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/suspend.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24 #include <linux/sunrpc/xprt.h>
  25
  26 #ifdef RPC_DEBUG
  27 #define RPCDBG_FACILITY         RPCDBG_SCHED
  28 static int                      rpc_task_id;
  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 kmem_cache_t     *rpc_task_slabp;
  38 static kmem_cache_t     *rpc_buffer_slabp;
  39 static mempool_t        *rpc_task_mempool;
  40 static mempool_t        *rpc_buffer_mempool;
  41
  42 static void                     __rpc_default_timer(struct rpc_task *task);
  43 static void                     rpciod_killall(void);
  44
  45 /*
  46  * When an asynchronous RPC task is activated within a bottom half
  47  * handler, or while executing another RPC task, it is put on
  48  * schedq, and rpciod is woken up.
  49  */
  50 static RPC_WAITQ(schedq, "schedq");
  51
  52 /*
  53  * RPC tasks that create another task (e.g. for contacting the portmapper)
  54  * will wait on this queue for their child's completion
  55  */
  56 static RPC_WAITQ(childq, "childq");
  57
  58 /*
  59  * RPC tasks sit here while waiting for conditions to improve.
  60  */
  61 static RPC_WAITQ(delay_queue, "delayq");
  62
  63 /*
  64  * All RPC tasks are linked into this list
  65  */
  66 static LIST_HEAD(all_tasks);
  67
  68 /*
  69  * rpciod-related stuff
  70  */
  71 static DECLARE_WAIT_QUEUE_HEAD(rpciod_idle);
  72 static DECLARE_COMPLETION(rpciod_killer);
  73 static DECLARE_MUTEX(rpciod_sema);
  74 static unsigned int             rpciod_users;
  75 static pid_t                    rpciod_pid;
  76 static int                      rpc_inhibit;
  77
  78 /*
  79  * Spinlock for wait queues. Access to the latter also has to be
  80  * interrupt-safe in order to allow timers to wake up sleeping tasks.
  81  */
  82 static spinlock_t rpc_queue_lock = SPIN_LOCK_UNLOCKED;
  83 /*
  84  * Spinlock for other critical sections of code.
  85  */
  86 static spinlock_t rpc_sched_lock = SPIN_LOCK_UNLOCKED;
  87
  88 /*
  89  * Disable the timer for a given RPC task. Should be called with
  90  * rpc_queue_lock and bh_disabled in order to avoid races within
  91  * rpc_run_timer().
  92  */
  93 static inline void
  94 __rpc_disable_timer(struct rpc_task *task)
  95 {
  96         dprintk("RPC: %4d disabling timer\n", task->tk_pid);
  97         task->tk_timeout_fn = NULL;
  98         task->tk_timeout = 0;
  99 }
 100
 101 /*
 102  * Run a timeout function.
 103  * We use the callback in order to allow __rpc_wake_up_task()
 104  * and friends to disable the timer synchronously on SMP systems
 105  * without calling del_timer_sync(). The latter could cause a
 106  * deadlock if called while we're holding spinlocks...
 107  */
 108 static void
 109 rpc_run_timer(struct rpc_task *task)
 110 {
 111         void (*callback)(struct rpc_task *);
 112
 113         spin_lock_bh(&rpc_queue_lock);
 114         callback = task->tk_timeout_fn;
 115         task->tk_timeout_fn = NULL;
 116         spin_unlock_bh(&rpc_queue_lock);
 117         if (callback) {
 118                 dprintk("RPC: %4d running timer\n", task->tk_pid);
 119                 callback(task);
 120         }
 121 }
 122
 123 /*
 124  * Set up a timer for the current task.
 125  */
 126 static inline void
 127 __rpc_add_timer(struct rpc_task *task, rpc_action timer)
 128 {
 129         if (!task->tk_timeout)
 130                 return;
 131
 132         dprintk("RPC: %4d setting alarm for %lu ms\n",
 133                         task->tk_pid, task->tk_timeout * 1000 / HZ);
 134
 135         if (timer)
 136                 task->tk_timeout_fn = timer;
 137         else
 138                 task->tk_timeout_fn = __rpc_default_timer;
 139         mod_timer(&task->tk_timer, jiffies + task->tk_timeout);
 140 }
 141
 142 /*
 143  * Set up a timer for an already sleeping task.
 144  */
 145 void rpc_add_timer(struct rpc_task *task, rpc_action timer)
 146 {
 147         spin_lock_bh(&rpc_queue_lock);
 148         if (!RPC_IS_RUNNING(task))
 149                 __rpc_add_timer(task, timer);
 150         spin_unlock_bh(&rpc_queue_lock);
 151 }
 152
 153 /*
 154  * Delete any timer for the current task. Because we use del_timer_sync(),
 155  * this function should never be called while holding rpc_queue_lock.
 156  */
 157 static inline void
 158 rpc_delete_timer(struct rpc_task *task)
 159 {
 160         if (del_timer_sync(&task->tk_timer))
 161                 dprintk("RPC: %4d deleting timer\n", task->tk_pid);
 162 }
 163
 164 /*
 165  * Add new request to a priority queue.
 166  */
 167 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue, struct rpc_task *task)
 168 {
 169         struct list_head *q;
 170         struct rpc_task *t;
 171
 172         q = &queue->tasks[task->tk_priority];
 173         if (unlikely(task->tk_priority > queue->maxpriority))
 174                 q = &queue->tasks[queue->maxpriority];
 175         list_for_each_entry(t, q, tk_list) {
 176                 if (t->tk_cookie == task->tk_cookie) {
 177                         list_add_tail(&task->tk_list, &t->tk_links);
 178                         return;
 179                 }
 180         }
 181         list_add_tail(&task->tk_list, q);
 182 }
 183
 184 /*
 185  * Add new request to wait queue.
