net/sunrpc/sched.c

   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)
  43 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  44 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  45 static mempool_t        *rpc_task_mempool __read_mostly;
  46 static mempool_t        *rpc_buffer_mempool __read_mostly;
  47
  48 static void                     rpc_async_schedule(struct work_struct *);
  49 static void                      rpc_release_task(struct rpc_task *task);
  50 static void __rpc_queue_timer_fn(struct work_struct *);
  51
  52 /*
  53  * RPC tasks sit here while waiting for conditions to improve.
  54  */
  55 static struct rpc_wait_queue delay_queue;
  56
  57 /*
  58  * rpciod-related stuff
  59  */
  60 struct workqueue_struct *rpciod_workqueue __read_mostly;
  61 struct workqueue_struct *xprtiod_workqueue __read_mostly;
  62 EXPORT_SYMBOL_GPL(xprtiod_workqueue);
  63
  64 unsigned long
  65 rpc_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 }
  76 EXPORT_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  */
  83 static 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
  95 static void
  96 rpc_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  */
 110 static 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
 123 static 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
 131 static 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  */
 139 static 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  */
 161 static 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  */
 186 static 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  */
 203 static 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  */
 227 static 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  */
 236 static 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
 248 static 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
 264 void 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 }
 268 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 269
 270 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 271 {
 272         __rpc_init_priority_wait_queue(queue, qname, 1);
 273 }
 274 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 275
 276 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 277 {
 278         cancel_delayed_work_sync(&queue->timer_list.dwork);
 279 }
 280 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 281
 282 static 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)
 291 static 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
 298 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 299 {
 300 }
 301 #endif
 302
 303 static 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  */
 314 static 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  */
 340 int __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 }
 347 EXPORT_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  */
 360 static 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  */
 381 static 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
 394 static 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
 403 static 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
 416 static 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
 422 static 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
 433 void 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 }
 448 EXPORT_SYMBOL_GPL(rpc_sleep_on_timeout);
 449
 450 void 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 }
 466 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 467
 468 void 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 }
 482 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority_timeout);
 483
 484 void 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 }
 499 EXPORT_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  */
 509 static 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  */
 534 static struct rpc_task *
 535 rpc_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  */
 554 static 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  */
 564 void 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 }
 572 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 573
 574 static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
 575 {
 576         task->tk_status = *(int *)status;
 577         return true;
 578 }
 579
 580 static void
 581 rpc_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  */
 597 void
 598 rpc_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  */
 611 static 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 the privileged queue.
 618          */
 619         q = &queue->tasks[RPC_NR_PRIORITY - 1];
 620         if (queue->maxpriority > RPC_PRIORITY_PRIVILEGED && !list_empty(q)) {
 621                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 622                 goto out;
 623         }
 624
 625         /*
 626          * Service a batch of tasks from a single owner.
 627          */
 628         q = &queue->tasks[queue->priority];
 629         if (!list_empty(q) && queue->nr) {
 630                 queue->nr--;
 631                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 632                 goto out;
 633         }
 634
 635         /*
 636          * Service the next queue.
 637          */
 638         do {
 639                 if (q == &queue->tasks[0])
 640                         q = &queue->tasks[queue->maxpriority];
 641                 else
 642                         q = q - 1;
 643                 if (!list_empty(q)) {
 644                         task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 645                         goto new_queue;
 646                 }
 647         } while (q != &queue->tasks[queue->priority]);
 648
 649         rpc_reset_waitqueue_priority(queue);
 650         return NULL;
 651
 652 new_queue:
 653         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 654 out:
 655         return task;
 656 }
 657
 658 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 659 {
 660         if (RPC_IS_PRIORITY(queue))
 661                 return __rpc_find_next_queued_priority(queue);
 662         if (!list_empty(&queue->tasks[0]))
 663                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 664         return NULL;
 665 }
 666
 667 /*
 668  * Wake up the first task on the wait queue.
 669  */
 670 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 671                 struct rpc_wait_queue *queue,
 672                 bool (*func)(struct rpc_task *, void *), void *data)
 673 {
 674         struct rpc_task *task = NULL;
 675
 676         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 677                         queue, rpc_qname(queue));
 678         spin_lock(&queue->lock);
 679         task = __rpc_find_next_queued(queue);
 680         if (task != NULL)
 681                 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
 682                                 task, func, data);
 683         spin_unlock(&queue->lock);
 684
 685         return task;
 686 }
 687
 688 /*
 689  * Wake up the first task on the wait queue.
