SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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/smp_lock.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/mutex.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #ifdef RPC_DEBUG
  26 #define RPCDBG_FACILITY         RPCDBG_SCHED
  27 #define RPC_TASK_MAGIC_ID       0xf00baa
  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
  44 /*
  45  * RPC tasks sit here while waiting for conditions to improve.
  46  */
  47 static struct rpc_wait_queue delay_queue;
  48
  49 /*
  50  * rpciod-related stuff
  51  */
  52 struct workqueue_struct *rpciod_workqueue;
  53
  54 /*
  55  * Disable the timer for a given RPC task. Should be called with
  56  * queue->lock and bh_disabled in order to avoid races within
  57  * rpc_run_timer().
  58  */
  59 static inline void
  60 __rpc_disable_timer(struct rpc_task *task)
  61 {
  62         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  63         task->tk_timeout_fn = NULL;
  64         task->tk_timeout = 0;
  65 }
  66
  67 /*
  68  * Default timeout handler if none specified by user
  69  */
  70 static void
  71 __rpc_default_timer(struct rpc_task *task)
  72 {
  73         dprintk("RPC: %5u timeout (default timer)\n", task->tk_pid);
  74         task->tk_status = -ETIMEDOUT;
  75 }
  76
  77 /*
  78  * Set up a timer for the current task.
  79  */
  80 static inline void
  81 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
  82 {
  83         if (!task->tk_timeout)
  84                 return;
  85
  86         dprintk("RPC: %5u setting alarm for %lu ms\n",
  87                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  88
  89         if (timer)
  90                 task->tk_timeout_fn = timer;
  91         else
  92                 task->tk_timeout_fn = __rpc_default_timer;
  93         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
  94         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
  95 }
  96
  97 /*
  98  * Delete any timer for the current task. Because we use del_timer_sync(),
  99  * this function should never be called while holding queue->lock.
 100  */
 101 static void
 102 rpc_delete_timer(struct rpc_task *task)
 103 {
 104         if (RPC_IS_QUEUED(task))
 105                 return;
 106         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 107                 del_singleshot_timer_sync(&task->tk_timer);
 108                 dprintk("RPC: %5u deleting timer\n", task->tk_pid);
 109         }
 110 }
 111
 112 /*
 113  * Add new request to a priority queue.
 114  */
 115 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 116 {
 117         struct list_head *q;
 118         struct rpc_task *t;
 119
 120         INIT_LIST_HEAD(&task->u.tk_wait.links);
 121         q = &queue->tasks[task->tk_priority];
 122         if (unlikely(task->tk_priority > queue->maxpriority))
 123                 q = &queue->tasks[queue->maxpriority];
 124         list_for_each_entry(t, q, u.tk_wait.list) {
 125                 if (t->tk_owner == task->tk_owner) {
 126                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 127                         return;
 128                 }
 129         }
 130         list_add_tail(&task->u.tk_wait.list, q);
 131 }
 132
 133 /*
 134  * Add new request to wait queue.
 135  *
 136  * Swapper tasks always get inserted at the head of the queue.
 137  * This should avoid many nasty memory deadlocks and hopefully
 138  * improve overall performance.
 139  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 140  */
 141 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 142 {
 143         BUG_ON (RPC_IS_QUEUED(task));
 144
 145         if (RPC_IS_PRIORITY(queue))
 146                 __rpc_add_wait_queue_priority(queue, task);
 147         else if (RPC_IS_SWAPPER(task))
 148                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 149         else
 150                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 151         task->u.tk_wait.rpc_waitq = queue;
 152         queue->qlen++;
 153         rpc_set_queued(task);
 154
 155         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 156                         task->tk_pid, queue, rpc_qname(queue));
 157 }
 158
 159 /*
 160  * Remove request from a priority queue.
 161  */
 162 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 163 {
 164         struct rpc_task *t;
 165
 166         if (!list_empty(&task->u.tk_wait.links)) {
 167                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 168                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 169                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 170         }
 171         list_del(&task->u.tk_wait.list);
 172 }
 173
 174 /*
 175  * Remove request from queue.
 176  * Note: must be called with spin lock held.
