net/sunrpc/sched.c

   1 /*
   2  * linux/net/sunrpc/sched.c
   3  *
   4  * Scheduling for synchronous and asynchronous RPC requests.
   5  *
   6  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  10  */
  11
  12 #include <linux/module.h>
  13
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/slab.h>
  17 #include <linux/mempool.h>
  18 #include <linux/smp.h>
  19 #include <linux/smp_lock.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/mutex.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #ifdef RPC_DEBUG
  26 #define RPCDBG_FACILITY         RPCDBG_SCHED
  27 #define RPC_TASK_MAGIC_ID       0xf00baa
  28 static int                      rpc_task_id;
  29 #endif
  30
  31 /*
  32  * RPC slabs and memory pools
  33  */
  34 #define RPC_BUFFER_MAXSIZE      (2048)
  35 #define RPC_BUFFER_POOLSIZE     (8)
  36 #define RPC_TASK_POOLSIZE       (8)
  37 static kmem_cache_t     *rpc_task_slabp __read_mostly;
  38 static kmem_cache_t     *rpc_buffer_slabp __read_mostly;
  39 static mempool_t        *rpc_task_mempool __read_mostly;
  40 static mempool_t        *rpc_buffer_mempool __read_mostly;
  41
  42 static void                     __rpc_default_timer(struct rpc_task *task);
  43 static void                     rpciod_killall(void);
  44 static void                     rpc_async_schedule(void *);
  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 tasks are linked into this list
  53  */
  54 static LIST_HEAD(all_tasks);
  55
  56 /*
  57  * rpciod-related stuff
  58  */
  59 static DEFINE_MUTEX(rpciod_mutex);
  60 static unsigned int             rpciod_users;
  61 struct workqueue_struct *rpciod_workqueue;
  62
  63 /*
  64  * Spinlock for other critical sections of code.
  65  */
  66 static DEFINE_SPINLOCK(rpc_sched_lock);
  67
  68 /*
  69  * Disable the timer for a given RPC task. Should be called with
  70  * queue->lock and bh_disabled in order to avoid races within
  71  * rpc_run_timer().
  72  */
  73 static inline void
  74 __rpc_disable_timer(struct rpc_task *task)
  75 {
  76         dprintk("RPC: %4d disabling timer\n", task->tk_pid);
  77         task->tk_timeout_fn = NULL;
  78         task->tk_timeout = 0;
  79 }
  80
  81 /*
  82  * Run a timeout function.
  83  * We use the callback in order to allow __rpc_wake_up_task()
  84  * and friends to disable the timer synchronously on SMP systems
  85  * without calling del_timer_sync(). The latter could cause a
  86  * deadlock if called while we're holding spinlocks...
  87  */
  88 static void rpc_run_timer(struct rpc_task *task)
  89 {
  90         void (*callback)(struct rpc_task *);
  91
  92         callback = task->tk_timeout_fn;
  93         task->tk_timeout_fn = NULL;
  94         if (callback && RPC_IS_QUEUED(task)) {
  95                 dprintk("RPC: %4d running timer\n", task->tk_pid);
  96                 callback(task);
  97         }
  98         smp_mb__before_clear_bit();
  99         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 100         smp_mb__after_clear_bit();
 101 }
 102
 103 /*
 104  * Set up a timer for the current task.
 105  */
 106 static inline void
 107 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
 108 {
 109         if (!task->tk_timeout)
 110                 return;
 111
 112         dprintk("RPC: %4d setting alarm for %lu ms\n",
 113                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 114
 115         if (timer)
 116                 task->tk_timeout_fn = timer;
 117         else
 118                 task->tk_timeout_fn = __rpc_default_timer;
 119         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 120         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 121 }
 122
 123 /*
 124  * Delete any timer for the current task. Because we use del_timer_sync(),
 125  * this function should never be called while holding queue->lock.
 126  */
 127 static void
 128 rpc_delete_timer(struct rpc_task *task)
 129 {
 130         if (RPC_IS_QUEUED(task))
 131                 return;
 132         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 133                 del_singleshot_timer_sync(&task->tk_timer);
 134                 dprintk("RPC: %4d deleting timer\n", task->tk_pid);
 135         }
 136 }
 137
 138 /*
 139  * Add new request to a priority queue.
