1 // SPDX-License-Identifier: GPL-2.0-only
3 * linux/net/sunrpc/sched.c
5 * Scheduling for synchronous and asynchronous RPC requests.
7 * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
9 * TCP NFS related read + write fixes
10 * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
13 #include <linux/module.h>
15 #include <linux/sched.h>
16 #include <linux/interrupt.h>
17 #include <linux/slab.h>
18 #include <linux/mempool.h>
19 #include <linux/smp.h>
20 #include <linux/spinlock.h>
21 #include <linux/mutex.h>
22 #include <linux/freezer.h>
23 #include <linux/sched/mm.h>
25 #include <linux/sunrpc/clnt.h>
26 #include <linux/sunrpc/metrics.h>
30 #define CREATE_TRACE_POINTS
31 #include <trace/events/sunrpc.h>
34 * RPC slabs and memory pools
36 #define RPC_BUFFER_MAXSIZE (2048)
37 #define RPC_BUFFER_POOLSIZE (8)
38 #define RPC_TASK_POOLSIZE (8)
39 static struct kmem_cache *rpc_task_slabp __read_mostly;
40 static struct kmem_cache *rpc_buffer_slabp __read_mostly;
41 static mempool_t *rpc_task_mempool __read_mostly;
42 static mempool_t *rpc_buffer_mempool __read_mostly;
44 static void rpc_async_schedule(struct work_struct *);
45 static void rpc_release_task(struct rpc_task *task);
46 static void __rpc_queue_timer_fn(struct work_struct *);
49 * RPC tasks sit here while waiting for conditions to improve.
51 static struct rpc_wait_queue delay_queue;
54 * rpciod-related stuff
56 struct workqueue_struct *rpciod_workqueue __read_mostly;
57 struct workqueue_struct *xprtiod_workqueue __read_mostly;
58 EXPORT_SYMBOL_GPL(xprtiod_workqueue);
61 rpc_task_timeout(const struct rpc_task *task)
63 unsigned long timeout = READ_ONCE(task->tk_timeout);
66 unsigned long now = jiffies;
67 if (time_before(now, timeout))
72 EXPORT_SYMBOL_GPL(rpc_task_timeout);
75 * Disable the timer for a given RPC task. Should be called with
76 * queue->lock and bh_disabled in order to avoid races within
80 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
82 if (list_empty(&task->u.tk_wait.timer_list))
85 list_del(&task->u.tk_wait.timer_list);
86 if (list_empty(&queue->timer_list.list))
87 cancel_delayed_work(&queue->timer_list.dwork);
91 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
93 unsigned long now = jiffies;
94 queue->timer_list.expires = expires;
95 if (time_before_eq(expires, now))
99 mod_delayed_work(rpciod_workqueue, &queue->timer_list.dwork, expires);
103 * Set up a timer for the current task.
106 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task,
107 unsigned long timeout)
109 task->tk_timeout = timeout;
110 if (list_empty(&queue->timer_list.list) || time_before(timeout, queue->timer_list.expires))
111 rpc_set_queue_timer(queue, timeout);
112 list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
115 static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
117 if (queue->priority != priority) {
118 queue->priority = priority;
119 queue->nr = 1U << priority;
123 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
125 rpc_set_waitqueue_priority(queue, queue->maxpriority);
129 * Add a request to a queue list
132 __rpc_list_enqueue_task(struct list_head *q, struct rpc_task *task)
136 list_for_each_entry(t, q, u.tk_wait.list) {
137 if (t->tk_owner == task->tk_owner) {
138 list_add_tail(&task->u.tk_wait.links,
139 &t->u.tk_wait.links);
140 /* Cache the queue head in task->u.tk_wait.list */
141 task->u.tk_wait.list.next = q;
142 task->u.tk_wait.list.prev = NULL;
146 INIT_LIST_HEAD(&task->u.tk_wait.links);
147 list_add_tail(&task->u.tk_wait.list, q);
151 * Remove request from a queue list
154 __rpc_list_dequeue_task(struct rpc_task *task)
159 if (task->u.tk_wait.list.prev == NULL) {
160 list_del(&task->u.tk_wait.links);
163 if (!list_empty(&task->u.tk_wait.links)) {
164 t = list_first_entry(&task->u.tk_wait.links,
167 /* Assume __rpc_list_enqueue_task() cached the queue head */
168 q = t->u.tk_wait.list.next;
169 list_add_tail(&t->u.tk_wait.list, q);
170 list_del(&task->u.tk_wait.links);
172 list_del(&task->u.tk_wait.list);
176 * Add new request to a priority queue.
