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 #include <linux/freezer.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #include "sunrpc.h"
  26
  27 #ifdef RPC_DEBUG
  28 #define RPCDBG_FACILITY         RPCDBG_SCHED
  29 #endif
  30
  31 #define CREATE_TRACE_POINTS
  32 #include <trace/events/sunrpc.h>
  33
  34 /*
  35  * RPC slabs and memory pools
  36  */
  37 #define RPC_BUFFER_MAXSIZE      (2048)
  38 #define RPC_BUFFER_POOLSIZE     (8)
  39 #define RPC_TASK_POOLSIZE       (8)
  40 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  41 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  42 static mempool_t        *rpc_task_mempool __read_mostly;
  43 static mempool_t        *rpc_buffer_mempool __read_mostly;
  44
  45 static void                     rpc_async_schedule(struct work_struct *);
  46 static void                      rpc_release_task(struct rpc_task *task);
  47 static void __rpc_queue_timer_fn(unsigned long ptr);
  48
  49 /*
  50  * RPC tasks sit here while waiting for conditions to improve.
  51  */
  52 static struct rpc_wait_queue delay_queue;
  53
  54 /*
  55  * rpciod-related stuff
  56  */
  57 struct workqueue_struct *rpciod_workqueue;
  58
  59 /*
  60  * Disable the timer for a given RPC task. Should be called with
  61  * queue->lock and bh_disabled in order to avoid races within
  62  * rpc_run_timer().
  63  */
  64 static void
  65 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  66 {
  67         if (task->tk_timeout == 0)
  68                 return;
  69         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  70         task->tk_timeout = 0;
  71         list_del(&task->u.tk_wait.timer_list);
  72         if (list_empty(&queue->timer_list.list))
  73                 del_timer(&queue->timer_list.timer);
  74 }
  75
  76 static void
  77 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  78 {
  79         queue->timer_list.expires = expires;
  80         mod_timer(&queue->timer_list.timer, expires);
  81 }
  82
  83 /*
  84  * Set up a timer for the current task.
  85  */
  86 static void
  87 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  88 {
  89         if (!task->tk_timeout)
  90                 return;
  91
  92         dprintk("RPC: %5u setting alarm for %lu ms\n",
  93                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  94
  95         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  96         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  97                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  98         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  99 }
 100
 101 static void rpc_rotate_queue_owner(struct rpc_wait_queue *queue)
 102 {
 103         struct list_head *q = &queue->tasks[queue->priority];
 104         struct rpc_task *task;
 105
 106         if (!list_empty(q)) {
 107                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 108                 if (task->tk_owner == queue->owner)
 109                         list_move_tail(&task->u.tk_wait.list, q);
 110         }
 111 }
 112
 113 static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 114 {
 115         if (queue->priority != priority) {
 116                 /* Fairness: rotate the list when changing priority */
 117                 rpc_rotate_queue_owner(queue);
 118                 queue->priority = priority;
 119         }
 120 }
 121
 122 static void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 123 {
 124         queue->owner = pid;
 125         queue->nr = RPC_BATCH_COUNT;
 126 }
 127
 128 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 129 {
 130         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 131         rpc_set_waitqueue_owner(queue, 0);
 132 }
 133
 134 /*
 135  * Add new request to a priority queue.
 136  */
 137 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 138                 struct rpc_task *task,
 139                 unsigned char queue_priority)
 140 {
 141         struct list_head *q;
 142         struct rpc_task *t;
 143
 144         INIT_LIST_HEAD(&task->u.tk_wait.links);
 145         if (unlikely(queue_priority > queue->maxpriority))
 146                 queue_priority = queue->maxpriority;
 147         if (queue_priority > queue->priority)
 148                 rpc_set_waitqueue_priority(queue, queue_priority);
 149         q = &queue->tasks[queue_priority];
 150         list_for_each_entry(t, q, u.tk_wait.list) {
 151                 if (t->tk_owner == task->tk_owner) {
 152                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 153                         return;
 154                 }
 155         }
 156         list_add_tail(&task->u.tk_wait.list, q);
 157 }
 158
 159 /*
 160  * Add new request to wait queue.
 161  *
 162  * Swapper tasks always get inserted at the head of the queue.
 163  * This should avoid many nasty memory deadlocks and hopefully
 164  * improve overall performance.
