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 static int                      rpc_task_id;
  29 #endif
  30
  31 /*
  32  * RPC slabs and memory pools
  33  */
  34 #define RPC_BUFFER_MAXSIZE      (2048)
  35 #define RPC_BUFFER_POOLSIZE     (8)
  36 #define RPC_TASK_POOLSIZE       (8)
  37 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  38 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  39 static mempool_t        *rpc_task_mempool __read_mostly;
  40 static mempool_t        *rpc_buffer_mempool __read_mostly;
  41
  42 static void                     __rpc_default_timer(struct rpc_task *task);
  43 static void                     rpciod_killall(void);
  44 static void                     rpc_async_schedule(struct work_struct *);
  45
  46 /*
  47  * RPC tasks sit here while waiting for conditions to improve.
  48  */
  49 static RPC_WAITQ(delay_queue, "delayq");
  50
  51 /*
  52  * All RPC tasks are linked into this list
  53  */
  54 static LIST_HEAD(all_tasks);
  55
  56 /*
  57  * rpciod-related stuff
  58  */
  59 static DEFINE_MUTEX(rpciod_mutex);
  60 static unsigned int             rpciod_users;
  61 struct workqueue_struct *rpciod_workqueue;
  62
  63 /*
  64  * Spinlock for other critical sections of code.
  65  */
  66 static DEFINE_SPINLOCK(rpc_sched_lock);
  67
  68 /*
  69  * Disable the timer for a given RPC task. Should be called with
  70  * queue->lock and bh_disabled in order to avoid races within
  71  * rpc_run_timer().
  72  */
  73 static inline void
  74 __rpc_disable_timer(struct rpc_task *task)
  75 {
  76         dprintk("RPC: %4d disabling timer\n", task->tk_pid);
  77         task->tk_timeout_fn = NULL;
  78         task->tk_timeout = 0;
  79 }
  80
  81 /*
  82  * Run a timeout function.
  83  * We use the callback in order to allow __rpc_wake_up_task()
  84  * and friends to disable the timer synchronously on SMP systems
  85  * without calling del_timer_sync(). The latter could cause a
  86  * deadlock if called while we're holding spinlocks...
  87  */
  88 static void rpc_run_timer(struct rpc_task *task)
  89 {
  90         void (*callback)(struct rpc_task *);
  91
  92         callback = task->tk_timeout_fn;
  93         task->tk_timeout_fn = NULL;
  94         if (callback && RPC_IS_QUEUED(task)) {
  95                 dprintk("RPC: %4d running timer\n", task->tk_pid);
  96                 callback(task);
  97         }
  98         smp_mb__before_clear_bit();
  99         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 100         smp_mb__after_clear_bit();
 101 }
 102
 103 /*
 104  * Set up a timer for the current task.
 105  */
 106 static inline void
 107 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
 108 {
 109         if (!task->tk_timeout)
 110                 return;
 111
 112         dprintk("RPC: %4d setting alarm for %lu ms\n",
 113                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 114
 115         if (timer)
 116                 task->tk_timeout_fn = timer;
 117         else
 118                 task->tk_timeout_fn = __rpc_default_timer;
 119         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 120         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 121 }
 122
 123 /*
 124  * Delete any timer for the current task. Because we use del_timer_sync(),
 125  * this function should never be called while holding queue->lock.
 126  */
 127 static void
 128 rpc_delete_timer(struct rpc_task *task)
 129 {
 130         if (RPC_IS_QUEUED(task))
 131                 return;
 132         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 133                 del_singleshot_timer_sync(&task->tk_timer);
 134                 dprintk("RPC: %4d deleting timer\n", task->tk_pid);
 135         }
 136 }
 137
 138 /*
 139  * Add new request to a priority queue.
 140  */
 141 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 142 {
 143         struct list_head *q;
 144         struct rpc_task *t;
 145
 146         INIT_LIST_HEAD(&task->u.tk_wait.links);
 147         q = &queue->tasks[task->tk_priority];
 148         if (unlikely(task->tk_priority > queue->maxpriority))
 149                 q = &queue->tasks[queue->maxpriority];
 150         list_for_each_entry(t, q, u.tk_wait.list) {
 151                 if (t->tk_cookie == task->tk_cookie) {
 152                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 153                         return;
 154                 }
 155         }
 156         list_add_tail(&task->u.tk_wait.list, q);
 157 }
 158
 159 /*
 160  * Add new request to wait queue.
