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_async_schedule(struct work_struct *);
  42 static void                      rpc_release_task(struct rpc_task *task);
  43
  44 /*
  45  * RPC tasks sit here while waiting for conditions to improve.
  46  */
  47 static struct rpc_wait_queue delay_queue;
  48
  49 /*
  50  * rpciod-related stuff
  51  */
  52 struct workqueue_struct *rpciod_workqueue;
  53
  54 /*
  55  * Disable the timer for a given RPC task. Should be called with
  56  * queue->lock and bh_disabled in order to avoid races within
  57  * rpc_run_timer().
  58  */
  59 static inline void
  60 __rpc_disable_timer(struct rpc_task *task)
  61 {
  62         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  63         task->tk_timeout_fn = NULL;
  64         task->tk_timeout = 0;
  65 }
  66
  67 /*
  68  * Default timeout handler if none specified by user
  69  */
  70 static void
  71 __rpc_default_timer(struct rpc_task *task)
  72 {
  73         dprintk("RPC: %5u timeout (default timer)\n", task->tk_pid);
  74         task->tk_status = -ETIMEDOUT;
  75 }
  76
  77 /*
  78  * Set up a timer for the current task.
  79  */
  80 static inline void
  81 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
  82 {
  83         if (!task->tk_timeout)
  84                 return;
  85
  86         dprintk("RPC: %5u setting alarm for %lu ms\n",
  87                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  88
  89         if (timer)
  90                 task->tk_timeout_fn = timer;
  91         else
  92                 task->tk_timeout_fn = __rpc_default_timer;
  93         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
  94         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
  95 }
  96
  97 /*
  98  * Delete any timer for the current task. Because we use del_timer_sync(),
  99  * this function should never be called while holding queue->lock.
 100  */
 101 static void
 102 rpc_delete_timer(struct rpc_task *task)
 103 {
 104         if (RPC_IS_QUEUED(task))
 105                 return;
 106         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 107                 del_singleshot_timer_sync(&task->tk_timer);
 108                 dprintk("RPC: %5u deleting timer\n", task->tk_pid);
 109         }
 110 }
 111
 112 /*
 113  * Add new request to a priority queue.
 114  */
 115 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 116 {
 117         struct list_head *q;
 118         struct rpc_task *t;
 119
 120         INIT_LIST_HEAD(&task->u.tk_wait.links);
 121         q = &queue->tasks[task->tk_priority];
 122         if (unlikely(task->tk_priority > queue->maxpriority))
 123                 q = &queue->tasks[queue->maxpriority];
 124         list_for_each_entry(t, q, u.tk_wait.list) {
 125                 if (t->tk_owner == task->tk_owner) {
 126                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 127                         return;
 128                 }
 129         }
 130         list_add_tail(&task->u.tk_wait.list, q);
 131 }
 132
 133 /*
 134  * Add new request to wait queue.
 135  *
 136  * Swapper tasks always get inserted at the head of the queue.
 137  * This should avoid many nasty memory deadlocks and hopefully
 138  * improve overall performance.
 139  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 140  */
 141 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 142 {
 143         BUG_ON (RPC_IS_QUEUED(task));
 144
 145         if (RPC_IS_PRIORITY(queue))
 146                 __rpc_add_wait_queue_priority(queue, task);
 147         else if (RPC_IS_SWAPPER(task))
 148                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 149         else
 150                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 151         task->tk_waitqueue = queue;
 152         queue->qlen++;
 153         rpc_set_queued(task);
 154
 155         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 156                         task->tk_pid, queue, rpc_qname(queue));
 157 }
 158
 159 /*
 160  * Remove request from a priority queue.
 161  */
 162 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 163 {
 164         struct rpc_task *t;
 165
 166         if (!list_empty(&task->u.tk_wait.links)) {
 167                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 168                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 169                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 170         }
 171         list_del(&task->u.tk_wait.list);
 172 }
 173
 174 /*
 175  * Remove request from queue.
