Loading...
1/*
2 * linux/net/sunrpc/sched.c
3 *
4 * Scheduling for synchronous and asynchronous RPC requests.
5 *
6 * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
7 *
8 * TCP NFS related read + write fixes
9 * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
10 */
11
12#include <linux/module.h>
13
14#include <linux/sched.h>
15#include <linux/interrupt.h>
16#include <linux/slab.h>
17#include <linux/mempool.h>
18#include <linux/smp.h>
19#include <linux/spinlock.h>
20#include <linux/mutex.h>
21#include <linux/freezer.h>
22
23#include <linux/sunrpc/clnt.h>
24
25#include "sunrpc.h"
26
27#if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
28#define RPCDBG_FACILITY RPCDBG_SCHED
29#endif
30
31#define CREATE_TRACE_POINTS
32#include <trace/events/sunrpc.h>
33
34/*
35 * RPC slabs and memory pools
36 */
37#define RPC_BUFFER_MAXSIZE (2048)
38#define RPC_BUFFER_POOLSIZE (8)
39#define RPC_TASK_POOLSIZE (8)
40static struct kmem_cache *rpc_task_slabp __read_mostly;
41static struct kmem_cache *rpc_buffer_slabp __read_mostly;
42static mempool_t *rpc_task_mempool __read_mostly;
43static mempool_t *rpc_buffer_mempool __read_mostly;
44
45static void rpc_async_schedule(struct work_struct *);
46static void rpc_release_task(struct rpc_task *task);
47static void __rpc_queue_timer_fn(struct timer_list *t);
48
49/*
50 * RPC tasks sit here while waiting for conditions to improve.
51 */
52static struct rpc_wait_queue delay_queue;
53
54/*
55 * rpciod-related stuff
56 */
57struct workqueue_struct *rpciod_workqueue __read_mostly;
58struct workqueue_struct *xprtiod_workqueue __read_mostly;
59
60/*
61 * Disable the timer for a given RPC task. Should be called with
62 * queue->lock and bh_disabled in order to avoid races within
63 * rpc_run_timer().
64 */
65static void
66__rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
67{
68 if (task->tk_timeout == 0)
69 return;
70 dprintk("RPC: %5u disabling timer\n", task->tk_pid);
71 task->tk_timeout = 0;
72 list_del(&task->u.tk_wait.timer_list);
73 if (list_empty(&queue->timer_list.list))
74 del_timer(&queue->timer_list.timer);
75}
76
77static void
78rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
79{
80 queue->timer_list.expires = expires;
81 mod_timer(&queue->timer_list.timer, expires);
82}
83
84/*
85 * Set up a timer for the current task.
86 */
87static void
88__rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
89{
90 if (!task->tk_timeout)
91 return;
92
93 dprintk("RPC: %5u setting alarm for %u ms\n",
94 task->tk_pid, jiffies_to_msecs(task->tk_timeout));
95
96 task->u.tk_wait.expires = jiffies + task->tk_timeout;
97 if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
98 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
99 list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
100}
101
102static void rpc_rotate_queue_owner(struct rpc_wait_queue *queue)
103{
104 struct list_head *q = &queue->tasks[queue->priority];
105 struct rpc_task *task;
106
107 if (!list_empty(q)) {
108 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
109 if (task->tk_owner == queue->owner)
110 list_move_tail(&task->u.tk_wait.list, q);
111 }
112}
113
114static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
115{
116 if (queue->priority != priority) {
117 /* Fairness: rotate the list when changing priority */
118 rpc_rotate_queue_owner(queue);
119 queue->priority = priority;
120 }
121}
122
123static void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
124{
125 queue->owner = pid;
126 queue->nr = RPC_BATCH_COUNT;
127}
128
129static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
130{
131 rpc_set_waitqueue_priority(queue, queue->maxpriority);
132 rpc_set_waitqueue_owner(queue, 0);
133}
134
135/*
136 * Add new request to a priority queue.
137 */
138static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
139 struct rpc_task *task,
140 unsigned char queue_priority)
141{
142 struct list_head *q;
143 struct rpc_task *t;
144
145 INIT_LIST_HEAD(&task->u.tk_wait.links);
146 if (unlikely(queue_priority > queue->maxpriority))
147 queue_priority = queue->maxpriority;
148 if (queue_priority > queue->priority)
149 rpc_set_waitqueue_priority(queue, queue_priority);
150 q = &queue->tasks[queue_priority];
151 list_for_each_entry(t, q, u.tk_wait.list) {
152 if (t->tk_owner == task->tk_owner) {
153 list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
154 return;
155 }
156 }
157 list_add_tail(&task->u.tk_wait.list, q);
158}
159
160/*
161 * Add new request to wait queue.
162 *
163 * Swapper tasks always get inserted at the head of the queue.
164 * This should avoid many nasty memory deadlocks and hopefully
165 * improve overall performance.
166 * Everyone else gets appended to the queue to ensure proper FIFO behavior.
167 */
168static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
169 struct rpc_task *task,
170 unsigned char queue_priority)
171{
172 WARN_ON_ONCE(RPC_IS_QUEUED(task));
173 if (RPC_IS_QUEUED(task))
174 return;
175
176 if (RPC_IS_PRIORITY(queue))
177 __rpc_add_wait_queue_priority(queue, task, queue_priority);
178 else if (RPC_IS_SWAPPER(task))
179 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
180 else
181 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
182 task->tk_waitqueue = queue;
183 queue->qlen++;
184 /* barrier matches the read in rpc_wake_up_task_queue_locked() */
185 smp_wmb();
186 rpc_set_queued(task);
187
188 dprintk("RPC: %5u added to queue %p \"%s\"\n",
189 task->tk_pid, queue, rpc_qname(queue));
190}
191
192/*
193 * Remove request from a priority queue.
194 */
195static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
196{
197 struct rpc_task *t;
198
199 if (!list_empty(&task->u.tk_wait.links)) {
200 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
201 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
202 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
203 }
204}
205
206/*
207 * Remove request from queue.
208 * Note: must be called with spin lock held.
209 */
210static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
211{
212 __rpc_disable_timer(queue, task);
213 if (RPC_IS_PRIORITY(queue))
214 __rpc_remove_wait_queue_priority(task);
215 list_del(&task->u.tk_wait.list);
216 queue->qlen--;
217 dprintk("RPC: %5u removed from queue %p \"%s\"\n",
218 task->tk_pid, queue, rpc_qname(queue));
219}
220
221static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
222{
223 int i;
224
225 spin_lock_init(&queue->lock);
226 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
227 INIT_LIST_HEAD(&queue->tasks[i]);
228 queue->maxpriority = nr_queues - 1;
229 rpc_reset_waitqueue_priority(queue);
230 queue->qlen = 0;
231 timer_setup(&queue->timer_list.timer, __rpc_queue_timer_fn, 0);
232 INIT_LIST_HEAD(&queue->timer_list.list);
233 rpc_assign_waitqueue_name(queue, qname);
234}
235
236void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
237{
238 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
239}
240EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
241
242void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
243{
244 __rpc_init_priority_wait_queue(queue, qname, 1);
245}
246EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
247
248void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
249{
250 del_timer_sync(&queue->timer_list.timer);
251}
252EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
253
254static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
255{
256 freezable_schedule_unsafe();
257 if (signal_pending_state(mode, current))
258 return -ERESTARTSYS;
259 return 0;
260}
261
262#if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
263static void rpc_task_set_debuginfo(struct rpc_task *task)
264{
265 static atomic_t rpc_pid;
266
267 task->tk_pid = atomic_inc_return(&rpc_pid);
268}
269#else
270static inline void rpc_task_set_debuginfo(struct rpc_task *task)
271{
272}
273#endif
274
275static void rpc_set_active(struct rpc_task *task)
276{
277 rpc_task_set_debuginfo(task);
278 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
279 trace_rpc_task_begin(task, NULL);
280}
281
282/*
283 * Mark an RPC call as having completed by clearing the 'active' bit
284 * and then waking up all tasks that were sleeping.
285 */
286static int rpc_complete_task(struct rpc_task *task)
287{
288 void *m = &task->tk_runstate;
289 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
290 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
291 unsigned long flags;
292 int ret;
293
294 trace_rpc_task_complete(task, NULL);
295
296 spin_lock_irqsave(&wq->lock, flags);
297 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
298 ret = atomic_dec_and_test(&task->tk_count);
299 if (waitqueue_active(wq))
300 __wake_up_locked_key(wq, TASK_NORMAL, &k);
301 spin_unlock_irqrestore(&wq->lock, flags);
302 return ret;
303}
304
305/*
306 * Allow callers to wait for completion of an RPC call
307 *
308 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
309 * to enforce taking of the wq->lock and hence avoid races with
310 * rpc_complete_task().
