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 #include <linux/freezer.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #include "sunrpc.h"
  26
  27 #ifdef RPC_DEBUG
  28 #define RPCDBG_FACILITY         RPCDBG_SCHED
  29 #endif
  30
  31 #define CREATE_TRACE_POINTS
  32 #include <trace/events/sunrpc.h>
  33
  34 /*
  35  * RPC slabs and memory pools
  36  */
  37 #define RPC_BUFFER_MAXSIZE      (2048)
  38 #define RPC_BUFFER_POOLSIZE     (8)
  39 #define RPC_TASK_POOLSIZE       (8)
  40 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  41 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  42 static mempool_t        *rpc_task_mempool __read_mostly;
  43 static mempool_t        *rpc_buffer_mempool __read_mostly;
  44
  45 static void                     rpc_async_schedule(struct work_struct *);
  46 static void                      rpc_release_task(struct rpc_task *task);
  47 static void __rpc_queue_timer_fn(unsigned long ptr);
  48
  49 /*
  50  * RPC tasks sit here while waiting for conditions to improve.
  51  */
  52 static struct rpc_wait_queue delay_queue;
  53
  54 /*
  55  * rpciod-related stuff
  56  */
  57 struct workqueue_struct *rpciod_workqueue;
  58
  59 /*
  60  * Disable the timer for a given RPC task. Should be called with
  61  * queue->lock and bh_disabled in order to avoid races within
  62  * rpc_run_timer().
  63  */
  64 static void
  65 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  66 {
  67         if (task->tk_timeout == 0)
  68                 return;
  69         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  70         task->tk_timeout = 0;
  71         list_del(&task->u.tk_wait.timer_list);
  72         if (list_empty(&queue->timer_list.list))
  73                 del_timer(&queue->timer_list.timer);
  74 }
  75
  76 static void
  77 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  78 {
  79         queue->timer_list.expires = expires;
  80         mod_timer(&queue->timer_list.timer, expires);
  81 }
  82
  83 /*
  84  * Set up a timer for the current task.
  85  */
  86 static void
  87 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  88 {
  89         if (!task->tk_timeout)
  90                 return;
  91
  92         dprintk("RPC: %5u setting alarm for %lu ms\n",
  93                         task->tk_pid, task->tk_timeout * 1000 / HZ);
  94
  95         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  96         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  97                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  98         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
  99 }
 100
 101 /*
 102  * Add new request to a priority queue.
 103  */
 104 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 105                 struct rpc_task *task,
 106                 unsigned char queue_priority)
 107 {
 108         struct list_head *q;
 109         struct rpc_task *t;
 110
 111         INIT_LIST_HEAD(&task->u.tk_wait.links);
 112         q = &queue->tasks[queue_priority];
 113         if (unlikely(queue_priority > queue->maxpriority))
 114                 q = &queue->tasks[queue->maxpriority];
 115         list_for_each_entry(t, q, u.tk_wait.list) {
 116                 if (t->tk_owner == task->tk_owner) {
 117                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 118                         return;
 119                 }
 120         }
 121         list_add_tail(&task->u.tk_wait.list, q);
 122 }
 123
 124 /*
 125  * Add new request to wait queue.
 126  *
 127  * Swapper tasks always get inserted at the head of the queue.
 128  * This should avoid many nasty memory deadlocks and hopefully
 129  * improve overall performance.
 130  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 131  */
 132 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 133                 struct rpc_task *task,
 134                 unsigned char queue_priority)
 135 {
 136         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 137         if (RPC_IS_QUEUED(task))
 138                 return;
 139
 140         if (RPC_IS_PRIORITY(queue))
 141                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 142         else if (RPC_IS_SWAPPER(task))
 143                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 144         else
 145                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 146         task->tk_waitqueue = queue;
 147         queue->qlen++;
 148         rpc_set_queued(task);
 149
 150         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 151                         task->tk_pid, queue, rpc_qname(queue));
 152 }
 153
 154 /*
 155  * Remove request from a priority queue.
 156  */
 157 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 158 {
 159         struct rpc_task *t;
 160
 161         if (!list_empty(&task->u.tk_wait.links)) {
 162                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 163                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 164                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 165         }
 166 }
 167
 168 /*
 169  * Remove request from queue.