 140  */
 141 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 142 {
 143         struct list_head *q;
 144         struct rpc_task *t;
 145
 146         INIT_LIST_HEAD(&task->u.tk_wait.links);
 147         q = &queue->tasks[task->tk_priority];
 148         if (unlikely(task->tk_priority > queue->maxpriority))
 149                 q = &queue->tasks[queue->maxpriority];
 150         list_for_each_entry(t, q, u.tk_wait.list) {
 151                 if (t->tk_cookie == task->tk_cookie) {
 152                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 153                         return;
 154                 }
 155         }
 156         list_add_tail(&task->u.tk_wait.list, q);
 157 }
 158
 159 /*
 160  * Add new request to wait queue.
 161  *
 162  * Swapper tasks always get inserted at the head of the queue.
 163  * This should avoid many nasty memory deadlocks and hopefully
 164  * improve overall performance.
 165  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 166  */
 167 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 168 {
 169         BUG_ON (RPC_IS_QUEUED(task));
 170
 171         if (RPC_IS_PRIORITY(queue))
 172                 __rpc_add_wait_queue_priority(queue, task);
 173         else if (RPC_IS_SWAPPER(task))
 174                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 175         else
 176                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 177         task->u.tk_wait.rpc_waitq = queue;
 178         queue->qlen++;
 179         rpc_set_queued(task);
 180
 181         dprintk("RPC: %4d added to queue %p \"%s\"\n",
 182                                 task->tk_pid, queue, rpc_qname(queue));
 183 }
 184
 185 /*
 186  * Remove request from a priority queue.
 187  */
 188 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 189 {
 190         struct rpc_task *t;
 191
 192         if (!list_empty(&task->u.tk_wait.links)) {
 193                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 194                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 195                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 196         }
 197         list_del(&task->u.tk_wait.list);
 198 }
 199
 200 /*
 201  * Remove request from queue.
 202  * Note: must be called with spin lock held.
 203  */
 204 static void __rpc_remove_wait_queue(struct rpc_task *task)
 205 {
 206         struct rpc_wait_queue *queue;
 207         queue = task->u.tk_wait.rpc_waitq;
 208
 209         if (RPC_IS_PRIORITY(queue))
 210                 __rpc_remove_wait_queue_priority(task);
 211         else
 212                 list_del(&task->u.tk_wait.list);
 213         queue->qlen--;
 214         dprintk("RPC: %4d removed from queue %p \"%s\"\n",
 215                                 task->tk_pid, queue, rpc_qname(queue));
 216 }
 217
 218 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 219 {
 220         queue->priority = priority;
 221         queue->count = 1 << (priority * 2);
 222 }
 223
 224 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
 225 {
 226         queue->cookie = cookie;
 227         queue->nr = RPC_BATCH_COUNT;
 228 }
 229
 230 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 231 {
 232         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 233         rpc_set_waitqueue_cookie(queue, 0);
 234 }
 235
 236 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
 237 {
 238         int i;
 239
 240         spin_lock_init(&queue->lock);
 241         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 242                 INIT_LIST_HEAD(&queue->tasks[i]);
 243         queue->maxpriority = maxprio;
 244         rpc_reset_waitqueue_priority(queue);
 245 #ifdef RPC_DEBUG
 246         queue->name = qname;
 247 #endif
 248 }
 249
 250 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 251 {
 252         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
 253 }
 254
 255 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 256 {
 257         __rpc_init_priority_wait_queue(queue, qname, 0);
 258 }
 259 EXPORT_SYMBOL(rpc_init_wait_queue);
 260
 261 static int rpc_wait_bit_interruptible(void *word)
 262 {
 263         if (signal_pending(current))
 264                 return -ERESTARTSYS;
 265         schedule();
 266         return 0;
 267 }
 268
 269 /*
 270  * Mark an RPC call as having completed by clearing the 'active' bit
 271  */
 272 static inline void rpc_mark_complete_task(struct rpc_task *task)
 273 {
 274         rpc_clear_active(task);
 275         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 276 }
 277
 278 /*
 279  * Allow callers to wait for completion of an RPC call
 280  */
 281 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 282 {
 283         if (action == NULL)
 284                 action = rpc_wait_bit_interruptible;
 285         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 286                         action, TASK_INTERRUPTIBLE);
 287 }
 288 EXPORT_SYMBOL(__rpc_wait_for_completion_task);
 289
 290 /*
 291  * Make an RPC task runnable.
 292  *
 293  * Note: If the task is ASYNC, this must be called with
 294  * the spinlock held to protect the wait queue operation.
