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 #include <linux/mutex.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #ifdef RPC_DEBUG
  26 #define RPCDBG_FACILITY         RPCDBG_SCHED
  27 #define RPC_TASK_MAGIC_ID       0xf00baa
  28 #endif
  29
  30 /*
  31  * RPC slabs and memory pools
  32  */
  33 #define RPC_BUFFER_MAXSIZE      (2048)
  34 #define RPC_BUFFER_POOLSIZE     (8)
  35 #define RPC_TASK_POOLSIZE       (8)
  36 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  37 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  38 static mempool_t        *rpc_task_mempool __read_mostly;
  39 static mempool_t        *rpc_buffer_mempool __read_mostly;
  40
  41 static void                     __rpc_default_timer(struct rpc_task *task);
  42 static void                     rpc_async_schedule(struct work_struct *);
  43 static void                      rpc_release_task(struct rpc_task *task);
  44
  45 /*
  46  * RPC tasks sit here while waiting for conditions to improve.
  47  */
  48 static struct rpc_wait_queue delay_queue;
  49
  50 /*
  51  * rpciod-related stuff
  52  */
  53 struct workqueue_struct *rpciod_workqueue;
  54
  55 /*
  56  * Disable the timer for a given RPC task. Should be called with
  57  * queue->lock and bh_disabled in order to avoid races within
  58  * rpc_run_timer().
  59  */
  60 static inline void
  61 __rpc_disable_timer(struct rpc_task *task)
  62 {
  63         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  64         task->tk_timeout_fn = NULL;
  65         task->tk_timeout = 0;
  66 }
  67
  68 /*
  69  * Run a timeout function.
  70  * We use the callback in order to allow __rpc_wake_up_task()
  71  * and friends to disable the timer synchronously on SMP systems
  72  * without calling del_timer_sync(). The latter could cause a
  73  * deadlock if called while we're holding spinlocks...
  74  */
  75 static void rpc_run_timer(struct rpc_task *task)
  76 {
  77         void (*callback)(struct rpc_task *);
  78
  79         callback = task->tk_timeout_fn;
  80         task->tk_timeout_fn = NULL;
  81         if (callback && RPC_IS_QUEUED(task)) {
  82                 dprintk("RPC: %5u running timer\n", task->tk_pid);
  83                 callback(task);
  84         }
  85         smp_mb__before_clear_bit();
  86         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
  87         smp_mb__after_clear_bit();
  88 }
  89
  90 /*
  91  * Set up a timer for the current task.
  92  */
  93 static inline void
  94 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
  95 {
  96         if (!task->tk_timeout)
  97                 return;
  98
  99         dprintk("RPC: %5u setting alarm for %lu ms\n",
 100                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 101
 102         if (timer)
 103                 task->tk_timeout_fn = timer;
 104         else
 105                 task->tk_timeout_fn = __rpc_default_timer;
 106         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 107         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 108 }
 109
 110 /*
 111  * Delete any timer for the current task. Because we use del_timer_sync(),
 112  * this function should never be called while holding queue->lock.
 113  */
 114 static void
 115 rpc_delete_timer(struct rpc_task *task)
 116 {
 117         if (RPC_IS_QUEUED(task))
 118                 return;
 119         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 120                 del_singleshot_timer_sync(&task->tk_timer);
 121                 dprintk("RPC: %5u deleting timer\n", task->tk_pid);
 122         }
 123 }
 124
 125 /*
 126  * Add new request to a priority queue.
 127  */
 128 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 129 {
 130         struct list_head *q;
 131         struct rpc_task *t;
 132
 133         INIT_LIST_HEAD(&task->u.tk_wait.links);
 134         q = &queue->tasks[task->tk_priority];
 135         if (unlikely(task->tk_priority > queue->maxpriority))
 136                 q = &queue->tasks[queue->maxpriority];
 137         list_for_each_entry(t, q, u.tk_wait.list) {
 138                 if (t->tk_owner == task->tk_owner) {
 139                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 140                         return;
 141                 }
 142         }
 143         list_add_tail(&task->u.tk_wait.list, q);
 144 }
 145
 146 /*
 147  * Add new request to wait queue.
 148  *
 149  * Swapper tasks always get inserted at the head of the queue.
 150  * This should avoid many nasty memory deadlocks and hopefully
 151  * improve overall performance.