 177  */
 178 static void __rpc_remove_wait_queue(struct rpc_task *task)
 179 {
 180         struct rpc_wait_queue *queue;
 181         queue = task->u.tk_wait.rpc_waitq;
 182
 183         if (RPC_IS_PRIORITY(queue))
 184                 __rpc_remove_wait_queue_priority(task);
 185         else
 186                 list_del(&task->u.tk_wait.list);
 187         queue->qlen--;
 188         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 189                         task->tk_pid, queue, rpc_qname(queue));
 190 }
 191
 192 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 193 {
 194         queue->priority = priority;
 195         queue->count = 1 << (priority * 2);
 196 }
 197
 198 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 199 {
 200         queue->owner = pid;
 201         queue->nr = RPC_BATCH_COUNT;
 202 }
 203
 204 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 205 {
 206         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 207         rpc_set_waitqueue_owner(queue, 0);
 208 }
 209
 210 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 211 {
 212         int i;
 213
 214         spin_lock_init(&queue->lock);
 215         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 216                 INIT_LIST_HEAD(&queue->tasks[i]);
 217         queue->maxpriority = nr_queues - 1;
 218         rpc_reset_waitqueue_priority(queue);
 219 #ifdef RPC_DEBUG
 220         queue->name = qname;
 221 #endif
 222 }
 223
 224 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 225 {
 226         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 227 }
 228
 229 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 230 {
 231         __rpc_init_priority_wait_queue(queue, qname, 1);
 232 }
 233 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 234
 235 static int rpc_wait_bit_killable(void *word)
 236 {
 237         if (fatal_signal_pending(current))
 238                 return -ERESTARTSYS;
 239         schedule();
 240         return 0;
 241 }
 242
 243 #ifdef RPC_DEBUG
 244 static void rpc_task_set_debuginfo(struct rpc_task *task)
 245 {
 246         static atomic_t rpc_pid;
 247
 248         task->tk_magic = RPC_TASK_MAGIC_ID;
 249         task->tk_pid = atomic_inc_return(&rpc_pid);
 250 }
 251 #else
 252 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 253 {
 254 }
 255 #endif
 256
 257 static void rpc_set_active(struct rpc_task *task)
 258 {
 259         struct rpc_clnt *clnt;
 260         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 261                 return;
 262         rpc_task_set_debuginfo(task);
 263         /* Add to global list of all tasks */
 264         clnt = task->tk_client;
 265         if (clnt != NULL) {
 266                 spin_lock(&clnt->cl_lock);
 267                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 268                 spin_unlock(&clnt->cl_lock);
 269         }
 270 }
 271
 272 /*
 273  * Mark an RPC call as having completed by clearing the 'active' bit
 274  */
 275 static void rpc_mark_complete_task(struct rpc_task *task)
 276 {
 277         smp_mb__before_clear_bit();
 278         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 279         smp_mb__after_clear_bit();
 280         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 281 }
 282
 283 /*
 284  * Allow callers to wait for completion of an RPC call
 285  */
 286 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 287 {
 288         if (action == NULL)
 289                 action = rpc_wait_bit_killable;
 290         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 291                         action, TASK_KILLABLE);
 292 }
 293 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 294
 295 /*
 296  * Make an RPC task runnable.
 297  *
 298  * Note: If the task is ASYNC, this must be called with
 299  * the spinlock held to protect the wait queue operation.
 300  */
 301 static void rpc_make_runnable(struct rpc_task *task)
 302 {
 303         BUG_ON(task->tk_timeout_fn);
 304         rpc_clear_queued(task);
 305         if (rpc_test_and_set_running(task))
 306                 return;
 307         /* We might have raced */
 308         if (RPC_IS_QUEUED(task)) {
 309                 rpc_clear_running(task);
 310                 return;
 311         }
 312         if (RPC_IS_ASYNC(task)) {
 313                 int status;
 314
 315                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 316                 status = queue_work(rpciod_workqueue, &task->u.tk_work);
 317                 if (status < 0) {
 318                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 319                         task->tk_status = status;
 320                         return;
 321                 }
 322         } else
 323                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 324 }
 325
 326 /*
 327  * Prepare for sleeping on a wait queue.
 328  * By always appending tasks to the list we ensure FIFO behavior.
 329  * NB: An RPC task will only receive interrupt-driven events as long
 330  * as it's on a wait queue.
