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 #define CREATE_TRACE_POINTS
  31 #include <trace/events/sunrpc.h>
  32
  33 /*
  34  * RPC slabs and memory pools
  35  */
  36 #define RPC_BUFFER_MAXSIZE      (2048)
  37 #define RPC_BUFFER_POOLSIZE     (8)
  38 #define RPC_TASK_POOLSIZE       (8)
  39 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  40 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  41 static mempool_t        *rpc_task_mempool __read_mostly;
  42 static mempool_t        *rpc_buffer_mempool __read_mostly;
  43
  44 static void                     rpc_async_schedule(struct work_struct *);
  45 static void                      rpc_release_task(struct rpc_task *task);
  46 static void __rpc_queue_timer_fn(struct work_struct *);
  47
  48 /*
  49  * RPC tasks sit here while waiting for conditions to improve.
  50  */
  51 static struct rpc_wait_queue delay_queue;
  52
  53 /*
  54  * rpciod-related stuff
  55  */
  56 struct workqueue_struct *rpciod_workqueue __read_mostly;
  57 struct workqueue_struct *xprtiod_workqueue __read_mostly;
  58 EXPORT_SYMBOL_GPL(xprtiod_workqueue);
  59
  60 gfp_t rpc_task_gfp_mask(void)
  61 {
  62         if (current->flags & PF_WQ_WORKER)
  63                 return GFP_KERNEL | __GFP_NORETRY | __GFP_NOWARN;
  64         return GFP_KERNEL;
  65 }
  66
  67 unsigned long
  68 rpc_task_timeout(const struct rpc_task *task)
  69 {
  70         unsigned long timeout = READ_ONCE(task->tk_timeout);
  71
  72         if (timeout != 0) {
  73                 unsigned long now = jiffies;
  74                 if (time_before(now, timeout))
  75                         return timeout - now;
  76         }
  77         return 0;
  78 }
  79 EXPORT_SYMBOL_GPL(rpc_task_timeout);
  80
  81 /*
  82  * Disable the timer for a given RPC task. Should be called with
  83  * queue->lock and bh_disabled in order to avoid races within
  84  * rpc_run_timer().
  85  */
  86 static void
  87 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  88 {
  89         if (list_empty(&task->u.tk_wait.timer_list))
  90                 return;
  91         task->tk_timeout = 0;
  92         list_del(&task->u.tk_wait.timer_list);
  93         if (list_empty(&queue->timer_list.list))
  94                 cancel_delayed_work(&queue->timer_list.dwork);
  95 }
  96
  97 static void
  98 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  99 {
 100         unsigned long now = jiffies;
 101         queue->timer_list.expires = expires;
 102         if (time_before_eq(expires, now))
 103                 expires = 0;
 104         else
 105                 expires -= now;
 106         mod_delayed_work(rpciod_workqueue, &queue->timer_list.dwork, expires);
 107 }
 108
 109 /*
 110  * Set up a timer for the current task.
 111  */
 112 static void
 113 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task,
 114                 unsigned long timeout)
 115 {
 116         task->tk_timeout = timeout;
 117         if (list_empty(&queue->timer_list.list) || time_before(timeout, queue->timer_list.expires))
 118                 rpc_set_queue_timer(queue, timeout);
 119         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
 120 }
 121
 122 static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 123 {
 124         if (queue->priority != priority) {
 125                 queue->priority = priority;
 126                 queue->nr = 1U << priority;
 127         }
 128 }
 129
 130 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 131 {
 132         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 133 }
 134
 135 /*
 136  * Add a request to a queue list
 137  */
 138 static void
 139 __rpc_list_enqueue_task(struct list_head *q, struct rpc_task *task)
 140 {
 141         struct rpc_task *t;
 142
 143         list_for_each_entry(t, q, u.tk_wait.list) {
 144                 if (t->tk_owner == task->tk_owner) {
 145                         list_add_tail(&task->u.tk_wait.links,
 146                                         &t->u.tk_wait.links);
 147                         /* Cache the queue head in task->u.tk_wait.list */
 148                         task->u.tk_wait.list.next = q;
 149                         task->u.tk_wait.list.prev = NULL;
 150                         return;
 151                 }
 152         }
 153         INIT_LIST_HEAD(&task->u.tk_wait.links);
 154         list_add_tail(&task->u.tk_wait.list, q);
 155 }
 156
 157 /*
 158  * Remove request from a queue list
 159  */
 160 static void
 161 __rpc_list_dequeue_task(struct rpc_task *task)
 162 {
 163         struct list_head *q;
 164         struct rpc_task *t;
 165
 166         if (task->u.tk_wait.list.prev == NULL) {
 167                 list_del(&task->u.tk_wait.links);
 168                 return;
 169         }
 170         if (!list_empty(&task->u.tk_wait.links)) {
 171                 t = list_first_entry(&task->u.tk_wait.links,
 172                                 struct rpc_task,
 173                                 u.tk_wait.links);
 174                 /* Assume __rpc_list_enqueue_task() cached the queue head */
 175                 q = t->u.tk_wait.list.next;
 176                 list_add_tail(&t->u.tk_wait.list, q);
 177                 list_del(&task->u.tk_wait.links);
 178         }
 179         list_del(&task->u.tk_wait.list);
 180 }
 181
 182 /*
 183  * Add new request to a priority queue.
