SAFE public projects git trees. - safe/jmp/linux-2.6/blob - net/sunrpc/sched.c

   1 /*
   2  * linux/net/sunrpc/sched.c
   3  *
   4  * Scheduling for synchronous and asynchronous RPC requests.
   5  *
   6  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  10  */
  11
  12 #include <linux/module.h>
  13
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/slab.h>
  17 #include <linux/mempool.h>
  18 #include <linux/smp.h>
  19 #include <linux/smp_lock.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/mutex.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #ifdef RPC_DEBUG
  26 #define RPCDBG_FACILITY         RPCDBG_SCHED
  27 #define RPC_TASK_MAGIC_ID       0xf00baa
  28 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 static void                      rpc_release_task(struct rpc_task *task);
  46
  47 /*
  48  * RPC tasks sit here while waiting for conditions to improve.
  49  */
  50 static RPC_WAITQ(delay_queue, "delayq");
  51
  52 /*
  53  * All RPC clients are linked into this list
  54  */
  55 static LIST_HEAD(all_clients);
  56 static DECLARE_WAIT_QUEUE_HEAD(client_kill_wait);
  57
  58 /*
  59  * rpciod-related stuff
  60  */
  61 static DEFINE_MUTEX(rpciod_mutex);
  62 static unsigned int             rpciod_users;
  63 struct workqueue_struct *rpciod_workqueue;
  64
  65 /*
  66  * Spinlock for other critical sections of code.
  67  */
  68 static DEFINE_SPINLOCK(rpc_sched_lock);
  69
  70 /*
  71  * Disable the timer for a given RPC task. Should be called with
  72  * queue->lock and bh_disabled in order to avoid races within
  73  * rpc_run_timer().
  74  */
  75 static inline void
  76 __rpc_disable_timer(struct rpc_task *task)
  77 {
  78         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  79         task->tk_timeout_fn = NULL;
  80         task->tk_timeout = 0;
  81 }
  82
  83 /*
  84  * Run a timeout function.
  85  * We use the callback in order to allow __rpc_wake_up_task()
  86  * and friends to disable the timer synchronously on SMP systems
  87  * without calling del_timer_sync(). The latter could cause a
  88  * deadlock if called while we're holding spinlocks...
  89  */
  90 static void rpc_run_timer(struct rpc_task *task)
  91 {
  92         void (*callback)(struct rpc_task *);
  93
  94         callback = task->tk_timeout_fn;
  95         task->tk_timeout_fn = NULL;
  96         if (callback && RPC_IS_QUEUED(task)) {
  97                 dprintk("RPC: %5u running timer\n", task->tk_pid);
  98                 callback(task);
  99         }
 100         smp_mb__before_clear_bit();
 101         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 102         smp_mb__after_clear_bit();
 103 }
 104
 105 /*
 106  * Set up a timer for the current task.
 107  */
 108 static inline void
 109 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
 110 {
 111         if (!task->tk_timeout)
 112                 return;
 113
 114         dprintk("RPC: %5u setting alarm for %lu ms\n",
 115                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 116
 117         if (timer)
 118                 task->tk_timeout_fn = timer;
 119         else
 120                 task->tk_timeout_fn = __rpc_default_timer;
 121         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 122         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 123 }
 124
 125 /*
 126  * Delete any timer for the current task. Because we use del_timer_sync(),
 127  * this function should never be called while holding queue->lock.
 128  */
 129 static void
 130 rpc_delete_timer(struct rpc_task *task)
 131 {
 132         if (RPC_IS_QUEUED(task))
 133                 return;
 134         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 135                 del_singleshot_timer_sync(&task->tk_timer);
 136                 dprintk("RPC: %5u deleting timer\n", task->tk_pid);
 137         }
 138 }
 139
 140 /*
 141  * Add new request to a priority queue.
 142  */
 143 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 144 {
 145         struct list_head *q;
 146         struct rpc_task *t;
 147
 148         INIT_LIST_HEAD(&task->u.tk_wait.links);
 149         q = &queue->tasks[task->tk_priority];
 150         if (unlikely(task->tk_priority > queue->maxpriority))
 151                 q = &queue->tasks[queue->maxpriority];
 152         list_for_each_entry(t, q, u.tk_wait.list) {
 153                 if (t->tk_cookie == task->tk_cookie) {
 154                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 155                         return;
 156                 }
 157         }
 158         list_add_tail(&task->u.tk_wait.list, q);
 159 }
 160
 161 /*
 162  * Add new request to wait queue.
 163  *
 164  * Swapper tasks always get inserted at the head of the queue.
