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/smp_lock.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/mutex.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #ifdef RPC_DEBUG
  26 #define RPCDBG_FACILITY         RPCDBG_SCHED
  27 #define RPC_TASK_MAGIC_ID       0xf00baa
  28 #endif
  29
  30 /*
  31  * RPC slabs and memory pools
  32  */
  33 #define RPC_BUFFER_MAXSIZE      (2048)
  34 #define RPC_BUFFER_POOLSIZE     (8)
  35 #define RPC_TASK_POOLSIZE       (8)
  36 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  37 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  38 static mempool_t        *rpc_task_mempool __read_mostly;
  39 static mempool_t        *rpc_buffer_mempool __read_mostly;
  40
  41 static void                     rpc_async_schedule(struct work_struct *);
  42 static void                      rpc_release_task(struct rpc_task *task);
  43 static void __rpc_queue_timer_fn(unsigned long ptr);
  44
  45 /*
  46  * RPC tasks sit here while waiting for conditions to improve.
  47  */
  48 static struct rpc_wait_queue delay_queue;
  49
  50 /*
  51  * rpciod-related stuff
  52  */
  53 struct workqueue_struct *rpciod_workqueue;
  54
  55 /*
  56  * Disable the timer for a given RPC task. Should be called with
  57  * queue->lock and bh_disabled in order to avoid races within
  58  * rpc_run_timer().
  59  */
  60 static void
  61 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  62 {
  63         if (task->tk_timeout == 0)
  64                 return;
  65         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  66         task->tk_timeout = 0;
  67         list_del(&task->u.tk_wait.timer_list);
  68         if (list_empty(&queue->timer_list.list))
  69                 del_timer(&queue->timer_list.timer);
  70 }
  71
  72 static void
  73 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  74 {
  75         queue->timer_list.expires = expires;
  76         mod_timer(&queue->timer_list.timer, expires);
  77 }
  78
  79 /*
  80  * Set up a timer for the current task.
  81  */
  82 static void
  83 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  84 {
  85         if (!task->tk_timeout)
  86                 return;
  87
  88         dprintk("RPC: %5u setting alarm for %lu ms\n",
  89                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  90
  91         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  92         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  93                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  94         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  95 }
  96
  97 /*
  98  * Add new request to a priority queue.
  99  */
 100 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 101 {
 102         struct list_head *q;
 103         struct rpc_task *t;
 104
 105         INIT_LIST_HEAD(&task->u.tk_wait.links);
 106         q = &queue->tasks[task->tk_priority];
 107         if (unlikely(task->tk_priority > queue->maxpriority))
 108                 q = &queue->tasks[queue->maxpriority];
 109         list_for_each_entry(t, q, u.tk_wait.list) {
 110                 if (t->tk_owner == task->tk_owner) {
 111                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 112                         return;
 113                 }
 114         }
 115         list_add_tail(&task->u.tk_wait.list, q);
 116 }
 117
 118 /*
 119  * Add new request to wait queue.
 120  *
 121  * Swapper tasks always get inserted at the head of the queue.
 122  * This should avoid many nasty memory deadlocks and hopefully
 123  * improve overall performance.
 124  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 125  */
 126 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 127 {
 128         BUG_ON (RPC_IS_QUEUED(task));
 129
 130         if (RPC_IS_PRIORITY(queue))
 131                 __rpc_add_wait_queue_priority(queue, task);
 132         else if (RPC_IS_SWAPPER(task))
 133                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 134         else
 135                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 136         task->tk_waitqueue = queue;
 137         queue->qlen++;
 138         rpc_set_queued(task);
 139
 140         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 141                         task->tk_pid, queue, rpc_qname(queue));
 142 }
 143
 144 /*
 145  * Remove request from a priority queue.
 146  */
 147 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 148 {
 149         struct rpc_task *t;
 150
 151         if (!list_empty(&task->u.tk_wait.links)) {
 152                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 153                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 154                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 155         }
 156 }
 157
 158 /*
 159  * Remove request from queue.
