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 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}
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}