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(head: &task->u.tk_wait.timer_list))
98 return;
99 task->tk_timeout = 0;
100 list_del(entry: &task->u.tk_wait.timer_list);
101 if (list_empty(head: &queue->timer_list.list))
102 cancel_delayed_work(dwork: &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(wq: rpciod_workqueue, dwork: &queue->timer_list.dwork, delay: 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(head: &queue->timer_list.list) || time_before(timeout, queue->timer_list.expires))
126 rpc_set_queue_timer(queue, expires: timeout);
127 list_add(new: &task->u.tk_wait.timer_list, head: &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, priority: 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(new: &task->u.tk_wait.links,
154 head: &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(list: &task->u.tk_wait.links);
162 list_add_tail(new: &task->u.tk_wait.list, head: 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(entry: &task->u.tk_wait.links);
176 return;
177 }
178 if (!list_empty(head: &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(new: &t->u.tk_wait.list, head: q);
185 list_del(entry: &task->u.tk_wait.links);
186 }
187 list_del(entry: &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(q: &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(list: &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(new: &task->u.tk_wait.list, head: &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(entry: &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(list: &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(list: &queue->timer_list.list);
256 rpc_assign_waitqueue_name(q: queue, name: 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, nr_queues: 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(dwork: &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(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(v: &rpc_pid);
295 return;
296 }
297
298 task->tk_pid = atomic_inc_return(v: &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, addr: &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(word: 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, addr: &task->tk_runstate);
329 ret = atomic_dec_and_test(v: &task->tk_count);
330 if (waitqueue_active(wq_head: wq))
331 __wake_up_locked_key(wq_head: wq, TASK_NORMAL, key: &k);
332 spin_unlock_irqrestore(lock: &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(word: &task->tk_runstate, RPC_TASK_ACTIVE,
346 action: 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, work: &task->u.tk_work);
372 } else
373 wake_up_bit(word: &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(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(queue: 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(lock: &q->lock);
442 __rpc_sleep_on_priority_timeout(q, task, timeout, queue_priority: task->tk_priority);
443 spin_unlock(lock: &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(lock: &q->lock);
460 __rpc_sleep_on_priority(q, task, queue_priority: task->tk_priority);
461 spin_unlock(lock: &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(lock: &q->lock);
476 __rpc_sleep_on_priority_timeout(q, task, timeout, queue_priority: priority);
477 spin_unlock(lock: &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(lock: &q->lock);
493 __rpc_sleep_on_priority(q, task, queue_priority: priority);
494 spin_unlock(lock: &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, q: 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(wq: 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(lock: &queue->lock);
561 rpc_wake_up_task_queue_locked(queue, task);
562 spin_unlock(lock: &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(wq: rpciod_workqueue, queue,
577 task, action: rpc_task_action_set_status, data: &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(lock: &queue->lock);
596 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
597 spin_unlock(lock: &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(head: 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(head: 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(head: 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, priority: (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(head: &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(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, action: func, data);
673 spin_unlock(lock: &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(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(lock: &queue->lock);
728 rpc_wake_up_locked(queue);
729 spin_unlock(lock: &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(lock: &queue->lock);
760 rpc_wake_up_status_locked(queue, status);
761 spin_unlock(lock: &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(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, action: 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(head: &queue->timer_list.list))
787 rpc_set_queue_timer(queue, expires);
788 spin_unlock(lock: &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
