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
28 struct padata_work {
29 struct work_struct pw_work;
30 struct list_head pw_list; /* padata_free_works linkage */
31 void *pw_data;
32 };
33
34 static DEFINE_SPINLOCK(padata_works_lock);
35 static struct padata_work *padata_works;
36 static LIST_HEAD(padata_free_works);
37
38 struct 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
47 static void padata_free_pd(struct parallel_data *pd);
48 static void __init padata_mt_helper(struct work_struct *work);
49
50 static 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
61 static 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
72 static 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 */
94 static 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
104 static 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
124 static 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
130 static 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
145 static 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 */
175 int 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;
231 out:
232 rcu_read_unlock_bh();
233
234 return err;
235 }
236 EXPORT_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 */
248 static 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
284 static 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
343 static 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
353 static 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 */
395 void 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 /* Compare by difference to consider integer wrap around */
408 if ((signed int)(cur->seq_nr - padata->seq_nr) < 0)
409 break;
410 }
411 list_add(&padata->list, pos);
412 spin_unlock(&reorder->lock);
413
414 /*
415 * Ensure the addition to the reorder list is ordered correctly
416 * with the trylock of pd->lock in padata_reorder. Pairs with smp_mb
417 * in padata_reorder.
418 */
419 smp_mb();
420
421 padata_reorder(pd);
422 }
423 EXPORT_SYMBOL(padata_do_serial);
424
425 static int padata_setup_cpumasks(struct padata_instance *pinst)
426 {
427 struct workqueue_attrs *attrs;
428 int err;
429
430 attrs = alloc_workqueue_attrs();
431 if (!attrs)
432 return -ENOMEM;
433
434 /* Restrict parallel_wq workers to pd->cpumask.pcpu. */
435 cpumask_copy(attrs->cpumask, pinst->cpumask.pcpu);
436 err = apply_workqueue_attrs(pinst->parallel_wq, attrs);
437 free_workqueue_attrs(attrs);
438
439 return err;
440 }
441
442 static void __init padata_mt_helper(struct work_struct *w)
443 {
444 struct padata_work *pw = container_of(w, struct padata_work, pw_work);
445 struct padata_mt_job_state *ps = pw->pw_data;
446 struct padata_mt_job *job = ps->job;
447 bool done;
448
449 spin_lock(&ps->lock);
450
451 while (job->size > 0) {
452 unsigned long start, size, end;
453
454 start = job->start;
455 /* So end is chunk size aligned if enough work remains. */
456 size = roundup(start + 1, ps->chunk_size) - start;
457 size = min(size, job->size);
458 end = start + size;
459
460 job->start = end;
461 job->size -= size;
462
463 spin_unlock(&ps->lock);
464 job->thread_fn(start, end, job->fn_arg);
465 spin_lock(&ps->lock);
466 }
467
468 ++ps->nworks_fini;
469 done = (ps->nworks_fini == ps->nworks);
470 spin_unlock(&ps->lock);
471
472 if (done)
473 complete(&ps->completion);
474 }
475
476 /**
477 * padata_do_multithreaded - run a multithreaded job
478 * @job: Description of the job.
479 *
480 * See the definition of struct padata_mt_job for more details.
481 */
482 void __init padata_do_multithreaded(struct padata_mt_job *job)
483 {
484 /* In case threads finish at different times. */
485 static const unsigned long load_balance_factor = 4;
486 struct padata_work my_work, *pw;
487 struct padata_mt_job_state ps;
488 LIST_HEAD(works);
489 int nworks, nid;
490 static atomic_t last_used_nid __initdata;
491
492 if (job->size == 0)
493 return;
494
495 /* Ensure at least one thread when size < min_chunk. */
496 nworks = max(job->size / max(job->min_chunk, job->align), 1ul);
497 nworks = min(nworks, job->max_threads);
498
499 if (nworks == 1) {
500 /* Single thread, no coordination needed, cut to the chase. */
501 job->thread_fn(job->start, job->start + job->size, job->fn_arg);
502 return;
503 }
504
505 spin_lock_init(&ps.lock);
506 init_completion(&ps.completion);
507 ps.job = job;
508 ps.nworks = padata_work_alloc_mt(nworks, &ps, &works);
509 ps.nworks_fini = 0;
510
511 /*
512 * Chunk size is the amount of work a helper does per call to the
513 * thread function. Load balance large jobs between threads by
514 * increasing the number of chunks, guarantee at least the minimum
515 * chunk size from the caller, and honor the caller's alignment.