 152  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 153  */
 154 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 155 {
 156         BUG_ON (RPC_IS_QUEUED(task));
 157
 158         if (RPC_IS_PRIORITY(queue))
 159                 __rpc_add_wait_queue_priority(queue, task);
 160         else if (RPC_IS_SWAPPER(task))
 161                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 162         else
 163                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 164         task->u.tk_wait.rpc_waitq = queue;
 165         queue->qlen++;
 166         rpc_set_queued(task);
 167
 168         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 169                         task->tk_pid, queue, rpc_qname(queue));
 170 }
 171
 172 /*
 173  * Remove request from a priority queue.
 174  */
 175 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 176 {
 177         struct rpc_task *t;
 178
 179         if (!list_empty(&task->u.tk_wait.links)) {
 180                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 181                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 182                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 183         }
 184         list_del(&task->u.tk_wait.list);
 185 }
 186
 187 /*
 188  * Remove request from queue.
 189  * Note: must be called with spin lock held.
 190  */
 191 static void __rpc_remove_wait_queue(struct rpc_task *task)
 192 {
 193         struct rpc_wait_queue *queue;
 194         queue = task->u.tk_wait.rpc_waitq;
 195
 196         if (RPC_IS_PRIORITY(queue))
 197                 __rpc_remove_wait_queue_priority(task);
 198         else
 199                 list_del(&task->u.tk_wait.list);
 200         queue->qlen--;
 201         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 202                         task->tk_pid, queue, rpc_qname(queue));
 203 }
 204
 205 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 206 {
 207         queue->priority = priority;
 208         queue->count = 1 << (priority * 2);
 209 }
 210
 211 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 212 {
 213         queue->owner = pid;
 214         queue->nr = RPC_BATCH_COUNT;
 215 }
 216
 217 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 218 {
 219         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 220         rpc_set_waitqueue_owner(queue, 0);
 221 }
 222
 223 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 224 {
 225         int i;
 226
 227         spin_lock_init(&queue->lock);
 228         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 229                 INIT_LIST_HEAD(&queue->tasks[i]);
 230         queue->maxpriority = nr_queues - 1;
 231         rpc_reset_waitqueue_priority(queue);
 232 #ifdef RPC_DEBUG
 233         queue->name = qname;
 234 #endif
 235 }
 236
 237 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 238 {
 239         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 240 }
 241
 242 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 243 {
 244         __rpc_init_priority_wait_queue(queue, qname, 1);
 245 }
 246 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 247
 248 static int rpc_wait_bit_interruptible(void *word)
 249 {
 250         if (signal_pending(current))
 251                 return -ERESTARTSYS;
 252         schedule();
 253         return 0;
 254 }
 255
 256 #ifdef RPC_DEBUG
 257 static void rpc_task_set_debuginfo(struct rpc_task *task)
 258 {
 259         static atomic_t rpc_pid;
 260
 261         task->tk_magic = RPC_TASK_MAGIC_ID;
 262         task->tk_pid = atomic_inc_return(&rpc_pid);
 263 }
 264 #else
 265 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 266 {
 267 }
 268 #endif
 269
 270 static void rpc_set_active(struct rpc_task *task)
 271 {
 272         struct rpc_clnt *clnt;
 273         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 274                 return;
 275         rpc_task_set_debuginfo(task);
 276         /* Add to global list of all tasks */
 277         clnt = task->tk_client;
 278         if (clnt != NULL) {
 279                 spin_lock(&clnt->cl_lock);
 280                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 281                 spin_unlock(&clnt->cl_lock);
 282         }
 283 }
 284
 285 /*
 286  * Mark an RPC call as having completed by clearing the 'active' bit
 287  */
 288 static void rpc_mark_complete_task(struct rpc_task *task)
 289 {
 290         smp_mb__before_clear_bit();
 291         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 292         smp_mb__after_clear_bit();
 293         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 294 }
 295
 296 /*
 297  * Allow callers to wait for completion of an RPC call
 298  */
 299 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 300 {
 301         if (action == NULL)
 302                 action = rpc_wait_bit_interruptible;
 303         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 304                         action, TASK_INTERRUPTIBLE);
 305 }
 306 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 307
 308 /*
 309  * Make an RPC task runnable.
 310  *
 311  * Note: If the task is ASYNC, this must be called with
 312  * the spinlock held to protect the wait queue operation.
