net/sunrpc/sched.c

   1 /*
   2  * linux/net/sunrpc/sched.c
   3  *
   4  * Scheduling for synchronous and asynchronous RPC requests.
   5  *
   6  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  10  */
  11
  12 #include <linux/module.h>
  13
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/slab.h>
  17 #include <linux/mempool.h>
  18 #include <linux/smp.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/mutex.h>
  21 #include <linux/freezer.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #include "sunrpc.h"
  26
  27 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  28 #define RPCDBG_FACILITY         RPCDBG_SCHED
  29 #endif
  30
  31 #define CREATE_TRACE_POINTS
  32 #include <trace/events/sunrpc.h>
  33
  34 /*
  35  * RPC slabs and memory pools
  36  */
  37 #define RPC_BUFFER_MAXSIZE      (2048)
  38 #define RPC_BUFFER_POOLSIZE     (8)
  39 #define RPC_TASK_POOLSIZE       (8)
  40 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  41 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  42 static mempool_t        *rpc_task_mempool __read_mostly;
  43 static mempool_t        *rpc_buffer_mempool __read_mostly;
  44
  45 static void                     rpc_async_schedule(struct work_struct *);
  46 static void                      rpc_release_task(struct rpc_task *task);
  47 static void __rpc_queue_timer_fn(unsigned long ptr);
  48
  49 /*
  50  * RPC tasks sit here while waiting for conditions to improve.
  51  */
  52 static struct rpc_wait_queue delay_queue;
  53
  54 /*
  55  * rpciod-related stuff
  56  */
  57 struct workqueue_struct *rpciod_workqueue __read_mostly;
  58 struct workqueue_struct *xprtiod_workqueue __read_mostly;
  59
  60 /*
  61  * Disable the timer for a given RPC task. Should be called with
  62  * queue->lock and bh_disabled in order to avoid races within
  63  * rpc_run_timer().
  64  */
  65 static void
  66 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  67 {
  68         if (task->tk_timeout == 0)
  69                 return;
  70         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  71         task->tk_timeout = 0;
  72         list_del(&task->u.tk_wait.timer_list);
  73         if (list_empty(&queue->timer_list.list))
  74                 del_timer(&queue->timer_list.timer);
  75 }
  76
  77 static void
  78 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  79 {
  80         queue->timer_list.expires = expires;
  81         mod_timer(&queue->timer_list.timer, expires);
  82 }
  83
  84 /*
  85  * Set up a timer for the current task.
  86  */
  87 static void
  88 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  89 {
  90         if (!task->tk_timeout)
  91                 return;
  92
  93         dprintk("RPC: %5u setting alarm for %u ms\n",
  94                 task->tk_pid, jiffies_to_msecs(task->tk_timeout));
  95
  96         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  97         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  98                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  99         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
 100 }
 101
 102 static void rpc_rotate_queue_owner(struct rpc_wait_queue *queue)
 103 {
 104         struct list_head *q = &queue->tasks[queue->priority];
 105         struct rpc_task *task;
 106
 107         if (!list_empty(q)) {
 108                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 109                 if (task->tk_owner == queue->owner)
 110                         list_move_tail(&task->u.tk_wait.list, q);
 111         }
 112 }
 113
 114 static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 115 {
 116         if (queue->priority != priority) {
 117                 /* Fairness: rotate the list when changing priority */
 118                 rpc_rotate_queue_owner(queue);
 119                 queue->priority = priority;
 120         }
 121 }
 122
 123 static void rpc_set_waitqueue_owner(struct rpc_wait_queue *queue, pid_t pid)
 124 {
 125         queue->owner = pid;
 126         queue->nr = RPC_BATCH_COUNT;
 127 }
 128
 129 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 130 {
 131         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 132         rpc_set_waitqueue_owner(queue, 0);
 133 }
 134
 135 /*
 136  * Add new request to a priority queue.
 137  */
 138 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 139                 struct rpc_task *task,
 140                 unsigned char queue_priority)
 141 {
 142         struct list_head *q;
 143         struct rpc_task *t;
 144
 145         INIT_LIST_HEAD(&task->u.tk_wait.links);
 146         if (unlikely(queue_priority > queue->maxpriority))
 147                 queue_priority = queue->maxpriority;
 148         if (queue_priority > queue->priority)
 149                 rpc_set_waitqueue_priority(queue, queue_priority);
 150         q = &queue->tasks[queue_priority];
 151         list_for_each_entry(t, q, u.tk_wait.list) {
 152                 if (t->tk_owner == task->tk_owner) {
 153                         list_add_tail(&task->u.tk_wait.list, &t->u.tk_wait.links);
 154                         return;
 155                 }
 156         }
 157         list_add_tail(&task->u.tk_wait.list, q);
 158 }
 159
 160 /*
 161  * Add new request to wait queue.
