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 #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(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 __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_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 103 {
 104         if (queue->priority != priority) {
 105                 queue->priority = priority;
 106                 queue->nr = 1U << priority;
 107         }
 108 }
 109
 110 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 111 {
 112         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 113 }
 114
 115 /*
 116  * Add a request to a queue list
 117  */
 118 static void
 119 __rpc_list_enqueue_task(struct list_head *q, struct rpc_task *task)
 120 {
 121         struct rpc_task *t;
 122
 123         list_for_each_entry(t, q, u.tk_wait.list) {
 124                 if (t->tk_owner == task->tk_owner) {
 125                         list_add_tail(&task->u.tk_wait.links,
 126                                         &t->u.tk_wait.links);
 127                         /* Cache the queue head in task->u.tk_wait.list */
 128                         task->u.tk_wait.list.next = q;
 129                         task->u.tk_wait.list.prev = NULL;
 130                         return;
 131                 }
 132         }
 133         INIT_LIST_HEAD(&task->u.tk_wait.links);
 134         list_add_tail(&task->u.tk_wait.list, q);
 135 }
 136
 137 /*
 138  * Remove request from a queue list
 139  */
 140 static void
 141 __rpc_list_dequeue_task(struct rpc_task *task)
 142 {
 143         struct list_head *q;
 144         struct rpc_task *t;
 145
 146         if (task->u.tk_wait.list.prev == NULL) {
 147                 list_del(&task->u.tk_wait.links);
 148                 return;
 149         }
 150         if (!list_empty(&task->u.tk_wait.links)) {
 151                 t = list_first_entry(&task->u.tk_wait.links,
 152                                 struct rpc_task,
 153                                 u.tk_wait.links);
 154                 /* Assume __rpc_list_enqueue_task() cached the queue head */
 155                 q = t->u.tk_wait.list.next;
 156                 list_add_tail(&t->u.tk_wait.list, q);
 157                 list_del(&task->u.tk_wait.links);
 158         }
 159         list_del(&task->u.tk_wait.list);
 160 }
 161
 162 /*
 163  * Add new request to a priority queue.
 164  */
 165 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 166                 struct rpc_task *task,
 167                 unsigned char queue_priority)
 168 {
 169         if (unlikely(queue_priority > queue->maxpriority))
 170                 queue_priority = queue->maxpriority;
 171         __rpc_list_enqueue_task(&queue->tasks[queue_priority], task);
 172 }
 173
 174 /*
 175  * Add new request to wait queue.
 176  *
 177  * Swapper tasks always get inserted at the head of the queue.
 178  * This should avoid many nasty memory deadlocks and hopefully
 179  * improve overall performance.
 180  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 181  */
 182 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 183                 struct rpc_task *task,
 184                 unsigned char queue_priority)
 185 {
 186         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 187         if (RPC_IS_QUEUED(task))
 188                 return;
 189
 190         if (RPC_IS_PRIORITY(queue))
 191                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 192         else if (RPC_IS_SWAPPER(task))
 193                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 194         else
 195                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 196         task->tk_waitqueue = queue;
 197         queue->qlen++;
 198         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 199         smp_wmb();
 200         rpc_set_queued(task);
 201
 202         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 203                         task->tk_pid, queue, rpc_qname(queue));
 204 }
 205
 206 /*
 207  * Remove request from a priority queue.
 208  */
 209 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 210 {
 211         __rpc_list_dequeue_task(task);
 212 }
 213
 214 /*
 215  * Remove request from queue.
 216  * Note: must be called with spin lock held.
