git.oblomov.eu Git - 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
  22 #include <linux/sunrpc/clnt.h>
  23 #include <linux/sunrpc/xprt.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 kmem_cache_t     *rpc_task_slabp __read_mostly;
  38 static kmem_cache_t     *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_free(struct rpc_task *task);
  45
  46 static void                     rpc_async_schedule(void *);
  47
  48 /*
  49  * RPC tasks that create another task (e.g. for contacting the portmapper)
  50  * will wait on this queue for their child's completion
  51  */
  52 static RPC_WAITQ(childq, "childq");
  53
  54 /*
  55  * RPC tasks sit here while waiting for conditions to improve.
  56  */
  57 static RPC_WAITQ(delay_queue, "delayq");
  58
  59 /*
  60  * All RPC tasks are linked into this list
  61  */
  62 static LIST_HEAD(all_tasks);
  63
  64 /*
  65  * rpciod-related stuff
  66  */
  67 static DECLARE_MUTEX(rpciod_sema);
  68 static unsigned int             rpciod_users;
  69 static struct workqueue_struct *rpciod_workqueue;
  70
  71 /*
  72  * Spinlock for other critical sections of code.
  73  */
  74 static DEFINE_SPINLOCK(rpc_sched_lock);
  75
  76 /*
  77  * Disable the timer for a given RPC task. Should be called with
  78  * queue->lock and bh_disabled in order to avoid races within
  79  * rpc_run_timer().
  80  */
  81 static inline void
  82 __rpc_disable_timer(struct rpc_task *task)
  83 {
  84         dprintk("RPC: %4d disabling timer\n", task->tk_pid);
  85         task->tk_timeout_fn = NULL;
  86         task->tk_timeout = 0;
  87 }
  88
  89 /*
  90  * Run a timeout function.
  91  * We use the callback in order to allow __rpc_wake_up_task()
  92  * and friends to disable the timer synchronously on SMP systems
  93  * without calling del_timer_sync(). The latter could cause a
  94  * deadlock if called while we're holding spinlocks...
  95  */
  96 static void rpc_run_timer(struct rpc_task *task)
  97 {
  98         void (*callback)(struct rpc_task *);
  99
 100         callback = task->tk_timeout_fn;
 101         task->tk_timeout_fn = NULL;
 102         if (callback && RPC_IS_QUEUED(task)) {
 103                 dprintk("RPC: %4d running timer\n", task->tk_pid);
 104                 callback(task);
 105         }
 106         smp_mb__before_clear_bit();
 107         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 108         smp_mb__after_clear_bit();
 109 }
 110
 111 /*
 112  * Set up a timer for the current task.
 113  */
 114 static inline void
 115 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
 116 {
 117         if (!task->tk_timeout)
 118                 return;
 119
 120         dprintk("RPC: %4d setting alarm for %lu ms\n",
 121                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 122
 123         if (timer)
 124                 task->tk_timeout_fn = timer;
 125         else
 126                 task->tk_timeout_fn = __rpc_default_timer;
 127         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 128         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 129 }
 130
 131 /*
 132  * Delete any timer for the current task. Because we use del_timer_sync(),
 133  * this function should never be called while holding queue->lock.
 134  */
 135 static void
 136 rpc_delete_timer(struct rpc_task *task)
 137 {
 138         if (RPC_IS_QUEUED(task))
 139                 return;
 140         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 141                 del_singleshot_timer_sync(&task->tk_timer);
 142                 dprintk("RPC: %4d deleting timer\n", task->tk_pid);
 143         }
 144 }
 145
 146 /*
 147  * Add new request to a priority queue.
 148  */
 149 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 150 {
 151         struct list_head *q;
 152         struct rpc_task *t;
 153
 154         INIT_LIST_HEAD(&task->u.tk_wait.links);
 155         q = &queue->tasks[task->tk_priority];
 156         if (unlikely(task->tk_priority > queue->maxpriority))
 157                 q = &queue->tasks[queue->maxpriority];
 158         list_for_each_entry(t, q, u.tk_wait.list) {
 159                 if (t->tk_cookie == task->tk_cookie) {
 160                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 161                         return;
 162                 }
 163         }
 164         list_add_tail(&task->u.tk_wait.list, q);
 165 }
 166
 167 /*
 168  * Add new request to wait queue.
 169  *
 170  * Swapper tasks always get inserted at the head of the queue.
 171  * This should avoid many nasty memory deadlocks and hopefully
 172  * improve overall performance.