 331  */
 332 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 333                         rpc_action action, rpc_action timer)
 334 {
 335         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 336                         task->tk_pid, rpc_qname(q), jiffies);
 337
 338         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 339                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 340                 return;
 341         }
 342
 343         __rpc_add_wait_queue(q, task);
 344
 345         BUG_ON(task->tk_callback != NULL);
 346         task->tk_callback = action;
 347         __rpc_add_timer(task, timer);
 348 }
 349
 350 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 351                                 rpc_action action, rpc_action timer)
 352 {
 353         /* Mark the task as being activated if so needed */
 354         rpc_set_active(task);
 355
 356         /*
 357          * Protect the queue operations.
 358          */
 359         spin_lock_bh(&q->lock);
 360         __rpc_sleep_on(q, task, action, timer);
 361         spin_unlock_bh(&q->lock);
 362 }
 363 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 364
 365 /**
 366  * __rpc_do_wake_up_task - wake up a single rpc_task
 367  * @task: task to be woken up
 368  *
 369  * Caller must hold queue->lock, and have cleared the task queued flag.
 370  */
 371 static void __rpc_do_wake_up_task(struct rpc_task *task)
 372 {
 373         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 374                         task->tk_pid, jiffies);
 375
 376 #ifdef RPC_DEBUG
 377         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 378 #endif
 379         /* Has the task been executed yet? If not, we cannot wake it up! */
 380         if (!RPC_IS_ACTIVATED(task)) {
 381                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 382                 return;
 383         }
 384
 385         __rpc_disable_timer(task);
 386         __rpc_remove_wait_queue(task);
 387
 388         rpc_make_runnable(task);
 389
 390         dprintk("RPC:       __rpc_wake_up_task done\n");
 391 }
 392
 393 /*
 394  * Wake up the specified task
 395  */
 396 static void __rpc_wake_up_task(struct rpc_task *task)
 397 {
 398         if (rpc_start_wakeup(task)) {
 399                 if (RPC_IS_QUEUED(task))
 400                         __rpc_do_wake_up_task(task);
 401                 rpc_finish_wakeup(task);
 402         }
 403 }
 404
 405 /*
 406  * Wake up the specified task
 407  */
 408 void rpc_wake_up_task(struct rpc_task *task)
 409 {
 410         rcu_read_lock_bh();
 411         if (rpc_start_wakeup(task)) {
 412                 if (RPC_IS_QUEUED(task)) {
 413                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 414
 415                         /* Note: we're already in a bh-safe context */
 416                         spin_lock(&queue->lock);
 417                         __rpc_do_wake_up_task(task);
 418                         spin_unlock(&queue->lock);
 419                 }
 420                 rpc_finish_wakeup(task);
 421         }
 422         rcu_read_unlock_bh();
 423 }
 424 EXPORT_SYMBOL_GPL(rpc_wake_up_task);
 425
 426 /*
 427  * Wake up the next task on a priority queue.
 428  */
 429 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 430 {
 431         struct list_head *q;
 432         struct rpc_task *task;
 433
 434         /*
 435          * Service a batch of tasks from a single owner.
 436          */
 437         q = &queue->tasks[queue->priority];
 438         if (!list_empty(q)) {
 439                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 440                 if (queue->owner == task->tk_owner) {
 441                         if (--queue->nr)
 442                                 goto out;
 443                         list_move_tail(&task->u.tk_wait.list, q);
 444                 }
 445                 /*
 446                  * Check if we need to switch queues.
 447                  */
 448                 if (--queue->count)
 449                         goto new_owner;
 450         }
 451
 452         /*
 453          * Service the next queue.
 454          */
 455         do {
 456                 if (q == &queue->tasks[0])
 457                         q = &queue->tasks[queue->maxpriority];
 458                 else
 459                         q = q - 1;
 460                 if (!list_empty(q)) {
 461                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 462                         goto new_queue;
 463                 }
 464         } while (q != &queue->tasks[queue->priority]);
 465
 466         rpc_reset_waitqueue_priority(queue);
 467         return NULL;
 468
 469 new_queue:
 470         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 471 new_owner:
 472         rpc_set_waitqueue_owner(queue, task->tk_owner);
 473 out:
 474         __rpc_wake_up_task(task);
 475         return task;
 476 }
 477
 478 /*
 479  * Wake up the next task on the wait queue.