 217  */
 218 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 219 {
 220         __rpc_disable_timer(queue, task);
 221         if (RPC_IS_PRIORITY(queue))
 222                 __rpc_remove_wait_queue_priority(task);
 223         else
 224                 list_del(&task->u.tk_wait.list);
 225         queue->qlen--;
 226         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 227                         task->tk_pid, queue, rpc_qname(queue));
 228 }
 229
 230 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 231 {
 232         int i;
 233
 234         spin_lock_init(&queue->lock);
 235         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 236                 INIT_LIST_HEAD(&queue->tasks[i]);
 237         queue->maxpriority = nr_queues - 1;
 238         rpc_reset_waitqueue_priority(queue);
 239         queue->qlen = 0;
 240         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 241         INIT_LIST_HEAD(&queue->timer_list.list);
 242         rpc_assign_waitqueue_name(queue, qname);
 243 }
 244
 245 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 246 {
 247         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 248 }
 249 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 250
 251 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 252 {
 253         __rpc_init_priority_wait_queue(queue, qname, 1);
 254 }
 255 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 256
 257 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 258 {
 259         del_timer_sync(&queue->timer_list.timer);
 260 }
 261 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 262
 263 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
 264 {
 265         freezable_schedule_unsafe();
 266         if (signal_pending_state(mode, current))
 267                 return -ERESTARTSYS;
 268         return 0;
 269 }
 270
 271 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
 272 static void rpc_task_set_debuginfo(struct rpc_task *task)
 273 {
 274         static atomic_t rpc_pid;
 275
 276         task->tk_pid = atomic_inc_return(&rpc_pid);
 277 }
 278 #else
 279 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 280 {
 281 }
 282 #endif
 283
 284 static void rpc_set_active(struct rpc_task *task)
 285 {
 286         rpc_task_set_debuginfo(task);
 287         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 288         trace_rpc_task_begin(task->tk_client, task, NULL);
 289 }
 290
 291 /*
 292  * Mark an RPC call as having completed by clearing the 'active' bit
 293  * and then waking up all tasks that were sleeping.
 294  */
 295 static int rpc_complete_task(struct rpc_task *task)
 296 {
 297         void *m = &task->tk_runstate;
 298         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 299         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 300         unsigned long flags;
 301         int ret;
 302
 303         trace_rpc_task_complete(task->tk_client, task, NULL);
 304
 305         spin_lock_irqsave(&wq->lock, flags);
 306         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 307         ret = atomic_dec_and_test(&task->tk_count);
 308         if (waitqueue_active(wq))
 309                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 310         spin_unlock_irqrestore(&wq->lock, flags);
 311         return ret;
 312 }
 313
 314 /*
 315  * Allow callers to wait for completion of an RPC call
 316  *
 317  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 318  * to enforce taking of the wq->lock and hence avoid races with
 319  * rpc_complete_task().
 320  */
 321 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
 322 {
 323         if (action == NULL)
 324                 action = rpc_wait_bit_killable;
 325         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 326                         action, TASK_KILLABLE);
 327 }
 328 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 329
 330 /*
 331  * Make an RPC task runnable.
 332  *
 333  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 334  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 335  * the wait queue operation.
 336  * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
 337  * which is needed to ensure that __rpc_execute() doesn't loop (due to the
 338  * lockless RPC_IS_QUEUED() test) before we've had a chance to test
 339  * the RPC_TASK_RUNNING flag.
 340  */
 341 static void rpc_make_runnable(struct workqueue_struct *wq,
 342                 struct rpc_task *task)
 343 {
 344         bool need_wakeup = !rpc_test_and_set_running(task);
 345
 346         rpc_clear_queued(task);
 347         if (!need_wakeup)
 348                 return;
 349         if (RPC_IS_ASYNC(task)) {
 350                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 351                 queue_work(wq, &task->u.tk_work);
 352         } else
 353                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 354 }
 355
 356 /*
 357  * Prepare for sleeping on a wait queue.
 358  * By always appending tasks to the list we ensure FIFO behavior.
 359  * NB: An RPC task will only receive interrupt-driven events as long
 360  * as it's on a wait queue.
 361  */
 362 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 363                 struct rpc_task *task,
 364                 rpc_action action,
 365                 unsigned char queue_priority)
 366 {
 367         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 368                         task->tk_pid, rpc_qname(q), jiffies);
 369
 370         trace_rpc_task_sleep(task->tk_client, task, q);
 371
 372         __rpc_add_wait_queue(q, task, queue_priority);
 373
 374         WARN_ON_ONCE(task->tk_callback != NULL);
 375         task->tk_callback = action;
 376         __rpc_add_timer(q, task);
 377 }
 378
 379 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 380                                 rpc_action action)
 381 {
 382         /* We shouldn't ever put an inactive task to sleep */
 383         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 384         if (!RPC_IS_ACTIVATED(task)) {
 385                 task->tk_status = -EIO;
 386                 rpc_put_task_async(task);
 387                 return;
 388         }
 389
 390         /*
 391          * Protect the queue operations.
