SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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 #include "sunrpc.h"
  25
  26 #ifdef RPC_DEBUG
  27 #define RPCDBG_FACILITY         RPCDBG_SCHED
  28 #define RPC_TASK_MAGIC_ID       0xf00baa
  29 #endif
  30
  31 /*
  32  * RPC slabs and memory pools
  33  */
  34 #define RPC_BUFFER_MAXSIZE      (2048)
  35 #define RPC_BUFFER_POOLSIZE     (8)
  36 #define RPC_TASK_POOLSIZE       (8)
  37 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  38 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  39 static mempool_t        *rpc_task_mempool __read_mostly;
  40 static mempool_t        *rpc_buffer_mempool __read_mostly;
  41
  42 static void                     rpc_async_schedule(struct work_struct *);
  43 static void                      rpc_release_task(struct rpc_task *task);
  44 static void __rpc_queue_timer_fn(unsigned long ptr);
  45
  46 /*
  47  * RPC tasks sit here while waiting for conditions to improve.
  48  */
  49 static struct rpc_wait_queue delay_queue;
  50
  51 /*
  52  * rpciod-related stuff
  53  */
  54 struct workqueue_struct *rpciod_workqueue;
  55
  56 /*
  57  * Disable the timer for a given RPC task. Should be called with
  58  * queue->lock and bh_disabled in order to avoid races within
  59  * rpc_run_timer().
  60  */
  61 static void
  62 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  63 {
  64         if (task->tk_timeout == 0)
  65                 return;
  66         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  67         task->tk_timeout = 0;
  68         list_del(&task->u.tk_wait.timer_list);
  69         if (list_empty(&queue->timer_list.list))
  70                 del_timer(&queue->timer_list.timer);
  71 }
  72
  73 static void
  74 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  75 {
  76         queue->timer_list.expires = expires;
  77         mod_timer(&queue->timer_list.timer, expires);
  78 }
  79
  80 /*
  81  * Set up a timer for the current task.
  82  */
  83 static void
  84 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  85 {
  86         if (!task->tk_timeout)
  87                 return;
  88
  89         dprintk("RPC: %5u setting alarm for %lu ms\n",
  90                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  91
  92         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  93         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  94                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  95         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  96 }
  97
  98 /*
  99  * Add new request to a priority queue.
 100  */
 101 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 102 {
 103         struct list_head *q;
 104         struct rpc_task *t;
 105
 106         INIT_LIST_HEAD(&task->u.tk_wait.links);
 107         q = &queue->tasks[task->tk_priority];
 108         if (unlikely(task->tk_priority > queue->maxpriority))
 109                 q = &queue->tasks[queue->maxpriority];
 110         list_for_each_entry(t, q, u.tk_wait.list) {
 111                 if (t->tk_owner == task->tk_owner) {
 112                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 113                         return;
 114                 }
 115         }
 116         list_add_tail(&task->u.tk_wait.list, q);
 117 }
 118
 119 /*
 120  * Add new request to wait queue.
 121  *
 122  * Swapper tasks always get inserted at the head of the queue.
 123  * This should avoid many nasty memory deadlocks and hopefully
 124  * improve overall performance.
 125  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 126  */
 127 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 128 {
 129         BUG_ON (RPC_IS_QUEUED(task));
 130
 131         if (RPC_IS_PRIORITY(queue))
 132                 __rpc_add_wait_queue_priority(queue, task);
 133         else if (RPC_IS_SWAPPER(task))
 134                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 135         else
 136                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 137         task->tk_waitqueue = queue;
 138         queue->qlen++;
 139         rpc_set_queued(task);
 140
 141         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 142                         task->tk_pid, queue, rpc_qname(queue));
 143 }
 144
 145 /*
 146  * Remove request from a priority queue.
 147  */
 148 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 149 {
 150         struct rpc_task *t;
 151
 152         if (!list_empty(&task->u.tk_wait.links)) {
 153                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 154                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 155                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 156         }
 157 }
 158
 159 /*
 160  * Remove request from queue.
 161  * Note: must be called with spin lock held.