 184  */
 185 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 186                 struct rpc_task *task,
 187                 unsigned char queue_priority)
 188 {
 189         if (unlikely(queue_priority > queue->maxpriority))
 190                 queue_priority = queue->maxpriority;
 191         __rpc_list_enqueue_task(&queue->tasks[queue_priority], task);
 192 }
 193
 194 /*
 195  * Add new request to wait queue.
 196  */
 197 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 198                 struct rpc_task *task,
 199                 unsigned char queue_priority)
 200 {
 201         INIT_LIST_HEAD(&task->u.tk_wait.timer_list);
 202         if (RPC_IS_PRIORITY(queue))
 203                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 204         else
 205                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 206         task->tk_waitqueue = queue;
 207         queue->qlen++;
 208         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 209         smp_wmb();
 210         rpc_set_queued(task);
 211 }
 212
 213 /*
 214  * Remove request from a priority queue.
 215  */
 216 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 217 {
 218         __rpc_list_dequeue_task(task);
 219 }
 220
 221 /*
 222  * Remove request from queue.
 223  * Note: must be called with spin lock held.
 224  */
 225 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 226 {
 227         __rpc_disable_timer(queue, task);
 228         if (RPC_IS_PRIORITY(queue))
 229                 __rpc_remove_wait_queue_priority(task);
 230         else
 231                 list_del(&task->u.tk_wait.list);
 232         queue->qlen--;
 233 }
 234
 235 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 236 {
 237         int i;
 238
 239         spin_lock_init(&queue->lock);
 240         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 241                 INIT_LIST_HEAD(&queue->tasks[i]);
 242         queue->maxpriority = nr_queues - 1;
 243         rpc_reset_waitqueue_priority(queue);
 244         queue->qlen = 0;
 245         queue->timer_list.expires = 0;
 246         INIT_DELAYED_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
 247         INIT_LIST_HEAD(&queue->timer_list.list);
 248         rpc_assign_waitqueue_name(queue, qname);
 249 }
 250
 251 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 252 {
 253         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 254 }
 255 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 256
 257 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 258 {
 259         __rpc_init_priority_wait_queue(queue, qname, 1);
 260 }
 261 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 262
 263 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 264 {
 265         cancel_delayed_work_sync(&queue->timer_list.dwork);
 266 }
 267 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 268
 269 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
 270 {
 271         freezable_schedule_unsafe();
 272         if (signal_pending_state(mode, current))
 273                 return -ERESTARTSYS;
 274         return 0;
 275 }
 276
 277 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
 278 static void rpc_task_set_debuginfo(struct rpc_task *task)
 279 {
 280         struct rpc_clnt *clnt = task->tk_client;
 281
 282         /* Might be a task carrying a reverse-direction operation */
 283         if (!clnt) {
 284                 static atomic_t rpc_pid;
 285
 286                 task->tk_pid = atomic_inc_return(&rpc_pid);
 287                 return;
 288         }
 289
 290         task->tk_pid = atomic_inc_return(&clnt->cl_pid);
 291 }
 292 #else
 293 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 294 {
 295 }
 296 #endif
 297
 298 static void rpc_set_active(struct rpc_task *task)
 299 {
 300         rpc_task_set_debuginfo(task);
 301         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 302         trace_rpc_task_begin(task, NULL);
 303 }
 304
 305 /*
 306  * Mark an RPC call as having completed by clearing the 'active' bit
 307  * and then waking up all tasks that were sleeping.
 308  */
 309 static int rpc_complete_task(struct rpc_task *task)
 310 {
 311         void *m = &task->tk_runstate;
 312         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 313         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 314         unsigned long flags;
 315         int ret;
 316
 317         trace_rpc_task_complete(task, NULL);
 318
 319         spin_lock_irqsave(&wq->lock, flags);
 320         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 321         ret = atomic_dec_and_test(&task->tk_count);
 322         if (waitqueue_active(wq))
 323                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 324         spin_unlock_irqrestore(&wq->lock, flags);
 325         return ret;
 326 }
 327
 328 /*
 329  * Allow callers to wait for completion of an RPC call
 330  *
 331  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 332  * to enforce taking of the wq->lock and hence avoid races with
 333  * rpc_complete_task().
 334  */
 335 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
 336 {
 337         if (action == NULL)
 338                 action = rpc_wait_bit_killable;
 339         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 340                         action, TASK_KILLABLE);
 341 }
 342 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 343
 344 /*
 345  * Make an RPC task runnable.
 346  *
 347  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 348  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 349  * the wait queue operation.
 350  * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
 351  * which is needed to ensure that __rpc_execute() doesn't loop (due to the
 352  * lockless RPC_IS_QUEUED() test) before we've had a chance to test
 353  * the RPC_TASK_RUNNING flag.
 354  */
 355 static void rpc_make_runnable(struct workqueue_struct *wq,
 356                 struct rpc_task *task)
 357 {
 358         bool need_wakeup = !rpc_test_and_set_running(task);
 359
 360         rpc_clear_queued(task);
 361         if (!need_wakeup)
 362                 return;
 363         if (RPC_IS_ASYNC(task)) {
 364                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 365                 queue_work(wq, &task->u.tk_work);
 366         } else
 367                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 368 }
 369
 370 /*
 371  * Prepare for sleeping on a wait queue.
