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 #include <linux/sunrpc/xprt.h>
  25
  26 #ifdef RPC_DEBUG
  27 #define RPCDBG_FACILITY         RPCDBG_SCHED
  28 #define RPC_TASK_MAGIC_ID       0xf00baa
  29 static int                      rpc_task_id;
  30 #endif
  31
  32 /*
  33  * RPC slabs and memory pools
  34  */
  35 #define RPC_BUFFER_MAXSIZE      (2048)
  36 #define RPC_BUFFER_POOLSIZE     (8)
  37 #define RPC_TASK_POOLSIZE       (8)
  38 static kmem_cache_t     *rpc_task_slabp __read_mostly;
  39 static kmem_cache_t     *rpc_buffer_slabp __read_mostly;
  40 static mempool_t        *rpc_task_mempool __read_mostly;
  41 static mempool_t        *rpc_buffer_mempool __read_mostly;
  42
  43 static void                     __rpc_default_timer(struct rpc_task *task);
  44 static void                     rpciod_killall(void);
  45 static void                     rpc_async_schedule(void *);
  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 tasks are linked into this list
  54  */
  55 static LIST_HEAD(all_tasks);
  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: %4d 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: %4d 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: %4d 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: %4d 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: %4d 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: %4d 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 /*
 271  * Mark an RPC call as having completed by clearing the 'active' bit
 272  */
 273 static inline void rpc_mark_complete_task(struct rpc_task *task)
 274 {
 275         rpc_clear_active(task);
 276         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 277 }
 278
 279 /*
 280  * Allow callers to wait for completion of an RPC call
 281  */
 282 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 283 {
 284         if (action == NULL)
 285                 action = rpc_wait_bit_interruptible;
 286         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 287                         action, TASK_INTERRUPTIBLE);
 288 }
 289 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
 290
 291 /*
 292  * Make an RPC task runnable.
 293  *
 294  * Note: If the task is ASYNC, this must be called with
 295  * the spinlock held to protect the wait queue operation.
 296  */
 297 static void rpc_make_runnable(struct rpc_task *task)
 298 {
 299         int do_ret;
 300
 301         BUG_ON(task->tk_timeout_fn);
 302         do_ret = rpc_test_and_set_running(task);
 303         rpc_clear_queued(task);
 304         if (do_ret)
 305                 return;
 306         if (RPC_IS_ASYNC(task)) {
 307                 int status;
 308
 309                 INIT_WORK(&task->u.tk_work, rpc_async_schedule, (void *)task);
 310                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 311                 if (status < 0) {
 312                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 313                         task->tk_status = status;
 314                         return;
 315                 }
 316         } else
 317                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 318 }
 319
 320 /*
 321  * Prepare for sleeping on a wait queue.
 322  * By always appending tasks to the list we ensure FIFO behavior.
 323  * NB: An RPC task will only receive interrupt-driven events as long
 324  * as it's on a wait queue.
 325  */
 326 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 327                         rpc_action action, rpc_action timer)
 328 {
 329         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
 330                                 rpc_qname(q), jiffies);
 331
 332         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 333                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 334                 return;
 335         }
 336
 337         /* Mark the task as being activated if so needed */
 338         rpc_set_active(task);
 339
 340         __rpc_add_wait_queue(q, task);
 341
 342         BUG_ON(task->tk_callback != NULL);
 343         task->tk_callback = action;
 344         __rpc_add_timer(task, timer);
 345 }
 346
 347 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 348                                 rpc_action action, rpc_action timer)
 349 {
 350         /*
 351          * Protect the queue operations.
 352          */
 353         spin_lock_bh(&q->lock);
 354         __rpc_sleep_on(q, task, action, timer);
 355         spin_unlock_bh(&q->lock);
 356 }
 357
 358 /**
 359  * __rpc_do_wake_up_task - wake up a single rpc_task
 360  * @task: task to be woken up
 361  *
 362  * Caller must hold queue->lock, and have cleared the task queued flag.
