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