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         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 143         smp_wmb();
 144         rpc_set_queued(task);
 145
 146         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 147                         task->tk_pid, queue, rpc_qname(queue));
 148 }
 149
 150 /*
 151  * Remove request from a priority queue.
 152  */
 153 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 154 {
 155         struct rpc_task *t;
 156
 157         if (!list_empty(&task->u.tk_wait.links)) {
 158                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 159                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 160                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 161         }
 162 }
 163
 164 /*
 165  * Remove request from queue.
 166  * Note: must be called with spin lock held.
 167  */
 168 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 169 {
 170         __rpc_disable_timer(queue, task);
 171         if (RPC_IS_PRIORITY(queue))
 172                 __rpc_remove_wait_queue_priority(task);
 173         list_del(&task->u.tk_wait.list);
 174         queue->qlen--;
 175         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 176                         task->tk_pid, queue, rpc_qname(queue));
 177 }
 178
 179 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 180 {
 181         queue->priority = priority;
 182         queue->count = 1 << (priority * 2);
 183 }
 184
 185 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 186 {
 187         queue->owner = pid;
 188         queue->nr = RPC_BATCH_COUNT;
 189 }
 190
 191 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 192 {
 193         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 194         rpc_set_waitqueue_owner(queue, 0);
 195 }
 196
 197 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 198 {
 199         int i;
 200
 201         spin_lock_init(&queue->lock);
 202         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 203                 INIT_LIST_HEAD(&queue->tasks[i]);
 204         queue->maxpriority = nr_queues - 1;
 205         rpc_reset_waitqueue_priority(queue);
 206         queue->qlen = 0;
 207         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 208         INIT_LIST_HEAD(&queue->timer_list.list);
 209 #ifdef RPC_DEBUG
 210         queue->name = qname;
 211 #endif
 212 }
 213
 214 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 215 {
 216         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 217 }
 218 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 219
 220 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 221 {
 222         __rpc_init_priority_wait_queue(queue, qname, 1);
 223 }
 224 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 225
 226 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 227 {
 228         del_timer_sync(&queue->timer_list.timer);
 229 }
 230 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 231
 232 static int rpc_wait_bit_killable(void *word)
 233 {
 234         if (fatal_signal_pending(current))
 235                 return -ERESTARTSYS;
 236         schedule();
 237         return 0;
 238 }
 239
 240 #ifdef RPC_DEBUG
 241 static void rpc_task_set_debuginfo(struct rpc_task *task)
 242 {
 243         static atomic_t rpc_pid;
 244
 245         task->tk_pid = atomic_inc_return(&rpc_pid);
 246 }
 247 #else
 248 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 249 {
 250 }
 251 #endif
 252
 253 static void rpc_set_active(struct rpc_task *task)
 254 {
 255         rpc_task_set_debuginfo(task);
 256         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 257 }
 258
 259 /*
 260  * Mark an RPC call as having completed by clearing the 'active' bit
 261  * and then waking up all tasks that were sleeping.
 262  */
 263 static int rpc_complete_task(struct rpc_task *task)
 264 {
 265         void *m = &task->tk_runstate;
 266         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 267         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 268         unsigned long flags;
 269         int ret;
 270
 271         spin_lock_irqsave(&wq->lock, flags);
 272         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 273         ret = atomic_dec_and_test(&task->tk_count);
 274         if (waitqueue_active(wq))
 275                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 276         spin_unlock_irqrestore(&wq->lock, flags);
 277         return ret;
 278 }
 279
 280 /*
 281  * Allow callers to wait for completion of an RPC call
 282  *
 283  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 284  * to enforce taking of the wq->lock and hence avoid races with
 285  * rpc_complete_task().
 286  */
 287 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 288 {
 289         if (action == NULL)
 290                 action = rpc_wait_bit_killable;
 291         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 292                         action, TASK_KILLABLE);
 293 }
 294 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 295
 296 /*
 297  * Make an RPC task runnable.
 298  *
 299  * Note: If the task is ASYNC, this must be called with
 300  * the spinlock held to protect the wait queue operation.