 165  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 166  */
 167 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 168                 struct rpc_task *task,
 169                 unsigned char queue_priority)
 170 {
 171         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 172         if (RPC_IS_QUEUED(task))
 173                 return;
 174
 175         if (RPC_IS_PRIORITY(queue))
 176                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 177         else if (RPC_IS_SWAPPER(task))
 178                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 179         else
 180                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 181         task->tk_waitqueue = queue;
 182         queue->qlen++;
 183         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 184         smp_wmb();
 185         rpc_set_queued(task);
 186
 187         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 188                         task->tk_pid, queue, rpc_qname(queue));
 189 }
 190
 191 /*
 192  * Remove request from a priority queue.
 193  */
 194 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 195 {
 196         struct rpc_task *t;
 197
 198         if (!list_empty(&task->u.tk_wait.links)) {
 199                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 200                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 201                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 202         }
 203 }
 204
 205 /*
 206  * Remove request from queue.
 207  * Note: must be called with spin lock held.
 208  */
 209 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 210 {
 211         __rpc_disable_timer(queue, task);
 212         if (RPC_IS_PRIORITY(queue))
 213                 __rpc_remove_wait_queue_priority(task);
 214         list_del(&task->u.tk_wait.list);
 215         queue->qlen--;
 216         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 217                         task->tk_pid, queue, rpc_qname(queue));
 218 }
 219
 220 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 221 {
 222         int i;
 223
 224         spin_lock_init(&queue->lock);
 225         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 226                 INIT_LIST_HEAD(&queue->tasks[i]);
 227         queue->maxpriority = nr_queues - 1;
 228         rpc_reset_waitqueue_priority(queue);
 229         queue->qlen = 0;
 230         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 231         INIT_LIST_HEAD(&queue->timer_list.list);
 232         rpc_assign_waitqueue_name(queue, qname);
 233 }
 234
 235 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 236 {
 237         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 238 }
 239 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 240
 241 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 242 {
 243         __rpc_init_priority_wait_queue(queue, qname, 1);
 244 }
 245 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 246
 247 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 248 {
 249         del_timer_sync(&queue->timer_list.timer);
 250 }
 251 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 252
 253 static int rpc_wait_bit_killable(void *word)
 254 {
 255         if (fatal_signal_pending(current))
 256                 return -ERESTARTSYS;
 257         freezable_schedule_unsafe();
 258         return 0;
 259 }
 260
 261 #ifdef RPC_DEBUG
 262 static void rpc_task_set_debuginfo(struct rpc_task *task)
 263 {
 264         static atomic_t rpc_pid;
 265
 266         task->tk_pid = atomic_inc_return(&rpc_pid);
 267 }
 268 #else
 269 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 270 {
 271 }
 272 #endif
 273
 274 static void rpc_set_active(struct rpc_task *task)
 275 {
 276         trace_rpc_task_begin(task->tk_client, task, NULL);
 277
 278         rpc_task_set_debuginfo(task);
 279         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 280 }
 281
 282 /*
 283  * Mark an RPC call as having completed by clearing the 'active' bit
 284  * and then waking up all tasks that were sleeping.
 285  */
 286 static int rpc_complete_task(struct rpc_task *task)
 287 {
 288         void *m = &task->tk_runstate;
 289         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 290         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 291         unsigned long flags;
 292         int ret;
 293
 294         trace_rpc_task_complete(task->tk_client, task, NULL);
 295
 296         spin_lock_irqsave(&wq->lock, flags);
 297         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 298         ret = atomic_dec_and_test(&task->tk_count);
 299         if (waitqueue_active(wq))
 300                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 301         spin_unlock_irqrestore(&wq->lock, flags);
 302         return ret;
 303 }
 304
 305 /*
 306  * Allow callers to wait for completion of an RPC call
 307  *
 308  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 309  * to enforce taking of the wq->lock and hence avoid races with
 310  * rpc_complete_task().
 311  */
 312 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 313 {
 314         if (action == NULL)
 315                 action = rpc_wait_bit_killable;
 316         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 317                         action, TASK_KILLABLE);
 318 }
 319 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 320
 321 /*
 322  * Make an RPC task runnable.
 323  *
 324  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 325  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 326  * the wait queue operation.
