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_default_timer(struct rpc_task *task);
  42 static void                     rpciod_killall(void);
  43 static void                     rpc_async_schedule(struct work_struct *);
  44 static void                      rpc_release_task(struct rpc_task *task);
  45
  46 /*
  47  * RPC tasks sit here while waiting for conditions to improve.
  48  */
  49 static RPC_WAITQ(delay_queue, "delayq");
  50
  51 /*
  52  * All RPC clients are linked into this list
  53  */
  54 static LIST_HEAD(all_clients);
  55 static DECLARE_WAIT_QUEUE_HEAD(client_kill_wait);
  56
  57 /*
  58  * rpciod-related stuff
  59  */
  60 static DEFINE_MUTEX(rpciod_mutex);
  61 static unsigned int             rpciod_users;
  62 struct workqueue_struct *rpciod_workqueue;
  63
  64 /*
  65  * Spinlock for other critical sections of code.
  66  */
  67 static DEFINE_SPINLOCK(rpc_sched_lock);
  68
  69 /*
  70  * Disable the timer for a given RPC task. Should be called with
  71  * queue->lock and bh_disabled in order to avoid races within
  72  * rpc_run_timer().
  73  */
  74 static inline void
  75 __rpc_disable_timer(struct rpc_task *task)
  76 {
  77         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  78         task->tk_timeout_fn = NULL;
  79         task->tk_timeout = 0;
  80 }
  81
  82 /*
  83  * Run a timeout function.
  84  * We use the callback in order to allow __rpc_wake_up_task()
  85  * and friends to disable the timer synchronously on SMP systems
  86  * without calling del_timer_sync(). The latter could cause a
  87  * deadlock if called while we're holding spinlocks...
  88  */
  89 static void rpc_run_timer(struct rpc_task *task)
  90 {
  91         void (*callback)(struct rpc_task *);
  92
  93         callback = task->tk_timeout_fn;
  94         task->tk_timeout_fn = NULL;
  95         if (callback && RPC_IS_QUEUED(task)) {
  96                 dprintk("RPC: %5u running timer\n", task->tk_pid);
  97                 callback(task);
  98         }
  99         smp_mb__before_clear_bit();
 100         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 101         smp_mb__after_clear_bit();
 102 }
 103
 104 /*
 105  * Set up a timer for the current task.
 106  */
 107 static inline void
 108 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
 109 {
 110         if (!task->tk_timeout)
 111                 return;
 112
 113         dprintk("RPC: %5u setting alarm for %lu ms\n",
 114                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 115
 116         if (timer)
 117                 task->tk_timeout_fn = timer;
 118         else
 119                 task->tk_timeout_fn = __rpc_default_timer;
 120         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 121         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 122 }
 123
 124 /*
 125  * Delete any timer for the current task. Because we use del_timer_sync(),
 126  * this function should never be called while holding queue->lock.
 127  */
 128 static void
 129 rpc_delete_timer(struct rpc_task *task)
 130 {
 131         if (RPC_IS_QUEUED(task))
 132                 return;
 133         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 134                 del_singleshot_timer_sync(&task->tk_timer);
 135                 dprintk("RPC: %5u deleting timer\n", task->tk_pid);
 136         }
 137 }
 138
 139 /*
 140  * Add new request to a priority queue.
 141  */
 142 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 143 {
 144         struct list_head *q;
 145         struct rpc_task *t;
 146
 147         INIT_LIST_HEAD(&task->u.tk_wait.links);
 148         q = &queue->tasks[task->tk_priority];
 149         if (unlikely(task->tk_priority > queue->maxpriority))
 150                 q = &queue->tasks[queue->maxpriority];
 151         list_for_each_entry(t, q, u.tk_wait.list) {
 152                 if (t->tk_cookie == task->tk_cookie) {
 153                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 154                         return;
 155                 }
 156         }
 157         list_add_tail(&task->u.tk_wait.list, q);
 158 }
 159
 160 /*
 161  * Add new request to wait queue.
 162  *
 163  * Swapper tasks always get inserted at the head of the queue.
 164  * This should avoid many nasty memory deadlocks and hopefully
 165  * improve overall performance.
 166  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 167  */
 168 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 169 {
 170         BUG_ON (RPC_IS_QUEUED(task));
 171
 172         if (RPC_IS_PRIORITY(queue))
 173                 __rpc_add_wait_queue_priority(queue, task);
 174         else if (RPC_IS_SWAPPER(task))
 175                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 176         else
 177                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 178         task->u.tk_wait.rpc_waitq = queue;
 179         queue->qlen++;
 180         rpc_set_queued(task);
 181
 182         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 183                         task->tk_pid, queue, rpc_qname(queue));
 184 }
 185
 186 /*
 187  * Remove request from a priority queue.