 301  */
 302 static void rpc_make_runnable(struct rpc_task *task)
 303 {
 304         rpc_clear_queued(task);
 305         if (rpc_test_and_set_running(task))
 306                 return;
 307         if (RPC_IS_ASYNC(task)) {
 308                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 309                 queue_work(rpciod_workqueue, &task->u.tk_work);
 310         } else
 311                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 312 }
 313
 314 /*
 315  * Prepare for sleeping on a wait queue.
 316  * By always appending tasks to the list we ensure FIFO behavior.
 317  * NB: An RPC task will only receive interrupt-driven events as long
 318  * as it's on a wait queue.
 319  */
 320 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 321                 struct rpc_task *task,
 322                 rpc_action action,
 323                 unsigned char queue_priority)
 324 {
 325         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 326                         task->tk_pid, rpc_qname(q), jiffies);
 327
 328         __rpc_add_wait_queue(q, task, queue_priority);
 329
 330         BUG_ON(task->tk_callback != NULL);
 331         task->tk_callback = action;
 332         __rpc_add_timer(q, task);
 333 }
 334
 335 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 336                                 rpc_action action)
 337 {
 338         /* We shouldn't ever put an inactive task to sleep */
 339         BUG_ON(!RPC_IS_ACTIVATED(task));
 340
 341         /*
 342          * Protect the queue operations.
 343          */
 344         spin_lock_bh(&q->lock);
 345         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 346         spin_unlock_bh(&q->lock);
 347 }
 348 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 349
 350 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 351                 rpc_action action, int priority)
 352 {
 353         /* We shouldn't ever put an inactive task to sleep */
 354         BUG_ON(!RPC_IS_ACTIVATED(task));
 355
 356         /*
 357          * Protect the queue operations.
 358          */
 359         spin_lock_bh(&q->lock);
 360         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 361         spin_unlock_bh(&q->lock);
 362 }
 363
 364 /**
 365  * __rpc_do_wake_up_task - wake up a single rpc_task
 366  * @queue: wait queue
 367  * @task: task to be woken up
 368  *
 369  * Caller must hold queue->lock, and have cleared the task queued flag.
 370  */
 371 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 372 {
 373         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 374                         task->tk_pid, jiffies);
 375
 376         /* Has the task been executed yet? If not, we cannot wake it up! */
 377         if (!RPC_IS_ACTIVATED(task)) {
 378                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 379                 return;
 380         }
 381
 382         __rpc_remove_wait_queue(queue, task);
 383
 384         rpc_make_runnable(task);
 385
 386         dprintk("RPC:       __rpc_wake_up_task done\n");
 387 }
 388
 389 /*
 390  * Wake up a queued task while the queue lock is being held
 391  */
 392 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 393 {
 394         if (RPC_IS_QUEUED(task)) {
 395                 smp_rmb();
 396                 if (task->tk_waitqueue == queue)
 397                         __rpc_do_wake_up_task(queue, task);
 398         }
 399 }
 400
 401 /*
 402  * Tests whether rpc queue is empty
 403  */
 404 int rpc_queue_empty(struct rpc_wait_queue *queue)
 405 {
 406         int res;
 407
 408         spin_lock_bh(&queue->lock);
 409         res = queue->qlen;
 410         spin_unlock_bh(&queue->lock);
 411         return res == 0;
 412 }
 413 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 414
 415 /*
 416  * Wake up a task on a specific queue
 417  */
 418 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 419 {
 420         spin_lock_bh(&queue->lock);
 421         rpc_wake_up_task_queue_locked(queue, task);
 422         spin_unlock_bh(&queue->lock);
 423 }
 424 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 425
 426 /*
 427  * Wake up the next task on a priority queue.
 428  */
 429 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 430 {
 431         struct list_head *q;
 432         struct rpc_task *task;
 433
 434         /*
 435          * Service a batch of tasks from a single owner.
 436          */
 437         q = &queue->tasks[queue->priority];
 438         if (!list_empty(q)) {
 439                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 440                 if (queue->owner == task->tk_owner) {
 441                         if (--queue->nr)
 442                                 goto out;
 443                         list_move_tail(&task->u.tk_wait.list, q);
 444                 }
 445                 /*
 446                  * Check if we need to switch queues.
