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, this must be called with
 304  * the spinlock held to protect the wait queue operation.
 305  */
 306 static void rpc_make_runnable(struct rpc_task *task)
 307 {
 308         rpc_clear_queued(task);
 309         if (rpc_test_and_set_running(task))
 310                 return;
 311         if (RPC_IS_ASYNC(task)) {
 312                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 313                 queue_work(rpciod_workqueue, &task->u.tk_work);
 314         } else
 315                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 316 }
 317
 318 /*
 319  * Prepare for sleeping on a wait queue.
 320  * By always appending tasks to the list we ensure FIFO behavior.
 321  * NB: An RPC task will only receive interrupt-driven events as long
 322  * as it's on a wait queue.
 323  */
 324 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 325                 struct rpc_task *task,
 326                 rpc_action action,
 327                 unsigned char queue_priority)
 328 {
 329         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 330                         task->tk_pid, rpc_qname(q), jiffies);
 331
 332         trace_rpc_task_sleep(task->tk_client, task, q);
 333
 334         __rpc_add_wait_queue(q, task, queue_priority);
 335
 336         BUG_ON(task->tk_callback != NULL);
 337         task->tk_callback = action;
 338         __rpc_add_timer(q, task);
 339 }
 340
 341 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 342                                 rpc_action action)
 343 {
 344         /* We shouldn't ever put an inactive task to sleep */
 345         BUG_ON(!RPC_IS_ACTIVATED(task));
 346
 347         /*
 348          * Protect the queue operations.
 349          */
 350         spin_lock_bh(&q->lock);
 351         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 352         spin_unlock_bh(&q->lock);
 353 }
 354 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 355
 356 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 357                 rpc_action action, int priority)
 358 {
 359         /* We shouldn't ever put an inactive task to sleep */
 360         BUG_ON(!RPC_IS_ACTIVATED(task));
 361
 362         /*
 363          * Protect the queue operations.
 364          */
 365         spin_lock_bh(&q->lock);
 366         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 367         spin_unlock_bh(&q->lock);
 368 }
 369
 370 /**
 371  * __rpc_do_wake_up_task - wake up a single rpc_task
 372  * @queue: wait queue
 373  * @task: task to be woken up
 374  *
 375  * Caller must hold queue->lock, and have cleared the task queued flag.
 376  */
 377 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 378 {
 379         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 380                         task->tk_pid, jiffies);
 381
 382         /* Has the task been executed yet? If not, we cannot wake it up! */
 383         if (!RPC_IS_ACTIVATED(task)) {
 384                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 385                 return;
 386         }
 387
 388         trace_rpc_task_wakeup(task->tk_client, task, queue);
 389
 390         __rpc_remove_wait_queue(queue, task);
 391
 392         rpc_make_runnable(task);
 393
 394         dprintk("RPC:       __rpc_wake_up_task done\n");
 395 }
 396
 397 /*
 398  * Wake up a queued task while the queue lock is being held
 399  */
 400 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 401 {
 402         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 403                 __rpc_do_wake_up_task(queue, task);
 404 }
 405
 406 /*
 407  * Tests whether rpc queue is empty
 408  */
 409 int rpc_queue_empty(struct rpc_wait_queue *queue)
 410 {
 411         int res;
 412
 413         spin_lock_bh(&queue->lock);
 414         res = queue->qlen;
 415         spin_unlock_bh(&queue->lock);
 416         return res == 0;
 417 }
 418 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 419
 420 /*
 421  * Wake up a task on a specific queue
 422  */
 423 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 424 {
 425         spin_lock_bh(&queue->lock);
 426         rpc_wake_up_task_queue_locked(queue, task);
 427         spin_unlock_bh(&queue->lock);
 428 }
 429 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 430
 431 /*
 432  * Wake up the next task on a priority queue.
 433  */
 434 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 435 {
 436         struct list_head *q;
 437         struct rpc_task *task;
 438
 439         /*
 440          * Service a batch of tasks from a single owner.