 162  *
 163  * Swapper tasks always get inserted at the head of the queue.
 164  * This should avoid many nasty memory deadlocks and hopefully
 165  * improve overall performance.
 166  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 167  */
 168 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 169                 struct rpc_task *task,
 170                 unsigned char queue_priority)
 171 {
 172         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 173         if (RPC_IS_QUEUED(task))
 174                 return;
 175
 176         if (RPC_IS_PRIORITY(queue))
 177                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 178         else if (RPC_IS_SWAPPER(task))
 179                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 180         else
 181                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 182         task->tk_waitqueue = queue;
 183         queue->qlen++;
 184         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 185         smp_wmb();
 186         rpc_set_queued(task);
 187
 188         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 189                         task->tk_pid, queue, rpc_qname(queue));
 190 }
 191
 192 /*
 193  * Remove request from a priority queue.
 194  */
 195 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 196 {
 197         struct rpc_task *t;
 198
 199         if (!list_empty(&task->u.tk_wait.links)) {
 200                 t = list_entry(task->u.tk_wait.links.next, struct rpc_task, u.tk_wait.list);
 201                 list_move(&t->u.tk_wait.list, &task->u.tk_wait.list);
 202                 list_splice_init(&task->u.tk_wait.links, &t->u.tk_wait.links);
 203         }
 204 }
 205
 206 /*
 207  * Remove request from queue.
 208  * Note: must be called with spin lock held.
 209  */
 210 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 211 {
 212         __rpc_disable_timer(queue, task);
 213         if (RPC_IS_PRIORITY(queue))
 214                 __rpc_remove_wait_queue_priority(task);
 215         list_del(&task->u.tk_wait.list);
 216         queue->qlen--;
 217         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 218                         task->tk_pid, queue, rpc_qname(queue));
 219 }
 220
 221 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 222 {
 223         int i;
 224
 225         spin_lock_init(&queue->lock);
 226         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 227                 INIT_LIST_HEAD(&queue->tasks[i]);
 228         queue->maxpriority = nr_queues - 1;
 229         rpc_reset_waitqueue_priority(queue);
 230         queue->qlen = 0;
 231         setup_timer(&queue->timer_list.timer, __rpc_queue_timer_fn, (unsigned long)queue);
 232         INIT_LIST_HEAD(&queue->timer_list.list);
 233         rpc_assign_waitqueue_name(queue, qname);
 234 }
 235
 236 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 237 {
 238         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 239 }
 240 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 241
 242 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 243 {
 244         __rpc_init_priority_wait_queue(queue, qname, 1);
 245 }
 246 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 247
 248 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 249 {
 250         del_timer_sync(&queue->timer_list.timer);
 251 }
 252 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 253
 254 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
 255 {
 256         freezable_schedule_unsafe();
 257         if (signal_pending_state(mode, current))
 258                 return -ERESTARTSYS;
 259         return 0;
 260 }
 261
 262 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
 263 static void rpc_task_set_debuginfo(struct rpc_task *task)
 264 {
 265         static atomic_t rpc_pid;
 266
 267         task->tk_pid = atomic_inc_return(&rpc_pid);
 268 }
 269 #else
 270 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 271 {
 272 }
 273 #endif
 274
 275 static void rpc_set_active(struct rpc_task *task)
 276 {
 277         trace_rpc_task_begin(task->tk_client, task, NULL);
 278
 279         rpc_task_set_debuginfo(task);
 280         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 281 }
 282
 283 /*
 284  * Mark an RPC call as having completed by clearing the 'active' bit
 285  * and then waking up all tasks that were sleeping.
 286  */
 287 static int rpc_complete_task(struct rpc_task *task)
 288 {
 289         void *m = &task->tk_runstate;
 290         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 291         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 292         unsigned long flags;
 293         int ret;
 294
 295         trace_rpc_task_complete(task->tk_client, task, NULL);
 296
 297         spin_lock_irqsave(&wq->lock, flags);
 298         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 299         ret = atomic_dec_and_test(&task->tk_count);
 300         if (waitqueue_active(wq))
 301                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 302         spin_unlock_irqrestore(&wq->lock, flags);
 303         return ret;
 304 }
 305
 306 /*
 307  * Allow callers to wait for completion of an RPC call
 308  *
 309  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 310  * to enforce taking of the wq->lock and hence avoid races with
 311  * rpc_complete_task().
