Loading...
1/*
2 * linux/net/sunrpc/sched.c
3 *
4 * Scheduling for synchronous and asynchronous RPC requests.
5 *
6 * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
7 *
8 * TCP NFS related read + write fixes
9 * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
10 */
11
12#include <linux/module.h>
13
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
22#include <linux/sunrpc/clnt.h>
23
24#include "sunrpc.h"
25
26#ifdef RPC_DEBUG
27#define RPCDBG_FACILITY RPCDBG_SCHED
28#endif
29
30/*
31 * RPC slabs and memory pools
32 */
33#define RPC_BUFFER_MAXSIZE (2048)
34#define RPC_BUFFER_POOLSIZE (8)
35#define RPC_TASK_POOLSIZE (8)
36static struct kmem_cache *rpc_task_slabp __read_mostly;
37static struct kmem_cache *rpc_buffer_slabp __read_mostly;
38static mempool_t *rpc_task_mempool __read_mostly;
39static mempool_t *rpc_buffer_mempool __read_mostly;
40
41static void rpc_async_schedule(struct work_struct *);
42static void rpc_release_task(struct rpc_task *task);
43static void __rpc_queue_timer_fn(unsigned long ptr);
44
45/*
46 * RPC tasks sit here while waiting for conditions to improve.
47 */
48static struct rpc_wait_queue delay_queue;
49
50/*
51 * rpciod-related stuff
52 */
53struct workqueue_struct *rpciod_workqueue;
54
55/*
56 * Disable the timer for a given RPC task. Should be called with
57 * queue->lock and bh_disabled in order to avoid races within
58 * rpc_run_timer().
59 */
60static void
61__rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
62{
63 if (task->tk_timeout == 0)
64 return;
65 dprintk("RPC: %5u disabling timer\n", task->tk_pid);
66 task->tk_timeout = 0;
67 list_del(&task->u.tk_wait.timer_list);
68 if (list_empty(&queue->timer_list.list))
69 del_timer(&queue->timer_list.timer);
70}
71
72static void
73rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
74{
75 queue->timer_list.expires = expires;
76 mod_timer(&queue->timer_list.timer, expires);
77}
78
79/*
80 * Set up a timer for the current task.
81 */
82static void
83__rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
84{
85 if (!task->tk_timeout)
86 return;
87
88 dprintk("RPC: %5u setting alarm for %lu ms\n",
89 task->tk_pid, task->tk_timeout * 1000 / HZ);
90
91 task->u.tk_wait.expires = jiffies + task->tk_timeout;
92 if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
93 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
94 list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
95}
96
97/*
98 * Add new request to a priority queue.
99 */
100static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
101 struct rpc_task *task,
102 unsigned char queue_priority)
103{
104 struct list_head *q;
105 struct rpc_task *t;
106
107 INIT_LIST_HEAD(&task->u.tk_wait.links);
108 q = &queue->tasks[queue_priority];
109 if (unlikely(queue_priority > queue->maxpriority))
110 q = &queue->tasks[queue->maxpriority];
111 list_for_each_entry(t, q, u.tk_wait.list) {
112 if (t->tk_owner == task->tk_owner) {
113 list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
114 return;
115 }
116 }
117 list_add_tail(&task->u.tk_wait.list, q);
118}
119
120/*
121 * Add new request to wait queue.
122 *
123 * Swapper tasks always get inserted at the head of the queue.
124 * This should avoid many nasty memory deadlocks and hopefully
125 * improve overall performance.
126 * Everyone else gets appended to the queue to ensure proper FIFO behavior.
127 */
128static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
129 struct rpc_task *task,
130 unsigned char queue_priority)
131{
132 BUG_ON (RPC_IS_QUEUED(task));
133
134 if (RPC_IS_PRIORITY(queue))
135 __rpc_add_wait_queue_priority(queue, task, queue_priority);
136 else if (RPC_IS_SWAPPER(task))
137 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
138 else
139 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
140 task->tk_waitqueue = queue;
141 queue->qlen++;
142 rpc_set_queued(task);
143
144 dprintk("RPC: %5u added to queue %p \"%s\"\n",
145 task->tk_pid, queue, rpc_qname(queue));
146}
147
148/*
149 * Remove request from a priority queue.
150 */
151static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
152{
153 struct rpc_task *t;
154
155 if (!list_empty(&task->u.tk_wait.links)) {
156 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
157 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
158 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
159 }
160}
161
162/*
163 * Remove request from queue.
164 * Note: must be called with spin lock held.
165 */
166static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
167{
168 __rpc_disable_timer(queue, task);
169 if (RPC_IS_PRIORITY(queue))
170 __rpc_remove_wait_queue_priority(task);
171 list_del(&task->u.tk_wait.list);
172 queue->qlen--;
173 dprintk("RPC: %5u removed from queue %p \"%s\"\n",
174 task->tk_pid, queue, rpc_qname(queue));
175}
176
177static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
178{
179 queue->priority = priority;
180 queue->count = 1 << (priority * 2);
181}
182
183static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
184{
185 queue->owner = pid;
186 queue->nr = RPC_BATCH_COUNT;
187}
188
189static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
190{
191 rpc_set_waitqueue_priority(queue, queue->maxpriority);
192 rpc_set_waitqueue_owner(queue, 0);
193}
194
195static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
196{
197 int i;
198
199 spin_lock_init(&queue->lock);
200 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
201 INIT_LIST_HEAD(&queue->tasks[i]);
202 queue->maxpriority = nr_queues - 1;
203 rpc_reset_waitqueue_priority(queue);
204 queue->qlen = 0;
205 setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
206 INIT_LIST_HEAD(&queue->timer_list.list);
207#ifdef RPC_DEBUG
208 queue->name = qname;
209#endif
210}
211
212void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
213{
214 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
215}
216EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
217
218void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
219{
220 __rpc_init_priority_wait_queue(queue, qname, 1);
221}
222EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
223
224void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
225{
226 del_timer_sync(&queue->timer_list.timer);
227}
228EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
229
230static int rpc_wait_bit_killable(void *word)
231{
232 if (fatal_signal_pending(current))
233 return -ERESTARTSYS;
234 schedule();
235 return 0;
236}
237
238#ifdef RPC_DEBUG
239static void rpc_task_set_debuginfo(struct rpc_task *task)
240{
241 static atomic_t rpc_pid;
242
243 task->tk_pid = atomic_inc_return(&rpc_pid);
244}
245#else
246static inline void rpc_task_set_debuginfo(struct rpc_task *task)
247{
248}
249#endif
250
251static void rpc_set_active(struct rpc_task *task)
252{
253 rpc_task_set_debuginfo(task);
254 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
255}
256
257/*
258 * Mark an RPC call as having completed by clearing the 'active' bit
259 * and then waking up all tasks that were sleeping.
260 */
261static int rpc_complete_task(struct rpc_task *task)
262{
263 void *m = &task->tk_runstate;
264 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
265 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
266 unsigned long flags;
267 int ret;
268
269 spin_lock_irqsave(&wq->lock, flags);
270 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
271 ret = atomic_dec_and_test(&task->tk_count);
272 if (waitqueue_active(wq))
273 __wake_up_locked_key(wq, TASK_NORMAL, &k);
274 spin_unlock_irqrestore(&wq->lock, flags);
275 return ret;
276}
277
278/*
279 * Allow callers to wait for completion of an RPC call
280 *
281 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
282 * to enforce taking of the wq->lock and hence avoid races with
283 * rpc_complete_task().
284 */
285int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
286{
287 if (action == NULL)
288 action = rpc_wait_bit_killable;
289 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
290 action, TASK_KILLABLE);
291}
292EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
293
294/*
295 * Make an RPC task runnable.
296 *
297 * Note: If the task is ASYNC, this must be called with
298 * the spinlock held to protect the wait queue operation.
