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
  22 #include <linux/sunrpc/clnt.h>
  23 #include <linux/sunrpc/xprt.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_free(struct rpc_task *task);
  45
  46 static void                     rpc_async_schedule(void *);
  47
  48 /*
  49  * RPC tasks that create another task (e.g. for contacting the portmapper)
  50  * will wait on this queue for their child's completion
  51  */
  52 static RPC_WAITQ(childq, "childq");
  53
  54 /*
  55  * RPC tasks sit here while waiting for conditions to improve.
  56  */
  57 static RPC_WAITQ(delay_queue, "delayq");
  58
  59 /*
  60  * All RPC tasks are linked into this list
  61  */
  62 static LIST_HEAD(all_tasks);
  63
  64 /*
  65  * rpciod-related stuff
  66  */
  67 static DECLARE_MUTEX(rpciod_sema);
  68 static unsigned int             rpciod_users;
  69 static struct workqueue_struct *rpciod_workqueue;
  70
  71 /*
  72  * Spinlock for other critical sections of code.
  73  */
  74 static DEFINE_SPINLOCK(rpc_sched_lock);
  75
  76 /*
  77  * Disable the timer for a given RPC task. Should be called with
  78  * queue->lock and bh_disabled in order to avoid races within
  79  * rpc_run_timer().
  80  */
  81 static inline void
  82 __rpc_disable_timer(struct rpc_task *task)
  83 {
  84         dprintk("RPC: %4d disabling timer\n", task->tk_pid);
  85         task->tk_timeout_fn = NULL;
  86         task->tk_timeout = 0;
  87 }
  88
  89 /*
  90  * Run a timeout function.
  91  * We use the callback in order to allow __rpc_wake_up_task()
  92  * and friends to disable the timer synchronously on SMP systems
  93  * without calling del_timer_sync(). The latter could cause a
  94  * deadlock if called while we're holding spinlocks...
  95  */
  96 static void rpc_run_timer(struct rpc_task *task)
  97 {
  98         void (*callback)(struct rpc_task *);
  99
 100         callback = task->tk_timeout_fn;
 101         task->tk_timeout_fn = NULL;
 102         if (callback && RPC_IS_QUEUED(task)) {
 103                 dprintk("RPC: %4d running timer\n", task->tk_pid);
 104                 callback(task);
 105         }
 106         smp_mb__before_clear_bit();
 107         clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 108         smp_mb__after_clear_bit();
 109 }
 110
 111 /*
 112  * Set up a timer for the current task.
 113  */
 114 static inline void
 115 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
 116 {
 117         if (!task->tk_timeout)
 118                 return;
 119
 120         dprintk("RPC: %4d setting alarm for %lu ms\n",
 121                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 122
 123         if (timer)
 124                 task->tk_timeout_fn = timer;
 125         else
 126                 task->tk_timeout_fn = __rpc_default_timer;
 127         set_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate);
 128         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 129 }
 130
 131 /*
 132  * Delete any timer for the current task. Because we use del_timer_sync(),
 133  * this function should never be called while holding queue->lock.
 134  */
 135 static void
 136 rpc_delete_timer(struct rpc_task *task)
 137 {
 138         if (RPC_IS_QUEUED(task))
 139                 return;
 140         if (test_and_clear_bit(RPC_TASK_HAS_TIMER, &task->tk_runstate)) {
 141                 del_singleshot_timer_sync(&task->tk_timer);
 142                 dprintk("RPC: %4d deleting timer\n", task->tk_pid);
 143         }
 144 }
 145
 146 /*
 147  * Add new request to a priority queue.
 148  */
 149 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 150 {
 151         struct list_head *q;
 152         struct rpc_task *t;
 153
 154         INIT_LIST_HEAD(&task->u.tk_wait.links);
 155         q = &queue->tasks[task->tk_priority];
 156         if (unlikely(task->tk_priority > queue->maxpriority))
 157                 q = &queue->tasks[queue->maxpriority];
 158         list_for_each_entry(t, q, u.tk_wait.list) {
 159                 if (t->tk_cookie == task->tk_cookie) {
 160                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 161                         return;
 162                 }
 163         }
 164         list_add_tail(&task->u.tk_wait.list, q);
 165 }
 166
 167 /*
 168  * Add new request to wait queue.
 169  *
 170  * Swapper tasks always get inserted at the head of the queue.
 171  * This should avoid many nasty memory deadlocks and hopefully
 172  * improve overall performance.
