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/spinlock.h>
  20 #include <linux/mutex.h>
  21
  22 #include <linux/sunrpc/clnt.h>
  23
  24 #include "sunrpc.h"
  25
  26 #ifdef RPC_DEBUG
  27 #define RPCDBG_FACILITY         RPCDBG_SCHED
  28 #endif
  29
  30 /*
  31  * RPC slabs and memory pools
  32  */
  33 #define RPC_BUFFER_MAXSIZE      (2048)
  34 #define RPC_BUFFER_POOLSIZE     (8)
  35 #define RPC_TASK_POOLSIZE       (8)
  36 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  37 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  38 static mempool_t        *rpc_task_mempool __read_mostly;
  39 static mempool_t        *rpc_buffer_mempool __read_mostly;
  40
  41 static void                     rpc_async_schedule(struct work_struct *);
  42 static void                      rpc_release_task(struct rpc_task *task);
  43 static void __rpc_queue_timer_fn(unsigned long ptr);
  44
  45 /*
  46  * RPC tasks sit here while waiting for conditions to improve.
  47  */
  48 static struct rpc_wait_queue delay_queue;
  49
  50 /*
  51  * rpciod-related stuff
  52  */
  53 struct workqueue_struct *rpciod_workqueue;
  54
  55 /*
  56  * Disable the timer for a given RPC task. Should be called with
  57  * queue->lock and bh_disabled in order to avoid races within
  58  * rpc_run_timer().
  59  */
  60 static void
  61 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  62 {
  63         if (task->tk_timeout == 0)
  64                 return;
  65         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  66         task->tk_timeout = 0;
  67         list_del(&task->u.tk_wait.timer_list);
  68         if (list_empty(&queue->timer_list.list))
  69                 del_timer(&queue->timer_list.timer);
  70 }
  71
  72 static void
  73 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  74 {
  75         queue->timer_list.expires = expires;
  76         mod_timer(&queue->timer_list.timer, expires);
  77 }
  78
  79 /*
  80  * Set up a timer for the current task.
  81  */
  82 static void
  83 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  84 {
  85         if (!task->tk_timeout)
  86                 return;
  87
  88         dprintk("RPC: %5u setting alarm for %lu ms\n",
  89                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  90
  91         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  92         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  93                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  94         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  95 }
  96
  97 /*
  98  * Add new request to a priority queue.
  99  */
 100 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 101 {
 102         struct list_head *q;
 103         struct rpc_task *t;
 104
 105         INIT_LIST_HEAD(&task->u.tk_wait.links);
 106         q = &queue->tasks[task->tk_priority];
 107         if (unlikely(task->tk_priority > queue->maxpriority))
 108                 q = &queue->tasks[queue->maxpriority];
 109         list_for_each_entry(t, q, u.tk_wait.list) {
 110                 if (t->tk_owner == task->tk_owner) {
 111                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 112                         return;
 113                 }
 114         }
 115         list_add_tail(&task->u.tk_wait.list, q);
 116 }
 117
 118 /*
 119  * Add new request to wait queue.
 120  *
 121  * Swapper tasks always get inserted at the head of the queue.
 122  * This should avoid many nasty memory deadlocks and hopefully
 123  * improve overall performance.
 124  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 125  */
 126 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 127 {
 128         BUG_ON (RPC_IS_QUEUED(task));
 129
 130         if (RPC_IS_PRIORITY(queue))
 131                 __rpc_add_wait_queue_priority(queue, task);
 132         else if (RPC_IS_SWAPPER(task))
 133                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 134         else
 135                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 136         task->tk_waitqueue = queue;
 137         queue->qlen++;
 138         rpc_set_queued(task);
 139
 140         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 141                         task->tk_pid, queue, rpc_qname(queue));
 142 }
 143
 144 /*
 145  * Remove request from a priority queue.
