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#define CREATE_TRACE_POINTS
31#include <trace/events/sunrpc.h>
32
33/*
34 * RPC slabs and memory pools
35 */
36#define RPC_BUFFER_MAXSIZE (2048)
37#define RPC_BUFFER_POOLSIZE (8)
38#define RPC_TASK_POOLSIZE (8)
39static struct kmem_cache *rpc_task_slabp __read_mostly;
40static struct kmem_cache *rpc_buffer_slabp __read_mostly;
41static mempool_t *rpc_task_mempool __read_mostly;
42static mempool_t *rpc_buffer_mempool __read_mostly;
43
44static void rpc_async_schedule(struct work_struct *);
45static void rpc_release_task(struct rpc_task *task);
46static void __rpc_queue_timer_fn(struct work_struct *);
47
48/*
49 * RPC tasks sit here while waiting for conditions to improve.
50 */
51static struct rpc_wait_queue delay_queue;
52
53/*
54 * rpciod-related stuff
55 */
56struct workqueue_struct *rpciod_workqueue __read_mostly;
57struct workqueue_struct *xprtiod_workqueue __read_mostly;
58EXPORT_SYMBOL_GPL(xprtiod_workqueue);
59
60gfp_t rpc_task_gfp_mask(void)
61{
62 if (current->flags & PF_WQ_WORKER)
63 return GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
64 return GFP_KERNEL;
65}
66EXPORT_SYMBOL_GPL(rpc_task_gfp_mask);
67
68bool rpc_task_set_rpc_status(struct rpc_task *task, int rpc_status)
69{
70 if (cmpxchg(&task->tk_rpc_status, 0, rpc_status) == 0)
71 return true;
72 return false;
73}
74
75unsigned long
76rpc_task_timeout(const struct rpc_task *task)
77{
78 unsigned long timeout = READ_ONCE(task->tk_timeout);
79
80 if (timeout != 0) {
81 unsigned long now = jiffies;
82 if (time_before(now, timeout))
83 return timeout - now;
84 }
85 return 0;
86}
87EXPORT_SYMBOL_GPL(rpc_task_timeout);
88
89/*
90 * Disable the timer for a given RPC task. Should be called with
91 * queue->lock and bh_disabled in order to avoid races within
92 * rpc_run_timer().
93 */
94static void
95__rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
96{
97 if (list_empty(&task->u.tk_wait.timer_list))
98 return;
99 task->tk_timeout = 0;
100 list_del(&task->u.tk_wait.timer_list);
101 if (list_empty(&queue->timer_list.list))
102 cancel_delayed_work(&queue->timer_list.dwork);
103}
104
105static void
106rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
107{
108 unsigned long now = jiffies;
109 queue->timer_list.expires = expires;
110 if (time_before_eq(expires, now))
111 expires = 0;
112 else
113 expires -= now;
114 mod_delayed_work(rpciod_workqueue, &queue->timer_list.dwork, expires);
115}
116
117/*
118 * Set up a timer for the current task.
119 */
120static void
121__rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task,
122 unsigned long timeout)
123{
124 task->tk_timeout = timeout;
125 if (list_empty(&queue->timer_list.list) || time_before(timeout, queue->timer_list.expires))
126 rpc_set_queue_timer(queue, timeout);
127 list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
128}
129
130static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
131{
132 if (queue->priority != priority) {
133 queue->priority = priority;
134 queue->nr = 1U << priority;
135 }
136}
137
138static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
139{
140 rpc_set_waitqueue_priority(queue, queue->maxpriority);
141}
142
143/*
144 * Add a request to a queue list
145 */
146static void
147__rpc_list_enqueue_task(struct list_head *q, struct rpc_task *task)
148{
149 struct rpc_task *t;
150
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.links,
154 &t->u.tk_wait.links);
155 /* Cache the queue head in task->u.tk_wait.list */
156 task->u.tk_wait.list.next = q;
157 task->u.tk_wait.list.prev = NULL;
158 return;
159 }
160 }
161 INIT_LIST_HEAD(&task->u.tk_wait.links);
162 list_add_tail(&task->u.tk_wait.list, q);
163}
164
165/*
166 * Remove request from a queue list
167 */
168static void
169__rpc_list_dequeue_task(struct rpc_task *task)
170{
171 struct list_head *q;
172 struct rpc_task *t;
173
174 if (task->u.tk_wait.list.prev == NULL) {
175 list_del(&task->u.tk_wait.links);
176 return;
177 }
178 if (!list_empty(&task->u.tk_wait.links)) {
179 t = list_first_entry(&task->u.tk_wait.links,
180 struct rpc_task,
181 u.tk_wait.links);
182 /* Assume __rpc_list_enqueue_task() cached the queue head */
183 q = t->u.tk_wait.list.next;
184 list_add_tail(&t->u.tk_wait.list, q);
185 list_del(&task->u.tk_wait.links);
186 }
187 list_del(&task->u.tk_wait.list);
188}
189
190/*
191 * Add new request to a priority queue.
