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 #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 void rpc_wake_up_task(struct rpc_task *task)
 421 {
 422         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 423 }
 424 EXPORT_SYMBOL_GPL(rpc_wake_up_task);
 425
 426 /*
 427  * Wake up the next task on a priority queue.
 428  */
 429 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 430 {
 431         struct list_head *q;
 432         struct rpc_task *task;
 433
 434         /*
 435          * Service a batch of tasks from a single owner.
 436          */
 437         q = &queue->tasks[queue->priority];
 438         if (!list_empty(q)) {
 439                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 440                 if (queue->owner == task->tk_owner) {
 441                         if (--queue->nr)
 442                                 goto out;
 443                         list_move_tail(&task->u.tk_wait.list, q);
 444                 }
 445                 /*
 446                  * Check if we need to switch queues.
 447                  */
 448                 if (--queue->count)
 449                         goto new_owner;
 450         }
 451
 452         /*
 453          * Service the next queue.
 454          */
 455         do {
 456                 if (q == &queue->tasks[0])
 457                         q = &queue->tasks[queue->maxpriority];
 458                 else
 459                         q = q - 1;
 460                 if (!list_empty(q)) {
 461                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 462                         goto new_queue;
 463                 }
 464         } while (q != &queue->tasks[queue->priority]);
 465
 466         rpc_reset_waitqueue_priority(queue);
 467         return NULL;
 468
 469 new_queue:
 470         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 471 new_owner:
 472         rpc_set_waitqueue_owner(queue, task->tk_owner);
 473 out:
 474         rpc_wake_up_task_queue_locked(queue, task);
 475         return task;
 476 }
 477
 478 /*
 479  * Wake up the next task on the wait queue.
 480  */
 481 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 482 {
 483         struct rpc_task *task = NULL;
 484
 485         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 486                         queue, rpc_qname(queue));
 487         rcu_read_lock_bh();
 488         spin_lock(&queue->lock);
 489         if (RPC_IS_PRIORITY(queue))
 490                 task = __rpc_wake_up_next_priority(queue);
 491         else {
 492                 task_for_first(task, &queue->tasks[0])
 493                         rpc_wake_up_task_queue_locked(queue, task);
 494         }
 495         spin_unlock(&queue->lock);
 496         rcu_read_unlock_bh();
 497
 498         return task;
 499 }
 500 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 501
 502 /**
 503  * rpc_wake_up - wake up all rpc_tasks
 504  * @queue: rpc_wait_queue on which the tasks are sleeping
 505  *
 506  * Grabs queue->lock
 507  */
 508 void rpc_wake_up(struct rpc_wait_queue *queue)
 509 {
 510         struct rpc_task *task, *next;
 511         struct list_head *head;
 512
 513         rcu_read_lock_bh();
 514         spin_lock(&queue->lock);
 515         head = &queue->tasks[queue->maxpriority];
 516         for (;;) {
 517                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 518                         rpc_wake_up_task_queue_locked(queue, task);
 519                 if (head == &queue->tasks[0])
 520                         break;
 521                 head--;
 522         }
 523         spin_unlock(&queue->lock);
 524         rcu_read_unlock_bh();
 525 }
 526 EXPORT_SYMBOL_GPL(rpc_wake_up);
 527
 528 /**
 529  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 530  * @queue: rpc_wait_queue on which the tasks are sleeping
 531  * @status: status value to set
 532  *
 533  * Grabs queue->lock
 534  */
 535 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 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                         task->tk_status = status;
 546                         rpc_wake_up_task_queue_locked(queue, task);
 547                 }
 548                 if (head == &queue->tasks[0])
 549                         break;
 550                 head--;
 551         }
 552         spin_unlock(&queue->lock);
 553         rcu_read_unlock_bh();
 554 }
 555 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 556
 557 /*
 558  * Run a timeout function.