516 * Ensure chunk_size is at least 1 to prevent divide-by-0
517 * panic in padata_mt_helper().
518 */
519 ps.chunk_size = job->size / (ps.nworks * load_balance_factor);
520 ps.chunk_size = max(ps.chunk_size, job->min_chunk);
521 ps.chunk_size = max(ps.chunk_size, 1ul);
522 ps.chunk_size = roundup(ps.chunk_size, job->align);
523
524 /*
525 * chunk_size can be 0 if the caller sets min_chunk to 0. So force it
526 * to at least 1 to prevent divide-by-0 panic in padata_mt_helper().`
527 */
528 if (!ps.chunk_size)
529 ps.chunk_size = 1U;
530
531 list_for_each_entry(pw, &works, pw_list)
532 if (job->numa_aware) {
533 int old_node = atomic_read(&last_used_nid);
534
535 do {
536 nid = next_node_in(old_node, node_states[N_CPU]);
537 } while (!atomic_try_cmpxchg(&last_used_nid, &old_node, nid));
538 queue_work_node(nid, system_unbound_wq, &pw->pw_work);
539 } else {
540 queue_work(system_unbound_wq, &pw->pw_work);
541 }
542
543 /* Use the current thread, which saves starting a workqueue worker. */
544 padata_work_init(&my_work, padata_mt_helper, &ps, PADATA_WORK_ONSTACK);
545 padata_mt_helper(&my_work.pw_work);
546
547 /* Wait for all the helpers to finish. */
548 wait_for_completion(&ps.completion);
549
550 destroy_work_on_stack(&my_work.pw_work);
551 padata_works_free(&works);
552 }
553
554 static void __padata_list_init(struct padata_list *pd_list)
555 {
556 INIT_LIST_HEAD(&pd_list->list);
557 spin_lock_init(&pd_list->lock);
558 }
559
560 /* Initialize all percpu queues used by serial workers */
561 static void padata_init_squeues(struct parallel_data *pd)
562 {
563 int cpu;
564 struct padata_serial_queue *squeue;
565
566 for_each_cpu(cpu, pd->cpumask.cbcpu) {
567 squeue = per_cpu_ptr(pd->squeue, cpu);
568 squeue->pd = pd;
569 __padata_list_init(&squeue->serial);
570 INIT_WORK(&squeue->work, padata_serial_worker);
571 }
572 }
573
574 /* Initialize per-CPU reorder lists */
575 static void padata_init_reorder_list(struct parallel_data *pd)
576 {
577 int cpu;
578 struct padata_list *list;
579
580 for_each_cpu(cpu, pd->cpumask.pcpu) {
581 list = per_cpu_ptr(pd->reorder_list, cpu);
582 __padata_list_init(list);
583 }
584 }
585
586 /* Allocate and initialize the internal cpumask dependend resources. */
587 static struct parallel_data *padata_alloc_pd(struct padata_shell *ps)
588 {
589 struct padata_instance *pinst = ps->pinst;
590 struct parallel_data *pd;
591
592 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
593 if (!pd)
594 goto err;
595
596 pd->reorder_list = alloc_percpu(struct padata_list);
597 if (!pd->reorder_list)
598 goto err_free_pd;
599
600 pd->squeue = alloc_percpu(struct padata_serial_queue);
601 if (!pd->squeue)
602 goto err_free_reorder_list;
603
604 pd->ps = ps;
605
606 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
607 goto err_free_squeue;
608 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
609 goto err_free_pcpu;
610
611 cpumask_and(pd->cpumask.pcpu, pinst->cpumask.pcpu, cpu_online_mask);
612 cpumask_and(pd->cpumask.cbcpu, pinst->cpumask.cbcpu, cpu_online_mask);
613
614 padata_init_reorder_list(pd);
615 padata_init_squeues(pd);
616 pd->seq_nr = -1;
