Loading...
1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * linux/net/sunrpc/sched.c
4 *
5 * Scheduling for synchronous and asynchronous RPC requests.
6 *
7 * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
8 *
9 * TCP NFS related read + write fixes
10 * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
11 */
12
13#include <linux/module.h>
14
15#include <linux/sched.h>
16#include <linux/interrupt.h>
17#include <linux/slab.h>
18#include <linux/mempool.h>
19#include <linux/smp.h>
20#include <linux/spinlock.h>
21#include <linux/mutex.h>
22#include <linux/freezer.h>
23#include <linux/sched/mm.h>
24
25#include <linux/sunrpc/clnt.h>
26#include <linux/sunrpc/metrics.h>
27
28#include "sunrpc.h"
29
30#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
31#define RPCDBG_FACILITY RPCDBG_SCHED
32#endif
33
34#define CREATE_TRACE_POINTS
35#include <trace/events/sunrpc.h>
36
37/*
38 * RPC slabs and memory pools
39 */
40#define RPC_BUFFER_MAXSIZE (2048)
41#define RPC_BUFFER_POOLSIZE (8)
42#define RPC_TASK_POOLSIZE (8)
43static struct kmem_cache *rpc_task_slabp __read_mostly;
44static struct kmem_cache *rpc_buffer_slabp __read_mostly;
45static mempool_t *rpc_task_mempool __read_mostly;
46static mempool_t *rpc_buffer_mempool __read_mostly;
47
48static void rpc_async_schedule(struct work_struct *);
49static void rpc_release_task(struct rpc_task *task);
50static void __rpc_queue_timer_fn(struct work_struct *);
51
52/*
53 * RPC tasks sit here while waiting for conditions to improve.
54 */
55static struct rpc_wait_queue delay_queue;
56
57/*
58 * rpciod-related stuff
59 */
60struct workqueue_struct *rpciod_workqueue __read_mostly;
61struct workqueue_struct *xprtiod_workqueue __read_mostly;
62EXPORT_SYMBOL_GPL(xprtiod_workqueue);
63
64unsigned long
65rpc_task_timeout(const struct rpc_task *task)
66{
67 unsigned long timeout = READ_ONCE(task->tk_timeout);
68
69 if (timeout != 0) {
70 unsigned long now = jiffies;
71 if (time_before(now, timeout))
72 return timeout - now;
73 }
74 return 0;
75}
76EXPORT_SYMBOL_GPL(rpc_task_timeout);
77
78/*
79 * Disable the timer for a given RPC task. Should be called with
80 * queue->lock and bh_disabled in order to avoid races within
81 * rpc_run_timer().
82 */
83static void
84__rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
85{
86 if (list_empty(&task->u.tk_wait.timer_list))
87 return;
88 dprintk("RPC: %5u disabling timer\n", task->tk_pid);
89 task->tk_timeout = 0;
90 list_del(&task->u.tk_wait.timer_list);
91 if (list_empty(&queue->timer_list.list))
92 cancel_delayed_work(&queue->timer_list.dwork);
93}
94
95static void
96rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
97{
98 unsigned long now = jiffies;
99 queue->timer_list.expires = expires;
100 if (time_before_eq(expires, now))
101 expires = 0;
102 else
103 expires -= now;
104 mod_delayed_work(rpciod_workqueue, &queue->timer_list.dwork, expires);
105}
106
107/*
108 * Set up a timer for the current task.
109 */
110static void
111__rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task,
112 unsigned long timeout)
113{
114 dprintk("RPC: %5u setting alarm for %u ms\n",
115 task->tk_pid, jiffies_to_msecs(timeout - jiffies));
116
117 task->tk_timeout = timeout;
118 if (list_empty(&queue->timer_list.list) || time_before(timeout, queue->timer_list.expires))
119 rpc_set_queue_timer(queue, timeout);
120 list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
121}
122
123static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
124{
125 if (queue->priority != priority) {
126 queue->priority = priority;
127 queue->nr = 1U << priority;
128 }
129}
130
131static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
132{
133 rpc_set_waitqueue_priority(queue, queue->maxpriority);
134}
135
136/*
137 * Add a request to a queue list
138 */
139static void
140__rpc_list_enqueue_task(struct list_head *q, struct rpc_task *task)
141{
142 struct rpc_task *t;
143
144 list_for_each_entry(t, q, u.tk_wait.list) {
145 if (t->tk_owner == task->tk_owner) {
146 list_add_tail(&task->u.tk_wait.links,
147 &t->u.tk_wait.links);
148 /* Cache the queue head in task->u.tk_wait.list */
149 task->u.tk_wait.list.next = q;
150 task->u.tk_wait.list.prev = NULL;
151 return;
152 }
153 }
154 INIT_LIST_HEAD(&task->u.tk_wait.links);
155 list_add_tail(&task->u.tk_wait.list, q);
156}
157
158/*
159 * Remove request from a queue list
160 */
161static void
162__rpc_list_dequeue_task(struct rpc_task *task)
163{
164 struct list_head *q;
165 struct rpc_task *t;
166
167 if (task->u.tk_wait.list.prev == NULL) {
168 list_del(&task->u.tk_wait.links);
169 return;
170 }
171 if (!list_empty(&task->u.tk_wait.links)) {
172 t = list_first_entry(&task->u.tk_wait.links,
173 struct rpc_task,
174 u.tk_wait.links);
175 /* Assume __rpc_list_enqueue_task() cached the queue head */
176 q = t->u.tk_wait.list.next;
177 list_add_tail(&t->u.tk_wait.list, q);
178 list_del(&task->u.tk_wait.links);
179 }
180 list_del(&task->u.tk_wait.list);
181}
182
183/*
184 * Add new request to a priority queue.
