projects/linux·blob v6.11 · refs

1175 lines · 28.5 KiB

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 <[email protected]>
9 *
10 * Copyright (c) 2020 Oracle and/or its affiliates.
11 * Author: Daniel Jordan <[email protected]>
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_bh(&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_bh(&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_bh(&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_bh(&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 /*
521 * chunk_size can be 0 if the caller sets min_chunk to 0. So force it
522 * to at least 1 to prevent divide-by-0 panic in padata_mt_helper().`
523 */
524 if (!ps.chunk_size)
525 ps.chunk_size = 1U;
526
527 list_for_each_entry(pw, &works, pw_list)
528 if (job->numa_aware) {
529 int old_node = atomic_read(&last_used_nid);
530
531 do {
532 nid = next_node_in(old_node, node_states[N_CPU]);
533 } while (!atomic_try_cmpxchg(&last_used_nid, &old_node, nid));
534 queue_work_node(nid, system_unbound_wq, &pw->pw_work);
535 } else {
536 queue_work(system_unbound_wq, &pw->pw_work);
537 }
538
539 /* Use the current thread, which saves starting a workqueue worker. */
540 padata_work_init(&my_work, padata_mt_helper, &ps, PADATA_WORK_ONSTACK);
541 padata_mt_helper(&my_work.pw_work);
542
543 /* Wait for all the helpers to finish. */
544 wait_for_completion(&ps.completion);
545
546 destroy_work_on_stack(&my_work.pw_work);
547 padata_works_free(&works);
548}
549
550static void __padata_list_init(struct padata_list *pd_list)
551{
552 INIT_LIST_HEAD(&pd_list->list);
553 spin_lock_init(&pd_list->lock);
554}
555
556/* Initialize all percpu queues used by serial workers */
557static void padata_init_squeues(struct parallel_data *pd)
558{
559 int cpu;
560 struct padata_serial_queue *squeue;
561
562 for_each_cpu(cpu, pd->cpumask.cbcpu) {
563 squeue = per_cpu_ptr(pd->squeue, cpu);
564 squeue->pd = pd;
565 __padata_list_init(&squeue->serial);
566 INIT_WORK(&squeue->work, padata_serial_worker);
567 }
568}
569
570/* Initialize per-CPU reorder lists */
571static void padata_init_reorder_list(struct parallel_data *pd)
572{
573 int cpu;
574 struct padata_list *list;
575
576 for_each_cpu(cpu, pd->cpumask.pcpu) {
577 list = per_cpu_ptr(pd->reorder_list, cpu);
578 __padata_list_init(list);
579 }
580}
581
582/* Allocate and initialize the internal cpumask dependend resources. */
583static struct parallel_data *padata_alloc_pd(struct padata_shell *ps)
584{
585 struct padata_instance *pinst = ps->pinst;
586 struct parallel_data *pd;
587
588 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
589 if (!pd)
590 goto err;
591
592 pd->reorder_list = alloc_percpu(struct padata_list);
593 if (!pd->reorder_list)
594 goto err_free_pd;
595
596 pd->squeue = alloc_percpu(struct padata_serial_queue);
597 if (!pd->squeue)
598 goto err_free_reorder_list;
599
600 pd->ps = ps;
601
602 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
603 goto err_free_squeue;
604 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
605 goto err_free_pcpu;
606
607 cpumask_and(pd->cpumask.pcpu, pinst->cpumask.pcpu, cpu_online_mask);
608 cpumask_and(pd->cpumask.cbcpu, pinst->cpumask.cbcpu, cpu_online_mask);