299 */
300static void rpc_make_runnable(struct rpc_task *task)
301{
302 rpc_clear_queued(task);
303 if (rpc_test_and_set_running(task))
304 return;
305 if (RPC_IS_ASYNC(task)) {
306 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
307 queue_work(rpciod_workqueue, &task->u.tk_work);
308 } else
309 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
310}
311
312/*
313 * Prepare for sleeping on a wait queue.
314 * By always appending tasks to the list we ensure FIFO behavior.
315 * NB: An RPC task will only receive interrupt-driven events as long
316 * as it's on a wait queue.
317 */
318static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
319 struct rpc_task *task,
320 rpc_action action,
321 unsigned char queue_priority)
322{
323 dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
324 task->tk_pid, rpc_qname(q), jiffies);
325
326 __rpc_add_wait_queue(q, task, queue_priority);
327
328 BUG_ON(task->tk_callback != NULL);
329 task->tk_callback = action;
330 __rpc_add_timer(q, task);
331}
332
333void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
334 rpc_action action)
335{
336 /* We shouldn't ever put an inactive task to sleep */
337 BUG_ON(!RPC_IS_ACTIVATED(task));
338
339 /*
340 * Protect the queue operations.
341 */
342 spin_lock_bh(&q->lock);
343 __rpc_sleep_on_priority(q, task, action, task->tk_priority);
344 spin_unlock_bh(&q->lock);
345}
346EXPORT_SYMBOL_GPL(rpc_sleep_on);
347
348void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
349 rpc_action action, int priority)
350{
351 /* We shouldn't ever put an inactive task to sleep */
352 BUG_ON(!RPC_IS_ACTIVATED(task));
353
354 /*
355 * Protect the queue operations.
356 */
357 spin_lock_bh(&q->lock);
358 __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
359 spin_unlock_bh(&q->lock);
360}
361
362/**
363 * __rpc_do_wake_up_task - wake up a single rpc_task
364 * @queue: wait queue
365 * @task: task to be woken up
366 *
367 * Caller must hold queue->lock, and have cleared the task queued flag.
368 */
369static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
370{
371 dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
372 task->tk_pid, jiffies);
373
374 /* Has the task been executed yet? If not, we cannot wake it up! */
375 if (!RPC_IS_ACTIVATED(task)) {
376 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
377 return;
378 }
379
380 __rpc_remove_wait_queue(queue, task);
381
382 rpc_make_runnable(task);
383
384 dprintk("RPC: __rpc_wake_up_task done\n");
385}
386
387/*
388 * Wake up a queued task while the queue lock is being held
389 */
390static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
391{
392 if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
393 __rpc_do_wake_up_task(queue, task);
394}
395
396/*
397 * Tests whether rpc queue is empty
398 */
399int rpc_queue_empty(struct rpc_wait_queue *queue)
400{
401 int res;
402
403 spin_lock_bh(&queue->lock);
404 res = queue->qlen;
405 spin_unlock_bh(&queue->lock);
406 return res == 0;
407}
408EXPORT_SYMBOL_GPL(rpc_queue_empty);
409
410/*
411 * Wake up a task on a specific queue
412 */
413void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
414{
415 spin_lock_bh(&queue->lock);
416 rpc_wake_up_task_queue_locked(queue, task);
417 spin_unlock_bh(&queue->lock);
418}
419EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
420
421/*
422 * Wake up the next task on a priority queue.
423 */
424static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
425{
426 struct list_head *q;
427 struct rpc_task *task;
428
429 /*
430 * Service a batch of tasks from a single owner.
431 */
432 q = &queue->tasks[queue->priority];
433 if (!list_empty(q)) {
434 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
435 if (queue->owner == task->tk_owner) {
436 if (--queue->nr)
437 goto out;
438 list_move_tail(&task->u.tk_wait.list, q);
439 }
440 /*
441 * Check if we need to switch queues.
442 */
443 if (--queue->count)
444 goto new_owner;
445 }
446
447 /*
448 * Service the next queue.
449 */
450 do {
451 if (q == &queue->tasks[0])
452 q = &queue->tasks[queue->maxpriority];
453 else
454 q = q - 1;
455 if (!list_empty(q)) {
456 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
457 goto new_queue;
458 }
459 } while (q != &queue->tasks[queue->priority]);
460
461 rpc_reset_waitqueue_priority(queue);
462 return NULL;
463
464new_queue:
465 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
466new_owner:
467 rpc_set_waitqueue_owner(queue, task->tk_owner);
468out:
469 rpc_wake_up_task_queue_locked(queue, task);
470 return task;
471}
472
473/*
474 * Wake up the next task on the wait queue.
475 */
476struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
477{
478 struct rpc_task *task = NULL;
479
480 dprintk("RPC: wake_up_next(%p \"%s\")\n",
481 queue, rpc_qname(queue));
482 spin_lock_bh(&queue->lock);
483 if (RPC_IS_PRIORITY(queue))
484 task = __rpc_wake_up_next_priority(queue);
485 else {
486 task_for_first(task, &queue->tasks[0])
487 rpc_wake_up_task_queue_locked(queue, task);
488 }
489 spin_unlock_bh(&queue->lock);
490
491 return task;
492}
493EXPORT_SYMBOL_GPL(rpc_wake_up_next);
494
495/**
496 * rpc_wake_up - wake up all rpc_tasks
497 * @queue: rpc_wait_queue on which the tasks are sleeping
498 *
499 * Grabs queue->lock
500 */
501void rpc_wake_up(struct rpc_wait_queue *queue)
502{
503 struct rpc_task *task, *next;
504 struct list_head *head;
505
506 spin_lock_bh(&queue->lock);
507 head = &queue->tasks[queue->maxpriority];
508 for (;;) {
509 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
510 rpc_wake_up_task_queue_locked(queue, task);
511 if (head == &queue->tasks[0])
512 break;
513 head--;
514 }
515 spin_unlock_bh(&queue->lock);
516}
517EXPORT_SYMBOL_GPL(rpc_wake_up);
518
519/**
520 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
521 * @queue: rpc_wait_queue on which the tasks are sleeping
522 * @status: status value to set
523 *
524 * Grabs queue->lock
525 */
526void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
527{
528 struct rpc_task *task, *next;
529 struct list_head *head;
530
531 spin_lock_bh(&queue->lock);
532 head = &queue->tasks[queue->maxpriority];
533 for (;;) {
534 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
535 task->tk_status = status;
536 rpc_wake_up_task_queue_locked(queue, task);
537 }
538 if (head == &queue->tasks[0])
539 break;
540 head--;
541 }
542 spin_unlock_bh(&queue->lock);
543}
544EXPORT_SYMBOL_GPL(rpc_wake_up_status);
545
546static void __rpc_queue_timer_fn(unsigned long ptr)
547{
548 struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
549 struct rpc_task *task, *n;
550 unsigned long expires, now, timeo;
551
552 spin_lock(&queue->lock);
553 expires = now = jiffies;
554 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
555 timeo = task->u.tk_wait.expires;
556 if (time_after_eq(now, timeo)) {
557 dprintk("RPC: %5u timeout\n", task->tk_pid);
558 task->tk_status = -ETIMEDOUT;
559 rpc_wake_up_task_queue_locked(queue, task);
560 continue;
561 }
562 if (expires == now || time_after(expires, timeo))
563 expires = timeo;
564 }
565 if (!list_empty(&queue->timer_list.list))
566 rpc_set_queue_timer(queue, expires);
567 spin_unlock(&queue->lock);
568}
569
570static void __rpc_atrun(struct rpc_task *task)
571{
572 task->tk_status = 0;
573}
574
575/*
576 * Run a task at a later time
577 */
578void rpc_delay(struct rpc_task *task, unsigned long delay)
579{
580 task->tk_timeout = delay;
581 rpc_sleep_on(&delay_queue, task, __rpc_atrun);
582}
583EXPORT_SYMBOL_GPL(rpc_delay);
584
585/*
586 * Helper to call task->tk_ops->rpc_call_prepare
587 */
588void rpc_prepare_task(struct rpc_task *task)
589{
590 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
591}
592
593/*
594 * Helper that calls task->tk_ops->rpc_call_done if it exists
595 */
596void rpc_exit_task(struct rpc_task *task)
597{
598 task->tk_action = NULL;
599 if (task->tk_ops->rpc_call_done != NULL) {
600 task->tk_ops->rpc_call_done(task, task->tk_calldata);
601 if (task->tk_action != NULL) {
602 WARN_ON(RPC_ASSASSINATED(task));
603 /* Always release the RPC slot and buffer memory */
604 xprt_release(task);
605 }
606 }
607}
608
609void rpc_exit(struct rpc_task *task, int status)
610{
611 task->tk_status = status;
612 task->tk_action = rpc_exit_task;
613 if (RPC_IS_QUEUED(task))
614 rpc_wake_up_queued_task(task->tk_waitqueue, task);
615}
616EXPORT_SYMBOL_GPL(rpc_exit);
617
618void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
619{
620 if (ops->rpc_release != NULL)
621 ops->rpc_release(calldata);
622}
623
624/*
625 * This is the RPC `scheduler' (or rather, the finite state machine).
