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 void
  60 __rpc_disable_timer(struct rpc_task *task)
  61 {
  62         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  63         task->tk_timeout = 0;
  64 }
  65
  66 /*
  67  * Set up a timer for the current task.
  68  */
  69 static void
  70 __rpc_add_timer(struct rpc_task *task)
  71 {
  72         if (!task->tk_timeout)
  73                 return;
  74
  75         dprintk("RPC: %5u setting alarm for %lu ms\n",
  76                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  77
  78         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
  79         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
  80 }
  81
  82 /*
  83  * Delete any timer for the current task. Because we use del_timer_sync(),
  84  * this function should never be called while holding queue->lock.
  85  */
  86 static void
  87 rpc_delete_timer(struct rpc_task *task)
  88 {
  89         if (RPC_IS_QUEUED(task))
  90                 return;
  91         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
  92                 del_singleshot_timer_sync(&task->tk_timer);
  93                 dprintk("RPC: %5u deleting timer\n", task->tk_pid);
  94         }
  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, struct rpc_task *task)
 101 {
 102         struct list_head *q;
 103         struct rpc_task *t;
 104
 105         INIT_LIST_HEAD(&task->u.tk_wait.links);
 106         q = &queue->tasks[task->tk_priority];
 107         if (unlikely(task->tk_priority > queue->maxpriority))
 108                 q = &queue->tasks[queue->maxpriority];
 109         list_for_each_entry(t, q, u.tk_wait.list) {
 110                 if (t->tk_owner == task->tk_owner) {
 111                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 112                         return;
 113                 }
 114         }
 115         list_add_tail(&task->u.tk_wait.list, q);
 116 }
 117
 118 /*
 119  * Add new request to wait queue.
 120  *
 121  * Swapper tasks always get inserted at the head of the queue.
 122  * This should avoid many nasty memory deadlocks and hopefully
 123  * improve overall performance.
 124  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 125  */
 126 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 127 {
 128         BUG_ON (RPC_IS_QUEUED(task));
 129
 130         if (RPC_IS_PRIORITY(queue))
 131                 __rpc_add_wait_queue_priority(queue, task);
 132         else if (RPC_IS_SWAPPER(task))
 133                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 134         else
 135                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 136         task->tk_waitqueue = queue;
 137         queue->qlen++;
 138         rpc_set_queued(task);
 139
 140         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 141                         task->tk_pid, queue, rpc_qname(queue));
 142 }
 143
 144 /*
 145  * Remove request from a priority queue.
 146  */
 147 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 148 {
 149         struct rpc_task *t;
 150
 151         if (!list_empty(&task->u.tk_wait.links)) {
 152                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 153                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 154                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 155         }
 156         list_del(&task->u.tk_wait.list);
 157 }
 158
 159 /*
 160  * Remove request from queue.
 161  * Note: must be called with spin lock held.
 162  */
 163 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 164 {
 165         if (RPC_IS_PRIORITY(queue))
 166                 __rpc_remove_wait_queue_priority(task);
 167         else
 168                 list_del(&task->u.tk_wait.list);
 169         queue->qlen--;
 170         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 171                         task->tk_pid, queue, rpc_qname(queue));
 172 }
 173
 174 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 175 {
 176         queue->priority = priority;
 177         queue->count = 1 << (priority * 2);
 178 }
 179
 180 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 181 {
 182         queue->owner = pid;
 183         queue->nr = RPC_BATCH_COUNT;
 184 }
 185
 186 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 187 {
 188         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 189         rpc_set_waitqueue_owner(queue, 0);
 190 }
 191
 192 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 193 {
 194         int i;
 195
 196         spin_lock_init(&queue->lock);
 197         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 198                 INIT_LIST_HEAD(&queue->tasks[i]);
 199         queue->maxpriority = nr_queues - 1;
 200         rpc_reset_waitqueue_priority(queue);
