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>
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;
60 * Disable the timer for a given RPC task. Should be called with
61 * queue->lock and bh_disabled in order to avoid races within
65 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
67 if (task->tk_timeout == 0)
69 dprintk("RPC: %5u disabling timer\n", task->tk_pid);
71 list_del(&task->u.tk_wait.timer_list);
72 if (list_empty(&queue->timer_list.list))
73 del_timer(&queue->timer_list.timer);
77 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
79 queue->timer_list.expires = expires;
80 mod_timer(&queue->timer_list.timer, expires);
84 * Set up a timer for the current task.
87 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
89 if (!task->tk_timeout)
92 dprintk("RPC: %5u setting alarm for %lu ms\n",
93 task->tk_pid, task->tk_timeout * 1000 / HZ);
95 task->u.tk_wait.expires = jiffies + task->tk_timeout;
96 if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
97 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
98 list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
102 * Add new request to a priority queue.
104 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
105 struct rpc_task *task,
106 unsigned char queue_priority)
111 INIT_LIST_HEAD(&task->u.tk_wait.links);
112 q = &queue->tasks[queue_priority];
113 if (unlikely(queue_priority > queue->maxpriority))
114 q = &queue->tasks[queue->maxpriority];
115 list_for_each_entry(t, q, u.tk_wait.list) {
116 if (t->tk_owner == task->tk_owner) {
117 list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
121 list_add_tail(&task->u.tk_wait.list, q);
125 * Add new request to wait queue.
127 * Swapper tasks always get inserted at the head of the queue.
128 * This should avoid many nasty memory deadlocks and hopefully
129 * improve overall performance.
130 * Everyone else gets appended to the queue to ensure proper FIFO behavior.
132 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
133 struct rpc_task *task,
134 unsigned char queue_priority)
136 WARN_ON_ONCE(RPC_IS_QUEUED(task));
137 if (RPC_IS_QUEUED(task))
140 if (RPC_IS_PRIORITY(queue))
141 __rpc_add_wait_queue_priority(queue, task, queue_priority);
142 else if (RPC_IS_SWAPPER(task))
143 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
145 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
146 task->tk_waitqueue = queue;
148 rpc_set_queued(task);
150 dprintk("RPC: %5u added to queue %p \"%s\"\n",
151 task->tk_pid, queue, rpc_qname(queue));
155 * Remove request from a priority queue.
157 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
161 if (!list_empty(&task->u.tk_wait.links)) {
162 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
163 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
164 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
169 * Remove request from queue.
170 * Note: must be called with spin lock held.
172 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
174 __rpc_disable_timer(queue, task);
175 if (RPC_IS_PRIORITY(queue))
176 __rpc_remove_wait_queue_priority(task);
177 list_del(&task->u.tk_wait.list);
179 dprintk("RPC: %5u removed from queue %p \"%s\"\n",
180 task->tk_pid, queue, rpc_qname(queue));
183 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
185 queue->priority = priority;
186 queue->count = 1 << (priority * 2);
189 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
192 queue->nr = RPC_BATCH_COUNT;
195 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
197 rpc_set_waitqueue_priority(queue, queue->maxpriority);
198 rpc_set_waitqueue_owner(queue, 0);
201 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
205 spin_lock_init(&queue->lock);
206 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
207 INIT_LIST_HEAD(&queue->tasks[i]);
208 queue->maxpriority = nr_queues - 1;
209 rpc_reset_waitqueue_priority(queue);
211 setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
212 INIT_LIST_HEAD(&queue->timer_list.list);
213 rpc_assign_waitqueue_name(queue, qname);
216 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
218 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
220 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
222 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
224 __rpc_init_priority_wait_queue(queue, qname, 1);
226 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
228 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
230 del_timer_sync(&queue->timer_list.timer);
232 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
234 static int rpc_wait_bit_killable(void *word)
236 if (fatal_signal_pending(current))
238 freezable_schedule();
243 static void rpc_task_set_debuginfo(struct rpc_task *task)
245 static atomic_t rpc_pid;
247 task->tk_pid = atomic_inc_return(&rpc_pid);
250 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
255 static void rpc_set_active(struct rpc_task *task)
257 trace_rpc_task_begin(task->tk_client, task, NULL);
259 rpc_task_set_debuginfo(task);
260 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
264 * Mark an RPC call as having completed by clearing the 'active' bit
265 * and then waking up all tasks that were sleeping.