 176  * Note: must be called with spin lock held.
 177  */
 178 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 179 {
 180         if (RPC_IS_PRIORITY(queue))
 181                 __rpc_remove_wait_queue_priority(task);
 182         else
 183                 list_del(&task->u.tk_wait.list);
 184         queue->qlen--;
 185         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 186                         task->tk_pid, queue, rpc_qname(queue));
 187 }
 188
 189 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 190 {
 191         queue->priority = priority;
 192         queue->count = 1 << (priority * 2);
 193 }
 194
 195 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 196 {
 197         queue->owner = pid;
 198         queue->nr = RPC_BATCH_COUNT;
 199 }
 200
 201 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 202 {
 203         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 204         rpc_set_waitqueue_owner(queue, 0);
 205 }
 206
 207 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 208 {
 209         int i;
 210
 211         spin_lock_init(&queue->lock);
 212         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 213                 INIT_LIST_HEAD(&queue->tasks[i]);
 214         queue->maxpriority = nr_queues - 1;
 215         rpc_reset_waitqueue_priority(queue);
 216 #ifdef RPC_DEBUG
 217         queue->name = qname;
 218 #endif
 219 }
 220
 221 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 222 {
 223         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 224 }
 225
 226 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 227 {
 228         __rpc_init_priority_wait_queue(queue, qname, 1);
 229 }
 230 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 231
 232 static int rpc_wait_bit_killable(void *word)
 233 {
 234         if (fatal_signal_pending(current))
 235                 return -ERESTARTSYS;
 236         schedule();
 237         return 0;
 238 }
 239
 240 #ifdef RPC_DEBUG
 241 static void rpc_task_set_debuginfo(struct rpc_task *task)
 242 {
 243         static atomic_t rpc_pid;
 244
 245         task->tk_magic = RPC_TASK_MAGIC_ID;
 246         task->tk_pid = atomic_inc_return(&rpc_pid);
 247 }
 248 #else
 249 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 250 {
 251 }
 252 #endif
 253
 254 static void rpc_set_active(struct rpc_task *task)
 255 {
 256         struct rpc_clnt *clnt;
 257         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 258                 return;
 259         rpc_task_set_debuginfo(task);
 260         /* Add to global list of all tasks */
 261         clnt = task->tk_client;
 262         if (clnt != NULL) {
 263                 spin_lock(&clnt->cl_lock);
 264                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 265                 spin_unlock(&clnt->cl_lock);
 266         }
 267 }
 268
 269 /*
 270  * Mark an RPC call as having completed by clearing the 'active' bit
 271  */
 272 static void rpc_mark_complete_task(struct rpc_task *task)
 273 {
 274         smp_mb__before_clear_bit();
 275         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 276         smp_mb__after_clear_bit();
 277         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 278 }
 279
 280 /*
 281  * Allow callers to wait for completion of an RPC call
 282  */
 283 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 284 {
 285         if (action == NULL)
 286                 action = rpc_wait_bit_killable;
 287         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 288                         action, TASK_KILLABLE);
 289 }
 290 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 291
 292 /*
 293  * Make an RPC task runnable.
 294  *
 295  * Note: If the task is ASYNC, this must be called with
 296  * the spinlock held to protect the wait queue operation.
 297  */
 298 static void rpc_make_runnable(struct rpc_task *task)
 299 {
 300         BUG_ON(task->tk_timeout_fn);
 301         rpc_clear_queued(task);
 302         if (rpc_test_and_set_running(task))
 303                 return;
 304         /* We might have raced */
 305         if (RPC_IS_QUEUED(task)) {
 306                 rpc_clear_running(task);
 307                 return;
 308         }
 309         if (RPC_IS_ASYNC(task)) {
 310                 int status;
 311
 312                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 313                 status = queue_work(rpciod_workqueue, &task->u.tk_work);
 314                 if (status < 0) {
 315                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 316                         task->tk_status = status;
 317                         return;
 318                 }
 319         } else
 320                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 321 }
 322
 323 /*
 324  * Prepare for sleeping on a wait queue.