311 */
312int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
313{
314 if (action == NULL)
315 action = rpc_wait_bit_killable;
316 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
317 action, TASK_KILLABLE);
318}
319EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
320
321/*
322 * Make an RPC task runnable.
323 *
324 * Note: If the task is ASYNC, and is being made runnable after sitting on an
325 * rpc_wait_queue, this must be called with the queue spinlock held to protect
326 * the wait queue operation.
327 * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
328 * which is needed to ensure that __rpc_execute() doesn't loop (due to the
329 * lockless RPC_IS_QUEUED() test) before we've had a chance to test
330 * the RPC_TASK_RUNNING flag.
331 */
332static void rpc_make_runnable(struct workqueue_struct *wq,
333 struct rpc_task *task)
334{
335 bool need_wakeup = !rpc_test_and_set_running(task);
336
337 rpc_clear_queued(task);
338 if (!need_wakeup)
339 return;
340 if (RPC_IS_ASYNC(task)) {
341 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
342 queue_work(wq, &task->u.tk_work);
343 } else
344 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
345}
346
347/*
348 * Prepare for sleeping on a wait queue.
349 * By always appending tasks to the list we ensure FIFO behavior.
350 * NB: An RPC task will only receive interrupt-driven events as long
351 * as it's on a wait queue.
352 */
353static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
354 struct rpc_task *task,
355 rpc_action action,
356 unsigned char queue_priority)
357{
358 dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
359 task->tk_pid, rpc_qname(q), jiffies);
360
361 trace_rpc_task_sleep(task, q);
362
363 __rpc_add_wait_queue(q, task, queue_priority);
364
365 WARN_ON_ONCE(task->tk_callback != NULL);
366 task->tk_callback = action;
367 __rpc_add_timer(q, task);
368}
369
370void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
371 rpc_action action)
372{
373 /* We shouldn't ever put an inactive task to sleep */
374 WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
375 if (!RPC_IS_ACTIVATED(task)) {
376 task->tk_status = -EIO;
377 rpc_put_task_async(task);
378 return;
379 }
380
381 /*
382 * Protect the queue operations.
383 */
384 spin_lock_bh(&q->lock);
385 __rpc_sleep_on_priority(q, task, action, task->tk_priority);
386 spin_unlock_bh(&q->lock);
387}
388EXPORT_SYMBOL_GPL(rpc_sleep_on);
389
390void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
391 rpc_action action, int priority)
392{
393 /* We shouldn't ever put an inactive task to sleep */
394 WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
395 if (!RPC_IS_ACTIVATED(task)) {
396 task->tk_status = -EIO;
397 rpc_put_task_async(task);
398 return;
399 }
400
401 /*
402 * Protect the queue operations.
403 */
404 spin_lock_bh(&q->lock);
405 __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
406 spin_unlock_bh(&q->lock);
407}
408EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
409
410/**
411 * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
412 * @wq: workqueue on which to run task
413 * @queue: wait queue
414 * @task: task to be woken up
415 *
416 * Caller must hold queue->lock, and have cleared the task queued flag.
417 */
418static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
419 struct rpc_wait_queue *queue,
420 struct rpc_task *task)
421{
422 dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
423 task->tk_pid, jiffies);
424
425 /* Has the task been executed yet? If not, we cannot wake it up! */
426 if (!RPC_IS_ACTIVATED(task)) {
427 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
428 return;
429 }
430
431 trace_rpc_task_wakeup(task, queue);
432
433 __rpc_remove_wait_queue(queue, task);
434
435 rpc_make_runnable(wq, task);
436
437 dprintk("RPC: __rpc_wake_up_task done\n");
438}
439
440/*
441 * Wake up a queued task while the queue lock is being held
442 */
443static void rpc_wake_up_task_on_wq_queue_locked(struct workqueue_struct *wq,
444 struct rpc_wait_queue *queue, struct rpc_task *task)
445{
446 if (RPC_IS_QUEUED(task)) {
447 smp_rmb();
448 if (task->tk_waitqueue == queue)
449 __rpc_do_wake_up_task_on_wq(wq, queue, task);
450 }
451}
452
453/*
454 * Wake up a queued task while the queue lock is being held
455 */
456static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
457{
458 rpc_wake_up_task_on_wq_queue_locked(rpciod_workqueue, queue, task);
459}
460
461/*
462 * Wake up a task on a specific queue
463 */
464void rpc_wake_up_queued_task_on_wq(struct workqueue_struct *wq,
465 struct rpc_wait_queue *queue,
466 struct rpc_task *task)
467{
468 spin_lock_bh(&queue->lock);
469 rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
470 spin_unlock_bh(&queue->lock);
471}
472
473/*
474 * Wake up a task on a specific queue
475 */
476void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
477{
478 spin_lock_bh(&queue->lock);
479 rpc_wake_up_task_queue_locked(queue, task);
480 spin_unlock_bh(&queue->lock);
481}
482EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
483
484/*
485 * Wake up the next task on a priority queue.
486 */
487static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
488{
489 struct list_head *q;
490 struct rpc_task *task;
491
492 /*
493 * Service a batch of tasks from a single owner.
494 */
495 q = &queue->tasks[queue->priority];
496 if (!list_empty(q)) {
497 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
498 if (queue->owner == task->tk_owner) {
499 if (--queue->nr)
500 goto out;
501 list_move_tail(&task->u.tk_wait.list, q);
502 }
503 /*
504 * Check if we need to switch queues.
505 */
506 goto new_owner;
507 }
508
509 /*
510 * Service the next queue.
511 */
512 do {
513 if (q == &queue->tasks[0])
514 q = &queue->tasks[queue->maxpriority];
515 else
516 q = q - 1;
517 if (!list_empty(q)) {
518 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
519 goto new_queue;
520 }
521 } while (q != &queue->tasks[queue->priority]);
522
523 rpc_reset_waitqueue_priority(queue);
524 return NULL;
525
526new_queue:
527 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
528new_owner:
529 rpc_set_waitqueue_owner(queue, task->tk_owner);
530out:
531 return task;
532}
533
534static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
535{
536 if (RPC_IS_PRIORITY(queue))
537 return __rpc_find_next_queued_priority(queue);
538 if (!list_empty(&queue->tasks[0]))
539 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
540 return NULL;
541}
542
543/*
544 * Wake up the first task on the wait queue.
545 */
546struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
547 struct rpc_wait_queue *queue,
548 bool (*func)(struct rpc_task *, void *), void *data)
549{
550 struct rpc_task *task = NULL;
551
552 dprintk("RPC: wake_up_first(%p \"%s\")\n",
553 queue, rpc_qname(queue));
554 spin_lock_bh(&queue->lock);
555 task = __rpc_find_next_queued(queue);
556 if (task != NULL) {
557 if (func(task, data))
558 rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
559 else
560 task = NULL;
561 }
562 spin_unlock_bh(&queue->lock);
563
564 return task;
565}
566
567/*
568 * Wake up the first task on the wait queue.
569 */
570struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
571 bool (*func)(struct rpc_task *, void *), void *data)
572{
573 return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
574}
575EXPORT_SYMBOL_GPL(rpc_wake_up_first);
576
577static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
578{
579 return true;
580}
581
582/*
583 * Wake up the next task on the wait queue.