 480  */
 481 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 482 {
 483         struct rpc_task *task = NULL;
 484
 485         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 486                         queue, rpc_qname(queue));
 487         rcu_read_lock_bh();
 488         spin_lock(&queue->lock);
 489         if (RPC_IS_PRIORITY(queue))
 490                 task = __rpc_wake_up_next_priority(queue);
 491         else {
 492                 task_for_first(task, &queue->tasks[0])
 493                         __rpc_wake_up_task(task);
 494         }
 495         spin_unlock(&queue->lock);
 496         rcu_read_unlock_bh();
 497
 498         return task;
 499 }
 500 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 501
 502 /**
 503  * rpc_wake_up - wake up all rpc_tasks
 504  * @queue: rpc_wait_queue on which the tasks are sleeping
 505  *
 506  * Grabs queue->lock
 507  */
 508 void rpc_wake_up(struct rpc_wait_queue *queue)
 509 {
 510         struct rpc_task *task, *next;
 511         struct list_head *head;
 512
 513         rcu_read_lock_bh();
 514         spin_lock(&queue->lock);
 515         head = &queue->tasks[queue->maxpriority];
 516         for (;;) {
 517                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 518                         __rpc_wake_up_task(task);
 519                 if (head == &queue->tasks[0])
 520                         break;
 521                 head--;
 522         }
 523         spin_unlock(&queue->lock);
 524         rcu_read_unlock_bh();
 525 }
 526 EXPORT_SYMBOL_GPL(rpc_wake_up);
 527
 528 /**
 529  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 530  * @queue: rpc_wait_queue on which the tasks are sleeping
 531  * @status: status value to set
 532  *
 533  * Grabs queue->lock
 534  */
 535 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 536 {
 537         struct rpc_task *task, *next;
 538         struct list_head *head;
 539
 540         rcu_read_lock_bh();
 541         spin_lock(&queue->lock);
 542         head = &queue->tasks[queue->maxpriority];
 543         for (;;) {
 544                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 545                         task->tk_status = status;
 546                         __rpc_wake_up_task(task);
 547                 }
 548                 if (head == &queue->tasks[0])
 549                         break;
 550                 head--;
 551         }
 552         spin_unlock(&queue->lock);
 553         rcu_read_unlock_bh();
 554 }
 555 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 556
 557 /*
 558  * Run a timeout function.
 559  */
 560 static void rpc_run_timer(unsigned long ptr)
 561 {
 562         struct rpc_task *task = (struct rpc_task *)ptr;
 563         void (*callback)(struct rpc_task *);
 564
 565         if (!rpc_start_wakeup(task))
 566                 goto out;
 567         if (RPC_IS_QUEUED(task)) {
 568                 struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 569                 callback = task->tk_timeout_fn;
 570
 571                 dprintk("RPC: %5u running timer\n", task->tk_pid);
 572                 if (callback != NULL)
 573                         callback(task);
 574                 /* Note: we're already in a bh-safe context */
 575                 spin_lock(&queue->lock);
 576                 __rpc_do_wake_up_task(task);
 577                 spin_unlock(&queue->lock);
 578         }
 579         rpc_finish_wakeup(task);
 580 out:
 581         smp_mb__before_clear_bit();
 582         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 583         smp_mb__after_clear_bit();
 584 }
 585
 586 static void __rpc_atrun(struct rpc_task *task)
 587 {
 588 }
 589
 590 /*
 591  * Run a task at a later time
 592  */
 593 void rpc_delay(struct rpc_task *task, unsigned long delay)
 594 {
 595         task->tk_timeout = delay;
 596         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 597 }
 598 EXPORT_SYMBOL_GPL(rpc_delay);
 599
 600 /*
 601  * Helper to call task->tk_ops->rpc_call_prepare
 602  */
 603 static void rpc_prepare_task(struct rpc_task *task)
 604 {
 605         lock_kernel();
 606         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 607         unlock_kernel();
 608 }
 609
 610 /*
 611  * Helper that calls task->tk_ops->rpc_call_done if it exists
 612  */
 613 void rpc_exit_task(struct rpc_task *task)
 614 {
 615         task->tk_action = NULL;
 616         if (task->tk_ops->rpc_call_done != NULL) {
 617                 lock_kernel();
 618                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 619                 unlock_kernel();
 620                 if (task->tk_action != NULL) {
 621                         WARN_ON(RPC_ASSASSINATED(task));
 622                         /* Always release the RPC slot and buffer memory */
 623                         xprt_release(task);
 624                 }
 625         }
 626 }
 627 EXPORT_SYMBOL_GPL(rpc_exit_task);
 628
 629 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 630 {
 631         if (ops->rpc_release != NULL) {
 632                 lock_kernel();
 633                 ops->rpc_release(calldata);
 634                 unlock_kernel();
 635         }
 636 }
 637
 638 /*
 639  * This is the RPC `scheduler' (or rather, the finite state machine).