 173  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 174  */
 175 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 176 {
 177         BUG_ON (RPC_IS_QUEUED(task));
 178
 179         if (RPC_IS_PRIORITY(queue))
 180                 __rpc_add_wait_queue_priority(queue, task);
 181         else if (RPC_IS_SWAPPER(task))
 182                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 183         else
 184                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 185         task->u.tk_wait.rpc_waitq = queue;
 186         rpc_set_queued(task);
 187
 188         dprintk("RPC: %4d added to queue %p \"%s\"\n",
 189                                 task->tk_pid, queue, rpc_qname(queue));
 190 }
 191
 192 /*
 193  * Remove request from a priority queue.
 194  */
 195 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 196 {
 197         struct rpc_task *t;
 198
 199         if (!list_empty(&task->u.tk_wait.links)) {
 200                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 201                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 202                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 203         }
 204         list_del(&task->u.tk_wait.list);
 205 }
 206
 207 /*
 208  * Remove request from queue.
 209  * Note: must be called with spin lock held.
 210  */
 211 static void __rpc_remove_wait_queue(struct rpc_task *task)
 212 {
 213         struct rpc_wait_queue *queue;
 214         queue = task->u.tk_wait.rpc_waitq;
 215
 216         if (RPC_IS_PRIORITY(queue))
 217                 __rpc_remove_wait_queue_priority(task);
 218         else
 219                 list_del(&task->u.tk_wait.list);
 220         dprintk("RPC: %4d removed from queue %p \"%s\"\n",
 221                                 task->tk_pid, queue, rpc_qname(queue));
 222 }
 223
 224 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 225 {
 226         queue->priority = priority;
 227         queue->count = 1 << (priority * 2);
 228 }
 229
 230 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
 231 {
 232         queue->cookie = cookie;
 233         queue->nr = RPC_BATCH_COUNT;
 234 }
 235
 236 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 237 {
 238         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 239         rpc_set_waitqueue_cookie(queue, 0);
 240 }
 241
 242 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
 243 {
 244         int i;
 245
 246         spin_lock_init(&queue->lock);
 247         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 248                 INIT_LIST_HEAD(&queue->tasks[i]);
 249         queue->maxpriority = maxprio;
 250         rpc_reset_waitqueue_priority(queue);
 251 #ifdef RPC_DEBUG
 252         queue->name = qname;
 253 #endif
 254 }
 255
 256 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 257 {
 258         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
 259 }
 260
 261 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 262 {
 263         __rpc_init_priority_wait_queue(queue, qname, 0);
 264 }
 265 EXPORT_SYMBOL(rpc_init_wait_queue);
 266
 267 /*
 268  * Make an RPC task runnable.
 269  *
 270  * Note: If the task is ASYNC, this must be called with
 271  * the spinlock held to protect the wait queue operation.
 272  */
 273 static void rpc_make_runnable(struct rpc_task *task)
 274 {
 275         int do_ret;
 276
 277         BUG_ON(task->tk_timeout_fn);
 278         do_ret = rpc_test_and_set_running(task);
 279         rpc_clear_queued(task);
 280         if (do_ret)
 281                 return;
 282         if (RPC_IS_ASYNC(task)) {
 283                 int status;
 284
 285                 INIT_WORK(&task->u.tk_work, rpc_async_schedule, (void *)task);
 286                 status = queue_work(task->tk_workqueue, &task->u.tk_work);
 287                 if (status < 0) {
 288                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 289                         task->tk_status = status;
 290                         return;
 291                 }
 292         } else
 293                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 294 }
 295
 296 /*
 297  * Place a newly initialized task on the workqueue.
 298  */
 299 static inline void
 300 rpc_schedule_run(struct rpc_task *task)
 301 {
 302         /* Don't run a child twice! */
 303         if (RPC_IS_ACTIVATED(task))
 304                 return;
 305         task->tk_active = 1;
 306         rpc_make_runnable(task);
 307 }
 308
 309 /*
 310  * Prepare for sleeping on a wait queue.
 311  * By always appending tasks to the list we ensure FIFO behavior.
 312  * NB: An RPC task will only receive interrupt-driven events as long
 313  * as it's on a wait queue.
 314  */
 315 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 316                         rpc_action action, rpc_action timer)
 317 {
 318         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
 319                                 rpc_qname(q), jiffies);
 320
 321         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 322                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 323                 return;
 324         }
 325
 326         /* Mark the task as being activated if so needed */
 327         if (!RPC_IS_ACTIVATED(task))
 328                 task->tk_active = 1;
 329
 330         __rpc_add_wait_queue(q, task);
 331
 332         BUG_ON(task->tk_callback != NULL);
 333         task->tk_callback = action;
 334         __rpc_add_timer(task, timer);
 335 }
 336
 337 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 338                                 rpc_action action, rpc_action timer)
 339 {
 340         /*
 341          * Protect the queue operations.
