2 * linux/net/sunrpc/sched.c
4 * Scheduling for synchronous and asynchronous RPC requests.
6 * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
8 * TCP NFS related read + write fixes
9 * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
12 #include <linux/module.h>
14 #include <linux/sched.h>
15 #include <linux/interrupt.h>
16 #include <linux/slab.h>
17 #include <linux/mempool.h>
18 #include <linux/smp.h>
19 #include <linux/spinlock.h>
20 #include <linux/mutex.h>
21 #include <linux/freezer.h>
23 #include <linux/sunrpc/clnt.h>
27 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
28 #define RPCDBG_FACILITY RPCDBG_SCHED
31 #define CREATE_TRACE_POINTS
32 #include <trace/events/sunrpc.h>
35 * RPC slabs and memory pools
37 #define RPC_BUFFER_MAXSIZE (2048)
38 #define RPC_BUFFER_POOLSIZE (8)
39 #define RPC_TASK_POOLSIZE (8)
40 static struct kmem_cache *rpc_task_slabp __read_mostly;
41 static struct kmem_cache *rpc_buffer_slabp __read_mostly;
42 static mempool_t *rpc_task_mempool __read_mostly;
43 static mempool_t *rpc_buffer_mempool __read_mostly;
45 static void rpc_async_schedule(struct work_struct *);
46 static void rpc_release_task(struct rpc_task *task);
47 static void __rpc_queue_timer_fn(unsigned long ptr);
50 * RPC tasks sit here while waiting for conditions to improve.
52 static struct rpc_wait_queue delay_queue;
55 * rpciod-related stuff
57 struct workqueue_struct *rpciod_workqueue __read_mostly;
58 struct workqueue_struct *xprtiod_workqueue __read_mostly;
61 * Disable the timer for a given RPC task. Should be called with
62 * queue->lock and bh_disabled in order to avoid races within
66 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
68 if (task->tk_timeout == 0)
70 dprintk("RPC: %5u disabling timer\n", task->tk_pid);
72 list_del(&task->u.tk_wait.timer_list);
73 if (list_empty(&queue->timer_list.list))
74 del_timer(&queue->timer_list.timer);
78 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
80 queue->timer_list.expires = expires;
81 mod_timer(&queue->timer_list.timer, expires);
85 * Set up a timer for the current task.
88 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
90 if (!task->tk_timeout)
93 dprintk("RPC: %5u setting alarm for %u ms\n",
94 task->tk_pid, jiffies_to_msecs(task->tk_timeout));
96 task->u.tk_wait.expires = jiffies + task->tk_timeout;
97 if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
98 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
99 list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
102 static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
104 if (queue->priority != priority) {
105 queue->priority = priority;
106 queue->nr = 1U << priority;
110 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
112 rpc_set_waitqueue_priority(queue, queue->maxpriority);
116 * Add a request to a queue list
119 __rpc_list_enqueue_task(struct list_head *q, struct rpc_task *task)
123 list_for_each_entry(t, q, u.tk_wait.list) {
124 if (t->tk_owner == task->tk_owner) {
125 list_add_tail(&task->u.tk_wait.links,
126 &t->u.tk_wait.links);
127 /* Cache the queue head in task->u.tk_wait.list */
128 task->u.tk_wait.list.next = q;
129 task->u.tk_wait.list.prev = NULL;
133 INIT_LIST_HEAD(&task->u.tk_wait.links);
134 list_add_tail(&task->u.tk_wait.list, q);
138 * Remove request from a queue list
141 __rpc_list_dequeue_task(struct rpc_task *task)
146 if (task->u.tk_wait.list.prev == NULL) {
147 list_del(&task->u.tk_wait.links);
150 if (!list_empty(&task->u.tk_wait.links)) {
151 t = list_first_entry(&task->u.tk_wait.links,
154 /* Assume __rpc_list_enqueue_task() cached the queue head */
155 q = t->u.tk_wait.list.next;
156 list_add_tail(&t->u.tk_wait.list, q);
157 list_del(&task->u.tk_wait.links);
159 list_del(&task->u.tk_wait.list);
163 * Add new request to a priority queue.
