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         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 347         if (!RPC_IS_ACTIVATED(task)) {
 348                 task->tk_status = -EIO;
 349                 rpc_put_task_async(task);
 350                 return;
 351         }
 352
 353         /*
 354          * Protect the queue operations.
 355          */
 356         spin_lock_bh(&q->lock);
 357         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 358         spin_unlock_bh(&q->lock);
 359 }
 360 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 361
 362 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 363                 rpc_action action, int priority)
 364 {
 365         /* We shouldn't ever put an inactive task to sleep */
 366         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 367         if (!RPC_IS_ACTIVATED(task)) {
 368                 task->tk_status = -EIO;
 369                 rpc_put_task_async(task);
 370                 return;
 371         }
 372
 373         /*
 374          * Protect the queue operations.
 375          */
 376         spin_lock_bh(&q->lock);
 377         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 378         spin_unlock_bh(&q->lock);
 379 }
 380
 381 /**
 382  * __rpc_do_wake_up_task - wake up a single rpc_task
 383  * @queue: wait queue
 384  * @task: task to be woken up
 385  *
 386  * Caller must hold queue->lock, and have cleared the task queued flag.
 387  */
 388 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 389 {
 390         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 391                         task->tk_pid, jiffies);
 392
 393         /* Has the task been executed yet? If not, we cannot wake it up! */
 394         if (!RPC_IS_ACTIVATED(task)) {
 395                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 396                 return;
 397         }
 398
 399         trace_rpc_task_wakeup(task->tk_client, task, queue);
 400
 401         __rpc_remove_wait_queue(queue, task);
 402
 403         rpc_make_runnable(task);
 404
 405         dprintk("RPC:       __rpc_wake_up_task done\n");
 406 }
 407
 408 /*
 409  * Wake up a queued task while the queue lock is being held
 410  */
 411 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 412 {
 413         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 414                 __rpc_do_wake_up_task(queue, task);
 415 }
 416
 417 /*
 418  * Tests whether rpc queue is empty
 419  */
 420 int rpc_queue_empty(struct rpc_wait_queue *queue)
 421 {
 422         int res;
 423
 424         spin_lock_bh(&queue->lock);
 425         res = queue->qlen;
 426         spin_unlock_bh(&queue->lock);
 427         return res == 0;
 428 }
 429 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 430
 431 /*
 432  * Wake up a task on a specific queue
 433  */
 434 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 435 {
 436         spin_lock_bh(&queue->lock);
 437         rpc_wake_up_task_queue_locked(queue, task);
 438         spin_unlock_bh(&queue->lock);
 439 }
 440 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 441
 442 /*
 443  * Wake up the next task on a priority queue.
 444  */
 445 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 446 {
 447         struct list_head *q;
 448         struct rpc_task *task;
 449
 450         /*
 451          * Service a batch of tasks from a single owner.
 452          */
 453         q = &queue->tasks[queue->priority];
 454         if (!list_empty(q)) {
 455                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 456                 if (queue->owner == task->tk_owner) {
 457                         if (--queue->nr)
 458                                 goto out;
 459                         list_move_tail(&task->u.tk_wait.list, q);
 460                 }
 461                 /*
 462                  * Check if we need to switch queues.
 463                  */
 464                 if (--queue->count)
 465                         goto new_owner;
 466         }
 467
 468         /*
 469          * Service the next queue.
 470          */
 471         do {
 472                 if (q == &queue->tasks[0])
 473                         q = &queue->tasks[queue->maxpriority];
 474                 else
 475                         q = q - 1;
 476                 if (!list_empty(q)) {
 477                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 478                         goto new_queue;
 479                 }
 480         } while (q != &queue->tasks[queue->priority]);
 481
 482         rpc_reset_waitqueue_priority(queue);
 483         return NULL;
 484
 485 new_queue:
 486         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 487 new_owner:
 488         rpc_set_waitqueue_owner(queue, task->tk_owner);
 489 out:
 490         return task;
 491 }
 492
 493 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 494 {
 495         if (RPC_IS_PRIORITY(queue))
 496                 return __rpc_find_next_queued_priority(queue);
 497         if (!list_empty(&queue->tasks[0]))
 498                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 499         return NULL;
 500 }
 501
 502 /*
 503  * Wake up the first task on the wait queue.
