INTENSIVE, 1, name);
1035 if (!pinst->serial_wq)
1036 goto err_put_cpus;
1037
1038 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
1039 goto err_free_serial_wq;
1040 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
1041 free_cpumask_var(pinst->cpumask.pcpu);
1042 goto err_free_serial_wq;
1043 }
1044
1045 INIT_LIST_HEAD(&pinst->pslist);
1046
1047 cpumask_copy(pinst->cpumask.pcpu, cpu_possible_mask);
1048 cpumask_copy(pinst->cpumask.cbcpu, cpu_possible_mask);
1049
1050 if (padata_setup_cpumasks(pinst))
1051 goto err_free_masks;
1052
1053 __padata_start(pinst);
1054
1055 kobject_init(&pinst->kobj, &padata_attr_type);
1056 mutex_init(&pinst->lock);
1057
1058#ifdef CONFIG_HOTPLUG_CPU
1059 cpuhp_state_add_instance_nocalls_cpuslocked(hp_online,
1060 &pinst->cpu_online_node);
1061 cpuhp_state_add_instance_nocalls_cpuslocked(CPUHP_PADATA_DEAD,
1062 &pinst->cpu_dead_node);
1063#endif
1064
1065 cpus_read_unlock();
1066
1067 return pinst;
1068
1069err_free_masks:
1070 free_cpumask_var(pinst->cpumask.pcpu);
1071 free_cpumask_var(pinst->cpumask.cbcpu);
1072err_free_serial_wq:
1073 destroy_workqueue(pinst->serial_wq);
1074err_put_cpus:
1075 cpus_read_unlock();
1076 destroy_workqueue(pinst->parallel_wq);
1077err_free_inst:
1078 kfree(pinst);
1079err:
1080 return NULL;
1081}
1082EXPORT_SYMBOL(padata_alloc);
1083
1084/**
1085 * padata_free - free a padata instance
1086 *
1087 * @pinst: padata instance to free
1088 */
1089void padata_free(struct padata_instance *pinst)
1090{
1091 kobject_put(&pinst->kobj);
1092}
1093EXPORT_SYMBOL(padata_free);
1094
1095/**
1096 * padata_alloc_shell - Allocate and initialize padata shell.
1097 *
1098 * @pinst: Parent padata_instance object.
1099 *
1100 * Return: new shell on success, NULL on error
1101 */
1102struct padata_shell *padata_alloc_shell(struct padata_instance *pinst)
1103{
1104 struct parallel_data *pd;
1105 struct padata_shell *ps;
1106
1107 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
1108 if (!ps)
1109 goto out;
1110
1111 ps->pinst = pinst;
1112
1113 cpus_read_lock();
1114 pd = padata_alloc_pd(ps);
1115 cpus_read_unlock();
1116
1117 if (!pd)
1118 goto out_free_ps;
1119
1120 mutex_lock(&pinst->lock);
1121 RCU_INIT_POINTER(ps->pd, pd);
1122 list_add(&ps->list, &pinst->pslist);
1123 mutex_unlock(&pinst->lock);
1124
1125 return ps;
1126
1127out_free_ps:
1128 kfree(ps);
1129out:
1130 return NULL;
1131}
1132EXPORT_SYMBOL(padata_alloc_shell);
1133
1134/**
1135 * padata_free_shell - free a padata shell
1136 *
1137 * @ps: padata shell to free
1138 */
1139void padata_free_shell(struct padata_shell *ps)
1140{
1141 struct parallel_data *pd;
1142
1143 if (!ps)
1144 return;
1145
1146 /*
1147 * Wait for all _do_serial calls to finish to avoid touching
1148 * freed pd's and ps's.
1149 */
1150 synchronize_rcu();
1151
1152 mutex_lock(&ps->pinst->lock);
1153 list_del(&ps->list);
1154 pd = rcu_dereference_protected(ps->pd, 1);
1155 padata_put_pd(pd);
1156 mutex_unlock(&ps->pinst->lock);
1157
1158 kfree(ps);
1159}
1160EXPORT_SYMBOL(padata_free_shell);
1161
1162void __init padata_init(void)
1163{
1164 unsigned int i, possible_cpus;
1165#ifdef CONFIG_HOTPLUG_CPU
1166 int ret;
1167
1168 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1169 padata_cpu_online, NULL);
1170 if (ret < 0)
1171 goto err;
1172 hp_online = ret;
1173
1174 ret = cpuhp_setup_state_multi(CPUHP_PADATA_DEAD, "padata:dead",
1175 NULL, padata_cpu_dead);
1176 if (ret < 0)
1177 goto remove_online_state;
1178#endif
1179
1180 possible_cpus = num_possible_cpus();
1181 padata_works = kmalloc_array(possible_cpus, sizeof(struct padata_work),
1182 GFP_KERNEL);
1183 if (!padata_works)
1184 goto remove_dead_state;
1185
1186 for (i = 0; i < possible_cpus; ++i)
1187 list_add(&padata_works[i].pw_list, &padata_free_works);
1188
1189 return;
1190
1191remove_dead_state:
1192#ifdef CONFIG_HOTPLUG_CPU
1193 cpuhp_remove_multi_state(CPUHP_PADATA_DEAD);
1194remove_online_state:
1195 cpuhp_remove_multi_state(hp_online);
1196err:
1197#endif
1198 pr_warn("padata: initialization failed\n");
1199}
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);