 186  *
 187  * Swapper tasks always get inserted at the head of the queue.
 188  * This should avoid many nasty memory deadlocks and hopefully
 189  * improve overall performance.
 190  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 191  */
 192 static int __rpc_add_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 193 {
 194         if (task->tk_rpcwait == queue)
 195                 return 0;
 196
 197         if (task->tk_rpcwait) {
 198                 printk(KERN_WARNING "RPC: doubly enqueued task!\n");
 199                 return -EWOULDBLOCK;
 200         }
 201         if (RPC_IS_PRIORITY(queue))
 202                 __rpc_add_wait_queue_priority(queue, task);
 203         else if (RPC_IS_SWAPPER(task))
 204                 list_add(&task->tk_list, &queue->tasks[0]);
 205         else
 206                 list_add_tail(&task->tk_list, &queue->tasks[0]);
 207         task->tk_rpcwait = queue;
 208
 209         dprintk("RPC: %4d added to queue %p \"%s\"\n",
 210                                 task->tk_pid, queue, rpc_qname(queue));
 211
 212         return 0;
 213 }
 214
 215 int rpc_add_wait_queue(struct rpc_wait_queue *q, struct rpc_task *task)
 216 {
 217         int             result;
 218
 219         spin_lock_bh(&rpc_queue_lock);
 220         result = __rpc_add_wait_queue(q, task);
 221         spin_unlock_bh(&rpc_queue_lock);
 222         return result;
 223 }
 224
 225 /*
 226  * Remove request from a priority queue.
 227  */
 228 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 229 {
 230         struct rpc_task *t;
 231
 232         if (!list_empty(&task->tk_links)) {
 233                 t = list_entry(task->tk_links.next, struct rpc_task, tk_list);
 234                 list_move(&t->tk_list, &task->tk_list);
 235                 list_splice_init(&task->tk_links, &t->tk_links);
 236         }
 237         list_del(&task->tk_list);
 238 }
 239
 240 /*
 241  * Remove request from queue.
 242  * Note: must be called with spin lock held.
 243  */
 244 static void __rpc_remove_wait_queue(struct rpc_task *task)
 245 {
 246         struct rpc_wait_queue *queue = task->tk_rpcwait;
 247
 248         if (!queue)
 249                 return;
 250
 251         if (RPC_IS_PRIORITY(queue))
 252                 __rpc_remove_wait_queue_priority(task);
 253         else
 254                 list_del(&task->tk_list);
 255         task->tk_rpcwait = NULL;
 256
 257         dprintk("RPC: %4d removed from queue %p \"%s\"\n",
 258                                 task->tk_pid, queue, rpc_qname(queue));
 259 }
 260
 261 void
 262 rpc_remove_wait_queue(struct rpc_task *task)
 263 {
 264         if (!task->tk_rpcwait)
 265                 return;
 266         spin_lock_bh(&rpc_queue_lock);
 267         __rpc_remove_wait_queue(task);
 268         spin_unlock_bh(&rpc_queue_lock);
 269 }
 270
 271 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 272 {
 273         queue->priority = priority;
 274         queue->count = 1 << (priority * 2);
 275 }
 276
 277 static inline void rpc_set_waitqueue_cookie(struct rpc_wait_queue *queue, unsigned long cookie)
 278 {
 279         queue->cookie = cookie;
 280         queue->nr = RPC_BATCH_COUNT;
 281 }
 282
 283 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 284 {
 285         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 286         rpc_set_waitqueue_cookie(queue, 0);
 287 }
 288
 289 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, int maxprio)
 290 {
 291         int i;
 292
 293         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 294                 INIT_LIST_HEAD(&queue->tasks[i]);
 295         queue->maxpriority = maxprio;
 296         rpc_reset_waitqueue_priority(queue);
 297 #ifdef RPC_DEBUG
 298         queue->name = qname;
 299 #endif
 300 }
 301
 302 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 303 {
 304         __rpc_init_priority_wait_queue(queue, qname, RPC_PRIORITY_HIGH);
 305 }
 306
 307 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 308 {
 309         __rpc_init_priority_wait_queue(queue, qname, 0);
 310 }
 311 EXPORT_SYMBOL(rpc_init_wait_queue);
 312
 313 /*
 314  * Make an RPC task runnable.
 315  *
 316  * Note: If the task is ASYNC, this must be called with
 317  * the spinlock held to protect the wait queue operation.
 318  */
 319 static inline void
 320 rpc_make_runnable(struct rpc_task *task)
 321 {
 322         if (task->tk_timeout_fn) {
 323                 printk(KERN_ERR "RPC: task w/ running timer in rpc_make_runnable!!\n");
 324                 return;
 325         }
 326         rpc_set_running(task);
 327         if (RPC_IS_ASYNC(task)) {
 328                 if (RPC_IS_SLEEPING(task)) {
 329                         int status;
 330                         status = __rpc_add_wait_queue(&schedq, task);
 331                         if (status < 0) {
 332                                 printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 333                                 task->tk_status = status;
 334                                 return;
 335                         }
 336                         rpc_clear_sleeping(task);
 337                         wake_up(&rpciod_idle);
 338                 }
 339         } else {
 340                 rpc_clear_sleeping(task);
 341                 wake_up(&task->tk_wait);
 342         }
 343 }
 344
 345 /*
 346  * Place a newly initialized task on the schedq.