 342          */
 343         spin_lock_bh(&q->lock);
 344         __rpc_sleep_on(q, task, action, timer);
 345         spin_unlock_bh(&q->lock);
 346 }
 347
 348 /**
 349  * __rpc_do_wake_up_task - wake up a single rpc_task
 350  * @task: task to be woken up
 351  *
 352  * Caller must hold queue->lock, and have cleared the task queued flag.
 353  */
 354 static void __rpc_do_wake_up_task(struct rpc_task *task)
 355 {
 356         dprintk("RPC: %4d __rpc_wake_up_task (now %ld)\n", task->tk_pid, jiffies);
 357
 358 #ifdef RPC_DEBUG
 359         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 360 #endif
 361         /* Has the task been executed yet? If not, we cannot wake it up! */
 362         if (!RPC_IS_ACTIVATED(task)) {
 363                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 364                 return;
 365         }
 366
 367         __rpc_disable_timer(task);
 368         __rpc_remove_wait_queue(task);
 369
 370         rpc_make_runnable(task);
 371
 372         dprintk("RPC:      __rpc_wake_up_task done\n");
 373 }
 374
 375 /*
 376  * Wake up the specified task
 377  */
 378 static void __rpc_wake_up_task(struct rpc_task *task)
 379 {
 380         if (rpc_start_wakeup(task)) {
 381                 if (RPC_IS_QUEUED(task))
 382                         __rpc_do_wake_up_task(task);
 383                 rpc_finish_wakeup(task);
 384         }
 385 }
 386
 387 /*
 388  * Default timeout handler if none specified by user
 389  */
 390 static void
 391 __rpc_default_timer(struct rpc_task *task)
 392 {
 393         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
 394         task->tk_status = -ETIMEDOUT;
 395         rpc_wake_up_task(task);
 396 }
 397
 398 /*
 399  * Wake up the specified task
 400  */
 401 void rpc_wake_up_task(struct rpc_task *task)
 402 {
 403         if (rpc_start_wakeup(task)) {
 404                 if (RPC_IS_QUEUED(task)) {
 405                         struct rpc_wait_queue *queue = task->u.tk_wait.rpc_waitq;
 406
 407                         spin_lock_bh(&queue->lock);
 408                         __rpc_do_wake_up_task(task);
 409                         spin_unlock_bh(&queue->lock);
 410                 }
 411                 rpc_finish_wakeup(task);
 412         }
 413 }
 414
 415 /*
 416  * Wake up the next task on a priority queue.
 417  */
 418 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 419 {
 420         struct list_head *q;
 421         struct rpc_task *task;
 422
 423         /*
 424          * Service a batch of tasks from a single cookie.
 425          */
 426         q = &queue->tasks[queue->priority];
 427         if (!list_empty(q)) {
 428                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 429                 if (queue->cookie == task->tk_cookie) {
 430                         if (--queue->nr)
 431                                 goto out;
 432                         list_move_tail(&task->u.tk_wait.list, q);
 433                 }
 434                 /*
 435                  * Check if we need to switch queues.
 436                  */
 437                 if (--queue->count)
 438                         goto new_cookie;
 439         }
 440
 441         /*
 442          * Service the next queue.
 443          */
 444         do {
 445                 if (q == &queue->tasks[0])
 446                         q = &queue->tasks[queue->maxpriority];
 447                 else
 448                         q = q - 1;
 449                 if (!list_empty(q)) {
 450                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 451                         goto new_queue;
 452                 }
 453         } while (q != &queue->tasks[queue->priority]);
 454
 455         rpc_reset_waitqueue_priority(queue);
 456         return NULL;
 457
 458 new_queue:
 459         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 460 new_cookie:
 461         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 462 out:
 463         __rpc_wake_up_task(task);
 464         return task;
 465 }
 466
 467 /*
 468  * Wake up the next task on the wait queue.