 170  * Note: must be called with spin lock held.
 171  */
 172 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 173 {
 174         __rpc_disable_timer(queue, task);
 175         if (RPC_IS_PRIORITY(queue))
 176                 __rpc_remove_wait_queue_priority(task);
 177         list_del(&task->u.tk_wait.list);
 178         queue->qlen--;
 179         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 180                         task->tk_pid, queue, rpc_qname(queue));
 181 }
 182
 183 static inline void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 184 {
 185         queue->priority = priority;
 186         queue->count = 1 << (priority * 2);
 187 }
 188
 189 static inline void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 190 {
 191         queue->owner = pid;
 192         queue->nr = RPC_BATCH_COUNT;
 193 }
 194
 195 static inline void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 196 {
 197         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 198         rpc_set_waitqueue_owner(queue, 0);
 199 }
 200
 201 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 202 {
 203         int i;
 204
 205         spin_lock_init(&queue->lock);
 206         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 207                 INIT_LIST_HEAD(&queue->tasks[i]);
 208         queue->maxpriority = nr_queues - 1;
 209         rpc_reset_waitqueue_priority(queue);
 210         queue->qlen = 0;
 211         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 212         INIT_LIST_HEAD(&queue->timer_list.list);
 213         rpc_assign_waitqueue_name(queue, qname);
 214 }
 215
 216 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 217 {
 218         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 219 }
 220 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 221
 222 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 223 {
 224         __rpc_init_priority_wait_queue(queue, qname, 1);
 225 }
 226 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 227
 228 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 229 {
 230         del_timer_sync(&queue->timer_list.timer);
 231 }
 232 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 233
 234 static int rpc_wait_bit_killable(void *word)
 235 {
 236         if (fatal_signal_pending(current))
 237                 return -ERESTARTSYS;
 238         freezable_schedule();
 239         return 0;
 240 }
 241
 242 #ifdef RPC_DEBUG
 243 static void rpc_task_set_debuginfo(struct rpc_task *task)
 244 {
 245         static atomic_t rpc_pid;
 246
 247         task->tk_pid = atomic_inc_return(&rpc_pid);
 248 }
 249 #else
 250 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 251 {
 252 }
 253 #endif
 254
 255 static void rpc_set_active(struct rpc_task *task)
 256 {
 257         trace_rpc_task_begin(task->tk_client, task, NULL);
 258
 259         rpc_task_set_debuginfo(task);
 260         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 261 }
 262
 263 /*
 264  * Mark an RPC call as having completed by clearing the 'active' bit
 265  * and then waking up all tasks that were sleeping.
 266  */
 267 static int rpc_complete_task(struct rpc_task *task)
 268 {
 269         void *m = &task->tk_runstate;
 270         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 271         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 272         unsigned long flags;
 273         int ret;
 274
 275         trace_rpc_task_complete(task->tk_client, task, NULL);
 276
 277         spin_lock_irqsave(&wq->lock, flags);
 278         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 279         ret = atomic_dec_and_test(&task->tk_count);
 280         if (waitqueue_active(wq))
 281                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 282         spin_unlock_irqrestore(&wq->lock, flags);
 283         return ret;
 284 }
 285
 286 /*
 287  * Allow callers to wait for completion of an RPC call
 288  *
 289  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 290  * to enforce taking of the wq->lock and hence avoid races with
 291  * rpc_complete_task().
 292  */
 293 int __rpc_wait_for_completion_task(struct rpc_task *task, int (*action)(void *))
 294 {
 295         if (action == NULL)
 296                 action = rpc_wait_bit_killable;
 297         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 298                         action, TASK_KILLABLE);
 299 }
 300 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 301
 302 /*
 303  * Make an RPC task runnable.
 304  *
 305  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 306  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 307  * the wait queue operation.
 308  */
 309 static void rpc_make_runnable(struct rpc_task *task)
 310 {
 311         rpc_clear_queued(task);
 312         if (rpc_test_and_set_running(task))
 313                 return;
 314         if (RPC_IS_ASYNC(task)) {
 315                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 316                 queue_work(rpciod_workqueue, &task->u.tk_work);
 317         } else
 318                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 319 }
 320
 321 /*
 322  * Prepare for sleeping on a wait queue.
