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

   1 /*
   2  * linux/net/sunrpc/sched.c
   3  *
   4  * Scheduling for synchronous and asynchronous RPC requests.
   5  *
   6  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  10  */
  11
  12 #include <linux/module.h>
  13
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/slab.h>
  17 #include <linux/mempool.h>
  18 #include <linux/smp.h>
  19 #include <linux/smp_lock.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/mutex.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #ifdef RPC_DEBUG
  26 #define RPCDBG_FACILITY         RPCDBG_SCHED
  27 #define RPC_TASK_MAGIC_ID       0xf00baa
  28 static int                      rpc_task_id;
  29 #endif
  30
  31 /*
  32  * RPC slabs and memory pools
  33  */
  34 #define RPC_BUFFER_MAXSIZE      (2048)
  35 #define RPC_BUFFER_POOLSIZE     (8)
  36 #define RPC_TASK_POOLSIZE       (8)
  37 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  38 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  39 static mempool_t        *rpc_task_mempool __read_mostly;
  40 static mempool_t        *rpc_buffer_mempool __read_mostly;
  41
  42 static void                     __rpc_default_timer(struct rpc_task *task);
  43 static void                     rpciod_killall(void);
  44 static void                     rpc_async_schedule(struct work_struct *);
  45 static void                      rpc_release_task(struct rpc_task *task);
  46
  47 /*
  48  * RPC tasks sit here while waiting for conditions to improve.
  49  */
  50 static RPC_WAITQ(delay_queue, "delayq");
  51
  52 /*
  53  * All RPC clients are linked into this list
  54  */
  55 static LIST_HEAD(all_clients);
  56 static DECLARE_WAIT_QUEUE_HEAD(client_kill_wait);
  57
  58 /*
  59  * rpciod-related stuff
  60  */
  61 static DEFINE_MUTEX(rpciod_mutex);
  62 static unsigned int             rpciod_users;
  63 struct workqueue_struct *rpciod_workqueue;
  64
  65 /*
  66  * Spinlock for other critical sections of code.
  67  */
  68 static DEFINE_SPINLOCK(rpc_sched_lock);
  69
  70 /*
  71  * Disable the timer for a given RPC task. Should be called with
  72  * queue->lock and bh_disabled in order to avoid races within
  73  * rpc_run_timer().
  74  */
  75 static inline void
  76 __rpc_disable_timer(struct rpc_task *task)
  77 {
  78         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  79         task->tk_timeout_fn = NULL;
  80         task->tk_timeout = 0;
  81 }
  82
  83 /*
  84  * Run a timeout function.
  85  * We use the callback in order to allow __rpc_wake_up_task()
  86  * and friends to disable the timer synchronously on SMP systems
  87  * without calling del_timer_sync(). The latter could cause a
  88  * deadlock if called while we're holding spinlocks...
  89  */
  90 static void rpc_run_timer(struct rpc_task *task)
  91 {
  92         void (*callback)(struct rpc_task *);
  93
  94         callback = task->tk_timeout_fn;
  95         task->tk_timeout_fn = NULL;
  96         if (callback && RPC_IS_QUEUED(task)) {
  97                 dprintk("RPC: %5u running timer\n", task->tk_pid);
  98                 callback(task);
  99         }
 100         smp_mb__before_clear_bit();
 101         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 102         smp_mb__after_clear_bit();
 103 }
 104
 105 /*
 106  * Set up a timer for the current task.
 107  */
 108 static inline void
 109 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
 110 {
 111         if (!task->tk_timeout)
 112                 return;
 113
 114         dprintk("RPC: %5u setting alarm for %lu ms\n",
 115                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 116
 117         if (timer)
 118                 task->tk_timeout_fn = timer;
 119         else
 120                 task->tk_timeout_fn = __rpc_default_timer;
 121         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 122         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 123 }
 124
 125 /*
 126  * Delete any timer for the current task. Because we use del_timer_sync(),
 127  * this function should never be called while holding queue->lock.
 128  */
 129 static void
 130 rpc_delete_timer(struct rpc_task *task)
 131 {
 132         if (RPC_IS_QUEUED(task))
 133                 return;
 134         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 135                 del_singleshot_timer_sync(&task->tk_timer);
 136                 dprintk("RPC: %5u deleting timer\n", task->tk_pid);
 137         }
 138 }
 139
 140 /*
 141  * Add new request to a priority queue.
