xref: /openbmc/linux/kernel/padata.c (revision fadbafc1)
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 static void padata_work_init(struct padata_work *pw, work_func_t work_fn,
87 			     void *data, int flags)
88 {
89 	if (flags & PADATA_WORK_ONSTACK)
90 		INIT_WORK_ONSTACK(&pw->pw_work, work_fn);
91 	else
92 		INIT_WORK(&pw->pw_work, work_fn);
93 	pw->pw_data = data;
94 }
95 
96 static int __init padata_work_alloc_mt(int nworks, void *data,
97 				       struct list_head *head)
98 {
99 	int i;
100 
101 	spin_lock(&padata_works_lock);
102 	/* Start at 1 because the current task participates in the job. */
103 	for (i = 1; i < nworks; ++i) {
104 		struct padata_work *pw = padata_work_alloc();
105 
106 		if (!pw)
107 			break;
108 		padata_work_init(pw, padata_mt_helper, data, 0);
109 		list_add(&pw->pw_list, head);
110 	}
111 	spin_unlock(&padata_works_lock);
112 
113 	return i;
114 }
115 
116 static void padata_work_free(struct padata_work *pw)
117 {
118 	lockdep_assert_held(&padata_works_lock);
119 	list_add(&pw->pw_list, &padata_free_works);
120 }
121 
122 static void __init padata_works_free(struct list_head *works)
123 {
124 	struct padata_work *cur, *next;
125 
126 	if (list_empty(works))
127 		return;
128 
129 	spin_lock(&padata_works_lock);
130 	list_for_each_entry_safe(cur, next, works, pw_list) {
131 		list_del(&cur->pw_list);
132 		padata_work_free(cur);
133 	}
134 	spin_unlock(&padata_works_lock);
135 }
136 
137 static void padata_parallel_worker(struct work_struct *parallel_work)
138 {
139 	struct padata_work *pw = container_of(parallel_work, struct padata_work,
140 					      pw_work);
141 	struct padata_priv *padata = pw->pw_data;
142 
143 	local_bh_disable();
144 	padata->parallel(padata);
145 	spin_lock(&padata_works_lock);
146 	padata_work_free(pw);
147 	spin_unlock(&padata_works_lock);
148 	local_bh_enable();
149 }
150 
151 /**
152  * padata_do_parallel - padata parallelization function
153  *
154  * @ps: padatashell
155  * @padata: object to be parallelized
156  * @cb_cpu: pointer to the CPU that the serialization callback function should
157  *          run on.  If it's not in the serial cpumask of @pinst
158  *          (i.e. cpumask.cbcpu), this function selects a fallback CPU and if
159  *          none found, returns -EINVAL.
160  *
161  * The parallelization callback function will run with BHs off.
162  * Note: Every object which is parallelized by padata_do_parallel
163  * must be seen by padata_do_serial.
164  *
165  * Return: 0 on success or else negative error code.
166  */
167 int padata_do_parallel(struct padata_shell *ps,
168 		       struct padata_priv *padata, int *cb_cpu)
169 {
170 	struct padata_instance *pinst = ps->pinst;
171 	int i, cpu, cpu_index, err;
172 	struct parallel_data *pd;
173 	struct padata_work *pw;
174 
175 	rcu_read_lock_bh();
176 
177 	pd = rcu_dereference_bh(ps->pd);
178 
179 	err = -EINVAL;
180 	if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
181 		goto out;
182 
183 	if (!cpumask_test_cpu(*cb_cpu, pd->cpumask.cbcpu)) {
184 		if (cpumask_empty(pd->cpumask.cbcpu))
185 			goto out;
186 
187 		/* Select an alternate fallback CPU and notify the caller. */
188 		cpu_index = *cb_cpu % cpumask_weight(pd->cpumask.cbcpu);
189 
190 		cpu = cpumask_first(pd->cpumask.cbcpu);
191 		for (i = 0; i < cpu_index; i++)
192 			cpu = cpumask_next(cpu, pd->cpumask.cbcpu);
193 
194 		*cb_cpu = cpu;
195 	}
196 
197 	err =  -EBUSY;
198 	if ((pinst->flags & PADATA_RESET))
199 		goto out;
200 
201 	refcount_inc(&pd->refcnt);
202 	padata->pd = pd;
203 	padata->cb_cpu = *cb_cpu;
204 
205 	spin_lock(&padata_works_lock);
206 	padata->seq_nr = ++pd->seq_nr;
207 	pw = padata_work_alloc();
208 	spin_unlock(&padata_works_lock);
209 
210 	rcu_read_unlock_bh();
211 
212 	if (pw) {
213 		padata_work_init(pw, padata_parallel_worker, padata, 0);
214 		queue_work(pinst->parallel_wq, &pw->pw_work);
215 	} else {
216 		/* Maximum works limit exceeded, run in the current task. */
217 		padata->parallel(padata);
218 	}
219 
220 	return 0;
221 out:
222 	rcu_read_unlock_bh();
223 
224 	return err;
225 }
226 EXPORT_SYMBOL(padata_do_parallel);
227 
228 /*
229  * padata_find_next - Find the next object that needs serialization.