 441          */
 442         q = &queue->tasks[queue->priority];
 443         if (!list_empty(q)) {
 444                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 445                 if (queue->owner == task->tk_owner) {
 446                         if (--queue->nr)
 447                                 goto out;
 448                         list_move_tail(&task->u.tk_wait.list, q);
 449                 }
 450                 /*
 451                  * Check if we need to switch queues.
 452                  */
 453                 if (--queue->count)
 454                         goto new_owner;
 455         }
 456
 457         /*
 458          * Service the next queue.
 459          */
 460         do {
 461                 if (q == &queue->tasks[0])
 462                         q = &queue->tasks[queue->maxpriority];
 463                 else
 464                         q = q - 1;
 465                 if (!list_empty(q)) {
 466                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 467                         goto new_queue;
 468                 }
 469         } while (q != &queue->tasks[queue->priority]);
 470
 471         rpc_reset_waitqueue_priority(queue);
 472         return NULL;
 473
 474 new_queue:
 475         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 476 new_owner:
 477         rpc_set_waitqueue_owner(queue, task->tk_owner);
 478 out:
 479         return task;
 480 }
 481
 482 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 483 {
 484         if (RPC_IS_PRIORITY(queue))
 485                 return __rpc_find_next_queued_priority(queue);
 486         if (!list_empty(&queue->tasks[0]))
 487                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 488         return NULL;
 489 }
 490
 491 /*
 492  * Wake up the first task on the wait queue.
 493  */
 494 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 495                 bool (*func)(struct rpc_task *, void *), void *data)
 496 {
 497         struct rpc_task *task = NULL;
 498
 499         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 500                         queue, rpc_qname(queue));
 501         spin_lock_bh(&queue->lock);
 502         task = __rpc_find_next_queued(queue);
 503         if (task != NULL) {
 504                 if (func(task, data))
 505                         rpc_wake_up_task_queue_locked(queue, task);
 506                 else
 507                         task = NULL;
 508         }
 509         spin_unlock_bh(&queue->lock);
 510
 511         return task;
 512 }
 513 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 514
 515 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 516 {
 517         return true;
 518 }
 519
 520 /*
 521  * Wake up the next task on the wait queue.
 522 */
 523 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 524 {
 525         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 526 }
 527 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 528
 529 /**
 530  * rpc_wake_up - wake up all rpc_tasks
 531  * @queue: rpc_wait_queue on which the tasks are sleeping
 532  *
 533  * Grabs queue->lock
 534  */
 535 void rpc_wake_up(struct rpc_wait_queue *queue)
 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                         rpc_wake_up_task_queue_locked(queue, task);
 548                 }
 549                 if (head == &queue->tasks[0])
 550                         break;
 551                 head--;
 552         }
 553         spin_unlock_bh(&queue->lock);
 554 }
 555 EXPORT_SYMBOL_GPL(rpc_wake_up);
 556
 557 /**
 558  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 559  * @queue: rpc_wait_queue on which the tasks are sleeping
 560  * @status: status value to set
 561  *
 562  * Grabs queue->lock
 563  */
 564 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 565 {
 566         struct list_head *head;
 567
 568         spin_lock_bh(&queue->lock);
 569         head = &queue->tasks[queue->maxpriority];
 570         for (;;) {
 571                 while (!list_empty(head)) {
 572                         struct rpc_task *task;
 573                         task = list_first_entry(head,
 574                                         struct rpc_task,
 575                                         u.tk_wait.list);
 576                         task->tk_status = status;
 577                         rpc_wake_up_task_queue_locked(queue, task);
 578                 }
 579                 if (head == &queue->tasks[0])
 580                         break;
 581                 head--;
 582         }
 583         spin_unlock_bh(&queue->lock);
 584 }
 585 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 586
 587 /**
 588  * rpc_wake_up_softconn_status - wake up all SOFTCONN rpc_tasks and set their
 589  * status value.