165 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
166 struct rpc_task *task,
167 unsigned char queue_priority)
169 if (unlikely(queue_priority > queue->maxpriority))
170 queue_priority = queue->maxpriority;
171 __rpc_list_enqueue_task(&queue->tasks[queue_priority], task);
175 * Add new request to wait queue.
177 * Swapper tasks always get inserted at the head of the queue.
178 * This should avoid many nasty memory deadlocks and hopefully
179 * improve overall performance.
180 * Everyone else gets appended to the queue to ensure proper FIFO behavior.
182 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
183 struct rpc_task *task,
184 unsigned char queue_priority)
186 WARN_ON_ONCE(RPC_IS_QUEUED(task));
187 if (RPC_IS_QUEUED(task))
190 if (RPC_IS_PRIORITY(queue))
191 __rpc_add_wait_queue_priority(queue, task, queue_priority);
192 else if (RPC_IS_SWAPPER(task))
193 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
195 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
196 task->tk_waitqueue = queue;
198 /* barrier matches the read in rpc_wake_up_task_queue_locked() */
200 rpc_set_queued(task);
202 dprintk("RPC: %5u added to queue %p \"%s\"\n",
203 task->tk_pid, queue, rpc_qname(queue));
207 * Remove request from a priority queue.
209 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
211 __rpc_list_dequeue_task(task);
215 * Remove request from queue.
216 * Note: must be called with spin lock held.
218 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
220 __rpc_disable_timer(queue, task);
221 if (RPC_IS_PRIORITY(queue))
222 __rpc_remove_wait_queue_priority(task);
224 list_del(&task->u.tk_wait.list);
226 dprintk("RPC: %5u removed from queue %p \"%s\"\n",
227 task->tk_pid, queue, rpc_qname(queue));
230 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
234 spin_lock_init(&queue->lock);
235 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
236 INIT_LIST_HEAD(&queue->tasks[i]);
237 queue->maxpriority = nr_queues - 1;
238 rpc_reset_waitqueue_priority(queue);
240 setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
241 INIT_LIST_HEAD(&queue->timer_list.list);
242 rpc_assign_waitqueue_name(queue, qname);
245 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
247 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
249 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
251 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
253 __rpc_init_priority_wait_queue(queue, qname, 1);
255 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
257 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
259 del_timer_sync(&queue->timer_list.timer);
261 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
263 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
265 freezable_schedule_unsafe();
266 if (signal_pending_state(mode, current))
271 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
272 static void rpc_task_set_debuginfo(struct rpc_task *task)
274 static atomic_t rpc_pid;
276 task->tk_pid = atomic_inc_return(&rpc_pid);
279 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
284 static void rpc_set_active(struct rpc_task *task)
286 rpc_task_set_debuginfo(task);
287 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
288 trace_rpc_task_begin(task->tk_client, task, NULL);
292 * Mark an RPC call as having completed by clearing the 'active' bit
293 * and then waking up all tasks that were sleeping.
295 static int rpc_complete_task(struct rpc_task *task)
297 void *m = &task->tk_runstate;
298 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
299 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
303 trace_rpc_task_complete(task->tk_client, task, NULL);
305 spin_lock_irqsave(&wq->lock, flags);
306 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
307 ret = atomic_dec_and_test(&task->tk_count);
308 if (waitqueue_active(wq))
309 __wake_up_locked_key(wq, TASK_NORMAL, &k);
310 spin_unlock_irqrestore(&wq->lock, flags);
315 * Allow callers to wait for completion of an RPC call
317 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
318 * to enforce taking of the wq->lock and hence avoid races with
319 * rpc_complete_task().
321 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
324 action = rpc_wait_bit_killable;
325 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
326 action, TASK_KILLABLE);
328 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
331 * Make an RPC task runnable.
333 * Note: If the task is ASYNC, and is being made runnable after sitting on an
334 * rpc_wait_queue, this must be called with the queue spinlock held to protect
335 * the wait queue operation.
336 * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
337 * which is needed to ensure that __rpc_execute() doesn't loop (due to the
338 * lockless RPC_IS_QUEUED() test) before we've had a chance to test
339 * the RPC_TASK_RUNNING flag.
