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 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 191                 struct rpc_task *task,
 192                 unsigned char queue_priority)
 193 {
 194         INIT_LIST_HEAD(&task->u.tk_wait.timer_list);
 195         if (RPC_IS_PRIORITY(queue))
 196                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 197         else
 198                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 199         task->tk_waitqueue = queue;
 200         queue->qlen++;
 201         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 202         smp_wmb();
 203         rpc_set_queued(task);
 204 }
 205
 206 /*
 207  * Remove request from a priority queue.
 208  */
 209 static void __rpc_remove_wait_queue_priority(struct rpc_task *task)
 210 {
 211         __rpc_list_dequeue_task(task);
 212 }
 213
 214 /*
 215  * Remove request from queue.
 216  * Note: must be called with spin lock held.
 217  */
 218 static void __rpc_remove_wait_queue(struct rpc_wait_queue *queue, struct rpc_task *task)
 219 {
 220         __rpc_disable_timer(queue, task);
 221         if (RPC_IS_PRIORITY(queue))
 222                 __rpc_remove_wait_queue_priority(task);
 223         else
 224                 list_del(&task->u.tk_wait.list);
 225         queue->qlen--;
 226 }
 227
 228 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 229 {
 230         int i;
 231
 232         spin_lock_init(&queue->lock);
 233         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 234                 INIT_LIST_HEAD(&queue->tasks[i]);
 235         queue->maxpriority = nr_queues - 1;
 236         rpc_reset_waitqueue_priority(queue);
 237         queue->qlen = 0;
 238         queue->timer_list.expires = 0;
 239         INIT_DELAYED_WORK(&queue->timer_list.dwork, __rpc_queue_timer_fn);
 240         INIT_LIST_HEAD(&queue->timer_list.list);
 241         rpc_assign_waitqueue_name(queue, qname);
 242 }
 243
 244 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 245 {
 246         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 247 }
 248 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 249
 250 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 251 {
 252         __rpc_init_priority_wait_queue(queue, qname, 1);
 253 }
 254 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 255
 256 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 257 {
 258         cancel_delayed_work_sync(&queue->timer_list.dwork);
 259 }
 260 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 261
 262 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
 263 {
 264         freezable_schedule_unsafe();
 265         if (signal_pending_state(mode, current))
 266                 return -ERESTARTSYS;
 267         return 0;
 268 }
 269
 270 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
 271 static void rpc_task_set_debuginfo(struct rpc_task *task)
 272 {
 273         static atomic_t rpc_pid;
 274
 275         task->tk_pid = atomic_inc_return(&rpc_pid);
 276 }
 277 #else
 278 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 279 {
 280 }
 281 #endif
 282
 283 static void rpc_set_active(struct rpc_task *task)
 284 {
 285         rpc_task_set_debuginfo(task);
 286         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 287         trace_rpc_task_begin(task, NULL);
 288 }
 289
 290 /*
 291  * Mark an RPC call as having completed by clearing the 'active' bit
 292  * and then waking up all tasks that were sleeping.
 293  */
 294 static int rpc_complete_task(struct rpc_task *task)
 295 {
 296         void *m = &task->tk_runstate;
 297         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 298         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 299         unsigned long flags;
 300         int ret;
 301
 302         trace_rpc_task_complete(task, NULL);
 303
 304         spin_lock_irqsave(&wq->lock, flags);
 305         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 306         ret = atomic_dec_and_test(&task->tk_count);
 307         if (waitqueue_active(wq))
 308                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 309         spin_unlock_irqrestore(&wq->lock, flags);
 310         return ret;
 311 }
 312
 313 /*
 314  * Allow callers to wait for completion of an RPC call
 315  *
 316  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 317  * to enforce taking of the wq->lock and hence avoid races with
 318  * rpc_complete_task().
 319  */
 320 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
 321 {
 322         if (action == NULL)
 323                 action = rpc_wait_bit_killable;
 324         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 325                         action, TASK_KILLABLE);
 326 }
 327 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 328
 329 /*
 330  * Make an RPC task runnable.
 331  *
 332  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 333  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 334  * the wait queue operation.
 335  * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
 336  * which is needed to ensure that __rpc_execute() doesn't loop (due to the
 337  * lockless RPC_IS_QUEUED() test) before we've had a chance to test
 338  * the RPC_TASK_RUNNING flag.