 201 #ifdef RPC_DEBUG
 202         queue->name = qname;
 203 #endif
 204 }
 205
 206 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 207 {
 208         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 209 }
 210
 211 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 212 {
 213         __rpc_init_priority_wait_queue(queue, qname, 1);
 214 }
 215 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 216
 217 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 218 {
 219 }
 220 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 221
 222 static int rpc_wait_bit_killable(void *word)
 223 {
 224         if (fatal_signal_pending(current))
 225                 return -ERESTARTSYS;
 226         schedule();
 227         return 0;
 228 }
 229
 230 #ifdef RPC_DEBUG
 231 static void rpc_task_set_debuginfo(struct rpc_task *task)
 232 {
 233         static atomic_t rpc_pid;
 234
 235         task->tk_magic = RPC_TASK_MAGIC_ID;
 236         task->tk_pid = atomic_inc_return(&rpc_pid);
 237 }
 238 #else
 239 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 240 {
 241 }
 242 #endif
 243
 244 static void rpc_set_active(struct rpc_task *task)
 245 {
 246         struct rpc_clnt *clnt;
 247         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 248                 return;
 249         rpc_task_set_debuginfo(task);
 250         /* Add to global list of all tasks */
 251         clnt = task->tk_client;
 252         if (clnt != NULL) {
 253                 spin_lock(&clnt->cl_lock);
 254                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 255                 spin_unlock(&clnt->cl_lock);
 256         }
 257 }
 258
 259 /*
 260  * Mark an RPC call as having completed by clearing the 'active' bit
 261  */
 262 static void rpc_mark_complete_task(struct rpc_task *task)
 263 {
 264         smp_mb__before_clear_bit();
 265         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 266         smp_mb__after_clear_bit();
 267         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 268 }
 269
 270 /*
 271  * Allow callers to wait for completion of an RPC call
 272  */
 273 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 274 {
 275         if (action == NULL)
 276                 action = rpc_wait_bit_killable;
 277         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 278                         action, TASK_KILLABLE);
 279 }
 280 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 281
 282 /*
 283  * Make an RPC task runnable.
 284  *
 285  * Note: If the task is ASYNC, this must be called with
 286  * the spinlock held to protect the wait queue operation.
 287  */
 288 static void rpc_make_runnable(struct rpc_task *task)
 289 {
 290         rpc_clear_queued(task);
 291         if (rpc_test_and_set_running(task))
 292                 return;
 293         /* We might have raced */
 294         if (RPC_IS_QUEUED(task)) {
 295                 rpc_clear_running(task);
 296                 return;
 297         }
 298         if (RPC_IS_ASYNC(task)) {
 299                 int status;
 300
 301                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 302                 status = queue_work(rpciod_workqueue, &task->u.tk_work);
 303                 if (status < 0) {
 304                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 305                         task->tk_status = status;
 306                         return;
 307                 }
 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(struct rpc_wait_queue *q, struct rpc_task *task,
 319                         rpc_action action)
 320 {
 321         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 322                         task->tk_pid, rpc_qname(q), jiffies);
 323
 324         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 325                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 326                 return;
 327         }
 328
 329         __rpc_add_wait_queue(q, task);
 330
 331         BUG_ON(task->tk_callback != NULL);
 332         task->tk_callback = action;
 333         __rpc_add_timer(task);
 334 }
 335
 336 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 337                                 rpc_action action)
 338 {
 339         /* Mark the task as being activated if so needed */
 340         rpc_set_active(task);
 341
 342         /*
 343          * Protect the queue operations.
 344          */
 345         spin_lock_bh(&q->lock);
 346         __rpc_sleep_on(q, task, action);
 347         spin_unlock_bh(&q->lock);
 348 }
 349 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 350
 351 /**
 352  * __rpc_do_wake_up_task - wake up a single rpc_task
 353  * @queue: wait queue
 354  * @task: task to be woken up
 355  *
 356  * Caller must hold queue->lock, and have cleared the task queued flag.