609
610 padata_init_reorder_list(pd);
611 padata_init_squeues(pd);
612 pd->seq_nr = -1;
613 refcount_set(&pd->refcnt, 1);
614 spin_lock_init(&pd->lock);
615 pd->cpu = cpumask_first(pd->cpumask.pcpu);
616 INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
617
618 return pd;
619
620err_free_pcpu:
621 free_cpumask_var(pd->cpumask.pcpu);
622err_free_squeue:
623 free_percpu(pd->squeue);
624err_free_reorder_list:
625 free_percpu(pd->reorder_list);
626err_free_pd:
627 kfree(pd);
628err:
629 return NULL;
630}
631
632static void padata_free_pd(struct parallel_data *pd)
633{
634 free_cpumask_var(pd->cpumask.pcpu);
635 free_cpumask_var(pd->cpumask.cbcpu);
636 free_percpu(pd->reorder_list);
637 free_percpu(pd->squeue);
638 kfree(pd);
639}
640
641static void __padata_start(struct padata_instance *pinst)
642{
643 pinst->flags |= PADATA_INIT;
644}
645
646static void __padata_stop(struct padata_instance *pinst)
647{
648 if (!(pinst->flags & PADATA_INIT))
649 return;
650
651 pinst->flags &= ~PADATA_INIT;
652
653 synchronize_rcu();
654}
655
656/* Replace the internal control structure with a new one. */
657static int padata_replace_one(struct padata_shell *ps)
658{
659 struct parallel_data *pd_new;
660
661 pd_new = padata_alloc_pd(ps);
662 if (!pd_new)
663 return -ENOMEM;
664
665 ps->opd = rcu_dereference_protected(ps->pd, 1);
666 rcu_assign_pointer(ps->pd, pd_new);
667
668 return 0;
669}
670
671static int padata_replace(struct padata_instance *pinst)
672{
673 struct padata_shell *ps;
674 int err = 0;
675
676 pinst->flags |= PADATA_RESET;
677
678 list_for_each_entry(ps, &pinst->pslist, list) {
679 err = padata_replace_one(ps);
680 if (err)
681 break;
682 }
683
684 synchronize_rcu();
685
686 list_for_each_entry_continue_reverse(ps, &pinst->pslist, list)
687 if (refcount_dec_and_test(&ps->opd->refcnt))
688 padata_free_pd(ps->opd);
689
690 pinst->flags &= ~PADATA_RESET;
691
692 return err;
693}
694
695/* If cpumask contains no active cpu, we mark the instance as invalid. */
696static bool padata_validate_cpumask(struct padata_instance *pinst,
697 const struct cpumask *cpumask)
698{
699 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
700 pinst->flags |= PADATA_INVALID;
701 return false;
702 }
703
704 pinst->flags &= ~PADATA_INVALID;
705 return true;
706}
707
708static int __padata_set_cpumasks(struct padata_instance *pinst,
709 cpumask_var_t pcpumask,
710 cpumask_var_t cbcpumask)
711{
712 int valid;
713 int err;
714
715 valid = padata_validate_cpumask(pinst, pcpumask);
716 if (!valid) {
717 __padata_stop(pinst);
718 goto out_replace;
719 }
720
721 valid = padata_validate_cpumask(pinst, cbcpumask);
722 if (!valid)
723 __padata_stop(pinst);
724
725out_replace:
726 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
727 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
728
729 err = padata_setup_cpumasks(pinst) ?: padata_replace(pinst);
730
731 if (valid)
732 __padata_start(pinst);
733
734 return err;
735}
736
737/**
738 * padata_set_cpumask - Sets specified by @cpumask_type cpumask to the value
739 * equivalent to @cpumask.
740 * @pinst: padata instance
741 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
742 * to parallel and serial cpumasks respectively.