 173  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 174  */
 175 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 176 {
 177         BUG_ON (RPC_IS_QUEUED(task));
 178
 179         if (RPC_IS_PRIORITY(queue))
 180                 __rpc_add_wait_queue_priority(queue, task);
 181         else if (RPC_IS_SWAPPER(task))
 182                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 183         else
 184                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 185         task->u.tk_wait.rpc_waitq = queue;
 186         rpc_set_queued(task);
 187
 188         dprintk("RPC: %4d added to queue %p \"%s\"\n",
 189                                 task->tk_pid, queue, rpc_qname(queue));
 190 }
 191
 192 /*
 193  * Remove request from a priority queue.
 194  */
 195 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 196 {
 197         struct rpc_task *t;
 198
 199         if (!list_empty(&task->u.tk_wait.links)) {
 200                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 201                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 202                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 203         }
 204         list_del(&task->u.tk_wait.list);
 205 }
 206
 207 /*
 208  * Remove request from queue.
 209  * Note: must be called with spin lock held.
 210  */
 211 static void __rpc_remove_wait_queue(struct rpc_task *task)
 212 {
 213         struct rpc_wait_queue *queue;
 214         queue = task->u.tk_wait.rpc_waitq;
 215
 216         if (RPC_IS_PRIORITY(queue))
 217                 __rpc_remove_wait_queue_priority(task);
 218         else
 219                 list_del(&task->u.tk_wait.list);
 220         dprintk("RPC: %4d removed from queue %p \"%s\"\n",
 221                                 task->tk_pid, queue, rpc_qname(queue));
 222 }
 223
 224 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 225 {
 226         queue->priority = priority;
 227         queue->count = 1 << (priority * 2);
 228 }
 229
 230 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
 231 {
 232         queue->cookie = cookie;
 233         queue->nr = RPC_BATCH_COUNT;
 234 }
 235
 236 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 237 {
 238         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 239         rpc_set_waitqueue_cookie(queue, 0);
 240 }
 241
 242 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
 243 {
 244         int i;
 245
 246         spin_lock_init(&queue->lock);
 247         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 248                 INIT_LIST_HEAD(&queue->tasks[i]);
 249         queue->maxpriority = maxprio;
 250         rpc_reset_waitqueue_priority(queue);
 251 #ifdef RPC_DEBUG
 252         queue->name = qname;
 253 #endif
 254 }
 255
 256 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 257 {
 258         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
 259 }
 260
 261 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 262 {
 263         __rpc_init_priority_wait_queue(queue, qname, 0);
 264 }
 265 EXPORT_SYMBOL(rpc_init_wait_queue);
 266
 267 static int rpc_wait_bit_interruptible(void *word)
 268 {
 269         if (signal_pending(current))
 270                 return -ERESTARTSYS;
 271         schedule();
 272         return 0;
 273 }
 274
 275 /*
 276  * Mark an RPC call as having completed by clearing the 'active' bit
 277  */
 278 static inline void rpc_mark_complete_task(struct rpc_task *task)
 279 {
 280         rpc_clear_active(task);
 281         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 282 }
 283
 284 /*
 285  * Allow callers to wait for completion of an RPC call
 286  */
 287 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 288 {
 289         if (action == NULL)
 290                 action = rpc_wait_bit_interruptible;
 291         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 292                         action, TASK_INTERRUPTIBLE);
 293 }
 294 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
 295
 296 /*
 297  * Make an RPC task runnable.
 298  *
 299  * Note: If the task is ASYNC, this must be called with
 300  * the spinlock held to protect the wait queue operation.
 301  */
 302 static void rpc_make_runnable(struct rpc_task *task)
 303 {
 304         int do_ret;
 305
 306         BUG_ON(task->tk_timeout_fn);
 307         do_ret = rpc_test_and_set_running(task);
 308         rpc_clear_queued(task);
 309         if (do_ret)
 310                 return;
 311         if (RPC_IS_ASYNC(task)) {
 312                 int status;
 313
 314                 INIT_WORK(&task->u.tk_work, rpc_async_schedule, (void *)task);
 315                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 316                 if (status < 0) {
 317                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 318                         task->tk_status = status;
 319                         return;
 320                 }
 321         } else
 322                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 323 }
 324
 325 /*
 326  * Place a newly initialized task on the workqueue.
 327  */
 328 static inline void
 329 rpc_schedule_run(struct rpc_task *task)
 330 {
 331         rpc_set_active(task);
 332         rpc_make_runnable(task);
 333 }
 334
 335 /*
 336  * Prepare for sleeping on a wait queue.
 337  * By always appending tasks to the list we ensure FIFO behavior.
 338  * NB: An RPC task will only receive interrupt-driven events as long
 339  * as it's on a wait queue.