230  *
231  * Return:
232  * * A pointer to the control struct of the next object that needs
233  *   serialization, if present in one of the percpu reorder queues.
234  * * NULL, if the next object that needs serialization will
235  *   be parallel processed by another cpu and is not yet present in
236  *   the cpu's reorder queue.
237  */
238 static struct padata_priv *padata_find_next(struct parallel_data *pd,
239 					    bool remove_object)
240 {
241 	struct padata_priv *padata;
242 	struct padata_list *reorder;
243 	int cpu = pd->cpu;
244 
245 	reorder = per_cpu_ptr(pd->reorder_list, cpu);
246 
247 	spin_lock(&reorder->lock);
248 	if (list_empty(&reorder->list)) {
249 		spin_unlock(&reorder->lock);
250 		return NULL;
251 	}
252 
253 	padata = list_entry(reorder->list.next, struct padata_priv, list);
254 
255 	/*
256 	 * Checks the rare case where two or more parallel jobs have hashed to
257 	 * the same CPU and one of the later ones finishes first.
258 	 */
259 	if (padata->seq_nr != pd->processed) {
260 		spin_unlock(&reorder->lock);
261 		return NULL;
262 	}
263 
264 	if (remove_object) {
265 		list_del_init(&padata->list);
266 		++pd->processed;
267 		pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, -1, false);
268 	}
269 
270 	spin_unlock(&reorder->lock);
271 	return padata;
272 }
273 
274 static void padata_reorder(struct parallel_data *pd)
275 {
276 	struct padata_instance *pinst = pd->ps->pinst;
277 	int cb_cpu;
278 	struct padata_priv *padata;
279 	struct padata_serial_queue *squeue;
280 	struct padata_list *reorder;
281 
282 	/*
283 	 * We need to ensure that only one cpu can work on dequeueing of
284 	 * the reorder queue the time. Calculating in which percpu reorder
285 	 * queue the next object will arrive takes some time. A spinlock
286 	 * would be highly contended. Also it is not clear in which order
287 	 * the objects arrive to the reorder queues. So a cpu could wait to
288 	 * get the lock just to notice that there is nothing to do at the
289 	 * moment. Therefore we use a trylock and let the holder of the lock
290 	 * care for all the objects enqueued during the holdtime of the lock.
291 	 */
292 	if (!spin_trylock_bh(&pd->lock))
293 		return;
294 
295 	while (1) {
296 		padata = padata_find_next(pd, true);
297 
298 		/*
299 		 * If the next object that needs serialization is parallel
300 		 * processed by another cpu and is still on it's way to the
301 		 * cpu's reorder queue, nothing to do for now.
302 		 */
303 		if (!padata)
304 			break;
305 
306 		cb_cpu = padata->cb_cpu;
307 		squeue = per_cpu_ptr(pd->squeue, cb_cpu);
308 
309 		spin_lock(&squeue->serial.lock);
310 		list_add_tail(&padata->list, &squeue->serial.list);
311 		spin_unlock(&squeue->serial.lock);
312 
313 		queue_work_on(cb_cpu, pinst->serial_wq, &squeue->work);
314 	}
315 
316 	spin_unlock_bh(&pd->lock);
317 
318 	/*
319 	 * The next object that needs serialization might have arrived to
320 	 * the reorder queues in the meantime.
321 	 *
322 	 * Ensure reorder queue is read after pd->lock is dropped so we see
323 	 * new objects from another task in padata_do_serial.  Pairs with
324 	 * smp_mb in padata_do_serial.