 160  * Note: must be called with spin lock held.
 161  */
 162 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 163 {
 164         __rpc_disable_timer(queue, task);
 165         if (RPC_IS_PRIORITY(queue))
 166                 __rpc_remove_wait_queue_priority(task);
 167         list_del(&task->u.tk_wait.list);
 168         queue->qlen--;
 169         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 170                         task->tk_pid, queue, rpc_qname(queue));
 171 }
 172
 173 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 174 {
 175         queue->priority = priority;
 176         queue->count = 1 << (priority * 2);
 177 }
 178
 179 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 180 {
 181         queue->owner = pid;
 182         queue->nr = RPC_BATCH_COUNT;
 183 }
 184
 185 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 186 {
 187         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 188         rpc_set_waitqueue_owner(queue, 0);
 189 }
 190
 191 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 192 {
 193         int i;
 194
 195         spin_lock_init(&queue->lock);
 196         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 197                 INIT_LIST_HEAD(&queue->tasks[i]);
 198         queue->maxpriority = nr_queues - 1;
 199         rpc_reset_waitqueue_priority(queue);
 200         queue->qlen = 0;
 201         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 202         INIT_LIST_HEAD(&queue->timer_list.list);
 203 #ifdef RPC_DEBUG
 204         queue->name = qname;
 205 #endif
 206 }
 207
 208 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 209 {
 210         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 211 }
 212
 213 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 214 {
 215         __rpc_init_priority_wait_queue(queue, qname, 1);
 216 }
 217 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 218
 219 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 220 {
 221         del_timer_sync(&queue->timer_list.timer);
 222 }
 223 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 224
 225 static int rpc_wait_bit_killable(void *word)
 226 {
 227         if (fatal_signal_pending(current))
 228                 return -ERESTARTSYS;
 229         schedule();
 230         return 0;
 231 }
 232
 233 #ifdef RPC_DEBUG
 234 static void rpc_task_set_debuginfo(struct rpc_task *task)
 235 {
 236         static atomic_t rpc_pid;
 237
 238         task->tk_magic = RPC_TASK_MAGIC_ID;
 239         task->tk_pid = atomic_inc_return(&rpc_pid);
 240 }
 241 #else
 242 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 243 {
 244 }
 245 #endif
 246
 247 static void rpc_set_active(struct rpc_task *task)
 248 {
 249         struct rpc_clnt *clnt;
 250         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 251                 return;
 252         rpc_task_set_debuginfo(task);
 253         /* Add to global list of all tasks */
 254         clnt = task->tk_client;
 255         if (clnt != NULL) {
 256                 spin_lock(&clnt->cl_lock);
 257                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 258                 spin_unlock(&clnt->cl_lock);
 259         }
 260 }
 261
 262 /*
 263  * Mark an RPC call as having completed by clearing the 'active' bit
 264  */
 265 static void rpc_mark_complete_task(struct rpc_task *task)
 266 {
 267         smp_mb__before_clear_bit();
 268         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 269         smp_mb__after_clear_bit();
 270         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 271 }
 272
 273 /*
 274  * Allow callers to wait for completion of an RPC call
 275  */
 276 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 277 {
 278         if (action == NULL)
 279                 action = rpc_wait_bit_killable;
 280         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 281                         action, TASK_KILLABLE);
 282 }
 283 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 284
 285 /*
 286  * Make an RPC task runnable.
 287  *
 288  * Note: If the task is ASYNC, this must be called with
 289  * the spinlock held to protect the wait queue operation.