 142  */
 143 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 144 {
 145         struct list_head *q;
 146         struct rpc_task *t;
 147
 148         INIT_LIST_HEAD(&task->u.tk_wait.links);
 149         q = &queue->tasks[task->tk_priority];
 150         if (unlikely(task->tk_priority > queue->maxpriority))
 151                 q = &queue->tasks[queue->maxpriority];
 152         list_for_each_entry(t, q, u.tk_wait.list) {
 153                 if (t->tk_cookie == task->tk_cookie) {
 154                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 155                         return;
 156                 }
 157         }
 158         list_add_tail(&task->u.tk_wait.list, q);
 159 }
 160
 161 /*
 162  * Add new request to wait queue.
 163  *
 164  * Swapper tasks always get inserted at the head of the queue.
 165  * This should avoid many nasty memory deadlocks and hopefully
 166  * improve overall performance.
 167  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 168  */
 169 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 170 {
 171         BUG_ON (RPC_IS_QUEUED(task));
 172
 173         if (RPC_IS_PRIORITY(queue))
 174                 __rpc_add_wait_queue_priority(queue, task);
 175         else if (RPC_IS_SWAPPER(task))
 176                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 177         else
 178                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 179         task->u.tk_wait.rpc_waitq = queue;
 180         queue->qlen++;
 181         rpc_set_queued(task);
 182
 183         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 184                         task->tk_pid, queue, rpc_qname(queue));
 185 }
 186
 187 /*
 188  * Remove request from a priority queue.
 189  */
 190 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 191 {
 192         struct rpc_task *t;
 193
 194         if (!list_empty(&task->u.tk_wait.links)) {
 195                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 196                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 197                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 198         }
 199         list_del(&task->u.tk_wait.list);
 200 }
 201
 202 /*
 203  * Remove request from queue.
 204  * Note: must be called with spin lock held.
 205  */
 206 static void __rpc_remove_wait_queue(struct rpc_task *task)
 207 {
 208         struct rpc_wait_queue *queue;
 209         queue = task->u.tk_wait.rpc_waitq;
 210
 211         if (RPC_IS_PRIORITY(queue))
 212                 __rpc_remove_wait_queue_priority(task);
 213         else
 214                 list_del(&task->u.tk_wait.list);
 215         queue->qlen--;
 216         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 217                         task->tk_pid, queue, rpc_qname(queue));
 218 }
 219
 220 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 221 {
 222         queue->priority = priority;
 223         queue->count = 1 << (priority * 2);
 224 }
 225
 226 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
 227 {
 228         queue->cookie = cookie;
 229         queue->nr = RPC_BATCH_COUNT;
 230 }
 231
 232 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 233 {
 234         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 235         rpc_set_waitqueue_cookie(queue, 0);
 236 }
 237
 238 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
 239 {
 240         int i;
 241
 242         spin_lock_init(&queue->lock);
 243         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 244                 INIT_LIST_HEAD(&queue->tasks[i]);
 245         queue->maxpriority = maxprio;
 246         rpc_reset_waitqueue_priority(queue);
 247 #ifdef RPC_DEBUG
 248         queue->name = qname;
 249 #endif
 250 }
 251
 252 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 253 {
 254         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
 255 }
 256
 257 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 258 {
 259         __rpc_init_priority_wait_queue(queue, qname, 0);
 260 }
 261 EXPORT_SYMBOL(rpc_init_wait_queue);
 262
 263 static int rpc_wait_bit_interruptible(void *word)
 264 {
 265         if (signal_pending(current))
 266                 return -ERESTARTSYS;
 267         schedule();
 268         return 0;
 269 }
 270
 271 static void rpc_set_active(struct rpc_task *task)
 272 {
 273         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 274                 return;
 275         spin_lock(&rpc_sched_lock);
 276 #ifdef RPC_DEBUG
 277         task->tk_magic = RPC_TASK_MAGIC_ID;
 278         task->tk_pid = rpc_task_id++;
 279 #endif
 280         /* Add to global list of all tasks */
 281         if (task->tk_client)
 282                 list_add_tail(&task->tk_task, &task->tk_client->cl_tasks);
 283         spin_unlock(&rpc_sched_lock);
 284 }
 285
 286 /*
 287  * Mark an RPC call as having completed by clearing the 'active' bit
 288  */
 289 static void rpc_mark_complete_task(struct rpc_task *task)
 290 {
 291         smp_mb__before_clear_bit();
 292         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 293         smp_mb__after_clear_bit();
 294         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 295 }
 296
 297 /*
 298  * Allow callers to wait for completion of an RPC call
 299  */
 300 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 301 {
 302         if (action == NULL)
 303                 action = rpc_wait_bit_interruptible;
 304         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 305                         action, TASK_INTERRUPTIBLE);
 306 }
 307 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
 308
 309 /*
 310  * Make an RPC task runnable.