 312  */
 313 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
 314 {
 315         if (action == NULL)
 316                 action = rpc_wait_bit_killable;
 317         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 318                         action, TASK_KILLABLE);
 319 }
 320 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 321
 322 /*
 323  * Make an RPC task runnable.
 324  *
 325  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 326  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 327  * the wait queue operation.
 328  * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
 329  * which is needed to ensure that __rpc_execute() doesn't loop (due to the
 330  * lockless RPC_IS_QUEUED() test) before we've had a chance to test
 331  * the RPC_TASK_RUNNING flag.
 332  */
 333 static void rpc_make_runnable(struct workqueue_struct *wq,
 334                 struct rpc_task *task)
 335 {
 336         bool need_wakeup = !rpc_test_and_set_running(task);
 337
 338         rpc_clear_queued(task);
 339         if (!need_wakeup)
 340                 return;
 341         if (RPC_IS_ASYNC(task)) {
 342                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 343                 queue_work(wq, &task->u.tk_work);
 344         } else
 345                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 346 }
 347
 348 /*
 349  * Prepare for sleeping on a wait queue.
 350  * By always appending tasks to the list we ensure FIFO behavior.
 351  * NB: An RPC task will only receive interrupt-driven events as long
 352  * as it's on a wait queue.
 353  */
 354 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 355                 struct rpc_task *task,
 356                 rpc_action action,
 357                 unsigned char queue_priority)
 358 {
 359         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 360                         task->tk_pid, rpc_qname(q), jiffies);
 361
 362         trace_rpc_task_sleep(task->tk_client, task, q);
 363
 364         __rpc_add_wait_queue(q, task, queue_priority);
 365
 366         WARN_ON_ONCE(task->tk_callback != NULL);
 367         task->tk_callback = action;
 368         __rpc_add_timer(q, task);
 369 }
 370
 371 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 372                                 rpc_action action)
 373 {
 374         /* We shouldn't ever put an inactive task to sleep */
 375         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 376         if (!RPC_IS_ACTIVATED(task)) {
 377                 task->tk_status = -EIO;
 378                 rpc_put_task_async(task);
 379                 return;
 380         }
 381
 382         /*
 383          * Protect the queue operations.
 384          */
 385         spin_lock_bh(&q->lock);
 386         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 387         spin_unlock_bh(&q->lock);
 388 }
 389 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 390
 391 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 392                 rpc_action action, int priority)
 393 {
 394         /* We shouldn't ever put an inactive task to sleep */
 395         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 396         if (!RPC_IS_ACTIVATED(task)) {
 397                 task->tk_status = -EIO;
 398                 rpc_put_task_async(task);
 399                 return;
 400         }
 401
 402         /*
 403          * Protect the queue operations.
 404          */
 405         spin_lock_bh(&q->lock);
 406         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 407         spin_unlock_bh(&q->lock);
 408 }
 409 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 410
 411 /**
 412  * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
 413  * @wq: workqueue on which to run task
 414  * @queue: wait queue
 415  * @task: task to be woken up
 416  *
 417  * Caller must hold queue->lock, and have cleared the task queued flag.
 418  */
 419 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
 420                 struct rpc_wait_queue *queue,
 421                 struct rpc_task *task)
 422 {
 423         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 424                         task->tk_pid, jiffies);
 425
 426         /* Has the task been executed yet? If not, we cannot wake it up! */
 427         if (!RPC_IS_ACTIVATED(task)) {
 428                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 429                 return;
 430         }
 431
 432         trace_rpc_task_wakeup(task->tk_client, task, queue);
 433
 434         __rpc_remove_wait_queue(queue, task);
 435
 436         rpc_make_runnable(wq, task);
 437
 438         dprintk("RPC:       __rpc_wake_up_task done\n");
 439 }
 440
 441 /*
 442  * Wake up a queued task while the queue lock is being held
 443  */
 444 static void rpc_wake_up_task_on_wq_queue_locked(struct workqueue_struct *wq,
 445                 struct rpc_wait_queue *queue, struct rpc_task *task)
 446 {
 447         if (RPC_IS_QUEUED(task)) {
 448                 smp_rmb();
 449                 if (task->tk_waitqueue == queue)
 450                         __rpc_do_wake_up_task_on_wq(wq, queue, task);
 451         }
 452 }
 453
 454 /*
 455  * Wake up a queued task while the queue lock is being held
 456  */
 457 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 458 {
 459         rpc_wake_up_task_on_wq_queue_locked(rpciod_workqueue, queue, task);
 460 }
 461
 462 /*
 463  * Wake up a task on a specific queue
 464  */
 465 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 466 {
 467         spin_lock_bh(&queue->lock);
 468         rpc_wake_up_task_queue_locked(queue, task);
 469         spin_unlock_bh(&queue->lock);
 470 }
 471 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 472
 473 /*
 474  * Wake up the next task on a priority queue.
