net/sunrpc/sched.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * linux/net/sunrpc/sched.c
   4  *
   5  * Scheduling for synchronous and asynchronous RPC requests.
   6  *
   7  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
   8  *
   9  * TCP NFS related read + write fixes
  10  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  11  */
  12
  13 #include <linux/module.h>
  14
  15 #include <linux/sched.h>
  16 #include <linux/interrupt.h>
  17 #include <linux/slab.h>
  18 #include <linux/mempool.h>
  19 #include <linux/smp.h>
  20 #include <linux/spinlock.h>
  21 #include <linux/mutex.h>
  22 #include <linux/freezer.h>
  23 #include <linux/sched/mm.h>
  24
  25 #include <linux/sunrpc/clnt.h>
  26 #include <linux/sunrpc/metrics.h>
  27
  28 #include "sunrpc.h"
  29
  30 #define CREATE_TRACE_POINTS
  31 #include <trace/events/sunrpc.h>
  32
  33 /*
  34  * RPC slabs and memory pools
  35  */
  36 #define RPC_BUFFER_MAXSIZE      (2048)
  37 #define RPC_BUFFER_POOLSIZE     (8)
  38 #define RPC_TASK_POOLSIZE       (8)
  39 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  40 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  41 static mempool_t        *rpc_task_mempool __read_mostly;
  42 static mempool_t        *rpc_buffer_mempool __read_mostly;
  43
  44 static void                     rpc_async_schedule(struct work_struct *);
  45 static void                      rpc_release_task(struct rpc_task *task);
  46 static void __rpc_queue_timer_fn(struct work_struct *);
  47
  48 /*
  49  * RPC tasks sit here while waiting for conditions to improve.
  50  */
  51 static struct rpc_wait_queue delay_queue;
  52
  53 /*
  54  * rpciod-related stuff
  55  */
  56 struct workqueue_struct *rpciod_workqueue __read_mostly;
  57 struct workqueue_struct *xprtiod_workqueue __read_mostly;
  58 EXPORT_SYMBOL_GPL(xprtiod_workqueue);
  59
  60 unsigned long
  61 rpc_task_timeout(const struct rpc_task *task)
  62 {
  63         unsigned long timeout = READ_ONCE(task->tk_timeout);
  64
  65         if (timeout != 0) {
  66                 unsigned long now = jiffies;
  67                 if (time_before(now, timeout))
  68                         return timeout - now;
  69         }
  70         return 0;
  71 }
  72 EXPORT_SYMBOL_GPL(rpc_task_timeout);
  73
  74 /*
  75  * Disable the timer for a given RPC task. Should be called with
  76  * queue->lock and bh_disabled in order to avoid races within
  77  * rpc_run_timer().
  78  */
  79 static void
  80 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  81 {
  82         if (list_empty(&task->u.tk_wait.timer_list))
  83                 return;
  84         task->tk_timeout = 0;
  85         list_del(&task->u.tk_wait.timer_list);
  86         if (list_empty(&queue->timer_list.list))
  87                 cancel_delayed_work(&queue->timer_list.dwork);
  88 }
  89
  90 static void
  91 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  92 {
  93         unsigned long now = jiffies;
  94         queue->timer_list.expires = expires;
  95         if (time_before_eq(expires, now))
  96                 expires = 0;
  97         else
  98                 expires -= now;
  99         mod_delayed_work(rpciod_workqueue, &queue->timer_list.dwork, expires);
 100 }
 101
 102 /*
 103  * Set up a timer for the current task.
 104  */
 105 static void
 106 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task,
 107                 unsigned long timeout)
 108 {
 109         task->tk_timeout = timeout;
 110         if (list_empty(&queue->timer_list.list) || time_before(timeout, queue->timer_list.expires))
 111                 rpc_set_queue_timer(queue, timeout);
 112         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
 113 }
 114
 115 static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 116 {
 117         if (queue->priority != priority) {
 118                 queue->priority = priority;
 119                 queue->nr = 1U << priority;
 120         }
 121 }
 122
 123 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 124 {
 125         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 126 }
 127
 128 /*
 129  * Add a request to a queue list
 130  */
 131 static void
 132 __rpc_list_enqueue_task(struct list_head *q, struct rpc_task *task)
 133 {
 134         struct rpc_task *t;
 135
 136         list_for_each_entry(t, q, u.tk_wait.list) {
 137                 if (t->tk_owner == task->tk_owner) {
 138                         list_add_tail(&task->u.tk_wait.links,
 139                                         &t->u.tk_wait.links);
 140                         /* Cache the queue head in task->u.tk_wait.list */
 141                         task->u.tk_wait.list.next = q;
 142                         task->u.tk_wait.list.prev = NULL;
 143                         return;
 144                 }
 145         }
 146         INIT_LIST_HEAD(&task->u.tk_wait.links);
 147         list_add_tail(&task->u.tk_wait.list, q);
 148 }
 149
 150 /*
 151  * Remove request from a queue list
 152  */
 153 static void
 154 __rpc_list_dequeue_task(struct rpc_task *task)
 155 {
 156         struct list_head *q;
 157         struct rpc_task *t;
 158
 159         if (task->u.tk_wait.list.prev == NULL) {
 160                 list_del(&task->u.tk_wait.links);
 161                 return;
 162         }
 163         if (!list_empty(&task->u.tk_wait.links)) {
 164                 t = list_first_entry(&task->u.tk_wait.links,
 165                                 struct rpc_task,
 166                                 u.tk_wait.links);
 167                 /* Assume __rpc_list_enqueue_task() cached the queue head */
 168                 q = t->u.tk_wait.list.next;
 169                 list_add_tail(&t->u.tk_wait.list, q);
 170                 list_del(&task->u.tk_wait.links);
 171         }
 172         list_del(&task->u.tk_wait.list);
 173 }
 174
 175 /*
 176  * Add new request to a priority queue.