185 */
186static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
187 struct rpc_task *task,
188 unsigned char queue_priority)
189{
190 if (unlikely(queue_priority > queue->maxpriority))
191 queue_priority = queue->maxpriority;
192 __rpc_list_enqueue_task(&queue->tasks[queue_priority], task);
193}
194
195/*
196 * Add new request to wait queue.
197 *
198 * Swapper tasks always get inserted at the head of the queue.
199 * This should avoid many nasty memory deadlocks and hopefully
200 * improve overall performance.
201 * Everyone else gets appended to the queue to ensure proper FIFO behavior.
202 */
203static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
204 struct rpc_task *task,
205 unsigned char queue_priority)
206{
207 WARN_ON_ONCE(RPC_IS_QUEUED(task));
208 if (RPC_IS_QUEUED(task))
209 return;
210
211 INIT_LIST_HEAD(&task->u.tk_wait.timer_list);
212 if (RPC_IS_PRIORITY(queue))
213 __rpc_add_wait_queue_priority(queue, task, queue_priority);
214 else if (RPC_IS_SWAPPER(task))
215 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
216 else
217 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
218 task->tk_waitqueue = queue;
219 queue->qlen++;
220 /* barrier matches the read in rpc_wake_up_task_queue_locked() */
221 smp_wmb();
222 rpc_set_queued(task);
223
224 dprintk("RPC: %5u added to queue %p \"%s\"\n",
225 task->tk_pid, queue, rpc_qname(queue));
226}
227
228/*
229 * Remove request from a priority queue.
230 */
231static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
232{
233 __rpc_list_dequeue_task(task);
234}
235
236/*
237 * Remove request from queue.
238 * Note: must be called with spin lock held.
239 */
240static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
241{
242 __rpc_disable_timer(queue, task);
243 if (RPC_IS_PRIORITY(queue))
244 __rpc_remove_wait_queue_priority(task);
245 else
246 list_del(&task->u.tk_wait.list);
247 queue->qlen--;
248 dprintk("RPC: %5u removed from queue %p \"%s\"\n",
249 task->tk_pid, queue, rpc_qname(queue));
250}
251
252static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
253{
254 int i;
255
256 spin_lock_init(&queue->lock);
257 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
258 INIT_LIST_HEAD(&queue->tasks[i]);
259 queue->maxpriority = nr_queues - 1;
260 rpc_reset_waitqueue_priority(queue);
261 queue->qlen = 0;
262 queue->timer_list.expires = 0;
263 INIT_DEFERRABLE_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
264 INIT_LIST_HEAD(&queue->timer_list.list);
265 rpc_assign_waitqueue_name(queue, qname);
266}
267
268void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
269{
270 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
271}
272EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
273
274void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
275{
276 __rpc_init_priority_wait_queue(queue, qname, 1);
277}
278EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
279
280void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
281{
282 cancel_delayed_work_sync(&queue->timer_list.dwork);
283}
284EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
285
286static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
287{
288 freezable_schedule_unsafe();
289 if (signal_pending_state(mode, current))
290 return -ERESTARTSYS;
291 return 0;
292}
293
294#if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
295static void rpc_task_set_debuginfo(struct rpc_task *task)
296{
297 static atomic_t rpc_pid;
298
299 task->tk_pid = atomic_inc_return(&rpc_pid);
300}
301#else
302static inline void rpc_task_set_debuginfo(struct rpc_task *task)
303{
304}
305#endif
306
307static void rpc_set_active(struct rpc_task *task)
308{
309 rpc_task_set_debuginfo(task);
310 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
311 trace_rpc_task_begin(task, NULL);
312}
313
314/*
315 * Mark an RPC call as having completed by clearing the 'active' bit
316 * and then waking up all tasks that were sleeping.
317 */
318static int rpc_complete_task(struct rpc_task *task)
319{
320 void *m = &task->tk_runstate;
321 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
322 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
323 unsigned long flags;
324 int ret;
325
326 trace_rpc_task_complete(task, NULL);
327
328 spin_lock_irqsave(&wq->lock, flags);
329 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
330 ret = atomic_dec_and_test(&task->tk_count);
331 if (waitqueue_active(wq))
332 __wake_up_locked_key(wq, TASK_NORMAL, &k);
333 spin_unlock_irqrestore(&wq->lock, flags);
334 return ret;
335}
336
337/*
338 * Allow callers to wait for completion of an RPC call
339 *
340 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
341 * to enforce taking of the wq->lock and hence avoid races with
342 * rpc_complete_task().