 146  */
 147 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 148 {
 149         struct rpc_task *t;
 150
 151         if (!list_empty(&task->u.tk_wait.links)) {
 152                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 153                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 154                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 155         }
 156 }
 157
 158 /*
 159  * Remove request from queue.
 160  * Note: must be called with spin lock held.
 161  */
 162 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 163 {
 164         __rpc_disable_timer(queue, task);
 165         if (RPC_IS_PRIORITY(queue))
 166                 __rpc_remove_wait_queue_priority(task);
 167         list_del(&task->u.tk_wait.list);
 168         queue->qlen--;
 169         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 170                         task->tk_pid, queue, rpc_qname(queue));
 171 }
 172
 173 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 174 {
 175         queue->priority = priority;
 176         queue->count = 1 << (priority * 2);
 177 }
 178
 179 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 180 {
 181         queue->owner = pid;
 182         queue->nr = RPC_BATCH_COUNT;
 183 }
 184
 185 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 186 {
 187         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 188         rpc_set_waitqueue_owner(queue, 0);
 189 }
 190
 191 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 192 {
 193         int i;
 194
 195         spin_lock_init(&queue->lock);
 196         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 197                 INIT_LIST_HEAD(&queue->tasks[i]);
 198         queue->maxpriority = nr_queues - 1;
 199         rpc_reset_waitqueue_priority(queue);
 200         queue->qlen = 0;
 201         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 202         INIT_LIST_HEAD(&queue->timer_list.list);
 203 #ifdef RPC_DEBUG
 204         queue->name = qname;
 205 #endif
 206 }
 207
 208 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 209 {
 210         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 211 }
 212 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 213
 214 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 215 {
 216         __rpc_init_priority_wait_queue(queue, qname, 1);
 217 }
 218 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 219
 220 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 221 {
 222         del_timer_sync(&queue->timer_list.timer);
 223 }
 224 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 225
 226 static int rpc_wait_bit_killable(void *word)
 227 {
 228         if (fatal_signal_pending(current))
 229                 return -ERESTARTSYS;
 230         schedule();
 231         return 0;
 232 }
 233
 234 #ifdef RPC_DEBUG
 235 static void rpc_task_set_debuginfo(struct rpc_task *task)
 236 {
 237         static atomic_t rpc_pid;
 238
 239         task->tk_pid = atomic_inc_return(&rpc_pid);
 240 }
 241 #else
 242 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 243 {
 244 }
 245 #endif
 246
 247 static void rpc_set_active(struct rpc_task *task)
 248 {
 249         struct rpc_clnt *clnt;
 250         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 251                 return;
 252         rpc_task_set_debuginfo(task);
 253         /* Add to global list of all tasks */
 254         clnt = task->tk_client;
 255         if (clnt != NULL) {
 256                 spin_lock(&clnt->cl_lock);
 257                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 258                 spin_unlock(&clnt->cl_lock);
 259         }
 260 }
 261
 262 /*
 263  * Mark an RPC call as having completed by clearing the 'active' bit
 264  */
 265 static void rpc_mark_complete_task(struct rpc_task *task)
 266 {
 267         smp_mb__before_clear_bit();
 268         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 269         smp_mb__after_clear_bit();
 270         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 271 }
 272
 273 /*
 274  * Allow callers to wait for completion of an RPC call
 275  */
 276 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 277 {
 278         if (action == NULL)
 279                 action = rpc_wait_bit_killable;
 280         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 281                         action, TASK_KILLABLE);
 282 }
 283 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 284
 285 /*
 286  * Make an RPC task runnable.
 287  *
 288  * Note: If the task is ASYNC, this must be called with
 289  * the spinlock held to protect the wait queue operation.