 161  *
 162  * Swapper tasks always get inserted at the head of the queue.
 163  * This should avoid many nasty memory deadlocks and hopefully
 164  * improve overall performance.
 165  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 166  */
 167 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 168 {
 169         BUG_ON (RPC_IS_QUEUED(task));
 170
 171         if (RPC_IS_PRIORITY(queue))
 172                 __rpc_add_wait_queue_priority(queue, task);
 173         else if (RPC_IS_SWAPPER(task))
 174                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 175         else
 176                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 177         task->u.tk_wait.rpc_waitq = queue;
 178         queue->qlen++;
 179         rpc_set_queued(task);
 180
 181         dprintk("RPC: %4d added to queue %p \"%s\"\n",
 182                                 task->tk_pid, queue, rpc_qname(queue));
 183 }
 184
 185 /*
 186  * Remove request from a priority queue.
 187  */
 188 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 189 {
 190         struct rpc_task *t;
 191
 192         if (!list_empty(&task->u.tk_wait.links)) {
 193                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 194                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 195                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 196         }
 197         list_del(&task->u.tk_wait.list);
 198 }
 199
 200 /*
 201  * Remove request from queue.
 202  * Note: must be called with spin lock held.
 203  */
 204 static void __rpc_remove_wait_queue(struct rpc_task *task)
 205 {
 206         struct rpc_wait_queue *queue;
 207         queue = task->u.tk_wait.rpc_waitq;
 208
 209         if (RPC_IS_PRIORITY(queue))
 210                 __rpc_remove_wait_queue_priority(task);
 211         else
 212                 list_del(&task->u.tk_wait.list);
 213         queue->qlen--;
 214         dprintk("RPC: %4d removed from queue %p \"%s\"\n",
 215                                 task->tk_pid, queue, rpc_qname(queue));
 216 }
 217
 218 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 219 {
 220         queue->priority = priority;
 221         queue->count = 1 << (priority * 2);
 222 }
 223
 224 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
 225 {
 226         queue->cookie = cookie;
 227         queue->nr = RPC_BATCH_COUNT;
 228 }
 229
 230 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 231 {
 232         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 233         rpc_set_waitqueue_cookie(queue, 0);
 234 }
 235
 236 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
 237 {
 238         int i;
 239
 240         spin_lock_init(&queue->lock);
 241         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 242                 INIT_LIST_HEAD(&queue->tasks[i]);
 243         queue->maxpriority = maxprio;
 244         rpc_reset_waitqueue_priority(queue);
 245 #ifdef RPC_DEBUG
 246         queue->name = qname;
 247 #endif
 248 }
 249
 250 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 251 {
 252         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
 253 }
 254
 255 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 256 {
 257         __rpc_init_priority_wait_queue(queue, qname, 0);
 258 }
 259 EXPORT_SYMBOL(rpc_init_wait_queue);
 260
 261 static int rpc_wait_bit_interruptible(void *word)
 262 {
 263         if (signal_pending(current))
 264                 return -ERESTARTSYS;
 265         schedule();
 266         return 0;
 267 }
 268
 269 static void rpc_set_active(struct rpc_task *task)
 270 {
 271         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 272                 return;
 273         spin_lock(&rpc_sched_lock);
 274 #ifdef RPC_DEBUG
 275         task->tk_magic = RPC_TASK_MAGIC_ID;
 276         task->tk_pid = rpc_task_id++;
 277 #endif
 278         /* Add to global list of all tasks */
 279         list_add_tail(&task->tk_task, &all_tasks);
 280         spin_unlock(&rpc_sched_lock);
 281 }
 282
 283 /*
 284  * Mark an RPC call as having completed by clearing the 'active' bit
 285  */
 286 static void rpc_mark_complete_task(struct rpc_task *task)
 287 {
 288         smp_mb__before_clear_bit();
 289         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 290         smp_mb__after_clear_bit();
 291         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 292 }
 293
 294 /*
 295  * Allow callers to wait for completion of an RPC call
 296  */
 297 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 298 {
 299         if (action == NULL)
 300                 action = rpc_wait_bit_interruptible;
 301         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 302                         action, TASK_INTERRUPTIBLE);
 303 }
 304 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
 305
 306 /*
 307  * Make an RPC task runnable.