192 */
193static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
194 struct rpc_task *task,
195 unsigned char queue_priority)
196{
197 if (unlikely(queue_priority > queue->maxpriority))
198 queue_priority = queue->maxpriority;
199 __rpc_list_enqueue_task(&queue->tasks[queue_priority], task);
200}
201
202/*
203 * Add new request to wait queue.
204 */
205static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
206 struct rpc_task *task,
207 unsigned char queue_priority)
208{
209 INIT_LIST_HEAD(&task->u.tk_wait.timer_list);
210 if (RPC_IS_PRIORITY(queue))
211 __rpc_add_wait_queue_priority(queue, task, queue_priority);
212 else
213 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
214 task->tk_waitqueue = queue;
215 queue->qlen++;
216 /* barrier matches the read in rpc_wake_up_task_queue_locked() */
217 smp_wmb();
218 rpc_set_queued(task);
219}
220
221/*
222 * Remove request from a priority queue.
223 */
224static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
225{
226 __rpc_list_dequeue_task(task);
227}
228
229/*
230 * Remove request from queue.
231 * Note: must be called with spin lock held.
232 */
233static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
234{
235 __rpc_disable_timer(queue, task);
236 if (RPC_IS_PRIORITY(queue))
237 __rpc_remove_wait_queue_priority(task);
238 else
239 list_del(&task->u.tk_wait.list);
240 queue->qlen--;
241}
242
243static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
244{
245 int i;
246
247 spin_lock_init(&queue->lock);
248 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
249 INIT_LIST_HEAD(&queue->tasks[i]);
250 queue->maxpriority = nr_queues - 1;
251 rpc_reset_waitqueue_priority(queue);
252 queue->qlen = 0;
253 queue->timer_list.expires = 0;
254 INIT_DELAYED_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
255 INIT_LIST_HEAD(&queue->timer_list.list);
256 rpc_assign_waitqueue_name(queue, qname);
257}
258
259void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
260{
261 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
262}
263EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
264
265void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
266{
267 __rpc_init_priority_wait_queue(queue, qname, 1);
268}
269EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
270
271void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
272{
273 cancel_delayed_work_sync(&queue->timer_list.dwork);
274}
275EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
276
277static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
278{
279 schedule();
280 if (signal_pending_state(mode, current))
281 return -ERESTARTSYS;
282 return 0;
283}
284
285#if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
286static void rpc_task_set_debuginfo(struct rpc_task *task)
287{
288 struct rpc_clnt *clnt = task->tk_client;
289
290 /* Might be a task carrying a reverse-direction operation */
291 if (!clnt) {
292 static atomic_t rpc_pid;
293
294 task->tk_pid = atomic_inc_return(&rpc_pid);
295 return;
296 }
297
298 task->tk_pid = atomic_inc_return(&clnt->cl_pid);
299}
300#else
301static inline void rpc_task_set_debuginfo(struct rpc_task *task)
302{
303}
304#endif
305
306static void rpc_set_active(struct rpc_task *task)
307{
308 rpc_task_set_debuginfo(task);
309 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
310 trace_rpc_task_begin(task, NULL);
311}
312
313/*
314 * Mark an RPC call as having completed by clearing the 'active' bit
315 * and then waking up all tasks that were sleeping.
316 */
317static int rpc_complete_task(struct rpc_task *task)
318{
319 void *m = &task->tk_runstate;
320 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
321 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
322 unsigned long flags;
323 int ret;
324
325 trace_rpc_task_complete(task, NULL);
326
327 spin_lock_irqsave(&wq->lock, flags);
328 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
329 ret = atomic_dec_and_test(&task->tk_count);
330 if (waitqueue_active(wq))
331 __wake_up_locked_key(wq, TASK_NORMAL, &k);
332 spin_unlock_irqrestore(&wq->lock, flags);
333 return ret;
334}
335
336/*
337 * Allow callers to wait for completion of an RPC call
338 *
339 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
340 * to enforce taking of the wq->lock and hence avoid races with
341 * rpc_complete_task().
342 */
343int rpc_wait_for_completion_task(struct rpc_task *task)
344{
345 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
346 rpc_wait_bit_killable, TASK_KILLABLE|TASK_FREEZABLE_UNSAFE);
347}
348EXPORT_SYMBOL_GPL(rpc_wait_for_completion_task);
349
350/*
351 * Make an RPC task runnable.
352 *
353 * Note: If the task is ASYNC, and is being made runnable after sitting on an
354 * rpc_wait_queue, this must be called with the queue spinlock held to protect
355 * the wait queue operation.
356 * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
357 * which is needed to ensure that __rpc_execute() doesn't loop (due to the
358 * lockless RPC_IS_QUEUED() test) before we've had a chance to test
359 * the RPC_TASK_RUNNING flag.
360 */
361static void rpc_make_runnable(struct workqueue_struct *wq,
362 struct rpc_task *task)
363{
364 bool need_wakeup = !rpc_test_and_set_running(task);
365
366 rpc_clear_queued(task);
367 if (!need_wakeup)
368 return;
369 if (RPC_IS_ASYNC(task)) {
370 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
371 queue_work(wq, &task->u.tk_work);
372 } else
373 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
374}
375
376/*
377 * Prepare for sleeping on a wait queue.