821 /* starting timestamp */
822 task->tk_start = ktime_get();
823}
824
825static void
826rpc_reset_task_statistics(struct rpc_task *task)
827{
828 task->tk_timeouts = 0;
829 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
830 rpc_init_task_statistics(task);
831}
832
833/*
834 * Helper that calls task->tk_ops->rpc_call_done if it exists
835 */
836void rpc_exit_task(struct rpc_task *task)
837{
838 trace_rpc_task_end(task, action: task->tk_action);
839 task->tk_action = NULL;
840 if (task->tk_ops->rpc_count_stats)
841 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
842 else if (task->tk_client)
843 rpc_count_iostats(task, task->tk_client->cl_metrics);
844 if (task->tk_ops->rpc_call_done != NULL) {
845 trace_rpc_task_call_done(task, action: task->tk_ops->rpc_call_done);
846 task->tk_ops->rpc_call_done(task, task->tk_calldata);
847 if (task->tk_action != NULL) {
848 /* Always release the RPC slot and buffer memory */
849 xprt_release(task);
850 rpc_reset_task_statistics(task);
851 }
852 }
853}
854
855void rpc_signal_task(struct rpc_task *task)
856{
857 struct rpc_wait_queue *queue;
858
859 if (!RPC_IS_ACTIVATED(task))
860 return;
861
862 if (!rpc_task_set_rpc_status(task, rpc_status: -ERESTARTSYS))
863 return;
864 trace_rpc_task_signalled(task, action: task->tk_action);
865 set_bit(RPC_TASK_SIGNALLED, addr: &task->tk_runstate);
866 smp_mb__after_atomic();
867 queue = READ_ONCE(task->tk_waitqueue);
868 if (queue)
869 rpc_wake_up_queued_task(queue, task);
870}
871
872void rpc_task_try_cancel(struct rpc_task *task, int error)
873{
874 struct rpc_wait_queue *queue;
875
876 if (!rpc_task_set_rpc_status(task, rpc_status: error))
877 return;
878 queue = READ_ONCE(task->tk_waitqueue);
879 if (queue)
880 rpc_wake_up_queued_task(queue, task);
881}
882
883void rpc_exit(struct rpc_task *task, int status)
884{
885 task->tk_status = status;
886 task->tk_action = rpc_exit_task;
887 rpc_wake_up_queued_task(task->tk_waitqueue, task);
888}
889EXPORT_SYMBOL_GPL(rpc_exit);
890
891void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
892{
893 if (ops->rpc_release != NULL)
894 ops->rpc_release(calldata);
895}
896
897static bool xprt_needs_memalloc(struct rpc_xprt *xprt, struct rpc_task *tk)
898{
899 if (!xprt)
900 return false;
901 if (!atomic_read(v: &xprt->swapper))
902 return false;
903 return test_bit(XPRT_LOCKED, &xprt->state) && xprt->snd_task == tk;
904}
905
906/*
907 * This is the RPC `scheduler' (or rather, the finite state machine).
908 */
909static void __rpc_execute(struct rpc_task *task)
910{
911 struct rpc_wait_queue *queue;
912 int task_is_async = RPC_IS_ASYNC(task);
913 int status = 0;
914 unsigned long pflags = current->flags;
915
916 WARN_ON_ONCE(RPC_IS_QUEUED(task));
917 if (RPC_IS_QUEUED(task))
918 return;
919
920 for (;;) {
921 void (*do_action)(struct rpc_task *);
922
923 /*
924 * Perform the next FSM step or a pending callback.
925 *
926 * tk_action may be NULL if the task has been killed.
927 */
928 do_action = task->tk_action;
929 /* Tasks with an RPC error status should exit */
930 if (do_action && do_action != rpc_exit_task &&
931 (status = READ_ONCE(task->tk_rpc_status)) != 0) {
932 task->tk_status = status;
933 do_action = rpc_exit_task;
934 }
935 /* Callbacks override all actions */
936 if (task->tk_callback) {
937 do_action = task->tk_callback;
938 task->tk_callback = NULL;
939 }
940 if (!do_action)
941 break;
942 if (RPC_IS_SWAPPER(task) ||
943 xprt_needs_memalloc(xprt: task->tk_xprt, tk: task))
944 current->flags |= PF_MEMALLOC;
945
946 trace_rpc_task_run_action(task, action: do_action);
947 do_action(task);
948
949 /*
950 * Lockless check for whether task is sleeping or not.
951 */
952 if (!RPC_IS_QUEUED(task)) {
953 cond_resched();
954 continue;
955 }
956
957 /*
958 * The queue->lock protects against races with
959 * rpc_make_runnable().
960 *
961 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
962 * rpc_task, rpc_make_runnable() can assign it to a
963 * different workqueue. We therefore cannot assume that the
964 * rpc_task pointer may still be dereferenced.
965 */
966 queue = task->tk_waitqueue;
967 spin_lock(lock: &queue->lock);
968 if (!RPC_IS_QUEUED(task)) {
969 spin_unlock(lock: &queue->lock);
970 continue;
971 }
972 /* Wake up any task that has an exit status */
973 if (READ_ONCE(task->tk_rpc_status) != 0) {
974 rpc_wake_up_task_queue_locked(queue, task);
975 spin_unlock(lock: &queue->lock);
976 continue;
977 }
978 rpc_clear_running(task);
979 spin_unlock(lock: &queue->lock);
980 if (task_is_async)
981 goto out;
982
983 /* sync task: sleep here */
984 trace_rpc_task_sync_sleep(task, action: task->tk_action);
985 status = out_of_line_wait_on_bit(word: &task->tk_runstate,
986 RPC_TASK_QUEUED, action: rpc_wait_bit_killable,
987 TASK_KILLABLE|TASK_FREEZABLE);
988 if (status < 0) {
989 /*
990 * When a sync task receives a signal, it exits with
991 * -ERESTARTSYS. In order to catch any callbacks that
992 * clean up after sleeping on some queue, we don't
993 * break the loop here, but go around once more.