 311  *
 312  * Note: If the task is ASYNC, this must be called with
 313  * the spinlock held to protect the wait queue operation.
 314  */
 315 static void rpc_make_runnable(struct rpc_task *task)
 316 {
 317         BUG_ON(task->tk_timeout_fn);
 318         rpc_clear_queued(task);
 319         if (rpc_test_and_set_running(task))
 320                 return;
 321         /* We might have raced */
 322         if (RPC_IS_QUEUED(task)) {
 323                 rpc_clear_running(task);
 324                 return;
 325         }
 326         if (RPC_IS_ASYNC(task)) {
 327                 int status;
 328
 329                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 330                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 331                 if (status < 0) {
 332                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 333                         task->tk_status = status;
 334                         return;
 335                 }
 336         } else
 337                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 338 }
 339
 340 /*
 341  * Prepare for sleeping on a wait queue.
 342  * By always appending tasks to the list we ensure FIFO behavior.
 343  * NB: An RPC task will only receive interrupt-driven events as long
 344  * as it's on a wait queue.
 345  */
 346 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 347                         rpc_action action, rpc_action timer)
 348 {
 349         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 350                         task->tk_pid, rpc_qname(q), jiffies);
 351
 352         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 353                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 354                 return;
 355         }
 356
 357         __rpc_add_wait_queue(q, task);
 358
 359         BUG_ON(task->tk_callback != NULL);
 360         task->tk_callback = action;
 361         __rpc_add_timer(task, timer);
 362 }
 363
 364 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 365                                 rpc_action action, rpc_action timer)
 366 {
 367         /* Mark the task as being activated if so needed */
 368         rpc_set_active(task);
 369
 370         /*
 371          * Protect the queue operations.
 372          */
 373         spin_lock_bh(&q->lock);
 374         __rpc_sleep_on(q, task, action, timer);
 375         spin_unlock_bh(&q->lock);
 376 }
 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
 449 /*
 450  * Wake up the next task on a priority queue.
 451  */
 452 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 453 {
 454         struct list_head *q;
 455         struct rpc_task *task;
 456
 457         /*
 458          * Service a batch of tasks from a single cookie.
 459          */
 460         q = &queue->tasks[queue->priority];
 461         if (!list_empty(q)) {
 462                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 463                 if (queue->cookie == task->tk_cookie) {
 464                         if (--queue->nr)
 465                                 goto out;
 466                         list_move_tail(&task->u.tk_wait.list, q);
 467                 }
 468                 /*
 469                  * Check if we need to switch queues.
 470                  */
 471                 if (--queue->count)
 472                         goto new_cookie;
 473         }
 474
 475         /*
 476          * Service the next queue.
 477          */
 478         do {
 479                 if (q == &queue->tasks[0])
 480                         q = &queue->tasks[queue->maxpriority];
 481                 else
 482                         q = q - 1;
 483                 if (!list_empty(q)) {
 484                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 485                         goto new_queue;
 486                 }
 487         } while (q != &queue->tasks[queue->priority]);
 488
 489         rpc_reset_waitqueue_priority(queue);
 490         return NULL;
 491
 492 new_queue:
 493         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 494 new_cookie:
 495         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 496 out:
 497         __rpc_wake_up_task(task);
 498         return task;
 499 }
 500
 501 /*
 502  * Wake up the next task on the wait queue.