378 * By always appending tasks to the list we ensure FIFO behavior.
379 * NB: An RPC task will only receive interrupt-driven events as long
380 * as it's on a wait queue.
381 */
382static void __rpc_do_sleep_on_priority(struct rpc_wait_queue *q,
383 struct rpc_task *task,
384 unsigned char queue_priority)
385{
386 trace_rpc_task_sleep(task, q);
387
388 __rpc_add_wait_queue(q, task, queue_priority);
389}
390
391static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
392 struct rpc_task *task,
393 unsigned char queue_priority)
394{
395 if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
396 return;
397 __rpc_do_sleep_on_priority(q, task, queue_priority);
398}
399
400static void __rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
401 struct rpc_task *task, unsigned long timeout,
402 unsigned char queue_priority)
403{
404 if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
405 return;
406 if (time_is_after_jiffies(timeout)) {
407 __rpc_do_sleep_on_priority(q, task, queue_priority);
408 __rpc_add_timer(q, task, timeout);
409 } else
410 task->tk_status = -ETIMEDOUT;
411}
412
413static void rpc_set_tk_callback(struct rpc_task *task, rpc_action action)
414{
415 if (action && !WARN_ON_ONCE(task->tk_callback != NULL))
416 task->tk_callback = action;
417}
418
419static bool rpc_sleep_check_activated(struct rpc_task *task)
420{
421 /* We shouldn't ever put an inactive task to sleep */
422 if (WARN_ON_ONCE(!RPC_IS_ACTIVATED(task))) {
423 task->tk_status = -EIO;
424 rpc_put_task_async(task);
425 return false;
426 }
427 return true;
428}
429
430void rpc_sleep_on_timeout(struct rpc_wait_queue *q, struct rpc_task *task,
431 rpc_action action, unsigned long timeout)
432{
433 if (!rpc_sleep_check_activated(task))
434 return;
435
436 rpc_set_tk_callback(task, action);
437
438 /*
439 * Protect the queue operations.
440 */
441 spin_lock(&q->lock);
442 __rpc_sleep_on_priority_timeout(q, task, timeout, task->tk_priority);
443 spin_unlock(&q->lock);
444}
445EXPORT_SYMBOL_GPL(rpc_sleep_on_timeout);
446
447void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
448 rpc_action action)
449{
450 if (!rpc_sleep_check_activated(task))
451 return;
452
453 rpc_set_tk_callback(task, action);
454
455 WARN_ON_ONCE(task->tk_timeout != 0);
456 /*
457 * Protect the queue operations.
458 */
459 spin_lock(&q->lock);
460 __rpc_sleep_on_priority(q, task, task->tk_priority);
461 spin_unlock(&q->lock);
462}
463EXPORT_SYMBOL_GPL(rpc_sleep_on);
464
465void rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
466 struct rpc_task *task, unsigned long timeout, int priority)
467{
468 if (!rpc_sleep_check_activated(task))
469 return;
470
471 priority -= RPC_PRIORITY_LOW;
472 /*
473 * Protect the queue operations.
474 */
475 spin_lock(&q->lock);
476 __rpc_sleep_on_priority_timeout(q, task, timeout, priority);
477 spin_unlock(&q->lock);
478}
479EXPORT_SYMBOL_GPL(rpc_sleep_on_priority_timeout);
480
481void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
482 int priority)
483{
484 if (!rpc_sleep_check_activated(task))
485 return;
486
487 WARN_ON_ONCE(task->tk_timeout != 0);
488 priority -= RPC_PRIORITY_LOW;
489 /*
490 * Protect the queue operations.
491 */
492 spin_lock(&q->lock);
493 __rpc_sleep_on_priority(q, task, priority);
494 spin_unlock(&q->lock);
495}
496EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
497
498/**
499 * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
500 * @wq: workqueue on which to run task
501 * @queue: wait queue
502 * @task: task to be woken up
503 *
504 * Caller must hold queue->lock, and have cleared the task queued flag.
505 */
506static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
507 struct rpc_wait_queue *queue,
508 struct rpc_task *task)
509{
510 /* Has the task been executed yet? If not, we cannot wake it up! */
511 if (!RPC_IS_ACTIVATED(task)) {
512 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
513 return;
514 }
515
516 trace_rpc_task_wakeup(task, queue);
517
518 __rpc_remove_wait_queue(queue, task);
519
520 rpc_make_runnable(wq, task);
521}
522
523/*
524 * Wake up a queued task while the queue lock is being held
525 */
526static struct rpc_task *
527rpc_wake_up_task_on_wq_queue_action_locked(struct workqueue_struct *wq,
528 struct rpc_wait_queue *queue, struct rpc_task *task,
529 bool (*action)(struct rpc_task *, void *), void *data)
530{
531 if (RPC_IS_QUEUED(task)) {
532 smp_rmb();
533 if (task->tk_waitqueue == queue) {
534 if (action == NULL || action(task, data)) {
535 __rpc_do_wake_up_task_on_wq(wq, queue, task);
536 return task;
537 }
538 }
539 }
540 return NULL;
541}
542
543/*
544 * Wake up a queued task while the queue lock is being held
545 */
546static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue,
547 struct rpc_task *task)
548{
549 rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
550 task, NULL, NULL);
551}
552
553/*
554 * Wake up a task on a specific queue
555 */
556void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
557{
558 if (!RPC_IS_QUEUED(task))
559 return;
560 spin_lock(&queue->lock);
561 rpc_wake_up_task_queue_locked(queue, task);
562 spin_unlock(&queue->lock);
563}
564EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
565
566static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
567{
568 task->tk_status = *(int *)status;
569 return true;
570}
571
572static void
573rpc_wake_up_task_queue_set_status_locked(struct rpc_wait_queue *queue,
574 struct rpc_task *task, int status)
575{
576 rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
577 task, rpc_task_action_set_status, &status);
578}
579
580/**
581 * rpc_wake_up_queued_task_set_status - wake up a task and set task->tk_status
582 * @queue: pointer to rpc_wait_queue
583 * @task: pointer to rpc_task
584 * @status: integer error value
585 *
586 * If @task is queued on @queue, then it is woken up, and @task->tk_status is
587 * set to the value of @status.