178 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
179 struct rpc_task *task,
180 unsigned char queue_priority)
182 if (unlikely(queue_priority > queue->maxpriority))
183 queue_priority = queue->maxpriority;
184 __rpc_list_enqueue_task(&queue->tasks[queue_priority], task);
188 * Add new request to wait queue.
190 * Swapper tasks always get inserted at the head of the queue.
191 * This should avoid many nasty memory deadlocks and hopefully
192 * improve overall performance.
193 * Everyone else gets appended to the queue to ensure proper FIFO behavior.
195 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
196 struct rpc_task *task,
197 unsigned char queue_priority)
199 INIT_LIST_HEAD(&task->u.tk_wait.timer_list);
200 if (RPC_IS_PRIORITY(queue))
201 __rpc_add_wait_queue_priority(queue, task, queue_priority);
202 else if (RPC_IS_SWAPPER(task))
203 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
205 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
206 task->tk_waitqueue = queue;
208 /* barrier matches the read in rpc_wake_up_task_queue_locked() */
210 rpc_set_queued(task);
214 * Remove request from a priority queue.
216 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
218 __rpc_list_dequeue_task(task);
222 * Remove request from queue.
223 * Note: must be called with spin lock held.
225 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
227 __rpc_disable_timer(queue, task);
228 if (RPC_IS_PRIORITY(queue))
229 __rpc_remove_wait_queue_priority(task);
231 list_del(&task->u.tk_wait.list);
235 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
239 spin_lock_init(&queue->lock);
240 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
241 INIT_LIST_HEAD(&queue->tasks[i]);
242 queue->maxpriority = nr_queues - 1;
243 rpc_reset_waitqueue_priority(queue);
245 queue->timer_list.expires = 0;
246 INIT_DELAYED_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
247 INIT_LIST_HEAD(&queue->timer_list.list);
248 rpc_assign_waitqueue_name(queue, qname);
251 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
253 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
255 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
257 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
259 __rpc_init_priority_wait_queue(queue, qname, 1);
261 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
263 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
265 cancel_delayed_work_sync(&queue->timer_list.dwork);
267 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
269 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
271 freezable_schedule_unsafe();
272 if (signal_pending_state(mode, current))
277 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
278 static void rpc_task_set_debuginfo(struct rpc_task *task)
280 static atomic_t rpc_pid;
282 task->tk_pid = atomic_inc_return(&rpc_pid);
285 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
290 static void rpc_set_active(struct rpc_task *task)
292 rpc_task_set_debuginfo(task);
293 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
294 trace_rpc_task_begin(task, NULL);
298 * Mark an RPC call as having completed by clearing the 'active' bit
299 * and then waking up all tasks that were sleeping.
301 static int rpc_complete_task(struct rpc_task *task)
303 void *m = &task->tk_runstate;
304 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
305 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
309 trace_rpc_task_complete(task, NULL);
311 spin_lock_irqsave(&wq->lock, flags);
312 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
313 ret = atomic_dec_and_test(&task->tk_count);
314 if (waitqueue_active(wq))
315 __wake_up_locked_key(wq, TASK_NORMAL, &k);
316 spin_unlock_irqrestore(&wq->lock, flags);
321 * Allow callers to wait for completion of an RPC call
323 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
324 * to enforce taking of the wq->lock and hence avoid races with
325 * rpc_complete_task().
