net/sunrpc/sched.c

   1 /*
   2  * linux/net/sunrpc/sched.c
   3  *
   4  * Scheduling for synchronous and asynchronous RPC requests.
   5  *
   6  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  10  */
  11
  12 #include <linux/module.h>
  13
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/slab.h>
  17 #include <linux/mempool.h>
  18 #include <linux/smp.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/mutex.h>
  21
  22 #include <linux/sunrpc/clnt.h>
  23
  24 #ifdef RPC_DEBUG
  25 #define RPCDBG_FACILITY         RPCDBG_SCHED
  26 #define RPC_TASK_MAGIC_ID       0xf00baa
  27 #endif
  28
  29 /*
  30  * RPC slabs and memory pools
  31  */
  32 #define RPC_BUFFER_MAXSIZE      (2048)
  33 #define RPC_BUFFER_POOLSIZE     (8)
  34 #define RPC_TASK_POOLSIZE       (8)
  35 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  36 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  37 static mempool_t        *rpc_task_mempool __read_mostly;
  38 static mempool_t        *rpc_buffer_mempool __read_mostly;
  39
  40 static void                     rpc_async_schedule(struct work_struct *);
  41 static void                      rpc_release_task(struct rpc_task *task);
  42 static void __rpc_queue_timer_fn(unsigned long ptr);
  43
  44 /*
  45  * RPC tasks sit here while waiting for conditions to improve.
  46  */
  47 static struct rpc_wait_queue delay_queue;
  48
  49 /*
  50  * rpciod-related stuff
  51  */
  52 struct workqueue_struct *rpciod_workqueue;
  53
  54 /*
  55  * Disable the timer for a given RPC task. Should be called with
  56  * queue->lock and bh_disabled in order to avoid races within
  57  * rpc_run_timer().
  58  */
  59 static void
  60 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  61 {
  62         if (task->tk_timeout == 0)
  63                 return;
  64         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  65         task->tk_timeout = 0;
  66         list_del(&task->u.tk_wait.timer_list);
  67         if (list_empty(&queue->timer_list.list))
  68                 del_timer(&queue->timer_list.timer);
  69 }
  70
  71 static void
  72 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  73 {
  74         queue->timer_list.expires = expires;
  75         mod_timer(&queue->timer_list.timer, expires);
  76 }
  77
  78 /*
  79  * Set up a timer for the current task.
  80  */
  81 static void
  82 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  83 {
  84         if (!task->tk_timeout)
  85                 return;
  86
  87         dprintk("RPC: %5u setting alarm for %lu ms\n",
  88                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  89
  90         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  91         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  92                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  93         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  94 }
  95
  96 /*
  97  * Add new request to a priority queue.
  98  */
  99 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 100 {
 101         struct list_head *q;
 102         struct rpc_task *t;
 103
 104         INIT_LIST_HEAD(&task->u.tk_wait.links);
 105         q = &queue->tasks[task->tk_priority];
 106         if (unlikely(task->tk_priority > queue->maxpriority))
 107                 q = &queue->tasks[queue->maxpriority];
 108         list_for_each_entry(t, q, u.tk_wait.list) {
 109                 if (t->tk_owner == task->tk_owner) {
 110                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 111                         return;
 112                 }
 113         }
 114         list_add_tail(&task->u.tk_wait.list, q);
 115 }
 116
 117 /*
 118  * Add new request to wait queue.
 119  *
 120  * Swapper tasks always get inserted at the head of the queue.
 121  * This should avoid many nasty memory deadlocks and hopefully
 122  * improve overall performance.
 123  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 124  */
 125 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 126 {
 127         BUG_ON (RPC_IS_QUEUED(task));
 128
 129         if (RPC_IS_PRIORITY(queue))
 130                 __rpc_add_wait_queue_priority(queue, task);
 131         else if (RPC_IS_SWAPPER(task))
 132                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 133         else
 134                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 135         task->tk_waitqueue = queue;
 136         queue->qlen++;
 137         rpc_set_queued(task);
 138
 139         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 140                         task->tk_pid, queue, rpc_qname(queue));
 141 }
 142
 143 /*
 144  * Remove request from a priority queue.