626 */
627static void __rpc_execute(struct rpc_task *task)
628{
629 struct rpc_wait_queue *queue;
630 int task_is_async = RPC_IS_ASYNC(task);
631 int status = 0;
632
633 dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
634 task->tk_pid, task->tk_flags);
635
636 BUG_ON(RPC_IS_QUEUED(task));
637
638 for (;;) {
639 void (*do_action)(struct rpc_task *);
640
641 /*
642 * Execute any pending callback first.
643 */
644 do_action = task->tk_callback;
645 task->tk_callback = NULL;
646 if (do_action == NULL) {
647 /*
648 * Perform the next FSM step.
649 * tk_action may be NULL if the task has been killed.
650 * In particular, note that rpc_killall_tasks may
651 * do this at any time, so beware when dereferencing.
652 */
653 do_action = task->tk_action;
654 if (do_action == NULL)
655 break;
656 }
657 do_action(task);
658
659 /*
660 * Lockless check for whether task is sleeping or not.
661 */
662 if (!RPC_IS_QUEUED(task))
663 continue;
664 /*
665 * The queue->lock protects against races with
666 * rpc_make_runnable().
667 *
668 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
669 * rpc_task, rpc_make_runnable() can assign it to a
670 * different workqueue. We therefore cannot assume that the
671 * rpc_task pointer may still be dereferenced.
672 */
673 queue = task->tk_waitqueue;
674 spin_lock_bh(&queue->lock);
675 if (!RPC_IS_QUEUED(task)) {
676 spin_unlock_bh(&queue->lock);
677 continue;
678 }
679 rpc_clear_running(task);
680 spin_unlock_bh(&queue->lock);
681 if (task_is_async)
682 return;
683
684 /* sync task: sleep here */
685 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
686 status = out_of_line_wait_on_bit(&task->tk_runstate,
687 RPC_TASK_QUEUED, rpc_wait_bit_killable,
688 TASK_KILLABLE);
689 if (status == -ERESTARTSYS) {
690 /*
691 * When a sync task receives a signal, it exits with
692 * -ERESTARTSYS. In order to catch any callbacks that
693 * clean up after sleeping on some queue, we don't
694 * break the loop here, but go around once more.
695 */
696 dprintk("RPC: %5u got signal\n", task->tk_pid);
697 task->tk_flags |= RPC_TASK_KILLED;
698 rpc_exit(task, -ERESTARTSYS);
699 }
700 rpc_set_running(task);
701 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
702 }
703
704 dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
705 task->tk_status);
706 /* Release all resources associated with the task */
707 rpc_release_task(task);
708}
709
710/*
711 * User-visible entry point to the scheduler.
712 *
713 * This may be called recursively if e.g. an async NFS task updates
714 * the attributes and finds that dirty pages must be flushed.
715 * NOTE: Upon exit of this function the task is guaranteed to be
716 * released. In particular note that tk_release() will have
717 * been called, so your task memory may have been freed.
718 */
719void rpc_execute(struct rpc_task *task)
720{
721 rpc_set_active(task);
722 rpc_make_runnable(task);
723 if (!RPC_IS_ASYNC(task))
724 __rpc_execute(task);
725}
726
727static void rpc_async_schedule(struct work_struct *work)
728{
729 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
730}
731
732/**
733 * rpc_malloc - allocate an RPC buffer
734 * @task: RPC task that will use this buffer
735 * @size: requested byte size
736 *
737 * To prevent rpciod from hanging, this allocator never sleeps,
738 * returning NULL if the request cannot be serviced immediately.
739 * The caller can arrange to sleep in a way that is safe for rpciod.
740 *
741 * Most requests are 'small' (under 2KiB) and can be serviced from a
742 * mempool, ensuring that NFS reads and writes can always proceed,
743 * and that there is good locality of reference for these buffers.
744 *
745 * In order to avoid memory starvation triggering more writebacks of
746 * NFS requests, we avoid using GFP_KERNEL.
747 */
748void *rpc_malloc(struct rpc_task *task, size_t size)
749{
750 struct rpc_buffer *buf;
751 gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
752
753 size += sizeof(struct rpc_buffer);
754 if (size <= RPC_BUFFER_MAXSIZE)
755 buf = mempool_alloc(rpc_buffer_mempool, gfp);
756 else
757 buf = kmalloc(size, gfp);
758
759 if (!buf)
760 return NULL;
761
762 buf->len = size;
763 dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
764 task->tk_pid, size, buf);
765 return &buf->data;
766}
767EXPORT_SYMBOL_GPL(rpc_malloc);
768
769/**
770 * rpc_free - free buffer allocated via rpc_malloc
771 * @buffer: buffer to free
772 *
773 */
774void rpc_free(void *buffer)
775{
776 size_t size;
777 struct rpc_buffer *buf;
778
779 if (!buffer)
780 return;
781
782 buf = container_of(buffer, struct rpc_buffer, data);
783 size = buf->len;
784
785 dprintk("RPC: freeing buffer of size %zu at %p\n",
786 size, buf);
787
788 if (size <= RPC_BUFFER_MAXSIZE)
789 mempool_free(buf, rpc_buffer_mempool);
790 else
791 kfree(buf);
792}
793EXPORT_SYMBOL_GPL(rpc_free);
794
795/*
796 * Creation and deletion of RPC task structures
797 */
798static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
799{
800 memset(task, 0, sizeof(*task));
801 atomic_set(&task->tk_count, 1);
802 task->tk_flags = task_setup_data->flags;
803 task->tk_ops = task_setup_data->callback_ops;
804 task->tk_calldata = task_setup_data->callback_data;
805 INIT_LIST_HEAD(&task->tk_task);
806
807 /* Initialize retry counters */
808 task->tk_garb_retry = 2;
809 task->tk_cred_retry = 2;
810 task->tk_rebind_retry = 2;
811
812 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
813 task->tk_owner = current->tgid;
814
815 /* Initialize workqueue for async tasks */
816 task->tk_workqueue = task_setup_data->workqueue;
817
818 if (task->tk_ops->rpc_call_prepare != NULL)
819 task->tk_action = rpc_prepare_task;
820
821 /* starting timestamp */
822 task->tk_start = ktime_get();
823
824 dprintk("RPC: new task initialized, procpid %u\n",
825 task_pid_nr(current));
826}
827
828static struct rpc_task *
829rpc_alloc_task(void)
830{
831 return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
832}
833
834/*
835 * Create a new task for the specified client.
