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