 372  * By always appending tasks to the list we ensure FIFO behavior.
 373  * NB: An RPC task will only receive interrupt-driven events as long
 374  * as it's on a wait queue.
 375  */
 376 static void __rpc_do_sleep_on_priority(struct rpc_wait_queue *q,
 377                 struct rpc_task *task,
 378                 unsigned char queue_priority)
 379 {
 380         trace_rpc_task_sleep(task, q);
 381
 382         __rpc_add_wait_queue(q, task, queue_priority);
 383 }
 384
 385 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 386                 struct rpc_task *task,
 387                 unsigned char queue_priority)
 388 {
 389         if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
 390                 return;
 391         __rpc_do_sleep_on_priority(q, task, queue_priority);
 392 }
 393
 394 static void __rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
 395                 struct rpc_task *task, unsigned long timeout,
 396                 unsigned char queue_priority)
 397 {
 398         if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
 399                 return;
 400         if (time_is_after_jiffies(timeout)) {
 401                 __rpc_do_sleep_on_priority(q, task, queue_priority);
 402                 __rpc_add_timer(q, task, timeout);
 403         } else
 404                 task->tk_status = -ETIMEDOUT;
 405 }
 406
 407 static void rpc_set_tk_callback(struct rpc_task *task, rpc_action action)
 408 {
 409         if (action && !WARN_ON_ONCE(task->tk_callback != NULL))
 410                 task->tk_callback = action;
 411 }
 412
 413 static bool rpc_sleep_check_activated(struct rpc_task *task)
 414 {
 415         /* We shouldn't ever put an inactive task to sleep */
 416         if (WARN_ON_ONCE(!RPC_IS_ACTIVATED(task))) {
 417                 task->tk_status = -EIO;
 418                 rpc_put_task_async(task);
 419                 return false;
 420         }
 421         return true;
 422 }
 423
 424 void rpc_sleep_on_timeout(struct rpc_wait_queue *q, struct rpc_task *task,
 425                                 rpc_action action, unsigned long timeout)
 426 {
 427         if (!rpc_sleep_check_activated(task))
 428                 return;
 429
 430         rpc_set_tk_callback(task, action);
 431
 432         /*
 433          * Protect the queue operations.
 434          */
 435         spin_lock(&q->lock);
 436         __rpc_sleep_on_priority_timeout(q, task, timeout, task->tk_priority);
 437         spin_unlock(&q->lock);
 438 }
 439 EXPORT_SYMBOL_GPL(rpc_sleep_on_timeout);
 440
 441 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 442                                 rpc_action action)
 443 {
 444         if (!rpc_sleep_check_activated(task))
 445                 return;
 446
 447         rpc_set_tk_callback(task, action);
 448
 449         WARN_ON_ONCE(task->tk_timeout != 0);
 450         /*
 451          * Protect the queue operations.
 452          */
 453         spin_lock(&q->lock);
 454         __rpc_sleep_on_priority(q, task, task->tk_priority);
 455         spin_unlock(&q->lock);
 456 }
 457 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 458
 459 void rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
 460                 struct rpc_task *task, unsigned long timeout, int priority)
 461 {
 462         if (!rpc_sleep_check_activated(task))
 463                 return;
 464
 465         priority -= RPC_PRIORITY_LOW;
 466         /*
 467          * Protect the queue operations.
 468          */
 469         spin_lock(&q->lock);
 470         __rpc_sleep_on_priority_timeout(q, task, timeout, priority);
 471         spin_unlock(&q->lock);
 472 }
 473 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority_timeout);
 474
 475 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 476                 int priority)
 477 {
 478         if (!rpc_sleep_check_activated(task))
 479                 return;
 480
 481         WARN_ON_ONCE(task->tk_timeout != 0);
 482         priority -= RPC_PRIORITY_LOW;
 483         /*
 484          * Protect the queue operations.
 485          */
 486         spin_lock(&q->lock);
 487         __rpc_sleep_on_priority(q, task, priority);
 488         spin_unlock(&q->lock);
 489 }
 490 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 491
 492 /**
 493  * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
 494  * @wq: workqueue on which to run task
 495  * @queue: wait queue
 496  * @task: task to be woken up
 497  *
 498  * Caller must hold queue->lock, and have cleared the task queued flag.