 363  */
 364 static void __rpc_do_wake_up_task(struct rpc_task *task)
 365 {
 366         dprintk("RPC: %4d __rpc_wake_up_task (now %ld)\n", task->tk_pid, jiffies);
 367
 368 #ifdef RPC_DEBUG
 369         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 370 #endif
 371         /* Has the task been executed yet? If not, we cannot wake it up! */
 372         if (!RPC_IS_ACTIVATED(task)) {
 373                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 374                 return;
 375         }
 376
 377         __rpc_disable_timer(task);
 378         __rpc_remove_wait_queue(task);
 379
 380         rpc_make_runnable(task);
 381
 382         dprintk("RPC:      __rpc_wake_up_task done\n");
 383 }
 384
 385 /*
 386  * Wake up the specified task
 387  */
 388 static void __rpc_wake_up_task(struct rpc_task *task)
 389 {
 390         if (rpc_start_wakeup(task)) {
 391                 if (RPC_IS_QUEUED(task))
 392                         __rpc_do_wake_up_task(task);
 393                 rpc_finish_wakeup(task);
 394         }
 395 }
 396
 397 /*
 398  * Default timeout handler if none specified by user
 399  */
 400 static void
 401 __rpc_default_timer(struct rpc_task *task)
 402 {
 403         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
 404         task->tk_status = -ETIMEDOUT;
 405         rpc_wake_up_task(task);
 406 }
 407
 408 /*
 409  * Wake up the specified task
 410  */
 411 void rpc_wake_up_task(struct rpc_task *task)
 412 {
 413         if (rpc_start_wakeup(task)) {
 414                 if (RPC_IS_QUEUED(task)) {
 415                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 416
 417                         spin_lock_bh(&queue->lock);
 418                         __rpc_do_wake_up_task(task);
 419                         spin_unlock_bh(&queue->lock);
 420                 }
 421                 rpc_finish_wakeup(task);
 422         }
 423 }
 424
 425 /*
 426  * Wake up the next task on a priority queue.
 427  */
 428 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 429 {
 430         struct list_head *q;
 431         struct rpc_task *task;
 432
 433         /*
 434          * Service a batch of tasks from a single cookie.
 435          */
 436         q = &queue->tasks[queue->priority];
 437         if (!list_empty(q)) {
 438                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 439                 if (queue->cookie == task->tk_cookie) {
 440                         if (--queue->nr)
 441                                 goto out;
 442                         list_move_tail(&task->u.tk_wait.list, q);
 443                 }
 444                 /*
 445                  * Check if we need to switch queues.
 446                  */
 447                 if (--queue->count)
 448                         goto new_cookie;
 449         }
 450
 451         /*
 452          * Service the next queue.
 453          */
 454         do {
 455                 if (q == &queue->tasks[0])
 456                         q = &queue->tasks[queue->maxpriority];
 457                 else
 458                         q = q - 1;
 459                 if (!list_empty(q)) {
 460                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 461                         goto new_queue;
 462                 }
 463         } while (q != &queue->tasks[queue->priority]);
 464
 465         rpc_reset_waitqueue_priority(queue);
 466         return NULL;
 467
 468 new_queue:
 469         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 470 new_cookie:
 471         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 472 out:
 473         __rpc_wake_up_task(task);
 474         return task;
 475 }
 476
 477 /*
 478  * Wake up the next task on the wait queue.