588 */
589void
590rpc_wake_up_queued_task_set_status(struct rpc_wait_queue *queue,
591 struct rpc_task *task, int status)
592{
593 if (!RPC_IS_QUEUED(task))
594 return;
595 spin_lock(&queue->lock);
596 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
597 spin_unlock(&queue->lock);
598}
599
600/*
601 * Wake up the next task on a priority queue.
602 */
603static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
604{
605 struct list_head *q;
606 struct rpc_task *task;
607
608 /*
609 * Service the privileged queue.
610 */
611 q = &queue->tasks[RPC_NR_PRIORITY - 1];
612 if (queue->maxpriority > RPC_PRIORITY_PRIVILEGED && !list_empty(q)) {
613 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
614 goto out;
615 }
616
617 /*
618 * Service a batch of tasks from a single owner.
619 */
620 q = &queue->tasks[queue->priority];
621 if (!list_empty(q) && queue->nr) {
622 queue->nr--;
623 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
624 goto out;
625 }
626
627 /*
628 * Service the next queue.
629 */
630 do {
631 if (q == &queue->tasks[0])
632 q = &queue->tasks[queue->maxpriority];
633 else
634 q = q - 1;
635 if (!list_empty(q)) {
636 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
637 goto new_queue;
638 }
639 } while (q != &queue->tasks[queue->priority]);
640
641 rpc_reset_waitqueue_priority(queue);
642 return NULL;
643
644new_queue:
645 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
646out:
647 return task;
648}
649
650static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
651{
652 if (RPC_IS_PRIORITY(queue))
653 return __rpc_find_next_queued_priority(queue);
654 if (!list_empty(&queue->tasks[0]))
655 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
656 return NULL;
657}
658
659/*
660 * Wake up the first task on the wait queue.
661 */
662struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
663 struct rpc_wait_queue *queue,
664 bool (*func)(struct rpc_task *, void *), void *data)
665{
666 struct rpc_task *task = NULL;
667
668 spin_lock(&queue->lock);
669 task = __rpc_find_next_queued(queue);
670 if (task != NULL)
671 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
672 task, func, data);
673 spin_unlock(&queue->lock);
674
675 return task;
676}
677
678/*
679 * Wake up the first task on the wait queue.
680 */
681struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
682 bool (*func)(struct rpc_task *, void *), void *data)
683{
684 return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
685}
686EXPORT_SYMBOL_GPL(rpc_wake_up_first);
687
688static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
689{
690 return true;
691}
692
693/*
694 * Wake up the next task on the wait queue.
695*/
696struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
697{
698 return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
699}
700EXPORT_SYMBOL_GPL(rpc_wake_up_next);
701
702/**
703 * rpc_wake_up_locked - wake up all rpc_tasks
704 * @queue: rpc_wait_queue on which the tasks are sleeping
705 *
706 */
707static void rpc_wake_up_locked(struct rpc_wait_queue *queue)
708{
709 struct rpc_task *task;
710
711 for (;;) {
712 task = __rpc_find_next_queued(queue);
713 if (task == NULL)
714 break;
715 rpc_wake_up_task_queue_locked(queue, task);
716 }
717}
718
719/**
720 * rpc_wake_up - wake up all rpc_tasks
721 * @queue: rpc_wait_queue on which the tasks are sleeping
722 *
723 * Grabs queue->lock
724 */
725void rpc_wake_up(struct rpc_wait_queue *queue)
726{
727 spin_lock(&queue->lock);
728 rpc_wake_up_locked(queue);
729 spin_unlock(&queue->lock);
730}
731EXPORT_SYMBOL_GPL(rpc_wake_up);
732
733/**
734 * rpc_wake_up_status_locked - wake up all rpc_tasks and set their status value.