 165  * This should avoid many nasty memory deadlocks and hopefully
 166  * improve overall performance.
 167  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 168  */
 169 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 170 {
 171         BUG_ON (RPC_IS_QUEUED(task));
 172
 173         if (RPC_IS_PRIORITY(queue))
 174                 __rpc_add_wait_queue_priority(queue, task);
 175         else if (RPC_IS_SWAPPER(task))
 176                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 177         else
 178                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 179         task->u.tk_wait.rpc_waitq = queue;
 180         queue->qlen++;
 181         rpc_set_queued(task);
 182
 183         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 184                         task->tk_pid, queue, rpc_qname(queue));
 185 }
 186
 187 /*
 188  * Remove request from a priority queue.
 189  */
 190 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 191 {
 192         struct rpc_task *t;
 193
 194         if (!list_empty(&task->u.tk_wait.links)) {
 195                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 196                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 197                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 198         }
 199         list_del(&task->u.tk_wait.list);
 200 }
 201
 202 /*
 203  * Remove request from queue.
 204  * Note: must be called with spin lock held.
 205  */
 206 static void __rpc_remove_wait_queue(struct rpc_task *task)
 207 {
 208         struct rpc_wait_queue *queue;
 209         queue = task->u.tk_wait.rpc_waitq;
 210
 211         if (RPC_IS_PRIORITY(queue))
 212                 __rpc_remove_wait_queue_priority(task);
 213         else
 214                 list_del(&task->u.tk_wait.list);
 215         queue->qlen--;
 216         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 217                         task->tk_pid, queue, rpc_qname(queue));
 218 }
 219
 220 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 221 {
 222         queue->priority = priority;
 223         queue->count = 1 << (priority * 2);
 224 }
 225
 226 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
 227 {
 228         queue->cookie = cookie;
 229         queue->nr = RPC_BATCH_COUNT;
 230 }
 231
 232 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 233 {
 234         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 235         rpc_set_waitqueue_cookie(queue, 0);
 236 }
 237
 238 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
 239 {
 240         int i;
 241
 242         spin_lock_init(&queue->lock);
 243         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 244                 INIT_LIST_HEAD(&queue->tasks[i]);
 245         queue->maxpriority = maxprio;
 246         rpc_reset_waitqueue_priority(queue);
 247 #ifdef RPC_DEBUG
 248         queue->name = qname;
 249 #endif
 250 }
 251
 252 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 253 {
 254         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
 255 }
 256
 257 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 258 {
 259         __rpc_init_priority_wait_queue(queue, qname, 0);
 260 }
 261 EXPORT_SYMBOL(rpc_init_wait_queue);
 262
 263 static int rpc_wait_bit_interruptible(void *word)
 264 {
 265         if (signal_pending(current))
 266                 return -ERESTARTSYS;
 267         schedule();
 268         return 0;
 269 }
 270
 271 static void rpc_set_active(struct rpc_task *task)
 272 {
 273         struct rpc_clnt *clnt;
 274         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 275                 return;
 276 #ifdef RPC_DEBUG
 277         task->tk_magic = RPC_TASK_MAGIC_ID;
 278         spin_lock(&rpc_sched_lock);
 279         task->tk_pid = rpc_task_id++;
 280         spin_unlock(&rpc_sched_lock);
 281 #endif
 282         /* Add to global list of all tasks */
 283         clnt = task->tk_client;
 284         if (clnt != NULL) {
 285                 spin_lock(&clnt->cl_lock);
 286                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 287                 spin_unlock(&clnt->cl_lock);
 288         }
 289 }
 290
 291 /*
 292  * Mark an RPC call as having completed by clearing the 'active' bit
 293  */
 294 static void rpc_mark_complete_task(struct rpc_task *task)
 295 {
 296         smp_mb__before_clear_bit();
 297         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 298         smp_mb__after_clear_bit();
 299         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 300 }
 301
 302 /*
 303  * Allow callers to wait for completion of an RPC call
 304  */
 305 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 306 {
 307         if (action == NULL)
 308                 action = rpc_wait_bit_interruptible;
 309         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 310                         action, TASK_INTERRUPTIBLE);
 311 }
 312 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
 313
 314 /*
 315  * Make an RPC task runnable.
 316  *
 317  * Note: If the task is ASYNC, this must be called with
 318  * the spinlock held to protect the wait queue operation.