 479  */
 480 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 481 {
 482         struct rpc_task *task = NULL;
 483
 484         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
 485         spin_lock_bh(&queue->lock);
 486         if (RPC_IS_PRIORITY(queue))
 487                 task = __rpc_wake_up_next_priority(queue);
 488         else {
 489                 task_for_first(task, &queue->tasks[0])
 490                         __rpc_wake_up_task(task);
 491         }
 492         spin_unlock_bh(&queue->lock);
 493
 494         return task;
 495 }
 496
 497 /**
 498  * rpc_wake_up - wake up all rpc_tasks
 499  * @queue: rpc_wait_queue on which the tasks are sleeping
 500  *
 501  * Grabs queue->lock
 502  */
 503 void rpc_wake_up(struct rpc_wait_queue *queue)
 504 {
 505         struct rpc_task *task, *next;
 506         struct list_head *head;
 507
 508         spin_lock_bh(&queue->lock);
 509         head = &queue->tasks[queue->maxpriority];
 510         for (;;) {
 511                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 512                         __rpc_wake_up_task(task);
 513                 if (head == &queue->tasks[0])
 514                         break;
 515                 head--;
 516         }
 517         spin_unlock_bh(&queue->lock);
 518 }
 519
 520 /**
 521  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 522  * @queue: rpc_wait_queue on which the tasks are sleeping
 523  * @status: status value to set
 524  *
 525  * Grabs queue->lock
 526  */
 527 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 528 {
 529         struct rpc_task *task, *next;
 530         struct list_head *head;
 531
 532         spin_lock_bh(&queue->lock);
 533         head = &queue->tasks[queue->maxpriority];
 534         for (;;) {
 535                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 536                         task->tk_status = status;
 537                         __rpc_wake_up_task(task);
 538                 }
 539                 if (head == &queue->tasks[0])
 540                         break;
 541                 head--;
 542         }
 543         spin_unlock_bh(&queue->lock);
 544 }
 545
 546 /*
 547  * Run a task at a later time
 548  */
 549 static void     __rpc_atrun(struct rpc_task *);
 550 void
 551 rpc_delay(struct rpc_task *task, unsigned long delay)
 552 {
 553         task->tk_timeout = delay;
 554         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 555 }
 556
 557 static void
 558 __rpc_atrun(struct rpc_task *task)
 559 {
 560         task->tk_status = 0;
 561         rpc_wake_up_task(task);
 562 }
 563
 564 /*
 565  * Helper to call task->tk_ops->rpc_call_prepare
 566  */
 567 static void rpc_prepare_task(struct rpc_task *task)
 568 {
 569         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 570 }
 571
 572 /*
 573  * Helper that calls task->tk_ops->rpc_call_done if it exists
 574  */
 575 void rpc_exit_task(struct rpc_task *task)
 576 {
 577         task->tk_action = NULL;
 578         if (task->tk_ops->rpc_call_done != NULL) {
 579                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 580                 if (task->tk_action != NULL) {
 581                         WARN_ON(RPC_ASSASSINATED(task));
 582                         /* Always release the RPC slot and buffer memory */
 583                         xprt_release(task);
 584                 }
 585         }
 586 }
 587 EXPORT_SYMBOL(rpc_exit_task);
 588
 589 /*
 590  * This is the RPC `scheduler' (or rather, the finite state machine).
 591  */
 592 static int __rpc_execute(struct rpc_task *task)
 593 {
 594         int             status = 0;
 595
 596         dprintk("RPC: %4d rpc_execute flgs %x\n",
 597                                 task->tk_pid, task->tk_flags);
 598
 599         BUG_ON(RPC_IS_QUEUED(task));
 600
 601         for (;;) {
 602                 /*
 603                  * Garbage collection of pending timers...
 604                  */
 605                 rpc_delete_timer(task);
 606
 607                 /*
 608                  * Execute any pending callback.
 609                  */
 610                 if (RPC_DO_CALLBACK(task)) {
 611                         /* Define a callback save pointer */
 612                         void (*save_callback)(struct rpc_task *);
 613
 614                         /*
 615                          * If a callback exists, save it, reset it,
 616                          * call it.
 617                          * The save is needed to stop from resetting
 618                          * another callback set within the callback handler
 619                          * - Dave
 620                          */
 621                         save_callback=task->tk_callback;
 622                         task->tk_callback=NULL;
 623                         lock_kernel();
 624                         save_callback(task);
 625                         unlock_kernel();
 626                 }
 627
 628                 /*
 629                  * Perform the next FSM step.
 630                  * tk_action may be NULL when the task has been killed
 631                  * by someone else.