735 * @queue: rpc_wait_queue on which the tasks are sleeping
736 * @status: status value to set
737 */
738static void rpc_wake_up_status_locked(struct rpc_wait_queue *queue, int status)
739{
740 struct rpc_task *task;
741
742 for (;;) {
743 task = __rpc_find_next_queued(queue);
744 if (task == NULL)
745 break;
746 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
747 }
748}
749
750/**
751 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
752 * @queue: rpc_wait_queue on which the tasks are sleeping
753 * @status: status value to set
754 *
755 * Grabs queue->lock
756 */
757void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
758{
759 spin_lock(&queue->lock);
760 rpc_wake_up_status_locked(queue, status);
761 spin_unlock(&queue->lock);
762}
763EXPORT_SYMBOL_GPL(rpc_wake_up_status);
764
765static void __rpc_queue_timer_fn(struct work_struct *work)
766{
767 struct rpc_wait_queue *queue = container_of(work,
768 struct rpc_wait_queue,
769 timer_list.dwork.work);
770 struct rpc_task *task, *n;
771 unsigned long expires, now, timeo;
772
773 spin_lock(&queue->lock);
774 expires = now = jiffies;
775 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
776 timeo = task->tk_timeout;
777 if (time_after_eq(now, timeo)) {
778 trace_rpc_task_timeout(task, task->tk_action);
779 task->tk_status = -ETIMEDOUT;
780 rpc_wake_up_task_queue_locked(queue, task);
781 continue;
782 }
783 if (expires == now || time_after(expires, timeo))
784 expires = timeo;
785 }
786 if (!list_empty(&queue->timer_list.list))
787 rpc_set_queue_timer(queue, expires);
788 spin_unlock(&queue->lock);
789}
790
791static void __rpc_atrun(struct rpc_task *task)
792{
793 if (task->tk_status == -ETIMEDOUT)
794 task->tk_status = 0;
795}
796
797/*
798 * Run a task at a later time
799 */
800void rpc_delay(struct rpc_task *task, unsigned long delay)
801{
802 rpc_sleep_on_timeout(&delay_queue, task, __rpc_atrun, jiffies + delay);
803}
804EXPORT_SYMBOL_GPL(rpc_delay);
805
806/*
807 * Helper to call task->tk_ops->rpc_call_prepare
808 */
809void rpc_prepare_task(struct rpc_task *task)
810{
811 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
812}
813
814static void
815rpc_init_task_statistics(struct rpc_task *task)
816{
817 /* Initialize retry counters */
818 task->tk_garb_retry = 2;
819 task->tk_cred_retry = 2;
820 task->tk_rebind_retry = 2;
821
822 /* starting timestamp */
823 task->tk_start = ktime_get();
824}
825
826static void
827rpc_reset_task_statistics(struct rpc_task *task)
828{
829 task->tk_timeouts = 0;
830 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
831 rpc_init_task_statistics(task);
832}
833
834/*
835 * Helper that calls task->tk_ops->rpc_call_done if it exists
836 */
837void rpc_exit_task(struct rpc_task *task)
838{
839 trace_rpc_task_end(task, task->tk_action);
840 task->tk_action = NULL;
841 if (task->tk_ops->rpc_count_stats)
842 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
843 else if (task->tk_client)
844 rpc_count_iostats(task, task->tk_client->cl_metrics);
845 if (task->tk_ops->rpc_call_done != NULL) {
846 trace_rpc_task_call_done(task, task->tk_ops->rpc_call_done);
847 task->tk_ops->rpc_call_done(task, task->tk_calldata);
848 if (task->tk_action != NULL) {
849 /* Always release the RPC slot and buffer memory */
850 xprt_release(task);
851 rpc_reset_task_statistics(task);
852 }
853 }
854}
855
856void rpc_signal_task(struct rpc_task *task)
857{
858 struct rpc_wait_queue *queue;
859
860 if (!RPC_IS_ACTIVATED(task))
861 return;
862
863 if (!rpc_task_set_rpc_status(task, -ERESTARTSYS))
864 return;
865 trace_rpc_task_signalled(task, task->tk_action);
866 set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
867 smp_mb__after_atomic();
868 queue = READ_ONCE(task->tk_waitqueue);
869 if (queue)
870 rpc_wake_up_queued_task(queue, task);
871}
872
873void rpc_task_try_cancel(struct rpc_task *task, int error)
874{
875 struct rpc_wait_queue *queue;
876
877 if (!rpc_task_set_rpc_status(task, error))
878 return;
879 queue = READ_ONCE(task->tk_waitqueue);
880 if (queue)
881 rpc_wake_up_queued_task(queue, task);
882}
883
884void rpc_exit(struct rpc_task *task, int status)
885{
886 task->tk_status = status;
887 task->tk_action = rpc_exit_task;
888 rpc_wake_up_queued_task(task->tk_waitqueue, task);
889}
890EXPORT_SYMBOL_GPL(rpc_exit);
891
892void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
893{
894 if (ops->rpc_release != NULL)
895 ops->rpc_release(calldata);
896}
897
898static bool xprt_needs_memalloc(struct rpc_xprt *xprt, struct rpc_task *tk)
899{
900 if (!xprt)
901 return false;
902 if (!atomic_read(&xprt->swapper))
903 return false;
904 return test_bit(XPRT_LOCKED, &xprt->state) && xprt->snd_task == tk;
905}
906
907/*
908 * This is the RPC `scheduler' (or rather, the finite state machine).