 347  */
 348 static inline void
 349 rpc_schedule_run(struct rpc_task *task)
 350 {
 351         /* Don't run a child twice! */
 352         if (RPC_IS_ACTIVATED(task))
 353                 return;
 354         task->tk_active = 1;
 355         rpc_set_sleeping(task);
 356         rpc_make_runnable(task);
 357 }
 358
 359 /*
 360  *      For other people who may need to wake the I/O daemon
 361  *      but should (for now) know nothing about its innards
 362  */
 363 void rpciod_wake_up(void)
 364 {
 365         if(rpciod_pid==0)
 366                 printk(KERN_ERR "rpciod: wot no daemon?\n");
 367         wake_up(&rpciod_idle);
 368 }
 369
 370 /*
 371  * Prepare for sleeping on a wait queue.
 372  * By always appending tasks to the list we ensure FIFO behavior.
 373  * NB: An RPC task will only receive interrupt-driven events as long
 374  * as it's on a wait queue.
 375  */
 376 static void
 377 __rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 378                         rpc_action action, rpc_action timer)
 379 {
 380         int status;
 381
 382         dprintk("RPC: %4d sleep_on(queue \"%s\" time %ld)\n", task->tk_pid,
 383                                 rpc_qname(q), jiffies);
 384
 385         if (!RPC_IS_ASYNC(task) && !RPC_IS_ACTIVATED(task)) {
 386                 printk(KERN_ERR "RPC: Inactive synchronous task put to sleep!\n");
 387                 return;
 388         }
 389
 390         /* Mark the task as being activated if so needed */
 391         if (!RPC_IS_ACTIVATED(task)) {
 392                 task->tk_active = 1;
 393                 rpc_set_sleeping(task);
 394         }
 395
 396         status = __rpc_add_wait_queue(q, task);
 397         if (status) {
 398                 printk(KERN_WARNING "RPC: failed to add task to queue: error: %d!\n", status);
 399                 task->tk_status = status;
 400         } else {
 401                 rpc_clear_running(task);
 402                 if (task->tk_callback) {
 403                         dprintk(KERN_ERR "RPC: %4d overwrites an active callback\n", task->tk_pid);
 404                         BUG();
 405                 }
 406                 task->tk_callback = action;
 407                 __rpc_add_timer(task, timer);
 408         }
 409 }
 410
 411 void
 412 rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 413                                 rpc_action action, rpc_action timer)
 414 {
 415         /*
 416          * Protect the queue operations.
 417          */
 418         spin_lock_bh(&rpc_queue_lock);
 419         __rpc_sleep_on(q, task, action, timer);
 420         spin_unlock_bh(&rpc_queue_lock);
 421 }
 422
 423 /**
 424  * __rpc_wake_up_task - wake up a single rpc_task
 425  * @task: task to be woken up
 426  *
 427  * Caller must hold rpc_queue_lock
 428  */
 429 static void
 430 __rpc_wake_up_task(struct rpc_task *task)
 431 {
 432         dprintk("RPC: %4d __rpc_wake_up_task (now %ld inh %d)\n",
 433                                         task->tk_pid, jiffies, rpc_inhibit);
 434
 435 #ifdef RPC_DEBUG
 436         if (task->tk_magic != 0xf00baa) {
 437                 printk(KERN_ERR "RPC: attempt to wake up non-existing task!\n");
 438                 rpc_debug = ~0;
 439                 rpc_show_tasks();
 440                 return;
 441         }
 442 #endif
 443         /* Has the task been executed yet? If not, we cannot wake it up! */
 444         if (!RPC_IS_ACTIVATED(task)) {
 445                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 446                 return;
 447         }
 448         if (RPC_IS_RUNNING(task))
 449                 return;
 450
 451         __rpc_disable_timer(task);
 452         if (task->tk_rpcwait != &schedq)
 453                 __rpc_remove_wait_queue(task);
 454
 455         rpc_make_runnable(task);
 456
 457         dprintk("RPC:      __rpc_wake_up_task done\n");
 458 }
 459
 460 /*
 461  * Default timeout handler if none specified by user
 462  */
 463 static void
 464 __rpc_default_timer(struct rpc_task *task)
 465 {
 466         dprintk("RPC: %d timeout (default timer)\n", task->tk_pid);
 467         task->tk_status = -ETIMEDOUT;
 468         rpc_wake_up_task(task);
 469 }
 470
 471 /*
 472  * Wake up the specified task
 473  */
 474 void
 475 rpc_wake_up_task(struct rpc_task *task)
 476 {
 477         if (RPC_IS_RUNNING(task))
 478                 return;
 479         spin_lock_bh(&rpc_queue_lock);
 480         __rpc_wake_up_task(task);
 481         spin_unlock_bh(&rpc_queue_lock);
 482 }
 483
 484 /*
 485  * Wake up the next task on a priority queue.
 486  */
 487 static struct rpc_task * __rpc_wake_up_next_priority(struct rpc_wait_queue *queue)
 488 {
 489         struct list_head *q;
 490         struct rpc_task *task;
 491
 492         /*
 493          * Service a batch of tasks from a single cookie.
 494          */
 495         q = &queue->tasks[queue->priority];
 496         if (!list_empty(q)) {
 497                 task = list_entry(q->next, struct rpc_task, tk_list);
 498                 if (queue->cookie == task->tk_cookie) {
 499                         if (--queue->nr)
 500                                 goto out;
 501                         list_move_tail(&task->tk_list, q);
 502                 }
 503                 /*
 504                  * Check if we need to switch queues.
 505                  */
 506                 if (--queue->count)
 507                         goto new_cookie;
 508         }
 509
 510         /*
 511          * Service the next queue.
 512          */
 513         do {
 514                 if (q == &queue->tasks[0])
 515                         q = &queue->tasks[queue->maxpriority];
 516                 else
 517                         q = q - 1;
 518                 if (!list_empty(q)) {
 519                         task = list_entry(q->next, struct rpc_task, tk_list);
 520                         goto new_queue;
 521                 }
 522         } while (q != &queue->tasks[queue->priority]);
 523
 524         rpc_reset_waitqueue_priority(queue);
 525         return NULL;
 526
 527 new_queue:
 528         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 529 new_cookie:
 530         rpc_set_waitqueue_cookie(queue, task->tk_cookie);
 531 out:
 532         __rpc_wake_up_task(task);
 533         return task;
 534 }
 535
 536 /*
 537  * Wake up the next task on the wait queue.