 392          */
 393         spin_lock_bh(&q->lock);
 394         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 395         spin_unlock_bh(&q->lock);
 396 }
 397 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 398
 399 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 400                 rpc_action action, int priority)
 401 {
 402         /* We shouldn't ever put an inactive task to sleep */
 403         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 404         if (!RPC_IS_ACTIVATED(task)) {
 405                 task->tk_status = -EIO;
 406                 rpc_put_task_async(task);
 407                 return;
 408         }
 409
 410         /*
 411          * Protect the queue operations.
 412          */
 413         spin_lock_bh(&q->lock);
 414         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 415         spin_unlock_bh(&q->lock);
 416 }
 417 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 418
 419 /**
 420  * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
 421  * @wq: workqueue on which to run task
 422  * @queue: wait queue
 423  * @task: task to be woken up
 424  *
 425  * Caller must hold queue->lock, and have cleared the task queued flag.
 426  */
 427 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
 428                 struct rpc_wait_queue *queue,
 429                 struct rpc_task *task)
 430 {
 431         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 432                         task->tk_pid, jiffies);
 433
 434         /* Has the task been executed yet? If not, we cannot wake it up! */
 435         if (!RPC_IS_ACTIVATED(task)) {
 436                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 437                 return;
 438         }
 439
 440         trace_rpc_task_wakeup(task->tk_client, task, queue);
 441
 442         __rpc_remove_wait_queue(queue, task);
 443
 444         rpc_make_runnable(wq, task);
 445
 446         dprintk("RPC:       __rpc_wake_up_task done\n");
 447 }
 448
 449 /*
 450  * Wake up a queued task while the queue lock is being held
 451  */
 452 static void rpc_wake_up_task_on_wq_queue_locked(struct workqueue_struct *wq,
 453                 struct rpc_wait_queue *queue, struct rpc_task *task)
 454 {
 455         if (RPC_IS_QUEUED(task)) {
 456                 smp_rmb();
 457                 if (task->tk_waitqueue == queue)
 458                         __rpc_do_wake_up_task_on_wq(wq, queue, task);
 459         }
 460 }
 461
 462 /*
 463  * Wake up a queued task while the queue lock is being held
 464  */
 465 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 466 {
 467         rpc_wake_up_task_on_wq_queue_locked(rpciod_workqueue, queue, task);
 468 }
 469
 470 /*
 471  * Wake up a task on a specific queue
 472  */
 473 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 474 {
 475         spin_lock_bh(&queue->lock);
 476         rpc_wake_up_task_queue_locked(queue, task);
 477         spin_unlock_bh(&queue->lock);
 478 }
 479 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 480
 481 /*
 482  * Wake up the next task on a priority queue.
 483  */
 484 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 485 {
 486         struct list_head *q;
 487         struct rpc_task *task;
 488
 489         /*
 490          * Service the privileged queue.
 491          */
 492         q = &queue->tasks[RPC_NR_PRIORITY - 1];
 493         if (queue->maxpriority > RPC_PRIORITY_PRIVILEGED && !list_empty(q)) {
 494                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 495                 goto out;
 496         }
 497
 498         /*
 499          * Service a batch of tasks from a single owner.
 500          */
 501         q = &queue->tasks[queue->priority];
 502         if (!list_empty(q) && queue->nr) {
 503                 queue->nr--;
 504                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 505                 goto out;
 506         }
 507
 508         /*
 509          * Service the next queue.
 510          */
 511         do {
 512                 if (q == &queue->tasks[0])
 513                         q = &queue->tasks[queue->maxpriority];
 514                 else
 515                         q = q - 1;
 516                 if (!list_empty(q)) {
 517                         task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 518                         goto new_queue;
 519                 }
 520         } while (q != &queue->tasks[queue->priority]);
 521
 522         rpc_reset_waitqueue_priority(queue);
 523         return NULL;
 524
 525 new_queue:
 526         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 527 out:
 528         return task;
 529 }
 530
 531 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 532 {
 533         if (RPC_IS_PRIORITY(queue))
 534                 return __rpc_find_next_queued_priority(queue);
 535         if (!list_empty(&queue->tasks[0]))
 536                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 537         return NULL;
 538 }
 539
 540 /*
 541  * Wake up the first task on the wait queue.