 188  */
 189 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 190 {
 191         struct rpc_task *t;
 192
 193         if (!list_empty(&task->u.tk_wait.links)) {
 194                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 195                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 196                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 197         }
 198         list_del(&task->u.tk_wait.list);
 199 }
 200
 201 /*
 202  * Remove request from queue.
 203  * Note: must be called with spin lock held.
 204  */
 205 static void __rpc_remove_wait_queue(struct rpc_task *task)
 206 {
 207         struct rpc_wait_queue *queue;
 208         queue = task->u.tk_wait.rpc_waitq;
 209
 210         if (RPC_IS_PRIORITY(queue))
 211                 __rpc_remove_wait_queue_priority(task);
 212         else
 213                 list_del(&task->u.tk_wait.list);
 214         queue->qlen--;
 215         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 216                         task->tk_pid, queue, rpc_qname(queue));
 217 }
 218
 219 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 220 {
 221         queue->priority = priority;
 222         queue->count = 1 << (priority * 2);
 223 }
 224
 225 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
 226 {
 227         queue->cookie = cookie;
 228         queue->nr = RPC_BATCH_COUNT;
 229 }
 230
 231 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 232 {
 233         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 234         rpc_set_waitqueue_cookie(queue, 0);
 235 }
 236
 237 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
 238 {
 239         int i;
 240
 241         spin_lock_init(&queue->lock);
 242         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 243                 INIT_LIST_HEAD(&queue->tasks[i]);
 244         queue->maxpriority = maxprio;
 245         rpc_reset_waitqueue_priority(queue);
 246 #ifdef RPC_DEBUG
 247         queue->name = qname;
 248 #endif
 249 }
 250
 251 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 252 {
 253         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
 254 }
 255
 256 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 257 {
 258         __rpc_init_priority_wait_queue(queue, qname, 0);
 259 }
 260 EXPORT_SYMBOL(rpc_init_wait_queue);
 261
 262 static int rpc_wait_bit_interruptible(void *word)
 263 {
 264         if (signal_pending(current))
 265                 return -ERESTARTSYS;
 266         schedule();
 267         return 0;
 268 }
 269
 270 #ifdef RPC_DEBUG
 271 static void rpc_task_set_debuginfo(struct rpc_task *task)
 272 {
 273         static atomic_t rpc_pid;
 274
 275         task->tk_magic = RPC_TASK_MAGIC_ID;
 276         task->tk_pid = atomic_inc_return(&rpc_pid);
 277 }
 278 #else
 279 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 280 {
 281 }
 282 #endif
 283
 284 static void rpc_set_active(struct rpc_task *task)
 285 {
 286         struct rpc_clnt *clnt;
 287         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 288                 return;
 289         rpc_task_set_debuginfo(task);
 290         /* Add to global list of all tasks */
 291         clnt = task->tk_client;
 292         if (clnt != NULL) {
 293                 spin_lock(&clnt->cl_lock);
 294                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 295                 spin_unlock(&clnt->cl_lock);
 296         }
 297 }
 298
 299 /*
 300  * Mark an RPC call as having completed by clearing the 'active' bit
 301  */
 302 static void rpc_mark_complete_task(struct rpc_task *task)
 303 {
 304         smp_mb__before_clear_bit();
 305         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 306         smp_mb__after_clear_bit();
 307         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 308 }
 309
 310 /*
 311  * Allow callers to wait for completion of an RPC call
 312  */
 313 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 314 {
 315         if (action == NULL)
 316                 action = rpc_wait_bit_interruptible;
 317         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 318                         action, TASK_INTERRUPTIBLE);
 319 }
 320 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
 321
 322 /*
 323  * Make an RPC task runnable.
 324  *
 325  * Note: If the task is ASYNC, this must be called with
 326  * the spinlock held to protect the wait queue operation.
 327  */
 328 static void rpc_make_runnable(struct rpc_task *task)
 329 {
 330         BUG_ON(task->tk_timeout_fn);
 331         rpc_clear_queued(task);
 332         if (rpc_test_and_set_running(task))
 333                 return;
 334         /* We might have raced */
 335         if (RPC_IS_QUEUED(task)) {
 336                 rpc_clear_running(task);
 337                 return;
 338         }
 339         if (RPC_IS_ASYNC(task)) {
 340                 int status;
 341
 342                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 343                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 344                 if (status < 0) {
 345                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 346                         task->tk_status = status;
 347                         return;
 348                 }
 349         } else
 350                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 351 }
 352
 353 /*
 354  * Prepare for sleeping on a wait queue.