 504  */
 505 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 506                 bool (*func)(struct rpc_task *, void *), void *data)
 507 {
 508         struct rpc_task *task = NULL;
 509
 510         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 511                         queue, rpc_qname(queue));
 512         spin_lock_bh(&queue->lock);
 513         task = __rpc_find_next_queued(queue);
 514         if (task != NULL) {
 515                 if (func(task, data))
 516                         rpc_wake_up_task_queue_locked(queue, task);
 517                 else
 518                         task = NULL;
 519         }
 520         spin_unlock_bh(&queue->lock);
 521
 522         return task;
 523 }
 524 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 525
 526 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 527 {
 528         return true;
 529 }
 530
 531 /*
 532  * Wake up the next task on the wait queue.
 533 */
 534 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 535 {
 536         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 537 }
 538 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 539
 540 /**
 541  * rpc_wake_up - wake up all rpc_tasks
 542  * @queue: rpc_wait_queue on which the tasks are sleeping
 543  *
 544  * Grabs queue->lock
 545  */
 546 void rpc_wake_up(struct rpc_wait_queue *queue)
 547 {
 548         struct list_head *head;
 549
 550         spin_lock_bh(&queue->lock);
 551         head = &queue->tasks[queue->maxpriority];
 552         for (;;) {
 553                 while (!list_empty(head)) {
 554                         struct rpc_task *task;
 555                         task = list_first_entry(head,
 556                                         struct rpc_task,
 557                                         u.tk_wait.list);
 558                         rpc_wake_up_task_queue_locked(queue, task);
 559                 }
 560                 if (head == &queue->tasks[0])
 561                         break;
 562                 head--;
 563         }
 564         spin_unlock_bh(&queue->lock);
 565 }
 566 EXPORT_SYMBOL_GPL(rpc_wake_up);
 567
 568 /**
 569  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 570  * @queue: rpc_wait_queue on which the tasks are sleeping
 571  * @status: status value to set
 572  *
 573  * Grabs queue->lock
 574  */
 575 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 576 {
 577         struct list_head *head;
 578
 579         spin_lock_bh(&queue->lock);
 580         head = &queue->tasks[queue->maxpriority];
 581         for (;;) {
 582                 while (!list_empty(head)) {
 583                         struct rpc_task *task;
 584                         task = list_first_entry(head,
 585                                         struct rpc_task,
 586                                         u.tk_wait.list);
 587                         task->tk_status = status;
 588                         rpc_wake_up_task_queue_locked(queue, task);
 589                 }
 590                 if (head == &queue->tasks[0])
 591                         break;
 592                 head--;
 593         }
 594         spin_unlock_bh(&queue->lock);
 595 }
 596 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 597
 598 static void __rpc_queue_timer_fn(unsigned long ptr)
 599 {
 600         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 601         struct rpc_task *task, *n;
 602         unsigned long expires, now, timeo;
 603
 604         spin_lock(&queue->lock);
 605         expires = now = jiffies;
 606         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 607                 timeo = task->u.tk_wait.expires;
 608                 if (time_after_eq(now, timeo)) {
 609                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 610                         task->tk_status = -ETIMEDOUT;
 611                         rpc_wake_up_task_queue_locked(queue, task);
 612                         continue;
 613                 }
 614                 if (expires == now || time_after(expires, timeo))
 615                         expires = timeo;
 616         }
 617         if (!list_empty(&queue->timer_list.list))
 618                 rpc_set_queue_timer(queue, expires);
 619         spin_unlock(&queue->lock);
 620 }
 621
 622 static void __rpc_atrun(struct rpc_task *task)
 623 {