327 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
330 action = rpc_wait_bit_killable;
331 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
332 action, TASK_KILLABLE);
334 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
337 * Make an RPC task runnable.
339 * Note: If the task is ASYNC, and is being made runnable after sitting on an
340 * rpc_wait_queue, this must be called with the queue spinlock held to protect
341 * the wait queue operation.
342 * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
343 * which is needed to ensure that __rpc_execute() doesn't loop (due to the
344 * lockless RPC_IS_QUEUED() test) before we've had a chance to test
345 * the RPC_TASK_RUNNING flag.
347 static void rpc_make_runnable(struct workqueue_struct *wq,
348 struct rpc_task *task)
350 bool need_wakeup = !rpc_test_and_set_running(task);
352 rpc_clear_queued(task);
355 if (RPC_IS_ASYNC(task)) {
356 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
357 queue_work(wq, &task->u.tk_work);
359 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
363 * Prepare for sleeping on a wait queue.
364 * By always appending tasks to the list we ensure FIFO behavior.
365 * NB: An RPC task will only receive interrupt-driven events as long
366 * as it's on a wait queue.
368 static void __rpc_do_sleep_on_priority(struct rpc_wait_queue *q,
369 struct rpc_task *task,
370 unsigned char queue_priority)
372 trace_rpc_task_sleep(task, q);
374 __rpc_add_wait_queue(q, task, queue_priority);
377 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
378 struct rpc_task *task,
379 unsigned char queue_priority)
381 if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
383 __rpc_do_sleep_on_priority(q, task, queue_priority);
386 static void __rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
387 struct rpc_task *task, unsigned long timeout,
388 unsigned char queue_priority)
390 if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
392 if (time_is_after_jiffies(timeout)) {
393 __rpc_do_sleep_on_priority(q, task, queue_priority);
394 __rpc_add_timer(q, task, timeout);
396 task->tk_status = -ETIMEDOUT;
399 static void rpc_set_tk_callback(struct rpc_task *task, rpc_action action)
401 if (action && !WARN_ON_ONCE(task->tk_callback != NULL))
402 task->tk_callback = action;
405 static bool rpc_sleep_check_activated(struct rpc_task *task)
407 /* We shouldn't ever put an inactive task to sleep */
408 if (WARN_ON_ONCE(!RPC_IS_ACTIVATED(task))) {
409 task->tk_status = -EIO;
410 rpc_put_task_async(task);
416 void rpc_sleep_on_timeout(struct rpc_wait_queue *q, struct rpc_task *task,
417 rpc_action action, unsigned long timeout)
419 if (!rpc_sleep_check_activated(task))
422 rpc_set_tk_callback(task, action);
425 * Protect the queue operations.
428 __rpc_sleep_on_priority_timeout(q, task, timeout, task->tk_priority);
429 spin_unlock(&q->lock);
431 EXPORT_SYMBOL_GPL(rpc_sleep_on_timeout);
433 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
436 if (!rpc_sleep_check_activated(task))
439 rpc_set_tk_callback(task, action);
441 WARN_ON_ONCE(task->tk_timeout != 0);
443 * Protect the queue operations.
446 __rpc_sleep_on_priority(q, task, task->tk_priority);
447 spin_unlock(&q->lock);
449 EXPORT_SYMBOL_GPL(rpc_sleep_on);
451 void rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
452 struct rpc_task *task, unsigned long timeout, int priority)
454 if (!rpc_sleep_check_activated(task))
457 priority -= RPC_PRIORITY_LOW;
459 * Protect the queue operations.
462 __rpc_sleep_on_priority_timeout(q, task, timeout, priority);
463 spin_unlock(&q->lock);
465 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority_timeout);
467 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
470 if (!rpc_sleep_check_activated(task))
473 WARN_ON_ONCE(task->tk_timeout != 0);
474 priority -= RPC_PRIORITY_LOW;
476 * Protect the queue operations.
