1 // SPDX-License-Identifier: GPL-2.0
3 * padata.c - generic interface to process data streams in parallel
5 * See Documentation/core-api/padata.rst for more information.
7 * Copyright (C) 2008, 2009 secunet Security Networks AG
8 * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
10 * Copyright (c) 2020 Oracle and/or its affiliates.
11 * Author: Daniel Jordan <daniel.m.jordan@oracle.com>
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>
26 #define PADATA_WORK_ONSTACK 1 /* Work's memory is on stack */
29 struct work_struct pw_work;
30 struct list_head pw_list; /* padata_free_works linkage */
34 static DEFINE_SPINLOCK(padata_works_lock);
35 static struct padata_work *padata_works;
36 static LIST_HEAD(padata_free_works);
38 struct padata_mt_job_state {
40 struct completion completion;
41 struct padata_mt_job *job;
44 unsigned long chunk_size;
47 static void padata_free_pd(struct parallel_data *pd);
48 static void __init padata_mt_helper(struct work_struct *work);
50 static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
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);
61 static int padata_cpu_hash(struct parallel_data *pd, unsigned int seq_nr)
64 * Hash the sequence numbers to the cpus by taking
65 * seq_nr mod. number of cpus in use.
67 int cpu_index = seq_nr % cpumask_weight(pd->cpumask.pcpu);
69 return padata_index_to_cpu(pd, cpu_index);
72 static struct padata_work *padata_work_alloc(void)
74 struct padata_work *pw;
76 lockdep_assert_held(&padata_works_lock);
78 if (list_empty(&padata_free_works))
79 return NULL; /* No more work items allowed to be queued. */
81 pw = list_first_entry(&padata_free_works, struct padata_work, pw_list);
82 list_del(&pw->pw_list);
86 static void padata_work_init(struct padata_work *pw, work_func_t work_fn,
87 void *data, int flags)
89 if (flags & PADATA_WORK_ONSTACK)
90 INIT_WORK_ONSTACK(&pw->pw_work, work_fn);
92 INIT_WORK(&pw->pw_work, work_fn);
96 static int __init padata_work_alloc_mt(int nworks, void *data,
97 struct list_head *head)
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();
108 padata_work_init(pw, padata_mt_helper, data, 0);
109 list_add(&pw->pw_list, head);
111 spin_unlock(&padata_works_lock);
116 static void padata_work_free(struct padata_work *pw)
118 lockdep_assert_held(&padata_works_lock);
119 list_add(&pw->pw_list, &padata_free_works);
122 static void __init padata_works_free(struct list_head *works)
124 struct padata_work *cur, *next;
126 if (list_empty(works))
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);
134 spin_unlock(&padata_works_lock);
137 static void padata_parallel_worker(struct work_struct *parallel_work)
139 struct padata_work *pw = container_of(parallel_work, struct padata_work,
141 struct padata_priv *padata = pw->pw_data;
144 padata->parallel(padata);
145 spin_lock(&padata_works_lock);
146 padata_work_free(pw);
147 spin_unlock(&padata_works_lock);
152 * padata_do_parallel - padata parallelization function
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.
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.
165 * Return: 0 on success or else negative error code.
167 int padata_do_parallel(struct padata_shell *ps,
168 struct padata_priv *padata, int *cb_cpu)
170 struct padata_instance *pinst = ps->pinst;
171 int i, cpu, cpu_index, err;
172 struct parallel_data *pd;
173 struct padata_work *pw;
177 pd = rcu_dereference_bh(ps->pd);
180 if (!(pinst->flags & PADATA_INIT) || pinst->flags & PADATA_INVALID)
183 if (!cpumask_test_cpu(*cb_cpu, pd->cpumask.cbcpu)) {
184 if (cpumask_empty(pd->cpumask.cbcpu))
187 /* Select an alternate fallback CPU and notify the caller. */
188 cpu_index = *cb_cpu % cpumask_weight(pd->cpumask.cbcpu);
190 cpu = cpumask_first(pd->cpumask.cbcpu);
191 for (i = 0; i < cpu_index; i++)
192 cpu = cpumask_next(cpu, pd->cpumask.cbcpu);
198 if ((pinst->flags & PADATA_RESET))
201 refcount_inc(&pd->refcnt);
203 padata->cb_cpu = *cb_cpu;
205 spin_lock(&padata_works_lock);
206 padata->seq_nr = ++pd->seq_nr;
207 pw = padata_work_alloc();
208 spin_unlock(&padata_works_lock);
211 /* Maximum works limit exceeded, run in the current task. */
212 padata->parallel(padata);
215 rcu_read_unlock_bh();
218 padata_work_init(pw, padata_parallel_worker, padata, 0);
219 queue_work(pinst->parallel_wq, &pw->pw_work);
224 rcu_read_unlock_bh();
228 EXPORT_SYMBOL(padata_do_parallel);
231 * padata_find_next - Find the next object that needs serialization.