 475  */
 476 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 477 {
 478         struct list_head *q;
 479         struct rpc_task *task;
 480
 481         /*
 482          * Service a batch of tasks from a single owner.
 483          */
 484         q = &queue->tasks[queue->priority];
 485         if (!list_empty(q)) {
 486                 task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 487                 if (queue->owner == task->tk_owner) {
 488                         if (--queue->nr)
 489                                 goto out;
 490                         list_move_tail(&task->u.tk_wait.list, q);
 491                 }
 492                 /*
 493                  * Check if we need to switch queues.
 494                  */
 495                 goto new_owner;
 496         }
 497
 498         /*
 499          * Service the next queue.
 500          */
 501         do {
 502                 if (q == &queue->tasks[0])
 503                         q = &queue->tasks[queue->maxpriority];
 504                 else
 505                         q = q - 1;
 506                 if (!list_empty(q)) {
 507                         task = list_entry(q->next, struct rpc_task, u.tk_wait.list);
 508                         goto new_queue;
 509                 }
 510         } while (q != &queue->tasks[queue->priority]);
 511
 512         rpc_reset_waitqueue_priority(queue);
 513         return NULL;
 514
 515 new_queue:
 516         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 517 new_owner:
 518         rpc_set_waitqueue_owner(queue, task->tk_owner);
 519 out:
 520         return task;
 521 }
 522
 523 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 524 {
 525         if (RPC_IS_PRIORITY(queue))
 526                 return __rpc_find_next_queued_priority(queue);
 527         if (!list_empty(&queue->tasks[0]))
 528                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 529         return NULL;
 530 }
 531
 532 /*
 533  * Wake up the first task on the wait queue.
 534  */
 535 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 536                 struct rpc_wait_queue *queue,
 537                 bool (*func)(struct rpc_task *, void *), void *data)
 538 {
 539         struct rpc_task *task = NULL;
 540
 541         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 542                         queue, rpc_qname(queue));
 543         spin_lock_bh(&queue->lock);
 544         task = __rpc_find_next_queued(queue);
 545         if (task != NULL) {
 546                 if (func(task, data))
 547                         rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
 548                 else
 549                         task = NULL;
 550         }
 551         spin_unlock_bh(&queue->lock);
 552
 553         return task;
 554 }
 555
 556 /*
 557  * Wake up the first task on the wait queue.
 558  */
 559 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 560                 bool (*func)(struct rpc_task *, void *), void *data)
 561 {
 562         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 563 }
 564 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 565
 566 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 567 {
 568         return true;
 569 }
 570
 571 /*
 572  * Wake up the next task on the wait queue.
 573 */
 574 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 575 {
 576         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 577 }
 578 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 579
 580 /**
 581  * rpc_wake_up - wake up all rpc_tasks
 582  * @queue: rpc_wait_queue on which the tasks are sleeping
 583  *
 584  * Grabs queue->lock
 585  */
 586 void rpc_wake_up(struct rpc_wait_queue *queue)
 587 {
 588         struct list_head *head;
 589
 590         spin_lock_bh(&queue->lock);
 591         head = &queue->tasks[queue->maxpriority];
 592         for (;;) {
 593                 while (!list_empty(head)) {
 594                         struct rpc_task *task;
 595                         task = list_first_entry(head,
 596                                         struct rpc_task,
 597                                         u.tk_wait.list);
 598                         rpc_wake_up_task_queue_locked(queue, task);
 599                 }
 600                 if (head == &queue->tasks[0])
 601                         break;
 602                 head--;
 603         }
 604         spin_unlock_bh(&queue->lock);
 605 }
 606 EXPORT_SYMBOL_GPL(rpc_wake_up);
 607
 608 /**
 609  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 610  * @queue: rpc_wait_queue on which the tasks are sleeping
 611  * @status: status value to set
 612  *
 613  * Grabs queue->lock
 614  */
 615 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 616 {
 617         struct list_head *head;
 618
 619         spin_lock_bh(&queue->lock);
 620         head = &queue->tasks[queue->maxpriority];
 621         for (;;) {
 622                 while (!list_empty(head)) {
 623                         struct rpc_task *task;