584*/
585struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
586{
587 return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
588}
589EXPORT_SYMBOL_GPL(rpc_wake_up_next);
590
591/**
592 * rpc_wake_up - wake up all rpc_tasks
593 * @queue: rpc_wait_queue on which the tasks are sleeping
594 *
595 * Grabs queue->lock
596 */
597void rpc_wake_up(struct rpc_wait_queue *queue)
598{
599 struct list_head *head;
600
601 spin_lock_bh(&queue->lock);
602 head = &queue->tasks[queue->maxpriority];
603 for (;;) {
604 while (!list_empty(head)) {
605 struct rpc_task *task;
606 task = list_first_entry(head,
607 struct rpc_task,
608 u.tk_wait.list);
609 rpc_wake_up_task_queue_locked(queue, task);
610 }
611 if (head == &queue->tasks[0])
612 break;
613 head--;
614 }
615 spin_unlock_bh(&queue->lock);
616}
617EXPORT_SYMBOL_GPL(rpc_wake_up);
618
619/**
620 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
621 * @queue: rpc_wait_queue on which the tasks are sleeping
622 * @status: status value to set
623 *
624 * Grabs queue->lock
625 */
626void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
627{
628 struct list_head *head;
629
630 spin_lock_bh(&queue->lock);
631 head = &queue->tasks[queue->maxpriority];
632 for (;;) {
633 while (!list_empty(head)) {
634 struct rpc_task *task;
635 task = list_first_entry(head,
636 struct rpc_task,
637 u.tk_wait.list);
638 task->tk_status = status;
639 rpc_wake_up_task_queue_locked(queue, task);
640 }
641 if (head == &queue->tasks[0])
642 break;
643 head--;
644 }
645 spin_unlock_bh(&queue->lock);
646}
647EXPORT_SYMBOL_GPL(rpc_wake_up_status);
648
649static void __rpc_queue_timer_fn(struct timer_list *t)
650{
651 struct rpc_wait_queue *queue = from_timer(queue, t, timer_list.timer);
652 struct rpc_task *task, *n;
653 unsigned long expires, now, timeo;
654
655 spin_lock(&queue->lock);
656 expires = now = jiffies;
657 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
658 timeo = task->u.tk_wait.expires;
659 if (time_after_eq(now, timeo)) {
660 dprintk("RPC: %5u timeout\n", task->tk_pid);
661 task->tk_status = -ETIMEDOUT;
662 rpc_wake_up_task_queue_locked(queue, task);
663 continue;
664 }
665 if (expires == now || time_after(expires, timeo))
666 expires = timeo;
667 }
668 if (!list_empty(&queue->timer_list.list))
669 rpc_set_queue_timer(queue, expires);
670 spin_unlock(&queue->lock);
671}
672
673static void __rpc_atrun(struct rpc_task *task)
674{
675 if (task->tk_status == -ETIMEDOUT)
676 task->tk_status = 0;
677}
678
679/*
680 * Run a task at a later time
681 */
682void rpc_delay(struct rpc_task *task, unsigned long delay)
683{
684 task->tk_timeout = delay;
685 rpc_sleep_on(&delay_queue, task, __rpc_atrun);
686}
687EXPORT_SYMBOL_GPL(rpc_delay);
688
689/*
690 * Helper to call task->tk_ops->rpc_call_prepare
691 */
692void rpc_prepare_task(struct rpc_task *task)
693{
694 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
695}
696
697static void
698rpc_init_task_statistics(struct rpc_task *task)
699{
700 /* Initialize retry counters */
701 task->tk_garb_retry = 2;
702 task->tk_cred_retry = 2;
703 task->tk_rebind_retry = 2;
704
705 /* starting timestamp */
706 task->tk_start = ktime_get();
707}
708
709static void
710rpc_reset_task_statistics(struct rpc_task *task)
711{
712 task->tk_timeouts = 0;
713 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
714
715 rpc_init_task_statistics(task);
716}
717
718/*
719 * Helper that calls task->tk_ops->rpc_call_done if it exists
720 */
721void rpc_exit_task(struct rpc_task *task)
722{
723 task->tk_action = NULL;
724 if (task->tk_ops->rpc_call_done != NULL) {
725 task->tk_ops->rpc_call_done(task, task->tk_calldata);
726 if (task->tk_action != NULL) {
727 WARN_ON(RPC_ASSASSINATED(task));
728 /* Always release the RPC slot and buffer memory */
729 xprt_release(task);
730 rpc_reset_task_statistics(task);
731 }
732 }
733}
734
735void rpc_exit(struct rpc_task *task, int status)
736{
737 task->tk_status = status;
738 task->tk_action = rpc_exit_task;
739 if (RPC_IS_QUEUED(task))
740 rpc_wake_up_queued_task(task->tk_waitqueue, task);
741}
742EXPORT_SYMBOL_GPL(rpc_exit);
743
744void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
745{
746 if (ops->rpc_release != NULL)
747 ops->rpc_release(calldata);
748}
749
750/*
751 * This is the RPC `scheduler' (or rather, the finite state machine).
752 */
753static void __rpc_execute(struct rpc_task *task)
754{
755 struct rpc_wait_queue *queue;
756 int task_is_async = RPC_IS_ASYNC(task);
757 int status = 0;
758
759 dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
760 task->tk_pid, task->tk_flags);
761
762 WARN_ON_ONCE(RPC_IS_QUEUED(task));
763 if (RPC_IS_QUEUED(task))
764 return;
765
766 for (;;) {
767 void (*do_action)(struct rpc_task *);
768
769 /*
770 * Perform the next FSM step or a pending callback.
771 *
772 * tk_action may be NULL if the task has been killed.
773 * In particular, note that rpc_killall_tasks may
774 * do this at any time, so beware when dereferencing.
775 */
776 do_action = task->tk_action;
777 if (task->tk_callback) {
778 do_action = task->tk_callback;
779 task->tk_callback = NULL;
780 }
781 if (!do_action)
782 break;
783 trace_rpc_task_run_action(task, do_action);
784 do_action(task);
785
786 /*
787 * Lockless check for whether task is sleeping or not.
788 */
789 if (!RPC_IS_QUEUED(task))
790 continue;
791 /*
792 * The queue->lock protects against races with
793 * rpc_make_runnable().
794 *
795 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
796 * rpc_task, rpc_make_runnable() can assign it to a
797 * different workqueue. We therefore cannot assume that the
798 * rpc_task pointer may still be dereferenced.
799 */
800 queue = task->tk_waitqueue;
801 spin_lock_bh(&queue->lock);
802 if (!RPC_IS_QUEUED(task)) {
803 spin_unlock_bh(&queue->lock);
804 continue;
805 }
806 rpc_clear_running(task);
807 spin_unlock_bh(&queue->lock);
808 if (task_is_async)
809 return;
810
811 /* sync task: sleep here */
812 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
813 status = out_of_line_wait_on_bit(&task->tk_runstate,
814 RPC_TASK_QUEUED, rpc_wait_bit_killable,
815 TASK_KILLABLE);
816 if (status == -ERESTARTSYS) {
817 /*
818 * When a sync task receives a signal, it exits with
819 * -ERESTARTSYS. In order to catch any callbacks that
820 * clean up after sleeping on some queue, we don't
821 * break the loop here, but go around once more.
822 */
823 dprintk("RPC: %5u got signal\n", task->tk_pid);
824 task->tk_flags |= RPC_TASK_KILLED;
825 rpc_exit(task, -ERESTARTSYS);
826 }
827 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
828 }
829
830 dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
831 task->tk_status);
832 /* Release all resources associated with the task */
833 rpc_release_task(task);
834}
835
836/*
837 * User-visible entry point to the scheduler.
838 *
839 * This may be called recursively if e.g. an async NFS task updates
840 * the attributes and finds that dirty pages must be flushed.
841 * NOTE: Upon exit of this function the task is guaranteed to be
842 * released. In particular note that tk_release() will have
843 * been called, so your task memory may have been freed.
844 */
845void rpc_execute(struct rpc_task *task)
846{
847 bool is_async = RPC_IS_ASYNC(task);
848
849 rpc_set_active(task);
850 rpc_make_runnable(rpciod_workqueue, task);
851 if (!is_async)
852 __rpc_execute(task);
853}
854
855static void rpc_async_schedule(struct work_struct *work)
856{
857 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
858}
859
860/**
861 * rpc_malloc - allocate RPC buffer resources
862 * @task: RPC task
863 *
864 * A single memory region is allocated, which is split between the
865 * RPC call and RPC reply that this task is being used for. When
866 * this RPC is retired, the memory is released by calling rpc_free.
867 *
868 * To prevent rpciod from hanging, this allocator never sleeps,
869 * returning -ENOMEM and suppressing warning if the request cannot
870 * be serviced immediately. The caller can arrange to sleep in a
871 * way that is safe for rpciod.
872 *
873 * Most requests are 'small' (under 2KiB) and can be serviced from a
874 * mempool, ensuring that NFS reads and writes can always proceed,
875 * and that there is good locality of reference for these buffers.
876 *
877 * In order to avoid memory starvation triggering more writebacks of
878 * NFS requests, we avoid using GFP_KERNEL.