 469  */
 470 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 471 {
 472         struct rpc_task *task = NULL;
 473
 474         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
 475         spin_lock_bh(&queue->lock);
 476         if (RPC_IS_PRIORITY(queue))
 477                 task = __rpc_wake_up_next_priority(queue);
 478         else {
 479                 task_for_first(task, &queue->tasks[0])
 480                         __rpc_wake_up_task(task);
 481         }
 482         spin_unlock_bh(&queue->lock);
 483
 484         return task;
 485 }
 486
 487 /**
 488  * rpc_wake_up - wake up all rpc_tasks
 489  * @queue: rpc_wait_queue on which the tasks are sleeping
 490  *
 491  * Grabs queue->lock
 492  */
 493 void rpc_wake_up(struct rpc_wait_queue *queue)
 494 {
 495         struct rpc_task *task;
 496
 497         struct list_head *head;
 498         spin_lock_bh(&queue->lock);
 499         head = &queue->tasks[queue->maxpriority];
 500         for (;;) {
 501                 while (!list_empty(head)) {
 502                         task = list_entry(head->next, struct rpc_task, u.tk_wait.list);
 503                         __rpc_wake_up_task(task);
 504                 }
 505                 if (head == &queue->tasks[0])
 506                         break;
 507                 head--;
 508         }
 509         spin_unlock_bh(&queue->lock);
 510 }
 511
 512 /**
 513  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 514  * @queue: rpc_wait_queue on which the tasks are sleeping
 515  * @status: status value to set
 516  *
 517  * Grabs queue->lock
 518  */
 519 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 520 {
 521         struct list_head *head;
 522         struct rpc_task *task;
 523
 524         spin_lock_bh(&queue->lock);
 525         head = &queue->tasks[queue->maxpriority];
 526         for (;;) {
 527                 while (!list_empty(head)) {
 528                         task = list_entry(head->next, struct rpc_task, u.tk_wait.list);
 529                         task->tk_status = status;
 530                         __rpc_wake_up_task(task);
 531                 }
 532                 if (head == &queue->tasks[0])
 533                         break;
 534                 head--;
 535         }
 536         spin_unlock_bh(&queue->lock);
 537 }
 538
 539 /*
 540  * Run a task at a later time
 541  */
 542 static void     __rpc_atrun(struct rpc_task *);
 543 void
 544 rpc_delay(struct rpc_task *task, unsigned long delay)
 545 {
 546         task->tk_timeout = delay;
 547         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 548 }
 549
 550 static void
 551 __rpc_atrun(struct rpc_task *task)
 552 {
 553         task->tk_status = 0;
 554         rpc_wake_up_task(task);
 555 }
 556
 557 /*
 558  * Helper that calls task->tk_ops->rpc_call_done if it exists
 559  */
 560 void rpc_exit_task(struct rpc_task *task)
 561 {
 562         task->tk_action = NULL;
 563         if (task->tk_ops->rpc_call_done != NULL) {
 564                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 565                 if (task->tk_action != NULL) {
 566                         WARN_ON(RPC_ASSASSINATED(task));
 567                         /* Always release the RPC slot and buffer memory */
 568                         xprt_release(task);
 569                         rpc_free(task);
 570                 }
 571         }
 572 }
 573 EXPORT_SYMBOL(rpc_exit_task);
 574
 575 static int rpc_wait_bit_interruptible(void *word)
 576 {
 577         if (signal_pending(current))
 578                 return -ERESTARTSYS;
 579         schedule();
 580         return 0;
 581 }
 582
 583 /*
 584  * This is the RPC `scheduler' (or rather, the finite state machine).
 585  */
 586 static int __rpc_execute(struct rpc_task *task)
 587 {
 588         int             status = 0;
 589
 590         dprintk("RPC: %4d rpc_execute flgs %x\n",
 591                                 task->tk_pid, task->tk_flags);
 592
 593         BUG_ON(RPC_IS_QUEUED(task));
 594
 595         for (;;) {
 596                 /*
 597                  * Garbage collection of pending timers...
 598                  */
 599                 rpc_delete_timer(task);
 600
 601                 /*
 602                  * Execute any pending callback.
 603                  */
 604                 if (RPC_DO_CALLBACK(task)) {
 605                         /* Define a callback save pointer */
 606                         void (*save_callback)(struct rpc_task *);
 607
 608                         /*
 609                          * If a callback exists, save it, reset it,
 610                          * call it.
 611                          * The save is needed to stop from resetting
 612                          * another callback set within the callback handler
 613                          * - Dave
 614                          */
 615                         save_callback=task->tk_callback;
 616                         task->tk_callback=NULL;
 617                         lock_kernel();
 618                         save_callback(task);
 619                         unlock_kernel();
 620                 }
 621
 622                 /*
 623                  * Perform the next FSM step.
 624                  * tk_action may be NULL when the task has been killed
 625                  * by someone else.
