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