879 */
880int rpc_malloc(struct rpc_task *task)
881{
882 struct rpc_rqst *rqst = task->tk_rqstp;
883 size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
884 struct rpc_buffer *buf;
885 gfp_t gfp = GFP_NOIO | __GFP_NOWARN;
886
887 if (RPC_IS_SWAPPER(task))
888 gfp = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
889
890 size += sizeof(struct rpc_buffer);
891 if (size <= RPC_BUFFER_MAXSIZE)
892 buf = mempool_alloc(rpc_buffer_mempool, gfp);
893 else
894 buf = kmalloc(size, gfp);
895
896 if (!buf)
897 return -ENOMEM;
898
899 buf->len = size;
900 dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
901 task->tk_pid, size, buf);
902 rqst->rq_buffer = buf->data;
903 rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
904 return 0;
905}
906EXPORT_SYMBOL_GPL(rpc_malloc);
907
908/**
909 * rpc_free - free RPC buffer resources allocated via rpc_malloc
910 * @task: RPC task
911 *
912 */
913void rpc_free(struct rpc_task *task)
914{
915 void *buffer = task->tk_rqstp->rq_buffer;
916 size_t size;
917 struct rpc_buffer *buf;
918
919 buf = container_of(buffer, struct rpc_buffer, data);
920 size = buf->len;
921
922 dprintk("RPC: freeing buffer of size %zu at %p\n",
923 size, buf);
924
925 if (size <= RPC_BUFFER_MAXSIZE)
926 mempool_free(buf, rpc_buffer_mempool);
927 else
928 kfree(buf);
929}
930EXPORT_SYMBOL_GPL(rpc_free);
931
932/*
933 * Creation and deletion of RPC task structures
934 */
935static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
936{
937 memset(task, 0, sizeof(*task));
938 atomic_set(&task->tk_count, 1);
939 task->tk_flags = task_setup_data->flags;
940 task->tk_ops = task_setup_data->callback_ops;
941 task->tk_calldata = task_setup_data->callback_data;
942 INIT_LIST_HEAD(&task->tk_task);
943
944 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
945 task->tk_owner = current->tgid;
946
947 /* Initialize workqueue for async tasks */
948 task->tk_workqueue = task_setup_data->workqueue;
949
950 task->tk_xprt = xprt_get(task_setup_data->rpc_xprt);
951
952 if (task->tk_ops->rpc_call_prepare != NULL)
953 task->tk_action = rpc_prepare_task;
954
955 rpc_init_task_statistics(task);
956
957 dprintk("RPC: new task initialized, procpid %u\n",
958 task_pid_nr(current));
959}
960
961static struct rpc_task *
962rpc_alloc_task(void)
963{
964 return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
965}
966
967/*
968 * Create a new task for the specified client.
969 */
970struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
971{
972 struct rpc_task *task = setup_data->task;
973 unsigned short flags = 0;
974
975 if (task == NULL) {
976 task = rpc_alloc_task();
977 flags = RPC_TASK_DYNAMIC;
978 }
979
980 rpc_init_task(task, setup_data);
981 task->tk_flags |= flags;
982 dprintk("RPC: allocated task %p\n", task);
983 return task;
984}
985
986/*
987 * rpc_free_task - release rpc task and perform cleanups
988 *
989 * Note that we free up the rpc_task _after_ rpc_release_calldata()
990 * in order to work around a workqueue dependency issue.
991 *
992 * Tejun Heo states:
993 * "Workqueue currently considers two work items to be the same if they're
994 * on the same address and won't execute them concurrently - ie. it
995 * makes a work item which is queued again while being executed wait
996 * for the previous execution to complete.
997 *
998 * If a work function frees the work item, and then waits for an event
999 * which should be performed by another work item and *that* work item
1000 * recycles the freed work item, it can create a false dependency loop.
1001 * There really is no reliable way to detect this short of verifying
1002 * every memory free."
1003 *
1004 */
1005static void rpc_free_task(struct rpc_task *task)
1006{
1007 unsigned short tk_flags = task->tk_flags;
1008
1009 rpc_release_calldata(task->tk_ops, task->tk_calldata);
1010
1011 if (tk_flags & RPC_TASK_DYNAMIC) {
1012 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1013 mempool_free(task, rpc_task_mempool);
1014 }
1015}
1016
1017static void rpc_async_release(struct work_struct *work)
1018{
1019 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1020}
1021
1022static void rpc_release_resources_task(struct rpc_task *task)
1023{
1024 xprt_release(task);
1025 if (task->tk_msg.rpc_cred) {
1026 put_rpccred(task->tk_msg.rpc_cred);
1027 task->tk_msg.rpc_cred = NULL;
1028 }
1029 rpc_task_release_client(task);
1030}
1031
1032static void rpc_final_put_task(struct rpc_task *task,
1033 struct workqueue_struct *q)
1034{
1035 if (q != NULL) {
1036 INIT_WORK(&task->u.tk_work, rpc_async_release);
1037 queue_work(q, &task->u.tk_work);
1038 } else
1039 rpc_free_task(task);
1040}
1041
1042static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1043{
1044 if (atomic_dec_and_test(&task->tk_count)) {
1045 rpc_release_resources_task(task);
1046 rpc_final_put_task(task, q);
1047 }
1048}
1049
1050void rpc_put_task(struct rpc_task *task)
1051{
1052 rpc_do_put_task(task, NULL);
1053}
1054EXPORT_SYMBOL_GPL(rpc_put_task);
1055
1056void rpc_put_task_async(struct rpc_task *task)
1057{
1058 rpc_do_put_task(task, task->tk_workqueue);
1059}
1060EXPORT_SYMBOL_GPL(rpc_put_task_async);
1061
1062static void rpc_release_task(struct rpc_task *task)
1063{
1064 dprintk("RPC: %5u release task\n", task->tk_pid);
1065
1066 WARN_ON_ONCE(RPC_IS_QUEUED(task));
1067
1068 rpc_release_resources_task(task);
1069
1070 /*
1071 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1072 * so it should be safe to use task->tk_count as a test for whether
1073 * or not any other processes still hold references to our rpc_task.
1074 */
1075 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1076 /* Wake up anyone who may be waiting for task completion */
1077 if (!rpc_complete_task(task))
1078 return;
1079 } else {
1080 if (!atomic_dec_and_test(&task->tk_count))
1081 return;
1082 }
1083 rpc_final_put_task(task, task->tk_workqueue);
1084}
1085
1086int rpciod_up(void)
1087{
1088 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1089}
1090
1091void rpciod_down(void)
1092{
1093 module_put(THIS_MODULE);
1094}
1095
1096/*
1097 * Start up the rpciod workqueue.
1098 */
1099static int rpciod_start(void)
1100{
1101 struct workqueue_struct *wq;
1102
1103 /*
1104 * Create the rpciod thread and wait for it to start.
1105 */
1106 dprintk("RPC: creating workqueue rpciod\n");
1107 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1108 if (!wq)
1109 goto out_failed;
1110 rpciod_workqueue = wq;
1111 /* Note: highpri because network receive is latency sensitive */
1112 wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1113 if (!wq)
1114 goto free_rpciod;
1115 xprtiod_workqueue = wq;
1116 return 1;
1117free_rpciod:
1118 wq = rpciod_workqueue;
1119 rpciod_workqueue = NULL;
1120 destroy_workqueue(wq);
1121out_failed:
1122 return 0;
1123}
1124
1125static void rpciod_stop(void)
1126{
1127 struct workqueue_struct *wq = NULL;
1128
1129 if (rpciod_workqueue == NULL)
1130 return;
1131 dprintk("RPC: destroying workqueue rpciod\n");
1132
1133 wq = rpciod_workqueue;
1134 rpciod_workqueue = NULL;
1135 destroy_workqueue(wq);
1136 wq = xprtiod_workqueue;
1137 xprtiod_workqueue = NULL;
1138 destroy_workqueue(wq);
1139}
1140
1141void
1142rpc_destroy_mempool(void)
1143{
1144 rpciod_stop();
1145 mempool_destroy(rpc_buffer_mempool);
1146 mempool_destroy(rpc_task_mempool);
1147 kmem_cache_destroy(rpc_task_slabp);
1148 kmem_cache_destroy(rpc_buffer_slabp);
1149 rpc_destroy_wait_queue(&delay_queue);
1150}
1151
1152int
1153rpc_init_mempool(void)
1154{
1155 /*
1156 * The following is not strictly a mempool initialisation,
1157 * but there is no harm in doing it here
1158 */
1159 rpc_init_wait_queue(&delay_queue, "delayq");
1160 if (!rpciod_start())
1161 goto err_nomem;
1162
1163 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1164 sizeof(struct rpc_task),
1165 0, SLAB_HWCACHE_ALIGN,
1166 NULL);
1167 if (!rpc_task_slabp)
1168 goto err_nomem;
1169 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1170 RPC_BUFFER_MAXSIZE,
1171 0, SLAB_HWCACHE_ALIGN,
1172 NULL);
1173 if (!rpc_buffer_slabp)
1174 goto err_nomem;
1175 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1176 rpc_task_slabp);
1177 if (!rpc_task_mempool)
1178 goto err_nomem;
1179 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1180 rpc_buffer_slabp);
1181 if (!rpc_buffer_mempool)
1182 goto err_nomem;
1183 return 0;
1184err_nomem:
1185 rpc_destroy_mempool();
1186 return -ENOMEM;
1187}
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 INIT_LIST_HEAD(&task->u.tk_wait.timer_list);
208 if (RPC_IS_PRIORITY(queue))
209 __rpc_add_wait_queue_priority(queue, task, queue_priority);
210 else if (RPC_IS_SWAPPER(task))
211 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
212 else
213 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
214 task->tk_waitqueue = queue;
215 queue->qlen++;
216 /* barrier matches the read in rpc_wake_up_task_queue_locked() */
217 smp_wmb();
218 rpc_set_queued(task);
219
220 dprintk("RPC: %5u added to queue %p \"%s\"\n",
221 task->tk_pid, queue, rpc_qname(queue));
222}
223
224/*
225 * Remove request from a priority queue.
