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
  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, return %d, status %d\n", task->tk_pid, status, task->tk_status);
 698         /* Wake up anyone who is waiting for task completion */
 699         rpc_mark_complete_task(task);
 700         /* Release all resources associated with the task */
 701         rpc_release_task(task);
 702         return status;
 703 }
 704
 705 /*
 706  * User-visible entry point to the scheduler.
 707  *
 708  * This may be called recursively if e.g. an async NFS task updates
 709  * the attributes and finds that dirty pages must be flushed.
 710  * NOTE: Upon exit of this function the task is guaranteed to be
 711  *       released. In particular note that tk_release() will have
 712  *       been called, so your task memory may have been freed.
 713  */
 714 int
 715 rpc_execute(struct rpc_task *task)
 716 {
 717         rpc_set_active(task);
 718         rpc_set_running(task);
 719         return __rpc_execute(task);
 720 }
 721
 722 static void rpc_async_schedule(void *arg)
 723 {
 724         __rpc_execute((struct rpc_task *)arg);
 725 }
 726
 727 /*
 728  * Allocate memory for RPC purposes.
 729  *
 730  * We try to ensure that some NFS reads and writes can always proceed
 731  * by using a mempool when allocating 'small' buffers.
 732  * In order to avoid memory starvation triggering more writebacks of
 733  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
 734  */
 735 void *
 736 rpc_malloc(struct rpc_task *task, size_t size)
 737 {
 738         gfp_t   gfp;
 739
 740         if (task->tk_flags & RPC_TASK_SWAPPER)
 741                 gfp = GFP_ATOMIC;
 742         else
 743                 gfp = GFP_NOFS;
 744
 745         if (size > RPC_BUFFER_MAXSIZE) {
 746                 task->tk_buffer =  kmalloc(size, gfp);
 747                 if (task->tk_buffer)
 748                         task->tk_bufsize = size;
 749         } else {
 750                 task->tk_buffer =  mempool_alloc(rpc_buffer_mempool, gfp);
 751                 if (task->tk_buffer)
 752                         task->tk_bufsize = RPC_BUFFER_MAXSIZE;
 753         }
 754         return task->tk_buffer;
 755 }
 756
 757 static void
 758 rpc_free(struct rpc_task *task)
 759 {
 760         if (task->tk_buffer) {
 761                 if (task->tk_bufsize == RPC_BUFFER_MAXSIZE)
 762                         mempool_free(task->tk_buffer, rpc_buffer_mempool);
 763                 else
 764                         kfree(task->tk_buffer);
 765                 task->tk_buffer = NULL;
 766                 task->tk_bufsize = 0;
 767         }
 768 }
 769
 770 /*
 771  * Creation and deletion of RPC task structures
 772  */
 773 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 774 {
 775         memset(task, 0, sizeof(*task));
 776         init_timer(&task->tk_timer);
 777         task->tk_timer.data     = (unsigned long) task;
 778         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 779         atomic_set(&task->tk_count, 1);
 780         task->tk_client = clnt;
 781         task->tk_flags  = flags;
 782         task->tk_ops = tk_ops;
 783         if (tk_ops->rpc_call_prepare != NULL)
 784                 task->tk_action = rpc_prepare_task;
 785         task->tk_calldata = calldata;
 786
 787         /* Initialize retry counters */
 788         task->tk_garb_retry = 2;
 789         task->tk_cred_retry = 2;
 790
 791         task->tk_priority = RPC_PRIORITY_NORMAL;
 792         task->tk_cookie = (unsigned long)current;
 793
 794         /* Initialize workqueue for async tasks */
 795         task->tk_workqueue = rpciod_workqueue;
 796
 797         if (clnt) {
 798                 atomic_inc(&clnt->cl_users);
 799                 if (clnt->cl_softrtry)
 800                         task->tk_flags |= RPC_TASK_SOFT;
 801                 if (!clnt->cl_intr)
 802                         task->tk_flags |= RPC_TASK_NOINTR;
 803         }
 804
 805 #ifdef RPC_DEBUG
 806         task->tk_magic = RPC_TASK_MAGIC_ID;
 807         task->tk_pid = rpc_task_id++;
 808 #endif
 809         /* Add to global list of all tasks */
 810         spin_lock(&rpc_sched_lock);
 811         list_add_tail(&task->tk_task, &all_tasks);
 812         spin_unlock(&rpc_sched_lock);
 813
 814         BUG_ON(task->tk_ops == NULL);
 815
 816         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
 817                                 current->pid);
 818 }
 819
 820 static struct rpc_task *
 821 rpc_alloc_task(void)
 822 {
 823         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 824 }
 825
 826 static void rpc_free_task(struct rpc_task *task)
 827 {
 828         dprintk("RPC: %4d freeing task\n", task->tk_pid);
 829         mempool_free(task, rpc_task_mempool);
 830 }
 831
 832 /*
 833  * Create a new task for the specified client.  We have to
 834  * clean up after an allocation failure, as the client may
 835  * have specified "oneshot".