 145  */
 146 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 147 {
 148         struct rpc_task *t;
 149
 150         if (!list_empty(&task->u.tk_wait.links)) {
 151                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 152                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 153                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 154         }
 155 }
 156
 157 /*
 158  * Remove request from queue.
 159  * Note: must be called with spin lock held.
 160  */
 161 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 162 {
 163         __rpc_disable_timer(queue, task);
 164         if (RPC_IS_PRIORITY(queue))
 165                 __rpc_remove_wait_queue_priority(task);
 166         list_del(&task->u.tk_wait.list);
 167         queue->qlen--;
 168         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 169                         task->tk_pid, queue, rpc_qname(queue));
 170 }
 171
 172 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 173 {
 174         queue->priority = priority;
 175         queue->count = 1 << (priority * 2);
 176 }
 177
 178 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 179 {
 180         queue->owner = pid;
 181         queue->nr = RPC_BATCH_COUNT;
 182 }
 183
 184 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 185 {
 186         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 187         rpc_set_waitqueue_owner(queue, 0);
 188 }
 189
 190 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 191 {
 192         int i;
 193
 194         spin_lock_init(&queue->lock);
 195         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 196                 INIT_LIST_HEAD(&queue->tasks[i]);
 197         queue->maxpriority = nr_queues - 1;
 198         rpc_reset_waitqueue_priority(queue);
 199         queue->qlen = 0;
 200         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 201         INIT_LIST_HEAD(&queue->timer_list.list);
 202 #ifdef RPC_DEBUG
 203         queue->name = qname;
 204 #endif
 205 }
 206
 207 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 208 {
 209         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 210 }
 211
 212 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 213 {
 214         __rpc_init_priority_wait_queue(queue, qname, 1);
 215 }
 216 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 217
 218 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 219 {
 220         del_timer_sync(&queue->timer_list.timer);
 221 }
 222 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 223
 224 static int rpc_wait_bit_killable(void *word)
 225 {
 226         if (fatal_signal_pending(current))
 227                 return -ERESTARTSYS;
 228         schedule();
 229         return 0;
 230 }
 231
 232 #ifdef RPC_DEBUG
 233 static void rpc_task_set_debuginfo(struct rpc_task *task)
 234 {
 235         static atomic_t rpc_pid;
 236
 237         task->tk_magic = RPC_TASK_MAGIC_ID;
 238         task->tk_pid = atomic_inc_return(&rpc_pid);
 239 }
 240 #else
 241 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 242 {
 243 }
 244 #endif
 245
 246 static void rpc_set_active(struct rpc_task *task)
 247 {
 248         struct rpc_clnt *clnt;
 249         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 250                 return;
 251         rpc_task_set_debuginfo(task);
 252         /* Add to global list of all tasks */
 253         clnt = task->tk_client;
 254         if (clnt != NULL) {
 255                 spin_lock(&clnt->cl_lock);
 256                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 257                 spin_unlock(&clnt->cl_lock);
 258         }
 259 }
 260
 261 /*
 262  * Mark an RPC call as having completed by clearing the 'active' bit
 263  */
 264 static void rpc_mark_complete_task(struct rpc_task *task)
 265 {
 266         smp_mb__before_clear_bit();
 267         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 268         smp_mb__after_clear_bit();
 269         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 270 }
 271
 272 /*
 273  * Allow callers to wait for completion of an RPC call
 274  */
 275 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 276 {
 277         if (action == NULL)
 278                 action = rpc_wait_bit_killable;
 279         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 280                         action, TASK_KILLABLE);
 281 }
 282 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 283
 284 /*
 285  * Make an RPC task runnable.
 286  *
 287  * Note: If the task is ASYNC, this must be called with
 288  * the spinlock held to protect the wait queue operation.