 327  * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
 328  * which is needed to ensure that __rpc_execute() doesn't loop (due to the
 329  * lockless RPC_IS_QUEUED() test) before we've had a chance to test
 330  * the RPC_TASK_RUNNING flag.
 331  */
 332 static void rpc_make_runnable(struct rpc_task *task)
 333 {
 334         bool need_wakeup = !rpc_test_and_set_running(task);
 335
 336         rpc_clear_queued(task);
 337         if (!need_wakeup)
 338                 return;
 339         if (RPC_IS_ASYNC(task)) {
 340                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 341                 queue_work(rpciod_workqueue, &task->u.tk_work);
 342         } else
 343                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 344 }
 345
 346 /*
 347  * Prepare for sleeping on a wait queue.
 348  * By always appending tasks to the list we ensure FIFO behavior.
 349  * NB: An RPC task will only receive interrupt-driven events as long
 350  * as it's on a wait queue.
 351  */
 352 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 353                 struct rpc_task *task,
 354                 rpc_action action,
 355                 unsigned char queue_priority)
 356 {
 357         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 358                         task->tk_pid, rpc_qname(q), jiffies);
 359
 360         trace_rpc_task_sleep(task->tk_client, task, q);
 361
 362         __rpc_add_wait_queue(q, task, queue_priority);
 363
 364         WARN_ON_ONCE(task->tk_callback != NULL);
 365         task->tk_callback = action;
 366         __rpc_add_timer(q, task);
 367 }
 368
 369 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 370                                 rpc_action action)
 371 {
 372         /* We shouldn't ever put an inactive task to sleep */
 373         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 374         if (!RPC_IS_ACTIVATED(task)) {
 375                 task->tk_status = -EIO;
 376                 rpc_put_task_async(task);
 377                 return;
 378         }
 379
 380         /*
 381          * Protect the queue operations.
 382          */
 383         spin_lock_bh(&q->lock);
 384         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 385         spin_unlock_bh(&q->lock);
 386 }
 387 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 388
 389 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 390                 rpc_action action, int priority)
 391 {
 392         /* We shouldn't ever put an inactive task to sleep */
 393         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 394         if (!RPC_IS_ACTIVATED(task)) {
 395                 task->tk_status = -EIO;
 396                 rpc_put_task_async(task);
 397                 return;
 398         }
 399
 400         /*
 401          * Protect the queue operations.
 402          */
 403         spin_lock_bh(&q->lock);
 404         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 405         spin_unlock_bh(&q->lock);
 406 }
 407 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 408
 409 /**
 410  * __rpc_do_wake_up_task - wake up a single rpc_task
 411  * @queue: wait queue
 412  * @task: task to be woken up
 413  *
 414  * Caller must hold queue->lock, and have cleared the task queued flag.
 415  */
 416 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 417 {
 418         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 419                         task->tk_pid, jiffies);
 420
 421         /* Has the task been executed yet? If not, we cannot wake it up! */
 422         if (!RPC_IS_ACTIVATED(task)) {
 423                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 424                 return;
 425         }
 426
 427         trace_rpc_task_wakeup(task->tk_client, task, queue);
 428
 429         __rpc_remove_wait_queue(queue, task);
 430
 431         rpc_make_runnable(task);
 432
 433         dprintk("RPC:       __rpc_wake_up_task done\n");
 434 }
 435
 436 /*
 437  * Wake up a queued task while the queue lock is being held
 438  */
 439 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 440 {
 441         if (RPC_IS_QUEUED(task)) {
 442                 smp_rmb();
 443                 if (task->tk_waitqueue == queue)
 444                         __rpc_do_wake_up_task(queue, task);
 445         }
 446 }
 447
 448 /*
 449  * Tests whether rpc queue is empty
 450  */
 451 int rpc_queue_empty(struct rpc_wait_queue *queue)
 452 {
 453         int res;
 454
 455         spin_lock_bh(&queue->lock);
 456         res = queue->qlen;
 457         spin_unlock_bh(&queue->lock);
 458         return res == 0;
 459 }
 460 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 461
 462 /*
 463  * Wake up a task on a specific queue
 464  */
 465 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 466 {
 467         spin_lock_bh(&queue->lock);
 468         rpc_wake_up_task_queue_locked(queue, task);
 469         spin_unlock_bh(&queue->lock);
 470 }
 471 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 472
 473 /*
 474  * Wake up the next task on a priority queue.
