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