 632                  */
 633                 if (!RPC_IS_QUEUED(task)) {
 634                         if (task->tk_action == NULL)
 635                                 break;
 636                         lock_kernel();
 637                         task->tk_action(task);
 638                         unlock_kernel();
 639                 }
 640
 641                 /*
 642                  * Lockless check for whether task is sleeping or not.
 643                  */
 644                 if (!RPC_IS_QUEUED(task))
 645                         continue;
 646                 rpc_clear_running(task);
 647                 if (RPC_IS_ASYNC(task)) {
 648                         /* Careful! we may have raced... */
 649                         if (RPC_IS_QUEUED(task))
 650                                 return 0;
 651                         if (rpc_test_and_set_running(task))
 652                                 return 0;
 653                         continue;
 654                 }
 655
 656                 /* sync task: sleep here */
 657                 dprintk("RPC: %4d sync task going to sleep\n", task->tk_pid);
 658                 /* Note: Caller should be using rpc_clnt_sigmask() */
 659                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 660                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
 661                                 TASK_INTERRUPTIBLE);
 662                 if (status == -ERESTARTSYS) {
 663                         /*
 664                          * When a sync task receives a signal, it exits with
 665                          * -ERESTARTSYS. In order to catch any callbacks that
 666                          * clean up after sleeping on some queue, we don't
 667                          * break the loop here, but go around once more.
 668                          */
 669                         dprintk("RPC: %4d got signal\n", task->tk_pid);
 670                         task->tk_flags |= RPC_TASK_KILLED;
 671                         rpc_exit(task, -ERESTARTSYS);
 672                         rpc_wake_up_task(task);
 673                 }
 674                 rpc_set_running(task);
 675                 dprintk("RPC: %4d sync task resuming\n", task->tk_pid);
 676         }
 677
 678         dprintk("RPC: %4d, return %d, status %d\n", task->tk_pid, status, task->tk_status);
 679         /* Wake up anyone who is waiting for task completion */
 680         rpc_mark_complete_task(task);
 681         /* Release all resources associated with the task */
 682         rpc_release_task(task);
 683         return status;
 684 }
 685
 686 /*
 687  * User-visible entry point to the scheduler.
 688  *
 689  * This may be called recursively if e.g. an async NFS task updates
 690  * the attributes and finds that dirty pages must be flushed.
 691  * NOTE: Upon exit of this function the task is guaranteed to be
 692  *       released. In particular note that tk_release() will have
 693  *       been called, so your task memory may have been freed.
 694  */
 695 int
 696 rpc_execute(struct rpc_task *task)
 697 {
 698         rpc_set_active(task);
 699         rpc_set_running(task);
 700         return __rpc_execute(task);
 701 }
 702
 703 static void rpc_async_schedule(void *arg)
 704 {
 705         __rpc_execute((struct rpc_task *)arg);
 706 }
 707
 708 /**
 709  * rpc_malloc - allocate an RPC buffer
 710  * @task: RPC task that will use this buffer
 711  * @size: requested byte size
 712  *
 713  * We try to ensure that some NFS reads and writes can always proceed
 714  * by using a mempool when allocating 'small' buffers.