 499  */
 500 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
 501                 struct rpc_wait_queue *queue,
 502                 struct rpc_task *task)
 503 {
 504         /* Has the task been executed yet? If not, we cannot wake it up! */
 505         if (!RPC_IS_ACTIVATED(task)) {
 506                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 507                 return;
 508         }
 509
 510         trace_rpc_task_wakeup(task, queue);
 511
 512         __rpc_remove_wait_queue(queue, task);
 513
 514         rpc_make_runnable(wq, task);
 515 }
 516
 517 /*
 518  * Wake up a queued task while the queue lock is being held
 519  */
 520 static struct rpc_task *
 521 rpc_wake_up_task_on_wq_queue_action_locked(struct workqueue_struct *wq,
 522                 struct rpc_wait_queue *queue, struct rpc_task *task,
 523                 bool (*action)(struct rpc_task *, void *), void *data)
 524 {
 525         if (RPC_IS_QUEUED(task)) {
 526                 smp_rmb();
 527                 if (task->tk_waitqueue == queue) {
 528                         if (action == NULL || action(task, data)) {
 529                                 __rpc_do_wake_up_task_on_wq(wq, queue, task);
 530                                 return task;
 531                         }
 532                 }
 533         }
 534         return NULL;
 535 }
 536
 537 /*
 538  * Wake up a queued task while the queue lock is being held
 539  */
 540 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue,
 541                                           struct rpc_task *task)
 542 {
 543         rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
 544                                                    task, NULL, NULL);
 545 }
 546
 547 /*
 548  * Wake up a task on a specific queue
 549  */
 550 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 551 {
 552         if (!RPC_IS_QUEUED(task))
 553                 return;
 554         spin_lock(&queue->lock);
 555         rpc_wake_up_task_queue_locked(queue, task);
 556         spin_unlock(&queue->lock);
 557 }
 558 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 559
 560 static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
 561 {
 562         task->tk_status = *(int *)status;
 563         return true;
 564 }
 565
 566 static void
 567 rpc_wake_up_task_queue_set_status_locked(struct rpc_wait_queue *queue,
 568                 struct rpc_task *task, int status)
 569 {
 570         rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
 571                         task, rpc_task_action_set_status, &status);
 572 }
 573
 574 /**
 575  * rpc_wake_up_queued_task_set_status - wake up a task and set task->tk_status
 576  * @queue: pointer to rpc_wait_queue
 577  * @task: pointer to rpc_task
 578  * @status: integer error value
 579  *
 580  * If @task is queued on @queue, then it is woken up, and @task->tk_status is
 581  * set to the value of @status.
 582  */
 583 void
 584 rpc_wake_up_queued_task_set_status(struct rpc_wait_queue *queue,
 585                 struct rpc_task *task, int status)
 586 {
 587         if (!RPC_IS_QUEUED(task))
 588                 return;
 589         spin_lock(&queue->lock);
 590         rpc_wake_up_task_queue_set_status_locked(queue, task, status);
 591         spin_unlock(&queue->lock);
 592 }
 593
 594 /*
 595  * Wake up the next task on a priority queue.
 596  */
 597 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 598 {
 599         struct list_head *q;
 600         struct rpc_task *task;
 601
 602         /*
 603          * Service the privileged queue.
 604          */
 605         q = &queue->tasks[RPC_NR_PRIORITY - 1];
 606         if (queue->maxpriority > RPC_PRIORITY_PRIVILEGED && !list_empty(q)) {
 607                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 608                 goto out;
 609         }
 610
 611         /*
 612          * Service a batch of tasks from a single owner.
 613          */
 614         q = &queue->tasks[queue->priority];
 615         if (!list_empty(q) && queue->nr) {
 616                 queue->nr--;
 617                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 618                 goto out;
 619         }
 620
 621         /*
 622          * Service the next queue.
 623          */
 624         do {
 625                 if (q == &queue->tasks[0])
 626                         q = &queue->tasks[queue->maxpriority];
 627                 else
 628                         q = q - 1;
 629                 if (!list_empty(q)) {
 630                         task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 631                         goto new_queue;
 632                 }
 633         } while (q != &queue->tasks[queue->priority]);
 634
 635         rpc_reset_waitqueue_priority(queue);
 636         return NULL;
 637
 638 new_queue:
 639         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 640 out:
 641         return task;
 642 }
 643
 644 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 645 {
 646         if (RPC_IS_PRIORITY(queue))
 647                 return __rpc_find_next_queued_priority(queue);
 648         if (!list_empty(&queue->tasks[0]))
 649                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 650         return NULL;
 651 }
 652
 653 /*
 654  * Wake up the first task on the wait queue.
 655  */
 656 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 657                 struct rpc_wait_queue *queue,
 658                 bool (*func)(struct rpc_task *, void *), void *data)
 659 {
 660         struct rpc_task *task = NULL;
 661
 662         spin_lock(&queue->lock);
 663         task = __rpc_find_next_queued(queue);
 664         if (task != NULL)
 665                 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
 666                                 task, func, data);
 667         spin_unlock(&queue->lock);
 668
 669         return task;
 670 }
 671
 672 /*
 673  * Wake up the first task on the wait queue.
 674  */
 675 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 676                 bool (*func)(struct rpc_task *, void *), void *data)
 677 {
 678         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 679 }
 680 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 681
 682 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 683 {
 684         return true;
 685 }
 686
 687 /*
 688  * Wake up the next task on the wait queue.
