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,
 101                 struct rpc_task *task,
 102                 unsigned char queue_priority)
 103 {
 104         struct list_head *q;
 105         struct rpc_task *t;
 106
 107         INIT_LIST_HEAD(&task->u.tk_wait.links);
 108         q = &queue->tasks[queue_priority];
 109         if (unlikely(queue_priority > queue->maxpriority))
 110                 q = &queue->tasks[queue->maxpriority];
 111         list_for_each_entry(t, q, u.tk_wait.list) {
 112                 if (t->tk_owner == task->tk_owner) {
 113                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 114                         return;
 115                 }
 116         }
 117         list_add_tail(&task->u.tk_wait.list, q);
 118 }
 119
 120 /*
 121  * Add new request to wait queue.
 122  *
 123  * Swapper tasks always get inserted at the head of the queue.
 124  * This should avoid many nasty memory deadlocks and hopefully
 125  * improve overall performance.
 126  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 127  */
 128 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 129                 struct rpc_task *task,
 130                 unsigned char queue_priority)
 131 {
 132         BUG_ON (RPC_IS_QUEUED(task));
 133
 134         if (RPC_IS_PRIORITY(queue))
 135                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 136         else if (RPC_IS_SWAPPER(task))
 137                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 138         else
 139                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 140         task->tk_waitqueue = queue;
 141         queue->qlen++;
 142         rpc_set_queued(task);
 143
 144         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 145                         task->tk_pid, queue, rpc_qname(queue));
 146 }
 147
 148 /*
 149  * Remove request from a priority queue.
 150  */
 151 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 152 {
 153         struct rpc_task *t;
 154
 155         if (!list_empty(&task->u.tk_wait.links)) {
 156                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 157                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 158                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 159         }
 160 }
 161
 162 /*
 163  * Remove request from queue.
 164  * Note: must be called with spin lock held.
 165  */
 166 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 167 {
 168         __rpc_disable_timer(queue, task);
 169         if (RPC_IS_PRIORITY(queue))
 170                 __rpc_remove_wait_queue_priority(task);
 171         list_del(&task->u.tk_wait.list);
 172         queue->qlen--;
 173         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 174                         task->tk_pid, queue, rpc_qname(queue));
 175 }
 176
 177 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 178 {
 179         queue->priority = priority;
 180         queue->count = 1 << (priority * 2);
 181 }
 182
 183 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 184 {
 185         queue->owner = pid;
 186         queue->nr = RPC_BATCH_COUNT;
 187 }
 188
 189 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 190 {
 191         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 192         rpc_set_waitqueue_owner(queue, 0);
 193 }
 194
 195 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 196 {
 197         int i;
 198
 199         spin_lock_init(&queue->lock);
 200         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 201                 INIT_LIST_HEAD(&queue->tasks[i]);
 202         queue->maxpriority = nr_queues - 1;
 203         rpc_reset_waitqueue_priority(queue);
 204         queue->qlen = 0;
 205         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 206         INIT_LIST_HEAD(&queue->timer_list.list);
 207 #ifdef RPC_DEBUG
 208         queue->name = qname;
 209 #endif
 210 }
 211
 212 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 213 {
 214         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 215 }
 216 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 217
 218 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 219 {
 220         __rpc_init_priority_wait_queue(queue, qname, 1);
 221 }
 222 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 223
 224 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 225 {
 226         del_timer_sync(&queue->timer_list.timer);
 227 }
 228 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 229
 230 static int rpc_wait_bit_killable(void *word)
 231 {
 232         if (fatal_signal_pending(current))
 233                 return -ERESTARTSYS;
 234         schedule();
 235         return 0;
 236 }
 237
 238 #ifdef RPC_DEBUG
 239 static void rpc_task_set_debuginfo(struct rpc_task *task)
 240 {
 241         static atomic_t rpc_pid;
 242
 243         task->tk_pid = atomic_inc_return(&rpc_pid);
 244 }
 245 #else
 246 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 247 {
 248 }
 249 #endif
 250
 251 static void rpc_set_active(struct rpc_task *task)
 252 {
 253         rpc_task_set_debuginfo(task);
 254         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 255 }
 256
 257 /*
 258  * Mark an RPC call as having completed by clearing the 'active' bit
 259  * and then waking up all tasks that were sleeping.