 626                  */
 627                 if (!RPC_IS_QUEUED(task)) {
 628                         if (task->tk_action == NULL)
 629                                 break;
 630                         lock_kernel();
 631                         task->tk_action(task);
 632                         unlock_kernel();
 633                 }
 634
 635                 /*
 636                  * Lockless check for whether task is sleeping or not.
 637                  */
 638                 if (!RPC_IS_QUEUED(task))
 639                         continue;
 640                 rpc_clear_running(task);
 641                 if (RPC_IS_ASYNC(task)) {
 642                         /* Careful! we may have raced... */
 643                         if (RPC_IS_QUEUED(task))
 644                                 return 0;
 645                         if (rpc_test_and_set_running(task))
 646                                 return 0;
 647                         continue;
 648                 }
 649
 650                 /* sync task: sleep here */
 651                 dprintk("RPC: %4d sync task going to sleep\n", task->tk_pid);
 652                 /* Note: Caller should be using rpc_clnt_sigmask() */
 653                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 654                                 RPC_TASK_QUEUED, rpc_wait_bit_interruptible,
 655                                 TASK_INTERRUPTIBLE);
 656                 if (status == -ERESTARTSYS) {
 657                         /*
 658                          * When a sync task receives a signal, it exits with
 659                          * -ERESTARTSYS. In order to catch any callbacks that
 660                          * clean up after sleeping on some queue, we don't
 661                          * break the loop here, but go around once more.
 662                          */
 663                         dprintk("RPC: %4d got signal\n", task->tk_pid);
 664                         task->tk_flags |= RPC_TASK_KILLED;
 665                         rpc_exit(task, -ERESTARTSYS);
 666                         rpc_wake_up_task(task);
 667                 }
 668                 rpc_set_running(task);
 669                 dprintk("RPC: %4d sync task resuming\n", task->tk_pid);
 670         }
 671
 672         dprintk("RPC: %4d exit() = %d\n", task->tk_pid, task->tk_status);
 673         status = task->tk_status;
 674
 675         /* Release all resources associated with the task */
 676         rpc_release_task(task);
 677         return status;
 678 }
 679
 680 /*
 681  * User-visible entry point to the scheduler.
 682  *
 683  * This may be called recursively if e.g. an async NFS task updates
 684  * the attributes and finds that dirty pages must be flushed.
 685  * NOTE: Upon exit of this function the task is guaranteed to be
 686  *       released. In particular note that tk_release() will have
 687  *       been called, so your task memory may have been freed.
 688  */
 689 int
 690 rpc_execute(struct rpc_task *task)
 691 {
 692         BUG_ON(task->tk_active);
 693
 694         task->tk_active = 1;
 695         rpc_set_running(task);
 696         return __rpc_execute(task);
 697 }
 698
 699 static void rpc_async_schedule(void *arg)
 700 {
 701         __rpc_execute((struct rpc_task *)arg);
 702 }
 703
 704 /*
 705  * Allocate memory for RPC purposes.
 706  *
 707  * We try to ensure that some NFS reads and writes can always proceed
 708  * by using a mempool when allocating 'small' buffers.
 709  * In order to avoid memory starvation triggering more writebacks of
 710  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
 711  */
 712 void *
 713 rpc_malloc(struct rpc_task *task, size_t size)
 714 {
 715         gfp_t   gfp;
 716
 717         if (task->tk_flags & RPC_TASK_SWAPPER)
 718                 gfp = GFP_ATOMIC;
 719         else
 720                 gfp = GFP_NOFS;
 721
 722         if (size > RPC_BUFFER_MAXSIZE) {
 723                 task->tk_buffer =  kmalloc(size, gfp);
 724                 if (task->tk_buffer)
 725                         task->tk_bufsize = size;
 726         } else {
 727                 task->tk_buffer =  mempool_alloc(rpc_buffer_mempool, gfp);
 728                 if (task->tk_buffer)
 729                         task->tk_bufsize = RPC_BUFFER_MAXSIZE;
 730         }
 731         return task->tk_buffer;
 732 }
 733
 734 static void
 735 rpc_free(struct rpc_task *task)
 736 {
 737         if (task->tk_buffer) {
 738                 if (task->tk_bufsize == RPC_BUFFER_MAXSIZE)
 739                         mempool_free(task->tk_buffer, rpc_buffer_mempool);
 740                 else
