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();
 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  */
 328 static void rpc_make_runnable(struct rpc_task *task)
 329 {
 330         rpc_clear_queued(task);
 331         if (rpc_test_and_set_running(task))
 332                 return;
 333         if (RPC_IS_ASYNC(task)) {
 334                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 335                 queue_work(rpciod_workqueue, &task->u.tk_work);
 336         } else
 337                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 338 }
 339
 340 /*
 341  * Prepare for sleeping on a wait queue.
 342  * By always appending tasks to the list we ensure FIFO behavior.
 343  * NB: An RPC task will only receive interrupt-driven events as long
 344  * as it's on a wait queue.
 345  */
 346 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 347                 struct rpc_task *task,
 348                 rpc_action action,
 349                 unsigned char queue_priority)
 350 {
 351         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 352                         task->tk_pid, rpc_qname(q), jiffies);
 353
 354         trace_rpc_task_sleep(task->tk_client, task, q);
 355
 356         __rpc_add_wait_queue(q, task, queue_priority);
 357
 358         WARN_ON_ONCE(task->tk_callback != NULL);
 359         task->tk_callback = action;
 360         __rpc_add_timer(q, task);
 361 }
 362
 363 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 364                                 rpc_action action)
 365 {
 366         /* We shouldn't ever put an inactive task to sleep */
 367         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 368         if (!RPC_IS_ACTIVATED(task)) {
 369                 task->tk_status = -EIO;
 370                 rpc_put_task_async(task);
 371                 return;
 372         }
 373
 374         /*
 375          * Protect the queue operations.
 376          */
 377         spin_lock_bh(&q->lock);
 378         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 379         spin_unlock_bh(&q->lock);
 380 }
 381 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 382
 383 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 384                 rpc_action action, int priority)
 385 {
 386         /* We shouldn't ever put an inactive task to sleep */
 387         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 388         if (!RPC_IS_ACTIVATED(task)) {
 389                 task->tk_status = -EIO;
 390                 rpc_put_task_async(task);
 391                 return;
 392         }
 393
 394         /*
 395          * Protect the queue operations.
 396          */
 397         spin_lock_bh(&q->lock);
 398         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 399         spin_unlock_bh(&q->lock);
 400 }
 401 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 402
 403 /**
 404  * __rpc_do_wake_up_task - wake up a single rpc_task
 405  * @queue: wait queue
 406  * @task: task to be woken up
 407  *
 408  * Caller must hold queue->lock, and have cleared the task queued flag.
 409  */
 410 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 411 {
 412         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 413                         task->tk_pid, jiffies);
 414
 415         /* Has the task been executed yet? If not, we cannot wake it up! */
 416         if (!RPC_IS_ACTIVATED(task)) {
 417                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 418                 return;
 419         }
 420
 421         trace_rpc_task_wakeup(task->tk_client, task, queue);
 422
 423         __rpc_remove_wait_queue(queue, task);
 424
 425         rpc_make_runnable(task);
 426
 427         dprintk("RPC:       __rpc_wake_up_task done\n");
 428 }
 429
 430 /*
 431  * Wake up a queued task while the queue lock is being held
 432  */
 433 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 434 {
 435         if (RPC_IS_QUEUED(task)) {
 436                 smp_rmb();
 437                 if (task->tk_waitqueue == queue)
 438                         __rpc_do_wake_up_task(queue, task);
 439         }
 440 }
 441
 442 /*
 443  * Tests whether rpc queue is empty
 444  */
 445 int rpc_queue_empty(struct rpc_wait_queue *queue)
 446 {
 447         int res;
 448
 449         spin_lock_bh(&queue->lock);
 450         res = queue->qlen;
 451         spin_unlock_bh(&queue->lock);
 452         return res == 0;
 453 }
 454 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 455
 456 /*
 457  * Wake up a task on a specific queue
 458  */
 459 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 460 {
 461         spin_lock_bh(&queue->lock);
 462         rpc_wake_up_task_queue_locked(queue, task);
 463         spin_unlock_bh(&queue->lock);
 464 }
 465 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 466
 467 /*
 468  * Wake up the next task on a priority queue.
 469  */
 470 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 471 {
 472         struct list_head *q;
 473         struct rpc_task *task;
 474
 475         /*
 476          * Service a batch of tasks from a single owner.