 690  */
 691 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 692                 bool (*func)(struct rpc_task *, void *), void *data)
 693 {
 694         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 695 }
 696 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 697
 698 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 699 {
 700         return true;
 701 }
 702
 703 /*
 704  * Wake up the next task on the wait queue.
 705 */
 706 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 707 {
 708         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 709 }
 710 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 711
 712 /**
 713  * rpc_wake_up_locked - wake up all rpc_tasks
 714  * @queue: rpc_wait_queue on which the tasks are sleeping
 715  *
 716  */
 717 static void rpc_wake_up_locked(struct rpc_wait_queue *queue)
 718 {
 719         struct rpc_task *task;
 720
 721         for (;;) {
 722                 task = __rpc_find_next_queued(queue);
 723                 if (task == NULL)
 724                         break;
 725                 rpc_wake_up_task_queue_locked(queue, task);
 726         }
 727 }
 728
 729 /**
 730  * rpc_wake_up - wake up all rpc_tasks
 731  * @queue: rpc_wait_queue on which the tasks are sleeping
 732  *
 733  * Grabs queue->lock
 734  */
 735 void rpc_wake_up(struct rpc_wait_queue *queue)
 736 {
 737         spin_lock(&queue->lock);
 738         rpc_wake_up_locked(queue);
 739         spin_unlock(&queue->lock);
 740 }
 741 EXPORT_SYMBOL_GPL(rpc_wake_up);
 742
 743 /**
 744  * rpc_wake_up_status_locked - wake up all rpc_tasks and set their status value.
 745  * @queue: rpc_wait_queue on which the tasks are sleeping
 746  * @status: status value to set
 747  */
 748 static void rpc_wake_up_status_locked(struct rpc_wait_queue *queue, int status)
 749 {
 750         struct rpc_task *task;
 751
 752         for (;;) {
 753                 task = __rpc_find_next_queued(queue);
 754                 if (task == NULL)
 755                         break;
 756                 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
 757         }
 758 }
 759
 760 /**
 761  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 762  * @queue: rpc_wait_queue on which the tasks are sleeping
 763  * @status: status value to set
 764  *
 765  * Grabs queue->lock
 766  */
 767 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 768 {
 769         spin_lock(&queue->lock);
 770         rpc_wake_up_status_locked(queue, status);
 771         spin_unlock(&queue->lock);
 772 }
 773 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 774
 775 static void __rpc_queue_timer_fn(struct work_struct *work)
 776 {
 777         struct rpc_wait_queue *queue = container_of(work,
 778                         struct rpc_wait_queue,
 779                         timer_list.dwork.work);
 780         struct rpc_task *task, *n;
 781         unsigned long expires, now, timeo;
 782
 783         spin_lock(&queue->lock);
 784         expires = now = jiffies;
 785         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 786                 timeo = task->tk_timeout;
 787                 if (time_after_eq(now, timeo)) {
 788                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 789                         task->tk_status = -ETIMEDOUT;
 790                         rpc_wake_up_task_queue_locked(queue, task);
 791                         continue;
 792                 }
 793                 if (expires == now || time_after(expires, timeo))
 794                         expires = timeo;
 795         }
 796         if (!list_empty(&queue->timer_list.list))
 797                 rpc_set_queue_timer(queue, expires);
 798         spin_unlock(&queue->lock);
 799 }
 800
 801 static void __rpc_atrun(struct rpc_task *task)
 802 {
 803         if (task->tk_status == -ETIMEDOUT)
 804                 task->tk_status = 0;
 805 }
 806
 807 /*
 808  * Run a task at a later time
 809  */
 810 void rpc_delay(struct rpc_task *task, unsigned long delay)
 811 {
 812         rpc_sleep_on_timeout(&delay_queue, task, __rpc_atrun, jiffies + delay);
 813 }
 814 EXPORT_SYMBOL_GPL(rpc_delay);
 815
 816 /*
 817  * Helper to call task->tk_ops->rpc_call_prepare
 818  */
 819 void rpc_prepare_task(struct rpc_task *task)
 820 {
 821         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 822 }
 823
 824 static void
 825 rpc_init_task_statistics(struct rpc_task *task)
 826 {
 827         /* Initialize retry counters */
 828         task->tk_garb_retry = 2;