 260  */
 261 static int rpc_complete_task(struct rpc_task *task)
 262 {
 263         void *m = &task->tk_runstate;
 264         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 265         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 266         unsigned long flags;
 267         int ret;
 268
 269         spin_lock_irqsave(&wq->lock, flags);
 270         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 271         ret = atomic_dec_and_test(&task->tk_count);
 272         if (waitqueue_active(wq))
 273                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 274         spin_unlock_irqrestore(&wq->lock, flags);
 275         return ret;
 276 }
 277
 278 /*
 279  * Allow callers to wait for completion of an RPC call
 280  *
 281  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 282  * to enforce taking of the wq->lock and hence avoid races with
 283  * rpc_complete_task().
 284  */
 285 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 286 {
 287         if (action == NULL)
 288                 action = rpc_wait_bit_killable;
 289         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 290                         action, TASK_KILLABLE);
 291 }
 292 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 293
 294 /*
 295  * Make an RPC task runnable.
 296  *
 297  * Note: If the task is ASYNC, this must be called with
 298  * the spinlock held to protect the wait queue operation.
 299  */
 300 static void rpc_make_runnable(struct rpc_task *task)
 301 {
 302         rpc_clear_queued(task);
 303         if (rpc_test_and_set_running(task))
 304                 return;
 305         if (RPC_IS_ASYNC(task)) {
 306                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 307                 queue_work(rpciod_workqueue, &task->u.tk_work);
 308         } else
 309                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 310 }
 311
 312 /*
 313  * Prepare for sleeping on a wait queue.
 314  * By always appending tasks to the list we ensure FIFO behavior.
 315  * NB: An RPC task will only receive interrupt-driven events as long
 316  * as it's on a wait queue.
 317  */
 318 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 319                 struct rpc_task *task,
 320                 rpc_action action,
 321                 unsigned char queue_priority)
 322 {
 323         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 324                         task->tk_pid, rpc_qname(q), jiffies);
 325
 326         __rpc_add_wait_queue(q, task, queue_priority);
 327
 328         BUG_ON(task->tk_callback != NULL);
 329         task->tk_callback = action;
 330         __rpc_add_timer(q, task);
 331 }
 332
 333 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 334                                 rpc_action action)
 335 {
 336         /* We shouldn't ever put an inactive task to sleep */
 337         BUG_ON(!RPC_IS_ACTIVATED(task));
 338
 339         /*
 340          * Protect the queue operations.
 341          */
 342         spin_lock_bh(&q->lock);
 343         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 344         spin_unlock_bh(&q->lock);
 345 }
 346 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 347
 348 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 349                 rpc_action action, int priority)
 350 {
 351         /* We shouldn't ever put an inactive task to sleep */
 352         BUG_ON(!RPC_IS_ACTIVATED(task));
 353
 354         /*
 355          * Protect the queue operations.
 356          */
 357         spin_lock_bh(&q->lock);
 358         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 359         spin_unlock_bh(&q->lock);
 360 }
 361
 362 /**
 363  * __rpc_do_wake_up_task - wake up a single rpc_task
 364  * @queue: wait queue
 365  * @task: task to be woken up
 366  *
 367  * Caller must hold queue->lock, and have cleared the task queued flag.
 368  */
 369 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 370 {
 371         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 372                         task->tk_pid, jiffies);
 373
 374         /* Has the task been executed yet? If not, we cannot wake it up! */
 375         if (!RPC_IS_ACTIVATED(task)) {
 376                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 377                 return;
 378         }
 379
 380         __rpc_remove_wait_queue(queue, task);
 381
 382         rpc_make_runnable(task);
 383
 384         dprintk("RPC:       __rpc_wake_up_task done\n");
 385 }
 386
 387 /*
 388  * Wake up a queued task while the queue lock is being held
 389  */
 390 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 391 {
 392         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 393                 __rpc_do_wake_up_task(queue, task);
 394 }
 395
 396 /*
 397  * Tests whether rpc queue is empty
 398  */
 399 int rpc_queue_empty(struct rpc_wait_queue *queue)
 400 {
 401         int res;
 402
 403         spin_lock_bh(&queue->lock);
 404         res = queue->qlen;
 405         spin_unlock_bh(&queue->lock);
 406         return res == 0;
 407 }
 408 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 409
 410 /*
 411  * Wake up a task on a specific queue
 412  */
 413 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 414 {
 415         spin_lock_bh(&queue->lock);
 416         rpc_wake_up_task_queue_locked(queue, task);
 417         spin_unlock_bh(&queue->lock);
 418 }
 419 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 420
 421 /*
 422  * Wake up the next task on a priority queue.