 477          */
 478         q = &queue->tasks[queue->priority];
 479         if (!list_empty(q)) {
 480                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 481                 if (queue->owner == task->tk_owner) {
 482                         if (--queue->nr)
 483                                 goto out;
 484                         list_move_tail(&task->u.tk_wait.list, q);
 485                 }
 486                 /*
 487                  * Check if we need to switch queues.
 488                  */
 489                 goto new_owner;
 490         }
 491
 492         /*
 493          * Service the next queue.
 494          */
 495         do {
 496                 if (q == &queue->tasks[0])
 497                         q = &queue->tasks[queue->maxpriority];
 498                 else
 499                         q = q - 1;
 500                 if (!list_empty(q)) {
 501                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 502                         goto new_queue;
 503                 }
 504         } while (q != &queue->tasks[queue->priority]);
 505
 506         rpc_reset_waitqueue_priority(queue);
 507         return NULL;
 508
 509 new_queue:
 510         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 511 new_owner:
 512         rpc_set_waitqueue_owner(queue, task->tk_owner);
 513 out:
 514         return task;
 515 }
 516
 517 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 518 {
 519         if (RPC_IS_PRIORITY(queue))
 520                 return __rpc_find_next_queued_priority(queue);
 521         if (!list_empty(&queue->tasks[0]))
 522                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 523         return NULL;
 524 }
 525
 526 /*
 527  * Wake up the first task on the wait queue.
 528  */
 529 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 530                 bool (*func)(struct rpc_task *, void *), void *data)
 531 {
 532         struct rpc_task *task = NULL;
 533
 534         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 535                         queue, rpc_qname(queue));
 536         spin_lock_bh(&queue->lock);
 537         task = __rpc_find_next_queued(queue);
 538         if (task != NULL) {
 539                 if (func(task, data))
 540                         rpc_wake_up_task_queue_locked(queue, task);
 541                 else
 542                         task = NULL;
 543         }
 544         spin_unlock_bh(&queue->lock);
 545
 546         return task;
 547 }
 548 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 549
 550 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 551 {
 552         return true;
 553 }
 554
 555 /*
 556  * Wake up the next task on the wait queue.
 557 */
 558 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 559 {
 560         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 561 }
 562 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 563
 564 /**
 565  * rpc_wake_up - wake up all rpc_tasks
 566  * @queue: rpc_wait_queue on which the tasks are sleeping
 567  *
 568  * Grabs queue->lock
 569  */
 570 void rpc_wake_up(struct rpc_wait_queue *queue)
 571 {
 572         struct list_head *head;
 573
 574         spin_lock_bh(&queue->lock);
 575         head = &queue->tasks[queue->maxpriority];
 576         for (;;) {
 577                 while (!list_empty(head)) {
 578                         struct rpc_task *task;
 579                         task = list_first_entry(head,
 580                                         struct rpc_task,
 581                                         u.tk_wait.list);
 582                         rpc_wake_up_task_queue_locked(queue, task);
 583                 }
 584                 if (head == &queue->tasks[0])
 585                         break;
 586                 head--;
 587         }
 588         spin_unlock_bh(&queue->lock);
 589 }
 590 EXPORT_SYMBOL_GPL(rpc_wake_up);
 591
 592 /**
 593  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 594  * @queue: rpc_wait_queue on which the tasks are sleeping
 595  * @status: status value to set
 596  *
 597  * Grabs queue->lock
 598  */
 599 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 600 {
 601         struct list_head *head;
 602
 603         spin_lock_bh(&queue->lock);
 604         head = &queue->tasks[queue->maxpriority];
 605         for (;;) {
 606                 while (!list_empty(head)) {
 607                         struct rpc_task *task;
 608                         task = list_first_entry(head,
 609                                         struct rpc_task,
 610                                         u.tk_wait.list);
 611                         task->tk_status = status;
 612                         rpc_wake_up_task_queue_locked(queue, task);
 613                 }
 614                 if (head == &queue->tasks[0])
 615                         break;
 616                 head--;
 617         }
 618         spin_unlock_bh(&queue->lock);
 619 }
 620 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 621