 829         task->tk_cred_retry = 2;
 830
 831         /* starting timestamp */
 832         task->tk_start = ktime_get();
 833 }
 834
 835 static void
 836 rpc_reset_task_statistics(struct rpc_task *task)
 837 {
 838         task->tk_timeouts = 0;
 839         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
 840         rpc_init_task_statistics(task);
 841 }
 842
 843 /*
 844  * Helper that calls task->tk_ops->rpc_call_done if it exists
 845  */
 846 void rpc_exit_task(struct rpc_task *task)
 847 {
 848         trace_rpc_task_end(task, task->tk_action);
 849         task->tk_action = NULL;
 850         if (task->tk_ops->rpc_count_stats)
 851                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
 852         else if (task->tk_client)
 853                 rpc_count_iostats(task, task->tk_client->cl_metrics);
 854         if (task->tk_ops->rpc_call_done != NULL) {
 855                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 856                 if (task->tk_action != NULL) {
 857                         /* Always release the RPC slot and buffer memory */
 858                         xprt_release(task);
 859                         rpc_reset_task_statistics(task);
 860                 }
 861         }
 862 }
 863
 864 void rpc_signal_task(struct rpc_task *task)
 865 {
 866         struct rpc_wait_queue *queue;
 867
 868         if (!RPC_IS_ACTIVATED(task))
 869                 return;
 870         set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
 871         smp_mb__after_atomic();
 872         queue = READ_ONCE(task->tk_waitqueue);
 873         if (queue)
 874                 rpc_wake_up_queued_task_set_status(queue, task, -ERESTARTSYS);
 875 }
 876
 877 void rpc_exit(struct rpc_task *task, int status)
 878 {
 879         task->tk_status = status;
 880         task->tk_action = rpc_exit_task;
 881         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 882 }
 883 EXPORT_SYMBOL_GPL(rpc_exit);
 884
 885 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 886 {
 887         if (ops->rpc_release != NULL)
 888                 ops->rpc_release(calldata);
 889 }
 890
 891 /*
 892  * This is the RPC `scheduler' (or rather, the finite state machine).
 893  */
 894 static void __rpc_execute(struct rpc_task *task)
 895 {
 896         struct rpc_wait_queue *queue;
 897         int task_is_async = RPC_IS_ASYNC(task);
 898         int status = 0;
 899
 900         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 901                         task->tk_pid, task->tk_flags);
 902
 903         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 904         if (RPC_IS_QUEUED(task))
 905                 return;
 906
 907         for (;;) {
 908                 void (*do_action)(struct rpc_task *);
 909
 910                 /*
 911                  * Perform the next FSM step or a pending callback.
 912                  *
 913                  * tk_action may be NULL if the task has been killed.
 914                  * In particular, note that rpc_killall_tasks may
 915                  * do this at any time, so beware when dereferencing.
 916                  */
 917                 do_action = task->tk_action;
 918                 if (task->tk_callback) {
 919                         do_action = task->tk_callback;
 920                         task->tk_callback = NULL;
 921                 }
 922                 if (!do_action)
 923                         break;
 924                 trace_rpc_task_run_action(task, do_action);
 925                 do_action(task);
 926
 927                 /*
 928                  * Lockless check for whether task is sleeping or not.
 929                  */
 930                 if (!RPC_IS_QUEUED(task))
 931                         continue;
 932
 933                 /*
 934                  * Signalled tasks should exit rather than sleep.
 935                  */
 936                 if (RPC_SIGNALLED(task)) {
 937                         task->tk_rpc_status = -ERESTARTSYS;
 938                         rpc_exit(task, -ERESTARTSYS);
 939                 }
 940
 941                 /*
 942                  * The queue->lock protects against races with
 943                  * rpc_make_runnable().
 944                  *
 945                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 946                  * rpc_task, rpc_make_runnable() can assign it to a
 947                  * different workqueue. We therefore cannot assume that the
 948                  * rpc_task pointer may still be dereferenced.