 290  */
 291 static void rpc_make_runnable(struct rpc_task *task)
 292 {
 293         rpc_clear_queued(task);
 294         if (rpc_test_and_set_running(task))
 295                 return;
 296         if (RPC_IS_ASYNC(task)) {
 297                 int status;
 298
 299                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 300                 status = queue_work(rpciod_workqueue, &task->u.tk_work);
 301                 if (status < 0) {
 302                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 303                         task->tk_status = status;
 304                         return;
 305                 }
 306         } else
 307                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 308 }
 309
 310 /*
 311  * Prepare for sleeping on a wait queue.
 312  * By always appending tasks to the list we ensure FIFO behavior.
 313  * NB: An RPC task will only receive interrupt-driven events as long
 314  * as it's on a wait queue.
 315  */
 316 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 317                         rpc_action action)
 318 {
 319         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 320                         task->tk_pid, rpc_qname(q), jiffies);
 321
 322         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 323                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 324                 return;
 325         }
 326
 327         __rpc_add_wait_queue(q, task);
 328
 329         BUG_ON(task->tk_callback != NULL);
 330         task->tk_callback = action;
 331         __rpc_add_timer(q, task);
 332 }
 333
 334 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 335                                 rpc_action action)
 336 {
 337         /* Mark the task as being activated if so needed */
 338         rpc_set_active(task);
 339
 340         /*
 341          * Protect the queue operations.
 342          */
 343         spin_lock_bh(&q->lock);
 344         __rpc_sleep_on(q, task, action);
 345         spin_unlock_bh(&q->lock);
 346 }
 347 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 348
 349 /**
 350  * __rpc_do_wake_up_task - wake up a single rpc_task
 351  * @queue: wait queue
 352  * @task: task to be woken up
 353  *
 354  * Caller must hold queue->lock, and have cleared the task queued flag.
 355  */
 356 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 357 {
 358         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 359                         task->tk_pid, jiffies);
 360
 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_remove_wait_queue(queue, task);
 368
 369         rpc_make_runnable(task);
 370
 371         dprintk("RPC:       __rpc_wake_up_task done\n");
 372 }
 373
 374 /*
 375  * Wake up a queued task while the queue lock is being held
 376  */
 377 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 378 {
 379         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 380                 __rpc_do_wake_up_task(queue, task);
 381 }
 382
 383 /*
 384  * Tests whether rpc queue is empty
 385  */
 386 int rpc_queue_empty(struct rpc_wait_queue *queue)
 387 {
 388         int res;
 389
 390         spin_lock_bh(&queue->lock);
 391         res = queue->qlen;
 392         spin_unlock_bh(&queue->lock);
 393         return (res == 0);
 394 }
 395 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 396
 397 /*
 398  * Wake up a task on a specific queue
 399  */
 400 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 401 {
 402         spin_lock_bh(&queue->lock);
 403         rpc_wake_up_task_queue_locked(queue, task);
 404         spin_unlock_bh(&queue->lock);
 405 }
 406 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 407
 408 /*
 409  * Wake up the specified task
 410  */
 411 static void rpc_wake_up_task(struct rpc_task *task)
 412 {
 413         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 414 }
 415
 416 /*
 417  * Wake up the next task on a priority queue.
 418  */
 419 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 420 {
 421         struct list_head *q;
 422         struct rpc_task *task;
 423
 424         /*
 425          * Service a batch of tasks from a single owner.
 426          */
 427         q = &queue->tasks[queue->priority];
 428         if (!list_empty(q)) {
 429                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 430                 if (queue->owner == task->tk_owner) {
 431                         if (--queue->nr)
 432                                 goto out;
 433                         list_move_tail(&task->u.tk_wait.list, q);
 434                 }
 435                 /*
 436                  * Check if we need to switch queues.
 437                  */
 438                 if (--queue->count)
 439                         goto new_owner;
 440         }
 441
 442         /*
 443          * Service the next queue.