 447                  */
 448                 if (--queue->count)
 449                         goto new_owner;
 450         }
 451
 452         /*
 453          * Service the next queue.
 454          */
 455         do {
 456                 if (q == &queue->tasks[0])
 457                         q = &queue->tasks[queue->maxpriority];
 458                 else
 459                         q = q - 1;
 460                 if (!list_empty(q)) {
 461                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 462                         goto new_queue;
 463                 }
 464         } while (q != &queue->tasks[queue->priority]);
 465
 466         rpc_reset_waitqueue_priority(queue);
 467         return NULL;
 468
 469 new_queue:
 470         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 471 new_owner:
 472         rpc_set_waitqueue_owner(queue, task->tk_owner);
 473 out:
 474         rpc_wake_up_task_queue_locked(queue, task);
 475         return task;
 476 }
 477
 478 /*
 479  * Wake up the next task on the wait queue.
 480  */
 481 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 482 {
 483         struct rpc_task *task = NULL;
 484
 485         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 486                         queue, rpc_qname(queue));
 487         spin_lock_bh(&queue->lock);
 488         if (RPC_IS_PRIORITY(queue))
 489                 task = __rpc_wake_up_next_priority(queue);
 490         else {
 491                 task_for_first(task, &queue->tasks[0])
 492                         rpc_wake_up_task_queue_locked(queue, task);
 493         }
 494         spin_unlock_bh(&queue->lock);
 495
 496         return task;
 497 }
 498 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 499
 500 /**
 501  * rpc_wake_up - wake up all rpc_tasks
 502  * @queue: rpc_wait_queue on which the tasks are sleeping
 503  *
 504  * Grabs queue->lock
 505  */
 506 void rpc_wake_up(struct rpc_wait_queue *queue)
 507 {
 508         struct list_head *head;
 509
 510         spin_lock_bh(&queue->lock);
 511         head = &queue->tasks[queue->maxpriority];
 512         for (;;) {
 513                 while (!list_empty(head)) {
 514                         struct rpc_task *task;
 515                         task = list_first_entry(head,
 516                                         struct rpc_task,
 517                                         u.tk_wait.list);
 518                         rpc_wake_up_task_queue_locked(queue, task);
 519                 }
 520                 if (head == &queue->tasks[0])
 521                         break;
 522                 head--;
 523         }
 524         spin_unlock_bh(&queue->lock);
 525 }
 526 EXPORT_SYMBOL_GPL(rpc_wake_up);
 527
 528 /**
 529  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 530  * @queue: rpc_wait_queue on which the tasks are sleeping
 531  * @status: status value to set
 532  *
 533  * Grabs queue->lock
 534  */
 535 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 536 {
 537         struct list_head *head;
 538
 539         spin_lock_bh(&queue->lock);
 540         head = &queue->tasks[queue->maxpriority];
 541         for (;;) {
 542                 while (!list_empty(head)) {
 543                         struct rpc_task *task;
 544                         task = list_first_entry(head,
 545                                         struct rpc_task,
 546                                         u.tk_wait.list);
 547                         task->tk_status = status;
 548                         rpc_wake_up_task_queue_locked(queue, task);
 549                 }
 550                 if (head == &queue->tasks[0])
 551                         break;
 552                 head--;
 553         }
 554         spin_unlock_bh(&queue->lock);
 555 }
 556 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 557
 558 static void __rpc_queue_timer_fn(unsigned long ptr)
 559 {
 560         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 561         struct rpc_task *task, *n;
 562         unsigned long expires, now, timeo;
 563
 564         spin_lock(&queue->lock);
 565         expires = now = jiffies;
 566         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 567                 timeo = task->u.tk_wait.expires;
 568                 if (time_after_eq(now, timeo)) {
 569                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 570                         task->tk_status = -ETIMEDOUT;
 571                         rpc_wake_up_task_queue_locked(queue, task);
 572                         continue;
 573                 }