 295  */
 296 static void rpc_make_runnable(struct rpc_task *task)
 297 {
 298         BUG_ON(task->tk_timeout_fn);
 299         rpc_clear_queued(task);
 300         if (rpc_test_and_set_running(task))
 301                 return;
 302         /* We might have raced */
 303         if (RPC_IS_QUEUED(task)) {
 304                 rpc_clear_running(task);
 305                 return;
 306         }
 307         if (RPC_IS_ASYNC(task)) {
 308                 int status;
 309
 310                 INIT_WORK(&task->u.tk_work, rpc_async_schedule, (void *)task);
 311                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 312                 if (status < 0) {
 313                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 314                         task->tk_status = status;
 315                         return;
 316                 }
 317         } else
 318                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 319 }
 320
 321 /*
 322  * Prepare for sleeping on a wait queue.
 323  * By always appending tasks to the list we ensure FIFO behavior.
 324  * NB: An RPC task will only receive interrupt-driven events as long
 325  * as it's on a wait queue.
 326  */
 327 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 328                         rpc_action action, rpc_action timer)
 329 {
 330         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
 331                                 rpc_qname(q), jiffies);
 332
 333         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 334                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 335                 return;
 336         }
 337
 338         /* Mark the task as being activated if so needed */
 339         rpc_set_active(task);
 340
 341         __rpc_add_wait_queue(q, task);
 342
 343         BUG_ON(task->tk_callback != NULL);
 344         task->tk_callback = action;
 345         __rpc_add_timer(task, timer);
 346 }
 347
 348 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 349                                 rpc_action action, rpc_action timer)
 350 {
 351         /*
 352          * Protect the queue operations.
 353          */
 354         spin_lock_bh(&q->lock);
 355         __rpc_sleep_on(q, task, action, timer);
 356         spin_unlock_bh(&q->lock);
 357 }
 358
 359 /**
 360  * __rpc_do_wake_up_task - wake up a single rpc_task
 361  * @task: task to be woken up
 362  *
 363  * Caller must hold queue->lock, and have cleared the task queued flag.
 364  */
 365 static void __rpc_do_wake_up_task(struct rpc_task *task)
 366 {
 367         dprintk("RPC: %4d __rpc_wake_up_task (now %ld)\n", task->tk_pid, jiffies);
 368
 369 #ifdef RPC_DEBUG
 370         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 371 #endif
 372         /* Has the task been executed yet? If not, we cannot wake it up! */
 373         if (!RPC_IS_ACTIVATED(task)) {
 374                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 375                 return;
 376         }
 377
 378         __rpc_disable_timer(task);
 379         __rpc_remove_wait_queue(task);
 380
 381         rpc_make_runnable(task);
 382
 383         dprintk("RPC:      __rpc_wake_up_task done\n");
 384 }
 385
 386 /*
 387  * Wake up the specified task
 388  */
 389 static void __rpc_wake_up_task(struct rpc_task *task)
 390 {
 391         if (rpc_start_wakeup(task)) {
 392                 if (RPC_IS_QUEUED(task))
 393                         __rpc_do_wake_up_task(task);
 394                 rpc_finish_wakeup(task);
 395         }
 396 }
 397
 398 /*
 399  * Default timeout handler if none specified by user
 400  */
 401 static void
 402 __rpc_default_timer(struct rpc_task *task)
 403 {
 404         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
 405         task->tk_status = -ETIMEDOUT;
 406         rpc_wake_up_task(task);
 407 }
 408
 409 /*
 410  * Wake up the specified task
 411  */
 412 void rpc_wake_up_task(struct rpc_task *task)
 413 {
 414         if (rpc_start_wakeup(task)) {
 415                 if (RPC_IS_QUEUED(task)) {
 416                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 417
 418                         spin_lock_bh(&queue->lock);
 419                         __rpc_do_wake_up_task(task);
 420                         spin_unlock_bh(&queue->lock);
 421                 }
 422                 rpc_finish_wakeup(task);
 423         }
 424 }
 425
 426 /*
 427  * Wake up the next task on a priority queue.
 428  */
 429 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 430 {
 431         struct list_head *q;
 432         struct rpc_task *task;
 433
 434         /*
 435          * Service a batch of tasks from a single cookie.