 340  */
 341 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 342                         rpc_action action, rpc_action timer)
 343 {
 344         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
 345                                 rpc_qname(q), jiffies);
 346
 347         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 348                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 349                 return;
 350         }
 351
 352         /* Mark the task as being activated if so needed */
 353         rpc_set_active(task);
 354
 355         __rpc_add_wait_queue(q, task);
 356
 357         BUG_ON(task->tk_callback != NULL);
 358         task->tk_callback = action;
 359         __rpc_add_timer(task, timer);
 360 }
 361
 362 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 363                                 rpc_action action, rpc_action timer)
 364 {
 365         /*
 366          * Protect the queue operations.
 367          */
 368         spin_lock_bh(&q->lock);
 369         __rpc_sleep_on(q, task, action, timer);
 370         spin_unlock_bh(&q->lock);
 371 }
 372
 373 /**
 374  * __rpc_do_wake_up_task - wake up a single rpc_task
 375  * @task: task to be woken up
 376  *
 377  * Caller must hold queue->lock, and have cleared the task queued flag.
 378  */
 379 static void __rpc_do_wake_up_task(struct rpc_task *task)
 380 {
 381         dprintk("RPC: %4d __rpc_wake_up_task (now %ld)\n", task->tk_pid, jiffies);
 382
 383 #ifdef RPC_DEBUG
 384         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 385 #endif
 386         /* Has the task been executed yet? If not, we cannot wake it up! */
 387         if (!RPC_IS_ACTIVATED(task)) {
 388                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 389                 return;
 390         }
 391
 392         __rpc_disable_timer(task);
 393         __rpc_remove_wait_queue(task);
 394
 395         rpc_make_runnable(task);
 396
 397         dprintk("RPC:      __rpc_wake_up_task done\n");
 398 }
 399
 400 /*
 401  * Wake up the specified task
 402  */
 403 static void __rpc_wake_up_task(struct rpc_task *task)
 404 {
 405         if (rpc_start_wakeup(task)) {
 406                 if (RPC_IS_QUEUED(task))
 407                         __rpc_do_wake_up_task(task);
 408                 rpc_finish_wakeup(task);
 409         }
 410 }
 411
 412 /*
 413  * Default timeout handler if none specified by user
 414  */
 415 static void
 416 __rpc_default_timer(struct rpc_task *task)
 417 {
 418         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
 419         task->tk_status = -ETIMEDOUT;
 420         rpc_wake_up_task(task);
 421 }
 422
 423 /*
 424  * Wake up the specified task
 425  */
 426 void rpc_wake_up_task(struct rpc_task *task)
 427 {
 428         if (rpc_start_wakeup(task)) {
 429                 if (RPC_IS_QUEUED(task)) {
 430                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 431
 432                         spin_lock_bh(&queue->lock);
 433                         __rpc_do_wake_up_task(task);
 434                         spin_unlock_bh(&queue->lock);
 435                 }
 436                 rpc_finish_wakeup(task);
 437         }
 438 }
 439
 440 /*
 441  * Wake up the next task on a priority queue.
 442  */
 443 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 444 {
 445         struct list_head *q;
 446         struct rpc_task *task;
 447
 448         /*
 449          * Service a batch of tasks from a single cookie.
 450          */
 451         q = &queue->tasks[queue->priority];
 452         if (!list_empty(q)) {
 453                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 454                 if (queue->cookie == task->tk_cookie) {
 455                         if (--queue->nr)
 456                                 goto out;
 457                         list_move_tail(&task->u.tk_wait.list, q);
 458                 }
 459                 /*
 460                  * Check if we need to switch queues.
 461                  */
 462                 if (--queue->count)
 463                         goto new_cookie;
 464         }
 465
 466         /*
 467          * Service the next queue.
 468          */
 469         do {
 470                 if (q == &queue->tasks[0])
 471                         q = &queue->tasks[queue->maxpriority];
 472                 else
 473                         q = q - 1;
 474                 if (!list_empty(q)) {
 475                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 476                         goto new_queue;
 477                 }
 478         } while (q != &queue->tasks[queue->priority]);
 479
 480         rpc_reset_waitqueue_priority(queue);
 481         return NULL;
 482
 483 new_queue:
 484         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 485 new_cookie:
 486         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 487 out:
 488         __rpc_wake_up_task(task);
 489         return task;
 490 }
 491
 492 /*
 493  * Wake up the next task on the wait queue.