 177  */
 178 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 179                 struct rpc_task *task,
 180                 unsigned char queue_priority)
 181 {
 182         if (unlikely(queue_priority > queue->maxpriority))
 183                 queue_priority = queue->maxpriority;
 184         __rpc_list_enqueue_task(&queue->tasks[queue_priority], task);
 185 }
 186
 187 /*
 188  * Add new request to wait queue.
 189  *
 190  * Swapper tasks always get inserted at the head of the queue.
 191  * This should avoid many nasty memory deadlocks and hopefully
 192  * improve overall performance.
 193  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 194  */
 195 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 196                 struct rpc_task *task,
 197                 unsigned char queue_priority)
 198 {
 199         INIT_LIST_HEAD(&task->u.tk_wait.timer_list);
 200         if (RPC_IS_PRIORITY(queue))
 201                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 202         else if (RPC_IS_SWAPPER(task))
 203                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 204         else
 205                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 206         task->tk_waitqueue = queue;
 207         queue->qlen++;
 208         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 209         smp_wmb();
 210         rpc_set_queued(task);
 211 }
 212
 213 /*
 214  * Remove request from a priority queue.
 215  */
 216 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 217 {
 218         __rpc_list_dequeue_task(task);
 219 }
 220
 221 /*
 222  * Remove request from queue.
 223  * Note: must be called with spin lock held.
 224  */
 225 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 226 {
 227         __rpc_disable_timer(queue, task);
 228         if (RPC_IS_PRIORITY(queue))
 229                 __rpc_remove_wait_queue_priority(task);
 230         else
 231                 list_del(&task->u.tk_wait.list);
 232         queue->qlen--;
 233 }
 234
 235 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 236 {
 237         int i;
 238
 239         spin_lock_init(&queue->lock);
 240         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 241                 INIT_LIST_HEAD(&queue->tasks[i]);
 242         queue->maxpriority = nr_queues - 1;
 243         rpc_reset_waitqueue_priority(queue);
 244         queue->qlen = 0;
 245         queue->timer_list.expires = 0;
 246         INIT_DELAYED_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
 247         INIT_LIST_HEAD(&queue->timer_list.list);
 248         rpc_assign_waitqueue_name(queue, qname);
 249 }
 250
 251 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 252 {
 253         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 254 }
 255 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 256
 257 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 258 {
 259         __rpc_init_priority_wait_queue(queue, qname, 1);
 260 }
 261 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 262
 263 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 264 {
 265         cancel_delayed_work_sync(&queue->timer_list.dwork);
 266 }
 267 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 268
 269 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
 270 {
 271         freezable_schedule_unsafe();
 272         if (signal_pending_state(mode, current))
 273                 return -ERESTARTSYS;
 274         return 0;
 275 }
 276
 277 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
 278 static void rpc_task_set_debuginfo(struct rpc_task *task)
 279 {
 280         static atomic_t rpc_pid;
 281
 282         task->tk_pid = atomic_inc_return(&rpc_pid);
 283 }
 284 #else
 285 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 286 {
 287 }
 288 #endif
 289
 290 static void rpc_set_active(struct rpc_task *task)
 291 {
 292         rpc_task_set_debuginfo(task);
 293         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 294         trace_rpc_task_begin(task, NULL);
 295 }
 296
 297 /*
 298  * Mark an RPC call as having completed by clearing the 'active' bit
 299  * and then waking up all tasks that were sleeping.
 300  */
 301 static int rpc_complete_task(struct rpc_task *task)
 302 {
 303         void *m = &task->tk_runstate;
 304         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 305         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 306         unsigned long flags;
 307         int ret;
 308
 309         trace_rpc_task_complete(task, NULL);
 310
 311         spin_lock_irqsave(&wq->lock, flags);
 312         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 313         ret = atomic_dec_and_test(&task->tk_count);
 314         if (waitqueue_active(wq))
 315                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 316         spin_unlock_irqrestore(&wq->lock, flags);
 317         return ret;
 318 }
 319
 320 /*
 321  * Allow callers to wait for completion of an RPC call
 322  *
 323  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 324  * to enforce taking of the wq->lock and hence avoid races with
 325  * rpc_complete_task().
 326  */
 327 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
 328 {
 329         if (action == NULL)
 330                 action = rpc_wait_bit_killable;
 331         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 332                         action, TASK_KILLABLE);
 333 }
 334 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 335
 336 /*
 337  * Make an RPC task runnable.