 542  */
 543 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 544                 struct rpc_wait_queue *queue,
 545                 bool (*func)(struct rpc_task *, void *), void *data)
 546 {
 547         struct rpc_task *task = NULL;
 548
 549         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 550                         queue, rpc_qname(queue));
 551         spin_lock_bh(&queue->lock);
 552         task = __rpc_find_next_queued(queue);
 553         if (task != NULL) {
 554                 if (func(task, data))
 555                         rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
 556                 else
 557                         task = NULL;
 558         }
 559         spin_unlock_bh(&queue->lock);
 560
 561         return task;
 562 }
 563
 564 /*
 565  * Wake up the first task on the wait queue.
 566  */
 567 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 568                 bool (*func)(struct rpc_task *, void *), void *data)
 569 {
 570         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 571 }
 572 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 573
 574 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 575 {
 576         return true;
 577 }
 578
 579 /*
 580  * Wake up the next task on the wait queue.
 581 */
 582 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 583 {
 584         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 585 }
 586 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 587
 588 /**
 589  * rpc_wake_up - wake up all rpc_tasks
 590  * @queue: rpc_wait_queue on which the tasks are sleeping
 591  *
 592  * Grabs queue->lock
 593  */
 594 void rpc_wake_up(struct rpc_wait_queue *queue)
 595 {
 596         struct list_head *head;
 597
 598         spin_lock_bh(&queue->lock);
 599         head = &queue->tasks[queue->maxpriority];
 600         for (;;) {
 601                 while (!list_empty(head)) {
 602                         struct rpc_task *task;
 603                         task = list_first_entry(head,
 604                                         struct rpc_task,
 605                                         u.tk_wait.list);
 606                         rpc_wake_up_task_queue_locked(queue, task);
 607                 }
 608                 if (head == &queue->tasks[0])
 609                         break;
 610                 head--;
 611         }
 612         spin_unlock_bh(&queue->lock);
 613 }
 614 EXPORT_SYMBOL_GPL(rpc_wake_up);
 615
 616 /**
 617  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 618  * @queue: rpc_wait_queue on which the tasks are sleeping
 619  * @status: status value to set
 620  *
 621  * Grabs queue->lock
 622  */
 623 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 624 {
 625         struct list_head *head;
 626
 627         spin_lock_bh(&queue->lock);
 628         head = &queue->tasks[queue->maxpriority];
 629         for (;;) {
 630                 while (!list_empty(head)) {
 631                         struct rpc_task *task;
 632                         task = list_first_entry(head,
 633                                         struct rpc_task,
 634                                         u.tk_wait.list);
 635                         task->tk_status = status;
 636                         rpc_wake_up_task_queue_locked(queue, task);
 637                 }
 638                 if (head == &queue->tasks[0])
 639                         break;
 640                 head--;
 641         }
 642         spin_unlock_bh(&queue->lock);
 643 }
 644 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 645
 646 static void __rpc_queue_timer_fn(unsigned long ptr)
 647 {
 648         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 649         struct rpc_task *task, *n;
 650         unsigned long expires, now, timeo;
 651
 652         spin_lock(&queue->lock);
 653         expires = now = jiffies;
 654         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 655                 timeo = task->u.tk_wait.expires;
 656                 if (time_after_eq(now, timeo)) {
 657                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 658                         task->tk_status = -ETIMEDOUT;
 659                         rpc_wake_up_task_queue_locked(queue, task);
 660                         continue;
 661                 }
 662                 if (expires == now || time_after(expires, timeo))
 663                         expires = timeo;
 664         }
 665         if (!list_empty(&queue->timer_list.list))
 666                 rpc_set_queue_timer(queue, expires);