 624                         task = list_first_entry(head,
 625                                         struct rpc_task,
 626                                         u.tk_wait.list);
 627                         task->tk_status = status;
 628                         rpc_wake_up_task_queue_locked(queue, task);
 629                 }
 630                 if (head == &queue->tasks[0])
 631                         break;
 632                 head--;
 633         }
 634         spin_unlock_bh(&queue->lock);
 635 }
 636 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 637
 638 static void __rpc_queue_timer_fn(unsigned long ptr)
 639 {
 640         struct rpc_wait_queue *queue = (struct rpc_wait_queue *)ptr;
 641         struct rpc_task *task, *n;
 642         unsigned long expires, now, timeo;
 643
 644         spin_lock(&queue->lock);
 645         expires = now = jiffies;
 646         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 647                 timeo = task->u.tk_wait.expires;
 648                 if (time_after_eq(now, timeo)) {
 649                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 650                         task->tk_status = -ETIMEDOUT;
 651                         rpc_wake_up_task_queue_locked(queue, task);
 652                         continue;
 653                 }
 654                 if (expires == now || time_after(expires, timeo))
 655                         expires = timeo;
 656         }
 657         if (!list_empty(&queue->timer_list.list))
 658                 rpc_set_queue_timer(queue, expires);
 659         spin_unlock(&queue->lock);
 660 }
 661
 662 static void __rpc_atrun(struct rpc_task *task)
 663 {
 664         if (task->tk_status == -ETIMEDOUT)
 665                 task->tk_status = 0;
 666 }
 667
 668 /*
 669  * Run a task at a later time
 670  */
 671 void rpc_delay(struct rpc_task *task, unsigned long delay)
 672 {
 673         task->tk_timeout = delay;
 674         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 675 }
 676 EXPORT_SYMBOL_GPL(rpc_delay);
 677
 678 /*
 679  * Helper to call task->tk_ops->rpc_call_prepare
 680  */
 681 void rpc_prepare_task(struct rpc_task *task)
 682 {
 683         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 684 }
 685
 686 static void
 687 rpc_init_task_statistics(struct rpc_task *task)
 688 {
 689         /* Initialize retry counters */
 690         task->tk_garb_retry = 2;
 691         task->tk_cred_retry = 2;
 692         task->tk_rebind_retry = 2;
 693
 694         /* starting timestamp */
 695         task->tk_start = ktime_get();
 696 }
 697
 698 static void
 699 rpc_reset_task_statistics(struct rpc_task *task)
 700 {
 701         task->tk_timeouts = 0;
 702         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 703
 704         rpc_init_task_statistics(task);
 705 }
 706
 707 /*
 708  * Helper that calls task->tk_ops->rpc_call_done if it exists
 709  */
 710 void rpc_exit_task(struct rpc_task *task)
 711 {
 712         task->tk_action = NULL;
 713         if (task->tk_ops->rpc_call_done != NULL) {
 714                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 715                 if (task->tk_action != NULL) {
 716                         WARN_ON(RPC_ASSASSINATED(task));
 717                         /* Always release the RPC slot and buffer memory */
 718                         xprt_release(task);
 719                         rpc_reset_task_statistics(task);
 720                 }
 721         }
 722 }
 723
 724 void rpc_exit(struct rpc_task *task, int status)
 725 {
 726         task->tk_status = status;
 727         task->tk_action = rpc_exit_task;
 728         if (RPC_IS_QUEUED(task))
 729                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 730 }
 731 EXPORT_SYMBOL_GPL(rpc_exit);
 732
 733 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 734 {
 735         if (ops->rpc_release != NULL)
 736                 ops->rpc_release(calldata);
 737 }
 738
 739 /*
 740  * This is the RPC `scheduler' (or rather, the finite state machine).
 741  */
 742 static void __rpc_execute(struct rpc_task *task)
 743 {
 744         struct rpc_wait_queue *queue;
 745         int task_is_async = RPC_IS_ASYNC(task);
 746         int status = 0;
 747
 748         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 749                         task->tk_pid, task->tk_flags);
 750
 751         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 752         if (RPC_IS_QUEUED(task))
 753                 return;
 754
 755         for (;;) {
 756                 void (*do_action)(struct rpc_task *);
 757
 758                 /*
 759                  * Execute any pending callback first.
 760                  */
 761                 do_action = task->tk_callback;
 762                 task->tk_callback = NULL;
 763                 if (do_action == NULL) {
 764                         /*
 765                          * Perform the next FSM step.
 766                          * tk_action may be NULL if the task has been killed.
