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, <[email protected]>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <[email protected]>
  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 #if IS_ENABLED(CONFIG_SUNRPC_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(struct timer_list *t);
  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 __read_mostly;
  58 struct workqueue_struct *xprtiod_workqueue __read_mostly;
  59
  60 /*
  61  * Disable the timer for a given RPC task. Should be called with
  62  * queue->lock and bh_disabled in order to avoid races within
  63  * rpc_run_timer().
  64  */
  65 static void
  66 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  67 {
  68         if (task->tk_timeout == 0)
  69                 return;
  70         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  71         task->tk_timeout = 0;
  72         list_del(&task->u.tk_wait.timer_list);
  73         if (list_empty(&queue->timer_list.list))
  74                 del_timer(&queue->timer_list.timer);
  75 }
  76
  77 static void
  78 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  79 {
  80         queue->timer_list.expires = expires;
  81         mod_timer(&queue->timer_list.timer, expires);
  82 }
  83
  84 /*
  85  * Set up a timer for the current task.
  86  */
  87 static void
  88 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  89 {
  90         if (!task->tk_timeout)
  91                 return;
  92
  93         dprintk("RPC: %5u setting alarm for %u ms\n",
  94                 task->tk_pid, jiffies_to_msecs(task->tk_timeout));
  95
  96         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  97         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  98                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  99         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
 100 }
 101
 102 static void rpc_rotate_queue_owner(struct rpc_wait_queue *queue)
 103 {
 104         struct list_head *q = &queue->tasks[queue->priority];
 105         struct rpc_task *task;
 106
 107         if (!list_empty(q)) {
 108                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 109                 if (task->tk_owner == queue->owner)
 110                         list_move_tail(&task->u.tk_wait.list, q);
 111         }
 112 }
 113
 114 static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 115 {
 116         if (queue->priority != priority) {
 117                 /* Fairness: rotate the list when changing priority */
 118                 rpc_rotate_queue_owner(queue);
 119                 queue->priority = priority;
 120         }
 121 }
 122
 123 static void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 124 {
 125         queue->owner = pid;
 126         queue->nr = RPC_BATCH_COUNT;
 127 }
 128
 129 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 130 {
 131         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 132         rpc_set_waitqueue_owner(queue, 0);
 133 }
 134
 135 /*
 136  * Add new request to a priority queue.
 137  */
 138 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 139                 struct rpc_task *task,
 140                 unsigned char queue_priority)
 141 {
 142         struct list_head *q;
 143         struct rpc_task *t;
 144
 145         INIT_LIST_HEAD(&task->u.tk_wait.links);
 146         if (unlikely(queue_priority > queue->maxpriority))
 147                 queue_priority = queue->maxpriority;
 148         if (queue_priority > queue->priority)
 149                 rpc_set_waitqueue_priority(queue, queue_priority);
 150         q = &queue->tasks[queue_priority];
 151         list_for_each_entry(t, q, u.tk_wait.list) {
 152                 if (t->tk_owner == task->tk_owner) {
 153                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 154                         return;
 155                 }
 156         }
 157         list_add_tail(&task->u.tk_wait.list, q);
 158 }
 159
 160 /*
 161  * Add new request to wait queue.
 162  *
 163  * Swapper tasks always get inserted at the head of the queue.
 164  * This should avoid many nasty memory deadlocks and hopefully
 165  * improve overall performance.
 166  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 167  */
 168 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 169                 struct rpc_task *task,
 170                 unsigned char queue_priority)
 171 {
 172         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 173         if (RPC_IS_QUEUED(task))
 174                 return;
 175
 176         if (RPC_IS_PRIORITY(queue))
 177                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 178         else if (RPC_IS_SWAPPER(task))
 179                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 180         else
 181                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 182         task->tk_waitqueue = queue;
 183         queue->qlen++;
 184         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 185         smp_wmb();
 186         rpc_set_queued(task);
 187
 188         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 189                         task->tk_pid, queue, rpc_qname(queue));
 190 }
 191
 192 /*
 193  * Remove request from a priority queue.
 194  */
 195 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 196 {
 197         struct rpc_task *t;
 198
 199         if (!list_empty(&task->u.tk_wait.links)) {
 200                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 201                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 202                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 203         }
 204 }
 205
 206 /*
 207  * Remove request from queue.