 338  *
 339  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 340  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 341  * the wait queue operation.
 342  * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
 343  * which is needed to ensure that __rpc_execute() doesn't loop (due to the
 344  * lockless RPC_IS_QUEUED() test) before we've had a chance to test
 345  * the RPC_TASK_RUNNING flag.
 346  */
 347 static void rpc_make_runnable(struct workqueue_struct *wq,
 348                 struct rpc_task *task)
 349 {
 350         bool need_wakeup = !rpc_test_and_set_running(task);
 351
 352         rpc_clear_queued(task);
 353         if (!need_wakeup)
 354                 return;
 355         if (RPC_IS_ASYNC(task)) {
 356                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 357                 queue_work(wq, &task->u.tk_work);
 358         } else
 359                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 360 }
 361
 362 /*
 363  * Prepare for sleeping on a wait queue.
 364  * By always appending tasks to the list we ensure FIFO behavior.
 365  * NB: An RPC task will only receive interrupt-driven events as long
 366  * as it's on a wait queue.
 367  */
 368 static void __rpc_do_sleep_on_priority(struct rpc_wait_queue *q,
 369                 struct rpc_task *task,
 370                 unsigned char queue_priority)
 371 {
 372         trace_rpc_task_sleep(task, q);
 373
 374         __rpc_add_wait_queue(q, task, queue_priority);
 375 }
 376
 377 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 378                 struct rpc_task *task,
 379                 unsigned char queue_priority)
 380 {
 381         if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
 382                 return;
 383         __rpc_do_sleep_on_priority(q, task, queue_priority);
 384 }
 385
 386 static void __rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
 387                 struct rpc_task *task, unsigned long timeout,
 388                 unsigned char queue_priority)
 389 {
 390         if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
 391                 return;
 392         if (time_is_after_jiffies(timeout)) {
 393                 __rpc_do_sleep_on_priority(q, task, queue_priority);
 394                 __rpc_add_timer(q, task, timeout);
 395         } else
 396                 task->tk_status = -ETIMEDOUT;
 397 }
 398
 399 static void rpc_set_tk_callback(struct rpc_task *task, rpc_action action)
 400 {
 401         if (action && !WARN_ON_ONCE(task->tk_callback != NULL))
 402                 task->tk_callback = action;
 403 }
 404
 405 static bool rpc_sleep_check_activated(struct rpc_task *task)
 406 {
 407         /* We shouldn't ever put an inactive task to sleep */
 408         if (WARN_ON_ONCE(!RPC_IS_ACTIVATED(task))) {
 409                 task->tk_status = -EIO;
 410                 rpc_put_task_async(task);
 411                 return false;
 412         }
 413         return true;
 414 }
 415
 416 void rpc_sleep_on_timeout(struct rpc_wait_queue *q, struct rpc_task *task,
 417                                 rpc_action action, unsigned long timeout)
 418 {
 419         if (!rpc_sleep_check_activated(task))
 420                 return;
 421
 422         rpc_set_tk_callback(task, action);
 423
 424         /*
 425          * Protect the queue operations.
 426          */
 427         spin_lock(&q->lock);
 428         __rpc_sleep_on_priority_timeout(q, task, timeout, task->tk_priority);
 429         spin_unlock(&q->lock);
 430 }
 431 EXPORT_SYMBOL_GPL(rpc_sleep_on_timeout);
 432
 433 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 434                                 rpc_action action)
 435 {
 436         if (!rpc_sleep_check_activated(task))
 437                 return;
 438
 439         rpc_set_tk_callback(task, action);
 440
 441         WARN_ON_ONCE(task->tk_timeout != 0);
 442         /*
 443          * Protect the queue operations.
 444          */
 445         spin_lock(&q->lock);
 446         __rpc_sleep_on_priority(q, task, task->tk_priority);
 447         spin_unlock(&q->lock);
 448 }
 449 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 450
 451 void rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
 452                 struct rpc_task *task, unsigned long timeout, int priority)
 453 {
 454         if (!rpc_sleep_check_activated(task))
 455                 return;
 456
 457         priority -= RPC_PRIORITY_LOW;
 458         /*
 459          * Protect the queue operations.
 460          */
 461         spin_lock(&q->lock);
 462         __rpc_sleep_on_priority_timeout(q, task, timeout, priority);
 463         spin_unlock(&q->lock);
 464 }
 465 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority_timeout);
 466
 467 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 468                 int priority)
 469 {
 470         if (!rpc_sleep_check_activated(task))
 471                 return;
 472
 473         WARN_ON_ONCE(task->tk_timeout != 0);
 474         priority -= RPC_PRIORITY_LOW;
 475         /*
 476          * Protect the queue operations.
 477          */
 478         spin_lock(&q->lock);
 479         __rpc_sleep_on_priority(q, task, priority);
 480         spin_unlock(&q->lock);
 481 }
 482 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 483
 484 /**
 485  * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
 486  * @wq: workqueue on which to run task
 487  * @queue: wait queue
 488  * @task: task to be woken up
 489  *
 490  * Caller must hold queue->lock, and have cleared the task queued flag.