 313  */
 314 static void rpc_make_runnable(struct rpc_task *task)
 315 {
 316         BUG_ON(task->tk_timeout_fn);
 317         rpc_clear_queued(task);
 318         if (rpc_test_and_set_running(task))
 319                 return;
 320         /* We might have raced */
 321         if (RPC_IS_QUEUED(task)) {
 322                 rpc_clear_running(task);
 323                 return;
 324         }
 325         if (RPC_IS_ASYNC(task)) {
 326                 int status;
 327
 328                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 329                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 330                 if (status < 0) {
 331                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 332                         task->tk_status = status;
 333                         return;
 334                 }
 335         } else
 336                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 337 }
 338
 339 /*
 340  * Prepare for sleeping on a wait queue.
 341  * By always appending tasks to the list we ensure FIFO behavior.
 342  * NB: An RPC task will only receive interrupt-driven events as long
 343  * as it's on a wait queue.
 344  */
 345 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 346                         rpc_action action, rpc_action timer)
 347 {
 348         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 349                         task->tk_pid, rpc_qname(q), jiffies);
 350
 351         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 352                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 353                 return;
 354         }
 355
 356         __rpc_add_wait_queue(q, task);
 357
 358         BUG_ON(task->tk_callback != NULL);
 359         task->tk_callback = action;
 360         __rpc_add_timer(task, timer);
 361 }
 362
 363 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 364                                 rpc_action action, rpc_action timer)
 365 {
 366         /* Mark the task as being activated if so needed */
 367         rpc_set_active(task);
 368
 369         /*
 370          * Protect the queue operations.
 371          */
 372         spin_lock_bh(&q->lock);
 373         __rpc_sleep_on(q, task, action, timer);
 374         spin_unlock_bh(&q->lock);
 375 }
 376 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 377
 378 /**
 379  * __rpc_do_wake_up_task - wake up a single rpc_task
 380  * @task: task to be woken up
 381  *
 382  * Caller must hold queue->lock, and have cleared the task queued flag.
 383  */
 384 static void __rpc_do_wake_up_task(struct rpc_task *task)
 385 {
 386         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 387                         task->tk_pid, jiffies);
 388
 389 #ifdef RPC_DEBUG
 390         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 391 #endif
 392         /* Has the task been executed yet? If not, we cannot wake it up! */
 393         if (!RPC_IS_ACTIVATED(task)) {
 394                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 395                 return;
 396         }
 397
 398         __rpc_disable_timer(task);
 399         __rpc_remove_wait_queue(task);
 400
 401         rpc_make_runnable(task);
 402
 403         dprintk("RPC:       __rpc_wake_up_task done\n");
 404 }
 405
 406 /*
 407  * Wake up the specified task
 408  */
 409 static void __rpc_wake_up_task(struct rpc_task *task)
 410 {
 411         if (rpc_start_wakeup(task)) {
 412                 if (RPC_IS_QUEUED(task))
 413                         __rpc_do_wake_up_task(task);
 414                 rpc_finish_wakeup(task);
 415         }
 416 }
 417
 418 /*
 419  * Default timeout handler if none specified by user
 420  */
 421 static void
 422 __rpc_default_timer(struct rpc_task *task)
 423 {
 424         dprintk("RPC: %5u timeout (default timer)\n", task->tk_pid);
 425         task->tk_status = -ETIMEDOUT;
 426         rpc_wake_up_task(task);
 427 }
 428
 429 /*
 430  * Wake up the specified task
 431  */
 432 void rpc_wake_up_task(struct rpc_task *task)
 433 {
 434         rcu_read_lock_bh();
 435         if (rpc_start_wakeup(task)) {
 436                 if (RPC_IS_QUEUED(task)) {
 437                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 438
 439                         /* Note: we're already in a bh-safe context */
 440                         spin_lock(&queue->lock);
 441                         __rpc_do_wake_up_task(task);
 442                         spin_unlock(&queue->lock);
 443                 }
 444                 rpc_finish_wakeup(task);
 445         }
 446         rcu_read_unlock_bh();
 447 }
 448 EXPORT_SYMBOL_GPL(rpc_wake_up_task);
 449
 450 /*
 451  * Wake up the next task on a priority queue.
 452  */
 453 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 454 {
 455         struct list_head *q;
 456         struct rpc_task *task;
 457
 458         /*
 459          * Service a batch of tasks from a single owner.