 208  * Note: must be called with spin lock held.
 209  */
 210 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 211 {
 212         __rpc_disable_timer(queue, task);
 213         if (RPC_IS_PRIORITY(queue))
 214                 __rpc_remove_wait_queue_priority(task);
 215         list_del(&task->u.tk_wait.list);
 216         queue->qlen--;
 217         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 218                         task->tk_pid, queue, rpc_qname(queue));
 219 }
 220
 221 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 222 {
 223         int i;
 224
 225         spin_lock_init(&queue->lock);
 226         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 227                 INIT_LIST_HEAD(&queue->tasks[i]);
 228         queue->maxpriority = nr_queues - 1;
 229         rpc_reset_waitqueue_priority(queue);
 230         queue->qlen = 0;
 231         timer_setup(&queue->timer_list.timer, __rpc_queue_timer_fn, 0);
 232         INIT_LIST_HEAD(&queue->timer_list.list);
 233         rpc_assign_waitqueue_name(queue, qname);
 234 }
 235
 236 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 237 {
 238         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 239 }
 240 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 241
 242 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 243 {
 244         __rpc_init_priority_wait_queue(queue, qname, 1);
 245 }
 246 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 247
 248 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 249 {
 250         del_timer_sync(&queue->timer_list.timer);
 251 }
 252 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 253
 254 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
 255 {
 256         freezable_schedule_unsafe();
 257         if (signal_pending_state(mode, current))
 258                 return -ERESTARTSYS;
 259         return 0;
 260 }
 261
 262 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
 263 static void rpc_task_set_debuginfo(struct rpc_task *task)
 264 {
 265         static atomic_t rpc_pid;
 266
 267         task->tk_pid = atomic_inc_return(&rpc_pid);
 268 }
 269 #else
 270 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 271 {
 272 }
 273 #endif
 274
 275 static void rpc_set_active(struct rpc_task *task)
 276 {
 277         rpc_task_set_debuginfo(task);
 278         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 279         trace_rpc_task_begin(task->tk_client, task, NULL);
 280 }
 281
 282 /*
 283  * Mark an RPC call as having completed by clearing the 'active' bit
 284  * and then waking up all tasks that were sleeping.
 285  */
 286 static int rpc_complete_task(struct rpc_task *task)
 287 {
 288         void *m = &task->tk_runstate;
 289         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 290         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 291         unsigned long flags;
 292         int ret;
 293
 294         trace_rpc_task_complete(task->tk_client, task, NULL);
 295
 296         spin_lock_irqsave(&wq->lock, flags);
 297         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 298         ret = atomic_dec_and_test(&task->tk_count);
 299         if (waitqueue_active(wq))
 300                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 301         spin_unlock_irqrestore(&wq->lock, flags);
 302         return ret;
 303 }
 304
 305 /*
 306  * Allow callers to wait for completion of an RPC call
 307  *
 308  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 309  * to enforce taking of the wq->lock and hence avoid races with
 310  * rpc_complete_task().
 311  */
 312 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
 313 {
 314         if (action == NULL)
 315                 action = rpc_wait_bit_killable;
 316         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 317                         action, TASK_KILLABLE);
 318 }
 319 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 320
 321 /*
 322  * Make an RPC task runnable.
 323  *
 324  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 325  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 326  * the wait queue operation.
 327  * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
 328  * which is needed to ensure that __rpc_execute() doesn't loop (due to the
 329  * lockless RPC_IS_QUEUED() test) before we've had a chance to test
 330  * the RPC_TASK_RUNNING flag.
 331  */
 332 static void rpc_make_runnable(struct workqueue_struct *wq,
 333                 struct rpc_task *task)
 334 {
 335         bool need_wakeup = !rpc_test_and_set_running(task);
 336
 337         rpc_clear_queued(task);
 338         if (!need_wakeup)
 339                 return;
 340         if (RPC_IS_ASYNC(task)) {
 341                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 342                 queue_work(wq, &task->u.tk_work);
 343         } else
 344                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 345 }
 346
 347 /*
 348  * Prepare for sleeping on a wait queue.
 349  * By always appending tasks to the list we ensure FIFO behavior.
 350  * NB: An RPC task will only receive interrupt-driven events as long
 351  * as it's on a wait queue.