 319  */
 320 static void rpc_make_runnable(struct rpc_task *task)
 321 {
 322         BUG_ON(task->tk_timeout_fn);
 323         rpc_clear_queued(task);
 324         if (rpc_test_and_set_running(task))
 325                 return;
 326         /* We might have raced */
 327         if (RPC_IS_QUEUED(task)) {
 328                 rpc_clear_running(task);
 329                 return;
 330         }
 331         if (RPC_IS_ASYNC(task)) {
 332                 int status;
 333
 334                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 335                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 336                 if (status < 0) {
 337                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 338                         task->tk_status = status;
 339                         return;
 340                 }
 341         } else
 342                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 343 }
 344
 345 /*
 346  * Prepare for sleeping on a wait queue.
 347  * By always appending tasks to the list we ensure FIFO behavior.
 348  * NB: An RPC task will only receive interrupt-driven events as long
 349  * as it's on a wait queue.
 350  */
 351 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 352                         rpc_action action, rpc_action timer)
 353 {
 354         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 355                         task->tk_pid, rpc_qname(q), jiffies);
 356
 357         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 358                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 359                 return;
 360         }
 361
 362         __rpc_add_wait_queue(q, task);
 363
 364         BUG_ON(task->tk_callback != NULL);
 365         task->tk_callback = action;
 366         __rpc_add_timer(task, timer);
 367 }
 368
 369 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 370                                 rpc_action action, rpc_action timer)
 371 {
 372         /* Mark the task as being activated if so needed */
 373         rpc_set_active(task);
 374
 375         /*
 376          * Protect the queue operations.
 377          */
 378         spin_lock_bh(&q->lock);
 379         __rpc_sleep_on(q, task, action, timer);
 380         spin_unlock_bh(&q->lock);
 381 }
 382
 383 /**
 384  * __rpc_do_wake_up_task - wake up a single rpc_task
 385  * @task: task to be woken up
 386  *
 387  * Caller must hold queue->lock, and have cleared the task queued flag.
 388  */
 389 static void __rpc_do_wake_up_task(struct rpc_task *task)
 390 {
 391         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 392                         task->tk_pid, jiffies);
 393
 394 #ifdef RPC_DEBUG
 395         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 396 #endif
 397         /* Has the task been executed yet? If not, we cannot wake it up! */
 398         if (!RPC_IS_ACTIVATED(task)) {
 399                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 400                 return;
 401         }
 402
 403         __rpc_disable_timer(task);
 404         __rpc_remove_wait_queue(task);
 405
 406         rpc_make_runnable(task);
 407
 408         dprintk("RPC:       __rpc_wake_up_task done\n");
 409 }
 410
 411 /*
 412  * Wake up the specified task
 413  */
 414 static void __rpc_wake_up_task(struct rpc_task *task)
 415 {
 416         if (rpc_start_wakeup(task)) {
 417                 if (RPC_IS_QUEUED(task))
 418                         __rpc_do_wake_up_task(task);
 419                 rpc_finish_wakeup(task);
 420         }
 421 }
 422
 423 /*
 424  * Default timeout handler if none specified by user
 425  */
 426 static void
 427 __rpc_default_timer(struct rpc_task *task)
 428 {
 429         dprintk("RPC: %5u timeout (default timer)\n", task->tk_pid);
 430         task->tk_status = -ETIMEDOUT;
 431         rpc_wake_up_task(task);
 432 }
 433
 434 /*
 435  * Wake up the specified task
 436  */
 437 void rpc_wake_up_task(struct rpc_task *task)
 438 {
 439         rcu_read_lock_bh();
 440         if (rpc_start_wakeup(task)) {
 441                 if (RPC_IS_QUEUED(task)) {
 442                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 443
 444                         /* Note: we're already in a bh-safe context */
 445                         spin_lock(&queue->lock);
 446                         __rpc_do_wake_up_task(task);
 447                         spin_unlock(&queue->lock);
 448                 }
 449                 rpc_finish_wakeup(task);
 450         }
 451         rcu_read_unlock_bh();
 452 }
 453
 454 /*
 455  * Wake up the next task on a priority queue.
 456  */
 457 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 458 {
 459         struct list_head *q;
 460         struct rpc_task *task;
 461
 462         /*
 463          * Service a batch of tasks from a single cookie.
 464          */
 465         q = &queue->tasks[queue->priority];
 466         if (!list_empty(q)) {
 467                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 468                 if (queue->cookie == task->tk_cookie) {
 469                         if (--queue->nr)
 470                                 goto out;
 471                         list_move_tail(&task->u.tk_wait.list, q);
 472                 }
 473                 /*
 474                  * Check if we need to switch queues.