 667         spin_unlock(&queue->lock);
 668 }
 669
 670 static void __rpc_atrun(struct rpc_task *task)
 671 {
 672         if (task->tk_status == -ETIMEDOUT)
 673                 task->tk_status = 0;
 674 }
 675
 676 /*
 677  * Run a task at a later time
 678  */
 679 void rpc_delay(struct rpc_task *task, unsigned long delay)
 680 {
 681         task->tk_timeout = delay;
 682         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 683 }
 684 EXPORT_SYMBOL_GPL(rpc_delay);
 685
 686 /*
 687  * Helper to call task->tk_ops->rpc_call_prepare
 688  */
 689 void rpc_prepare_task(struct rpc_task *task)
 690 {
 691         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 692 }
 693
 694 static void
 695 rpc_init_task_statistics(struct rpc_task *task)
 696 {
 697         /* Initialize retry counters */
 698         task->tk_garb_retry = 2;
 699         task->tk_cred_retry = 2;
 700
 701         /* starting timestamp */
 702         task->tk_start = ktime_get();
 703 }
 704
 705 static void
 706 rpc_reset_task_statistics(struct rpc_task *task)
 707 {
 708         task->tk_timeouts = 0;
 709         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 710
 711         rpc_init_task_statistics(task);
 712 }
 713
 714 /*
 715  * Helper that calls task->tk_ops->rpc_call_done if it exists
 716  */
 717 void rpc_exit_task(struct rpc_task *task)
 718 {
 719         task->tk_action = NULL;
 720         if (task->tk_ops->rpc_call_done != NULL) {
 721                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 722                 if (task->tk_action != NULL) {
 723                         WARN_ON(RPC_ASSASSINATED(task));
 724                         /* Always release the RPC slot and buffer memory */
 725                         xprt_release(task);
 726                         rpc_reset_task_statistics(task);
 727                 }
 728         }
 729 }
 730
 731 void rpc_exit(struct rpc_task *task, int status)
 732 {
 733         task->tk_status = status;
 734         task->tk_action = rpc_exit_task;
 735         if (RPC_IS_QUEUED(task))
 736                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 737 }
 738 EXPORT_SYMBOL_GPL(rpc_exit);
 739
 740 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 741 {
 742         if (ops->rpc_release != NULL)
 743                 ops->rpc_release(calldata);
 744 }
 745
 746 /*
 747  * This is the RPC `scheduler' (or rather, the finite state machine).
 748  */
 749 static void __rpc_execute(struct rpc_task *task)
 750 {
 751         struct rpc_wait_queue *queue;
 752         int task_is_async = RPC_IS_ASYNC(task);
 753         int status = 0;
 754
 755         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 756                         task->tk_pid, task->tk_flags);
 757
 758         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 759         if (RPC_IS_QUEUED(task))
 760                 return;
 761
 762         for (;;) {
 763                 void (*do_action)(struct rpc_task *);
 764
 765                 /*
 766                  * Execute any pending callback first.
 767                  */
 768                 do_action = task->tk_callback;
 769                 task->tk_callback = NULL;
 770                 if (do_action == NULL) {
 771                         /*
 772                          * Perform the next FSM step.
 773                          * tk_action may be NULL if the task has been killed.
 774                          * In particular, note that rpc_killall_tasks may
 775                          * do this at any time, so beware when dereferencing.
 776                          */
 777                         do_action = task->tk_action;
 778                         if (do_action == NULL)
 779                                 break;
 780                 }
 781                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 782                 do_action(task);
 783
 784                 /*
 785                  * Lockless check for whether task is sleeping or not.
 786                  */
 787                 if (!RPC_IS_QUEUED(task))
 788                         continue;
 789                 /*
 790                  * The queue->lock protects against races with
 791                  * rpc_make_runnable().
 792                  *
 793                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 794                  * rpc_task, rpc_make_runnable() can assign it to a
 795                  * different workqueue. We therefore cannot assume that the
 796                  * rpc_task pointer may still be dereferenced.