 339  */
 340 static void rpc_make_runnable(struct workqueue_struct *wq,
 341                 struct rpc_task *task)
 342 {
 343         bool need_wakeup = !rpc_test_and_set_running(task);
 344
 345         rpc_clear_queued(task);
 346         if (!need_wakeup)
 347                 return;
 348         if (RPC_IS_ASYNC(task)) {
 349                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 350                 queue_work(wq, &task->u.tk_work);
 351         } else
 352                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 353 }
 354
 355 /*
 356  * Prepare for sleeping on a wait queue.
 357  * By always appending tasks to the list we ensure FIFO behavior.
 358  * NB: An RPC task will only receive interrupt-driven events as long
 359  * as it's on a wait queue.
 360  */
 361 static void __rpc_do_sleep_on_priority(struct rpc_wait_queue *q,
 362                 struct rpc_task *task,
 363                 unsigned char queue_priority)
 364 {
 365         trace_rpc_task_sleep(task, q);
 366
 367         __rpc_add_wait_queue(q, task, queue_priority);
 368 }
 369
 370 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 371                 struct rpc_task *task,
 372                 unsigned char queue_priority)
 373 {
 374         if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
 375                 return;
 376         __rpc_do_sleep_on_priority(q, task, queue_priority);
 377 }
 378
 379 static void __rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
 380                 struct rpc_task *task, unsigned long timeout,
 381                 unsigned char queue_priority)
 382 {
 383         if (WARN_ON_ONCE(RPC_IS_QUEUED(task)))
 384                 return;
 385         if (time_is_after_jiffies(timeout)) {
 386                 __rpc_do_sleep_on_priority(q, task, queue_priority);
 387                 __rpc_add_timer(q, task, timeout);
 388         } else
 389                 task->tk_status = -ETIMEDOUT;
 390 }
 391
 392 static void rpc_set_tk_callback(struct rpc_task *task, rpc_action action)
 393 {
 394         if (action && !WARN_ON_ONCE(task->tk_callback != NULL))
 395                 task->tk_callback = action;
 396 }
 397
 398 static bool rpc_sleep_check_activated(struct rpc_task *task)
 399 {
 400         /* We shouldn't ever put an inactive task to sleep */
 401         if (WARN_ON_ONCE(!RPC_IS_ACTIVATED(task))) {
 402                 task->tk_status = -EIO;
 403                 rpc_put_task_async(task);
 404                 return false;
 405         }
 406         return true;
 407 }
 408
 409 void rpc_sleep_on_timeout(struct rpc_wait_queue *q, struct rpc_task *task,
 410                                 rpc_action action, unsigned long timeout)
 411 {
 412         if (!rpc_sleep_check_activated(task))
 413                 return;
 414
 415         rpc_set_tk_callback(task, action);
 416
 417         /*
 418          * Protect the queue operations.
 419          */
 420         spin_lock(&q->lock);
 421         __rpc_sleep_on_priority_timeout(q, task, timeout, task->tk_priority);
 422         spin_unlock(&q->lock);
 423 }
 424 EXPORT_SYMBOL_GPL(rpc_sleep_on_timeout);
 425
 426 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 427                                 rpc_action action)
 428 {
 429         if (!rpc_sleep_check_activated(task))
 430                 return;
 431
 432         rpc_set_tk_callback(task, action);
 433
 434         WARN_ON_ONCE(task->tk_timeout != 0);
 435         /*
 436          * Protect the queue operations.
 437          */
 438         spin_lock(&q->lock);
 439         __rpc_sleep_on_priority(q, task, task->tk_priority);
 440         spin_unlock(&q->lock);
 441 }
 442 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 443
 444 void rpc_sleep_on_priority_timeout(struct rpc_wait_queue *q,
 445                 struct rpc_task *task, unsigned long timeout, int priority)
 446 {
 447         if (!rpc_sleep_check_activated(task))
 448                 return;
 449
 450         priority -= RPC_PRIORITY_LOW;
 451         /*
 452          * Protect the queue operations.
 453          */
 454         spin_lock(&q->lock);
 455         __rpc_sleep_on_priority_timeout(q, task, timeout, priority);
 456         spin_unlock(&q->lock);
 457 }
 458 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority_timeout);
 459
 460 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 461                 int priority)
 462 {
 463         if (!rpc_sleep_check_activated(task))
 464                 return;
 465
 466         WARN_ON_ONCE(task->tk_timeout != 0);
 467         priority -= RPC_PRIORITY_LOW;
 468         /*
 469          * Protect the queue operations.