479 __rpc_sleep_on_priority(q, task, priority);
480 spin_unlock(&q->lock);
482 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
485 * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
486 * @wq: workqueue on which to run task
488 * @task: task to be woken up
490 * Caller must hold queue->lock, and have cleared the task queued flag.
492 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
493 struct rpc_wait_queue *queue,
494 struct rpc_task *task)
496 /* Has the task been executed yet? If not, we cannot wake it up! */
497 if (!RPC_IS_ACTIVATED(task)) {
498 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
502 trace_rpc_task_wakeup(task, queue);
504 __rpc_remove_wait_queue(queue, task);
506 rpc_make_runnable(wq, task);
510 * Wake up a queued task while the queue lock is being held
512 static struct rpc_task *
513 rpc_wake_up_task_on_wq_queue_action_locked(struct workqueue_struct *wq,
514 struct rpc_wait_queue *queue, struct rpc_task *task,
515 bool (*action)(struct rpc_task *, void *), void *data)
517 if (RPC_IS_QUEUED(task)) {
519 if (task->tk_waitqueue == queue) {
520 if (action == NULL || action(task, data)) {
521 __rpc_do_wake_up_task_on_wq(wq, queue, task);
530 * Wake up a queued task while the queue lock is being held
532 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue,
533 struct rpc_task *task)
535 rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
540 * Wake up a task on a specific queue
542 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
544 if (!RPC_IS_QUEUED(task))
546 spin_lock(&queue->lock);
547 rpc_wake_up_task_queue_locked(queue, task);
548 spin_unlock(&queue->lock);
550 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
552 static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
554 task->tk_status = *(int *)status;
559 rpc_wake_up_task_queue_set_status_locked(struct rpc_wait_queue *queue,
560 struct rpc_task *task, int status)
562 rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
563 task, rpc_task_action_set_status, &status);
567 * rpc_wake_up_queued_task_set_status - wake up a task and set task->tk_status
568 * @queue: pointer to rpc_wait_queue
569 * @task: pointer to rpc_task
570 * @status: integer error value
572 * If @task is queued on @queue, then it is woken up, and @task->tk_status is
573 * set to the value of @status.
576 rpc_wake_up_queued_task_set_status(struct rpc_wait_queue *queue,
577 struct rpc_task *task, int status)
579 if (!RPC_IS_QUEUED(task))
581 spin_lock(&queue->lock);
582 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
583 spin_unlock(&queue->lock);
587 * Wake up the next task on a priority queue.
589 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
592 struct rpc_task *task;
595 * Service a batch of tasks from a single owner.
597 q = &queue->tasks[queue->priority];
598 if (!list_empty(q) && --queue->nr) {
599 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
604 * Service the next queue.
607 if (q == &queue->tasks[0])
608 q = &queue->tasks[queue->maxpriority];
611 if (!list_empty(q)) {
612 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
615 } while (q != &queue->tasks[queue->priority]);
617 rpc_reset_waitqueue_priority(queue);
621 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
626 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
628 if (RPC_IS_PRIORITY(queue))
629 return __rpc_find_next_queued_priority(queue);
630 if (!list_empty(&queue->tasks[0]))
631 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
636 * Wake up the first task on the wait queue.
638 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
639 struct rpc_wait_queue *queue,
640 bool (*func)(struct rpc_task *, void *), void *data)
642 struct rpc_task *task = NULL;
644 spin_lock(&queue->lock);
645 task = __rpc_find_next_queued(queue);
647 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
649 spin_unlock(&queue->lock);
655 * Wake up the first task on the wait queue.
657 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
658 bool (*func)(struct rpc_task *, void *), void *data)
660 return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
662 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
664 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
670 * Wake up the next task on the wait queue.
