err.no Git - linux-2.6/blob - net/sunrpc/sched.c

   1 /*
   2  * linux/net/sunrpc/sched.c
   3  *
   4  * Scheduling for synchronous and asynchronous RPC requests.
   5  *
   6  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  10  */
  11
  12 #include <linux/module.h>
  13
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/slab.h>
  17 #include <linux/mempool.h>
  18 #include <linux/smp.h>
  19 #include <linux/smp_lock.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/mutex.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #ifdef RPC_DEBUG
  26 #define RPCDBG_FACILITY         RPCDBG_SCHED
  27 #define RPC_TASK_MAGIC_ID       0xf00baa
  28 #endif
  29
  30 /*
  31  * RPC slabs and memory pools
  32  */
  33 #define RPC_BUFFER_MAXSIZE      (2048)
  34 #define RPC_BUFFER_POOLSIZE     (8)
  35 #define RPC_TASK_POOLSIZE       (8)
  36 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  37 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  38 static mempool_t        *rpc_task_mempool __read_mostly;
  39 static mempool_t        *rpc_buffer_mempool __read_mostly;
  40
  41 static void                     rpc_async_schedule(struct work_struct *);
  42 static void                      rpc_release_task(struct rpc_task *task);
  43 static void __rpc_queue_timer_fn(unsigned long ptr);
  44
  45 /*
  46  * RPC tasks sit here while waiting for conditions to improve.
  47  */
  48 static struct rpc_wait_queue delay_queue;
  49
  50 /*
  51  * rpciod-related stuff
  52  */
  53 struct workqueue_struct *rpciod_workqueue;
  54
  55 /*
  56  * Disable the timer for a given RPC task. Should be called with
  57  * queue->lock and bh_disabled in order to avoid races within
  58  * rpc_run_timer().
  59  */
  60 static void
  61 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  62 {
  63         if (task->tk_timeout == 0)
  64                 return;
  65         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  66         task->tk_timeout = 0;
  67         list_del(&task->u.tk_wait.timer_list);
  68         if (list_empty(&queue->timer_list.list))
  69                 del_timer(&queue->timer_list.timer);
  70 }
  71
  72 static void
  73 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  74 {
  75         queue->timer_list.expires = expires;
  76         mod_timer(&queue->timer_list.timer, expires);
  77 }
  78
  79 /*
  80  * Set up a timer for the current task.
  81  */
  82 static void
  83 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  84 {
  85         if (!task->tk_timeout)
  86                 return;
  87
  88         dprintk("RPC: %5u setting alarm for %lu ms\n",
  89                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  90
  91         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  92         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  93                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  94         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  95 }
  96
  97 /*
  98  * Add new request to a priority queue.
  99  */
 100 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 101 {
 102         struct list_head *q;
 103         struct rpc_task *t;
 104
 105         INIT_LIST_HEAD(&task->u.tk_wait.links);
 106         q = &queue->tasks[task->tk_priority];
 107         if (unlikely(task->tk_priority > queue->maxpriority))
 108                 q = &queue->tasks[queue->maxpriority];
 109         list_for_each_entry(t, q, u.tk_wait.list) {
 110                 if (t->tk_owner == task->tk_owner) {
 111                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 112                         return;
 113                 }
 114         }
 115         list_add_tail(&task->u.tk_wait.list, q);
 116 }
 117
 118 /*
 119  * Add new request to wait queue.
 120  *
 121  * Swapper tasks always get inserted at the head of the queue.
 122  * This should avoid many nasty memory deadlocks and hopefully
 123  * improve overall performance.
 124  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 125  */
 126 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 127 {
 128         BUG_ON (RPC_IS_QUEUED(task));
 129
 130         if (RPC_IS_PRIORITY(queue))
 131                 __rpc_add_wait_queue_priority(queue, task);
 132         else if (RPC_IS_SWAPPER(task))
 133                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 134         else
 135                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 136         task->tk_waitqueue = queue;
 137         queue->qlen++;
 138         rpc_set_queued(task);
 139
 140         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 141                         task->tk_pid, queue, rpc_qname(queue));
 142 }
 143
 144 /*
 145  * Remove request from a priority queue.