234 * * A pointer to the control struct of the next object that needs
235 * serialization, if present in one of the percpu reorder queues.
236 * * NULL, if the next object that needs serialization will
237 * be parallel processed by another cpu and is not yet present in
238 * the cpu's reorder queue.
240 static struct padata_priv *padata_find_next(struct parallel_data *pd,
243 struct padata_priv *padata;
244 struct padata_list *reorder;
247 reorder = per_cpu_ptr(pd->reorder_list, cpu);
249 spin_lock(&reorder->lock);
250 if (list_empty(&reorder->list)) {
251 spin_unlock(&reorder->lock);
255 padata = list_entry(reorder->list.next, struct padata_priv, list);
258 * Checks the rare case where two or more parallel jobs have hashed to
259 * the same CPU and one of the later ones finishes first.
261 if (padata->seq_nr != pd->processed) {
262 spin_unlock(&reorder->lock);
267 list_del_init(&padata->list);
269 pd->cpu = cpumask_next_wrap(cpu, pd->cpumask.pcpu, -1, false);
272 spin_unlock(&reorder->lock);
276 static void padata_reorder(struct parallel_data *pd)
278 struct padata_instance *pinst = pd->ps->pinst;
280 struct padata_priv *padata;
281 struct padata_serial_queue *squeue;
282 struct padata_list *reorder;
285 * We need to ensure that only one cpu can work on dequeueing of
286 * the reorder queue the time. Calculating in which percpu reorder
287 * queue the next object will arrive takes some time. A spinlock
288 * would be highly contended. Also it is not clear in which order
289 * the objects arrive to the reorder queues. So a cpu could wait to
290 * get the lock just to notice that there is nothing to do at the
291 * moment. Therefore we use a trylock and let the holder of the lock
292 * care for all the objects enqueued during the holdtime of the lock.
294 if (!spin_trylock_bh(&pd->lock))
298 padata = padata_find_next(pd, true);
301 * If the next object that needs serialization is parallel
302 * processed by another cpu and is still on it's way to the
303 * cpu's reorder queue, nothing to do for now.
308 cb_cpu = padata->cb_cpu;
309 squeue = per_cpu_ptr(pd->squeue, cb_cpu);
311 spin_lock(&squeue->serial.lock);
312 list_add_tail(&padata->list, &squeue->serial.list);
313 spin_unlock(&squeue->serial.lock);
315 queue_work_on(cb_cpu, pinst->serial_wq, &squeue->work);
318 spin_unlock_bh(&pd->lock);
321 * The next object that needs serialization might have arrived to
322 * the reorder queues in the meantime.
324 * Ensure reorder queue is read after pd->lock is dropped so we see
325 * new objects from another task in padata_do_serial. Pairs with
326 * smp_mb in padata_do_serial.
330 reorder = per_cpu_ptr(pd->reorder_list, pd->cpu);
331 if (!list_empty(&reorder->list) && padata_find_next(pd, false))
332 queue_work(pinst->serial_wq, &pd->reorder_work);
335 static void invoke_padata_reorder(struct work_struct *work)
337 struct parallel_data *pd;
340 pd = container_of(work, struct parallel_data, reorder_work);
345 static void padata_serial_worker(struct work_struct *serial_work)
347 struct padata_serial_queue *squeue;
348 struct parallel_data *pd;
349 LIST_HEAD(local_list);
353 squeue = container_of(serial_work, struct padata_serial_queue, work);
356 spin_lock(&squeue->serial.lock);
357 list_replace_init(&squeue->serial.list, &local_list);
358 spin_unlock(&squeue->serial.lock);
362 while (!list_empty(&local_list)) {
363 struct padata_priv *padata;
365 padata = list_entry(local_list.next,
366 struct padata_priv, list);
368 list_del_init(&padata->list);
370 padata->serial(padata);
375 if (refcount_sub_and_test(cnt, &pd->refcnt))
380 * padata_do_serial - padata serialization function
382 * @padata: object to be serialized.