 352  */
 353 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 354                 struct rpc_task *task,
 355                 rpc_action action,
 356                 unsigned char queue_priority)
 357 {
 358         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 359                         task->tk_pid, rpc_qname(q), jiffies);
 360
 361         trace_rpc_task_sleep(task->tk_client, task, q);
 362
 363         __rpc_add_wait_queue(q, task, queue_priority);
 364
 365         WARN_ON_ONCE(task->tk_callback != NULL);
 366         task->tk_callback = action;
 367         __rpc_add_timer(q, task);
 368 }
 369
 370 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 371                                 rpc_action action)
 372 {
 373         /* We shouldn't ever put an inactive task to sleep */
 374         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 375         if (!RPC_IS_ACTIVATED(task)) {
 376                 task->tk_status = -EIO;
 377                 rpc_put_task_async(task);
 378                 return;
 379         }
 380
 381         /*
 382          * Protect the queue operations.
 383          */
 384         spin_lock_bh(&q->lock);
 385         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 386         spin_unlock_bh(&q->lock);
 387 }
 388 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 389
 390 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 391                 rpc_action action, int priority)
 392 {
 393         /* We shouldn't ever put an inactive task to sleep */
 394         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 395         if (!RPC_IS_ACTIVATED(task)) {
 396                 task->tk_status = -EIO;
 397                 rpc_put_task_async(task);
 398                 return;
 399         }
 400
 401         /*
 402          * Protect the queue operations.
 403          */
 404         spin_lock_bh(&q->lock);
 405         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 406         spin_unlock_bh(&q->lock);
 407 }
 408 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 409
 410 /**
 411  * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
 412  * @wq: workqueue on which to run task
 413  * @queue: wait queue
 414  * @task: task to be woken up
 415  *
 416  * Caller must hold queue->lock, and have cleared the task queued flag.
 417  */
 418 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
 419                 struct rpc_wait_queue *queue,
 420                 struct rpc_task *task)
 421 {
 422         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 423                         task->tk_pid, jiffies);
 424
 425         /* Has the task been executed yet? If not, we cannot wake it up! */
 426         if (!RPC_IS_ACTIVATED(task)) {
 427                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 428                 return;
 429         }
 430
 431         trace_rpc_task_wakeup(task->tk_client, task, queue);
 432
 433         __rpc_remove_wait_queue(queue, task);
 434
 435         rpc_make_runnable(wq, task);
 436
 437         dprintk("RPC:       __rpc_wake_up_task done\n");
 438 }
 439
 440 /*
 441  * Wake up a queued task while the queue lock is being held
 442  */
 443 static void rpc_wake_up_task_on_wq_queue_locked(struct workqueue_struct *wq,
 444                 struct rpc_wait_queue *queue, struct rpc_task *task)
 445 {
 446         if (RPC_IS_QUEUED(task)) {
 447                 smp_rmb();
 448                 if (task->tk_waitqueue == queue)
 449                         __rpc_do_wake_up_task_on_wq(wq, queue, task);
 450         }
 451 }
 452
 453 /*
 454  * Wake up a queued task while the queue lock is being held
 455  */
 456 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 457 {
 458         rpc_wake_up_task_on_wq_queue_locked(rpciod_workqueue, queue, task);
 459 }
 460
 461 /*
 462  * Wake up a task on a specific queue
 463  */
 464 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 465 {
 466         spin_lock_bh(&queue->lock);
 467         rpc_wake_up_task_queue_locked(queue, task);
 468         spin_unlock_bh(&queue->lock);
 469 }
 470 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 471
 472 /*
 473  * Wake up the next task on a priority queue.
 474  */
 475 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 476 {
 477         struct list_head *q;
 478         struct rpc_task *task;
 479
 480         /*
 481          * Service a batch of tasks from a single owner.
 482          */
 483         q = &queue->tasks[queue->priority];
 484         if (!list_empty(q)) {
 485                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 486                 if (queue->owner == task->tk_owner) {
 487                         if (--queue->nr)
 488                                 goto out;
 489                         list_move_tail(&task->u.tk_wait.list, q);
 490                 }
 491                 /*
 492                  * Check if we need to switch queues.
 493                  */
 494                 goto new_owner;
 495         }
 496
 497         /*
 498          * Service the next queue.