226 */
227static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
228{
229 __rpc_list_dequeue_task(task);
230}
231
232/*
233 * Remove request from queue.
234 * Note: must be called with spin lock held.
235 */
236static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
237{
238 __rpc_disable_timer(queue, task);
239 if (RPC_IS_PRIORITY(queue))
240 __rpc_remove_wait_queue_priority(task);
241 else
242 list_del(&task->u.tk_wait.list);
243 queue->qlen--;
244 dprintk("RPC: %5u removed from queue %p \"%s\"\n",
245 task->tk_pid, queue, rpc_qname(queue));
246}
247
248static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
249{
250 int i;
251
252 spin_lock_init(&queue->lock);
253 for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
254 INIT_LIST_HEAD(&queue->tasks[i]);
255 queue->maxpriority = nr_queues - 1;
256 rpc_reset_waitqueue_priority(queue);
257 queue->qlen = 0;
258 queue->timer_list.expires = 0;
259 INIT_DELAYED_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
260 INIT_LIST_HEAD(&queue->timer_list.list);
261 rpc_assign_waitqueue_name(queue, qname);
262}
263
264void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
265{
266 __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
267}
268EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
269
270void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
271{
272 __rpc_init_priority_wait_queue(queue, qname, 1);
273}
274EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
275
276void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
277{
278 cancel_delayed_work_sync(&queue->timer_list.dwork);
279}
280EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
281
282static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
283{
284 freezable_schedule_unsafe();
285 if (signal_pending_state(mode, current))
286 return -ERESTARTSYS;
287 return 0;
288}
289
290#if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
291static void rpc_task_set_debuginfo(struct rpc_task *task)
292{
293 static atomic_t rpc_pid;
294
295 task->tk_pid = atomic_inc_return(&rpc_pid);
296}
297#else
298static inline void rpc_task_set_debuginfo(struct rpc_task *task)
299{
300}
301#endif
302
303static void rpc_set_active(struct rpc_task *task)
304{
305 rpc_task_set_debuginfo(task);
306 set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
307 trace_rpc_task_begin(task, NULL);
308}
309
310/*
311 * Mark an RPC call as having completed by clearing the 'active' bit
312 * and then waking up all tasks that were sleeping.
313 */
314static int rpc_complete_task(struct rpc_task *task)
315{
316 void *m = &task->tk_runstate;
317 wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
318 struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
319 unsigned long flags;
320 int ret;
321
322 trace_rpc_task_complete(task, NULL);
323
324 spin_lock_irqsave(&wq->lock, flags);
325 clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
326 ret = atomic_dec_and_test(&task->tk_count);
327 if (waitqueue_active(wq))
328 __wake_up_locked_key(wq, TASK_NORMAL, &k);
329 spin_unlock_irqrestore(&wq->lock, flags);
330 return ret;
331}
332
333/*
334 * Allow callers to wait for completion of an RPC call
335 *
336 * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
337 * to enforce taking of the wq->lock and hence avoid races with
338 * rpc_complete_task().
339 */
340int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
341{
342 if (action == NULL)
343 action = rpc_wait_bit_killable;
344 return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
345 action, TASK_KILLABLE);
346}
347EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
348
349/*
350 * Make an RPC task runnable.
351 *
352 * Note: If the task is ASYNC, and is being made runnable after sitting on an
353 * rpc_wait_queue, this must be called with the queue spinlock held to protect
354 * the wait queue operation.
355 * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
356 * which is needed to ensure that __rpc_execute() doesn't loop (due to the
357 * lockless RPC_IS_QUEUED() test) before we've had a chance to test
358 * the RPC_TASK_RUNNING flag.
359 */
360static void rpc_make_runnable(struct workqueue_struct *wq,
361 struct rpc_task *task)
362{
363 bool need_wakeup = !rpc_test_and_set_running(task);
364
365 rpc_clear_queued(task);
366 if (!need_wakeup)
367 return;
368 if (RPC_IS_ASYNC(task)) {
369 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
370 queue_work(wq, &task->u.tk_work);
371 } else
372 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
373}
374
375/*
376 * Prepare for sleeping on a wait queue.
377 * By always appending tasks to the list we ensure FIFO behavior.
378 * NB: An RPC task will only receive interrupt-driven events as long
379 * as it's on a wait queue.
380 */
381static void __rpc_do_sleep_on_priority(struct rpc_wait_queue *q,
382 struct rpc_task *task,
383 unsigned char queue_priority)
384{
385 dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
386 task->tk_pid, rpc_qname(q), jiffies);
387
388 trace_rpc_task_sleep(task, q);
389
390 __rpc_add_wait_queue(q, task, queue_priority);
391
392}
393
394static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
395 struct rpc_task *task,
396 unsigned char queue_priority)
397{
398 if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
399 return;
400 __rpc_do_sleep_on_priority(q, task, queue_priority);
401}
402
403static void __rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
404 struct rpc_task *task, unsigned long timeout,
405 unsigned char queue_priority)
406{
407 if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
408 return;
409 if (time_is_after_jiffies(timeout)) {
410 __rpc_do_sleep_on_priority(q, task, queue_priority);
411 __rpc_add_timer(q, task, timeout);
412 } else
413 task->tk_status = -ETIMEDOUT;
414}
415
416static void rpc_set_tk_callback(struct rpc_task *task, rpc_action action)
417{
418 if (action && !WARN_ON_ONCE(task->tk_callback != NULL))
419 task->tk_callback = action;
420}
421
422static bool rpc_sleep_check_activated(struct rpc_task *task)
423{
424 /* We shouldn't ever put an inactive task to sleep */
425 if (WARN_ON_ONCE(!RPC_IS_ACTIVATED(task))) {
426 task->tk_status = -EIO;
427 rpc_put_task_async(task);
428 return false;
429 }
430 return true;
431}
432
433void rpc_sleep_on_timeout(struct rpc_wait_queue *q, struct rpc_task *task,
434 rpc_action action, unsigned long timeout)
435{
436 if (!rpc_sleep_check_activated(task))
437 return;
438
439 rpc_set_tk_callback(task, action);
440
441 /*
442 * Protect the queue operations.
443 */
444 spin_lock(&q->lock);
445 __rpc_sleep_on_priority_timeout(q, task, timeout, task->tk_priority);
446 spin_unlock(&q->lock);
447}
448EXPORT_SYMBOL_GPL(rpc_sleep_on_timeout);
449
450void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
451 rpc_action action)
452{
453 if (!rpc_sleep_check_activated(task))
454 return;
455
456 rpc_set_tk_callback(task, action);
457
458 WARN_ON_ONCE(task->tk_timeout != 0);
459 /*
460 * Protect the queue operations.
461 */
462 spin_lock(&q->lock);
463 __rpc_sleep_on_priority(q, task, task->tk_priority);
464 spin_unlock(&q->lock);
465}
466EXPORT_SYMBOL_GPL(rpc_sleep_on);
467
468void rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
469 struct rpc_task *task, unsigned long timeout, int priority)
470{
471 if (!rpc_sleep_check_activated(task))
472 return;
473
474 priority -= RPC_PRIORITY_LOW;
475 /*
476 * Protect the queue operations.
477 */
478 spin_lock(&q->lock);
479 __rpc_sleep_on_priority_timeout(q, task, timeout, priority);
480 spin_unlock(&q->lock);
481}
482EXPORT_SYMBOL_GPL(rpc_sleep_on_priority_timeout);
483
484void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
485 int priority)
486{
487 if (!rpc_sleep_check_activated(task))
488 return;
489
490 WARN_ON_ONCE(task->tk_timeout != 0);
491 priority -= RPC_PRIORITY_LOW;
492 /*
493 * Protect the queue operations.