909 */
910static void __rpc_execute(struct rpc_task *task)
911{
912 struct rpc_wait_queue *queue;
913 int task_is_async = RPC_IS_ASYNC(task);
914 int status = 0;
915 unsigned long pflags = current->flags;
916
917 WARN_ON_ONCE(RPC_IS_QUEUED(task));
918 if (RPC_IS_QUEUED(task))
919 return;
920
921 for (;;) {
922 void (*do_action)(struct rpc_task *);
923
924 /*
925 * Perform the next FSM step or a pending callback.
926 *
927 * tk_action may be NULL if the task has been killed.
928 */
929 do_action = task->tk_action;
930 /* Tasks with an RPC error status should exit */
931 if (do_action != rpc_exit_task &&
932 (status = READ_ONCE(task->tk_rpc_status)) != 0) {
933 task->tk_status = status;
934 if (do_action != NULL)
935 do_action = rpc_exit_task;
936 }
937 /* Callbacks override all actions */
938 if (task->tk_callback) {
939 do_action = task->tk_callback;
940 task->tk_callback = NULL;
941 }
942 if (!do_action)
943 break;
944 if (RPC_IS_SWAPPER(task) ||
945 xprt_needs_memalloc(task->tk_xprt, task))
946 current->flags |= PF_MEMALLOC;
947
948 trace_rpc_task_run_action(task, do_action);
949 do_action(task);
950
951 /*
952 * Lockless check for whether task is sleeping or not.
953 */
954 if (!RPC_IS_QUEUED(task)) {
955 cond_resched();
956 continue;
957 }
958
959 /*
960 * The queue->lock protects against races with
961 * rpc_make_runnable().
962 *
963 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
964 * rpc_task, rpc_make_runnable() can assign it to a
965 * different workqueue. We therefore cannot assume that the
966 * rpc_task pointer may still be dereferenced.
967 */
968 queue = task->tk_waitqueue;
969 spin_lock(&queue->lock);
970 if (!RPC_IS_QUEUED(task)) {
971 spin_unlock(&queue->lock);
972 continue;
973 }
974 /* Wake up any task that has an exit status */
975 if (READ_ONCE(task->tk_rpc_status) != 0) {
976 rpc_wake_up_task_queue_locked(queue, task);
977 spin_unlock(&queue->lock);
978 continue;
979 }
980 rpc_clear_running(task);
981 spin_unlock(&queue->lock);
982 if (task_is_async)
983 goto out;
984
985 /* sync task: sleep here */
986 trace_rpc_task_sync_sleep(task, task->tk_action);
987 status = out_of_line_wait_on_bit(&task->tk_runstate,
988 RPC_TASK_QUEUED, rpc_wait_bit_killable,
989 TASK_KILLABLE|TASK_FREEZABLE);
990 if (status < 0) {
991 /*
992 * When a sync task receives a signal, it exits with
993 * -ERESTARTSYS. In order to catch any callbacks that
994 * clean up after sleeping on some queue, we don't
995 * break the loop here, but go around once more.
996 */
997 rpc_signal_task(task);
998 }
999 trace_rpc_task_sync_wake(task, task->tk_action);
1000 }
1001
1002 /* Release all resources associated with the task */
1003 rpc_release_task(task);
1004out:
1005 current_restore_flags(pflags, PF_MEMALLOC);
1006}
1007
1008/*
1009 * User-visible entry point to the scheduler.
1010 *
1011 * This may be called recursively if e.g. an async NFS task updates
1012 * the attributes and finds that dirty pages must be flushed.
1013 * NOTE: Upon exit of this function the task is guaranteed to be
1014 * released. In particular note that tk_release() will have
1015 * been called, so your task memory may have been freed.
1016 */
1017void rpc_execute(struct rpc_task *task)
1018{
1019 bool is_async = RPC_IS_ASYNC(task);
1020
1021 rpc_set_active(task);
1022 rpc_make_runnable(rpciod_workqueue, task);
1023 if (!is_async) {
1024 unsigned int pflags = memalloc_nofs_save();
1025 __rpc_execute(task);
1026 memalloc_nofs_restore(pflags);
1027 }
1028}
1029
1030static void rpc_async_schedule(struct work_struct *work)
1031{
1032 unsigned int pflags = memalloc_nofs_save();
1033
1034 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
1035 memalloc_nofs_restore(pflags);
1036}
1037
1038/**
1039 * rpc_malloc - allocate RPC buffer resources
1040 * @task: RPC task
1041 *
1042 * A single memory region is allocated, which is split between the
1043 * RPC call and RPC reply that this task is being used for. When
1044 * this RPC is retired, the memory is released by calling rpc_free.
1045 *
1046 * To prevent rpciod from hanging, this allocator never sleeps,
1047 * returning -ENOMEM and suppressing warning if the request cannot
1048 * be serviced immediately. The caller can arrange to sleep in a
1049 * way that is safe for rpciod.
1050 *
1051 * Most requests are 'small' (under 2KiB) and can be serviced from a
1052 * mempool, ensuring that NFS reads and writes can always proceed,
1053 * and that there is good locality of reference for these buffers.