 767                          * In particular, note that rpc_killall_tasks may
 768                          * do this at any time, so beware when dereferencing.
 769                          */
 770                         do_action = task->tk_action;
 771                         if (do_action == NULL)
 772                                 break;
 773                 }
 774                 trace_rpc_task_run_action(task->tk_client, task, task->tk_action);
 775                 do_action(task);
 776
 777                 /*
 778                  * Lockless check for whether task is sleeping or not.
 779                  */
 780                 if (!RPC_IS_QUEUED(task))
 781                         continue;
 782                 /*
 783                  * The queue->lock protects against races with
 784                  * rpc_make_runnable().
 785                  *
 786                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 787                  * rpc_task, rpc_make_runnable() can assign it to a
 788                  * different workqueue. We therefore cannot assume that the
 789                  * rpc_task pointer may still be dereferenced.
 790                  */
 791                 queue = task->tk_waitqueue;
 792                 spin_lock_bh(&queue->lock);
 793                 if (!RPC_IS_QUEUED(task)) {
 794                         spin_unlock_bh(&queue->lock);
 795                         continue;
 796                 }
 797                 rpc_clear_running(task);
 798                 spin_unlock_bh(&queue->lock);
 799                 if (task_is_async)
 800                         return;
 801
 802                 /* sync task: sleep here */
 803                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 804                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 805                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 806                                 TASK_KILLABLE);
 807                 if (status == -ERESTARTSYS) {
 808                         /*
 809                          * When a sync task receives a signal, it exits with
 810                          * -ERESTARTSYS. In order to catch any callbacks that
 811                          * clean up after sleeping on some queue, we don't
 812                          * break the loop here, but go around once more.
 813                          */
 814                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 815                         task->tk_flags |= RPC_TASK_KILLED;
 816                         rpc_exit(task, -ERESTARTSYS);
 817                 }
 818                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 819         }
 820
 821         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 822                         task->tk_status);
 823         /* Release all resources associated with the task */
 824         rpc_release_task(task);
 825 }
 826
 827 /*
 828  * User-visible entry point to the scheduler.
 829  *
 830  * This may be called recursively if e.g. an async NFS task updates
 831  * the attributes and finds that dirty pages must be flushed.
 832  * NOTE: Upon exit of this function the task is guaranteed to be
 833  *       released. In particular note that tk_release() will have
 834  *       been called, so your task memory may have been freed.
 835  */
 836 void rpc_execute(struct rpc_task *task)
 837 {
 838         bool is_async = RPC_IS_ASYNC(task);
 839
 840         rpc_set_active(task);
 841         rpc_make_runnable(rpciod_workqueue, task);
 842         if (!is_async)
 843                 __rpc_execute(task);
 844 }
 845
 846 static void rpc_async_schedule(struct work_struct *work)
 847 {
 848         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 849 }
 850
 851 /**
 852  * rpc_malloc - allocate RPC buffer resources
 853  * @task: RPC task
 854  *
 855  * A single memory region is allocated, which is split between the
 856  * RPC call and RPC reply that this task is being used for. When
 857  * this RPC is retired, the memory is released by calling rpc_free.
 858  *
 859  * To prevent rpciod from hanging, this allocator never sleeps,
 860  * returning -ENOMEM and suppressing warning if the request cannot
 861  * be serviced immediately. The caller can arrange to sleep in a
 862  * way that is safe for rpciod.
 863  *
 864  * Most requests are 'small' (under 2KiB) and can be serviced from a
 865  * mempool, ensuring that NFS reads and writes can always proceed,
 866  * and that there is good locality of reference for these buffers.
 867  *
 868  * In order to avoid memory starvation triggering more writebacks of
 869  * NFS requests, we avoid using GFP_KERNEL.