 436          */
 437         q = &queue->tasks[queue->priority];
 438         if (!list_empty(q)) {
 439                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 440                 if (queue->cookie == task->tk_cookie) {
 441                         if (--queue->nr)
 442                                 goto out;
 443                         list_move_tail(&task->u.tk_wait.list, q);
 444                 }
 445                 /*
 446                  * Check if we need to switch queues.
 447                  */
 448                 if (--queue->count)
 449                         goto new_cookie;
 450         }
 451
 452         /*
 453          * Service the next queue.
 454          */
 455         do {
 456                 if (q == &queue->tasks[0])
 457                         q = &queue->tasks[queue->maxpriority];
 458                 else
 459                         q = q - 1;
 460                 if (!list_empty(q)) {
 461                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 462                         goto new_queue;
 463                 }
 464         } while (q != &queue->tasks[queue->priority]);
 465
 466         rpc_reset_waitqueue_priority(queue);
 467         return NULL;
 468
 469 new_queue:
 470         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 471 new_cookie:
 472         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 473 out:
 474         __rpc_wake_up_task(task);
 475         return task;
 476 }
 477
 478 /*
 479  * Wake up the next task on the wait queue.
 480  */
 481 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 482 {
 483         struct rpc_task *task = NULL;
 484
 485         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
 486         spin_lock_bh(&queue->lock);
 487         if (RPC_IS_PRIORITY(queue))
 488                 task = __rpc_wake_up_next_priority(queue);
 489         else {
 490                 task_for_first(task, &queue->tasks[0])
 491                         __rpc_wake_up_task(task);
 492         }
 493         spin_unlock_bh(&queue->lock);
 494
 495         return task;
 496 }
 497
 498 /**
 499  * rpc_wake_up - wake up all rpc_tasks
 500  * @queue: rpc_wait_queue on which the tasks are sleeping
 501  *
 502  * Grabs queue->lock
 503  */
 504 void rpc_wake_up(struct rpc_wait_queue *queue)
 505 {
 506         struct rpc_task *task, *next;
 507         struct list_head *head;
 508
 509         spin_lock_bh(&queue->lock);
 510         head = &queue->tasks[queue->maxpriority];
 511         for (;;) {
 512                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 513                         __rpc_wake_up_task(task);
 514                 if (head == &queue->tasks[0])
 515                         break;
 516                 head--;
 517         }
 518         spin_unlock_bh(&queue->lock);
 519 }
 520
 521 /**
 522  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 523  * @queue: rpc_wait_queue on which the tasks are sleeping
 524  * @status: status value to set
 525  *
 526  * Grabs queue->lock
 527  */
 528 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 529 {
 530         struct rpc_task *task, *next;
 531         struct list_head *head;
 532
 533         spin_lock_bh(&queue->lock);
 534         head = &queue->tasks[queue->maxpriority];
 535         for (;;) {
 536                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 537                         task->tk_status = status;
 538                         __rpc_wake_up_task(task);
 539                 }
 540                 if (head == &queue->tasks[0])
 541                         break;
 542                 head--;
 543         }
 544         spin_unlock_bh(&queue->lock);
 545 }
 546
 547 static void __rpc_atrun(struct rpc_task *task)
 548 {
 549         rpc_wake_up_task(task);
 550 }
 551
 552 /*
 553  * Run a task at a later time
 554  */
 555 void rpc_delay(struct rpc_task *task, unsigned long delay)
 556 {
 557         task->tk_timeout = delay;
 558         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 559 }
 560
 561 /*
 562  * Helper to call task->tk_ops->rpc_call_prepare
 563  */
 564 static void rpc_prepare_task(struct rpc_task *task)
 565 {
 566         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 567 }
 568
 569 /*
 570  * Helper that calls task->tk_ops->rpc_call_done if it exists
 571  */
 572 void rpc_exit_task(struct rpc_task *task)
 573 {
 574         task->tk_action = NULL;
 575         if (task->tk_ops->rpc_call_done != NULL) {
 576                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 577                 if (task->tk_action != NULL) {
 578                         WARN_ON(RPC_ASSASSINATED(task));
 579                         /* Always release the RPC slot and buffer memory */
 580                         xprt_release(task);
 581                 }
 582         }
 583 }
 584 EXPORT_SYMBOL(rpc_exit_task);
 585
 586 /*
 587  * This is the RPC `scheduler' (or rather, the finite state machine).
