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 #include <linux/freezer.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #include "sunrpc.h"
  26
  27 #ifdef RPC_DEBUG
  28 #define RPCDBG_FACILITY         RPCDBG_SCHED
  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 struct kmem_cache        *rpc_task_slabp __read_mostly;
  38 static struct kmem_cache        *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_async_schedule(struct work_struct *);
  43 static void                      rpc_release_task(struct rpc_task *task);
  44 static void __rpc_queue_timer_fn(unsigned long ptr);
  45
  46 /*
  47  * RPC tasks sit here while waiting for conditions to improve.
  48  */
  49 static struct rpc_wait_queue delay_queue;
  50
  51 /*
  52  * rpciod-related stuff
  53  */
  54 struct workqueue_struct *rpciod_workqueue;
  55
  56 /*
  57  * Disable the timer for a given RPC task. Should be called with
  58  * queue->lock and bh_disabled in order to avoid races within
  59  * rpc_run_timer().
  60  */
  61 static void
  62 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  63 {
  64         if (task->tk_timeout == 0)
  65                 return;
  66         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  67         task->tk_timeout = 0;
  68         list_del(&task->u.tk_wait.timer_list);
  69         if (list_empty(&queue->timer_list.list))
  70                 del_timer(&queue->timer_list.timer);
  71 }
  72
  73 static void
  74 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  75 {
  76         queue->timer_list.expires = expires;
  77         mod_timer(&queue->timer_list.timer, expires);
  78 }
  79
  80 /*
  81  * Set up a timer for the current task.
  82  */
  83 static void
  84 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  85 {
  86         if (!task->tk_timeout)
  87                 return;
  88
  89         dprintk("RPC: %5u setting alarm for %lu ms\n",
  90                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  91
  92         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  93         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  94                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  95         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  96 }
  97
  98 /*
  99  * Add new request to a priority queue.
 100  */
 101 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 102                 struct rpc_task *task,
 103                 unsigned char queue_priority)
 104 {
 105         struct list_head *q;
 106         struct rpc_task *t;
 107
 108         INIT_LIST_HEAD(&task->u.tk_wait.links);
 109         q = &queue->tasks[queue_priority];
 110         if (unlikely(queue_priority > queue->maxpriority))
 111                 q = &queue->tasks[queue->maxpriority];
 112         list_for_each_entry(t, q, u.tk_wait.list) {
 113                 if (t->tk_owner == task->tk_owner) {
 114                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 115                         return;
 116                 }
 117         }
 118         list_add_tail(&task->u.tk_wait.list, q);
 119 }
 120
 121 /*
 122  * Add new request to wait queue.
 123  *
 124  * Swapper tasks always get inserted at the head of the queue.
 125  * This should avoid many nasty memory deadlocks and hopefully
 126  * improve overall performance.
 127  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 128  */
 129 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 130                 struct rpc_task *task,
 131                 unsigned char queue_priority)
 132 {
 133         BUG_ON (RPC_IS_QUEUED(task));
 134
 135         if (RPC_IS_PRIORITY(queue))
 136                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 137         else if (RPC_IS_SWAPPER(task))
 138                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 139         else
 140                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 141         task->tk_waitqueue = queue;
 142         queue->qlen++;
 143         rpc_set_queued(task);
 144
 145         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 146                         task->tk_pid, queue, rpc_qname(queue));
 147 }
 148
 149 /*
 150  * Remove request from a priority queue.
 151  */
 152 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 153 {
 154         struct rpc_task *t;
 155
 156         if (!list_empty(&task->u.tk_wait.links)) {
 157                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 158                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 159                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 160         }
 161 }
 162
 163 /*
 164  * Remove request from queue.
 165  * Note: must be called with spin lock held.