 460          */
 461         q = &queue->tasks[queue->priority];
 462         if (!list_empty(q)) {
 463                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 464                 if (queue->owner == task->tk_owner) {
 465                         if (--queue->nr)
 466                                 goto out;
 467                         list_move_tail(&task->u.tk_wait.list, q);
 468                 }
 469                 /*
 470                  * Check if we need to switch queues.
 471                  */
 472                 if (--queue->count)
 473                         goto new_owner;
 474         }
 475
 476         /*
 477          * Service the next queue.
 478          */
 479         do {
 480                 if (q == &queue->tasks[0])
 481                         q = &queue->tasks[queue->maxpriority];
 482                 else
 483                         q = q - 1;
 484                 if (!list_empty(q)) {
 485                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 486                         goto new_queue;
 487                 }
 488         } while (q != &queue->tasks[queue->priority]);
 489
 490         rpc_reset_waitqueue_priority(queue);
 491         return NULL;
 492
 493 new_queue:
 494         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 495 new_owner:
 496         rpc_set_waitqueue_owner(queue, task->tk_owner);
 497 out:
 498         __rpc_wake_up_task(task);
 499         return task;
 500 }
 501
 502 /*
 503  * Wake up the next task on the wait queue.
 504  */
 505 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 506 {
 507         struct rpc_task *task = NULL;
 508
 509         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 510                         queue, rpc_qname(queue));
 511         rcu_read_lock_bh();
 512         spin_lock(&queue->lock);
 513         if (RPC_IS_PRIORITY(queue))
 514                 task = __rpc_wake_up_next_priority(queue);
 515         else {
 516                 task_for_first(task, &queue->tasks[0])
 517                         __rpc_wake_up_task(task);
 518         }
 519         spin_unlock(&queue->lock);
 520         rcu_read_unlock_bh();
 521
 522         return task;
 523 }
 524 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 525
 526 /**
 527  * rpc_wake_up - wake up all rpc_tasks
 528  * @queue: rpc_wait_queue on which the tasks are sleeping
 529  *
 530  * Grabs queue->lock
 531  */
 532 void rpc_wake_up(struct rpc_wait_queue *queue)
 533 {
 534         struct rpc_task *task, *next;
 535         struct list_head *head;
 536
 537         rcu_read_lock_bh();
 538         spin_lock(&queue->lock);
 539         head = &queue->tasks[queue->maxpriority];
 540         for (;;) {
 541                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 542                         __rpc_wake_up_task(task);
 543                 if (head == &queue->tasks[0])
 544                         break;
 545                 head--;
 546         }
 547         spin_unlock(&queue->lock);
 548         rcu_read_unlock_bh();
 549 }
 550 EXPORT_SYMBOL_GPL(rpc_wake_up);
 551
 552 /**
 553  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 554  * @queue: rpc_wait_queue on which the tasks are sleeping
 555  * @status: status value to set
 556  *
 557  * Grabs queue->lock
 558  */
 559 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 560 {
 561         struct rpc_task *task, *next;
 562         struct list_head *head;
 563
 564         rcu_read_lock_bh();
 565         spin_lock(&queue->lock);
 566         head = &queue->tasks[queue->maxpriority];
 567         for (;;) {
 568                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 569                         task->tk_status = status;
 570                         __rpc_wake_up_task(task);
 571                 }
 572                 if (head == &queue->tasks[0])
 573                         break;
 574                 head--;
 575         }
 576         spin_unlock(&queue->lock);
 577         rcu_read_unlock_bh();
 578 }
 579 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 580
 581 static void __rpc_atrun(struct rpc_task *task)
 582 {
 583         rpc_wake_up_task(task);
 584 }
 585
 586 /*
 587  * Run a task at a later time
 588  */
 589 void rpc_delay(struct rpc_task *task, unsigned long delay)
 590 {
 591         task->tk_timeout = delay;
 592         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 593 }
 594 EXPORT_SYMBOL_GPL(rpc_delay);
 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_GPL(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 EXPORT_SYMBOL_GPL(rpc_malloc);
 787
 788 /**
 789  * rpc_free - free buffer allocated via rpc_malloc
 790  * @buffer: buffer to free
 791  *
 792  */
 793 void rpc_free(void *buffer)
 794 {
 795         size_t size;
 796         struct rpc_buffer *buf;
 797
 798         if (!buffer)
 799                 return;
 800
 801         buf = container_of(buffer, struct rpc_buffer, data);
 802         size = buf->len;
 803
 804         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 805                         size, buf);
 806
 807         if (size <= RPC_BUFFER_MAXSIZE)
 808                 mempool_free(buf, rpc_buffer_mempool);
