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 list_head *head;
 504
 505         spin_lock_bh(&queue->lock);
 506         head = &queue->tasks[queue->maxpriority];
 507         for (;;) {
 508                 while (!list_empty(head)) {
 509                         struct rpc_task *task;
 510                         task = list_first_entry(head,
 511                                         struct rpc_task,
 512                                         u.tk_wait.list);
 513                         rpc_wake_up_task_queue_locked(queue, task);
 514                 }
 515                 if (head == &queue->tasks[0])
 516                         break;
 517                 head--;
 518         }
 519         spin_unlock_bh(&queue->lock);
 520 }
 521 EXPORT_SYMBOL_GPL(rpc_wake_up);
 522
 523 /**
 524  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 525  * @queue: rpc_wait_queue on which the tasks are sleeping
 526  * @status: status value to set
 527  *
 528  * Grabs queue->lock
 529  */
 530 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 531 {
 532         struct list_head *head;
 533
 534         spin_lock_bh(&queue->lock);
 535         head = &queue->tasks[queue->maxpriority];
 536         for (;;) {
 537                 while (!list_empty(head)) {
 538                         struct rpc_task *task;
 539                         task = list_first_entry(head,
 540                                         struct rpc_task,
 541                                         u.tk_wait.list);
 542                         task->tk_status = status;
 543                         rpc_wake_up_task_queue_locked(queue, task);
 544                 }
 545                 if (head == &queue->tasks[0])
 546                         break;
 547                 head--;
 548         }
 549         spin_unlock_bh(&queue->lock);
 550 }
 551 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 552
 553 static void __rpc_queue_timer_fn(unsigned long ptr)
 554 {
 555         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 556         struct rpc_task *task, *n;
 557         unsigned long expires, now, timeo;
 558
 559         spin_lock(&queue->lock);
 560         expires = now = jiffies;
 561         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 562                 timeo = task->u.tk_wait.expires;
 563                 if (time_after_eq(now, timeo)) {
 564                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 565                         task->tk_status = -ETIMEDOUT;
 566                         rpc_wake_up_task_queue_locked(queue, task);
 567                         continue;
 568                 }
 569                 if (expires == now || time_after(expires, timeo))
 570                         expires = timeo;
 571         }
 572         if (!list_empty(&queue->timer_list.list))
 573                 rpc_set_queue_timer(queue, expires);
 574         spin_unlock(&queue->lock);
 575 }
 576
 577 static void __rpc_atrun(struct rpc_task *task)
 578 {
 579         task->tk_status = 0;
 580 }
 581
 582 /*
 583  * Run a task at a later time
 584  */
 585 void rpc_delay(struct rpc_task *task, unsigned long delay)
 586 {
 587         task->tk_timeout = delay;
 588         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 589 }
 590 EXPORT_SYMBOL_GPL(rpc_delay);
 591
 592 /*
 593  * Helper to call task->tk_ops->rpc_call_prepare
 594  */
 595 void rpc_prepare_task(struct rpc_task *task)
 596 {
 597         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 598 }
 599
 600 static void
 601 rpc_init_task_statistics(struct rpc_task *task)
 602 {
 603         /* Initialize retry counters */
 604         task->tk_garb_retry = 2;
 605         task->tk_cred_retry = 2;
 606         task->tk_rebind_retry = 2;
 607
 608         /* starting timestamp */
 609         task->tk_start = ktime_get();
 610 }
 611
 612 static void
 613 rpc_reset_task_statistics(struct rpc_task *task)
 614 {
 615         task->tk_timeouts = 0;
 616         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 617
 618         rpc_init_task_statistics(task);
 619 }
 620
 621 /*
 622  * Helper that calls task->tk_ops->rpc_call_done if it exists
 623  */
 624 void rpc_exit_task(struct rpc_task *task)
 625 {
 626         task->tk_action = NULL;
 627         if (task->tk_ops->rpc_call_done != NULL) {
 628                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 629                 if (task->tk_action != NULL) {
 630                         WARN_ON(RPC_ASSASSINATED(task));
 631                         /* Always release the RPC slot and buffer memory */
 632                         xprt_release(task);
 633                         rpc_reset_task_statistics(task);
 634                 }
 635         }
 636 }
 637
 638 void rpc_exit(struct rpc_task *task, int status)
 639 {
 640         task->tk_status = status;
 641         task->tk_action = rpc_exit_task;
 642         if (RPC_IS_QUEUED(task))
 643                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 644 }
 645 EXPORT_SYMBOL_GPL(rpc_exit);
 646
 647 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 648 {
 649         if (ops->rpc_release != NULL)
 650                 ops->rpc_release(calldata);
 651 }
 652
 653 /*
 654  * This is the RPC `scheduler' (or rather, the finite state machine).