 423  */
 424 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 425 {
 426         struct list_head *q;
 427         struct rpc_task *task;
 428
 429         /*
 430          * Service a batch of tasks from a single owner.
 431          */
 432         q = &queue->tasks[queue->priority];
 433         if (!list_empty(q)) {
 434                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 435                 if (queue->owner == task->tk_owner) {
 436                         if (--queue->nr)
 437                                 goto out;
 438                         list_move_tail(&task->u.tk_wait.list, q);
 439                 }
 440                 /*
 441                  * Check if we need to switch queues.
 442                  */
 443                 if (--queue->count)
 444                         goto new_owner;
 445         }
 446
 447         /*
 448          * Service the next queue.
 449          */
 450         do {
 451                 if (q == &queue->tasks[0])
 452                         q = &queue->tasks[queue->maxpriority];
 453                 else
 454                         q = q - 1;
 455                 if (!list_empty(q)) {
 456                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 457                         goto new_queue;
 458                 }
 459         } while (q != &queue->tasks[queue->priority]);
 460
 461         rpc_reset_waitqueue_priority(queue);
 462         return NULL;
 463
 464 new_queue:
 465         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 466 new_owner:
 467         rpc_set_waitqueue_owner(queue, task->tk_owner);
 468 out:
 469         rpc_wake_up_task_queue_locked(queue, task);
 470         return task;
 471 }
 472
 473 /*
 474  * Wake up the next task on the wait queue.
 475  */
 476 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 477 {
 478         struct rpc_task *task = NULL;
 479
 480         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 481                         queue, rpc_qname(queue));
 482         spin_lock_bh(&queue->lock);
 483         if (RPC_IS_PRIORITY(queue))
 484                 task = __rpc_wake_up_next_priority(queue);
 485         else {
 486                 task_for_first(task, &queue->tasks[0])
 487                         rpc_wake_up_task_queue_locked(queue, task);
 488         }
 489         spin_unlock_bh(&queue->lock);
 490
 491         return task;
 492 }
 493 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 494
 495 /**
 496  * rpc_wake_up - wake up all rpc_tasks
 497  * @queue: rpc_wait_queue on which the tasks are sleeping
 498  *
 499  * Grabs queue->lock
 500  */
 501 void rpc_wake_up(struct rpc_wait_queue *queue)
 502 {
 503         struct rpc_task *task, *next;
 504         struct list_head *head;
 505
 506         spin_lock_bh(&queue->lock);
 507         head = &queue->tasks[queue->maxpriority];
 508         for (;;) {
 509                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 510                         rpc_wake_up_task_queue_locked(queue, task);
 511                 if (head == &queue->tasks[0])
 512                         break;
 513                 head--;
 514         }
 515         spin_unlock_bh(&queue->lock);
 516 }
 517 EXPORT_SYMBOL_GPL(rpc_wake_up);
 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_queue_locked(queue, task);
 537                 }
 538                 if (head == &queue->tasks[0])
 539                         break;
 540                 head--;
 541         }
 542         spin_unlock_bh(&queue->lock);
 543 }
 544 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 545
 546 /**
 547  * rpc_wake_up_softconn_status - wake up all SOFTCONN rpc_tasks and set their
 548  * status value.