1054 */
1055int rpc_malloc(struct rpc_task *task)
1056{
1057 struct rpc_rqst *rqst = task->tk_rqstp;
1058 size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
1059 struct rpc_buffer *buf;
1060 gfp_t gfp = rpc_task_gfp_mask();
1061
1062 size += sizeof(struct rpc_buffer);
1063 if (size <= RPC_BUFFER_MAXSIZE) {
1064 buf = kmem_cache_alloc(rpc_buffer_slabp, gfp);
1065 /* Reach for the mempool if dynamic allocation fails */
1066 if (!buf && RPC_IS_ASYNC(task))
1067 buf = mempool_alloc(rpc_buffer_mempool, GFP_NOWAIT);
1068 } else
1069 buf = kmalloc(size, gfp);
1070
1071 if (!buf)
1072 return -ENOMEM;
1073
1074 buf->len = size;
1075 rqst->rq_buffer = buf->data;
1076 rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1077 return 0;
1078}
1079EXPORT_SYMBOL_GPL(rpc_malloc);
1080
1081/**
1082 * rpc_free - free RPC buffer resources allocated via rpc_malloc
1083 * @task: RPC task
1084 *
1085 */
1086void rpc_free(struct rpc_task *task)
1087{
1088 void *buffer = task->tk_rqstp->rq_buffer;
1089 size_t size;
1090 struct rpc_buffer *buf;
1091
1092 buf = container_of(buffer, struct rpc_buffer, data);
1093 size = buf->len;
1094
1095 if (size <= RPC_BUFFER_MAXSIZE)
1096 mempool_free(buf, rpc_buffer_mempool);
1097 else
1098 kfree(buf);
1099}
1100EXPORT_SYMBOL_GPL(rpc_free);
1101
1102/*
1103 * Creation and deletion of RPC task structures
1104 */
1105static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1106{
1107 memset(task, 0, sizeof(*task));
1108 atomic_set(&task->tk_count, 1);
1109 task->tk_flags = task_setup_data->flags;
1110 task->tk_ops = task_setup_data->callback_ops;
1111 task->tk_calldata = task_setup_data->callback_data;
1112 INIT_LIST_HEAD(&task->tk_task);
1113
1114 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1115 task->tk_owner = current->tgid;
1116
1117 /* Initialize workqueue for async tasks */
1118 task->tk_workqueue = task_setup_data->workqueue;
1119
1120 task->tk_xprt = rpc_task_get_xprt(task_setup_data->rpc_client,
1121 xprt_get(task_setup_data->rpc_xprt));
1122
1123 task->tk_op_cred = get_rpccred(task_setup_data->rpc_op_cred);
1124
1125 if (task->tk_ops->rpc_call_prepare != NULL)
1126 task->tk_action = rpc_prepare_task;
1127
1128 rpc_init_task_statistics(task);
1129}
1130
1131static struct rpc_task *rpc_alloc_task(void)
1132{
1133 struct rpc_task *task;
1134
1135 task = kmem_cache_alloc(rpc_task_slabp, rpc_task_gfp_mask());
1136 if (task)
1137 return task;
1138 return mempool_alloc(rpc_task_mempool, GFP_NOWAIT);
1139}
1140
1141/*
1142 * Create a new task for the specified client.
1143 */
1144struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1145{
1146 struct rpc_task *task = setup_data->task;
1147 unsigned short flags = 0;
1148
1149 if (task == NULL) {
1150 task = rpc_alloc_task();
1151 if (task == NULL) {
1152 rpc_release_calldata(setup_data->callback_ops,
1153 setup_data->callback_data);
1154 return ERR_PTR(-ENOMEM);
1155 }
1156 flags = RPC_TASK_DYNAMIC;
1157 }
1158
1159 rpc_init_task(task, setup_data);
1160 task->tk_flags |= flags;
1161 return task;
1162}
1163
1164/*
1165 * rpc_free_task - release rpc task and perform cleanups
1166 *
1167 * Note that we free up the rpc_task _after_ rpc_release_calldata()
1168 * in order to work around a workqueue dependency issue.
1169 *
1170 * Tejun Heo states:
1171 * "Workqueue currently considers two work items to be the same if they're
1172 * on the same address and won't execute them concurrently - ie. it
1173 * makes a work item which is queued again while being executed wait
1174 * for the previous execution to complete.
1175 *
1176 * If a work function frees the work item, and then waits for an event
1177 * which should be performed by another work item and *that* work item
1178 * recycles the freed work item, it can create a false dependency loop.
1179 * There really is no reliable way to detect this short of verifying
1180 * every memory free."