 655  */
 656 static void __rpc_execute(struct rpc_task *task)
 657 {
 658         struct rpc_wait_queue *queue;
 659         int task_is_async = RPC_IS_ASYNC(task);
 660         int status = 0;
 661
 662         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 663                         task->tk_pid, task->tk_flags);
 664
 665         BUG_ON(RPC_IS_QUEUED(task));
 666
 667         for (;;) {
 668                 void (*do_action)(struct rpc_task *);
 669
 670                 /*
 671                  * Execute any pending callback first.
 672                  */
 673                 do_action = task->tk_callback;
 674                 task->tk_callback = NULL;
 675                 if (do_action == NULL) {
 676                         /*
 677                          * Perform the next FSM step.
 678                          * tk_action may be NULL if the task has been killed.
 679                          * In particular, note that rpc_killall_tasks may
 680                          * do this at any time, so beware when dereferencing.
 681                          */
 682                         do_action = task->tk_action;
 683                         if (do_action == NULL)
 684                                 break;
 685                 }
 686                 do_action(task);
 687
 688                 /*
 689                  * Lockless check for whether task is sleeping or not.
 690                  */
 691                 if (!RPC_IS_QUEUED(task))
 692                         continue;
 693                 /*
 694                  * The queue->lock protects against races with
 695                  * rpc_make_runnable().
 696                  *
 697                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 698                  * rpc_task, rpc_make_runnable() can assign it to a
 699                  * different workqueue. We therefore cannot assume that the
 700                  * rpc_task pointer may still be dereferenced.
 701                  */
 702                 queue = task->tk_waitqueue;
 703                 spin_lock_bh(&queue->lock);
 704                 if (!RPC_IS_QUEUED(task)) {
 705                         spin_unlock_bh(&queue->lock);
 706                         continue;
 707                 }
 708                 rpc_clear_running(task);
 709                 spin_unlock_bh(&queue->lock);
 710                 if (task_is_async)
 711                         return;
 712
 713                 /* sync task: sleep here */
 714                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 715                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 716                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 717                                 TASK_KILLABLE);
 718                 if (status == -ERESTARTSYS) {
 719                         /*
 720                          * When a sync task receives a signal, it exits with
 721                          * -ERESTARTSYS. In order to catch any callbacks that
 722                          * clean up after sleeping on some queue, we don't
 723                          * break the loop here, but go around once more.
 724                          */
 725                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 726                         task->tk_flags |= RPC_TASK_KILLED;
 727                         rpc_exit(task, -ERESTARTSYS);
 728                 }
 729                 rpc_set_running(task);
 730                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 731         }
 732
 733         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 734                         task->tk_status);
 735         /* Release all resources associated with the task */
 736         rpc_release_task(task);
 737 }
 738
 739 /*
 740  * User-visible entry point to the scheduler.
 741  *
 742  * This may be called recursively if e.g. an async NFS task updates
 743  * the attributes and finds that dirty pages must be flushed.
 744  * NOTE: Upon exit of this function the task is guaranteed to be
 745  *       released. In particular note that tk_release() will have
 746  *       been called, so your task memory may have been freed.
 747  */
 748 void rpc_execute(struct rpc_task *task)
 749 {
 750         rpc_set_active(task);
 751         rpc_make_runnable(task);
 752         if (!RPC_IS_ASYNC(task))
 753                 __rpc_execute(task);
 754 }
 755
 756 static void rpc_async_schedule(struct work_struct *work)
 757 {
 758         current->flags |= PF_FSTRANS;
 759         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 760         current->flags &= ~PF_FSTRANS;
 761 }
 762
 763 /**
 764  * rpc_malloc - allocate an RPC buffer
 765  * @task: RPC task that will use this buffer
 766  * @size: requested byte size
 767  *
 768  * To prevent rpciod from hanging, this allocator never sleeps,
 769  * returning NULL if the request cannot be serviced immediately.
 770  * The caller can arrange to sleep in a way that is safe for rpciod.
 771  *
 772  * Most requests are 'small' (under 2KiB) and can be serviced from a
 773  * mempool, ensuring that NFS reads and writes can always proceed,
 774  * and that there is good locality of reference for these buffers.