267 static int rpc_complete_task(struct rpc_task *task)
269 void *m = &task->tk_runstate;
270 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
271 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
275 trace_rpc_task_complete(task->tk_client, task, NULL);
277 spin_lock_irqsave(&wq->lock, flags);
278 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
279 ret = atomic_dec_and_test(&task->tk_count);
280 if (waitqueue_active(wq))
281 __wake_up_locked_key(wq, TASK_NORMAL, &k);
282 spin_unlock_irqrestore(&wq->lock, flags);
287 * Allow callers to wait for completion of an RPC call
289 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
290 * to enforce taking of the wq->lock and hence avoid races with
291 * rpc_complete_task().
293 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
296 action = rpc_wait_bit_killable;
297 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
298 action, TASK_KILLABLE);
300 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
303 * Make an RPC task runnable.
305 * Note: If the task is ASYNC, and is being made runnable after sitting on an
306 * rpc_wait_queue, this must be called with the queue spinlock held to protect
307 * the wait queue operation.
309 static void rpc_make_runnable(struct rpc_task *task)
311 rpc_clear_queued(task);
312 if (rpc_test_and_set_running(task))
314 if (RPC_IS_ASYNC(task)) {
315 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
316 queue_work(rpciod_workqueue, &task->u.tk_work);
318 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
322 * Prepare for sleeping on a wait queue.
323 * By always appending tasks to the list we ensure FIFO behavior.
324 * NB: An RPC task will only receive interrupt-driven events as long
325 * as it's on a wait queue.
327 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
328 struct rpc_task *task,
330 unsigned char queue_priority)
332 dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
333 task->tk_pid, rpc_qname(q), jiffies);
335 trace_rpc_task_sleep(task->tk_client, task, q);
337 __rpc_add_wait_queue(q, task, queue_priority);
339 WARN_ON_ONCE(task->tk_callback != NULL);
340 task->tk_callback = action;
341 __rpc_add_timer(q, task);
344 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
347 /* We shouldn't ever put an inactive task to sleep */
348 WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
349 if (!RPC_IS_ACTIVATED(task)) {
350 task->tk_status = -EIO;
351 rpc_put_task_async(task);
356 * Protect the queue operations.
358 spin_lock_bh(&q->lock);
359 __rpc_sleep_on_priority(q, task, action, task->tk_priority);
360 spin_unlock_bh(&q->lock);
362 EXPORT_SYMBOL_GPL(rpc_sleep_on);
364 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
365 rpc_action action, int priority)
367 /* We shouldn't ever put an inactive task to sleep */
368 WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
369 if (!RPC_IS_ACTIVATED(task)) {
370 task->tk_status = -EIO;
371 rpc_put_task_async(task);
376 * Protect the queue operations.
378 spin_lock_bh(&q->lock);
379 __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
380 spin_unlock_bh(&q->lock);
384 * __rpc_do_wake_up_task - wake up a single rpc_task
386 * @task: task to be woken up
388 * Caller must hold queue->lock, and have cleared the task queued flag.
390 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
392 dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
393 task->tk_pid, jiffies);
395 /* Has the task been executed yet? If not, we cannot wake it up! */
396 if (!RPC_IS_ACTIVATED(task)) {
397 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
401 trace_rpc_task_wakeup(task->tk_client, task, queue);
403 __rpc_remove_wait_queue(queue, task);
405 rpc_make_runnable(task);
407 dprintk("RPC: __rpc_wake_up_task done\n");
411 * Wake up a queued task while the queue lock is being held
413 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
415 if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
416 __rpc_do_wake_up_task(queue, task);
420 * Tests whether rpc queue is empty
422 int rpc_queue_empty(struct rpc_wait_queue *queue)
426 spin_lock_bh(&queue->lock);
428 spin_unlock_bh(&queue->lock);
431 EXPORT_SYMBOL_GPL(rpc_queue_empty);
434 * Wake up a task on a specific queue
436 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
438 spin_lock_bh(&queue->lock);
439 rpc_wake_up_task_queue_locked(queue, task);
440 spin_unlock_bh(&queue->lock);
442 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
445 * Wake up the next task on a priority queue.