 494  */
 495 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 496 {
 497         struct rpc_task *task = NULL;
 498
 499         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
 500         spin_lock_bh(&queue->lock);
 501         if (RPC_IS_PRIORITY(queue))
 502                 task = __rpc_wake_up_next_priority(queue);
 503         else {
 504                 task_for_first(task, &queue->tasks[0])
 505                         __rpc_wake_up_task(task);
 506         }
 507         spin_unlock_bh(&queue->lock);
 508
 509         return task;
 510 }
 511
 512 /**
 513  * rpc_wake_up - wake up all rpc_tasks
 514  * @queue: rpc_wait_queue on which the tasks are sleeping
 515  *
 516  * Grabs queue->lock
 517  */
 518 void rpc_wake_up(struct rpc_wait_queue *queue)
 519 {
 520         struct rpc_task *task;
 521
 522         struct list_head *head;
 523         spin_lock_bh(&queue->lock);
 524         head = &queue->tasks[queue->maxpriority];
 525         for (;;) {
 526                 while (!list_empty(head)) {
 527                         task = list_entry(head->next, struct rpc_task, u.tk_wait.list);
 528                         __rpc_wake_up_task(task);
 529                 }
 530                 if (head == &queue->tasks[0])
 531                         break;
 532                 head--;
 533         }
 534         spin_unlock_bh(&queue->lock);
 535 }
 536
 537 /**
 538  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 539  * @queue: rpc_wait_queue on which the tasks are sleeping
 540  * @status: status value to set
 541  *
 542  * Grabs queue->lock
 543  */
 544 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 545 {
 546         struct list_head *head;
 547         struct rpc_task *task;
 548
 549         spin_lock_bh(&queue->lock);
 550         head = &queue->tasks[queue->maxpriority];
 551         for (;;) {
 552                 while (!list_empty(head)) {
 553                         task = list_entry(head->next, struct rpc_task, u.tk_wait.list);
 554                         task->tk_status = status;
 555                         __rpc_wake_up_task(task);
 556                 }
 557                 if (head == &queue->tasks[0])
 558                         break;
 559                 head--;
 560         }
 561         spin_unlock_bh(&queue->lock);
 562 }
 563
 564 /*
 565  * Run a task at a later time
 566  */
 567 static void     __rpc_atrun(struct rpc_task *);
 568 void
 569 rpc_delay(struct rpc_task *task, unsigned long delay)
 570 {
 571         task->tk_timeout = delay;
 572         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 573 }
 574
 575 static void
 576 __rpc_atrun(struct rpc_task *task)
 577 {
 578         task->tk_status = 0;
 579         rpc_wake_up_task(task);
 580 }
 581
 582 /*
 583  * Helper to call task->tk_ops->rpc_call_prepare
 584  */
 585 static void rpc_prepare_task(struct rpc_task *task)
 586 {
 587         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 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                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 598                 if (task->tk_action != NULL) {
 599                         WARN_ON(RPC_ASSASSINATED(task));
 600                         /* Always release the RPC slot and buffer memory */
 601                         xprt_release(task);
 602                         rpc_free(task);
 603                 }
 604         }
 605 }
 606 EXPORT_SYMBOL(rpc_exit_task);
 607
 608 /*
 609  * This is the RPC `scheduler' (or rather, the finite state machine).
 610  */
 611 static int __rpc_execute(struct rpc_task *task)
 612 {
 613         int             status = 0;
 614
 615         dprintk("RPC: %4d rpc_execute flgs %x\n",
 616                                 task->tk_pid, task->tk_flags);
 617
 618         BUG_ON(RPC_IS_QUEUED(task));
 619
 620         for (;;) {
 621                 /*
 622                  * Garbage collection of pending timers...
 623                  */
 624                 rpc_delete_timer(task);
 625
 626                 /*
 627                  * Execute any pending callback.
 628                  */
 629                 if (RPC_DO_CALLBACK(task)) {
 630                         /* Define a callback save pointer */
 631                         void (*save_callback)(struct rpc_task *);
 632
 633                         /*
 634                          * If a callback exists, save it, reset it,
 635                          * call it.
 636                          * The save is needed to stop from resetting
 637                          * another callback set within the callback handler
 638                          * - Dave
 639                          */
 640                         save_callback=task->tk_callback;
 641                         task->tk_callback=NULL;
 642                         lock_kernel();
 643                         save_callback(task);
 644                         unlock_kernel();
 645                 }
 646
 647                 /*
 648                  * Perform the next FSM step.
 649                  * tk_action may be NULL when the task has been killed
 650                  * by someone else.