 775  *
 776  * In order to avoid memory starvation triggering more writebacks of
 777  * NFS requests, we avoid using GFP_KERNEL.
 778  */
 779 void *rpc_malloc(struct rpc_task *task, size_t size)
 780 {
 781         struct rpc_buffer *buf;
 782         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 783
 784         size += sizeof(struct rpc_buffer);
 785         if (size <= RPC_BUFFER_MAXSIZE)
 786                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 787         else
 788                 buf = kmalloc(size, gfp);
 789
 790         if (!buf)
 791                 return NULL;
 792
 793         buf->len = size;
 794         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 795                         task->tk_pid, size, buf);
 796         return &buf->data;
 797 }
 798 EXPORT_SYMBOL_GPL(rpc_malloc);
 799
 800 /**
 801  * rpc_free - free buffer allocated via rpc_malloc
 802  * @buffer: buffer to free
 803  *
 804  */
 805 void rpc_free(void *buffer)
 806 {
 807         size_t size;
 808         struct rpc_buffer *buf;
 809
 810         if (!buffer)
 811                 return;
 812
 813         buf = container_of(buffer, struct rpc_buffer, data);
 814         size = buf->len;
 815
 816         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 817                         size, buf);
 818
 819         if (size <= RPC_BUFFER_MAXSIZE)
 820                 mempool_free(buf, rpc_buffer_mempool);
 821         else
 822                 kfree(buf);
 823 }
 824 EXPORT_SYMBOL_GPL(rpc_free);
 825
 826 /*
 827  * Creation and deletion of RPC task structures
 828  */
 829 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 830 {
 831         memset(task, 0, sizeof(*task));
 832         atomic_set(&task->tk_count, 1);
 833         task->tk_flags  = task_setup_data->flags;
 834         task->tk_ops = task_setup_data->callback_ops;
 835         task->tk_calldata = task_setup_data->callback_data;
 836         INIT_LIST_HEAD(&task->tk_task);
 837
 838         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 839         task->tk_owner = current->tgid;
 840
 841         /* Initialize workqueue for async tasks */
 842         task->tk_workqueue = task_setup_data->workqueue;
 843
 844         if (task->tk_ops->rpc_call_prepare != NULL)
 845                 task->tk_action = rpc_prepare_task;
 846
 847         rpc_init_task_statistics(task);
 848
 849         dprintk("RPC:       new task initialized, procpid %u\n",
 850                                 task_pid_nr(current));
 851 }
 852
 853 static struct rpc_task *
 854 rpc_alloc_task(void)
 855 {
 856         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 857 }
 858
 859 /*
 860  * Create a new task for the specified client.
 861  */
 862 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 863 {
 864         struct rpc_task *task = setup_data->task;
 865         unsigned short flags = 0;
 866
 867         if (task == NULL) {
 868                 task = rpc_alloc_task();
 869                 if (task == NULL) {
 870                         rpc_release_calldata(setup_data->callback_ops,
 871                                         setup_data->callback_data);
 872                         return ERR_PTR(-ENOMEM);
 873                 }
 874                 flags = RPC_TASK_DYNAMIC;
 875         }
 876
 877         rpc_init_task(task, setup_data);
 878         task->tk_flags |= flags;
 879         dprintk("RPC:       allocated task %p\n", task);
 880         return task;
 881 }
 882
 883 static void rpc_free_task(struct rpc_task *task)
 884 {
 885         const struct rpc_call_ops *tk_ops = task->tk_ops;
 886         void *calldata = task->tk_calldata;
 887
 888         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 889                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 890                 mempool_free(task, rpc_task_mempool);
 891         }
 892         rpc_release_calldata(tk_ops, calldata);
 893 }
 894
 895 static void rpc_async_release(struct work_struct *work)
 896 {
 897         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 898 }
 899
 900 static void rpc_release_resources_task(struct rpc_task *task)
 901 {
 902         if (task->tk_rqstp)
 903                 xprt_release(task);
 904         if (task->tk_msg.rpc_cred) {
 905                 put_rpccred(task->tk_msg.rpc_cred);
 906                 task->tk_msg.rpc_cred = NULL;
 907         }
 908         rpc_task_release_client(task);
 909 }
 910
 911 static void rpc_final_put_task(struct rpc_task *task,
 912                 struct workqueue_struct *q)
 913 {
 914         if (q != NULL) {
 915                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 916                 queue_work(q, &task->u.tk_work);
 917         } else
 918                 rpc_free_task(task);
 919 }
 920
 921 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
 922 {
 923         if (atomic_dec_and_test(&task->tk_count)) {
 924                 rpc_release_resources_task(task);
 925                 rpc_final_put_task(task, q);
 926         }
 927 }
 928
 929 void rpc_put_task(struct rpc_task *task)
 930 {
 931         rpc_do_put_task(task, NULL);
 932 }
 933 EXPORT_SYMBOL_GPL(rpc_put_task);
 934
 935 void rpc_put_task_async(struct rpc_task *task)
 936 {
 937         rpc_do_put_task(task, task->tk_workqueue);
 938 }
 939 EXPORT_SYMBOL_GPL(rpc_put_task_async);
 940
 941 static void rpc_release_task(struct rpc_task *task)
 942 {
 943         dprintk("RPC: %5u release task\n", task->tk_pid);
 944
 945         BUG_ON (RPC_IS_QUEUED(task));
 946
 947         rpc_release_resources_task(task);
 948
 949         /*
 950          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
 951          * so it should be safe to use task->tk_count as a test for whether
 952          * or not any other processes still hold references to our rpc_task.