 949                  */
 950                 queue = task->tk_waitqueue;
 951                 spin_lock(&queue->lock);
 952                 if (!RPC_IS_QUEUED(task)) {
 953                         spin_unlock(&queue->lock);
 954                         continue;
 955                 }
 956                 rpc_clear_running(task);
 957                 spin_unlock(&queue->lock);
 958                 if (task_is_async)
 959                         return;
 960
 961                 /* sync task: sleep here */
 962                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 963                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 964                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 965                                 TASK_KILLABLE);
 966                 if (status < 0) {
 967                         /*
 968                          * When a sync task receives a signal, it exits with
 969                          * -ERESTARTSYS. In order to catch any callbacks that
 970                          * clean up after sleeping on some queue, we don't
 971                          * break the loop here, but go around once more.
 972                          */
 973                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 974                         set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
 975                         task->tk_rpc_status = -ERESTARTSYS;
 976                         rpc_exit(task, -ERESTARTSYS);
 977                 }
 978                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 979         }
 980
 981         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 982                         task->tk_status);
 983         /* Release all resources associated with the task */
 984         rpc_release_task(task);
 985 }
 986
 987 /*
 988  * User-visible entry point to the scheduler.
 989  *
 990  * This may be called recursively if e.g. an async NFS task updates
 991  * the attributes and finds that dirty pages must be flushed.
 992  * NOTE: Upon exit of this function the task is guaranteed to be
 993  *       released. In particular note that tk_release() will have
 994  *       been called, so your task memory may have been freed.
 995  */
 996 void rpc_execute(struct rpc_task *task)
 997 {
 998         bool is_async = RPC_IS_ASYNC(task);
 999
1000         rpc_set_active(task);
1001         rpc_make_runnable(rpciod_workqueue, task);
1002         if (!is_async) {
1003                 unsigned int pflags = memalloc_nofs_save();
1004                 __rpc_execute(task);
1005                 memalloc_nofs_restore(pflags);
1006         }
1007 }
1008
1009 static void rpc_async_schedule(struct work_struct *work)
1010 {
1011         unsigned int pflags = memalloc_nofs_save();
1012
1013         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
1014         memalloc_nofs_restore(pflags);
1015 }
1016
1017 /**
1018  * rpc_malloc - allocate RPC buffer resources
1019  * @task: RPC task
1020  *
1021  * A single memory region is allocated, which is split between the
1022  * RPC call and RPC reply that this task is being used for. When
1023  * this RPC is retired, the memory is released by calling rpc_free.
1024  *
1025  * To prevent rpciod from hanging, this allocator never sleeps,
1026  * returning -ENOMEM and suppressing warning if the request cannot
1027  * be serviced immediately. The caller can arrange to sleep in a
1028  * way that is safe for rpciod.
1029  *
1030  * Most requests are 'small' (under 2KiB) and can be serviced from a
1031  * mempool, ensuring that NFS reads and writes can always proceed,
1032  * and that there is good locality of reference for these buffers.
1033  */
1034 int rpc_malloc(struct rpc_task *task)
1035 {
1036         struct rpc_rqst *rqst = task->tk_rqstp;
1037         size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
1038         struct rpc_buffer *buf;
1039         gfp_t gfp = GFP_NOFS;
1040
1041         if (RPC_IS_ASYNC(task))
1042                 gfp = GFP_NOWAIT | __GFP_NOWARN;
1043         if (RPC_IS_SWAPPER(task))
1044                 gfp |= __GFP_MEMALLOC;
1045
1046         size += sizeof(struct rpc_buffer);
1047         if (size <= RPC_BUFFER_MAXSIZE)
1048                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
1049         else
1050                 buf = kmalloc(size, gfp);
1051
1052         if (!buf)
1053                 return -ENOMEM;
1054
1055         buf->len = size;
1056         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
1057                         task->tk_pid, size, buf);
1058         rqst->rq_buffer = buf->data;
1059         rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1060         return 0;
1061 }
1062 EXPORT_SYMBOL_GPL(rpc_malloc);
1063
1064 /**
1065  * rpc_free - free RPC buffer resources allocated via rpc_malloc
1066  * @task: RPC task
1067  *
1068  */
1069 void rpc_free(struct rpc_task *task)
1070 {
1071         void *buffer = task->tk_rqstp->rq_buffer;
1072         size_t size;
1073         struct rpc_buffer *buf;
1074
1075         buf = container_of(buffer, struct rpc_buffer, data);
1076         size = buf->len;
1077
1078         dprintk("RPC:       freeing buffer of size %zu at %p\n",
1079                         size, buf);
1080
1081         if (size <= RPC_BUFFER_MAXSIZE)
1082                 mempool_free(buf, rpc_buffer_mempool);
