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 #define CREATE_TRACE_POINTS
  32 #include <trace/events/sunrpc.h>
  33
  34 /*
  35  * RPC slabs and memory pools
  36  */
  37 #define RPC_BUFFER_MAXSIZE      (2048)
  38 #define RPC_BUFFER_POOLSIZE     (8)
  39 #define RPC_TASK_POOLSIZE       (8)
  40 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  41 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  42 static mempool_t        *rpc_task_mempool __read_mostly;
  43 static mempool_t        *rpc_buffer_mempool __read_mostly;
  44
  45 static void                     rpc_async_schedule(struct work_struct *);
  46 static void                      rpc_release_task(struct rpc_task *task);
  47 static void __rpc_queue_timer_fn(unsigned long ptr);
  48
  49 /*
  50  * RPC tasks sit here while waiting for conditions to improve.
  51  */
  52 static struct rpc_wait_queue delay_queue;
  53
  54 /*
  55  * rpciod-related stuff
  56  */
  57 struct workqueue_struct *rpciod_workqueue;
  58
  59 /*
  60  * Disable the timer for a given RPC task. Should be called with
  61  * queue->lock and bh_disabled in order to avoid races within
  62  * rpc_run_timer().
  63  */
  64 static void
  65 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  66 {
  67         if (task->tk_timeout == 0)
  68                 return;
  69         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  70         task->tk_timeout = 0;
  71         list_del(&task->u.tk_wait.timer_list);
  72         if (list_empty(&queue->timer_list.list))
  73                 del_timer(&queue->timer_list.timer);
  74 }
  75
  76 static void
  77 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  78 {
  79         queue->timer_list.expires = expires;
  80         mod_timer(&queue->timer_list.timer, expires);
  81 }
  82
  83 /*
  84  * Set up a timer for the current task.
  85  */
  86 static void
  87 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  88 {
  89         if (!task->tk_timeout)
  90                 return;
  91
  92         dprintk("RPC: %5u setting alarm for %lu ms\n",
  93                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  94
  95         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  96         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  97                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  98         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  99 }
 100
 101 /*
 102  * Add new request to a priority queue.
 103  */
 104 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 105                 struct rpc_task *task,
 106                 unsigned char queue_priority)
 107 {
 108         struct list_head *q;
 109         struct rpc_task *t;
 110
 111         INIT_LIST_HEAD(&task->u.tk_wait.links);
 112         q = &queue->tasks[queue_priority];
 113         if (unlikely(queue_priority > queue->maxpriority))
 114                 q = &queue->tasks[queue->maxpriority];
 115         list_for_each_entry(t, q, u.tk_wait.list) {
 116                 if (t->tk_owner == task->tk_owner) {
 117                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 118                         return;
 119                 }
 120         }
 121         list_add_tail(&task->u.tk_wait.list, q);
 122 }
 123
 124 /*
 125  * Add new request to wait queue.
 126  *
 127  * Swapper tasks always get inserted at the head of the queue.
 128  * This should avoid many nasty memory deadlocks and hopefully
 129  * improve overall performance.
 130  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 131  */
 132 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 133                 struct rpc_task *task,
 134                 unsigned char queue_priority)
 135 {
 136         BUG_ON (RPC_IS_QUEUED(task));
 137
 138         if (RPC_IS_PRIORITY(queue))
 139                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 140         else if (RPC_IS_SWAPPER(task))
 141                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 142         else
 143                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 144         task->tk_waitqueue = queue;
 145         queue->qlen++;
 146         rpc_set_queued(task);
 147
 148         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 149                         task->tk_pid, queue, rpc_qname(queue));
 150 }
 151
 152 /*
 153  * Remove request from a priority queue.
 154  */
 155 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 156 {
 157         struct rpc_task *t;
 158
 159         if (!list_empty(&task->u.tk_wait.links)) {
 160                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 161                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 162                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 163         }
 164 }
 165
 166 /*
 167  * Remove request from queue.
 168  * Note: must be called with spin lock held.