 357  */
 358 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 359 {
 360         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 361                         task->tk_pid, jiffies);
 362
 363 #ifdef RPC_DEBUG
 364         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 365 #endif
 366         /* Has the task been executed yet? If not, we cannot wake it up! */
 367         if (!RPC_IS_ACTIVATED(task)) {
 368                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 369                 return;
 370         }
 371
 372         __rpc_disable_timer(task);
 373         __rpc_remove_wait_queue(queue, task);
 374
 375         rpc_make_runnable(task);
 376
 377         dprintk("RPC:       __rpc_wake_up_task done\n");
 378 }
 379
 380 /*
 381  * Wake up a queued task while the queue lock is being held
 382  */
 383 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 384 {
 385         if (!RPC_IS_QUEUED(task) || task->tk_waitqueue != queue)
 386                 return;
 387         if (rpc_start_wakeup(task)) {
 388                         __rpc_do_wake_up_task(queue, task);
 389                 rpc_finish_wakeup(task);
 390         }
 391 }
 392
 393 /*
 394  * Wake up a task on a specific queue
 395  */
 396 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 397 {
 398         rcu_read_lock_bh();
 399         spin_lock(&queue->lock);
 400         rpc_wake_up_task_queue_locked(queue, task);
 401         spin_unlock(&queue->lock);
 402         rcu_read_unlock_bh();
 403 }
 404 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 405
 406 /*
 407  * Wake up the specified task
 408  */
 409 static void rpc_wake_up_task(struct rpc_task *task)
 410 {
 411         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 412 }
 413
 414 /*
 415  * Wake up the next task on a priority queue.
 416  */
 417 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 418 {
 419         struct list_head *q;
 420         struct rpc_task *task;
 421
 422         /*
 423          * Service a batch of tasks from a single owner.
 424          */
 425         q = &queue->tasks[queue->priority];
 426         if (!list_empty(q)) {
 427                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 428                 if (queue->owner == task->tk_owner) {
 429                         if (--queue->nr)
 430                                 goto out;
 431                         list_move_tail(&task->u.tk_wait.list, q);
 432                 }
 433                 /*
 434                  * Check if we need to switch queues.
 435                  */
 436                 if (--queue->count)
 437                         goto new_owner;
 438         }
 439
 440         /*
 441          * Service the next queue.
 442          */
 443         do {
 444                 if (q == &queue->tasks[0])
 445                         q = &queue->tasks[queue->maxpriority];
 446                 else
 447                         q = q - 1;
 448                 if (!list_empty(q)) {
 449                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 450                         goto new_queue;
 451                 }
 452         } while (q != &queue->tasks[queue->priority]);
 453
 454         rpc_reset_waitqueue_priority(queue);
 455         return NULL;
 456
 457 new_queue:
 458         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 459 new_owner:
 460         rpc_set_waitqueue_owner(queue, task->tk_owner);
 461 out:
 462         rpc_wake_up_task_queue_locked(queue, task);
 463         return task;
 464 }
 465
 466 /*
 467  * Wake up the next task on the wait queue.
 468  */
 469 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 470 {
 471         struct rpc_task *task = NULL;
 472
 473         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 474                         queue, rpc_qname(queue));
 475         rcu_read_lock_bh();
 476         spin_lock(&queue->lock);
 477         if (RPC_IS_PRIORITY(queue))
 478                 task = __rpc_wake_up_next_priority(queue);
 479         else {
 480                 task_for_first(task, &queue->tasks[0])
 481                         rpc_wake_up_task_queue_locked(queue, task);
 482         }
 483         spin_unlock(&queue->lock);
 484         rcu_read_unlock_bh();
 485
 486         return task;
 487 }
 488 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 489
 490 /**
 491  * rpc_wake_up - wake up all rpc_tasks
 492  * @queue: rpc_wait_queue on which the tasks are sleeping
 493  *
 494  * Grabs queue->lock
 495  */
 496 void rpc_wake_up(struct rpc_wait_queue *queue)
 497 {
 498         struct rpc_task *task, *next;
 499         struct list_head *head;
 500
 501         rcu_read_lock_bh();
 502         spin_lock(&queue->lock);
 503         head = &queue->tasks[queue->maxpriority];
 504         for (;;) {
 505                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 506                         rpc_wake_up_task_queue_locked(queue, task);
 507                 if (head == &queue->tasks[0])
 508                         break;
 509                 head--;
 510         }
 511         spin_unlock(&queue->lock);
 512         rcu_read_unlock_bh();
 513 }
 514 EXPORT_SYMBOL_GPL(rpc_wake_up);
 515
 516 /**
 517  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 518  * @queue: rpc_wait_queue on which the tasks are sleeping
 519  * @status: status value to set
 520  *
 521  * Grabs queue->lock
 522  */
 523 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 524 {
 525         struct rpc_task *task, *next;
 526         struct list_head *head;
 527
 528         rcu_read_lock_bh();
 529         spin_lock(&queue->lock);
 530         head = &queue->tasks[queue->maxpriority];
 531         for (;;) {
 532                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 533                         task->tk_status = status;
 534                         rpc_wake_up_task_queue_locked(queue, task);
 535                 }
 536                 if (head == &queue->tasks[0])
 537                         break;
 538                 head--;
 539         }
 540         spin_unlock(&queue->lock);
 541         rcu_read_unlock_bh();
 542 }
 543 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 544
 545 /*
 546  * Run a timeout function.