 590  * @queue: rpc_wait_queue on which the tasks are sleeping
 591  * @status: status value to set
 592  *
 593  * Grabs queue->lock
 594  */
 595 void rpc_wake_up_softconn_status(struct rpc_wait_queue *queue, int status)
 596 {
 597         struct rpc_task *task, *next;
 598         struct list_head *head;
 599
 600         spin_lock_bh(&queue->lock);
 601         head = &queue->tasks[queue->maxpriority];
 602         for (;;) {
 603                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 604                         if (RPC_IS_SOFTCONN(task)) {
 605                                 task->tk_status = status;
 606                                 rpc_wake_up_task_queue_locked(queue, task);
 607                         }
 608                 if (head == &queue->tasks[0])
 609                         break;
 610                 head--;
 611         }
 612         spin_unlock_bh(&queue->lock);
 613 }
 614 EXPORT_SYMBOL_GPL(rpc_wake_up_softconn_status);
 615
 616 static void __rpc_queue_timer_fn(unsigned long ptr)
 617 {
 618         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 619         struct rpc_task *task, *n;
 620         unsigned long expires, now, timeo;
 621
 622         spin_lock(&queue->lock);
 623         expires = now = jiffies;
 624         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 625                 timeo = task->u.tk_wait.expires;
 626                 if (time_after_eq(now, timeo)) {
 627                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 628                         task->tk_status = -ETIMEDOUT;
 629                         rpc_wake_up_task_queue_locked(queue, task);
 630                         continue;
 631                 }
 632                 if (expires == now || time_after(expires, timeo))
 633                         expires = timeo;
 634         }
 635         if (!list_empty(&queue->timer_list.list))
 636                 rpc_set_queue_timer(queue, expires);
 637         spin_unlock(&queue->lock);
 638 }
 639
 640 static void __rpc_atrun(struct rpc_task *task)
 641 {
 642         task->tk_status = 0;
 643 }
 644
 645 /*
 646  * Run a task at a later time
 647  */
 648 void rpc_delay(struct rpc_task *task, unsigned long delay)
 649 {
 650         task->tk_timeout = delay;
 651         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 652 }
 653 EXPORT_SYMBOL_GPL(rpc_delay);
 654
 655 /*
 656  * Helper to call task->tk_ops->rpc_call_prepare
 657  */
 658 void rpc_prepare_task(struct rpc_task *task)
 659 {
 660         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 661 }
 662
 663 static void
 664 rpc_init_task_statistics(struct rpc_task *task)
 665 {
 666         /* Initialize retry counters */
 667         task->tk_garb_retry = 2;
 668         task->tk_cred_retry = 2;
 669         task->tk_rebind_retry = 2;
 670
 671         /* starting timestamp */
 672         task->tk_start = ktime_get();
 673 }
 674
 675 static void
 676 rpc_reset_task_statistics(struct rpc_task *task)
 677 {
 678         task->tk_timeouts = 0;
 679         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 680
 681         rpc_init_task_statistics(task);
 682 }
 683
 684 /*
 685  * Helper that calls task->tk_ops->rpc_call_done if it exists
 686  */
 687 void rpc_exit_task(struct rpc_task *task)
 688 {
 689         task->tk_action = NULL;
 690         if (task->tk_ops->rpc_call_done != NULL) {
 691                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 692                 if (task->tk_action != NULL) {
 693                         WARN_ON(RPC_ASSASSINATED(task));
 694                         /* Always release the RPC slot and buffer memory */
 695                         xprt_release(task);
 696                         rpc_reset_task_statistics(task);
 697                 }
 698         }
 699 }
 700
 701 void rpc_exit(struct rpc_task *task, int status)
 702 {
 703         task->tk_status = status;
 704         task->tk_action = rpc_exit_task;
 705         if (RPC_IS_QUEUED(task))
 706                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 707 }
 708 EXPORT_SYMBOL_GPL(rpc_exit);
 709
 710 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 711 {
 712         if (ops->rpc_release != NULL)
 713                 ops->rpc_release(calldata);
 714 }
 715
 716 /*
 717  * This is the RPC `scheduler' (or rather, the finite state machine).
 718  */
 719 static void __rpc_execute(struct rpc_task *task)
 720 {
 721         struct rpc_wait_queue *queue;
 722         int task_is_async = RPC_IS_ASYNC(task);
 723         int status = 0;
 724
 725         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 726                         task->tk_pid, task->tk_flags);
 727
 728         BUG_ON(RPC_IS_QUEUED(task));
 729
 730         for (;;) {
 731                 void (*do_action)(struct rpc_task *);
 732
 733                 /*
 734                  * Execute any pending callback first.