 491  */
 492 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
 493                 struct rpc_wait_queue *queue,
 494                 struct rpc_task *task)
 495 {
 496         /* Has the task been executed yet? If not, we cannot wake it up! */
 497         if (!RPC_IS_ACTIVATED(task)) {
 498                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 499                 return;
 500         }
 501
 502         trace_rpc_task_wakeup(task, queue);
 503
 504         __rpc_remove_wait_queue(queue, task);
 505
 506         rpc_make_runnable(wq, task);
 507 }
 508
 509 /*
 510  * Wake up a queued task while the queue lock is being held
 511  */
 512 static struct rpc_task *
 513 rpc_wake_up_task_on_wq_queue_action_locked(struct workqueue_struct *wq,
 514                 struct rpc_wait_queue *queue, struct rpc_task *task,
 515                 bool (*action)(struct rpc_task *, void *), void *data)
 516 {
 517         if (RPC_IS_QUEUED(task)) {
 518                 smp_rmb();
 519                 if (task->tk_waitqueue == queue) {
 520                         if (action == NULL || action(task, data)) {
 521                                 __rpc_do_wake_up_task_on_wq(wq, queue, task);
 522                                 return task;
 523                         }
 524                 }
 525         }
 526         return NULL;
 527 }
 528
 529 /*
 530  * Wake up a queued task while the queue lock is being held
 531  */
 532 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue,
 533                                           struct rpc_task *task)
 534 {
 535         rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
 536                                                    task, NULL, NULL);
 537 }
 538
 539 /*
 540  * Wake up a task on a specific queue
 541  */
 542 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 543 {
 544         if (!RPC_IS_QUEUED(task))
 545                 return;
 546         spin_lock(&queue->lock);
 547         rpc_wake_up_task_queue_locked(queue, task);
 548         spin_unlock(&queue->lock);
 549 }
 550 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 551
 552 static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
 553 {
 554         task->tk_status = *(int *)status;
 555         return true;
 556 }
 557
 558 static void
 559 rpc_wake_up_task_queue_set_status_locked(struct rpc_wait_queue *queue,
 560                 struct rpc_task *task, int status)
 561 {
 562         rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
 563                         task, rpc_task_action_set_status, &status);
 564 }
 565
 566 /**
 567  * rpc_wake_up_queued_task_set_status - wake up a task and set task->tk_status
 568  * @queue: pointer to rpc_wait_queue
 569  * @task: pointer to rpc_task
 570  * @status: integer error value
 571  *
 572  * If @task is queued on @queue, then it is woken up, and @task->tk_status is
 573  * set to the value of @status.
 574  */
 575 void
 576 rpc_wake_up_queued_task_set_status(struct rpc_wait_queue *queue,
 577                 struct rpc_task *task, int status)
 578 {
 579         if (!RPC_IS_QUEUED(task))
 580                 return;
 581         spin_lock(&queue->lock);
 582         rpc_wake_up_task_queue_set_status_locked(queue, task, status);
 583         spin_unlock(&queue->lock);
 584 }
 585
 586 /*
 587  * Wake up the next task on a priority queue.
 588  */
 589 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 590 {
 591         struct list_head *q;
 592         struct rpc_task *task;
 593
 594         /*
 595          * Service a batch of tasks from a single owner.
 596          */
 597         q = &queue->tasks[queue->priority];
 598         if (!list_empty(q) && --queue->nr) {
 599                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 600                 goto out;
 601         }
 602
 603         /*
 604          * Service the next queue.
 605          */
 606         do {
 607                 if (q == &queue->tasks[0])
 608                         q = &queue->tasks[queue->maxpriority];
 609                 else
 610                         q = q - 1;
 611                 if (!list_empty(q)) {
 612                         task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 613                         goto new_queue;
 614                 }
 615         } while (q != &queue->tasks[queue->priority]);
 616
 617         rpc_reset_waitqueue_priority(queue);
 618         return NULL;
 619
 620 new_queue:
 621         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 622 out:
 623         return task;
 624 }
 625
 626 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 627 {
 628         if (RPC_IS_PRIORITY(queue))
 629                 return __rpc_find_next_queued_priority(queue);
 630         if (!list_empty(&queue->tasks[0]))
 631                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 632         return NULL;
 633 }
 634
 635 /*
 636  * Wake up the first task on the wait queue.
 637  */
 638 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 639                 struct rpc_wait_queue *queue,
 640                 bool (*func)(struct rpc_task *, void *), void *data)
 641 {
 642         struct rpc_task *task = NULL;
 643
 644         spin_lock(&queue->lock);
 645         task = __rpc_find_next_queued(queue);
 646         if (task != NULL)
 647                 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
 648                                 task, func, data);
 649         spin_unlock(&queue->lock);
 650
 651         return task;
 652 }
 653
 654 /*
 655  * Wake up the first task on the wait queue.
 656  */
 657 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 658                 bool (*func)(struct rpc_task *, void *), void *data)
 659 {
 660         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 661 }
 662 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 663
 664 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 665 {
 666         return true;
 667 }
 668
 669 /*
 670  * Wake up the next task on the wait queue.