 538  */
 539 struct rpc_task * rpc_wake_up_next(struct rpc_wait_queue *queue)
 540 {
 541         struct rpc_task *task = NULL;
 542
 543         dprintk("RPC:      wake_up_next(%p \"%s\")\n", queue, rpc_qname(queue));
 544         spin_lock_bh(&rpc_queue_lock);
 545         if (RPC_IS_PRIORITY(queue))
 546                 task = __rpc_wake_up_next_priority(queue);
 547         else {
 548                 task_for_first(task, &queue->tasks[0])
 549                         __rpc_wake_up_task(task);
 550         }
 551         spin_unlock_bh(&rpc_queue_lock);
 552
 553         return task;
 554 }
 555
 556 /**
 557  * rpc_wake_up - wake up all rpc_tasks
 558  * @queue: rpc_wait_queue on which the tasks are sleeping
 559  *
 560  * Grabs rpc_queue_lock
 561  */
 562 void rpc_wake_up(struct rpc_wait_queue *queue)
 563 {
 564         struct rpc_task *task;
 565
 566         struct list_head *head;
 567         spin_lock_bh(&rpc_queue_lock);
 568         head = &queue->tasks[queue->maxpriority];
 569         for (;;) {
 570                 while (!list_empty(head)) {
 571                         task = list_entry(head->next, struct rpc_task, tk_list);
 572                         __rpc_wake_up_task(task);
 573                 }
 574                 if (head == &queue->tasks[0])
 575                         break;
 576                 head--;
 577         }
 578         spin_unlock_bh(&rpc_queue_lock);
 579 }
 580
 581 /**
 582  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 583  * @queue: rpc_wait_queue on which the tasks are sleeping
 584  * @status: status value to set
 585  *
 586  * Grabs rpc_queue_lock
 587  */
 588 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 589 {
 590         struct list_head *head;
 591         struct rpc_task *task;
 592
 593         spin_lock_bh(&rpc_queue_lock);
 594         head = &queue->tasks[queue->maxpriority];
 595         for (;;) {
 596                 while (!list_empty(head)) {
 597                         task = list_entry(head->next, struct rpc_task, tk_list);
 598                         task->tk_status = status;
 599                         __rpc_wake_up_task(task);
 600                 }
 601                 if (head == &queue->tasks[0])
 602                         break;
 603                 head--;
 604         }
 605         spin_unlock_bh(&rpc_queue_lock);
 606 }
 607
 608 /*
 609  * Run a task at a later time
 610  */
 611 static void     __rpc_atrun(struct rpc_task *);
 612 void
 613 rpc_delay(struct rpc_task *task, unsigned long delay)
 614 {
 615         task->tk_timeout = delay;
 616         rpc_sleep_on(&delay_queue, task, NULL, __rpc_atrun);
 617 }
 618
 619 static void
 620 __rpc_atrun(struct rpc_task *task)
 621 {
 622         task->tk_status = 0;
 623         rpc_wake_up_task(task);
 624 }
 625
 626 /*
 627  * This is the RPC `scheduler' (or rather, the finite state machine).
 628  */
 629 static int
 630 __rpc_execute(struct rpc_task *task)
 631 {
 632         int             status = 0;
 633
 634         dprintk("RPC: %4d rpc_execute flgs %x\n",
 635                                 task->tk_pid, task->tk_flags);
 636
 637         if (!RPC_IS_RUNNING(task)) {
 638                 printk(KERN_WARNING "RPC: rpc_execute called for sleeping task!!\n");
 639                 return 0;
 640         }
 641
 642  restarted:
 643         while (1) {
 644                 /*
 645                  * Execute any pending callback.
 646                  */
 647                 if (RPC_DO_CALLBACK(task)) {
 648                         /* Define a callback save pointer */
 649                         void (*save_callback)(struct rpc_task *);
 650
 651                         /*
 652                          * If a callback exists, save it, reset it,
 653                          * call it.
 654                          * The save is needed to stop from resetting
 655                          * another callback set within the callback handler
 656                          * - Dave
 657                          */
 658                         save_callback=task->tk_callback;
 659                         task->tk_callback=NULL;
 660                         save_callback(task);
 661                 }
 662
 663                 /*
 664                  * Perform the next FSM step.
 665                  * tk_action may be NULL when the task has been killed
 666                  * by someone else.
 667                  */
 668                 if (RPC_IS_RUNNING(task)) {
 669                         /*
 670                          * Garbage collection of pending timers...
 671                          */
 672                         rpc_delete_timer(task);
 673                         if (!task->tk_action)
 674                                 break;
 675                         task->tk_action(task);
 676                         /* micro-optimization to avoid spinlock */
 677                         if (RPC_IS_RUNNING(task))
 678                                 continue;
 679                 }
 680
 681                 /*
 682                  * Check whether task is sleeping.