 547  */
 548 static void rpc_run_timer(unsigned long ptr)
 549 {
 550         struct rpc_task *task = (struct rpc_task *)ptr;
 551         struct rpc_wait_queue *queue = task->tk_waitqueue;
 552
 553         spin_lock(&queue->lock);
 554         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue) {
 555                 dprintk("RPC: %5u timeout\n", task->tk_pid);
 556                 task->tk_status = -ETIMEDOUT;
 557                 rpc_wake_up_task_queue_locked(queue, task);
 558         }
 559         spin_unlock(&queue->lock);
 560         smp_mb__before_clear_bit();
 561         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 562         smp_mb__after_clear_bit();
 563 }
 564
 565 static void __rpc_atrun(struct rpc_task *task)
 566 {
 567         task->tk_status = 0;
 568 }
 569
 570 /*
 571  * Run a task at a later time
 572  */
 573 void rpc_delay(struct rpc_task *task, unsigned long delay)
 574 {
 575         task->tk_timeout = delay;
 576         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 577 }
 578 EXPORT_SYMBOL_GPL(rpc_delay);
 579
 580 /*
 581  * Helper to call task->tk_ops->rpc_call_prepare
 582  */
 583 static void rpc_prepare_task(struct rpc_task *task)
 584 {
 585         lock_kernel();
 586         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 587         unlock_kernel();
 588 }
 589
 590 /*
 591  * Helper that calls task->tk_ops->rpc_call_done if it exists
 592  */
 593 void rpc_exit_task(struct rpc_task *task)
 594 {
 595         task->tk_action = NULL;
 596         if (task->tk_ops->rpc_call_done != NULL) {
 597                 lock_kernel();
 598                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 599                 unlock_kernel();
 600                 if (task->tk_action != NULL) {
 601                         WARN_ON(RPC_ASSASSINATED(task));
 602                         /* Always release the RPC slot and buffer memory */
 603                         xprt_release(task);
 604                 }
 605         }
 606 }
 607 EXPORT_SYMBOL_GPL(rpc_exit_task);
 608
 609 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 610 {
 611         if (ops->rpc_release != NULL) {
 612                 lock_kernel();
 613                 ops->rpc_release(calldata);
 614                 unlock_kernel();
 615         }
 616 }
 617
 618 /*
 619  * This is the RPC `scheduler' (or rather, the finite state machine).
 620  */
 621 static void __rpc_execute(struct rpc_task *task)
 622 {
 623         int             status = 0;
 624
 625         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 626                         task->tk_pid, task->tk_flags);
 627
 628         BUG_ON(RPC_IS_QUEUED(task));
 629
 630         for (;;) {
 631                 /*
 632                  * Garbage collection of pending timers...
 633                  */
 634                 rpc_delete_timer(task);
 635
 636                 /*
 637                  * Execute any pending callback.
 638                  */
 639                 if (RPC_DO_CALLBACK(task)) {
 640                         /* Define a callback save pointer */
 641                         void (*save_callback)(struct rpc_task *);
 642
 643                         /*
 644                          * If a callback exists, save it, reset it,
 645                          * call it.
 646                          * The save is needed to stop from resetting
 647                          * another callback set within the callback handler
 648                          * - Dave
 649                          */
 650                         save_callback=task->tk_callback;
 651                         task->tk_callback=NULL;
 652                         save_callback(task);
 653                 }
 654
 655                 /*
 656                  * Perform the next FSM step.
 657                  * tk_action may be NULL when the task has been killed
 658                  * by someone else.
 659                  */
 660                 if (!RPC_IS_QUEUED(task)) {
 661                         if (task->tk_action == NULL)
 662                                 break;
 663                         task->tk_action(task);
 664                 }
 665
 666                 /*
 667                  * Lockless check for whether task is sleeping or not.