 444          */
 445         do {
 446                 if (q == &queue->tasks[0])
 447                         q = &queue->tasks[queue->maxpriority];
 448                 else
 449                         q = q - 1;
 450                 if (!list_empty(q)) {
 451                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 452                         goto new_queue;
 453                 }
 454         } while (q != &queue->tasks[queue->priority]);
 455
 456         rpc_reset_waitqueue_priority(queue);
 457         return NULL;
 458
 459 new_queue:
 460         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 461 new_owner:
 462         rpc_set_waitqueue_owner(queue, task->tk_owner);
 463 out:
 464         rpc_wake_up_task_queue_locked(queue, task);
 465         return task;
 466 }
 467
 468 /*
 469  * Wake up the next task on the wait queue.
 470  */
 471 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 472 {
 473         struct rpc_task *task = NULL;
 474
 475         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 476                         queue, rpc_qname(queue));
 477         spin_lock_bh(&queue->lock);
 478         if (RPC_IS_PRIORITY(queue))
 479                 task = __rpc_wake_up_next_priority(queue);
 480         else {
 481                 task_for_first(task, &queue->tasks[0])
 482                         rpc_wake_up_task_queue_locked(queue, task);
 483         }
 484         spin_unlock_bh(&queue->lock);
 485
 486         return task;
 487 }
 488 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 489
 490 /**
 491  * rpc_wake_up - wake up all rpc_tasks
 492  * @queue: rpc_wait_queue on which the tasks are sleeping
 493  *
 494  * Grabs queue->lock
 495  */
 496 void rpc_wake_up(struct rpc_wait_queue *queue)
 497 {
 498         struct rpc_task *task, *next;
 499         struct list_head *head;
 500
 501         spin_lock_bh(&queue->lock);
 502         head = &queue->tasks[queue->maxpriority];
 503         for (;;) {
 504                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 505                         rpc_wake_up_task_queue_locked(queue, task);
 506                 if (head == &queue->tasks[0])
 507                         break;
 508                 head--;
 509         }
 510         spin_unlock_bh(&queue->lock);
 511 }
 512 EXPORT_SYMBOL_GPL(rpc_wake_up);
 513
 514 /**
 515  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 516  * @queue: rpc_wait_queue on which the tasks are sleeping
 517  * @status: status value to set
 518  *
 519  * Grabs queue->lock
 520  */
 521 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 522 {
 523         struct rpc_task *task, *next;
 524         struct list_head *head;
 525
 526         spin_lock_bh(&queue->lock);
 527         head = &queue->tasks[queue->maxpriority];
 528         for (;;) {
 529                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 530                         task->tk_status = status;
 531                         rpc_wake_up_task_queue_locked(queue, task);
 532                 }
 533                 if (head == &queue->tasks[0])
 534                         break;
 535                 head--;
 536         }
 537         spin_unlock_bh(&queue->lock);
 538 }
 539 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 540
 541 static void __rpc_queue_timer_fn(unsigned long ptr)
 542 {
 543         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 544         struct rpc_task *task, *n;
 545         unsigned long expires, now, timeo;
 546
 547         spin_lock(&queue->lock);
 548         expires = now = jiffies;
 549         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 550                 timeo = task->u.tk_wait.expires;
 551                 if (time_after_eq(now, timeo)) {
 552                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 553                         task->tk_status = -ETIMEDOUT;
 554                         rpc_wake_up_task_queue_locked(queue, task);
 555                         continue;
 556                 }
 557                 if (expires == now || time_after(expires, timeo))
 558                         expires = timeo;
 559         }
 560         if (!list_empty(&queue->timer_list.list))
 561                 rpc_set_queue_timer(queue, expires);
 562         spin_unlock(&queue->lock);
 563 }
 564
 565 static void __rpc_atrun(struct rpc_task *task)
 566 {
 567         task->tk_status = 0;
 568 }
 569
 570 /*
 571  * Run a task at a later time
 572  */
 573 void rpc_delay(struct rpc_task *task, unsigned long delay)
 574 {
 575         task->tk_timeout = delay;
 576         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 577 }
 578 EXPORT_SYMBOL_GPL(rpc_delay);
 579
 580 /*
 581  * Helper to call task->tk_ops->rpc_call_prepare
 582  */
 583 void rpc_prepare_task(struct rpc_task *task)
 584 {
 585         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 586 }
 587
 588 /*
 589  * Helper that calls task->tk_ops->rpc_call_done if it exists
 590  */
 591 void rpc_exit_task(struct rpc_task *task)
 592 {
 593         task->tk_action = NULL;
 594         if (task->tk_ops->rpc_call_done != NULL) {
 595                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 596                 if (task->tk_action != NULL) {
 597                         WARN_ON(RPC_ASSASSINATED(task));
 598                         /* Always release the RPC slot and buffer memory */
 599                         xprt_release(task);
 600                 }
 601         }
 602 }
 603 EXPORT_SYMBOL_GPL(rpc_exit_task);
 604
 605 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 606 {
 607         if (ops->rpc_release != NULL)
 608                 ops->rpc_release(calldata);
 609 }
 610
 611 /*
 612  * This is the RPC `scheduler' (or rather, the finite state machine).