 715  * In order to avoid memory starvation triggering more writebacks of
 716  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
 717  */
 718 void * rpc_malloc(struct rpc_task *task, size_t size)
 719 {
 720         struct rpc_rqst *req = task->tk_rqstp;
 721         gfp_t   gfp;
 722
 723         if (task->tk_flags & RPC_TASK_SWAPPER)
 724                 gfp = GFP_ATOMIC;
 725         else
 726                 gfp = GFP_NOFS;
 727
 728         if (size > RPC_BUFFER_MAXSIZE) {
 729                 req->rq_buffer = kmalloc(size, gfp);
 730                 if (req->rq_buffer)
 731                         req->rq_bufsize = size;
 732         } else {
 733                 req->rq_buffer = mempool_alloc(rpc_buffer_mempool, gfp);
 734                 if (req->rq_buffer)
 735                         req->rq_bufsize = RPC_BUFFER_MAXSIZE;
 736         }
 737         return req->rq_buffer;
 738 }
 739
 740 /**
 741  * rpc_free - free buffer allocated via rpc_malloc
 742  * @task: RPC task with a buffer to be freed
 743  *
 744  */
 745 void rpc_free(struct rpc_task *task)
 746 {
 747         struct rpc_rqst *req = task->tk_rqstp;
 748
 749         if (req->rq_buffer) {
 750                 if (req->rq_bufsize == RPC_BUFFER_MAXSIZE)
 751                         mempool_free(req->rq_buffer, rpc_buffer_mempool);
 752                 else
 753                         kfree(req->rq_buffer);
 754                 req->rq_buffer = NULL;
 755                 req->rq_bufsize = 0;
 756         }
 757 }
 758
 759 /*
 760  * Creation and deletion of RPC task structures
 761  */
 762 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 763 {
 764         memset(task, 0, sizeof(*task));
 765         init_timer(&task->tk_timer);
 766         task->tk_timer.data     = (unsigned long) task;
 767         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 768         atomic_set(&task->tk_count, 1);
 769         task->tk_client = clnt;
 770         task->tk_flags  = flags;
 771         task->tk_ops = tk_ops;
 772         if (tk_ops->rpc_call_prepare != NULL)
 773                 task->tk_action = rpc_prepare_task;
 774         task->tk_calldata = calldata;
 775
 776         /* Initialize retry counters */
 777         task->tk_garb_retry = 2;
 778         task->tk_cred_retry = 2;
 779
 780         task->tk_priority = RPC_PRIORITY_NORMAL;
 781         task->tk_cookie = (unsigned long)current;
 782
 783         /* Initialize workqueue for async tasks */
 784         task->tk_workqueue = rpciod_workqueue;
 785
 786         if (clnt) {
 787                 atomic_inc(&clnt->cl_users);
 788                 if (clnt->cl_softrtry)
 789                         task->tk_flags |= RPC_TASK_SOFT;
 790                 if (!clnt->cl_intr)
 791                         task->tk_flags |= RPC_TASK_NOINTR;
 792         }
 793
 794 #ifdef RPC_DEBUG
 795         task->tk_magic = RPC_TASK_MAGIC_ID;
 796         task->tk_pid = rpc_task_id++;
 797 #endif
 798         /* Add to global list of all tasks */
 799         spin_lock(&rpc_sched_lock);
 800         list_add_tail(&task->tk_task, &all_tasks);
 801         spin_unlock(&rpc_sched_lock);
 802
 803         BUG_ON(task->tk_ops == NULL);
 804
 805         /* starting timestamp */
 806         task->tk_start = jiffies;
 807
 808         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
 809                                 current->pid);
 810 }
 811
 812 static struct rpc_task *
 813 rpc_alloc_task(void)
 814 {
 815         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 816 }
 817
 818 static void rpc_free_task(struct rpc_task *task)
 819 {
 820         dprintk("RPC: %4d freeing task\n", task->tk_pid);
 821         mempool_free(task, rpc_task_mempool);
 822 }
 823
 824 /*
 825  * Create a new task for the specified client.  We have to
 826  * clean up after an allocation failure, as the client may
 827  * have specified "oneshot".