 475                  */
 476                 if (--queue->count)
 477                         goto new_cookie;
 478         }
 479
 480         /*
 481          * Service the next queue.
 482          */
 483         do {
 484                 if (q == &queue->tasks[0])
 485                         q = &queue->tasks[queue->maxpriority];
 486                 else
 487                         q = q - 1;
 488                 if (!list_empty(q)) {
 489                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 490                         goto new_queue;
 491                 }
 492         } while (q != &queue->tasks[queue->priority]);
 493
 494         rpc_reset_waitqueue_priority(queue);
 495         return NULL;
 496
 497 new_queue:
 498         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 499 new_cookie:
 500         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 501 out:
 502         __rpc_wake_up_task(task);
 503         return task;
 504 }
 505
 506 /*
 507  * Wake up the next task on the wait queue.
 508  */
 509 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 510 {
 511         struct rpc_task *task = NULL;
 512
 513         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 514                         queue, rpc_qname(queue));
 515         rcu_read_lock_bh();
 516         spin_lock(&queue->lock);
 517         if (RPC_IS_PRIORITY(queue))
 518                 task = __rpc_wake_up_next_priority(queue);
 519         else {
 520                 task_for_first(task, &queue->tasks[0])
 521                         __rpc_wake_up_task(task);
 522         }
 523         spin_unlock(&queue->lock);
 524         rcu_read_unlock_bh();
 525
 526         return task;
 527 }
 528
 529 /**
 530  * rpc_wake_up - wake up all rpc_tasks
 531  * @queue: rpc_wait_queue on which the tasks are sleeping
 532  *
 533  * Grabs queue->lock
 534  */
 535 void rpc_wake_up(struct rpc_wait_queue *queue)
 536 {
 537         struct rpc_task *task, *next;
 538         struct list_head *head;
 539
 540         rcu_read_lock_bh();
 541         spin_lock(&queue->lock);
 542         head = &queue->tasks[queue->maxpriority];
 543         for (;;) {
 544                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 545                         __rpc_wake_up_task(task);
 546                 if (head == &queue->tasks[0])
 547                         break;
 548                 head--;
 549         }
 550         spin_unlock(&queue->lock);
 551         rcu_read_unlock_bh();
 552 }
 553
 554 /**
 555  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 556  * @queue: rpc_wait_queue on which the tasks are sleeping
 557  * @status: status value to set
 558  *
 559  * Grabs queue->lock
 560  */
 561 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 562 {
 563         struct rpc_task *task, *next;
 564         struct list_head *head;
 565
 566         rcu_read_lock_bh();
 567         spin_lock(&queue->lock);
 568         head = &queue->tasks[queue->maxpriority];
 569         for (;;) {
 570                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 571                         task->tk_status = status;
 572                         __rpc_wake_up_task(task);
 573                 }
 574                 if (head == &queue->tasks[0])
 575                         break;
 576                 head--;
 577         }
 578         spin_unlock(&queue->lock);
 579         rcu_read_unlock_bh();
 580 }
 581
 582 static void __rpc_atrun(struct rpc_task *task)
 583 {
 584         rpc_wake_up_task(task);
 585 }
 586
 587 /*
 588  * Run a task at a later time
 589  */
 590 void rpc_delay(struct rpc_task *task, unsigned long delay)
 591 {
 592         task->tk_timeout = delay;
 593         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 594 }
 595
 596 /*
 597  * Helper to call task->tk_ops->rpc_call_prepare
 598  */
 599 static void rpc_prepare_task(struct rpc_task *task)
 600 {
 601         lock_kernel();
 602         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 603         unlock_kernel();
 604 }
 605
 606 /*
 607  * Helper that calls task->tk_ops->rpc_call_done if it exists
 608  */
 609 void rpc_exit_task(struct rpc_task *task)
 610 {
 611         task->tk_action = NULL;
 612         if (task->tk_ops->rpc_call_done != NULL) {
 613                 lock_kernel();
 614                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 615                 unlock_kernel();
 616                 if (task->tk_action != NULL) {
 617                         WARN_ON(RPC_ASSASSINATED(task));
 618                         /* Always release the RPC slot and buffer memory */
 619                         xprt_release(task);
 620                 }
 621         }
 622 }
 623 EXPORT_SYMBOL(rpc_exit_task);
 624
 625 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 626 {
 627         if (ops->rpc_release != NULL) {
 628                 lock_kernel();
 629                 ops->rpc_release(calldata);
 630                 unlock_kernel();
 631         }
 632 }
 633
 634 /*
 635  * This is the RPC `scheduler' (or rather, the finite state machine).