 503  */
 504 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 505 {
 506         struct rpc_task *task = NULL;
 507
 508         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 509                         queue, rpc_qname(queue));
 510         rcu_read_lock_bh();
 511         spin_lock(&queue->lock);
 512         if (RPC_IS_PRIORITY(queue))
 513                 task = __rpc_wake_up_next_priority(queue);
 514         else {
 515                 task_for_first(task, &queue->tasks[0])
 516                         __rpc_wake_up_task(task);
 517         }
 518         spin_unlock(&queue->lock);
 519         rcu_read_unlock_bh();
 520
 521         return task;
 522 }
 523
 524 /**
 525  * rpc_wake_up - wake up all rpc_tasks
 526  * @queue: rpc_wait_queue on which the tasks are sleeping
 527  *
 528  * Grabs queue->lock
 529  */
 530 void rpc_wake_up(struct rpc_wait_queue *queue)
 531 {
 532         struct rpc_task *task, *next;
 533         struct list_head *head;
 534
 535         rcu_read_lock_bh();
 536         spin_lock(&queue->lock);
 537         head = &queue->tasks[queue->maxpriority];
 538         for (;;) {
 539                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 540                         __rpc_wake_up_task(task);
 541                 if (head == &queue->tasks[0])
 542                         break;
 543                 head--;
 544         }
 545         spin_unlock(&queue->lock);
 546         rcu_read_unlock_bh();
 547 }
 548
 549 /**
 550  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 551  * @queue: rpc_wait_queue on which the tasks are sleeping
 552  * @status: status value to set
 553  *
 554  * Grabs queue->lock
 555  */
 556 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 557 {
 558         struct rpc_task *task, *next;
 559         struct list_head *head;
 560
 561         rcu_read_lock_bh();
 562         spin_lock(&queue->lock);
 563         head = &queue->tasks[queue->maxpriority];
 564         for (;;) {
 565                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 566                         task->tk_status = status;
 567                         __rpc_wake_up_task(task);
 568                 }
 569                 if (head == &queue->tasks[0])
 570                         break;
 571                 head--;
 572         }
 573         spin_unlock(&queue->lock);
 574         rcu_read_unlock_bh();
 575 }
 576
 577 static void __rpc_atrun(struct rpc_task *task)
 578 {
 579         rpc_wake_up_task(task);
 580 }
 581
 582 /*
 583  * Run a task at a later time
 584  */
 585 void rpc_delay(struct rpc_task *task, unsigned long delay)
 586 {
 587         task->tk_timeout = delay;
 588         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 589 }
 590
 591 /*
 592  * Helper to call task->tk_ops->rpc_call_prepare
 593  */
 594 static void rpc_prepare_task(struct rpc_task *task)
 595 {
 596         lock_kernel();
 597         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 598         unlock_kernel();
 599 }
 600
 601 /*
 602  * Helper that calls task->tk_ops->rpc_call_done if it exists
 603  */
 604 void rpc_exit_task(struct rpc_task *task)
 605 {
 606         task->tk_action = NULL;
 607         if (task->tk_ops->rpc_call_done != NULL) {
 608                 lock_kernel();
 609                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 610                 unlock_kernel();
 611                 if (task->tk_action != NULL) {
 612                         WARN_ON(RPC_ASSASSINATED(task));
 613                         /* Always release the RPC slot and buffer memory */
 614                         xprt_release(task);
 615                 }
 616         }
 617 }
 618 EXPORT_SYMBOL(rpc_exit_task);
 619
 620 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 621 {
 622         if (ops->rpc_release != NULL) {
 623                 lock_kernel();
 624                 ops->rpc_release(calldata);
 625                 unlock_kernel();
 626         }
 627 }
 628
 629 /*
 630  * This is the RPC `scheduler' (or rather, the finite state machine).
 631  */
 632 static void __rpc_execute(struct rpc_task *task)
 633 {
 634         int             status = 0;
 635
 636         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 637                         task->tk_pid, task->tk_flags);
 638
 639         BUG_ON(RPC_IS_QUEUED(task));
 640
 641         for (;;) {
 642                 /*
 643                  * Garbage collection of pending timers...
 644                  */
 645                 rpc_delete_timer(task);
 646
 647                 /*
 648                  * Execute any pending callback.
 649                  */
 650                 if (RPC_DO_CALLBACK(task)) {
 651                         /* Define a callback save pointer */
 652                         void (*save_callback)(struct rpc_task *);
 653
 654                         /*
 655                          * If a callback exists, save it, reset it,
 656                          * call it.