 683                  */
 684                 spin_lock_bh(&rpc_queue_lock);
 685                 if (!RPC_IS_RUNNING(task)) {
 686                         rpc_set_sleeping(task);
 687                         if (RPC_IS_ASYNC(task)) {
 688                                 spin_unlock_bh(&rpc_queue_lock);
 689                                 return 0;
 690                         }
 691                 }
 692                 spin_unlock_bh(&rpc_queue_lock);
 693
 694                 if (!RPC_IS_SLEEPING(task))
 695                         continue;
 696                 /* sync task: sleep here */
 697                 dprintk("RPC: %4d sync task going to sleep\n", task->tk_pid);
 698                 if (current->pid == rpciod_pid)
 699                         printk(KERN_ERR "RPC: rpciod waiting on sync task!\n");
 700
 701                 if (!task->tk_client->cl_intr) {
 702                         __wait_event(task->tk_wait, !RPC_IS_SLEEPING(task));
 703                 } else {
 704                         __wait_event_interruptible(task->tk_wait, !RPC_IS_SLEEPING(task), status);
 705                         /*
 706                          * When a sync task receives a signal, it exits with
 707                          * -ERESTARTSYS. In order to catch any callbacks that
 708                          * clean up after sleeping on some queue, we don't
 709                          * break the loop here, but go around once more.
 710                          */
 711                         if (status == -ERESTARTSYS) {
 712                                 dprintk("RPC: %4d got signal\n", task->tk_pid);
 713                                 task->tk_flags |= RPC_TASK_KILLED;
 714                                 rpc_exit(task, -ERESTARTSYS);
 715                                 rpc_wake_up_task(task);
 716                         }
 717                 }
 718                 dprintk("RPC: %4d sync task resuming\n", task->tk_pid);
 719         }
 720
 721         if (task->tk_exit) {
 722                 task->tk_exit(task);
 723                 /* If tk_action is non-null, the user wants us to restart */
 724                 if (task->tk_action) {
 725                         if (!RPC_ASSASSINATED(task)) {
 726                                 /* Release RPC slot and buffer memory */
 727                                 if (task->tk_rqstp)
 728                                         xprt_release(task);
 729                                 rpc_free(task);
 730                                 goto restarted;
 731                         }
 732                         printk(KERN_ERR "RPC: dead task tries to walk away.\n");
 733                 }
 734         }
 735
 736         dprintk("RPC: %4d exit() = %d\n", task->tk_pid, task->tk_status);
 737         status = task->tk_status;
 738
 739         /* Release all resources associated with the task */
 740         rpc_release_task(task);
 741
 742         return status;
 743 }
 744
 745 /*
 746  * User-visible entry point to the scheduler.
 747  *
 748  * This may be called recursively if e.g. an async NFS task updates
 749  * the attributes and finds that dirty pages must be flushed.
 750  * NOTE: Upon exit of this function the task is guaranteed to be
 751  *       released. In particular note that tk_release() will have
 752  *       been called, so your task memory may have been freed.
 753  */
 754 int
 755 rpc_execute(struct rpc_task *task)
 756 {
 757         int status = -EIO;
 758         if (rpc_inhibit) {
 759                 printk(KERN_INFO "RPC: execution inhibited!\n");
 760                 goto out_release;
 761         }
 762
 763         status = -EWOULDBLOCK;
 764         if (task->tk_active) {
 765                 printk(KERN_ERR "RPC: active task was run twice!\n");
 766                 goto out_err;
 767         }
 768
 769         task->tk_active = 1;
 770         rpc_set_running(task);
 771         return __rpc_execute(task);
 772  out_release:
 773         rpc_release_task(task);
 774  out_err:
 775         return status;
 776 }
 777
 778 /*
 779  * This is our own little scheduler for async RPC tasks.
 780  */
 781 static void
 782 __rpc_schedule(void)
 783 {
 784         struct rpc_task *task;
 785         int             count = 0;
 786
 787         dprintk("RPC:      rpc_schedule enter\n");
 788         while (1) {
 789
 790                 task_for_first(task, &schedq.tasks[0]) {
 791                         __rpc_remove_wait_queue(task);
 792                         spin_unlock_bh(&rpc_queue_lock);
 793
 794                         __rpc_execute(task);
 795                         spin_lock_bh(&rpc_queue_lock);
 796                 } else {
 797                         break;
 798                 }
 799
 800                 if (++count >= 200 || need_resched()) {
 801                         count = 0;
 802                         spin_unlock_bh(&rpc_queue_lock);
 803                         schedule();
 804                         spin_lock_bh(&rpc_queue_lock);
 805                 }
 806         }
 807         dprintk("RPC:      rpc_schedule leave\n");
 808 }
 809
 810 /*
 811  * Allocate memory for RPC purposes.
 812  *
 813  * We try to ensure that some NFS reads and writes can always proceed
 814  * by using a mempool when allocating 'small' buffers.
 815  * In order to avoid memory starvation triggering more writebacks of
 816  * NFS requests, we use GFP_NOFS rather than GFP_KERNEL.