384 * padata_do_serial must be called for every parallelized object.
385 * The serialization callback function will run with BHs off.
387 void padata_do_serial(struct padata_priv *padata)
389 struct parallel_data *pd = padata->pd;
390 int hashed_cpu = padata_cpu_hash(pd, padata->seq_nr);
391 struct padata_list *reorder = per_cpu_ptr(pd->reorder_list, hashed_cpu);
392 struct padata_priv *cur;
394 spin_lock(&reorder->lock);
395 /* Sort in ascending order of sequence number. */
396 list_for_each_entry_reverse(cur, &reorder->list, list)
397 if (cur->seq_nr < padata->seq_nr)
399 list_add(&padata->list, &cur->list);
400 spin_unlock(&reorder->lock);
403 * Ensure the addition to the reorder list is ordered correctly
404 * with the trylock of pd->lock in padata_reorder. Pairs with smp_mb
411 EXPORT_SYMBOL(padata_do_serial);
413 static int padata_setup_cpumasks(struct padata_instance *pinst)
415 struct workqueue_attrs *attrs;
418 attrs = alloc_workqueue_attrs();
422 /* Restrict parallel_wq workers to pd->cpumask.pcpu. */
423 cpumask_copy(attrs->cpumask, pinst->cpumask.pcpu);
424 err = apply_workqueue_attrs(pinst->parallel_wq, attrs);
425 free_workqueue_attrs(attrs);
430 static void __init padata_mt_helper(struct work_struct *w)
432 struct padata_work *pw = container_of(w, struct padata_work, pw_work);
433 struct padata_mt_job_state *ps = pw->pw_data;
434 struct padata_mt_job *job = ps->job;
437 spin_lock(&ps->lock);
439 while (job->size > 0) {
440 unsigned long start, size, end;
443 /* So end is chunk size aligned if enough work remains. */
444 size = roundup(start + 1, ps->chunk_size) - start;
445 size = min(size, job->size);
451 spin_unlock(&ps->lock);
452 job->thread_fn(start, end, job->fn_arg);
453 spin_lock(&ps->lock);
457 done = (ps->nworks_fini == ps->nworks);
458 spin_unlock(&ps->lock);
461 complete(&ps->completion);
465 * padata_do_multithreaded - run a multithreaded job
466 * @job: Description of the job.
468 * See the definition of struct padata_mt_job for more details.
470 void __init padata_do_multithreaded(struct padata_mt_job *job)
472 /* In case threads finish at different times. */
473 static const unsigned long load_balance_factor = 4;
474 struct padata_work my_work, *pw;
475 struct padata_mt_job_state ps;
482 /* Ensure at least one thread when size < min_chunk. */
483 nworks = max(job->size / job->min_chunk, 1ul);
484 nworks = min(nworks, job->max_threads);
487 /* Single thread, no coordination needed, cut to the chase. */
488 job->thread_fn(job->start, job->start + job->size, job->fn_arg);
492 spin_lock_init(&ps.lock);
493 init_completion(&ps.completion);
495 ps.nworks = padata_work_alloc_mt(nworks, &ps, &works);
499 * Chunk size is the amount of work a helper does per call to the
500 * thread function. Load balance large jobs between threads by
501 * increasing the number of chunks, guarantee at least the minimum
502 * chunk size from the caller, and honor the caller's alignment.