 355  * By always appending tasks to the list we ensure FIFO behavior.
 356  * NB: An RPC task will only receive interrupt-driven events as long
 357  * as it's on a wait queue.
 358  */
 359 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 360                         rpc_action action, rpc_action timer)
 361 {
 362         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 363                         task->tk_pid, rpc_qname(q), jiffies);
 364
 365         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 366                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 367                 return;
 368         }
 369
 370         __rpc_add_wait_queue(q, task);
 371
 372         BUG_ON(task->tk_callback != NULL);
 373         task->tk_callback = action;
 374         __rpc_add_timer(task, timer);
 375 }
 376
 377 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 378                                 rpc_action action, rpc_action timer)
 379 {
 380         /* Mark the task as being activated if so needed */
 381         rpc_set_active(task);
 382
 383         /*
 384          * Protect the queue operations.
 385          */
 386         spin_lock_bh(&q->lock);
 387         __rpc_sleep_on(q, task, action, timer);
 388         spin_unlock_bh(&q->lock);
 389 }
 390
 391 /**
 392  * __rpc_do_wake_up_task - wake up a single rpc_task
 393  * @task: task to be woken up
 394  *
 395  * Caller must hold queue->lock, and have cleared the task queued flag.
 396  */
 397 static void __rpc_do_wake_up_task(struct rpc_task *task)
 398 {
 399         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 400                         task->tk_pid, jiffies);
 401
 402 #ifdef RPC_DEBUG
 403         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 404 #endif
 405         /* Has the task been executed yet? If not, we cannot wake it up! */
 406         if (!RPC_IS_ACTIVATED(task)) {
 407                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 408                 return;
 409         }
 410
 411         __rpc_disable_timer(task);
 412         __rpc_remove_wait_queue(task);
 413
 414         rpc_make_runnable(task);
 415
 416         dprintk("RPC:       __rpc_wake_up_task done\n");
 417 }
 418
 419 /*
 420  * Wake up the specified task
 421  */
 422 static void __rpc_wake_up_task(struct rpc_task *task)
 423 {
 424         if (rpc_start_wakeup(task)) {
 425                 if (RPC_IS_QUEUED(task))
 426                         __rpc_do_wake_up_task(task);
 427                 rpc_finish_wakeup(task);
 428         }
 429 }
 430
 431 /*
 432  * Default timeout handler if none specified by user
 433  */
 434 static void
 435 __rpc_default_timer(struct rpc_task *task)
 436 {
 437         dprintk("RPC: %5u timeout (default timer)\n", task->tk_pid);
 438         task->tk_status = -ETIMEDOUT;
 439         rpc_wake_up_task(task);
 440 }
 441
 442 /*
 443  * Wake up the specified task
 444  */
 445 void rpc_wake_up_task(struct rpc_task *task)
 446 {
 447         rcu_read_lock_bh();
 448         if (rpc_start_wakeup(task)) {
 449                 if (RPC_IS_QUEUED(task)) {
 450                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 451
 452                         /* Note: we're already in a bh-safe context */
 453                         spin_lock(&queue->lock);
 454                         __rpc_do_wake_up_task(task);
 455                         spin_unlock(&queue->lock);
 456                 }
 457                 rpc_finish_wakeup(task);
 458         }
 459         rcu_read_unlock_bh();
 460 }
 461
 462 /*
 463  * Wake up the next task on a priority queue.
 464  */
 465 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 466 {
 467         struct list_head *q;
 468         struct rpc_task *task;
 469
 470         /*
 471          * Service a batch of tasks from a single cookie.
 472          */
 473         q = &queue->tasks[queue->priority];
 474         if (!list_empty(q)) {
 475                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 476                 if (queue->cookie == task->tk_cookie) {
 477                         if (--queue->nr)
 478                                 goto out;
 479                         list_move_tail(&task->u.tk_wait.list, q);
 480                 }
 481                 /*
 482                  * Check if we need to switch queues.
 483                  */
 484                 if (--queue->count)
 485                         goto new_cookie;
 486         }
 487
 488         /*
 489          * Service the next queue.