 146  */
 147 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 148 {
 149         struct rpc_task *t;
 150
 151         if (!list_empty(&task->u.tk_wait.links)) {
 152                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 153                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 154                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 155         }
 156 }
 157
 158 /*
 159  * Remove request from queue.
 160  * Note: must be called with spin lock held.
 161  */
 162 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 163 {
 164         __rpc_disable_timer(queue, task);
 165         if (RPC_IS_PRIORITY(queue))
 166                 __rpc_remove_wait_queue_priority(task);
 167         list_del(&task->u.tk_wait.list);
 168         queue->qlen--;
 169         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 170                         task->tk_pid, queue, rpc_qname(queue));
 171 }
 172
 173 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 174 {
 175         queue->priority = priority;
 176         queue->count = 1 << (priority * 2);
 177 }
 178
 179 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 180 {
 181         queue->owner = pid;
 182         queue->nr = RPC_BATCH_COUNT;
 183 }
 184
 185 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 186 {
 187         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 188         rpc_set_waitqueue_owner(queue, 0);
 189 }
 190
 191 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 192 {
 193         int i;
 194
 195         spin_lock_init(&queue->lock);
 196         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 197                 INIT_LIST_HEAD(&queue->tasks[i]);
 198         queue->maxpriority = nr_queues - 1;
 199         rpc_reset_waitqueue_priority(queue);
 200         queue->qlen = 0;
 201         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 202         INIT_LIST_HEAD(&queue->timer_list.list);
 203 #ifdef RPC_DEBUG
 204         queue->name = qname;
 205 #endif
 206 }
 207
 208 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 209 {
 210         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 211 }
 212
 213 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 214 {
 215         __rpc_init_priority_wait_queue(queue, qname, 1);
 216 }
 217 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 218
 219 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 220 {
 221         del_timer_sync(&queue->timer_list.timer);
 222 }
 223 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 224
 225 static int rpc_wait_bit_killable(void *word)
 226 {
 227         if (fatal_signal_pending(current))
 228                 return -ERESTARTSYS;
 229         schedule();
 230         return 0;
 231 }
 232
 233 #ifdef RPC_DEBUG
 234 static void rpc_task_set_debuginfo(struct rpc_task *task)
 235 {
 236         static atomic_t rpc_pid;
 237
 238         task->tk_magic = RPC_TASK_MAGIC_ID;
 239         task->tk_pid = atomic_inc_return(&rpc_pid);
 240 }
 241 #else
 242 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 243 {
 244 }
 245 #endif
 246
 247 static void rpc_set_active(struct rpc_task *task)
 248 {
 249         struct rpc_clnt *clnt;
 250         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 251                 return;
 252         rpc_task_set_debuginfo(task);
 253         /* Add to global list of all tasks */
 254         clnt = task->tk_client;
 255         if (clnt != NULL) {
 256                 spin_lock(&clnt->cl_lock);
 257                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 258                 spin_unlock(&clnt->cl_lock);
 259         }
 260 }
 261
 262 /*
 263  * Mark an RPC call as having completed by clearing the 'active' bit
 264  */
 265 static void rpc_mark_complete_task(struct rpc_task *task)
 266 {
 267         smp_mb__before_clear_bit();
 268         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 269         smp_mb__after_clear_bit();
 270         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 271 }
 272
 273 /*
 274  * Allow callers to wait for completion of an RPC call
 275  */
 276 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 277 {
 278         if (action == NULL)
 279                 action = rpc_wait_bit_killable;
 280         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 281                         action, TASK_KILLABLE);
 282 }
 283 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 284
 285 /*
 286  * Make an RPC task runnable.