 308  *
 309  * Note: If the task is ASYNC, this must be called with
 310  * the spinlock held to protect the wait queue operation.
 311  */
 312 static void rpc_make_runnable(struct rpc_task *task)
 313 {
 314         BUG_ON(task->tk_timeout_fn);
 315         rpc_clear_queued(task);
 316         if (rpc_test_and_set_running(task))
 317                 return;
 318         /* We might have raced */
 319         if (RPC_IS_QUEUED(task)) {
 320                 rpc_clear_running(task);
 321                 return;
 322         }
 323         if (RPC_IS_ASYNC(task)) {
 324                 int status;
 325
 326                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 327                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 328                 if (status < 0) {
 329                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 330                         task->tk_status = status;
 331                         return;
 332                 }
 333         } else
 334                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 335 }
 336
 337 /*
 338  * Prepare for sleeping on a wait queue.
 339  * By always appending tasks to the list we ensure FIFO behavior.
 340  * NB: An RPC task will only receive interrupt-driven events as long
 341  * as it's on a wait queue.
 342  */
 343 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 344                         rpc_action action, rpc_action timer)
 345 {
 346         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
 347                                 rpc_qname(q), jiffies);
 348
 349         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 350                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 351                 return;
 352         }
 353
 354         __rpc_add_wait_queue(q, task);
 355
 356         BUG_ON(task->tk_callback != NULL);
 357         task->tk_callback = action;
 358         __rpc_add_timer(task, timer);
 359 }
 360
 361 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 362                                 rpc_action action, rpc_action timer)
 363 {
 364         /* Mark the task as being activated if so needed */
 365         rpc_set_active(task);
 366
 367         /*
 368          * Protect the queue operations.
 369          */
 370         spin_lock_bh(&q->lock);
 371         __rpc_sleep_on(q, task, action, timer);
 372         spin_unlock_bh(&q->lock);
 373 }
 374
 375 /**
 376  * __rpc_do_wake_up_task - wake up a single rpc_task
 377  * @task: task to be woken up
 378  *
 379  * Caller must hold queue->lock, and have cleared the task queued flag.
 380  */
 381 static void __rpc_do_wake_up_task(struct rpc_task *task)
 382 {
 383         dprintk("RPC: %4d __rpc_wake_up_task (now %ld)\n", task->tk_pid, jiffies);
 384
 385 #ifdef RPC_DEBUG
 386         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 387 #endif
 388         /* Has the task been executed yet? If not, we cannot wake it up! */
 389         if (!RPC_IS_ACTIVATED(task)) {
 390                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 391                 return;
 392         }
 393
 394         __rpc_disable_timer(task);
 395         __rpc_remove_wait_queue(task);
 396
 397         rpc_make_runnable(task);
 398
 399         dprintk("RPC:      __rpc_wake_up_task done\n");
 400 }
 401
 402 /*
 403  * Wake up the specified task
 404  */
 405 static void __rpc_wake_up_task(struct rpc_task *task)
 406 {
 407         if (rpc_start_wakeup(task)) {
 408                 if (RPC_IS_QUEUED(task))
 409                         __rpc_do_wake_up_task(task);
 410                 rpc_finish_wakeup(task);
 411         }
 412 }
 413
 414 /*
 415  * Default timeout handler if none specified by user
 416  */
 417 static void
 418 __rpc_default_timer(struct rpc_task *task)
 419 {
 420         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
 421         task->tk_status = -ETIMEDOUT;
 422         rpc_wake_up_task(task);
 423 }
 424
 425 /*
 426  * Wake up the specified task
 427  */
 428 void rpc_wake_up_task(struct rpc_task *task)
 429 {
 430         rcu_read_lock_bh();
 431         if (rpc_start_wakeup(task)) {
 432                 if (RPC_IS_QUEUED(task)) {
 433                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 434
 435                         /* Note: we're already in a bh-safe context */
 436                         spin_lock(&queue->lock);
 437                         __rpc_do_wake_up_task(task);
 438                         spin_unlock(&queue->lock);
 439                 }
 440                 rpc_finish_wakeup(task);
 441         }
 442         rcu_read_unlock_bh();
 443 }
 444
 445 /*
 446  * Wake up the next task on a priority queue.
 447  */
 448 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 449 {
 450         struct list_head *q;
 451         struct rpc_task *task;
 452
 453         /*
 454          * Service a batch of tasks from a single cookie.