447 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
450 struct rpc_task *task;
453 * Service a batch of tasks from a single owner.
455 q = &queue->tasks[queue->priority];
456 if (!list_empty(q)) {
457 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
458 if (queue->owner == task->tk_owner) {
461 list_move_tail(&task->u.tk_wait.list, q);
464 * Check if we need to switch queues.
471 * Service the next queue.
474 if (q == &queue->tasks[0])
475 q = &queue->tasks[queue->maxpriority];
478 if (!list_empty(q)) {
479 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
482 } while (q != &queue->tasks[queue->priority]);
484 rpc_reset_waitqueue_priority(queue);
488 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
490 rpc_set_waitqueue_owner(queue, task->tk_owner);
495 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
497 if (RPC_IS_PRIORITY(queue))
498 return __rpc_find_next_queued_priority(queue);
499 if (!list_empty(&queue->tasks[0]))
500 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
505 * Wake up the first task on the wait queue.
507 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
508 bool (*func)(struct rpc_task *, void *), void *data)
510 struct rpc_task *task = NULL;
512 dprintk("RPC: wake_up_first(%p \"%s\")\n",
513 queue, rpc_qname(queue));
514 spin_lock_bh(&queue->lock);
515 task = __rpc_find_next_queued(queue);
517 if (func(task, data))
518 rpc_wake_up_task_queue_locked(queue, task);
522 spin_unlock_bh(&queue->lock);
526 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
528 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
534 * Wake up the next task on the wait queue.
536 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
538 return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
540 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
543 * rpc_wake_up - wake up all rpc_tasks
544 * @queue: rpc_wait_queue on which the tasks are sleeping
548 void rpc_wake_up(struct rpc_wait_queue *queue)
550 struct list_head *head;
552 spin_lock_bh(&queue->lock);
553 head = &queue->tasks[queue->maxpriority];
555 while (!list_empty(head)) {
556 struct rpc_task *task;
557 task = list_first_entry(head,
560 rpc_wake_up_task_queue_locked(queue, task);
562 if (head == &queue->tasks[0])
566 spin_unlock_bh(&queue->lock);
568 EXPORT_SYMBOL_GPL(rpc_wake_up);
571 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
572 * @queue: rpc_wait_queue on which the tasks are sleeping
573 * @status: status value to set
577 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
579 struct list_head *head;
581 spin_lock_bh(&queue->lock);
582 head = &queue->tasks[queue->maxpriority];
584 while (!list_empty(head)) {
585 struct rpc_task *task;
586 task = list_first_entry(head,
589 task->tk_status = status;
590 rpc_wake_up_task_queue_locked(queue, task);
592 if (head == &queue->tasks[0])
596 spin_unlock_bh(&queue->lock);
598 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
600 static void __rpc_queue_timer_fn(unsigned long ptr)
602 struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
603 struct rpc_task *task, *n;
604 unsigned long expires, now, timeo;
606 spin_lock(&queue->lock);
607 expires = now = jiffies;
608 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
609 timeo = task->u.tk_wait.expires;
610 if (time_after_eq(now, timeo)) {
611 dprintk("RPC: %5u timeout\n", task->tk_pid);
612 task->tk_status = -ETIMEDOUT;
613 rpc_wake_up_task_queue_locked(queue, task);
616 if (expires == now || time_after(expires, timeo))
619 if (!list_empty(&queue->timer_list.list))
620 rpc_set_queue_timer(queue, expires);
621 spin_unlock(&queue->lock);
624 static void __rpc_atrun(struct rpc_task *task)
630 * Run a task at a later time
632 void rpc_delay(struct rpc_task *task, unsigned long delay)