1181 *
1182 */
1183static void rpc_free_task(struct rpc_task *task)
1184{
1185 unsigned short tk_flags = task->tk_flags;
1186
1187 put_rpccred(task->tk_op_cred);
1188 rpc_release_calldata(task->tk_ops, task->tk_calldata);
1189
1190 if (tk_flags & RPC_TASK_DYNAMIC)
1191 mempool_free(task, rpc_task_mempool);
1192}
1193
1194static void rpc_async_release(struct work_struct *work)
1195{
1196 unsigned int pflags = memalloc_nofs_save();
1197
1198 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1199 memalloc_nofs_restore(pflags);
1200}
1201
1202static void rpc_release_resources_task(struct rpc_task *task)
1203{
1204 xprt_release(task);
1205 if (task->tk_msg.rpc_cred) {
1206 if (!(task->tk_flags & RPC_TASK_CRED_NOREF))
1207 put_cred(task->tk_msg.rpc_cred);
1208 task->tk_msg.rpc_cred = NULL;
1209 }
1210 rpc_task_release_client(task);
1211}
1212
1213static void rpc_final_put_task(struct rpc_task *task,
1214 struct workqueue_struct *q)
1215{
1216 if (q != NULL) {
1217 INIT_WORK(&task->u.tk_work, rpc_async_release);
1218 queue_work(q, &task->u.tk_work);
1219 } else
1220 rpc_free_task(task);
1221}
1222
1223static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1224{
1225 if (atomic_dec_and_test(&task->tk_count)) {
1226 rpc_release_resources_task(task);
1227 rpc_final_put_task(task, q);
1228 }
1229}
1230
1231void rpc_put_task(struct rpc_task *task)
1232{
1233 rpc_do_put_task(task, NULL);
1234}
1235EXPORT_SYMBOL_GPL(rpc_put_task);
1236
1237void rpc_put_task_async(struct rpc_task *task)
1238{
1239 rpc_do_put_task(task, task->tk_workqueue);
1240}
1241EXPORT_SYMBOL_GPL(rpc_put_task_async);
1242
1243static void rpc_release_task(struct rpc_task *task)
1244{
1245 WARN_ON_ONCE(RPC_IS_QUEUED(task));
1246
1247 rpc_release_resources_task(task);
1248
1249 /*
1250 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1251 * so it should be safe to use task->tk_count as a test for whether
1252 * or not any other processes still hold references to our rpc_task.
1253 */
1254 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1255 /* Wake up anyone who may be waiting for task completion */
1256 if (!rpc_complete_task(task))
1257 return;
1258 } else {
1259 if (!atomic_dec_and_test(&task->tk_count))
1260 return;
1261 }
1262 rpc_final_put_task(task, task->tk_workqueue);
1263}
1264
1265int rpciod_up(void)
1266{
1267 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1268}
1269
1270void rpciod_down(void)
1271{
1272 module_put(THIS_MODULE);
1273}
1274
1275/*
1276 * Start up the rpciod workqueue.
1277 */
1278static int rpciod_start(void)
1279{
1280 struct workqueue_struct *wq;
1281
1282 /*
1283 * Create the rpciod thread and wait for it to start.
1284 */
1285 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1286 if (!wq)
1287 goto out_failed;
1288 rpciod_workqueue = wq;
1289 wq = alloc_workqueue("xprtiod", WQ_UNBOUND | WQ_MEM_RECLAIM, 0);
1290 if (!wq)
1291 goto free_rpciod;
1292 xprtiod_workqueue = wq;
1293 return 1;
1294free_rpciod:
1295 wq = rpciod_workqueue;
1296 rpciod_workqueue = NULL;
1297 destroy_workqueue(wq);
1298out_failed:
1299 return 0;
1300}
1301
1302static void rpciod_stop(void)
1303{
1304 struct workqueue_struct *wq = NULL;
1305
1306 if (rpciod_workqueue == NULL)
1307 return;
1308
1309 wq = rpciod_workqueue;
1310 rpciod_workqueue = NULL;
1311 destroy_workqueue(wq);
1312 wq = xprtiod_workqueue;
1313 xprtiod_workqueue = NULL;
1314 destroy_workqueue(wq);
1315}
1316
1317void
1318rpc_destroy_mempool(void)
1319{
1320 rpciod_stop();
1321 mempool_destroy(rpc_buffer_mempool);
1322 mempool_destroy(rpc_task_mempool);
1323 kmem_cache_destroy(rpc_task_slabp);
1324 kmem_cache_destroy(rpc_buffer_slabp);
1325 rpc_destroy_wait_queue(&delay_queue);
1326}
1327
1328int
1329rpc_init_mempool(void)
1330{
1331 /*
1332 * The following is not strictly a mempool initialisation,
1333 * but there is no harm in doing it here
1334 */
1335 rpc_init_wait_queue(&delay_queue, "delayq");
1336 if (!rpciod_start())
1337 goto err_nomem;
1338
1339 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1340 sizeof(struct rpc_task),
1341 0, SLAB_HWCACHE_ALIGN,
1342 NULL);
1343 if (!rpc_task_slabp)
1344 goto err_nomem;
1345 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1346 RPC_BUFFER_MAXSIZE,
1347 0, SLAB_HWCACHE_ALIGN,
1348 NULL);
1349 if (!rpc_buffer_slabp)
1350 goto err_nomem;
1351 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1352 rpc_task_slabp);
1353 if (!rpc_task_mempool)
1354 goto err_nomem;
1355 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1356 rpc_buffer_slabp);
1357 if (!rpc_buffer_mempool)
1358 goto err_nomem;
1359 return 0;
1360err_nomem:
1361 rpc_destroy_mempool();
1362 return -ENOMEM;
1363}
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}