504 ps.chunk_size = job->size / (ps.nworks * load_balance_factor);
505 ps.chunk_size = max(ps.chunk_size, job->min_chunk);
506 ps.chunk_size = roundup(ps.chunk_size, job->align);
508 list_for_each_entry(pw, &works, pw_list)
509 queue_work(system_unbound_wq, &pw->pw_work);
511 /* Use the current thread, which saves starting a workqueue worker. */
512 padata_work_init(&my_work, padata_mt_helper, &ps, PADATA_WORK_ONSTACK);
513 padata_mt_helper(&my_work.pw_work);
515 /* Wait for all the helpers to finish. */
516 wait_for_completion(&ps.completion);
518 destroy_work_on_stack(&my_work.pw_work);
519 padata_works_free(&works);
522 static void __padata_list_init(struct padata_list *pd_list)
524 INIT_LIST_HEAD(&pd_list->list);
525 spin_lock_init(&pd_list->lock);
528 /* Initialize all percpu queues used by serial workers */
529 static void padata_init_squeues(struct parallel_data *pd)
532 struct padata_serial_queue *squeue;
534 for_each_cpu(cpu, pd->cpumask.cbcpu) {
535 squeue = per_cpu_ptr(pd->squeue, cpu);
537 __padata_list_init(&squeue->serial);
538 INIT_WORK(&squeue->work, padata_serial_worker);
542 /* Initialize per-CPU reorder lists */
543 static void padata_init_reorder_list(struct parallel_data *pd)
546 struct padata_list *list;
548 for_each_cpu(cpu, pd->cpumask.pcpu) {
549 list = per_cpu_ptr(pd->reorder_list, cpu);
550 __padata_list_init(list);
554 /* Allocate and initialize the internal cpumask dependend resources. */
555 static struct parallel_data *padata_alloc_pd(struct padata_shell *ps)
557 struct padata_instance *pinst = ps->pinst;
558 struct parallel_data *pd;
560 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
564 pd->reorder_list = alloc_percpu(struct padata_list);
565 if (!pd->reorder_list)
568 pd->squeue = alloc_percpu(struct padata_serial_queue);
570 goto err_free_reorder_list;
574 if (!alloc_cpumask_var(&pd->cpumask.pcpu, GFP_KERNEL))
575 goto err_free_squeue;
576 if (!alloc_cpumask_var(&pd->cpumask.cbcpu, GFP_KERNEL))
579 cpumask_and(pd->cpumask.pcpu, pinst->cpumask.pcpu, cpu_online_mask);
580 cpumask_and(pd->cpumask.cbcpu, pinst->cpumask.cbcpu, cpu_online_mask);
582 padata_init_reorder_list(pd);
583 padata_init_squeues(pd);
585 refcount_set(&pd->refcnt, 1);
586 spin_lock_init(&pd->lock);
587 pd->cpu = cpumask_first(pd->cpumask.pcpu);
588 INIT_WORK(&pd->reorder_work, invoke_padata_reorder);
593 free_cpumask_var(pd->cpumask.pcpu);
595 free_percpu(pd->squeue);
596 err_free_reorder_list:
597 free_percpu(pd->reorder_list);
604 static void padata_free_pd(struct parallel_data *pd)
606 free_cpumask_var(pd->cpumask.pcpu);
607 free_cpumask_var(pd->cpumask.cbcpu);
608 free_percpu(pd->reorder_list);
609 free_percpu(pd->squeue);
613 static void __padata_start(struct padata_instance *pinst)
615 pinst->flags |= PADATA_INIT;
618 static void __padata_stop(struct padata_instance *pinst)
620 if (!(pinst->flags & PADATA_INIT))
623 pinst->flags &= ~PADATA_INIT;
628 /* Replace the internal control structure with a new one. */
629 static int padata_replace_one(struct padata_shell *ps)
631 struct parallel_data *pd_new;
633 pd_new = padata_alloc_pd(ps);
637 ps->opd = rcu_dereference_protected(ps->pd, 1);
638 rcu_assign_pointer(ps->pd, pd_new);
643 static int padata_replace(struct padata_instance *pinst)
645 struct padata_shell *ps;
648 pinst->flags |= PADATA_RESET;
650 list_for_each_entry(ps, &pinst->pslist, list) {
651 err = padata_replace_one(ps);
658 list_for_each_entry_continue_reverse(ps, &pinst->pslist, list)
659 if (refcount_dec_and_test(&ps->opd->refcnt))
660 padata_free_pd(ps->opd);
662 pinst->flags &= ~PADATA_RESET;
667 /* If cpumask contains no active cpu, we mark the instance as invalid. */
668 static bool padata_validate_cpumask(struct padata_instance *pinst,
669 const struct cpumask *cpumask)
671 if (!cpumask_intersects(cpumask, cpu_online_mask)) {
672 pinst->flags |= PADATA_INVALID;
676 pinst->flags &= ~PADATA_INVALID;
680 static int __padata_set_cpumasks(struct padata_instance *pinst,
681 cpumask_var_t pcpumask,
682 cpumask_var_t cbcpumask)
687 valid = padata_validate_cpumask(pinst, pcpumask);
689 __padata_stop(pinst);
693 valid = padata_validate_cpumask(pinst, cbcpumask);
695 __padata_stop(pinst);
698 cpumask_copy(pinst->cpumask.pcpu, pcpumask);
699 cpumask_copy(pinst->cpumask.cbcpu, cbcpumask);
701 err = padata_setup_cpumasks(pinst) ?: padata_replace(pinst);
704 __padata_start(pinst);
710 * padata_set_cpumask - Sets specified by @cpumask_type cpumask to the value
711 * equivalent to @cpumask.