 499          */
 500         do {
 501                 if (q == &queue->tasks[0])
 502                         q = &queue->tasks[queue->maxpriority];
 503                 else
 504                         q = q - 1;
 505                 if (!list_empty(q)) {
 506                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 507                         goto new_queue;
 508                 }
 509         } while (q != &queue->tasks[queue->priority]);
 510
 511         rpc_reset_waitqueue_priority(queue);
 512         return NULL;
 513
 514 new_queue:
 515         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 516 new_owner:
 517         rpc_set_waitqueue_owner(queue, task->tk_owner);
 518 out:
 519         return task;
 520 }
 521
 522 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 523 {
 524         if (RPC_IS_PRIORITY(queue))
 525                 return __rpc_find_next_queued_priority(queue);
 526         if (!list_empty(&queue->tasks[0]))
 527                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 528         return NULL;
 529 }
 530
 531 /*
 532  * Wake up the first task on the wait queue.
 533  */
 534 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 535                 struct rpc_wait_queue *queue,
 536                 bool (*func)(struct rpc_task *, void *), void *data)
 537 {
 538         struct rpc_task *task = NULL;
 539
 540         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 541                         queue, rpc_qname(queue));
 542         spin_lock_bh(&queue->lock);
 543         task = __rpc_find_next_queued(queue);
 544         if (task != NULL) {
 545                 if (func(task, data))
 546                         rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
 547                 else
 548                         task = NULL;
 549         }
 550         spin_unlock_bh(&queue->lock);
 551
 552         return task;
 553 }
 554
 555 /*
 556  * Wake up the first task on the wait queue.
 557  */
 558 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 559                 bool (*func)(struct rpc_task *, void *), void *data)
 560 {
 561         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 562 }
 563 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 564
 565 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 566 {
 567         return true;
 568 }
 569
 570 /*
 571  * Wake up the next task on the wait queue.
 572 */
 573 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 574 {
 575         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 576 }
 577 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 578
 579 /**
 580  * rpc_wake_up - wake up all rpc_tasks
 581  * @queue: rpc_wait_queue on which the tasks are sleeping
 582  *
 583  * Grabs queue->lock
 584  */
 585 void rpc_wake_up(struct rpc_wait_queue *queue)
 586 {
 587         struct list_head *head;
 588
 589         spin_lock_bh(&queue->lock);
 590         head = &queue->tasks[queue->maxpriority];
 591         for (;;) {
 592                 while (!list_empty(head)) {
 593                         struct rpc_task *task;
 594                         task = list_first_entry(head,
 595                                         struct rpc_task,
 596                                         u.tk_wait.list);
 597                         rpc_wake_up_task_queue_locked(queue, task);
 598                 }
 599                 if (head == &queue->tasks[0])
 600                         break;
 601                 head--;
 602         }
 603         spin_unlock_bh(&queue->lock);
 604 }
 605 EXPORT_SYMBOL_GPL(rpc_wake_up);
 606
 607 /**
 608  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 609  * @queue: rpc_wait_queue on which the tasks are sleeping
 610  * @status: status value to set
 611  *
 612  * Grabs queue->lock
 613  */
 614 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 615 {
 616         struct list_head *head;
 617
 618         spin_lock_bh(&queue->lock);
 619         head = &queue->tasks[queue->maxpriority];
 620         for (;;) {
 621                 while (!list_empty(head)) {
 622                         struct rpc_task *task;
 623                         task = list_first_entry(head,
 624                                         struct rpc_task,
 625                                         u.tk_wait.list);
 626                         task->tk_status = status;
 627                         rpc_wake_up_task_queue_locked(queue, task);
 628                 }
 629                 if (head == &queue->tasks[0])
 630                         break;
 631                 head--;
 632         }
 633         spin_unlock_bh(&queue->lock);
 634 }
 635 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 636
 637 static void __rpc_queue_timer_fn(struct timer_list *t)
 638 {
 639         struct rpc_wait_queue *queue = from_timer(queue, t, timer_list.timer);
 640         struct rpc_task *task, *n;
 641         unsigned long expires, now, timeo;
 642
 643         spin_lock(&queue->lock);
 644         expires = now = jiffies;
 645         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 646                 timeo = task->u.tk_wait.expires;