 323  * By always appending tasks to the list we ensure FIFO behavior.
 324  * NB: An RPC task will only receive interrupt-driven events as long
 325  * as it's on a wait queue.
 326  */
 327 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 328                 struct rpc_task *task,
 329                 rpc_action action,
 330                 unsigned char queue_priority)
 331 {
 332         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 333                         task->tk_pid, rpc_qname(q), jiffies);
 334
 335         trace_rpc_task_sleep(task->tk_client, task, q);
 336
 337         __rpc_add_wait_queue(q, task, queue_priority);
 338
 339         WARN_ON_ONCE(task->tk_callback != NULL);
 340         task->tk_callback = action;
 341         __rpc_add_timer(q, task);
 342 }
 343
 344 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 345                                 rpc_action action)
 346 {
 347         /* We shouldn't ever put an inactive task to sleep */
 348         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 349         if (!RPC_IS_ACTIVATED(task)) {
 350                 task->tk_status = -EIO;
 351                 rpc_put_task_async(task);
 352                 return;
 353         }
 354
 355         /*
 356          * Protect the queue operations.
 357          */
 358         spin_lock_bh(&q->lock);
 359         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 360         spin_unlock_bh(&q->lock);
 361 }
 362 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 363
 364 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 365                 rpc_action action, int priority)
 366 {
 367         /* We shouldn't ever put an inactive task to sleep */
 368         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 369         if (!RPC_IS_ACTIVATED(task)) {
 370                 task->tk_status = -EIO;
 371                 rpc_put_task_async(task);
 372                 return;
 373         }
 374
 375         /*
 376          * Protect the queue operations.
 377          */
 378         spin_lock_bh(&q->lock);
 379         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 380         spin_unlock_bh(&q->lock);
 381 }
 382
 383 /**
 384  * __rpc_do_wake_up_task - wake up a single rpc_task
 385  * @queue: wait queue
 386  * @task: task to be woken up
 387  *
 388  * Caller must hold queue->lock, and have cleared the task queued flag.
 389  */
 390 static void __rpc_do_wake_up_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 391 {
 392         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 393                         task->tk_pid, jiffies);
 394
 395         /* Has the task been executed yet? If not, we cannot wake it up! */
 396         if (!RPC_IS_ACTIVATED(task)) {
 397                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 398                 return;
 399         }
 400
 401         trace_rpc_task_wakeup(task->tk_client, task, queue);
 402
 403         __rpc_remove_wait_queue(queue, task);
 404
 405         rpc_make_runnable(task);
 406
 407         dprintk("RPC:       __rpc_wake_up_task done\n");
 408 }
 409
 410 /*
 411  * Wake up a queued task while the queue lock is being held
 412  */
 413 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 414 {
 415         if (RPC_IS_QUEUED(task) && task->tk_waitqueue == queue)
 416                 __rpc_do_wake_up_task(queue, task);
 417 }
 418
 419 /*
 420  * Tests whether rpc queue is empty
 421  */
 422 int rpc_queue_empty(struct rpc_wait_queue *queue)
 423 {
 424         int res;
 425
 426         spin_lock_bh(&queue->lock);
 427         res = queue->qlen;
 428         spin_unlock_bh(&queue->lock);
 429         return res == 0;
 430 }
 431 EXPORT_SYMBOL_GPL(rpc_queue_empty);
 432
 433 /*
 434  * Wake up a task on a specific queue
 435  */
 436 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 437 {
 438         spin_lock_bh(&queue->lock);
 439         rpc_wake_up_task_queue_locked(queue, task);
 440         spin_unlock_bh(&queue->lock);
 441 }
 442 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 443
 444 /*
 445  * Wake up the next task on a priority queue.
 446  */
 447 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 448 {
 449         struct list_head *q;
 450         struct rpc_task *task;
 451
 452         /*
 453          * Service a batch of tasks from a single owner.