 624         task->tk_status = 0;
 625 }
 626
 627 /*
 628  * Run a task at a later time
 629  */
 630 void rpc_delay(struct rpc_task *task, unsigned long delay)
 631 {
 632         task->tk_timeout = delay;
 633         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 634 }
 635 EXPORT_SYMBOL_GPL(rpc_delay);
 636
 637 /*
 638  * Helper to call task->tk_ops->rpc_call_prepare
 639  */
 640 void rpc_prepare_task(struct rpc_task *task)
 641 {
 642         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 643 }
 644
 645 static void
 646 rpc_init_task_statistics(struct rpc_task *task)
 647 {
 648         /* Initialize retry counters */
 649         task->tk_garb_retry = 2;
 650         task->tk_cred_retry = 2;
 651         task->tk_rebind_retry = 2;
 652
 653         /* starting timestamp */
 654         task->tk_start = ktime_get();
 655 }
 656
 657 static void
 658 rpc_reset_task_statistics(struct rpc_task *task)
 659 {
 660         task->tk_timeouts = 0;
 661         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 662
 663         rpc_init_task_statistics(task);
 664 }
 665
 666 /*
 667  * Helper that calls task->tk_ops->rpc_call_done if it exists
 668  */
 669 void rpc_exit_task(struct rpc_task *task)
 670 {
 671         task->tk_action = NULL;
 672         if (task->tk_ops->rpc_call_done != NULL) {
 673                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 674                 if (task->tk_action != NULL) {
 675                         WARN_ON(RPC_ASSASSINATED(task));
 676                         /* Always release the RPC slot and buffer memory */
 677                         xprt_release(task);
 678                         rpc_reset_task_statistics(task);
 679                 }
 680         }
 681 }
 682
 683 void rpc_exit(struct rpc_task *task, int status)
 684 {
 685         task->tk_status = status;
 686         task->tk_action = rpc_exit_task;
 687         if (RPC_IS_QUEUED(task))
 688                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 689 }
 690 EXPORT_SYMBOL_GPL(rpc_exit);
 691
 692 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 693 {
 694         if (ops->rpc_release != NULL)
 695                 ops->rpc_release(calldata);
 696 }
 697
 698 /*
 699  * This is the RPC `scheduler' (or rather, the finite state machine).
 700  */
 701 static void __rpc_execute(struct rpc_task *task)
 702 {
 703         struct rpc_wait_queue *queue;
 704         int task_is_async = RPC_IS_ASYNC(task);
 705         int status = 0;
 706
 707         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 708                         task->tk_pid, task->tk_flags);
 709
 710         BUG_ON(RPC_IS_QUEUED(task));
 711
 712         for (;;) {
 713                 void (*do_action)(struct rpc_task *);
 714
 715                 /*
 716                  * Execute any pending callback first.
 717                  */
 718                 do_action = task->tk_callback;
 719                 task->tk_callback = NULL;
 720                 if (do_action == NULL) {
 721                         /*
 722                          * Perform the next FSM step.
 723                          * tk_action may be NULL if the task has been killed.
 724                          * In particular, note that rpc_killall_tasks may
 725                          * do this at any time, so beware when dereferencing.
 726                          */
 727                         do_action = task->tk_action;
 728                         if (do_action == NULL)
 729                                 break;
 730                 }
 731                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 732                 do_action(task);
 733
 734                 /*
 735                  * Lockless check for whether task is sleeping or not.
 736                  */
 737                 if (!RPC_IS_QUEUED(task))
 738                         continue;
 739                 /*
 740                  * The queue->lock protects against races with
 741                  * rpc_make_runnable().
 742                  *
 743                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 744                  * rpc_task, rpc_make_runnable() can assign it to a
 745                  * different workqueue. We therefore cannot assume that the
 746                  * rpc_task pointer may still be dereferenced.