 166  */
 167 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 168 {
 169         __rpc_disable_timer(queue, task);
 170         if (RPC_IS_PRIORITY(queue))
 171                 __rpc_remove_wait_queue_priority(task);
 172         list_del(&task->u.tk_wait.list);
 173         queue->qlen--;
 174         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 175                         task->tk_pid, queue, rpc_qname(queue));
 176 }
 177
 178 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 179 {
 180         queue->priority = priority;
 181         queue->count = 1 << (priority * 2);
 182 }
 183
 184 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 185 {
 186         queue->owner = pid;
 187         queue->nr = RPC_BATCH_COUNT;
 188 }
 189
 190 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 191 {
 192         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 193         rpc_set_waitqueue_owner(queue, 0);
 194 }
 195
 196 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 197 {
 198         int i;
 199
 200         spin_lock_init(&queue->lock);
 201         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 202                 INIT_LIST_HEAD(&queue->tasks[i]);
 203         queue->maxpriority = nr_queues - 1;
 204         rpc_reset_waitqueue_priority(queue);
 205         queue->qlen = 0;
 206         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 207         INIT_LIST_HEAD(&queue->timer_list.list);
 208 #ifdef RPC_DEBUG
 209         queue->name = qname;
 210 #endif
 211 }
 212
 213 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 214 {
 215         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 216 }
 217 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 218
 219 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 220 {
 221         __rpc_init_priority_wait_queue(queue, qname, 1);
 222 }
 223 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 224
 225 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 226 {
 227         del_timer_sync(&queue->timer_list.timer);
 228 }
 229 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 230
 231 static int rpc_wait_bit_killable(void *word)
 232 {
 233         if (fatal_signal_pending(current))
 234                 return -ERESTARTSYS;
 235         freezable_schedule();
 236         return 0;
 237 }
 238
 239 #ifdef RPC_DEBUG
 240 static void rpc_task_set_debuginfo(struct rpc_task *task)
 241 {
 242         static atomic_t rpc_pid;
 243
 244         task->tk_pid = atomic_inc_return(&rpc_pid);
 245 }
 246 #else
 247 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 248 {
 249 }
 250 #endif
 251
 252 static void rpc_set_active(struct rpc_task *task)
 253 {
 254         rpc_task_set_debuginfo(task);
 255         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 256 }
 257
 258 /*
 259  * Mark an RPC call as having completed by clearing the 'active' bit
 260  * and then waking up all tasks that were sleeping.
 261  */
 262 static int rpc_complete_task(struct rpc_task *task)
 263 {
 264         void *m = &task->tk_runstate;
 265         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 266         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 267         unsigned long flags;
 268         int ret;
 269
 270         spin_lock_irqsave(&wq->lock, flags);
 271         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 272         ret = atomic_dec_and_test(&task->tk_count);
 273         if (waitqueue_active(wq))
 274                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 275         spin_unlock_irqrestore(&wq->lock, flags);
 276         return ret;
 277 }
 278
 279 /*
 280  * Allow callers to wait for completion of an RPC call
 281  *
 282  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 283  * to enforce taking of the wq->lock and hence avoid races with
 284  * rpc_complete_task().
 285  */
 286 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 287 {
 288         if (action == NULL)
 289                 action = rpc_wait_bit_killable;
 290         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 291                         action, TASK_KILLABLE);
 292 }
 293 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 294
 295 /*
 296  * Make an RPC task runnable.
 297  *
 298  * Note: If the task is ASYNC, this must be called with
 299  * the spinlock held to protect the wait queue operation.
 300  */
 301 static void rpc_make_runnable(struct rpc_task *task)
 302 {
 303         rpc_clear_queued(task);
 304         if (rpc_test_and_set_running(task))
 305                 return;
 306         if (RPC_IS_ASYNC(task)) {
 307                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 308                 queue_work(rpciod_workqueue, &task->u.tk_work);
 309         } else
 310                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 311 }
 312
 313 /*
 314  * Prepare for sleeping on a wait queue.
 315  * By always appending tasks to the list we ensure FIFO behavior.
 316  * NB: An RPC task will only receive interrupt-driven events as long
 317  * as it's on a wait queue.
 318  */
 319 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 320                 struct rpc_task *task,
 321                 rpc_action action,
 322                 unsigned char queue_priority)
 323 {
 324         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 325                         task->tk_pid, rpc_qname(q), jiffies);
 326
 327         __rpc_add_wait_queue(q, task, queue_priority);
 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         /* We shouldn't ever put an inactive task to sleep */
 338         BUG_ON(!RPC_IS_ACTIVATED(task));
 339
 340         /*
 341          * Protect the queue operations.
 342          */
 343         spin_lock_bh(&q->lock);
 344         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 345         spin_unlock_bh(&q->lock);
 346 }
 347 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 348
 349 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 350                 rpc_action action, int priority)
 351 {
 352         /* We shouldn't ever put an inactive task to sleep */
 353         BUG_ON(!RPC_IS_ACTIVATED(task));
 354
 355         /*
 356          * Protect the queue operations.