 836  */
 837 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 838 {
 839         struct rpc_task *task;
 840
 841         task = rpc_alloc_task();
 842         if (!task)
 843                 goto cleanup;
 844
 845         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 846
 847         dprintk("RPC: %4d allocated task\n", task->tk_pid);
 848         task->tk_flags |= RPC_TASK_DYNAMIC;
 849 out:
 850         return task;
 851
 852 cleanup:
 853         /* Check whether to release the client */
 854         if (clnt) {
 855                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 856                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 857                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 858                 rpc_release_client(clnt);
 859         }
 860         goto out;
 861 }
 862
 863 void rpc_release_task(struct rpc_task *task)
 864 {
 865         const struct rpc_call_ops *tk_ops = task->tk_ops;
 866         void *calldata = task->tk_calldata;
 867
 868 #ifdef RPC_DEBUG
 869         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 870 #endif
 871         if (!atomic_dec_and_test(&task->tk_count))
 872                 return;
 873         dprintk("RPC: %4d release task\n", task->tk_pid);
 874
 875         /* Remove from global task list */
 876         spin_lock(&rpc_sched_lock);
 877         list_del(&task->tk_task);
 878         spin_unlock(&rpc_sched_lock);
 879
 880         BUG_ON (RPC_IS_QUEUED(task));
 881
 882         /* Synchronously delete any running timer */
 883         rpc_delete_timer(task);
 884
 885         /* Release resources */
 886         if (task->tk_rqstp)
 887                 xprt_release(task);
 888         if (task->tk_msg.rpc_cred)
 889                 rpcauth_unbindcred(task);
 890         rpc_free(task);
 891         if (task->tk_client) {
 892                 rpc_release_client(task->tk_client);
 893                 task->tk_client = NULL;
 894         }
 895
 896 #ifdef RPC_DEBUG
 897         task->tk_magic = 0;
 898 #endif
 899         if (task->tk_flags & RPC_TASK_DYNAMIC)
 900                 rpc_free_task(task);
 901         if (tk_ops->rpc_release)
 902                 tk_ops->rpc_release(calldata);
 903 }
 904
 905 /**
 906  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
 907  * @clnt - pointer to RPC client
 908  * @flags - RPC flags
 909  * @ops - RPC call ops
 910  * @data - user call data
 911  */
 912 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
 913                                         const struct rpc_call_ops *ops,
 914                                         void *data)
 915 {
 916         struct rpc_task *task;
 917         task = rpc_new_task(clnt, flags, ops, data);
 918         if (task == NULL)
 919                 return ERR_PTR(-ENOMEM);
 920         atomic_inc(&task->tk_count);
 921         rpc_execute(task);
 922         return task;
 923 }
 924 EXPORT_SYMBOL(rpc_run_task);
 925
 926 /**
 927  * rpc_find_parent - find the parent of a child task.
 928  * @child: child task
 929  *
 930  * Checks that the parent task is still sleeping on the
 931  * queue 'childq'. If so returns a pointer to the parent.
 932  * Upon failure returns NULL.