 647                 if (time_after_eq(now, timeo)) {
 648                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 649                         task->tk_status = -ETIMEDOUT;
 650                         rpc_wake_up_task_queue_locked(queue, task);
 651                         continue;
 652                 }
 653                 if (expires == now || time_after(expires, timeo))
 654                         expires = timeo;
 655         }
 656         if (!list_empty(&queue->timer_list.list))
 657                 rpc_set_queue_timer(queue, expires);
 658         spin_unlock(&queue->lock);
 659 }
 660
 661 static void __rpc_atrun(struct rpc_task *task)
 662 {
 663         if (task->tk_status == -ETIMEDOUT)
 664                 task->tk_status = 0;
 665 }
 666
 667 /*
 668  * Run a task at a later time
 669  */
 670 void rpc_delay(struct rpc_task *task, unsigned long delay)
 671 {
 672         task->tk_timeout = delay;
 673         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 674 }
 675 EXPORT_SYMBOL_GPL(rpc_delay);
 676
 677 /*
 678  * Helper to call task->tk_ops->rpc_call_prepare
 679  */
 680 void rpc_prepare_task(struct rpc_task *task)
 681 {
 682         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 683 }
 684
 685 static void
 686 rpc_init_task_statistics(struct rpc_task *task)
 687 {
 688         /* Initialize retry counters */
 689         task->tk_garb_retry = 2;
 690         task->tk_cred_retry = 2;
 691         task->tk_rebind_retry = 2;
 692
 693         /* starting timestamp */
 694         task->tk_start = ktime_get();
 695 }
 696
 697 static void
 698 rpc_reset_task_statistics(struct rpc_task *task)
 699 {
 700         task->tk_timeouts = 0;
 701         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 702
 703         rpc_init_task_statistics(task);
 704 }
 705
 706 /*
 707  * Helper that calls task->tk_ops->rpc_call_done if it exists
 708  */
 709 void rpc_exit_task(struct rpc_task *task)
 710 {
 711         task->tk_action = NULL;
 712         if (task->tk_ops->rpc_call_done != NULL) {
 713                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 714                 if (task->tk_action != NULL) {
 715                         WARN_ON(RPC_ASSASSINATED(task));
 716                         /* Always release the RPC slot and buffer memory */
 717                         xprt_release(task);
 718                         rpc_reset_task_statistics(task);
 719                 }
 720         }
 721 }
 722
 723 void rpc_exit(struct rpc_task *task, int status)
 724 {
 725         task->tk_status = status;
 726         task->tk_action = rpc_exit_task;
 727         if (RPC_IS_QUEUED(task))
 728                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 729 }
 730 EXPORT_SYMBOL_GPL(rpc_exit);
 731
 732 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 733 {
 734         if (ops->rpc_release != NULL)
 735                 ops->rpc_release(calldata);
 736 }
 737
 738 /*
 739  * This is the RPC `scheduler' (or rather, the finite state machine).
 740  */
 741 static void __rpc_execute(struct rpc_task *task)
 742 {
 743         struct rpc_wait_queue *queue;
 744         int task_is_async = RPC_IS_ASYNC(task);
 745         int status = 0;
 746
 747         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 748                         task->tk_pid, task->tk_flags);
 749
 750         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 751         if (RPC_IS_QUEUED(task))
 752                 return;
 753
 754         for (;;) {
 755                 void (*do_action)(struct rpc_task *);
 756
 757                 /*
 758                  * Execute any pending callback first.
 759                  */
 760                 do_action = task->tk_callback;
 761                 task->tk_callback = NULL;
 762                 if (do_action == NULL) {
 763                         /*
 764                          * Perform the next FSM step.
 765                          * tk_action may be NULL if the task has been killed.
 766                          * In particular, note that rpc_killall_tasks may
 767                          * do this at any time, so beware when dereferencing.
 768                          */
 769                         do_action = task->tk_action;
 770                         if (do_action == NULL)
 771                                 break;
 772                 }
 773                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 774                 do_action(task);
 775
 776                 /*
 777                  * Lockless check for whether task is sleeping or not.
 778                  */
 779                 if (!RPC_IS_QUEUED(task))
 780                         continue;
 781                 /*
 782                  * The queue->lock protects against races with
 783                  * rpc_make_runnable().
 784                  *
 785                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 786                  * rpc_task, rpc_make_runnable() can assign it to a
 787                  * different workqueue. We therefore cannot assume that the
 788                  * rpc_task pointer may still be dereferenced.