343 */
344int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
345{
346 if (action == NULL)
347 action = rpc_wait_bit_killable;
348 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
349 action, TASK_KILLABLE);
350}
351EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
352
353/*
354 * Make an RPC task runnable.
355 *
356 * Note: If the task is ASYNC, and is being made runnable after sitting on an
357 * rpc_wait_queue, this must be called with the queue spinlock held to protect
358 * the wait queue operation.
359 * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
360 * which is needed to ensure that __rpc_execute() doesn't loop (due to the
361 * lockless RPC_IS_QUEUED() test) before we've had a chance to test
362 * the RPC_TASK_RUNNING flag.
363 */
364static void rpc_make_runnable(struct workqueue_struct *wq,
365 struct rpc_task *task)
366{
367 bool need_wakeup = !rpc_test_and_set_running(task);
368
369 rpc_clear_queued(task);
370 if (!need_wakeup)
371 return;
372 if (RPC_IS_ASYNC(task)) {
373 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
374 queue_work(wq, &task->u.tk_work);
375 } else
376 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
377}
378
379/*
380 * Prepare for sleeping on a wait queue.
381 * By always appending tasks to the list we ensure FIFO behavior.
382 * NB: An RPC task will only receive interrupt-driven events as long
383 * as it's on a wait queue.
384 */
385static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
386 struct rpc_task *task,
387 unsigned char queue_priority)
388{
389 dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
390 task->tk_pid, rpc_qname(q), jiffies);
391
392 trace_rpc_task_sleep(task, q);
393
394 __rpc_add_wait_queue(q, task, queue_priority);
395
396}
397
398static void __rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
399 struct rpc_task *task, unsigned long timeout,
400 unsigned char queue_priority)
401{
402 if (time_is_after_jiffies(timeout)) {
403 __rpc_sleep_on_priority(q, task, queue_priority);
404 __rpc_add_timer(q, task, timeout);
405 } else
406 task->tk_status = -ETIMEDOUT;
407}
408
409static void rpc_set_tk_callback(struct rpc_task *task, rpc_action action)
410{
411 if (action && !WARN_ON_ONCE(task->tk_callback != NULL))
412 task->tk_callback = action;
413}
414
415static bool rpc_sleep_check_activated(struct rpc_task *task)
416{
417 /* We shouldn't ever put an inactive task to sleep */
418 if (WARN_ON_ONCE(!RPC_IS_ACTIVATED(task))) {
419 task->tk_status = -EIO;
420 rpc_put_task_async(task);
421 return false;
422 }
423 return true;
424}
425
426void rpc_sleep_on_timeout(struct rpc_wait_queue *q, struct rpc_task *task,
427 rpc_action action, unsigned long timeout)
428{
429 if (!rpc_sleep_check_activated(task))
430 return;
431
432 rpc_set_tk_callback(task, action);
433
434 /*
435 * Protect the queue operations.
436 */
437 spin_lock(&q->lock);
438 __rpc_sleep_on_priority_timeout(q, task, timeout, task->tk_priority);
439 spin_unlock(&q->lock);
440}
441EXPORT_SYMBOL_GPL(rpc_sleep_on_timeout);
442
443void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
444 rpc_action action)
445{
446 if (!rpc_sleep_check_activated(task))
447 return;
448
449 rpc_set_tk_callback(task, action);
450
451 WARN_ON_ONCE(task->tk_timeout != 0);
452 /*
453 * Protect the queue operations.
454 */
455 spin_lock(&q->lock);
456 __rpc_sleep_on_priority(q, task, task->tk_priority);
457 spin_unlock(&q->lock);
458}
459EXPORT_SYMBOL_GPL(rpc_sleep_on);
460
461void rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
462 struct rpc_task *task, unsigned long timeout, int priority)
463{
464 if (!rpc_sleep_check_activated(task))
465 return;
466
467 priority -= RPC_PRIORITY_LOW;
468 /*
469 * Protect the queue operations.
470 */
471 spin_lock(&q->lock);
472 __rpc_sleep_on_priority_timeout(q, task, timeout, priority);
473 spin_unlock(&q->lock);
474}
475EXPORT_SYMBOL_GPL(rpc_sleep_on_priority_timeout);
476
477void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
478 int priority)
479{
480 if (!rpc_sleep_check_activated(task))
481 return;
482
483 WARN_ON_ONCE(task->tk_timeout != 0);
484 priority -= RPC_PRIORITY_LOW;
485 /*
486 * Protect the queue operations.
487 */
488 spin_lock(&q->lock);
489 __rpc_sleep_on_priority(q, task, priority);
490 spin_unlock(&q->lock);
491}
492EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
493
494/**
495 * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
496 * @wq: workqueue on which to run task
497 * @queue: wait queue
498 * @task: task to be woken up
499 *
500 * Caller must hold queue->lock, and have cleared the task queued flag.