617 refcount_set(&pd->refcnt, 1);
618 spin_lock_init(&pd->lock);
619 pd->cpu = cpumask_first(pd->cpumask.pcpu);
620 INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
621
622 return pd;
623
624 err_free_pcpu:
625 free_cpumask_var(pd->cpumask.pcpu);
626 err_free_squeue:
627 free_percpu(pd->squeue);
628 err_free_reorder_list:
629 free_percpu(pd->reorder_list);
630 err_free_pd:
631 kfree(pd);
632 err:
633 return NULL;
634 }
635
636 static void padata_free_pd(struct parallel_data *pd)
637 {
638 free_cpumask_var(pd->cpumask.pcpu);
639 free_cpumask_var(pd->cpumask.cbcpu);
640 free_percpu(pd->reorder_list);
641 free_percpu(pd->squeue);
642 kfree(pd);
643 }
644
645 static void __padata_start(struct padata_instance *pinst)
646 {
647 pinst->flags |= PADATA_INIT;
648 }
649
650 static void __padata_stop(struct padata_instance *pinst)
651 {
652 if (!(pinst->flags & PADATA_INIT))
653 return;
654
655 pinst->flags &= ~PADATA_INIT;
656
657 synchronize_rcu();
658 }
659
660 /* Replace the internal control structure with a new one. */
661 static int padata_replace_one(struct padata_shell *ps)
662 {
663 struct parallel_data *pd_new;
664
665 pd_new = padata_alloc_pd(ps);
666 if (!pd_new)
667 return -ENOMEM;
668
669 ps->opd = rcu_dereference_protected(ps->pd, 1);
670 rcu_assign_pointer(ps->pd, pd_new);
671
672 return 0;
673 }
674
675 static int padata_replace(struct padata_instance *pinst)
676 {
677 struct padata_shell *ps;
678 int err = 0;
679
680 pinst->flags |= PADATA_RESET;
681
682 list_for_each_entry(ps, &pinst->pslist, list) {
683 err = padata_replace_one(ps);
684 if (err)
685 break;
686 }
687
688 synchronize_rcu();
689
690 list_for_each_entry_continue_reverse(ps, &pinst->pslist, list)
691 if (refcount_dec_and_test(&ps->opd->refcnt))
692 padata_free_pd(ps->opd);
693
694 pinst->flags &= ~PADATA_RESET;
695
696 return err;
697 }
698
699 /* If cpumask contains no active cpu, we mark the instance as invalid. */
700 static bool padata_validate_cpumask(struct padata_instance *pinst,
701 const struct cpumask *cpumask)
702 {
703 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
704 pinst->flags |= PADATA_INVALID;
705 return false;
706 }
707
708 pinst->flags &= ~PADATA_INVALID;
709 return true;
710 }
711
712 static int __padata_set_cpumasks(struct padata_instance *pinst,
713 cpumask_var_t pcpumask,
714 cpumask_var_t cbcpumask)
715 {
716 int valid;
717 int err;
718
719 valid = padata_validate_cpumask(pinst, pcpumask);
720 if (!valid) {
721 __padata_stop(pinst);
722 goto out_replace;
723 }
724
725 valid = padata_validate_cpumask(pinst, cbcpumask);
726 if (!valid)
727 __padata_stop(pinst);
728
729 out_replace:
730 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
731 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
732
733 err = padata_setup_cpumasks(pinst) ?: padata_replace(pinst);
734
735 if (valid)
736 __padata_start(pinst);
737
738 return err;
739 }
740
741 /**
742 * padata_set_cpumask - Sets specified by @cpumask_type cpumask to the value
743 * equivalent to @cpumask.
744 * @pinst: padata instance
745 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
746 * to parallel and serial cpumasks respectively.