325 	 */
326 	smp_mb();
327 
328 	reorder = per_cpu_ptr(pd->reorder_list, pd->cpu);
329 	if (!list_empty(&reorder->list) && padata_find_next(pd, false))
330 		queue_work(pinst->serial_wq, &pd->reorder_work);
331 }
332 
333 static void invoke_padata_reorder(struct work_struct *work)
334 {
335 	struct parallel_data *pd;
336 
337 	local_bh_disable();
338 	pd = container_of(work, struct parallel_data, reorder_work);
339 	padata_reorder(pd);
340 	local_bh_enable();
341 }
342 
343 static void padata_serial_worker(struct work_struct *serial_work)
344 {
345 	struct padata_serial_queue *squeue;
346 	struct parallel_data *pd;
347 	LIST_HEAD(local_list);
348 	int cnt;
349 
350 	local_bh_disable();
351 	squeue = container_of(serial_work, struct padata_serial_queue, work);
352 	pd = squeue->pd;
353 
354 	spin_lock(&squeue->serial.lock);
355 	list_replace_init(&squeue->serial.list, &local_list);
356 	spin_unlock(&squeue->serial.lock);
357 
358 	cnt = 0;
359 
360 	while (!list_empty(&local_list)) {
361 		struct padata_priv *padata;
362 
363 		padata = list_entry(local_list.next,
364 				    struct padata_priv, list);
365 
366 		list_del_init(&padata->list);
367 
368 		padata->serial(padata);
369 		cnt++;
370 	}
371 	local_bh_enable();
372 
373 	if (refcount_sub_and_test(cnt, &pd->refcnt))
374 		padata_free_pd(pd);
375 }
376 
377 /**
378  * padata_do_serial - padata serialization function
379  *
380  * @padata: object to be serialized.
381  *
382  * padata_do_serial must be called for every parallelized object.
383  * The serialization callback function will run with BHs off.
384  */
385 void padata_do_serial(struct padata_priv *padata)
386 {
387 	struct parallel_data *pd = padata->pd;
388 	int hashed_cpu = padata_cpu_hash(pd, padata->seq_nr);
389 	struct padata_list *reorder = per_cpu_ptr(pd->reorder_list, hashed_cpu);
390 	struct padata_priv *cur;
391 
392 	spin_lock(&reorder->lock);
393 	/* Sort in ascending order of sequence number. */
394 	list_for_each_entry_reverse(cur, &reorder->list, list)
395 		if (cur->seq_nr < padata->seq_nr)
396 			break;
397 	list_add(&padata->list, &cur->list);
398 	spin_unlock(&reorder->lock);
399 
400 	/*
401 	 * Ensure the addition to the reorder list is ordered correctly
402 	 * with the trylock of pd->lock in padata_reorder.  Pairs with smp_mb
403 	 * in padata_reorder.
404 	 */
405 	smp_mb();
406 
407 	padata_reorder(pd);
408 }
409 EXPORT_SYMBOL(padata_do_serial);
410 
411 static int padata_setup_cpumasks(struct padata_instance *pinst)
412 {
413 	struct workqueue_attrs *attrs;
414 	int err;
415 
416 	attrs = alloc_workqueue_attrs();
417 	if (!attrs)
418 		return -ENOMEM;
419 
420 	/* Restrict parallel_wq workers to pd->cpumask.pcpu. */
421 	cpumask_copy(attrs->cpumask, pinst->cpumask.pcpu);
422 	err = apply_workqueue_attrs(pinst->parallel_wq, attrs);
423 	free_workqueue_attrs(attrs);
424 
425 	return err;
426 }
427 
428 static void __init padata_mt_helper(struct work_struct *w)
429 {
430 	struct padata_work *pw = container_of(w, struct padata_work, pw_work);
431 	struct padata_mt_job_state *ps = pw->pw_data;
432 	struct padata_mt_job *job = ps->job;
433 	bool done;
434 
435 	spin_lock(&ps->lock);
436 
437 	while (job->size > 0) {
438 		unsigned long start, size, end;
439 
440 		start = job->start;
441 		/* So end is chunk size aligned if enough work remains. */
442 		size = roundup(start + 1, ps->chunk_size) - start;
443 		size = min(size, job->size);
444 		end = start + size;
445 
446 		job->start = end;
447 		job->size -= size;
448 
449 		spin_unlock(&ps->lock);
450 		job->thread_fn(start, end, job->fn_arg);
451 		spin_lock(&ps->lock);
452 	}
453 
454 	++ps->nworks_fini;
455 	done = (ps->nworks_fini == ps->nworks);
456 	spin_unlock(&ps->lock);
457 
458 	if (done)
459 		complete(&ps->completion);
460 }
461 
462 /**
463  * padata_do_multithreaded - run a multithreaded job
464  * @job: Description of the job.
465  *
466  * See the definition of struct padata_mt_job for more details.