 689 */
 690 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 691 {
 692         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 693 }
 694 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 695
 696 /**
 697  * rpc_wake_up_locked - wake up all rpc_tasks
 698  * @queue: rpc_wait_queue on which the tasks are sleeping
 699  *
 700  */
 701 static void rpc_wake_up_locked(struct rpc_wait_queue *queue)
 702 {
 703         struct rpc_task *task;
 704
 705         for (;;) {
 706                 task = __rpc_find_next_queued(queue);
 707                 if (task == NULL)
 708                         break;
 709                 rpc_wake_up_task_queue_locked(queue, task);
 710         }
 711 }
 712
 713 /**
 714  * rpc_wake_up - wake up all rpc_tasks
 715  * @queue: rpc_wait_queue on which the tasks are sleeping
 716  *
 717  * Grabs queue->lock
 718  */
 719 void rpc_wake_up(struct rpc_wait_queue *queue)
 720 {
 721         spin_lock(&queue->lock);
 722         rpc_wake_up_locked(queue);
 723         spin_unlock(&queue->lock);
 724 }
 725 EXPORT_SYMBOL_GPL(rpc_wake_up);
 726
 727 /**
 728  * rpc_wake_up_status_locked - wake up all rpc_tasks and set their status value.
 729  * @queue: rpc_wait_queue on which the tasks are sleeping
 730  * @status: status value to set
 731  */
 732 static void rpc_wake_up_status_locked(struct rpc_wait_queue *queue, int status)
 733 {
 734         struct rpc_task *task;
 735
 736         for (;;) {
 737                 task = __rpc_find_next_queued(queue);
 738                 if (task == NULL)
 739                         break;
 740                 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
 741         }
 742 }
 743
 744 /**
 745  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 746  * @queue: rpc_wait_queue on which the tasks are sleeping
 747  * @status: status value to set
 748  *
 749  * Grabs queue->lock
 750  */
 751 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 752 {
 753         spin_lock(&queue->lock);
 754         rpc_wake_up_status_locked(queue, status);
 755         spin_unlock(&queue->lock);
 756 }
 757 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 758
 759 static void __rpc_queue_timer_fn(struct work_struct *work)
 760 {
 761         struct rpc_wait_queue *queue = container_of(work,
 762                         struct rpc_wait_queue,
 763                         timer_list.dwork.work);
 764         struct rpc_task *task, *n;
 765         unsigned long expires, now, timeo;
 766
 767         spin_lock(&queue->lock);
 768         expires = now = jiffies;
 769         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 770                 timeo = task->tk_timeout;
 771                 if (time_after_eq(now, timeo)) {
 772                         trace_rpc_task_timeout(task, task->tk_action);
 773                         task->tk_status = -ETIMEDOUT;
 774                         rpc_wake_up_task_queue_locked(queue, task);
 775                         continue;
 776                 }
 777                 if (expires == now || time_after(expires, timeo))
 778                         expires = timeo;
 779         }
 780         if (!list_empty(&queue->timer_list.list))
 781                 rpc_set_queue_timer(queue, expires);
 782         spin_unlock(&queue->lock);
 783 }
 784
 785 static void __rpc_atrun(struct rpc_task *task)
 786 {
 787         if (task->tk_status == -ETIMEDOUT)
 788                 task->tk_status = 0;
 789 }
 790
 791 /*
 792  * Run a task at a later time
 793  */
 794 void rpc_delay(struct rpc_task *task, unsigned long delay)
 795 {
 796         rpc_sleep_on_timeout(&delay_queue, task, __rpc_atrun, jiffies + delay);
 797 }
 798 EXPORT_SYMBOL_GPL(rpc_delay);
 799
 800 /*
 801  * Helper to call task->tk_ops->rpc_call_prepare
 802  */
 803 void rpc_prepare_task(struct rpc_task *task)
 804 {
 805         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 806 }
 807
 808 static void
 809 rpc_init_task_statistics(struct rpc_task *task)
 810 {
 811         /* Initialize retry counters */
 812         task->tk_garb_retry = 2;
 813         task->tk_cred_retry = 2;
 814         task->tk_rebind_retry = 2;
 815
 816         /* starting timestamp */
 817         task->tk_start = ktime_get();
 818 }
 819
 820 static void
 821 rpc_reset_task_statistics(struct rpc_task *task)
 822 {
 823         task->tk_timeouts = 0;
 824         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
 825         rpc_init_task_statistics(task);
 826 }
 827
 828 /*
 829  * Helper that calls task->tk_ops->rpc_call_done if it exists
 830  */
 831 void rpc_exit_task(struct rpc_task *task)
 832 {
 833         trace_rpc_task_end(task, task->tk_action);
 834         task->tk_action = NULL;
 835         if (task->tk_ops->rpc_count_stats)
 836                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
 837         else if (task->tk_client)
 838                 rpc_count_iostats(task, task->tk_client->cl_metrics);
 839         if (task->tk_ops->rpc_call_done != NULL) {
 840                 trace_rpc_task_call_done(task, task->tk_ops->rpc_call_done);
 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
 850 void 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
 865 void 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 }
 871 EXPORT_SYMBOL_GPL(rpc_exit);
 872
 873 void 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 static bool xprt_needs_memalloc(struct rpc_xprt *xprt, struct rpc_task *tk)
 880 {
 881         if (!xprt)
 882                 return false;
 883         if (!atomic_read(&xprt->swapper))
 884                 return false;
 885         return test_bit(XPRT_LOCKED, &xprt->state) && xprt->snd_task == tk;
 886 }
 887
 888 /*
 889  * This is the RPC `scheduler' (or rather, the finite state machine).