 490          */
 491         do {
 492                 if (q == &queue->tasks[0])
 493                         q = &queue->tasks[queue->maxpriority];
 494                 else
 495                         q = q - 1;
 496                 if (!list_empty(q)) {
 497                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 498                         goto new_queue;
 499                 }
 500         } while (q != &queue->tasks[queue->priority]);
 501
 502         rpc_reset_waitqueue_priority(queue);
 503         return NULL;
 504
 505 new_queue:
 506         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 507 new_cookie:
 508         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 509 out:
 510         __rpc_wake_up_task(task);
 511         return task;
 512 }
 513
 514 /*
 515  * Wake up the next task on the wait queue.
 516  */
 517 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 518 {
 519         struct rpc_task *task = NULL;
 520
 521         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 522                         queue, rpc_qname(queue));
 523         rcu_read_lock_bh();
 524         spin_lock(&queue->lock);
 525         if (RPC_IS_PRIORITY(queue))
 526                 task = __rpc_wake_up_next_priority(queue);
 527         else {
 528                 task_for_first(task, &queue->tasks[0])
 529                         __rpc_wake_up_task(task);
 530         }
 531         spin_unlock(&queue->lock);
 532         rcu_read_unlock_bh();
 533
 534         return task;
 535 }
 536
 537 /**
 538  * rpc_wake_up - wake up all rpc_tasks
 539  * @queue: rpc_wait_queue on which the tasks are sleeping
 540  *
 541  * Grabs queue->lock
 542  */
 543 void rpc_wake_up(struct rpc_wait_queue *queue)
 544 {
 545         struct rpc_task *task, *next;
 546         struct list_head *head;
 547
 548         rcu_read_lock_bh();
 549         spin_lock(&queue->lock);
 550         head = &queue->tasks[queue->maxpriority];
 551         for (;;) {
 552                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 553                         __rpc_wake_up_task(task);
 554                 if (head == &queue->tasks[0])
 555                         break;
 556                 head--;
 557         }
 558         spin_unlock(&queue->lock);
 559         rcu_read_unlock_bh();
 560 }
 561
 562 /**
 563  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 564  * @queue: rpc_wait_queue on which the tasks are sleeping
 565  * @status: status value to set
 566  *
 567  * Grabs queue->lock
 568  */
 569 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 570 {
 571         struct rpc_task *task, *next;
 572         struct list_head *head;
 573
 574         rcu_read_lock_bh();
 575         spin_lock(&queue->lock);
 576         head = &queue->tasks[queue->maxpriority];
 577         for (;;) {
 578                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 579                         task->tk_status = status;
 580                         __rpc_wake_up_task(task);
 581                 }
 582                 if (head == &queue->tasks[0])
 583                         break;
 584                 head--;
 585         }
 586         spin_unlock(&queue->lock);
 587         rcu_read_unlock_bh();
 588 }
 589
 590 static void __rpc_atrun(struct rpc_task *task)
 591 {
 592         rpc_wake_up_task(task);
 593 }
 594
 595 /*
 596  * Run a task at a later time
 597  */
 598 void rpc_delay(struct rpc_task *task, unsigned long delay)
 599 {
 600         task->tk_timeout = delay;
 601         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 602 }
 603
 604 /*
 605  * Helper to call task->tk_ops->rpc_call_prepare
 606  */
 607 static void rpc_prepare_task(struct rpc_task *task)
 608 {
 609         lock_kernel();
 610         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 611         unlock_kernel();
 612 }
 613
 614 /*
 615  * Helper that calls task->tk_ops->rpc_call_done if it exists
 616  */
 617 void rpc_exit_task(struct rpc_task *task)
 618 {
 619         task->tk_action = NULL;
 620         if (task->tk_ops->rpc_call_done != NULL) {
 621                 lock_kernel();
 622                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 623                 unlock_kernel();
 624                 if (task->tk_action != NULL) {
 625                         WARN_ON(RPC_ASSASSINATED(task));
 626                         /* Always release the RPC slot and buffer memory */
 627                         xprt_release(task);
 628                 }
 629         }
 630 }
 631 EXPORT_SYMBOL(rpc_exit_task);
 632
 633 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 634 {
 635         if (ops->rpc_release != NULL) {
 636                 lock_kernel();
 637                 ops->rpc_release(calldata);
 638                 unlock_kernel();
 639         }
 640 }
 641
 642 /*
 643  * This is the RPC `scheduler' (or rather, the finite state machine).