 162  */
 163 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 164 {
 165         __rpc_disable_timer(queue, task);
 166         if (RPC_IS_PRIORITY(queue))
 167                 __rpc_remove_wait_queue_priority(task);
 168         list_del(&task->u.tk_wait.list);
 169         queue->qlen--;
 170         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 171                         task->tk_pid, queue, rpc_qname(queue));
 172 }
 173
 174 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 175 {
 176         queue->priority = priority;
 177         queue->count = 1 << (priority * 2);
 178 }
 179
 180 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 181 {
 182         queue->owner = pid;
 183         queue->nr = RPC_BATCH_COUNT;
 184 }
 185
 186 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 187 {
 188         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 189         rpc_set_waitqueue_owner(queue, 0);
 190 }
 191
 192 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 193 {
 194         int i;
 195
 196         spin_lock_init(&queue->lock);
 197         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 198                 INIT_LIST_HEAD(&queue->tasks[i]);
 199         queue->maxpriority = nr_queues - 1;
 200         rpc_reset_waitqueue_priority(queue);
 201         queue->qlen = 0;
 202         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 203         INIT_LIST_HEAD(&queue->timer_list.list);
 204 #ifdef RPC_DEBUG
 205         queue->name = qname;
 206 #endif
 207 }
 208
 209 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 210 {
 211         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 212 }
 213
 214 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 215 {
 216         __rpc_init_priority_wait_queue(queue, qname, 1);
 217 }
 218 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 219
 220 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 221 {
 222         del_timer_sync(&queue->timer_list.timer);
 223 }
 224 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 225
 226 static int rpc_wait_bit_killable(void *word)
 227 {
 228         if (fatal_signal_pending(current))
 229                 return -ERESTARTSYS;
 230         schedule();
 231         return 0;
 232 }
 233
 234 #ifdef RPC_DEBUG
 235 static void rpc_task_set_debuginfo(struct rpc_task *task)
 236 {
 237         static atomic_t rpc_pid;
 238
 239         task->tk_magic = RPC_TASK_MAGIC_ID;
 240         task->tk_pid = atomic_inc_return(&rpc_pid);
 241 }
 242 #else
 243 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 244 {
 245 }
 246 #endif
 247
 248 static void rpc_set_active(struct rpc_task *task)
 249 {
 250         struct rpc_clnt *clnt;
 251         if (test_and_set_bit(RPC_TASK_ACTIVE, &task->tk_runstate) != 0)
 252                 return;
 253         rpc_task_set_debuginfo(task);
 254         /* Add to global list of all tasks */
 255         clnt = task->tk_client;
 256         if (clnt != NULL) {
 257                 spin_lock(&clnt->cl_lock);
 258                 list_add_tail(&task->tk_task, &clnt->cl_tasks);
 259                 spin_unlock(&clnt->cl_lock);
 260         }
 261 }
 262
 263 /*
 264  * Mark an RPC call as having completed by clearing the 'active' bit
 265  */
 266 static void rpc_mark_complete_task(struct rpc_task *task)
 267 {
 268         smp_mb__before_clear_bit();
 269         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 270         smp_mb__after_clear_bit();
 271         wake_up_bit(&task->tk_runstate, RPC_TASK_ACTIVE);
 272 }
 273
 274 /*
 275  * Allow callers to wait for completion of an RPC call
 276  */
 277 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 278 {
 279         if (action == NULL)
 280                 action = rpc_wait_bit_killable;
 281         return wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 282                         action, TASK_KILLABLE);
 283 }
 284 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 285
 286 /*
 287  * Make an RPC task runnable.
 288  *
 289  * Note: If the task is ASYNC, this must be called with
 290  * the spinlock held to protect the wait queue operation.
 291  */
 292 static void rpc_make_runnable(struct rpc_task *task)
 293 {
 294         rpc_clear_queued(task);
 295         if (rpc_test_and_set_running(task))
 296                 return;
 297         if (RPC_IS_ASYNC(task)) {
 298                 int status;
 299
 300                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 301                 status = queue_work(rpciod_workqueue, &task->u.tk_work);
 302                 if (status < 0) {
 303                         printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 304                         task->tk_status = status;
 305                         return;
 306                 }
 307         } else
 308                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 309 }
 310
 311 /*
 312  * Prepare for sleeping on a wait queue.