 325  * By always appending tasks to the list we ensure FIFO behavior.
 326  * NB: An RPC task will only receive interrupt-driven events as long
 327  * as it's on a wait queue.
 328  */
 329 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 330                         rpc_action action, rpc_action timer)
 331 {
 332         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 333                         task->tk_pid, rpc_qname(q), jiffies);
 334
 335         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 336                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 337                 return;
 338         }
 339
 340         __rpc_add_wait_queue(q, task);
 341
 342         BUG_ON(task->tk_callback != NULL);
 343         task->tk_callback = action;
 344         __rpc_add_timer(task, timer);
 345 }
 346
 347 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 348                                 rpc_action action, rpc_action timer)
 349 {
 350         /* Mark the task as being activated if so needed */
 351         rpc_set_active(task);
 352
 353         /*
 354          * Protect the queue operations.
 355          */
 356         spin_lock_bh(&q->lock);
 357         __rpc_sleep_on(q, task, action, timer);
 358         spin_unlock_bh(&q->lock);
 359 }
 360 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 361
 362 /**
 363  * __rpc_do_wake_up_task - wake up a single rpc_task
 364  * @queue: wait queue
 365  * @task: task to be woken up
 366  *
 367  * Caller must hold queue->lock, and have cleared the task queued flag.
 368  */
 369 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 370 {
 371         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 372                         task->tk_pid, jiffies);
 373
 374 #ifdef RPC_DEBUG
 375         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 376 #endif
 377         /* Has the task been executed yet? If not, we cannot wake it up! */
 378         if (!RPC_IS_ACTIVATED(task)) {
 379                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 380                 return;
 381         }
 382
 383         __rpc_disable_timer(task);
 384         __rpc_remove_wait_queue(queue, task);
 385
 386         rpc_make_runnable(task);
 387
 388         dprintk("RPC:       __rpc_wake_up_task done\n");
 389 }
 390
 391 /*
 392  * Wake up a queued task while the queue lock is being held
 393  */
 394 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 395 {
 396         if (!RPC_IS_QUEUED(task) || task->tk_waitqueue != queue)
 397                 return;
 398         if (rpc_start_wakeup(task)) {
 399                         __rpc_do_wake_up_task(queue, task);
 400                 rpc_finish_wakeup(task);
 401         }
 402 }
 403
 404 /*
 405  * Wake up a task on a specific queue
 406  */
 407 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 408 {
 409         rcu_read_lock_bh();
 410         spin_lock(&queue->lock);
 411         rpc_wake_up_task_queue_locked(queue, task);
 412         spin_unlock(&queue->lock);
 413         rcu_read_unlock_bh();
 414 }
 415 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 416
 417 /*
 418  * Wake up the specified task
 419  */
 420 static void rpc_wake_up_task(struct rpc_task *task)
 421 {
 422         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 423 }
 424
 425 /*
 426  * Wake up the next task on a priority queue.
 427  */
 428 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 429 {
 430         struct list_head *q;
 431         struct rpc_task *task;
 432
 433         /*
 434          * Service a batch of tasks from a single owner.
 435          */
 436         q = &queue->tasks[queue->priority];
 437         if (!list_empty(q)) {
 438                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 439                 if (queue->owner == task->tk_owner) {
 440                         if (--queue->nr)
 441                                 goto out;
 442                         list_move_tail(&task->u.tk_wait.list, q);
 443                 }
 444                 /*
 445                  * Check if we need to switch queues.
 446                  */
 447                 if (--queue->count)
 448                         goto new_owner;
 449         }
 450
 451         /*
 452          * Service the next queue.