 817  */
 818 void *
 819 rpc_malloc(struct rpc_task *task, size_t size)
 820 {
 821         int     gfp;
 822
 823         if (task->tk_flags & RPC_TASK_SWAPPER)
 824                 gfp = GFP_ATOMIC;
 825         else
 826                 gfp = GFP_NOFS;
 827
 828         if (size > RPC_BUFFER_MAXSIZE) {
 829                 task->tk_buffer =  kmalloc(size, gfp);
 830                 if (task->tk_buffer)
 831                         task->tk_bufsize = size;
 832         } else {
 833                 task->tk_buffer =  mempool_alloc(rpc_buffer_mempool, gfp);
 834                 if (task->tk_buffer)
 835                         task->tk_bufsize = RPC_BUFFER_MAXSIZE;
 836         }
 837         return task->tk_buffer;
 838 }
 839
 840 void
 841 rpc_free(struct rpc_task *task)
 842 {
 843         if (task->tk_buffer) {
 844                 if (task->tk_bufsize == RPC_BUFFER_MAXSIZE)
 845                         mempool_free(task->tk_buffer, rpc_buffer_mempool);
 846                 else
 847                         kfree(task->tk_buffer);
 848                 task->tk_buffer = NULL;
 849                 task->tk_bufsize = 0;
 850         }
 851 }
 852
 853 /*
 854  * Creation and deletion of RPC task structures
 855  */
 856 void rpc_init_task(struct rpc_task *task, struct rpc_clnt *clnt, rpc_action callback, int flags)
 857 {
 858         memset(task, 0, sizeof(*task));
 859         init_timer(&task->tk_timer);
 860         task->tk_timer.data     = (unsigned long) task;
 861         task->tk_timer.function = (void (*)(unsigned long)) rpc_run_timer;
 862         task->tk_client = clnt;
 863         task->tk_flags  = flags;
 864         task->tk_exit   = callback;
 865         init_waitqueue_head(&task->tk_wait);
 866         if (current->uid != current->fsuid || current->gid != current->fsgid)
 867                 task->tk_flags |= RPC_TASK_SETUID;
 868
 869         /* Initialize retry counters */
 870         task->tk_garb_retry = 2;
 871         task->tk_cred_retry = 2;
 872         task->tk_suid_retry = 1;
 873
 874         task->tk_priority = RPC_PRIORITY_NORMAL;
 875         task->tk_cookie = (unsigned long)current;
 876         INIT_LIST_HEAD(&task->tk_links);
 877
 878         /* Add to global list of all tasks */
 879         spin_lock(&rpc_sched_lock);
 880         list_add(&task->tk_task, &all_tasks);
 881         spin_unlock(&rpc_sched_lock);
 882
 883         if (clnt) {
 884                 atomic_inc(&clnt->cl_users);
 885                 if (clnt->cl_softrtry)
 886                         task->tk_flags |= RPC_TASK_SOFT;
 887         }
 888
 889 #ifdef RPC_DEBUG
 890         task->tk_magic = 0xf00baa;
 891         task->tk_pid = rpc_task_id++;
 892 #endif
 893         dprintk("RPC: %4d new task procpid %d\n", task->tk_pid,
 894                                 current->pid);
 895 }
 896
 897 static struct rpc_task *
 898 rpc_alloc_task(void)
 899 {
 900         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
 901 }
 902
 903 static void
 904 rpc_default_free_task(struct rpc_task *task)
 905 {
 906         dprintk("RPC: %4d freeing task\n", task->tk_pid);
 907         mempool_free(task, rpc_task_mempool);
 908 }
 909
 910 /*
 911  * Create a new task for the specified client.  We have to
 912  * clean up after an allocation failure, as the client may
 913  * have specified "oneshot".
 914  */
 915 struct rpc_task *
 916 rpc_new_task(struct rpc_clnt *clnt, rpc_action callback, int flags)
 917 {
 918         struct rpc_task *task;
 919
 920         task = rpc_alloc_task();
 921         if (!task)
 922                 goto cleanup;
 923
 924         rpc_init_task(task, clnt, callback, flags);
 925
 926         /* Replace tk_release */
 927         task->tk_release = rpc_default_free_task;
 928
 929         dprintk("RPC: %4d allocated task\n", task->tk_pid);
 930         task->tk_flags |= RPC_TASK_DYNAMIC;
 931 out:
 932         return task;
 933
 934 cleanup:
 935         /* Check whether to release the client */
 936         if (clnt) {
 937                 printk("rpc_new_task: failed, users=%d, oneshot=%d\n",
 938                         atomic_read(&clnt->cl_users), clnt->cl_oneshot);
 939                 atomic_inc(&clnt->cl_users); /* pretend we were used ... */
 940                 rpc_release_client(clnt);
 941         }
 942         goto out;
 943 }
 944
 945 void
 946 rpc_release_task(struct rpc_task *task)
 947 {
 948         dprintk("RPC: %4d release task\n", task->tk_pid);
 949
 950 #ifdef RPC_DEBUG
 951         if (task->tk_magic != 0xf00baa) {
 952                 printk(KERN_ERR "RPC: attempt to release a non-existing task!\n");
 953                 rpc_debug = ~0;
 954                 rpc_show_tasks();
 955                 return;
 956         }
 957 #endif
 958
 959         /* Remove from global task list */
 960         spin_lock(&rpc_sched_lock);
 961         list_del(&task->tk_task);
 962         spin_unlock(&rpc_sched_lock);
 963
 964         /* Protect the execution below. */
 965         spin_lock_bh(&rpc_queue_lock);
 966
 967         /* Disable timer to prevent zombie wakeup */
 968         __rpc_disable_timer(task);
 969
 970         /* Remove from any wait queue we're still on */
 971         __rpc_remove_wait_queue(task);
 972
 973         task->tk_active = 0;
 974
 975         spin_unlock_bh(&rpc_queue_lock);
 976
 977         /* Synchronously delete any running timer */
 978         rpc_delete_timer(task);
 979
 980         /* Release resources */
 981         if (task->tk_rqstp)
 982                 xprt_release(task);
 983         if (task->tk_msg.rpc_cred)
 984                 rpcauth_unbindcred(task);
 985         rpc_free(task);
 986         if (task->tk_client) {
 987                 rpc_release_client(task->tk_client);
 988                 task->tk_client = NULL;
 989         }
 990
 991 #ifdef RPC_DEBUG
 992         task->tk_magic = 0;
 993 #endif
 994         if (task->tk_release)
 995                 task->tk_release(task);
 996 }
 997
 998 /**
 999  * rpc_find_parent - find the parent of a child task.
1000  * @child: child task
1001  *
1002  * Checks that the parent task is still sleeping on the
1003  * queue 'childq'. If so returns a pointer to the parent.
1004  * Upon failure returns NULL.