 636  */
 637 static void __rpc_execute(struct rpc_task *task)
 638 {
 639         int             status = 0;
 640
 641         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 642                         task->tk_pid, task->tk_flags);
 643
 644         BUG_ON(RPC_IS_QUEUED(task));
 645
 646         for (;;) {
 647                 /*
 648                  * Garbage collection of pending timers...
 649                  */
 650                 rpc_delete_timer(task);
 651
 652                 /*
 653                  * Execute any pending callback.
 654                  */
 655                 if (RPC_DO_CALLBACK(task)) {
 656                         /* Define a callback save pointer */
 657                         void (*save_callback)(struct rpc_task *);
 658
 659                         /*
 660                          * If a callback exists, save it, reset it,
 661                          * call it.
 662                          * The save is needed to stop from resetting
 663                          * another callback set within the callback handler
 664                          * - Dave
 665                          */
 666                         save_callback=task->tk_callback;
 667                         task->tk_callback=NULL;
 668                         save_callback(task);
 669                 }
 670
 671                 /*
 672                  * Perform the next FSM step.
 673                  * tk_action may be NULL when the task has been killed
 674                  * by someone else.
 675                  */
 676                 if (!RPC_IS_QUEUED(task)) {
 677                         if (task->tk_action == NULL)
 678                                 break;
 679                         task->tk_action(task);
 680                 }
 681
 682                 /*
 683                  * Lockless check for whether task is sleeping or not.
 684                  */
 685                 if (!RPC_IS_QUEUED(task))
 686                         continue;
 687                 rpc_clear_running(task);
 688                 if (RPC_IS_ASYNC(task)) {
 689                         /* Careful! we may have raced... */
 690                         if (RPC_IS_QUEUED(task))
 691                                 return;
 692                         if (rpc_test_and_set_running(task))
 693                                 return;
 694                         continue;
 695                 }
 696
 697                 /* sync task: sleep here */
 698                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 699                 /* Note: Caller should be using rpc_clnt_sigmask() */
 700                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 701                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
 702                                 TASK_INTERRUPTIBLE);
 703                 if (status == -ERESTARTSYS) {
 704                         /*
 705                          * When a sync task receives a signal, it exits with
 706                          * -ERESTARTSYS. In order to catch any callbacks that
 707                          * clean up after sleeping on some queue, we don't
 708                          * break the loop here, but go around once more.
 709                          */
 710                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 711                         task->tk_flags |= RPC_TASK_KILLED;
 712                         rpc_exit(task, -ERESTARTSYS);
 713                         rpc_wake_up_task(task);
 714                 }
 715                 rpc_set_running(task);
 716                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 717         }
 718
 719         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 720                         task->tk_status);
 721         /* Release all resources associated with the task */
 722         rpc_release_task(task);
 723 }
 724
 725 /*
 726  * User-visible entry point to the scheduler.
 727  *
 728  * This may be called recursively if e.g. an async NFS task updates
 729  * the attributes and finds that dirty pages must be flushed.
 730  * NOTE: Upon exit of this function the task is guaranteed to be
 731  *       released. In particular note that tk_release() will have
 732  *       been called, so your task memory may have been freed.
 733  */
 734 void rpc_execute(struct rpc_task *task)
 735 {
 736         rpc_set_active(task);
 737         rpc_set_running(task);
 738         __rpc_execute(task);
 739 }
 740
 741 static void rpc_async_schedule(struct work_struct *work)
 742 {
 743         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 744 }
 745
 746 struct rpc_buffer {
 747         size_t  len;
 748         char    data[];
 749 };
 750
 751 /**
 752  * rpc_malloc - allocate an RPC buffer
 753  * @task: RPC task that will use this buffer
 754  * @size: requested byte size
 755  *
 756  * To prevent rpciod from hanging, this allocator never sleeps,
 757  * returning NULL if the request cannot be serviced immediately.
 758  * The caller can arrange to sleep in a way that is safe for rpciod.
 759  *
 760  * Most requests are 'small' (under 2KiB) and can be serviced from a
 761  * mempool, ensuring that NFS reads and writes can always proceed,
 762  * and that there is good locality of reference for these buffers.
 763  *
 764  * In order to avoid memory starvation triggering more writebacks of
 765  * NFS requests, we avoid using GFP_KERNEL.