 475  */
 476 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 477 {
 478         struct list_head *q;
 479         struct rpc_task *task;
 480
 481         /*
 482          * Service a batch of tasks from a single owner.
 483          */
 484         q = &queue->tasks[queue->priority];
 485         if (!list_empty(q)) {
 486                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 487                 if (queue->owner == task->tk_owner) {
 488                         if (--queue->nr)
 489                                 goto out;
 490                         list_move_tail(&task->u.tk_wait.list, q);
 491                 }
 492                 /*
 493                  * Check if we need to switch queues.
 494                  */
 495                 goto new_owner;
 496         }
 497
 498         /*
 499          * Service the next queue.
 500          */
 501         do {
 502                 if (q == &queue->tasks[0])
 503                         q = &queue->tasks[queue->maxpriority];
 504                 else
 505                         q = q - 1;
 506                 if (!list_empty(q)) {
 507                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 508                         goto new_queue;
 509                 }
 510         } while (q != &queue->tasks[queue->priority]);
 511
 512         rpc_reset_waitqueue_priority(queue);
 513         return NULL;
 514
 515 new_queue:
 516         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 517 new_owner:
 518         rpc_set_waitqueue_owner(queue, task->tk_owner);
 519 out:
 520         return task;
 521 }
 522
 523 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 524 {
 525         if (RPC_IS_PRIORITY(queue))
 526                 return __rpc_find_next_queued_priority(queue);
 527         if (!list_empty(&queue->tasks[0]))
 528                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 529         return NULL;
 530 }
 531
 532 /*
 533  * Wake up the first task on the wait queue.
 534  */
 535 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 536                 bool (*func)(struct rpc_task *, void *), void *data)
 537 {
 538         struct rpc_task *task = NULL;
 539
 540         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 541                         queue, rpc_qname(queue));
 542         spin_lock_bh(&queue->lock);
 543         task = __rpc_find_next_queued(queue);
 544         if (task != NULL) {
 545                 if (func(task, data))
 546                         rpc_wake_up_task_queue_locked(queue, task);
 547                 else
 548                         task = NULL;
 549         }
 550         spin_unlock_bh(&queue->lock);
 551
 552         return task;
 553 }
 554 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 555
 556 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 557 {
 558         return true;
 559 }
 560
 561 /*
 562  * Wake up the next task on the wait queue.
 563 */
 564 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 565 {
 566         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 567 }
 568 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 569
 570 /**
 571  * rpc_wake_up - wake up all rpc_tasks
 572  * @queue: rpc_wait_queue on which the tasks are sleeping
 573  *
 574  * Grabs queue->lock
 575  */
 576 void rpc_wake_up(struct rpc_wait_queue *queue)
 577 {
 578         struct list_head *head;
 579
 580         spin_lock_bh(&queue->lock);
 581         head = &queue->tasks[queue->maxpriority];
 582         for (;;) {
 583                 while (!list_empty(head)) {
 584                         struct rpc_task *task;
 585                         task = list_first_entry(head,
 586                                         struct rpc_task,
 587                                         u.tk_wait.list);
 588                         rpc_wake_up_task_queue_locked(queue, task);
 589                 }
 590                 if (head == &queue->tasks[0])
 591                         break;
 592                 head--;
 593         }
 594         spin_unlock_bh(&queue->lock);
 595 }
 596 EXPORT_SYMBOL_GPL(rpc_wake_up);
 597
 598 /**
 599  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 600  * @queue: rpc_wait_queue on which the tasks are sleeping
 601  * @status: status value to set
 602  *
 603  * Grabs queue->lock
 604  */
 605 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 606 {
 607         struct list_head *head;
 608
 609         spin_lock_bh(&queue->lock);
 610         head = &queue->tasks[queue->maxpriority];
 611         for (;;) {
 612                 while (!list_empty(head)) {
 613                         struct rpc_task *task;
 614                         task = list_first_entry(head,
 615                                         struct rpc_task,
 616                                         u.tk_wait.list);
 617                         task->tk_status = status;