 651                  */
 652                 if (!RPC_IS_QUEUED(task)) {
 653                         if (task->tk_action == NULL)
 654                                 break;
 655                         lock_kernel();
 656                         task->tk_action(task);
 657                         unlock_kernel();
 658                 }
 659
 660                 /*
 661                  * Lockless check for whether task is sleeping or not.
 662                  */
 663                 if (!RPC_IS_QUEUED(task))
 664                         continue;
 665                 rpc_clear_running(task);
 666                 if (RPC_IS_ASYNC(task)) {
 667                         /* Careful! we may have raced... */
 668                         if (RPC_IS_QUEUED(task))
 669                                 return 0;
 670                         if (rpc_test_and_set_running(task))
 671                                 return 0;
 672                         continue;
 673                 }
 674
 675                 /* sync task: sleep here */
 676                 dprintk("RPC: %4d sync task going to sleep\n", task->tk_pid);
 677                 /* Note: Caller should be using rpc_clnt_sigmask() */
 678                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 679                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
 680                                 TASK_INTERRUPTIBLE);
 681                 if (status == -ERESTARTSYS) {
 682                         /*
 683                          * When a sync task receives a signal, it exits with
 684                          * -ERESTARTSYS. In order to catch any callbacks that
 685                          * clean up after sleeping on some queue, we don't
 686                          * break the loop here, but go around once more.
 687                          */
 688                         dprintk("RPC: %4d got signal\n", task->tk_pid);
 689                         task->tk_flags |= RPC_TASK_KILLED;
 690                         rpc_exit(task, -ERESTARTSYS);
 691                         rpc_wake_up_task(task);
 692                 }
 693                 rpc_set_running(task);
 694                 dprintk("RPC: %4d sync task resuming\n", task->tk_pid);
 695         }
 696
 697         dprintk("RPC: %4d exit() = %d\n", task->tk_pid, task->tk_status);
 698         status = task->tk_status;
 699
 700         /* Wake up anyone who is waiting for task completion */
 701         rpc_mark_complete_task(task);
 702         /* Release all resources associated with the task */
 703         rpc_release_task(task);
 704         return status;
 705 }
 706
 707 /*
 708  * User-visible entry point to the scheduler.
 709  *
 710  * This may be called recursively if e.g. an async NFS task updates
 711  * the attributes and finds that dirty pages must be flushed.
 712  * NOTE: Upon exit of this function the task is guaranteed to be
 713  *       released. In particular note that tk_release() will have
 714  *       been called, so your task memory may have been freed.
 715  */
 716 int
 717 rpc_execute(struct rpc_task *task)
 718 {
 719         rpc_set_active(task);
 720         rpc_set_running(task);
 721         return __rpc_execute(task);
 722 }
 723
 724 static void rpc_async_schedule(void *arg)
 725 {
 726         __rpc_execute((struct rpc_task *)arg);
 727 }
 728
 729 /*
 730  * Allocate memory for RPC purposes.
 731  *
 732  * We try to ensure that some NFS reads and writes can always proceed
 733  * by using a mempool when allocating 'small' buffers.
 734  * In order to avoid memory starvation triggering more writebacks of
 735  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
 736  */
 737 void *
 738 rpc_malloc(struct rpc_task *task, size_t size)
 739 {
 740         gfp_t   gfp;
 741
 742         if (task->tk_flags & RPC_TASK_SWAPPER)
 743                 gfp = GFP_ATOMIC;
 744         else
 745                 gfp = GFP_NOFS;
 746
 747         if (size > RPC_BUFFER_MAXSIZE) {
 748                 task->tk_buffer =  kmalloc(size, gfp);
 749                 if (task->tk_buffer)
 750                         task->tk_bufsize = size;
 751         } else {
 752                 task->tk_buffer =  mempool_alloc(rpc_buffer_mempool, gfp);
 753                 if (task->tk_buffer)
 754                         task->tk_bufsize = RPC_BUFFER_MAXSIZE;
 755         }
 756         return task->tk_buffer;
 757 }
 758
 759 static void
 760 rpc_free(struct rpc_task *task)
 761 {
 762         if (task->tk_buffer) {
 763                 if (task->tk_bufsize == RPC_BUFFER_MAXSIZE)
 764                         mempool_free(task->tk_buffer, rpc_buffer_mempool);
 765                 else
 766                         kfree(task->tk_buffer);
 767                 task->tk_buffer = NULL;
 768                 task->tk_bufsize = 0;
 769         }
 770 }
 771
 772 /*
 773  * Creation and deletion of RPC task structures
 774  */
 775 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 776 {
 777         memset(task, 0, sizeof(*task));
 778         init_timer(&task->tk_timer);