 668                  */
 669                 if (!RPC_IS_QUEUED(task))
 670                         continue;
 671                 rpc_clear_running(task);
 672                 if (RPC_IS_ASYNC(task)) {
 673                         /* Careful! we may have raced... */
 674                         if (RPC_IS_QUEUED(task))
 675                                 return;
 676                         if (rpc_test_and_set_running(task))
 677                                 return;
 678                         continue;
 679                 }
 680
 681                 /* sync task: sleep here */
 682                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 683                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 684                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 685                                 TASK_KILLABLE);
 686                 if (status == -ERESTARTSYS) {
 687                         /*
 688                          * When a sync task receives a signal, it exits with
 689                          * -ERESTARTSYS. In order to catch any callbacks that
 690                          * clean up after sleeping on some queue, we don't
 691                          * break the loop here, but go around once more.
 692                          */
 693                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 694                         task->tk_flags |= RPC_TASK_KILLED;
 695                         rpc_exit(task, -ERESTARTSYS);
 696                         rpc_wake_up_task(task);
 697                 }
 698                 rpc_set_running(task);
 699                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 700         }
 701
 702         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 703                         task->tk_status);
 704         /* Release all resources associated with the task */
 705         rpc_release_task(task);
 706 }
 707
 708 /*
 709  * User-visible entry point to the scheduler.
 710  *
 711  * This may be called recursively if e.g. an async NFS task updates
 712  * the attributes and finds that dirty pages must be flushed.
 713  * NOTE: Upon exit of this function the task is guaranteed to be
 714  *       released. In particular note that tk_release() will have
 715  *       been called, so your task memory may have been freed.
 716  */
 717 void rpc_execute(struct rpc_task *task)
 718 {
 719         rpc_set_active(task);
 720         rpc_set_running(task);
 721         __rpc_execute(task);
 722 }
 723
 724 static void rpc_async_schedule(struct work_struct *work)
 725 {
 726         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 727 }
 728
 729 struct rpc_buffer {
 730         size_t  len;
 731         char    data[];
 732 };
 733
 734 /**
 735  * rpc_malloc - allocate an RPC buffer
 736  * @task: RPC task that will use this buffer
 737  * @size: requested byte size
 738  *
 739  * To prevent rpciod from hanging, this allocator never sleeps,
 740  * returning NULL if the request cannot be serviced immediately.
 741  * The caller can arrange to sleep in a way that is safe for rpciod.
 742  *
 743  * Most requests are 'small' (under 2KiB) and can be serviced from a
 744  * mempool, ensuring that NFS reads and writes can always proceed,
 745  * and that there is good locality of reference for these buffers.
 746  *
 747  * In order to avoid memory starvation triggering more writebacks of
 748  * NFS requests, we avoid using GFP_KERNEL.
 749  */
 750 void *rpc_malloc(struct rpc_task *task, size_t size)
 751 {
 752         struct rpc_buffer *buf;
 753         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 754
 755         size += sizeof(struct rpc_buffer);
 756         if (size <= RPC_BUFFER_MAXSIZE)
 757                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 758         else
 759                 buf = kmalloc(size, gfp);
 760
 761         if (!buf)
 762                 return NULL;
 763
 764         buf->len = size;
 765         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 766                         task->tk_pid, size, buf);
 767         return &buf->data;
 768 }
 769 EXPORT_SYMBOL_GPL(rpc_malloc);
 770
 771 /**
 772  * rpc_free - free buffer allocated via rpc_malloc
 773  * @buffer: buffer to free
 774  *
 775  */
 776 void rpc_free(void *buffer)
 777 {
 778         size_t size;
 779         struct rpc_buffer *buf;
 780
 781         if (!buffer)
 782                 return;
 783
 784         buf = container_of(buffer, struct rpc_buffer, data);
 785         size = buf->len;
 786
 787         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 788                         size, buf);
 789
 790         if (size <= RPC_BUFFER_MAXSIZE)
 791                 mempool_free(buf, rpc_buffer_mempool);
 792         else
 793                 kfree(buf);
 794 }
 795 EXPORT_SYMBOL_GPL(rpc_free);
 796
 797 /*
 798  * Creation and deletion of RPC task structures
 799  */
 800 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 801 {
 802         memset(task, 0, sizeof(*task));
 803         setup_timer(&task->tk_timer, rpc_run_timer, (unsigned long)task);