 588  */
 589 static int __rpc_execute(struct rpc_task *task)
 590 {
 591         int             status = 0;
 592
 593         dprintk("RPC: %4d rpc_execute flgs %x\n",
 594                                 task->tk_pid, task->tk_flags);
 595
 596         BUG_ON(RPC_IS_QUEUED(task));
 597
 598         for (;;) {
 599                 /*
 600                  * Garbage collection of pending timers...
 601                  */
 602                 rpc_delete_timer(task);
 603
 604                 /*
 605                  * Execute any pending callback.
 606                  */
 607                 if (RPC_DO_CALLBACK(task)) {
 608                         /* Define a callback save pointer */
 609                         void (*save_callback)(struct rpc_task *);
 610
 611                         /*
 612                          * If a callback exists, save it, reset it,
 613                          * call it.
 614                          * The save is needed to stop from resetting
 615                          * another callback set within the callback handler
 616                          * - Dave
 617                          */
 618                         save_callback=task->tk_callback;
 619                         task->tk_callback=NULL;
 620                         lock_kernel();
 621                         save_callback(task);
 622                         unlock_kernel();
 623                 }
 624
 625                 /*
 626                  * Perform the next FSM step.
 627                  * tk_action may be NULL when the task has been killed
 628                  * by someone else.
 629                  */
 630                 if (!RPC_IS_QUEUED(task)) {
 631                         if (task->tk_action == NULL)
 632                                 break;
 633                         lock_kernel();
 634                         task->tk_action(task);
 635                         unlock_kernel();
 636                 }
 637
 638                 /*
 639                  * Lockless check for whether task is sleeping or not.
 640                  */
 641                 if (!RPC_IS_QUEUED(task))
 642                         continue;
 643                 rpc_clear_running(task);
 644                 if (RPC_IS_ASYNC(task)) {
 645                         /* Careful! we may have raced... */
 646                         if (RPC_IS_QUEUED(task))
 647                                 return 0;
 648                         if (rpc_test_and_set_running(task))
 649                                 return 0;
 650                         continue;
 651                 }
 652
 653                 /* sync task: sleep here */
 654                 dprintk("RPC: %4d sync task going to sleep\n", task->tk_pid);
 655                 /* Note: Caller should be using rpc_clnt_sigmask() */
 656                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 657                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
 658                                 TASK_INTERRUPTIBLE);
 659                 if (status == -ERESTARTSYS) {
 660                         /*
 661                          * When a sync task receives a signal, it exits with
 662                          * -ERESTARTSYS. In order to catch any callbacks that
 663                          * clean up after sleeping on some queue, we don't
 664                          * break the loop here, but go around once more.
 665                          */
 666                         dprintk("RPC: %4d got signal\n", task->tk_pid);
 667                         task->tk_flags |= RPC_TASK_KILLED;
 668                         rpc_exit(task, -ERESTARTSYS);
 669                         rpc_wake_up_task(task);
 670                 }
 671                 rpc_set_running(task);
 672                 dprintk("RPC: %4d sync task resuming\n", task->tk_pid);
 673         }
 674
 675         dprintk("RPC: %4d, return %d, status %d\n", task->tk_pid, status, task->tk_status);
 676         /* Wake up anyone who is waiting for task completion */
 677         rpc_mark_complete_task(task);
 678         /* Release all resources associated with the task */
 679         rpc_release_task(task);
 680         return status;
 681 }
 682
 683 /*
 684  * User-visible entry point to the scheduler.
 685  *
 686  * This may be called recursively if e.g. an async NFS task updates
 687  * the attributes and finds that dirty pages must be flushed.
 688  * NOTE: Upon exit of this function the task is guaranteed to be
 689  *       released. In particular note that tk_release() will have
 690  *       been called, so your task memory may have been freed.
 691  */
 692 int
 693 rpc_execute(struct rpc_task *task)
 694 {
 695         rpc_set_active(task);
 696         rpc_set_running(task);
 697         return __rpc_execute(task);
 698 }
 699
 700 static void rpc_async_schedule(void *arg)
 701 {
 702         __rpc_execute((struct rpc_task *)arg);
 703 }
 704
 705 /**
 706  * rpc_malloc - allocate an RPC buffer
 707  * @task: RPC task that will use this buffer
 708  * @size: requested byte size
 709  *
 710  * We try to ensure that some NFS reads and writes can always proceed
 711  * by using a mempool when allocating 'small' buffers.