747 * @cpumask: the cpumask to use
748 *
749 * Return: 0 on success or negative error code
750 */
751 int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
752 cpumask_var_t cpumask)
753 {
754 struct cpumask *serial_mask, *parallel_mask;
755 int err = -EINVAL;
756
757 cpus_read_lock();
758 mutex_lock(&pinst->lock);
759
760 switch (cpumask_type) {
761 case PADATA_CPU_PARALLEL:
762 serial_mask = pinst->cpumask.cbcpu;
763 parallel_mask = cpumask;
764 break;
765 case PADATA_CPU_SERIAL:
766 parallel_mask = pinst->cpumask.pcpu;
767 serial_mask = cpumask;
768 break;
769 default:
770 goto out;
771 }
772
773 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
774
775 out:
776 mutex_unlock(&pinst->lock);
777 cpus_read_unlock();
778
779 return err;
780 }
781 EXPORT_SYMBOL(padata_set_cpumask);
782
783 #ifdef CONFIG_HOTPLUG_CPU
784
785 static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
786 {
787 int err = 0;
788
789 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
790 err = padata_replace(pinst);
791
792 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
793 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
794 __padata_start(pinst);
795 }
796
797 return err;
798 }
799
800 static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
801 {
802 int err = 0;
803
804 if (!cpumask_test_cpu(cpu, cpu_online_mask)) {
805 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
806 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
807 __padata_stop(pinst);
808
809 err = padata_replace(pinst);
810 }
811
812 return err;
813 }
814
815 static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
816 {
817 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
818 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
819 }
820
821 static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
822 {
823 struct padata_instance *pinst;
824 int ret;
825
826 pinst = hlist_entry_safe(node, struct padata_instance, cpu_online_node);
827 if (!pinst_has_cpu(pinst, cpu))
828 return 0;
829
830 mutex_lock(&pinst->lock);
831 ret = __padata_add_cpu(pinst, cpu);
832 mutex_unlock(&pinst->lock);
833 return ret;
834 }
835
836 static int padata_cpu_dead(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_dead_node);
842 if (!pinst_has_cpu(pinst, cpu))
843 return 0;
844
845 mutex_lock(&pinst->lock);
846 ret = __padata_remove_cpu(pinst, cpu);
847 mutex_unlock(&pinst->lock);
848 return ret;
849 }
850
851 static enum cpuhp_state hp_online;
852 #endif
853
854 static void __padata_free(struct padata_instance *pinst)
855 {
856 #ifdef CONFIG_HOTPLUG_CPU
857 cpuhp_state_remove_instance_nocalls(CPUHP_PADATA_DEAD,
858 &pinst->cpu_dead_node);
859 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->cpu_online_node);
860 #endif
861
862 WARN_ON(!list_empty(&pinst->pslist));
863
864 free_cpumask_var(pinst->cpumask.pcpu);
865 free_cpumask_var(pinst->cpumask.cbcpu);
866 destroy_workqueue(pinst->serial_wq);
867 destroy_workqueue(pinst->parallel_wq);
868 kfree(pinst);
869 }
870
871 #define kobj2pinst(_kobj) \
872 container_of(_kobj, struct padata_instance, kobj)
873 #define attr2pentry(_attr) \
874 container_of(_attr, struct padata_sysfs_entry, attr)
875
876 static void padata_sysfs_release(struct kobject *kobj)
877 {
878 struct padata_instance *pinst = kobj2pinst(kobj);
879 __padata_free(pinst);
880 }
881
882 struct padata_sysfs_entry {
883 struct attribute attr;
884 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
885 ssize_t (*store)(struct padata_instance *, struct attribute *,
886 const char *, size_t);
887 };
888
889 static ssize_t show_cpumask(struct padata_instance *pinst,
890 struct attribute *attr, char *buf)
891 {
892 struct cpumask *cpumask;
893 ssize_t len;
894
895 mutex_lock(&pinst->lock);
896 if (!strcmp(attr->name, "serial_cpumask"))
897 cpumask = pinst->cpumask.cbcpu;
898 else
899 cpumask = pinst->cpumask.pcpu;
900
901 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
902 nr_cpu_ids, cpumask_bits(cpumask));
903 mutex_unlock(&pinst->lock);
904 return len < PAGE_SIZE ? len : -EINVAL;
905 }
906
907 static ssize_t store_cpumask(struct padata_instance *pinst,
908 struct attribute *attr,
909 const char *buf, size_t count)
910 {
911 cpumask_var_t new_cpumask;
912 ssize_t ret;
913 int mask_type;
914
915 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
916 return -ENOMEM;
917
918 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
919 nr_cpumask_bits);
920 if (ret < 0)
921 goto out;
922
923 mask_type = !strcmp(attr->name, "serial_cpumask") ?