 766  */
 767 void *rpc_malloc(struct rpc_task *task, size_t size)
 768 {
 769         struct rpc_buffer *buf;
 770         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 771
 772         size += sizeof(struct rpc_buffer);
 773         if (size <= RPC_BUFFER_MAXSIZE)
 774                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 775         else
 776                 buf = kmalloc(size, gfp);
 777
 778         if (!buf)
 779                 return NULL;
 780
 781         buf->len = size;
 782         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 783                         task->tk_pid, size, buf);
 784         return &buf->data;
 785 }
 786
 787 /**
 788  * rpc_free - free buffer allocated via rpc_malloc
 789  * @buffer: buffer to free
 790  *
 791  */
 792 void rpc_free(void *buffer)
 793 {
 794         size_t size;
 795         struct rpc_buffer *buf;
 796
 797         if (!buffer)
 798                 return;
 799
 800         buf = container_of(buffer, struct rpc_buffer, data);
 801         size = buf->len;
 802
 803         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 804                         size, buf);
 805
 806         if (size <= RPC_BUFFER_MAXSIZE)
 807                 mempool_free(buf, rpc_buffer_mempool);
 808         else
 809                 kfree(buf);
 810 }
 811
 812 /*
 813  * Creation and deletion of RPC task structures
 814  */
 815 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 816 {
 817         memset(task, 0, sizeof(*task));
 818         init_timer(&task->tk_timer);
 819         task->tk_timer.data     = (unsigned long) task;
 820         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 821         atomic_set(&task->tk_count, 1);
 822         task->tk_client = clnt;
 823         task->tk_flags  = flags;
 824         task->tk_ops = tk_ops;
 825         if (tk_ops->rpc_call_prepare != NULL)
 826                 task->tk_action = rpc_prepare_task;
 827         task->tk_calldata = calldata;
 828         INIT_LIST_HEAD(&task->tk_task);
 829
 830         /* Initialize retry counters */
 831         task->tk_garb_retry = 2;
 832         task->tk_cred_retry = 2;
 833
 834         task->tk_priority = RPC_PRIORITY_NORMAL;
 835         task->tk_cookie = (unsigned long)current;
 836
 837         /* Initialize workqueue for async tasks */
 838         task->tk_workqueue = rpciod_workqueue;
 839
 840         if (clnt) {
 841                 atomic_inc(&clnt->cl_users);
 842                 if (clnt->cl_softrtry)
 843                         task->tk_flags |= RPC_TASK_SOFT;
 844                 if (!clnt->cl_intr)
 845                         task->tk_flags |= RPC_TASK_NOINTR;
 846         }
 847
 848         BUG_ON(task->tk_ops == NULL);
 849
 850         /* starting timestamp */
 851         task->tk_start = jiffies;
 852
 853         dprintk("RPC:       new task initialized, procpid %u\n",
 854                                 current->pid);
 855 }
 856
 857 static struct rpc_task *
 858 rpc_alloc_task(void)
 859 {
 860         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 861 }
 862
 863 static void rpc_free_task(struct rcu_head *rcu)
 864 {
 865         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 866         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 867         mempool_free(task, rpc_task_mempool);
 868 }
 869
 870 /*
 871  * Create a new task for the specified client.  We have to
 872  * clean up after an allocation failure, as the client may
 873  * have specified "oneshot".
 874  */
 875 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 876 {
 877         struct rpc_task *task;
 878
 879         task = rpc_alloc_task();
 880         if (!task)
 881                 goto cleanup;
 882
 883         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 884
 885         dprintk("RPC:       allocated task %p\n", task);
 886         task->tk_flags |= RPC_TASK_DYNAMIC;
 887 out:
 888         return task;
 889
 890 cleanup:
 891         /* Check whether to release the client */
 892         if (clnt) {
 893                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 894                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 895                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 896                 rpc_release_client(clnt);
 897         }
 898         goto out;
 899 }
 900
 901
 902 void rpc_put_task(struct rpc_task *task)
 903 {
 904         const struct rpc_call_ops *tk_ops = task->tk_ops;
 905         void *calldata = task->tk_calldata;
 906
 907         if (!atomic_dec_and_test(&task->tk_count))
 908                 return;
 909         /* Release resources */
 910         if (task->tk_rqstp)
 911                 xprt_release(task);
 912         if (task->tk_msg.rpc_cred)
 913                 rpcauth_unbindcred(task);
 914         if (task->tk_client) {
 915                 rpc_release_client(task->tk_client);
 916                 task->tk_client = NULL;