 169  */
 170 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 171 {
 172         __rpc_disable_timer(queue, task);
 173         if (RPC_IS_PRIORITY(queue))
 174                 __rpc_remove_wait_queue_priority(task);
 175         list_del(&task->u.tk_wait.list);
 176         queue->qlen--;
 177         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 178                         task->tk_pid, queue, rpc_qname(queue));
 179 }
 180
 181 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 182 {
 183         queue->priority = priority;
 184         queue->count = 1 << (priority * 2);
 185 }
 186
 187 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 188 {
 189         queue->owner = pid;
 190         queue->nr = RPC_BATCH_COUNT;
 191 }
 192
 193 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 194 {
 195         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 196         rpc_set_waitqueue_owner(queue, 0);
 197 }
 198
 199 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 200 {
 201         int i;
 202
 203         spin_lock_init(&queue->lock);
 204         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 205                 INIT_LIST_HEAD(&queue->tasks[i]);
 206         queue->maxpriority = nr_queues - 1;
 207         rpc_reset_waitqueue_priority(queue);
 208         queue->qlen = 0;
 209         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 210         INIT_LIST_HEAD(&queue->timer_list.list);
 211         rpc_assign_waitqueue_name(queue, qname);
 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         freezable_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         trace_rpc_task_begin(task->tk_client, task, NULL);
 256
 257         rpc_task_set_debuginfo(task);
 258         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 259 }
 260
 261 /*
 262  * Mark an RPC call as having completed by clearing the 'active' bit
 263  * and then waking up all tasks that were sleeping.
 264  */
 265 static int rpc_complete_task(struct rpc_task *task)
 266 {
 267         void *m = &task->tk_runstate;
 268         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 269         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 270         unsigned long flags;
 271         int ret;
 272
 273         trace_rpc_task_complete(task->tk_client, task, NULL);
 274
 275         spin_lock_irqsave(&wq->lock, flags);
 276         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 277         ret = atomic_dec_and_test(&task->tk_count);
 278         if (waitqueue_active(wq))
 279                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 280         spin_unlock_irqrestore(&wq->lock, flags);
 281         return ret;
 282 }
 283
 284 /*
 285  * Allow callers to wait for completion of an RPC call
 286  *
 287  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 288  * to enforce taking of the wq->lock and hence avoid races with
 289  * rpc_complete_task().
 290  */
 291 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 292 {
 293         if (action == NULL)
 294                 action = rpc_wait_bit_killable;
 295         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 296                         action, TASK_KILLABLE);
 297 }
 298 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 299
 300 /*
 301  * Make an RPC task runnable.
 302  *
 303  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 304  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 305  * the wait queue operation.
 306  */
 307 static void rpc_make_runnable(struct rpc_task *task)
 308 {
 309         rpc_clear_queued(task);
 310         if (rpc_test_and_set_running(task))
 311                 return;
 312         if (RPC_IS_ASYNC(task)) {
 313                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 314                 queue_work(rpciod_workqueue, &task->u.tk_work);
 315         } else
 316                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 317 }
 318
 319 /*
 320  * Prepare for sleeping on a wait queue.
 321  * By always appending tasks to the list we ensure FIFO behavior.
 322  * NB: An RPC task will only receive interrupt-driven events as long
 323  * as it's on a wait queue.
 324  */
 325 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 326                 struct rpc_task *task,
 327                 rpc_action action,
 328                 unsigned char queue_priority)
 329 {
 330         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 331                         task->tk_pid, rpc_qname(q), jiffies);
 332
 333         trace_rpc_task_sleep(task->tk_client, task, q);
 334
 335         __rpc_add_wait_queue(q, task, queue_priority);
 336
 337         BUG_ON(task->tk_callback != NULL);
 338         task->tk_callback = action;
 339         __rpc_add_timer(q, task);
 340 }
 341
 342 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 343                                 rpc_action action)
 344 {
 345         /* We shouldn't ever put an inactive task to sleep */
 346         BUG_ON(!RPC_IS_ACTIVATED(task));
 347
 348         /*
 349          * Protect the queue operations.
 350          */
 351         spin_lock_bh(&q->lock);
 352         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 353         spin_unlock_bh(&q->lock);
 354 }
 355 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 356
 357 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 358                 rpc_action action, int priority)
 359 {
 360         /* We shouldn't ever put an inactive task to sleep */
 361         BUG_ON(!RPC_IS_ACTIVATED(task));
 362
 363         /*
 364          * Protect the queue operations.
 365          */
 366         spin_lock_bh(&q->lock);
 367         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 368         spin_unlock_bh(&q->lock);
 369 }
 370
 371 /**
 372  * __rpc_do_wake_up_task - wake up a single rpc_task
 373  * @queue: wait queue
 374  * @task: task to be woken up
 375  *
 376  * Caller must hold queue->lock, and have cleared the task queued flag.