 289  */
 290 static void rpc_make_runnable(struct rpc_task *task)
 291 {
 292         rpc_clear_queued(task);
 293         if (rpc_test_and_set_running(task))
 294                 return;
 295         if (RPC_IS_ASYNC(task)) {
 296                 int status;
 297
 298                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 299                 status = queue_work(rpciod_workqueue, &task->u.tk_work);
 300                 if (status < 0) {
 301                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 302                         task->tk_status = status;
 303                         return;
 304                 }
 305         } else
 306                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 307 }
 308
 309 /*
 310  * Prepare for sleeping on a wait queue.
 311  * By always appending tasks to the list we ensure FIFO behavior.
 312  * NB: An RPC task will only receive interrupt-driven events as long
 313  * as it's on a wait queue.
 314  */
 315 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 316                         rpc_action action)
 317 {
 318         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 319                         task->tk_pid, rpc_qname(q), jiffies);
 320
 321         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 322                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 323                 return;
 324         }
 325
 326         __rpc_add_wait_queue(q, task);
 327
 328         BUG_ON(task->tk_callback != NULL);
 329         task->tk_callback = action;
 330         __rpc_add_timer(q, task);
 331 }
 332
 333 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 334                                 rpc_action action)
 335 {
 336         /* Mark the task as being activated if so needed */
 337         rpc_set_active(task);
 338
 339         /*
 340          * Protect the queue operations.
 341          */
 342         spin_lock_bh(&q->lock);
 343         __rpc_sleep_on(q, task, action);
 344         spin_unlock_bh(&q->lock);
 345 }
 346 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 347
 348 /**
 349  * __rpc_do_wake_up_task - wake up a single rpc_task
 350  * @queue: wait queue
 351  * @task: task to be woken up
 352  *
 353  * Caller must hold queue->lock, and have cleared the task queued flag.
 354  */
 355 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 356 {
 357         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 358                         task->tk_pid, jiffies);
 359
 360 #ifdef RPC_DEBUG
 361         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 362 #endif
 363         /* Has the task been executed yet? If not, we cannot wake it up! */
 364         if (!RPC_IS_ACTIVATED(task)) {
 365                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 366                 return;
 367         }
 368
 369         __rpc_remove_wait_queue(queue, task);
 370
 371         rpc_make_runnable(task);
 372
 373         dprintk("RPC:       __rpc_wake_up_task done\n");
 374 }
 375
 376 /*
 377  * Wake up a queued task while the queue lock is being held
 378  */
 379 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 380 {
 381         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 382                 __rpc_do_wake_up_task(queue, task);
 383 }
 384
 385 /*
 386  * Wake up a task on a specific queue
 387  */
 388 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 389 {
 390         spin_lock_bh(&queue->lock);
 391         rpc_wake_up_task_queue_locked(queue, task);
 392         spin_unlock_bh(&queue->lock);
 393 }
 394 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 395
 396 /*
 397  * Wake up the specified task
 398  */
 399 static void rpc_wake_up_task(struct rpc_task *task)
 400 {
 401         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 402 }
 403
 404 /*
 405  * Wake up the next task on a priority queue.
 406  */
 407 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 408 {
 409         struct list_head *q;
 410         struct rpc_task *task;
 411
 412         /*
 413          * Service a batch of tasks from a single owner.
 414          */
 415         q = &queue->tasks[queue->priority];
 416         if (!list_empty(q)) {
 417                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 418                 if (queue->owner == task->tk_owner) {
 419                         if (--queue->nr)
 420                                 goto out;
 421                         list_move_tail(&task->u.tk_wait.list, q);
 422                 }
 423                 /*
 424                  * Check if we need to switch queues.
 425                  */
 426                 if (--queue->count)
 427                         goto new_owner;
 428         }
 429
 430         /*
 431          * Service the next queue.