 741                         kfree(task->tk_buffer);
 742                 task->tk_buffer = NULL;
 743                 task->tk_bufsize = 0;
 744         }
 745 }
 746
 747 /*
 748  * Creation and deletion of RPC task structures
 749  */
 750 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 751 {
 752         memset(task, 0, sizeof(*task));
 753         init_timer(&task->tk_timer);
 754         task->tk_timer.data     = (unsigned long) task;
 755         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 756         task->tk_client = clnt;
 757         task->tk_flags  = flags;
 758         task->tk_ops = tk_ops;
 759         task->tk_calldata = calldata;
 760
 761         /* Initialize retry counters */
 762         task->tk_garb_retry = 2;
 763         task->tk_cred_retry = 2;
 764
 765         task->tk_priority = RPC_PRIORITY_NORMAL;
 766         task->tk_cookie = (unsigned long)current;
 767
 768         /* Initialize workqueue for async tasks */
 769         task->tk_workqueue = rpciod_workqueue;
 770
 771         if (clnt) {
 772                 atomic_inc(&clnt->cl_users);
 773                 if (clnt->cl_softrtry)
 774                         task->tk_flags |= RPC_TASK_SOFT;
 775                 if (!clnt->cl_intr)
 776                         task->tk_flags |= RPC_TASK_NOINTR;
 777         }
 778
 779 #ifdef RPC_DEBUG
 780         task->tk_magic = RPC_TASK_MAGIC_ID;
 781         task->tk_pid = rpc_task_id++;
 782 #endif
 783         /* Add to global list of all tasks */
 784         spin_lock(&rpc_sched_lock);
 785         list_add_tail(&task->tk_task, &all_tasks);
 786         spin_unlock(&rpc_sched_lock);
 787
 788         BUG_ON(task->tk_ops == NULL);
 789
 790         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
 791                                 current->pid);
 792 }
 793
 794 static struct rpc_task *
 795 rpc_alloc_task(void)
 796 {
 797         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 798 }
 799
 800 static void rpc_free_task(struct rpc_task *task)
 801 {
 802         dprintk("RPC: %4d freeing task\n", task->tk_pid);
 803         mempool_free(task, rpc_task_mempool);
 804 }
 805
 806 /*
 807  * Create a new task for the specified client.  We have to
 808  * clean up after an allocation failure, as the client may
 809  * have specified "oneshot".
 810  */
 811 struct rpc_task *rpc_new_task(struct rpc_clnt *clnt, int flags, const struct rpc_call_ops *tk_ops, void *calldata)
 812 {
 813         struct rpc_task *task;
 814
 815         task = rpc_alloc_task();
 816         if (!task)
 817                 goto cleanup;
 818
 819         rpc_init_task(task, clnt, flags, tk_ops, calldata);
 820
 821         dprintk("RPC: %4d allocated task\n", task->tk_pid);
 822         task->tk_flags |= RPC_TASK_DYNAMIC;
 823 out:
 824         return task;
 825
 826 cleanup:
 827         /* Check whether to release the client */
 828         if (clnt) {
 829                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 830                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 831                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 832                 rpc_release_client(clnt);
 833         }
 834         goto out;
 835 }
 836
 837 void rpc_release_task(struct rpc_task *task)
 838 {
 839         const struct rpc_call_ops *tk_ops = task->tk_ops;
 840         void *calldata = task->tk_calldata;
 841         dprintk("RPC: %4d release task\n", task->tk_pid);
 842
 843 #ifdef RPC_DEBUG
 844         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 845 #endif
 846
 847         /* Remove from global task list */
 848         spin_lock(&rpc_sched_lock);
 849         list_del(&task->tk_task);
 850         spin_unlock(&rpc_sched_lock);
 851
 852         BUG_ON (RPC_IS_QUEUED(task));
 853         task->tk_active = 0;
 854
 855         /* Synchronously delete any running timer */
 856         rpc_delete_timer(task);
 857
 858         /* Release resources */
 859         if (task->tk_rqstp)
 860                 xprt_release(task);
 861         if (task->tk_msg.rpc_cred)
 862                 rpcauth_unbindcred(task);
 863         rpc_free(task);
 864         if (task->tk_client) {
 865                 rpc_release_client(task->tk_client);
 866                 task->tk_client = NULL;
 867         }
 868
 869 #ifdef RPC_DEBUG
 870         task->tk_magic = 0;
 871 #endif
 872         if (task->tk_flags & RPC_TASK_DYNAMIC)
 873                 rpc_free_task(task);
 874         if (tk_ops->rpc_release)
 875                 tk_ops->rpc_release(calldata);
 876 }
 877
 878 /**
 879  * rpc_find_parent - find the parent of a child task.