836 */
837struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
838{
839 struct rpc_task *task = setup_data->task;
840 unsigned short flags = 0;
841
842 if (task == NULL) {
843 task = rpc_alloc_task();
844 if (task == NULL) {
845 rpc_release_calldata(setup_data->callback_ops,
846 setup_data->callback_data);
847 return ERR_PTR(-ENOMEM);
848 }
849 flags = RPC_TASK_DYNAMIC;
850 }
851
852 rpc_init_task(task, setup_data);
853 task->tk_flags |= flags;
854 dprintk("RPC: allocated task %p\n", task);
855 return task;
856}
857
858static void rpc_free_task(struct rpc_task *task)
859{
860 const struct rpc_call_ops *tk_ops = task->tk_ops;
861 void *calldata = task->tk_calldata;
862
863 if (task->tk_flags & RPC_TASK_DYNAMIC) {
864 dprintk("RPC: %5u freeing task\n", task->tk_pid);
865 mempool_free(task, rpc_task_mempool);
866 }
867 rpc_release_calldata(tk_ops, calldata);
868}
869
870static void rpc_async_release(struct work_struct *work)
871{
872 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
873}
874
875static void rpc_release_resources_task(struct rpc_task *task)
876{
877 if (task->tk_rqstp)
878 xprt_release(task);
879 if (task->tk_msg.rpc_cred) {
880 put_rpccred(task->tk_msg.rpc_cred);
881 task->tk_msg.rpc_cred = NULL;
882 }
883 rpc_task_release_client(task);
884}
885
886static void rpc_final_put_task(struct rpc_task *task,
887 struct workqueue_struct *q)
888{
889 if (q != NULL) {
890 INIT_WORK(&task->u.tk_work, rpc_async_release);
891 queue_work(q, &task->u.tk_work);
892 } else
893 rpc_free_task(task);
894}
895
896static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
897{
898 if (atomic_dec_and_test(&task->tk_count)) {
899 rpc_release_resources_task(task);
900 rpc_final_put_task(task, q);
901 }
902}
903
904void rpc_put_task(struct rpc_task *task)
905{
906 rpc_do_put_task(task, NULL);
907}
908EXPORT_SYMBOL_GPL(rpc_put_task);
909
910void rpc_put_task_async(struct rpc_task *task)
911{
912 rpc_do_put_task(task, task->tk_workqueue);
913}
914EXPORT_SYMBOL_GPL(rpc_put_task_async);
915
916static void rpc_release_task(struct rpc_task *task)
917{
918 dprintk("RPC: %5u release task\n", task->tk_pid);
919
920 BUG_ON (RPC_IS_QUEUED(task));
921
922 rpc_release_resources_task(task);
923
924 /*
925 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
926 * so it should be safe to use task->tk_count as a test for whether
927 * or not any other processes still hold references to our rpc_task.
928 */
929 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
930 /* Wake up anyone who may be waiting for task completion */
931 if (!rpc_complete_task(task))
932 return;
933 } else {
934 if (!atomic_dec_and_test(&task->tk_count))
935 return;
936 }
937 rpc_final_put_task(task, task->tk_workqueue);
938}
939
940int rpciod_up(void)
941{
942 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
943}
944
945void rpciod_down(void)
946{
947 module_put(THIS_MODULE);
948}
949
950/*
951 * Start up the rpciod workqueue.
952 */
953static int rpciod_start(void)
954{
955 struct workqueue_struct *wq;
956
957 /*
958 * Create the rpciod thread and wait for it to start.
959 */
960 dprintk("RPC: creating workqueue rpciod\n");
961 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
962 rpciod_workqueue = wq;
963 return rpciod_workqueue != NULL;
964}
965
966static void rpciod_stop(void)
967{
968 struct workqueue_struct *wq = NULL;
969
970 if (rpciod_workqueue == NULL)
971 return;
972 dprintk("RPC: destroying workqueue rpciod\n");
973
974 wq = rpciod_workqueue;
975 rpciod_workqueue = NULL;
976 destroy_workqueue(wq);
977}
978
979void
980rpc_destroy_mempool(void)
981{
982 rpciod_stop();
983 if (rpc_buffer_mempool)
984 mempool_destroy(rpc_buffer_mempool);
985 if (rpc_task_mempool)
986 mempool_destroy(rpc_task_mempool);
987 if (rpc_task_slabp)
988 kmem_cache_destroy(rpc_task_slabp);
989 if (rpc_buffer_slabp)
990 kmem_cache_destroy(rpc_buffer_slabp);
991 rpc_destroy_wait_queue(&delay_queue);
992}
993
994int
995rpc_init_mempool(void)
996{
997 /*
998 * The following is not strictly a mempool initialisation,
999 * but there is no harm in doing it here
1000 */
1001 rpc_init_wait_queue(&delay_queue, "delayq");
1002 if (!rpciod_start())
1003 goto err_nomem;
1004
1005 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1006 sizeof(struct rpc_task),
1007 0, SLAB_HWCACHE_ALIGN,
1008 NULL);
1009 if (!rpc_task_slabp)
1010 goto err_nomem;
1011 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1012 RPC_BUFFER_MAXSIZE,
1013 0, SLAB_HWCACHE_ALIGN,
1014 NULL);
1015 if (!rpc_buffer_slabp)
1016 goto err_nomem;
1017 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1018 rpc_task_slabp);
1019 if (!rpc_task_mempool)
1020 goto err_nomem;
1021 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1022 rpc_buffer_slabp);
1023 if (!rpc_buffer_mempool)
1024 goto err_nomem;
1025 return 0;
1026err_nomem:
1027 rpc_destroy_mempool();
1028 return -ENOMEM;
1029}
1/*
2 * linux/net/sunrpc/sched.c
3 *
4 * Scheduling for synchronous and asynchronous RPC requests.
5 *
6 * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
7 *
8 * TCP NFS related read + write fixes
9 * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
10 */
11
12#include <linux/module.h>
13
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>
22
23#include <linux/sunrpc/clnt.h>
24
25#include "sunrpc.h"
26
27#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
28#define RPCDBG_FACILITY RPCDBG_SCHED
29#endif
30
31#define CREATE_TRACE_POINTS
32#include <trace/events/sunrpc.h>
33
34/*
35 * RPC slabs and memory pools
36 */
37#define RPC_BUFFER_MAXSIZE (2048)
38#define RPC_BUFFER_POOLSIZE (8)
39#define RPC_TASK_POOLSIZE (8)
40static struct kmem_cache *rpc_task_slabp __read_mostly;
41static struct kmem_cache *rpc_buffer_slabp __read_mostly;
42static mempool_t *rpc_task_mempool __read_mostly;
43static mempool_t *rpc_buffer_mempool __read_mostly;
44
45static void rpc_async_schedule(struct work_struct *);
46static void rpc_release_task(struct rpc_task *task);
47static void __rpc_queue_timer_fn(struct timer_list *t);
48
49/*
50 * RPC tasks sit here while waiting for conditions to improve.
51 */
52static struct rpc_wait_queue delay_queue;
53
54/*
55 * rpciod-related stuff
56 */
57struct workqueue_struct *rpciod_workqueue __read_mostly;
58struct workqueue_struct *xprtiod_workqueue __read_mostly;
59
60/*
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
63 * rpc_run_timer().
64 */
65static void
66__rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
67{
68 if (task->tk_timeout == 0)
69 return;
70 dprintk("RPC: %5u disabling timer\n", task->tk_pid);
71 task->tk_timeout = 0;
72 list_del(&task->u.tk_wait.timer_list);
73 if (list_empty(&queue->timer_list.list))
74 del_timer(&queue->timer_list.timer);
75}
76
77static void
78rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
79{
80 queue->timer_list.expires = expires;
81 mod_timer(&queue->timer_list.timer, expires);
82}
83
84/*
85 * Set up a timer for the current task.