 618                         rpc_wake_up_task_queue_locked(queue, task);
 619                 }
 620                 if (head == &queue->tasks[0])
 621                         break;
 622                 head--;
 623         }
 624         spin_unlock_bh(&queue->lock);
 625 }
 626 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 627
 628 static void __rpc_queue_timer_fn(unsigned long ptr)
 629 {
 630         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 631         struct rpc_task *task, *n;
 632         unsigned long expires, now, timeo;
 633
 634         spin_lock(&queue->lock);
 635         expires = now = jiffies;
 636         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 637                 timeo = task->u.tk_wait.expires;
 638                 if (time_after_eq(now, timeo)) {
 639                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 640                         task->tk_status = -ETIMEDOUT;
 641                         rpc_wake_up_task_queue_locked(queue, task);
 642                         continue;
 643                 }
 644                 if (expires == now || time_after(expires, timeo))
 645                         expires = timeo;
 646         }
 647         if (!list_empty(&queue->timer_list.list))
 648                 rpc_set_queue_timer(queue, expires);
 649         spin_unlock(&queue->lock);
 650 }
 651
 652 static void __rpc_atrun(struct rpc_task *task)
 653 {
 654         task->tk_status = 0;
 655 }
 656
 657 /*
 658  * Run a task at a later time
 659  */
 660 void rpc_delay(struct rpc_task *task, unsigned long delay)
 661 {
 662         task->tk_timeout = delay;
 663         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 664 }
 665 EXPORT_SYMBOL_GPL(rpc_delay);
 666
 667 /*
 668  * Helper to call task->tk_ops->rpc_call_prepare
 669  */
 670 void rpc_prepare_task(struct rpc_task *task)
 671 {
 672         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 673 }
 674
 675 static void
 676 rpc_init_task_statistics(struct rpc_task *task)
 677 {
 678         /* Initialize retry counters */
 679         task->tk_garb_retry = 2;
 680         task->tk_cred_retry = 2;
 681         task->tk_rebind_retry = 2;
 682
 683         /* starting timestamp */
 684         task->tk_start = ktime_get();
 685 }
 686
 687 static void
 688 rpc_reset_task_statistics(struct rpc_task *task)
 689 {
 690         task->tk_timeouts = 0;
 691         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 692
 693         rpc_init_task_statistics(task);
 694 }
 695
 696 /*
 697  * Helper that calls task->tk_ops->rpc_call_done if it exists
 698  */
 699 void rpc_exit_task(struct rpc_task *task)
 700 {
 701         task->tk_action = NULL;
 702         if (task->tk_ops->rpc_call_done != NULL) {
 703                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 704                 if (task->tk_action != NULL) {
 705                         WARN_ON(RPC_ASSASSINATED(task));
 706                         /* Always release the RPC slot and buffer memory */
 707                         xprt_release(task);
 708                         rpc_reset_task_statistics(task);
 709                 }
 710         }
 711 }
 712
 713 void rpc_exit(struct rpc_task *task, int status)
 714 {
 715         task->tk_status = status;
 716         task->tk_action = rpc_exit_task;
 717         if (RPC_IS_QUEUED(task))
 718                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 719 }
 720 EXPORT_SYMBOL_GPL(rpc_exit);
 721
 722 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 723 {
 724         if (ops->rpc_release != NULL)
 725                 ops->rpc_release(calldata);
 726 }
 727
 728 /*
 729  * This is the RPC `scheduler' (or rather, the finite state machine).
 730  */
 731 static void __rpc_execute(struct rpc_task *task)
 732 {
 733         struct rpc_wait_queue *queue;
 734         int task_is_async = RPC_IS_ASYNC(task);
 735         int status = 0;
 736
 737         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 738                         task->tk_pid, task->tk_flags);
 739
 740         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 741         if (RPC_IS_QUEUED(task))
 742                 return;
 743
 744         for (;;) {
 745                 void (*do_action)(struct rpc_task *);
 746
 747                 /*
 748                  * Execute any pending callback first.
 749                  */
 750                 do_action = task->tk_callback;
 751                 task->tk_callback = NULL;
 752                 if (do_action == NULL) {
 753                         /*
 754                          * Perform the next FSM step.
 755                          * tk_action may be NULL if the task has been killed.
 756                          * In particular, note that rpc_killall_tasks may
 757                          * do this at any time, so beware when dereferencing.