 644  */
 645 static void __rpc_execute(struct rpc_task *task)
 646 {
 647         int             status = 0;
 648
 649         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 650                         task->tk_pid, task->tk_flags);
 651
 652         BUG_ON(RPC_IS_QUEUED(task));
 653
 654         for (;;) {
 655                 /*
 656                  * Garbage collection of pending timers...
 657                  */
 658                 rpc_delete_timer(task);
 659
 660                 /*
 661                  * Execute any pending callback.
 662                  */
 663                 if (RPC_DO_CALLBACK(task)) {
 664                         /* Define a callback save pointer */
 665                         void (*save_callback)(struct rpc_task *);
 666
 667                         /*
 668                          * If a callback exists, save it, reset it,
 669                          * call it.
 670                          * The save is needed to stop from resetting
 671                          * another callback set within the callback handler
 672                          * - Dave
 673                          */
 674                         save_callback=task->tk_callback;
 675                         task->tk_callback=NULL;
 676                         save_callback(task);
 677                 }
 678
 679                 /*
 680                  * Perform the next FSM step.
 681                  * tk_action may be NULL when the task has been killed
 682                  * by someone else.
 683                  */
 684                 if (!RPC_IS_QUEUED(task)) {
 685                         if (task->tk_action == NULL)
 686                                 break;
 687                         task->tk_action(task);
 688                 }
 689
 690                 /*
 691                  * Lockless check for whether task is sleeping or not.
 692                  */
 693                 if (!RPC_IS_QUEUED(task))
 694                         continue;
 695                 rpc_clear_running(task);
 696                 if (RPC_IS_ASYNC(task)) {
 697                         /* Careful! we may have raced... */
 698                         if (RPC_IS_QUEUED(task))
 699                                 return;
 700                         if (rpc_test_and_set_running(task))
 701                                 return;
 702                         continue;
 703                 }
 704
 705                 /* sync task: sleep here */
 706                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 707                 /* Note: Caller should be using rpc_clnt_sigmask() */
 708                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 709                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
 710                                 TASK_INTERRUPTIBLE);
 711                 if (status == -ERESTARTSYS) {
 712                         /*
 713                          * When a sync task receives a signal, it exits with
 714                          * -ERESTARTSYS. In order to catch any callbacks that
 715                          * clean up after sleeping on some queue, we don't
 716                          * break the loop here, but go around once more.
 717                          */
 718                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 719                         task->tk_flags |= RPC_TASK_KILLED;
 720                         rpc_exit(task, -ERESTARTSYS);
 721                         rpc_wake_up_task(task);
 722                 }
 723                 rpc_set_running(task);
 724                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 725         }
 726
 727         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 728                         task->tk_status);
 729         /* Release all resources associated with the task */
 730         rpc_release_task(task);
 731 }
 732
 733 /*
 734  * User-visible entry point to the scheduler.
 735  *
 736  * This may be called recursively if e.g. an async NFS task updates
 737  * the attributes and finds that dirty pages must be flushed.
 738  * NOTE: Upon exit of this function the task is guaranteed to be
 739  *       released. In particular note that tk_release() will have
 740  *       been called, so your task memory may have been freed.
 741  */
 742 void rpc_execute(struct rpc_task *task)
 743 {
 744         rpc_set_active(task);
 745         rpc_set_running(task);
 746         __rpc_execute(task);
 747 }
 748
 749 static void rpc_async_schedule(struct work_struct *work)
 750 {
 751         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 752 }
 753
 754 struct rpc_buffer {
 755         size_t  len;
 756         char    data[];
 757 };
 758
 759 /**
 760  * rpc_malloc - allocate an RPC buffer
 761  * @task: RPC task that will use this buffer
 762  * @size: requested byte size
 763  *
 764  * To prevent rpciod from hanging, this allocator never sleeps,
 765  * returning NULL if the request cannot be serviced immediately.
 766  * The caller can arrange to sleep in a way that is safe for rpciod.
 767  *
 768  * Most requests are 'small' (under 2KiB) and can be serviced from a
 769  * mempool, ensuring that NFS reads and writes can always proceed,
 770  * and that there is good locality of reference for these buffers.
 771  *
 772  * In order to avoid memory starvation triggering more writebacks of
 773  * NFS requests, we avoid using GFP_KERNEL.