 287  *
 288  * Note: If the task is ASYNC, this must be called with
 289  * the spinlock held to protect the wait queue operation.
 290  */
 291 static void rpc_make_runnable(struct rpc_task *task)
 292 {
 293         rpc_clear_queued(task);
 294         if (rpc_test_and_set_running(task))
 295                 return;
 296         /* We might have raced */
 297         if (RPC_IS_QUEUED(task)) {
 298                 rpc_clear_running(task);
 299                 return;
 300         }
 301         if (RPC_IS_ASYNC(task)) {
 302                 int status;
 303
 304                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 305                 status = queue_work(rpciod_workqueue, &task->u.tk_work);
 306                 if (status < 0) {
 307                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 308                         task->tk_status = status;
 309                         return;
 310                 }
 311         } else
 312                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 313 }
 314
 315 /*
 316  * Prepare for sleeping on a wait queue.
 317  * By always appending tasks to the list we ensure FIFO behavior.
 318  * NB: An RPC task will only receive interrupt-driven events as long
 319  * as it's on a wait queue.
 320  */
 321 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 322                         rpc_action action)
 323 {
 324         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 325                         task->tk_pid, rpc_qname(q), jiffies);
 326
 327         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 328                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 329                 return;
 330         }
 331
 332         __rpc_add_wait_queue(q, task);
 333
 334         BUG_ON(task->tk_callback != NULL);
 335         task->tk_callback = action;
 336         __rpc_add_timer(q, task);
 337 }
 338
 339 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 340                                 rpc_action action)
 341 {
 342         /* Mark the task as being activated if so needed */
 343         rpc_set_active(task);
 344
 345         /*
 346          * Protect the queue operations.
 347          */
 348         spin_lock_bh(&q->lock);
 349         __rpc_sleep_on(q, task, action);
 350         spin_unlock_bh(&q->lock);
 351 }
 352 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 353
 354 /**
 355  * __rpc_do_wake_up_task - wake up a single rpc_task
 356  * @queue: wait queue
 357  * @task: task to be woken up
 358  *
 359  * Caller must hold queue->lock, and have cleared the task queued flag.
 360  */
 361 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 362 {
 363         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 364                         task->tk_pid, jiffies);
 365
 366 #ifdef RPC_DEBUG
 367         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 368 #endif
 369         /* Has the task been executed yet? If not, we cannot wake it up! */
 370         if (!RPC_IS_ACTIVATED(task)) {
 371                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 372                 return;
 373         }
 374
 375         __rpc_remove_wait_queue(queue, task);
 376
 377         rpc_make_runnable(task);
 378
 379         dprintk("RPC:       __rpc_wake_up_task done\n");
 380 }
 381
 382 /*
 383  * Wake up a queued task while the queue lock is being held
 384  */
 385 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 386 {
 387         if (!RPC_IS_QUEUED(task) || task->tk_waitqueue != queue)
 388                 return;
 389         if (rpc_start_wakeup(task)) {
 390                         __rpc_do_wake_up_task(queue, task);
 391                 rpc_finish_wakeup(task);
 392         }
 393 }
 394
 395 /*
 396  * Wake up a task on a specific queue
 397  */
 398 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 399 {
 400         rcu_read_lock_bh();
 401         spin_lock(&queue->lock);
 402         rpc_wake_up_task_queue_locked(queue, task);
 403         spin_unlock(&queue->lock);
 404         rcu_read_unlock_bh();
 405 }
 406 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 407
 408 /*
 409  * Wake up the specified task
 410  */
 411 static void rpc_wake_up_task(struct rpc_task *task)
 412 {
 413         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 414 }
 415
 416 /*
 417  * Wake up the next task on a priority queue.
 418  */
 419 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 420 {
 421         struct list_head *q;
 422         struct rpc_task *task;
 423
 424         /*
 425          * Service a batch of tasks from a single owner.