 377  */
 378 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 379 {
 380         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 381                         task->tk_pid, jiffies);
 382
 383         /* Has the task been executed yet? If not, we cannot wake it up! */
 384         if (!RPC_IS_ACTIVATED(task)) {
 385                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 386                 return;
 387         }
 388
 389         trace_rpc_task_wakeup(task->tk_client, task, queue);
 390
 391         __rpc_remove_wait_queue(queue, task);
 392
 393         rpc_make_runnable(task);
 394
 395         dprintk("RPC:       __rpc_wake_up_task done\n");
 396 }
 397
 398 /*
 399  * Wake up a queued task while the queue lock is being held
 400  */
 401 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 402 {
 403         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 404                 __rpc_do_wake_up_task(queue, task);
 405 }
 406
 407 /*
 408  * Tests whether rpc queue is empty
 409  */
 410 int rpc_queue_empty(struct rpc_wait_queue *queue)
 411 {
 412         int res;
 413
 414         spin_lock_bh(&queue->lock);
 415         res = queue->qlen;
 416         spin_unlock_bh(&queue->lock);
 417         return res == 0;
 418 }
 419 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 420
 421 /*
 422  * Wake up a task on a specific queue
 423  */
 424 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 425 {
 426         spin_lock_bh(&queue->lock);
 427         rpc_wake_up_task_queue_locked(queue, task);
 428         spin_unlock_bh(&queue->lock);
 429 }
 430 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 431
 432 /*
 433  * Wake up the next task on a priority queue.
 434  */
 435 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 436 {
 437         struct list_head *q;
 438         struct rpc_task *task;
 439
 440         /*
 441          * Service a batch of tasks from a single owner.
 442          */
 443         q = &queue->tasks[queue->priority];
 444         if (!list_empty(q)) {
 445                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 446                 if (queue->owner == task->tk_owner) {
 447                         if (--queue->nr)
 448                                 goto out;
 449                         list_move_tail(&task->u.tk_wait.list, q);
 450                 }
 451                 /*
 452                  * Check if we need to switch queues.
 453                  */
 454                 if (--queue->count)
 455                         goto new_owner;
 456         }
 457
 458         /*
 459          * Service the next queue.
 460          */
 461         do {
 462                 if (q == &queue->tasks[0])
 463                         q = &queue->tasks[queue->maxpriority];
 464                 else
 465                         q = q - 1;
 466                 if (!list_empty(q)) {
 467                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 468                         goto new_queue;
 469                 }
 470         } while (q != &queue->tasks[queue->priority]);
 471
 472         rpc_reset_waitqueue_priority(queue);
 473         return NULL;
 474
 475 new_queue:
 476         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 477 new_owner:
 478         rpc_set_waitqueue_owner(queue, task->tk_owner);
 479 out:
 480         return task;
 481 }
 482
 483 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 484 {
 485         if (RPC_IS_PRIORITY(queue))
 486                 return __rpc_find_next_queued_priority(queue);
 487         if (!list_empty(&queue->tasks[0]))
 488                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 489         return NULL;
 490 }
 491
 492 /*
 493  * Wake up the first task on the wait queue.
 494  */
 495 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 496                 bool (*func)(struct rpc_task *, void *), void *data)
 497 {
 498         struct rpc_task *task = NULL;
 499
 500         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 501                         queue, rpc_qname(queue));
 502         spin_lock_bh(&queue->lock);
 503         task = __rpc_find_next_queued(queue);
 504         if (task != NULL) {
 505                 if (func(task, data))
 506                         rpc_wake_up_task_queue_locked(queue, task);
 507                 else
 508                         task = NULL;
 509         }
 510         spin_unlock_bh(&queue->lock);
 511
 512         return task;
 513 }
 514 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 515
 516 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 517 {
 518         return true;
 519 }
 520
 521 /*
 522  * Wake up the next task on the wait queue.