1083         else
1084                 kfree(buf);
1085 }
1086 EXPORT_SYMBOL_GPL(rpc_free);
1087
1088 /*
1089  * Creation and deletion of RPC task structures
1090  */
1091 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1092 {
1093         memset(task, 0, sizeof(*task));
1094         atomic_set(&task->tk_count, 1);
1095         task->tk_flags  = task_setup_data->flags;
1096         task->tk_ops = task_setup_data->callback_ops;
1097         task->tk_calldata = task_setup_data->callback_data;
1098         INIT_LIST_HEAD(&task->tk_task);
1099
1100         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1101         task->tk_owner = current->tgid;
1102
1103         /* Initialize workqueue for async tasks */
1104         task->tk_workqueue = task_setup_data->workqueue;
1105
1106         task->tk_xprt = rpc_task_get_xprt(task_setup_data->rpc_client,
1107                         xprt_get(task_setup_data->rpc_xprt));
1108
1109         task->tk_op_cred = get_rpccred(task_setup_data->rpc_op_cred);
1110
1111         if (task->tk_ops->rpc_call_prepare != NULL)
1112                 task->tk_action = rpc_prepare_task;
1113
1114         rpc_init_task_statistics(task);
1115
1116         dprintk("RPC:       new task initialized, procpid %u\n",
1117                                 task_pid_nr(current));
1118 }
1119
1120 static struct rpc_task *
1121 rpc_alloc_task(void)
1122 {
1123         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
1124 }
1125
1126 /*
1127  * Create a new task for the specified client.
1128  */
1129 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1130 {
1131         struct rpc_task *task = setup_data->task;
1132         unsigned short flags = 0;
1133
1134         if (task == NULL) {
1135                 task = rpc_alloc_task();
1136                 flags = RPC_TASK_DYNAMIC;
1137         }
1138
1139         rpc_init_task(task, setup_data);
1140         task->tk_flags |= flags;
1141         dprintk("RPC:       allocated task %p\n", task);
1142         return task;
1143 }
1144
1145 /*
1146  * rpc_free_task - release rpc task and perform cleanups
1147  *
1148  * Note that we free up the rpc_task _after_ rpc_release_calldata()
1149  * in order to work around a workqueue dependency issue.
1150  *
1151  * Tejun Heo states:
1152  * "Workqueue currently considers two work items to be the same if they're
1153  * on the same address and won't execute them concurrently - ie. it
1154  * makes a work item which is queued again while being executed wait
1155  * for the previous execution to complete.
1156  *
1157  * If a work function frees the work item, and then waits for an event
1158  * which should be performed by another work item and *that* work item
1159  * recycles the freed work item, it can create a false dependency loop.
1160  * There really is no reliable way to detect this short of verifying
1161  * every memory free."
1162  *
1163  */
1164 static void rpc_free_task(struct rpc_task *task)
1165 {
1166         unsigned short tk_flags = task->tk_flags;
1167
1168         put_rpccred(task->tk_op_cred);
1169         rpc_release_calldata(task->tk_ops, task->tk_calldata);
1170
1171         if (tk_flags & RPC_TASK_DYNAMIC) {
1172                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1173                 mempool_free(task, rpc_task_mempool);
1174         }
1175 }
1176
1177 static void rpc_async_release(struct work_struct *work)
1178 {
1179         unsigned int pflags = memalloc_nofs_save();
1180
1181         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1182         memalloc_nofs_restore(pflags);
1183 }
1184
1185 static void rpc_release_resources_task(struct rpc_task *task)
1186 {
1187         xprt_release(task);
1188         if (task->tk_msg.rpc_cred) {
1189                 put_cred(task->tk_msg.rpc_cred);
1190                 task->tk_msg.rpc_cred = NULL;
1191         }
1192         rpc_task_release_client(task);
1193 }
1194
1195 static void rpc_final_put_task(struct rpc_task *task,
1196                 struct workqueue_struct *q)
1197 {
1198         if (q != NULL) {
1199                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1200                 queue_work(q, &task->u.tk_work);
1201         } else
1202                 rpc_free_task(task);
1203 }
1204
1205 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1206 {
1207         if (atomic_dec_and_test(&task->tk_count)) {
1208                 rpc_release_resources_task(task);
1209                 rpc_final_put_task(task, q);
1210         }
1211 }
1212
1213 void rpc_put_task(struct rpc_task *task)
1214 {
1215         rpc_do_put_task(task, NULL);
1216 }
1217 EXPORT_SYMBOL_GPL(rpc_put_task);
1218
1219 void rpc_put_task_async(struct rpc_task *task)
1220 {
1221         rpc_do_put_task(task, task->tk_workqueue);
1222 }
1223 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1224
1225 static void rpc_release_task(struct rpc_task *task)
1226 {
1227         dprintk("RPC: %5u release task\n", task->tk_pid);
1228
1229         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1230
1231         rpc_release_resources_task(task);
1232
1233         /*
1234          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1235          * so it should be safe to use task->tk_count as a test for whether
1236          * or not any other processes still hold references to our rpc_task.
