net/sunrpc/sched.c

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