672 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
674 return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
676 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
679 * rpc_wake_up_locked - wake up all rpc_tasks
680 * @queue: rpc_wait_queue on which the tasks are sleeping
683 static void rpc_wake_up_locked(struct rpc_wait_queue *queue)
685 struct rpc_task *task;
688 task = __rpc_find_next_queued(queue);
691 rpc_wake_up_task_queue_locked(queue, task);
696 * rpc_wake_up - wake up all rpc_tasks
697 * @queue: rpc_wait_queue on which the tasks are sleeping
701 void rpc_wake_up(struct rpc_wait_queue *queue)
703 spin_lock(&queue->lock);
704 rpc_wake_up_locked(queue);
705 spin_unlock(&queue->lock);
707 EXPORT_SYMBOL_GPL(rpc_wake_up);
710 * rpc_wake_up_status_locked - wake up all rpc_tasks and set their status value.
711 * @queue: rpc_wait_queue on which the tasks are sleeping
712 * @status: status value to set
714 static void rpc_wake_up_status_locked(struct rpc_wait_queue *queue, int status)
716 struct rpc_task *task;
719 task = __rpc_find_next_queued(queue);
722 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
727 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
728 * @queue: rpc_wait_queue on which the tasks are sleeping
729 * @status: status value to set
733 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
735 spin_lock(&queue->lock);
736 rpc_wake_up_status_locked(queue, status);
737 spin_unlock(&queue->lock);
739 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
741 static void __rpc_queue_timer_fn(struct work_struct *work)
743 struct rpc_wait_queue *queue = container_of(work,
744 struct rpc_wait_queue,
745 timer_list.dwork.work);
746 struct rpc_task *task, *n;
747 unsigned long expires, now, timeo;
749 spin_lock(&queue->lock);
750 expires = now = jiffies;
751 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
752 timeo = task->tk_timeout;
753 if (time_after_eq(now, timeo)) {
754 trace_rpc_task_timeout(task, task->tk_action);
755 task->tk_status = -ETIMEDOUT;
756 rpc_wake_up_task_queue_locked(queue, task);
759 if (expires == now || time_after(expires, timeo))
762 if (!list_empty(&queue->timer_list.list))
763 rpc_set_queue_timer(queue, expires);
764 spin_unlock(&queue->lock);
767 static void __rpc_atrun(struct rpc_task *task)
769 if (task->tk_status == -ETIMEDOUT)
774 * Run a task at a later time
776 void rpc_delay(struct rpc_task *task, unsigned long delay)
778 rpc_sleep_on_timeout(&delay_queue, task, __rpc_atrun, jiffies + delay);
780 EXPORT_SYMBOL_GPL(rpc_delay);
783 * Helper to call task->tk_ops->rpc_call_prepare
785 void rpc_prepare_task(struct rpc_task *task)
787 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
791 rpc_init_task_statistics(struct rpc_task *task)
793 /* Initialize retry counters */
794 task->tk_garb_retry = 2;
795 task->tk_cred_retry = 2;
796 task->tk_rebind_retry = 2;
798 /* starting timestamp */
799 task->tk_start = ktime_get();
803 rpc_reset_task_statistics(struct rpc_task *task)
805 task->tk_timeouts = 0;
806 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
807 rpc_init_task_statistics(task);
811 * Helper that calls task->tk_ops->rpc_call_done if it exists
813 void rpc_exit_task(struct rpc_task *task)
815 trace_rpc_task_end(task, task->tk_action);
816 task->tk_action = NULL;
817 if (task->tk_ops->rpc_count_stats)
818 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
819 else if (task->tk_client)
820 rpc_count_iostats(task, task->tk_client->cl_metrics);
821 if (task->tk_ops->rpc_call_done != NULL) {
822 task->tk_ops->rpc_call_done(task, task->tk_calldata);
823 if (task->tk_action != NULL) {
824 /* Always release the RPC slot and buffer memory */
826 rpc_reset_task_statistics(task);
831 void rpc_signal_task(struct rpc_task *task)
833 struct rpc_wait_queue *queue;
835 if (!RPC_IS_ACTIVATED(task))
838 trace_rpc_task_signalled(task, task->tk_action);
839 set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
840 smp_mb__after_atomic();
841 queue = READ_ONCE(task->tk_waitqueue);
843 rpc_wake_up_queued_task_set_status(queue, task, -ERESTARTSYS);
846 void rpc_exit(struct rpc_task *task, int status)
848 task->tk_status = status;
849 task->tk_action = rpc_exit_task;
850 rpc_wake_up_queued_task(task->tk_waitqueue, task);
852 EXPORT_SYMBOL_GPL(rpc_exit);
854 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
856 if (ops->rpc_release != NULL)
857 ops->rpc_release(calldata);
861 * This is the RPC `scheduler' (or rather, the finite state machine).