 454          */
 455         q = &queue->tasks[queue->priority];
 456         if (!list_empty(q)) {
 457                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 458                 if (queue->owner == task->tk_owner) {
 459                         if (--queue->nr)
 460                                 goto out;
 461                         list_move_tail(&task->u.tk_wait.list, q);
 462                 }
 463                 /*
 464                  * Check if we need to switch queues.
 465                  */
 466                 if (--queue->count)
 467                         goto new_owner;
 468         }
 469
 470         /*
 471          * Service the next queue.
 472          */
 473         do {
 474                 if (q == &queue->tasks[0])
 475                         q = &queue->tasks[queue->maxpriority];
 476                 else
 477                         q = q - 1;
 478                 if (!list_empty(q)) {
 479                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 480                         goto new_queue;
 481                 }
 482         } while (q != &queue->tasks[queue->priority]);
 483
 484         rpc_reset_waitqueue_priority(queue);
 485         return NULL;
 486
 487 new_queue:
 488         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 489 new_owner:
 490         rpc_set_waitqueue_owner(queue, task->tk_owner);
 491 out:
 492         return task;
 493 }
 494
 495 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 496 {
 497         if (RPC_IS_PRIORITY(queue))
 498                 return __rpc_find_next_queued_priority(queue);
 499         if (!list_empty(&queue->tasks[0]))
 500                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 501         return NULL;
 502 }
 503
 504 /*
 505  * Wake up the first task on the wait queue.
 506  */
 507 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 508                 bool (*func)(struct rpc_task *, void *), void *data)
 509 {
 510         struct rpc_task *task = NULL;
 511
 512         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 513                         queue, rpc_qname(queue));
 514         spin_lock_bh(&queue->lock);
 515         task = __rpc_find_next_queued(queue);
 516         if (task != NULL) {
 517                 if (func(task, data))
 518                         rpc_wake_up_task_queue_locked(queue, task);
 519                 else
 520                         task = NULL;
 521         }
 522         spin_unlock_bh(&queue->lock);
 523
 524         return task;
 525 }
 526 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 527
 528 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 529 {
 530         return true;
 531 }
 532
 533 /*
 534  * Wake up the next task on the wait queue.
 535 */
 536 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 537 {
 538         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 539 }
 540 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 541
 542 /**
 543  * rpc_wake_up - wake up all rpc_tasks
 544  * @queue: rpc_wait_queue on which the tasks are sleeping
 545  *
 546  * Grabs queue->lock
 547  */
 548 void rpc_wake_up(struct rpc_wait_queue *queue)
 549 {
 550         struct list_head *head;
 551
 552         spin_lock_bh(&queue->lock);
 553         head = &queue->tasks[queue->maxpriority];
 554         for (;;) {
 555                 while (!list_empty(head)) {
 556                         struct rpc_task *task;
 557                         task = list_first_entry(head,
 558                                         struct rpc_task,
 559                                         u.tk_wait.list);
 560                         rpc_wake_up_task_queue_locked(queue, task);
 561                 }
 562                 if (head == &queue->tasks[0])
 563                         break;
 564                 head--;
 565         }
 566         spin_unlock_bh(&queue->lock);
 567 }
 568 EXPORT_SYMBOL_GPL(rpc_wake_up);
 569
 570 /**
 571  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 572  * @queue: rpc_wait_queue on which the tasks are sleeping
 573  * @status: status value to set
 574  *
 575  * Grabs queue->lock
 576  */
 577 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 578 {
 579         struct list_head *head;
 580
 581         spin_lock_bh(&queue->lock);
 582         head = &queue->tasks[queue->maxpriority];
 583         for (;;) {
 584                 while (!list_empty(head)) {
 585                         struct rpc_task *task;
 586                         task = list_first_entry(head,
 587                                         struct rpc_task,
 588                                         u.tk_wait.list);
 589                         task->tk_status = status;
 590                         rpc_wake_up_task_queue_locked(queue, task);