 712  * In order to avoid memory starvation triggering more writebacks of
 713  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
 714  */
 715 void * rpc_malloc(struct rpc_task *task, size_t size)
 716 {
 717         struct rpc_rqst *req = task->tk_rqstp;
 718         gfp_t   gfp;
 719
 720         if (task->tk_flags & RPC_TASK_SWAPPER)
 721                 gfp = GFP_ATOMIC;
 722         else
 723                 gfp = GFP_NOFS;
 724
 725         if (size > RPC_BUFFER_MAXSIZE) {
 726                 req->rq_buffer = kmalloc(size, gfp);
 727                 if (req->rq_buffer)
 728                         req->rq_bufsize = size;
 729         } else {
 730                 req->rq_buffer = mempool_alloc(rpc_buffer_mempool, gfp);
 731                 if (req->rq_buffer)
 732                         req->rq_bufsize = RPC_BUFFER_MAXSIZE;
 733         }
 734         return req->rq_buffer;
 735 }
 736
 737 /**
 738  * rpc_free - free buffer allocated via rpc_malloc
 739  * @task: RPC task with a buffer to be freed
 740  *
 741  */
 742 void rpc_free(struct rpc_task *task)
 743 {
 744         struct rpc_rqst *req = task->tk_rqstp;
 745
 746         if (req->rq_buffer) {
 747                 if (req->rq_bufsize == RPC_BUFFER_MAXSIZE)
 748                         mempool_free(req->rq_buffer, rpc_buffer_mempool);
 749                 else
 750                         kfree(req->rq_buffer);
 751                 req->rq_buffer = NULL;
 752                 req->rq_bufsize = 0;
 753         }
 754 }
 755
 756 /*
 757  * Creation and deletion of RPC task structures
 758  */
 759 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 760 {
 761         memset(task, 0, sizeof(*task));
 762         init_timer(&task->tk_timer);
 763         task->tk_timer.data     = (unsigned long) task;
 764         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 765         atomic_set(&task->tk_count, 1);
 766         task->tk_client = clnt;
 767         task->tk_flags  = flags;
 768         task->tk_ops = tk_ops;
 769         if (tk_ops->rpc_call_prepare != NULL)
 770                 task->tk_action = rpc_prepare_task;
 771         task->tk_calldata = calldata;
 772
 773         /* Initialize retry counters */
 774         task->tk_garb_retry = 2;
 775         task->tk_cred_retry = 2;
 776
 777         task->tk_priority = RPC_PRIORITY_NORMAL;
 778         task->tk_cookie = (unsigned long)current;
 779
 780         /* Initialize workqueue for async tasks */
 781         task->tk_workqueue = rpciod_workqueue;
 782
 783         if (clnt) {
 784                 atomic_inc(&clnt->cl_users);
 785                 if (clnt->cl_softrtry)
 786                         task->tk_flags |= RPC_TASK_SOFT;
 787                 if (!clnt->cl_intr)
 788                         task->tk_flags |= RPC_TASK_NOINTR;
 789         }
 790
 791 #ifdef RPC_DEBUG
 792         task->tk_magic = RPC_TASK_MAGIC_ID;
 793         task->tk_pid = rpc_task_id++;
 794 #endif
 795         /* Add to global list of all tasks */
 796         spin_lock(&rpc_sched_lock);
 797         list_add_tail(&task->tk_task, &all_tasks);
 798         spin_unlock(&rpc_sched_lock);
 799
 800         BUG_ON(task->tk_ops == NULL);
 801
 802         /* starting timestamp */
 803         task->tk_start = jiffies;
 804
 805         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
 806                                 current->pid);
 807 }
 808
 809 static struct rpc_task *
 810 rpc_alloc_task(void)
 811 {
 812         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 813 }
 814
 815 static void rpc_free_task(struct rpc_task *task)
 816 {
 817         dprintk("RPC: %4d freeing task\n", task->tk_pid);
 818         mempool_free(task, rpc_task_mempool);
 819 }
 820
 821 /*
 822  * Create a new task for the specified client.  We have to
 823  * clean up after an allocation failure, as the client may
 824  * have specified "oneshot".