 549  * @queue: rpc_wait_queue on which the tasks are sleeping
 550  * @status: status value to set
 551  *
 552  * Grabs queue->lock
 553  */
 554 void rpc_wake_up_softconn_status(struct rpc_wait_queue *queue, int status)
 555 {
 556         struct rpc_task *task, *next;
 557         struct list_head *head;
 558
 559         spin_lock_bh(&queue->lock);
 560         head = &queue->tasks[queue->maxpriority];
 561         for (;;) {
 562                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 563                         if (RPC_IS_SOFTCONN(task)) {
 564                                 task->tk_status = status;
 565                                 rpc_wake_up_task_queue_locked(queue, task);
 566                         }
 567                 if (head == &queue->tasks[0])
 568                         break;
 569                 head--;
 570         }
 571         spin_unlock_bh(&queue->lock);
 572 }
 573 EXPORT_SYMBOL_GPL(rpc_wake_up_softconn_status);
 574
 575 static void __rpc_queue_timer_fn(unsigned long ptr)
 576 {
 577         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 578         struct rpc_task *task, *n;
 579         unsigned long expires, now, timeo;
 580
 581         spin_lock(&queue->lock);
 582         expires = now = jiffies;
 583         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 584                 timeo = task->u.tk_wait.expires;
 585                 if (time_after_eq(now, timeo)) {
 586                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 587                         task->tk_status = -ETIMEDOUT;
 588                         rpc_wake_up_task_queue_locked(queue, task);
 589                         continue;
 590                 }
 591                 if (expires == now || time_after(expires, timeo))
 592                         expires = timeo;
 593         }
 594         if (!list_empty(&queue->timer_list.list))
 595                 rpc_set_queue_timer(queue, expires);
 596         spin_unlock(&queue->lock);
 597 }
 598
 599 static void __rpc_atrun(struct rpc_task *task)
 600 {
 601         task->tk_status = 0;
 602 }
 603
 604 /*
 605  * Run a task at a later time
 606  */
 607 void rpc_delay(struct rpc_task *task, unsigned long delay)
 608 {
 609         task->tk_timeout = delay;
 610         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 611 }
 612 EXPORT_SYMBOL_GPL(rpc_delay);
 613
 614 /*
 615  * Helper to call task->tk_ops->rpc_call_prepare
 616  */
 617 void rpc_prepare_task(struct rpc_task *task)
 618 {
 619         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 620 }
 621
 622 static void
 623 rpc_init_task_statistics(struct rpc_task *task)
 624 {
 625         /* Initialize retry counters */
 626         task->tk_garb_retry = 2;
 627         task->tk_cred_retry = 2;
 628         task->tk_rebind_retry = 2;
 629
 630         /* starting timestamp */
 631         task->tk_start = ktime_get();
 632 }
 633
 634 static void
 635 rpc_reset_task_statistics(struct rpc_task *task)
 636 {
 637         task->tk_timeouts = 0;
 638         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 639
 640         rpc_init_task_statistics(task);
 641 }
 642
 643 /*
 644  * Helper that calls task->tk_ops->rpc_call_done if it exists
 645  */
 646 void rpc_exit_task(struct rpc_task *task)
 647 {
 648         task->tk_action = NULL;
 649         if (task->tk_ops->rpc_call_done != NULL) {
 650                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 651                 if (task->tk_action != NULL) {
 652                         WARN_ON(RPC_ASSASSINATED(task));
 653                         /* Always release the RPC slot and buffer memory */
 654                         xprt_release(task);
 655                         rpc_reset_task_statistics(task);
 656                 }
 657         }
 658 }
 659
 660 void rpc_exit(struct rpc_task *task, int status)
 661 {
 662         task->tk_status = status;
 663         task->tk_action = rpc_exit_task;
 664         if (RPC_IS_QUEUED(task))
 665                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 666 }
 667 EXPORT_SYMBOL_GPL(rpc_exit);
 668
 669 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 670 {
 671         if (ops->rpc_release != NULL)
 672                 ops->rpc_release(calldata);
 673 }
 674
 675 /*
 676  * This is the RPC `scheduler' (or rather, the finite state machine).
 677  */
 678 static void __rpc_execute(struct rpc_task *task)
 679 {
 680         struct rpc_wait_queue *queue;
 681         int task_is_async = RPC_IS_ASYNC(task);
 682         int status = 0;
 683
 684         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 685                         task->tk_pid, task->tk_flags);
 686
 687         BUG_ON(RPC_IS_QUEUED(task));
 688
 689         for (;;) {
 690                 void (*do_action)(struct rpc_task *);
 691
 692                 /*
 693                  * Execute any pending callback first.