 432          */
 433         do {
 434                 if (q == &queue->tasks[0])
 435                         q = &queue->tasks[queue->maxpriority];
 436                 else
 437                         q = q - 1;
 438                 if (!list_empty(q)) {
 439                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 440                         goto new_queue;
 441                 }
 442         } while (q != &queue->tasks[queue->priority]);
 443
 444         rpc_reset_waitqueue_priority(queue);
 445         return NULL;
 446
 447 new_queue:
 448         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 449 new_owner:
 450         rpc_set_waitqueue_owner(queue, task->tk_owner);
 451 out:
 452         rpc_wake_up_task_queue_locked(queue, task);
 453         return task;
 454 }
 455
 456 /*
 457  * Wake up the next task on the wait queue.
 458  */
 459 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 460 {
 461         struct rpc_task *task = NULL;
 462
 463         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 464                         queue, rpc_qname(queue));
 465         spin_lock_bh(&queue->lock);
 466         if (RPC_IS_PRIORITY(queue))
 467                 task = __rpc_wake_up_next_priority(queue);
 468         else {
 469                 task_for_first(task, &queue->tasks[0])
 470                         rpc_wake_up_task_queue_locked(queue, task);
 471         }
 472         spin_unlock_bh(&queue->lock);
 473
 474         return task;
 475 }
 476 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 477
 478 /**
 479  * rpc_wake_up - wake up all rpc_tasks
 480  * @queue: rpc_wait_queue on which the tasks are sleeping
 481  *
 482  * Grabs queue->lock
 483  */
 484 void rpc_wake_up(struct rpc_wait_queue *queue)
 485 {
 486         struct rpc_task *task, *next;
 487         struct list_head *head;
 488
 489         spin_lock_bh(&queue->lock);
 490         head = &queue->tasks[queue->maxpriority];
 491         for (;;) {
 492                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 493                         rpc_wake_up_task_queue_locked(queue, task);
 494                 if (head == &queue->tasks[0])
 495                         break;
 496                 head--;
 497         }
 498         spin_unlock_bh(&queue->lock);
 499 }
 500 EXPORT_SYMBOL_GPL(rpc_wake_up);
 501
 502 /**
 503  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 504  * @queue: rpc_wait_queue on which the tasks are sleeping
 505  * @status: status value to set
 506  *
 507  * Grabs queue->lock
 508  */
 509 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 510 {
 511         struct rpc_task *task, *next;
 512         struct list_head *head;
 513
 514         spin_lock_bh(&queue->lock);
 515         head = &queue->tasks[queue->maxpriority];
 516         for (;;) {
 517                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 518                         task->tk_status = status;
 519                         rpc_wake_up_task_queue_locked(queue, task);
 520                 }
 521                 if (head == &queue->tasks[0])
 522                         break;
 523                 head--;
 524         }
 525         spin_unlock_bh(&queue->lock);
 526 }
 527 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 528
 529 static void __rpc_queue_timer_fn(unsigned long ptr)
 530 {
 531         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 532         struct rpc_task *task, *n;
 533         unsigned long expires, now, timeo;
 534
 535         spin_lock(&queue->lock);
 536         expires = now = jiffies;
 537         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 538                 timeo = task->u.tk_wait.expires;
 539                 if (time_after_eq(now, timeo)) {
 540                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 541                         task->tk_status = -ETIMEDOUT;
 542                         rpc_wake_up_task_queue_locked(queue, task);
 543                         continue;
 544                 }
 545                 if (expires == now || time_after(expires, timeo))
 546                         expires = timeo;
 547         }
 548         if (!list_empty(&queue->timer_list.list))
 549                 rpc_set_queue_timer(queue, expires);
 550         spin_unlock(&queue->lock);
 551 }
 552
 553 static void __rpc_atrun(struct rpc_task *task)
 554 {
 555         task->tk_status = 0;
 556 }
 557
 558 /*
 559  * Run a task at a later time
 560  */
 561 void rpc_delay(struct rpc_task *task, unsigned long delay)
 562 {