 640  */
 641 static void __rpc_execute(struct rpc_task *task)
 642 {
 643         int             status = 0;
 644
 645         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 646                         task->tk_pid, task->tk_flags);
 647
 648         BUG_ON(RPC_IS_QUEUED(task));
 649
 650         for (;;) {
 651                 /*
 652                  * Garbage collection of pending timers...
 653                  */
 654                 rpc_delete_timer(task);
 655
 656                 /*
 657                  * Execute any pending callback.
 658                  */
 659                 if (RPC_DO_CALLBACK(task)) {
 660                         /* Define a callback save pointer */
 661                         void (*save_callback)(struct rpc_task *);
 662
 663                         /*
 664                          * If a callback exists, save it, reset it,
 665                          * call it.
 666                          * The save is needed to stop from resetting
 667                          * another callback set within the callback handler
 668                          * - Dave
 669                          */
 670                         save_callback=task->tk_callback;
 671                         task->tk_callback=NULL;
 672                         save_callback(task);
 673                 }
 674
 675                 /*
 676                  * Perform the next FSM step.
 677                  * tk_action may be NULL when the task has been killed
 678                  * by someone else.
 679                  */
 680                 if (!RPC_IS_QUEUED(task)) {
 681                         if (task->tk_action == NULL)
 682                                 break;
 683                         task->tk_action(task);
 684                 }
 685
 686                 /*
 687                  * Lockless check for whether task is sleeping or not.
 688                  */
 689                 if (!RPC_IS_QUEUED(task))
 690                         continue;
 691                 rpc_clear_running(task);
 692                 if (RPC_IS_ASYNC(task)) {
 693                         /* Careful! we may have raced... */
 694                         if (RPC_IS_QUEUED(task))
 695                                 return;
 696                         if (rpc_test_and_set_running(task))
 697                                 return;
 698                         continue;
 699                 }
 700
 701                 /* sync task: sleep here */
 702                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 703                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 704                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 705                                 TASK_KILLABLE);
 706                 if (status == -ERESTARTSYS) {
 707                         /*
 708                          * When a sync task receives a signal, it exits with
 709                          * -ERESTARTSYS. In order to catch any callbacks that
 710                          * clean up after sleeping on some queue, we don't
 711                          * break the loop here, but go around once more.
 712                          */
 713                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 714                         task->tk_flags |= RPC_TASK_KILLED;
 715                         rpc_exit(task, -ERESTARTSYS);
 716                         rpc_wake_up_task(task);
 717                 }
 718                 rpc_set_running(task);
 719                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 720         }
 721
 722         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 723                         task->tk_status);
 724         /* Release all resources associated with the task */
 725         rpc_release_task(task);
 726 }
 727
 728 /*
 729  * User-visible entry point to the scheduler.
 730  *
 731  * This may be called recursively if e.g. an async NFS task updates
 732  * the attributes and finds that dirty pages must be flushed.
 733  * NOTE: Upon exit of this function the task is guaranteed to be
 734  *       released. In particular note that tk_release() will have
 735  *       been called, so your task memory may have been freed.