 804         atomic_set(&task->tk_count, 1);
 805         task->tk_flags  = task_setup_data->flags;
 806         task->tk_ops = task_setup_data->callback_ops;
 807         task->tk_calldata = task_setup_data->callback_data;
 808         INIT_LIST_HEAD(&task->tk_task);
 809
 810         /* Initialize retry counters */
 811         task->tk_garb_retry = 2;
 812         task->tk_cred_retry = 2;
 813
 814         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 815         task->tk_owner = current->tgid;
 816
 817         /* Initialize workqueue for async tasks */
 818         task->tk_workqueue = task_setup_data->workqueue;
 819
 820         task->tk_client = task_setup_data->rpc_client;
 821         if (task->tk_client != NULL) {
 822                 kref_get(&task->tk_client->cl_kref);
 823                 if (task->tk_client->cl_softrtry)
 824                         task->tk_flags |= RPC_TASK_SOFT;
 825         }
 826
 827         if (task->tk_ops->rpc_call_prepare != NULL)
 828                 task->tk_action = rpc_prepare_task;
 829
 830         if (task_setup_data->rpc_message != NULL) {
 831                 memcpy(&task->tk_msg, task_setup_data->rpc_message, sizeof(task->tk_msg));
 832                 /* Bind the user cred */
 833                 if (task->tk_msg.rpc_cred != NULL)
 834                         rpcauth_holdcred(task);
 835                 else
 836                         rpcauth_bindcred(task);
 837                 if (task->tk_action == NULL)
 838                         rpc_call_start(task);
 839         }
 840
 841         /* starting timestamp */
 842         task->tk_start = jiffies;
 843
 844         dprintk("RPC:       new task initialized, procpid %u\n",
 845                                 task_pid_nr(current));
 846 }
 847
 848 static struct rpc_task *
 849 rpc_alloc_task(void)
 850 {
 851         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 852 }
 853
 854 static void rpc_free_task_rcu(struct rcu_head *rcu)
 855 {
 856         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 857         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 858         mempool_free(task, rpc_task_mempool);
 859 }
 860
 861 /*
 862  * Create a new task for the specified client.
 863  */
 864 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 865 {
 866         struct rpc_task *task = setup_data->task;
 867         unsigned short flags = 0;
 868
 869         if (task == NULL) {
 870                 task = rpc_alloc_task();
 871                 if (task == NULL)
 872                         goto out;
 873                 flags = RPC_TASK_DYNAMIC;
 874         }
 875
 876         rpc_init_task(task, setup_data);
 877
 878         task->tk_flags |= flags;
 879         dprintk("RPC:       allocated task %p\n", task);
 880 out:
 881         return task;
 882 }
 883
 884 static void rpc_free_task(struct rpc_task *task)
 885 {
 886         const struct rpc_call_ops *tk_ops = task->tk_ops;
 887         void *calldata = task->tk_calldata;
 888
 889         if (task->tk_flags & RPC_TASK_DYNAMIC)
 890                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task_rcu);
 891         rpc_release_calldata(tk_ops, calldata);
 892 }
 893
 894 static void rpc_async_release(struct work_struct *work)
 895 {
 896         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 897 }
 898
 899 void rpc_put_task(struct rpc_task *task)
 900 {
 901         if (!atomic_dec_and_test(&task->tk_count))
 902                 return;
 903         /* Release resources */
 904         if (task->tk_rqstp)
 905                 xprt_release(task);
 906         if (task->tk_msg.rpc_cred)
 907                 rpcauth_unbindcred(task);
 908         if (task->tk_client) {
 909                 rpc_release_client(task->tk_client);
 910                 task->tk_client = NULL;
 911         }
 912         if (task->tk_workqueue != NULL) {
 913                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 914                 queue_work(task->tk_workqueue, &task->u.tk_work);
 915         } else
 916                 rpc_free_task(task);
 917 }
 918 EXPORT_SYMBOL_GPL(rpc_put_task);
 919
 920 static void rpc_release_task(struct rpc_task *task)
 921 {
 922 #ifdef RPC_DEBUG
 923         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 924 #endif
 925         dprintk("RPC: %5u release task\n", task->tk_pid);
 926
 927         if (!list_empty(&task->tk_task)) {
 928                 struct rpc_clnt *clnt = task->tk_client;
 929                 /* Remove from client task list */
 930                 spin_lock(&clnt->cl_lock);
 931                 list_del(&task->tk_task);
 932                 spin_unlock(&clnt->cl_lock);
 933         }
 934         BUG_ON (RPC_IS_QUEUED(task));
 935
 936         /* Synchronously delete any running timer */
 937         rpc_delete_timer(task);
 938
 939 #ifdef RPC_DEBUG
 940         task->tk_magic = 0;
 941 #endif
 942         /* Wake up anyone who is waiting for task completion */
 943         rpc_mark_complete_task(task);
 944
 945         rpc_put_task(task);
 946 }
 947
 948 /*
 949  * Kill all tasks for the given client.