 622 static void __rpc_queue_timer_fn(unsigned long ptr)
 623 {
 624         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 625         struct rpc_task *task, *n;
 626         unsigned long expires, now, timeo;
 627
 628         spin_lock(&queue->lock);
 629         expires = now = jiffies;
 630         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 631                 timeo = task->u.tk_wait.expires;
 632                 if (time_after_eq(now, timeo)) {
 633                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 634                         task->tk_status = -ETIMEDOUT;
 635                         rpc_wake_up_task_queue_locked(queue, task);
 636                         continue;
 637                 }
 638                 if (expires == now || time_after(expires, timeo))
 639                         expires = timeo;
 640         }
 641         if (!list_empty(&queue->timer_list.list))
 642                 rpc_set_queue_timer(queue, expires);
 643         spin_unlock(&queue->lock);
 644 }
 645
 646 static void __rpc_atrun(struct rpc_task *task)
 647 {
 648         task->tk_status = 0;
 649 }
 650
 651 /*
 652  * Run a task at a later time
 653  */
 654 void rpc_delay(struct rpc_task *task, unsigned long delay)
 655 {
 656         task->tk_timeout = delay;
 657         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 658 }
 659 EXPORT_SYMBOL_GPL(rpc_delay);
 660
 661 /*
 662  * Helper to call task->tk_ops->rpc_call_prepare
 663  */
 664 void rpc_prepare_task(struct rpc_task *task)
 665 {
 666         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 667 }
 668
 669 static void
 670 rpc_init_task_statistics(struct rpc_task *task)
 671 {
 672         /* Initialize retry counters */
 673         task->tk_garb_retry = 2;
 674         task->tk_cred_retry = 2;
 675         task->tk_rebind_retry = 2;
 676
 677         /* starting timestamp */
 678         task->tk_start = ktime_get();
 679 }
 680
 681 static void
 682 rpc_reset_task_statistics(struct rpc_task *task)
 683 {
 684         task->tk_timeouts = 0;
 685         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 686
 687         rpc_init_task_statistics(task);
 688 }
 689
 690 /*
 691  * Helper that calls task->tk_ops->rpc_call_done if it exists
 692  */
 693 void rpc_exit_task(struct rpc_task *task)
 694 {
 695         task->tk_action = NULL;
 696         if (task->tk_ops->rpc_call_done != NULL) {
 697                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 698                 if (task->tk_action != NULL) {
 699                         WARN_ON(RPC_ASSASSINATED(task));
 700                         /* Always release the RPC slot and buffer memory */
 701                         xprt_release(task);
 702                         rpc_reset_task_statistics(task);
 703                 }
 704         }
 705 }
 706
 707 void rpc_exit(struct rpc_task *task, int status)
 708 {
 709         task->tk_status = status;
 710         task->tk_action = rpc_exit_task;
 711         if (RPC_IS_QUEUED(task))
 712                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 713 }
 714 EXPORT_SYMBOL_GPL(rpc_exit);
 715
 716 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 717 {
 718         if (ops->rpc_release != NULL)
 719                 ops->rpc_release(calldata);
 720 }
 721
 722 /*
 723  * This is the RPC `scheduler' (or rather, the finite state machine).
 724  */
 725 static void __rpc_execute(struct rpc_task *task)
 726 {
 727         struct rpc_wait_queue *queue;
 728         int task_is_async = RPC_IS_ASYNC(task);
 729         int status = 0;
 730
 731         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 732                         task->tk_pid, task->tk_flags);
 733
 734         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 735         if (RPC_IS_QUEUED(task))
 736                 return;
 737
 738         for (;;) {
 739                 void (*do_action)(struct rpc_task *);
 740
 741                 /*
 742                  * Execute any pending callback first.
 743                  */
 744                 do_action = task->tk_callback;
 745                 task->tk_callback = NULL;
 746                 if (do_action == NULL) {
 747                         /*
 748                          * Perform the next FSM step.
 749                          * tk_action may be NULL if the task has been killed.
 750                          * In particular, note that rpc_killall_tasks may
 751                          * do this at any time, so beware when dereferencing.
 752                          */
 753                         do_action = task->tk_action;
 754                         if (do_action == NULL)
 755                                 break;
 756                 }
 757                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 758                 do_action(task);
 759
 760                 /*
 761                  * Lockless check for whether task is sleeping or not.