712 * @pinst: padata instance
713 * @cpumask_type: PADATA_CPU_SERIAL or PADATA_CPU_PARALLEL corresponding
714 * to parallel and serial cpumasks respectively.
715 * @cpumask: the cpumask to use
717 * Return: 0 on success or negative error code
719 int padata_set_cpumask(struct padata_instance *pinst, int cpumask_type,
720 cpumask_var_t cpumask)
722 struct cpumask *serial_mask, *parallel_mask;
726 mutex_lock(&pinst->lock);
728 switch (cpumask_type) {
729 case PADATA_CPU_PARALLEL:
730 serial_mask = pinst->cpumask.cbcpu;
731 parallel_mask = cpumask;
733 case PADATA_CPU_SERIAL:
734 parallel_mask = pinst->cpumask.pcpu;
735 serial_mask = cpumask;
741 err = __padata_set_cpumasks(pinst, parallel_mask, serial_mask);
744 mutex_unlock(&pinst->lock);
749 EXPORT_SYMBOL(padata_set_cpumask);
751 #ifdef CONFIG_HOTPLUG_CPU
753 static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
757 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
758 err = padata_replace(pinst);
760 if (padata_validate_cpumask(pinst, pinst->cpumask.pcpu) &&
761 padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
762 __padata_start(pinst);
768 static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
772 if (!cpumask_test_cpu(cpu, cpu_online_mask)) {
773 if (!padata_validate_cpumask(pinst, pinst->cpumask.pcpu) ||
774 !padata_validate_cpumask(pinst, pinst->cpumask.cbcpu))
775 __padata_stop(pinst);
777 err = padata_replace(pinst);
783 static inline int pinst_has_cpu(struct padata_instance *pinst, int cpu)
785 return cpumask_test_cpu(cpu, pinst->cpumask.pcpu) ||
786 cpumask_test_cpu(cpu, pinst->cpumask.cbcpu);
789 static int padata_cpu_online(unsigned int cpu, struct hlist_node *node)
791 struct padata_instance *pinst;
794 pinst = hlist_entry_safe(node, struct padata_instance, cpu_online_node);
795 if (!pinst_has_cpu(pinst, cpu))
798 mutex_lock(&pinst->lock);
799 ret = __padata_add_cpu(pinst, cpu);
800 mutex_unlock(&pinst->lock);
804 static int padata_cpu_dead(unsigned int cpu, struct hlist_node *node)
806 struct padata_instance *pinst;
809 pinst = hlist_entry_safe(node, struct padata_instance, cpu_dead_node);
810 if (!pinst_has_cpu(pinst, cpu))
813 mutex_lock(&pinst->lock);
814 ret = __padata_remove_cpu(pinst, cpu);
815 mutex_unlock(&pinst->lock);
819 static enum cpuhp_state hp_online;
822 static void __padata_free(struct padata_instance *pinst)
824 #ifdef CONFIG_HOTPLUG_CPU
825 cpuhp_state_remove_instance_nocalls(CPUHP_PADATA_DEAD,
826 &pinst->cpu_dead_node);
827 cpuhp_state_remove_instance_nocalls(hp_online, &pinst->cpu_online_node);
830 WARN_ON(!list_empty(&pinst->pslist));
832 free_cpumask_var(pinst->cpumask.pcpu);
833 free_cpumask_var(pinst->cpumask.cbcpu);
834 destroy_workqueue(pinst->serial_wq);
835 destroy_workqueue(pinst->parallel_wq);
839 #define kobj2pinst(_kobj) \
840 container_of(_kobj, struct padata_instance, kobj)
841 #define attr2pentry(_attr) \
842 container_of(_attr, struct padata_sysfs_entry, attr)
844 static void padata_sysfs_release(struct kobject *kobj)
846 struct padata_instance *pinst = kobj2pinst(kobj);
847 __padata_free(pinst);