86 */
87static void
88__rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
89{
90 if (!task->tk_timeout)
91 return;
92
93 dprintk("RPC: %5u setting alarm for %u ms\n",
94 task->tk_pid, jiffies_to_msecs(task->tk_timeout));
95
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);
100}
101
102static void rpc_rotate_queue_owner(struct rpc_wait_queue *queue)
103{
104 struct list_head *q = &queue->tasks[queue->priority];
105 struct rpc_task *task;
106
107 if (!list_empty(q)) {
108 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
109 if (task->tk_owner == queue->owner)
110 list_move_tail(&task->u.tk_wait.list, q);
111 }
112}
113
114static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
115{
116 if (queue->priority != priority) {
117 /* Fairness: rotate the list when changing priority */
118 rpc_rotate_queue_owner(queue);
119 queue->priority = priority;
120 }
121}
122
123static void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
124{
125 queue->owner = pid;
126 queue->nr = RPC_BATCH_COUNT;
127}
128
129static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
130{
131 rpc_set_waitqueue_priority(queue, queue->maxpriority);
132 rpc_set_waitqueue_owner(queue, 0);
133}
134
135/*
136 * Add new request to a priority queue.
137 */
138static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
139 struct rpc_task *task,
140 unsigned char queue_priority)
141{
142 struct list_head *q;
143 struct rpc_task *t;
144
145 INIT_LIST_HEAD(&task->u.tk_wait.links);
146 if (unlikely(queue_priority > queue->maxpriority))
147 queue_priority = queue->maxpriority;
148 if (queue_priority > queue->priority)
149 rpc_set_waitqueue_priority(queue, queue_priority);
150 q = &queue->tasks[queue_priority];
151 list_for_each_entry(t, q, u.tk_wait.list) {
152 if (t->tk_owner == task->tk_owner) {
153 list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
154 return;
155 }
156 }
157 list_add_tail(&task->u.tk_wait.list, q);
158}
159
160/*
161 * Add new request to wait queue.
162 *
163 * Swapper tasks always get inserted at the head of the queue.
164 * This should avoid many nasty memory deadlocks and hopefully
165 * improve overall performance.
166 * Everyone else gets appended to the queue to ensure proper FIFO behavior.
167 */
168static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
169 struct rpc_task *task,
170 unsigned char queue_priority)
171{
172 WARN_ON_ONCE(RPC_IS_QUEUED(task));
173 if (RPC_IS_QUEUED(task))
174 return;
175
176 if (RPC_IS_PRIORITY(queue))
177 __rpc_add_wait_queue_priority(queue, task, queue_priority);
178 else if (RPC_IS_SWAPPER(task))
179 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
180 else
181 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
182 task->tk_waitqueue = queue;
183 queue->qlen++;
184 /* barrier matches the read in rpc_wake_up_task_queue_locked() */
185 smp_wmb();
186 rpc_set_queued(task);
187
188 dprintk("RPC: %5u added to queue %p \"%s\"\n",
189 task->tk_pid, queue, rpc_qname(queue));
190}
191
192/*
193 * Remove request from a priority queue.
194 */
195static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
196{
197 struct rpc_task *t;
198
199 if (!list_empty(&task->u.tk_wait.links)) {
200 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
201 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
202 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
203 }
204}
205
206/*
207 * Remove request from queue.
208 * Note: must be called with spin lock held.
209 */
210static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
211{
212 __rpc_disable_timer(queue, task);
213 if (RPC_IS_PRIORITY(queue))
214 __rpc_remove_wait_queue_priority(task);
215 list_del(&task->u.tk_wait.list);
216 queue->qlen--;
217 dprintk("RPC: %5u removed from queue %p \"%s\"\n",
218 task->tk_pid, queue, rpc_qname(queue));
219}
220
221static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
222{
223 int i;
224
225 spin_lock_init(&queue->lock);
226 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
227 INIT_LIST_HEAD(&queue->tasks[i]);
228 queue->maxpriority = nr_queues - 1;
229 rpc_reset_waitqueue_priority(queue);
230 queue->qlen = 0;
231 timer_setup(&queue->timer_list.timer, __rpc_queue_timer_fn, 0);
232 INIT_LIST_HEAD(&queue->timer_list.list);
233 rpc_assign_waitqueue_name(queue, qname);
234}
235
236void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
237{
238 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
239}
240EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
241
242void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
243{
244 __rpc_init_priority_wait_queue(queue, qname, 1);
245}
246EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
247
248void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
249{
250 del_timer_sync(&queue->timer_list.timer);
251}
252EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
253
254static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
255{
256 freezable_schedule_unsafe();
257 if (signal_pending_state(mode, current))
258 return -ERESTARTSYS;
259 return 0;
260}
261
262#if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
263static void rpc_task_set_debuginfo(struct rpc_task *task)
264{
265 static atomic_t rpc_pid;
266
267 task->tk_pid = atomic_inc_return(&rpc_pid);
268}
269#else
270static inline void rpc_task_set_debuginfo(struct rpc_task *task)
271{
272}
273#endif
274
275static void rpc_set_active(struct rpc_task *task)
276{
277 rpc_task_set_debuginfo(task);
278 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
279 trace_rpc_task_begin(task, NULL);
280}
281
282/*
283 * Mark an RPC call as having completed by clearing the 'active' bit
284 * and then waking up all tasks that were sleeping.
285 */
286static int rpc_complete_task(struct rpc_task *task)
287{
288 void *m = &task->tk_runstate;
289 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
290 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
291 unsigned long flags;
292 int ret;
293
294 trace_rpc_task_complete(task, NULL);
295
296 spin_lock_irqsave(&wq->lock, flags);
297 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
298 ret = atomic_dec_and_test(&task->tk_count);
299 if (waitqueue_active(wq))
300 __wake_up_locked_key(wq, TASK_NORMAL, &k);
301 spin_unlock_irqrestore(&wq->lock, flags);
302 return ret;
303}
304
305/*
306 * Allow callers to wait for completion of an RPC call
307 *
308 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
309 * to enforce taking of the wq->lock and hence avoid races with
310 * rpc_complete_task().
311 */
312int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
313{
314 if (action == NULL)
315 action = rpc_wait_bit_killable;
316 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
317 action, TASK_KILLABLE);
318}
319EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
320
321/*
322 * Make an RPC task runnable.
323 *
324 * Note: If the task is ASYNC, and is being made runnable after sitting on an
325 * rpc_wait_queue, this must be called with the queue spinlock held to protect
326 * the wait queue operation.
327 * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
328 * which is needed to ensure that __rpc_execute() doesn't loop (due to the
329 * lockless RPC_IS_QUEUED() test) before we've had a chance to test
330 * the RPC_TASK_RUNNING flag.
331 */
332static void rpc_make_runnable(struct workqueue_struct *wq,
333 struct rpc_task *task)
334{
335 bool need_wakeup = !rpc_test_and_set_running(task);
336
337 rpc_clear_queued(task);
338 if (!need_wakeup)
339 return;
340 if (RPC_IS_ASYNC(task)) {
341 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
342 queue_work(wq, &task->u.tk_work);
343 } else
344 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
345}
346
347/*
348 * Prepare for sleeping on a wait queue.
349 * By always appending tasks to the list we ensure FIFO behavior.
350 * NB: An RPC task will only receive interrupt-driven events as long
351 * as it's on a wait queue.
352 */
353static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
354 struct rpc_task *task,
355 rpc_action action,
356 unsigned char queue_priority)
357{
358 dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
359 task->tk_pid, rpc_qname(q), jiffies);
360
361 trace_rpc_task_sleep(task, q);
362
363 __rpc_add_wait_queue(q, task, queue_priority);
364
365 WARN_ON_ONCE(task->tk_callback != NULL);
366 task->tk_callback = action;
367 __rpc_add_timer(q, task);
368}
369
370void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
371 rpc_action action)
372{
373 /* We shouldn't ever put an inactive task to sleep */
374 WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
375 if (!RPC_IS_ACTIVATED(task)) {
376 task->tk_status = -EIO;
377 rpc_put_task_async(task);
378 return;
379 }
380
381 /*
382 * Protect the queue operations.