 736  */
 737 void rpc_execute(struct rpc_task *task)
 738 {
 739         rpc_set_active(task);
 740         rpc_set_running(task);
 741         __rpc_execute(task);
 742 }
 743
 744 static void rpc_async_schedule(struct work_struct *work)
 745 {
 746         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 747 }
 748
 749 struct rpc_buffer {
 750         size_t  len;
 751         char    data[];
 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         setup_timer(&task->tk_timer, rpc_run_timer, (unsigned long)task);
 824         atomic_set(&task->tk_count, 1);
 825         task->tk_flags  = task_setup_data->flags;
 826         task->tk_ops = task_setup_data->callback_ops;
 827         task->tk_calldata = task_setup_data->callback_data;
 828         INIT_LIST_HEAD(&task->tk_task);
 829
 830         /* Initialize retry counters */
 831         task->tk_garb_retry = 2;
 832         task->tk_cred_retry = 2;
 833
 834         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 835         task->tk_owner = current->tgid;
 836
 837         /* Initialize workqueue for async tasks */
 838         task->tk_workqueue = task_setup_data->workqueue;
 839
 840         task->tk_client = task_setup_data->rpc_client;
 841         if (task->tk_client != NULL) {
 842                 kref_get(&task->tk_client->cl_kref);
 843                 if (task->tk_client->cl_softrtry)
 844                         task->tk_flags |= RPC_TASK_SOFT;
 845         }
 846
 847         if (task->tk_ops->rpc_call_prepare != NULL)
 848                 task->tk_action = rpc_prepare_task;
 849
 850         if (task_setup_data->rpc_message != NULL) {
 851                 memcpy(&task->tk_msg, task_setup_data->rpc_message, sizeof(task->tk_msg));
 852                 /* Bind the user cred */
 853                 if (task->tk_msg.rpc_cred != NULL)
 854                         rpcauth_holdcred(task);
 855                 else
 856                         rpcauth_bindcred(task);
 857                 if (task->tk_action == NULL)
 858                         rpc_call_start(task);
 859         }
 860
 861         /* starting timestamp */
 862         task->tk_start = jiffies;
 863
 864         dprintk("RPC:       new task initialized, procpid %u\n",
 865                                 task_pid_nr(current));
 866 }
 867
 868 static struct rpc_task *
 869 rpc_alloc_task(void)
 870 {
 871         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 872 }
 873
 874 static void rpc_free_task_rcu(struct rcu_head *rcu)
 875 {
 876         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 877         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 878         mempool_free(task, rpc_task_mempool);
 879 }
 880
 881 /*
 882  * Create a new task for the specified client.
 883  */
 884 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 885 {
 886         struct rpc_task *task = setup_data->task;
 887         unsigned short flags = 0;
 888
 889         if (task == NULL) {
 890                 task = rpc_alloc_task();
 891                 if (task == NULL)
 892                         goto out;
 893                 flags = RPC_TASK_DYNAMIC;
 894         }
 895
 896         rpc_init_task(task, setup_data);
 897
 898         task->tk_flags |= flags;
 899         dprintk("RPC:       allocated task %p\n", task);
 900 out:
 901         return task;
 902 }
 903
 904 static void rpc_free_task(struct rpc_task *task)
 905 {
 906         const struct rpc_call_ops *tk_ops = task->tk_ops;
 907         void *calldata = task->tk_calldata;
 908
 909         if (task->tk_flags & RPC_TASK_DYNAMIC)
 910                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task_rcu);
 911         rpc_release_calldata(tk_ops, calldata);
 912 }
 913
 914 static void rpc_async_release(struct work_struct *work)
 915 {
 916         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 917 }
 918
 919 void rpc_put_task(struct rpc_task *task)
 920 {
 921         if (!atomic_dec_and_test(&task->tk_count))
 922                 return;
 923         /* Release resources */
 924         if (task->tk_rqstp)
 925                 xprt_release(task);
 926         if (task->tk_msg.rpc_cred)
 927                 rpcauth_unbindcred(task);
 928         if (task->tk_client) {
 929                 rpc_release_client(task->tk_client);
 930                 task->tk_client = NULL;
 931         }
 932         if (task->tk_workqueue != NULL) {
 933                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 934                 queue_work(task->tk_workqueue, &task->u.tk_work);
 935         } else
 936                 rpc_free_task(task);
 937 }
 938 EXPORT_SYMBOL_GPL(rpc_put_task);
 939
 940 static void rpc_release_task(struct rpc_task *task)
 941 {
 942 #ifdef RPC_DEBUG
 943         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 944 #endif
 945         dprintk("RPC: %5u release task\n", task->tk_pid);
 946
 947         if (!list_empty(&task->tk_task)) {
 948                 struct rpc_clnt *clnt = task->tk_client;
 949                 /* Remove from client task list */
 950                 spin_lock(&clnt->cl_lock);
 951                 list_del(&task->tk_task);
 952                 spin_unlock(&clnt->cl_lock);
 953         }
 954         BUG_ON (RPC_IS_QUEUED(task));
 955
 956         /* Synchronously delete any running timer */
 957         rpc_delete_timer(task);
 958
 959 #ifdef RPC_DEBUG
 960         task->tk_magic = 0;
 961 #endif
 962         /* Wake up anyone who is waiting for task completion */
 963         rpc_mark_complete_task(task);
 964
 965         rpc_put_task(task);
 966 }
 967
 968 /*
 969  * Kill all tasks for the given client.