 563         task->tk_timeout = delay;
 564         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 565 }
 566 EXPORT_SYMBOL_GPL(rpc_delay);
 567
 568 /*
 569  * Helper to call task->tk_ops->rpc_call_prepare
 570  */
 571 void rpc_prepare_task(struct rpc_task *task)
 572 {
 573         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 574 }
 575
 576 /*
 577  * Helper that calls task->tk_ops->rpc_call_done if it exists
 578  */
 579 void rpc_exit_task(struct rpc_task *task)
 580 {
 581         task->tk_action = NULL;
 582         if (task->tk_ops->rpc_call_done != NULL) {
 583                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 584                 if (task->tk_action != NULL) {
 585                         WARN_ON(RPC_ASSASSINATED(task));
 586                         /* Always release the RPC slot and buffer memory */
 587                         xprt_release(task);
 588                 }
 589         }
 590 }
 591 EXPORT_SYMBOL_GPL(rpc_exit_task);
 592
 593 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 594 {
 595         if (ops->rpc_release != NULL)
 596                 ops->rpc_release(calldata);
 597 }
 598
 599 /*
 600  * This is the RPC `scheduler' (or rather, the finite state machine).
 601  */
 602 static void __rpc_execute(struct rpc_task *task)
 603 {
 604         struct rpc_wait_queue *queue;
 605         int task_is_async = RPC_IS_ASYNC(task);
 606         int status = 0;
 607
 608         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 609                         task->tk_pid, task->tk_flags);
 610
 611         BUG_ON(RPC_IS_QUEUED(task));
 612
 613         for (;;) {
 614
 615                 /*
 616                  * Execute any pending callback.
 617                  */
 618                 if (task->tk_callback) {
 619                         void (*save_callback)(struct rpc_task *);
 620
 621                         /*
 622                          * We set tk_callback to NULL before calling it,
 623                          * in case it sets the tk_callback field itself:
 624                          */
 625                         save_callback = task->tk_callback;
 626                         task->tk_callback = NULL;
 627                         save_callback(task);
 628                 }
 629
 630                 /*
 631                  * Perform the next FSM step.
 632                  * tk_action may be NULL when the task has been killed
 633                  * by someone else.
 634                  */
 635                 if (!RPC_IS_QUEUED(task)) {
 636                         if (task->tk_action == NULL)
 637                                 break;
 638                         task->tk_action(task);
 639                 }
 640
 641                 /*
 642                  * Lockless check for whether task is sleeping or not.
 643                  */
 644                 if (!RPC_IS_QUEUED(task))
 645                         continue;
 646                 /*
 647                  * The queue->lock protects against races with
 648                  * rpc_make_runnable().
 649                  *
 650                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 651                  * rpc_task, rpc_make_runnable() can assign it to a
 652                  * different workqueue. We therefore cannot assume that the
 653                  * rpc_task pointer may still be dereferenced.
 654                  */
 655                 queue = task->tk_waitqueue;
 656                 spin_lock_bh(&queue->lock);
 657                 if (!RPC_IS_QUEUED(task)) {
 658                         spin_unlock_bh(&queue->lock);
 659                         continue;
 660                 }
 661                 rpc_clear_running(task);
 662                 spin_unlock_bh(&queue->lock);
 663                 if (task_is_async)
 664                         return;
 665
 666                 /* sync task: sleep here */
 667                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 668                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 669                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 670                                 TASK_KILLABLE);
 671                 if (status == -ERESTARTSYS) {
 672                         /*
 673                          * When a sync task receives a signal, it exits with
 674                          * -ERESTARTSYS. In order to catch any callbacks that
 675                          * clean up after sleeping on some queue, we don't
 676                          * break the loop here, but go around once more.