 591                 }
 592                 if (head == &queue->tasks[0])
 593                         break;
 594                 head--;
 595         }
 596         spin_unlock_bh(&queue->lock);
 597 }
 598 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 599
 600 static void __rpc_queue_timer_fn(unsigned long ptr)
 601 {
 602         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 603         struct rpc_task *task, *n;
 604         unsigned long expires, now, timeo;
 605
 606         spin_lock(&queue->lock);
 607         expires = now = jiffies;
 608         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 609                 timeo = task->u.tk_wait.expires;
 610                 if (time_after_eq(now, timeo)) {
 611                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 612                         task->tk_status = -ETIMEDOUT;
 613                         rpc_wake_up_task_queue_locked(queue, task);
 614                         continue;
 615                 }
 616                 if (expires == now || time_after(expires, timeo))
 617                         expires = timeo;
 618         }
 619         if (!list_empty(&queue->timer_list.list))
 620                 rpc_set_queue_timer(queue, expires);
 621         spin_unlock(&queue->lock);
 622 }
 623
 624 static void __rpc_atrun(struct rpc_task *task)
 625 {
 626         task->tk_status = 0;
 627 }
 628
 629 /*
 630  * Run a task at a later time
 631  */
 632 void rpc_delay(struct rpc_task *task, unsigned long delay)
 633 {
 634         task->tk_timeout = delay;
 635         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 636 }
 637 EXPORT_SYMBOL_GPL(rpc_delay);
 638
 639 /*
 640  * Helper to call task->tk_ops->rpc_call_prepare
 641  */
 642 void rpc_prepare_task(struct rpc_task *task)
 643 {
 644         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 645 }
 646
 647 static void
 648 rpc_init_task_statistics(struct rpc_task *task)
 649 {
 650         /* Initialize retry counters */
 651         task->tk_garb_retry = 2;
 652         task->tk_cred_retry = 2;
 653         task->tk_rebind_retry = 2;
 654
 655         /* starting timestamp */
 656         task->tk_start = ktime_get();
 657 }
 658
 659 static void
 660 rpc_reset_task_statistics(struct rpc_task *task)
 661 {
 662         task->tk_timeouts = 0;
 663         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 664
 665         rpc_init_task_statistics(task);
 666 }
 667
 668 /*
 669  * Helper that calls task->tk_ops->rpc_call_done if it exists
 670  */
 671 void rpc_exit_task(struct rpc_task *task)
 672 {
 673         task->tk_action = NULL;
 674         if (task->tk_ops->rpc_call_done != NULL) {
 675                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 676                 if (task->tk_action != NULL) {
 677                         WARN_ON(RPC_ASSASSINATED(task));
 678                         /* Always release the RPC slot and buffer memory */
 679                         xprt_release(task);
 680                         rpc_reset_task_statistics(task);
 681                 }
 682         }
 683 }
 684
 685 void rpc_exit(struct rpc_task *task, int status)
 686 {
 687         task->tk_status = status;
 688         task->tk_action = rpc_exit_task;
 689         if (RPC_IS_QUEUED(task))
 690                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 691 }
 692 EXPORT_SYMBOL_GPL(rpc_exit);
 693
 694 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 695 {
 696         if (ops->rpc_release != NULL)
 697                 ops->rpc_release(calldata);
 698 }
 699
 700 /*
 701  * This is the RPC `scheduler' (or rather, the finite state machine).
 702  */
 703 static void __rpc_execute(struct rpc_task *task)
 704 {
 705         struct rpc_wait_queue *queue;
 706         int task_is_async = RPC_IS_ASYNC(task);
 707         int status = 0;
 708
 709         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 710                         task->tk_pid, task->tk_flags);
 711
 712         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 713         if (RPC_IS_QUEUED(task))
 714                 return;
 715
 716         for (;;) {
 717                 void (*do_action)(struct rpc_task *);
 718
 719                 /*
 720                  * Execute any pending callback first.
 721                  */
 722                 do_action = task->tk_callback;
 723                 task->tk_callback = NULL;
 724                 if (do_action == NULL) {
 725                         /*
 726                          * Perform the next FSM step.
 727                          * tk_action may be NULL if the task has been killed.
 728                          * In particular, note that rpc_killall_tasks may
 729                          * do this at any time, so beware when dereferencing.