 774  */
 775 void *rpc_malloc(struct rpc_task *task, size_t size)
 776 {
 777         struct rpc_buffer *buf;
 778         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 779
 780         size += sizeof(struct rpc_buffer);
 781         if (size <= RPC_BUFFER_MAXSIZE)
 782                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 783         else
 784                 buf = kmalloc(size, gfp);
 785
 786         if (!buf)
 787                 return NULL;
 788
 789         buf->len = size;
 790         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 791                         task->tk_pid, size, buf);
 792         return &buf->data;
 793 }
 794
 795 /**
 796  * rpc_free - free buffer allocated via rpc_malloc
 797  * @buffer: buffer to free
 798  *
 799  */
 800 void rpc_free(void *buffer)
 801 {
 802         size_t size;
 803         struct rpc_buffer *buf;
 804
 805         if (!buffer)
 806                 return;
 807
 808         buf = container_of(buffer, struct rpc_buffer, data);
 809         size = buf->len;
 810
 811         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 812                         size, buf);
 813
 814         if (size <= RPC_BUFFER_MAXSIZE)
 815                 mempool_free(buf, rpc_buffer_mempool);
 816         else
 817                 kfree(buf);
 818 }
 819
 820 /*
 821  * Creation and deletion of RPC task structures
 822  */
 823 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 824 {
 825         memset(task, 0, sizeof(*task));
 826         init_timer(&task->tk_timer);
 827         task->tk_timer.data     = (unsigned long) task;
 828         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 829         atomic_set(&task->tk_count, 1);
 830         task->tk_client = clnt;
 831         task->tk_flags  = flags;
 832         task->tk_ops = tk_ops;
 833         if (tk_ops->rpc_call_prepare != NULL)
 834                 task->tk_action = rpc_prepare_task;
 835         task->tk_calldata = calldata;
 836         INIT_LIST_HEAD(&task->tk_task);
 837
 838         /* Initialize retry counters */
 839         task->tk_garb_retry = 2;
 840         task->tk_cred_retry = 2;
 841
 842         task->tk_priority = RPC_PRIORITY_NORMAL;
 843         task->tk_cookie = (unsigned long)current;
 844
 845         /* Initialize workqueue for async tasks */
 846         task->tk_workqueue = rpciod_workqueue;
 847
 848         if (clnt) {
 849                 kref_get(&clnt->cl_kref);
 850                 if (clnt->cl_softrtry)
 851                         task->tk_flags |= RPC_TASK_SOFT;
 852                 if (!clnt->cl_intr)
 853                         task->tk_flags |= RPC_TASK_NOINTR;
 854         }
 855
 856         BUG_ON(task->tk_ops == NULL);
 857
 858         /* starting timestamp */
 859         task->tk_start = jiffies;
 860
 861         dprintk("RPC:       new task initialized, procpid %u\n",
 862                                 current->pid);
 863 }
 864
 865 static struct rpc_task *
 866 rpc_alloc_task(void)
 867 {
 868         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 869 }
 870
 871 static void rpc_free_task(struct rcu_head *rcu)
 872 {
 873         struct rpc_task *task = container_of(rcu, struct rpc_task, u.tk_rcu);
 874         dprintk("RPC: %5u freeing task\n", task->tk_pid);
 875         mempool_free(task, rpc_task_mempool);
 876 }
 877
 878 /*
 879  * Create a new task for the specified client.
 880  */
 881 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 882 {
 883         struct rpc_task *task;
 884
 885         task = rpc_alloc_task();
 886         if (!task)
 887                 goto out;
 888
 889         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 890
 891         dprintk("RPC:       allocated task %p\n", task);
 892         task->tk_flags |= RPC_TASK_DYNAMIC;
 893 out:
 894         return task;
 895 }
 896
 897
 898 void rpc_put_task(struct rpc_task *task)
 899 {
 900         const struct rpc_call_ops *tk_ops = task->tk_ops;
 901         void *calldata = task->tk_calldata;
 902
 903         if (!atomic_dec_and_test(&task->tk_count))
 904                 return;
 905         /* Release resources */
 906         if (task->tk_rqstp)
 907                 xprt_release(task);
 908         if (task->tk_msg.rpc_cred)
 909                 rpcauth_unbindcred(task);
 910         if (task->tk_client) {
 911                 rpc_release_client(task->tk_client);
 912                 task->tk_client = NULL;
 913         }
 914         if (task->tk_flags & RPC_TASK_DYNAMIC)
 915                 call_rcu_bh(&task->u.tk_rcu, rpc_free_task);
 916         rpc_release_calldata(tk_ops, calldata);
 917 }
 918 EXPORT_SYMBOL(rpc_put_task);
 919