494 */
495 spin_lock(&q->lock);
496 __rpc_sleep_on_priority(q, task, priority);
497 spin_unlock(&q->lock);
498}
499EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
500
501/**
502 * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
503 * @wq: workqueue on which to run task
504 * @queue: wait queue
505 * @task: task to be woken up
506 *
507 * Caller must hold queue->lock, and have cleared the task queued flag.
508 */
509static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
510 struct rpc_wait_queue *queue,
511 struct rpc_task *task)
512{
513 dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
514 task->tk_pid, jiffies);
515
516 /* Has the task been executed yet? If not, we cannot wake it up! */
517 if (!RPC_IS_ACTIVATED(task)) {
518 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
519 return;
520 }
521
522 trace_rpc_task_wakeup(task, queue);
523
524 __rpc_remove_wait_queue(queue, task);
525
526 rpc_make_runnable(wq, task);
527
528 dprintk("RPC: __rpc_wake_up_task done\n");
529}
530
531/*
532 * Wake up a queued task while the queue lock is being held
533 */
534static struct rpc_task *
535rpc_wake_up_task_on_wq_queue_action_locked(struct workqueue_struct *wq,
536 struct rpc_wait_queue *queue, struct rpc_task *task,
537 bool (*action)(struct rpc_task *, void *), void *data)
538{
539 if (RPC_IS_QUEUED(task)) {
540 smp_rmb();
541 if (task->tk_waitqueue == queue) {
542 if (action == NULL || action(task, data)) {
543 __rpc_do_wake_up_task_on_wq(wq, queue, task);
544 return task;
545 }
546 }
547 }
548 return NULL;
549}
550
551/*
552 * Wake up a queued task while the queue lock is being held
553 */
554static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue,
555 struct rpc_task *task)
556{
557 rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
558 task, NULL, NULL);
559}
560
561/*
562 * Wake up a task on a specific queue
563 */
564void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
565{
566 if (!RPC_IS_QUEUED(task))
567 return;
568 spin_lock(&queue->lock);
569 rpc_wake_up_task_queue_locked(queue, task);
570 spin_unlock(&queue->lock);
571}
572EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
573
574static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
575{
576 task->tk_status = *(int *)status;
577 return true;
578}
579
580static void
581rpc_wake_up_task_queue_set_status_locked(struct rpc_wait_queue *queue,
582 struct rpc_task *task, int status)
583{
584 rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
585 task, rpc_task_action_set_status, &status);
586}
587
588/**
589 * rpc_wake_up_queued_task_set_status - wake up a task and set task->tk_status
590 * @queue: pointer to rpc_wait_queue
591 * @task: pointer to rpc_task
592 * @status: integer error value
593 *
594 * If @task is queued on @queue, then it is woken up, and @task->tk_status is
595 * set to the value of @status.
596 */
597void
598rpc_wake_up_queued_task_set_status(struct rpc_wait_queue *queue,
599 struct rpc_task *task, int status)
600{
601 if (!RPC_IS_QUEUED(task))
602 return;
603 spin_lock(&queue->lock);
604 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
605 spin_unlock(&queue->lock);
606}
607
608/*
609 * Wake up the next task on a priority queue.
610 */
611static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
612{
613 struct list_head *q;
614 struct rpc_task *task;
615
616 /*
617 * Service a batch of tasks from a single owner.
618 */
619 q = &queue->tasks[queue->priority];
620 if (!list_empty(q) && --queue->nr) {
621 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
622 goto out;
623 }
624
625 /*
626 * Service the next queue.
627 */
628 do {
629 if (q == &queue->tasks[0])
630 q = &queue->tasks[queue->maxpriority];
631 else
632 q = q - 1;
633 if (!list_empty(q)) {
634 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
635 goto new_queue;
636 }
637 } while (q != &queue->tasks[queue->priority]);
638
639 rpc_reset_waitqueue_priority(queue);
640 return NULL;
641
642new_queue:
643 rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
644out:
645 return task;
646}
647
648static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
649{
650 if (RPC_IS_PRIORITY(queue))
651 return __rpc_find_next_queued_priority(queue);
652 if (!list_empty(&queue->tasks[0]))
653 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
654 return NULL;
655}
656
657/*
658 * Wake up the first task on the wait queue.
659 */
660struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
661 struct rpc_wait_queue *queue,
662 bool (*func)(struct rpc_task *, void *), void *data)
663{
664 struct rpc_task *task = NULL;
665
666 dprintk("RPC: wake_up_first(%p \"%s\")\n",
667 queue, rpc_qname(queue));
668 spin_lock(&queue->lock);
669 task = __rpc_find_next_queued(queue);
670 if (task != NULL)
671 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
672 task, func, data);
673 spin_unlock(&queue->lock);
674
675 return task;
676}
677
678/*
679 * Wake up the first task on the wait queue.
680 */
681struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
682 bool (*func)(struct rpc_task *, void *), void *data)
683{
684 return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
685}
686EXPORT_SYMBOL_GPL(rpc_wake_up_first);
687
688static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
689{
690 return true;
691}
692
693/*
694 * Wake up the next task on the wait queue.
695*/
696struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
697{
698 return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
699}
700EXPORT_SYMBOL_GPL(rpc_wake_up_next);
701
702/**
703 * rpc_wake_up - wake up all rpc_tasks
704 * @queue: rpc_wait_queue on which the tasks are sleeping
705 *
706 * Grabs queue->lock
707 */
708void rpc_wake_up(struct rpc_wait_queue *queue)
709{
710 struct list_head *head;
711
712 spin_lock(&queue->lock);
713 head = &queue->tasks[queue->maxpriority];
714 for (;;) {
715 while (!list_empty(head)) {
716 struct rpc_task *task;
717 task = list_first_entry(head,
718 struct rpc_task,
719 u.tk_wait.list);
720 rpc_wake_up_task_queue_locked(queue, task);
721 }
722 if (head == &queue->tasks[0])
723 break;
724 head--;
725 }
726 spin_unlock(&queue->lock);
727}
728EXPORT_SYMBOL_GPL(rpc_wake_up);
729
730/**
731 * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
732 * @queue: rpc_wait_queue on which the tasks are sleeping
733 * @status: status value to set
734 *
735 * Grabs queue->lock
736 */
737void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
738{
739 struct list_head *head;
740
741 spin_lock(&queue->lock);
742 head = &queue->tasks[queue->maxpriority];
743 for (;;) {
744 while (!list_empty(head)) {
745 struct rpc_task *task;
746 task = list_first_entry(head,
747 struct rpc_task,
748 u.tk_wait.list);
749 task->tk_status = status;
750 rpc_wake_up_task_queue_locked(queue, task);
751 }
752 if (head == &queue->tasks[0])
753 break;
754 head--;
755 }
756 spin_unlock(&queue->lock);
757}
758EXPORT_SYMBOL_GPL(rpc_wake_up_status);
759
760static void __rpc_queue_timer_fn(struct work_struct *work)
761{
762 struct rpc_wait_queue *queue = container_of(work,
763 struct rpc_wait_queue,
764 timer_list.dwork.work);
765 struct rpc_task *task, *n;
766 unsigned long expires, now, timeo;
767
768 spin_lock(&queue->lock);
769 expires = now = jiffies;
770 list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
771 timeo = task->tk_timeout;
772 if (time_after_eq(now, timeo)) {
773 dprintk("RPC: %5u timeout\n", task->tk_pid);
774 task->tk_status = -ETIMEDOUT;
775 rpc_wake_up_task_queue_locked(queue, task);
776 continue;
777 }
778 if (expires == now || time_after(expires, timeo))
779 expires = timeo;
780 }
781 if (!list_empty(&queue->timer_list.list))
782 rpc_set_queue_timer(queue, expires);
783 spin_unlock(&queue->lock);
784}
785
786static void __rpc_atrun(struct rpc_task *task)
787{
788 if (task->tk_status == -ETIMEDOUT)
789 task->tk_status = 0;
790}
791
792/*
793 * Run a task at a later time
794 */
795void rpc_delay(struct rpc_task *task, unsigned long delay)
796{
797 rpc_sleep_on_timeout(&delay_queue, task, __rpc_atrun, jiffies + delay);
798}
799EXPORT_SYMBOL_GPL(rpc_delay);
800
801/*
802 * Helper to call task->tk_ops->rpc_call_prepare
803 */
804void rpc_prepare_task(struct rpc_task *task)
805{
806 task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
807}
808
809static void
810rpc_init_task_statistics(struct rpc_task *task)
811{
812 /* Initialize retry counters */
813 task->tk_garb_retry = 2;
814 task->tk_cred_retry = 2;
815 task->tk_rebind_retry = 2;
816
817 /* starting timestamp */
818 task->tk_start = ktime_get();
819}
820
821static void
822rpc_reset_task_statistics(struct rpc_task *task)
823{
824 task->tk_timeouts = 0;
825 task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
826 rpc_init_task_statistics(task);
827}
828
829/*
830 * Helper that calls task->tk_ops->rpc_call_done if it exists
831 */
832void rpc_exit_task(struct rpc_task *task)
833{
834 trace_rpc_task_end(task, task->tk_action);
835 task->tk_action = NULL;
836 if (task->tk_ops->rpc_count_stats)
837 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
838 else if (task->tk_client)
839 rpc_count_iostats(task, task->tk_client->cl_metrics);
840 if (task->tk_ops->rpc_call_done != NULL) {
841 task->tk_ops->rpc_call_done(task, task->tk_calldata);
842 if (task->tk_action != NULL) {
843 /* Always release the RPC slot and buffer memory */
844 xprt_release(task);
845 rpc_reset_task_statistics(task);
846 }
847 }
848}
849
850void rpc_signal_task(struct rpc_task *task)
851{
852 struct rpc_wait_queue *queue;
853
854 if (!RPC_IS_ACTIVATED(task))
855 return;
856
857 trace_rpc_task_signalled(task, task->tk_action);
858 set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
859 smp_mb__after_atomic();
860 queue = READ_ONCE(task->tk_waitqueue);
861 if (queue)
862 rpc_wake_up_queued_task_set_status(queue, task, -ERESTARTSYS);
863}
864
865void rpc_exit(struct rpc_task *task, int status)
866{
867 task->tk_status = status;
868 task->tk_action = rpc_exit_task;
869 rpc_wake_up_queued_task(task->tk_waitqueue, task);
870}
871EXPORT_SYMBOL_GPL(rpc_exit);
872
873void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
874{
875 if (ops->rpc_release != NULL)
876 ops->rpc_release(calldata);
877}
878
879/*
880 * This is the RPC `scheduler' (or rather, the finite state machine).