 747                  */
 748                 queue = task->tk_waitqueue;
 749                 spin_lock_bh(&queue->lock);
 750                 if (!RPC_IS_QUEUED(task)) {
 751                         spin_unlock_bh(&queue->lock);
 752                         continue;
 753                 }
 754                 rpc_clear_running(task);
 755                 spin_unlock_bh(&queue->lock);
 756                 if (task_is_async)
 757                         return;
 758
 759                 /* sync task: sleep here */
 760                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 761                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 762                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 763                                 TASK_KILLABLE);
 764                 if (status == -ERESTARTSYS) {
 765                         /*
 766                          * When a sync task receives a signal, it exits with
 767                          * -ERESTARTSYS. In order to catch any callbacks that
 768                          * clean up after sleeping on some queue, we don't
 769                          * break the loop here, but go around once more.
 770                          */
 771                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 772                         task->tk_flags |= RPC_TASK_KILLED;
 773                         rpc_exit(task, -ERESTARTSYS);
 774                 }
 775                 rpc_set_running(task);
 776                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 777         }
 778
 779         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 780                         task->tk_status);
 781         /* Release all resources associated with the task */
 782         rpc_release_task(task);
 783 }
 784
 785 /*
 786  * User-visible entry point to the scheduler.
 787  *
 788  * This may be called recursively if e.g. an async NFS task updates
 789  * the attributes and finds that dirty pages must be flushed.
 790  * NOTE: Upon exit of this function the task is guaranteed to be
 791  *       released. In particular note that tk_release() will have
 792  *       been called, so your task memory may have been freed.
 793  */
 794 void rpc_execute(struct rpc_task *task)
 795 {
 796         rpc_set_active(task);
 797         rpc_make_runnable(task);
 798         if (!RPC_IS_ASYNC(task))
 799                 __rpc_execute(task);
 800 }
 801
 802 static void rpc_async_schedule(struct work_struct *work)
 803 {
 804         current->flags |= PF_FSTRANS;
 805         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 806         current->flags &= ~PF_FSTRANS;
 807 }
 808
 809 /**
 810  * rpc_malloc - allocate an RPC buffer
 811  * @task: RPC task that will use this buffer
 812  * @size: requested byte size
 813  *
 814  * To prevent rpciod from hanging, this allocator never sleeps,
 815  * returning NULL if the request cannot be serviced immediately.
 816  * The caller can arrange to sleep in a way that is safe for rpciod.
 817  *
 818  * Most requests are 'small' (under 2KiB) and can be serviced from a
 819  * mempool, ensuring that NFS reads and writes can always proceed,
 820  * and that there is good locality of reference for these buffers.
 821  *
 822  * In order to avoid memory starvation triggering more writebacks of
 823  * NFS requests, we avoid using GFP_KERNEL.