 426          */
 427         q = &queue->tasks[queue->priority];
 428         if (!list_empty(q)) {
 429                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 430                 if (queue->owner == task->tk_owner) {
 431                         if (--queue->nr)
 432                                 goto out;
 433                         list_move_tail(&task->u.tk_wait.list, q);
 434                 }
 435                 /*
 436                  * Check if we need to switch queues.
 437                  */
 438                 if (--queue->count)
 439                         goto new_owner;
 440         }
 441
 442         /*
 443          * Service the next queue.
 444          */
 445         do {
 446                 if (q == &queue->tasks[0])
 447                         q = &queue->tasks[queue->maxpriority];
 448                 else
 449                         q = q - 1;
 450                 if (!list_empty(q)) {
 451                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 452                         goto new_queue;
 453                 }
 454         } while (q != &queue->tasks[queue->priority]);
 455
 456         rpc_reset_waitqueue_priority(queue);
 457         return NULL;
 458
 459 new_queue:
 460         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 461 new_owner:
 462         rpc_set_waitqueue_owner(queue, task->tk_owner);
 463 out:
 464         rpc_wake_up_task_queue_locked(queue, task);
 465         return task;
 466 }
 467
 468 /*
 469  * Wake up the next task on the wait queue.
 470  */
 471 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 472 {
 473         struct rpc_task *task = NULL;
 474
 475         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 476                         queue, rpc_qname(queue));
 477         rcu_read_lock_bh();
 478         spin_lock(&queue->lock);
 479         if (RPC_IS_PRIORITY(queue))
 480                 task = __rpc_wake_up_next_priority(queue);
 481         else {
 482                 task_for_first(task, &queue->tasks[0])
 483                         rpc_wake_up_task_queue_locked(queue, task);
 484         }
 485         spin_unlock(&queue->lock);
 486         rcu_read_unlock_bh();
 487
 488         return task;
 489 }
 490 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 491
 492 /**
 493  * rpc_wake_up - wake up all rpc_tasks
 494  * @queue: rpc_wait_queue on which the tasks are sleeping
 495  *
 496  * Grabs queue->lock
 497  */
 498 void rpc_wake_up(struct rpc_wait_queue *queue)
 499 {
 500         struct rpc_task *task, *next;
 501         struct list_head *head;
 502
 503         rcu_read_lock_bh();
 504         spin_lock(&queue->lock);
 505         head = &queue->tasks[queue->maxpriority];
 506         for (;;) {
 507                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 508                         rpc_wake_up_task_queue_locked(queue, task);
 509                 if (head == &queue->tasks[0])
 510                         break;
 511                 head--;
 512         }
 513         spin_unlock(&queue->lock);
 514         rcu_read_unlock_bh();
 515 }
 516 EXPORT_SYMBOL_GPL(rpc_wake_up);
 517
 518 /**
 519  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 520  * @queue: rpc_wait_queue on which the tasks are sleeping
 521  * @status: status value to set
 522  *
 523  * Grabs queue->lock
 524  */
 525 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 526 {
 527         struct rpc_task *task, *next;
 528         struct list_head *head;
 529
 530         rcu_read_lock_bh();
 531         spin_lock(&queue->lock);
 532         head = &queue->tasks[queue->maxpriority];
 533         for (;;) {
 534                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 535                         task->tk_status = status;
 536                         rpc_wake_up_task_queue_locked(queue, task);
 537                 }
 538                 if (head == &queue->tasks[0])
 539                         break;
 540                 head--;
 541         }
 542         spin_unlock(&queue->lock);
 543         rcu_read_unlock_bh();
 544 }
 545 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 546
 547 static void __rpc_queue_timer_fn(unsigned long ptr)
 548 {
 549         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 550         struct rpc_task *task, *n;
 551         unsigned long expires, now, timeo;
 552
 553         spin_lock(&queue->lock);
 554         expires = now = jiffies;
 555         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 556                 timeo = task->u.tk_wait.expires;
 557                 if (time_after_eq(now, timeo)) {
 558                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 559                         task->tk_status = -ETIMEDOUT;
 560                         rpc_wake_up_task_queue_locked(queue, task);
 561                         continue;
 562                 }
 563                 if (expires == now || time_after(expires, timeo))