501 */
502static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
503 struct rpc_wait_queue *queue,
504 struct rpc_task *task)
505{
506 dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
507 task->tk_pid, jiffies);
508
509 /* Has the task been executed yet? If not, we cannot wake it up! */
510 if (!RPC_IS_ACTIVATED(task)) {
511 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
512 return;
513 }
514
515 trace_rpc_task_wakeup(task, queue);
516
517 __rpc_remove_wait_queue(queue, task);
518
519 rpc_make_runnable(wq, task);
520
521 dprintk("RPC: __rpc_wake_up_task done\n");
522}
523
524/*
525 * Wake up a queued task while the queue lock is being held
526 */
527static struct rpc_task *
528rpc_wake_up_task_on_wq_queue_action_locked(struct workqueue_struct *wq,
529 struct rpc_wait_queue *queue, struct rpc_task *task,
530 bool (*action)(struct rpc_task *, void *), void *data)
531{
532 if (RPC_IS_QUEUED(task)) {
533 smp_rmb();
534 if (task->tk_waitqueue == queue) {
535 if (action == NULL || action(task, data)) {
536 __rpc_do_wake_up_task_on_wq(wq, queue, task);
537 return task;
538 }
539 }
540 }
541 return NULL;
542}
543
544/*
545 * Wake up a queued task while the queue lock is being held
546 */
547static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue,
548 struct rpc_task *task)
549{
550 rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
551 task, NULL, NULL);
552}
553
554/*
555 * Wake up a task on a specific queue
556 */
557void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
558{
559 if (!RPC_IS_QUEUED(task))
560 return;
561 spin_lock(&queue->lock);
562 rpc_wake_up_task_queue_locked(queue, task);
563 spin_unlock(&queue->lock);
564}
565EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
566
567static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
568{
569 task->tk_status = *(int *)status;
570 return true;
571}
572
573static void
574rpc_wake_up_task_queue_set_status_locked(struct rpc_wait_queue *queue,
575 struct rpc_task *task, int status)
576{
577 rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
578 task, rpc_task_action_set_status, &status);
579}
580
581/**
582 * rpc_wake_up_queued_task_set_status - wake up a task and set task->tk_status
583 * @queue: pointer to rpc_wait_queue
584 * @task: pointer to rpc_task
585 * @status: integer error value
586 *
587 * If @task is queued on @queue, then it is woken up, and @task->tk_status is
588 * set to the value of @status.
589 */
590void
591rpc_wake_up_queued_task_set_status(struct rpc_wait_queue *queue,
592 struct rpc_task *task, int status)
593{
594 if (!RPC_IS_QUEUED(task))
595 return;
596 spin_lock(&queue->lock);
597 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
598 spin_unlock(&queue->lock);
599}
600
601/*
602 * Wake up the next task on a priority queue.
603 */
604static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
605{
606 struct list_head *q;
607 struct rpc_task *task;
608
609 /*
610 * Service a batch of tasks from a single owner.
611 */
612 q = &queue->tasks[queue->priority];
613 if (!list_empty(q) && --queue->nr) {
614 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
615 goto out;
616 }
617
618 /*
619 * Service the next queue.
620 */
621 do {
622 if (q == &queue->tasks[0])
623 q = &queue->tasks[queue->maxpriority];
624 else
625 q = q - 1;
626 if (!list_empty(q)) {
627 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
628 goto new_queue;
629 }
630 } while (q != &queue->tasks[queue->priority]);
631
632 rpc_reset_waitqueue_priority(queue);
633 return NULL;
634
635new_queue:
636 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
637out:
638 return task;
639}
640
641static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
642{
643 if (RPC_IS_PRIORITY(queue))
644 return __rpc_find_next_queued_priority(queue);
645 if (!list_empty(&queue->tasks[0]))
646 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
647 return NULL;
648}
649
650/*
651 * Wake up the first task on the wait queue.
652 */
653struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
654 struct rpc_wait_queue *queue,
655 bool (*func)(struct rpc_task *, void *), void *data)
656{
657 struct rpc_task *task = NULL;
658
659 dprintk("RPC: wake_up_first(%p \"%s\")\n",
660 queue, rpc_qname(queue));
661 spin_lock(&queue->lock);
662 task = __rpc_find_next_queued(queue);
663 if (task != NULL)
664 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
665 task, func, data);
666 spin_unlock(&queue->lock);
667
668 return task;
669}
670
671/*
672 * Wake up the first task on the wait queue.
673 */
674struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
675 bool (*func)(struct rpc_task *, void *), void *data)
676{
677 return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
678}
679EXPORT_SYMBOL_GPL(rpc_wake_up_first);
680
681static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
682{
683 return true;
684}
685
686/*
687 * Wake up the next task on the wait queue.