 677                          */
 678                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 679                         task->tk_flags |= RPC_TASK_KILLED;
 680                         rpc_exit(task, -ERESTARTSYS);
 681                         rpc_wake_up_task(task);
 682                 }
 683                 rpc_set_running(task);
 684                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 685         }
 686
 687         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 688                         task->tk_status);
 689         /* Release all resources associated with the task */
 690         rpc_release_task(task);
 691 }
 692
 693 /*
 694  * User-visible entry point to the scheduler.
 695  *
 696  * This may be called recursively if e.g. an async NFS task updates
 697  * the attributes and finds that dirty pages must be flushed.
 698  * NOTE: Upon exit of this function the task is guaranteed to be
 699  *       released. In particular note that tk_release() will have
 700  *       been called, so your task memory may have been freed.
 701  */
 702 void rpc_execute(struct rpc_task *task)
 703 {
 704         rpc_set_active(task);
 705         rpc_set_running(task);
 706         __rpc_execute(task);
 707 }
 708
 709 static void rpc_async_schedule(struct work_struct *work)
 710 {
 711         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 712 }
 713
 714 struct rpc_buffer {
 715         size_t  len;
 716         char    data[];
 717 };
 718
 719 /**
 720  * rpc_malloc - allocate an RPC buffer
 721  * @task: RPC task that will use this buffer
 722  * @size: requested byte size
 723  *
 724  * To prevent rpciod from hanging, this allocator never sleeps,
 725  * returning NULL if the request cannot be serviced immediately.
 726  * The caller can arrange to sleep in a way that is safe for rpciod.
 727  *
 728  * Most requests are 'small' (under 2KiB) and can be serviced from a
 729  * mempool, ensuring that NFS reads and writes can always proceed,
 730  * and that there is good locality of reference for these buffers.
 731  *
 732  * In order to avoid memory starvation triggering more writebacks of
 733  * NFS requests, we avoid using GFP_KERNEL.
 734  */
 735 void *rpc_malloc(struct rpc_task *task, size_t size)
 736 {
 737         struct rpc_buffer *buf;
 738         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 739
 740         size += sizeof(struct rpc_buffer);
 741         if (size <= RPC_BUFFER_MAXSIZE)
 742                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 743         else
 744                 buf = kmalloc(size, gfp);
 745
 746         if (!buf)
 747                 return NULL;
 748
 749         buf->len = size;
 750         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 751                         task->tk_pid, size, buf);
 752         return &buf->data;
 753 }
 754 EXPORT_SYMBOL_GPL(rpc_malloc);
 755
 756 /**
 757  * rpc_free - free buffer allocated via rpc_malloc
 758  * @buffer: buffer to free
 759  *
 760  */
 761 void rpc_free(void *buffer)
 762 {
 763         size_t size;
 764         struct rpc_buffer *buf;
 765
 766         if (!buffer)
 767                 return;
 768
 769         buf = container_of(buffer, struct rpc_buffer, data);
 770         size = buf->len;
 771
 772         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 773                         size, buf);
 774
 775         if (size <= RPC_BUFFER_MAXSIZE)
 776                 mempool_free(buf, rpc_buffer_mempool);
 777         else
 778                 kfree(buf);
 779 }
 780 EXPORT_SYMBOL_GPL(rpc_free);
 781
 782 /*
 783  * Creation and deletion of RPC task structures
 784  */
 785 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 786 {
 787         memset(task, 0, sizeof(*task));
 788         atomic_set(&task->tk_count, 1);
 789         task->tk_flags  = task_setup_data->flags;
 790         task->tk_ops = task_setup_data->callback_ops;
 791         task->tk_calldata = task_setup_data->callback_data;
 792         INIT_LIST_HEAD(&task->tk_task);
 793
 794         /* Initialize retry counters */
 795         task->tk_garb_retry = 2;
 796         task->tk_cred_retry = 2;
 797
 798         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 799         task->tk_owner = current->tgid;
 800
 801         /* Initialize workqueue for async tasks */