 880  * @child: child task
 881  *
 882  * Checks that the parent task is still sleeping on the
 883  * queue 'childq'. If so returns a pointer to the parent.
 884  * Upon failure returns NULL.
 885  *
 886  * Caller must hold childq.lock
 887  */
 888 static inline struct rpc_task *rpc_find_parent(struct rpc_task *child, struct rpc_task *parent)
 889 {
 890         struct rpc_task *task;
 891         struct list_head *le;
 892
 893         task_for_each(task, le, &childq.tasks[0])
 894                 if (task == parent)
 895                         return parent;
 896
 897         return NULL;
 898 }
 899
 900 static void rpc_child_exit(struct rpc_task *child, void *calldata)
 901 {
 902         struct rpc_task *parent;
 903
 904         spin_lock_bh(&childq.lock);
 905         if ((parent = rpc_find_parent(child, calldata)) != NULL) {
 906                 parent->tk_status = child->tk_status;
 907                 __rpc_wake_up_task(parent);
 908         }
 909         spin_unlock_bh(&childq.lock);
 910 }
 911
 912 static const struct rpc_call_ops rpc_child_ops = {
 913         .rpc_call_done = rpc_child_exit,
 914 };
 915
 916 /*
 917  * Note: rpc_new_task releases the client after a failure.
 918  */
 919 struct rpc_task *
 920 rpc_new_child(struct rpc_clnt *clnt, struct rpc_task *parent)
 921 {
 922         struct rpc_task *task;
 923
 924         task = rpc_new_task(clnt, RPC_TASK_ASYNC | RPC_TASK_CHILD, &rpc_child_ops, parent);
 925         if (!task)
 926                 goto fail;
 927         return task;
 928
 929 fail:
 930         parent->tk_status = -ENOMEM;
 931         return NULL;
 932 }
 933
 934 void rpc_run_child(struct rpc_task *task, struct rpc_task *child, rpc_action func)
 935 {
 936         spin_lock_bh(&childq.lock);
 937         /* N.B. Is it possible for the child to have already finished? */
 938         __rpc_sleep_on(&childq, task, func, NULL);
 939         rpc_schedule_run(child);
 940         spin_unlock_bh(&childq.lock);
 941 }
 942
 943 /*
 944  * Kill all tasks for the given client.
 945  * XXX: kill their descendants as well?
 946  */
 947 void rpc_killall_tasks(struct rpc_clnt *clnt)
 948 {
 949         struct rpc_task *rovr;
 950         struct list_head *le;
 951
 952         dprintk("RPC:      killing all tasks for client %p\n", clnt);
 953
 954         /*
 955          * Spin lock all_tasks to prevent changes...
 956          */
 957         spin_lock(&rpc_sched_lock);
 958         alltask_for_each(rovr, le, &all_tasks) {
 959                 if (! RPC_IS_ACTIVATED(rovr))
 960                         continue;
 961                 if (!clnt || rovr->tk_client == clnt) {
 962                         rovr->tk_flags |= RPC_TASK_KILLED;
 963                         rpc_exit(rovr, -EIO);
 964                         rpc_wake_up_task(rovr);
 965                 }
 966         }
 967         spin_unlock(&rpc_sched_lock);
 968 }
 969
 970 static DECLARE_MUTEX_LOCKED(rpciod_running);
 971
 972 static void rpciod_killall(void)
 973 {
 974         unsigned long flags;
 975
 976         while (!list_empty(&all_tasks)) {
 977                 clear_thread_flag(TIF_SIGPENDING);
 978                 rpc_killall_tasks(NULL);
 979                 flush_workqueue(rpciod_workqueue);
 980                 if (!list_empty(&all_tasks)) {
 981                         dprintk("rpciod_killall: waiting for tasks to exit\n");
 982                         yield();
 983                 }
 984         }
 985
 986         spin_lock_irqsave(&current->sighand->siglock, flags);
 987         recalc_sigpending();
 988         spin_unlock_irqrestore(&current->sighand->siglock, flags);
 989 }
 990
 991 /*
 992  * Start up the rpciod process if it's not already running.
 993  */
 994 int
 995 rpciod_up(void)
 996 {
 997         struct workqueue_struct *wq;
 998         int error = 0;
 999
1000         down(&rpciod_sema);
1001         dprintk("rpciod_up: users %d\n", rpciod_users);
1002         rpciod_users++;
1003         if (rpciod_workqueue)
1004                 goto out;
1005         /*
1006          * If there's no pid, we should be the first user.