1237          */
1238         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1239                 /* Wake up anyone who may be waiting for task completion */
1240                 if (!rpc_complete_task(task))
1241                         return;
1242         } else {
1243                 if (!atomic_dec_and_test(&task->tk_count))
1244                         return;
1245         }
1246         rpc_final_put_task(task, task->tk_workqueue);
1247 }
1248
1249 int rpciod_up(void)
1250 {
1251         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1252 }
1253
1254 void rpciod_down(void)
1255 {
1256         module_put(THIS_MODULE);
1257 }
1258
1259 /*
1260  * Start up the rpciod workqueue.
1261  */
1262 static int rpciod_start(void)
1263 {
1264         struct workqueue_struct *wq;
1265
1266         /*
1267          * Create the rpciod thread and wait for it to start.
1268          */
1269         dprintk("RPC:       creating workqueue rpciod\n");
1270         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1271         if (!wq)
1272                 goto out_failed;
1273         rpciod_workqueue = wq;
1274         /* Note: highpri because network receive is latency sensitive */
1275         wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1276         if (!wq)
1277                 goto free_rpciod;
1278         xprtiod_workqueue = wq;
1279         return 1;
1280 free_rpciod:
1281         wq = rpciod_workqueue;
1282         rpciod_workqueue = NULL;
1283         destroy_workqueue(wq);
1284 out_failed:
1285         return 0;
1286 }
1287
1288 static void rpciod_stop(void)
1289 {
1290         struct workqueue_struct *wq = NULL;
1291
1292         if (rpciod_workqueue == NULL)
1293                 return;
1294         dprintk("RPC:       destroying workqueue rpciod\n");
1295
1296         wq = rpciod_workqueue;
1297         rpciod_workqueue = NULL;
1298         destroy_workqueue(wq);
1299         wq = xprtiod_workqueue;
1300         xprtiod_workqueue = NULL;
1301         destroy_workqueue(wq);
1302 }
1303
1304 void
1305 rpc_destroy_mempool(void)
1306 {
1307         rpciod_stop();
1308         mempool_destroy(rpc_buffer_mempool);
1309         mempool_destroy(rpc_task_mempool);
1310         kmem_cache_destroy(rpc_task_slabp);
1311         kmem_cache_destroy(rpc_buffer_slabp);
1312         rpc_destroy_wait_queue(&delay_queue);
1313 }
1314
1315 int
1316 rpc_init_mempool(void)
1317 {
1318         /*
1319          * The following is not strictly a mempool initialisation,
1320          * but there is no harm in doing it here
1321          */
1322         rpc_init_wait_queue(&delay_queue, "delayq");
1323         if (!rpciod_start())
1324                 goto err_nomem;
1325
1326         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1327                                              sizeof(struct rpc_task),
1328                                              0, SLAB_HWCACHE_ALIGN,
1329                                              NULL);
1330         if (!rpc_task_slabp)
1331                 goto err_nomem;
1332         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1333                                              RPC_BUFFER_MAXSIZE,
1334                                              0, SLAB_HWCACHE_ALIGN,
1335                                              NULL);
1336         if (!rpc_buffer_slabp)
1337                 goto err_nomem;
1338         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1339                                                     rpc_task_slabp);
1340         if (!rpc_task_mempool)
1341                 goto err_nomem;
1342         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1343                                                       rpc_buffer_slabp);
1344         if (!rpc_buffer_mempool)
1345                 goto err_nomem;
1346         return 0;
1347 err_nomem:
1348         rpc_destroy_mempool();
1349         return -ENOMEM;
1350 }