383 */
384 spin_lock_bh(&q->lock);
385 __rpc_sleep_on_priority(q, task, action, task->tk_priority);
386 spin_unlock_bh(&q->lock);
387}
388EXPORT_SYMBOL_GPL(rpc_sleep_on);
389
390void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
391 rpc_action action, int priority)
392{
393 /* We shouldn't ever put an inactive task to sleep */
394 WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
395 if (!RPC_IS_ACTIVATED(task)) {
396 task->tk_status = -EIO;
397 rpc_put_task_async(task);
398 return;
399 }
400
401 /*
402 * Protect the queue operations.
403 */
404 spin_lock_bh(&q->lock);
405 __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
406 spin_unlock_bh(&q->lock);
407}
408EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
409
410/**
411 * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
412 * @wq: workqueue on which to run task
413 * @queue: wait queue
414 * @task: task to be woken up
415 *
416 * Caller must hold queue->lock, and have cleared the task queued flag.
417 */
418static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
419 struct rpc_wait_queue *queue,
420 struct rpc_task *task)
421{
422 dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
423 task->tk_pid, jiffies);
424
425 /* Has the task been executed yet? If not, we cannot wake it up! */
426 if (!RPC_IS_ACTIVATED(task)) {
427 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
428 return;
429 }
430
431 trace_rpc_task_wakeup(task, queue);
432
433 __rpc_remove_wait_queue(queue, task);
434
435 rpc_make_runnable(wq, task);
436
437 dprintk("RPC: __rpc_wake_up_task done\n");
438}
439
440/*
441 * Wake up a queued task while the queue lock is being held
442 */
443static void rpc_wake_up_task_on_wq_queue_locked(struct workqueue_struct *wq,
444 struct rpc_wait_queue *queue, struct rpc_task *task)
445{
446 if (RPC_IS_QUEUED(task)) {
447 smp_rmb();
448 if (task->tk_waitqueue == queue)
449 __rpc_do_wake_up_task_on_wq(wq, queue, task);
450 }
451}
452
453/*
454 * Wake up a queued task while the queue lock is being held
455 */
456static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
457{
458 rpc_wake_up_task_on_wq_queue_locked(rpciod_workqueue, queue, task);
459}
460
461/*
462 * Wake up a task on a specific queue
463 */
464void rpc_wake_up_queued_task_on_wq(struct workqueue_struct *wq,
465 struct rpc_wait_queue *queue,
466 struct rpc_task *task)
467{
468 spin_lock_bh(&queue->lock);
469 rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
470 spin_unlock_bh(&queue->lock);
471}
472
473/*
474 * Wake up a task on a specific queue
475 */
476void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
477{
478 spin_lock_bh(&queue->lock);
479 rpc_wake_up_task_queue_locked(queue, task);
480 spin_unlock_bh(&queue->lock);
481}
482EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
483
484/*
485 * Wake up the next task on a priority queue.
486 */
487static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
488{
489 struct list_head *q;
490 struct rpc_task *task;
491
492 /*
493 * Service a batch of tasks from a single owner.
494 */
495 q = &queue->tasks[queue->priority];
496 if (!list_empty(q)) {
497 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
498 if (queue->owner == task->tk_owner) {
499 if (--queue->nr)
500 goto out;
501 list_move_tail(&task->u.tk_wait.list, q);
502 }
503 /*
504 * Check if we need to switch queues.
505 */
506 goto new_owner;
507 }
508
509 /*
510 * Service the next queue.
511 */
512 do {
513 if (q == &queue->tasks[0])
514 q = &queue->tasks[queue->maxpriority];
515 else
516 q = q - 1;
517 if (!list_empty(q)) {
518 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
519 goto new_queue;
520 }
521 } while (q != &queue->tasks[queue->priority]);
522
523 rpc_reset_waitqueue_priority(queue);
524 return NULL;
525
526new_queue:
527 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
528new_owner:
529 rpc_set_waitqueue_owner(queue, task->tk_owner);
530out:
531 return task;
532}
533
534static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
535{
536 if (RPC_IS_PRIORITY(queue))
537 return __rpc_find_next_queued_priority(queue);
538 if (!list_empty(&queue->tasks[0]))
539 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
540 return NULL;
541}
542
543/*
544 * Wake up the first task on the wait queue.
545 */
546struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
547 struct rpc_wait_queue *queue,
548 bool (*func)(struct rpc_task *, void *), void *data)
549{
550 struct rpc_task *task = NULL;
551
552 dprintk("RPC: wake_up_first(%p \"%s\")\n",
553 queue, rpc_qname(queue));
554 spin_lock_bh(&queue->lock);
555 task = __rpc_find_next_queued(queue);
556 if (task != NULL) {
557 if (func(task, data))
558 rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
559 else
560 task = NULL;
561 }
562 spin_unlock_bh(&queue->lock);
563
564 return task;
565}
566
567/*
568 * Wake up the first task on the wait queue.
569 */
570struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
571 bool (*func)(struct rpc_task *, void *), void *data)
572{
573 return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
574}
575EXPORT_SYMBOL_GPL(rpc_wake_up_first);
576
577static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
578{
579 return true;
580}
581
582/*
583 * Wake up the next task on the wait queue.
584*/
585struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
586{
587 return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
588}
589EXPORT_SYMBOL_GPL(rpc_wake_up_next);
590
591/**
592 * rpc_wake_up - wake up all rpc_tasks
593 * @queue: rpc_wait_queue on which the tasks are sleeping
594 *
595 * Grabs queue->lock
596 */
597void rpc_wake_up(struct rpc_wait_queue *queue)
598{
599 struct list_head *head;
600
601 spin_lock_bh(&queue->lock);
602 head = &queue->tasks[queue->maxpriority];
603 for (;;) {
604 while (!list_empty(head)) {
605 struct rpc_task *task;
606 task = list_first_entry(head,
607 struct rpc_task,
608 u.tk_wait.list);
609 rpc_wake_up_task_queue_locked(queue, task);
610 }
611 if (head == &queue->tasks[0])
612 break;
613 head--;
614 }
615 spin_unlock_bh(&queue->lock);
616}
617EXPORT_SYMBOL_GPL(rpc_wake_up);
618
619/**
620 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
621 * @queue: rpc_wait_queue on which the tasks are sleeping
622 * @status: status value to set
623 *
624 * Grabs queue->lock
625 */
626void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
627{
628 struct list_head *head;
629
630 spin_lock_bh(&queue->lock);
631 head = &queue->tasks[queue->maxpriority];
632 for (;;) {
633 while (!list_empty(head)) {
634 struct rpc_task *task;
635 task = list_first_entry(head,
636 struct rpc_task,
637 u.tk_wait.list);
638 task->tk_status = status;
639 rpc_wake_up_task_queue_locked(queue, task);
640 }
641 if (head == &queue->tasks[0])
642 break;
643 head--;
644 }
645 spin_unlock_bh(&queue->lock);
646}
647EXPORT_SYMBOL_GPL(rpc_wake_up_status);
648
649static void __rpc_queue_timer_fn(struct timer_list *t)
650{
651 struct rpc_wait_queue *queue = from_timer(queue, t, timer_list.timer);
652 struct rpc_task *task, *n;
653 unsigned long expires, now, timeo;
654
655 spin_lock(&queue->lock);
656 expires = now = jiffies;
657 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
658 timeo = task->u.tk_wait.expires;
659 if (time_after_eq(now, timeo)) {
660 dprintk("RPC: %5u timeout\n", task->tk_pid);
661 task->tk_status = -ETIMEDOUT;
662 rpc_wake_up_task_queue_locked(queue, task);
663 continue;
664 }
665 if (expires == now || time_after(expires, timeo))
666 expires = timeo;
667 }
668 if (!list_empty(&queue->timer_list.list))
669 rpc_set_queue_timer(queue, expires);
670 spin_unlock(&queue->lock);
671}
672
673static void __rpc_atrun(struct rpc_task *task)
674{
675 if (task->tk_status == -ETIMEDOUT)
676 task->tk_status = 0;
677}
678
679/*
680 * Run a task at a later time
681 */
682void rpc_delay(struct rpc_task *task, unsigned long delay)
683{
684 task->tk_timeout = delay;
685 rpc_sleep_on(&delay_queue, task, __rpc_atrun);
686}
687EXPORT_SYMBOL_GPL(rpc_delay);
688
689/*
690 * Helper to call task->tk_ops->rpc_call_prepare
691 */
692void rpc_prepare_task(struct rpc_task *task)
693{
694 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
695}
696
697static void
698rpc_init_task_statistics(struct rpc_task *task)
699{
700 /* Initialize retry counters */
701 task->tk_garb_retry = 2;
702 task->tk_cred_retry = 2;
703 task->tk_rebind_retry = 2;
704
705 /* starting timestamp */
706 task->tk_start = ktime_get();
707}
708
709static void
710rpc_reset_task_statistics(struct rpc_task *task)
711{
712 task->tk_timeouts = 0;
713 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
714
715 rpc_init_task_statistics(task);
716}
717
718/*
719 * Helper that calls task->tk_ops->rpc_call_done if it exists
720 */
721void rpc_exit_task(struct rpc_task *task)
722{
723 task->tk_action = NULL;
724 if (task->tk_ops->rpc_call_done != NULL) {
725 task->tk_ops->rpc_call_done(task, task->tk_calldata);
726 if (task->tk_action != NULL) {
727 WARN_ON(RPC_ASSASSINATED(task));
728 /* Always release the RPC slot and buffer memory */
729 xprt_release(task);
730 rpc_reset_task_statistics(task);
731 }
732 }
733}
734
735void rpc_exit(struct rpc_task *task, int status)
736{
737 task->tk_status = status;
738 task->tk_action = rpc_exit_task;
739 if (RPC_IS_QUEUED(task))
740 rpc_wake_up_queued_task(task->tk_waitqueue, task);
741}
742EXPORT_SYMBOL_GPL(rpc_exit);
743
744void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
745{
746 if (ops->rpc_release != NULL)
747 ops->rpc_release(calldata);
748}
749
750/*
751 * This is the RPC `scheduler' (or rather, the finite state machine).