 290  */
 291 static void rpc_make_runnable(struct rpc_task *task)
 292 {
 293         rpc_clear_queued(task);
 294         if (rpc_test_and_set_running(task))
 295                 return;
 296         /* We might have raced */
 297         if (RPC_IS_QUEUED(task)) {
 298                 rpc_clear_running(task);
 299                 return;
 300         }
 301         if (RPC_IS_ASYNC(task)) {
 302                 int status;
 303
 304                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 305                 status = queue_work(rpciod_workqueue, &task->u.tk_work);
 306                 if (status < 0) {
 307                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 308                         task->tk_status = status;
 309                         return;
 310                 }
 311         } else
 312                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 313 }
 314
 315 /*
 316  * Prepare for sleeping on a wait queue.
 317  * By always appending tasks to the list we ensure FIFO behavior.
 318  * NB: An RPC task will only receive interrupt-driven events as long
 319  * as it's on a wait queue.
 320  */
 321 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 322                         rpc_action action)
 323 {
 324         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 325                         task->tk_pid, rpc_qname(q), jiffies);
 326
 327         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 328                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 329                 return;
 330         }
 331
 332         __rpc_add_wait_queue(q, task);
 333
 334         BUG_ON(task->tk_callback != NULL);
 335         task->tk_callback = action;
 336         __rpc_add_timer(q, task);
 337 }
 338
 339 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 340                                 rpc_action action)
 341 {
 342         /* Mark the task as being activated if so needed */
 343         rpc_set_active(task);
 344
 345         /*
 346          * Protect the queue operations.
 347          */
 348         spin_lock_bh(&q->lock);
 349         __rpc_sleep_on(q, task, action);
 350         spin_unlock_bh(&q->lock);
 351 }
 352 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 353
 354 /**
 355  * __rpc_do_wake_up_task - wake up a single rpc_task
 356  * @queue: wait queue
 357  * @task: task to be woken up
 358  *
 359  * Caller must hold queue->lock, and have cleared the task queued flag.
 360  */
 361 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 362 {
 363         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 364                         task->tk_pid, jiffies);
 365
 366 #ifdef RPC_DEBUG
 367         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 368 #endif
 369         /* Has the task been executed yet? If not, we cannot wake it up! */
 370         if (!RPC_IS_ACTIVATED(task)) {
 371                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 372                 return;
 373         }
 374
 375         __rpc_remove_wait_queue(queue, task);
 376
 377         rpc_make_runnable(task);
 378
 379         dprintk("RPC:       __rpc_wake_up_task done\n");
 380 }
 381
 382 /*
 383  * Wake up a queued task while the queue lock is being held
 384  */
 385 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 386 {
 387         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 388                 __rpc_do_wake_up_task(queue, task);
 389 }
 390
 391 /*
 392  * Wake up a task on a specific queue
 393  */
 394 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 395 {
 396         spin_lock_bh(&queue->lock);
 397         rpc_wake_up_task_queue_locked(queue, task);
 398         spin_unlock_bh(&queue->lock);
 399 }
 400 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 401
 402 /*
 403  * Wake up the specified task
 404  */
 405 static void rpc_wake_up_task(struct rpc_task *task)
 406 {
 407         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 408 }
 409
 410 /*
 411  * Wake up the next task on a priority queue.
 412  */
 413 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 414 {
 415         struct list_head *q;
 416         struct rpc_task *task;
 417
 418         /*
 419          * Service a batch of tasks from a single owner.
 420          */
 421         q = &queue->tasks[queue->priority];
 422         if (!list_empty(q)) {
 423                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 424                 if (queue->owner == task->tk_owner) {
 425                         if (--queue->nr)
 426                                 goto out;
 427                         list_move_tail(&task->u.tk_wait.list, q);
 428                 }
 429                 /*
 430                  * Check if we need to switch queues.
 431                  */
 432                 if (--queue->count)
 433                         goto new_owner;
 434         }
 435
 436         /*
 437          * Service the next queue.