 313  * By always appending tasks to the list we ensure FIFO behavior.
 314  * NB: An RPC task will only receive interrupt-driven events as long
 315  * as it's on a wait queue.
 316  */
 317 static void __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 318                         rpc_action action)
 319 {
 320         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 321                         task->tk_pid, rpc_qname(q), jiffies);
 322
 323         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 324                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 325                 return;
 326         }
 327
 328         __rpc_add_wait_queue(q, task);
 329
 330         BUG_ON(task->tk_callback != NULL);
 331         task->tk_callback = action;
 332         __rpc_add_timer(q, task);
 333 }
 334
 335 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 336                                 rpc_action action)
 337 {
 338         /* Mark the task as being activated if so needed */
 339         rpc_set_active(task);
 340
 341         /*
 342          * Protect the queue operations.
 343          */
 344         spin_lock_bh(&q->lock);
 345         __rpc_sleep_on(q, task, action);
 346         spin_unlock_bh(&q->lock);
 347 }
 348 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 349
 350 /**
 351  * __rpc_do_wake_up_task - wake up a single rpc_task
 352  * @queue: wait queue
 353  * @task: task to be woken up
 354  *
 355  * Caller must hold queue->lock, and have cleared the task queued flag.
 356  */
 357 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 358 {
 359         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 360                         task->tk_pid, jiffies);
 361
 362 #ifdef RPC_DEBUG
 363         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 364 #endif
 365         /* Has the task been executed yet? If not, we cannot wake it up! */
 366         if (!RPC_IS_ACTIVATED(task)) {
 367                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 368                 return;
 369         }
 370
 371         __rpc_remove_wait_queue(queue, task);
 372
 373         rpc_make_runnable(task);
 374
 375         dprintk("RPC:       __rpc_wake_up_task done\n");
 376 }
 377
 378 /*
 379  * Wake up a queued task while the queue lock is being held
 380  */
 381 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 382 {
 383         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 384                 __rpc_do_wake_up_task(queue, task);
 385 }
 386
 387 /*
 388  * Tests whether rpc queue is empty
 389  */
 390 int rpc_queue_empty(struct rpc_wait_queue *queue)
 391 {
 392         int res;
 393
 394         spin_lock_bh(&queue->lock);
 395         res = queue->qlen;
 396         spin_unlock_bh(&queue->lock);
 397         return (res == 0);
 398 }
 399 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 400
 401 /*
 402  * Wake up a task on a specific queue
 403  */
 404 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 405 {
 406         spin_lock_bh(&queue->lock);
 407         rpc_wake_up_task_queue_locked(queue, task);
 408         spin_unlock_bh(&queue->lock);
 409 }
 410 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 411
 412 /*
 413  * Wake up the specified task
 414  */
 415 static void rpc_wake_up_task(struct rpc_task *task)
 416 {
 417         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 418 }
 419
 420 /*
 421  * Wake up the next task on a priority queue.
 422  */
 423 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 424 {
 425         struct list_head *q;
 426         struct rpc_task *task;
 427
 428         /*
 429          * Service a batch of tasks from a single owner.
 430          */
 431         q = &queue->tasks[queue->priority];
 432         if (!list_empty(q)) {
 433                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 434                 if (queue->owner == task->tk_owner) {
 435                         if (--queue->nr)
 436                                 goto out;
 437                         list_move_tail(&task->u.tk_wait.list, q);
 438                 }
 439                 /*
 440                  * Check if we need to switch queues.
 441                  */
 442                 if (--queue->count)
 443                         goto new_owner;
 444         }
 445
 446         /*
 447          * Service the next queue.