 917         }
 918         if (task->tk_flags & RPC_TASK_DYNAMIC)
 919                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task);
 920         rpc_release_calldata(tk_ops, calldata);
 921 }
 922 EXPORT_SYMBOL(rpc_put_task);
 923
 924 static void rpc_release_task(struct rpc_task *task)
 925 {
 926 #ifdef RPC_DEBUG
 927         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 928 #endif
 929         dprintk("RPC: %5u release task\n", task->tk_pid);
 930
 931         if (!list_empty(&task->tk_task)) {
 932                 struct rpc_clnt *clnt = task->tk_client;
 933                 /* Remove from client task list */
 934                 spin_lock(&clnt->cl_lock);
 935                 list_del(&task->tk_task);
 936                 spin_unlock(&clnt->cl_lock);
 937         }
 938         BUG_ON (RPC_IS_QUEUED(task));
 939
 940         /* Synchronously delete any running timer */
 941         rpc_delete_timer(task);
 942
 943 #ifdef RPC_DEBUG
 944         task->tk_magic = 0;
 945 #endif
 946         /* Wake up anyone who is waiting for task completion */
 947         rpc_mark_complete_task(task);
 948
 949         rpc_put_task(task);
 950 }
 951
 952 /**
 953  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
 954  * @clnt: pointer to RPC client
 955  * @flags: RPC flags
 956  * @ops: RPC call ops
 957  * @data: user call data
 958  */
 959 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
 960                                         const struct rpc_call_ops *ops,
 961                                         void *data)
 962 {
 963         struct rpc_task *task;
 964         task = rpc_new_task(clnt, flags, ops, data);
 965         if (task == NULL) {
 966                 rpc_release_calldata(ops, data);
 967                 return ERR_PTR(-ENOMEM);
 968         }
 969         atomic_inc(&task->tk_count);
 970         rpc_execute(task);
 971         return task;
 972 }
 973 EXPORT_SYMBOL(rpc_run_task);
 974
 975 /*
 976  * Kill all tasks for the given client.
 977  * XXX: kill their descendants as well?
 978  */
 979 void rpc_killall_tasks(struct rpc_clnt *clnt)
 980 {
 981         struct rpc_task *rovr;
 982
 983
 984         if (list_empty(&clnt->cl_tasks))
 985                 return;
 986         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 987         /*
 988          * Spin lock all_tasks to prevent changes...
 989          */
 990         spin_lock(&clnt->cl_lock);
 991         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 992                 if (! RPC_IS_ACTIVATED(rovr))
 993                         continue;
 994                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 995                         rovr->tk_flags |= RPC_TASK_KILLED;
 996                         rpc_exit(rovr, -EIO);
 997                         rpc_wake_up_task(rovr);
 998                 }
 999         }
1000         spin_unlock(&clnt->cl_lock);
1001 }
1002
1003 static void rpciod_killall(void)
1004 {
1005         struct rpc_clnt *clnt;
1006         unsigned long flags;
1007
1008         for(;;) {
1009                 clear_thread_flag(TIF_SIGPENDING);
1010
1011                 spin_lock(&rpc_sched_lock);
1012                 list_for_each_entry(clnt, &all_clients, cl_clients)
1013                         rpc_killall_tasks(clnt);
1014                 spin_unlock(&rpc_sched_lock);
1015                 flush_workqueue(rpciod_workqueue);
1016                 if (!list_empty(&all_clients))
1017                         break;
1018                 dprintk("RPC:       rpciod_killall: waiting for tasks "
1019                                         "to exit\n");
1020                 wait_event_timeout(client_kill_wait,
1021                                 list_empty(&all_clients), 1*HZ);
1022         }
1023
1024         spin_lock_irqsave(&current->sighand->siglock, flags);
1025         recalc_sigpending();
1026         spin_unlock_irqrestore(&current->sighand->siglock, flags);
1027 }
1028
1029 void rpc_register_client(struct rpc_clnt *clnt)
1030 {
1031         spin_lock(&rpc_sched_lock);
1032         list_add(&clnt->cl_clients, &all_clients);
1033         spin_unlock(&rpc_sched_lock);
1034 }
1035
1036 void rpc_unregister_client(struct rpc_clnt *clnt)
1037 {
1038         spin_lock(&rpc_sched_lock);
1039         list_del(&clnt->cl_clients);
1040         if (list_empty(&all_clients))
1041                 wake_up(&client_kill_wait);
1042         spin_unlock(&rpc_sched_lock);
1043 }
1044
1045 /*
1046  * Start up the rpciod process if it's not already running.
1047  */
1048 int
1049 rpciod_up(void)
1050 {
1051         struct workqueue_struct *wq;
1052         int error = 0;
1053
1054         mutex_lock(&rpciod_mutex);
1055         dprintk("RPC:       rpciod_up: users %u\n", rpciod_users);
1056         rpciod_users++;
1057         if (rpciod_workqueue)
1058                 goto out;
1059         /*
1060          * If there's no pid, we should be the first user.