881 */
882static void __rpc_execute(struct rpc_task *task)
883{
884 struct rpc_wait_queue *queue;
885 int task_is_async = RPC_IS_ASYNC(task);
886 int status = 0;
887
888 dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
889 task->tk_pid, task->tk_flags);
890
891 WARN_ON_ONCE(RPC_IS_QUEUED(task));
892 if (RPC_IS_QUEUED(task))
893 return;
894
895 for (;;) {
896 void (*do_action)(struct rpc_task *);
897
898 /*
899 * Perform the next FSM step or a pending callback.
900 *
901 * tk_action may be NULL if the task has been killed.
902 * In particular, note that rpc_killall_tasks may
903 * do this at any time, so beware when dereferencing.
904 */
905 do_action = task->tk_action;
906 if (task->tk_callback) {
907 do_action = task->tk_callback;
908 task->tk_callback = NULL;
909 }
910 if (!do_action)
911 break;
912 trace_rpc_task_run_action(task, do_action);
913 do_action(task);
914
915 /*
916 * Lockless check for whether task is sleeping or not.
917 */
918 if (!RPC_IS_QUEUED(task))
919 continue;
920
921 /*
922 * Signalled tasks should exit rather than sleep.
923 */
924 if (RPC_SIGNALLED(task)) {
925 task->tk_rpc_status = -ERESTARTSYS;
926 rpc_exit(task, -ERESTARTSYS);
927 }
928
929 /*
930 * The queue->lock protects against races with
931 * rpc_make_runnable().
932 *
933 * Note that once we clear RPC_TASK_RUNNING on an asynchronous
934 * rpc_task, rpc_make_runnable() can assign it to a
935 * different workqueue. We therefore cannot assume that the
936 * rpc_task pointer may still be dereferenced.
937 */
938 queue = task->tk_waitqueue;
939 spin_lock(&queue->lock);
940 if (!RPC_IS_QUEUED(task)) {
941 spin_unlock(&queue->lock);
942 continue;
943 }
944 rpc_clear_running(task);
945 spin_unlock(&queue->lock);
946 if (task_is_async)
947 return;
948
949 /* sync task: sleep here */
950 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
951 status = out_of_line_wait_on_bit(&task->tk_runstate,
952 RPC_TASK_QUEUED, rpc_wait_bit_killable,
953 TASK_KILLABLE);
954 if (status < 0) {
955 /*
956 * When a sync task receives a signal, it exits with
957 * -ERESTARTSYS. In order to catch any callbacks that
958 * clean up after sleeping on some queue, we don't
959 * break the loop here, but go around once more.
960 */
961 trace_rpc_task_signalled(task, task->tk_action);
962 set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
963 task->tk_rpc_status = -ERESTARTSYS;
964 rpc_exit(task, -ERESTARTSYS);
965 }
966 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
967 }
968
969 dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
970 task->tk_status);
971 /* Release all resources associated with the task */
972 rpc_release_task(task);
973}
974
975/*
976 * User-visible entry point to the scheduler.
977 *
978 * This may be called recursively if e.g. an async NFS task updates
979 * the attributes and finds that dirty pages must be flushed.
980 * NOTE: Upon exit of this function the task is guaranteed to be
981 * released. In particular note that tk_release() will have
982 * been called, so your task memory may have been freed.
983 */
984void rpc_execute(struct rpc_task *task)
985{
986 bool is_async = RPC_IS_ASYNC(task);
987
988 rpc_set_active(task);
989 rpc_make_runnable(rpciod_workqueue, task);
990 if (!is_async)
991 __rpc_execute(task);
992}
993
994static void rpc_async_schedule(struct work_struct *work)
995{
996 unsigned int pflags = memalloc_nofs_save();
997
998 __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
999 memalloc_nofs_restore(pflags);
1000}
1001
1002/**
1003 * rpc_malloc - allocate RPC buffer resources
1004 * @task: RPC task
1005 *
1006 * A single memory region is allocated, which is split between the
1007 * RPC call and RPC reply that this task is being used for. When
1008 * this RPC is retired, the memory is released by calling rpc_free.
1009 *
1010 * To prevent rpciod from hanging, this allocator never sleeps,
1011 * returning -ENOMEM and suppressing warning if the request cannot
1012 * be serviced immediately. The caller can arrange to sleep in a
1013 * way that is safe for rpciod.
1014 *
1015 * Most requests are 'small' (under 2KiB) and can be serviced from a
1016 * mempool, ensuring that NFS reads and writes can always proceed,
1017 * and that there is good locality of reference for these buffers.
1018 */
1019int rpc_malloc(struct rpc_task *task)
1020{
1021 struct rpc_rqst *rqst = task->tk_rqstp;
1022 size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
1023 struct rpc_buffer *buf;
1024 gfp_t gfp = GFP_NOFS;
1025
1026 if (RPC_IS_SWAPPER(task))
1027 gfp = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
1028
1029 size += sizeof(struct rpc_buffer);
1030 if (size <= RPC_BUFFER_MAXSIZE)
1031 buf = mempool_alloc(rpc_buffer_mempool, gfp);
1032 else
1033 buf = kmalloc(size, gfp);
1034
1035 if (!buf)
1036 return -ENOMEM;
1037
1038 buf->len = size;
1039 dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
1040 task->tk_pid, size, buf);
1041 rqst->rq_buffer = buf->data;
1042 rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1043 return 0;
1044}
1045EXPORT_SYMBOL_GPL(rpc_malloc);
1046
1047/**
1048 * rpc_free - free RPC buffer resources allocated via rpc_malloc
1049 * @task: RPC task
1050 *
1051 */
1052void rpc_free(struct rpc_task *task)
1053{
1054 void *buffer = task->tk_rqstp->rq_buffer;
1055 size_t size;
1056 struct rpc_buffer *buf;
1057
1058 buf = container_of(buffer, struct rpc_buffer, data);
1059 size = buf->len;
1060
1061 dprintk("RPC: freeing buffer of size %zu at %p\n",
1062 size, buf);
1063
1064 if (size <= RPC_BUFFER_MAXSIZE)
1065 mempool_free(buf, rpc_buffer_mempool);
1066 else
1067 kfree(buf);
1068}
1069EXPORT_SYMBOL_GPL(rpc_free);
1070
1071/*
1072 * Creation and deletion of RPC task structures
1073 */
1074static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1075{
1076 memset(task, 0, sizeof(*task));
1077 atomic_set(&task->tk_count, 1);
1078 task->tk_flags = task_setup_data->flags;
1079 task->tk_ops = task_setup_data->callback_ops;
1080 task->tk_calldata = task_setup_data->callback_data;
1081 INIT_LIST_HEAD(&task->tk_task);
1082
1083 task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1084 task->tk_owner = current->tgid;
1085
1086 /* Initialize workqueue for async tasks */
1087 task->tk_workqueue = task_setup_data->workqueue;
1088
1089 task->tk_xprt = rpc_task_get_xprt(task_setup_data->rpc_client,
1090 xprt_get(task_setup_data->rpc_xprt));
1091
1092 task->tk_op_cred = get_rpccred(task_setup_data->rpc_op_cred);
1093
1094 if (task->tk_ops->rpc_call_prepare != NULL)
1095 task->tk_action = rpc_prepare_task;
1096
1097 rpc_init_task_statistics(task);
1098
1099 dprintk("RPC: new task initialized, procpid %u\n",
1100 task_pid_nr(current));
1101}
1102
1103static struct rpc_task *
1104rpc_alloc_task(void)
1105{
1106 return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
1107}
1108
1109/*
1110 * Create a new task for the specified client.