 671 */
 672 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 673 {
 674         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 675 }
 676 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 677
 678 /**
 679  * rpc_wake_up - wake up all rpc_tasks
 680  * @queue: rpc_wait_queue on which the tasks are sleeping
 681  *
 682  * Grabs queue->lock
 683  */
 684 void rpc_wake_up(struct rpc_wait_queue *queue)
 685 {
 686         struct list_head *head;
 687
 688         spin_lock(&queue->lock);
 689         head = &queue->tasks[queue->maxpriority];
 690         for (;;) {
 691                 while (!list_empty(head)) {
 692                         struct rpc_task *task;
 693                         task = list_first_entry(head,
 694                                         struct rpc_task,
 695                                         u.tk_wait.list);
 696                         rpc_wake_up_task_queue_locked(queue, task);
 697                 }
 698                 if (head == &queue->tasks[0])
 699                         break;
 700                 head--;
 701         }
 702         spin_unlock(&queue->lock);
 703 }
 704 EXPORT_SYMBOL_GPL(rpc_wake_up);
 705
 706 /**
 707  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 708  * @queue: rpc_wait_queue on which the tasks are sleeping
 709  * @status: status value to set
 710  *
 711  * Grabs queue->lock
 712  */
 713 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 714 {
 715         struct list_head *head;
 716
 717         spin_lock(&queue->lock);
 718         head = &queue->tasks[queue->maxpriority];
 719         for (;;) {
 720                 while (!list_empty(head)) {
 721                         struct rpc_task *task;
 722                         task = list_first_entry(head,
 723                                         struct rpc_task,
 724                                         u.tk_wait.list);
 725                         task->tk_status = status;
 726                         rpc_wake_up_task_queue_locked(queue, task);
 727                 }
 728                 if (head == &queue->tasks[0])
 729                         break;
 730                 head--;
 731         }
 732         spin_unlock(&queue->lock);
 733 }
 734 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 735
 736 static void __rpc_queue_timer_fn(struct work_struct *work)
 737 {
 738         struct rpc_wait_queue *queue = container_of(work,
 739                         struct rpc_wait_queue,
 740                         timer_list.dwork.work);
 741         struct rpc_task *task, *n;
 742         unsigned long expires, now, timeo;
 743
 744         spin_lock(&queue->lock);
 745         expires = now = jiffies;
 746         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 747                 timeo = task->tk_timeout;
 748                 if (time_after_eq(now, timeo)) {
 749                         trace_rpc_task_timeout(task, task->tk_action);
 750                         task->tk_status = -ETIMEDOUT;
 751                         rpc_wake_up_task_queue_locked(queue, task);
 752                         continue;
 753                 }
 754                 if (expires == now || time_after(expires, timeo))
 755                         expires = timeo;
 756         }
 757         if (!list_empty(&queue->timer_list.list))
 758                 rpc_set_queue_timer(queue, expires);
 759         spin_unlock(&queue->lock);
 760 }
 761
 762 static void __rpc_atrun(struct rpc_task *task)
 763 {
 764         if (task->tk_status == -ETIMEDOUT)
 765                 task->tk_status = 0;
 766 }
 767
 768 /*
 769  * Run a task at a later time
 770  */
 771 void rpc_delay(struct rpc_task *task, unsigned long delay)
 772 {
 773         rpc_sleep_on_timeout(&delay_queue, task, __rpc_atrun, jiffies + delay);
 774 }
 775 EXPORT_SYMBOL_GPL(rpc_delay);
 776
 777 /*
 778  * Helper to call task->tk_ops->rpc_call_prepare
 779  */
 780 void rpc_prepare_task(struct rpc_task *task)
 781 {
 782         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 783 }
 784
 785 static void
 786 rpc_init_task_statistics(struct rpc_task *task)
 787 {
 788         /* Initialize retry counters */
 789         task->tk_garb_retry = 2;
 790         task->tk_cred_retry = 2;
 791         task->tk_rebind_retry = 2;
 792
 793         /* starting timestamp */
 794         task->tk_start = ktime_get();
 795 }
 796
 797 static void
 798 rpc_reset_task_statistics(struct rpc_task *task)
 799 {
 800         task->tk_timeouts = 0;
 801         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
 802         rpc_init_task_statistics(task);
 803 }
 804
 805 /*
 806  * Helper that calls task->tk_ops->rpc_call_done if it exists
 807  */
 808 void rpc_exit_task(struct rpc_task *task)
 809 {
 810         trace_rpc_task_end(task, task->tk_action);
 811         task->tk_action = NULL;
 812         if (task->tk_ops->rpc_count_stats)
 813                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
 814         else if (task->tk_client)
 815                 rpc_count_iostats(task, task->tk_client->cl_metrics);
 816         if (task->tk_ops->rpc_call_done != NULL) {
 817                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 818                 if (task->tk_action != NULL) {
 819                         /* Always release the RPC slot and buffer memory */
 820                         xprt_release(task);
 821                         rpc_reset_task_statistics(task);
 822                 }
 823         }
 824 }
 825
 826 void rpc_signal_task(struct rpc_task *task)
 827 {
 828         struct rpc_wait_queue *queue;
 829
 830         if (!RPC_IS_ACTIVATED(task))
 831                 return;
 832
 833         trace_rpc_task_signalled(task, task->tk_action);
 834         set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
 835         smp_mb__after_atomic();
 836         queue = READ_ONCE(task->tk_waitqueue);
 837         if (queue)
 838                 rpc_wake_up_queued_task_set_status(queue, task, -ERESTARTSYS);
 839 }
 840
 841 void rpc_exit(struct rpc_task *task, int status)
 842 {
 843         task->tk_status = status;
 844         task->tk_action = rpc_exit_task;
 845         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 846 }
 847 EXPORT_SYMBOL_GPL(rpc_exit);
 848
 849 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 850 {
 851         if (ops->rpc_release != NULL)
 852                 ops->rpc_release(calldata);
 853 }
 854
 855 /*
 856  * This is the RPC `scheduler' (or rather, the finite state machine).