 559  */
 560 static void rpc_run_timer(unsigned long ptr)
 561 {
 562         struct rpc_task *task = (struct rpc_task *)ptr;
 563         void (*callback)(struct rpc_task *);
 564
 565         if (RPC_IS_QUEUED(task)) {
 566                 struct rpc_wait_queue *queue = task->tk_waitqueue;
 567                 callback = task->tk_timeout_fn;
 568
 569                 dprintk("RPC: %5u running timer\n", task->tk_pid);
 570                 if (callback != NULL)
 571                         callback(task);
 572                 /* Note: we're already in a bh-safe context */
 573                 spin_lock(&queue->lock);
 574                 rpc_wake_up_task_queue_locked(queue, task);
 575                 spin_unlock(&queue->lock);
 576         }
 577         smp_mb__before_clear_bit();
 578         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 579         smp_mb__after_clear_bit();
 580 }
 581
 582 static void __rpc_atrun(struct rpc_task *task)
 583 {
 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                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 700                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 701                                 TASK_KILLABLE);
 702                 if (status == -ERESTARTSYS) {
 703                         /*
 704                          * When a sync task receives a signal, it exits with
 705                          * -ERESTARTSYS. In order to catch any callbacks that
 706                          * clean up after sleeping on some queue, we don't
 707                          * break the loop here, but go around once more.
 708                          */
 709                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 710                         task->tk_flags |= RPC_TASK_KILLED;
 711                         rpc_exit(task, -ERESTARTSYS);
 712                         rpc_wake_up_task(task);
 713                 }
 714                 rpc_set_running(task);
 715                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 716         }
 717
 718         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 719                         task->tk_status);
 720         /* Release all resources associated with the task */
 721         rpc_release_task(task);
 722 }
 723
 724 /*
 725  * User-visible entry point to the scheduler.
 726  *
 727  * This may be called recursively if e.g. an async NFS task updates
 728  * the attributes and finds that dirty pages must be flushed.
 729  * NOTE: Upon exit of this function the task is guaranteed to be
 730  *       released. In particular note that tk_release() will have
 731  *       been called, so your task memory may have been freed.
 732  */
 733 void rpc_execute(struct rpc_task *task)
 734 {
 735         rpc_set_active(task);
 736         rpc_set_running(task);
 737         __rpc_execute(task);
 738 }
 739
 740 static void rpc_async_schedule(struct work_struct *work)
 741 {
 742         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 743 }
 744
 745 struct rpc_buffer {
 746         size_t  len;
 747         char    data[];
 748 };
 749
 750 /**
 751  * rpc_malloc - allocate an RPC buffer
 752  * @task: RPC task that will use this buffer
 753  * @size: requested byte size
 754  *
 755  * To prevent rpciod from hanging, this allocator never sleeps,
 756  * returning NULL if the request cannot be serviced immediately.
 757  * The caller can arrange to sleep in a way that is safe for rpciod.
 758  *
 759  * Most requests are 'small' (under 2KiB) and can be serviced from a
 760  * mempool, ensuring that NFS reads and writes can always proceed,
 761  * and that there is good locality of reference for these buffers.
 762  *
 763  * In order to avoid memory starvation triggering more writebacks of
 764  * NFS requests, we avoid using GFP_KERNEL.