634 task->tk_timeout = delay;
635 rpc_sleep_on(&delay_queue, task, __rpc_atrun);
637 EXPORT_SYMBOL_GPL(rpc_delay);
640 * Helper to call task->tk_ops->rpc_call_prepare
642 void rpc_prepare_task(struct rpc_task *task)
644 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
648 rpc_init_task_statistics(struct rpc_task *task)
650 /* Initialize retry counters */
651 task->tk_garb_retry = 2;
652 task->tk_cred_retry = 2;
653 task->tk_rebind_retry = 2;
655 /* starting timestamp */
656 task->tk_start = ktime_get();
660 rpc_reset_task_statistics(struct rpc_task *task)
662 task->tk_timeouts = 0;
663 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
665 rpc_init_task_statistics(task);
669 * Helper that calls task->tk_ops->rpc_call_done if it exists
671 void rpc_exit_task(struct rpc_task *task)
673 task->tk_action = NULL;
674 if (task->tk_ops->rpc_call_done != NULL) {
675 task->tk_ops->rpc_call_done(task, task->tk_calldata);
676 if (task->tk_action != NULL) {
677 WARN_ON(RPC_ASSASSINATED(task));
678 /* Always release the RPC slot and buffer memory */
680 rpc_reset_task_statistics(task);
685 void rpc_exit(struct rpc_task *task, int status)
687 task->tk_status = status;
688 task->tk_action = rpc_exit_task;
689 if (RPC_IS_QUEUED(task))
690 rpc_wake_up_queued_task(task->tk_waitqueue, task);
692 EXPORT_SYMBOL_GPL(rpc_exit);
694 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
696 if (ops->rpc_release != NULL)
697 ops->rpc_release(calldata);
701 * This is the RPC `scheduler' (or rather, the finite state machine).
703 static void __rpc_execute(struct rpc_task *task)
705 struct rpc_wait_queue *queue;
706 int task_is_async = RPC_IS_ASYNC(task);
709 dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
710 task->tk_pid, task->tk_flags);
712 WARN_ON_ONCE(RPC_IS_QUEUED(task));
713 if (RPC_IS_QUEUED(task))
717 void (*do_action)(struct rpc_task *);
720 * Execute any pending callback first.
722 do_action = task->tk_callback;
723 task->tk_callback = NULL;
724 if (do_action == NULL) {
726 * Perform the next FSM step.
727 * tk_action may be NULL if the task has been killed.
728 * In particular, note that rpc_killall_tasks may
729 * do this at any time, so beware when dereferencing.
731 do_action = task->tk_action;
732 if (do_action == NULL)
735 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
739 * Lockless check for whether task is sleeping or not.
741 if (!RPC_IS_QUEUED(task))
744 * The queue->lock protects against races with
745 * rpc_make_runnable().
747 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
748 * rpc_task, rpc_make_runnable() can assign it to a
749 * different workqueue. We therefore cannot assume that the
750 * rpc_task pointer may still be dereferenced.
752 queue = task->tk_waitqueue;
753 spin_lock_bh(&queue->lock);
754 if (!RPC_IS_QUEUED(task)) {
755 spin_unlock_bh(&queue->lock);
758 rpc_clear_running(task);
759 spin_unlock_bh(&queue->lock);
763 /* sync task: sleep here */
764 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
765 status = out_of_line_wait_on_bit(&task->tk_runstate,
766 RPC_TASK_QUEUED, rpc_wait_bit_killable,
768 if (status == -ERESTARTSYS) {
770 * When a sync task receives a signal, it exits with
771 * -ERESTARTSYS. In order to catch any callbacks that
772 * clean up after sleeping on some queue, we don't
773 * break the loop here, but go around once more.
775 dprintk("RPC: %5u got signal\n", task->tk_pid);
776 task->tk_flags |= RPC_TASK_KILLED;
777 rpc_exit(task, -ERESTARTSYS);
779 rpc_set_running(task);
780 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
783 dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
785 /* Release all resources associated with the task */
786 rpc_release_task(task);
790 * User-visible entry point to the scheduler.