 779         task->tk_timer.data     = (unsigned long) task;
 780         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 781         atomic_set(&task->tk_count, 1);
 782         task->tk_client = clnt;
 783         task->tk_flags  = flags;
 784         task->tk_ops = tk_ops;
 785         if (tk_ops->rpc_call_prepare != NULL)
 786                 task->tk_action = rpc_prepare_task;
 787         task->tk_calldata = calldata;
 788
 789         /* Initialize retry counters */
 790         task->tk_garb_retry = 2;
 791         task->tk_cred_retry = 2;
 792
 793         task->tk_priority = RPC_PRIORITY_NORMAL;
 794         task->tk_cookie = (unsigned long)current;
 795
 796         /* Initialize workqueue for async tasks */
 797         task->tk_workqueue = rpciod_workqueue;
 798
 799         if (clnt) {
 800                 atomic_inc(&clnt->cl_users);
 801                 if (clnt->cl_softrtry)
 802                         task->tk_flags |= RPC_TASK_SOFT;
 803                 if (!clnt->cl_intr)
 804                         task->tk_flags |= RPC_TASK_NOINTR;
 805         }
 806
 807 #ifdef RPC_DEBUG
 808         task->tk_magic = RPC_TASK_MAGIC_ID;
 809         task->tk_pid = rpc_task_id++;
 810 #endif
 811         /* Add to global list of all tasks */
 812         spin_lock(&rpc_sched_lock);
 813         list_add_tail(&task->tk_task, &all_tasks);
 814         spin_unlock(&rpc_sched_lock);
 815
 816         BUG_ON(task->tk_ops == NULL);
 817
 818         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
 819                                 current->pid);
 820 }
 821
 822 static struct rpc_task *
 823 rpc_alloc_task(void)
 824 {
 825         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 826 }
 827
 828 static void rpc_free_task(struct rpc_task *task)
 829 {
 830         dprintk("RPC: %4d freeing task\n", task->tk_pid);
 831         mempool_free(task, rpc_task_mempool);
 832 }
 833
 834 /*
 835  * Create a new task for the specified client.  We have to
 836  * clean up after an allocation failure, as the client may
 837  * have specified "oneshot".
 838  */
 839 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 840 {
 841         struct rpc_task *task;
 842
 843         task = rpc_alloc_task();
 844         if (!task)
 845                 goto cleanup;
 846
 847         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 848
 849         dprintk("RPC: %4d allocated task\n", task->tk_pid);
 850         task->tk_flags |= RPC_TASK_DYNAMIC;
 851 out:
 852         return task;
 853
 854 cleanup:
 855         /* Check whether to release the client */
 856         if (clnt) {
 857                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 858                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 859                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 860                 rpc_release_client(clnt);
 861         }
 862         goto out;
 863 }
 864
 865 void rpc_release_task(struct rpc_task *task)
 866 {
 867         const struct rpc_call_ops *tk_ops = task->tk_ops;
 868         void *calldata = task->tk_calldata;
 869
 870 #ifdef RPC_DEBUG
 871         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 872 #endif
 873         if (!atomic_dec_and_test(&task->tk_count))
 874                 return;
 875         dprintk("RPC: %4d release task\n", task->tk_pid);
 876
 877         /* Remove from global task list */
 878         spin_lock(&rpc_sched_lock);
 879         list_del(&task->tk_task);
 880         spin_unlock(&rpc_sched_lock);
 881
 882         BUG_ON (RPC_IS_QUEUED(task));
 883
 884         /* Synchronously delete any running timer */
 885         rpc_delete_timer(task);
 886
 887         /* Release resources */
 888         if (task->tk_rqstp)
 889                 xprt_release(task);
 890         if (task->tk_msg.rpc_cred)
 891                 rpcauth_unbindcred(task);
 892         rpc_free(task);
 893         if (task->tk_client) {
 894                 rpc_release_client(task->tk_client);
 895                 task->tk_client = NULL;
 896         }
 897
 898 #ifdef RPC_DEBUG
 899         task->tk_magic = 0;
 900 #endif
 901         if (task->tk_flags & RPC_TASK_DYNAMIC)
 902                 rpc_free_task(task);
 903         if (tk_ops->rpc_release)
 904                 tk_ops->rpc_release(calldata);
 905 }
 906
 907 /**
 908  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
 909  * @clnt - pointer to RPC client
 910  * @flags - RPC flags
 911  * @ops - RPC call ops
 912  * @data - user call data
 913  */
 914 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
 915                                         const struct rpc_call_ops *ops,
 916                                         void *data)
 917 {
 918         struct rpc_task *task;
 919         task = rpc_new_task(clnt, flags, ops, data);
 920         if (task == NULL)
 921                 return ERR_PTR(-ENOMEM);
 922         atomic_inc(&task->tk_count);
 923         rpc_execute(task);
 924         return task;
 925 }
 926 EXPORT_SYMBOL(rpc_run_task);
 927
 928 /**
 929  * rpc_find_parent - find the parent of a child task.