 824  */
 825 void *rpc_malloc(struct rpc_task *task, size_t size)
 826 {
 827         struct rpc_buffer *buf;
 828         gfp_t gfp = GFP_NOWAIT;
 829
 830         if (RPC_IS_SWAPPER(task))
 831                 gfp |= __GFP_MEMALLOC;
 832
 833         size += sizeof(struct rpc_buffer);
 834         if (size <= RPC_BUFFER_MAXSIZE)
 835                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 836         else
 837                 buf = kmalloc(size, gfp);
 838
 839         if (!buf)
 840                 return NULL;
 841
 842         buf->len = size;
 843         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 844                         task->tk_pid, size, buf);
 845         return &buf->data;
 846 }
 847 EXPORT_SYMBOL_GPL(rpc_malloc);
 848
 849 /**
 850  * rpc_free - free buffer allocated via rpc_malloc
 851  * @buffer: buffer to free
 852  *
 853  */
 854 void rpc_free(void *buffer)
 855 {
 856         size_t size;
 857         struct rpc_buffer *buf;
 858
 859         if (!buffer)
 860                 return;
 861
 862         buf = container_of(buffer, struct rpc_buffer, data);
 863         size = buf->len;
 864
 865         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 866                         size, buf);
 867
 868         if (size <= RPC_BUFFER_MAXSIZE)
 869                 mempool_free(buf, rpc_buffer_mempool);
 870         else
 871                 kfree(buf);
 872 }
 873 EXPORT_SYMBOL_GPL(rpc_free);
 874
 875 /*
 876  * Creation and deletion of RPC task structures
 877  */
 878 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 879 {
 880         memset(task, 0, sizeof(*task));
 881         atomic_set(&task->tk_count, 1);
 882         task->tk_flags  = task_setup_data->flags;
 883         task->tk_ops = task_setup_data->callback_ops;
 884         task->tk_calldata = task_setup_data->callback_data;
 885         INIT_LIST_HEAD(&task->tk_task);
 886
 887         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 888         task->tk_owner = current->tgid;
 889
 890         /* Initialize workqueue for async tasks */
 891         task->tk_workqueue = task_setup_data->workqueue;
 892
 893         if (task->tk_ops->rpc_call_prepare != NULL)
 894                 task->tk_action = rpc_prepare_task;
 895
 896         rpc_init_task_statistics(task);
 897
 898         dprintk("RPC:       new task initialized, procpid %u\n",
 899                                 task_pid_nr(current));
 900 }
 901
 902 static struct rpc_task *
 903 rpc_alloc_task(void)
 904 {
 905         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 906 }
 907
 908 /*
 909  * Create a new task for the specified client.
 910  */
 911 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 912 {
 913         struct rpc_task *task = setup_data->task;
 914         unsigned short flags = 0;
 915
 916         if (task == NULL) {
 917                 task = rpc_alloc_task();
 918                 if (task == NULL) {
 919                         rpc_release_calldata(setup_data->callback_ops,
 920                                         setup_data->callback_data);
 921                         return ERR_PTR(-ENOMEM);
 922                 }
 923                 flags = RPC_TASK_DYNAMIC;
 924         }
 925
 926         rpc_init_task(task, setup_data);
 927         task->tk_flags |= flags;
 928         dprintk("RPC:       allocated task %p\n", task);
 929         return task;
 930 }
 931
 932 static void rpc_free_task(struct rpc_task *task)
 933 {
 934         const struct rpc_call_ops *tk_ops = task->tk_ops;
 935         void *calldata = task->tk_calldata;
 936
 937         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 938                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 939                 mempool_free(task, rpc_task_mempool);
 940         }
 941         rpc_release_calldata(tk_ops, calldata);
 942 }
 943
 944 static void rpc_async_release(struct work_struct *work)
 945 {
 946         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 947 }
 948
 949 static void rpc_release_resources_task(struct rpc_task *task)
 950 {
 951         if (task->tk_rqstp)
 952                 xprt_release(task);
 953         if (task->tk_msg.rpc_cred) {
 954                 put_rpccred(task->tk_msg.rpc_cred);
 955                 task->tk_msg.rpc_cred = NULL;
 956         }
 957         rpc_task_release_client(task);
 958 }
 959
 960 static void rpc_final_put_task(struct rpc_task *task,
 961                 struct workqueue_struct *q)
 962 {
 963         if (q != NULL) {
 964                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 965                 queue_work(q, &task->u.tk_work);
 966         } else
 967                 rpc_free_task(task);
 968 }
 969
 970 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
 971 {
 972         if (atomic_dec_and_test(&task->tk_count)) {
 973                 rpc_release_resources_task(task);
 974                 rpc_final_put_task(task, q);
 975         }
 976 }
 977
 978 void rpc_put_task(struct rpc_task *task)
 979 {
 980         rpc_do_put_task(task, NULL);
 981 }
 982 EXPORT_SYMBOL_GPL(rpc_put_task);
 983
 984 void rpc_put_task_async(struct rpc_task *task)
 985 {
 986         rpc_do_put_task(task, task->tk_workqueue);
 987 }
 988 EXPORT_SYMBOL_GPL(rpc_put_task_async);
 989
 990 static void rpc_release_task(struct rpc_task *task)
 991 {
 992         dprintk("RPC: %5u release task\n", task->tk_pid);
 993
 994         BUG_ON (RPC_IS_QUEUED(task));
 995
 996         rpc_release_resources_task(task);
 997
 998         /*
 999          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1000          * so it should be safe to use task->tk_count as a test for whether
1001          * or not any other processes still hold references to our rpc_task.