 357          */
 358         spin_lock_bh(&q->lock);
 359         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 360         spin_unlock_bh(&q->lock);
 361 }
 362
 363 /**
 364  * __rpc_do_wake_up_task - wake up a single rpc_task
 365  * @queue: wait queue
 366  * @task: task to be woken up
 367  *
 368  * Caller must hold queue->lock, and have cleared the task queued flag.
 369  */
 370 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 371 {
 372         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 373                         task->tk_pid, jiffies);
 374
 375         /* Has the task been executed yet? If not, we cannot wake it up! */
 376         if (!RPC_IS_ACTIVATED(task)) {
 377                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 378                 return;
 379         }
 380
 381         __rpc_remove_wait_queue(queue, task);
 382
 383         rpc_make_runnable(task);
 384
 385         dprintk("RPC:       __rpc_wake_up_task done\n");
 386 }
 387
 388 /*
 389  * Wake up a queued task while the queue lock is being held
 390  */
 391 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 392 {
 393         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 394                 __rpc_do_wake_up_task(queue, task);
 395 }
 396
 397 /*
 398  * Tests whether rpc queue is empty
 399  */
 400 int rpc_queue_empty(struct rpc_wait_queue *queue)
 401 {
 402         int res;
 403
 404         spin_lock_bh(&queue->lock);
 405         res = queue->qlen;
 406         spin_unlock_bh(&queue->lock);
 407         return res == 0;
 408 }
 409 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 410
 411 /*
 412  * Wake up a task on a specific queue
 413  */
 414 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 415 {
 416         spin_lock_bh(&queue->lock);
 417         rpc_wake_up_task_queue_locked(queue, task);
 418         spin_unlock_bh(&queue->lock);
 419 }
 420 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 421
 422 /*
 423  * Wake up the next task on a priority queue.
 424  */
 425 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 426 {
 427         struct list_head *q;
 428         struct rpc_task *task;
 429
 430         /*
 431          * Service a batch of tasks from a single owner.
 432          */
 433         q = &queue->tasks[queue->priority];
 434         if (!list_empty(q)) {
 435                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 436                 if (queue->owner == task->tk_owner) {
 437                         if (--queue->nr)
 438                                 goto out;
 439                         list_move_tail(&task->u.tk_wait.list, q);
 440                 }
 441                 /*
 442                  * Check if we need to switch queues.
 443                  */
 444                 if (--queue->count)
 445                         goto new_owner;
 446         }
 447
 448         /*
 449          * Service the next queue.
 450          */
 451         do {
 452                 if (q == &queue->tasks[0])
 453                         q = &queue->tasks[queue->maxpriority];
 454                 else
 455                         q = q - 1;
 456                 if (!list_empty(q)) {
 457                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 458                         goto new_queue;
 459                 }
 460         } while (q != &queue->tasks[queue->priority]);
 461
 462         rpc_reset_waitqueue_priority(queue);
 463         return NULL;
 464
 465 new_queue:
 466         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 467 new_owner:
 468         rpc_set_waitqueue_owner(queue, task->tk_owner);
 469 out:
 470         rpc_wake_up_task_queue_locked(queue, task);
 471         return task;
 472 }
 473
 474 /*
 475  * Wake up the next task on the wait queue.
 476  */
 477 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 478 {
 479         struct rpc_task *task = NULL;
 480
 481         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 482                         queue, rpc_qname(queue));
 483         spin_lock_bh(&queue->lock);
 484         if (RPC_IS_PRIORITY(queue))
 485                 task = __rpc_wake_up_next_priority(queue);
 486         else {
 487                 task_for_first(task, &queue->tasks[0])
 488                         rpc_wake_up_task_queue_locked(queue, task);
 489         }
 490         spin_unlock_bh(&queue->lock);
 491
 492         return task;
 493 }
 494 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 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_queue_locked(queue, task);
 512                 if (head == &queue->tasks[0])
 513                         break;
 514                 head--;
 515         }
 516         spin_unlock_bh(&queue->lock);
 517 }
 518 EXPORT_SYMBOL_GPL(rpc_wake_up);
 519
 520 /**
 521  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 522  * @queue: rpc_wait_queue on which the tasks are sleeping
 523  * @status: status value to set
 524  *
 525  * Grabs queue->lock
 526  */
 527 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 528 {
 529         struct rpc_task *task, *next;
 530         struct list_head *head;
 531
 532         spin_lock_bh(&queue->lock);
 533         head = &queue->tasks[queue->maxpriority];
 534         for (;;) {
 535                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 536                         task->tk_status = status;
 537                         rpc_wake_up_task_queue_locked(queue, task);
 538                 }
 539                 if (head == &queue->tasks[0])
 540                         break;
 541                 head--;
 542         }
 543         spin_unlock_bh(&queue->lock);
 544 }
 545 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 546
 547 /**
 548  * rpc_wake_up_softconn_status - wake up all SOFTCONN rpc_tasks and set their
 549  * status value.
