2 * padata.c - generic interface to process data streams in parallel
4 * Copyright (C) 2008, 2009 secunet Security Networks AG
5 * Copyright (C) 2008, 2009 Steffen Klassert <steffen.klassert@secunet.com>
7 * This program is free software; you can redistribute it and/or modify it
8 * under the terms and conditions of the GNU General Public License,
9 * version 2, as published by the Free Software Foundation.
11 * This program is distributed in the hope it will be useful, but WITHOUT
12 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
13 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for
16 * You should have received a copy of the GNU General Public License along with
17 * this program; if not, write to the Free Software Foundation, Inc.,
18 * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
21 #include <linux/module.h>
22 #include <linux/cpumask.h>
23 #include <linux/err.h>
24 #include <linux/cpu.h>
25 #include <linux/padata.h>
26 #include <linux/mutex.h>
27 #include <linux/sched.h>
28 #include <linux/slab.h>
29 #include <linux/rcupdate.h>
31 #define MAX_SEQ_NR INT_MAX - NR_CPUS
32 #define MAX_OBJ_NUM 1000
34 static int padata_index_to_cpu(struct parallel_data *pd, int cpu_index)
38 target_cpu = cpumask_first(pd->cpumask);
39 for (cpu = 0; cpu < cpu_index; cpu++)
40 target_cpu = cpumask_next(target_cpu, pd->cpumask);
45 static int padata_cpu_hash(struct padata_priv *padata)
48 struct parallel_data *pd;
53 * Hash the sequence numbers to the cpus by taking
54 * seq_nr mod. number of cpus in use.
56 cpu_index = padata->seq_nr % cpumask_weight(pd->cpumask);
58 return padata_index_to_cpu(pd, cpu_index);
61 static void padata_parallel_worker(struct work_struct *work)
63 struct padata_queue *queue;
64 struct parallel_data *pd;
65 struct padata_instance *pinst;
66 LIST_HEAD(local_list);
69 queue = container_of(work, struct padata_queue, pwork);
73 spin_lock(&queue->parallel.lock);
74 list_replace_init(&queue->parallel.list, &local_list);
75 spin_unlock(&queue->parallel.lock);
77 while (!list_empty(&local_list)) {
78 struct padata_priv *padata;
80 padata = list_entry(local_list.next,
81 struct padata_priv, list);
83 list_del_init(&padata->list);
85 padata->parallel(padata);
92 * padata_do_parallel - padata parallelization function
94 * @pinst: padata instance
95 * @padata: object to be parallelized
96 * @cb_cpu: cpu the serialization callback function will run on,
97 * must be in the cpumask of padata.
99 * The parallelization callback function will run with BHs off.
100 * Note: Every object which is parallelized by padata_do_parallel
101 * must be seen by padata_do_serial.
103 int padata_do_parallel(struct padata_instance *pinst,
104 struct padata_priv *padata, int cb_cpu)
107 struct padata_queue *queue;
108 struct parallel_data *pd;
112 pd = rcu_dereference(pinst->pd);
115 if (!(pinst->flags & PADATA_INIT))
119 if ((pinst->flags & PADATA_RESET))
122 if (atomic_read(&pd->refcnt) >= MAX_OBJ_NUM)
126 if (!cpumask_test_cpu(cb_cpu, pd->cpumask))
130 atomic_inc(&pd->refcnt);
132 padata->cb_cpu = cb_cpu;
134 if (unlikely(atomic_read(&pd->seq_nr) == pd->max_seq_nr))
135 atomic_set(&pd->seq_nr, -1);
137 padata->seq_nr = atomic_inc_return(&pd->seq_nr);
139 target_cpu = padata_cpu_hash(padata);
140 queue = per_cpu_ptr(pd->queue, target_cpu);
142 spin_lock(&queue->parallel.lock);
143 list_add_tail(&padata->list, &queue->parallel.list);
144 spin_unlock(&queue->parallel.lock);
146 queue_work_on(target_cpu, pinst->wq, &queue->pwork);
149 rcu_read_unlock_bh();
153 EXPORT_SYMBOL(padata_do_parallel);
155 static struct padata_priv *padata_get_next(struct parallel_data *pd)
157 int cpu, num_cpus, empty, calc_seq_nr;
158 int seq_nr, next_nr, overrun, next_overrun;
159 struct padata_queue *queue, *next_queue;
160 struct padata_priv *padata;
161 struct padata_list *reorder;
168 num_cpus = cpumask_weight(pd->cpumask);
170 for_each_cpu(cpu, pd->cpumask) {
171 queue = per_cpu_ptr(pd->queue, cpu);
172 reorder = &queue->reorder;
175 * Calculate the seq_nr of the object that should be
176 * next in this queue.