467  */
468 void __init padata_do_multithreaded(struct padata_mt_job *job)
469 {
470 	/* In case threads finish at different times. */
471 	static const unsigned long load_balance_factor = 4;
472 	struct padata_work my_work, *pw;
473 	struct padata_mt_job_state ps;
474 	LIST_HEAD(works);
475 	int nworks;
476 
477 	if (job->size == 0)
478 		return;
479 
480 	/* Ensure at least one thread when size < min_chunk. */
481 	nworks = max(job->size / job->min_chunk, 1ul);
482 	nworks = min(nworks, job->max_threads);
483 
484 	if (nworks == 1) {
485 		/* Single thread, no coordination needed, cut to the chase. */
486 		job->thread_fn(job->start, job->start + job->size, job->fn_arg);
487 		return;
488 	}
489 
490 	spin_lock_init(&ps.lock);
491 	init_completion(&ps.completion);
492 	ps.job	       = job;
493 	ps.nworks      = padata_work_alloc_mt(nworks, &ps, &works);
494 	ps.nworks_fini = 0;
495 
496 	/*
497 	 * Chunk size is the amount of work a helper does per call to the
498 	 * thread function.  Load balance large jobs between threads by
499 	 * increasing the number of chunks, guarantee at least the minimum
500 	 * chunk size from the caller, and honor the caller's alignment.
501 	 */
502 	ps.chunk_size = job->size / (ps.nworks * load_balance_factor);
503 	ps.chunk_size = max(ps.chunk_size, job->min_chunk);
504 	ps.chunk_size = roundup(ps.chunk_size, job->align);
505 
506 	list_for_each_entry(pw, &works, pw_list)
507 		queue_work(system_unbound_wq, &pw->pw_work);
508 
509 	/* Use the current thread, which saves starting a workqueue worker. */
510 	padata_work_init(&my_work, padata_mt_helper, &ps, PADATA_WORK_ONSTACK);
511 	padata_mt_helper(&my_work.pw_work);
512 
513 	/* Wait for all the helpers to finish. */
514 	wait_for_completion(&ps.completion);
515 
516 	destroy_work_on_stack(&my_work.pw_work);
517 	padata_works_free(&works);
518 }
519 
520 static void __padata_list_init(struct padata_list *pd_list)
521 {
522 	INIT_LIST_HEAD(&pd_list->list);
523 	spin_lock_init(&pd_list->lock);
524 }
525 
526 /* Initialize all percpu queues used by serial workers */
527 static void padata_init_squeues(struct parallel_data *pd)
528 {
529 	int cpu;
530 	struct padata_serial_queue *squeue;
531 
532 	for_each_cpu(cpu, pd->cpumask.cbcpu) {
533 		squeue = per_cpu_ptr(pd->squeue, cpu);
534 		squeue->pd = pd;
535 		__padata_list_init(&squeue->serial);
536 		INIT_WORK(&squeue->work, padata_serial_worker);
537 	}
538 }
539 
540 /* Initialize per-CPU reorder lists */
541 static void padata_init_reorder_list(struct parallel_data *pd)
542 {
543 	int cpu;
544 	struct padata_list *list;
545 
546 	for_each_cpu(cpu, pd->cpumask.pcpu) {
547 		list = per_cpu_ptr(pd->reorder_list, cpu);
548 		__padata_list_init(list);
549 	}
550 }
551 
552 /* Allocate and initialize the internal cpumask dependend resources. */
553 static struct parallel_data *padata_alloc_pd(struct padata_shell *ps)
554 {
555 	struct padata_instance *pinst = ps->pinst;
556 	struct parallel_data *pd;
557 
558 	pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
559 	if (!pd)
560 		goto err;
561 
562 	pd->reorder_list = alloc_percpu(struct padata_list);
563 	if (!pd->reorder_list)
564 		goto err_free_pd;
565 
566 	pd->squeue = alloc_percpu(struct padata_serial_queue);
567 	if (!pd->squeue)
568 		goto err_free_reorder_list;
569 
570 	pd->ps = ps;
571 
572 	if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
573 		goto err_free_squeue;
574 	if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
575 		goto err_free_pcpu;
576 