 453          */
 454         do {
 455                 if (q == &queue->tasks[0])
 456                         q = &queue->tasks[queue->maxpriority];
 457                 else
 458                         q = q - 1;
 459                 if (!list_empty(q)) {
 460                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 461                         goto new_queue;
 462                 }
 463         } while (q != &queue->tasks[queue->priority]);
 464
 465         rpc_reset_waitqueue_priority(queue);
 466         return NULL;
 467
 468 new_queue:
 469         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 470 new_owner:
 471         rpc_set_waitqueue_owner(queue, task->tk_owner);
 472 out:
 473         rpc_wake_up_task_queue_locked(queue, task);
 474         return task;
 475 }
 476
 477 /*
 478  * Wake up the next task on the wait queue.
 479  */
 480 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 481 {
 482         struct rpc_task *task = NULL;
 483
 484         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 485                         queue, rpc_qname(queue));
 486         rcu_read_lock_bh();
 487         spin_lock(&queue->lock);
 488         if (RPC_IS_PRIORITY(queue))
 489                 task = __rpc_wake_up_next_priority(queue);
 490         else {
 491                 task_for_first(task, &queue->tasks[0])
 492                         rpc_wake_up_task_queue_locked(queue, task);
 493         }
 494         spin_unlock(&queue->lock);
 495         rcu_read_unlock_bh();
 496
 497         return task;
 498 }
 499 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 500
 501 /**
 502  * rpc_wake_up - wake up all rpc_tasks
 503  * @queue: rpc_wait_queue on which the tasks are sleeping
 504  *
 505  * Grabs queue->lock
 506  */
 507 void rpc_wake_up(struct rpc_wait_queue *queue)
 508 {
 509         struct rpc_task *task, *next;
 510         struct list_head *head;
 511
 512         rcu_read_lock_bh();
 513         spin_lock(&queue->lock);
 514         head = &queue->tasks[queue->maxpriority];
 515         for (;;) {
 516                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 517                         rpc_wake_up_task_queue_locked(queue, task);
 518                 if (head == &queue->tasks[0])
 519                         break;
 520                 head--;
 521         }
 522         spin_unlock(&queue->lock);
 523         rcu_read_unlock_bh();
 524 }
 525 EXPORT_SYMBOL_GPL(rpc_wake_up);
 526
 527 /**
 528  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 529  * @queue: rpc_wait_queue on which the tasks are sleeping
 530  * @status: status value to set
 531  *
 532  * Grabs queue->lock
 533  */
 534 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 535 {
 536         struct rpc_task *task, *next;
 537         struct list_head *head;
 538
 539         rcu_read_lock_bh();
 540         spin_lock(&queue->lock);
 541         head = &queue->tasks[queue->maxpriority];
 542         for (;;) {
 543                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 544                         task->tk_status = status;
 545                         rpc_wake_up_task_queue_locked(queue, task);
 546                 }
 547                 if (head == &queue->tasks[0])
 548                         break;
 549                 head--;
 550         }
 551         spin_unlock(&queue->lock);
 552         rcu_read_unlock_bh();
 553 }
 554 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 555
 556 /*
 557  * Run a timeout function.