 694                  */
 695                 do_action = task->tk_callback;
 696                 task->tk_callback = NULL;
 697                 if (do_action == NULL) {
 698                         /*
 699                          * Perform the next FSM step.
 700                          * tk_action may be NULL if the task has been killed.
 701                          * In particular, note that rpc_killall_tasks may
 702                          * do this at any time, so beware when dereferencing.
 703                          */
 704                         do_action = task->tk_action;
 705                         if (do_action == NULL)
 706                                 break;
 707                 }
 708                 do_action(task);
 709
 710                 /*
 711                  * Lockless check for whether task is sleeping or not.
 712                  */
 713                 if (!RPC_IS_QUEUED(task))
 714                         continue;
 715                 /*
 716                  * The queue->lock protects against races with
 717                  * rpc_make_runnable().
 718                  *
 719                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 720                  * rpc_task, rpc_make_runnable() can assign it to a
 721                  * different workqueue. We therefore cannot assume that the
 722                  * rpc_task pointer may still be dereferenced.
 723                  */
 724                 queue = task->tk_waitqueue;
 725                 spin_lock_bh(&queue->lock);
 726                 if (!RPC_IS_QUEUED(task)) {
 727                         spin_unlock_bh(&queue->lock);
 728                         continue;
 729                 }
 730                 rpc_clear_running(task);
 731                 spin_unlock_bh(&queue->lock);
 732                 if (task_is_async)
 733                         return;
 734
 735                 /* sync task: sleep here */
 736                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 737                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 738                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 739                                 TASK_KILLABLE);
 740                 if (status == -ERESTARTSYS) {
 741                         /*
 742                          * When a sync task receives a signal, it exits with
 743                          * -ERESTARTSYS. In order to catch any callbacks that
 744                          * clean up after sleeping on some queue, we don't
 745                          * break the loop here, but go around once more.
 746                          */
 747                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 748                         task->tk_flags |= RPC_TASK_KILLED;
 749                         rpc_exit(task, -ERESTARTSYS);
 750                 }
 751                 rpc_set_running(task);
 752                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 753         }
 754
 755         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 756                         task->tk_status);
 757         /* Release all resources associated with the task */
 758         rpc_release_task(task);
 759 }
 760
 761 /*
 762  * User-visible entry point to the scheduler.
 763  *
 764  * This may be called recursively if e.g. an async NFS task updates
 765  * the attributes and finds that dirty pages must be flushed.
 766  * NOTE: Upon exit of this function the task is guaranteed to be
 767  *       released. In particular note that tk_release() will have
 768  *       been called, so your task memory may have been freed.
 769  */
 770 void rpc_execute(struct rpc_task *task)
 771 {
 772         rpc_set_active(task);
 773         rpc_make_runnable(task);
 774         if (!RPC_IS_ASYNC(task))
 775                 __rpc_execute(task);
 776 }
 777
 778 static void rpc_async_schedule(struct work_struct *work)
 779 {
 780         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 781 }
 782
 783 /**
 784  * rpc_malloc - allocate an RPC buffer
 785  * @task: RPC task that will use this buffer
 786  * @size: requested byte size
 787  *
 788  * To prevent rpciod from hanging, this allocator never sleeps,
 789  * returning NULL if the request cannot be serviced immediately.
 790  * The caller can arrange to sleep in a way that is safe for rpciod.
 791  *
 792  * Most requests are 'small' (under 2KiB) and can be serviced from a
 793  * mempool, ensuring that NFS reads and writes can always proceed,
 794  * and that there is good locality of reference for these buffers.
 795  *
 796  * In order to avoid memory starvation triggering more writebacks of
 797  * NFS requests, we avoid using GFP_KERNEL.
