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