 762                  */
 763                 if (!RPC_IS_QUEUED(task))
 764                         continue;
 765                 /*
 766                  * The queue->lock protects against races with
 767                  * rpc_make_runnable().
 768                  *
 769                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 770                  * rpc_task, rpc_make_runnable() can assign it to a
 771                  * different workqueue. We therefore cannot assume that the
 772                  * rpc_task pointer may still be dereferenced.
 773                  */
 774                 queue = task->tk_waitqueue;
 775                 spin_lock_bh(&queue->lock);
 776                 if (!RPC_IS_QUEUED(task)) {
 777                         spin_unlock_bh(&queue->lock);
 778                         continue;
 779                 }
 780                 rpc_clear_running(task);
 781                 spin_unlock_bh(&queue->lock);
 782                 if (task_is_async)
 783                         return;
 784
 785                 /* sync task: sleep here */
 786                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 787                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 788                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 789                                 TASK_KILLABLE);
 790                 if (status == -ERESTARTSYS) {
 791                         /*
 792                          * When a sync task receives a signal, it exits with
 793                          * -ERESTARTSYS. In order to catch any callbacks that
 794                          * clean up after sleeping on some queue, we don't
 795                          * break the loop here, but go around once more.
 796                          */
 797                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 798                         task->tk_flags |= RPC_TASK_KILLED;
 799                         rpc_exit(task, -ERESTARTSYS);
 800                 }
 801                 rpc_set_running(task);
 802                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 803         }
 804
 805         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 806                         task->tk_status);
 807         /* Release all resources associated with the task */
 808         rpc_release_task(task);
 809 }
 810
 811 /*
 812  * User-visible entry point to the scheduler.
 813  *
 814  * This may be called recursively if e.g. an async NFS task updates
 815  * the attributes and finds that dirty pages must be flushed.
 816  * NOTE: Upon exit of this function the task is guaranteed to be
 817  *       released. In particular note that tk_release() will have
 818  *       been called, so your task memory may have been freed.
 819  */
 820 void rpc_execute(struct rpc_task *task)
 821 {
 822         rpc_set_active(task);
 823         rpc_make_runnable(task);
 824         if (!RPC_IS_ASYNC(task))
 825                 __rpc_execute(task);
 826 }
 827
 828 static void rpc_async_schedule(struct work_struct *work)
 829 {
 830         current->flags |= PF_FSTRANS;
 831         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 832         current->flags &= ~PF_FSTRANS;
 833 }
 834
 835 /**
 836  * rpc_malloc - allocate an RPC buffer
 837  * @task: RPC task that will use this buffer
 838  * @size: requested byte size
 839  *
 840  * To prevent rpciod from hanging, this allocator never sleeps,
 841  * returning NULL if the request cannot be serviced immediately.
 842  * The caller can arrange to sleep in a way that is safe for rpciod.
 843  *
 844  * Most requests are 'small' (under 2KiB) and can be serviced from a
 845  * mempool, ensuring that NFS reads and writes can always proceed,
 846  * and that there is good locality of reference for these buffers.
 847  *
 848  * In order to avoid memory starvation triggering more writebacks of
 849  * NFS requests, we avoid using GFP_KERNEL.