1007          */
1008         if (rpciod_users > 1)
1009                 printk(KERN_WARNING "rpciod_up: no workqueue, %d users??\n", rpciod_users);
1010         /*
1011          * Create the rpciod thread and wait for it to start.
1012          */
1013         error = -ENOMEM;
1014         wq = create_workqueue("rpciod");
1015         if (wq == NULL) {
1016                 printk(KERN_WARNING "rpciod_up: create workqueue failed, error=%d\n", error);
1017                 rpciod_users--;
1018                 goto out;
1019         }
1020         rpciod_workqueue = wq;
1021         error = 0;
1022 out:
1023         up(&rpciod_sema);
1024         return error;
1025 }
1026
1027 void
1028 rpciod_down(void)
1029 {
1030         down(&rpciod_sema);
1031         dprintk("rpciod_down sema %d\n", rpciod_users);
1032         if (rpciod_users) {
1033                 if (--rpciod_users)
1034                         goto out;
1035         } else
1036                 printk(KERN_WARNING "rpciod_down: no users??\n");
1037
1038         if (!rpciod_workqueue) {
1039                 dprintk("rpciod_down: Nothing to do!\n");
1040                 goto out;
1041         }
1042         rpciod_killall();
1043
1044         destroy_workqueue(rpciod_workqueue);
1045         rpciod_workqueue = NULL;
1046  out:
1047         up(&rpciod_sema);
1048 }
1049
1050 #ifdef RPC_DEBUG
1051 void rpc_show_tasks(void)
1052 {
1053         struct list_head *le;
1054         struct rpc_task *t;
1055
1056         spin_lock(&rpc_sched_lock);
1057         if (list_empty(&all_tasks)) {
1058                 spin_unlock(&rpc_sched_lock);
1059                 return;
1060         }
1061         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1062                 "-rpcwait -action- ---ops--\n");
1063         alltask_for_each(t, le, &all_tasks) {
1064                 const char *rpc_waitq = "none";
1065
1066                 if (RPC_IS_QUEUED(t))
1067                         rpc_waitq = rpc_qname(t->u.tk_wait.rpc_waitq);
1068
1069                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1070                         t->tk_pid,
1071                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1072                         t->tk_flags, t->tk_status,
1073                         t->tk_client,
1074                         (t->tk_client ? t->tk_client->cl_prog : 0),
1075                         t->tk_rqstp, t->tk_timeout,
1076                         rpc_waitq,
1077                         t->tk_action, t->tk_ops);
1078         }
1079         spin_unlock(&rpc_sched_lock);
1080 }
1081 #endif
1082
1083 void
1084 rpc_destroy_mempool(void)
1085 {
1086         if (rpc_buffer_mempool)
1087                 mempool_destroy(rpc_buffer_mempool);
1088         if (rpc_task_mempool)
1089                 mempool_destroy(rpc_task_mempool);
1090         if (rpc_task_slabp && kmem_cache_destroy(rpc_task_slabp))
1091                 printk(KERN_INFO "rpc_task: not all structures were freed\n");
1092         if (rpc_buffer_slabp && kmem_cache_destroy(rpc_buffer_slabp))
1093                 printk(KERN_INFO "rpc_buffers: not all structures were freed\n");
1094 }
1095
1096 int
1097 rpc_init_mempool(void)
1098 {
1099         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1100                                              sizeof(struct rpc_task),
1101                                              0, SLAB_HWCACHE_ALIGN,
1102                                              NULL, NULL);
1103         if (!rpc_task_slabp)
1104                 goto err_nomem;
1105         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1106                                              RPC_BUFFER_MAXSIZE,
1107                                              0, SLAB_HWCACHE_ALIGN,
1108                                              NULL, NULL);
1109         if (!rpc_buffer_slabp)
1110                 goto err_nomem;
1111         rpc_task_mempool = mempool_create(RPC_TASK_POOLSIZE,
1112                                             mempool_alloc_slab,
1113                                             mempool_free_slab,
1114                                             rpc_task_slabp);
1115         if (!rpc_task_mempool)
1116                 goto err_nomem;
1117         rpc_buffer_mempool = mempool_create(RPC_BUFFER_POOLSIZE,
1118                                             mempool_alloc_slab,
1119                                             mempool_free_slab,
1120                                             rpc_buffer_slabp);
1121         if (!rpc_buffer_mempool)
1122                 goto err_nomem;
1123         return 0;
1124 err_nomem:
1125         rpc_destroy_mempool();
1126         return -ENOMEM;
1127 }