688*/
689struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
690{
691 return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
692}
693EXPORT_SYMBOL_GPL(rpc_wake_up_next);
694
695/**
696 * rpc_wake_up - wake up all rpc_tasks
697 * @queue: rpc_wait_queue on which the tasks are sleeping
698 *
699 * Grabs queue->lock
700 */
701void rpc_wake_up(struct rpc_wait_queue *queue)
702{
703 struct list_head *head;
704
705 spin_lock(&queue->lock);
706 head = &queue->tasks[queue->maxpriority];
707 for (;;) {
708 while (!list_empty(head)) {
709 struct rpc_task *task;
710 task = list_first_entry(head,
711 struct rpc_task,
712 u.tk_wait.list);
713 rpc_wake_up_task_queue_locked(queue, task);
714 }
715 if (head == &queue->tasks[0])
716 break;
717 head--;
718 }
719 spin_unlock(&queue->lock);
720}
721EXPORT_SYMBOL_GPL(rpc_wake_up);
722
723/**
724 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
725 * @queue: rpc_wait_queue on which the tasks are sleeping
726 * @status: status value to set
727 *
728 * Grabs queue->lock
729 */
730void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
731{
732 struct list_head *head;
733
734 spin_lock(&queue->lock);
735 head = &queue->tasks[queue->maxpriority];
736 for (;;) {
737 while (!list_empty(head)) {
738 struct rpc_task *task;
739 task = list_first_entry(head,
740 struct rpc_task,
741 u.tk_wait.list);
742 task->tk_status = status;
743 rpc_wake_up_task_queue_locked(queue, task);
744 }
745 if (head == &queue->tasks[0])
746 break;
747 head--;
748 }
749 spin_unlock(&queue->lock);
750}
751EXPORT_SYMBOL_GPL(rpc_wake_up_status);
752
753static void __rpc_queue_timer_fn(struct work_struct *work)
754{
755 struct rpc_wait_queue *queue = container_of(work,
756 struct rpc_wait_queue,
757 timer_list.dwork.work);
758 struct rpc_task *task, *n;
759 unsigned long expires, now, timeo;
760
761 spin_lock(&queue->lock);
762 expires = now = jiffies;
763 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
764 timeo = task->tk_timeout;
765 if (time_after_eq(now, timeo)) {
766 dprintk("RPC: %5u timeout\n", task->tk_pid);
767 task->tk_status = -ETIMEDOUT;
768 rpc_wake_up_task_queue_locked(queue, task);
769 continue;
770 }
771 if (expires == now || time_after(expires, timeo))
772 expires = timeo;
773 }
774 if (!list_empty(&queue->timer_list.list))
775 rpc_set_queue_timer(queue, expires);
776 spin_unlock(&queue->lock);
777}
778
779static void __rpc_atrun(struct rpc_task *task)
780{
781 if (task->tk_status == -ETIMEDOUT)
782 task->tk_status = 0;
783}
784
785/*
786 * Run a task at a later time
787 */
788void rpc_delay(struct rpc_task *task, unsigned long delay)
789{
790 rpc_sleep_on_timeout(&delay_queue, task, __rpc_atrun, jiffies + delay);
791}
792EXPORT_SYMBOL_GPL(rpc_delay);
793
794/*
795 * Helper to call task->tk_ops->rpc_call_prepare
796 */
797void rpc_prepare_task(struct rpc_task *task)
798{
799 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
800}
801
802static void
803rpc_init_task_statistics(struct rpc_task *task)
804{
805 /* Initialize retry counters */
806 task->tk_garb_retry = 2;
807 task->tk_cred_retry = 2;
808 task->tk_rebind_retry = 2;
809
810 /* starting timestamp */
811 task->tk_start = ktime_get();
812}
813
814static void
815rpc_reset_task_statistics(struct rpc_task *task)
816{
817 task->tk_timeouts = 0;
818 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
819 rpc_init_task_statistics(task);
820}
821
822/*
823 * Helper that calls task->tk_ops->rpc_call_done if it exists
824 */
825void rpc_exit_task(struct rpc_task *task)
826{
827 task->tk_action = NULL;
828 if (task->tk_ops->rpc_count_stats)
829 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
830 else if (task->tk_client)
831 rpc_count_iostats(task, task->tk_client->cl_metrics);
832 if (task->tk_ops->rpc_call_done != NULL) {
833 task->tk_ops->rpc_call_done(task, task->tk_calldata);
834 if (task->tk_action != NULL) {
835 /* Always release the RPC slot and buffer memory */
836 xprt_release(task);
837 rpc_reset_task_statistics(task);
838 }
839 }
840}
841
842void rpc_signal_task(struct rpc_task *task)
843{
844 struct rpc_wait_queue *queue;
845
846 if (!RPC_IS_ACTIVATED(task))
847 return;
848 set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
849 smp_mb__after_atomic();
850 queue = READ_ONCE(task->tk_waitqueue);
851 if (queue)
852 rpc_wake_up_queued_task_set_status(queue, task, -ERESTARTSYS);
853}
854
855void rpc_exit(struct rpc_task *task, int status)
856{
857 task->tk_status = status;
858 task->tk_action = rpc_exit_task;
859 rpc_wake_up_queued_task(task->tk_waitqueue, task);
860}
861EXPORT_SYMBOL_GPL(rpc_exit);
862
863void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
864{
865 if (ops->rpc_release != NULL)
866 ops->rpc_release(calldata);
867}
868
869/*
870 * This is the RPC `scheduler' (or rather, the finite state machine).