 798  */
 799 void *rpc_malloc(struct rpc_task *task, size_t size)
 800 {
 801         struct rpc_buffer *buf;
 802         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 803
 804         size += sizeof(struct rpc_buffer);
 805         if (size <= RPC_BUFFER_MAXSIZE)
 806                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 807         else
 808                 buf = kmalloc(size, gfp);
 809
 810         if (!buf)
 811                 return NULL;
 812
 813         buf->len = size;
 814         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 815                         task->tk_pid, size, buf);
 816         return &buf->data;
 817 }
 818 EXPORT_SYMBOL_GPL(rpc_malloc);
 819
 820 /**
 821  * rpc_free - free buffer allocated via rpc_malloc
 822  * @buffer: buffer to free
 823  *
 824  */
 825 void rpc_free(void *buffer)
 826 {
 827         size_t size;
 828         struct rpc_buffer *buf;
 829
 830         if (!buffer)
 831                 return;
 832
 833         buf = container_of(buffer, struct rpc_buffer, data);
 834         size = buf->len;
 835
 836         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 837                         size, buf);
 838
 839         if (size <= RPC_BUFFER_MAXSIZE)
 840                 mempool_free(buf, rpc_buffer_mempool);
 841         else
 842                 kfree(buf);
 843 }
 844 EXPORT_SYMBOL_GPL(rpc_free);
 845
 846 /*
 847  * Creation and deletion of RPC task structures
 848  */
 849 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 850 {
 851         memset(task, 0, sizeof(*task));
 852         atomic_set(&task->tk_count, 1);
 853         task->tk_flags  = task_setup_data->flags;
 854         task->tk_ops = task_setup_data->callback_ops;
 855         task->tk_calldata = task_setup_data->callback_data;
 856         INIT_LIST_HEAD(&task->tk_task);
 857
 858         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 859         task->tk_owner = current->tgid;
 860
 861         /* Initialize workqueue for async tasks */
 862         task->tk_workqueue = task_setup_data->workqueue;
 863
 864         if (task->tk_ops->rpc_call_prepare != NULL)
 865                 task->tk_action = rpc_prepare_task;
 866
 867         rpc_init_task_statistics(task);
 868
 869         dprintk("RPC:       new task initialized, procpid %u\n",
 870                                 task_pid_nr(current));
 871 }
 872
 873 static struct rpc_task *
 874 rpc_alloc_task(void)
 875 {
 876         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 877 }
 878
 879 /*
 880  * Create a new task for the specified client.
 881  */
 882 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 883 {
 884         struct rpc_task *task = setup_data->task;
 885         unsigned short flags = 0;
 886
 887         if (task == NULL) {
 888                 task = rpc_alloc_task();
 889                 if (task == NULL) {
 890                         rpc_release_calldata(setup_data->callback_ops,
 891                                         setup_data->callback_data);
 892                         return ERR_PTR(-ENOMEM);
 893                 }
 894                 flags = RPC_TASK_DYNAMIC;
 895         }
 896
 897         rpc_init_task(task, setup_data);
 898         task->tk_flags |= flags;
 899         dprintk("RPC:       allocated task %p\n", task);
 900         return task;
 901 }
 902
 903 static void rpc_free_task(struct rpc_task *task)
 904 {
 905         const struct rpc_call_ops *tk_ops = task->tk_ops;
 906         void *calldata = task->tk_calldata;
 907
 908         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 909                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 910                 mempool_free(task, rpc_task_mempool);
 911         }
 912         rpc_release_calldata(tk_ops, calldata);
 913 }
 914
 915 static void rpc_async_release(struct work_struct *work)
 916 {
 917         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 918 }
 919
 920 static void rpc_release_resources_task(struct rpc_task *task)
 921 {
 922         if (task->tk_rqstp)
 923                 xprt_release(task);
 924         if (task->tk_msg.rpc_cred) {
 925                 put_rpccred(task->tk_msg.rpc_cred);
 926                 task->tk_msg.rpc_cred = NULL;
 927         }
 928         rpc_task_release_client(task);
 929 }
 930
 931 static void rpc_final_put_task(struct rpc_task *task,
 932                 struct workqueue_struct *q)
 933 {
 934         if (q != NULL) {
 935                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 936                 queue_work(q, &task->u.tk_work);
 937         } else
 938                 rpc_free_task(task);
 939 }
 940
 941 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
 942 {
 943         if (atomic_dec_and_test(&task->tk_count)) {
 944                 rpc_release_resources_task(task);
 945                 rpc_final_put_task(task, q);
 946         }
 947 }
 948
 949 void rpc_put_task(struct rpc_task *task)
 950 {
 951         rpc_do_put_task(task, NULL);
 952 }
 953 EXPORT_SYMBOL_GPL(rpc_put_task);
 954
 955 void rpc_put_task_async(struct rpc_task *task)
 956 {
 957         rpc_do_put_task(task, task->tk_workqueue);
 958 }
 959 EXPORT_SYMBOL_GPL(rpc_put_task_async);
 960
 961 static void rpc_release_task(struct rpc_task *task)
 962 {
 963         dprintk("RPC: %5u release task\n", task->tk_pid);
 964
 965         BUG_ON (RPC_IS_QUEUED(task));
 966
 967         rpc_release_resources_task(task);
 968
 969         /*
 970          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
 971          * so it should be safe to use task->tk_count as a test for whether
 972          * or not any other processes still hold references to our rpc_task.