 438          */
 439         do {
 440                 if (q == &queue->tasks[0])
 441                         q = &queue->tasks[queue->maxpriority];
 442                 else
 443                         q = q - 1;
 444                 if (!list_empty(q)) {
 445                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 446                         goto new_queue;
 447                 }
 448         } while (q != &queue->tasks[queue->priority]);
 449
 450         rpc_reset_waitqueue_priority(queue);
 451         return NULL;
 452
 453 new_queue:
 454         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 455 new_owner:
 456         rpc_set_waitqueue_owner(queue, task->tk_owner);
 457 out:
 458         rpc_wake_up_task_queue_locked(queue, task);
 459         return task;
 460 }
 461
 462 /*
 463  * Wake up the next task on the wait queue.
 464  */
 465 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 466 {
 467         struct rpc_task *task = NULL;
 468
 469         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 470                         queue, rpc_qname(queue));
 471         spin_lock_bh(&queue->lock);
 472         if (RPC_IS_PRIORITY(queue))
 473                 task = __rpc_wake_up_next_priority(queue);
 474         else {
 475                 task_for_first(task, &queue->tasks[0])
 476                         rpc_wake_up_task_queue_locked(queue, task);
 477         }
 478         spin_unlock_bh(&queue->lock);
 479
 480         return task;
 481 }
 482 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 483
 484 /**
 485  * rpc_wake_up - wake up all rpc_tasks
 486  * @queue: rpc_wait_queue on which the tasks are sleeping
 487  *
 488  * Grabs queue->lock
 489  */
 490 void rpc_wake_up(struct rpc_wait_queue *queue)
 491 {
 492         struct rpc_task *task, *next;
 493         struct list_head *head;
 494
 495         spin_lock_bh(&queue->lock);
 496         head = &queue->tasks[queue->maxpriority];
 497         for (;;) {
 498                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 499                         rpc_wake_up_task_queue_locked(queue, task);
 500                 if (head == &queue->tasks[0])
 501                         break;
 502                 head--;
 503         }
 504         spin_unlock_bh(&queue->lock);
 505 }
 506 EXPORT_SYMBOL_GPL(rpc_wake_up);
 507
 508 /**
 509  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 510  * @queue: rpc_wait_queue on which the tasks are sleeping
 511  * @status: status value to set
 512  *
 513  * Grabs queue->lock
 514  */
 515 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 516 {
 517         struct rpc_task *task, *next;
 518         struct list_head *head;
 519
 520         spin_lock_bh(&queue->lock);
 521         head = &queue->tasks[queue->maxpriority];
 522         for (;;) {
 523                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 524                         task->tk_status = status;
 525                         rpc_wake_up_task_queue_locked(queue, task);
 526                 }
 527                 if (head == &queue->tasks[0])
 528                         break;
 529                 head--;
 530         }
 531         spin_unlock_bh(&queue->lock);
 532 }
 533 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 534
 535 static void __rpc_queue_timer_fn(unsigned long ptr)
 536 {
 537         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 538         struct rpc_task *task, *n;
 539         unsigned long expires, now, timeo;
 540
 541         spin_lock(&queue->lock);
 542         expires = now = jiffies;
 543         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 544                 timeo = task->u.tk_wait.expires;
 545                 if (time_after_eq(now, timeo)) {
 546                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 547                         task->tk_status = -ETIMEDOUT;
 548                         rpc_wake_up_task_queue_locked(queue, task);
 549                         continue;
 550                 }
 551                 if (expires == now || time_after(expires, timeo))
 552                         expires = timeo;
 553         }
 554         if (!list_empty(&queue->timer_list.list))
 555                 rpc_set_queue_timer(queue, expires);
 556         spin_unlock(&queue->lock);
 557 }
 558
 559 static void __rpc_atrun(struct rpc_task *task)
 560 {
 561         task->tk_status = 0;
 562 }
 563
 564 /*
 565  * Run a task at a later time
 566  */
 567 void rpc_delay(struct rpc_task *task, unsigned long delay)
 568 {
 569         task->tk_timeout = delay;
 570         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 571 }
 572 EXPORT_SYMBOL_GPL(rpc_delay);
 573
 574 /*
 575  * Helper to call task->tk_ops->rpc_call_prepare
 576  */
 577 static void rpc_prepare_task(struct rpc_task *task)
 578 {
 579         lock_kernel();