 448          */
 449         do {
 450                 if (q == &queue->tasks[0])
 451                         q = &queue->tasks[queue->maxpriority];
 452                 else
 453                         q = q - 1;
 454                 if (!list_empty(q)) {
 455                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 456                         goto new_queue;
 457                 }
 458         } while (q != &queue->tasks[queue->priority]);
 459
 460         rpc_reset_waitqueue_priority(queue);
 461         return NULL;
 462
 463 new_queue:
 464         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 465 new_owner:
 466         rpc_set_waitqueue_owner(queue, task->tk_owner);
 467 out:
 468         rpc_wake_up_task_queue_locked(queue, task);
 469         return task;
 470 }
 471
 472 /*
 473  * Wake up the next task on the wait queue.
 474  */
 475 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 476 {
 477         struct rpc_task *task = NULL;
 478
 479         dprintk("RPC:       wake_up_next(%p \"%s\")\n",
 480                         queue, rpc_qname(queue));
 481         spin_lock_bh(&queue->lock);
 482         if (RPC_IS_PRIORITY(queue))
 483                 task = __rpc_wake_up_next_priority(queue);
 484         else {
 485                 task_for_first(task, &queue->tasks[0])
 486                         rpc_wake_up_task_queue_locked(queue, task);
 487         }
 488         spin_unlock_bh(&queue->lock);
 489
 490         return task;
 491 }
 492 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 493
 494 /**
 495  * rpc_wake_up - wake up all rpc_tasks
 496  * @queue: rpc_wait_queue on which the tasks are sleeping
 497  *
 498  * Grabs queue->lock
 499  */
 500 void rpc_wake_up(struct rpc_wait_queue *queue)
 501 {
 502         struct rpc_task *task, *next;
 503         struct list_head *head;
 504
 505         spin_lock_bh(&queue->lock);
 506         head = &queue->tasks[queue->maxpriority];
 507         for (;;) {
 508                 list_for_each_entry_safe(task, next, head, u.tk_wait.list)
 509                         rpc_wake_up_task_queue_locked(queue, task);
 510                 if (head == &queue->tasks[0])
 511                         break;
 512                 head--;
 513         }
 514         spin_unlock_bh(&queue->lock);
 515 }
 516 EXPORT_SYMBOL_GPL(rpc_wake_up);
 517
 518 /**
 519  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 520  * @queue: rpc_wait_queue on which the tasks are sleeping
 521  * @status: status value to set
 522  *
 523  * Grabs queue->lock
 524  */
 525 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 526 {
 527         struct rpc_task *task, *next;
 528         struct list_head *head;
 529
 530         spin_lock_bh(&queue->lock);
 531         head = &queue->tasks[queue->maxpriority];
 532         for (;;) {
 533                 list_for_each_entry_safe(task, next, head, u.tk_wait.list) {
 534                         task->tk_status = status;
 535                         rpc_wake_up_task_queue_locked(queue, task);
 536                 }
 537                 if (head == &queue->tasks[0])
 538                         break;
 539                 head--;
 540         }
 541         spin_unlock_bh(&queue->lock);
 542 }
 543 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 544
 545 static void __rpc_queue_timer_fn(unsigned long ptr)
 546 {
 547         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 548         struct rpc_task *task, *n;
 549         unsigned long expires, now, timeo;
 550
 551         spin_lock(&queue->lock);
 552         expires = now = jiffies;
 553         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 554                 timeo = task->u.tk_wait.expires;
 555                 if (time_after_eq(now, timeo)) {
 556                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 557                         task->tk_status = -ETIMEDOUT;
 558                         rpc_wake_up_task_queue_locked(queue, task);
 559                         continue;
 560                 }
 561                 if (expires == now || time_after(expires, timeo))
 562                         expires = timeo;
 563         }
 564         if (!list_empty(&queue->timer_list.list))
 565                 rpc_set_queue_timer(queue, expires);
 566         spin_unlock(&queue->lock);
 567 }
 568
 569 static void __rpc_atrun(struct rpc_task *task)
 570 {
 571         task->tk_status = 0;
 572 }
 573
 574 /*
 575  * Run a task at a later time
 576  */
 577 void rpc_delay(struct rpc_task *task, unsigned long delay)
 578 {
 579         task->tk_timeout = delay;
 580         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 581 }
 582 EXPORT_SYMBOL_GPL(rpc_delay);
 583
 584 /*
 585  * Helper to call task->tk_ops->rpc_call_prepare
 586  */
 587 void rpc_prepare_task(struct rpc_task *task)
 588 {
 589         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 590 }
 591
 592 /*