871 */
872static void __rpc_execute(struct rpc_task *task)
873{
874 struct rpc_wait_queue *queue;
875 int task_is_async = RPC_IS_ASYNC(task);
876 int status = 0;
877
878 dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
879 task->tk_pid, task->tk_flags);
880
881 WARN_ON_ONCE(RPC_IS_QUEUED(task));
882 if (RPC_IS_QUEUED(task))
883 return;
884
885 for (;;) {
886 void (*do_action)(struct rpc_task *);
887
888 /*
889 * Perform the next FSM step or a pending callback.
890 *
891 * tk_action may be NULL if the task has been killed.
892 * In particular, note that rpc_killall_tasks may
893 * do this at any time, so beware when dereferencing.
894 */
895 do_action = task->tk_action;
896 if (task->tk_callback) {
897 do_action = task->tk_callback;
898 task->tk_callback = NULL;
899 }
900 if (!do_action)
901 break;
902 trace_rpc_task_run_action(task, do_action);
903 do_action(task);
904
905 /*
906 * Lockless check for whether task is sleeping or not.
907 */
908 if (!RPC_IS_QUEUED(task))
909 continue;
910
911 /*
912 * Signalled tasks should exit rather than sleep.
913 */
914 if (RPC_SIGNALLED(task)) {
915 task->tk_rpc_status = -ERESTARTSYS;
916 rpc_exit(task, -ERESTARTSYS);
917 }
918
919 /*
920 * The queue->lock protects against races with
921 * rpc_make_runnable().
922 *
923 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
924 * rpc_task, rpc_make_runnable() can assign it to a
925 * different workqueue. We therefore cannot assume that the
926 * rpc_task pointer may still be dereferenced.
927 */
928 queue = task->tk_waitqueue;
929 spin_lock(&queue->lock);
930 if (!RPC_IS_QUEUED(task)) {
931 spin_unlock(&queue->lock);
932 continue;
933 }
934 rpc_clear_running(task);
935 spin_unlock(&queue->lock);
936 if (task_is_async)
937 return;
938
939 /* sync task: sleep here */
940 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
941 status = out_of_line_wait_on_bit(&task->tk_runstate,
942 RPC_TASK_QUEUED, rpc_wait_bit_killable,
943 TASK_KILLABLE);
944 if (status < 0) {
945 /*
946 * When a sync task receives a signal, it exits with
947 * -ERESTARTSYS. In order to catch any callbacks that
948 * clean up after sleeping on some queue, we don't
949 * break the loop here, but go around once more.
950 */
951 dprintk("RPC: %5u got signal\n", task->tk_pid);
952 set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
953 task->tk_rpc_status = -ERESTARTSYS;
954 rpc_exit(task, -ERESTARTSYS);
955 }
956 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
957 }
958
959 dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
960 task->tk_status);
961 /* Release all resources associated with the task */
962 rpc_release_task(task);
963}
964
965/*
966 * User-visible entry point to the scheduler.
967 *
968 * This may be called recursively if e.g. an async NFS task updates
969 * the attributes and finds that dirty pages must be flushed.
970 * NOTE: Upon exit of this function the task is guaranteed to be
971 * released. In particular note that tk_release() will have
972 * been called, so your task memory may have been freed.
973 */
974void rpc_execute(struct rpc_task *task)
975{
976 bool is_async = RPC_IS_ASYNC(task);
977
978 rpc_set_active(task);
979 rpc_make_runnable(rpciod_workqueue, task);
980 if (!is_async)
981 __rpc_execute(task);
982}
983
984static void rpc_async_schedule(struct work_struct *work)
985{
986 unsigned int pflags = memalloc_nofs_save();
987
988 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
989 memalloc_nofs_restore(pflags);
990}
991
992/**
993 * rpc_malloc - allocate RPC buffer resources
994 * @task: RPC task
995 *
996 * A single memory region is allocated, which is split between the
997 * RPC call and RPC reply that this task is being used for. When
998 * this RPC is retired, the memory is released by calling rpc_free.
999 *
1000 * To prevent rpciod from hanging, this allocator never sleeps,
1001 * returning -ENOMEM and suppressing warning if the request cannot
1002 * be serviced immediately. The caller can arrange to sleep in a
1003 * way that is safe for rpciod.
1004 *
1005 * Most requests are 'small' (under 2KiB) and can be serviced from a
1006 * mempool, ensuring that NFS reads and writes can always proceed,
1007 * and that there is good locality of reference for these buffers.