 523 */
 524 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 525 {
 526         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 527 }
 528 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 529
 530 /**
 531  * rpc_wake_up - wake up all rpc_tasks
 532  * @queue: rpc_wait_queue on which the tasks are sleeping
 533  *
 534  * Grabs queue->lock
 535  */
 536 void rpc_wake_up(struct rpc_wait_queue *queue)
 537 {
 538         struct list_head *head;
 539
 540         spin_lock_bh(&queue->lock);
 541         head = &queue->tasks[queue->maxpriority];
 542         for (;;) {
 543                 while (!list_empty(head)) {
 544                         struct rpc_task *task;
 545                         task = list_first_entry(head,
 546                                         struct rpc_task,
 547                                         u.tk_wait.list);
 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);
 557
 558 /**
 559  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 560  * @queue: rpc_wait_queue on which the tasks are sleeping
 561  * @status: status value to set
 562  *
 563  * Grabs queue->lock
 564  */
 565 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 566 {
 567         struct list_head *head;
 568
 569         spin_lock_bh(&queue->lock);
 570         head = &queue->tasks[queue->maxpriority];
 571         for (;;) {
 572                 while (!list_empty(head)) {
 573                         struct rpc_task *task;
 574                         task = list_first_entry(head,
 575                                         struct rpc_task,
 576                                         u.tk_wait.list);
 577                         task->tk_status = status;
 578                         rpc_wake_up_task_queue_locked(queue, task);
 579                 }
 580                 if (head == &queue->tasks[0])
 581                         break;
 582                 head--;
 583         }
 584         spin_unlock_bh(&queue->lock);
 585 }
 586 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 587
 588 static void __rpc_queue_timer_fn(unsigned long ptr)
 589 {
 590         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 591         struct rpc_task *task, *n;
 592         unsigned long expires, now, timeo;
 593
 594         spin_lock(&queue->lock);
 595         expires = now = jiffies;
 596         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 597                 timeo = task->u.tk_wait.expires;
 598                 if (time_after_eq(now, timeo)) {
 599                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 600                         task->tk_status = -ETIMEDOUT;
 601                         rpc_wake_up_task_queue_locked(queue, task);
 602                         continue;
 603                 }
 604                 if (expires == now || time_after(expires, timeo))
 605                         expires = timeo;
 606         }
 607         if (!list_empty(&queue->timer_list.list))
 608                 rpc_set_queue_timer(queue, expires);
 609         spin_unlock(&queue->lock);
 610 }
 611
 612 static void __rpc_atrun(struct rpc_task *task)
 613 {
 614         task->tk_status = 0;
 615 }
 616
 617 /*
 618  * Run a task at a later time
 619  */
 620 void rpc_delay(struct rpc_task *task, unsigned long delay)
 621 {
 622         task->tk_timeout = delay;
 623         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 624 }
 625 EXPORT_SYMBOL_GPL(rpc_delay);
 626
 627 /*
 628  * Helper to call task->tk_ops->rpc_call_prepare
 629  */
 630 void rpc_prepare_task(struct rpc_task *task)
 631 {
 632         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 633 }
 634
 635 static void
 636 rpc_init_task_statistics(struct rpc_task *task)
 637 {
 638         /* Initialize retry counters */
 639         task->tk_garb_retry = 2;
 640         task->tk_cred_retry = 2;
 641         task->tk_rebind_retry = 2;
 642
 643         /* starting timestamp */
 644         task->tk_start = ktime_get();
 645 }
 646
 647 static void
 648 rpc_reset_task_statistics(struct rpc_task *task)
 649 {
 650         task->tk_timeouts = 0;
 651         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 652
 653         rpc_init_task_statistics(task);
 654 }
 655
 656 /*
 657  * Helper that calls task->tk_ops->rpc_call_done if it exists
 658  */
 659 void rpc_exit_task(struct rpc_task *task)
 660 {
 661         task->tk_action = NULL;
 662         if (task->tk_ops->rpc_call_done != NULL) {
 663                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 664                 if (task->tk_action != NULL) {
 665                         WARN_ON(RPC_ASSASSINATED(task));
 666                         /* Always release the RPC slot and buffer memory */
 667                         xprt_release(task);
 668                         rpc_reset_task_statistics(task);
 669                 }
 670         }
 671 }
 672
 673 void rpc_exit(struct rpc_task *task, int status)
 674 {
 675         task->tk_status = status;
 676         task->tk_action = rpc_exit_task;
 677         if (RPC_IS_QUEUED(task))
 678                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 679 }
 680 EXPORT_SYMBOL_GPL(rpc_exit);
 681
 682 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 683 {
 684         if (ops->rpc_release != NULL)
 685                 ops->rpc_release(calldata);
 686 }
 687
 688 /*
 689  * This is the RPC `scheduler' (or rather, the finite state machine).