 765  */
 766 void *rpc_malloc(struct rpc_task *task, size_t size)
 767 {
 768         struct rpc_buffer *buf;
 769         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 770
 771         size += sizeof(struct rpc_buffer);
 772         if (size <= RPC_BUFFER_MAXSIZE)
 773                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 774         else
 775                 buf = kmalloc(size, gfp);
 776
 777         if (!buf)
 778                 return NULL;
 779
 780         buf->len = size;
 781         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 782                         task->tk_pid, size, buf);
 783         return &buf->data;
 784 }
 785 EXPORT_SYMBOL_GPL(rpc_malloc);
 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 EXPORT_SYMBOL_GPL(rpc_free);
 812
 813 /*
 814  * Creation and deletion of RPC task structures
 815  */
 816 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 817 {
 818         memset(task, 0, sizeof(*task));
 819         setup_timer(&task->tk_timer, rpc_run_timer, (unsigned long)task);
 820         atomic_set(&task->tk_count, 1);
 821         task->tk_flags  = task_setup_data->flags;
 822         task->tk_ops = task_setup_data->callback_ops;
 823         task->tk_calldata = task_setup_data->callback_data;
 824         INIT_LIST_HEAD(&task->tk_task);
 825
 826         /* Initialize retry counters */
 827         task->tk_garb_retry = 2;
 828         task->tk_cred_retry = 2;
 829
 830         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 831         task->tk_owner = current->tgid;
 832
 833         /* Initialize workqueue for async tasks */
 834         task->tk_workqueue = task_setup_data->workqueue;
 835
 836         task->tk_client = task_setup_data->rpc_client;
 837         if (task->tk_client != NULL) {
 838                 kref_get(&task->tk_client->cl_kref);
 839                 if (task->tk_client->cl_softrtry)
 840                         task->tk_flags |= RPC_TASK_SOFT;
 841         }
 842
 843         if (task->tk_ops->rpc_call_prepare != NULL)
 844                 task->tk_action = rpc_prepare_task;
 845
 846         if (task_setup_data->rpc_message != NULL) {
 847                 memcpy(&task->tk_msg, task_setup_data->rpc_message, sizeof(task->tk_msg));
 848                 /* Bind the user cred */
 849                 if (task->tk_msg.rpc_cred != NULL)
 850                         rpcauth_holdcred(task);
 851                 else
 852                         rpcauth_bindcred(task);
 853                 if (task->tk_action == NULL)
 854                         rpc_call_start(task);
 855         }
 856
 857         /* starting timestamp */
 858         task->tk_start = jiffies;
 859
 860         dprintk("RPC:       new task initialized, procpid %u\n",
 861                                 task_pid_nr(current));
 862 }
 863
 864 static struct rpc_task *
 865 rpc_alloc_task(void)
 866 {
 867         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 868 }
 869
 870 static void rpc_free_task_rcu(struct rcu_head *rcu)
 871 {
 872         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 873         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 874         mempool_free(task, rpc_task_mempool);
 875 }
 876
 877 /*
 878  * Create a new task for the specified client.
 879  */
 880 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 881 {
 882         struct rpc_task *task = setup_data->task;
 883         unsigned short flags = 0;
 884
 885         if (task == NULL) {
 886                 task = rpc_alloc_task();
 887                 if (task == NULL)
 888                         goto out;
 889                 flags = RPC_TASK_DYNAMIC;
 890         }
 891
 892         rpc_init_task(task, setup_data);
 893
 894         task->tk_flags |= flags;
 895         dprintk("RPC:       allocated task %p\n", task);
 896 out:
 897         return task;
 898 }
 899
 900 static void rpc_free_task(struct rpc_task *task)
 901 {
 902         const struct rpc_call_ops *tk_ops = task->tk_ops;
 903         void *calldata = task->tk_calldata;
 904
 905         if (task->tk_flags & RPC_TASK_DYNAMIC)
 906                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task_rcu);
 907         rpc_release_calldata(tk_ops, calldata);
 908 }
 909
 910 static void rpc_async_release(struct work_struct *work)
 911 {
 912         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 913 }
 914
 915 void rpc_put_task(struct rpc_task *task)
 916 {
 917         if (!atomic_dec_and_test(&task->tk_count))
 918                 return;
 919         /* Release resources */
 920         if (task->tk_rqstp)
 921                 xprt_release(task);
 922         if (task->tk_msg.rpc_cred)
 923                 rpcauth_unbindcred(task);
 924         if (task->tk_client) {
 925                 rpc_release_client(task->tk_client);
 926                 task->tk_client = NULL;
 927         }
 928         if (task->tk_workqueue != NULL) {
 929                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 930                 queue_work(task->tk_workqueue, &task->u.tk_work);
 931         } else
 932                 rpc_free_task(task);
 933 }
 934 EXPORT_SYMBOL_GPL(rpc_put_task);
 935
 936 static void rpc_release_task(struct rpc_task *task)
 937 {
 938 #ifdef RPC_DEBUG
 939         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 940 #endif
 941         dprintk("RPC: %5u release task\n", task->tk_pid);
 942
 943         if (!list_empty(&task->tk_task)) {
 944                 struct rpc_clnt *clnt = task->tk_client;
 945                 /* Remove from client task list */
 946                 spin_lock(&clnt->cl_lock);
 947                 list_del(&task->tk_task);
 948                 spin_unlock(&clnt->cl_lock);
 949         }
 950         BUG_ON (RPC_IS_QUEUED(task));
 951
 952         /* Synchronously delete any running timer */
 953         rpc_delete_timer(task);
 954
 955 #ifdef RPC_DEBUG
 956         task->tk_magic = 0;
 957 #endif
 958         /* Wake up anyone who is waiting for task completion */
 959         rpc_mark_complete_task(task);
 960
 961         rpc_put_task(task);
 962 }
 963
 964 /*
 965  * Kill all tasks for the given client.