 890  */
 891 static void __rpc_execute(struct rpc_task *task)
 892 {
 893         struct rpc_wait_queue *queue;
 894         int task_is_async = RPC_IS_ASYNC(task);
 895         int status = 0;
 896         unsigned long pflags = current->flags;
 897
 898         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 899         if (RPC_IS_QUEUED(task))
 900                 return;
 901
 902         for (;;) {
 903                 void (*do_action)(struct rpc_task *);
 904
 905                 /*
 906                  * Perform the next FSM step or a pending callback.
 907                  *
 908                  * tk_action may be NULL if the task has been killed.
 909                  * In particular, note that rpc_killall_tasks may
 910                  * do this at any time, so beware when dereferencing.
 911                  */
 912                 do_action = task->tk_action;
 913                 if (task->tk_callback) {
 914                         do_action = task->tk_callback;
 915                         task->tk_callback = NULL;
 916                 }
 917                 if (!do_action)
 918                         break;
 919                 if (RPC_IS_SWAPPER(task) ||
 920                     xprt_needs_memalloc(task->tk_xprt, task))
 921                         current->flags |= PF_MEMALLOC;
 922
 923                 trace_rpc_task_run_action(task, do_action);
 924                 do_action(task);
 925
 926                 /*
 927                  * Lockless check for whether task is sleeping or not.
 928                  */
 929                 if (!RPC_IS_QUEUED(task)) {
 930                         cond_resched();
 931                         continue;
 932                 }
 933
 934                 /*
 935                  * Signalled tasks should exit rather than sleep.
 936                  */
 937                 if (RPC_SIGNALLED(task)) {
 938                         task->tk_rpc_status = -ERESTARTSYS;
 939                         rpc_exit(task, -ERESTARTSYS);
 940                 }
 941
 942                 /*
 943                  * The queue->lock protects against races with
 944                  * rpc_make_runnable().
 945                  *
 946                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 947                  * rpc_task, rpc_make_runnable() can assign it to a
 948                  * different workqueue. We therefore cannot assume that the
 949                  * rpc_task pointer may still be dereferenced.
 950                  */
 951                 queue = task->tk_waitqueue;
 952                 spin_lock(&queue->lock);
 953                 if (!RPC_IS_QUEUED(task)) {
 954                         spin_unlock(&queue->lock);
 955                         continue;
 956                 }
 957                 rpc_clear_running(task);
 958                 spin_unlock(&queue->lock);
 959                 if (task_is_async)
 960                         goto out;
 961
 962                 /* sync task: sleep here */
 963                 trace_rpc_task_sync_sleep(task, task->tk_action);
 964                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 965                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 966                                 TASK_KILLABLE);
 967                 if (status < 0) {
 968                         /*
 969                          * When a sync task receives a signal, it exits with
 970                          * -ERESTARTSYS. In order to catch any callbacks that
 971                          * clean up after sleeping on some queue, we don't
 972                          * break the loop here, but go around once more.
 973                          */
 974                         trace_rpc_task_signalled(task, task->tk_action);
 975                         set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
 976                         task->tk_rpc_status = -ERESTARTSYS;
 977                         rpc_exit(task, -ERESTARTSYS);
 978                 }
 979                 trace_rpc_task_sync_wake(task, task->tk_action);
 980         }
 981
 982         /* Release all resources associated with the task */
 983         rpc_release_task(task);
 984 out:
 985         current_restore_flags(pflags, PF_MEMALLOC);
 986 }
 987
 988 /*
 989  * User-visible entry point to the scheduler.
 990  *
 991  * This may be called recursively if e.g. an async NFS task updates
 992  * the attributes and finds that dirty pages must be flushed.
 993  * NOTE: Upon exit of this function the task is guaranteed to be
 994  *       released. In particular note that tk_release() will have
 995  *       been called, so your task memory may have been freed.
 996  */
 997 void rpc_execute(struct rpc_task *task)
 998 {
 999         bool is_async = RPC_IS_ASYNC(task);
1000
1001         rpc_set_active(task);
1002         rpc_make_runnable(rpciod_workqueue, task);
1003         if (!is_async) {
1004                 unsigned int pflags = memalloc_nofs_save();
1005                 __rpc_execute(task);
1006                 memalloc_nofs_restore(pflags);
1007         }
1008 }
1009
1010 static void rpc_async_schedule(struct work_struct *work)
1011 {
1012         unsigned int pflags = memalloc_nofs_save();
1013
1014         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
1015         memalloc_nofs_restore(pflags);
1016 }
1017
1018 /**
1019  * rpc_malloc - allocate RPC buffer resources
1020  * @task: RPC task
1021  *
1022  * A single memory region is allocated, which is split between the
1023  * RPC call and RPC reply that this task is being used for. When
1024  * this RPC is retired, the memory is released by calling rpc_free.
1025  *
1026  * To prevent rpciod from hanging, this allocator never sleeps,
1027  * returning -ENOMEM and suppressing warning if the request cannot
1028  * be serviced immediately. The caller can arrange to sleep in a
1029  * way that is safe for rpciod.
