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