 789                  */
 790                 queue = task->tk_waitqueue;
 791                 spin_lock_bh(&queue->lock);
 792                 if (!RPC_IS_QUEUED(task)) {
 793                         spin_unlock_bh(&queue->lock);
 794                         continue;
 795                 }
 796                 rpc_clear_running(task);
 797                 spin_unlock_bh(&queue->lock);
 798                 if (task_is_async)
 799                         return;
 800
 801                 /* sync task: sleep here */
 802                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 803                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 804                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 805                                 TASK_KILLABLE);
 806                 if (status == -ERESTARTSYS) {
 807                         /*
 808                          * When a sync task receives a signal, it exits with
 809                          * -ERESTARTSYS. In order to catch any callbacks that
 810                          * clean up after sleeping on some queue, we don't
 811                          * break the loop here, but go around once more.
 812                          */
 813                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 814                         task->tk_flags |= RPC_TASK_KILLED;
 815                         rpc_exit(task, -ERESTARTSYS);
 816                 }
 817                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 818         }
 819
 820         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 821                         task->tk_status);
 822         /* Release all resources associated with the task */
 823         rpc_release_task(task);
 824 }
 825
 826 /*
 827  * User-visible entry point to the scheduler.
 828  *
 829  * This may be called recursively if e.g. an async NFS task updates
 830  * the attributes and finds that dirty pages must be flushed.
 831  * NOTE: Upon exit of this function the task is guaranteed to be
 832  *       released. In particular note that tk_release() will have
 833  *       been called, so your task memory may have been freed.
 834  */
 835 void rpc_execute(struct rpc_task *task)
 836 {
 837         bool is_async = RPC_IS_ASYNC(task);
 838
 839         rpc_set_active(task);
 840         rpc_make_runnable(rpciod_workqueue, task);
 841         if (!is_async)
 842                 __rpc_execute(task);
 843 }
 844
 845 static void rpc_async_schedule(struct work_struct *work)
 846 {
 847         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 848 }
 849
 850 /**
 851  * rpc_malloc - allocate RPC buffer resources
 852  * @task: RPC task
 853  *
 854  * A single memory region is allocated, which is split between the
 855  * RPC call and RPC reply that this task is being used for. When
 856  * this RPC is retired, the memory is released by calling rpc_free.
 857  *
 858  * To prevent rpciod from hanging, this allocator never sleeps,
 859  * returning -ENOMEM and suppressing warning if the request cannot
 860  * be serviced immediately. The caller can arrange to sleep in a
 861  * way that is safe for rpciod.
 862  *
 863  * Most requests are 'small' (under 2KiB) and can be serviced from a
 864  * mempool, ensuring that NFS reads and writes can always proceed,
 865  * and that there is good locality of reference for these buffers.
 866  *
 867  * In order to avoid memory starvation triggering more writebacks of
 868  * NFS requests, we avoid using GFP_KERNEL.
 869  */
 870 int rpc_malloc(struct rpc_task *task)
 871 {
 872         struct rpc_rqst *rqst = task->tk_rqstp;
 873         size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
 874         struct rpc_buffer *buf;
 875         gfp_t gfp = GFP_NOIO | __GFP_NOWARN;
 876
 877         if (RPC_IS_SWAPPER(task))
 878                 gfp = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 879
 880         size += sizeof(struct rpc_buffer);
 881         if (size <= RPC_BUFFER_MAXSIZE)
 882                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 883         else
 884                 buf = kmalloc(size, gfp);
 885
 886         if (!buf)
 887                 return -ENOMEM;
 888
 889         buf->len = size;
 890         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 891                         task->tk_pid, size, buf);
 892         rqst->rq_buffer = buf->data;
 893         rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
 894         return 0;
 895 }
 896 EXPORT_SYMBOL_GPL(rpc_malloc);
 897
 898 /**
 899  * rpc_free - free RPC buffer resources allocated via rpc_malloc
 900  * @task: RPC task
 901  *
 902  */
 903 void rpc_free(struct rpc_task *task)
 904 {
 905         void *buffer = task->tk_rqstp->rq_buffer;
 906         size_t size;
 907         struct rpc_buffer *buf;
 908
 909         buf = container_of(buffer, struct rpc_buffer, data);
 910         size = buf->len;
 911
 912         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 913                         size, buf);
 914
 915         if (size <= RPC_BUFFER_MAXSIZE)