 730                          */
 731                         do_action = task->tk_action;
 732                         if (do_action == NULL)
 733                                 break;
 734                 }
 735                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 736                 do_action(task);
 737
 738                 /*
 739                  * Lockless check for whether task is sleeping or not.
 740                  */
 741                 if (!RPC_IS_QUEUED(task))
 742                         continue;
 743                 /*
 744                  * The queue->lock protects against races with
 745                  * rpc_make_runnable().
 746                  *
 747                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 748                  * rpc_task, rpc_make_runnable() can assign it to a
 749                  * different workqueue. We therefore cannot assume that the
 750                  * rpc_task pointer may still be dereferenced.
 751                  */
 752                 queue = task->tk_waitqueue;
 753                 spin_lock_bh(&queue->lock);
 754                 if (!RPC_IS_QUEUED(task)) {
 755                         spin_unlock_bh(&queue->lock);
 756                         continue;
 757                 }
 758                 rpc_clear_running(task);
 759                 spin_unlock_bh(&queue->lock);
 760                 if (task_is_async)
 761                         return;
 762
 763                 /* sync task: sleep here */
 764                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 765                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 766                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 767                                 TASK_KILLABLE);
 768                 if (status == -ERESTARTSYS) {
 769                         /*
 770                          * When a sync task receives a signal, it exits with
 771                          * -ERESTARTSYS. In order to catch any callbacks that
 772                          * clean up after sleeping on some queue, we don't
 773                          * break the loop here, but go around once more.
 774                          */
 775                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 776                         task->tk_flags |= RPC_TASK_KILLED;
 777                         rpc_exit(task, -ERESTARTSYS);
 778                 }
 779                 rpc_set_running(task);
 780                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 781         }
 782
 783         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 784                         task->tk_status);
 785         /* Release all resources associated with the task */
 786         rpc_release_task(task);
 787 }
 788
 789 /*
 790  * User-visible entry point to the scheduler.
 791  *
 792  * This may be called recursively if e.g. an async NFS task updates
 793  * the attributes and finds that dirty pages must be flushed.
 794  * NOTE: Upon exit of this function the task is guaranteed to be
 795  *       released. In particular note that tk_release() will have
 796  *       been called, so your task memory may have been freed.
 797  */
 798 void rpc_execute(struct rpc_task *task)
 799 {
 800         rpc_set_active(task);
 801         rpc_make_runnable(task);
 802         if (!RPC_IS_ASYNC(task))
 803                 __rpc_execute(task);
 804 }
 805
 806 static void rpc_async_schedule(struct work_struct *work)
 807 {
 808         current->flags |= PF_FSTRANS;
 809         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 810         current->flags &= ~PF_FSTRANS;
 811 }
 812
 813 /**
 814  * rpc_malloc - allocate an RPC buffer
 815  * @task: RPC task that will use this buffer
 816  * @size: requested byte size
 817  *
 818  * To prevent rpciod from hanging, this allocator never sleeps,
 819  * returning NULL if the request cannot be serviced immediately.
 820  * The caller can arrange to sleep in a way that is safe for rpciod.
 821  *
 822  * Most requests are 'small' (under 2KiB) and can be serviced from a
 823  * mempool, ensuring that NFS reads and writes can always proceed,
 824  * and that there is good locality of reference for these buffers.
 825  *
 826  * In order to avoid memory starvation triggering more writebacks of
 827  * NFS requests, we avoid using GFP_KERNEL.