 593  * Helper that calls task->tk_ops->rpc_call_done if it exists
 594  */
 595 void rpc_exit_task(struct rpc_task *task)
 596 {
 597         task->tk_action = NULL;
 598         if (task->tk_ops->rpc_call_done != NULL) {
 599                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 600                 if (task->tk_action != NULL) {
 601                         WARN_ON(RPC_ASSASSINATED(task));
 602                         /* Always release the RPC slot and buffer memory */
 603                         xprt_release(task);
 604                 }
 605         }
 606 }
 607 EXPORT_SYMBOL_GPL(rpc_exit_task);
 608
 609 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 610 {
 611         if (ops->rpc_release != NULL)
 612                 ops->rpc_release(calldata);
 613 }
 614
 615 /*
 616  * This is the RPC `scheduler' (or rather, the finite state machine).
 617  */
 618 static void __rpc_execute(struct rpc_task *task)
 619 {
 620         struct rpc_wait_queue *queue;
 621         int task_is_async = RPC_IS_ASYNC(task);
 622         int status = 0;
 623
 624         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 625                         task->tk_pid, task->tk_flags);
 626
 627         BUG_ON(RPC_IS_QUEUED(task));
 628
 629         for (;;) {
 630
 631                 /*
 632                  * Execute any pending callback.
 633                  */
 634                 if (task->tk_callback) {
 635                         void (*save_callback)(struct rpc_task *);
 636
 637                         /*
 638                          * We set tk_callback to NULL before calling it,
 639                          * in case it sets the tk_callback field itself:
 640                          */
 641                         save_callback = task->tk_callback;
 642                         task->tk_callback = NULL;
 643                         save_callback(task);
 644                 }
 645
 646                 /*
 647                  * Perform the next FSM step.
 648                  * tk_action may be NULL when the task has been killed
 649                  * by someone else.
 650                  */
 651                 if (!RPC_IS_QUEUED(task)) {
 652                         if (task->tk_action == NULL)
 653                                 break;
 654                         task->tk_action(task);
 655                 }
 656
 657                 /*
 658                  * Lockless check for whether task is sleeping or not.
 659                  */
 660                 if (!RPC_IS_QUEUED(task))
 661                         continue;
 662                 /*
 663                  * The queue->lock protects against races with
 664                  * rpc_make_runnable().
 665                  *
 666                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 667                  * rpc_task, rpc_make_runnable() can assign it to a
 668                  * different workqueue. We therefore cannot assume that the
 669                  * rpc_task pointer may still be dereferenced.
 670                  */
 671                 queue = task->tk_waitqueue;
 672                 spin_lock_bh(&queue->lock);
 673                 if (!RPC_IS_QUEUED(task)) {
 674                         spin_unlock_bh(&queue->lock);
 675                         continue;
 676                 }
 677                 rpc_clear_running(task);
 678                 spin_unlock_bh(&queue->lock);
 679                 if (task_is_async)
 680                         return;
 681
 682                 /* sync task: sleep here */
 683                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 684                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 685                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 686                                 TASK_KILLABLE);
 687                 if (status == -ERESTARTSYS) {
 688                         /*
 689                          * When a sync task receives a signal, it exits with
 690                          * -ERESTARTSYS. In order to catch any callbacks that
 691                          * clean up after sleeping on some queue, we don't
 692                          * break the loop here, but go around once more.
 693                          */
 694                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 695                         task->tk_flags |= RPC_TASK_KILLED;
 696                         rpc_exit(task, -ERESTARTSYS);
 697                         rpc_wake_up_task(task);
 698                 }
 699                 rpc_set_running(task);
 700                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 701         }
 702
 703         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 704                         task->tk_status);
 705         /* Release all resources associated with the task */
 706         rpc_release_task(task);
 707 }
 708
 709 /*
 710  * User-visible entry point to the scheduler.