994 */
995 rpc_signal_task(task);
996 }
997 trace_rpc_task_sync_wake(task, action: task->tk_action);
998 }
999
1000 /* Release all resources associated with the task */
1001 rpc_release_task(task);
1002out:
1003 current_restore_flags(orig_flags: pflags, PF_MEMALLOC);
1004}
1005
1006/*
1007 * User-visible entry point to the scheduler.
1008 *
1009 * This may be called recursively if e.g. an async NFS task updates
1010 * the attributes and finds that dirty pages must be flushed.
1011 * NOTE: Upon exit of this function the task is guaranteed to be
1012 * released. In particular note that tk_release() will have
1013 * been called, so your task memory may have been freed.
1014 */
1015void rpc_execute(struct rpc_task *task)
1016{
1017 bool is_async = RPC_IS_ASYNC(task);
1018
1019 rpc_set_active(task);
1020 rpc_make_runnable(wq: rpciod_workqueue, task);
1021 if (!is_async) {
1022 unsigned int pflags = memalloc_nofs_save();
1023 __rpc_execute(task);
1024 memalloc_nofs_restore(flags: pflags);
1025 }
1026}
1027
1028static void rpc_async_schedule(struct work_struct *work)
1029{
1030 unsigned int pflags = memalloc_nofs_save();
1031
1032 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
1033 memalloc_nofs_restore(flags: pflags);
1034}
1035
1036/**
1037 * rpc_malloc - allocate RPC buffer resources
1038 * @task: RPC task
1039 *
1040 * A single memory region is allocated, which is split between the
1041 * RPC call and RPC reply that this task is being used for. When
1042 * this RPC is retired, the memory is released by calling rpc_free.
1043 *
1044 * To prevent rpciod from hanging, this allocator never sleeps,
1045 * returning -ENOMEM and suppressing warning if the request cannot
1046 * be serviced immediately. The caller can arrange to sleep in a
1047 * way that is safe for rpciod.
1048 *
1049 * Most requests are 'small' (under 2KiB) and can be serviced from a
1050 * mempool, ensuring that NFS reads and writes can always proceed,
1051 * and that there is good locality of reference for these buffers.
1052 */
1053int rpc_malloc(struct rpc_task *task)
1054{
1055 struct rpc_rqst *rqst = task->tk_rqstp;
1056 size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
1057 struct rpc_buffer *buf;
1058 gfp_t gfp = rpc_task_gfp_mask();
1059
1060 size += sizeof(struct rpc_buffer);
1061 if (size <= RPC_BUFFER_MAXSIZE) {
1062 buf = kmem_cache_alloc(cachep: rpc_buffer_slabp, flags: gfp);
1063 /* Reach for the mempool if dynamic allocation fails */
1064 if (!buf && RPC_IS_ASYNC(task))
1065 buf = mempool_alloc(pool: rpc_buffer_mempool, GFP_NOWAIT);
1066 } else
1067 buf = kmalloc(size, flags: gfp);
1068
1069 if (!buf)
1070 return -ENOMEM;
1071
1072 buf->len = size;
1073 rqst->rq_buffer = buf->data;
1074 rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1075 return 0;
1076}
1077EXPORT_SYMBOL_GPL(rpc_malloc);
1078
1079/**
1080 * rpc_free - free RPC buffer resources allocated via rpc_malloc
1081 * @task: RPC task
1082 *
1083 */
1084void rpc_free(struct rpc_task *task)
1085{
1086 void *buffer = task->tk_rqstp->rq_buffer;
1087 size_t size;
1088 struct rpc_buffer *buf;
1089
1090 buf = container_of(buffer, struct rpc_buffer, data);
1091 size = buf->len;
1092
1093 if (size <= RPC_BUFFER_MAXSIZE)
1094 mempool_free(element: buf, pool: rpc_buffer_mempool);
1095 else
1096 kfree(objp: buf);
1097}
1098EXPORT_SYMBOL_GPL(rpc_free);
1099
1100/*
1101 * Creation and deletion of RPC task structures
1102 */
1103static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1104{
1105 memset(task, 0, sizeof(*task));
1106 atomic_set(v: &task->tk_count, i: 1);
1107 task->tk_flags = task_setup_data->flags;
1108 task->tk_ops = task_setup_data->callback_ops;
1109 task->tk_calldata = task_setup_data->callback_data;
1110 INIT_LIST_HEAD(list: &task->tk_task);
1111
1112 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1113 task->tk_owner = current->tgid;
1114
1115 /* Initialize workqueue for async tasks */
1116 task->tk_workqueue = task_setup_data->workqueue;
1117
1118 task->tk_xprt = rpc_task_get_xprt(clnt: task_setup_data->rpc_client,