 857  */
 858 static void __rpc_execute(struct rpc_task *task)
 859 {
 860         struct rpc_wait_queue *queue;
 861         int task_is_async = RPC_IS_ASYNC(task);
 862         int status = 0;
 863
 864         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 865         if (RPC_IS_QUEUED(task))
 866                 return;
 867
 868         for (;;) {
 869                 void (*do_action)(struct rpc_task *);
 870
 871                 /*
 872                  * Perform the next FSM step or a pending callback.
 873                  *
 874                  * tk_action may be NULL if the task has been killed.
 875                  * In particular, note that rpc_killall_tasks may
 876                  * do this at any time, so beware when dereferencing.
 877                  */
 878                 do_action = task->tk_action;
 879                 if (task->tk_callback) {
 880                         do_action = task->tk_callback;
 881                         task->tk_callback = NULL;
 882                 }
 883                 if (!do_action)
 884                         break;
 885                 trace_rpc_task_run_action(task, do_action);
 886                 do_action(task);
 887
 888                 /*
 889                  * Lockless check for whether task is sleeping or not.
 890                  */
 891                 if (!RPC_IS_QUEUED(task))
 892                         continue;
 893
 894                 /*
 895                  * Signalled tasks should exit rather than sleep.
 896                  */
 897                 if (RPC_SIGNALLED(task)) {
 898                         task->tk_rpc_status = -ERESTARTSYS;
 899                         rpc_exit(task, -ERESTARTSYS);
 900                 }
 901
 902                 /*
 903                  * The queue->lock protects against races with
 904                  * rpc_make_runnable().
 905                  *
 906                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 907                  * rpc_task, rpc_make_runnable() can assign it to a
 908                  * different workqueue. We therefore cannot assume that the
 909                  * rpc_task pointer may still be dereferenced.
 910                  */
 911                 queue = task->tk_waitqueue;
 912                 spin_lock(&queue->lock);
 913                 if (!RPC_IS_QUEUED(task)) {
 914                         spin_unlock(&queue->lock);
 915                         continue;
 916                 }
 917                 rpc_clear_running(task);
 918                 spin_unlock(&queue->lock);
 919                 if (task_is_async)
 920                         return;
 921
 922                 /* sync task: sleep here */
 923                 trace_rpc_task_sync_sleep(task, task->tk_action);
 924                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 925                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 926                                 TASK_KILLABLE);
 927                 if (status < 0) {
 928                         /*
 929                          * When a sync task receives a signal, it exits with
 930                          * -ERESTARTSYS. In order to catch any callbacks that
 931                          * clean up after sleeping on some queue, we don't
 932                          * break the loop here, but go around once more.
 933                          */
 934                         trace_rpc_task_signalled(task, task->tk_action);
 935                         set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
 936                         task->tk_rpc_status = -ERESTARTSYS;
 937                         rpc_exit(task, -ERESTARTSYS);
 938                 }
 939                 trace_rpc_task_sync_wake(task, task->tk_action);
 940         }
 941
 942         /* Release all resources associated with the task */
 943         rpc_release_task(task);
 944 }
 945
 946 /*
 947  * User-visible entry point to the scheduler.
 948  *
 949  * This may be called recursively if e.g. an async NFS task updates
 950  * the attributes and finds that dirty pages must be flushed.
 951  * NOTE: Upon exit of this function the task is guaranteed to be
 952  *       released. In particular note that tk_release() will have
 953  *       been called, so your task memory may have been freed.
 954  */
 955 void rpc_execute(struct rpc_task *task)
 956 {
 957         bool is_async = RPC_IS_ASYNC(task);
 958
 959         rpc_set_active(task);
 960         rpc_make_runnable(rpciod_workqueue, task);
 961         if (!is_async)
 962                 __rpc_execute(task);
 963 }
 964
 965 static void rpc_async_schedule(struct work_struct *work)
 966 {
 967         unsigned int pflags = memalloc_nofs_save();
 968
 969         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 970         memalloc_nofs_restore(pflags);
 971 }
 972
 973 /**
 974  * rpc_malloc - allocate RPC buffer resources
 975  * @task: RPC task
 976  *
 977  * A single memory region is allocated, which is split between the
 978  * RPC call and RPC reply that this task is being used for. When
 979  * this RPC is retired, the memory is released by calling rpc_free.
