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