752 */
753static void __rpc_execute(struct rpc_task *task)
754{
755 struct rpc_wait_queue *queue;
756 int task_is_async = RPC_IS_ASYNC(task);
757 int status = 0;
758
759 dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
760 task->tk_pid, task->tk_flags);
761
762 WARN_ON_ONCE(RPC_IS_QUEUED(task));
763 if (RPC_IS_QUEUED(task))
764 return;
765
766 for (;;) {
767 void (*do_action)(struct rpc_task *);
768
769 /*
770 * Perform the next FSM step or a pending callback.
771 *
772 * tk_action may be NULL if the task has been killed.
773 * In particular, note that rpc_killall_tasks may
774 * do this at any time, so beware when dereferencing.
775 */
776 do_action = task->tk_action;
777 if (task->tk_callback) {
778 do_action = task->tk_callback;
779 task->tk_callback = NULL;
780 }
781 if (!do_action)
782 break;
783 trace_rpc_task_run_action(task, do_action);
784 do_action(task);
785
786 /*
787 * Lockless check for whether task is sleeping or not.
788 */
789 if (!RPC_IS_QUEUED(task))
790 continue;
791 /*
792 * The queue->lock protects against races with
793 * rpc_make_runnable().
794 *
795 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
796 * rpc_task, rpc_make_runnable() can assign it to a
797 * different workqueue. We therefore cannot assume that the
798 * rpc_task pointer may still be dereferenced.
799 */
800 queue = task->tk_waitqueue;
801 spin_lock_bh(&queue->lock);
802 if (!RPC_IS_QUEUED(task)) {
803 spin_unlock_bh(&queue->lock);
804 continue;
805 }
806 rpc_clear_running(task);
807 spin_unlock_bh(&queue->lock);
808 if (task_is_async)
809 return;
810
811 /* sync task: sleep here */
812 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
813 status = out_of_line_wait_on_bit(&task->tk_runstate,
814 RPC_TASK_QUEUED, rpc_wait_bit_killable,
815 TASK_KILLABLE);
816 if (status == -ERESTARTSYS) {
817 /*
818 * When a sync task receives a signal, it exits with
819 * -ERESTARTSYS. In order to catch any callbacks that
820 * clean up after sleeping on some queue, we don't
821 * break the loop here, but go around once more.
822 */
823 dprintk("RPC: %5u got signal\n", task->tk_pid);
824 task->tk_flags |= RPC_TASK_KILLED;
825 rpc_exit(task, -ERESTARTSYS);
826 }
827 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
828 }
829
830 dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
831 task->tk_status);
832 /* Release all resources associated with the task */
833 rpc_release_task(task);
834}
835
836/*
837 * User-visible entry point to the scheduler.
838 *
839 * This may be called recursively if e.g. an async NFS task updates
840 * the attributes and finds that dirty pages must be flushed.
841 * NOTE: Upon exit of this function the task is guaranteed to be
842 * released. In particular note that tk_release() will have
843 * been called, so your task memory may have been freed.
844 */
845void rpc_execute(struct rpc_task *task)
846{
847 bool is_async = RPC_IS_ASYNC(task);
848
849 rpc_set_active(task);
850 rpc_make_runnable(rpciod_workqueue, task);
851 if (!is_async)
852 __rpc_execute(task);
853}
854
855static void rpc_async_schedule(struct work_struct *work)
856{
857 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
858}
859
860/**
861 * rpc_malloc - allocate RPC buffer resources
862 * @task: RPC task
863 *
864 * A single memory region is allocated, which is split between the
865 * RPC call and RPC reply that this task is being used for. When
866 * this RPC is retired, the memory is released by calling rpc_free.
867 *
868 * To prevent rpciod from hanging, this allocator never sleeps,
869 * returning -ENOMEM and suppressing warning if the request cannot
870 * be serviced immediately. The caller can arrange to sleep in a
871 * way that is safe for rpciod.
872 *
873 * Most requests are 'small' (under 2KiB) and can be serviced from a
874 * mempool, ensuring that NFS reads and writes can always proceed,
875 * and that there is good locality of reference for these buffers.
876 *
877 * In order to avoid memory starvation triggering more writebacks of
878 * NFS requests, we avoid using GFP_KERNEL.
879 */
880int rpc_malloc(struct rpc_task *task)
881{
882 struct rpc_rqst *rqst = task->tk_rqstp;
883 size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
884 struct rpc_buffer *buf;
885 gfp_t gfp = GFP_NOIO | __GFP_NOWARN;
886
887 if (RPC_IS_SWAPPER(task))
888 gfp = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
889
890 size += sizeof(struct rpc_buffer);
891 if (size <= RPC_BUFFER_MAXSIZE)
892 buf = mempool_alloc(rpc_buffer_mempool, gfp);
893 else
894 buf = kmalloc(size, gfp);
895
896 if (!buf)
897 return -ENOMEM;
898
899 buf->len = size;
900 dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
901 task->tk_pid, size, buf);
902 rqst->rq_buffer = buf->data;
903 rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
904 return 0;
905}
906EXPORT_SYMBOL_GPL(rpc_malloc);
907
908/**
909 * rpc_free - free RPC buffer resources allocated via rpc_malloc
910 * @task: RPC task
911 *
912 */
913void rpc_free(struct rpc_task *task)
914{
915 void *buffer = task->tk_rqstp->rq_buffer;
916 size_t size;
917 struct rpc_buffer *buf;
918
919 buf = container_of(buffer, struct rpc_buffer, data);
920 size = buf->len;
921
922 dprintk("RPC: freeing buffer of size %zu at %p\n",
923 size, buf);
924
925 if (size <= RPC_BUFFER_MAXSIZE)
926 mempool_free(buf, rpc_buffer_mempool);
927 else
928 kfree(buf);
929}
930EXPORT_SYMBOL_GPL(rpc_free);
931
932/*
933 * Creation and deletion of RPC task structures
934 */
935static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
936{
937 memset(task, 0, sizeof(*task));
938 atomic_set(&task->tk_count, 1);
939 task->tk_flags = task_setup_data->flags;
940 task->tk_ops = task_setup_data->callback_ops;
941 task->tk_calldata = task_setup_data->callback_data;
942 INIT_LIST_HEAD(&task->tk_task);
943
944 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
945 task->tk_owner = current->tgid;
946
947 /* Initialize workqueue for async tasks */
948 task->tk_workqueue = task_setup_data->workqueue;
949
950 task->tk_xprt = xprt_get(task_setup_data->rpc_xprt);
951
952 if (task->tk_ops->rpc_call_prepare != NULL)
953 task->tk_action = rpc_prepare_task;
954
955 rpc_init_task_statistics(task);
956
957 dprintk("RPC: new task initialized, procpid %u\n",
958 task_pid_nr(current));
959}
960
961static struct rpc_task *
962rpc_alloc_task(void)
963{
964 return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
965}
966
967/*
968 * Create a new task for the specified client.