 920 static void rpc_release_task(struct rpc_task *task)
 921 {
 922 #ifdef RPC_DEBUG
 923         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 924 #endif
 925         dprintk("RPC: %5u release task\n", task->tk_pid);
 926
 927         if (!list_empty(&task->tk_task)) {
 928                 struct rpc_clnt *clnt = task->tk_client;
 929                 /* Remove from client task list */
 930                 spin_lock(&clnt->cl_lock);
 931                 list_del(&task->tk_task);
 932                 spin_unlock(&clnt->cl_lock);
 933         }
 934         BUG_ON (RPC_IS_QUEUED(task));
 935
 936         /* Synchronously delete any running timer */
 937         rpc_delete_timer(task);
 938
 939 #ifdef RPC_DEBUG
 940         task->tk_magic = 0;
 941 #endif
 942         /* Wake up anyone who is waiting for task completion */
 943         rpc_mark_complete_task(task);
 944
 945         rpc_put_task(task);
 946 }
 947
 948 /**
 949  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
 950  * @clnt: pointer to RPC client
 951  * @flags: RPC flags
 952  * @ops: RPC call ops
 953  * @data: user call data
 954  */
 955 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
 956                                         const struct rpc_call_ops *ops,
 957                                         void *data)
 958 {
 959         struct rpc_task *task;
 960         task = rpc_new_task(clnt, flags, ops, data);
 961         if (task == NULL) {
 962                 rpc_release_calldata(ops, data);
 963                 return ERR_PTR(-ENOMEM);
 964         }
 965         atomic_inc(&task->tk_count);
 966         rpc_execute(task);
 967         return task;
 968 }
 969 EXPORT_SYMBOL(rpc_run_task);
 970
 971 /*
 972  * Kill all tasks for the given client.
 973  * XXX: kill their descendants as well?
 974  */
 975 void rpc_killall_tasks(struct rpc_clnt *clnt)
 976 {
 977         struct rpc_task *rovr;
 978
 979
 980         if (list_empty(&clnt->cl_tasks))
 981                 return;
 982         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 983         /*
 984          * Spin lock all_tasks to prevent changes...
 985          */
 986         spin_lock(&clnt->cl_lock);
 987         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 988                 if (! RPC_IS_ACTIVATED(rovr))
 989                         continue;
 990                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 991                         rovr->tk_flags |= RPC_TASK_KILLED;
 992                         rpc_exit(rovr, -EIO);
 993                         rpc_wake_up_task(rovr);
 994                 }
 995         }
 996         spin_unlock(&clnt->cl_lock);
 997 }
 998
 999 static void rpciod_killall(void)
1000 {
1001         struct rpc_clnt *clnt;
1002         unsigned long flags;
1003
1004         for(;;) {
1005                 clear_thread_flag(TIF_SIGPENDING);
1006
1007                 spin_lock(&rpc_sched_lock);
1008                 list_for_each_entry(clnt, &all_clients, cl_clients)
1009                         rpc_killall_tasks(clnt);
1010                 spin_unlock(&rpc_sched_lock);
1011                 flush_workqueue(rpciod_workqueue);
1012                 if (!list_empty(&all_clients))
1013                         break;
1014                 dprintk("RPC:       rpciod_killall: waiting for tasks "
1015                                         "to exit\n");
1016                 wait_event_timeout(client_kill_wait,
1017                                 list_empty(&all_clients), 1*HZ);
1018         }
1019
1020         spin_lock_irqsave(&current->sighand->siglock, flags);
1021         recalc_sigpending();
1022         spin_unlock_irqrestore(&current->sighand->siglock, flags);
1023 }
1024
1025 void rpc_register_client(struct rpc_clnt *clnt)
1026 {
1027         spin_lock(&rpc_sched_lock);
1028         list_add(&clnt->cl_clients, &all_clients);
1029         spin_unlock(&rpc_sched_lock);
1030 }
1031
1032 void rpc_unregister_client(struct rpc_clnt *clnt)
1033 {
1034         spin_lock(&rpc_sched_lock);
1035         list_del(&clnt->cl_clients);
1036         if (list_empty(&all_clients))
1037                 wake_up(&client_kill_wait);
1038         spin_unlock(&rpc_sched_lock);
1039 }
1040
1041 /*
1042  * Start up the rpciod process if it's not already running.
1043  */
1044 int
1045 rpciod_up(void)
1046 {
1047         struct workqueue_struct *wq;
1048         int error = 0;
1049
1050         mutex_lock(&rpciod_mutex);
1051         dprintk("RPC:       rpciod_up: users %u\n", rpciod_users);
1052         rpciod_users++;
1053         if (rpciod_workqueue)
1054                 goto out;
1055         /*
1056          * If there's no pid, we should be the first user.