 550  * @queue: rpc_wait_queue on which the tasks are sleeping
 551  * @status: status value to set
 552  *
 553  * Grabs queue->lock
 554  */
 555 void rpc_wake_up_softconn_status(struct rpc_wait_queue *queue, int status)
 556 {
 557         struct rpc_task *task, *next;
 558         struct list_head *head;
 559
 560         spin_lock_bh(&queue->lock);
 561         head = &queue->tasks[queue->maxpriority];
 562         for (;;) {
 563                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 564                         if (RPC_IS_SOFTCONN(task)) {
 565                                 task->tk_status = status;
 566                                 rpc_wake_up_task_queue_locked(queue, task);
 567                         }
 568                 if (head == &queue->tasks[0])
 569                         break;
 570                 head--;
 571         }
 572         spin_unlock_bh(&queue->lock);
 573 }
 574 EXPORT_SYMBOL_GPL(rpc_wake_up_softconn_status);
 575
 576 static void __rpc_queue_timer_fn(unsigned long ptr)
 577 {
 578         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 579         struct rpc_task *task, *n;
 580         unsigned long expires, now, timeo;
 581
 582         spin_lock(&queue->lock);
 583         expires = now = jiffies;
 584         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 585                 timeo = task->u.tk_wait.expires;
 586                 if (time_after_eq(now, timeo)) {
 587                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 588                         task->tk_status = -ETIMEDOUT;
 589                         rpc_wake_up_task_queue_locked(queue, task);
 590                         continue;
 591                 }
 592                 if (expires == now || time_after(expires, timeo))
 593                         expires = timeo;
 594         }
 595         if (!list_empty(&queue->timer_list.list))
 596                 rpc_set_queue_timer(queue, expires);
 597         spin_unlock(&queue->lock);
 598 }
 599
 600 static void __rpc_atrun(struct rpc_task *task)
 601 {
 602         task->tk_status = 0;
 603 }
 604
 605 /*
 606  * Run a task at a later time
 607  */
 608 void rpc_delay(struct rpc_task *task, unsigned long delay)
 609 {
 610         task->tk_timeout = delay;
 611         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 612 }
 613 EXPORT_SYMBOL_GPL(rpc_delay);
 614
 615 /*
 616  * Helper to call task->tk_ops->rpc_call_prepare
 617  */
 618 void rpc_prepare_task(struct rpc_task *task)
 619 {
 620         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 621 }
 622
 623 static void
 624 rpc_init_task_statistics(struct rpc_task *task)
 625 {
 626         /* Initialize retry counters */
 627         task->tk_garb_retry = 2;
 628         task->tk_cred_retry = 2;
 629         task->tk_rebind_retry = 2;
 630
 631         /* starting timestamp */
 632         task->tk_start = ktime_get();
 633 }
 634
 635 static void
 636 rpc_reset_task_statistics(struct rpc_task *task)
 637 {
 638         task->tk_timeouts = 0;
 639         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 640
 641         rpc_init_task_statistics(task);
 642 }
 643
 644 /*
 645  * Helper that calls task->tk_ops->rpc_call_done if it exists
 646  */
 647 void rpc_exit_task(struct rpc_task *task)
 648 {
 649         task->tk_action = NULL;
 650         if (task->tk_ops->rpc_call_done != NULL) {
 651                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 652                 if (task->tk_action != NULL) {
 653                         WARN_ON(RPC_ASSASSINATED(task));
 654                         /* Always release the RPC slot and buffer memory */
 655                         xprt_release(task);
 656                         rpc_reset_task_statistics(task);
 657                 }
 658         }
 659 }
 660
 661 void rpc_exit(struct rpc_task *task, int status)
 662 {
 663         task->tk_status = status;
 664         task->tk_action = rpc_exit_task;
 665         if (RPC_IS_QUEUED(task))
 666                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 667 }
 668 EXPORT_SYMBOL_GPL(rpc_exit);
 669
 670 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 671 {
 672         if (ops->rpc_release != NULL)
 673                 ops->rpc_release(calldata);
 674 }
 675
 676 /*
 677  * This is the RPC `scheduler' (or rather, the finite state machine).