 564                         expires = timeo;
 565         }
 566         if (!list_empty(&queue->timer_list.list))
 567                 rpc_set_queue_timer(queue, expires);
 568         spin_unlock(&queue->lock);
 569 }
 570
 571 static void __rpc_atrun(struct rpc_task *task)
 572 {
 573         task->tk_status = 0;
 574 }
 575
 576 /*
 577  * Run a task at a later time
 578  */
 579 void rpc_delay(struct rpc_task *task, unsigned long delay)
 580 {
 581         task->tk_timeout = delay;
 582         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 583 }
 584 EXPORT_SYMBOL_GPL(rpc_delay);
 585
 586 /*
 587  * Helper to call task->tk_ops->rpc_call_prepare
 588  */
 589 static void rpc_prepare_task(struct rpc_task *task)
 590 {
 591         lock_kernel();
 592         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 593         unlock_kernel();
 594 }
 595
 596 /*
 597  * Helper that calls task->tk_ops->rpc_call_done if it exists
 598  */
 599 void rpc_exit_task(struct rpc_task *task)
 600 {
 601         task->tk_action = NULL;
 602         if (task->tk_ops->rpc_call_done != NULL) {
 603                 lock_kernel();
 604                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 605                 unlock_kernel();
 606                 if (task->tk_action != NULL) {
 607                         WARN_ON(RPC_ASSASSINATED(task));
 608                         /* Always release the RPC slot and buffer memory */
 609                         xprt_release(task);
 610                 }
 611         }
 612 }
 613 EXPORT_SYMBOL_GPL(rpc_exit_task);
 614
 615 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 616 {
 617         if (ops->rpc_release != NULL) {
 618                 lock_kernel();
 619                 ops->rpc_release(calldata);
 620                 unlock_kernel();
 621         }
 622 }
 623
 624 /*
 625  * This is the RPC `scheduler' (or rather, the finite state machine).
 626  */
 627 static void __rpc_execute(struct rpc_task *task)
 628 {
 629         int             status = 0;
 630
 631         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 632                         task->tk_pid, task->tk_flags);
 633
 634         BUG_ON(RPC_IS_QUEUED(task));
 635
 636         for (;;) {
 637
 638                 /*
 639                  * Execute any pending callback.
 640                  */
 641                 if (RPC_DO_CALLBACK(task)) {
 642                         /* Define a callback save pointer */
 643                         void (*save_callback)(struct rpc_task *);
 644
 645                         /*
 646                          * If a callback exists, save it, reset it,
 647                          * call it.
 648                          * The save is needed to stop from resetting
 649                          * another callback set within the callback handler
 650                          * - Dave
 651                          */
 652                         save_callback=task->tk_callback;
 653                         task->tk_callback=NULL;
 654                         save_callback(task);
 655                 }
 656
 657                 /*
 658                  * Perform the next FSM step.
 659                  * tk_action may be NULL when the task has been killed
 660                  * by someone else.
 661                  */
 662                 if (!RPC_IS_QUEUED(task)) {
 663                         if (task->tk_action == NULL)
 664                                 break;
 665                         task->tk_action(task);
 666                 }
 667
 668                 /*
 669                  * Lockless check for whether task is sleeping or not.
 670                  */
 671                 if (!RPC_IS_QUEUED(task))
 672                         continue;
 673                 rpc_clear_running(task);
 674                 if (RPC_IS_ASYNC(task)) {
 675                         /* Careful! we may have raced... */
 676                         if (RPC_IS_QUEUED(task))
 677                                 return;
 678                         if (rpc_test_and_set_running(task))
 679                                 return;
 680                         continue;
 681                 }
 682
 683                 /* sync task: sleep here */
 684                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 685                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 686                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 687                                 TASK_KILLABLE);
 688                 if (status == -ERESTARTSYS) {
 689                         /*
 690                          * When a sync task receives a signal, it exits with
 691                          * -ERESTARTSYS. In order to catch any callbacks that
 692                          * clean up after sleeping on some queue, we don't
 693                          * break the loop here, but go around once more.
