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