 953          */
 954         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
 955                 /* Wake up anyone who may be waiting for task completion */
 956                 if (!rpc_complete_task(task))
 957                         return;
 958         } else {
 959                 if (!atomic_dec_and_test(&task->tk_count))
 960                         return;
 961         }
 962         rpc_final_put_task(task, task->tk_workqueue);
 963 }
 964
 965 int rpciod_up(void)
 966 {
 967         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 968 }
 969
 970 void rpciod_down(void)
 971 {
 972         module_put(THIS_MODULE);
 973 }
 974
 975 /*
 976  * Start up the rpciod workqueue.
 977  */
 978 static int rpciod_start(void)
 979 {
 980         struct workqueue_struct *wq;
 981
 982         /*
 983          * Create the rpciod thread and wait for it to start.
 984          */
 985         dprintk("RPC:       creating workqueue rpciod\n");
 986         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
 987         rpciod_workqueue = wq;
 988         return rpciod_workqueue != NULL;
 989 }
 990
 991 static void rpciod_stop(void)
 992 {
 993         struct workqueue_struct *wq = NULL;
 994
 995         if (rpciod_workqueue == NULL)
 996                 return;
 997         dprintk("RPC:       destroying workqueue rpciod\n");
 998
 999         wq = rpciod_workqueue;
1000         rpciod_workqueue = NULL;
1001         destroy_workqueue(wq);
1002 }
1003
1004 void
1005 rpc_destroy_mempool(void)
1006 {
1007         rpciod_stop();
1008         if (rpc_buffer_mempool)
1009                 mempool_destroy(rpc_buffer_mempool);
1010         if (rpc_task_mempool)
1011                 mempool_destroy(rpc_task_mempool);
1012         if (rpc_task_slabp)
1013                 kmem_cache_destroy(rpc_task_slabp);
1014         if (rpc_buffer_slabp)
1015                 kmem_cache_destroy(rpc_buffer_slabp);
1016         rpc_destroy_wait_queue(&delay_queue);
1017 }
1018
1019 int
1020 rpc_init_mempool(void)
1021 {
1022         /*
1023          * The following is not strictly a mempool initialisation,
1024          * but there is no harm in doing it here
1025          */
1026         rpc_init_wait_queue(&delay_queue, "delayq");
1027         if (!rpciod_start())
1028                 goto err_nomem;
1029
1030         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1031                                              sizeof(struct rpc_task),
1032                                              0, SLAB_HWCACHE_ALIGN,
1033                                              NULL);
1034         if (!rpc_task_slabp)
1035                 goto err_nomem;
1036         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1037                                              RPC_BUFFER_MAXSIZE,
1038                                              0, SLAB_HWCACHE_ALIGN,
1039                                              NULL);
1040         if (!rpc_buffer_slabp)
1041                 goto err_nomem;
1042         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1043                                                     rpc_task_slabp);
1044         if (!rpc_task_mempool)
1045                 goto err_nomem;
1046         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1047                                                       rpc_buffer_slabp);
1048         if (!rpc_buffer_mempool)
1049                 goto err_nomem;
1050         return 0;
1051 err_nomem:
1052         rpc_destroy_mempool();
1053         return -ENOMEM;
1054 }