743 * @cpumask: the cpumask to use
744 *
745 * Return: 0 on success or negative error code
746 */
747int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
748 cpumask_var_t cpumask)
749{
750 struct cpumask *serial_mask, *parallel_mask;
751 int err = -EINVAL;
752
753 cpus_read_lock();
754 mutex_lock(&pinst->lock);
755
756 switch (cpumask_type) {
757 case PADATA_CPU_PARALLEL:
758 serial_mask = pinst->cpumask.cbcpu;
759 parallel_mask = cpumask;
760 break;
761 case PADATA_CPU_SERIAL:
762 parallel_mask = pinst->cpumask.pcpu;
763 serial_mask = cpumask;
764 break;
765 default:
766 goto out;
767 }
768
769 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
770
771out:
772 mutex_unlock(&pinst->lock);
773 cpus_read_unlock();
774
775 return err;
776}
777EXPORT_SYMBOL(padata_set_cpumask);
778
779#ifdef CONFIG_HOTPLUG_CPU
780
781static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
782{
783 int err = 0;
784
785 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
786 err = padata_replace(pinst);
787
788 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
789 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
790 __padata_start(pinst);
791 }
792
793 return err;
794}
795
796static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
797{
798 int err = 0;
799
800 if (!cpumask_test_cpu(cpu, cpu_online_mask)) {
801 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
802 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
803 __padata_stop(pinst);
804
805 err = padata_replace(pinst);
806 }
807
808 return err;
809}
810
811static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
812{
813 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
814 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
815}
816
817static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
818{
819 struct padata_instance *pinst;
820 int ret;
821
822 pinst = hlist_entry_safe(node, struct padata_instance, cpu_online_node);
823 if (!pinst_has_cpu(pinst, cpu))
824 return 0;
825
826 mutex_lock(&pinst->lock);
827 ret = __padata_add_cpu(pinst, cpu);
828 mutex_unlock(&pinst->lock);
829 return ret;
830}
831
832static int padata_cpu_dead(unsigned int cpu, struct hlist_node *node)
833{
834 struct padata_instance *pinst;
835 int ret;
836
837 pinst = hlist_entry_safe(node, struct padata_instance, cpu_dead_node);
838 if (!pinst_has_cpu(pinst, cpu))
839 return 0;
840
841 mutex_lock(&pinst->lock);
842 ret = __padata_remove_cpu(pinst, cpu);
843 mutex_unlock(&pinst->lock);
844 return ret;
845}
846
847static enum cpuhp_state hp_online;
848#endif
849
850static void __padata_free(struct padata_instance *pinst)
851{
852#ifdef CONFIG_HOTPLUG_CPU
853 cpuhp_state_remove_instance_nocalls(CPUHP_PADATA_DEAD,
854 &pinst->cpu_dead_node);
855 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->cpu_online_node);
856#endif
857
858 WARN_ON(!list_empty(&pinst->pslist));
859
860 free_cpumask_var(pinst->cpumask.pcpu);
861 free_cpumask_var(pinst->cpumask.cbcpu);
862 destroy_workqueue(pinst->serial_wq);
863 destroy_workqueue(pinst->parallel_wq);
864 kfree(pinst);
865}
866
867#define kobj2pinst(_kobj) \
868 container_of(_kobj, struct padata_instance, kobj)
869#define attr2pentry(_attr) \
870 container_of(_attr, struct padata_sysfs_entry, attr)
871
872static void padata_sysfs_release(struct kobject *kobj)
873{
874 struct padata_instance *pinst = kobj2pinst(kobj);
875 __padata_free(pinst);
876}
877
878struct padata_sysfs_entry {
879 struct attribute attr;
880 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
881 ssize_t (*store)(struct padata_instance *, struct attribute *,
882 const char *, size_t);
883};
884
885static ssize_t show_cpumask(struct padata_instance *pinst,
886 struct attribute *attr, char *buf)
887{
888 struct cpumask *cpumask;
889 ssize_t len;
890
891 mutex_lock(&pinst->lock);
892 if (!strcmp(attr->name, "serial_cpumask"))
893 cpumask = pinst->cpumask.cbcpu;
894 else
895 cpumask = pinst->cpumask.pcpu;
896
897 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
898 nr_cpu_ids, cpumask_bits(cpumask));
899 mutex_unlock(&pinst->lock);
900 return len < PAGE_SIZE ? len : -EINVAL;
901}
902
903static ssize_t store_cpumask(struct padata_instance *pinst,
904 struct attribute *attr,
905 const char *buf, size_t count)
906{
907 cpumask_var_t new_cpumask;
908 ssize_t ret;
909 int mask_type;
910
911 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
912 return -ENOMEM;
913
914 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
915 nr_cpumask_bits);
916 if (ret < 0)
917 goto out;
918
919 mask_type = !strcmp(attr->name, "serial_cpumask") ?