 678  */
 679 static void __rpc_execute(struct rpc_task *task)
 680 {
 681         struct rpc_wait_queue *queue;
 682         int task_is_async = RPC_IS_ASYNC(task);
 683         int status = 0;
 684
 685         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 686                         task->tk_pid, task->tk_flags);
 687
 688         BUG_ON(RPC_IS_QUEUED(task));
 689
 690         for (;;) {
 691                 void (*do_action)(struct rpc_task *);
 692
 693                 /*
 694                  * Execute any pending callback first.
 695                  */
 696                 do_action = task->tk_callback;
 697                 task->tk_callback = NULL;
 698                 if (do_action == NULL) {
 699                         /*
 700                          * Perform the next FSM step.
 701                          * tk_action may be NULL if the task has been killed.
 702                          * In particular, note that rpc_killall_tasks may
 703                          * do this at any time, so beware when dereferencing.
 704                          */
 705                         do_action = task->tk_action;
 706                         if (do_action == NULL)
 707                                 break;
 708                 }
 709                 do_action(task);
 710
 711                 /*
 712                  * Lockless check for whether task is sleeping or not.
 713                  */
 714                 if (!RPC_IS_QUEUED(task))
 715                         continue;
 716                 /*
 717                  * The queue->lock protects against races with
 718                  * rpc_make_runnable().
 719                  *
 720                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 721                  * rpc_task, rpc_make_runnable() can assign it to a
 722                  * different workqueue. We therefore cannot assume that the
 723                  * rpc_task pointer may still be dereferenced.
 724                  */
 725                 queue = task->tk_waitqueue;
 726                 spin_lock_bh(&queue->lock);
 727                 if (!RPC_IS_QUEUED(task)) {
 728                         spin_unlock_bh(&queue->lock);
 729                         continue;
 730                 }
 731                 rpc_clear_running(task);
 732                 spin_unlock_bh(&queue->lock);
 733                 if (task_is_async)
 734                         return;
 735
 736                 /* sync task: sleep here */
 737                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 738                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 739                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 740                                 TASK_KILLABLE);
 741                 if (status == -ERESTARTSYS) {
 742                         /*
 743                          * When a sync task receives a signal, it exits with
 744                          * -ERESTARTSYS. In order to catch any callbacks that
 745                          * clean up after sleeping on some queue, we don't
 746                          * break the loop here, but go around once more.
 747                          */
 748                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 749                         task->tk_flags |= RPC_TASK_KILLED;
 750                         rpc_exit(task, -ERESTARTSYS);
 751                 }
 752                 rpc_set_running(task);
 753                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 754         }
 755
 756         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 757                         task->tk_status);
 758         /* Release all resources associated with the task */
 759         rpc_release_task(task);
 760 }
 761
 762 /*
 763  * User-visible entry point to the scheduler.
 764  *
 765  * This may be called recursively if e.g. an async NFS task updates
 766  * the attributes and finds that dirty pages must be flushed.
 767  * NOTE: Upon exit of this function the task is guaranteed to be
 768  *       released. In particular note that tk_release() will have
 769  *       been called, so your task memory may have been freed.
 770  */
 771 void rpc_execute(struct rpc_task *task)
 772 {
 773         rpc_set_active(task);
 774         rpc_make_runnable(task);
 775         if (!RPC_IS_ASYNC(task))
 776                 __rpc_execute(task);
 777 }
 778
 779 static void rpc_async_schedule(struct work_struct *work)
 780 {
 781         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 782 }
 783
 784 /**
 785  * rpc_malloc - allocate an RPC buffer
 786  * @task: RPC task that will use this buffer
 787  * @size: requested byte size
 788  *
 789  * To prevent rpciod from hanging, this allocator never sleeps,
 790  * returning NULL if the request cannot be serviced immediately.
 791  * The caller can arrange to sleep in a way that is safe for rpciod.
 792  *
 793  * Most requests are 'small' (under 2KiB) and can be serviced from a
 794  * mempool, ensuring that NFS reads and writes can always proceed,
 795  * and that there is good locality of reference for these buffers.