 455          */
 456         q = &queue->tasks[queue->priority];
 457         if (!list_empty(q)) {
 458                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 459                 if (queue->cookie == task->tk_cookie) {
 460                         if (--queue->nr)
 461                                 goto out;
 462                         list_move_tail(&task->u.tk_wait.list, q);
 463                 }
 464                 /*
 465                  * Check if we need to switch queues.
 466                  */
 467                 if (--queue->count)
 468                         goto new_cookie;
 469         }
 470
 471         /*
 472          * Service the next queue.
 473          */
 474         do {
 475                 if (q == &queue->tasks[0])
 476                         q = &queue->tasks[queue->maxpriority];
 477                 else
 478                         q = q - 1;
 479                 if (!list_empty(q)) {
 480                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 481                         goto new_queue;
 482                 }
 483         } while (q != &queue->tasks[queue->priority]);
 484
 485         rpc_reset_waitqueue_priority(queue);
 486         return NULL;
 487
 488 new_queue:
 489         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 490 new_cookie:
 491         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 492 out:
 493         __rpc_wake_up_task(task);
 494         return task;
 495 }
 496
 497 /*
 498  * Wake up the next task on the wait queue.
 499  */
 500 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 501 {
 502         struct rpc_task *task = NULL;
 503
 504         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
 505         rcu_read_lock_bh();
 506         spin_lock(&queue->lock);
 507         if (RPC_IS_PRIORITY(queue))
 508                 task = __rpc_wake_up_next_priority(queue);
 509         else {
 510                 task_for_first(task, &queue->tasks[0])
 511                         __rpc_wake_up_task(task);
 512         }
 513         spin_unlock(&queue->lock);
 514         rcu_read_unlock_bh();
 515
 516         return task;
 517 }
 518
 519 /**
 520  * rpc_wake_up - wake up all rpc_tasks
 521  * @queue: rpc_wait_queue on which the tasks are sleeping
 522  *
 523  * Grabs queue->lock
 524  */
 525 void rpc_wake_up(struct rpc_wait_queue *queue)
 526 {
 527         struct rpc_task *task, *next;
 528         struct list_head *head;
 529
 530         rcu_read_lock_bh();
 531         spin_lock(&queue->lock);
 532         head = &queue->tasks[queue->maxpriority];
 533         for (;;) {
 534                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 535                         __rpc_wake_up_task(task);
 536                 if (head == &queue->tasks[0])
 537                         break;
 538                 head--;
 539         }
 540         spin_unlock(&queue->lock);
 541         rcu_read_unlock_bh();
 542 }
 543
 544 /**
 545  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 546  * @queue: rpc_wait_queue on which the tasks are sleeping
 547  * @status: status value to set
 548  *
 549  * Grabs queue->lock
 550  */
 551 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 552 {
 553         struct rpc_task *task, *next;
 554         struct list_head *head;
 555
 556         rcu_read_lock_bh();
 557         spin_lock(&queue->lock);
 558         head = &queue->tasks[queue->maxpriority];
 559         for (;;) {
 560                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 561                         task->tk_status = status;
 562                         __rpc_wake_up_task(task);
 563                 }
 564                 if (head == &queue->tasks[0])
 565                         break;
 566                 head--;
 567         }
 568         spin_unlock(&queue->lock);
 569         rcu_read_unlock_bh();
 570 }
 571
 572 static void __rpc_atrun(struct rpc_task *task)
 573 {
 574         rpc_wake_up_task(task);
 575 }
 576
 577 /*
 578  * Run a task at a later time
 579  */
 580 void rpc_delay(struct rpc_task *task, unsigned long delay)
 581 {
 582         task->tk_timeout = delay;
 583         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 584 }
 585
 586 /*
 587  * Helper to call task->tk_ops->rpc_call_prepare
 588  */
 589 static void rpc_prepare_task(struct rpc_task *task)
 590 {
 591         lock_kernel();
 592         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 593         unlock_kernel();
 594 }
 595
 596 /*
 597  * Helper that calls task->tk_ops->rpc_call_done if it exists
 598  */
 599 void rpc_exit_task(struct rpc_task *task)
 600 {
 601         task->tk_action = NULL;
 602         if (task->tk_ops->rpc_call_done != NULL) {
 603                 lock_kernel();
 604                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 605                 unlock_kernel();
 606                 if (task->tk_action != NULL) {
 607                         WARN_ON(RPC_ASSASSINATED(task));
 608                         /* Always release the RPC slot and buffer memory */
 609                         xprt_release(task);
 610                 }
 611         }
 612 }
 613 EXPORT_SYMBOL(rpc_exit_task);
 614
 615 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 616 {
 617         if (ops->rpc_release != NULL) {
 618                 lock_kernel();
 619                 ops->rpc_release(calldata);
 620                 unlock_kernel();
 621         }
 622 }
 623
 624 /*
 625  * This is the RPC `scheduler' (or rather, the finite state machine).