 558  */
 559 static void rpc_run_timer(unsigned long ptr)
 560 {
 561         struct rpc_task *task = (struct rpc_task *)ptr;
 562         void (*callback)(struct rpc_task *);
 563
 564         if (RPC_IS_QUEUED(task)) {
 565                 struct rpc_wait_queue *queue = task->tk_waitqueue;
 566                 callback = task->tk_timeout_fn;
 567
 568                 dprintk("RPC: %5u running timer\n", task->tk_pid);
 569                 if (callback != NULL)
 570                         callback(task);
 571                 /* Note: we're already in a bh-safe context */
 572                 spin_lock(&queue->lock);
 573                 rpc_wake_up_task_queue_locked(queue, task);
 574                 spin_unlock(&queue->lock);
 575         }
 576         smp_mb__before_clear_bit();
 577         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 578         smp_mb__after_clear_bit();
 579 }
 580
 581 static void __rpc_atrun(struct rpc_task *task)
 582 {
 583 }
 584
 585 /*
 586  * Run a task at a later time
 587  */
 588 void rpc_delay(struct rpc_task *task, unsigned long delay)
 589 {
 590         task->tk_timeout = delay;
 591         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 592 }
 593 EXPORT_SYMBOL_GPL(rpc_delay);
 594
 595 /*
 596  * Helper to call task->tk_ops->rpc_call_prepare
 597  */
 598 static void rpc_prepare_task(struct rpc_task *task)
 599 {
 600         lock_kernel();
 601         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 602         unlock_kernel();
 603 }
 604
 605 /*
 606  * Helper that calls task->tk_ops->rpc_call_done if it exists
 607  */
 608 void rpc_exit_task(struct rpc_task *task)
 609 {
 610         task->tk_action = NULL;
 611         if (task->tk_ops->rpc_call_done != NULL) {
 612                 lock_kernel();
 613                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 614                 unlock_kernel();
 615                 if (task->tk_action != NULL) {
 616                         WARN_ON(RPC_ASSASSINATED(task));
 617                         /* Always release the RPC slot and buffer memory */
 618                         xprt_release(task);
 619                 }
 620         }
 621 }
 622 EXPORT_SYMBOL_GPL(rpc_exit_task);
 623
 624 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 625 {
 626         if (ops->rpc_release != NULL) {
 627                 lock_kernel();
 628                 ops->rpc_release(calldata);
 629                 unlock_kernel();
 630         }
 631 }
 632
 633 /*
 634  * This is the RPC `scheduler' (or rather, the finite state machine).
 635  */
 636 static void __rpc_execute(struct rpc_task *task)
 637 {
 638         int             status = 0;
 639
 640         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 641                         task->tk_pid, task->tk_flags);
 642
 643         BUG_ON(RPC_IS_QUEUED(task));
 644
 645         for (;;) {
 646                 /*
 647                  * Garbage collection of pending timers...
 648                  */
 649                 rpc_delete_timer(task);
 650
 651                 /*
 652                  * Execute any pending callback.
 653                  */
 654                 if (RPC_DO_CALLBACK(task)) {
 655                         /* Define a callback save pointer */
 656                         void (*save_callback)(struct rpc_task *);
 657
 658                         /*
 659                          * If a callback exists, save it, reset it,
 660                          * call it.
 661                          * The save is needed to stop from resetting
 662                          * another callback set within the callback handler
 663                          * - Dave
 664                          */
 665                         save_callback=task->tk_callback;
 666                         task->tk_callback=NULL;
 667                         save_callback(task);
 668                 }
 669
 670                 /*
 671                  * Perform the next FSM step.
 672                  * tk_action may be NULL when the task has been killed
 673                  * by someone else.
 674                  */
 675                 if (!RPC_IS_QUEUED(task)) {
 676                         if (task->tk_action == NULL)
 677                                 break;
 678                         task->tk_action(task);
 679                 }
 680
 681                 /*
 682                  * Lockless check for whether task is sleeping or not.
 683                  */
 684                 if (!RPC_IS_QUEUED(task))
 685                         continue;
 686                 rpc_clear_running(task);
 687                 if (RPC_IS_ASYNC(task)) {
 688                         /* Careful! we may have raced... */
 689                         if (RPC_IS_QUEUED(task))
 690                                 return;
 691                         if (rpc_test_and_set_running(task))
 692                                 return;
 693                         continue;
 694                 }
 695
 696                 /* sync task: sleep here */
 697                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 698                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 699                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 700                                 TASK_KILLABLE);
 701                 if (status == -ERESTARTSYS) {
 702                         /*
 703                          * When a sync task receives a signal, it exits with
 704                          * -ERESTARTSYS. In order to catch any callbacks that
 705                          * clean up after sleeping on some queue, we don't
 706                          * break the loop here, but go around once more.
 707                          */
 708                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 709                         task->tk_flags |= RPC_TASK_KILLED;
 710                         rpc_exit(task, -ERESTARTSYS);
 711                         rpc_wake_up_task(task);
 712                 }
 713                 rpc_set_running(task);
 714                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 715         }
 716
 717         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 718                         task->tk_status);
 719         /* Release all resources associated with the task */
 720         rpc_release_task(task);
 721 }
 722
 723 /*
 724  * User-visible entry point to the scheduler.