 711  *
 712  * This may be called recursively if e.g. an async NFS task updates
 713  * the attributes and finds that dirty pages must be flushed.
 714  * NOTE: Upon exit of this function the task is guaranteed to be
 715  *       released. In particular note that tk_release() will have
 716  *       been called, so your task memory may have been freed.
 717  */
 718 void rpc_execute(struct rpc_task *task)
 719 {
 720         rpc_set_active(task);
 721         rpc_set_running(task);
 722         __rpc_execute(task);
 723 }
 724
 725 static void rpc_async_schedule(struct work_struct *work)
 726 {
 727         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 728 }
 729
 730 /**
 731  * rpc_malloc - allocate an RPC buffer
 732  * @task: RPC task that will use this buffer
 733  * @size: requested byte size
 734  *
 735  * To prevent rpciod from hanging, this allocator never sleeps,
 736  * returning NULL if the request cannot be serviced immediately.
 737  * The caller can arrange to sleep in a way that is safe for rpciod.
 738  *
 739  * Most requests are 'small' (under 2KiB) and can be serviced from a
 740  * mempool, ensuring that NFS reads and writes can always proceed,
 741  * and that there is good locality of reference for these buffers.
 742  *
 743  * In order to avoid memory starvation triggering more writebacks of
 744  * NFS requests, we avoid using GFP_KERNEL.
 745  */
 746 void *rpc_malloc(struct rpc_task *task, size_t size)
 747 {
 748         struct rpc_buffer *buf;
 749         gfp_t gfp = RPC_IS_SWAPPER(task) ? GFP_ATOMIC : GFP_NOWAIT;
 750
 751         size += sizeof(struct rpc_buffer);
 752         if (size <= RPC_BUFFER_MAXSIZE)
 753                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 754         else
 755                 buf = kmalloc(size, gfp);
 756
 757         if (!buf)
 758                 return NULL;
 759
 760         buf->len = size;
 761         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 762                         task->tk_pid, size, buf);
 763         return &buf->data;
 764 }
 765 EXPORT_SYMBOL_GPL(rpc_malloc);
 766
 767 /**
 768  * rpc_free - free buffer allocated via rpc_malloc
 769  * @buffer: buffer to free
 770  *
 771  */
 772 void rpc_free(void *buffer)
 773 {
 774         size_t size;
 775         struct rpc_buffer *buf;
 776
 777         if (!buffer)
 778                 return;
 779
 780         buf = container_of(buffer, struct rpc_buffer, data);
 781         size = buf->len;
 782
 783         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 784                         size, buf);
 785
 786         if (size <= RPC_BUFFER_MAXSIZE)
 787                 mempool_free(buf, rpc_buffer_mempool);
 788         else
 789                 kfree(buf);
 790 }
 791 EXPORT_SYMBOL_GPL(rpc_free);
 792
 793 /*
 794  * Creation and deletion of RPC task structures
 795  */
 796 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 797 {
 798         memset(task, 0, sizeof(*task));
 799         atomic_set(&task->tk_count, 1);
 800         task->tk_flags  = task_setup_data->flags;
 801         task->tk_ops = task_setup_data->callback_ops;
 802         task->tk_calldata = task_setup_data->callback_data;
 803         INIT_LIST_HEAD(&task->tk_task);
 804
 805         /* Initialize retry counters */
 806         task->tk_garb_retry = 2;
 807         task->tk_cred_retry = 2;
 808
 809         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 810         task->tk_owner = current->tgid;
 811
 812         /* Initialize workqueue for async tasks */
 813         task->tk_workqueue = task_setup_data->workqueue;
 814
 815         task->tk_client = task_setup_data->rpc_client;
 816         if (task->tk_client != NULL) {
 817                 kref_get(&task->tk_client->cl_kref);
 818                 if (task->tk_client->cl_softrtry)
 819                         task->tk_flags |= RPC_TASK_SOFT;
 820         }
 821
 822         if (task->tk_ops->rpc_call_prepare != NULL)