 828  */
 829 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 830 {
 831         struct rpc_task *task;
 832
 833         task = rpc_alloc_task();
 834         if (!task)
 835                 goto cleanup;
 836
 837         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 838
 839         dprintk("RPC: %4d allocated task\n", task->tk_pid);
 840         task->tk_flags |= RPC_TASK_DYNAMIC;
 841 out:
 842         return task;
 843
 844 cleanup:
 845         /* Check whether to release the client */
 846         if (clnt) {
 847                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 848                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 849                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 850                 rpc_release_client(clnt);
 851         }
 852         goto out;
 853 }
 854
 855 void rpc_release_task(struct rpc_task *task)
 856 {
 857         const struct rpc_call_ops *tk_ops = task->tk_ops;
 858         void *calldata = task->tk_calldata;
 859
 860 #ifdef RPC_DEBUG
 861         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 862 #endif
 863         if (!atomic_dec_and_test(&task->tk_count))
 864                 return;
 865         dprintk("RPC: %4d release task\n", task->tk_pid);
 866
 867         /* Remove from global task list */
 868         spin_lock(&rpc_sched_lock);
 869         list_del(&task->tk_task);
 870         spin_unlock(&rpc_sched_lock);
 871
 872         BUG_ON (RPC_IS_QUEUED(task));
 873
 874         /* Synchronously delete any running timer */
 875         rpc_delete_timer(task);
 876
 877         /* Release resources */
 878         if (task->tk_rqstp)
 879                 xprt_release(task);
 880         if (task->tk_msg.rpc_cred)
 881                 rpcauth_unbindcred(task);
 882         if (task->tk_client) {
 883                 rpc_release_client(task->tk_client);
 884                 task->tk_client = NULL;
 885         }
 886
 887 #ifdef RPC_DEBUG
 888         task->tk_magic = 0;
 889 #endif
 890         if (task->tk_flags & RPC_TASK_DYNAMIC)
 891                 rpc_free_task(task);
 892         if (tk_ops->rpc_release)
 893                 tk_ops->rpc_release(calldata);
 894 }
 895
 896 /**
 897  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
 898  * @clnt: pointer to RPC client
 899  * @flags: RPC flags
 900  * @ops: RPC call ops
 901  * @data: user call data
 902  */
 903 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
 904                                         const struct rpc_call_ops *ops,
 905                                         void *data)
 906 {
 907         struct rpc_task *task;
 908         task = rpc_new_task(clnt, flags, ops, data);
 909         if (task == NULL) {
 910                 if (ops->rpc_release != NULL)
 911                         ops->rpc_release(data);
 912                 return ERR_PTR(-ENOMEM);
 913         }
 914         atomic_inc(&task->tk_count);
 915         rpc_execute(task);
 916         return task;
 917 }
 918 EXPORT_SYMBOL(rpc_run_task);
 919
 920 /*
 921  * Kill all tasks for the given client.
 922  * XXX: kill their descendants as well?
 923  */
 924 void rpc_killall_tasks(struct rpc_clnt *clnt)
 925 {
 926         struct rpc_task *rovr;
 927         struct list_head *le;
 928
 929         dprintk("RPC:      killing all tasks for client %p\n", clnt);
 930
 931         /*
 932          * Spin lock all_tasks to prevent changes...
 933          */
 934         spin_lock(&rpc_sched_lock);
 935         alltask_for_each(rovr, le, &all_tasks) {
 936                 if (! RPC_IS_ACTIVATED(rovr))
 937                         continue;
 938                 if (!clnt || rovr->tk_client == clnt) {
 939                         rovr->tk_flags |= RPC_TASK_KILLED;
 940                         rpc_exit(rovr, -EIO);
 941                         rpc_wake_up_task(rovr);
 942                 }
 943         }
 944         spin_unlock(&rpc_sched_lock);
 945 }
 946
 947 static DECLARE_MUTEX_LOCKED(rpciod_running);
 948
 949 static void rpciod_killall(void)
 950 {
 951         unsigned long flags;
 952
 953         while (!list_empty(&all_tasks)) {
 954                 clear_thread_flag(TIF_SIGPENDING);
 955                 rpc_killall_tasks(NULL);
 956                 flush_workqueue(rpciod_workqueue);
 957                 if (!list_empty(&all_tasks)) {
 958                         dprintk("rpciod_killall: waiting for tasks to exit\n");
 959                         yield();
 960                 }
 961         }
 962
 963         spin_lock_irqsave(&current->sighand->siglock, flags);
 964         recalc_sigpending();
 965         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 966 }
 967
 968 /*
 969  * Start up the rpciod process if it's not already running.
 970  */
 971 int
 972 rpciod_up(void)
 973 {
 974         struct workqueue_struct *wq;
 975         int error = 0;
 976
 977         mutex_lock(&rpciod_mutex);
 978         dprintk("rpciod_up: users %d\n", rpciod_users);
 979         rpciod_users++;
 980         if (rpciod_workqueue)
 981                 goto out;
 982         /*
 983          * If there's no pid, we should be the first user.