 933  *
 934  * Caller must hold childq.lock
 935  */
 936 static inline struct rpc_task *rpc_find_parent(struct rpc_task *child, struct rpc_task *parent)
 937 {
 938         struct rpc_task *task;
 939         struct list_head *le;
 940
 941         task_for_each(task, le, &childq.tasks[0])
 942                 if (task == parent)
 943                         return parent;
 944
 945         return NULL;
 946 }
 947
 948 static void rpc_child_exit(struct rpc_task *child, void *calldata)
 949 {
 950         struct rpc_task *parent;
 951
 952         spin_lock_bh(&childq.lock);
 953         if ((parent = rpc_find_parent(child, calldata)) != NULL) {
 954                 parent->tk_status = child->tk_status;
 955                 __rpc_wake_up_task(parent);
 956         }
 957         spin_unlock_bh(&childq.lock);
 958 }
 959
 960 static const struct rpc_call_ops rpc_child_ops = {
 961         .rpc_call_done = rpc_child_exit,
 962 };
 963
 964 /*
 965  * Note: rpc_new_task releases the client after a failure.
 966  */
 967 struct rpc_task *
 968 rpc_new_child(struct rpc_clnt *clnt, struct rpc_task *parent)
 969 {
 970         struct rpc_task *task;
 971
 972         task = rpc_new_task(clnt, RPC_TASK_ASYNC | RPC_TASK_CHILD, &rpc_child_ops, parent);
 973         if (!task)
 974                 goto fail;
 975         return task;
 976
 977 fail:
 978         parent->tk_status = -ENOMEM;
 979         return NULL;
 980 }
 981
 982 void rpc_run_child(struct rpc_task *task, struct rpc_task *child, rpc_action func)
 983 {
 984         spin_lock_bh(&childq.lock);
 985         /* N.B. Is it possible for the child to have already finished? */
 986         __rpc_sleep_on(&childq, task, func, NULL);
 987         rpc_schedule_run(child);
 988         spin_unlock_bh(&childq.lock);
 989 }
 990
 991 /*
 992  * Kill all tasks for the given client.
 993  * XXX: kill their descendants as well?
 994  */
 995 void rpc_killall_tasks(struct rpc_clnt *clnt)
 996 {
 997         struct rpc_task *rovr;
 998         struct list_head *le;
 999
1000         dprintk("RPC:      killing all tasks for client %p\n", clnt);
1001
1002         /*
1003          * Spin lock all_tasks to prevent changes...
1004          */
1005         spin_lock(&rpc_sched_lock);
1006         alltask_for_each(rovr, le, &all_tasks) {
1007                 if (! RPC_IS_ACTIVATED(rovr))
1008                         continue;
1009                 if (!clnt || rovr->tk_client == clnt) {
1010                         rovr->tk_flags |= RPC_TASK_KILLED;
1011                         rpc_exit(rovr, -EIO);
1012                         rpc_wake_up_task(rovr);
1013                 }
1014         }
1015         spin_unlock(&rpc_sched_lock);
1016 }
1017
1018 static DECLARE_MUTEX_LOCKED(rpciod_running);
1019
1020 static void rpciod_killall(void)
1021 {
1022         unsigned long flags;
1023
1024         while (!list_empty(&all_tasks)) {
1025                 clear_thread_flag(TIF_SIGPENDING);
1026                 rpc_killall_tasks(NULL);
1027                 flush_workqueue(rpciod_workqueue);
1028                 if (!list_empty(&all_tasks)) {
1029                         dprintk("rpciod_killall: waiting for tasks to exit\n");
1030                         yield();
1031                 }
1032         }
1033
1034         spin_lock_irqsave(&current->sighand->siglock, flags);
1035         recalc_sigpending();
1036         spin_unlock_irqrestore(&current->sighand->siglock, flags);
1037 }
1038
1039 /*
1040  * Start up the rpciod process if it's not already running.
1041  */
1042 int
1043 rpciod_up(void)
1044 {
1045         struct workqueue_struct *wq;
1046         int error = 0;
1047
1048         down(&rpciod_sema);
1049         dprintk("rpciod_up: users %d\n", rpciod_users);
1050         rpciod_users++;
1051         if (rpciod_workqueue)
1052                 goto out;
1053         /*
1054          * If there's no pid, we should be the first user.
1055          */
1056         if (rpciod_users > 1)
1057                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
1058         /*
1059          * Create the rpciod thread and wait for it to start.
