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