863 static void __rpc_execute(struct rpc_task *task)
865 struct rpc_wait_queue *queue;
866 int task_is_async = RPC_IS_ASYNC(task);
869 WARN_ON_ONCE(RPC_IS_QUEUED(task));
870 if (RPC_IS_QUEUED(task))
874 void (*do_action)(struct rpc_task *);
877 * Perform the next FSM step or a pending callback.
879 * tk_action may be NULL if the task has been killed.
880 * In particular, note that rpc_killall_tasks may
881 * do this at any time, so beware when dereferencing.
883 do_action = task->tk_action;
884 if (task->tk_callback) {
885 do_action = task->tk_callback;
886 task->tk_callback = NULL;
890 trace_rpc_task_run_action(task, do_action);
894 * Lockless check for whether task is sleeping or not.
896 if (!RPC_IS_QUEUED(task))
900 * Signalled tasks should exit rather than sleep.
902 if (RPC_SIGNALLED(task)) {
903 task->tk_rpc_status = -ERESTARTSYS;
904 rpc_exit(task, -ERESTARTSYS);
908 * The queue->lock protects against races with
909 * rpc_make_runnable().
911 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
912 * rpc_task, rpc_make_runnable() can assign it to a
913 * different workqueue. We therefore cannot assume that the
914 * rpc_task pointer may still be dereferenced.
916 queue = task->tk_waitqueue;
917 spin_lock(&queue->lock);
918 if (!RPC_IS_QUEUED(task)) {
919 spin_unlock(&queue->lock);
922 rpc_clear_running(task);
923 spin_unlock(&queue->lock);
927 /* sync task: sleep here */
928 trace_rpc_task_sync_sleep(task, task->tk_action);
929 status = out_of_line_wait_on_bit(&task->tk_runstate,
930 RPC_TASK_QUEUED, rpc_wait_bit_killable,
934 * When a sync task receives a signal, it exits with
935 * -ERESTARTSYS. In order to catch any callbacks that
936 * clean up after sleeping on some queue, we don't
937 * break the loop here, but go around once more.
939 trace_rpc_task_signalled(task, task->tk_action);
940 set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
941 task->tk_rpc_status = -ERESTARTSYS;
942 rpc_exit(task, -ERESTARTSYS);
944 trace_rpc_task_sync_wake(task, task->tk_action);
947 /* Release all resources associated with the task */
948 rpc_release_task(task);
952 * User-visible entry point to the scheduler.
954 * This may be called recursively if e.g. an async NFS task updates
955 * the attributes and finds that dirty pages must be flushed.
956 * NOTE: Upon exit of this function the task is guaranteed to be
957 * released. In particular note that tk_release() will have
958 * been called, so your task memory may have been freed.
960 void rpc_execute(struct rpc_task *task)
962 bool is_async = RPC_IS_ASYNC(task);
964 rpc_set_active(task);
965 rpc_make_runnable(rpciod_workqueue, task);
967 unsigned int pflags = memalloc_nofs_save();
969 memalloc_nofs_restore(pflags);
973 static void rpc_async_schedule(struct work_struct *work)
975 unsigned int pflags = memalloc_nofs_save();
977 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
978 memalloc_nofs_restore(pflags);
982 * rpc_malloc - allocate RPC buffer resources
985 * A single memory region is allocated, which is split between the
986 * RPC call and RPC reply that this task is being used for. When
987 * this RPC is retired, the memory is released by calling rpc_free.