 796  *
 797  * In order to avoid memory starvation triggering more writebacks of
 798  * NFS requests, we avoid using GFP_KERNEL.
 799  */
 800 void *rpc_malloc(struct rpc_task *task, size_t size)
 801 {
 802         struct rpc_buffer *buf;
 803         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 804
 805         size += sizeof(struct rpc_buffer);
 806         if (size <= RPC_BUFFER_MAXSIZE)
 807                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 808         else
 809                 buf = kmalloc(size, gfp);
 810
 811         if (!buf)
 812                 return NULL;
 813
 814         buf->len = size;
 815         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 816                         task->tk_pid, size, buf);
 817         return &buf->data;
 818 }
 819 EXPORT_SYMBOL_GPL(rpc_malloc);
 820
 821 /**
 822  * rpc_free - free buffer allocated via rpc_malloc
 823  * @buffer: buffer to free
 824  *
 825  */
 826 void rpc_free(void *buffer)
 827 {
 828         size_t size;
 829         struct rpc_buffer *buf;
 830
 831         if (!buffer)
 832                 return;
 833
 834         buf = container_of(buffer, struct rpc_buffer, data);
 835         size = buf->len;
 836
 837         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 838                         size, buf);
 839
 840         if (size <= RPC_BUFFER_MAXSIZE)
 841                 mempool_free(buf, rpc_buffer_mempool);
 842         else
 843                 kfree(buf);
 844 }
 845 EXPORT_SYMBOL_GPL(rpc_free);
 846
 847 /*
 848  * Creation and deletion of RPC task structures
 849  */
 850 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 851 {
 852         memset(task, 0, sizeof(*task));
 853         atomic_set(&task->tk_count, 1);
 854         task->tk_flags  = task_setup_data->flags;
 855         task->tk_ops = task_setup_data->callback_ops;
 856         task->tk_calldata = task_setup_data->callback_data;
 857         INIT_LIST_HEAD(&task->tk_task);
 858
 859         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 860         task->tk_owner = current->tgid;
 861
 862         /* Initialize workqueue for async tasks */
 863         task->tk_workqueue = task_setup_data->workqueue;
 864
 865         if (task->tk_ops->rpc_call_prepare != NULL)
 866                 task->tk_action = rpc_prepare_task;
 867
 868         rpc_init_task_statistics(task);
 869
 870         dprintk("RPC:       new task initialized, procpid %u\n",
 871                                 task_pid_nr(current));
 872 }
 873
 874 static struct rpc_task *
 875 rpc_alloc_task(void)
 876 {
 877         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 878 }
 879
 880 /*
 881  * Create a new task for the specified client.
 882  */
 883 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 884 {
 885         struct rpc_task *task = setup_data->task;
 886         unsigned short flags = 0;
 887
 888         if (task == NULL) {
 889                 task = rpc_alloc_task();
 890                 if (task == NULL) {
 891                         rpc_release_calldata(setup_data->callback_ops,
 892                                         setup_data->callback_data);
 893                         return ERR_PTR(-ENOMEM);
 894                 }
 895                 flags = RPC_TASK_DYNAMIC;
 896         }
 897
 898         rpc_init_task(task, setup_data);
 899         task->tk_flags |= flags;
 900         dprintk("RPC:       allocated task %p\n", task);
 901         return task;
 902 }
 903
 904 static void rpc_free_task(struct rpc_task *task)
 905 {
 906         const struct rpc_call_ops *tk_ops = task->tk_ops;
 907         void *calldata = task->tk_calldata;
 908
 909         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 910                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 911                 mempool_free(task, rpc_task_mempool);
 912         }
 913         rpc_release_calldata(tk_ops, calldata);
 914 }
 915
 916 static void rpc_async_release(struct work_struct *work)
 917 {
 918         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 919 }
 920
 921 static void rpc_release_resources_task(struct rpc_task *task)
 922 {
 923         if (task->tk_rqstp)
 924                 xprt_release(task);
 925         if (task->tk_msg.rpc_cred) {
 926                 put_rpccred(task->tk_msg.rpc_cred);
 927                 task->tk_msg.rpc_cred = NULL;
 928         }
 929         rpc_task_release_client(task);
 930 }
 931
 932 static void rpc_final_put_task(struct rpc_task *task,
 933                 struct workqueue_struct *q)
 934 {
 935         if (q != NULL) {
 936                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 937                 queue_work(q, &task->u.tk_work);
 938         } else
 939                 rpc_free_task(task);
 940 }
 941
 942 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
 943 {
 944         if (atomic_dec_and_test(&task->tk_count)) {
 945                 rpc_release_resources_task(task);
 946                 rpc_final_put_task(task, q);
 947         }
 948 }
 949
 950 void rpc_put_task(struct rpc_task *task)
 951 {
 952         rpc_do_put_task(task, NULL);
 953 }
 954 EXPORT_SYMBOL_GPL(rpc_put_task);
 955
 956 void rpc_put_task_async(struct rpc_task *task)
 957 {
 958         rpc_do_put_task(task, task->tk_workqueue);
 959 }
 960 EXPORT_SYMBOL_GPL(rpc_put_task_async);
 961
 962 static void rpc_release_task(struct rpc_task *task)
 963 {
 964         dprintk("RPC: %5u release task\n", task->tk_pid);
 965
 966         BUG_ON (RPC_IS_QUEUED(task));
 967
 968         rpc_release_resources_task(task);
 969
 970         /*
 971          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
 972          * so it should be safe to use task->tk_count as a test for whether
 973          * or not any other processes still hold references to our rpc_task.