 797                  */
 798                 queue = task->tk_waitqueue;
 799                 spin_lock_bh(&queue->lock);
 800                 if (!RPC_IS_QUEUED(task)) {
 801                         spin_unlock_bh(&queue->lock);
 802                         continue;
 803                 }
 804                 rpc_clear_running(task);
 805                 spin_unlock_bh(&queue->lock);
 806                 if (task_is_async)
 807                         return;
 808
 809                 /* sync task: sleep here */
 810                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 811                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 812                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 813                                 TASK_KILLABLE);
 814                 if (status == -ERESTARTSYS) {
 815                         /*
 816                          * When a sync task receives a signal, it exits with
 817                          * -ERESTARTSYS. In order to catch any callbacks that
 818                          * clean up after sleeping on some queue, we don't
 819                          * break the loop here, but go around once more.
 820                          */
 821                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 822                         task->tk_flags |= RPC_TASK_KILLED;
 823                         rpc_exit(task, -ERESTARTSYS);
 824                 }
 825                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 826         }
 827
 828         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 829                         task->tk_status);
 830         /* Release all resources associated with the task */
 831         rpc_release_task(task);
 832 }
 833
 834 /*
 835  * User-visible entry point to the scheduler.
 836  *
 837  * This may be called recursively if e.g. an async NFS task updates
 838  * the attributes and finds that dirty pages must be flushed.
 839  * NOTE: Upon exit of this function the task is guaranteed to be
 840  *       released. In particular note that tk_release() will have
 841  *       been called, so your task memory may have been freed.
 842  */
 843 void rpc_execute(struct rpc_task *task)
 844 {
 845         bool is_async = RPC_IS_ASYNC(task);
 846
 847         rpc_set_active(task);
 848         rpc_make_runnable(rpciod_workqueue, task);
 849         if (!is_async)
 850                 __rpc_execute(task);
 851 }
 852
 853 static void rpc_async_schedule(struct work_struct *work)
 854 {
 855         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 856 }
 857
 858 /**
 859  * rpc_malloc - allocate RPC buffer resources
 860  * @task: RPC task
 861  *
 862  * A single memory region is allocated, which is split between the
 863  * RPC call and RPC reply that this task is being used for. When
 864  * this RPC is retired, the memory is released by calling rpc_free.
 865  *
 866  * To prevent rpciod from hanging, this allocator never sleeps,
 867  * returning -ENOMEM and suppressing warning if the request cannot
 868  * be serviced immediately. The caller can arrange to sleep in a
 869  * way that is safe for rpciod.
 870  *
 871  * Most requests are 'small' (under 2KiB) and can be serviced from a
 872  * mempool, ensuring that NFS reads and writes can always proceed,
 873  * and that there is good locality of reference for these buffers.
 874  *
 875  * In order to avoid memory starvation triggering more writebacks of
 876  * NFS requests, we avoid using GFP_KERNEL.
 877  */
 878 int rpc_malloc(struct rpc_task *task)
 879 {
 880         struct rpc_rqst *rqst = task->tk_rqstp;
 881         size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
 882         struct rpc_buffer *buf;
 883         gfp_t gfp = GFP_NOIO | __GFP_NOWARN;
 884
 885         if (RPC_IS_ASYNC(task))
 886                 gfp = GFP_NOWAIT | __GFP_NOWARN;
 887         if (RPC_IS_SWAPPER(task))
 888                 gfp |= __GFP_MEMALLOC;
 889
 890         size += sizeof(struct rpc_buffer);
 891         if (size <= RPC_BUFFER_MAXSIZE)
 892                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 893         else
 894                 buf = kmalloc(size, gfp);
 895
 896         if (!buf)
 897                 return -ENOMEM;
 898
 899         buf->len = size;
 900         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 901                         task->tk_pid, size, buf);
 902         rqst->rq_buffer = buf->data;
 903         rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
 904         return 0;
 905 }
 906 EXPORT_SYMBOL_GPL(rpc_malloc);
 907
 908 /**
 909  * rpc_free - free RPC buffer resources allocated via rpc_malloc
 910  * @task: RPC task
 911  *
 912  */
 913 void rpc_free(struct rpc_task *task)
 914 {
 915         void *buffer = task->tk_rqstp->rq_buffer;
 916         size_t size;
 917         struct rpc_buffer *buf;
 918
 919         buf = container_of(buffer, struct rpc_buffer, data);
 920         size = buf->len;
 921
 922         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 923                         size, buf);
 924