 870  */
 871 int rpc_malloc(struct rpc_task *task)
 872 {
 873         struct rpc_rqst *rqst = task->tk_rqstp;
 874         size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
 875         struct rpc_buffer *buf;
 876         gfp_t gfp = GFP_NOIO | __GFP_NOWARN;
 877
 878         if (RPC_IS_SWAPPER(task))
 879                 gfp = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 880
 881         size += sizeof(struct rpc_buffer);
 882         if (size <= RPC_BUFFER_MAXSIZE)
 883                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 884         else
 885                 buf = kmalloc(size, gfp);
 886
 887         if (!buf)
 888                 return -ENOMEM;
 889
 890         buf->len = size;
 891         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 892                         task->tk_pid, size, buf);
 893         rqst->rq_buffer = buf->data;
 894         rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
 895         return 0;
 896 }
 897 EXPORT_SYMBOL_GPL(rpc_malloc);
 898
 899 /**
 900  * rpc_free - free RPC buffer resources allocated via rpc_malloc
 901  * @task: RPC task
 902  *
 903  */
 904 void rpc_free(struct rpc_task *task)
 905 {
 906         void *buffer = task->tk_rqstp->rq_buffer;
 907         size_t size;
 908         struct rpc_buffer *buf;
 909
 910         buf = container_of(buffer, struct rpc_buffer, data);
 911         size = buf->len;
 912
 913         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 914                         size, buf);
 915
 916         if (size <= RPC_BUFFER_MAXSIZE)
 917                 mempool_free(buf, rpc_buffer_mempool);
 918         else
 919                 kfree(buf);
 920 }
 921 EXPORT_SYMBOL_GPL(rpc_free);
 922
 923 /*
 924  * Creation and deletion of RPC task structures
 925  */
 926 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 927 {
 928         memset(task, 0, sizeof(*task));
 929         atomic_set(&task->tk_count, 1);
 930         task->tk_flags  = task_setup_data->flags;
 931         task->tk_ops = task_setup_data->callback_ops;
 932         task->tk_calldata = task_setup_data->callback_data;
 933         INIT_LIST_HEAD(&task->tk_task);
 934
 935         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 936         task->tk_owner = current->tgid;
 937
 938         /* Initialize workqueue for async tasks */
 939         task->tk_workqueue = task_setup_data->workqueue;
 940
 941         task->tk_xprt = xprt_get(task_setup_data->rpc_xprt);
 942
 943         if (task->tk_ops->rpc_call_prepare != NULL)
 944                 task->tk_action = rpc_prepare_task;
 945
 946         rpc_init_task_statistics(task);
 947
 948         dprintk("RPC:       new task initialized, procpid %u\n",
 949                                 task_pid_nr(current));
 950 }
 951
 952 static struct rpc_task *
 953 rpc_alloc_task(void)
 954 {
 955         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 956 }
 957
 958 /*
 959  * Create a new task for the specified client.
 960  */
 961 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 962 {
 963         struct rpc_task *task = setup_data->task;
 964         unsigned short flags = 0;
 965
 966         if (task == NULL) {
 967                 task = rpc_alloc_task();
 968                 flags = RPC_TASK_DYNAMIC;
 969         }
 970
 971         rpc_init_task(task, setup_data);
 972         task->tk_flags |= flags;
 973         dprintk("RPC:       allocated task %p\n", task);
 974         return task;
 975 }
 976
 977 /*
 978  * rpc_free_task - release rpc task and perform cleanups
 979  *
 980  * Note that we free up the rpc_task _after_ rpc_release_calldata()
 981  * in order to work around a workqueue dependency issue.
 982  *
 983  * Tejun Heo states:
 984  * "Workqueue currently considers two work items to be the same if they're
 985  * on the same address and won't execute them concurrently - ie. it
 986  * makes a work item which is queued again while being executed wait
 987  * for the previous execution to complete.
 988  *
 989  * If a work function frees the work item, and then waits for an event
 990  * which should be performed by another work item and *that* work item
 991  * recycles the freed work item, it can create a false dependency loop.
 992  * There really is no reliable way to detect this short of verifying
 993  * every memory free."
 994  *
 995  */
 996 static void rpc_free_task(struct rpc_task *task)
 997 {
 998         unsigned short tk_flags = task->tk_flags;
 999
1000         rpc_release_calldata(task->tk_ops, task->tk_calldata);
1001
1002         if (tk_flags & RPC_TASK_DYNAMIC) {
1003                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1004                 mempool_free(task, rpc_task_mempool);
1005         }
1006 }
1007
1008 static void rpc_async_release(struct work_struct *work)
1009 {
1010         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1011 }
1012
1013 static void rpc_release_resources_task(struct rpc_task *task)
1014 {
1015         xprt_release(task);
1016         if (task->tk_msg.rpc_cred) {
1017                 put_rpccred(task->tk_msg.rpc_cred);
1018                 task->tk_msg.rpc_cred = NULL;
1019         }
1020         rpc_task_release_client(task);
1021 }
1022
1023 static void rpc_final_put_task(struct rpc_task *task,
1024                 struct workqueue_struct *q)
1025 {
1026         if (q != NULL) {
1027                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1028                 queue_work(q, &task->u.tk_work);
1029         } else
1030                 rpc_free_task(task);
1031 }
1032
1033 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1034 {
1035         if (atomic_dec_and_test(&task->tk_count)) {
1036                 rpc_release_resources_task(task);
1037                 rpc_final_put_task(task, q);
1038         }
1039 }
1040
1041 void rpc_put_task(struct rpc_task *task)
1042 {
1043         rpc_do_put_task(task, NULL);
1044 }
1045 EXPORT_SYMBOL_GPL(rpc_put_task);
1046
1047 void rpc_put_task_async(struct rpc_task *task)
1048 {
1049         rpc_do_put_task(task, task->tk_workqueue);
1050 }
1051 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1052
1053 static void rpc_release_task(struct rpc_task *task)
1054 {
1055         dprintk("RPC: %5u release task\n", task->tk_pid);
1056
1057         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1058
1059         rpc_release_resources_task(task);
1060
1061         /*
1062          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1063          * so it should be safe to use task->tk_count as a test for whether
1064          * or not any other processes still hold references to our rpc_task.