 973          */
 974         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
 975                 /* Wake up anyone who may be waiting for task completion */
 976                 if (!rpc_complete_task(task))
 977                         return;
 978         } else {
 979                 if (!atomic_dec_and_test(&task->tk_count))
 980                         return;
 981         }
 982         rpc_final_put_task(task, task->tk_workqueue);
 983 }
 984
 985 int rpciod_up(void)
 986 {
 987         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 988 }
 989
 990 void rpciod_down(void)
 991 {
 992         module_put(THIS_MODULE);
 993 }
 994
 995 /*
 996  * Start up the rpciod workqueue.
 997  */
 998 static int rpciod_start(void)
 999 {
1000         struct workqueue_struct *wq;
1001
1002         /*
1003          * Create the rpciod thread and wait for it to start.
1004          */
1005         dprintk("RPC:       creating workqueue rpciod\n");
1006         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1007         rpciod_workqueue = wq;
1008         return rpciod_workqueue != NULL;
1009 }
1010
1011 static void rpciod_stop(void)
1012 {
1013         struct workqueue_struct *wq = NULL;
1014
1015         if (rpciod_workqueue == NULL)
1016                 return;
1017         dprintk("RPC:       destroying workqueue rpciod\n");
1018
1019         wq = rpciod_workqueue;
1020         rpciod_workqueue = NULL;
1021         destroy_workqueue(wq);
1022 }
1023
1024 void
1025 rpc_destroy_mempool(void)
1026 {
1027         rpciod_stop();
1028         if (rpc_buffer_mempool)
1029                 mempool_destroy(rpc_buffer_mempool);
1030         if (rpc_task_mempool)
1031                 mempool_destroy(rpc_task_mempool);
1032         if (rpc_task_slabp)
1033                 kmem_cache_destroy(rpc_task_slabp);
1034         if (rpc_buffer_slabp)
1035                 kmem_cache_destroy(rpc_buffer_slabp);
1036         rpc_destroy_wait_queue(&delay_queue);
1037 }
1038
1039 int
1040 rpc_init_mempool(void)
1041 {
1042         /*
1043          * The following is not strictly a mempool initialisation,
1044          * but there is no harm in doing it here
1045          */
1046         rpc_init_wait_queue(&delay_queue, "delayq");
1047         if (!rpciod_start())
1048                 goto err_nomem;
1049
1050         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1051                                              sizeof(struct rpc_task),
1052                                              0, SLAB_HWCACHE_ALIGN,
1053                                              NULL);
1054         if (!rpc_task_slabp)
1055                 goto err_nomem;
1056         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1057                                              RPC_BUFFER_MAXSIZE,
1058                                              0, SLAB_HWCACHE_ALIGN,
1059                                              NULL);
1060         if (!rpc_buffer_slabp)
1061                 goto err_nomem;
1062         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1063                                                     rpc_task_slabp);
1064         if (!rpc_task_mempool)
1065                 goto err_nomem;
1066         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1067                                                       rpc_buffer_slabp);
1068         if (!rpc_buffer_mempool)
1069                 goto err_nomem;
1070         return 0;
1071 err_nomem:
1072         rpc_destroy_mempool();
1073         return -ENOMEM;
1074 }