577 	cpumask_and(pd->cpumask.pcpu, pinst->cpumask.pcpu, cpu_online_mask);
578 	cpumask_and(pd->cpumask.cbcpu, pinst->cpumask.cbcpu, cpu_online_mask);
579 
580 	padata_init_reorder_list(pd);
581 	padata_init_squeues(pd);
582 	pd->seq_nr = -1;
583 	refcount_set(&pd->refcnt, 1);
584 	spin_lock_init(&pd->lock);
585 	pd->cpu = cpumask_first(pd->cpumask.pcpu);
586 	INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
587 
588 	return pd;
589 
590 err_free_pcpu:
591 	free_cpumask_var(pd->cpumask.pcpu);
592 err_free_squeue:
593 	free_percpu(pd->squeue);
594 err_free_reorder_list:
595 	free_percpu(pd->reorder_list);
596 err_free_pd:
597 	kfree(pd);
598 err:
599 	return NULL;
600 }
601 
602 static void padata_free_pd(struct parallel_data *pd)
603 {
604 	free_cpumask_var(pd->cpumask.pcpu);
605 	free_cpumask_var(pd->cpumask.cbcpu);
606 	free_percpu(pd->reorder_list);
607 	free_percpu(pd->squeue);
608 	kfree(pd);
609 }
610 
611 static void __padata_start(struct padata_instance *pinst)
612 {
613 	pinst->flags |= PADATA_INIT;
614 }
615 
616 static void __padata_stop(struct padata_instance *pinst)
617 {
618 	if (!(pinst->flags & PADATA_INIT))
619 		return;
620 
621 	pinst->flags &= ~PADATA_INIT;
622 
623 	synchronize_rcu();
624 }
625 
626 /* Replace the internal control structure with a new one. */
627 static int padata_replace_one(struct padata_shell *ps)
628 {
629 	struct parallel_data *pd_new;
630 
631 	pd_new = padata_alloc_pd(ps);
632 	if (!pd_new)
633 		return -ENOMEM;
634 
635 	ps->opd = rcu_dereference_protected(ps->pd, 1);
636 	rcu_assign_pointer(ps->pd, pd_new);
637 
638 	return 0;
639 }
640 
641 static int padata_replace(struct padata_instance *pinst)
642 {
643 	struct padata_shell *ps;
644 	int err = 0;
645 
646 	pinst->flags |= PADATA_RESET;
647 
648 	list_for_each_entry(ps, &pinst->pslist, list) {
649 		err = padata_replace_one(ps);
650 		if (err)
651 			break;
652 	}
653 
654 	synchronize_rcu();
655 
656 	list_for_each_entry_continue_reverse(ps, &pinst->pslist, list)
657 		if (refcount_dec_and_test(&ps->opd->refcnt))
658 			padata_free_pd(ps->opd);
659 
660 	pinst->flags &= ~PADATA_RESET;
661 
662 	return err;
663 }
664 
665 /* If cpumask contains no active cpu, we mark the instance as invalid. */
666 static bool padata_validate_cpumask(struct padata_instance *pinst,
667 				    const struct cpumask *cpumask)
668 {
669 	if (!cpumask_intersects(cpumask, cpu_online_mask)) {
670 		pinst->flags |= PADATA_INVALID;
671 		return false;
672 	}
673 
674 	pinst->flags &= ~PADATA_INVALID;
675 	return true;
676 }
677 
678 static int __padata_set_cpumasks(struct padata_instance *pinst,
679 				 cpumask_var_t pcpumask,
680 				 cpumask_var_t cbcpumask)
681 {
682 	int valid;
683 	int err;
684 
685 	valid = padata_validate_cpumask(pinst, pcpumask);
686 	if (!valid) {
687 		__padata_stop(pinst);
688 		goto out_replace;
689 	}
690 
691 	valid = padata_validate_cpumask(pinst, cbcpumask);
692 	if (!valid)
693 		__padata_stop(pinst);
694 
695 out_replace:
696 	cpumask_copy(pinst->cpumask.pcpu, pcpumask);
697 	cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
698 
699 	err = padata_setup_cpumasks(pinst) ?: padata_replace(pinst);
700 
701 	if (valid)
702 		__padata_start(pinst);
703 
704 	return err;
705 }
706 
707 /**
708  * padata_set_cpumask - Sets specified by @cpumask_type cpumask to the value
709  *                      equivalent to @cpumask.
710  * @pinst: padata instance
711  * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
712  *                to parallel and serial cpumasks respectively.