1111 */
1112struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1113{
1114 struct rpc_task *task = setup_data->task;
1115 unsigned short flags = 0;
1116
1117 if (task == NULL) {
1118 task = rpc_alloc_task();
1119 flags = RPC_TASK_DYNAMIC;
1120 }
1121
1122 rpc_init_task(task, setup_data);
1123 task->tk_flags |= flags;
1124 dprintk("RPC: allocated task %p\n", task);
1125 return task;
1126}
1127
1128/*
1129 * rpc_free_task - release rpc task and perform cleanups
1130 *
1131 * Note that we free up the rpc_task _after_ rpc_release_calldata()
1132 * in order to work around a workqueue dependency issue.
1133 *
1134 * Tejun Heo states:
1135 * "Workqueue currently considers two work items to be the same if they're
1136 * on the same address and won't execute them concurrently - ie. it
1137 * makes a work item which is queued again while being executed wait
1138 * for the previous execution to complete.
1139 *
1140 * If a work function frees the work item, and then waits for an event
1141 * which should be performed by another work item and *that* work item
1142 * recycles the freed work item, it can create a false dependency loop.
1143 * There really is no reliable way to detect this short of verifying
1144 * every memory free."
1145 *
1146 */
1147static void rpc_free_task(struct rpc_task *task)
1148{
1149 unsigned short tk_flags = task->tk_flags;
1150
1151 put_rpccred(task->tk_op_cred);
1152 rpc_release_calldata(task->tk_ops, task->tk_calldata);
1153
1154 if (tk_flags & RPC_TASK_DYNAMIC) {
1155 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1156 mempool_free(task, rpc_task_mempool);
1157 }
1158}
1159
1160static void rpc_async_release(struct work_struct *work)
1161{
1162 unsigned int pflags = memalloc_nofs_save();
1163
1164 rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1165 memalloc_nofs_restore(pflags);
1166}
1167
1168static void rpc_release_resources_task(struct rpc_task *task)
1169{
1170 xprt_release(task);
1171 if (task->tk_msg.rpc_cred) {
1172 if (!(task->tk_flags & RPC_TASK_CRED_NOREF))
1173 put_cred(task->tk_msg.rpc_cred);
1174 task->tk_msg.rpc_cred = NULL;
1175 }
1176 rpc_task_release_client(task);
1177}
1178
1179static void rpc_final_put_task(struct rpc_task *task,
1180 struct workqueue_struct *q)
1181{
1182 if (q != NULL) {
1183 INIT_WORK(&task->u.tk_work, rpc_async_release);
1184 queue_work(q, &task->u.tk_work);
1185 } else
1186 rpc_free_task(task);
1187}
1188
1189static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1190{
1191 if (atomic_dec_and_test(&task->tk_count)) {
1192 rpc_release_resources_task(task);
1193 rpc_final_put_task(task, q);
1194 }
1195}
1196
1197void rpc_put_task(struct rpc_task *task)
1198{
1199 rpc_do_put_task(task, NULL);
1200}
1201EXPORT_SYMBOL_GPL(rpc_put_task);
1202
1203void rpc_put_task_async(struct rpc_task *task)
1204{
1205 rpc_do_put_task(task, task->tk_workqueue);
1206}
1207EXPORT_SYMBOL_GPL(rpc_put_task_async);
1208
1209static void rpc_release_task(struct rpc_task *task)
1210{
1211 dprintk("RPC: %5u release task\n", task->tk_pid);
1212
1213 WARN_ON_ONCE(RPC_IS_QUEUED(task));
1214
1215 rpc_release_resources_task(task);
1216
1217 /*
1218 * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1219 * so it should be safe to use task->tk_count as a test for whether
1220 * or not any other processes still hold references to our rpc_task.
1221 */
1222 if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1223 /* Wake up anyone who may be waiting for task completion */
1224 if (!rpc_complete_task(task))
1225 return;
1226 } else {
1227 if (!atomic_dec_and_test(&task->tk_count))
1228 return;
1229 }
1230 rpc_final_put_task(task, task->tk_workqueue);
1231}
1232
1233int rpciod_up(void)
1234{
1235 return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1236}
1237
1238void rpciod_down(void)
1239{
1240 module_put(THIS_MODULE);
1241}
1242
1243/*
1244 * Start up the rpciod workqueue.
1245 */
1246static int rpciod_start(void)
1247{
1248 struct workqueue_struct *wq;
1249
1250 /*
1251 * Create the rpciod thread and wait for it to start.
1252 */
1253 dprintk("RPC: creating workqueue rpciod\n");
1254 wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1255 if (!wq)
1256 goto out_failed;
1257 rpciod_workqueue = wq;
1258 /* Note: highpri because network receive is latency sensitive */
1259 wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1260 if (!wq)
1261 goto free_rpciod;
1262 xprtiod_workqueue = wq;
1263 return 1;
1264free_rpciod:
1265 wq = rpciod_workqueue;
1266 rpciod_workqueue = NULL;
1267 destroy_workqueue(wq);
1268out_failed:
1269 return 0;
1270}
1271
1272static void rpciod_stop(void)
1273{
1274 struct workqueue_struct *wq = NULL;
1275
1276 if (rpciod_workqueue == NULL)
1277 return;
1278 dprintk("RPC: destroying workqueue rpciod\n");
1279
1280 wq = rpciod_workqueue;
1281 rpciod_workqueue = NULL;
1282 destroy_workqueue(wq);
1283 wq = xprtiod_workqueue;
1284 xprtiod_workqueue = NULL;
1285 destroy_workqueue(wq);
1286}
1287
1288void
1289rpc_destroy_mempool(void)
1290{
1291 rpciod_stop();
1292 mempool_destroy(rpc_buffer_mempool);
1293 mempool_destroy(rpc_task_mempool);
1294 kmem_cache_destroy(rpc_task_slabp);
1295 kmem_cache_destroy(rpc_buffer_slabp);
1296 rpc_destroy_wait_queue(&delay_queue);
1297}
1298
1299int
1300rpc_init_mempool(void)
1301{
1302 /*
1303 * The following is not strictly a mempool initialisation,
1304 * but there is no harm in doing it here
1305 */
1306 rpc_init_wait_queue(&delay_queue, "delayq");
1307 if (!rpciod_start())
1308 goto err_nomem;
1309
1310 rpc_task_slabp = kmem_cache_create("rpc_tasks",
1311 sizeof(struct rpc_task),
1312 0, SLAB_HWCACHE_ALIGN,
1313 NULL);
1314 if (!rpc_task_slabp)
1315 goto err_nomem;
1316 rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1317 RPC_BUFFER_MAXSIZE,
1318 0, SLAB_HWCACHE_ALIGN,
1319 NULL);
1320 if (!rpc_buffer_slabp)
1321 goto err_nomem;
1322 rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1323 rpc_task_slabp);
1324 if (!rpc_task_mempool)
1325 goto err_nomem;
1326 rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1327 rpc_buffer_slabp);
1328 if (!rpc_buffer_mempool)
1329 goto err_nomem;
1330 return 0;
1331err_nomem:
1332 rpc_destroy_mempool();
1333 return -ENOMEM;
1334}