1002          */
1003         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1004                 /* Wake up anyone who may be waiting for task completion */
1005                 if (!rpc_complete_task(task))
1006                         return;
1007         } else {
1008                 if (!atomic_dec_and_test(&task->tk_count))
1009                         return;
1010         }
1011         rpc_final_put_task(task, task->tk_workqueue);
1012 }
1013
1014 int rpciod_up(void)
1015 {
1016         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1017 }
1018
1019 void rpciod_down(void)
1020 {
1021         module_put(THIS_MODULE);
1022 }
1023
1024 /*
1025  * Start up the rpciod workqueue.
1026  */
1027 static int rpciod_start(void)
1028 {
1029         struct workqueue_struct *wq;
1030
1031         /*
1032          * Create the rpciod thread and wait for it to start.
1033          */
1034         dprintk("RPC:       creating workqueue rpciod\n");
1035         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 1);
1036         rpciod_workqueue = wq;
1037         return rpciod_workqueue != NULL;
1038 }
1039
1040 static void rpciod_stop(void)
1041 {
1042         struct workqueue_struct *wq = NULL;
1043
1044         if (rpciod_workqueue == NULL)
1045                 return;
1046         dprintk("RPC:       destroying workqueue rpciod\n");
1047
1048         wq = rpciod_workqueue;
1049         rpciod_workqueue = NULL;
1050         destroy_workqueue(wq);
1051 }
1052
1053 void
1054 rpc_destroy_mempool(void)
1055 {
1056         rpciod_stop();
1057         if (rpc_buffer_mempool)
1058                 mempool_destroy(rpc_buffer_mempool);
1059         if (rpc_task_mempool)
1060                 mempool_destroy(rpc_task_mempool);
1061         if (rpc_task_slabp)
1062                 kmem_cache_destroy(rpc_task_slabp);
1063         if (rpc_buffer_slabp)
1064                 kmem_cache_destroy(rpc_buffer_slabp);
1065         rpc_destroy_wait_queue(&delay_queue);
1066 }
1067
1068 int
1069 rpc_init_mempool(void)
1070 {
1071         /*
1072          * The following is not strictly a mempool initialisation,
1073          * but there is no harm in doing it here
1074          */
1075         rpc_init_wait_queue(&delay_queue, "delayq");
1076         if (!rpciod_start())
1077                 goto err_nomem;
1078
1079         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1080                                              sizeof(struct rpc_task),
1081                                              0, SLAB_HWCACHE_ALIGN,
1082                                              NULL);
1083         if (!rpc_task_slabp)
1084                 goto err_nomem;
1085         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1086                                              RPC_BUFFER_MAXSIZE,
1087                                              0, SLAB_HWCACHE_ALIGN,
1088                                              NULL);
1089         if (!rpc_buffer_slabp)
1090                 goto err_nomem;
1091         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1092                                                     rpc_task_slabp);
1093         if (!rpc_task_mempool)
1094                 goto err_nomem;
1095         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1096                                                       rpc_buffer_slabp);
1097         if (!rpc_buffer_mempool)
1098                 goto err_nomem;
1099         return 0;
1100 err_nomem:
1101         rpc_destroy_mempool();
1102         return -ENOMEM;
1103 }