 930  * @child: child task
 931  *
 932  * Checks that the parent task is still sleeping on the
 933  * queue 'childq'. If so returns a pointer to the parent.
 934  * Upon failure returns NULL.
 935  *
 936  * Caller must hold childq.lock
 937  */
 938 static inline struct rpc_task *rpc_find_parent(struct rpc_task *child, struct rpc_task *parent)
 939 {
 940         struct rpc_task *task;
 941         struct list_head *le;
 942
 943         task_for_each(task, le, &childq.tasks[0])
 944                 if (task == parent)
 945                         return parent;
 946
 947         return NULL;
 948 }
 949
 950 static void rpc_child_exit(struct rpc_task *child, void *calldata)
 951 {
 952         struct rpc_task *parent;
 953
 954         spin_lock_bh(&childq.lock);
 955         if ((parent = rpc_find_parent(child, calldata)) != NULL) {
 956                 parent->tk_status = child->tk_status;
 957                 __rpc_wake_up_task(parent);
 958         }
 959         spin_unlock_bh(&childq.lock);
 960 }
 961
 962 static const struct rpc_call_ops rpc_child_ops = {
 963         .rpc_call_done = rpc_child_exit,
 964 };
 965
 966 /*
 967  * Note: rpc_new_task releases the client after a failure.
 968  */
 969 struct rpc_task *
 970 rpc_new_child(struct rpc_clnt *clnt, struct rpc_task *parent)
 971 {
 972         struct rpc_task *task;
 973
 974         task = rpc_new_task(clnt, RPC_TASK_ASYNC | RPC_TASK_CHILD, &rpc_child_ops, parent);
 975         if (!task)
 976                 goto fail;
 977         return task;
 978
 979 fail:
 980         parent->tk_status = -ENOMEM;
 981         return NULL;
 982 }
 983
 984 void rpc_run_child(struct rpc_task *task, struct rpc_task *child, rpc_action func)
 985 {
 986         spin_lock_bh(&childq.lock);
 987         /* N.B. Is it possible for the child to have already finished? */
 988         __rpc_sleep_on(&childq, task, func, NULL);
 989         rpc_schedule_run(child);
 990         spin_unlock_bh(&childq.lock);
 991 }
 992
 993 /*
 994  * Kill all tasks for the given client.
 995  * XXX: kill their descendants as well?
 996  */
 997 void rpc_killall_tasks(struct rpc_clnt *clnt)
 998 {
 999         struct rpc_task *rovr;
1000         struct list_head *le;
1001
1002         dprintk("RPC:      killing all tasks for client %p\n", clnt);
1003
1004         /*
1005          * Spin lock all_tasks to prevent changes...
1006          */
1007         spin_lock(&rpc_sched_lock);
1008         alltask_for_each(rovr, le, &all_tasks) {
1009                 if (! RPC_IS_ACTIVATED(rovr))
1010                         continue;
1011                 if (!clnt || rovr->tk_client == clnt) {
1012                         rovr->tk_flags |= RPC_TASK_KILLED;
1013                         rpc_exit(rovr, -EIO);
1014                         rpc_wake_up_task(rovr);
1015                 }
1016         }
1017         spin_unlock(&rpc_sched_lock);
1018 }
1019
1020 static DECLARE_MUTEX_LOCKED(rpciod_running);
1021
1022 static void rpciod_killall(void)
1023 {
1024         unsigned long flags;
1025
1026         while (!list_empty(&all_tasks)) {
1027                 clear_thread_flag(TIF_SIGPENDING);
1028                 rpc_killall_tasks(NULL);
1029                 flush_workqueue(rpciod_workqueue);
1030                 if (!list_empty(&all_tasks)) {
1031                         dprintk("rpciod_killall: waiting for tasks to exit\n");
1032                         yield();
1033                 }
1034         }
1035
1036         spin_lock_irqsave(&current->sighand->siglock, flags);
1037         recalc_sigpending();
1038         spin_unlock_irqrestore(&current->sighand->siglock, flags);
1039 }
1040
1041 /*
1042  * Start up the rpciod process if it's not already running.