1005  *
1006  * Caller must hold rpc_queue_lock
1007  */
1008 static inline struct rpc_task *
1009 rpc_find_parent(struct rpc_task *child)
1010 {
1011         struct rpc_task *task, *parent;
1012         struct list_head *le;
1013
1014         parent = (struct rpc_task *) child->tk_calldata;
1015         task_for_each(task, le, &childq.tasks[0])
1016                 if (task == parent)
1017                         return parent;
1018
1019         return NULL;
1020 }
1021
1022 static void
1023 rpc_child_exit(struct rpc_task *child)
1024 {
1025         struct rpc_task *parent;
1026
1027         spin_lock_bh(&rpc_queue_lock);
1028         if ((parent = rpc_find_parent(child)) != NULL) {
1029                 parent->tk_status = child->tk_status;
1030                 __rpc_wake_up_task(parent);
1031         }
1032         spin_unlock_bh(&rpc_queue_lock);
1033 }
1034
1035 /*
1036  * Note: rpc_new_task releases the client after a failure.
1037  */
1038 struct rpc_task *
1039 rpc_new_child(struct rpc_clnt *clnt, struct rpc_task *parent)
1040 {
1041         struct rpc_task *task;
1042
1043         task = rpc_new_task(clnt, NULL, RPC_TASK_ASYNC | RPC_TASK_CHILD);
1044         if (!task)
1045                 goto fail;
1046         task->tk_exit = rpc_child_exit;
1047         task->tk_calldata = parent;
1048         return task;
1049
1050 fail:
1051         parent->tk_status = -ENOMEM;
1052         return NULL;
1053 }
1054
1055 void
1056 rpc_run_child(struct rpc_task *task, struct rpc_task *child, rpc_action func)
1057 {
1058         spin_lock_bh(&rpc_queue_lock);
1059         /* N.B. Is it possible for the child to have already finished? */
1060         __rpc_sleep_on(&childq, task, func, NULL);
1061         rpc_schedule_run(child);
1062         spin_unlock_bh(&rpc_queue_lock);
1063 }
1064
1065 /*
1066  * Kill all tasks for the given client.
1067  * XXX: kill their descendants as well?
1068  */
1069 void
1070 rpc_killall_tasks(struct rpc_clnt *clnt)
1071 {
1072         struct rpc_task *rovr;
1073         struct list_head *le;
1074
1075         dprintk("RPC:      killing all tasks for client %p\n", clnt);
1076
1077         /*
1078          * Spin lock all_tasks to prevent changes...
1079          */
1080         spin_lock(&rpc_sched_lock);
1081         alltask_for_each(rovr, le, &all_tasks)
1082                 if (!clnt || rovr->tk_client == clnt) {
1083                         rovr->tk_flags |= RPC_TASK_KILLED;
1084                         rpc_exit(rovr, -EIO);
1085                         rpc_wake_up_task(rovr);
1086                 }
1087         spin_unlock(&rpc_sched_lock);
1088 }
1089
1090 static DECLARE_MUTEX_LOCKED(rpciod_running);
1091
1092 static inline int
1093 rpciod_task_pending(void)
1094 {
1095         return !list_empty(&schedq.tasks[0]);
1096 }
1097
1098
1099 /*
1100  * This is the rpciod kernel thread
1101  */
1102 static int
1103 rpciod(void *ptr)
1104 {
1105         int             rounds = 0;
1106
1107         lock_kernel();
1108         /*
1109          * Let our maker know we're running ...
1110          */
1111         rpciod_pid = current->pid;
1112         up(&rpciod_running);
1113
1114         daemonize("rpciod");
1115         allow_signal(SIGKILL);
1116
1117         dprintk("RPC: rpciod starting (pid %d)\n", rpciod_pid);
1118         spin_lock_bh(&rpc_queue_lock);
1119         while (rpciod_users) {
1120                 DEFINE_WAIT(wait);
1121                 if (signalled()) {
1122                         spin_unlock_bh(&rpc_queue_lock);
1123                         rpciod_killall();
1124                         flush_signals(current);
1125                         spin_lock_bh(&rpc_queue_lock);
1126                 }
1127                 __rpc_schedule();
1128                 if (current->flags & PF_FREEZE) {
1129                         spin_unlock_bh(&rpc_queue_lock);
1130                         refrigerator(PF_FREEZE);
1131                         spin_lock_bh(&rpc_queue_lock);
1132                 }
1133
1134                 if (++rounds >= 64) {   /* safeguard */
1135                         spin_unlock_bh(&rpc_queue_lock);
1136                         schedule();
1137                         rounds = 0;
1138                         spin_lock_bh(&rpc_queue_lock);
1139                 }
1140
1141                 dprintk("RPC: rpciod back to sleep\n");
1142                 prepare_to_wait(&rpciod_idle, &wait, TASK_INTERRUPTIBLE);
1143                 if (!rpciod_task_pending() && !signalled()) {
1144                         spin_unlock_bh(&rpc_queue_lock);
1145                         schedule();
1146                         rounds = 0;
1147                         spin_lock_bh(&rpc_queue_lock);
1148                 }
1149                 finish_wait(&rpciod_idle, &wait);
1150                 dprintk("RPC: switch to rpciod\n");
1151         }
1152         spin_unlock_bh(&rpc_queue_lock);
1153
1154         dprintk("RPC: rpciod shutdown commences\n");
1155         if (!list_empty(&all_tasks)) {
1156                 printk(KERN_ERR "rpciod: active tasks at shutdown?!\n");
1157                 rpciod_killall();
1158         }
1159
1160         dprintk("RPC: rpciod exiting\n");
1161         unlock_kernel();
1162
1163         rpciod_pid = 0;
1164         complete_and_exit(&rpciod_killer, 0);
1165         return 0;
1166 }
1167
1168 static void
1169 rpciod_killall(void)
1170 {
1171         unsigned long flags;
1172
1173         while (!list_empty(&all_tasks)) {
1174                 clear_thread_flag(TIF_SIGPENDING);
1175                 rpc_killall_tasks(NULL);
1176                 spin_lock_bh(&rpc_queue_lock);
1177                 __rpc_schedule();
1178                 spin_unlock_bh(&rpc_queue_lock);
1179                 if (!list_empty(&all_tasks)) {
1180                         dprintk("rpciod_killall: waiting for tasks to exit\n");
1181                         yield();
1182                 }
1183         }
1184
1185         spin_lock_irqsave(&current->sighand->siglock, flags);
1186         recalc_sigpending();
1187         spin_unlock_irqrestore(&current->sighand->siglock, flags);
1188 }
1189
1190 /*
1191  * Start up the rpciod process if it's not already running.