 809         else
 810                 kfree(buf);
 811 }
 812 EXPORT_SYMBOL_GPL(rpc_free);
 813
 814 /*
 815  * Creation and deletion of RPC task structures
 816  */
 817 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 818 {
 819         memset(task, 0, sizeof(*task));
 820         setup_timer(&task->tk_timer, (void (*)(unsigned long))rpc_run_timer,
 821                         (unsigned long)task);
 822         atomic_set(&task->tk_count, 1);
 823         task->tk_flags  = task_setup_data->flags;
 824         task->tk_ops = task_setup_data->callback_ops;
 825         task->tk_calldata = task_setup_data->callback_data;
 826         INIT_LIST_HEAD(&task->tk_task);
 827
 828         /* Initialize retry counters */
 829         task->tk_garb_retry = 2;
 830         task->tk_cred_retry = 2;
 831
 832         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 833         task->tk_owner = current->tgid;
 834
 835         /* Initialize workqueue for async tasks */
 836         task->tk_workqueue = rpciod_workqueue;
 837
 838         task->tk_client = task_setup_data->rpc_client;
 839         if (task->tk_client != NULL) {
 840                 kref_get(&task->tk_client->cl_kref);
 841                 if (task->tk_client->cl_softrtry)
 842                         task->tk_flags |= RPC_TASK_SOFT;
 843                 if (!task->tk_client->cl_intr)
 844                         task->tk_flags |= RPC_TASK_NOINTR;
 845         }
 846
 847         if (task->tk_ops->rpc_call_prepare != NULL)
 848                 task->tk_action = rpc_prepare_task;
 849
 850         if (task_setup_data->rpc_message != NULL) {
 851                 memcpy(&task->tk_msg, task_setup_data->rpc_message, sizeof(task->tk_msg));
 852                 /* Bind the user cred */
 853                 if (task->tk_msg.rpc_cred != NULL)
 854                         rpcauth_holdcred(task);
 855                 else
 856                         rpcauth_bindcred(task);
 857                 if (task->tk_action == NULL)
 858                         rpc_call_start(task);
 859         }
 860
 861         /* starting timestamp */
 862         task->tk_start = jiffies;
 863
 864         dprintk("RPC:       new task initialized, procpid %u\n",
 865                                 task_pid_nr(current));
 866 }
 867
 868 static struct rpc_task *
 869 rpc_alloc_task(void)
 870 {
 871         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 872 }
 873
 874 static void rpc_free_task(struct rcu_head *rcu)
 875 {
 876         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 877         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 878         mempool_free(task, rpc_task_mempool);
 879 }
 880
 881 /*
 882  * Create a new task for the specified client.
 883  */
 884 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 885 {
 886         struct rpc_task *task = setup_data->task;
 887         unsigned short flags = 0;
 888
 889         if (task == NULL) {
 890                 task = rpc_alloc_task();
 891                 if (task == NULL)
 892                         goto out;
 893                 flags = RPC_TASK_DYNAMIC;
 894         }
 895
 896         rpc_init_task(task, setup_data);
 897
 898         task->tk_flags |= flags;
 899         dprintk("RPC:       allocated task %p\n", task);
 900 out:
 901         return task;
 902 }
 903
 904
 905 void rpc_put_task(struct rpc_task *task)
 906 {
 907         const struct rpc_call_ops *tk_ops = task->tk_ops;
 908         void *calldata = task->tk_calldata;
 909
 910         if (!atomic_dec_and_test(&task->tk_count))
 911                 return;
 912         /* Release resources */
 913         if (task->tk_rqstp)
 914                 xprt_release(task);
 915         if (task->tk_msg.rpc_cred)
 916                 rpcauth_unbindcred(task);
 917         if (task->tk_client) {
 918                 rpc_release_client(task->tk_client);
 919                 task->tk_client = NULL;
 920         }
 921         if (task->tk_flags & RPC_TASK_DYNAMIC)
 922                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task);
 923         rpc_release_calldata(tk_ops, calldata);
 924 }
 925 EXPORT_SYMBOL_GPL(rpc_put_task);
 926
 927 static void rpc_release_task(struct rpc_task *task)
 928 {
 929 #ifdef RPC_DEBUG
 930         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 931 #endif
 932         dprintk("RPC: %5u release task\n", task->tk_pid);
 933
 934         if (!list_empty(&task->tk_task)) {
 935                 struct rpc_clnt *clnt = task->tk_client;
 936                 /* Remove from client task list */
 937                 spin_lock(&clnt->cl_lock);
 938                 list_del(&task->tk_task);
 939                 spin_unlock(&clnt->cl_lock);
 940         }
 941         BUG_ON (RPC_IS_QUEUED(task));
 942
 943         /* Synchronously delete any running timer */
 944         rpc_delete_timer(task);
 945
 946 #ifdef RPC_DEBUG
 947         task->tk_magic = 0;
 948 #endif
 949         /* Wake up anyone who is waiting for task completion */
 950         rpc_mark_complete_task(task);
 951
 952         rpc_put_task(task);
 953 }
 954
 955 /*
 956  * Kill all tasks for the given client.