179 calc_seq_nr = (atomic_read(&queue->num_obj) * num_cpus)
182 if (unlikely(calc_seq_nr > pd->max_seq_nr)) {
183 calc_seq_nr = calc_seq_nr - pd->max_seq_nr - 1;
187 if (!list_empty(&reorder->list)) {
188 padata = list_entry(reorder->list.next,
189 struct padata_priv, list);
191 seq_nr = padata->seq_nr;
192 BUG_ON(calc_seq_nr != seq_nr);
194 seq_nr = calc_seq_nr;
198 if (next_nr < 0 || seq_nr < next_nr
199 || (next_overrun && !overrun)) {
201 next_overrun = overrun;
208 if (empty == num_cpus)
211 reorder = &next_queue->reorder;
213 if (!list_empty(&reorder->list)) {
214 padata = list_entry(reorder->list.next,
215 struct padata_priv, list);
217 if (unlikely(next_overrun)) {
218 for_each_cpu(cpu, pd->cpumask) {
219 queue = per_cpu_ptr(pd->queue, cpu);
220 atomic_set(&queue->num_obj, 0);
224 spin_lock(&reorder->lock);
225 list_del_init(&padata->list);
226 atomic_dec(&pd->reorder_objects);
227 spin_unlock(&reorder->lock);
229 atomic_inc(&next_queue->num_obj);
234 if (next_nr % num_cpus == next_queue->cpu_index) {
235 padata = ERR_PTR(-ENODATA);
239 padata = ERR_PTR(-EINPROGRESS);
244 static void padata_reorder(struct parallel_data *pd)
246 struct padata_priv *padata;
247 struct padata_queue *queue;
248 struct padata_instance *pinst = pd->pinst;
251 if (!spin_trylock_bh(&pd->lock))
255 padata = padata_get_next(pd);
257 if (!padata || PTR_ERR(padata) == -EINPROGRESS)
260 if (PTR_ERR(padata) == -ENODATA) {
261 spin_unlock_bh(&pd->lock);
265 queue = per_cpu_ptr(pd->queue, padata->cb_cpu);
267 spin_lock(&queue->serial.lock);
268 list_add_tail(&padata->list, &queue->serial.list);
269 spin_unlock(&queue->serial.lock);
271 queue_work_on(padata->cb_cpu, pinst->wq, &queue->swork);
274 spin_unlock_bh(&pd->lock);
276 if (atomic_read(&pd->reorder_objects))
283 static void padata_serial_worker(struct work_struct *work)
285 struct padata_queue *queue;
286 struct parallel_data *pd;
287 LIST_HEAD(local_list);
290 queue = container_of(work, struct padata_queue, swork);
293 spin_lock(&queue->serial.lock);
294 list_replace_init(&queue->serial.list, &local_list);
295 spin_unlock(&queue->serial.lock);
297 while (!list_empty(&local_list)) {
298 struct padata_priv *padata;
300 padata = list_entry(local_list.next,
301 struct padata_priv, list);
303 list_del_init(&padata->list);
305 padata->serial(padata);
306 atomic_dec(&pd->refcnt);
312 * padata_do_serial - padata serialization function
314 * @padata: object to be serialized.
316 * padata_do_serial must be called for every parallelized object.
317 * The serialization callback function will run with BHs off.