 825  */
 826 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 827 {
 828         struct rpc_task *task;
 829
 830         task = rpc_alloc_task();
 831         if (!task)
 832                 goto cleanup;
 833
 834         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 835
 836         dprintk("RPC: %4d allocated task\n", task->tk_pid);
 837         task->tk_flags |= RPC_TASK_DYNAMIC;
 838 out:
 839         return task;
 840
 841 cleanup:
 842         /* Check whether to release the client */
 843         if (clnt) {
 844                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 845                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 846                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 847                 rpc_release_client(clnt);
 848         }
 849         goto out;
 850 }
 851
 852 void rpc_release_task(struct rpc_task *task)
 853 {
 854         const struct rpc_call_ops *tk_ops = task->tk_ops;
 855         void *calldata = task->tk_calldata;
 856
 857 #ifdef RPC_DEBUG
 858         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 859 #endif
 860         if (!atomic_dec_and_test(&task->tk_count))
 861                 return;
 862         dprintk("RPC: %4d release task\n", task->tk_pid);
 863
 864         /* Remove from global task list */
 865         spin_lock(&rpc_sched_lock);
 866         list_del(&task->tk_task);
 867         spin_unlock(&rpc_sched_lock);
 868
 869         BUG_ON (RPC_IS_QUEUED(task));
 870
 871         /* Synchronously delete any running timer */
 872         rpc_delete_timer(task);
 873
 874         /* Release resources */
 875         if (task->tk_rqstp)
 876                 xprt_release(task);
 877         if (task->tk_msg.rpc_cred)
 878                 rpcauth_unbindcred(task);
 879         if (task->tk_client) {
 880                 rpc_release_client(task->tk_client);
 881                 task->tk_client = NULL;
 882         }
 883
 884 #ifdef RPC_DEBUG
 885         task->tk_magic = 0;
 886 #endif
 887         if (task->tk_flags & RPC_TASK_DYNAMIC)
 888                 rpc_free_task(task);
 889         if (tk_ops->rpc_release)
 890                 tk_ops->rpc_release(calldata);
 891 }
 892
 893 /**
 894  * rpc_run_task - Allocate a new RPC task, then run rpc_execute against it
 895  * @clnt: pointer to RPC client
 896  * @flags: RPC flags
 897  * @ops: RPC call ops
 898  * @data: user call data
 899  */
 900 struct rpc_task *rpc_run_task(struct rpc_clnt *clnt, int flags,
 901                                         const struct rpc_call_ops *ops,
 902                                         void *data)
 903 {
 904         struct rpc_task *task;
 905         task = rpc_new_task(clnt, flags, ops, data);
 906         if (task == NULL) {
 907                 if (ops->rpc_release != NULL)
 908                         ops->rpc_release(data);
 909                 return ERR_PTR(-ENOMEM);
 910         }
 911         atomic_inc(&task->tk_count);
 912         rpc_execute(task);
 913         return task;
 914 }
 915 EXPORT_SYMBOL(rpc_run_task);
 916
 917 /*
 918  * Kill all tasks for the given client.
 919  * XXX: kill their descendants as well?
 920  */
 921 void rpc_killall_tasks(struct rpc_clnt *clnt)
 922 {
 923         struct rpc_task *rovr;
 924         struct list_head *le;
 925
 926         dprintk("RPC:      killing all tasks for client %p\n", clnt);
 927
 928         /*
 929          * Spin lock all_tasks to prevent changes...
 930          */
 931         spin_lock(&rpc_sched_lock);
 932         alltask_for_each(rovr, le, &all_tasks) {
 933                 if (! RPC_IS_ACTIVATED(rovr))
 934                         continue;
 935                 if (!clnt || rovr->tk_client == clnt) {
 936                         rovr->tk_flags |= RPC_TASK_KILLED;
 937                         rpc_exit(rovr, -EIO);
 938                         rpc_wake_up_task(rovr);
 939                 }
 940         }
 941         spin_unlock(&rpc_sched_lock);
 942 }
 943
 944 static DECLARE_MUTEX_LOCKED(rpciod_running);
 945
 946 static void rpciod_killall(void)
 947 {
 948         unsigned long flags;
 949
 950         while (!list_empty(&all_tasks)) {
 951                 clear_thread_flag(TIF_SIGPENDING);
 952                 rpc_killall_tasks(NULL);
 953                 flush_workqueue(rpciod_workqueue);
 954                 if (!list_empty(&all_tasks)) {
 955                         dprintk("rpciod_killall: waiting for tasks to exit\n");
 956                         yield();
 957                 }
 958         }
 959
 960         spin_lock_irqsave(&current->sighand->siglock, flags);
 961         recalc_sigpending();
 962         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 963 }
 964
 965 /*
 966  * Start up the rpciod process if it's not already running.