 980  *
 981  * To prevent rpciod from hanging, this allocator never sleeps,
 982  * returning -ENOMEM and suppressing warning if the request cannot
 983  * be serviced immediately. The caller can arrange to sleep in a
 984  * way that is safe for rpciod.
 985  *
 986  * Most requests are 'small' (under 2KiB) and can be serviced from a
 987  * mempool, ensuring that NFS reads and writes can always proceed,
 988  * and that there is good locality of reference for these buffers.
 989  */
 990 int rpc_malloc(struct rpc_task *task)
 991 {
 992         struct rpc_rqst *rqst = task->tk_rqstp;
 993         size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
 994         struct rpc_buffer *buf;
 995         gfp_t gfp = GFP_NOFS;
 996
 997         if (RPC_IS_SWAPPER(task))
 998                 gfp = __GFP_MEMALLOC | GFP_NOWAIT | __GFP_NOWARN;
 999
1000         size += sizeof(struct rpc_buffer);
1001         if (size <= RPC_BUFFER_MAXSIZE)
1002                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
1003         else
1004                 buf = kmalloc(size, gfp);
1005
1006         if (!buf)
1007                 return -ENOMEM;
1008
1009         buf->len = size;
1010         rqst->rq_buffer = buf->data;
1011         rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1012         return 0;
1013 }
1014 EXPORT_SYMBOL_GPL(rpc_malloc);
1015
1016 /**
1017  * rpc_free - free RPC buffer resources allocated via rpc_malloc
1018  * @task: RPC task
1019  *
1020  */
1021 void rpc_free(struct rpc_task *task)
1022 {
1023         void *buffer = task->tk_rqstp->rq_buffer;
1024         size_t size;
1025         struct rpc_buffer *buf;
1026
1027         buf = container_of(buffer, struct rpc_buffer, data);
1028         size = buf->len;
1029
1030         if (size <= RPC_BUFFER_MAXSIZE)
1031                 mempool_free(buf, rpc_buffer_mempool);
1032         else
1033                 kfree(buf);
1034 }
1035 EXPORT_SYMBOL_GPL(rpc_free);
1036
1037 /*
1038  * Creation and deletion of RPC task structures
1039  */
1040 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1041 {
1042         memset(task, 0, sizeof(*task));
1043         atomic_set(&task->tk_count, 1);
1044         task->tk_flags  = task_setup_data->flags;
1045         task->tk_ops = task_setup_data->callback_ops;
1046         task->tk_calldata = task_setup_data->callback_data;
1047         INIT_LIST_HEAD(&task->tk_task);
1048
1049         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1050         task->tk_owner = current->tgid;
1051
1052         /* Initialize workqueue for async tasks */
1053         task->tk_workqueue = task_setup_data->workqueue;
1054
1055         task->tk_xprt = rpc_task_get_xprt(task_setup_data->rpc_client,
1056                         xprt_get(task_setup_data->rpc_xprt));
1057
1058         task->tk_op_cred = get_rpccred(task_setup_data->rpc_op_cred);
1059
1060         if (task->tk_ops->rpc_call_prepare != NULL)
1061                 task->tk_action = rpc_prepare_task;
1062
1063         rpc_init_task_statistics(task);
1064 }
1065
1066 static struct rpc_task *
1067 rpc_alloc_task(void)
1068 {
1069         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
1070 }
1071
1072 /*
1073  * Create a new task for the specified client.
1074  */
1075 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1076 {
1077         struct rpc_task *task = setup_data->task;
1078         unsigned short flags = 0;
1079
1080         if (task == NULL) {
1081                 task = rpc_alloc_task();
1082                 flags = RPC_TASK_DYNAMIC;
1083         }
1084
1085         rpc_init_task(task, setup_data);
1086         task->tk_flags |= flags;
1087         return task;
1088 }
1089
1090 /*
1091  * rpc_free_task - release rpc task and perform cleanups
1092  *
1093  * Note that we free up the rpc_task _after_ rpc_release_calldata()
1094  * in order to work around a workqueue dependency issue.
1095  *
1096  * Tejun Heo states:
1097  * "Workqueue currently considers two work items to be the same if they're
1098  * on the same address and won't execute them concurrently - ie. it
1099  * makes a work item which is queued again while being executed wait
1100  * for the previous execution to complete.
1101  *
1102  * If a work function frees the work item, and then waits for an event
1103  * which should be performed by another work item and *that* work item
1104  * recycles the freed work item, it can create a false dependency loop.
1105  * There really is no reliable way to detect this short of verifying
1106  * every memory free."