924 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
925 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
926 if (!ret)
927 ret = count;
928
929 out:
930 free_cpumask_var(new_cpumask);
931 return ret;
932 }
933
934 #define PADATA_ATTR_RW(_name, _show_name, _store_name) \
935 static struct padata_sysfs_entry _name##_attr = \
936 __ATTR(_name, 0644, _show_name, _store_name)
937 #define PADATA_ATTR_RO(_name, _show_name) \
938 static struct padata_sysfs_entry _name##_attr = \
939 __ATTR(_name, 0400, _show_name, NULL)
940
941 PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
942 PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
943
944 /*
945 * Padata sysfs provides the following objects:
946 * serial_cpumask [RW] - cpumask for serial workers
947 * parallel_cpumask [RW] - cpumask for parallel workers
948 */
949 static struct attribute *padata_default_attrs[] = {
950 &serial_cpumask_attr.attr,
951 ¶llel_cpumask_attr.attr,
952 NULL,
953 };
954 ATTRIBUTE_GROUPS(padata_default);
955
956 static ssize_t padata_sysfs_show(struct kobject *kobj,
957 struct attribute *attr, char *buf)
958 {
959 struct padata_instance *pinst;
960 struct padata_sysfs_entry *pentry;
961 ssize_t ret = -EIO;
962
963 pinst = kobj2pinst(kobj);
964 pentry = attr2pentry(attr);
965 if (pentry->show)
966 ret = pentry->show(pinst, attr, buf);
967
968 return ret;
969 }
970
971 static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
972 const char *buf, size_t count)
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->store(pinst, attr, buf, count);
982
983 return ret;
984 }
985
986 static const struct sysfs_ops padata_sysfs_ops = {
987 .show = padata_sysfs_show,
988 .store = padata_sysfs_store,
989 };
990
991 static const struct kobj_type padata_attr_type = {
992 .sysfs_ops = &padata_sysfs_ops,
993 .default_groups = padata_default_groups,
994 .release = padata_sysfs_release,
995 };
996
997 /**
998 * padata_alloc - allocate and initialize a padata instance
999 * @name: used to identify the instance
1000 *
1001 * Return: new instance on success, NULL on error
1002 */
1003 struct padata_instance *padata_alloc(const char *name)
1004 {
1005 struct padata_instance *pinst;
1006
1007 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
1008 if (!pinst)
1009 goto err;
1010
1011 pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
1012 name);
1013 if (!pinst->parallel_wq)
1014 goto err_free_inst;
1015
1016 cpus_read_lock();
1017
1018 pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
1019 WQ_CPU_INTENSIVE, 1, name);
1020 if (!pinst->serial_wq)
1021 goto err_put_cpus;
1022
1023 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
1024 goto err_free_serial_wq;
1025 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
1026 free_cpumask_var(pinst->cpumask.pcpu);
1027 goto err_free_serial_wq;
1028 }
1029
1030 INIT_LIST_HEAD(&pinst->pslist);
1031
1032 cpumask_copy(pinst->cpumask.pcpu, cpu_possible_mask);
1033 cpumask_copy(pinst->cpumask.cbcpu, cpu_possible_mask);
1034
1035 if (padata_setup_cpumasks(pinst))
1036 goto err_free_masks;
1037
1038 __padata_start(pinst);
1039
1040 kobject_init(&pinst->kobj, &padata_attr_type);
1041 mutex_init(&pinst->lock);
1042
1043 #ifdef CONFIG_HOTPLUG_CPU
1044 cpuhp_state_add_instance_nocalls_cpuslocked(hp_online,
1045 &pinst->cpu_online_node);
1046 cpuhp_state_add_instance_nocalls_cpuslocked(CPUHP_PADATA_DEAD,
1047 &pinst->cpu_dead_node);
1048 #endif
1049
1050 cpus_read_unlock();
1051
1052 return pinst;
1053
1054 err_free_masks:
1055 free_cpumask_var(pinst->cpumask.pcpu);
1056 free_cpumask_var(pinst->cpumask.cbcpu);
1057 err_free_serial_wq:
1058 destroy_workqueue(pinst->serial_wq);
1059 err_put_cpus:
1060 cpus_read_unlock();
1061 destroy_workqueue(pinst->parallel_wq);
1062 err_free_inst:
1063 kfree(pinst);
1064 err:
1065 return NULL;
1066 }
1067 EXPORT_SYMBOL(padata_alloc);
1068
1069 /**
1070 * padata_free - free a padata instance
1071 *
1072 * @pinst: padata instance to free
1073 */
1074 void padata_free(struct padata_instance *pinst)
1075 {
1076 kobject_put(&pinst->kobj);
1077 }
1078 EXPORT_SYMBOL(padata_free);
1079
1080 /**
1081 * padata_alloc_shell - Allocate and initialize padata shell.