792 * This may be called recursively if e.g. an async NFS task updates
793 * the attributes and finds that dirty pages must be flushed.
794 * NOTE: Upon exit of this function the task is guaranteed to be
795 * released. In particular note that tk_release() will have
796 * been called, so your task memory may have been freed.
798 void rpc_execute(struct rpc_task *task)
800 rpc_set_active(task);
801 rpc_make_runnable(task);
802 if (!RPC_IS_ASYNC(task))
806 static void rpc_async_schedule(struct work_struct *work)
808 current->flags |= PF_FSTRANS;
809 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
810 current->flags &= ~PF_FSTRANS;
814 * rpc_malloc - allocate an RPC buffer
815 * @task: RPC task that will use this buffer
816 * @size: requested byte size
818 * To prevent rpciod from hanging, this allocator never sleeps,
819 * returning NULL if the request cannot be serviced immediately.
820 * The caller can arrange to sleep in a way that is safe for rpciod.
822 * Most requests are 'small' (under 2KiB) and can be serviced from a
823 * mempool, ensuring that NFS reads and writes can always proceed,
824 * and that there is good locality of reference for these buffers.
826 * In order to avoid memory starvation triggering more writebacks of
827 * NFS requests, we avoid using GFP_KERNEL.
829 void *rpc_malloc(struct rpc_task *task, size_t size)
831 struct rpc_buffer *buf;
832 gfp_t gfp = GFP_NOWAIT;
834 if (RPC_IS_SWAPPER(task))
835 gfp |= __GFP_MEMALLOC;
837 size += sizeof(struct rpc_buffer);
838 if (size <= RPC_BUFFER_MAXSIZE)
839 buf = mempool_alloc(rpc_buffer_mempool, gfp);
841 buf = kmalloc(size, gfp);
847 dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
848 task->tk_pid, size, buf);
851 EXPORT_SYMBOL_GPL(rpc_malloc);
854 * rpc_free - free buffer allocated via rpc_malloc
855 * @buffer: buffer to free
858 void rpc_free(void *buffer)
861 struct rpc_buffer *buf;
866 buf = container_of(buffer, struct rpc_buffer, data);
869 dprintk("RPC: freeing buffer of size %zu at %p\n",
872 if (size <= RPC_BUFFER_MAXSIZE)
873 mempool_free(buf, rpc_buffer_mempool);
877 EXPORT_SYMBOL_GPL(rpc_free);
880 * Creation and deletion of RPC task structures
882 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
884 memset(task, 0, sizeof(*task));
885 atomic_set(&task->tk_count, 1);
886 task->tk_flags = task_setup_data->flags;
887 task->tk_ops = task_setup_data->callback_ops;
888 task->tk_calldata = task_setup_data->callback_data;
889 INIT_LIST_HEAD(&task->tk_task);
891 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
892 task->tk_owner = current->tgid;
894 /* Initialize workqueue for async tasks */
895 task->tk_workqueue = task_setup_data->workqueue;
897 if (task->tk_ops->rpc_call_prepare != NULL)
898 task->tk_action = rpc_prepare_task;
900 rpc_init_task_statistics(task);
902 dprintk("RPC: new task initialized, procpid %u\n",
903 task_pid_nr(current));
906 static struct rpc_task *
909 return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
913 * Create a new task for the specified client.