 580         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 581         unlock_kernel();
 582 }
 583
 584 /*
 585  * Helper that calls task->tk_ops->rpc_call_done if it exists
 586  */
 587 void rpc_exit_task(struct rpc_task *task)
 588 {
 589         task->tk_action = NULL;
 590         if (task->tk_ops->rpc_call_done != NULL) {
 591                 lock_kernel();
 592                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 593                 unlock_kernel();
 594                 if (task->tk_action != NULL) {
 595                         WARN_ON(RPC_ASSASSINATED(task));
 596                         /* Always release the RPC slot and buffer memory */
 597                         xprt_release(task);
 598                 }
 599         }
 600 }
 601 EXPORT_SYMBOL_GPL(rpc_exit_task);
 602
 603 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 604 {
 605         if (ops->rpc_release != NULL) {
 606                 lock_kernel();
 607                 ops->rpc_release(calldata);
 608                 unlock_kernel();
 609         }
 610 }
 611
 612 /*
 613  * This is the RPC `scheduler' (or rather, the finite state machine).
 614  */
 615 static void __rpc_execute(struct rpc_task *task)
 616 {
 617         int             status = 0;
 618
 619         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 620                         task->tk_pid, task->tk_flags);
 621
 622         BUG_ON(RPC_IS_QUEUED(task));
 623
 624         for (;;) {
 625
 626                 /*
 627                  * Execute any pending callback.
 628                  */
 629                 if (RPC_DO_CALLBACK(task)) {
 630                         /* Define a callback save pointer */
 631                         void (*save_callback)(struct rpc_task *);
 632
 633                         /*
 634                          * If a callback exists, save it, reset it,
 635                          * call it.
 636                          * The save is needed to stop from resetting
 637                          * another callback set within the callback handler
 638                          * - Dave
 639                          */
 640                         save_callback=task->tk_callback;
 641                         task->tk_callback=NULL;
 642                         save_callback(task);
 643                 }
 644
 645                 /*
 646                  * Perform the next FSM step.
 647                  * tk_action may be NULL when the task has been killed
 648                  * by someone else.
 649                  */
 650                 if (!RPC_IS_QUEUED(task)) {
 651                         if (task->tk_action == NULL)
 652                                 break;
 653                         task->tk_action(task);
 654                 }
 655
 656                 /*
 657                  * Lockless check for whether task is sleeping or not.
 658                  */
 659                 if (!RPC_IS_QUEUED(task))
 660                         continue;
 661                 rpc_clear_running(task);
 662                 if (RPC_IS_ASYNC(task)) {
 663                         /* Careful! we may have raced... */
 664                         if (RPC_IS_QUEUED(task))
 665                                 return;
 666                         if (rpc_test_and_set_running(task))
 667                                 return;
 668                         continue;
 669                 }
 670
 671                 /* sync task: sleep here */
 672                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 673                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 674                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 675                                 TASK_KILLABLE);
 676                 if (status == -ERESTARTSYS) {
 677                         /*
 678                          * When a sync task receives a signal, it exits with
 679                          * -ERESTARTSYS. In order to catch any callbacks that
 680                          * clean up after sleeping on some queue, we don't
 681                          * break the loop here, but go around once more.
 682                          */
 683                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 684                         task->tk_flags |= RPC_TASK_KILLED;
 685                         rpc_exit(task, -ERESTARTSYS);
 686                         rpc_wake_up_task(task);
 687                 }
 688                 rpc_set_running(task);
 689                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 690         }
 691
 692         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 693                         task->tk_status);
 694         /* Release all resources associated with the task */
 695         rpc_release_task(task);
 696 }
 697
 698 /*
 699  * User-visible entry point to the scheduler.