341 static void rpc_make_runnable(struct workqueue_struct *wq,
342 struct rpc_task *task)
344 bool need_wakeup = !rpc_test_and_set_running(task);
346 rpc_clear_queued(task);
349 if (RPC_IS_ASYNC(task)) {
350 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
351 queue_work(wq, &task->u.tk_work);
353 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
357 * Prepare for sleeping on a wait queue.
358 * By always appending tasks to the list we ensure FIFO behavior.
359 * NB: An RPC task will only receive interrupt-driven events as long
360 * as it's on a wait queue.
362 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
363 struct rpc_task *task,
365 unsigned char queue_priority)
367 dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
368 task->tk_pid, rpc_qname(q), jiffies);
370 trace_rpc_task_sleep(task->tk_client, task, q);
372 __rpc_add_wait_queue(q, task, queue_priority);
374 WARN_ON_ONCE(task->tk_callback != NULL);
375 task->tk_callback = action;
376 __rpc_add_timer(q, task);
379 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
382 /* We shouldn't ever put an inactive task to sleep */
383 WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
384 if (!RPC_IS_ACTIVATED(task)) {
385 task->tk_status = -EIO;
386 rpc_put_task_async(task);
391 * Protect the queue operations.
393 spin_lock_bh(&q->lock);
394 __rpc_sleep_on_priority(q, task, action, task->tk_priority);
395 spin_unlock_bh(&q->lock);
397 EXPORT_SYMBOL_GPL(rpc_sleep_on);
399 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
400 rpc_action action, int priority)
402 /* We shouldn't ever put an inactive task to sleep */
403 WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
404 if (!RPC_IS_ACTIVATED(task)) {
405 task->tk_status = -EIO;
406 rpc_put_task_async(task);
411 * Protect the queue operations.
413 spin_lock_bh(&q->lock);
414 __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
415 spin_unlock_bh(&q->lock);
417 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
420 * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
421 * @wq: workqueue on which to run task
423 * @task: task to be woken up
425 * Caller must hold queue->lock, and have cleared the task queued flag.
427 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
428 struct rpc_wait_queue *queue,
429 struct rpc_task *task)
431 dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
432 task->tk_pid, jiffies);
434 /* Has the task been executed yet? If not, we cannot wake it up! */
435 if (!RPC_IS_ACTIVATED(task)) {
436 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
440 trace_rpc_task_wakeup(task->tk_client, task, queue);
442 __rpc_remove_wait_queue(queue, task);
444 rpc_make_runnable(wq, task);
446 dprintk("RPC: __rpc_wake_up_task done\n");
450 * Wake up a queued task while the queue lock is being held
452 static void rpc_wake_up_task_on_wq_queue_locked(struct workqueue_struct *wq,
453 struct rpc_wait_queue *queue, struct rpc_task *task)
455 if (RPC_IS_QUEUED(task)) {
457 if (task->tk_waitqueue == queue)
458 __rpc_do_wake_up_task_on_wq(wq, queue, task);
463 * Wake up a queued task while the queue lock is being held
465 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
467 rpc_wake_up_task_on_wq_queue_locked(rpciod_workqueue, queue, task);
471 * Wake up a task on a specific queue
473 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
475 spin_lock_bh(&queue->lock);
476 rpc_wake_up_task_queue_locked(queue, task);
477 spin_unlock_bh(&queue->lock);
479 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
482 * Wake up the next task on a priority queue.
484 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
487 struct rpc_task *task;
490 * Service the privileged queue.
492 q = &queue->tasks[RPC_NR_PRIORITY - 1];
493 if (queue->maxpriority > RPC_PRIORITY_PRIVILEGED && !list_empty(q)) {
494 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
499 * Service a batch of tasks from a single owner.
501 q = &queue->tasks[queue->priority];
502 if (!list_empty(q) && queue->nr) {
504 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
509 * Service the next queue.