 823                 task->tk_action = rpc_prepare_task;
 824
 825         if (task_setup_data->rpc_message != NULL) {
 826                 task->tk_msg.rpc_proc = task_setup_data->rpc_message->rpc_proc;
 827                 task->tk_msg.rpc_argp = task_setup_data->rpc_message->rpc_argp;
 828                 task->tk_msg.rpc_resp = task_setup_data->rpc_message->rpc_resp;
 829                 /* Bind the user cred */
 830                 rpcauth_bindcred(task, task_setup_data->rpc_message->rpc_cred, task_setup_data->flags);
 831                 if (task->tk_action == NULL)
 832                         rpc_call_start(task);
 833         }
 834
 835         /* starting timestamp */
 836         task->tk_start = jiffies;
 837
 838         dprintk("RPC:       new task initialized, procpid %u\n",
 839                                 task_pid_nr(current));
 840 }
 841
 842 static struct rpc_task *
 843 rpc_alloc_task(void)
 844 {
 845         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 846 }
 847
 848 /*
 849  * Create a new task for the specified client.
 850  */
 851 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 852 {
 853         struct rpc_task *task = setup_data->task;
 854         unsigned short flags = 0;
 855
 856         if (task == NULL) {
 857                 task = rpc_alloc_task();
 858                 if (task == NULL)
 859                         goto out;
 860                 flags = RPC_TASK_DYNAMIC;
 861         }
 862
 863         rpc_init_task(task, setup_data);
 864
 865         task->tk_flags |= flags;
 866         dprintk("RPC:       allocated task %p\n", task);
 867 out:
 868         return task;
 869 }
 870
 871 static void rpc_free_task(struct rpc_task *task)
 872 {
 873         const struct rpc_call_ops *tk_ops = task->tk_ops;
 874         void *calldata = task->tk_calldata;
 875
 876         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 877                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 878                 mempool_free(task, rpc_task_mempool);
 879         }
 880         rpc_release_calldata(tk_ops, calldata);
 881 }
 882
 883 static void rpc_async_release(struct work_struct *work)
 884 {
 885         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 886 }
 887
 888 void rpc_put_task(struct rpc_task *task)
 889 {
 890         if (!atomic_dec_and_test(&task->tk_count))
 891                 return;
 892         /* Release resources */
 893         if (task->tk_rqstp)
 894                 xprt_release(task);
 895         if (task->tk_msg.rpc_cred)
 896                 rpcauth_unbindcred(task);
 897         if (task->tk_client) {
 898                 rpc_release_client(task->tk_client);
 899                 task->tk_client = NULL;
 900         }
 901         if (task->tk_workqueue != NULL) {
 902                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 903                 queue_work(task->tk_workqueue, &task->u.tk_work);
 904         } else
 905                 rpc_free_task(task);
 906 }
 907 EXPORT_SYMBOL_GPL(rpc_put_task);
 908
 909 static void rpc_release_task(struct rpc_task *task)
 910 {
 911 #ifdef RPC_DEBUG
 912         BUG_ON(task->tk_magic != RPC_TASK_MAGIC_ID);
 913 #endif
 914         dprintk("RPC: %5u release task\n", task->tk_pid);
 915
 916         if (!list_empty(&task->tk_task)) {
 917                 struct rpc_clnt *clnt = task->tk_client;
 918                 /* Remove from client task list */
 919                 spin_lock(&clnt->cl_lock);
 920                 list_del(&task->tk_task);
 921                 spin_unlock(&clnt->cl_lock);
 922         }
 923         BUG_ON (RPC_IS_QUEUED(task));
 924
 925 #ifdef RPC_DEBUG
 926         task->tk_magic = 0;
 927 #endif
 928         /* Wake up anyone who is waiting for task completion */
 929         rpc_mark_complete_task(task);
 930
 931         rpc_put_task(task);
 932 }
 933
 934 /*
 935  * Kill all tasks for the given client.