1008 */
1009int rpc_malloc(struct rpc_task *task)
1010{
1011 struct rpc_rqst *rqst = task->tk_rqstp;
1012 size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
1013 struct rpc_buffer *buf;
1014 gfp_t gfp = GFP_NOFS;
1015
1016 if (RPC_IS_SWAPPER(task))
1017 gfp = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
1018
1019 size += sizeof(struct rpc_buffer);
1020 if (size <= RPC_BUFFER_MAXSIZE)
1021 buf = mempool_alloc(rpc_buffer_mempool, gfp);
1022 else
1023 buf = kmalloc(size, gfp);
1024
1025 if (!buf)
1026 return -ENOMEM;
1027
1028 buf->len = size;
1029 dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
1030 task->tk_pid, size, buf);
1031 rqst->rq_buffer = buf->data;
1032 rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1033 return 0;
1034}
1035EXPORT_SYMBOL_GPL(rpc_malloc);
1036
1037/**
1038 * rpc_free - free RPC buffer resources allocated via rpc_malloc
1039 * @task: RPC task
1040 *
1041 */
1042void rpc_free(struct rpc_task *task)
1043{
1044 void *buffer = task->tk_rqstp->rq_buffer;
1045 size_t size;
1046 struct rpc_buffer *buf;
1047
1048 buf = container_of(buffer, struct rpc_buffer, data);
1049 size = buf->len;
1050
1051 dprintk("RPC: freeing buffer of size %zu at %p\n",
1052 size, buf);
1053
1054 if (size <= RPC_BUFFER_MAXSIZE)
1055 mempool_free(buf, rpc_buffer_mempool);
1056 else
1057 kfree(buf);
1058}
1059EXPORT_SYMBOL_GPL(rpc_free);
1060
1061/*
1062 * Creation and deletion of RPC task structures
1063 */
1064static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1065{
1066 memset(task, 0, sizeof(*task));
1067 atomic_set(&task->tk_count, 1);
1068 task->tk_flags = task_setup_data->flags;
1069 task->tk_ops = task_setup_data->callback_ops;
1070 task->tk_calldata = task_setup_data->callback_data;
1071 INIT_LIST_HEAD(&task->tk_task);
1072
1073 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1074 task->tk_owner = current->tgid;
1075
1076 /* Initialize workqueue for async tasks */
1077 task->tk_workqueue = task_setup_data->workqueue;
1078
1079 task->tk_xprt = rpc_task_get_xprt(task_setup_data->rpc_client,
1080 xprt_get(task_setup_data->rpc_xprt));
1081
1082 task->tk_op_cred = get_rpccred(task_setup_data->rpc_op_cred);
1083
1084 if (task->tk_ops->rpc_call_prepare != NULL)
1085 task->tk_action = rpc_prepare_task;
1086
1087 rpc_init_task_statistics(task);
1088
1089 dprintk("RPC: new task initialized, procpid %u\n",
1090 task_pid_nr(current));
1091}
1092
1093static struct rpc_task *
1094rpc_alloc_task(void)
1095{
1096 return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
1097}
1098
1099/*
1100 * Create a new task for the specified client.
1101 */
1102struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1103{
1104 struct rpc_task *task = setup_data->task;
1105 unsigned short flags = 0;
1106
1107 if (task == NULL) {
1108 task = rpc_alloc_task();
1109 flags = RPC_TASK_DYNAMIC;
1110 }
1111
1112 rpc_init_task(task, setup_data);
1113 task->tk_flags |= flags;
1114 dprintk("RPC: allocated task %p\n", task);
1115 return task;
1116}
1117
1118/*
1119 * rpc_free_task - release rpc task and perform cleanups
1120 *
1121 * Note that we free up the rpc_task _after_ rpc_release_calldata()
1122 * in order to work around a workqueue dependency issue.
1123 *
1124 * Tejun Heo states:
1125 * "Workqueue currently considers two work items to be the same if they're
1126 * on the same address and won't execute them concurrently - ie. it
1127 * makes a work item which is queued again while being executed wait
1128 * for the previous execution to complete.
1129 *
1130 * If a work function frees the work item, and then waits for an event
1131 * which should be performed by another work item and *that* work item
1132 * recycles the freed work item, it can create a false dependency loop.
1133 * There really is no reliable way to detect this short of verifying
1134 * every memory free."