 828  */
 829 void *rpc_malloc(struct rpc_task *task, size_t size)
 830 {
 831         struct rpc_buffer *buf;
 832         gfp_t gfp = GFP_NOWAIT;
 833
 834         if (RPC_IS_SWAPPER(task))
 835                 gfp |= __GFP_MEMALLOC;
 836
 837         size += sizeof(struct rpc_buffer);
 838         if (size <= RPC_BUFFER_MAXSIZE)
 839                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 840         else
 841                 buf = kmalloc(size, gfp);
 842
 843         if (!buf)
 844                 return NULL;
 845
 846         buf->len = size;
 847         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 848                         task->tk_pid, size, buf);
 849         return &buf->data;
 850 }
 851 EXPORT_SYMBOL_GPL(rpc_malloc);
 852
 853 /**
 854  * rpc_free - free buffer allocated via rpc_malloc
 855  * @buffer: buffer to free
 856  *
 857  */
 858 void rpc_free(void *buffer)
 859 {
 860         size_t size;
 861         struct rpc_buffer *buf;
 862
 863         if (!buffer)
 864                 return;
 865
 866         buf = container_of(buffer, struct rpc_buffer, data);
 867         size = buf->len;
 868
 869         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 870                         size, buf);
 871
 872         if (size <= RPC_BUFFER_MAXSIZE)
 873                 mempool_free(buf, rpc_buffer_mempool);
 874         else
 875                 kfree(buf);
 876 }
 877 EXPORT_SYMBOL_GPL(rpc_free);
 878
 879 /*
 880  * Creation and deletion of RPC task structures
 881  */
 882 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 883 {
 884         memset(task, 0, sizeof(*task));
 885         atomic_set(&task->tk_count, 1);
 886         task->tk_flags  = task_setup_data->flags;
 887         task->tk_ops = task_setup_data->callback_ops;
 888         task->tk_calldata = task_setup_data->callback_data;
 889         INIT_LIST_HEAD(&task->tk_task);
 890
 891         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 892         task->tk_owner = current->tgid;
 893
 894         /* Initialize workqueue for async tasks */
 895         task->tk_workqueue = task_setup_data->workqueue;
 896
 897         if (task->tk_ops->rpc_call_prepare != NULL)
 898                 task->tk_action = rpc_prepare_task;
 899
 900         rpc_init_task_statistics(task);
 901
 902         dprintk("RPC:       new task initialized, procpid %u\n",
 903                                 task_pid_nr(current));
 904 }
 905
 906 static struct rpc_task *
 907 rpc_alloc_task(void)
 908 {
 909         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 910 }
 911
 912 /*
 913  * Create a new task for the specified client.
 914  */
 915 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 916 {
 917         struct rpc_task *task = setup_data->task;
 918         unsigned short flags = 0;
 919
 920         if (task == NULL) {
 921                 task = rpc_alloc_task();
 922                 if (task == NULL) {
 923                         rpc_release_calldata(setup_data->callback_ops,
 924                                         setup_data->callback_data);
 925                         return ERR_PTR(-ENOMEM);
 926                 }
 927                 flags = RPC_TASK_DYNAMIC;
 928         }
 929
 930         rpc_init_task(task, setup_data);
 931         task->tk_flags |= flags;
 932         dprintk("RPC:       allocated task %p\n", task);
 933         return task;
 934 }
 935
 936 static void rpc_free_task(struct rpc_task *task)
 937 {
 938         const struct rpc_call_ops *tk_ops = task->tk_ops;
 939         void *calldata = task->tk_calldata;
 940
 941         if (task->tk_flags & RPC_TASK_DYNAMIC) {
 942                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
 943                 mempool_free(task, rpc_task_mempool);
 944         }
 945         rpc_release_calldata(tk_ops, calldata);
 946 }
 947
 948 static void rpc_async_release(struct work_struct *work)
 949 {
 950         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
 951 }
 952
 953 static void rpc_release_resources_task(struct rpc_task *task)
 954 {
 955         if (task->tk_rqstp)
 956                 xprt_release(task);
 957         if (task->tk_msg.rpc_cred) {
 958                 put_rpccred(task->tk_msg.rpc_cred);
 959                 task->tk_msg.rpc_cred = NULL;
 960         }
 961         rpc_task_release_client(task);
 962 }
 963
 964 static void rpc_final_put_task(struct rpc_task *task,
 965                 struct workqueue_struct *q)
 966 {
 967         if (q != NULL) {
 968                 INIT_WORK(&task->u.tk_work, rpc_async_release);
 969                 queue_work(q, &task->u.tk_work);
 970         } else
 971                 rpc_free_task(task);
 972 }
 973
 974 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
 975 {
 976         if (atomic_dec_and_test(&task->tk_count)) {
 977                 rpc_release_resources_task(task);
 978                 rpc_final_put_task(task, q);
 979         }
 980 }
 981
 982 void rpc_put_task(struct rpc_task *task)
 983 {
 984         rpc_do_put_task(task, NULL);
 985 }
 986 EXPORT_SYMBOL_GPL(rpc_put_task);
 987
 988 void rpc_put_task_async(struct rpc_task *task)
 989 {
 990         rpc_do_put_task(task, task->tk_workqueue);
 991 }
 992 EXPORT_SYMBOL_GPL(rpc_put_task_async);
 993
 994 static void rpc_release_task(struct rpc_task *task)
 995 {
 996         dprintk("RPC: %5u release task\n", task->tk_pid);
 997
 998         BUG_ON (RPC_IS_QUEUED(task));
 999
1000         rpc_release_resources_task(task);
1001
1002         /*
1003          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1004          * so it should be safe to use task->tk_count as a test for whether
1005          * or not any other processes still hold references to our rpc_task.