 657                          * The save is needed to stop from resetting
 658                          * another callback set within the callback handler
 659                          * - Dave
 660                          */
 661                         save_callback=task->tk_callback;
 662                         task->tk_callback=NULL;
 663                         save_callback(task);
 664                 }
 665
 666                 /*
 667                  * Perform the next FSM step.
 668                  * tk_action may be NULL when the task has been killed
 669                  * by someone else.
 670                  */
 671                 if (!RPC_IS_QUEUED(task)) {
 672                         if (task->tk_action == NULL)
 673                                 break;
 674                         task->tk_action(task);
 675                 }
 676
 677                 /*
 678                  * Lockless check for whether task is sleeping or not.
 679                  */
 680                 if (!RPC_IS_QUEUED(task))
 681                         continue;
 682                 rpc_clear_running(task);
 683                 if (RPC_IS_ASYNC(task)) {
 684                         /* Careful! we may have raced... */
 685                         if (RPC_IS_QUEUED(task))
 686                                 return;
 687                         if (rpc_test_and_set_running(task))
 688                                 return;
 689                         continue;
 690                 }
 691
 692                 /* sync task: sleep here */
 693                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 694                 /* Note: Caller should be using rpc_clnt_sigmask() */
 695                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 696                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
 697                                 TASK_INTERRUPTIBLE);
 698                 if (status == -ERESTARTSYS) {
 699                         /*
 700                          * When a sync task receives a signal, it exits with
 701                          * -ERESTARTSYS. In order to catch any callbacks that
 702                          * clean up after sleeping on some queue, we don't
 703                          * break the loop here, but go around once more.
 704                          */
 705                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 706                         task->tk_flags |= RPC_TASK_KILLED;
 707                         rpc_exit(task, -ERESTARTSYS);
 708                         rpc_wake_up_task(task);
 709                 }
 710                 rpc_set_running(task);
 711                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 712         }
 713
 714         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 715                         task->tk_status);
 716         /* Release all resources associated with the task */
 717         rpc_release_task(task);
 718 }
 719
 720 /*
 721  * User-visible entry point to the scheduler.
 722  *
 723  * This may be called recursively if e.g. an async NFS task updates
 724  * the attributes and finds that dirty pages must be flushed.
 725  * NOTE: Upon exit of this function the task is guaranteed to be
 726  *       released. In particular note that tk_release() will have
 727  *       been called, so your task memory may have been freed.
 728  */
 729 void rpc_execute(struct rpc_task *task)
 730 {
 731         rpc_set_active(task);
 732         rpc_set_running(task);
 733         __rpc_execute(task);
 734 }
 735
 736 static void rpc_async_schedule(struct work_struct *work)
 737 {
 738         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 739 }
 740
 741 struct rpc_buffer {
 742         size_t  len;
 743         char    data[];
 744 };
 745
 746 /**
 747  * rpc_malloc - allocate an RPC buffer
 748  * @task: RPC task that will use this buffer
 749  * @size: requested byte size
 750  *
 751  * To prevent rpciod from hanging, this allocator never sleeps,
 752  * returning NULL if the request cannot be serviced immediately.
 753  * The caller can arrange to sleep in a way that is safe for rpciod.
 754  *
 755  * Most requests are 'small' (under 2KiB) and can be serviced from a
 756  * mempool, ensuring that NFS reads and writes can always proceed,
 757  * and that there is good locality of reference for these buffers.
 758  *
 759  * In order to avoid memory starvation triggering more writebacks of
 760  * NFS requests, we avoid using GFP_KERNEL.