1057          */
1058         if (rpciod_users > 1)
1059                 printk(KERN_WARNING "rpciod_up: no workqueue, %u users??\n", rpciod_users);
1060         /*
1061          * Create the rpciod thread and wait for it to start.
1062          */
1063         error = -ENOMEM;
1064         wq = create_workqueue("rpciod");
1065         if (wq == NULL) {
1066                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1067                 rpciod_users--;
1068                 goto out;
1069         }
1070         rpciod_workqueue = wq;
1071         error = 0;
1072 out:
1073         mutex_unlock(&rpciod_mutex);
1074         return error;
1075 }
1076
1077 void
1078 rpciod_down(void)
1079 {
1080         mutex_lock(&rpciod_mutex);
1081         dprintk("RPC:       rpciod_down sema %u\n", rpciod_users);
1082         if (rpciod_users) {
1083                 if (--rpciod_users)
1084                         goto out;
1085         } else
1086                 printk(KERN_WARNING "rpciod_down: no users??\n");
1087
1088         if (!rpciod_workqueue) {
1089                 dprintk("RPC:       rpciod_down: Nothing to do!\n");
1090                 goto out;
1091         }
1092         rpciod_killall();
1093
1094         destroy_workqueue(rpciod_workqueue);
1095         rpciod_workqueue = NULL;
1096  out:
1097         mutex_unlock(&rpciod_mutex);
1098 }
1099
1100 #ifdef RPC_DEBUG
1101 void rpc_show_tasks(void)
1102 {
1103         struct rpc_clnt *clnt;
1104         struct rpc_task *t;
1105
1106         spin_lock(&rpc_sched_lock);
1107         if (list_empty(&all_clients))
1108                 goto out;
1109         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1110                 "-rpcwait -action- ---ops--\n");
1111         list_for_each_entry(clnt, &all_clients, cl_clients) {
1112                 if (list_empty(&clnt->cl_tasks))
1113                         continue;
1114                 spin_lock(&clnt->cl_lock);
1115                 list_for_each_entry(t, &clnt->cl_tasks, tk_task) {
1116                         const char *rpc_waitq = "none";
1117
1118                         if (RPC_IS_QUEUED(t))
1119                                 rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1120
1121                         printk("%5u %04d %04x %6d %8p %6d %8p %8ld %8s %8p %8p\n",
1122                                 t->tk_pid,
1123                                 (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1124                                 t->tk_flags, t->tk_status,
1125                                 t->tk_client,
1126                                 (t->tk_client ? t->tk_client->cl_prog : 0),
1127                                 t->tk_rqstp, t->tk_timeout,
1128                                 rpc_waitq,
1129                                 t->tk_action, t->tk_ops);
1130                 }
1131                 spin_unlock(&clnt->cl_lock);
1132         }
1133 out:
1134         spin_unlock(&rpc_sched_lock);
1135 }
1136 #endif
1137
1138 void
1139 rpc_destroy_mempool(void)
1140 {
1141         if (rpc_buffer_mempool)
1142                 mempool_destroy(rpc_buffer_mempool);
1143         if (rpc_task_mempool)
1144                 mempool_destroy(rpc_task_mempool);
1145         if (rpc_task_slabp)
1146                 kmem_cache_destroy(rpc_task_slabp);
1147         if (rpc_buffer_slabp)
1148                 kmem_cache_destroy(rpc_buffer_slabp);
1149 }
1150
1151 int
1152 rpc_init_mempool(void)
1153 {
1154         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1155                                              sizeof(struct rpc_task),
1156                                              0, SLAB_HWCACHE_ALIGN,
1157                                              NULL, NULL);
1158         if (!rpc_task_slabp)
1159                 goto err_nomem;
1160         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1161                                              RPC_BUFFER_MAXSIZE,
1162                                              0, SLAB_HWCACHE_ALIGN,
1163                                              NULL, NULL);
1164         if (!rpc_buffer_slabp)
1165                 goto err_nomem;
1166         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1167                                                     rpc_task_slabp);
1168         if (!rpc_task_mempool)
1169                 goto err_nomem;
1170         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1171                                                       rpc_buffer_slabp);
1172         if (!rpc_buffer_mempool)
1173                 goto err_nomem;
1174         return 0;
1175 err_nomem:
1176         rpc_destroy_mempool();
1177         return -ENOMEM;
1178 }