 966  * XXX: kill their descendants as well?
 967  */
 968 void rpc_killall_tasks(struct rpc_clnt *clnt)
 969 {
 970         struct rpc_task *rovr;
 971
 972
 973         if (list_empty(&clnt->cl_tasks))
 974                 return;
 975         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 976         /*
 977          * Spin lock all_tasks to prevent changes...
 978          */
 979         spin_lock(&clnt->cl_lock);
 980         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 981                 if (! RPC_IS_ACTIVATED(rovr))
 982                         continue;
 983                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 984                         rovr->tk_flags |= RPC_TASK_KILLED;
 985                         rpc_exit(rovr, -EIO);
 986                         rpc_wake_up_task(rovr);
 987                 }
 988         }
 989         spin_unlock(&clnt->cl_lock);
 990 }
 991 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 992
 993 int rpciod_up(void)
 994 {
 995         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 996 }
 997
 998 void rpciod_down(void)
 999 {
1000         module_put(THIS_MODULE);
1001 }
1002
1003 /*
1004  * Start up the rpciod workqueue.
1005  */
1006 static int rpciod_start(void)
1007 {
1008         struct workqueue_struct *wq;
1009
1010         /*
1011          * Create the rpciod thread and wait for it to start.
1012          */
1013         dprintk("RPC:       creating workqueue rpciod\n");
1014         wq = create_workqueue("rpciod");
1015         rpciod_workqueue = wq;
1016         return rpciod_workqueue != NULL;
1017 }
1018
1019 static void rpciod_stop(void)
1020 {
1021         struct workqueue_struct *wq = NULL;
1022
1023         if (rpciod_workqueue == NULL)
1024                 return;
1025         dprintk("RPC:       destroying workqueue rpciod\n");
1026
1027         wq = rpciod_workqueue;
1028         rpciod_workqueue = NULL;
1029         destroy_workqueue(wq);
1030 }
1031
1032 void
1033 rpc_destroy_mempool(void)
1034 {
1035         rpciod_stop();
1036         if (rpc_buffer_mempool)
1037                 mempool_destroy(rpc_buffer_mempool);
1038         if (rpc_task_mempool)
1039                 mempool_destroy(rpc_task_mempool);
1040         if (rpc_task_slabp)
1041                 kmem_cache_destroy(rpc_task_slabp);
1042         if (rpc_buffer_slabp)
1043                 kmem_cache_destroy(rpc_buffer_slabp);
1044 }
1045
1046 int
1047 rpc_init_mempool(void)
1048 {
1049         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1050                                              sizeof(struct rpc_task),
1051                                              0, SLAB_HWCACHE_ALIGN,
1052                                              NULL);
1053         if (!rpc_task_slabp)
1054                 goto err_nomem;
1055         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1056                                              RPC_BUFFER_MAXSIZE,
1057                                              0, SLAB_HWCACHE_ALIGN,
1058                                              NULL);
1059         if (!rpc_buffer_slabp)
1060                 goto err_nomem;
1061         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1062                                                     rpc_task_slabp);
1063         if (!rpc_task_mempool)
1064                 goto err_nomem;
1065         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1066                                                       rpc_buffer_slabp);
1067         if (!rpc_buffer_mempool)
1068                 goto err_nomem;
1069         if (!rpciod_start())
1070                 goto err_nomem;
1071         /*
1072          * The following is not strictly a mempool initialisation,
1073          * but there is no harm in doing it here
1074          */
1075         rpc_init_wait_queue(&delay_queue, "delayq");
1076         return 0;
1077 err_nomem:
1078         rpc_destroy_mempool();
1079         return -ENOMEM;
1080 }