 925         if (size <= RPC_BUFFER_MAXSIZE)
 926                 mempool_free(buf, rpc_buffer_mempool);
 927         else
 928                 kfree(buf);
 929 }
 930 EXPORT_SYMBOL_GPL(rpc_free);
 931
 932 /*
 933  * Creation and deletion of RPC task structures
 934  */
 935 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 936 {
 937         memset(task, 0, sizeof(*task));
 938         atomic_set(&task->tk_count, 1);
 939         task->tk_flags  = task_setup_data->flags;
 940         task->tk_ops = task_setup_data->callback_ops;
 941         task->tk_calldata = task_setup_data->callback_data;
 942         INIT_LIST_HEAD(&task->tk_task);
 943
 944         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 945         task->tk_owner = current->tgid;
 946
 947         /* Initialize workqueue for async tasks */
 948         task->tk_workqueue = task_setup_data->workqueue;
 949
 950         task->tk_xprt = xprt_get(task_setup_data->rpc_xprt);
 951
 952         if (task->tk_ops->rpc_call_prepare != NULL)
 953                 task->tk_action = rpc_prepare_task;
 954
 955         rpc_init_task_statistics(task);
 956
 957         dprintk("RPC:       new task initialized, procpid %u\n",
 958                                 task_pid_nr(current));
 959 }
 960
 961 static struct rpc_task *
 962 rpc_alloc_task(void)
 963 {
 964         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 965 }
 966
 967 /*
 968  * Create a new task for the specified client.
 969  */
 970 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 971 {
 972         struct rpc_task *task = setup_data->task;
 973         unsigned short flags = 0;
 974
 975         if (task == NULL) {
 976                 task = rpc_alloc_task();
 977                 flags = RPC_TASK_DYNAMIC;
 978         }
 979
 980         rpc_init_task(task, setup_data);
 981         task->tk_flags |= flags;
 982         dprintk("RPC:       allocated task %p\n", task);
 983         return task;
 984 }
 985
 986 /*
 987  * rpc_free_task - release rpc task and perform cleanups
 988  *
 989  * Note that we free up the rpc_task _after_ rpc_release_calldata()
 990  * in order to work around a workqueue dependency issue.
 991  *
 992  * Tejun Heo states:
 993  * "Workqueue currently considers two work items to be the same if they're
 994  * on the same address and won't execute them concurrently - ie. it
 995  * makes a work item which is queued again while being executed wait
 996  * for the previous execution to complete.
 997  *
 998  * If a work function frees the work item, and then waits for an event
 999  * which should be performed by another work item and *that* work item
1000  * recycles the freed work item, it can create a false dependency loop.
1001  * There really is no reliable way to detect this short of verifying
1002  * every memory free."
1003  *
1004  */
1005 static void rpc_free_task(struct rpc_task *task)
1006 {
1007         unsigned short tk_flags = task->tk_flags;
1008
1009         rpc_release_calldata(task->tk_ops, task->tk_calldata);
1010
1011         if (tk_flags & RPC_TASK_DYNAMIC) {
1012                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1013                 mempool_free(task, rpc_task_mempool);
1014         }
1015 }
1016
1017 static void rpc_async_release(struct work_struct *work)
1018 {
1019         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1020 }
1021
1022 static void rpc_release_resources_task(struct rpc_task *task)
1023 {
1024         xprt_release(task);
1025         if (task->tk_msg.rpc_cred) {
1026                 put_rpccred(task->tk_msg.rpc_cred);
1027                 task->tk_msg.rpc_cred = NULL;
1028         }
1029         rpc_task_release_client(task);
1030 }
1031
1032 static void rpc_final_put_task(struct rpc_task *task,
1033                 struct workqueue_struct *q)
1034 {
1035         if (q != NULL) {
1036                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1037                 queue_work(q, &task->u.tk_work);
1038         } else
1039                 rpc_free_task(task);
1040 }
1041
1042 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1043 {
1044         if (atomic_dec_and_test(&task->tk_count)) {
1045                 rpc_release_resources_task(task);
1046                 rpc_final_put_task(task, q);
1047         }
1048 }
1049
1050 void rpc_put_task(struct rpc_task *task)
1051 {
1052         rpc_do_put_task(task, NULL);
1053 }
1054 EXPORT_SYMBOL_GPL(rpc_put_task);
1055
1056 void rpc_put_task_async(struct rpc_task *task)
1057 {
1058         rpc_do_put_task(task, task->tk_workqueue);
1059 }
1060 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1061
1062 static void rpc_release_task(struct rpc_task *task)
1063 {
1064         dprintk("RPC: %5u release task\n", task->tk_pid);
1065
1066         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1067
1068         rpc_release_resources_task(task);
1069
1070         /*
1071          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1072          * so it should be safe to use task->tk_count as a test for whether
1073          * or not any other processes still hold references to our rpc_task.