 694                          */
 695                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 696                         task->tk_flags |= RPC_TASK_KILLED;
 697                         rpc_exit(task, -ERESTARTSYS);
 698                         rpc_wake_up_task(task);
 699                 }
 700                 rpc_set_running(task);
 701                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 702         }
 703
 704         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 705                         task->tk_status);
 706         /* Release all resources associated with the task */
 707         rpc_release_task(task);
 708 }
 709
 710 /*
 711  * User-visible entry point to the scheduler.
 712  *
 713  * This may be called recursively if e.g. an async NFS task updates
 714  * the attributes and finds that dirty pages must be flushed.
 715  * NOTE: Upon exit of this function the task is guaranteed to be
 716  *       released. In particular note that tk_release() will have
 717  *       been called, so your task memory may have been freed.
 718  */
 719 void rpc_execute(struct rpc_task *task)
 720 {
 721         rpc_set_active(task);
 722         rpc_set_running(task);
 723         __rpc_execute(task);
 724 }
 725
 726 static void rpc_async_schedule(struct work_struct *work)
 727 {
 728         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 729 }
 730
 731 struct rpc_buffer {
 732         size_t  len;
 733         char    data[];
 734 };
 735
 736 /**
 737  * rpc_malloc - allocate an RPC buffer
 738  * @task: RPC task that will use this buffer
 739  * @size: requested byte size
 740  *
 741  * To prevent rpciod from hanging, this allocator never sleeps,
 742  * returning NULL if the request cannot be serviced immediately.
 743  * The caller can arrange to sleep in a way that is safe for rpciod.
 744  *
 745  * Most requests are 'small' (under 2KiB) and can be serviced from a
 746  * mempool, ensuring that NFS reads and writes can always proceed,
 747  * and that there is good locality of reference for these buffers.
 748  *
 749  * In order to avoid memory starvation triggering more writebacks of
 750  * NFS requests, we avoid using GFP_KERNEL.
 751  */
 752 void *rpc_malloc(struct rpc_task *task, size_t size)
 753 {
 754         struct rpc_buffer *buf;
 755         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 756
 757         size += sizeof(struct rpc_buffer);
 758         if (size <= RPC_BUFFER_MAXSIZE)
 759                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 760         else
 761                 buf = kmalloc(size, gfp);
 762
 763         if (!buf)
 764                 return NULL;
 765
 766         buf->len = size;
 767         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 768                         task->tk_pid, size, buf);
 769         return &buf->data;
 770 }
 771 EXPORT_SYMBOL_GPL(rpc_malloc);
 772
 773 /**
 774  * rpc_free - free buffer allocated via rpc_malloc
 775  * @buffer: buffer to free
 776  *
 777  */
 778 void rpc_free(void *buffer)
 779 {
 780         size_t size;
 781         struct rpc_buffer *buf;
 782
 783         if (!buffer)
 784                 return;
 785
 786         buf = container_of(buffer, struct rpc_buffer, data);
 787         size = buf->len;
 788
 789         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 790                         size, buf);
 791
 792         if (size <= RPC_BUFFER_MAXSIZE)
 793                 mempool_free(buf, rpc_buffer_mempool);
 794         else
 795                 kfree(buf);
 796 }
 797 EXPORT_SYMBOL_GPL(rpc_free);
 798
 799 /*
 800  * Creation and deletion of RPC task structures
 801  */
 802 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 803 {
 804         memset(task, 0, sizeof(*task));
 805         atomic_set(&task->tk_count, 1);
 806         task->tk_flags  = task_setup_data->flags;
 807         task->tk_ops = task_setup_data->callback_ops;
 808         task->tk_calldata = task_setup_data->callback_data;
 809         INIT_LIST_HEAD(&task->tk_task);
 810
 811         /* Initialize retry counters */