 690  */
 691 static void __rpc_execute(struct rpc_task *task)
 692 {
 693         struct rpc_wait_queue *queue;
 694         int task_is_async = RPC_IS_ASYNC(task);
 695         int status = 0;
 696
 697         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 698                         task->tk_pid, task->tk_flags);
 699
 700         BUG_ON(RPC_IS_QUEUED(task));
 701
 702         for (;;) {
 703                 void (*do_action)(struct rpc_task *);
 704
 705                 /*
 706                  * Execute any pending callback first.
 707                  */
 708                 do_action = task->tk_callback;
 709                 task->tk_callback = NULL;
 710                 if (do_action == NULL) {
 711                         /*
 712                          * Perform the next FSM step.
 713                          * tk_action may be NULL if the task has been killed.
 714                          * In particular, note that rpc_killall_tasks may
 715                          * do this at any time, so beware when dereferencing.
 716                          */
 717                         do_action = task->tk_action;
 718                         if (do_action == NULL)
 719                                 break;
 720                 }
 721                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 722                 do_action(task);
 723
 724                 /*
 725                  * Lockless check for whether task is sleeping or not.
 726                  */
 727                 if (!RPC_IS_QUEUED(task))
 728                         continue;
 729                 /*
 730                  * The queue->lock protects against races with
 731                  * rpc_make_runnable().
 732                  *
 733                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 734                  * rpc_task, rpc_make_runnable() can assign it to a
 735                  * different workqueue. We therefore cannot assume that the
 736                  * rpc_task pointer may still be dereferenced.
 737                  */
 738                 queue = task->tk_waitqueue;
 739                 spin_lock_bh(&queue->lock);
 740                 if (!RPC_IS_QUEUED(task)) {
 741                         spin_unlock_bh(&queue->lock);
 742                         continue;
 743                 }
 744                 rpc_clear_running(task);
 745                 spin_unlock_bh(&queue->lock);
 746                 if (task_is_async)
 747                         return;
 748
 749                 /* sync task: sleep here */
 750                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 751                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 752                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 753                                 TASK_KILLABLE);
 754                 if (status == -ERESTARTSYS) {
 755                         /*
 756                          * When a sync task receives a signal, it exits with
 757                          * -ERESTARTSYS. In order to catch any callbacks that
 758                          * clean up after sleeping on some queue, we don't
 759                          * break the loop here, but go around once more.
 760                          */
 761                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 762                         task->tk_flags |= RPC_TASK_KILLED;
 763                         rpc_exit(task, -ERESTARTSYS);
 764                 }
 765                 rpc_set_running(task);
 766                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 767         }
 768
 769         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 770                         task->tk_status);
 771         /* Release all resources associated with the task */
 772         rpc_release_task(task);
 773 }
 774
 775 /*
 776  * User-visible entry point to the scheduler.
 777  *
 778  * This may be called recursively if e.g. an async NFS task updates
 779  * the attributes and finds that dirty pages must be flushed.
 780  * NOTE: Upon exit of this function the task is guaranteed to be
 781  *       released. In particular note that tk_release() will have
 782  *       been called, so your task memory may have been freed.
 783  */
 784 void rpc_execute(struct rpc_task *task)
 785 {
 786         rpc_set_active(task);
 787         rpc_make_runnable(task);
 788         if (!RPC_IS_ASYNC(task))
 789                 __rpc_execute(task);
 790 }
 791
 792 static void rpc_async_schedule(struct work_struct *work)
 793 {
 794         current->flags |= PF_FSTRANS;
 795         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 796         current->flags &= ~PF_FSTRANS;
 797 }
 798
 799 /**
 800  * rpc_malloc - allocate an RPC buffer
 801  * @task: RPC task that will use this buffer
 802  * @size: requested byte size
 803  *
 804  * To prevent rpciod from hanging, this allocator never sleeps,
 805  * returning NULL if the request cannot be serviced immediately.
 806  * The caller can arrange to sleep in a way that is safe for rpciod.
 807  *
 808  * Most requests are 'small' (under 2KiB) and can be serviced from a
 809  * mempool, ensuring that NFS reads and writes can always proceed,
 810  * and that there is good locality of reference for these buffers.
 811  *
 812  * In order to avoid memory starvation triggering more writebacks of
 813  * NFS requests, we avoid using GFP_KERNEL.