 761  */
 762 void *rpc_malloc(struct rpc_task *task, size_t size)
 763 {
 764         struct rpc_buffer *buf;
 765         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 766
 767         size += sizeof(struct rpc_buffer);
 768         if (size <= RPC_BUFFER_MAXSIZE)
 769                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 770         else
 771                 buf = kmalloc(size, gfp);
 772
 773         if (!buf)
 774                 return NULL;
 775
 776         buf->len = size;
 777         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 778                         task->tk_pid, size, buf);
 779         return &buf->data;
 780 }
 781
 782 /**
 783  * rpc_free - free buffer allocated via rpc_malloc
 784  * @buffer: buffer to free
 785  *
 786  */
 787 void rpc_free(void *buffer)
 788 {
 789         size_t size;
 790         struct rpc_buffer *buf;
 791
 792         if (!buffer)
 793                 return;
 794
 795         buf = container_of(buffer, struct rpc_buffer, data);
 796         size = buf->len;
 797
 798         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 799                         size, buf);
 800
 801         if (size <= RPC_BUFFER_MAXSIZE)
 802                 mempool_free(buf, rpc_buffer_mempool);
 803         else
 804                 kfree(buf);
 805 }
 806
 807 /*
 808  * Creation and deletion of RPC task structures
 809  */
 810 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 811 {
 812         memset(task, 0, sizeof(*task));
 813         init_timer(&task->tk_timer);
 814         task->tk_timer.data     = (unsigned long) task;
 815         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 816         atomic_set(&task->tk_count, 1);
 817         task->tk_client = clnt;
 818         task->tk_flags  = flags;
 819         task->tk_ops = tk_ops;
 820         if (tk_ops->rpc_call_prepare != NULL)
 821                 task->tk_action = rpc_prepare_task;
 822         task->tk_calldata = calldata;
 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 = RPC_PRIORITY_NORMAL;
 830         task->tk_cookie = (unsigned long)current;
 831
 832         /* Initialize workqueue for async tasks */
 833         task->tk_workqueue = rpciod_workqueue;
 834
 835         if (clnt) {
 836                 atomic_inc(&clnt->cl_users);
 837                 if (clnt->cl_softrtry)
 838                         task->tk_flags |= RPC_TASK_SOFT;
 839                 if (!clnt->cl_intr)
 840                         task->tk_flags |= RPC_TASK_NOINTR;
 841         }
 842
 843         BUG_ON(task->tk_ops == NULL);
 844
 845         /* starting timestamp */
 846         task->tk_start = jiffies;
 847
 848         dprintk("RPC:       new task initialized, procpid %u\n",
 849                                 current->pid);
 850 }
 851
 852 static struct rpc_task *
 853 rpc_alloc_task(void)
 854 {
 855         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 856 }
 857
 858 static void rpc_free_task(struct rcu_head *rcu)
 859 {
 860         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 861         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 862         mempool_free(task, rpc_task_mempool);
 863 }
 864
 865 /*
 866  * Create a new task for the specified client.  We have to
 867  * clean up after an allocation failure, as the client may
 868  * have specified "oneshot".
 869  */
 870 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 871 {
 872         struct rpc_task *task;
 873
 874         task = rpc_alloc_task();
 875         if (!task)
 876                 goto cleanup;
 877
 878         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 879
 880         dprintk("RPC:       allocated task %p\n", task);
 881         task->tk_flags |= RPC_TASK_DYNAMIC;
 882 out:
 883         return task;
 884
 885 cleanup:
 886         /* Check whether to release the client */
 887         if (clnt) {
 888                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 889                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 890                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 891                 rpc_release_client(clnt);
 892         }
 893         goto out;
 894 }
 895
 896
 897 void rpc_put_task(struct rpc_task *task)
 898 {
 899         const struct rpc_call_ops *tk_ops = task->tk_ops;
 900         void *calldata = task->tk_calldata;
 901
 902         if (!atomic_dec_and_test(&task->tk_count))
 903                 return;
 904         /* Release resources */
 905         if (task->tk_rqstp)
 906                 xprt_release(task);
 907         if (task->tk_msg.rpc_cred)
 908                 rpcauth_unbindcred(task);
 909         if (task->tk_client) {
 910                 rpc_release_client(task->tk_client);
 911                 task->tk_client = NULL;
 912         }
 913         if (task->tk_flags & RPC_TASK_DYNAMIC)
 914                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task);
 915         rpc_release_calldata(tk_ops, calldata);
 916 }
 917 EXPORT_SYMBOL(rpc_put_task);
 918
 919 static void rpc_release_task(struct rpc_task *task)
 920 {
 921 #ifdef RPC_DEBUG
 922         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 923 #endif
 924         dprintk("RPC: %5u release task\n", task->tk_pid);
 925
 926         if (!list_empty(&task->tk_task)) {
 927                 /* Remove from client task list */
 928                 spin_lock(&rpc_sched_lock);
 929                 list_del(&task->tk_task);
 930                 spin_unlock(&rpc_sched_lock);
 931         }
 932         BUG_ON (RPC_IS_QUEUED(task));
 933
 934         /* Synchronously delete any running timer */
 935         rpc_delete_timer(task);
 936
 937 #ifdef RPC_DEBUG
 938         task->tk_magic = 0;
 939 #endif
 940         /* Wake up anyone who is waiting for task completion */
 941         rpc_mark_complete_task(task);
 942
 943         rpc_put_task(task);
 944 }
 945
 946 /**
 947  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
 948  * @clnt: pointer to RPC client
 949  * @flags: RPC flags
 950  * @ops: RPC call ops
 951  * @data: user call data
 952  */
 953 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
 954                                         const struct rpc_call_ops *ops,
 955                                         void *data)
 956 {
 957         struct rpc_task *task;
 958         task = rpc_new_task(clnt, flags, ops, data);
 959         if (task == NULL) {
 960                 rpc_release_calldata(ops, data);
 961                 return ERR_PTR(-ENOMEM);
 962         }
 963         atomic_inc(&task->tk_count);
 964         rpc_execute(task);
 965         return task;
 966 }
 967 EXPORT_SYMBOL(rpc_run_task);
 968
 969 /*
 970  * Kill all tasks for the given client.