512 if (q == &queue->tasks[0])
513 q = &queue->tasks[queue->maxpriority];
516 if (!list_empty(q)) {
517 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
520 } while (q != &queue->tasks[queue->priority]);
522 rpc_reset_waitqueue_priority(queue);
526 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
531 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
533 if (RPC_IS_PRIORITY(queue))
534 return __rpc_find_next_queued_priority(queue);
535 if (!list_empty(&queue->tasks[0]))
536 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
541 * Wake up the first task on the wait queue.
543 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
544 struct rpc_wait_queue *queue,
545 bool (*func)(struct rpc_task *, void *), void *data)
547 struct rpc_task *task = NULL;
549 dprintk("RPC: wake_up_first(%p \"%s\")\n",
550 queue, rpc_qname(queue));
551 spin_lock_bh(&queue->lock);
552 task = __rpc_find_next_queued(queue);
554 if (func(task, data))
555 rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
559 spin_unlock_bh(&queue->lock);
565 * Wake up the first task on the wait queue.
567 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
568 bool (*func)(struct rpc_task *, void *), void *data)
570 return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
572 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
574 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
580 * Wake up the next task on the wait queue.
582 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
584 return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
586 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
589 * rpc_wake_up - wake up all rpc_tasks
590 * @queue: rpc_wait_queue on which the tasks are sleeping
594 void rpc_wake_up(struct rpc_wait_queue *queue)
596 struct list_head *head;
598 spin_lock_bh(&queue->lock);
599 head = &queue->tasks[queue->maxpriority];
601 while (!list_empty(head)) {
602 struct rpc_task *task;
603 task = list_first_entry(head,
606 rpc_wake_up_task_queue_locked(queue, task);
608 if (head == &queue->tasks[0])
612 spin_unlock_bh(&queue->lock);
614 EXPORT_SYMBOL_GPL(rpc_wake_up);
617 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
618 * @queue: rpc_wait_queue on which the tasks are sleeping
619 * @status: status value to set
623 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
625 struct list_head *head;
627 spin_lock_bh(&queue->lock);
628 head = &queue->tasks[queue->maxpriority];
630 while (!list_empty(head)) {
631 struct rpc_task *task;
632 task = list_first_entry(head,
635 task->tk_status = status;
636 rpc_wake_up_task_queue_locked(queue, task);
638 if (head == &queue->tasks[0])
642 spin_unlock_bh(&queue->lock);
644 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
646 static void __rpc_queue_timer_fn(unsigned long ptr)
648 struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
649 struct rpc_task *task, *n;
650 unsigned long expires, now, timeo;
652 spin_lock(&queue->lock);
653 expires = now = jiffies;
654 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
655 timeo = task->u.tk_wait.expires;
656 if (time_after_eq(now, timeo)) {
657 dprintk("RPC: %5u timeout\n", task->tk_pid);
658 task->tk_status = -ETIMEDOUT;
659 rpc_wake_up_task_queue_locked(queue, task);
662 if (expires == now || time_after(expires, timeo))
665 if (!list_empty(&queue->timer_list.list))
666 rpc_set_queue_timer(queue, expires);
667 spin_unlock(&queue->lock);
670 static void __rpc_atrun(struct rpc_task *task)
672 if (task->tk_status == -ETIMEDOUT)
677 * Run a task at a later time
679 void rpc_delay(struct rpc_task *task, unsigned long delay)
681 task->tk_timeout = delay;
682 rpc_sleep_on(&delay_queue, task, __rpc_atrun);
684 EXPORT_SYMBOL_GPL(rpc_delay);
687 * Helper to call task->tk_ops->rpc_call_prepare
689 void rpc_prepare_task(struct rpc_task *task)
691 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
695 rpc_init_task_statistics(struct rpc_task *task)
697 /* Initialize retry counters */
698 task->tk_garb_retry = 2;
699 task->tk_cred_retry = 2;
700 task->tk_rebind_retry = 2;
702 /* starting timestamp */
703 task->tk_start = ktime_get();
707 rpc_reset_task_statistics(struct rpc_task *task)
709 task->tk_timeouts = 0;
710 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
712 rpc_init_task_statistics(task);
716 * Helper that calls task->tk_ops->rpc_call_done if it exists
718 void rpc_exit_task(struct rpc_task *task)
720 task->tk_action = NULL;
721 if (task->tk_ops->rpc_call_done != NULL) {
722 task->tk_ops->rpc_call_done(task, task->tk_calldata);
723 if (task->tk_action != NULL) {
724 WARN_ON(RPC_ASSASSINATED(task));
725 /* Always release the RPC slot and buffer memory */
727 rpc_reset_task_statistics(task);
732 void rpc_exit(struct rpc_task *task, int status)
734 task->tk_status = status;
735 task->tk_action = rpc_exit_task;
736 if (RPC_IS_QUEUED(task))
737 rpc_wake_up_queued_task(task->tk_waitqueue, task);
739 EXPORT_SYMBOL_GPL(rpc_exit);
741 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
743 if (ops->rpc_release != NULL)
744 ops->rpc_release(calldata);
748 * This is the RPC `scheduler' (or rather, the finite state machine).