1006          */
1007         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1008                 /* Wake up anyone who may be waiting for task completion */
1009                 if (!rpc_complete_task(task))
1010                         return;
1011         } else {
1012                 if (!atomic_dec_and_test(&task->tk_count))
1013                         return;
1014         }
1015         rpc_final_put_task(task, task->tk_workqueue);
1016 }
1017
1018 int rpciod_up(void)
1019 {
1020         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1021 }
1022
1023 void rpciod_down(void)
1024 {
1025         module_put(THIS_MODULE);
1026 }
1027
1028 /*
1029  * Start up the rpciod workqueue.
1030  */
1031 static int rpciod_start(void)
1032 {
1033         struct workqueue_struct *wq;
1034
1035         /*
1036          * Create the rpciod thread and wait for it to start.
1037          */
1038         dprintk("RPC:       creating workqueue rpciod\n");
1039         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 1);
1040         rpciod_workqueue = wq;
1041         return rpciod_workqueue != NULL;
1042 }
1043
1044 static void rpciod_stop(void)
1045 {
1046         struct workqueue_struct *wq = NULL;
1047
1048         if (rpciod_workqueue == NULL)
1049                 return;
1050         dprintk("RPC:       destroying workqueue rpciod\n");
1051
1052         wq = rpciod_workqueue;
1053         rpciod_workqueue = NULL;
1054         destroy_workqueue(wq);
1055 }
1056
1057 void
1058 rpc_destroy_mempool(void)
1059 {
1060         rpciod_stop();
1061         if (rpc_buffer_mempool)
1062                 mempool_destroy(rpc_buffer_mempool);
1063         if (rpc_task_mempool)
1064                 mempool_destroy(rpc_task_mempool);
1065         if (rpc_task_slabp)
1066                 kmem_cache_destroy(rpc_task_slabp);
1067         if (rpc_buffer_slabp)
1068                 kmem_cache_destroy(rpc_buffer_slabp);
1069         rpc_destroy_wait_queue(&delay_queue);
1070 }
1071
1072 int
1073 rpc_init_mempool(void)
1074 {
1075         /*
1076          * The following is not strictly a mempool initialisation,
1077          * but there is no harm in doing it here
1078          */
1079         rpc_init_wait_queue(&delay_queue, "delayq");
1080         if (!rpciod_start())
1081                 goto err_nomem;
1082
1083         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1084                                              sizeof(struct rpc_task),
1085                                              0, SLAB_HWCACHE_ALIGN,
1086                                              NULL);
1087         if (!rpc_task_slabp)
1088                 goto err_nomem;
1089         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1090                                              RPC_BUFFER_MAXSIZE,
1091                                              0, SLAB_HWCACHE_ALIGN,
1092                                              NULL);
1093         if (!rpc_buffer_slabp)
1094                 goto err_nomem;
1095         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1096                                                     rpc_task_slabp);
1097         if (!rpc_task_mempool)
1098                 goto err_nomem;
1099         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1100                                                       rpc_buffer_slabp);
1101         if (!rpc_buffer_mempool)
1102                 goto err_nomem;
1103         return 0;
1104 err_nomem:
1105         rpc_destroy_mempool();
1106         return -ENOMEM;
1107 }