1119 xprt: xprt_get(xprt: task_setup_data->rpc_xprt));
1120
1121 task->tk_op_cred = get_rpccred(cred: task_setup_data->rpc_op_cred);
1122
1123 if (task->tk_ops->rpc_call_prepare != NULL)
1124 task->tk_action = rpc_prepare_task;
1125
1126 rpc_init_task_statistics(task);
1127}
1128
1129static struct rpc_task *rpc_alloc_task(void)
1130{
1131 struct rpc_task *task;
1132
1133 task = kmem_cache_alloc(cachep: rpc_task_slabp, flags: rpc_task_gfp_mask());
1134 if (task)
1135 return task;
1136 return mempool_alloc(pool: rpc_task_mempool, GFP_NOWAIT);
1137}
1138
1139/*
1140 * Create a new task for the specified client.
1141 */
1142struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1143{
1144 struct rpc_task *task = setup_data->task;
1145 unsigned short flags = 0;
1146
1147 if (task == NULL) {
1148 task = rpc_alloc_task();
1149 if (task == NULL) {
1150 rpc_release_calldata(ops: setup_data->callback_ops,
1151 calldata: setup_data->callback_data);
1152 return ERR_PTR(error: -ENOMEM);
1153 }
1154 flags = RPC_TASK_DYNAMIC;
1155 }
1156
1157 rpc_init_task(task, task_setup_data: setup_data);
1158 task->tk_flags |= flags;
1159 return task;
1160}
1161
1162/*
1163 * rpc_free_task - release rpc task and perform cleanups
1164 *
1165 * Note that we free up the rpc_task _after_ rpc_release_calldata()
1166 * in order to work around a workqueue dependency issue.
1167 *
1168 * Tejun Heo states:
1169 * "Workqueue currently considers two work items to be the same if they're
1170 * on the same address and won't execute them concurrently - ie. it
1171 * makes a work item which is queued again while being executed wait
1172 * for the previous execution to complete.
1173 *
1174 * If a work function frees the work item, and then waits for an event
1175 * which should be performed by another work item and *that* work item
1176 * recycles the freed work item, it can create a false dependency loop.
1177 * There really is no reliable way to detect this short of verifying
1178 * every memory free."
1179 *
1180 */
1181static void rpc_free_task(struct rpc_task *task)
1182{
1183 unsigned short tk_flags = task->tk_flags;
1184
1185 put_rpccred(task->tk_op_cred);
1186 rpc_release_calldata(ops: task->tk_ops, calldata: task->tk_calldata);
1187
1188 if (tk_flags & RPC_TASK_DYNAMIC)
1189 mempool_free(element: task, pool: rpc_task_mempool);
1190}
1191
1192static void rpc_async_release(struct work_struct *work)
1193{
1194 unsigned int pflags = memalloc_nofs_save();
1195
1196 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1197 memalloc_nofs_restore(flags: pflags);
1198}
1199
1200static void rpc_release_resources_task(struct rpc_task *task)
1201{
1202 xprt_release(task);
1203 if (task->tk_msg.rpc_cred) {
1204 if (!(task->tk_flags & RPC_TASK_CRED_NOREF))
1205 put_cred(cred: task->tk_msg.rpc_cred);
1206 task->tk_msg.rpc_cred = NULL;
1207 }
1208 rpc_task_release_client(task);
1209}
1210
1211static void rpc_final_put_task(struct rpc_task *task,
1212 struct workqueue_struct *q)
1213{
1214 if (q != NULL) {
1215 INIT_WORK(&task->u.tk_work, rpc_async_release);
1216 queue_work(wq: q, work: &task->u.tk_work);
1217 } else
1218 rpc_free_task(task);
1219}
1220
1221static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1222{
1223 if (atomic_dec_and_test(v: &task->tk_count)) {
1224 rpc_release_resources_task(task);
1225 rpc_final_put_task(task, q);
1226 }
1227}
1228
1229void rpc_put_task(struct rpc_task *task)
1230{
1231 rpc_do_put_task(task, NULL);
1232}
1233EXPORT_SYMBOL_GPL(rpc_put_task);
1234
1235void rpc_put_task_async(struct rpc_task *task)
1236{
1237 rpc_do_put_task(task, q: task->tk_workqueue);
1238}
1239EXPORT_SYMBOL_GPL(rpc_put_task_async);
1240
1241static void rpc_release_task(struct rpc_task *task)
1242{
1243 WARN_ON_ONCE(RPC_IS_QUEUED(task));
1244
1245 rpc_release_resources_task(task);
1246
1247 /*
1248 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1249 * so it should be safe to use task->tk_count as a test for whether
1250 * or not any other processes still hold references to our rpc_task.