920 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
921 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
922 if (!ret)
923 ret = count;
924
925out:
926 free_cpumask_var(new_cpumask);
927 return ret;
928}
929
930#define PADATA_ATTR_RW(_name, _show_name, _store_name) \
931 static struct padata_sysfs_entry _name##_attr = \
932 __ATTR(_name, 0644, _show_name, _store_name)
933#define PADATA_ATTR_RO(_name, _show_name) \
934 static struct padata_sysfs_entry _name##_attr = \
935 __ATTR(_name, 0400, _show_name, NULL)
936
937PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
938PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
939
940/*
941 * Padata sysfs provides the following objects:
942 * serial_cpumask [RW] - cpumask for serial workers
943 * parallel_cpumask [RW] - cpumask for parallel workers
944 */
945static struct attribute *padata_default_attrs[] = {
946 &serial_cpumask_attr.attr,
947 &parallel_cpumask_attr.attr,
948 NULL,
949};
950ATTRIBUTE_GROUPS(padata_default);
951
952static ssize_t padata_sysfs_show(struct kobject *kobj,
953 struct attribute *attr, char *buf)
954{
955 struct padata_instance *pinst;
956 struct padata_sysfs_entry *pentry;
957 ssize_t ret = -EIO;
958
959 pinst = kobj2pinst(kobj);
960 pentry = attr2pentry(attr);
961 if (pentry->show)
962 ret = pentry->show(pinst, attr, buf);
963
964 return ret;
965}
966
967static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
968 const char *buf, size_t count)
969{
970 struct padata_instance *pinst;
971 struct padata_sysfs_entry *pentry;
972 ssize_t ret = -EIO;
973
974 pinst = kobj2pinst(kobj);
975 pentry = attr2pentry(attr);
976 if (pentry->show)
977 ret = pentry->store(pinst, attr, buf, count);
978
979 return ret;
980}
981
982static const struct sysfs_ops padata_sysfs_ops = {
983 .show = padata_sysfs_show,
984 .store = padata_sysfs_store,
985};
986
987static const struct kobj_type padata_attr_type = {
988 .sysfs_ops = &padata_sysfs_ops,
989 .default_groups = padata_default_groups,
990 .release = padata_sysfs_release,
991};
992
993/**
994 * padata_alloc - allocate and initialize a padata instance
995 * @name: used to identify the instance
996 *
997 * Return: new instance on success, NULL on error
998 */
999struct padata_instance *padata_alloc(const char *name)
1000{
1001 struct padata_instance *pinst;
1002
1003 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
1004 if (!pinst)
1005 goto err;
1006
1007 pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
1008 name);
1009 if (!pinst->parallel_wq)
1010 goto err_free_inst;
1011
1012 cpus_read_lock();
1013
1014 pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
1015 WQ_CPU_INTENSIVE, 1, name);
1016 if (!pinst->serial_wq)
1017 goto err_put_cpus;
1018
1019 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
1020 goto err_free_serial_wq;
1021 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
1022 free_cpumask_var(pinst->cpumask.pcpu);
1023 goto err_free_serial_wq;
1024 }
1025
1026 INIT_LIST_HEAD(&pinst->pslist);
1027
1028 cpumask_copy(pinst->cpumask.pcpu, cpu_possible_mask);
1029 cpumask_copy(pinst->cpumask.cbcpu, cpu_possible_mask);
1030
1031 if (padata_setup_cpumasks(pinst))
1032 goto err_free_masks;
1033
1034 __padata_start(pinst);
1035
1036 kobject_init(&pinst->kobj, &padata_attr_type);
1037 mutex_init(&pinst->lock);
1038
1039#ifdef CONFIG_HOTPLUG_CPU
1040 cpuhp_state_add_instance_nocalls_cpuslocked(hp_online,
1041 &pinst->cpu_online_node);
1042 cpuhp_state_add_instance_nocalls_cpuslocked(CPUHP_PADATA_DEAD,
1043 &pinst->cpu_dead_node);
1044#endif
1045
1046 cpus_read_unlock();
1047
1048 return pinst;
1049
1050err_free_masks:
1051 free_cpumask_var(pinst->cpumask.pcpu);
1052 free_cpumask_var(pinst->cpumask.cbcpu);
1053err_free_serial_wq:
1054 destroy_workqueue(pinst->serial_wq);
1055err_put_cpus:
1056 cpus_read_unlock();
1057 destroy_workqueue(pinst->parallel_wq);
1058err_free_inst:
1059 kfree(pinst);
1060err:
1061 return NULL;
1062}
1063EXPORT_SYMBOL(padata_alloc);
1064
1065/**
1066 * padata_free - free a padata instance
1067 *
1068 * @pinst: padata instance to free
1069 */
1070void padata_free(struct padata_instance *pinst)
1071{
1072 kobject_put(&pinst->kobj);
1073}
1074EXPORT_SYMBOL(padata_free);
1075
1076/**
1077 * padata_alloc_shell - Allocate and initialize padata shell.