989 * To prevent rpciod from hanging, this allocator never sleeps,
990 * returning -ENOMEM and suppressing warning if the request cannot
991 * be serviced immediately. The caller can arrange to sleep in a
992 * way that is safe for rpciod.
994 * Most requests are 'small' (under 2KiB) and can be serviced from a
995 * mempool, ensuring that NFS reads and writes can always proceed,
996 * and that there is good locality of reference for these buffers.
998 int rpc_malloc(struct rpc_task *task)
1000 struct rpc_rqst *rqst = task->tk_rqstp;
1001 size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
1002 struct rpc_buffer *buf;
1003 gfp_t gfp = GFP_NOFS;
1005 if (RPC_IS_SWAPPER(task))
1006 gfp = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
1008 size += sizeof(struct rpc_buffer);
1009 if (size <= RPC_BUFFER_MAXSIZE)
1010 buf = mempool_alloc(rpc_buffer_mempool, gfp);
1012 buf = kmalloc(size, gfp);
1018 rqst->rq_buffer = buf->data;
1019 rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1022 EXPORT_SYMBOL_GPL(rpc_malloc);
1025 * rpc_free - free RPC buffer resources allocated via rpc_malloc
1029 void rpc_free(struct rpc_task *task)
1031 void *buffer = task->tk_rqstp->rq_buffer;
1033 struct rpc_buffer *buf;
1035 buf = container_of(buffer, struct rpc_buffer, data);
1038 if (size <= RPC_BUFFER_MAXSIZE)
1039 mempool_free(buf, rpc_buffer_mempool);
1043 EXPORT_SYMBOL_GPL(rpc_free);
1046 * Creation and deletion of RPC task structures
1048 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1050 memset(task, 0, sizeof(*task));
1051 atomic_set(&task->tk_count, 1);
1052 task->tk_flags = task_setup_data->flags;
1053 task->tk_ops = task_setup_data->callback_ops;
1054 task->tk_calldata = task_setup_data->callback_data;
1055 INIT_LIST_HEAD(&task->tk_task);
1057 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1058 task->tk_owner = current->tgid;
1060 /* Initialize workqueue for async tasks */
1061 task->tk_workqueue = task_setup_data->workqueue;
1063 task->tk_xprt = rpc_task_get_xprt(task_setup_data->rpc_client,
1064 xprt_get(task_setup_data->rpc_xprt));
1066 task->tk_op_cred = get_rpccred(task_setup_data->rpc_op_cred);
1068 if (task->tk_ops->rpc_call_prepare != NULL)
1069 task->tk_action = rpc_prepare_task;
1071 rpc_init_task_statistics(task);
1074 static struct rpc_task *
1075 rpc_alloc_task(void)
1077 return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
1081 * Create a new task for the specified client.
1083 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1085 struct rpc_task *task = setup_data->task;
1086 unsigned short flags = 0;
1089 task = rpc_alloc_task();
1090 flags = RPC_TASK_DYNAMIC;
1093 rpc_init_task(task, setup_data);
1094 task->tk_flags |= flags;
1099 * rpc_free_task - release rpc task and perform cleanups
1101 * Note that we free up the rpc_task _after_ rpc_release_calldata()
1102 * in order to work around a workqueue dependency issue.
1105 * "Workqueue currently considers two work items to be the same if they're
1106 * on the same address and won't execute them concurrently - ie. it
1107 * makes a work item which is queued again while being executed wait
1108 * for the previous execution to complete.
1110 * If a work function frees the work item, and then waits for an event
1111 * which should be performed by another work item and *that* work item
1112 * recycles the freed work item, it can create a false dependency loop.
1113 * There really is no reliable way to detect this short of verifying
1114 * every memory free."