 812         task->tk_garb_retry = 2;
 813         task->tk_cred_retry = 2;
 814
 815         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 816         task->tk_owner = current->tgid;
 817
 818         /* Initialize workqueue for async tasks */
 819         task->tk_workqueue = task_setup_data->workqueue;
 820
 821         task->tk_client = task_setup_data->rpc_client;
 822         if (task->tk_client != NULL) {
 823                 kref_get(&task->tk_client->cl_kref);
 824                 if (task->tk_client->cl_softrtry)
 825                         task->tk_flags |= RPC_TASK_SOFT;
 826         }
 827
 828         if (task->tk_ops->rpc_call_prepare != NULL)
 829                 task->tk_action = rpc_prepare_task;
 830
 831         if (task_setup_data->rpc_message != NULL) {
 832                 memcpy(&task->tk_msg, task_setup_data->rpc_message, sizeof(task->tk_msg));
 833                 /* Bind the user cred */
 834                 if (task->tk_msg.rpc_cred != NULL)
 835                         rpcauth_holdcred(task);
 836                 else
 837                         rpcauth_bindcred(task);
 838                 if (task->tk_action == NULL)
 839                         rpc_call_start(task);
 840         }
 841
 842         /* starting timestamp */
 843         task->tk_start = jiffies;
 844
 845         dprintk("RPC:       new task initialized, procpid %u\n",
 846                                 task_pid_nr(current));
 847 }
 848
 849 static struct rpc_task *
 850 rpc_alloc_task(void)
 851 {
 852         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 853 }
 854
 855 static void rpc_free_task_rcu(struct rcu_head *rcu)
 856 {
 857         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 858         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 859         mempool_free(task, rpc_task_mempool);
 860 }
 861
 862 /*
 863  * Create a new task for the specified client.
 864  */
 865 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 866 {
 867         struct rpc_task *task = setup_data->task;
 868         unsigned short flags = 0;
 869
 870         if (task == NULL) {
 871                 task = rpc_alloc_task();
 872                 if (task == NULL)
 873                         goto out;
 874                 flags = RPC_TASK_DYNAMIC;
 875         }
 876
 877         rpc_init_task(task, setup_data);
 878
 879         task->tk_flags |= flags;
 880         dprintk("RPC:       allocated task %p\n", task);
 881 out:
 882         return task;
 883 }
 884
 885 static void rpc_free_task(struct rpc_task *task)
 886 {
 887         const struct rpc_call_ops *tk_ops = task->tk_ops;
 888         void *calldata = task->tk_calldata;
 889
 890         if (task->tk_flags & RPC_TASK_DYNAMIC)
 891                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task_rcu);
 892         rpc_release_calldata(tk_ops, calldata);
 893 }
 894
 895 static void rpc_async_release(struct work_struct *work)
 896 {
 897         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 898 }
 899
 900 void rpc_put_task(struct rpc_task *task)
 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_workqueue != NULL) {
 914                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 915                 queue_work(task->tk_workqueue, &task->u.tk_work);
 916         } else
 917                 rpc_free_task(task);
 918 }
 919 EXPORT_SYMBOL_GPL(rpc_put_task);
 920
 921 static void rpc_release_task(struct rpc_task *task)
 922 {
 923 #ifdef RPC_DEBUG
 924         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 925 #endif
 926         dprintk("RPC: %5u release task\n", task->tk_pid);
 927
 928         if (!list_empty(&task->tk_task)) {
 929                 struct rpc_clnt *clnt = task->tk_client;
 930                 /* Remove from client task list */
 931                 spin_lock(&clnt->cl_lock);
 932                 list_del(&task->tk_task);
 933                 spin_unlock(&clnt->cl_lock);
 934         }
 935         BUG_ON (RPC_IS_QUEUED(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  * Kill all tasks for the given client.