 971  * XXX: kill their descendants as well?
 972  */
 973 static void rpc_killall_tasks_locked(struct list_head *head)
 974 {
 975         struct rpc_task *rovr;
 976
 977
 978         list_for_each_entry(rovr, head, tk_task) {
 979                 if (! RPC_IS_ACTIVATED(rovr))
 980                         continue;
 981                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 982                         rovr->tk_flags |= RPC_TASK_KILLED;
 983                         rpc_exit(rovr, -EIO);
 984                         rpc_wake_up_task(rovr);
 985                 }
 986         }
 987 }
 988
 989 void rpc_killall_tasks(struct rpc_clnt *clnt)
 990 {
 991         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 992         /*
 993          * Spin lock all_tasks to prevent changes...
 994          */
 995         spin_lock(&rpc_sched_lock);
 996         rpc_killall_tasks_locked(&clnt->cl_tasks);
 997         spin_unlock(&rpc_sched_lock);
 998 }
 999
1000 static void rpciod_killall(void)
1001 {
1002         struct rpc_clnt *clnt;
1003         unsigned long flags;
1004
1005         for(;;) {
1006                 clear_thread_flag(TIF_SIGPENDING);
1007
1008                 spin_lock(&rpc_sched_lock);
1009                 list_for_each_entry(clnt, &all_clients, cl_clients)
1010                         rpc_killall_tasks_locked(&clnt->cl_tasks);
1011                 spin_unlock(&rpc_sched_lock);
1012                 flush_workqueue(rpciod_workqueue);
1013                 if (!list_empty(&all_clients))
1014                         break;
1015                 dprintk("RPC:       rpciod_killall: waiting for tasks "
1016                                         "to exit\n");
1017                 wait_event_timeout(client_kill_wait,
1018                                 list_empty(&all_clients), 1*HZ);
1019         }
1020
1021         spin_lock_irqsave(&current->sighand->siglock, flags);
1022         recalc_sigpending();
1023         spin_unlock_irqrestore(&current->sighand->siglock, flags);
1024 }
1025
1026 void rpc_register_client(struct rpc_clnt *clnt)
1027 {
1028         spin_lock(&rpc_sched_lock);
1029         list_add(&clnt->cl_clients, &all_clients);
1030         spin_unlock(&rpc_sched_lock);
1031 }
1032
1033 void rpc_unregister_client(struct rpc_clnt *clnt)
1034 {
1035         spin_lock(&rpc_sched_lock);
1036         list_del(&clnt->cl_clients);
1037         if (list_empty(&all_clients))
1038                 wake_up(&client_kill_wait);
1039         spin_unlock(&rpc_sched_lock);
1040 }
1041
1042 /*
1043  * Start up the rpciod process if it's not already running.
1044  */
1045 int
1046 rpciod_up(void)
1047 {
1048         struct workqueue_struct *wq;
1049         int error = 0;
1050
1051         mutex_lock(&rpciod_mutex);
1052         dprintk("RPC:       rpciod_up: users %u\n", rpciod_users);
1053         rpciod_users++;
1054         if (rpciod_workqueue)
1055                 goto out;
1056         /*
1057          * If there's no pid, we should be the first user.