1060          */
1061         error = -ENOMEM;
1062         wq = create_workqueue("rpciod");
1063         if (wq == NULL) {
1064                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1065                 rpciod_users--;
1066                 goto out;
1067         }
1068         rpciod_workqueue = wq;
1069         error = 0;
1070 out:
1071         up(&rpciod_sema);
1072         return error;
1073 }
1074
1075 void
1076 rpciod_down(void)
1077 {
1078         down(&rpciod_sema);
1079         dprintk("rpciod_down sema %d\n", rpciod_users);
1080         if (rpciod_users) {
1081                 if (--rpciod_users)
1082                         goto out;
1083         } else
1084                 printk(KERN_WARNING "rpciod_down: no users??\n");
1085
1086         if (!rpciod_workqueue) {
1087                 dprintk("rpciod_down: Nothing to do!\n");
1088                 goto out;
1089         }
1090         rpciod_killall();
1091
1092         destroy_workqueue(rpciod_workqueue);
1093         rpciod_workqueue = NULL;
1094  out:
1095         up(&rpciod_sema);
1096 }
1097
1098 #ifdef RPC_DEBUG
1099 void rpc_show_tasks(void)
1100 {
1101         struct list_head *le;
1102         struct rpc_task *t;
1103
1104         spin_lock(&rpc_sched_lock);
1105         if (list_empty(&all_tasks)) {
1106                 spin_unlock(&rpc_sched_lock);
1107                 return;
1108         }
1109         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1110                 "-rpcwait -action- ---ops--\n");
1111         alltask_for_each(t, le, &all_tasks) {
1112                 const char *rpc_waitq = "none";
1113
1114                 if (RPC_IS_QUEUED(t))
1115                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1116
1117                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1118                         t->tk_pid,
1119                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1120                         t->tk_flags, t->tk_status,
1121                         t->tk_client,
1122                         (t->tk_client ? t->tk_client->cl_prog : 0),
1123                         t->tk_rqstp, t->tk_timeout,
1124                         rpc_waitq,
1125                         t->tk_action, t->tk_ops);
1126         }
1127         spin_unlock(&rpc_sched_lock);
1128 }
1129 #endif
1130
1131 void
1132 rpc_destroy_mempool(void)
1133 {
1134         if (rpc_buffer_mempool)
1135                 mempool_destroy(rpc_buffer_mempool);
1136         if (rpc_task_mempool)
1137                 mempool_destroy(rpc_task_mempool);
1138         if (rpc_task_slabp && kmem_cache_destroy(rpc_task_slabp))
1139                 printk(KERN_INFO "rpc_task: not all structures were freed\n");
1140         if (rpc_buffer_slabp && kmem_cache_destroy(rpc_buffer_slabp))
1141                 printk(KERN_INFO "rpc_buffers: not all structures were freed\n");
1142 }
1143
1144 int
1145 rpc_init_mempool(void)
1146 {
1147         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1148                                              sizeof(struct rpc_task),
1149                                              0, SLAB_HWCACHE_ALIGN,
1150                                              NULL, NULL);
1151         if (!rpc_task_slabp)
1152                 goto err_nomem;
1153         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1154                                              RPC_BUFFER_MAXSIZE,
1155                                              0, SLAB_HWCACHE_ALIGN,
1156                                              NULL, NULL);
1157         if (!rpc_buffer_slabp)
1158                 goto err_nomem;
1159         rpc_task_mempool = mempool_create(RPC_TASK_POOLSIZE,
1160                                             mempool_alloc_slab,
1161                                             mempool_free_slab,
1162                                             rpc_task_slabp);
1163         if (!rpc_task_mempool)
1164                 goto err_nomem;
1165         rpc_buffer_mempool = mempool_create(RPC_BUFFER_POOLSIZE,
1166                                             mempool_alloc_slab,
1167                                             mempool_free_slab,
1168                                             rpc_buffer_slabp);
1169         if (!rpc_buffer_mempool)
1170                 goto err_nomem;
1171         return 0;
1172 err_nomem:
1173         rpc_destroy_mempool();
1174         return -ENOMEM;
1175 }