750 static void __rpc_execute(struct rpc_task *task)
752 struct rpc_wait_queue *queue;
753 int task_is_async = RPC_IS_ASYNC(task);
756 dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
757 task->tk_pid, task->tk_flags);
759 WARN_ON_ONCE(RPC_IS_QUEUED(task));
760 if (RPC_IS_QUEUED(task))
764 void (*do_action)(struct rpc_task *);
767 * Execute any pending callback first.
769 do_action = task->tk_callback;
770 task->tk_callback = NULL;
771 if (do_action == NULL) {
773 * Perform the next FSM step.
774 * tk_action may be NULL if the task has been killed.
775 * In particular, note that rpc_killall_tasks may
776 * do this at any time, so beware when dereferencing.
778 do_action = task->tk_action;
779 if (do_action == NULL)
782 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
786 * Lockless check for whether task is sleeping or not.
788 if (!RPC_IS_QUEUED(task))
791 * The queue->lock protects against races with
792 * rpc_make_runnable().
794 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
795 * rpc_task, rpc_make_runnable() can assign it to a
796 * different workqueue. We therefore cannot assume that the
797 * rpc_task pointer may still be dereferenced.
799 queue = task->tk_waitqueue;
800 spin_lock_bh(&queue->lock);
801 if (!RPC_IS_QUEUED(task)) {
802 spin_unlock_bh(&queue->lock);
805 rpc_clear_running(task);
806 spin_unlock_bh(&queue->lock);
810 /* sync task: sleep here */
811 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
812 status = out_of_line_wait_on_bit(&task->tk_runstate,
813 RPC_TASK_QUEUED, rpc_wait_bit_killable,
815 if (status == -ERESTARTSYS) {
817 * When a sync task receives a signal, it exits with
818 * -ERESTARTSYS. In order to catch any callbacks that
819 * clean up after sleeping on some queue, we don't
820 * break the loop here, but go around once more.
822 dprintk("RPC: %5u got signal\n", task->tk_pid);
823 task->tk_flags |= RPC_TASK_KILLED;
824 rpc_exit(task, -ERESTARTSYS);
826 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
829 dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
831 /* Release all resources associated with the task */
832 rpc_release_task(task);
836 * User-visible entry point to the scheduler.
838 * This may be called recursively if e.g. an async NFS task updates
839 * the attributes and finds that dirty pages must be flushed.
840 * NOTE: Upon exit of this function the task is guaranteed to be
841 * released. In particular note that tk_release() will have
842 * been called, so your task memory may have been freed.
844 void rpc_execute(struct rpc_task *task)
846 bool is_async = RPC_IS_ASYNC(task);
848 rpc_set_active(task);
849 rpc_make_runnable(rpciod_workqueue, task);
854 static void rpc_async_schedule(struct work_struct *work)
856 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
860 * rpc_malloc - allocate RPC buffer resources
863 * A single memory region is allocated, which is split between the
864 * RPC call and RPC reply that this task is being used for. When
865 * this RPC is retired, the memory is released by calling rpc_free.
867 * To prevent rpciod from hanging, this allocator never sleeps,
868 * returning -ENOMEM and suppressing warning if the request cannot
869 * be serviced immediately. The caller can arrange to sleep in a
870 * way that is safe for rpciod.