 613  */
 614 static void __rpc_execute(struct rpc_task *task)
 615 {
 616         struct rpc_wait_queue *queue;
 617         int task_is_async = RPC_IS_ASYNC(task);
 618         int status = 0;
 619
 620         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 621                         task->tk_pid, task->tk_flags);
 622
 623         BUG_ON(RPC_IS_QUEUED(task));
 624
 625         for (;;) {
 626
 627                 /*
 628                  * Execute any pending callback.
 629                  */
 630                 if (task->tk_callback) {
 631                         void (*save_callback)(struct rpc_task *);
 632
 633                         /*
 634                          * We set tk_callback to NULL before calling it,
 635                          * in case it sets the tk_callback field itself:
 636                          */
 637                         save_callback = task->tk_callback;
 638                         task->tk_callback = NULL;
 639                         save_callback(task);
 640                 }
 641
 642                 /*
 643                  * Perform the next FSM step.
 644                  * tk_action may be NULL when the task has been killed
 645                  * by someone else.
 646                  */
 647                 if (!RPC_IS_QUEUED(task)) {
 648                         if (task->tk_action == NULL)
 649                                 break;
 650                         task->tk_action(task);
 651                 }
 652
 653                 /*
 654                  * Lockless check for whether task is sleeping or not.
 655                  */
 656                 if (!RPC_IS_QUEUED(task))
 657                         continue;
 658                 /*
 659                  * The queue->lock protects against races with
 660                  * rpc_make_runnable().
 661                  *
 662                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 663                  * rpc_task, rpc_make_runnable() can assign it to a
 664                  * different workqueue. We therefore cannot assume that the
 665                  * rpc_task pointer may still be dereferenced.
 666                  */
 667                 queue = task->tk_waitqueue;
 668                 spin_lock_bh(&queue->lock);
 669                 if (!RPC_IS_QUEUED(task)) {
 670                         spin_unlock_bh(&queue->lock);
 671                         continue;
 672                 }
 673                 rpc_clear_running(task);
 674                 spin_unlock_bh(&queue->lock);
 675                 if (task_is_async)
 676                         return;
 677
 678                 /* sync task: sleep here */
 679                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 680                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 681                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 682                                 TASK_KILLABLE);
 683                 if (status == -ERESTARTSYS) {
 684                         /*
 685                          * When a sync task receives a signal, it exits with
 686                          * -ERESTARTSYS. In order to catch any callbacks that
 687                          * clean up after sleeping on some queue, we don't
 688                          * break the loop here, but go around once more.
 689                          */
 690                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 691                         task->tk_flags |= RPC_TASK_KILLED;
 692                         rpc_exit(task, -ERESTARTSYS);
 693                         rpc_wake_up_task(task);
 694                 }
 695                 rpc_set_running(task);
 696                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 697         }
 698
 699         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 700                         task->tk_status);
 701         /* Release all resources associated with the task */
 702         rpc_release_task(task);
 703 }
 704
 705 /*
 706  * User-visible entry point to the scheduler.