1058          */
1059         if (rpciod_users > 1)
1060                 printk(KERN_WARNING "rpciod_up: no workqueue, %u users??\n", rpciod_users);
1061         /*
1062          * Create the rpciod thread and wait for it to start.
1063          */
1064         error = -ENOMEM;
1065         wq = create_workqueue("rpciod");
1066         if (wq == NULL) {
1067                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1068                 rpciod_users--;
1069                 goto out;
1070         }
1071         rpciod_workqueue = wq;
1072         error = 0;
1073 out:
1074         mutex_unlock(&rpciod_mutex);
1075         return error;
1076 }
1077
1078 void
1079 rpciod_down(void)
1080 {
1081         mutex_lock(&rpciod_mutex);
1082         dprintk("RPC:       rpciod_down sema %u\n", rpciod_users);
1083         if (rpciod_users) {
1084                 if (--rpciod_users)
1085                         goto out;
1086         } else
1087                 printk(KERN_WARNING "rpciod_down: no users??\n");
1088
1089         if (!rpciod_workqueue) {
1090                 dprintk("RPC:       rpciod_down: Nothing to do!\n");
1091                 goto out;
1092         }
1093         rpciod_killall();
1094
1095         destroy_workqueue(rpciod_workqueue);
1096         rpciod_workqueue = NULL;
1097  out:
1098         mutex_unlock(&rpciod_mutex);
1099 }
1100
1101 #ifdef RPC_DEBUG
1102 void rpc_show_tasks(void)
1103 {
1104         struct rpc_clnt *clnt;
1105         struct rpc_task *t;
1106
1107         spin_lock(&rpc_sched_lock);
1108         if (list_empty(&all_clients))
1109                 goto out;
1110         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1111                 "-rpcwait -action- ---ops--\n");
1112         list_for_each_entry(clnt, &all_clients, cl_clients) {
1113                 list_for_each_entry(t, &clnt->cl_tasks, tk_task) {
1114                         const char *rpc_waitq = "none";
1115
1116                         if (RPC_IS_QUEUED(t))
1117                                 rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1118
1119                         printk("%5u %04d %04x %6d %8p %6d %8p %8ld %8s %8p %8p\n",
1120                                 t->tk_pid,
1121                                 (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1122                                 t->tk_flags, t->tk_status,
1123                                 t->tk_client,
1124                                 (t->tk_client ? t->tk_client->cl_prog : 0),
1125                                 t->tk_rqstp, t->tk_timeout,
1126                                 rpc_waitq,
1127                                 t->tk_action, t->tk_ops);
1128                 }
1129         }
1130 out:
1131         spin_unlock(&rpc_sched_lock);
1132 }
1133 #endif
1134
1135 void
1136 rpc_destroy_mempool(void)
1137 {
1138         if (rpc_buffer_mempool)
1139                 mempool_destroy(rpc_buffer_mempool);
1140         if (rpc_task_mempool)
1141                 mempool_destroy(rpc_task_mempool);
1142         if (rpc_task_slabp)
1143                 kmem_cache_destroy(rpc_task_slabp);
1144         if (rpc_buffer_slabp)
1145                 kmem_cache_destroy(rpc_buffer_slabp);
1146 }
1147
1148 int
1149 rpc_init_mempool(void)
1150 {
1151         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1152                                              sizeof(struct rpc_task),
1153                                              0, SLAB_HWCACHE_ALIGN,
1154                                              NULL, NULL);
1155         if (!rpc_task_slabp)
1156                 goto err_nomem;
1157         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1158                                              RPC_BUFFER_MAXSIZE,
1159                                              0, SLAB_HWCACHE_ALIGN,
1160                                              NULL, NULL);
1161         if (!rpc_buffer_slabp)
1162                 goto err_nomem;
1163         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1164                                                     rpc_task_slabp);
1165         if (!rpc_task_mempool)
1166                 goto err_nomem;
1167         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1168                                                       rpc_buffer_slabp);
1169         if (!rpc_buffer_mempool)
1170                 goto err_nomem;
1171         return 0;
1172 err_nomem:
1173         rpc_destroy_mempool();
1174         return -ENOMEM;
1175 }