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