1251 */
1252 if (atomic_read(v: &task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1253 /* Wake up anyone who may be waiting for task completion */
1254 if (!rpc_complete_task(task))
1255 return;
1256 } else {
1257 if (!atomic_dec_and_test(v: &task->tk_count))
1258 return;
1259 }
1260 rpc_final_put_task(task, q: task->tk_workqueue);
1261}
1262
1263int rpciod_up(void)
1264{
1265 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1266}
1267
1268void rpciod_down(void)
1269{
1270 module_put(THIS_MODULE);
1271}
1272
1273/*
1274 * Start up the rpciod workqueue.
1275 */
1276static int rpciod_start(void)
1277{
1278 struct workqueue_struct *wq;
1279
1280 /*
1281 * Create the rpciod thread and wait for it to start.
1282 */
1283 wq = alloc_workqueue(fmt: "rpciod", flags: WQ_MEM_RECLAIM | WQ_UNBOUND, max_active: 0);
1284 if (!wq)
1285 goto out_failed;
1286 rpciod_workqueue = wq;
1287 wq = alloc_workqueue(fmt: "xprtiod", flags: WQ_UNBOUND | WQ_MEM_RECLAIM, max_active: 0);
1288 if (!wq)
1289 goto free_rpciod;
1290 xprtiod_workqueue = wq;
1291 return 1;
1292free_rpciod:
1293 wq = rpciod_workqueue;
1294 rpciod_workqueue = NULL;
1295 destroy_workqueue(wq);
1296out_failed:
1297 return 0;
1298}
1299
1300static void rpciod_stop(void)
1301{
1302 struct workqueue_struct *wq = NULL;
1303
1304 if (rpciod_workqueue == NULL)
1305 return;
1306
1307 wq = rpciod_workqueue;
1308 rpciod_workqueue = NULL;
1309 destroy_workqueue(wq);
1310 wq = xprtiod_workqueue;
1311 xprtiod_workqueue = NULL;
1312 destroy_workqueue(wq);
1313}
1314
1315void
1316rpc_destroy_mempool(void)
1317{
1318 rpciod_stop();
1319 mempool_destroy(pool: rpc_buffer_mempool);
1320 mempool_destroy(pool: rpc_task_mempool);
1321 kmem_cache_destroy(s: rpc_task_slabp);
1322 kmem_cache_destroy(s: rpc_buffer_slabp);
1323 rpc_destroy_wait_queue(&delay_queue);
1324}
1325
1326int
1327rpc_init_mempool(void)
1328{
1329 /*
1330 * The following is not strictly a mempool initialisation,
1331 * but there is no harm in doing it here
1332 */
1333 rpc_init_wait_queue(&delay_queue, "delayq");
1334 if (!rpciod_start())
1335 goto err_nomem;
1336
1337 rpc_task_slabp = kmem_cache_create(name: "rpc_tasks",
1338 size: sizeof(struct rpc_task),
1339 align: 0, SLAB_HWCACHE_ALIGN,
1340 NULL);
1341 if (!rpc_task_slabp)
1342 goto err_nomem;
1343 rpc_buffer_slabp = kmem_cache_create(name: "rpc_buffers",
1344 RPC_BUFFER_MAXSIZE,
1345 align: 0, SLAB_HWCACHE_ALIGN,
1346 NULL);
1347 if (!rpc_buffer_slabp)
1348 goto err_nomem;
1349 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1350 kc: rpc_task_slabp);
1351 if (!rpc_task_mempool)
1352 goto err_nomem;
1353 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1354 kc: rpc_buffer_slabp);
1355 if (!rpc_buffer_mempool)
1356 goto err_nomem;
1357 return 0;
1358err_nomem:
1359 rpc_destroy_mempool();
1360 return -ENOMEM;
1361}
1362

source code of linux/net/sunrpc/sched.c