872 * Most requests are 'small' (under 2KiB) and can be serviced from a
873 * mempool, ensuring that NFS reads and writes can always proceed,
874 * and that there is good locality of reference for these buffers.
876 * In order to avoid memory starvation triggering more writebacks of
877 * NFS requests, we avoid using GFP_KERNEL.
879 int rpc_malloc(struct rpc_task *task)
881 struct rpc_rqst *rqst = task->tk_rqstp;
882 size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
883 struct rpc_buffer *buf;
884 gfp_t gfp = GFP_NOIO | __GFP_NOWARN;
886 if (RPC_IS_ASYNC(task))
887 gfp = GFP_NOWAIT | __GFP_NOWARN;
888 if (RPC_IS_SWAPPER(task))
889 gfp |= __GFP_MEMALLOC;
891 size += sizeof(struct rpc_buffer);
892 if (size <= RPC_BUFFER_MAXSIZE)
893 buf = mempool_alloc(rpc_buffer_mempool, gfp);
895 buf = kmalloc(size, gfp);
901 dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
902 task->tk_pid, size, buf);
903 rqst->rq_buffer = buf->data;
904 rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
907 EXPORT_SYMBOL_GPL(rpc_malloc);
910 * rpc_free - free RPC buffer resources allocated via rpc_malloc
914 void rpc_free(struct rpc_task *task)
916 void *buffer = task->tk_rqstp->rq_buffer;
918 struct rpc_buffer *buf;
920 buf = container_of(buffer, struct rpc_buffer, data);
923 dprintk("RPC: freeing buffer of size %zu at %p\n",
926 if (size <= RPC_BUFFER_MAXSIZE)
927 mempool_free(buf, rpc_buffer_mempool);
931 EXPORT_SYMBOL_GPL(rpc_free);
934 * Creation and deletion of RPC task structures
936 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
938 memset(task, 0, sizeof(*task));
939 atomic_set(&task->tk_count, 1);
940 task->tk_flags = task_setup_data->flags;
941 task->tk_ops = task_setup_data->callback_ops;
942 task->tk_calldata = task_setup_data->callback_data;
943 INIT_LIST_HEAD(&task->tk_task);
945 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
946 task->tk_owner = current->tgid;
948 /* Initialize workqueue for async tasks */
949 task->tk_workqueue = task_setup_data->workqueue;
951 task->tk_xprt = xprt_get(task_setup_data->rpc_xprt);
953 if (task->tk_ops->rpc_call_prepare != NULL)
954 task->tk_action = rpc_prepare_task;
956 rpc_init_task_statistics(task);
958 dprintk("RPC: new task initialized, procpid %u\n",
959 task_pid_nr(current));
962 static struct rpc_task *
965 return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
969 * Create a new task for the specified client.
971 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
973 struct rpc_task *task = setup_data->task;
974 unsigned short flags = 0;
977 task = rpc_alloc_task();
979 rpc_release_calldata(setup_data->callback_ops,
980 setup_data->callback_data);
981 return ERR_PTR(-ENOMEM);
983 flags = RPC_TASK_DYNAMIC;
986 rpc_init_task(task, setup_data);
987 task->tk_flags |= flags;
988 dprintk("RPC: allocated task %p\n", task);
993 * rpc_free_task - release rpc task and perform cleanups
995 * Note that we free up the rpc_task _after_ rpc_release_calldata()
996 * in order to work around a workqueue dependency issue.
999 * "Workqueue currently considers two work items to be the same if they're
1000 * on the same address and won't execute them concurrently - ie. it
1001 * makes a work item which is queued again while being executed wait
1002 * for the previous execution to complete.
1004 * If a work function frees the work item, and then waits for an event
1005 * which should be performed by another work item and *that* work item
1006 * recycles the freed work item, it can create a false dependency loop.
1007 * There really is no reliable way to detect this short of verifying
1008 * every memory free."