 735                  */
 736                 do_action = task->tk_callback;
 737                 task->tk_callback = NULL;
 738                 if (do_action == NULL) {
 739                         /*
 740                          * Perform the next FSM step.
 741                          * tk_action may be NULL if the task has been killed.
 742                          * In particular, note that rpc_killall_tasks may
 743                          * do this at any time, so beware when dereferencing.
 744                          */
 745                         do_action = task->tk_action;
 746                         if (do_action == NULL)
 747                                 break;
 748                 }
 749                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 750                 do_action(task);
 751
 752                 /*
 753                  * Lockless check for whether task is sleeping or not.
 754                  */
 755                 if (!RPC_IS_QUEUED(task))
 756                         continue;
 757                 /*
 758                  * The queue->lock protects against races with
 759                  * rpc_make_runnable().
 760                  *
 761                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 762                  * rpc_task, rpc_make_runnable() can assign it to a
 763                  * different workqueue. We therefore cannot assume that the
 764                  * rpc_task pointer may still be dereferenced.
 765                  */
 766                 queue = task->tk_waitqueue;
 767                 spin_lock_bh(&queue->lock);
 768                 if (!RPC_IS_QUEUED(task)) {
 769                         spin_unlock_bh(&queue->lock);
 770                         continue;
 771                 }
 772                 rpc_clear_running(task);
 773                 spin_unlock_bh(&queue->lock);
 774                 if (task_is_async)
 775                         return;
 776
 777                 /* sync task: sleep here */
 778                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 779                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 780                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 781                                 TASK_KILLABLE);
 782                 if (status == -ERESTARTSYS) {
 783                         /*
 784                          * When a sync task receives a signal, it exits with
 785                          * -ERESTARTSYS. In order to catch any callbacks that
 786                          * clean up after sleeping on some queue, we don't
 787                          * break the loop here, but go around once more.
 788                          */
 789                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 790                         task->tk_flags |= RPC_TASK_KILLED;
 791                         rpc_exit(task, -ERESTARTSYS);
 792                 }
 793                 rpc_set_running(task);
 794                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 795         }
 796
 797         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 798                         task->tk_status);
 799         /* Release all resources associated with the task */
 800         rpc_release_task(task);
 801 }
 802
 803 /*
 804  * User-visible entry point to the scheduler.
 805  *
 806  * This may be called recursively if e.g. an async NFS task updates
 807  * the attributes and finds that dirty pages must be flushed.
 808  * NOTE: Upon exit of this function the task is guaranteed to be
 809  *       released. In particular note that tk_release() will have
 810  *       been called, so your task memory may have been freed.
 811  */
 812 void rpc_execute(struct rpc_task *task)
 813 {
 814         rpc_set_active(task);
 815         rpc_make_runnable(task);
 816         if (!RPC_IS_ASYNC(task))
 817                 __rpc_execute(task);
 818 }
 819
 820 static void rpc_async_schedule(struct work_struct *work)
 821 {
 822         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 823 }
 824
 825 /**
 826  * rpc_malloc - allocate an RPC buffer
 827  * @task: RPC task that will use this buffer
 828  * @size: requested byte size
 829  *
 830  * To prevent rpciod from hanging, this allocator never sleeps,
 831  * returning NULL if the request cannot be serviced immediately.
 832  * The caller can arrange to sleep in a way that is safe for rpciod.
 833  *
 834  * Most requests are 'small' (under 2KiB) and can be serviced from a
 835  * mempool, ensuring that NFS reads and writes can always proceed,
 836  * and that there is good locality of reference for these buffers.
 837  *
 838  * In order to avoid memory starvation triggering more writebacks of
 839  * NFS requests, we avoid using GFP_KERNEL.