1078 *
1079 * @pinst: Parent padata_instance object.
1080 *
1081 * Return: new shell on success, NULL on error
1082 */
1083struct padata_shell *padata_alloc_shell(struct padata_instance *pinst)
1084{
1085 struct parallel_data *pd;
1086 struct padata_shell *ps;
1087
1088 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
1089 if (!ps)
1090 goto out;
1091
1092 ps->pinst = pinst;
1093
1094 cpus_read_lock();
1095 pd = padata_alloc_pd(ps);
1096 cpus_read_unlock();
1097
1098 if (!pd)
1099 goto out_free_ps;
1100
1101 mutex_lock(&pinst->lock);
1102 RCU_INIT_POINTER(ps->pd, pd);
1103 list_add(&ps->list, &pinst->pslist);
1104 mutex_unlock(&pinst->lock);
1105
1106 return ps;
1107
1108out_free_ps:
1109 kfree(ps);
1110out:
1111 return NULL;
1112}
1113EXPORT_SYMBOL(padata_alloc_shell);
1114
1115/**
1116 * padata_free_shell - free a padata shell
1117 *
1118 * @ps: padata shell to free
1119 */
1120void padata_free_shell(struct padata_shell *ps)
1121{
1122 struct parallel_data *pd;
1123
1124 if (!ps)
1125 return;
1126
1127 mutex_lock(&ps->pinst->lock);
1128 list_del(&ps->list);
1129 pd = rcu_dereference_protected(ps->pd, 1);
1130 if (refcount_dec_and_test(&pd->refcnt))
1131 padata_free_pd(pd);
1132 mutex_unlock(&ps->pinst->lock);
1133
1134 kfree(ps);
1135}
1136EXPORT_SYMBOL(padata_free_shell);
1137
1138void __init padata_init(void)
1139{
1140 unsigned int i, possible_cpus;
1141#ifdef CONFIG_HOTPLUG_CPU
1142 int ret;
1143
1144 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1145 padata_cpu_online, NULL);
1146 if (ret < 0)
1147 goto err;
1148 hp_online = ret;
1149
1150 ret = cpuhp_setup_state_multi(CPUHP_PADATA_DEAD, "padata:dead",
1151 NULL, padata_cpu_dead);
1152 if (ret < 0)
1153 goto remove_online_state;
1154#endif
1155
1156 possible_cpus = num_possible_cpus();
1157 padata_works = kmalloc_array(possible_cpus, sizeof(struct padata_work),
1158 GFP_KERNEL);
1159 if (!padata_works)
1160 goto remove_dead_state;
1161
1162 for (i = 0; i < possible_cpus; ++i)
1163 list_add(&padata_works[i].pw_list, &padata_free_works);
1164
1165 return;
1166
1167remove_dead_state:
1168#ifdef CONFIG_HOTPLUG_CPU
1169 cpuhp_remove_multi_state(CPUHP_PADATA_DEAD);
1170remove_online_state:
1171 cpuhp_remove_multi_state(hp_online);
1172err:
1173#endif
1174 pr_warn("padata: initialization failed\n");
1175}