 948  * XXX: kill their descendants as well?
 949  */
 950 void rpc_killall_tasks(struct rpc_clnt *clnt)
 951 {
 952         struct rpc_task *rovr;
 953
 954
 955         if (list_empty(&clnt->cl_tasks))
 956                 return;
 957         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 958         /*
 959          * Spin lock all_tasks to prevent changes...
 960          */
 961         spin_lock(&clnt->cl_lock);
 962         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 963                 if (! RPC_IS_ACTIVATED(rovr))
 964                         continue;
 965                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 966                         rovr->tk_flags |= RPC_TASK_KILLED;
 967                         rpc_exit(rovr, -EIO);
 968                         rpc_wake_up_task(rovr);
 969                 }
 970         }
 971         spin_unlock(&clnt->cl_lock);
 972 }
 973 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 974
 975 int rpciod_up(void)
 976 {
 977         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 978 }
 979
 980 void rpciod_down(void)
 981 {
 982         module_put(THIS_MODULE);
 983 }
 984
 985 /*
 986  * Start up the rpciod workqueue.
 987  */
 988 static int rpciod_start(void)
 989 {
 990         struct workqueue_struct *wq;
 991
 992         /*
 993          * Create the rpciod thread and wait for it to start.
 994          */
 995         dprintk("RPC:       creating workqueue rpciod\n");
 996         wq = create_workqueue("rpciod");
 997         rpciod_workqueue = wq;
 998         return rpciod_workqueue != NULL;
 999 }
1000
1001 static void rpciod_stop(void)
1002 {
1003         struct workqueue_struct *wq = NULL;
1004
1005         if (rpciod_workqueue == NULL)
1006                 return;
1007         dprintk("RPC:       destroying workqueue rpciod\n");
1008
1009         wq = rpciod_workqueue;
1010         rpciod_workqueue = NULL;
1011         destroy_workqueue(wq);
1012 }
1013
1014 void
1015 rpc_destroy_mempool(void)
1016 {
1017         rpciod_stop();
1018         if (rpc_buffer_mempool)
1019                 mempool_destroy(rpc_buffer_mempool);
1020         if (rpc_task_mempool)
1021                 mempool_destroy(rpc_task_mempool);
1022         if (rpc_task_slabp)
1023                 kmem_cache_destroy(rpc_task_slabp);
1024         if (rpc_buffer_slabp)
1025                 kmem_cache_destroy(rpc_buffer_slabp);
1026         rpc_destroy_wait_queue(&delay_queue);
1027 }
1028
1029 int
1030 rpc_init_mempool(void)
1031 {
1032         /*
1033          * The following is not strictly a mempool initialisation,
1034          * but there is no harm in doing it here
1035          */
1036         rpc_init_wait_queue(&delay_queue, "delayq");
1037         if (!rpciod_start())
1038                 goto err_nomem;
1039
1040         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1041                                              sizeof(struct rpc_task),
1042                                              0, SLAB_HWCACHE_ALIGN,
1043                                              NULL);
1044         if (!rpc_task_slabp)
1045                 goto err_nomem;
1046         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1047                                              RPC_BUFFER_MAXSIZE,
1048                                              0, SLAB_HWCACHE_ALIGN,
1049                                              NULL);
1050         if (!rpc_buffer_slabp)
1051                 goto err_nomem;
1052         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1053                                                     rpc_task_slabp);
1054         if (!rpc_task_mempool)
1055                 goto err_nomem;
1056         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1057                                                       rpc_buffer_slabp);
1058         if (!rpc_buffer_mempool)
1059                 goto err_nomem;
1060         return 0;
1061 err_nomem:
1062         rpc_destroy_mempool();
1063         return -ENOMEM;
1064 }