 700  *
 701  * This may be called recursively if e.g. an async NFS task updates
 702  * the attributes and finds that dirty pages must be flushed.
 703  * NOTE: Upon exit of this function the task is guaranteed to be
 704  *       released. In particular note that tk_release() will have
 705  *       been called, so your task memory may have been freed.
 706  */
 707 void rpc_execute(struct rpc_task *task)
 708 {
 709         rpc_set_active(task);
 710         rpc_set_running(task);
 711         __rpc_execute(task);
 712 }
 713
 714 static void rpc_async_schedule(struct work_struct *work)
 715 {
 716         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 717 }
 718
 719 struct rpc_buffer {
 720         size_t  len;
 721         char    data[];
 722 };
 723
 724 /**
 725  * rpc_malloc - allocate an RPC buffer
 726  * @task: RPC task that will use this buffer
 727  * @size: requested byte size
 728  *
 729  * To prevent rpciod from hanging, this allocator never sleeps,
 730  * returning NULL if the request cannot be serviced immediately.
 731  * The caller can arrange to sleep in a way that is safe for rpciod.
 732  *
 733  * Most requests are 'small' (under 2KiB) and can be serviced from a
 734  * mempool, ensuring that NFS reads and writes can always proceed,
 735  * and that there is good locality of reference for these buffers.
 736  *
 737  * In order to avoid memory starvation triggering more writebacks of
 738  * NFS requests, we avoid using GFP_KERNEL.
 739  */
 740 void *rpc_malloc(struct rpc_task *task, size_t size)
 741 {
 742         struct rpc_buffer *buf;
 743         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 744
 745         size += sizeof(struct rpc_buffer);
 746         if (size <= RPC_BUFFER_MAXSIZE)
 747                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 748         else
 749                 buf = kmalloc(size, gfp);
 750
 751         if (!buf)
 752                 return NULL;
 753
 754         buf->len = size;
 755         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 756                         task->tk_pid, size, buf);
 757         return &buf->data;
 758 }
 759 EXPORT_SYMBOL_GPL(rpc_malloc);
 760
 761 /**
 762  * rpc_free - free buffer allocated via rpc_malloc
 763  * @buffer: buffer to free
 764  *
 765  */
 766 void rpc_free(void *buffer)
 767 {
 768         size_t size;
 769         struct rpc_buffer *buf;
 770
 771         if (!buffer)
 772                 return;
 773
 774         buf = container_of(buffer, struct rpc_buffer, data);
 775         size = buf->len;
 776
 777         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 778                         size, buf);
 779
 780         if (size <= RPC_BUFFER_MAXSIZE)
 781                 mempool_free(buf, rpc_buffer_mempool);
 782         else
 783                 kfree(buf);
 784 }
 785 EXPORT_SYMBOL_GPL(rpc_free);
 786
 787 /*
 788  * Creation and deletion of RPC task structures
 789  */
 790 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 791 {
 792         memset(task, 0, sizeof(*task));
 793         atomic_set(&task->tk_count, 1);
 794         task->tk_flags  = task_setup_data->flags;
 795         task->tk_ops = task_setup_data->callback_ops;
 796         task->tk_calldata = task_setup_data->callback_data;
 797         INIT_LIST_HEAD(&task->tk_task);
 798
 799         /* Initialize retry counters */
 800         task->tk_garb_retry = 2;
 801         task->tk_cred_retry = 2;
 802
 803         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 804         task->tk_owner = current->tgid;
 805
 806         /* Initialize workqueue for async tasks */
 807         task->tk_workqueue = task_setup_data->workqueue;
 808
 809         task->tk_client = task_setup_data->rpc_client;
 810         if (task->tk_client != NULL) {
 811                 kref_get(&task->tk_client->cl_kref);
 812                 if (task->tk_client->cl_softrtry)
 813                         task->tk_flags |= RPC_TASK_SOFT;
 814         }