 758                          */
 759                         do_action = task->tk_action;
 760                         if (do_action == NULL)
 761                                 break;
 762                 }
 763                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 764                 do_action(task);
 765
 766                 /*
 767                  * Lockless check for whether task is sleeping or not.
 768                  */
 769                 if (!RPC_IS_QUEUED(task))
 770                         continue;
 771                 /*
 772                  * The queue->lock protects against races with
 773                  * rpc_make_runnable().
 774                  *
 775                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 776                  * rpc_task, rpc_make_runnable() can assign it to a
 777                  * different workqueue. We therefore cannot assume that the
 778                  * rpc_task pointer may still be dereferenced.
 779                  */
 780                 queue = task->tk_waitqueue;
 781                 spin_lock_bh(&queue->lock);
 782                 if (!RPC_IS_QUEUED(task)) {
 783                         spin_unlock_bh(&queue->lock);
 784                         continue;
 785                 }
 786                 rpc_clear_running(task);
 787                 spin_unlock_bh(&queue->lock);
 788                 if (task_is_async)
 789                         return;
 790
 791                 /* sync task: sleep here */
 792                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 793                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 794                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 795                                 TASK_KILLABLE);
 796                 if (status == -ERESTARTSYS) {
 797                         /*
 798                          * When a sync task receives a signal, it exits with
 799                          * -ERESTARTSYS. In order to catch any callbacks that
 800                          * clean up after sleeping on some queue, we don't
 801                          * break the loop here, but go around once more.
 802                          */
 803                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 804                         task->tk_flags |= RPC_TASK_KILLED;
 805                         rpc_exit(task, -ERESTARTSYS);
 806                 }
 807                 rpc_set_running(task);
 808                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 809         }
 810
 811         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 812                         task->tk_status);
 813         /* Release all resources associated with the task */
 814         rpc_release_task(task);
 815 }
 816
 817 /*
 818  * User-visible entry point to the scheduler.
 819  *
 820  * This may be called recursively if e.g. an async NFS task updates
 821  * the attributes and finds that dirty pages must be flushed.
 822  * NOTE: Upon exit of this function the task is guaranteed to be
 823  *       released. In particular note that tk_release() will have
 824  *       been called, so your task memory may have been freed.
 825  */
 826 void rpc_execute(struct rpc_task *task)
 827 {
 828         rpc_set_active(task);
 829         rpc_make_runnable(task);
 830         if (!RPC_IS_ASYNC(task))
 831                 __rpc_execute(task);
 832 }
 833
 834 static void rpc_async_schedule(struct work_struct *work)
 835 {
 836         current->flags |= PF_FSTRANS;
 837         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 838         current->flags &= ~PF_FSTRANS;
 839 }
 840
 841 /**
 842  * rpc_malloc - allocate an RPC buffer
 843  * @task: RPC task that will use this buffer
 844  * @size: requested byte size
 845  *
 846  * To prevent rpciod from hanging, this allocator never sleeps,
 847  * returning NULL if the request cannot be serviced immediately.
 848  * The caller can arrange to sleep in a way that is safe for rpciod.
 849  *
 850  * Most requests are 'small' (under 2KiB) and can be serviced from a
 851  * mempool, ensuring that NFS reads and writes can always proceed,
 852  * and that there is good locality of reference for these buffers.
 853  *
 854  * In order to avoid memory starvation triggering more writebacks of
 855  * NFS requests, we avoid using GFP_KERNEL.