 707  *
 708  * This may be called recursively if e.g. an async NFS task updates
 709  * the attributes and finds that dirty pages must be flushed.
 710  * NOTE: Upon exit of this function the task is guaranteed to be
 711  *       released. In particular note that tk_release() will have
 712  *       been called, so your task memory may have been freed.
 713  */
 714 void rpc_execute(struct rpc_task *task)
 715 {
 716         rpc_set_active(task);
 717         rpc_set_running(task);
 718         __rpc_execute(task);
 719 }
 720
 721 static void rpc_async_schedule(struct work_struct *work)
 722 {
 723         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 724 }
 725
 726 /**
 727  * rpc_malloc - allocate an RPC buffer
 728  * @task: RPC task that will use this buffer
 729  * @size: requested byte size
 730  *
 731  * To prevent rpciod from hanging, this allocator never sleeps,
 732  * returning NULL if the request cannot be serviced immediately.
 733  * The caller can arrange to sleep in a way that is safe for rpciod.
 734  *
 735  * Most requests are 'small' (under 2KiB) and can be serviced from a
 736  * mempool, ensuring that NFS reads and writes can always proceed,
 737  * and that there is good locality of reference for these buffers.
 738  *
 739  * In order to avoid memory starvation triggering more writebacks of
 740  * NFS requests, we avoid using GFP_KERNEL.
 741  */
 742 void *rpc_malloc(struct rpc_task *task, size_t size)
 743 {
 744         struct rpc_buffer *buf;
 745         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 746
 747         size += sizeof(struct rpc_buffer);
 748         if (size <= RPC_BUFFER_MAXSIZE)
 749                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 750         else
 751                 buf = kmalloc(size, gfp);
 752
 753         if (!buf)
 754                 return NULL;
 755
 756         buf->len = size;
 757         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 758                         task->tk_pid, size, buf);
 759         return &buf->data;
 760 }
 761 EXPORT_SYMBOL_GPL(rpc_malloc);
 762
 763 /**
 764  * rpc_free - free buffer allocated via rpc_malloc
 765  * @buffer: buffer to free
 766  *
 767  */
 768 void rpc_free(void *buffer)
 769 {
 770         size_t size;
 771         struct rpc_buffer *buf;
 772
 773         if (!buffer)
 774                 return;
 775
 776         buf = container_of(buffer, struct rpc_buffer, data);
 777         size = buf->len;
 778
 779         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 780                         size, buf);
 781
 782         if (size <= RPC_BUFFER_MAXSIZE)
 783                 mempool_free(buf, rpc_buffer_mempool);
 784         else
 785                 kfree(buf);
 786 }
 787 EXPORT_SYMBOL_GPL(rpc_free);
 788
 789 /*
 790  * Creation and deletion of RPC task structures
 791  */
 792 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 793 {
 794         memset(task, 0, sizeof(*task));
 795         atomic_set(&task->tk_count, 1);
 796         task->tk_flags  = task_setup_data->flags;
 797         task->tk_ops = task_setup_data->callback_ops;
 798         task->tk_calldata = task_setup_data->callback_data;
 799         INIT_LIST_HEAD(&task->tk_task);
 800
 801         /* Initialize retry counters */
 802         task->tk_garb_retry = 2;
 803         task->tk_cred_retry = 2;
 804
 805         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 806         task->tk_owner = current->tgid;
 807
 808         /* Initialize workqueue for async tasks */
 809         task->tk_workqueue = task_setup_data->workqueue;
 810
 811         task->tk_client = task_setup_data->rpc_client;
 812         if (task->tk_client != NULL) {
 813                 kref_get(&task->tk_client->cl_kref);
 814                 if (task->tk_client->cl_softrtry)
 815                         task->tk_flags |= RPC_TASK_SOFT;
 816         }
 817
 818         if (task->tk_ops->rpc_call_prepare != NULL)
 819                 task->tk_action = rpc_prepare_task;
 820