713  * @cpumask: the cpumask to use
714  *
715  * Return: 0 on success or negative error code
716  */
717 int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
718 		       cpumask_var_t cpumask)
719 {
720 	struct cpumask *serial_mask, *parallel_mask;
721 	int err = -EINVAL;
722 
723 	cpus_read_lock();
724 	mutex_lock(&pinst->lock);
725 
726 	switch (cpumask_type) {
727 	case PADATA_CPU_PARALLEL:
728 		serial_mask = pinst->cpumask.cbcpu;
729 		parallel_mask = cpumask;
730 		break;
731 	case PADATA_CPU_SERIAL:
732 		parallel_mask = pinst->cpumask.pcpu;
733 		serial_mask = cpumask;
734 		break;
735 	default:
736 		 goto out;
737 	}
738 
739 	err =  __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
740 
741 out:
742 	mutex_unlock(&pinst->lock);
743 	cpus_read_unlock();
744 
745 	return err;
746 }
747 EXPORT_SYMBOL(padata_set_cpumask);
748 
749 #ifdef CONFIG_HOTPLUG_CPU
750 
751 static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
752 {
753 	int err = 0;
754 
755 	if (cpumask_test_cpu(cpu, cpu_online_mask)) {
756 		err = padata_replace(pinst);
757 
758 		if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
759 		    padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
760 			__padata_start(pinst);
761 	}
762 
763 	return err;
764 }
765 
766 static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
767 {
768 	int err = 0;
769 
770 	if (!cpumask_test_cpu(cpu, cpu_online_mask)) {
771 		if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
772 		    !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
773 			__padata_stop(pinst);
774 
775 		err = padata_replace(pinst);
776 	}
777 
778 	return err;
779 }
780 
781 static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
782 {
783 	return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
784 		cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
785 }
786 
787 static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
788 {
789 	struct padata_instance *pinst;
790 	int ret;
791 
792 	pinst = hlist_entry_safe(node, struct padata_instance, cpu_online_node);
793 	if (!pinst_has_cpu(pinst, cpu))
794 		return 0;
795 
796 	mutex_lock(&pinst->lock);
797 	ret = __padata_add_cpu(pinst, cpu);
798 	mutex_unlock(&pinst->lock);
799 	return ret;
800 }
801 
802 static int padata_cpu_dead(unsigned int cpu, struct hlist_node *node)
803 {
804 	struct padata_instance *pinst;
805 	int ret;
806 
807 	pinst = hlist_entry_safe(node, struct padata_instance, cpu_dead_node);
808 	if (!pinst_has_cpu(pinst, cpu))
809 		return 0;
810 
811 	mutex_lock(&pinst->lock);
812 	ret = __padata_remove_cpu(pinst, cpu);
813 	mutex_unlock(&pinst->lock);
814 	return ret;
815 }
816 
817 static enum cpuhp_state hp_online;
818 #endif
819 
820 static void __padata_free(struct padata_instance *pinst)
821 {
822 #ifdef CONFIG_HOTPLUG_CPU
823 	cpuhp_state_remove_instance_nocalls(CPUHP_PADATA_DEAD,
824 					    &pinst->cpu_dead_node);
825 	cpuhp_state_remove_instance_nocalls(hp_online, &pinst->cpu_online_node);
826 #endif
827 
828 	WARN_ON(!list_empty(&pinst->pslist));
829 
830 	free_cpumask_var(pinst->cpumask.pcpu);
831 	free_cpumask_var(pinst->cpumask.cbcpu);
832 	destroy_workqueue(pinst->serial_wq);
833 	destroy_workqueue(pinst->parallel_wq);
834 	kfree(pinst);
835 }
836 
837 #define kobj2pinst(_kobj)					\
838 	container_of(_kobj, struct padata_instance, kobj)
839 #define attr2pentry(_attr)					\
840 	container_of(_attr, struct padata_sysfs_entry, attr)
841 
842 static void padata_sysfs_release(struct kobject *kobj)
843 {
844 	struct padata_instance *pinst = kobj2pinst(kobj);
845 	__padata_free(pinst);
846 }
847 
848 struct padata_sysfs_entry {
849 	struct attribute attr;
850 	ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
851 	ssize_t (*store)(struct padata_instance *, struct attribute *,
852 			 const char *, size_t);
853 };
854 
855 static ssize_t show_cpumask(struct padata_instance *pinst,
856 			    struct attribute *attr,  char *buf)
857 {
858 	struct cpumask *cpumask;
859 	ssize_t len;
860 
861 	mutex_lock(&pinst->lock);
862 	if (!strcmp(attr->name, "serial_cpumask"))
863 		cpumask = pinst->cpumask.cbcpu;
864 	else
865 		cpumask = pinst->cpumask.pcpu;
866 
867 	len = snprintf(buf, PAGE_SIZE, "%*pb\n",
868 		       nr_cpu_ids, cpumask_bits(cpumask));
869 	mutex_unlock(&pinst->lock);
870 	return len < PAGE_SIZE ? len : -EINVAL;
871 }
872 
873 static ssize_t store_cpumask(struct padata_instance *pinst,
874 			     struct attribute *attr,
875 			     const char *buf, size_t count)
876 {
877 	cpumask_var_t new_cpumask;
878 	ssize_t ret;
879 	int mask_type;
880 
881 	if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
882 		return -ENOMEM;
883 
884 	ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
885 			   nr_cpumask_bits);
886 	if (ret < 0)
887 		goto out;
888 
889 	mask_type = !strcmp(attr->name, "serial_cpumask") ?