 574                 if (expires == now || time_after(expires, timeo))
 575                         expires = timeo;
 576         }
 577         if (!list_empty(&queue->timer_list.list))
 578                 rpc_set_queue_timer(queue, expires);
 579         spin_unlock(&queue->lock);
 580 }
 581
 582 static void __rpc_atrun(struct rpc_task *task)
 583 {
 584         task->tk_status = 0;
 585 }
 586
 587 /*
 588  * Run a task at a later time
 589  */
 590 void rpc_delay(struct rpc_task *task, unsigned long delay)
 591 {
 592         task->tk_timeout = delay;
 593         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 594 }
 595 EXPORT_SYMBOL_GPL(rpc_delay);
 596
 597 /*
 598  * Helper to call task->tk_ops->rpc_call_prepare
 599  */
 600 void rpc_prepare_task(struct rpc_task *task)
 601 {
 602         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 603 }
 604
 605 static void
 606 rpc_init_task_statistics(struct rpc_task *task)
 607 {
 608         /* Initialize retry counters */
 609         task->tk_garb_retry = 2;
 610         task->tk_cred_retry = 2;
 611         task->tk_rebind_retry = 2;
 612
 613         /* starting timestamp */
 614         task->tk_start = ktime_get();
 615 }
 616
 617 static void
 618 rpc_reset_task_statistics(struct rpc_task *task)
 619 {
 620         task->tk_timeouts = 0;
 621         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 622
 623         rpc_init_task_statistics(task);
 624 }
 625
 626 /*
 627  * Helper that calls task->tk_ops->rpc_call_done if it exists
 628  */
 629 void rpc_exit_task(struct rpc_task *task)
 630 {
 631         task->tk_action = NULL;
 632         if (task->tk_ops->rpc_call_done != NULL) {
 633                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 634                 if (task->tk_action != NULL) {
 635                         WARN_ON(RPC_ASSASSINATED(task));
 636                         /* Always release the RPC slot and buffer memory */
 637                         xprt_release(task);
 638                         rpc_reset_task_statistics(task);
 639                 }
 640         }
 641 }
 642
 643 void rpc_exit(struct rpc_task *task, int status)
 644 {
 645         task->tk_status = status;
 646         task->tk_action = rpc_exit_task;
 647         if (RPC_IS_QUEUED(task))
 648                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 649 }
 650 EXPORT_SYMBOL_GPL(rpc_exit);
 651
 652 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 653 {
 654         if (ops->rpc_release != NULL)
 655                 ops->rpc_release(calldata);
 656 }
 657
 658 /*
 659  * This is the RPC `scheduler' (or rather, the finite state machine).
 660  */
 661 static void __rpc_execute(struct rpc_task *task)
 662 {
 663         struct rpc_wait_queue *queue;
 664         int task_is_async = RPC_IS_ASYNC(task);
 665         int status = 0;
 666
 667         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 668                         task->tk_pid, task->tk_flags);
 669
 670         BUG_ON(RPC_IS_QUEUED(task));
 671
 672         for (;;) {
 673                 void (*do_action)(struct rpc_task *);
 674
 675                 /*
 676                  * Execute any pending callback first.
 677                  */
 678                 do_action = task->tk_callback;
 679                 task->tk_callback = NULL;
 680                 if (do_action == NULL) {
 681                         /*
 682                          * Perform the next FSM step.
 683                          * tk_action may be NULL if the task has been killed.
 684                          * In particular, note that rpc_killall_tasks may
 685                          * do this at any time, so beware when dereferencing.
 686                          */
 687                         do_action = task->tk_action;
 688                         if (do_action == NULL)
 689                                 break;
 690                 }
 691                 do_action(task);
 692
 693                 /*
 694                  * Lockless check for whether task is sleeping or not.
 695                  */
 696                 if (!RPC_IS_QUEUED(task))
 697                         continue;
 698                 /*
 699                  * The queue->lock protects against races with
 700                  * rpc_make_runnable().
 701                  *
 702                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 703                  * rpc_task, rpc_make_runnable() can assign it to a
 704                  * different workqueue. We therefore cannot assume that the
 705                  * rpc_task pointer may still be dereferenced.