319 void padata_do_serial(struct padata_priv *padata)
322 struct padata_queue *queue;
323 struct parallel_data *pd;
328 queue = per_cpu_ptr(pd->queue, cpu);
330 spin_lock(&queue->reorder.lock);
331 atomic_inc(&pd->reorder_objects);
332 list_add_tail(&padata->list, &queue->reorder.list);
333 spin_unlock(&queue->reorder.lock);
339 EXPORT_SYMBOL(padata_do_serial);
341 static struct parallel_data *padata_alloc_pd(struct padata_instance *pinst,
342 const struct cpumask *cpumask)
344 int cpu, cpu_index, num_cpus;
345 struct padata_queue *queue;
346 struct parallel_data *pd;
350 pd = kzalloc(sizeof(struct parallel_data), GFP_KERNEL);
354 pd->queue = alloc_percpu(struct padata_queue);
358 if (!alloc_cpumask_var(&pd->cpumask, GFP_KERNEL))
361 cpumask_and(pd->cpumask, cpumask, cpu_active_mask);
363 for_each_cpu(cpu, pd->cpumask) {
364 queue = per_cpu_ptr(pd->queue, cpu);
368 queue->cpu_index = cpu_index;
371 INIT_LIST_HEAD(&queue->reorder.list);
372 INIT_LIST_HEAD(&queue->parallel.list);
373 INIT_LIST_HEAD(&queue->serial.list);
374 spin_lock_init(&queue->reorder.lock);
375 spin_lock_init(&queue->parallel.lock);
376 spin_lock_init(&queue->serial.lock);
378 INIT_WORK(&queue->pwork, padata_parallel_worker);
379 INIT_WORK(&queue->swork, padata_serial_worker);
380 atomic_set(&queue->num_obj, 0);
383 num_cpus = cpumask_weight(pd->cpumask);
384 pd->max_seq_nr = (MAX_SEQ_NR / num_cpus) * num_cpus - 1;
386 atomic_set(&pd->seq_nr, -1);
387 atomic_set(&pd->reorder_objects, 0);
388 atomic_set(&pd->refcnt, 0);
390 spin_lock_init(&pd->lock);
395 free_percpu(pd->queue);
402 static void padata_free_pd(struct parallel_data *pd)
404 free_cpumask_var(pd->cpumask);
405 free_percpu(pd->queue);
409 static void padata_replace(struct padata_instance *pinst,
410 struct parallel_data *pd_new)
412 struct parallel_data *pd_old = pinst->pd;
414 pinst->flags |= PADATA_RESET;
416 rcu_assign_pointer(pinst->pd, pd_new);
420 while (atomic_read(&pd_old->refcnt) != 0)
423 flush_workqueue(pinst->wq);
425 padata_free_pd(pd_old);
427 pinst->flags &= ~PADATA_RESET;
431 * padata_set_cpumask - set the cpumask that padata should use
433 * @pinst: padata instance
434 * @cpumask: the cpumask to use
436 int padata_set_cpumask(struct padata_instance *pinst,
437 cpumask_var_t cpumask)
439 struct parallel_data *pd;
442 mutex_lock(&pinst->lock);
446 pd = padata_alloc_pd(pinst, cpumask);
452 cpumask_copy(pinst->cpumask, cpumask);
454 padata_replace(pinst, pd);
459 mutex_unlock(&pinst->lock);
463 EXPORT_SYMBOL(padata_set_cpumask);
465 static int __padata_add_cpu(struct padata_instance *pinst, int cpu)
467 struct parallel_data *pd;
469 if (cpumask_test_cpu(cpu, cpu_active_mask)) {
470 pd = padata_alloc_pd(pinst, pinst->cpumask);
474 padata_replace(pinst, pd);
481 * padata_add_cpu - add a cpu to the padata cpumask
483 * @pinst: padata instance
486 int padata_add_cpu(struct padata_instance *pinst, int cpu)
490 mutex_lock(&pinst->lock);
493 cpumask_set_cpu(cpu, pinst->cpumask);
494 err = __padata_add_cpu(pinst, cpu);
497 mutex_unlock(&pinst->lock);
501 EXPORT_SYMBOL(padata_add_cpu);
503 static int __padata_remove_cpu(struct padata_instance *pinst, int cpu)
505 struct parallel_data *pd;
507 if (cpumask_test_cpu(cpu, cpu_online_mask)) {
508 pd = padata_alloc_pd(pinst, pinst->cpumask);