890 		PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
891 	ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
892 	if (!ret)
893 		ret = count;
894 
895 out:
896 	free_cpumask_var(new_cpumask);
897 	return ret;
898 }
899 
900 #define PADATA_ATTR_RW(_name, _show_name, _store_name)		\
901 	static struct padata_sysfs_entry _name##_attr =		\
902 		__ATTR(_name, 0644, _show_name, _store_name)
903 #define PADATA_ATTR_RO(_name, _show_name)		\
904 	static struct padata_sysfs_entry _name##_attr = \
905 		__ATTR(_name, 0400, _show_name, NULL)
906 
907 PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
908 PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
909 
910 /*
911  * Padata sysfs provides the following objects:
912  * serial_cpumask   [RW] - cpumask for serial workers
913  * parallel_cpumask [RW] - cpumask for parallel workers
914  */
915 static struct attribute *padata_default_attrs[] = {
916 	&serial_cpumask_attr.attr,
917 	&parallel_cpumask_attr.attr,
918 	NULL,
919 };
920 ATTRIBUTE_GROUPS(padata_default);
921 
922 static ssize_t padata_sysfs_show(struct kobject *kobj,
923 				 struct attribute *attr, char *buf)
924 {
925 	struct padata_instance *pinst;
926 	struct padata_sysfs_entry *pentry;
927 	ssize_t ret = -EIO;
928 
929 	pinst = kobj2pinst(kobj);
930 	pentry = attr2pentry(attr);
931 	if (pentry->show)
932 		ret = pentry->show(pinst, attr, buf);
933 
934 	return ret;
935 }
936 
937 static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
938 				  const char *buf, size_t count)
939 {
940 	struct padata_instance *pinst;
941 	struct padata_sysfs_entry *pentry;
942 	ssize_t ret = -EIO;
943 
944 	pinst = kobj2pinst(kobj);
945 	pentry = attr2pentry(attr);
946 	if (pentry->show)
947 		ret = pentry->store(pinst, attr, buf, count);
948 
949 	return ret;
950 }
951 
952 static const struct sysfs_ops padata_sysfs_ops = {
953 	.show = padata_sysfs_show,
954 	.store = padata_sysfs_store,
955 };
956 
957 static struct kobj_type padata_attr_type = {
958 	.sysfs_ops = &padata_sysfs_ops,
959 	.default_groups = padata_default_groups,
960 	.release = padata_sysfs_release,
961 };
962 
963 /**
964  * padata_alloc - allocate and initialize a padata instance
965  * @name: used to identify the instance
966  *
967  * Return: new instance on success, NULL on error
968  */
969 struct padata_instance *padata_alloc(const char *name)
970 {
971 	struct padata_instance *pinst;
972 
973 	pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
974 	if (!pinst)
975 		goto err;
976 
977 	pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
978 					     name);
979 	if (!pinst->parallel_wq)
980 		goto err_free_inst;
981 
982 	cpus_read_lock();
983 
984 	pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
985 					   WQ_CPU_INTENSIVE, 1, name);
986 	if (!pinst->serial_wq)
987 		goto err_put_cpus;
988 
989 	if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
990 		goto err_free_serial_wq;
991 	if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
992 		free_cpumask_var(pinst->cpumask.pcpu);
993 		goto err_free_serial_wq;
994 	}
995 
996 	INIT_LIST_HEAD(&pinst->pslist);
997 
998 	cpumask_copy(pinst->cpumask.pcpu, cpu_possible_mask);
999 	cpumask_copy(pinst->cpumask.cbcpu, cpu_possible_mask);
1000 
1001 	if (padata_setup_cpumasks(pinst))
1002 		goto err_free_masks;
1003 
1004 	__padata_start(pinst);
1005 
1006 	kobject_init(&pinst->kobj, &padata_attr_type);
1007 	mutex_init(&pinst->lock);
1008 
1009 #ifdef CONFIG_HOTPLUG_CPU
1010 	cpuhp_state_add_instance_nocalls_cpuslocked(hp_online,
1011 						    &pinst->cpu_online_node);
1012 	cpuhp_state_add_instance_nocalls_cpuslocked(CPUHP_PADATA_DEAD,
1013 						    &pinst->cpu_dead_node);
1014 #endif
1015 
1016 	cpus_read_unlock();
1017 
1018 	return pinst;
1019 
1020 err_free_masks:
1021 	free_cpumask_var(pinst->cpumask.pcpu);
1022 	free_cpumask_var(pinst->cpumask.cbcpu);
1023 err_free_serial_wq:
1024 	destroy_workqueue(pinst->serial_wq);
1025 err_put_cpus:
1026 	cpus_read_unlock();
1027 	destroy_workqueue(pinst->parallel_wq);
1028 err_free_inst:
1029 	kfree(pinst);
1030 err:
1031 	return NULL;
1032 }
1033 EXPORT_SYMBOL(padata_alloc);
1034 
1035 /**
1036  * padata_free - free a padata instance
1037  *
1038  * @pinst: padata instance to free
1039  */
1040 void padata_free(struct padata_instance *pinst)
1041 {
1042 	kobject_put(&pinst->kobj);
1043 }
1044 EXPORT_SYMBOL(padata_free);
1045 
1046 /**
1047  * padata_alloc_shell - Allocate and initialize padata shell.