1030  *
1031  * Most requests are 'small' (under 2KiB) and can be serviced from a
1032  * mempool, ensuring that NFS reads and writes can always proceed,
1033  * and that there is good locality of reference for these buffers.
1034  */
1035 int rpc_malloc(struct rpc_task *task)
1036 {
1037         struct rpc_rqst *rqst = task->tk_rqstp;
1038         size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
1039         struct rpc_buffer *buf;
1040         gfp_t gfp = rpc_task_gfp_mask();
1041
1042         size += sizeof(struct rpc_buffer);
1043         if (size <= RPC_BUFFER_MAXSIZE) {
1044                 buf = kmem_cache_alloc(rpc_buffer_slabp, gfp);
1045                 /* Reach for the mempool if dynamic allocation fails */
1046                 if (!buf && RPC_IS_ASYNC(task))
1047                         buf = mempool_alloc(rpc_buffer_mempool, GFP_NOWAIT);
1048         } else
1049                 buf = kmalloc(size, gfp);
1050
1051         if (!buf)
1052                 return -ENOMEM;
1053
1054         buf->len = size;
1055         rqst->rq_buffer = buf->data;
1056         rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1057         return 0;
1058 }
1059 EXPORT_SYMBOL_GPL(rpc_malloc);
1060
1061 /**
1062  * rpc_free - free RPC buffer resources allocated via rpc_malloc
1063  * @task: RPC task
1064  *
1065  */
1066 void rpc_free(struct rpc_task *task)
1067 {
1068         void *buffer = task->tk_rqstp->rq_buffer;
1069         size_t size;
1070         struct rpc_buffer *buf;
1071
1072         buf = container_of(buffer, struct rpc_buffer, data);
1073         size = buf->len;
1074
1075         if (size <= RPC_BUFFER_MAXSIZE)
1076                 mempool_free(buf, rpc_buffer_mempool);
1077         else
1078                 kfree(buf);
1079 }
1080 EXPORT_SYMBOL_GPL(rpc_free);
1081
1082 /*
1083  * Creation and deletion of RPC task structures
1084  */
1085 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1086 {
1087         memset(task, 0, sizeof(*task));
1088         atomic_set(&task->tk_count, 1);
1089         task->tk_flags  = task_setup_data->flags;
1090         task->tk_ops = task_setup_data->callback_ops;
1091         task->tk_calldata = task_setup_data->callback_data;
1092         INIT_LIST_HEAD(&task->tk_task);
1093
1094         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1095         task->tk_owner = current->tgid;
1096
1097         /* Initialize workqueue for async tasks */
1098         task->tk_workqueue = task_setup_data->workqueue;
1099
1100         task->tk_xprt = rpc_task_get_xprt(task_setup_data->rpc_client,
1101                         xprt_get(task_setup_data->rpc_xprt));
1102
1103         task->tk_op_cred = get_rpccred(task_setup_data->rpc_op_cred);
1104
1105         if (task->tk_ops->rpc_call_prepare != NULL)
1106                 task->tk_action = rpc_prepare_task;
1107
1108         rpc_init_task_statistics(task);
1109 }
1110
1111 static struct rpc_task *rpc_alloc_task(void)
1112 {
1113         struct rpc_task *task;
1114
1115         task = kmem_cache_alloc(rpc_task_slabp, rpc_task_gfp_mask());
1116         if (task)
1117                 return task;
1118         return mempool_alloc(rpc_task_mempool, GFP_NOWAIT);
1119 }
1120
1121 /*
1122  * Create a new task for the specified client.
1123  */
1124 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1125 {
1126         struct rpc_task *task = setup_data->task;
1127         unsigned short flags = 0;
1128
1129         if (task == NULL) {
1130                 task = rpc_alloc_task();
1131                 flags = RPC_TASK_DYNAMIC;
1132         }
1133
1134         rpc_init_task(task, setup_data);
1135         task->tk_flags |= flags;
1136         return task;
1137 }
1138
1139 /*
1140  * rpc_free_task - release rpc task and perform cleanups
1141  *
1142  * Note that we free up the rpc_task _after_ rpc_release_calldata()
1143  * in order to work around a workqueue dependency issue.
1144  *
1145  * Tejun Heo states:
1146  * "Workqueue currently considers two work items to be the same if they're
1147  * on the same address and won't execute them concurrently - ie. it
1148  * makes a work item which is queued again while being executed wait
1149  * for the previous execution to complete.
1150  *
1151  * If a work function frees the work item, and then waits for an event
1152  * which should be performed by another work item and *that* work item
1153  * recycles the freed work item, it can create a false dependency loop.
1154  * There really is no reliable way to detect this short of verifying
1155  * every memory free."