 725  *
 726  * This may be called recursively if e.g. an async NFS task updates
 727  * the attributes and finds that dirty pages must be flushed.
 728  * NOTE: Upon exit of this function the task is guaranteed to be
 729  *       released. In particular note that tk_release() will have
 730  *       been called, so your task memory may have been freed.
 731  */
 732 void rpc_execute(struct rpc_task *task)
 733 {
 734         rpc_set_active(task);
 735         rpc_set_running(task);
 736         __rpc_execute(task);
 737 }
 738
 739 static void rpc_async_schedule(struct work_struct *work)
 740 {
 741         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 742 }
 743
 744 struct rpc_buffer {
 745         size_t  len;
 746         char    data[];
 747 };
 748
 749 /**
 750  * rpc_malloc - allocate an RPC buffer
 751  * @task: RPC task that will use this buffer
 752  * @size: requested byte size
 753  *
 754  * To prevent rpciod from hanging, this allocator never sleeps,
 755  * returning NULL if the request cannot be serviced immediately.
 756  * The caller can arrange to sleep in a way that is safe for rpciod.
 757  *
 758  * Most requests are 'small' (under 2KiB) and can be serviced from a
 759  * mempool, ensuring that NFS reads and writes can always proceed,
 760  * and that there is good locality of reference for these buffers.
 761  *
 762  * In order to avoid memory starvation triggering more writebacks of
 763  * NFS requests, we avoid using GFP_KERNEL.
 764  */
 765 void *rpc_malloc(struct rpc_task *task, size_t size)
 766 {
 767         struct rpc_buffer *buf;
 768         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 769
 770         size += sizeof(struct rpc_buffer);
 771         if (size <= RPC_BUFFER_MAXSIZE)
 772                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 773         else
 774                 buf = kmalloc(size, gfp);
 775
 776         if (!buf)
 777                 return NULL;
 778
 779         buf->len = size;
 780         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 781                         task->tk_pid, size, buf);
 782         return &buf->data;
 783 }
 784 EXPORT_SYMBOL_GPL(rpc_malloc);
 785
 786 /**
 787  * rpc_free - free buffer allocated via rpc_malloc
 788  * @buffer: buffer to free
 789  *
 790  */
 791 void rpc_free(void *buffer)
 792 {
 793         size_t size;
 794         struct rpc_buffer *buf;
 795
 796         if (!buffer)
 797                 return;
 798
 799         buf = container_of(buffer, struct rpc_buffer, data);
 800         size = buf->len;
 801
 802         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 803                         size, buf);
 804
 805         if (size <= RPC_BUFFER_MAXSIZE)
 806                 mempool_free(buf, rpc_buffer_mempool);
 807         else
 808                 kfree(buf);
 809 }
 810 EXPORT_SYMBOL_GPL(rpc_free);
 811
 812 /*
 813  * Creation and deletion of RPC task structures
 814  */
 815 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 816 {
 817         memset(task, 0, sizeof(*task));
 818         setup_timer(&task->tk_timer, rpc_run_timer, (unsigned long)task);
 819         atomic_set(&task->tk_count, 1);
 820         task->tk_flags  = task_setup_data->flags;
 821         task->tk_ops = task_setup_data->callback_ops;
 822         task->tk_calldata = task_setup_data->callback_data;
 823         INIT_LIST_HEAD(&task->tk_task);
 824
 825         /* Initialize retry counters */
 826         task->tk_garb_retry = 2;
 827         task->tk_cred_retry = 2;
 828
 829         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 830         task->tk_owner = current->tgid;
 831
 832         /* Initialize workqueue for async tasks */
 833         task->tk_workqueue = task_setup_data->workqueue;
 834
 835         task->tk_client = task_setup_data->rpc_client;
 836         if (task->tk_client != NULL) {
 837                 kref_get(&task->tk_client->cl_kref);
 838                 if (task->tk_client->cl_softrtry)
 839                         task->tk_flags |= RPC_TASK_SOFT;
 840         }
 841
 842         if (task->tk_ops->rpc_call_prepare != NULL)