 984          */
 985         if (rpciod_users > 1)
 986                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
 987         /*
 988          * Create the rpciod thread and wait for it to start.
 989          */
 990         error = -ENOMEM;
 991         wq = create_workqueue("rpciod");
 992         if (wq == NULL) {
 993                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
 994                 rpciod_users--;
 995                 goto out;
 996         }
 997         rpciod_workqueue = wq;
 998         error = 0;
 999 out:
1000         mutex_unlock(&rpciod_mutex);
1001         return error;
1002 }
1003
1004 void
1005 rpciod_down(void)
1006 {
1007         mutex_lock(&rpciod_mutex);
1008         dprintk("rpciod_down sema %d\n", rpciod_users);
1009         if (rpciod_users) {
1010                 if (--rpciod_users)
1011                         goto out;
1012         } else
1013                 printk(KERN_WARNING "rpciod_down: no users??\n");
1014
1015         if (!rpciod_workqueue) {
1016                 dprintk("rpciod_down: Nothing to do!\n");
1017                 goto out;
1018         }
1019         rpciod_killall();
1020
1021         destroy_workqueue(rpciod_workqueue);
1022         rpciod_workqueue = NULL;
1023  out:
1024         mutex_unlock(&rpciod_mutex);
1025 }
1026
1027 #ifdef RPC_DEBUG
1028 void rpc_show_tasks(void)
1029 {
1030         struct list_head *le;
1031         struct rpc_task *t;
1032
1033         spin_lock(&rpc_sched_lock);
1034         if (list_empty(&all_tasks)) {
1035                 spin_unlock(&rpc_sched_lock);
1036                 return;
1037         }
1038         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1039                 "-rpcwait -action- ---ops--\n");
1040         alltask_for_each(t, le, &all_tasks) {
1041                 const char *rpc_waitq = "none";
1042
1043                 if (RPC_IS_QUEUED(t))
1044                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1045
1046                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1047                         t->tk_pid,
1048                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1049                         t->tk_flags, t->tk_status,
1050                         t->tk_client,
1051                         (t->tk_client ? t->tk_client->cl_prog : 0),
1052                         t->tk_rqstp, t->tk_timeout,
1053                         rpc_waitq,
1054                         t->tk_action, t->tk_ops);
1055         }
1056         spin_unlock(&rpc_sched_lock);
1057 }
1058 #endif
1059
1060 void
1061 rpc_destroy_mempool(void)
1062 {
1063         if (rpc_buffer_mempool)
1064                 mempool_destroy(rpc_buffer_mempool);
1065         if (rpc_task_mempool)
1066                 mempool_destroy(rpc_task_mempool);
1067         if (rpc_task_slabp && kmem_cache_destroy(rpc_task_slabp))
1068                 printk(KERN_INFO "rpc_task: not all structures were freed\n");
1069         if (rpc_buffer_slabp && kmem_cache_destroy(rpc_buffer_slabp))
1070                 printk(KERN_INFO "rpc_buffers: not all structures were freed\n");
1071 }
1072
1073 int
1074 rpc_init_mempool(void)
1075 {
1076         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1077                                              sizeof(struct rpc_task),
1078                                              0, SLAB_HWCACHE_ALIGN,
1079                                              NULL, NULL);
1080         if (!rpc_task_slabp)
1081                 goto err_nomem;
1082         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1083                                              RPC_BUFFER_MAXSIZE,
1084                                              0, SLAB_HWCACHE_ALIGN,
1085                                              NULL, NULL);
1086         if (!rpc_buffer_slabp)
1087                 goto err_nomem;
1088         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1089                                                     rpc_task_slabp);
1090         if (!rpc_task_mempool)
1091                 goto err_nomem;
1092         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1093                                                       rpc_buffer_slabp);
1094         if (!rpc_buffer_mempool)
1095                 goto err_nomem;
1096         return 0;
1097 err_nomem:
1098         rpc_destroy_mempool();
1099         return -ENOMEM;
1100 }