 626  */
 627 static int __rpc_execute(struct rpc_task *task)
 628 {
 629         int             status = 0;
 630
 631         dprintk("RPC: %4d rpc_execute flgs %x\n",
 632                                 task->tk_pid, task->tk_flags);
 633
 634         BUG_ON(RPC_IS_QUEUED(task));
 635
 636         for (;;) {
 637                 /*
 638                  * Garbage collection of pending timers...
 639                  */
 640                 rpc_delete_timer(task);
 641
 642                 /*
 643                  * Execute any pending callback.
 644                  */
 645                 if (RPC_DO_CALLBACK(task)) {
 646                         /* Define a callback save pointer */
 647                         void (*save_callback)(struct rpc_task *);
 648
 649                         /*
 650                          * If a callback exists, save it, reset it,
 651                          * call it.
 652                          * The save is needed to stop from resetting
 653                          * another callback set within the callback handler
 654                          * - Dave
 655                          */
 656                         save_callback=task->tk_callback;
 657                         task->tk_callback=NULL;
 658                         save_callback(task);
 659                 }
 660
 661                 /*
 662                  * Perform the next FSM step.
 663                  * tk_action may be NULL when the task has been killed
 664                  * by someone else.
 665                  */
 666                 if (!RPC_IS_QUEUED(task)) {
 667                         if (task->tk_action == NULL)
 668                                 break;
 669                         task->tk_action(task);
 670                 }
 671
 672                 /*
 673                  * Lockless check for whether task is sleeping or not.
 674                  */
 675                 if (!RPC_IS_QUEUED(task))
 676                         continue;
 677                 rpc_clear_running(task);
 678                 if (RPC_IS_ASYNC(task)) {
 679                         /* Careful! we may have raced... */
 680                         if (RPC_IS_QUEUED(task))
 681                                 return 0;
 682                         if (rpc_test_and_set_running(task))
 683                                 return 0;
 684                         continue;
 685                 }
 686
 687                 /* sync task: sleep here */
 688                 dprintk("RPC: %4d sync task going to sleep\n", task->tk_pid);
 689                 /* Note: Caller should be using rpc_clnt_sigmask() */
 690                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 691                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
 692                                 TASK_INTERRUPTIBLE);
 693                 if (status == -ERESTARTSYS) {
 694                         /*
 695                          * When a sync task receives a signal, it exits with
 696                          * -ERESTARTSYS. In order to catch any callbacks that
 697                          * clean up after sleeping on some queue, we don't
 698                          * break the loop here, but go around once more.
 699                          */
 700                         dprintk("RPC: %4d got signal\n", task->tk_pid);
 701                         task->tk_flags |= RPC_TASK_KILLED;
 702                         rpc_exit(task, -ERESTARTSYS);
 703                         rpc_wake_up_task(task);
 704                 }
 705                 rpc_set_running(task);
 706                 dprintk("RPC: %4d sync task resuming\n", task->tk_pid);
 707         }
 708
 709         dprintk("RPC: %4d, return %d, status %d\n", task->tk_pid, status, task->tk_status);
 710         /* Release all resources associated with the task */
 711         rpc_release_task(task);
 712         return status;
 713 }
 714
 715 /*
 716  * User-visible entry point to the scheduler.
 717  *
 718  * This may be called recursively if e.g. an async NFS task updates
 719  * the attributes and finds that dirty pages must be flushed.
 720  * NOTE: Upon exit of this function the task is guaranteed to be
 721  *       released. In particular note that tk_release() will have
 722  *       been called, so your task memory may have been freed.