 957  * XXX: kill their descendants as well?
 958  */
 959 void rpc_killall_tasks(struct rpc_clnt *clnt)
 960 {
 961         struct rpc_task *rovr;
 962
 963
 964         if (list_empty(&clnt->cl_tasks))
 965                 return;
 966         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 967         /*
 968          * Spin lock all_tasks to prevent changes...
 969          */
 970         spin_lock(&clnt->cl_lock);
 971         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 972                 if (! RPC_IS_ACTIVATED(rovr))
 973                         continue;
 974                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 975                         rovr->tk_flags |= RPC_TASK_KILLED;
 976                         rpc_exit(rovr, -EIO);
 977                         rpc_wake_up_task(rovr);
 978                 }
 979         }
 980         spin_unlock(&clnt->cl_lock);
 981 }
 982 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 983
 984 int rpciod_up(void)
 985 {
 986         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 987 }
 988
 989 void rpciod_down(void)
 990 {
 991         module_put(THIS_MODULE);
 992 }
 993
 994 /*
 995  * Start up the rpciod workqueue.
 996  */
 997 static int rpciod_start(void)
 998 {
 999         struct workqueue_struct *wq;
1000
1001         /*
1002          * Create the rpciod thread and wait for it to start.
1003          */
1004         dprintk("RPC:       creating workqueue rpciod\n");
1005         wq = create_workqueue("rpciod");
1006         rpciod_workqueue = wq;
1007         return rpciod_workqueue != NULL;
1008 }
1009
1010 static void rpciod_stop(void)
1011 {
1012         struct workqueue_struct *wq = NULL;
1013
1014         if (rpciod_workqueue == NULL)
1015                 return;
1016         dprintk("RPC:       destroying workqueue rpciod\n");
1017
1018         wq = rpciod_workqueue;
1019         rpciod_workqueue = NULL;
1020         destroy_workqueue(wq);
1021 }
1022
1023 void
1024 rpc_destroy_mempool(void)
1025 {
1026         rpciod_stop();
1027         if (rpc_buffer_mempool)
1028                 mempool_destroy(rpc_buffer_mempool);
1029         if (rpc_task_mempool)
1030                 mempool_destroy(rpc_task_mempool);
1031         if (rpc_task_slabp)
1032                 kmem_cache_destroy(rpc_task_slabp);
1033         if (rpc_buffer_slabp)
1034                 kmem_cache_destroy(rpc_buffer_slabp);
1035 }
1036
1037 int
1038 rpc_init_mempool(void)
1039 {
1040         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1041                                              sizeof(struct rpc_task),
1042                                              0, SLAB_HWCACHE_ALIGN,
1043                                              NULL);
1044         if (!rpc_task_slabp)
1045                 goto err_nomem;
1046         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1047                                              RPC_BUFFER_MAXSIZE,
1048                                              0, SLAB_HWCACHE_ALIGN,
1049                                              NULL);
1050         if (!rpc_buffer_slabp)
1051                 goto err_nomem;
1052         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1053                                                     rpc_task_slabp);
1054         if (!rpc_task_mempool)
1055                 goto err_nomem;
1056         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1057                                                       rpc_buffer_slabp);
1058         if (!rpc_buffer_mempool)
1059                 goto err_nomem;
1060         if (!rpciod_start())
1061                 goto err_nomem;
1062         /*
1063          * The following is not strictly a mempool initialisation,
1064          * but there is no harm in doing it here
1065          */
1066         rpc_init_wait_queue(&delay_queue, "delayq");
1067         return 0;
1068 err_nomem:
1069         rpc_destroy_mempool();
1070         return -ENOMEM;
1071 }