 856  */
 857 void *rpc_malloc(struct rpc_task *task, size_t size)
 858 {
 859         struct rpc_buffer *buf;
 860         gfp_t gfp = GFP_NOWAIT;
 861
 862         if (RPC_IS_SWAPPER(task))
 863                 gfp |= __GFP_MEMALLOC;
 864
 865         size += sizeof(struct rpc_buffer);
 866         if (size <= RPC_BUFFER_MAXSIZE)
 867                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 868         else
 869                 buf = kmalloc(size, gfp);
 870
 871         if (!buf)
 872                 return NULL;
 873
 874         buf->len = size;
 875         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 876                         task->tk_pid, size, buf);
 877         return &buf->data;
 878 }
 879 EXPORT_SYMBOL_GPL(rpc_malloc);
 880
 881 /**
 882  * rpc_free - free buffer allocated via rpc_malloc
 883  * @buffer: buffer to free
 884  *
 885  */
 886 void rpc_free(void *buffer)
 887 {
 888         size_t size;
 889         struct rpc_buffer *buf;
 890
 891         if (!buffer)
 892                 return;
 893
 894         buf = container_of(buffer, struct rpc_buffer, data);
 895         size = buf->len;
 896
 897         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 898                         size, buf);
 899
 900         if (size <= RPC_BUFFER_MAXSIZE)
 901                 mempool_free(buf, rpc_buffer_mempool);
 902         else
 903                 kfree(buf);
 904 }
 905 EXPORT_SYMBOL_GPL(rpc_free);
 906
 907 /*
 908  * Creation and deletion of RPC task structures
 909  */
 910 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 911 {
 912         memset(task, 0, sizeof(*task));
 913         atomic_set(&task->tk_count, 1);
 914         task->tk_flags  = task_setup_data->flags;
 915         task->tk_ops = task_setup_data->callback_ops;
 916         task->tk_calldata = task_setup_data->callback_data;
 917         INIT_LIST_HEAD(&task->tk_task);
 918
 919         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 920         task->tk_owner = current->tgid;
 921
 922         /* Initialize workqueue for async tasks */
 923         task->tk_workqueue = task_setup_data->workqueue;
 924
 925         if (task->tk_ops->rpc_call_prepare != NULL)
 926                 task->tk_action = rpc_prepare_task;
 927
 928         rpc_init_task_statistics(task);
 929
 930         dprintk("RPC:       new task initialized, procpid %u\n",
 931                                 task_pid_nr(current));
 932 }
 933
 934 static struct rpc_task *
 935 rpc_alloc_task(void)
 936 {
 937         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 938 }
 939
 940 /*
 941  * Create a new task for the specified client.
 942  */
 943 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 944 {
 945         struct rpc_task *task = setup_data->task;
 946         unsigned short flags = 0;
 947
 948         if (task == NULL) {
 949                 task = rpc_alloc_task();
 950                 if (task == NULL) {
 951                         rpc_release_calldata(setup_data->callback_ops,
 952                                         setup_data->callback_data);
 953                         return ERR_PTR(-ENOMEM);
 954                 }
 955                 flags = RPC_TASK_DYNAMIC;
 956         }
 957
 958         rpc_init_task(task, setup_data);
 959         task->tk_flags |= flags;
 960         dprintk("RPC:       allocated task %p\n", task);
 961         return task;
 962 }
 963
 964 /*
 965  * rpc_free_task - release rpc task and perform cleanups
 966  *
 967  * Note that we free up the rpc_task _after_ rpc_release_calldata()
 968  * in order to work around a workqueue dependency issue.
 969  *
 970  * Tejun Heo states:
 971  * "Workqueue currently considers two work items to be the same if they're
 972  * on the same address and won't execute them concurrently - ie. it
 973  * makes a work item which is queued again while being executed wait
 974  * for the previous execution to complete.
 975  *
 976  * If a work function frees the work item, and then waits for an event
 977  * which should be performed by another work item and *that* work item
 978  * recycles the freed work item, it can create a false dependency loop.
 979  * There really is no reliable way to detect this short of verifying
 980  * every memory free."
 981  *
 982  */
 983 static void rpc_free_task(struct rpc_task *task)
 984 {
 985         unsigned short tk_flags = task->tk_flags;
 986
 987         rpc_release_calldata(task->tk_ops, task->tk_calldata);
 988
 989         if (tk_flags & RPC_TASK_DYNAMIC) {
 990                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 991                 mempool_free(task, rpc_task_mempool);
 992         }
 993 }
 994
 995 static void rpc_async_release(struct work_struct *work)
 996 {
 997         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 998 }
 999
1000 static void rpc_release_resources_task(struct rpc_task *task)
1001 {
1002         xprt_release(task);
1003         if (task->tk_msg.rpc_cred) {
1004                 put_rpccred(task->tk_msg.rpc_cred);
1005                 task->tk_msg.rpc_cred = NULL;
1006         }
1007         rpc_task_release_client(task);
1008 }
1009
1010 static void rpc_final_put_task(struct rpc_task *task,
1011                 struct workqueue_struct *q)
1012 {
1013         if (q != NULL) {
1014                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1015                 queue_work(q, &task->u.tk_work);
1016         } else
1017                 rpc_free_task(task);
1018 }
1019
1020 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1021 {
1022         if (atomic_dec_and_test(&task->tk_count)) {
1023                 rpc_release_resources_task(task);
1024                 rpc_final_put_task(task, q);
1025         }
1026 }
1027
1028 void rpc_put_task(struct rpc_task *task)
1029 {
1030         rpc_do_put_task(task, NULL);
1031 }
1032 EXPORT_SYMBOL_GPL(rpc_put_task);
1033
1034 void rpc_put_task_async(struct rpc_task *task)
1035 {
1036         rpc_do_put_task(task, task->tk_workqueue);
1037 }
1038 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1039
1040 static void rpc_release_task(struct rpc_task *task)
1041 {
1042         dprintk("RPC: %5u release task\n", task->tk_pid);
1043
1044         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1045
1046         rpc_release_resources_task(task);
1047
1048         /*
1049          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1050          * so it should be safe to use task->tk_count as a test for whether
1051          * or not any other processes still hold references to our rpc_task.