 723  */
 724 int
 725 rpc_execute(struct rpc_task *task)
 726 {
 727         rpc_set_active(task);
 728         rpc_set_running(task);
 729         return __rpc_execute(task);
 730 }
 731
 732 static void rpc_async_schedule(struct work_struct *work)
 733 {
 734         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 735 }
 736
 737 /**
 738  * rpc_malloc - allocate an RPC buffer
 739  * @task: RPC task that will use this buffer
 740  * @size: requested byte size
 741  *
 742  * We try to ensure that some NFS reads and writes can always proceed
 743  * by using a mempool when allocating 'small' buffers.
 744  * In order to avoid memory starvation triggering more writebacks of
 745  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
 746  */
 747 void * rpc_malloc(struct rpc_task *task, size_t size)
 748 {
 749         struct rpc_rqst *req = task->tk_rqstp;
 750         gfp_t   gfp;
 751
 752         if (task->tk_flags & RPC_TASK_SWAPPER)
 753                 gfp = GFP_ATOMIC;
 754         else
 755                 gfp = GFP_NOFS;
 756
 757         if (size > RPC_BUFFER_MAXSIZE) {
 758                 req->rq_buffer = kmalloc(size, gfp);
 759                 if (req->rq_buffer)
 760                         req->rq_bufsize = size;
 761         } else {
 762                 req->rq_buffer = mempool_alloc(rpc_buffer_mempool, gfp);
 763                 if (req->rq_buffer)
 764                         req->rq_bufsize = RPC_BUFFER_MAXSIZE;
 765         }
 766         return req->rq_buffer;
 767 }
 768
 769 /**
 770  * rpc_free - free buffer allocated via rpc_malloc
 771  * @task: RPC task with a buffer to be freed
 772  *
 773  */
 774 void rpc_free(struct rpc_task *task)
 775 {
 776         struct rpc_rqst *req = task->tk_rqstp;
 777
 778         if (req->rq_buffer) {
 779                 if (req->rq_bufsize == RPC_BUFFER_MAXSIZE)
 780                         mempool_free(req->rq_buffer, rpc_buffer_mempool);
 781                 else
 782                         kfree(req->rq_buffer);
 783                 req->rq_buffer = NULL;
 784                 req->rq_bufsize = 0;
 785         }
 786 }
 787
 788 /*
 789  * Creation and deletion of RPC task structures
 790  */
 791 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 792 {
 793         memset(task, 0, sizeof(*task));
 794         init_timer(&task->tk_timer);
 795         task->tk_timer.data     = (unsigned long) task;
 796         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 797         atomic_set(&task->tk_count, 1);
 798         task->tk_client = clnt;
 799         task->tk_flags  = flags;
 800         task->tk_ops = tk_ops;
 801         if (tk_ops->rpc_call_prepare != NULL)
 802                 task->tk_action = rpc_prepare_task;
 803         task->tk_calldata = calldata;
 804
 805         /* Initialize retry counters */
 806         task->tk_garb_retry = 2;
 807         task->tk_cred_retry = 2;
 808
 809         task->tk_priority = RPC_PRIORITY_NORMAL;
 810         task->tk_cookie = (unsigned long)current;
 811
 812         /* Initialize workqueue for async tasks */
 813         task->tk_workqueue = rpciod_workqueue;
 814
 815         if (clnt) {
 816                 atomic_inc(&clnt->cl_users);
 817                 if (clnt->cl_softrtry)
 818                         task->tk_flags |= RPC_TASK_SOFT;
 819                 if (!clnt->cl_intr)
 820                         task->tk_flags |= RPC_TASK_NOINTR;
 821         }
 822
 823         BUG_ON(task->tk_ops == NULL);
 824
 825         /* starting timestamp */
 826         task->tk_start = jiffies;
 827
 828         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
 829                                 current->pid);
 830 }
 831
 832 static struct rpc_task *
 833 rpc_alloc_task(void)
 834 {
 835         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 836 }
 837
 838 static void rpc_free_task(struct rcu_head *rcu)
 839 {
 840         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 841         dprintk("RPC: %4d freeing task\n", task->tk_pid);
 842         mempool_free(task, rpc_task_mempool);
 843 }
 844
 845 /*
 846  * Create a new task for the specified client.  We have to
 847  * clean up after an allocation failure, as the client may
 848  * have specified "oneshot".