 850  */
 851 void *rpc_malloc(struct rpc_task *task, size_t size)
 852 {
 853         struct rpc_buffer *buf;
 854         gfp_t gfp = GFP_NOWAIT;
 855
 856         if (RPC_IS_SWAPPER(task))
 857                 gfp |= __GFP_MEMALLOC;
 858
 859         size += sizeof(struct rpc_buffer);
 860         if (size <= RPC_BUFFER_MAXSIZE)
 861                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 862         else
 863                 buf = kmalloc(size, gfp);
 864
 865         if (!buf)
 866                 return NULL;
 867
 868         buf->len = size;
 869         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 870                         task->tk_pid, size, buf);
 871         return &buf->data;
 872 }
 873 EXPORT_SYMBOL_GPL(rpc_malloc);
 874
 875 /**
 876  * rpc_free - free buffer allocated via rpc_malloc
 877  * @buffer: buffer to free
 878  *
 879  */
 880 void rpc_free(void *buffer)
 881 {
 882         size_t size;
 883         struct rpc_buffer *buf;
 884
 885         if (!buffer)
 886                 return;
 887
 888         buf = container_of(buffer, struct rpc_buffer, data);
 889         size = buf->len;
 890
 891         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 892                         size, buf);
 893
 894         if (size <= RPC_BUFFER_MAXSIZE)
 895                 mempool_free(buf, rpc_buffer_mempool);
 896         else
 897                 kfree(buf);
 898 }
 899 EXPORT_SYMBOL_GPL(rpc_free);
 900
 901 /*
 902  * Creation and deletion of RPC task structures
 903  */
 904 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 905 {
 906         memset(task, 0, sizeof(*task));
 907         atomic_set(&task->tk_count, 1);
 908         task->tk_flags  = task_setup_data->flags;
 909         task->tk_ops = task_setup_data->callback_ops;
 910         task->tk_calldata = task_setup_data->callback_data;
 911         INIT_LIST_HEAD(&task->tk_task);
 912
 913         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 914         task->tk_owner = current->tgid;
 915
 916         /* Initialize workqueue for async tasks */
 917         task->tk_workqueue = task_setup_data->workqueue;
 918
 919         if (task->tk_ops->rpc_call_prepare != NULL)
 920                 task->tk_action = rpc_prepare_task;
 921
 922         rpc_init_task_statistics(task);
 923
 924         dprintk("RPC:       new task initialized, procpid %u\n",
 925                                 task_pid_nr(current));
 926 }
 927
 928 static struct rpc_task *
 929 rpc_alloc_task(void)
 930 {
 931         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 932 }
 933
 934 /*
 935  * Create a new task for the specified client.
 936  */
 937 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 938 {
 939         struct rpc_task *task = setup_data->task;
 940         unsigned short flags = 0;
 941
 942         if (task == NULL) {
 943                 task = rpc_alloc_task();
 944                 if (task == NULL) {
 945                         rpc_release_calldata(setup_data->callback_ops,
 946                                         setup_data->callback_data);
 947                         return ERR_PTR(-ENOMEM);
 948                 }
 949                 flags = RPC_TASK_DYNAMIC;
 950         }
 951
 952         rpc_init_task(task, setup_data);
 953         task->tk_flags |= flags;
 954         dprintk("RPC:       allocated task %p\n", task);
 955         return task;
 956 }
 957
 958 /*
 959  * rpc_free_task - release rpc task and perform cleanups
 960  *
 961  * Note that we free up the rpc_task _after_ rpc_release_calldata()
 962  * in order to work around a workqueue dependency issue.
 963  *
 964  * Tejun Heo states:
 965  * "Workqueue currently considers two work items to be the same if they're
 966  * on the same address and won't execute them concurrently - ie. it
 967  * makes a work item which is queued again while being executed wait
 968  * for the previous execution to complete.
 969  *
 970  * If a work function frees the work item, and then waits for an event
 971  * which should be performed by another work item and *that* work item
 972  * recycles the freed work item, it can create a false dependency loop.
 973  * There really is no reliable way to detect this short of verifying
 974  * every memory free."
 975  *
 976  */
 977 static void rpc_free_task(struct rpc_task *task)
 978 {
 979         unsigned short tk_flags = task->tk_flags;
 980
 981         rpc_release_calldata(task->tk_ops, task->tk_calldata);
 982
 983         if (tk_flags & RPC_TASK_DYNAMIC) {
 984                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 985                 mempool_free(task, rpc_task_mempool);
 986         }
 987 }
 988
 989 static void rpc_async_release(struct work_struct *work)
 990 {
 991         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 992 }
 993
 994 static void rpc_release_resources_task(struct rpc_task *task)
 995 {
 996         xprt_release(task);
 997         if (task->tk_msg.rpc_cred) {
 998                 put_rpccred(task->tk_msg.rpc_cred);
 999                 task->tk_msg.rpc_cred = NULL;
1000         }
1001         rpc_task_release_client(task);
1002 }
1003
1004 static void rpc_final_put_task(struct rpc_task *task,
1005                 struct workqueue_struct *q)
1006 {
1007         if (q != NULL) {
1008                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1009                 queue_work(q, &task->u.tk_work);
1010         } else
1011                 rpc_free_task(task);
1012 }
1013
1014 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1015 {
1016         if (atomic_dec_and_test(&task->tk_count)) {
1017                 rpc_release_resources_task(task);
1018                 rpc_final_put_task(task, q);
1019         }
1020 }
1021
1022 void rpc_put_task(struct rpc_task *task)
1023 {
1024         rpc_do_put_task(task, NULL);
1025 }
1026 EXPORT_SYMBOL_GPL(rpc_put_task);
1027
1028 void rpc_put_task_async(struct rpc_task *task)
1029 {
1030         rpc_do_put_task(task, task->tk_workqueue);
1031 }
1032 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1033
1034 static void rpc_release_task(struct rpc_task *task)
1035 {
1036         dprintk("RPC: %5u release task\n", task->tk_pid);
1037
1038         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1039
1040         rpc_release_resources_task(task);
1041
1042         /*
1043          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1044          * so it should be safe to use task->tk_count as a test for whether
1045          * or not any other processes still hold references to our rpc_task.