850 struct padata_sysfs_entry {
851 struct attribute attr;
852 ssize_t (*show)(struct padata_instance *, struct attribute *, char *);
853 ssize_t (*store)(struct padata_instance *, struct attribute *,
854 const char *, size_t);
857 static ssize_t show_cpumask(struct padata_instance *pinst,
858 struct attribute *attr, char *buf)
860 struct cpumask *cpumask;
863 mutex_lock(&pinst->lock);
864 if (!strcmp(attr->name, "serial_cpumask"))
865 cpumask = pinst->cpumask.cbcpu;
867 cpumask = pinst->cpumask.pcpu;
869 len = snprintf(buf, PAGE_SIZE, "%*pb\n",
870 nr_cpu_ids, cpumask_bits(cpumask));
871 mutex_unlock(&pinst->lock);
872 return len < PAGE_SIZE ? len : -EINVAL;
875 static ssize_t store_cpumask(struct padata_instance *pinst,
876 struct attribute *attr,
877 const char *buf, size_t count)
879 cpumask_var_t new_cpumask;
883 if (!alloc_cpumask_var(&new_cpumask, GFP_KERNEL))
886 ret = bitmap_parse(buf, count, cpumask_bits(new_cpumask),
891 mask_type = !strcmp(attr->name, "serial_cpumask") ?
892 PADATA_CPU_SERIAL : PADATA_CPU_PARALLEL;
893 ret = padata_set_cpumask(pinst, mask_type, new_cpumask);
898 free_cpumask_var(new_cpumask);
902 #define PADATA_ATTR_RW(_name, _show_name, _store_name) \
903 static struct padata_sysfs_entry _name##_attr = \
904 __ATTR(_name, 0644, _show_name, _store_name)
905 #define PADATA_ATTR_RO(_name, _show_name) \
906 static struct padata_sysfs_entry _name##_attr = \
907 __ATTR(_name, 0400, _show_name, NULL)
909 PADATA_ATTR_RW(serial_cpumask, show_cpumask, store_cpumask);
910 PADATA_ATTR_RW(parallel_cpumask, show_cpumask, store_cpumask);
913 * Padata sysfs provides the following objects:
914 * serial_cpumask [RW] - cpumask for serial workers
915 * parallel_cpumask [RW] - cpumask for parallel workers
917 static struct attribute *padata_default_attrs[] = {
918 &serial_cpumask_attr.attr,
919 ¶llel_cpumask_attr.attr,
922 ATTRIBUTE_GROUPS(padata_default);
924 static ssize_t padata_sysfs_show(struct kobject *kobj,
925 struct attribute *attr, char *buf)
927 struct padata_instance *pinst;
928 struct padata_sysfs_entry *pentry;
931 pinst = kobj2pinst(kobj);
932 pentry = attr2pentry(attr);
934 ret = pentry->show(pinst, attr, buf);
939 static ssize_t padata_sysfs_store(struct kobject *kobj, struct attribute *attr,
940 const char *buf, size_t count)
942 struct padata_instance *pinst;
943 struct padata_sysfs_entry *pentry;
946 pinst = kobj2pinst(kobj);
947 pentry = attr2pentry(attr);
949 ret = pentry->store(pinst, attr, buf, count);
954 static const struct sysfs_ops padata_sysfs_ops = {
955 .show = padata_sysfs_show,
956 .store = padata_sysfs_store,
959 static struct kobj_type padata_attr_type = {
960 .sysfs_ops = &padata_sysfs_ops,
961 .default_groups = padata_default_groups,
962 .release = padata_sysfs_release,
966 * padata_alloc - allocate and initialize a padata instance
967 * @name: used to identify the instance
969 * Return: new instance on success, NULL on error
971 struct padata_instance *padata_alloc(const char *name)
973 struct padata_instance *pinst;
975 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