 916                 mempool_free(buf, rpc_buffer_mempool);
 917         else
 918                 kfree(buf);
 919 }
 920 EXPORT_SYMBOL_GPL(rpc_free);
 921
 922 /*
 923  * Creation and deletion of RPC task structures
 924  */
 925 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 926 {
 927         memset(task, 0, sizeof(*task));
 928         atomic_set(&task->tk_count, 1);
 929         task->tk_flags  = task_setup_data->flags;
 930         task->tk_ops = task_setup_data->callback_ops;
 931         task->tk_calldata = task_setup_data->callback_data;
 932         INIT_LIST_HEAD(&task->tk_task);
 933
 934         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 935         task->tk_owner = current->tgid;
 936
 937         /* Initialize workqueue for async tasks */
 938         task->tk_workqueue = task_setup_data->workqueue;
 939
 940         task->tk_xprt = xprt_get(task_setup_data->rpc_xprt);
 941
 942         if (task->tk_ops->rpc_call_prepare != NULL)
 943                 task->tk_action = rpc_prepare_task;
 944
 945         rpc_init_task_statistics(task);
 946
 947         dprintk("RPC:       new task initialized, procpid %u\n",
 948                                 task_pid_nr(current));
 949 }
 950
 951 static struct rpc_task *
 952 rpc_alloc_task(void)
 953 {
 954         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 955 }
 956
 957 /*
 958  * Create a new task for the specified client.
 959  */
 960 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 961 {
 962         struct rpc_task *task = setup_data->task;
 963         unsigned short flags = 0;
 964
 965         if (task == NULL) {
 966                 task = rpc_alloc_task();
 967                 flags = RPC_TASK_DYNAMIC;
 968         }
 969
 970         rpc_init_task(task, setup_data);
 971         task->tk_flags |= flags;
 972         dprintk("RPC:       allocated task %p\n", task);
 973         return task;
 974 }
 975
 976 /*
 977  * rpc_free_task - release rpc task and perform cleanups
 978  *
 979  * Note that we free up the rpc_task _after_ rpc_release_calldata()
 980  * in order to work around a workqueue dependency issue.
 981  *
 982  * Tejun Heo states:
 983  * "Workqueue currently considers two work items to be the same if they're
 984  * on the same address and won't execute them concurrently - ie. it
 985  * makes a work item which is queued again while being executed wait
 986  * for the previous execution to complete.
 987  *
 988  * If a work function frees the work item, and then waits for an event
 989  * which should be performed by another work item and *that* work item
 990  * recycles the freed work item, it can create a false dependency loop.
 991  * There really is no reliable way to detect this short of verifying
 992  * every memory free."
 993  *
 994  */
 995 static void rpc_free_task(struct rpc_task *task)
 996 {
 997         unsigned short tk_flags = task->tk_flags;
 998
 999         rpc_release_calldata(task->tk_ops, task->tk_calldata);
1000
1001         if (tk_flags & RPC_TASK_DYNAMIC) {
1002                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1003                 mempool_free(task, rpc_task_mempool);
1004         }
1005 }
1006
1007 static void rpc_async_release(struct work_struct *work)
1008 {
1009         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1010 }
1011
1012 static void rpc_release_resources_task(struct rpc_task *task)
1013 {
1014         xprt_release(task);
1015         if (task->tk_msg.rpc_cred) {
1016                 put_rpccred(task->tk_msg.rpc_cred);
1017                 task->tk_msg.rpc_cred = NULL;
1018         }
1019         rpc_task_release_client(task);
1020 }
1021
1022 static void rpc_final_put_task(struct rpc_task *task,
1023                 struct workqueue_struct *q)
1024 {
1025         if (q != NULL) {
1026                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1027                 queue_work(q, &task->u.tk_work);
1028         } else
1029                 rpc_free_task(task);
1030 }
1031
1032 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1033 {
1034         if (atomic_dec_and_test(&task->tk_count)) {
1035                 rpc_release_resources_task(task);
1036                 rpc_final_put_task(task, q);
1037         }
1038 }
1039
1040 void rpc_put_task(struct rpc_task *task)
1041 {
1042         rpc_do_put_task(task, NULL);
1043 }
1044 EXPORT_SYMBOL_GPL(rpc_put_task);
1045
1046 void rpc_put_task_async(struct rpc_task *task)
1047 {
1048         rpc_do_put_task(task, task->tk_workqueue);
1049 }
1050 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1051
1052 static void rpc_release_task(struct rpc_task *task)
1053 {
1054         dprintk("RPC: %5u release task\n", task->tk_pid);
1055
1056         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1057
1058         rpc_release_resources_task(task);
1059
1060         /*
1061          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1062          * so it should be safe to use task->tk_count as a test for whether
1063          * or not any other processes still hold references to our rpc_task.