1074          */
1075         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1076                 /* Wake up anyone who may be waiting for task completion */
1077                 if (!rpc_complete_task(task))
1078                         return;
1079         } else {
1080                 if (!atomic_dec_and_test(&task->tk_count))
1081                         return;
1082         }
1083         rpc_final_put_task(task, task->tk_workqueue);
1084 }
1085
1086 int rpciod_up(void)
1087 {
1088         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1089 }
1090
1091 void rpciod_down(void)
1092 {
1093         module_put(THIS_MODULE);
1094 }
1095
1096 /*
1097  * Start up the rpciod workqueue.
1098  */
1099 static int rpciod_start(void)
1100 {
1101         struct workqueue_struct *wq;
1102
1103         /*
1104          * Create the rpciod thread and wait for it to start.
1105          */
1106         dprintk("RPC:       creating workqueue rpciod\n");
1107         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1108         if (!wq)
1109                 goto out_failed;
1110         rpciod_workqueue = wq;
1111         /* Note: highpri because network receive is latency sensitive */
1112         wq = alloc_workqueue("xprtiod", WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1113         if (!wq)
1114                 goto free_rpciod;
1115         xprtiod_workqueue = wq;
1116         return 1;
1117 free_rpciod:
1118         wq = rpciod_workqueue;
1119         rpciod_workqueue = NULL;
1120         destroy_workqueue(wq);
1121 out_failed:
1122         return 0;
1123 }
1124
1125 static void rpciod_stop(void)
1126 {
1127         struct workqueue_struct *wq = NULL;
1128
1129         if (rpciod_workqueue == NULL)
1130                 return;
1131         dprintk("RPC:       destroying workqueue rpciod\n");
1132
1133         wq = rpciod_workqueue;
1134         rpciod_workqueue = NULL;
1135         destroy_workqueue(wq);
1136         wq = xprtiod_workqueue;
1137         xprtiod_workqueue = NULL;
1138         destroy_workqueue(wq);
1139 }
1140
1141 void
1142 rpc_destroy_mempool(void)
1143 {
1144         rpciod_stop();
1145         mempool_destroy(rpc_buffer_mempool);
1146         mempool_destroy(rpc_task_mempool);
1147         kmem_cache_destroy(rpc_task_slabp);
1148         kmem_cache_destroy(rpc_buffer_slabp);
1149         rpc_destroy_wait_queue(&delay_queue);
1150 }
1151
1152 int
1153 rpc_init_mempool(void)
1154 {
1155         /*
1156          * The following is not strictly a mempool initialisation,
1157          * but there is no harm in doing it here
1158          */
1159         rpc_init_wait_queue(&delay_queue, "delayq");
1160         if (!rpciod_start())
1161                 goto err_nomem;
1162
1163         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1164                                              sizeof(struct rpc_task),
1165                                              0, SLAB_HWCACHE_ALIGN,
1166                                              NULL);
1167         if (!rpc_task_slabp)
1168                 goto err_nomem;
1169         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1170                                              RPC_BUFFER_MAXSIZE,
1171                                              0, SLAB_HWCACHE_ALIGN,
1172                                              NULL);
1173         if (!rpc_buffer_slabp)
1174                 goto err_nomem;
1175         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1176                                                     rpc_task_slabp);
1177         if (!rpc_task_mempool)
1178                 goto err_nomem;
1179         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1180                                                       rpc_buffer_slabp);
1181         if (!rpc_buffer_mempool)
1182                 goto err_nomem;
1183         return 0;
1184 err_nomem:
1185         rpc_destroy_mempool();
1186         return -ENOMEM;
1187 }