1048  *
1049  * @pinst: Parent padata_instance object.
1050  *
1051  * Return: new shell on success, NULL on error
1052  */
1053 struct padata_shell *padata_alloc_shell(struct padata_instance *pinst)
1054 {
1055 	struct parallel_data *pd;
1056 	struct padata_shell *ps;
1057 
1058 	ps = kzalloc(sizeof(*ps), GFP_KERNEL);
1059 	if (!ps)
1060 		goto out;
1061 
1062 	ps->pinst = pinst;
1063 
1064 	cpus_read_lock();
1065 	pd = padata_alloc_pd(ps);
1066 	cpus_read_unlock();
1067 
1068 	if (!pd)
1069 		goto out_free_ps;
1070 
1071 	mutex_lock(&pinst->lock);
1072 	RCU_INIT_POINTER(ps->pd, pd);
1073 	list_add(&ps->list, &pinst->pslist);
1074 	mutex_unlock(&pinst->lock);
1075 
1076 	return ps;
1077 
1078 out_free_ps:
1079 	kfree(ps);
1080 out:
1081 	return NULL;
1082 }
1083 EXPORT_SYMBOL(padata_alloc_shell);
1084 
1085 /**
1086  * padata_free_shell - free a padata shell
1087  *
1088  * @ps: padata shell to free
1089  */
1090 void padata_free_shell(struct padata_shell *ps)
1091 {
1092 	if (!ps)
1093 		return;
1094 
1095 	mutex_lock(&ps->pinst->lock);
1096 	list_del(&ps->list);
1097 	padata_free_pd(rcu_dereference_protected(ps->pd, 1));
1098 	mutex_unlock(&ps->pinst->lock);
1099 
1100 	kfree(ps);
1101 }
1102 EXPORT_SYMBOL(padata_free_shell);
1103 
1104 void __init padata_init(void)
1105 {
1106 	unsigned int i, possible_cpus;
1107 #ifdef CONFIG_HOTPLUG_CPU
1108 	int ret;
1109 
1110 	ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1111 				      padata_cpu_online, NULL);
1112 	if (ret < 0)
1113 		goto err;
1114 	hp_online = ret;
1115 
1116 	ret = cpuhp_setup_state_multi(CPUHP_PADATA_DEAD, "padata:dead",
1117 				      NULL, padata_cpu_dead);
1118 	if (ret < 0)
1119 		goto remove_online_state;
1120 #endif
1121 
1122 	possible_cpus = num_possible_cpus();
1123 	padata_works = kmalloc_array(possible_cpus, sizeof(struct padata_work),
1124 				     GFP_KERNEL);
1125 	if (!padata_works)
1126 		goto remove_dead_state;
1127 
1128 	for (i = 0; i < possible_cpus; ++i)
1129 		list_add(&padata_works[i].pw_list, &padata_free_works);
1130 
1131 	return;
1132 
1133 remove_dead_state:
1134 #ifdef CONFIG_HOTPLUG_CPU
1135 	cpuhp_remove_multi_state(CPUHP_PADATA_DEAD);
1136 remove_online_state:
1137 	cpuhp_remove_multi_state(hp_online);
1138 err:
1139 #endif
1140 	pr_warn("padata: initialization failed\n");
1141 }
1142