979 pinst->parallel_wq = alloc_workqueue("%s_parallel", WQ_UNBOUND, 0,
981 if (!pinst->parallel_wq)
986 pinst->serial_wq = alloc_workqueue("%s_serial", WQ_MEM_RECLAIM |
987 WQ_CPU_INTENSIVE, 1, name);
988 if (!pinst->serial_wq)
991 if (!alloc_cpumask_var(&pinst->cpumask.pcpu, GFP_KERNEL))
992 goto err_free_serial_wq;
993 if (!alloc_cpumask_var(&pinst->cpumask.cbcpu, GFP_KERNEL)) {
994 free_cpumask_var(pinst->cpumask.pcpu);
995 goto err_free_serial_wq;
998 INIT_LIST_HEAD(&pinst->pslist);
1000 cpumask_copy(pinst->cpumask.pcpu, cpu_possible_mask);
1001 cpumask_copy(pinst->cpumask.cbcpu, cpu_possible_mask);
1003 if (padata_setup_cpumasks(pinst))
1004 goto err_free_masks;
1006 __padata_start(pinst);
1008 kobject_init(&pinst->kobj, &padata_attr_type);
1009 mutex_init(&pinst->lock);
1011 #ifdef CONFIG_HOTPLUG_CPU
1012 cpuhp_state_add_instance_nocalls_cpuslocked(hp_online,
1013 &pinst->cpu_online_node);
1014 cpuhp_state_add_instance_nocalls_cpuslocked(CPUHP_PADATA_DEAD,
1015 &pinst->cpu_dead_node);
1023 free_cpumask_var(pinst->cpumask.pcpu);
1024 free_cpumask_var(pinst->cpumask.cbcpu);
1026 destroy_workqueue(pinst->serial_wq);
1029 destroy_workqueue(pinst->parallel_wq);
1035 EXPORT_SYMBOL(padata_alloc);
1038 * padata_free - free a padata instance
1040 * @pinst: padata instance to free
1042 void padata_free(struct padata_instance *pinst)
1044 kobject_put(&pinst->kobj);
1046 EXPORT_SYMBOL(padata_free);
1049 * padata_alloc_shell - Allocate and initialize padata shell.
1051 * @pinst: Parent padata_instance object.
1053 * Return: new shell on success, NULL on error
1055 struct padata_shell *padata_alloc_shell(struct padata_instance *pinst)
1057 struct parallel_data *pd;
1058 struct padata_shell *ps;
1060 ps = kzalloc(sizeof(*ps), GFP_KERNEL);
1067 pd = padata_alloc_pd(ps);
1073 mutex_lock(&pinst->lock);
1074 RCU_INIT_POINTER(ps->pd, pd);
1075 list_add(&ps->list, &pinst->pslist);
1076 mutex_unlock(&pinst->lock);
1085 EXPORT_SYMBOL(padata_alloc_shell);
1088 * padata_free_shell - free a padata shell
1090 * @ps: padata shell to free
1092 void padata_free_shell(struct padata_shell *ps)
1097 mutex_lock(&ps->pinst->lock);
1098 list_del(&ps->list);
1099 padata_free_pd(rcu_dereference_protected(ps->pd, 1));
1100 mutex_unlock(&ps->pinst->lock);
1104 EXPORT_SYMBOL(padata_free_shell);
1106 void __init padata_init(void)
1108 unsigned int i, possible_cpus;
1109 #ifdef CONFIG_HOTPLUG_CPU
1112 ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "padata:online",
1113 padata_cpu_online, NULL);
1118 ret = cpuhp_setup_state_multi(CPUHP_PADATA_DEAD, "padata:dead",
1119 NULL, padata_cpu_dead);
1121 goto remove_online_state;
1124 possible_cpus = num_possible_cpus();
1125 padata_works = kmalloc_array(possible_cpus, sizeof(struct padata_work),
1128 goto remove_dead_state;
1130 for (i = 0; i < possible_cpus; ++i)
1131 list_add(&padata_works[i].pw_list, &padata_free_works);
1136 #ifdef CONFIG_HOTPLUG_CPU
1137 cpuhp_remove_multi_state(CPUHP_PADATA_DEAD);
1138 remove_online_state:
1139 cpuhp_remove_multi_state(hp_online);
1142 pr_warn("padata: initialization failed\n");