1192  */
1193 int
1194 rpciod_up(void)
1195 {
1196         int error = 0;
1197
1198         down(&rpciod_sema);
1199         dprintk("rpciod_up: pid %d, users %d\n", rpciod_pid, rpciod_users);
1200         rpciod_users++;
1201         if (rpciod_pid)
1202                 goto out;
1203         /*
1204          * If there's no pid, we should be the first user.
1205          */
1206         if (rpciod_users > 1)
1207                 printk(KERN_WARNING "rpciod_up: no pid, %d users??\n", rpciod_users);
1208         /*
1209          * Create the rpciod thread and wait for it to start.
1210          */
1211         error = kernel_thread(rpciod, NULL, 0);
1212         if (error < 0) {
1213                 printk(KERN_WARNING "rpciod_up: create thread failed, error=%d\n", error);
1214                 rpciod_users--;
1215                 goto out;
1216         }
1217         down(&rpciod_running);
1218         error = 0;
1219 out:
1220         up(&rpciod_sema);
1221         return error;
1222 }
1223
1224 void
1225 rpciod_down(void)
1226 {
1227         down(&rpciod_sema);
1228         dprintk("rpciod_down pid %d sema %d\n", rpciod_pid, rpciod_users);
1229         if (rpciod_users) {
1230                 if (--rpciod_users)
1231                         goto out;
1232         } else
1233                 printk(KERN_WARNING "rpciod_down: pid=%d, no users??\n", rpciod_pid);
1234
1235         if (!rpciod_pid) {
1236                 dprintk("rpciod_down: Nothing to do!\n");
1237                 goto out;
1238         }
1239
1240         kill_proc(rpciod_pid, SIGKILL, 1);
1241         wait_for_completion(&rpciod_killer);
1242  out:
1243         up(&rpciod_sema);
1244 }
1245
1246 #ifdef RPC_DEBUG
1247 void rpc_show_tasks(void)
1248 {
1249         struct list_head *le;
1250         struct rpc_task *t;
1251
1252         spin_lock(&rpc_sched_lock);
1253         if (list_empty(&all_tasks)) {
1254                 spin_unlock(&rpc_sched_lock);
1255                 return;
1256         }
1257         printk("-pid- proc flgs status -client- -prog- --rqstp- -timeout "
1258                 "-rpcwait -action- --exit--\n");
1259         alltask_for_each(t, le, &all_tasks)
1260                 printk("%05d %04d %04x %06d %8p %6d %8p %08ld %8s %8p %8p\n",
1261                         t->tk_pid,
1262                         (t->tk_msg.rpc_proc ? t->tk_msg.rpc_proc->p_proc : -1),
1263                         t->tk_flags, t->tk_status,
1264                         t->tk_client,
1265                         (t->tk_client ? t->tk_client->cl_prog : 0),
1266                         t->tk_rqstp, t->tk_timeout,
1267                         rpc_qname(t->tk_rpcwait),
1268                         t->tk_action, t->tk_exit);
1269         spin_unlock(&rpc_sched_lock);
1270 }
1271 #endif
1272
1273 void
1274 rpc_destroy_mempool(void)
1275 {
1276         if (rpc_buffer_mempool)
1277                 mempool_destroy(rpc_buffer_mempool);
1278         if (rpc_task_mempool)
1279                 mempool_destroy(rpc_task_mempool);
1280         if (rpc_task_slabp && kmem_cache_destroy(rpc_task_slabp))
1281                 printk(KERN_INFO "rpc_task: not all structures were freed\n");
1282         if (rpc_buffer_slabp && kmem_cache_destroy(rpc_buffer_slabp))
1283                 printk(KERN_INFO "rpc_buffers: not all structures were freed\n");
1284 }
1285
1286 int
1287 rpc_init_mempool(void)
1288 {
1289         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1290                                              sizeof(struct rpc_task),
1291                                              0, SLAB_HWCACHE_ALIGN,
1292                                              NULL, NULL);
1293         if (!rpc_task_slabp)
1294                 goto err_nomem;
1295         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1296                                              RPC_BUFFER_MAXSIZE,
1297                                              0, SLAB_HWCACHE_ALIGN,
1298                                              NULL, NULL);
1299         if (!rpc_buffer_slabp)
1300                 goto err_nomem;
1301         rpc_task_mempool = mempool_create(RPC_TASK_POOLSIZE,
1302                                             mempool_alloc_slab,
1303                                             mempool_free_slab,
1304                                             rpc_task_slabp);
1305         if (!rpc_task_mempool)
1306                 goto err_nomem;
1307         rpc_buffer_mempool = mempool_create(RPC_BUFFER_POOLSIZE,
1308                                             mempool_alloc_slab,
1309                                             mempool_free_slab,
1310                                             rpc_buffer_slabp);
1311         if (!rpc_buffer_mempool)
1312                 goto err_nomem;
1313         return 0;
1314 err_nomem:
1315         rpc_destroy_mempool();
1316         return -ENOMEM;
1317 }