 970  * XXX: kill their descendants as well?
 971  */
 972 void rpc_killall_tasks(struct rpc_clnt *clnt)
 973 {
 974         struct rpc_task *rovr;
 975
 976
 977         if (list_empty(&clnt->cl_tasks))
 978                 return;
 979         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 980         /*
 981          * Spin lock all_tasks to prevent changes...
 982          */
 983         spin_lock(&clnt->cl_lock);
 984         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 985                 if (! RPC_IS_ACTIVATED(rovr))
 986                         continue;
 987                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 988                         rovr->tk_flags |= RPC_TASK_KILLED;
 989                         rpc_exit(rovr, -EIO);
 990                         rpc_wake_up_task(rovr);
 991                 }
 992         }
 993         spin_unlock(&clnt->cl_lock);
 994 }
 995 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 996
 997 int rpciod_up(void)
 998 {
 999         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1000 }
1001
1002 void rpciod_down(void)
1003 {
1004         module_put(THIS_MODULE);
1005 }
1006
1007 /*
1008  * Start up the rpciod workqueue.
1009  */
1010 static int rpciod_start(void)
1011 {
1012         struct workqueue_struct *wq;
1013
1014         /*
1015          * Create the rpciod thread and wait for it to start.
1016          */
1017         dprintk("RPC:       creating workqueue rpciod\n");
1018         wq = create_workqueue("rpciod");
1019         rpciod_workqueue = wq;
1020         return rpciod_workqueue != NULL;
1021 }
1022
1023 static void rpciod_stop(void)
1024 {
1025         struct workqueue_struct *wq = NULL;
1026
1027         if (rpciod_workqueue == NULL)
1028                 return;
1029         dprintk("RPC:       destroying workqueue rpciod\n");
1030
1031         wq = rpciod_workqueue;
1032         rpciod_workqueue = NULL;
1033         destroy_workqueue(wq);
1034 }
1035
1036 void
1037 rpc_destroy_mempool(void)
1038 {
1039         rpciod_stop();
1040         if (rpc_buffer_mempool)
1041                 mempool_destroy(rpc_buffer_mempool);
1042         if (rpc_task_mempool)
1043                 mempool_destroy(rpc_task_mempool);
1044         if (rpc_task_slabp)
1045                 kmem_cache_destroy(rpc_task_slabp);
1046         if (rpc_buffer_slabp)
1047                 kmem_cache_destroy(rpc_buffer_slabp);
1048 }
1049
1050 int
1051 rpc_init_mempool(void)
1052 {
1053         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1054                                              sizeof(struct rpc_task),
1055                                              0, SLAB_HWCACHE_ALIGN,
1056                                              NULL);
1057         if (!rpc_task_slabp)
1058                 goto err_nomem;
1059         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1060                                              RPC_BUFFER_MAXSIZE,
1061                                              0, SLAB_HWCACHE_ALIGN,
1062                                              NULL);
1063         if (!rpc_buffer_slabp)
1064                 goto err_nomem;
1065         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1066                                                     rpc_task_slabp);
1067         if (!rpc_task_mempool)
1068                 goto err_nomem;
1069         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1070                                                       rpc_buffer_slabp);
1071         if (!rpc_buffer_mempool)
1072                 goto err_nomem;
1073         if (!rpciod_start())
1074                 goto err_nomem;
1075         /*
1076          * The following is not strictly a mempool initialisation,
1077          * but there is no harm in doing it here
1078          */
1079         rpc_init_wait_queue(&delay_queue, "delayq");
1080         return 0;
1081 err_nomem:
1082         rpc_destroy_mempool();
1083         return -ENOMEM;
1084 }