 815
 816         if (task->tk_ops->rpc_call_prepare != NULL)
 817                 task->tk_action = rpc_prepare_task;
 818
 819         if (task_setup_data->rpc_message != NULL) {
 820                 task->tk_msg.rpc_proc = task_setup_data->rpc_message->rpc_proc;
 821                 task->tk_msg.rpc_argp = task_setup_data->rpc_message->rpc_argp;
 822                 task->tk_msg.rpc_resp = task_setup_data->rpc_message->rpc_resp;
 823                 /* Bind the user cred */
 824                 rpcauth_bindcred(task, task_setup_data->rpc_message->rpc_cred, task_setup_data->flags);
 825                 if (task->tk_action == NULL)
 826                         rpc_call_start(task);
 827         }
 828
 829         /* starting timestamp */
 830         task->tk_start = jiffies;
 831
 832         dprintk("RPC:       new task initialized, procpid %u\n",
 833                                 task_pid_nr(current));
 834 }
 835
 836 static struct rpc_task *
 837 rpc_alloc_task(void)
 838 {
 839         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 840 }
 841
 842 /*
 843  * Create a new task for the specified client.
 844  */
 845 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 846 {
 847         struct rpc_task *task = setup_data->task;
 848         unsigned short flags = 0;
 849
 850         if (task == NULL) {
 851                 task = rpc_alloc_task();
 852                 if (task == NULL)
 853                         goto out;
 854                 flags = RPC_TASK_DYNAMIC;
 855         }
 856
 857         rpc_init_task(task, setup_data);
 858
 859         task->tk_flags |= flags;
 860         dprintk("RPC:       allocated task %p\n", task);
 861 out:
 862         return task;
 863 }
 864
 865 static void rpc_free_task(struct rpc_task *task)
 866 {
 867         const struct rpc_call_ops *tk_ops = task->tk_ops;
 868         void *calldata = task->tk_calldata;
 869
 870         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 871                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 872                 mempool_free(task, rpc_task_mempool);
 873         }
 874         rpc_release_calldata(tk_ops, calldata);
 875 }
 876
 877 static void rpc_async_release(struct work_struct *work)
 878 {
 879         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 880 }
 881
 882 void rpc_put_task(struct rpc_task *task)
 883 {
 884         if (!atomic_dec_and_test(&task->tk_count))
 885                 return;
 886         /* Release resources */
 887         if (task->tk_rqstp)
 888                 xprt_release(task);
 889         if (task->tk_msg.rpc_cred)
 890                 rpcauth_unbindcred(task);
 891         if (task->tk_client) {
 892                 rpc_release_client(task->tk_client);
 893                 task->tk_client = NULL;
 894         }
 895         if (task->tk_workqueue != NULL) {
 896                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 897                 queue_work(task->tk_workqueue, &task->u.tk_work);
 898         } else
 899                 rpc_free_task(task);
 900 }
 901 EXPORT_SYMBOL_GPL(rpc_put_task);
 902
 903 static void rpc_release_task(struct rpc_task *task)
 904 {
 905 #ifdef RPC_DEBUG
 906         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 907 #endif
 908         dprintk("RPC: %5u release task\n", task->tk_pid);
 909
 910         if (!list_empty(&task->tk_task)) {
 911                 struct rpc_clnt *clnt = task->tk_client;
 912                 /* Remove from client task list */
 913                 spin_lock(&clnt->cl_lock);
 914                 list_del(&task->tk_task);
 915                 spin_unlock(&clnt->cl_lock);
 916         }
 917         BUG_ON (RPC_IS_QUEUED(task));
 918
 919 #ifdef RPC_DEBUG
 920         task->tk_magic = 0;
 921 #endif
 922         /* Wake up anyone who is waiting for task completion */
 923         rpc_mark_complete_task(task);
 924
 925         rpc_put_task(task);
 926 }
 927
 928 /*
 929  * Kill all tasks for the given client.