 814  */
 815 void *rpc_malloc(struct rpc_task *task, size_t size)
 816 {
 817         struct rpc_buffer *buf;
 818         gfp_t gfp = GFP_NOWAIT;
 819
 820         if (RPC_IS_SWAPPER(task))
 821                 gfp |= __GFP_MEMALLOC;
 822
 823         size += sizeof(struct rpc_buffer);
 824         if (size <= RPC_BUFFER_MAXSIZE)
 825                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 826         else
 827                 buf = kmalloc(size, gfp);
 828
 829         if (!buf)
 830                 return NULL;
 831
 832         buf->len = size;
 833         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 834                         task->tk_pid, size, buf);
 835         return &buf->data;
 836 }
 837 EXPORT_SYMBOL_GPL(rpc_malloc);
 838
 839 /**
 840  * rpc_free - free buffer allocated via rpc_malloc
 841  * @buffer: buffer to free
 842  *
 843  */
 844 void rpc_free(void *buffer)
 845 {
 846         size_t size;
 847         struct rpc_buffer *buf;
 848
 849         if (!buffer)
 850                 return;
 851
 852         buf = container_of(buffer, struct rpc_buffer, data);
 853         size = buf->len;
 854
 855         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 856                         size, buf);
 857
 858         if (size <= RPC_BUFFER_MAXSIZE)
 859                 mempool_free(buf, rpc_buffer_mempool);
 860         else
 861                 kfree(buf);
 862 }
 863 EXPORT_SYMBOL_GPL(rpc_free);
 864
 865 /*
 866  * Creation and deletion of RPC task structures
 867  */
 868 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 869 {
 870         memset(task, 0, sizeof(*task));
 871         atomic_set(&task->tk_count, 1);
 872         task->tk_flags  = task_setup_data->flags;
 873         task->tk_ops = task_setup_data->callback_ops;
 874         task->tk_calldata = task_setup_data->callback_data;
 875         INIT_LIST_HEAD(&task->tk_task);
 876
 877         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 878         task->tk_owner = current->tgid;
 879
 880         /* Initialize workqueue for async tasks */
 881         task->tk_workqueue = task_setup_data->workqueue;
 882
 883         if (task->tk_ops->rpc_call_prepare != NULL)
 884                 task->tk_action = rpc_prepare_task;
 885
 886         rpc_init_task_statistics(task);
 887
 888         dprintk("RPC:       new task initialized, procpid %u\n",
 889                                 task_pid_nr(current));
 890 }
 891
 892 static struct rpc_task *
 893 rpc_alloc_task(void)
 894 {
 895         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 896 }
 897
 898 /*
 899  * Create a new task for the specified client.
 900  */
 901 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 902 {
 903         struct rpc_task *task = setup_data->task;
 904         unsigned short flags = 0;
 905
 906         if (task == NULL) {
 907                 task = rpc_alloc_task();
 908                 if (task == NULL) {
 909                         rpc_release_calldata(setup_data->callback_ops,
 910                                         setup_data->callback_data);
 911                         return ERR_PTR(-ENOMEM);
 912                 }
 913                 flags = RPC_TASK_DYNAMIC;
 914         }
 915
 916         rpc_init_task(task, setup_data);
 917         task->tk_flags |= flags;
 918         dprintk("RPC:       allocated task %p\n", task);
 919         return task;
 920 }
 921
 922 static void rpc_free_task(struct rpc_task *task)
 923 {
 924         const struct rpc_call_ops *tk_ops = task->tk_ops;
 925         void *calldata = task->tk_calldata;
 926
 927         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 928                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 929                 mempool_free(task, rpc_task_mempool);
 930         }
 931         rpc_release_calldata(tk_ops, calldata);
 932 }
 933
 934 static void rpc_async_release(struct work_struct *work)
 935 {
 936         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 937 }
 938
 939 static void rpc_release_resources_task(struct rpc_task *task)
 940 {
 941         if (task->tk_rqstp)
 942                 xprt_release(task);
 943         if (task->tk_msg.rpc_cred) {
 944                 put_rpccred(task->tk_msg.rpc_cred);
 945                 task->tk_msg.rpc_cred = NULL;
 946         }
 947         rpc_task_release_client(task);
 948 }
 949
 950 static void rpc_final_put_task(struct rpc_task *task,
 951                 struct workqueue_struct *q)
 952 {
 953         if (q != NULL) {
 954                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 955                 queue_work(q, &task->u.tk_work);
 956         } else
 957                 rpc_free_task(task);
 958 }
 959
 960 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
 961 {
 962         if (atomic_dec_and_test(&task->tk_count)) {
 963                 rpc_release_resources_task(task);
 964                 rpc_final_put_task(task, q);
 965         }
 966 }
 967
 968 void rpc_put_task(struct rpc_task *task)
 969 {
 970         rpc_do_put_task(task, NULL);
 971 }
 972 EXPORT_SYMBOL_GPL(rpc_put_task);
 973
 974 void rpc_put_task_async(struct rpc_task *task)
 975 {
 976         rpc_do_put_task(task, task->tk_workqueue);
 977 }
 978 EXPORT_SYMBOL_GPL(rpc_put_task_async);
 979
 980 static void rpc_release_task(struct rpc_task *task)
 981 {
 982         dprintk("RPC: %5u release task\n", task->tk_pid);
 983
 984         BUG_ON (RPC_IS_QUEUED(task));
 985
 986         rpc_release_resources_task(task);
 987
 988         /*
 989          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
 990          * so it should be safe to use task->tk_count as a test for whether
 991          * or not any other processes still hold references to our rpc_task.