 936  * XXX: kill their descendants as well?
 937  */
 938 void rpc_killall_tasks(struct rpc_clnt *clnt)
 939 {
 940         struct rpc_task *rovr;
 941
 942
 943         if (list_empty(&clnt->cl_tasks))
 944                 return;
 945         dprintk("RPC:       killing all tasks for client %p\n", clnt);
 946         /*
 947          * Spin lock all_tasks to prevent changes...
 948          */
 949         spin_lock(&clnt->cl_lock);
 950         list_for_each_entry(rovr, &clnt->cl_tasks, tk_task) {
 951                 if (! RPC_IS_ACTIVATED(rovr))
 952                         continue;
 953                 if (!(rovr->tk_flags & RPC_TASK_KILLED)) {
 954                         rovr->tk_flags |= RPC_TASK_KILLED;
 955                         rpc_exit(rovr, -EIO);
 956                         rpc_wake_up_task(rovr);
 957                 }
 958         }
 959         spin_unlock(&clnt->cl_lock);
 960 }
 961 EXPORT_SYMBOL_GPL(rpc_killall_tasks);
 962
 963 int rpciod_up(void)
 964 {
 965         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
 966 }
 967
 968 void rpciod_down(void)
 969 {
 970         module_put(THIS_MODULE);
 971 }
 972
 973 /*
 974  * Start up the rpciod workqueue.
 975  */
 976 static int rpciod_start(void)
 977 {
 978         struct workqueue_struct *wq;
 979
 980         /*
 981          * Create the rpciod thread and wait for it to start.
 982          */
 983         dprintk("RPC:       creating workqueue rpciod\n");
 984         wq = create_workqueue("rpciod");
 985         rpciod_workqueue = wq;
 986         return rpciod_workqueue != NULL;
 987 }
 988
 989 static void rpciod_stop(void)
 990 {
 991         struct workqueue_struct *wq = NULL;
 992
 993         if (rpciod_workqueue == NULL)
 994                 return;
 995         dprintk("RPC:       destroying workqueue rpciod\n");
 996
 997         wq = rpciod_workqueue;
 998         rpciod_workqueue = NULL;
 999         destroy_workqueue(wq);
1000 }
1001
1002 void
1003 rpc_destroy_mempool(void)
1004 {
1005         rpciod_stop();
1006         if (rpc_buffer_mempool)
1007                 mempool_destroy(rpc_buffer_mempool);
1008         if (rpc_task_mempool)
1009                 mempool_destroy(rpc_task_mempool);
1010         if (rpc_task_slabp)
1011                 kmem_cache_destroy(rpc_task_slabp);
1012         if (rpc_buffer_slabp)
1013                 kmem_cache_destroy(rpc_buffer_slabp);
1014         rpc_destroy_wait_queue(&delay_queue);
1015 }
1016
1017 int
1018 rpc_init_mempool(void)
1019 {
1020         /*
1021          * The following is not strictly a mempool initialisation,
1022          * but there is no harm in doing it here
1023          */
1024         rpc_init_wait_queue(&delay_queue, "delayq");
1025         if (!rpciod_start())
1026                 goto err_nomem;
1027
1028         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1029                                              sizeof(struct rpc_task),
1030                                              0, SLAB_HWCACHE_ALIGN,
1031                                              NULL);
1032         if (!rpc_task_slabp)
1033                 goto err_nomem;
1034         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1035                                              RPC_BUFFER_MAXSIZE,
1036                                              0, SLAB_HWCACHE_ALIGN,
1037                                              NULL);
1038         if (!rpc_buffer_slabp)
1039                 goto err_nomem;
1040         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1041                                                     rpc_task_slabp);
1042         if (!rpc_task_mempool)
1043                 goto err_nomem;
1044         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1045                                                       rpc_buffer_slabp);
1046         if (!rpc_buffer_mempool)
1047                 goto err_nomem;
1048         return 0;
1049 err_nomem:
1050         rpc_destroy_mempool();
1051         return -ENOMEM;
1052 }