 974          */
 975         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
 976                 /* Wake up anyone who may be waiting for task completion */
 977                 if (!rpc_complete_task(task))
 978                         return;
 979         } else {
 980                 if (!atomic_dec_and_test(&task->tk_count))
 981                         return;
 982         }
 983         rpc_final_put_task(task, task->tk_workqueue);
 984 }
 985
 986 int rpciod_up(void)
 987 {
 988         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 989 }
 990
 991 void rpciod_down(void)
 992 {
 993         module_put(THIS_MODULE);
 994 }
 995
 996 /*
 997  * Start up the rpciod workqueue.
 998  */
 999 static int rpciod_start(void)
1000 {
1001         struct workqueue_struct *wq;
1002
1003         /*
1004          * Create the rpciod thread and wait for it to start.
1005          */
1006         dprintk("RPC:       creating workqueue rpciod\n");
1007         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1008         rpciod_workqueue = wq;
1009         return rpciod_workqueue != NULL;
1010 }
1011
1012 static void rpciod_stop(void)
1013 {
1014         struct workqueue_struct *wq = NULL;
1015
1016         if (rpciod_workqueue == NULL)
1017                 return;
1018         dprintk("RPC:       destroying workqueue rpciod\n");
1019
1020         wq = rpciod_workqueue;
1021         rpciod_workqueue = NULL;
1022         destroy_workqueue(wq);
1023 }
1024
1025 void
1026 rpc_destroy_mempool(void)
1027 {
1028         rpciod_stop();
1029         if (rpc_buffer_mempool)
1030                 mempool_destroy(rpc_buffer_mempool);
1031         if (rpc_task_mempool)
1032                 mempool_destroy(rpc_task_mempool);
1033         if (rpc_task_slabp)
1034                 kmem_cache_destroy(rpc_task_slabp);
1035         if (rpc_buffer_slabp)
1036                 kmem_cache_destroy(rpc_buffer_slabp);
1037         rpc_destroy_wait_queue(&delay_queue);
1038 }
1039
1040 int
1041 rpc_init_mempool(void)
1042 {
1043         /*
1044          * The following is not strictly a mempool initialisation,
1045          * but there is no harm in doing it here
1046          */
1047         rpc_init_wait_queue(&delay_queue, "delayq");
1048         if (!rpciod_start())
1049                 goto err_nomem;
1050
1051         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1052                                              sizeof(struct rpc_task),
1053                                              0, SLAB_HWCACHE_ALIGN,
1054                                              NULL);
1055         if (!rpc_task_slabp)
1056                 goto err_nomem;
1057         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1058                                              RPC_BUFFER_MAXSIZE,
1059                                              0, SLAB_HWCACHE_ALIGN,
1060                                              NULL);
1061         if (!rpc_buffer_slabp)
1062                 goto err_nomem;
1063         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1064                                                     rpc_task_slabp);
1065         if (!rpc_task_mempool)
1066                 goto err_nomem;
1067         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1068                                                       rpc_buffer_slabp);
1069         if (!rpc_buffer_mempool)
1070                 goto err_nomem;
1071         return 0;
1072 err_nomem:
1073         rpc_destroy_mempool();
1074         return -ENOMEM;
1075 }