1046          */
1047         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1048                 /* Wake up anyone who may be waiting for task completion */
1049                 if (!rpc_complete_task(task))
1050                         return;
1051         } else {
1052                 if (!atomic_dec_and_test(&task->tk_count))
1053                         return;
1054         }
1055         rpc_final_put_task(task, task->tk_workqueue);
1056 }
1057
1058 int rpciod_up(void)
1059 {
1060         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1061 }
1062
1063 void rpciod_down(void)
1064 {
1065         module_put(THIS_MODULE);
1066 }
1067
1068 /*
1069  * Start up the rpciod workqueue.
1070  */
1071 static int rpciod_start(void)
1072 {
1073         struct workqueue_struct *wq;
1074
1075         /*
1076          * Create the rpciod thread and wait for it to start.
1077          */
1078         dprintk("RPC:       creating workqueue rpciod\n");
1079         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 1);
1080         rpciod_workqueue = wq;
1081         return rpciod_workqueue != NULL;
1082 }
1083
1084 static void rpciod_stop(void)
1085 {
1086         struct workqueue_struct *wq = NULL;
1087
1088         if (rpciod_workqueue == NULL)
1089                 return;
1090         dprintk("RPC:       destroying workqueue rpciod\n");
1091
1092         wq = rpciod_workqueue;
1093         rpciod_workqueue = NULL;
1094         destroy_workqueue(wq);
1095 }
1096
1097 void
1098 rpc_destroy_mempool(void)
1099 {
1100         rpciod_stop();
1101         if (rpc_buffer_mempool)
1102                 mempool_destroy(rpc_buffer_mempool);
1103         if (rpc_task_mempool)
1104                 mempool_destroy(rpc_task_mempool);
1105         if (rpc_task_slabp)
1106                 kmem_cache_destroy(rpc_task_slabp);
1107         if (rpc_buffer_slabp)
1108                 kmem_cache_destroy(rpc_buffer_slabp);
1109         rpc_destroy_wait_queue(&delay_queue);
1110 }
1111
1112 int
1113 rpc_init_mempool(void)
1114 {
1115         /*
1116          * The following is not strictly a mempool initialisation,
1117          * but there is no harm in doing it here
1118          */
1119         rpc_init_wait_queue(&delay_queue, "delayq");
1120         if (!rpciod_start())
1121                 goto err_nomem;
1122
1123         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1124                                              sizeof(struct rpc_task),
1125                                              0, SLAB_HWCACHE_ALIGN,
1126                                              NULL);
1127         if (!rpc_task_slabp)
1128                 goto err_nomem;
1129         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1130                                              RPC_BUFFER_MAXSIZE,
1131                                              0, SLAB_HWCACHE_ALIGN,
1132                                              NULL);
1133         if (!rpc_buffer_slabp)
1134                 goto err_nomem;
1135         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1136                                                     rpc_task_slabp);
1137         if (!rpc_task_mempool)
1138                 goto err_nomem;
1139         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1140                                                       rpc_buffer_slabp);
1141         if (!rpc_buffer_mempool)
1142                 goto err_nomem;
1143         return 0;
1144 err_nomem:
1145         rpc_destroy_mempool();
1146         return -ENOMEM;
1147 }