1082 *
1083 * @pinst: Parent padata_instance object.
1084 *
1085 * Return: new shell on success, NULL on error
1086 */
1087 struct padata_shell *padata_alloc_shell(struct padata_instance *pinst)
1088 {
1089 struct parallel_data *pd;
1090 struct padata_shell *ps;
1091
1092 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
1093 if (!ps)
1094 goto out;
1095
1096 ps->pinst = pinst;
1097
1098 cpus_read_lock();
1099 pd = padata_alloc_pd(ps);
1100 cpus_read_unlock();
1101
1102 if (!pd)
1103 goto out_free_ps;
1104
1105 mutex_lock(&pinst->lock);
1106 RCU_INIT_POINTER(ps->pd, pd);
1107 list_add(&ps->list, &pinst->pslist);
1108 mutex_unlock(&pinst->lock);
1109
1110 return ps;
1111
1112 out_free_ps:
1113 kfree(ps);
1114 out:
1115 return NULL;
1116 }
1117 EXPORT_SYMBOL(padata_alloc_shell);
1118
1119 /**
1120 * padata_free_shell - free a padata shell
1121 *
1122 * @ps: padata shell to free
1123 */
1124 void padata_free_shell(struct padata_shell *ps)
1125 {
1126 struct parallel_data *pd;
1127
1128 if (!ps)
1129 return;
1130
1131 mutex_lock(&ps->pinst->lock);
1132 list_del(&ps->list);
1133 pd = rcu_dereference_protected(ps->pd, 1);
1134 if (refcount_dec_and_test(&pd->refcnt))
1135 padata_free_pd(pd);
1136 mutex_unlock(&ps->pinst->lock);
1137
1138 kfree(ps);
1139 }
1140 EXPORT_SYMBOL(padata_free_shell);
1141
1142 void __init padata_init(void)
1143 {
1144 unsigned int i, possible_cpus;
1145 #ifdef CONFIG_HOTPLUG_CPU
1146 int ret;
1147
1148 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1149 padata_cpu_online, NULL);
1150 if (ret < 0)
1151 goto err;
1152 hp_online = ret;
1153
1154 ret = cpuhp_setup_state_multi(CPUHP_PADATA_DEAD, "padata:dead",
1155 NULL, padata_cpu_dead);
1156 if (ret < 0)
1157 goto remove_online_state;
1158 #endif
1159
1160 possible_cpus = num_possible_cpus();
1161 padata_works = kmalloc_array(possible_cpus, sizeof(struct padata_work),
1162 GFP_KERNEL);
1163 if (!padata_works)
1164 goto remove_dead_state;
1165
1166 for (i = 0; i < possible_cpus; ++i)
1167 list_add(&padata_works[i].pw_list, &padata_free_works);
1168
1169 return;
1170
1171 remove_dead_state:
1172 #ifdef CONFIG_HOTPLUG_CPU
1173 cpuhp_remove_multi_state(CPUHP_PADATA_DEAD);
1174 remove_online_state:
1175 cpuhp_remove_multi_state(hp_online);
1176 err:
1177 #endif
1178 pr_warn("padata: initialization failed\n");
1179 }
1180
Linux® is a registered trademark of Linus Torvalds in the United States and other countries.
TOMOYO® is a registered trademark of NTT DATA CORPORATION.