969 */
970struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
971{
972 struct rpc_task *task = setup_data->task;
973 unsigned short flags = 0;
974
975 if (task == NULL) {
976 task = rpc_alloc_task();
977 flags = RPC_TASK_DYNAMIC;
978 }
979
980 rpc_init_task(task, setup_data);
981 task->tk_flags |= flags;
982 dprintk("RPC: allocated task %p\n", task);
983 return task;
984}
985
986/*
987 * rpc_free_task - release rpc task and perform cleanups
988 *
989 * Note that we free up the rpc_task _after_ rpc_release_calldata()
990 * in order to work around a workqueue dependency issue.
991 *
992 * Tejun Heo states:
993 * "Workqueue currently considers two work items to be the same if they're
994 * on the same address and won't execute them concurrently - ie. it
995 * makes a work item which is queued again while being executed wait
996 * for the previous execution to complete.
997 *
998 * If a work function frees the work item, and then waits for an event
999 * which should be performed by another work item and *that* work item
1000 * recycles the freed work item, it can create a false dependency loop.
1001 * There really is no reliable way to detect this short of verifying
1002 * every memory free."
1003 *
1004 */
1005static void rpc_free_task(struct rpc_task *task)
1006{
1007 unsigned short tk_flags = task->tk_flags;
1008
1009 rpc_release_calldata(task->tk_ops, task->tk_calldata);
1010
1011 if (tk_flags & RPC_TASK_DYNAMIC) {
1012 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1013 mempool_free(task, rpc_task_mempool);
1014 }
1015}
1016
1017static void rpc_async_release(struct work_struct *work)
1018{
1019 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1020}
1021
1022static void rpc_release_resources_task(struct rpc_task *task)
1023{
1024 xprt_release(task);
1025 if (task->tk_msg.rpc_cred) {
1026 put_rpccred(task->tk_msg.rpc_cred);
1027 task->tk_msg.rpc_cred = NULL;
1028 }
1029 rpc_task_release_client(task);
1030}
1031
1032static void rpc_final_put_task(struct rpc_task *task,
1033 struct workqueue_struct *q)
1034{
1035 if (q != NULL) {
1036 INIT_WORK(&task->u.tk_work, rpc_async_release);
1037 queue_work(q, &task->u.tk_work);
1038 } else
1039 rpc_free_task(task);
1040}
1041
1042static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1043{
1044 if (atomic_dec_and_test(&task->tk_count)) {
1045 rpc_release_resources_task(task);
1046 rpc_final_put_task(task, q);
1047 }
1048}
1049
1050void rpc_put_task(struct rpc_task *task)
1051{
1052 rpc_do_put_task(task, NULL);
1053}
1054EXPORT_SYMBOL_GPL(rpc_put_task);
1055
1056void rpc_put_task_async(struct rpc_task *task)
1057{
1058 rpc_do_put_task(task, task->tk_workqueue);
1059}
1060EXPORT_SYMBOL_GPL(rpc_put_task_async);
1061
1062static void rpc_release_task(struct rpc_task *task)
1063{
1064 dprintk("RPC: %5u release task\n", task->tk_pid);
1065
1066 WARN_ON_ONCE(RPC_IS_QUEUED(task));
1067
1068 rpc_release_resources_task(task);
1069
1070 /*
1071 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1072 * so it should be safe to use task->tk_count as a test for whether
1073 * or not any other processes still hold references to our rpc_task.
1074 */
1075 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1076 /* Wake up anyone who may be waiting for task completion */
1077 if (!rpc_complete_task(task))
1078 return;
1079 } else {
1080 if (!atomic_dec_and_test(&task->tk_count))
1081 return;
1082 }
1083 rpc_final_put_task(task, task->tk_workqueue);
1084}
1085
1086int rpciod_up(void)
1087{
1088 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1089}
1090
1091void rpciod_down(void)
1092{
1093 module_put(THIS_MODULE);
1094}
1095
1096/*
1097 * Start up the rpciod workqueue.
1098 */
1099static int rpciod_start(void)
1100{
1101 struct workqueue_struct *wq;
1102
1103 /*
1104 * Create the rpciod thread and wait for it to start.
1105 */
1106 dprintk("RPC: creating workqueue rpciod\n");
1107 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1108 if (!wq)
1109 goto out_failed;
1110 rpciod_workqueue = wq;
1111 /* Note: highpri because network receive is latency sensitive */
1112 wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1113 if (!wq)
1114 goto free_rpciod;
1115 xprtiod_workqueue = wq;
1116 return 1;
1117free_rpciod:
1118 wq = rpciod_workqueue;
1119 rpciod_workqueue = NULL;
1120 destroy_workqueue(wq);
1121out_failed:
1122 return 0;
1123}
1124
1125static void rpciod_stop(void)
1126{
1127 struct workqueue_struct *wq = NULL;
1128
1129 if (rpciod_workqueue == NULL)
1130 return;
1131 dprintk("RPC: destroying workqueue rpciod\n");
1132
1133 wq = rpciod_workqueue;
1134 rpciod_workqueue = NULL;
1135 destroy_workqueue(wq);
1136 wq = xprtiod_workqueue;
1137 xprtiod_workqueue = NULL;
1138 destroy_workqueue(wq);
1139}
1140
1141void
1142rpc_destroy_mempool(void)
1143{
1144 rpciod_stop();
1145 mempool_destroy(rpc_buffer_mempool);
1146 mempool_destroy(rpc_task_mempool);
1147 kmem_cache_destroy(rpc_task_slabp);
1148 kmem_cache_destroy(rpc_buffer_slabp);
1149 rpc_destroy_wait_queue(&delay_queue);
1150}
1151
1152int
1153rpc_init_mempool(void)
1154{
1155 /*
1156 * The following is not strictly a mempool initialisation,
1157 * but there is no harm in doing it here
1158 */
1159 rpc_init_wait_queue(&delay_queue, "delayq");
1160 if (!rpciod_start())
1161 goto err_nomem;
1162
1163 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1164 sizeof(struct rpc_task),
1165 0, SLAB_HWCACHE_ALIGN,
1166 NULL);
1167 if (!rpc_task_slabp)
1168 goto err_nomem;
1169 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1170 RPC_BUFFER_MAXSIZE,
1171 0, SLAB_HWCACHE_ALIGN,
1172 NULL);
1173 if (!rpc_buffer_slabp)
1174 goto err_nomem;
1175 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1176 rpc_task_slabp);
1177 if (!rpc_task_mempool)
1178 goto err_nomem;
1179 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1180 rpc_buffer_slabp);
1181 if (!rpc_buffer_mempool)
1182 goto err_nomem;
1183 return 0;
1184err_nomem:
1185 rpc_destroy_mempool();
1186 return -ENOMEM;
1187}