 950  * XXX: kill their descendants as well?
 951  */
 952 void rpc_killall_tasks(struct rpc_clnt *clnt)
 953 {
 954         struct rpc_task *rovr;
 955
 956
 957         if (list_empty(&clnt->cl_tasks))
 958                 return;
 959         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 960         /*
 961          * Spin lock all_tasks to prevent changes...
 962          */
 963         spin_lock(&clnt->cl_lock);
 964         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 965                 if (! RPC_IS_ACTIVATED(rovr))
 966                         continue;
 967                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 968                         rovr->tk_flags |= RPC_TASK_KILLED;
 969                         rpc_exit(rovr, -EIO);
 970                         rpc_wake_up_task(rovr);
 971                 }
 972         }
 973         spin_unlock(&clnt->cl_lock);
 974 }
 975 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 976
 977 int rpciod_up(void)
 978 {
 979         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 980 }
 981
 982 void rpciod_down(void)
 983 {
 984         module_put(THIS_MODULE);
 985 }
 986
 987 /*
 988  * Start up the rpciod workqueue.
 989  */
 990 static int rpciod_start(void)
 991 {
 992         struct workqueue_struct *wq;
 993
 994         /*
 995          * Create the rpciod thread and wait for it to start.
 996          */
 997         dprintk("RPC:       creating workqueue rpciod\n");
 998         wq = create_workqueue("rpciod");
 999         rpciod_workqueue = wq;
1000         return rpciod_workqueue != NULL;
1001 }
1002
1003 static void rpciod_stop(void)
1004 {
1005         struct workqueue_struct *wq = NULL;
1006
1007         if (rpciod_workqueue == NULL)
1008                 return;
1009         dprintk("RPC:       destroying workqueue rpciod\n");
1010
1011         wq = rpciod_workqueue;
1012         rpciod_workqueue = NULL;
1013         destroy_workqueue(wq);
1014 }
1015
1016 void
1017 rpc_destroy_mempool(void)
1018 {
1019         rpciod_stop();
1020         if (rpc_buffer_mempool)
1021                 mempool_destroy(rpc_buffer_mempool);
1022         if (rpc_task_mempool)
1023                 mempool_destroy(rpc_task_mempool);
1024         if (rpc_task_slabp)
1025                 kmem_cache_destroy(rpc_task_slabp);
1026         if (rpc_buffer_slabp)
1027                 kmem_cache_destroy(rpc_buffer_slabp);
1028         rpc_destroy_wait_queue(&delay_queue);
1029 }
1030
1031 int
1032 rpc_init_mempool(void)
1033 {
1034         /*
1035          * The following is not strictly a mempool initialisation,
1036          * but there is no harm in doing it here
1037          */
1038         rpc_init_wait_queue(&delay_queue, "delayq");
1039         if (!rpciod_start())
1040                 goto err_nomem;
1041
1042         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1043                                              sizeof(struct rpc_task),
1044                                              0, SLAB_HWCACHE_ALIGN,
1045                                              NULL);
1046         if (!rpc_task_slabp)
1047                 goto err_nomem;
1048         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1049                                              RPC_BUFFER_MAXSIZE,
1050                                              0, SLAB_HWCACHE_ALIGN,
1051                                              NULL);
1052         if (!rpc_buffer_slabp)
1053                 goto err_nomem;
1054         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1055                                                     rpc_task_slabp);
1056         if (!rpc_task_mempool)
1057                 goto err_nomem;
1058         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1059                                                       rpc_buffer_slabp);
1060         if (!rpc_buffer_mempool)
1061                 goto err_nomem;
1062         return 0;
1063 err_nomem:
1064         rpc_destroy_mempool();
1065         return -ENOMEM;
1066 }