 843                 task->tk_action = rpc_prepare_task;
 844
 845         if (task_setup_data->rpc_message != NULL) {
 846                 memcpy(&task->tk_msg, task_setup_data->rpc_message, sizeof(task->tk_msg));
 847                 /* Bind the user cred */
 848                 if (task->tk_msg.rpc_cred != NULL)
 849                         rpcauth_holdcred(task);
 850                 else
 851                         rpcauth_bindcred(task);
 852                 if (task->tk_action == NULL)
 853                         rpc_call_start(task);
 854         }
 855
 856         /* starting timestamp */
 857         task->tk_start = jiffies;
 858
 859         dprintk("RPC:       new task initialized, procpid %u\n",
 860                                 task_pid_nr(current));
 861 }
 862
 863 static struct rpc_task *
 864 rpc_alloc_task(void)
 865 {
 866         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 867 }
 868
 869 static void rpc_free_task_rcu(struct rcu_head *rcu)
 870 {
 871         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 872         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 873         mempool_free(task, rpc_task_mempool);
 874 }
 875
 876 /*
 877  * Create a new task for the specified client.
 878  */
 879 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 880 {
 881         struct rpc_task *task = setup_data->task;
 882         unsigned short flags = 0;
 883
 884         if (task == NULL) {
 885                 task = rpc_alloc_task();
 886                 if (task == NULL)
 887                         goto out;
 888                 flags = RPC_TASK_DYNAMIC;
 889         }
 890
 891         rpc_init_task(task, setup_data);
 892
 893         task->tk_flags |= flags;
 894         dprintk("RPC:       allocated task %p\n", task);
 895 out:
 896         return task;
 897 }
 898
 899 static void rpc_free_task(struct rpc_task *task)
 900 {
 901         const struct rpc_call_ops *tk_ops = task->tk_ops;
 902         void *calldata = task->tk_calldata;
 903
 904         if (task->tk_flags & RPC_TASK_DYNAMIC)
 905                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task_rcu);
 906         rpc_release_calldata(tk_ops, calldata);
 907 }
 908
 909 static void rpc_async_release(struct work_struct *work)
 910 {
 911         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 912 }
 913
 914 void rpc_put_task(struct rpc_task *task)
 915 {
 916         if (!atomic_dec_and_test(&task->tk_count))
 917                 return;
 918         /* Release resources */
 919         if (task->tk_rqstp)
 920                 xprt_release(task);
 921         if (task->tk_msg.rpc_cred)
 922                 rpcauth_unbindcred(task);
 923         if (task->tk_client) {
 924                 rpc_release_client(task->tk_client);
 925                 task->tk_client = NULL;
 926         }
 927         if (task->tk_workqueue != NULL) {
 928                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 929                 queue_work(task->tk_workqueue, &task->u.tk_work);
 930         } else
 931                 rpc_free_task(task);
 932 }
 933 EXPORT_SYMBOL_GPL(rpc_put_task);
 934
 935 static void rpc_release_task(struct rpc_task *task)
 936 {
 937 #ifdef RPC_DEBUG
 938         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 939 #endif
 940         dprintk("RPC: %5u release task\n", task->tk_pid);
 941
 942         if (!list_empty(&task->tk_task)) {
 943                 struct rpc_clnt *clnt = task->tk_client;
 944                 /* Remove from client task list */
 945                 spin_lock(&clnt->cl_lock);
 946                 list_del(&task->tk_task);
 947                 spin_unlock(&clnt->cl_lock);
 948         }
 949         BUG_ON (RPC_IS_QUEUED(task));
 950
 951         /* Synchronously delete any running timer */
 952         rpc_delete_timer(task);
 953
 954 #ifdef RPC_DEBUG
 955         task->tk_magic = 0;
 956 #endif
 957         /* Wake up anyone who is waiting for task completion */
 958         rpc_mark_complete_task(task);
 959
 960         rpc_put_task(task);
 961 }
 962
 963 /*
 964  * Kill all tasks for the given client.