1156  *
1157  */
1158 static void rpc_free_task(struct rpc_task *task)
1159 {
1160         unsigned short tk_flags = task->tk_flags;
1161
1162         put_rpccred(task->tk_op_cred);
1163         rpc_release_calldata(task->tk_ops, task->tk_calldata);
1164
1165         if (tk_flags & RPC_TASK_DYNAMIC)
1166                 mempool_free(task, rpc_task_mempool);
1167 }
1168
1169 static void rpc_async_release(struct work_struct *work)
1170 {
1171         unsigned int pflags = memalloc_nofs_save();
1172
1173         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1174         memalloc_nofs_restore(pflags);
1175 }
1176
1177 static void rpc_release_resources_task(struct rpc_task *task)
1178 {
1179         xprt_release(task);
1180         if (task->tk_msg.rpc_cred) {
1181                 if (!(task->tk_flags & RPC_TASK_CRED_NOREF))
1182                         put_cred(task->tk_msg.rpc_cred);
1183                 task->tk_msg.rpc_cred = NULL;
1184         }
1185         rpc_task_release_client(task);
1186 }
1187
1188 static void rpc_final_put_task(struct rpc_task *task,
1189                 struct workqueue_struct *q)
1190 {
1191         if (q != NULL) {
1192                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1193                 queue_work(q, &task->u.tk_work);
1194         } else
1195                 rpc_free_task(task);
1196 }
1197
1198 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1199 {
1200         if (atomic_dec_and_test(&task->tk_count)) {
1201                 rpc_release_resources_task(task);
1202                 rpc_final_put_task(task, q);
1203         }
1204 }
1205
1206 void rpc_put_task(struct rpc_task *task)
1207 {
1208         rpc_do_put_task(task, NULL);
1209 }
1210 EXPORT_SYMBOL_GPL(rpc_put_task);
1211
1212 void rpc_put_task_async(struct rpc_task *task)
1213 {
1214         rpc_do_put_task(task, task->tk_workqueue);
1215 }
1216 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1217
1218 static void rpc_release_task(struct rpc_task *task)
1219 {
1220         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1221
1222         rpc_release_resources_task(task);
1223
1224         /*
1225          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1226          * so it should be safe to use task->tk_count as a test for whether
1227          * or not any other processes still hold references to our rpc_task.
1228          */
1229         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1230                 /* Wake up anyone who may be waiting for task completion */
1231                 if (!rpc_complete_task(task))
1232                         return;
1233         } else {
1234                 if (!atomic_dec_and_test(&task->tk_count))
1235                         return;
1236         }
1237         rpc_final_put_task(task, task->tk_workqueue);
1238 }
1239
1240 int rpciod_up(void)
1241 {
1242         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1243 }
1244
1245 void rpciod_down(void)
1246 {
1247         module_put(THIS_MODULE);
1248 }
1249
1250 /*
1251  * Start up the rpciod workqueue.
1252  */
1253 static int rpciod_start(void)
1254 {
1255         struct workqueue_struct *wq;
1256
1257         /*
1258          * Create the rpciod thread and wait for it to start.
1259          */
1260         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1261         if (!wq)
1262                 goto out_failed;
1263         rpciod_workqueue = wq;
1264         wq = alloc_workqueue("xprtiod", WQ_UNBOUND | WQ_MEM_RECLAIM, 0);
1265         if (!wq)
1266                 goto free_rpciod;
1267         xprtiod_workqueue = wq;
1268         return 1;
1269 free_rpciod:
1270         wq = rpciod_workqueue;
1271         rpciod_workqueue = NULL;
1272         destroy_workqueue(wq);
1273 out_failed:
1274         return 0;
1275 }
1276
1277 static void rpciod_stop(void)
1278 {
1279         struct workqueue_struct *wq = NULL;
1280
1281         if (rpciod_workqueue == NULL)
1282                 return;
1283
1284         wq = rpciod_workqueue;
1285         rpciod_workqueue = NULL;
1286         destroy_workqueue(wq);
1287         wq = xprtiod_workqueue;
1288         xprtiod_workqueue = NULL;
1289         destroy_workqueue(wq);
1290 }
1291
1292 void
1293 rpc_destroy_mempool(void)
1294 {
1295         rpciod_stop();
1296         mempool_destroy(rpc_buffer_mempool);
1297         mempool_destroy(rpc_task_mempool);
1298         kmem_cache_destroy(rpc_task_slabp);
1299         kmem_cache_destroy(rpc_buffer_slabp);
1300         rpc_destroy_wait_queue(&delay_queue);
1301 }
1302
1303 int
1304 rpc_init_mempool(void)
1305 {
1306         /*
1307          * The following is not strictly a mempool initialisation,
1308          * but there is no harm in doing it here
1309          */
1310         rpc_init_wait_queue(&delay_queue, "delayq");
1311         if (!rpciod_start())
1312                 goto err_nomem;
1313
1314         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1315                                              sizeof(struct rpc_task),
1316                                              0, SLAB_HWCACHE_ALIGN,
1317                                              NULL);
1318         if (!rpc_task_slabp)
1319                 goto err_nomem;
1320         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1321                                              RPC_BUFFER_MAXSIZE,
1322                                              0, SLAB_HWCACHE_ALIGN,
1323                                              NULL);
1324         if (!rpc_buffer_slabp)
1325                 goto err_nomem;
1326         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1327                                                     rpc_task_slabp);
1328         if (!rpc_task_mempool)
1329                 goto err_nomem;
1330         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1331                                                       rpc_buffer_slabp);
1332         if (!rpc_buffer_mempool)
1333                 goto err_nomem;
1334         return 0;
1335 err_nomem:
1336         rpc_destroy_mempool();
1337         return -ENOMEM;
1338 }