1064          */
1065         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1066                 /* Wake up anyone who may be waiting for task completion */
1067                 if (!rpc_complete_task(task))
1068                         return;
1069         } else {
1070                 if (!atomic_dec_and_test(&task->tk_count))
1071                         return;
1072         }
1073         rpc_final_put_task(task, task->tk_workqueue);
1074 }
1075
1076 int rpciod_up(void)
1077 {
1078         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1079 }
1080
1081 void rpciod_down(void)
1082 {
1083         module_put(THIS_MODULE);
1084 }
1085
1086 /*
1087  * Start up the rpciod workqueue.
1088  */
1089 static int rpciod_start(void)
1090 {
1091         struct workqueue_struct *wq;
1092
1093         /*
1094          * Create the rpciod thread and wait for it to start.
1095          */
1096         dprintk("RPC:       creating workqueue rpciod\n");
1097         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1098         if (!wq)
1099                 goto out_failed;
1100         rpciod_workqueue = wq;
1101         /* Note: highpri because network receive is latency sensitive */
1102         wq = alloc_workqueue("xprtiod", WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1103         if (!wq)
1104                 goto free_rpciod;
1105         xprtiod_workqueue = wq;
1106         return 1;
1107 free_rpciod:
1108         wq = rpciod_workqueue;
1109         rpciod_workqueue = NULL;
1110         destroy_workqueue(wq);
1111 out_failed:
1112         return 0;
1113 }
1114
1115 static void rpciod_stop(void)
1116 {
1117         struct workqueue_struct *wq = NULL;
1118
1119         if (rpciod_workqueue == NULL)
1120                 return;
1121         dprintk("RPC:       destroying workqueue rpciod\n");
1122
1123         wq = rpciod_workqueue;
1124         rpciod_workqueue = NULL;
1125         destroy_workqueue(wq);
1126         wq = xprtiod_workqueue;
1127         xprtiod_workqueue = NULL;
1128         destroy_workqueue(wq);
1129 }
1130
1131 void
1132 rpc_destroy_mempool(void)
1133 {
1134         rpciod_stop();
1135         mempool_destroy(rpc_buffer_mempool);
1136         mempool_destroy(rpc_task_mempool);
1137         kmem_cache_destroy(rpc_task_slabp);
1138         kmem_cache_destroy(rpc_buffer_slabp);
1139         rpc_destroy_wait_queue(&delay_queue);
1140 }
1141
1142 int
1143 rpc_init_mempool(void)
1144 {
1145         /*
1146          * The following is not strictly a mempool initialisation,
1147          * but there is no harm in doing it here
1148          */
1149         rpc_init_wait_queue(&delay_queue, "delayq");
1150         if (!rpciod_start())
1151                 goto err_nomem;
1152
1153         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1154                                              sizeof(struct rpc_task),
1155                                              0, SLAB_HWCACHE_ALIGN,
1156                                              NULL);
1157         if (!rpc_task_slabp)
1158                 goto err_nomem;
1159         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1160                                              RPC_BUFFER_MAXSIZE,
1161                                              0, SLAB_HWCACHE_ALIGN,
1162                                              NULL);
1163         if (!rpc_buffer_slabp)
1164                 goto err_nomem;
1165         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1166                                                     rpc_task_slabp);
1167         if (!rpc_task_mempool)
1168                 goto err_nomem;
1169         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1170                                                       rpc_buffer_slabp);
1171         if (!rpc_buffer_mempool)
1172                 goto err_nomem;
1173         return 0;
1174 err_nomem:
1175         rpc_destroy_mempool();
1176         return -ENOMEM;
1177 }