 706                  */
 707                 queue = task->tk_waitqueue;
 708                 spin_lock_bh(&queue->lock);
 709                 if (!RPC_IS_QUEUED(task)) {
 710                         spin_unlock_bh(&queue->lock);
 711                         continue;
 712                 }
 713                 rpc_clear_running(task);
 714                 spin_unlock_bh(&queue->lock);
 715                 if (task_is_async)
 716                         return;
 717
 718                 /* sync task: sleep here */
 719                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 720                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 721                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 722                                 TASK_KILLABLE);
 723                 if (status == -ERESTARTSYS) {
 724                         /*
 725                          * When a sync task receives a signal, it exits with
 726                          * -ERESTARTSYS. In order to catch any callbacks that
 727                          * clean up after sleeping on some queue, we don't
 728                          * break the loop here, but go around once more.
 729                          */
 730                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 731                         task->tk_flags |= RPC_TASK_KILLED;
 732                         rpc_exit(task, -ERESTARTSYS);
 733                 }
 734                 rpc_set_running(task);
 735                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 736         }
 737
 738         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 739                         task->tk_status);
 740         /* Release all resources associated with the task */
 741         rpc_release_task(task);
 742 }
 743
 744 /*
 745  * User-visible entry point to the scheduler.
 746  *
 747  * This may be called recursively if e.g. an async NFS task updates
 748  * the attributes and finds that dirty pages must be flushed.
 749  * NOTE: Upon exit of this function the task is guaranteed to be
 750  *       released. In particular note that tk_release() will have
 751  *       been called, so your task memory may have been freed.
 752  */
 753 void rpc_execute(struct rpc_task *task)
 754 {
 755         rpc_set_active(task);
 756         rpc_make_runnable(task);
 757         if (!RPC_IS_ASYNC(task))
 758                 __rpc_execute(task);
 759 }
 760
 761 static void rpc_async_schedule(struct work_struct *work)
 762 {
 763         current->flags |= PF_FSTRANS;
 764         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 765         current->flags &= ~PF_FSTRANS;
 766 }
 767
 768 /**
 769  * rpc_malloc - allocate an RPC buffer
 770  * @task: RPC task that will use this buffer
 771  * @size: requested byte size
 772  *
 773  * To prevent rpciod from hanging, this allocator never sleeps,
 774  * returning NULL if the request cannot be serviced immediately.
 775  * The caller can arrange to sleep in a way that is safe for rpciod.
 776  *
 777  * Most requests are 'small' (under 2KiB) and can be serviced from a
 778  * mempool, ensuring that NFS reads and writes can always proceed,
 779  * and that there is good locality of reference for these buffers.
 780  *
 781  * In order to avoid memory starvation triggering more writebacks of
 782  * NFS requests, we avoid using GFP_KERNEL.
 783  */
 784 void *rpc_malloc(struct rpc_task *task, size_t size)
 785 {
 786         struct rpc_buffer *buf;
 787         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 788
 789         size += sizeof(struct rpc_buffer);
 790         if (size <= RPC_BUFFER_MAXSIZE)
 791                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 792         else
 793                 buf = kmalloc(size, gfp);
 794
 795         if (!buf)
 796                 return NULL;
 797
 798         buf->len = size;
 799         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 800                         task->tk_pid, size, buf);
 801         return &buf->data;
 802 }
 803 EXPORT_SYMBOL_GPL(rpc_malloc);
 804
 805 /**
 806  * rpc_free - free buffer allocated via rpc_malloc
 807  * @buffer: buffer to free
 808  *
 809  */
 810 void rpc_free(void *buffer)
 811 {
 812         size_t size;
 813         struct rpc_buffer *buf;
 814
 815         if (!buffer)
 816                 return;
 817
 818         buf = container_of(buffer, struct rpc_buffer, data);
 819         size = buf->len;
 820
 821         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 822                         size, buf);
 823
 824         if (size <= RPC_BUFFER_MAXSIZE)
 825                 mempool_free(buf, rpc_buffer_mempool);
 826         else
 827                 kfree(buf);
 828 }
 829 EXPORT_SYMBOL_GPL(rpc_free);
 830
 831 /*
 832  * Creation and deletion of RPC task structures
 833  */
 834 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 835 {