 849  */
 850 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 851 {
 852         struct rpc_task *task;
 853
 854         task = rpc_alloc_task();
 855         if (!task)
 856                 goto cleanup;
 857
 858         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 859
 860         dprintk("RPC: %4d allocated task\n", task->tk_pid);
 861         task->tk_flags |= RPC_TASK_DYNAMIC;
 862 out:
 863         return task;
 864
 865 cleanup:
 866         /* Check whether to release the client */
 867         if (clnt) {
 868                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 869                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 870                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 871                 rpc_release_client(clnt);
 872         }
 873         goto out;
 874 }
 875
 876
 877 void rpc_put_task(struct rpc_task *task)
 878 {
 879         const struct rpc_call_ops *tk_ops = task->tk_ops;
 880         void *calldata = task->tk_calldata;
 881
 882         if (!atomic_dec_and_test(&task->tk_count))
 883                 return;
 884         /* Release resources */
 885         if (task->tk_rqstp)
 886                 xprt_release(task);
 887         if (task->tk_msg.rpc_cred)
 888                 rpcauth_unbindcred(task);
 889         if (task->tk_client) {
 890                 rpc_release_client(task->tk_client);
 891                 task->tk_client = NULL;
 892         }
 893         if (task->tk_flags & RPC_TASK_DYNAMIC)
 894                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task);
 895         rpc_release_calldata(tk_ops, calldata);
 896 }
 897 EXPORT_SYMBOL(rpc_put_task);
 898
 899 void rpc_release_task(struct rpc_task *task)
 900 {
 901 #ifdef RPC_DEBUG
 902         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 903 #endif
 904         dprintk("RPC: %4d release task\n", task->tk_pid);
 905
 906         /* Remove from global task list */
 907         spin_lock(&rpc_sched_lock);
 908         list_del(&task->tk_task);
 909         spin_unlock(&rpc_sched_lock);
 910
 911         BUG_ON (RPC_IS_QUEUED(task));
 912
 913         /* Synchronously delete any running timer */
 914         rpc_delete_timer(task);
 915
 916 #ifdef RPC_DEBUG
 917         task->tk_magic = 0;
 918 #endif
 919         /* Wake up anyone who is waiting for task completion */
 920         rpc_mark_complete_task(task);
 921
 922         rpc_put_task(task);
 923 }
 924
 925 /**
 926  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
 927  * @clnt: pointer to RPC client
 928  * @flags: RPC flags
 929  * @ops: RPC call ops
 930  * @data: user call data
 931  */
 932 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
 933                                         const struct rpc_call_ops *ops,
 934                                         void *data)
 935 {
 936         struct rpc_task *task;
 937         task = rpc_new_task(clnt, flags, ops, data);
 938         if (task == NULL) {
 939                 rpc_release_calldata(ops, data);
 940                 return ERR_PTR(-ENOMEM);
 941         }
 942         atomic_inc(&task->tk_count);
 943         rpc_execute(task);
 944         return task;
 945 }
 946 EXPORT_SYMBOL(rpc_run_task);
 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         struct list_head *le;
 956
 957         dprintk("RPC:      killing all tasks for client %p\n", clnt);
 958
 959         /*
 960          * Spin lock all_tasks to prevent changes...
 961          */
 962         spin_lock(&rpc_sched_lock);
 963         alltask_for_each(rovr, le, &all_tasks) {
 964                 if (! RPC_IS_ACTIVATED(rovr))
 965                         continue;
 966                 if (!clnt || rovr->tk_client == clnt) {
 967                         rovr->tk_flags |= RPC_TASK_KILLED;
 968                         rpc_exit(rovr, -EIO);
 969                         rpc_wake_up_task(rovr);
 970                 }
 971         }
 972         spin_unlock(&rpc_sched_lock);
 973 }
 974
 975 static DECLARE_MUTEX_LOCKED(rpciod_running);
 976
 977 static void rpciod_killall(void)
 978 {
 979         unsigned long flags;
 980
 981         while (!list_empty(&all_tasks)) {
 982                 clear_thread_flag(TIF_SIGPENDING);
 983                 rpc_killall_tasks(NULL);
 984                 flush_workqueue(rpciod_workqueue);
 985                 if (!list_empty(&all_tasks)) {
 986                         dprintk("rpciod_killall: waiting for tasks to exit\n");
 987                         yield();
 988                 }
 989         }
 990
 991         spin_lock_irqsave(&current->sighand->siglock, flags);
 992         recalc_sigpending();
 993         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 994 }
 995
 996 /*
 997  * Start up the rpciod process if it's not already running.