 840  */
 841 void *rpc_malloc(struct rpc_task *task, size_t size)
 842 {
 843         struct rpc_buffer *buf;
 844         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 845
 846         size += sizeof(struct rpc_buffer);
 847         if (size <= RPC_BUFFER_MAXSIZE)
 848                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 849         else
 850                 buf = kmalloc(size, gfp);
 851
 852         if (!buf)
 853                 return NULL;
 854
 855         buf->len = size;
 856         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 857                         task->tk_pid, size, buf);
 858         return &buf->data;
 859 }
 860 EXPORT_SYMBOL_GPL(rpc_malloc);
 861
 862 /**
 863  * rpc_free - free buffer allocated via rpc_malloc
 864  * @buffer: buffer to free
 865  *
 866  */
 867 void rpc_free(void *buffer)
 868 {
 869         size_t size;
 870         struct rpc_buffer *buf;
 871
 872         if (!buffer)
 873                 return;
 874
 875         buf = container_of(buffer, struct rpc_buffer, data);
 876         size = buf->len;
 877
 878         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 879                         size, buf);
 880
 881         if (size <= RPC_BUFFER_MAXSIZE)
 882                 mempool_free(buf, rpc_buffer_mempool);
 883         else
 884                 kfree(buf);
 885 }
 886 EXPORT_SYMBOL_GPL(rpc_free);
 887
 888 /*
 889  * Creation and deletion of RPC task structures
 890  */
 891 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 892 {
 893         memset(task, 0, sizeof(*task));
 894         atomic_set(&task->tk_count, 1);
 895         task->tk_flags  = task_setup_data->flags;
 896         task->tk_ops = task_setup_data->callback_ops;
 897         task->tk_calldata = task_setup_data->callback_data;
 898         INIT_LIST_HEAD(&task->tk_task);
 899
 900         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 901         task->tk_owner = current->tgid;
 902
 903         /* Initialize workqueue for async tasks */
 904         task->tk_workqueue = task_setup_data->workqueue;
 905
 906         if (task->tk_ops->rpc_call_prepare != NULL)
 907                 task->tk_action = rpc_prepare_task;
 908
 909         rpc_init_task_statistics(task);
 910
 911         dprintk("RPC:       new task initialized, procpid %u\n",
 912                                 task_pid_nr(current));
 913 }
 914
 915 static struct rpc_task *
 916 rpc_alloc_task(void)
 917 {
 918         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 919 }
 920
 921 /*
 922  * Create a new task for the specified client.
 923  */
 924 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 925 {
 926         struct rpc_task *task = setup_data->task;
 927         unsigned short flags = 0;
 928
 929         if (task == NULL) {
 930                 task = rpc_alloc_task();
 931                 if (task == NULL) {
 932                         rpc_release_calldata(setup_data->callback_ops,
 933                                         setup_data->callback_data);
 934                         return ERR_PTR(-ENOMEM);
 935                 }
 936                 flags = RPC_TASK_DYNAMIC;
 937         }
 938
 939         rpc_init_task(task, setup_data);
 940         task->tk_flags |= flags;
 941         dprintk("RPC:       allocated task %p\n", task);
 942         return task;
 943 }
 944
 945 static void rpc_free_task(struct rpc_task *task)
 946 {
 947         const struct rpc_call_ops *tk_ops = task->tk_ops;
 948         void *calldata = task->tk_calldata;
 949
 950         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 951                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 952                 mempool_free(task, rpc_task_mempool);
 953         }
 954         rpc_release_calldata(tk_ops, calldata);
 955 }
 956
 957 static void rpc_async_release(struct work_struct *work)
 958 {
 959         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 960 }
 961
 962 static void rpc_release_resources_task(struct rpc_task *task)
 963 {
 964         if (task->tk_rqstp)
 965                 xprt_release(task);
 966         if (task->tk_msg.rpc_cred) {
 967                 put_rpccred(task->tk_msg.rpc_cred);
 968                 task->tk_msg.rpc_cred = NULL;
 969         }
 970         rpc_task_release_client(task);
 971 }
 972
 973 static void rpc_final_put_task(struct rpc_task *task,
 974                 struct workqueue_struct *q)
 975 {
 976         if (q != NULL) {
 977                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 978                 queue_work(q, &task->u.tk_work);
 979         } else
 980                 rpc_free_task(task);
 981 }
 982
 983 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
 984 {
 985         if (atomic_dec_and_test(&task->tk_count)) {
 986                 rpc_release_resources_task(task);
 987                 rpc_final_put_task(task, q);
 988         }
 989 }
 990
 991 void rpc_put_task(struct rpc_task *task)
 992 {
 993         rpc_do_put_task(task, NULL);
 994 }
 995 EXPORT_SYMBOL_GPL(rpc_put_task);
 996
 997 void rpc_put_task_async(struct rpc_task *task)
 998 {
 999         rpc_do_put_task(task, task->tk_workqueue);
1000 }
1001 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1002
1003 static void rpc_release_task(struct rpc_task *task)
1004 {
1005         dprintk("RPC: %5u release task\n", task->tk_pid);
1006
1007         BUG_ON (RPC_IS_QUEUED(task));
1008
1009         rpc_release_resources_task(task);
1010
1011         /*
1012          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1013          * so it should be safe to use task->tk_count as a test for whether
1014          * or not any other processes still hold references to our rpc_task.