1052          */
1053         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1054                 /* Wake up anyone who may be waiting for task completion */
1055                 if (!rpc_complete_task(task))
1056                         return;
1057         } else {
1058                 if (!atomic_dec_and_test(&task->tk_count))
1059                         return;
1060         }
1061         rpc_final_put_task(task, task->tk_workqueue);
1062 }
1063
1064 int rpciod_up(void)
1065 {
1066         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1067 }
1068
1069 void rpciod_down(void)
1070 {
1071         module_put(THIS_MODULE);
1072 }
1073
1074 /*
1075  * Start up the rpciod workqueue.
1076  */
1077 static int rpciod_start(void)
1078 {
1079         struct workqueue_struct *wq;
1080
1081         /*
1082          * Create the rpciod thread and wait for it to start.
1083          */
1084         dprintk("RPC:       creating workqueue rpciod\n");
1085         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 1);
1086         rpciod_workqueue = wq;
1087         return rpciod_workqueue != NULL;
1088 }
1089
1090 static void rpciod_stop(void)
1091 {
1092         struct workqueue_struct *wq = NULL;
1093
1094         if (rpciod_workqueue == NULL)
1095                 return;
1096         dprintk("RPC:       destroying workqueue rpciod\n");
1097
1098         wq = rpciod_workqueue;
1099         rpciod_workqueue = NULL;
1100         destroy_workqueue(wq);
1101 }
1102
1103 void
1104 rpc_destroy_mempool(void)
1105 {
1106         rpciod_stop();
1107         if (rpc_buffer_mempool)
1108                 mempool_destroy(rpc_buffer_mempool);
1109         if (rpc_task_mempool)
1110                 mempool_destroy(rpc_task_mempool);
1111         if (rpc_task_slabp)
1112                 kmem_cache_destroy(rpc_task_slabp);
1113         if (rpc_buffer_slabp)
1114                 kmem_cache_destroy(rpc_buffer_slabp);
1115         rpc_destroy_wait_queue(&delay_queue);
1116 }
1117
1118 int
1119 rpc_init_mempool(void)
1120 {
1121         /*
1122          * The following is not strictly a mempool initialisation,
1123          * but there is no harm in doing it here
1124          */
1125         rpc_init_wait_queue(&delay_queue, "delayq");
1126         if (!rpciod_start())
1127                 goto err_nomem;
1128
1129         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1130                                              sizeof(struct rpc_task),
1131                                              0, SLAB_HWCACHE_ALIGN,
1132                                              NULL);
1133         if (!rpc_task_slabp)
1134                 goto err_nomem;
1135         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1136                                              RPC_BUFFER_MAXSIZE,
1137                                              0, SLAB_HWCACHE_ALIGN,
1138                                              NULL);
1139         if (!rpc_buffer_slabp)
1140                 goto err_nomem;
1141         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1142                                                     rpc_task_slabp);
1143         if (!rpc_task_mempool)
1144                 goto err_nomem;
1145         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1146                                                       rpc_buffer_slabp);
1147         if (!rpc_buffer_mempool)
1148                 goto err_nomem;
1149         return 0;
1150 err_nomem:
1151         rpc_destroy_mempool();
1152         return -ENOMEM;
1153 }