 802         task->tk_workqueue = task_setup_data->workqueue;
 803
 804         task->tk_client = task_setup_data->rpc_client;
 805         if (task->tk_client != NULL) {
 806                 kref_get(&task->tk_client->cl_kref);
 807                 if (task->tk_client->cl_softrtry)
 808                         task->tk_flags |= RPC_TASK_SOFT;
 809         }
 810
 811         if (task->tk_ops->rpc_call_prepare != NULL)
 812                 task->tk_action = rpc_prepare_task;
 813
 814         if (task_setup_data->rpc_message != NULL) {
 815                 task->tk_msg.rpc_proc = task_setup_data->rpc_message->rpc_proc;
 816                 task->tk_msg.rpc_argp = task_setup_data->rpc_message->rpc_argp;
 817                 task->tk_msg.rpc_resp = task_setup_data->rpc_message->rpc_resp;
 818                 /* Bind the user cred */
 819                 rpcauth_bindcred(task, task_setup_data->rpc_message->rpc_cred, task_setup_data->flags);
 820                 if (task->tk_action == NULL)
 821                         rpc_call_start(task);
 822         }
 823
 824         /* starting timestamp */
 825         task->tk_start = jiffies;
 826
 827         dprintk("RPC:       new task initialized, procpid %u\n",
 828                                 task_pid_nr(current));
 829 }
 830
 831 static struct rpc_task *
 832 rpc_alloc_task(void)
 833 {
 834         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 835 }
 836
 837 /*
 838  * Create a new task for the specified client.
 839  */
 840 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 841 {
 842         struct rpc_task *task = setup_data->task;
 843         unsigned short flags = 0;
 844
 845         if (task == NULL) {
 846                 task = rpc_alloc_task();
 847                 if (task == NULL)
 848                         goto out;
 849                 flags = RPC_TASK_DYNAMIC;
 850         }
 851
 852         rpc_init_task(task, setup_data);
 853
 854         task->tk_flags |= flags;
 855         dprintk("RPC:       allocated task %p\n", task);
 856 out:
 857         return task;
 858 }
 859
 860 static void rpc_free_task(struct rpc_task *task)
 861 {
 862         const struct rpc_call_ops *tk_ops = task->tk_ops;
 863         void *calldata = task->tk_calldata;
 864
 865         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 866                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 867                 mempool_free(task, rpc_task_mempool);
 868         }
 869         rpc_release_calldata(tk_ops, calldata);
 870 }
 871
 872 static void rpc_async_release(struct work_struct *work)
 873 {
 874         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 875 }
 876
 877 void rpc_put_task(struct rpc_task *task)
 878 {
 879         if (!atomic_dec_and_test(&task->tk_count))
 880                 return;
 881         /* Release resources */
 882         if (task->tk_rqstp)
 883                 xprt_release(task);
 884         if (task->tk_msg.rpc_cred)
 885                 rpcauth_unbindcred(task);
 886         if (task->tk_client) {
 887                 rpc_release_client(task->tk_client);
 888                 task->tk_client = NULL;
 889         }
 890         if (task->tk_workqueue != NULL) {
 891                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 892                 queue_work(task->tk_workqueue, &task->u.tk_work);
 893         } else
 894                 rpc_free_task(task);
 895 }
 896 EXPORT_SYMBOL_GPL(rpc_put_task);
 897
 898 static void rpc_release_task(struct rpc_task *task)
 899 {
 900 #ifdef RPC_DEBUG
 901         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 902 #endif
 903         dprintk("RPC: %5u release task\n", task->tk_pid);
 904
 905         if (!list_empty(&task->tk_task)) {
 906                 struct rpc_clnt *clnt = task->tk_client;
 907                 /* Remove from client task list */
 908                 spin_lock(&clnt->cl_lock);
 909                 list_del(&task->tk_task);
 910                 spin_unlock(&clnt->cl_lock);
 911         }
 912         BUG_ON (RPC_IS_QUEUED(task));
 913
 914 #ifdef RPC_DEBUG
 915         task->tk_magic = 0;
 916 #endif
 917         /* Wake up anyone who is waiting for task completion */
 918         rpc_mark_complete_task(task);
 919
 920         rpc_put_task(task);
 921 }
 922
 923 /*
 924  * Kill all tasks for the given client.