1043  */
1044 int
1045 rpciod_up(void)
1046 {
1047         struct workqueue_struct *wq;
1048         int error = 0;
1049
1050         down(&rpciod_sema);
1051         dprintk("rpciod_up: users %d\n", rpciod_users);
1052         rpciod_users++;
1053         if (rpciod_workqueue)
1054                 goto out;
1055         /*
1056          * If there's no pid, we should be the first user.
1057          */
1058         if (rpciod_users > 1)
1059                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
1060         /*
1061          * Create the rpciod thread and wait for it to start.
1062          */
1063         error = -ENOMEM;
1064         wq = create_workqueue("rpciod");
1065         if (wq == NULL) {
1066                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1067                 rpciod_users--;
1068                 goto out;
1069         }
1070         rpciod_workqueue = wq;
1071         error = 0;
1072 out:
1073         up(&rpciod_sema);
1074         return error;
1075 }
1076
1077 void
1078 rpciod_down(void)
1079 {
1080         down(&rpciod_sema);
1081         dprintk("rpciod_down sema %d\n", rpciod_users);
1082         if (rpciod_users) {
1083                 if (--rpciod_users)
1084                         goto out;
1085         } else
1086                 printk(KERN_WARNING "rpciod_down: no users??\n");
1087
1088         if (!rpciod_workqueue) {
1089                 dprintk("rpciod_down: Nothing to do!\n");
1090                 goto out;
1091         }
1092         rpciod_killall();
1093
1094         destroy_workqueue(rpciod_workqueue);
1095         rpciod_workqueue = NULL;
1096  out:
1097         up(&rpciod_sema);
1098 }
1099
1100 #ifdef RPC_DEBUG
1101 void rpc_show_tasks(void)
1102 {
1103         struct list_head *le;
1104         struct rpc_task *t;
1105
1106         spin_lock(&rpc_sched_lock);
1107         if (list_empty(&all_tasks)) {
1108                 spin_unlock(&rpc_sched_lock);
1109                 return;
1110         }
1111         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1112                 "-rpcwait -action- ---ops--\n");
1113         alltask_for_each(t, le, &all_tasks) {
1114                 const char *rpc_waitq = "none";
1115
1116                 if (RPC_IS_QUEUED(t))
1117                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1118
1119                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1120                         t->tk_pid,
1121                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1122                         t->tk_flags, t->tk_status,
1123                         t->tk_client,
1124                         (t->tk_client ? t->tk_client->cl_prog : 0),
1125                         t->tk_rqstp, t->tk_timeout,
1126                         rpc_waitq,
1127                         t->tk_action, t->tk_ops);
1128         }
1129         spin_unlock(&rpc_sched_lock);
1130 }
1131 #endif
1132
1133 void
1134 rpc_destroy_mempool(void)
1135 {
1136         if (rpc_buffer_mempool)
1137                 mempool_destroy(rpc_buffer_mempool);
1138         if (rpc_task_mempool)
1139                 mempool_destroy(rpc_task_mempool);
1140         if (rpc_task_slabp && kmem_cache_destroy(rpc_task_slabp))
1141                 printk(KERN_INFO "rpc_task: not all structures were freed\n");
1142         if (rpc_buffer_slabp && kmem_cache_destroy(rpc_buffer_slabp))
1143                 printk(KERN_INFO "rpc_buffers: not all structures were freed\n");
1144 }
1145
1146 int
1147 rpc_init_mempool(void)
1148 {
1149         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1150                                              sizeof(struct rpc_task),
1151                                              0, SLAB_HWCACHE_ALIGN,
1152                                              NULL, NULL);
1153         if (!rpc_task_slabp)
1154                 goto err_nomem;
1155         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1156                                              RPC_BUFFER_MAXSIZE,
1157                                              0, SLAB_HWCACHE_ALIGN,
1158                                              NULL, NULL);
1159         if (!rpc_buffer_slabp)
1160                 goto err_nomem;
1161         rpc_task_mempool = mempool_create(RPC_TASK_POOLSIZE,
1162                                             mempool_alloc_slab,
1163                                             mempool_free_slab,
1164                                             rpc_task_slabp);
1165         if (!rpc_task_mempool)
1166                 goto err_nomem;
1167         rpc_buffer_mempool = mempool_create(RPC_BUFFER_POOLSIZE,
1168                                             mempool_alloc_slab,
1169                                             mempool_free_slab,
1170                                             rpc_buffer_slabp);
1171         if (!rpc_buffer_mempool)
1172                 goto err_nomem;
1173         return 0;
1174 err_nomem:
1175         rpc_destroy_mempool();
1176         return -ENOMEM;
1177 }