915 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
917 struct rpc_task *task = setup_data->task;
918 unsigned short flags = 0;
921 task = rpc_alloc_task();
923 rpc_release_calldata(setup_data->callback_ops,
924 setup_data->callback_data);
925 return ERR_PTR(-ENOMEM);
927 flags = RPC_TASK_DYNAMIC;
930 rpc_init_task(task, setup_data);
931 task->tk_flags |= flags;
932 dprintk("RPC: allocated task %p\n", task);
936 static void rpc_free_task(struct rpc_task *task)
938 const struct rpc_call_ops *tk_ops = task->tk_ops;
939 void *calldata = task->tk_calldata;
941 if (task->tk_flags & RPC_TASK_DYNAMIC) {
942 dprintk("RPC: %5u freeing task\n", task->tk_pid);
943 mempool_free(task, rpc_task_mempool);
945 rpc_release_calldata(tk_ops, calldata);
948 static void rpc_async_release(struct work_struct *work)
950 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
953 static void rpc_release_resources_task(struct rpc_task *task)
957 if (task->tk_msg.rpc_cred) {
958 put_rpccred(task->tk_msg.rpc_cred);
959 task->tk_msg.rpc_cred = NULL;
961 rpc_task_release_client(task);
964 static void rpc_final_put_task(struct rpc_task *task,
965 struct workqueue_struct *q)
968 INIT_WORK(&task->u.tk_work, rpc_async_release);
969 queue_work(q, &task->u.tk_work);
974 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
976 if (atomic_dec_and_test(&task->tk_count)) {
977 rpc_release_resources_task(task);
978 rpc_final_put_task(task, q);
982 void rpc_put_task(struct rpc_task *task)
984 rpc_do_put_task(task, NULL);
986 EXPORT_SYMBOL_GPL(rpc_put_task);
988 void rpc_put_task_async(struct rpc_task *task)
990 rpc_do_put_task(task, task->tk_workqueue);
992 EXPORT_SYMBOL_GPL(rpc_put_task_async);
994 static void rpc_release_task(struct rpc_task *task)
996 dprintk("RPC: %5u release task\n", task->tk_pid);
998 BUG_ON (RPC_IS_QUEUED(task));
1000 rpc_release_resources_task(task);
1003 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1004 * so it should be safe to use task->tk_count as a test for whether
1005 * or not any other processes still hold references to our rpc_task.
1007 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1008 /* Wake up anyone who may be waiting for task completion */
1009 if (!rpc_complete_task(task))
1012 if (!atomic_dec_and_test(&task->tk_count))
1015 rpc_final_put_task(task, task->tk_workqueue);
1020 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1023 void rpciod_down(void)
1025 module_put(THIS_MODULE);
1029 * Start up the rpciod workqueue.
1031 static int rpciod_start(void)
1033 struct workqueue_struct *wq;
1036 * Create the rpciod thread and wait for it to start.
1038 dprintk("RPC: creating workqueue rpciod\n");
1039 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 1);
1040 rpciod_workqueue = wq;
1041 return rpciod_workqueue != NULL;
1044 static void rpciod_stop(void)
1046 struct workqueue_struct *wq = NULL;
1048 if (rpciod_workqueue == NULL)
1050 dprintk("RPC: destroying workqueue rpciod\n");
1052 wq = rpciod_workqueue;
1053 rpciod_workqueue = NULL;
1054 destroy_workqueue(wq);
1058 rpc_destroy_mempool(void)
1061 if (rpc_buffer_mempool)
1062 mempool_destroy(rpc_buffer_mempool);
1063 if (rpc_task_mempool)
1064 mempool_destroy(rpc_task_mempool);
1066 kmem_cache_destroy(rpc_task_slabp);
1067 if (rpc_buffer_slabp)
1068 kmem_cache_destroy(rpc_buffer_slabp);
1069 rpc_destroy_wait_queue(&delay_queue);
1073 rpc_init_mempool(void)
1076 * The following is not strictly a mempool initialisation,
1077 * but there is no harm in doing it here
1079 rpc_init_wait_queue(&delay_queue, "delayq");
1080 if (!rpciod_start())
1083 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1084 sizeof(struct rpc_task),
1085 0, SLAB_HWCACHE_ALIGN,
1087 if (!rpc_task_slabp)
1089 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1091 0, SLAB_HWCACHE_ALIGN,
1093 if (!rpc_buffer_slabp)
1095 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1097 if (!rpc_task_mempool)
1099 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1101 if (!rpc_buffer_mempool)
1105 rpc_destroy_mempool();