 967  */
 968 int
 969 rpciod_up(void)
 970 {
 971         struct workqueue_struct *wq;
 972         int error = 0;
 973
 974         mutex_lock(&rpciod_mutex);
 975         dprintk("rpciod_up: users %d\n", rpciod_users);
 976         rpciod_users++;
 977         if (rpciod_workqueue)
 978                 goto out;
 979         /*
 980          * If there's no pid, we should be the first user.
 981          */
 982         if (rpciod_users > 1)
 983                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
 984         /*
 985          * Create the rpciod thread and wait for it to start.
 986          */
 987         error = -ENOMEM;
 988         wq = create_workqueue("rpciod");
 989         if (wq == NULL) {
 990                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
 991                 rpciod_users--;
 992                 goto out;
 993         }
 994         rpciod_workqueue = wq;
 995         error = 0;
 996 out:
 997         mutex_unlock(&rpciod_mutex);
 998         return error;
 999 }
1000
1001 void
1002 rpciod_down(void)
1003 {
1004         mutex_lock(&rpciod_mutex);
1005         dprintk("rpciod_down sema %d\n", rpciod_users);
1006         if (rpciod_users) {
1007                 if (--rpciod_users)
1008                         goto out;
1009         } else
1010                 printk(KERN_WARNING "rpciod_down: no users??\n");
1011
1012         if (!rpciod_workqueue) {
1013                 dprintk("rpciod_down: Nothing to do!\n");
1014                 goto out;
1015         }
1016         rpciod_killall();
1017
1018         destroy_workqueue(rpciod_workqueue);
1019         rpciod_workqueue = NULL;
1020  out:
1021         mutex_unlock(&rpciod_mutex);
1022 }
1023
1024 #ifdef RPC_DEBUG
1025 void rpc_show_tasks(void)
1026 {
1027         struct list_head *le;
1028         struct rpc_task *t;
1029
1030         spin_lock(&rpc_sched_lock);
1031         if (list_empty(&all_tasks)) {
1032                 spin_unlock(&rpc_sched_lock);
1033                 return;
1034         }
1035         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1036                 "-rpcwait -action- ---ops--\n");
1037         alltask_for_each(t, le, &all_tasks) {
1038                 const char *rpc_waitq = "none";
1039
1040                 if (RPC_IS_QUEUED(t))
1041                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1042
1043                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1044                         t->tk_pid,
1045                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1046                         t->tk_flags, t->tk_status,
1047                         t->tk_client,
1048                         (t->tk_client ? t->tk_client->cl_prog : 0),
1049                         t->tk_rqstp, t->tk_timeout,
1050                         rpc_waitq,
1051                         t->tk_action, t->tk_ops);
1052         }
1053         spin_unlock(&rpc_sched_lock);
1054 }
1055 #endif
1056
1057 void
1058 rpc_destroy_mempool(void)
1059 {
1060         if (rpc_buffer_mempool)
1061                 mempool_destroy(rpc_buffer_mempool);
1062         if (rpc_task_mempool)
1063                 mempool_destroy(rpc_task_mempool);
1064         if (rpc_task_slabp)
1065                 kmem_cache_destroy(rpc_task_slabp);
1066         if (rpc_buffer_slabp)
1067                 kmem_cache_destroy(rpc_buffer_slabp);
1068 }
1069
1070 int
1071 rpc_init_mempool(void)
1072 {
1073         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1074                                              sizeof(struct rpc_task),
1075                                              0, SLAB_HWCACHE_ALIGN,
1076                                              NULL, NULL);
1077         if (!rpc_task_slabp)
1078                 goto err_nomem;
1079         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1080                                              RPC_BUFFER_MAXSIZE,
1081                                              0, SLAB_HWCACHE_ALIGN,
1082                                              NULL, NULL);
1083         if (!rpc_buffer_slabp)
1084                 goto err_nomem;
1085         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1086                                                     rpc_task_slabp);
1087         if (!rpc_task_mempool)
1088                 goto err_nomem;
1089         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1090                                                       rpc_buffer_slabp);
1091         if (!rpc_buffer_mempool)
1092                 goto err_nomem;
1093         return 0;
1094 err_nomem:
1095         rpc_destroy_mempool();
1096         return -ENOMEM;
1097 }