 992          */
 993         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
 994                 /* Wake up anyone who may be waiting for task completion */
 995                 if (!rpc_complete_task(task))
 996                         return;
 997         } else {
 998                 if (!atomic_dec_and_test(&task->tk_count))
 999                         return;
1000         }
1001         rpc_final_put_task(task, task->tk_workqueue);
1002 }
1003
1004 int rpciod_up(void)
1005 {
1006         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1007 }
1008
1009 void rpciod_down(void)
1010 {
1011         module_put(THIS_MODULE);
1012 }
1013
1014 /*
1015  * Start up the rpciod workqueue.
1016  */
1017 static int rpciod_start(void)
1018 {
1019         struct workqueue_struct *wq;
1020
1021         /*
1022          * Create the rpciod thread and wait for it to start.
1023          */
1024         dprintk("RPC:       creating workqueue rpciod\n");
1025         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1026         rpciod_workqueue = wq;
1027         return rpciod_workqueue != NULL;
1028 }
1029
1030 static void rpciod_stop(void)
1031 {
1032         struct workqueue_struct *wq = NULL;
1033
1034         if (rpciod_workqueue == NULL)
1035                 return;
1036         dprintk("RPC:       destroying workqueue rpciod\n");
1037
1038         wq = rpciod_workqueue;
1039         rpciod_workqueue = NULL;
1040         destroy_workqueue(wq);
1041 }
1042
1043 void
1044 rpc_destroy_mempool(void)
1045 {
1046         rpciod_stop();
1047         if (rpc_buffer_mempool)
1048                 mempool_destroy(rpc_buffer_mempool);
1049         if (rpc_task_mempool)
1050                 mempool_destroy(rpc_task_mempool);
1051         if (rpc_task_slabp)
1052                 kmem_cache_destroy(rpc_task_slabp);
1053         if (rpc_buffer_slabp)
1054                 kmem_cache_destroy(rpc_buffer_slabp);
1055         rpc_destroy_wait_queue(&delay_queue);
1056 }
1057
1058 int
1059 rpc_init_mempool(void)
1060 {
1061         /*
1062          * The following is not strictly a mempool initialisation,
1063          * but there is no harm in doing it here
1064          */
1065         rpc_init_wait_queue(&delay_queue, "delayq");
1066         if (!rpciod_start())
1067                 goto err_nomem;
1068
1069         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1070                                              sizeof(struct rpc_task),
1071                                              0, SLAB_HWCACHE_ALIGN,
1072                                              NULL);
1073         if (!rpc_task_slabp)
1074                 goto err_nomem;
1075         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1076                                              RPC_BUFFER_MAXSIZE,
1077                                              0, SLAB_HWCACHE_ALIGN,
1078                                              NULL);
1079         if (!rpc_buffer_slabp)
1080                 goto err_nomem;
1081         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1082                                                     rpc_task_slabp);
1083         if (!rpc_task_mempool)
1084                 goto err_nomem;
1085         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1086                                                       rpc_buffer_slabp);
1087         if (!rpc_buffer_mempool)
1088                 goto err_nomem;
1089         return 0;
1090 err_nomem:
1091         rpc_destroy_mempool();
1092         return -ENOMEM;
1093 }