512 padata_replace(pinst, pd);
519 * padata_remove_cpu - remove a cpu from the padata cpumask
521 * @pinst: padata instance
522 * @cpu: cpu to remove
524 int padata_remove_cpu(struct padata_instance *pinst, int cpu)
528 mutex_lock(&pinst->lock);
531 cpumask_clear_cpu(cpu, pinst->cpumask);
532 err = __padata_remove_cpu(pinst, cpu);
535 mutex_unlock(&pinst->lock);
539 EXPORT_SYMBOL(padata_remove_cpu);
542 * padata_start - start the parallel processing
544 * @pinst: padata instance to start
546 void padata_start(struct padata_instance *pinst)
548 mutex_lock(&pinst->lock);
549 pinst->flags |= PADATA_INIT;
550 mutex_unlock(&pinst->lock);
552 EXPORT_SYMBOL(padata_start);
555 * padata_stop - stop the parallel processing
557 * @pinst: padata instance to stop
559 void padata_stop(struct padata_instance *pinst)
561 mutex_lock(&pinst->lock);
562 pinst->flags &= ~PADATA_INIT;
563 mutex_unlock(&pinst->lock);
565 EXPORT_SYMBOL(padata_stop);
567 #ifdef CONFIG_HOTPLUG_CPU
568 static int padata_cpu_callback(struct notifier_block *nfb,
569 unsigned long action, void *hcpu)
572 struct padata_instance *pinst;
573 int cpu = (unsigned long)hcpu;
575 pinst = container_of(nfb, struct padata_instance, cpu_notifier);
579 case CPU_ONLINE_FROZEN:
580 if (!cpumask_test_cpu(cpu, pinst->cpumask))
582 mutex_lock(&pinst->lock);
583 err = __padata_add_cpu(pinst, cpu);
584 mutex_unlock(&pinst->lock);
589 case CPU_DOWN_PREPARE:
590 case CPU_DOWN_PREPARE_FROZEN:
591 if (!cpumask_test_cpu(cpu, pinst->cpumask))
593 mutex_lock(&pinst->lock);
594 err = __padata_remove_cpu(pinst, cpu);
595 mutex_unlock(&pinst->lock);
600 case CPU_UP_CANCELED:
601 case CPU_UP_CANCELED_FROZEN:
602 if (!cpumask_test_cpu(cpu, pinst->cpumask))
604 mutex_lock(&pinst->lock);
605 __padata_remove_cpu(pinst, cpu);
606 mutex_unlock(&pinst->lock);
608 case CPU_DOWN_FAILED:
609 case CPU_DOWN_FAILED_FROZEN:
610 if (!cpumask_test_cpu(cpu, pinst->cpumask))
612 mutex_lock(&pinst->lock);
613 __padata_add_cpu(pinst, cpu);
614 mutex_unlock(&pinst->lock);
622 * padata_alloc - allocate and initialize a padata instance
624 * @cpumask: cpumask that padata uses for parallelization
625 * @wq: workqueue to use for the allocated padata instance
627 struct padata_instance *padata_alloc(const struct cpumask *cpumask,
628 struct workqueue_struct *wq)
630 struct padata_instance *pinst;
631 struct parallel_data *pd;
633 pinst = kzalloc(sizeof(struct padata_instance), GFP_KERNEL);
639 pd = padata_alloc_pd(pinst, cpumask);
643 if (!alloc_cpumask_var(&pinst->cpumask, GFP_KERNEL))
646 rcu_assign_pointer(pinst->pd, pd);
650 cpumask_copy(pinst->cpumask, cpumask);
654 #ifdef CONFIG_HOTPLUG_CPU
655 pinst->cpu_notifier.notifier_call = padata_cpu_callback;
656 pinst->cpu_notifier.priority = 0;
657 register_hotcpu_notifier(&pinst->cpu_notifier);
662 mutex_init(&pinst->lock);
674 EXPORT_SYMBOL(padata_alloc);
677 * padata_free - free a padata instance
679 * @ padata_inst: padata instance to free
681 void padata_free(struct padata_instance *pinst)
687 while (atomic_read(&pinst->pd->refcnt) != 0)
690 #ifdef CONFIG_HOTPLUG_CPU
691 unregister_hotcpu_notifier(&pinst->cpu_notifier);
693 padata_free_pd(pinst->pd);
694 free_cpumask_var(pinst->cpumask);
697 EXPORT_SYMBOL(padata_free);