 821         if (task_setup_data->rpc_message != NULL) {
 822                 task->tk_msg.rpc_proc = task_setup_data->rpc_message->rpc_proc;
 823                 task->tk_msg.rpc_argp = task_setup_data->rpc_message->rpc_argp;
 824                 task->tk_msg.rpc_resp = task_setup_data->rpc_message->rpc_resp;
 825                 /* Bind the user cred */
 826                 rpcauth_bindcred(task, task_setup_data->rpc_message->rpc_cred, task_setup_data->flags);
 827                 if (task->tk_action == NULL)
 828                         rpc_call_start(task);
 829         }
 830
 831         /* starting timestamp */
 832         task->tk_start = ktime_get();
 833
 834         dprintk("RPC:       new task initialized, procpid %u\n",
 835                                 task_pid_nr(current));
 836 }
 837
 838 static struct rpc_task *
 839 rpc_alloc_task(void)
 840 {
 841         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 842 }
 843
 844 /*
 845  * Create a new task for the specified client.
 846  */
 847 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 848 {
 849         struct rpc_task *task = setup_data->task;
 850         unsigned short flags = 0;
 851
 852         if (task == NULL) {
 853                 task = rpc_alloc_task();
 854                 if (task == NULL) {
 855                         rpc_release_calldata(setup_data->callback_ops,
 856                                         setup_data->callback_data);
 857                         return ERR_PTR(-ENOMEM);
 858                 }
 859                 flags = RPC_TASK_DYNAMIC;
 860         }
 861
 862         rpc_init_task(task, setup_data);
 863         if (task->tk_status < 0) {
 864                 int err = task->tk_status;
 865                 rpc_put_task(task);
 866                 return ERR_PTR(err);
 867         }
 868
 869         task->tk_flags |= flags;
 870         dprintk("RPC:       allocated task %p\n", task);
 871         return task;
 872 }
 873
 874 static void rpc_free_task(struct rpc_task *task)
 875 {
 876         const struct rpc_call_ops *tk_ops = task->tk_ops;
 877         void *calldata = task->tk_calldata;
 878
 879         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 880                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 881                 mempool_free(task, rpc_task_mempool);
 882         }
 883         rpc_release_calldata(tk_ops, calldata);
 884 }
 885
 886 static void rpc_async_release(struct work_struct *work)
 887 {
 888         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 889 }
 890
 891 void rpc_put_task(struct rpc_task *task)
 892 {
 893         if (!atomic_dec_and_test(&task->tk_count))
 894                 return;
 895         /* Release resources */
 896         if (task->tk_rqstp)
 897                 xprt_release(task);
 898         if (task->tk_msg.rpc_cred)
 899                 rpcauth_unbindcred(task);
 900         if (task->tk_client) {
 901                 rpc_release_client(task->tk_client);
 902                 task->tk_client = NULL;
 903         }
 904         if (task->tk_workqueue != NULL) {
 905                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 906                 queue_work(task->tk_workqueue, &task->u.tk_work);
 907         } else
 908                 rpc_free_task(task);
 909 }
 910 EXPORT_SYMBOL_GPL(rpc_put_task);
 911
 912 static void rpc_release_task(struct rpc_task *task)
 913 {
 914         dprintk("RPC: %5u release task\n", task->tk_pid);
 915
 916         if (!list_empty(&task->tk_task)) {
 917                 struct rpc_clnt *clnt = task->tk_client;
 918                 /* Remove from client task list */
 919                 spin_lock(&clnt->cl_lock);
 920                 list_del(&task->tk_task);
 921                 spin_unlock(&clnt->cl_lock);
 922         }
 923         BUG_ON (RPC_IS_QUEUED(task));
 924
 925         /* Wake up anyone who is waiting for task completion */
 926         rpc_mark_complete_task(task);
 927
 928         rpc_put_task(task);
 929 }
 930
 931 /*
 932  * Kill all tasks for the given client.