1061          */
1062         if (rpciod_users > 1)
1063                 printk(KERN_WARNING "rpciod_up: no workqueue, %u users??\n", rpciod_users);
1064         /*
1065          * Create the rpciod thread and wait for it to start.
1066          */
1067         error = -ENOMEM;
1068         wq = create_workqueue("rpciod");
1069         if (wq == NULL) {
1070                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1071                 rpciod_users--;
1072                 goto out;
1073         }
1074         rpciod_workqueue = wq;
1075         error = 0;
1076 out:
1077         mutex_unlock(&rpciod_mutex);
1078         return error;
1079 }
1080
1081 void
1082 rpciod_down(void)
1083 {
1084         mutex_lock(&rpciod_mutex);
1085         dprintk("RPC:       rpciod_down sema %u\n", rpciod_users);
1086         if (rpciod_users) {
1087                 if (--rpciod_users)
1088                         goto out;
1089         } else
1090                 printk(KERN_WARNING "rpciod_down: no users??\n");
1091
1092         if (!rpciod_workqueue) {
1093                 dprintk("RPC:       rpciod_down: Nothing to do!\n");
1094                 goto out;
1095         }
1096         rpciod_killall();
1097
1098         destroy_workqueue(rpciod_workqueue);
1099         rpciod_workqueue = NULL;
1100  out:
1101         mutex_unlock(&rpciod_mutex);
1102 }
1103
1104 #ifdef RPC_DEBUG
1105 void rpc_show_tasks(void)
1106 {
1107         struct rpc_clnt *clnt;
1108         struct rpc_task *t;
1109
1110         spin_lock(&rpc_sched_lock);
1111         if (list_empty(&all_clients))
1112                 goto out;
1113         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1114                 "-rpcwait -action- ---ops--\n");
1115         list_for_each_entry(clnt, &all_clients, cl_clients) {
1116                 if (list_empty(&clnt->cl_tasks))
1117                         continue;
1118                 spin_lock(&clnt->cl_lock);
1119                 list_for_each_entry(t, &clnt->cl_tasks, tk_task) {
1120                         const char *rpc_waitq = "none";
1121
1122                         if (RPC_IS_QUEUED(t))
1123                                 rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1124
1125                         printk("%5u %04d %04x %6d %8p %6d %8p %8ld %8s %8p %8p\n",
1126                                 t->tk_pid,
1127                                 (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1128                                 t->tk_flags, t->tk_status,
1129                                 t->tk_client,
1130                                 (t->tk_client ? t->tk_client->cl_prog : 0),
1131                                 t->tk_rqstp, t->tk_timeout,
1132                                 rpc_waitq,
1133                                 t->tk_action, t->tk_ops);
1134                 }
1135                 spin_unlock(&clnt->cl_lock);
1136         }
1137 out:
1138         spin_unlock(&rpc_sched_lock);
1139 }
1140 #endif
1141
1142 void
1143 rpc_destroy_mempool(void)
1144 {
1145         if (rpc_buffer_mempool)
1146                 mempool_destroy(rpc_buffer_mempool);
1147         if (rpc_task_mempool)
1148                 mempool_destroy(rpc_task_mempool);
1149         if (rpc_task_slabp)
1150                 kmem_cache_destroy(rpc_task_slabp);
1151         if (rpc_buffer_slabp)
1152                 kmem_cache_destroy(rpc_buffer_slabp);
1153 }
1154
1155 int
1156 rpc_init_mempool(void)
1157 {
1158         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1159                                              sizeof(struct rpc_task),
1160                                              0, SLAB_HWCACHE_ALIGN,
1161                                              NULL, NULL);
1162         if (!rpc_task_slabp)
1163                 goto err_nomem;
1164         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1165                                              RPC_BUFFER_MAXSIZE,
1166                                              0, SLAB_HWCACHE_ALIGN,
1167                                              NULL, NULL);
1168         if (!rpc_buffer_slabp)
1169                 goto err_nomem;
1170         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1171                                                     rpc_task_slabp);
1172         if (!rpc_task_mempool)
1173                 goto err_nomem;
1174         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1175                                                       rpc_buffer_slabp);
1176         if (!rpc_buffer_mempool)
1177                 goto err_nomem;
1178         return 0;
1179 err_nomem:
1180         rpc_destroy_mempool();
1181         return -ENOMEM;
1182 }