1011 static void rpc_free_task(struct rpc_task *task)
1013 unsigned short tk_flags = task->tk_flags;
1015 rpc_release_calldata(task->tk_ops, task->tk_calldata);
1017 if (tk_flags & RPC_TASK_DYNAMIC) {
1018 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1019 mempool_free(task, rpc_task_mempool);
1023 static void rpc_async_release(struct work_struct *work)
1025 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1028 static void rpc_release_resources_task(struct rpc_task *task)
1031 if (task->tk_msg.rpc_cred) {
1032 put_rpccred(task->tk_msg.rpc_cred);
1033 task->tk_msg.rpc_cred = NULL;
1035 rpc_task_release_client(task);
1038 static void rpc_final_put_task(struct rpc_task *task,
1039 struct workqueue_struct *q)
1042 INIT_WORK(&task->u.tk_work, rpc_async_release);
1043 queue_work(q, &task->u.tk_work);
1045 rpc_free_task(task);
1048 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1050 if (atomic_dec_and_test(&task->tk_count)) {
1051 rpc_release_resources_task(task);
1052 rpc_final_put_task(task, q);
1056 void rpc_put_task(struct rpc_task *task)
1058 rpc_do_put_task(task, NULL);
1060 EXPORT_SYMBOL_GPL(rpc_put_task);
1062 void rpc_put_task_async(struct rpc_task *task)
1064 rpc_do_put_task(task, task->tk_workqueue);
1066 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1068 static void rpc_release_task(struct rpc_task *task)
1070 dprintk("RPC: %5u release task\n", task->tk_pid);
1072 WARN_ON_ONCE(RPC_IS_QUEUED(task));
1074 rpc_release_resources_task(task);
1077 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1078 * so it should be safe to use task->tk_count as a test for whether
1079 * or not any other processes still hold references to our rpc_task.
1081 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1082 /* Wake up anyone who may be waiting for task completion */
1083 if (!rpc_complete_task(task))
1086 if (!atomic_dec_and_test(&task->tk_count))
1089 rpc_final_put_task(task, task->tk_workqueue);
1094 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1097 void rpciod_down(void)
1099 module_put(THIS_MODULE);
1103 * Start up the rpciod workqueue.
1105 static int rpciod_start(void)
1107 struct workqueue_struct *wq;
1110 * Create the rpciod thread and wait for it to start.
1112 dprintk("RPC: creating workqueue rpciod\n");
1113 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1116 rpciod_workqueue = wq;
1117 /* Note: highpri because network receive is latency sensitive */
1118 wq = alloc_workqueue("xprtiod", WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1121 xprtiod_workqueue = wq;
1124 wq = rpciod_workqueue;
1125 rpciod_workqueue = NULL;
1126 destroy_workqueue(wq);
1131 static void rpciod_stop(void)
1133 struct workqueue_struct *wq = NULL;
1135 if (rpciod_workqueue == NULL)
1137 dprintk("RPC: destroying workqueue rpciod\n");
1139 wq = rpciod_workqueue;
1140 rpciod_workqueue = NULL;
1141 destroy_workqueue(wq);
1142 wq = xprtiod_workqueue;
1143 xprtiod_workqueue = NULL;
1144 destroy_workqueue(wq);
1148 rpc_destroy_mempool(void)
1151 mempool_destroy(rpc_buffer_mempool);
1152 mempool_destroy(rpc_task_mempool);
1153 kmem_cache_destroy(rpc_task_slabp);
1154 kmem_cache_destroy(rpc_buffer_slabp);
1155 rpc_destroy_wait_queue(&delay_queue);
1159 rpc_init_mempool(void)
1162 * The following is not strictly a mempool initialisation,
1163 * but there is no harm in doing it here
1165 rpc_init_wait_queue(&delay_queue, "delayq");
1166 if (!rpciod_start())
1169 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1170 sizeof(struct rpc_task),
1171 0, SLAB_HWCACHE_ALIGN,
1173 if (!rpc_task_slabp)
1175 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1177 0, SLAB_HWCACHE_ALIGN,
1179 if (!rpc_buffer_slabp)
1181 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1183 if (!rpc_task_mempool)
1185 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1187 if (!rpc_buffer_mempool)
1191 rpc_destroy_mempool();