 925  * XXX: kill their descendants as well?
 926  */
 927 void rpc_killall_tasks(struct rpc_clnt *clnt)
 928 {
 929         struct rpc_task *rovr;
 930
 931
 932         if (list_empty(&clnt->cl_tasks))
 933                 return;
 934         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 935         /*
 936          * Spin lock all_tasks to prevent changes...
 937          */
 938         spin_lock(&clnt->cl_lock);
 939         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 940                 if (! RPC_IS_ACTIVATED(rovr))
 941                         continue;
 942                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 943                         rovr->tk_flags |= RPC_TASK_KILLED;
 944                         rpc_exit(rovr, -EIO);
 945                         rpc_wake_up_task(rovr);
 946                 }
 947         }
 948         spin_unlock(&clnt->cl_lock);
 949 }
 950 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 951
 952 int rpciod_up(void)
 953 {
 954         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 955 }
 956
 957 void rpciod_down(void)
 958 {
 959         module_put(THIS_MODULE);
 960 }
 961
 962 /*
 963  * Start up the rpciod workqueue.
 964  */
 965 static int rpciod_start(void)
 966 {
 967         struct workqueue_struct *wq;
 968
 969         /*
 970          * Create the rpciod thread and wait for it to start.
 971          */
 972         dprintk("RPC:       creating workqueue rpciod\n");
 973         wq = create_workqueue("rpciod");
 974         rpciod_workqueue = wq;
 975         return rpciod_workqueue != NULL;
 976 }
 977
 978 static void rpciod_stop(void)
 979 {
 980         struct workqueue_struct *wq = NULL;
 981
 982         if (rpciod_workqueue == NULL)
 983                 return;
 984         dprintk("RPC:       destroying workqueue rpciod\n");
 985
 986         wq = rpciod_workqueue;
 987         rpciod_workqueue = NULL;
 988         destroy_workqueue(wq);
 989 }
 990
 991 void
 992 rpc_destroy_mempool(void)
 993 {
 994         rpciod_stop();
 995         if (rpc_buffer_mempool)
 996                 mempool_destroy(rpc_buffer_mempool);
 997         if (rpc_task_mempool)
 998                 mempool_destroy(rpc_task_mempool);
 999         if (rpc_task_slabp)
1000                 kmem_cache_destroy(rpc_task_slabp);
1001         if (rpc_buffer_slabp)
1002                 kmem_cache_destroy(rpc_buffer_slabp);
1003         rpc_destroy_wait_queue(&delay_queue);
1004 }
1005
1006 int
1007 rpc_init_mempool(void)
1008 {
1009         /*
1010          * The following is not strictly a mempool initialisation,
1011          * but there is no harm in doing it here
1012          */
1013         rpc_init_wait_queue(&delay_queue, "delayq");
1014         if (!rpciod_start())
1015                 goto err_nomem;
1016
1017         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1018                                              sizeof(struct rpc_task),
1019                                              0, SLAB_HWCACHE_ALIGN,
1020                                              NULL);
1021         if (!rpc_task_slabp)
1022                 goto err_nomem;
1023         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1024                                              RPC_BUFFER_MAXSIZE,
1025                                              0, SLAB_HWCACHE_ALIGN,
1026                                              NULL);
1027         if (!rpc_buffer_slabp)
1028                 goto err_nomem;
1029         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1030                                                     rpc_task_slabp);
1031         if (!rpc_task_mempool)
1032                 goto err_nomem;
1033         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1034                                                       rpc_buffer_slabp);
1035         if (!rpc_buffer_mempool)
1036                 goto err_nomem;
1037         return 0;
1038 err_nomem:
1039         rpc_destroy_mempool();
1040         return -ENOMEM;
1041 }