 836         memset(task, 0, sizeof(*task));
 837         atomic_set(&task->tk_count, 1);
 838         task->tk_flags  = task_setup_data->flags;
 839         task->tk_ops = task_setup_data->callback_ops;
 840         task->tk_calldata = task_setup_data->callback_data;
 841         INIT_LIST_HEAD(&task->tk_task);
 842
 843         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 844         task->tk_owner = current->tgid;
 845
 846         /* Initialize workqueue for async tasks */
 847         task->tk_workqueue = task_setup_data->workqueue;
 848
 849         if (task->tk_ops->rpc_call_prepare != NULL)
 850                 task->tk_action = rpc_prepare_task;
 851
 852         rpc_init_task_statistics(task);
 853
 854         dprintk("RPC:       new task initialized, procpid %u\n",
 855                                 task_pid_nr(current));
 856 }
 857
 858 static struct rpc_task *
 859 rpc_alloc_task(void)
 860 {
 861         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 862 }
 863
 864 /*
 865  * Create a new task for the specified client.
 866  */
 867 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 868 {
 869         struct rpc_task *task = setup_data->task;
 870         unsigned short flags = 0;
 871
 872         if (task == NULL) {
 873                 task = rpc_alloc_task();
 874                 if (task == NULL) {
 875                         rpc_release_calldata(setup_data->callback_ops,
 876                                         setup_data->callback_data);
 877                         return ERR_PTR(-ENOMEM);
 878                 }
 879                 flags = RPC_TASK_DYNAMIC;
 880         }
 881
 882         rpc_init_task(task, setup_data);
 883         task->tk_flags |= flags;
 884         dprintk("RPC:       allocated task %p\n", task);
 885         return task;
 886 }
 887
 888 /*
 889  * rpc_free_task - release rpc task and perform cleanups
 890  *
 891  * Note that we free up the rpc_task _after_ rpc_release_calldata()
 892  * in order to work around a workqueue dependency issue.
 893  *
 894  * Tejun Heo states:
 895  * "Workqueue currently considers two work items to be the same if they're
 896  * on the same address and won't execute them concurrently - ie. it
 897  * makes a work item which is queued again while being executed wait
 898  * for the previous execution to complete.
 899  *
 900  * If a work function frees the work item, and then waits for an event
 901  * which should be performed by another work item and *that* work item
 902  * recycles the freed work item, it can create a false dependency loop.
 903  * There really is no reliable way to detect this short of verifying
 904  * every memory free."
 905  *
 906  */
 907 static void rpc_free_task(struct rpc_task *task)
 908 {
 909         unsigned short tk_flags = task->tk_flags;
 910
 911         rpc_release_calldata(task->tk_ops, task->tk_calldata);
 912
 913         if (tk_flags & RPC_TASK_DYNAMIC) {
 914                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 915                 mempool_free(task, rpc_task_mempool);
 916         }
 917 }
 918
 919 static void rpc_async_release(struct work_struct *work)
 920 {
 921         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 922 }
 923
 924 static void rpc_release_resources_task(struct rpc_task *task)
 925 {
 926         xprt_release(task);
 927         if (task->tk_msg.rpc_cred) {
 928                 put_rpccred(task->tk_msg.rpc_cred);
 929                 task->tk_msg.rpc_cred = NULL;
 930         }
 931         rpc_task_release_client(task);
 932 }
 933
 934 static void rpc_final_put_task(struct rpc_task *task,
 935                 struct workqueue_struct *q)
 936 {
 937         if (q != NULL) {
 938                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 939                 queue_work(q, &task->u.tk_work);
 940         } else
 941                 rpc_free_task(task);
 942 }
 943
 944 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
 945 {
 946         if (atomic_dec_and_test(&task->tk_count)) {
 947                 rpc_release_resources_task(task);
 948                 rpc_final_put_task(task, q);
 949         }
 950 }
 951
 952 void rpc_put_task(struct rpc_task *task)
 953 {
 954         rpc_do_put_task(task, NULL);
 955 }
 956 EXPORT_SYMBOL_GPL(rpc_put_task);
 957
 958 void rpc_put_task_async(struct rpc_task *task)
 959 {
 960         rpc_do_put_task(task, task->tk_workqueue);
 961 }
 962 EXPORT_SYMBOL_GPL(rpc_put_task_async);
 963
 964 static void rpc_release_task(struct rpc_task *task)
 965 {
 966         dprintk("RPC: %5u release task\n", task->tk_pid);
 967
 968         BUG_ON (RPC_IS_QUEUED(task));
 969
 970         rpc_release_resources_task(task);
 971
 972         /*
 973          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
 974          * so it should be safe to use task->tk_count as a test for whether
 975          * or not any other processes still hold references to our rpc_task.