 933  * XXX: kill their descendants as well?
 934  */
 935 void rpc_killall_tasks(struct rpc_clnt *clnt)
 936 {
 937         struct rpc_task *rovr;
 938
 939
 940         if (list_empty(&clnt->cl_tasks))
 941                 return;
 942         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 943         /*
 944          * Spin lock all_tasks to prevent changes...
 945          */
 946         spin_lock(&clnt->cl_lock);
 947         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 948                 if (! RPC_IS_ACTIVATED(rovr))
 949                         continue;
 950                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 951                         rovr->tk_flags |= RPC_TASK_KILLED;
 952                         rpc_exit(rovr, -EIO);
 953                         rpc_wake_up_task(rovr);
 954                 }
 955         }
 956         spin_unlock(&clnt->cl_lock);
 957 }
 958 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 959
 960 int rpciod_up(void)
 961 {
 962         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 963 }
 964
 965 void rpciod_down(void)
 966 {
 967         module_put(THIS_MODULE);
 968 }
 969
 970 /*
 971  * Start up the rpciod workqueue.
 972  */
 973 static int rpciod_start(void)
 974 {
 975         struct workqueue_struct *wq;
 976
 977         /*
 978          * Create the rpciod thread and wait for it to start.
 979          */
 980         dprintk("RPC:       creating workqueue rpciod\n");
 981         wq = create_workqueue("rpciod");
 982         rpciod_workqueue = wq;
 983         return rpciod_workqueue != NULL;
 984 }
 985
 986 static void rpciod_stop(void)
 987 {
 988         struct workqueue_struct *wq = NULL;
 989
 990         if (rpciod_workqueue == NULL)
 991                 return;
 992         dprintk("RPC:       destroying workqueue rpciod\n");
 993
 994         wq = rpciod_workqueue;
 995         rpciod_workqueue = NULL;
 996         destroy_workqueue(wq);
 997 }
 998
 999 void
1000 rpc_destroy_mempool(void)
1001 {
1002         rpciod_stop();
1003         if (rpc_buffer_mempool)
1004                 mempool_destroy(rpc_buffer_mempool);
1005         if (rpc_task_mempool)
1006                 mempool_destroy(rpc_task_mempool);
1007         if (rpc_task_slabp)
1008                 kmem_cache_destroy(rpc_task_slabp);
1009         if (rpc_buffer_slabp)
1010                 kmem_cache_destroy(rpc_buffer_slabp);
1011         rpc_destroy_wait_queue(&delay_queue);
1012 }
1013
1014 int
1015 rpc_init_mempool(void)
1016 {
1017         /*
1018          * The following is not strictly a mempool initialisation,
1019          * but there is no harm in doing it here
1020          */
1021         rpc_init_wait_queue(&delay_queue, "delayq");
1022         if (!rpciod_start())
1023                 goto err_nomem;
1024
1025         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1026                                              sizeof(struct rpc_task),
1027                                              0, SLAB_HWCACHE_ALIGN,
1028                                              NULL);
1029         if (!rpc_task_slabp)
1030                 goto err_nomem;
1031         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1032                                              RPC_BUFFER_MAXSIZE,
1033                                              0, SLAB_HWCACHE_ALIGN,
1034                                              NULL);
1035         if (!rpc_buffer_slabp)
1036                 goto err_nomem;
1037         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1038                                                     rpc_task_slabp);
1039         if (!rpc_task_mempool)
1040                 goto err_nomem;
1041         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1042                                                       rpc_buffer_slabp);
1043         if (!rpc_buffer_mempool)
1044                 goto err_nomem;
1045         return 0;
1046 err_nomem:
1047         rpc_destroy_mempool();
1048         return -ENOMEM;
1049 }