1015          */
1016         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1017                 /* Wake up anyone who may be waiting for task completion */
1018                 if (!rpc_complete_task(task))
1019                         return;
1020         } else {
1021                 if (!atomic_dec_and_test(&task->tk_count))
1022                         return;
1023         }
1024         rpc_final_put_task(task, task->tk_workqueue);
1025 }
1026
1027 int rpciod_up(void)
1028 {
1029         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1030 }
1031
1032 void rpciod_down(void)
1033 {
1034         module_put(THIS_MODULE);
1035 }
1036
1037 /*
1038  * Start up the rpciod workqueue.
1039  */
1040 static int rpciod_start(void)
1041 {
1042         struct workqueue_struct *wq;
1043
1044         /*
1045          * Create the rpciod thread and wait for it to start.
1046          */
1047         dprintk("RPC:       creating workqueue rpciod\n");
1048         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1049         rpciod_workqueue = wq;
1050         return rpciod_workqueue != NULL;
1051 }
1052
1053 static void rpciod_stop(void)
1054 {
1055         struct workqueue_struct *wq = NULL;
1056
1057         if (rpciod_workqueue == NULL)
1058                 return;
1059         dprintk("RPC:       destroying workqueue rpciod\n");
1060
1061         wq = rpciod_workqueue;
1062         rpciod_workqueue = NULL;
1063         destroy_workqueue(wq);
1064 }
1065
1066 void
1067 rpc_destroy_mempool(void)
1068 {
1069         rpciod_stop();
1070         if (rpc_buffer_mempool)
1071                 mempool_destroy(rpc_buffer_mempool);
1072         if (rpc_task_mempool)
1073                 mempool_destroy(rpc_task_mempool);
1074         if (rpc_task_slabp)
1075                 kmem_cache_destroy(rpc_task_slabp);
1076         if (rpc_buffer_slabp)
1077                 kmem_cache_destroy(rpc_buffer_slabp);
1078         rpc_destroy_wait_queue(&delay_queue);
1079 }
1080
1081 int
1082 rpc_init_mempool(void)
1083 {
1084         /*
1085          * The following is not strictly a mempool initialisation,
1086          * but there is no harm in doing it here
1087          */
1088         rpc_init_wait_queue(&delay_queue, "delayq");
1089         if (!rpciod_start())
1090                 goto err_nomem;
1091
1092         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1093                                              sizeof(struct rpc_task),
1094                                              0, SLAB_HWCACHE_ALIGN,
1095                                              NULL);
1096         if (!rpc_task_slabp)
1097                 goto err_nomem;
1098         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1099                                              RPC_BUFFER_MAXSIZE,
1100                                              0, SLAB_HWCACHE_ALIGN,
1101                                              NULL);
1102         if (!rpc_buffer_slabp)
1103                 goto err_nomem;
1104         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1105                                                     rpc_task_slabp);
1106         if (!rpc_task_mempool)
1107                 goto err_nomem;
1108         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1109                                                       rpc_buffer_slabp);
1110         if (!rpc_buffer_mempool)
1111                 goto err_nomem;
1112         return 0;
1113 err_nomem:
1114         rpc_destroy_mempool();
1115         return -ENOMEM;
1116 }