1107  *
1108  */
1109 static void rpc_free_task(struct rpc_task *task)
1110 {
1111         unsigned short tk_flags = task->tk_flags;
1112
1113         put_rpccred(task->tk_op_cred);
1114         rpc_release_calldata(task->tk_ops, task->tk_calldata);
1115
1116         if (tk_flags & RPC_TASK_DYNAMIC)
1117                 mempool_free(task, rpc_task_mempool);
1118 }
1119
1120 static void rpc_async_release(struct work_struct *work)
1121 {
1122         unsigned int pflags = memalloc_nofs_save();
1123
1124         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1125         memalloc_nofs_restore(pflags);
1126 }
1127
1128 static void rpc_release_resources_task(struct rpc_task *task)
1129 {
1130         xprt_release(task);
1131         if (task->tk_msg.rpc_cred) {
1132                 if (!(task->tk_flags & RPC_TASK_CRED_NOREF))
1133                         put_cred(task->tk_msg.rpc_cred);
1134                 task->tk_msg.rpc_cred = NULL;
1135         }
1136         rpc_task_release_client(task);
1137 }
1138
1139 static void rpc_final_put_task(struct rpc_task *task,
1140                 struct workqueue_struct *q)
1141 {
1142         if (q != NULL) {
1143                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1144                 queue_work(q, &task->u.tk_work);
1145         } else
1146                 rpc_free_task(task);
1147 }
1148
1149 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1150 {
1151         if (atomic_dec_and_test(&task->tk_count)) {
1152                 rpc_release_resources_task(task);
1153                 rpc_final_put_task(task, q);
1154         }
1155 }
1156
1157 void rpc_put_task(struct rpc_task *task)
1158 {
1159         rpc_do_put_task(task, NULL);
1160 }
1161 EXPORT_SYMBOL_GPL(rpc_put_task);
1162
1163 void rpc_put_task_async(struct rpc_task *task)
1164 {
1165         rpc_do_put_task(task, task->tk_workqueue);
1166 }
1167 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1168
1169 static void rpc_release_task(struct rpc_task *task)
1170 {
1171         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1172
1173         rpc_release_resources_task(task);
1174
1175         /*
1176          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1177          * so it should be safe to use task->tk_count as a test for whether
1178          * or not any other processes still hold references to our rpc_task.
1179          */
1180         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1181                 /* Wake up anyone who may be waiting for task completion */
1182                 if (!rpc_complete_task(task))
1183                         return;
1184         } else {
1185                 if (!atomic_dec_and_test(&task->tk_count))
1186                         return;
1187         }
1188         rpc_final_put_task(task, task->tk_workqueue);
1189 }
1190
1191 int rpciod_up(void)
1192 {
1193         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1194 }
1195
1196 void rpciod_down(void)
1197 {
1198         module_put(THIS_MODULE);
1199 }
1200
1201 /*
1202  * Start up the rpciod workqueue.
1203  */
1204 static int rpciod_start(void)
1205 {
1206         struct workqueue_struct *wq;
1207
1208         /*
1209          * Create the rpciod thread and wait for it to start.
1210          */
1211         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1212         if (!wq)
1213                 goto out_failed;
1214         rpciod_workqueue = wq;
1215         /* Note: highpri because network receive is latency sensitive */
1216         wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1217         if (!wq)
1218                 goto free_rpciod;
1219         xprtiod_workqueue = wq;
1220         return 1;
1221 free_rpciod:
1222         wq = rpciod_workqueue;
1223         rpciod_workqueue = NULL;
1224         destroy_workqueue(wq);
1225 out_failed:
1226         return 0;
1227 }
1228
1229 static void rpciod_stop(void)
1230 {
1231         struct workqueue_struct *wq = NULL;
1232
1233         if (rpciod_workqueue == NULL)
1234                 return;
1235
1236         wq = rpciod_workqueue;
1237         rpciod_workqueue = NULL;
1238         destroy_workqueue(wq);
1239         wq = xprtiod_workqueue;
1240         xprtiod_workqueue = NULL;
1241         destroy_workqueue(wq);
1242 }
1243
1244 void
1245 rpc_destroy_mempool(void)
1246 {
1247         rpciod_stop();
1248         mempool_destroy(rpc_buffer_mempool);
1249         mempool_destroy(rpc_task_mempool);
1250         kmem_cache_destroy(rpc_task_slabp);
1251         kmem_cache_destroy(rpc_buffer_slabp);
1252         rpc_destroy_wait_queue(&delay_queue);
1253 }
1254
1255 int
1256 rpc_init_mempool(void)
1257 {
1258         /*
1259          * The following is not strictly a mempool initialisation,
1260          * but there is no harm in doing it here
1261          */
1262         rpc_init_wait_queue(&delay_queue, "delayq");
1263         if (!rpciod_start())
1264                 goto err_nomem;
1265
1266         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1267                                              sizeof(struct rpc_task),
1268                                              0, SLAB_HWCACHE_ALIGN,
1269                                              NULL);
1270         if (!rpc_task_slabp)
1271                 goto err_nomem;
1272         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1273                                              RPC_BUFFER_MAXSIZE,
1274                                              0, SLAB_HWCACHE_ALIGN,
1275                                              NULL);
1276         if (!rpc_buffer_slabp)
1277                 goto err_nomem;
1278         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1279                                                     rpc_task_slabp);
1280         if (!rpc_task_mempool)
1281                 goto err_nomem;
1282         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1283                                                       rpc_buffer_slabp);
1284         if (!rpc_buffer_mempool)
1285                 goto err_nomem;
1286         return 0;
1287 err_nomem:
1288         rpc_destroy_mempool();
1289         return -ENOMEM;
1290 }