 930  * XXX: kill their descendants as well?
 931  */
 932 void rpc_killall_tasks(struct rpc_clnt *clnt)
 933 {
 934         struct rpc_task *rovr;
 935
 936
 937         if (list_empty(&clnt->cl_tasks))
 938                 return;
 939         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 940         /*
 941          * Spin lock all_tasks to prevent changes...
 942          */
 943         spin_lock(&clnt->cl_lock);
 944         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 945                 if (! RPC_IS_ACTIVATED(rovr))
 946                         continue;
 947                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 948                         rovr->tk_flags |= RPC_TASK_KILLED;
 949                         rpc_exit(rovr, -EIO);
 950                         rpc_wake_up_task(rovr);
 951                 }
 952         }
 953         spin_unlock(&clnt->cl_lock);
 954 }
 955 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 956
 957 int rpciod_up(void)
 958 {
 959         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 960 }
 961
 962 void rpciod_down(void)
 963 {
 964         module_put(THIS_MODULE);
 965 }
 966
 967 /*
 968  * Start up the rpciod workqueue.
 969  */
 970 static int rpciod_start(void)
 971 {
 972         struct workqueue_struct *wq;
 973
 974         /*
 975          * Create the rpciod thread and wait for it to start.
 976          */
 977         dprintk("RPC:       creating workqueue rpciod\n");
 978         wq = create_workqueue("rpciod");
 979         rpciod_workqueue = wq;
 980         return rpciod_workqueue != NULL;
 981 }
 982
 983 static void rpciod_stop(void)
 984 {
 985         struct workqueue_struct *wq = NULL;
 986
 987         if (rpciod_workqueue == NULL)
 988                 return;
 989         dprintk("RPC:       destroying workqueue rpciod\n");
 990
 991         wq = rpciod_workqueue;
 992         rpciod_workqueue = NULL;
 993         destroy_workqueue(wq);
 994 }
 995
 996 void
 997 rpc_destroy_mempool(void)
 998 {
 999         rpciod_stop();
1000         if (rpc_buffer_mempool)
1001                 mempool_destroy(rpc_buffer_mempool);
1002         if (rpc_task_mempool)
1003                 mempool_destroy(rpc_task_mempool);
1004         if (rpc_task_slabp)
1005                 kmem_cache_destroy(rpc_task_slabp);
1006         if (rpc_buffer_slabp)
1007                 kmem_cache_destroy(rpc_buffer_slabp);
1008         rpc_destroy_wait_queue(&delay_queue);
1009 }
1010
1011 int
1012 rpc_init_mempool(void)
1013 {
1014         /*
1015          * The following is not strictly a mempool initialisation,
1016          * but there is no harm in doing it here
1017          */
1018         rpc_init_wait_queue(&delay_queue, "delayq");
1019         if (!rpciod_start())
1020                 goto err_nomem;
1021
1022         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1023                                              sizeof(struct rpc_task),
1024                                              0, SLAB_HWCACHE_ALIGN,
1025                                              NULL);
1026         if (!rpc_task_slabp)
1027                 goto err_nomem;
1028         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1029                                              RPC_BUFFER_MAXSIZE,
1030                                              0, SLAB_HWCACHE_ALIGN,
1031                                              NULL);
1032         if (!rpc_buffer_slabp)
1033                 goto err_nomem;
1034         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1035                                                     rpc_task_slabp);
1036         if (!rpc_task_mempool)
1037                 goto err_nomem;
1038         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1039                                                       rpc_buffer_slabp);
1040         if (!rpc_buffer_mempool)
1041                 goto err_nomem;
1042         return 0;
1043 err_nomem:
1044         rpc_destroy_mempool();
1045         return -ENOMEM;
1046 }