1117 static void rpc_free_task(struct rpc_task *task)
1119 unsigned short tk_flags = task->tk_flags;
1121 put_rpccred(task->tk_op_cred);
1122 rpc_release_calldata(task->tk_ops, task->tk_calldata);
1124 if (tk_flags & RPC_TASK_DYNAMIC)
1125 mempool_free(task, rpc_task_mempool);
1128 static void rpc_async_release(struct work_struct *work)
1130 unsigned int pflags = memalloc_nofs_save();
1132 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1133 memalloc_nofs_restore(pflags);
1136 static void rpc_release_resources_task(struct rpc_task *task)
1139 if (task->tk_msg.rpc_cred) {
1140 if (!(task->tk_flags & RPC_TASK_CRED_NOREF))
1141 put_cred(task->tk_msg.rpc_cred);
1142 task->tk_msg.rpc_cred = NULL;
1144 rpc_task_release_client(task);
1147 static void rpc_final_put_task(struct rpc_task *task,
1148 struct workqueue_struct *q)
1151 INIT_WORK(&task->u.tk_work, rpc_async_release);
1152 queue_work(q, &task->u.tk_work);
1154 rpc_free_task(task);
1157 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1159 if (atomic_dec_and_test(&task->tk_count)) {
1160 rpc_release_resources_task(task);
1161 rpc_final_put_task(task, q);
1165 void rpc_put_task(struct rpc_task *task)
1167 rpc_do_put_task(task, NULL);
1169 EXPORT_SYMBOL_GPL(rpc_put_task);
1171 void rpc_put_task_async(struct rpc_task *task)
1173 rpc_do_put_task(task, task->tk_workqueue);
1175 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1177 static void rpc_release_task(struct rpc_task *task)
1179 WARN_ON_ONCE(RPC_IS_QUEUED(task));
1181 rpc_release_resources_task(task);
1184 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1185 * so it should be safe to use task->tk_count as a test for whether
1186 * or not any other processes still hold references to our rpc_task.
1188 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1189 /* Wake up anyone who may be waiting for task completion */
1190 if (!rpc_complete_task(task))
1193 if (!atomic_dec_and_test(&task->tk_count))
1196 rpc_final_put_task(task, task->tk_workqueue);
1201 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1204 void rpciod_down(void)
1206 module_put(THIS_MODULE);
1210 * Start up the rpciod workqueue.
1212 static int rpciod_start(void)
1214 struct workqueue_struct *wq;
1217 * Create the rpciod thread and wait for it to start.
1219 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1222 rpciod_workqueue = wq;
1223 /* Note: highpri because network receive is latency sensitive */
1224 wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1227 xprtiod_workqueue = wq;
1230 wq = rpciod_workqueue;
1231 rpciod_workqueue = NULL;
1232 destroy_workqueue(wq);
1237 static void rpciod_stop(void)
1239 struct workqueue_struct *wq = NULL;
1241 if (rpciod_workqueue == NULL)
1244 wq = rpciod_workqueue;
1245 rpciod_workqueue = NULL;
1246 destroy_workqueue(wq);
1247 wq = xprtiod_workqueue;
1248 xprtiod_workqueue = NULL;
1249 destroy_workqueue(wq);
1253 rpc_destroy_mempool(void)
1256 mempool_destroy(rpc_buffer_mempool);
1257 mempool_destroy(rpc_task_mempool);
1258 kmem_cache_destroy(rpc_task_slabp);
1259 kmem_cache_destroy(rpc_buffer_slabp);
1260 rpc_destroy_wait_queue(&delay_queue);
1264 rpc_init_mempool(void)
1267 * The following is not strictly a mempool initialisation,
1268 * but there is no harm in doing it here
1270 rpc_init_wait_queue(&delay_queue, "delayq");
1271 if (!rpciod_start())
1274 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1275 sizeof(struct rpc_task),
1276 0, SLAB_HWCACHE_ALIGN,
1278 if (!rpc_task_slabp)
1280 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1282 0, SLAB_HWCACHE_ALIGN,
1284 if (!rpc_buffer_slabp)
1286 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1288 if (!rpc_task_mempool)
1290 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1292 if (!rpc_buffer_mempool)
1296 rpc_destroy_mempool();