 976          */
 977         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
 978                 /* Wake up anyone who may be waiting for task completion */
 979                 if (!rpc_complete_task(task))
 980                         return;
 981         } else {
 982                 if (!atomic_dec_and_test(&task->tk_count))
 983                         return;
 984         }
 985         rpc_final_put_task(task, task->tk_workqueue);
 986 }
 987
 988 int rpciod_up(void)
 989 {
 990         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 991 }
 992
 993 void rpciod_down(void)
 994 {
 995         module_put(THIS_MODULE);
 996 }
 997
 998 /*
 999  * Start up the rpciod workqueue.
1000  */
1001 static int rpciod_start(void)
1002 {
1003         struct workqueue_struct *wq;
1004
1005         /*
1006          * Create the rpciod thread and wait for it to start.
1007          */
1008         dprintk("RPC:       creating workqueue rpciod\n");
1009         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1010         rpciod_workqueue = wq;
1011         return rpciod_workqueue != NULL;
1012 }
1013
1014 static void rpciod_stop(void)
1015 {
1016         struct workqueue_struct *wq = NULL;
1017
1018         if (rpciod_workqueue == NULL)
1019                 return;
1020         dprintk("RPC:       destroying workqueue rpciod\n");
1021
1022         wq = rpciod_workqueue;
1023         rpciod_workqueue = NULL;
1024         destroy_workqueue(wq);
1025 }
1026
1027 void
1028 rpc_destroy_mempool(void)
1029 {
1030         rpciod_stop();
1031         if (rpc_buffer_mempool)
1032                 mempool_destroy(rpc_buffer_mempool);
1033         if (rpc_task_mempool)
1034                 mempool_destroy(rpc_task_mempool);
1035         if (rpc_task_slabp)
1036                 kmem_cache_destroy(rpc_task_slabp);
1037         if (rpc_buffer_slabp)
1038                 kmem_cache_destroy(rpc_buffer_slabp);
1039         rpc_destroy_wait_queue(&delay_queue);
1040 }
1041
1042 int
1043 rpc_init_mempool(void)
1044 {
1045         /*
1046          * The following is not strictly a mempool initialisation,
1047          * but there is no harm in doing it here
1048          */
1049         rpc_init_wait_queue(&delay_queue, "delayq");
1050         if (!rpciod_start())
1051                 goto err_nomem;
1052
1053         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1054                                              sizeof(struct rpc_task),
1055                                              0, SLAB_HWCACHE_ALIGN,
1056                                              NULL);
1057         if (!rpc_task_slabp)
1058                 goto err_nomem;
1059         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1060                                              RPC_BUFFER_MAXSIZE,
1061                                              0, SLAB_HWCACHE_ALIGN,
1062                                              NULL);
1063         if (!rpc_buffer_slabp)
1064                 goto err_nomem;
1065         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1066                                                     rpc_task_slabp);
1067         if (!rpc_task_mempool)
1068                 goto err_nomem;
1069         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1070                                                       rpc_buffer_slabp);
1071         if (!rpc_buffer_mempool)
1072                 goto err_nomem;
1073         return 0;
1074 err_nomem:
1075         rpc_destroy_mempool();
1076         return -ENOMEM;
1077 }