1135 *
1136 */
1137static void rpc_free_task(struct rpc_task *task)
1138{
1139 unsigned short tk_flags = task->tk_flags;
1140
1141 put_rpccred(task->tk_op_cred);
1142 rpc_release_calldata(task->tk_ops, task->tk_calldata);
1143
1144 if (tk_flags & RPC_TASK_DYNAMIC) {
1145 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1146 mempool_free(task, rpc_task_mempool);
1147 }
1148}
1149
1150static void rpc_async_release(struct work_struct *work)
1151{
1152 unsigned int pflags = memalloc_nofs_save();
1153
1154 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1155 memalloc_nofs_restore(pflags);
1156}
1157
1158static void rpc_release_resources_task(struct rpc_task *task)
1159{
1160 xprt_release(task);
1161 if (task->tk_msg.rpc_cred) {
1162 put_cred(task->tk_msg.rpc_cred);
1163 task->tk_msg.rpc_cred = NULL;
1164 }
1165 rpc_task_release_client(task);
1166}
1167
1168static void rpc_final_put_task(struct rpc_task *task,
1169 struct workqueue_struct *q)
1170{
1171 if (q != NULL) {
1172 INIT_WORK(&task->u.tk_work, rpc_async_release);
1173 queue_work(q, &task->u.tk_work);
1174 } else
1175 rpc_free_task(task);
1176}
1177
1178static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1179{
1180 if (atomic_dec_and_test(&task->tk_count)) {
1181 rpc_release_resources_task(task);
1182 rpc_final_put_task(task, q);
1183 }
1184}
1185
1186void rpc_put_task(struct rpc_task *task)
1187{
1188 rpc_do_put_task(task, NULL);
1189}
1190EXPORT_SYMBOL_GPL(rpc_put_task);
1191
1192void rpc_put_task_async(struct rpc_task *task)
1193{
1194 rpc_do_put_task(task, task->tk_workqueue);
1195}
1196EXPORT_SYMBOL_GPL(rpc_put_task_async);
1197
1198static void rpc_release_task(struct rpc_task *task)
1199{
1200 dprintk("RPC: %5u release task\n", task->tk_pid);
1201
1202 WARN_ON_ONCE(RPC_IS_QUEUED(task));
1203
1204 rpc_release_resources_task(task);
1205
1206 /*
1207 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1208 * so it should be safe to use task->tk_count as a test for whether
1209 * or not any other processes still hold references to our rpc_task.
1210 */
1211 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1212 /* Wake up anyone who may be waiting for task completion */
1213 if (!rpc_complete_task(task))
1214 return;
1215 } else {
1216 if (!atomic_dec_and_test(&task->tk_count))
1217 return;
1218 }
1219 rpc_final_put_task(task, task->tk_workqueue);
1220}
1221
1222int rpciod_up(void)
1223{
1224 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1225}
1226
1227void rpciod_down(void)
1228{
1229 module_put(THIS_MODULE);
1230}
1231
1232/*
1233 * Start up the rpciod workqueue.
1234 */
1235static int rpciod_start(void)
1236{
1237 struct workqueue_struct *wq;
1238
1239 /*
1240 * Create the rpciod thread and wait for it to start.
1241 */
1242 dprintk("RPC: creating workqueue rpciod\n");
1243 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1244 if (!wq)
1245 goto out_failed;
1246 rpciod_workqueue = wq;
1247 /* Note: highpri because network receive is latency sensitive */
1248 wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1249 if (!wq)
1250 goto free_rpciod;
1251 xprtiod_workqueue = wq;
1252 return 1;
1253free_rpciod:
1254 wq = rpciod_workqueue;
1255 rpciod_workqueue = NULL;
1256 destroy_workqueue(wq);
1257out_failed:
1258 return 0;
1259}
1260
1261static void rpciod_stop(void)
1262{
1263 struct workqueue_struct *wq = NULL;
1264
1265 if (rpciod_workqueue == NULL)
1266 return;
1267 dprintk("RPC: destroying workqueue rpciod\n");
1268
1269 wq = rpciod_workqueue;
1270 rpciod_workqueue = NULL;
1271 destroy_workqueue(wq);
1272 wq = xprtiod_workqueue;
1273 xprtiod_workqueue = NULL;
1274 destroy_workqueue(wq);
1275}
1276
1277void
1278rpc_destroy_mempool(void)
1279{
1280 rpciod_stop();
1281 mempool_destroy(rpc_buffer_mempool);
1282 mempool_destroy(rpc_task_mempool);
1283 kmem_cache_destroy(rpc_task_slabp);
1284 kmem_cache_destroy(rpc_buffer_slabp);
1285 rpc_destroy_wait_queue(&delay_queue);
1286}
1287
1288int
1289rpc_init_mempool(void)
1290{
1291 /*
1292 * The following is not strictly a mempool initialisation,
1293 * but there is no harm in doing it here
1294 */
1295 rpc_init_wait_queue(&delay_queue, "delayq");
1296 if (!rpciod_start())
1297 goto err_nomem;
1298
1299 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1300 sizeof(struct rpc_task),
1301 0, SLAB_HWCACHE_ALIGN,
1302 NULL);
1303 if (!rpc_task_slabp)
1304 goto err_nomem;
1305 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1306 RPC_BUFFER_MAXSIZE,
1307 0, SLAB_HWCACHE_ALIGN,
1308 NULL);
1309 if (!rpc_buffer_slabp)
1310 goto err_nomem;
1311 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1312 rpc_task_slabp);
1313 if (!rpc_task_mempool)
1314 goto err_nomem;
1315 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1316 rpc_buffer_slabp);
1317 if (!rpc_buffer_mempool)
1318 goto err_nomem;
1319 return 0;
1320err_nomem:
1321 rpc_destroy_mempool();
1322 return -ENOMEM;
1323}
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}