1065          */
1066         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1067                 /* Wake up anyone who may be waiting for task completion */
1068                 if (!rpc_complete_task(task))
1069                         return;
1070         } else {
1071                 if (!atomic_dec_and_test(&task->tk_count))
1072                         return;
1073         }
1074         rpc_final_put_task(task, task->tk_workqueue);
1075 }
1076
1077 int rpciod_up(void)
1078 {
1079         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1080 }
1081
1082 void rpciod_down(void)
1083 {
1084         module_put(THIS_MODULE);
1085 }
1086
1087 /*
1088  * Start up the rpciod workqueue.
1089  */
1090 static int rpciod_start(void)
1091 {
1092         struct workqueue_struct *wq;
1093
1094         /*
1095          * Create the rpciod thread and wait for it to start.
1096          */
1097         dprintk("RPC:       creating workqueue rpciod\n");
1098         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM, 0);
1099         if (!wq)
1100                 goto out_failed;
1101         rpciod_workqueue = wq;
1102         /* Note: highpri because network receive is latency sensitive */
1103         wq = alloc_workqueue("xprtiod", WQ_MEM_RECLAIM | WQ_HIGHPRI, 0);
1104         if (!wq)
1105                 goto free_rpciod;
1106         xprtiod_workqueue = wq;
1107         return 1;
1108 free_rpciod:
1109         wq = rpciod_workqueue;
1110         rpciod_workqueue = NULL;
1111         destroy_workqueue(wq);
1112 out_failed:
1113         return 0;
1114 }
1115
1116 static void rpciod_stop(void)
1117 {
1118         struct workqueue_struct *wq = NULL;
1119
1120         if (rpciod_workqueue == NULL)
1121                 return;
1122         dprintk("RPC:       destroying workqueue rpciod\n");
1123
1124         wq = rpciod_workqueue;
1125         rpciod_workqueue = NULL;
1126         destroy_workqueue(wq);
1127         wq = xprtiod_workqueue;
1128         xprtiod_workqueue = NULL;
1129         destroy_workqueue(wq);
1130 }
1131
1132 void
1133 rpc_destroy_mempool(void)
1134 {
1135         rpciod_stop();
1136         mempool_destroy(rpc_buffer_mempool);
1137         mempool_destroy(rpc_task_mempool);
1138         kmem_cache_destroy(rpc_task_slabp);
1139         kmem_cache_destroy(rpc_buffer_slabp);
1140         rpc_destroy_wait_queue(&delay_queue);
1141 }
1142
1143 int
1144 rpc_init_mempool(void)
1145 {
1146         /*
1147          * The following is not strictly a mempool initialisation,
1148          * but there is no harm in doing it here
1149          */
1150         rpc_init_wait_queue(&delay_queue, "delayq");
1151         if (!rpciod_start())
1152                 goto err_nomem;
1153
1154         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1155                                              sizeof(struct rpc_task),
1156                                              0, SLAB_HWCACHE_ALIGN,
1157                                              NULL);
1158         if (!rpc_task_slabp)
1159                 goto err_nomem;
1160         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1161                                              RPC_BUFFER_MAXSIZE,
1162                                              0, SLAB_HWCACHE_ALIGN,
1163                                              NULL);
1164         if (!rpc_buffer_slabp)
1165                 goto err_nomem;
1166         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1167                                                     rpc_task_slabp);
1168         if (!rpc_task_mempool)
1169                 goto err_nomem;
1170         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1171                                                       rpc_buffer_slabp);
1172         if (!rpc_buffer_mempool)
1173                 goto err_nomem;
1174         return 0;
1175 err_nomem:
1176         rpc_destroy_mempool();
1177         return -ENOMEM;
1178 }