 998  */
 999 int
1000 rpciod_up(void)
1001 {
1002         struct workqueue_struct *wq;
1003         int error = 0;
1004
1005         mutex_lock(&rpciod_mutex);
1006         dprintk("rpciod_up: users %d\n", rpciod_users);
1007         rpciod_users++;
1008         if (rpciod_workqueue)
1009                 goto out;
1010         /*
1011          * If there's no pid, we should be the first user.
1012          */
1013         if (rpciod_users > 1)
1014                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
1015         /*
1016          * Create the rpciod thread and wait for it to start.
1017          */
1018         error = -ENOMEM;
1019         wq = create_workqueue("rpciod");
1020         if (wq == NULL) {
1021                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1022                 rpciod_users--;
1023                 goto out;
1024         }
1025         rpciod_workqueue = wq;
1026         error = 0;
1027 out:
1028         mutex_unlock(&rpciod_mutex);
1029         return error;
1030 }
1031
1032 void
1033 rpciod_down(void)
1034 {
1035         mutex_lock(&rpciod_mutex);
1036         dprintk("rpciod_down sema %d\n", rpciod_users);
1037         if (rpciod_users) {
1038                 if (--rpciod_users)
1039                         goto out;
1040         } else
1041                 printk(KERN_WARNING "rpciod_down: no users??\n");
1042
1043         if (!rpciod_workqueue) {
1044                 dprintk("rpciod_down: Nothing to do!\n");
1045                 goto out;
1046         }
1047         rpciod_killall();
1048
1049         destroy_workqueue(rpciod_workqueue);
1050         rpciod_workqueue = NULL;
1051  out:
1052         mutex_unlock(&rpciod_mutex);
1053 }
1054
1055 #ifdef RPC_DEBUG
1056 void rpc_show_tasks(void)
1057 {
1058         struct list_head *le;
1059         struct rpc_task *t;
1060
1061         spin_lock(&rpc_sched_lock);
1062         if (list_empty(&all_tasks)) {
1063                 spin_unlock(&rpc_sched_lock);
1064                 return;
1065         }
1066         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1067                 "-rpcwait -action- ---ops--\n");
1068         alltask_for_each(t, le, &all_tasks) {
1069                 const char *rpc_waitq = "none";
1070
1071                 if (RPC_IS_QUEUED(t))
1072                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1073
1074                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1075                         t->tk_pid,
1076                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1077                         t->tk_flags, t->tk_status,
1078                         t->tk_client,
1079                         (t->tk_client ? t->tk_client->cl_prog : 0),
1080                         t->tk_rqstp, t->tk_timeout,
1081                         rpc_waitq,
1082                         t->tk_action, t->tk_ops);
1083         }
1084         spin_unlock(&rpc_sched_lock);
1085 }
1086 #endif
1087
1088 void
1089 rpc_destroy_mempool(void)
1090 {
1091         if (rpc_buffer_mempool)
1092                 mempool_destroy(rpc_buffer_mempool);
1093         if (rpc_task_mempool)
1094                 mempool_destroy(rpc_task_mempool);
1095         if (rpc_task_slabp)
1096                 kmem_cache_destroy(rpc_task_slabp);
1097         if (rpc_buffer_slabp)
1098                 kmem_cache_destroy(rpc_buffer_slabp);
1099 }
1100
1101 int
1102 rpc_init_mempool(void)
1103 {
1104         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1105                                              sizeof(struct rpc_task),
1106                                              0, SLAB_HWCACHE_ALIGN,
1107                                              NULL, NULL);
1108         if (!rpc_task_slabp)
1109                 goto err_nomem;
1110         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1111                                              RPC_BUFFER_MAXSIZE,
1112                                              0, SLAB_HWCACHE_ALIGN,
1113                                              NULL, NULL);
1114         if (!rpc_buffer_slabp)
1115                 goto err_nomem;
1116         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1117                                                     rpc_task_slabp);
1118         if (!rpc_task_mempool)
1119                 goto err_nomem;
1120         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1121                                                       rpc_buffer_slabp);
1122         if (!rpc_buffer_mempool)
1123                 goto err_nomem;
1124         return 0;
1125 err_nomem:
1126         rpc_destroy_mempool();
1127         return -ENOMEM;
1128 }