 965  * XXX: kill their descendants as well?
 966  */
 967 void rpc_killall_tasks(struct rpc_clnt *clnt)
 968 {
 969         struct rpc_task *rovr;
 970
 971
 972         if (list_empty(&clnt->cl_tasks))
 973                 return;
 974         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 975         /*
 976          * Spin lock all_tasks to prevent changes...
 977          */
 978         spin_lock(&clnt->cl_lock);
 979         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 980                 if (! RPC_IS_ACTIVATED(rovr))
 981                         continue;
 982                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 983                         rovr->tk_flags |= RPC_TASK_KILLED;
 984                         rpc_exit(rovr, -EIO);
 985                         rpc_wake_up_task(rovr);
 986                 }
 987         }
 988         spin_unlock(&clnt->cl_lock);
 989 }
 990 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 991
 992 int rpciod_up(void)
 993 {
 994         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 995 }
 996
 997 void rpciod_down(void)
 998 {
 999         module_put(THIS_MODULE);
1000 }
1001
1002 /*
1003  * Start up the rpciod workqueue.
1004  */
1005 static int rpciod_start(void)
1006 {
1007         struct workqueue_struct *wq;
1008
1009         /*
1010          * Create the rpciod thread and wait for it to start.
1011          */
1012         dprintk("RPC:       creating workqueue rpciod\n");
1013         wq = create_workqueue("rpciod");
1014         rpciod_workqueue = wq;
1015         return rpciod_workqueue != NULL;
1016 }
1017
1018 static void rpciod_stop(void)
1019 {
1020         struct workqueue_struct *wq = NULL;
1021
1022         if (rpciod_workqueue == NULL)
1023                 return;
1024         dprintk("RPC:       destroying workqueue rpciod\n");
1025
1026         wq = rpciod_workqueue;
1027         rpciod_workqueue = NULL;
1028         destroy_workqueue(wq);
1029 }
1030
1031 void
1032 rpc_destroy_mempool(void)
1033 {
1034         rpciod_stop();
1035         if (rpc_buffer_mempool)
1036                 mempool_destroy(rpc_buffer_mempool);
1037         if (rpc_task_mempool)
1038                 mempool_destroy(rpc_task_mempool);
1039         if (rpc_task_slabp)
1040                 kmem_cache_destroy(rpc_task_slabp);
1041         if (rpc_buffer_slabp)
1042                 kmem_cache_destroy(rpc_buffer_slabp);
1043 }
1044
1045 int
1046 rpc_init_mempool(void)
1047 {
1048         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1049                                              sizeof(struct rpc_task),
1050                                              0, SLAB_HWCACHE_ALIGN,
1051                                              NULL);
1052         if (!rpc_task_slabp)
1053                 goto err_nomem;
1054         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1055                                              RPC_BUFFER_MAXSIZE,
1056                                              0, SLAB_HWCACHE_ALIGN,
1057                                              NULL);
1058         if (!rpc_buffer_slabp)
1059                 goto err_nomem;
1060         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1061                                                     rpc_task_slabp);
1062         if (!rpc_task_mempool)
1063                 goto err_nomem;
1064         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1065                                                       rpc_buffer_slabp);
1066         if (!rpc_buffer_mempool)
1067                 goto err_nomem;
1068         if (!rpciod_start())
1069                 goto err_nomem;
1070         /*
1071          * The following is not strictly a mempool initialisation,
1072          * but there is no harm in doing it here
1073          */
1074         rpc_init_wait_queue(&delay_queue, "delayq");
1075         return 0;
1076 err_nomem:
1077         rpc_destroy_mempool();
1078         return -ENOMEM;
1079 }