net/sunrpc/sched.c

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