 470          */
 471         spin_lock(&q->lock);
 472         __rpc_sleep_on_priority(q, task, priority);
 473         spin_unlock(&q->lock);
 474 }
 475 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 476
 477 /**
 478  * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
 479  * @wq: workqueue on which to run task
 480  * @queue: wait queue
 481  * @task: task to be woken up
 482  *
 483  * Caller must hold queue->lock, and have cleared the task queued flag.
 484  */
 485 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
 486                 struct rpc_wait_queue *queue,
 487                 struct rpc_task *task)
 488 {
 489         /* Has the task been executed yet? If not, we cannot wake it up! */
 490         if (!RPC_IS_ACTIVATED(task)) {
 491                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 492                 return;
 493         }
 494
 495         trace_rpc_task_wakeup(task, queue);
 496
 497         __rpc_remove_wait_queue(queue, task);
 498
 499         rpc_make_runnable(wq, task);
 500 }
 501
 502 /*
 503  * Wake up a queued task while the queue lock is being held
 504  */
 505 static struct rpc_task *
 506 rpc_wake_up_task_on_wq_queue_action_locked(struct workqueue_struct *wq,
 507                 struct rpc_wait_queue *queue, struct rpc_task *task,
 508                 bool (*action)(struct rpc_task *, void *), void *data)
 509 {
 510         if (RPC_IS_QUEUED(task)) {
 511                 smp_rmb();
 512                 if (task->tk_waitqueue == queue) {
 513                         if (action == NULL || action(task, data)) {
 514                                 __rpc_do_wake_up_task_on_wq(wq, queue, task);
 515                                 return task;
 516                         }
 517                 }
 518         }
 519         return NULL;
 520 }
 521
 522 /*
 523  * Wake up a queued task while the queue lock is being held
 524  */
 525 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue,
 526                                           struct rpc_task *task)
 527 {
 528         rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
 529                                                    task, NULL, NULL);
 530 }
 531
 532 /*
 533  * Wake up a task on a specific queue
 534  */
 535 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 536 {
 537         if (!RPC_IS_QUEUED(task))
 538                 return;
 539         spin_lock(&queue->lock);
 540         rpc_wake_up_task_queue_locked(queue, task);
 541         spin_unlock(&queue->lock);
 542 }
 543 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 544
 545 static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
 546 {
 547         task->tk_status = *(int *)status;
 548         return true;
 549 }
 550
 551 static void
 552 rpc_wake_up_task_queue_set_status_locked(struct rpc_wait_queue *queue,
 553                 struct rpc_task *task, int status)
 554 {
 555         rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
 556                         task, rpc_task_action_set_status, &status);
 557 }
 558
 559 /**
 560  * rpc_wake_up_queued_task_set_status - wake up a task and set task->tk_status
 561  * @queue: pointer to rpc_wait_queue
 562  * @task: pointer to rpc_task
 563  * @status: integer error value
 564  *
 565  * If @task is queued on @queue, then it is woken up, and @task->tk_status is
 566  * set to the value of @status.
 567  */
 568 void
 569 rpc_wake_up_queued_task_set_status(struct rpc_wait_queue *queue,
 570                 struct rpc_task *task, int status)
 571 {
 572         if (!RPC_IS_QUEUED(task))
 573                 return;
 574         spin_lock(&queue->lock);
 575         rpc_wake_up_task_queue_set_status_locked(queue, task, status);
 576         spin_unlock(&queue->lock);
 577 }
 578
 579 /*
 580  * Wake up the next task on a priority queue.
 581  */
 582 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 583 {
 584         struct list_head *q;
 585         struct rpc_task *task;
 586
 587         /*
 588          * Service the privileged queue.
 589          */
 590         q = &queue->tasks[RPC_NR_PRIORITY - 1];
 591         if (queue->maxpriority > RPC_PRIORITY_PRIVILEGED && !list_empty(q)) {
 592                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 593                 goto out;
 594         }
 595
 596         /*
 597          * Service a batch of tasks from a single owner.
 598          */
 599         q = &queue->tasks[queue->priority];
 600         if (!list_empty(q) && queue->nr) {
 601                 queue->nr--;
 602                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 603                 goto out;
 604         }
 605
 606         /*
 607          * Service the next queue.
 608          */
 609         do {
 610                 if (q == &queue->tasks[0])
 611                         q = &queue->tasks[queue->maxpriority];
 612                 else
 613                         q = q - 1;
 614                 if (!list_empty(q)) {
 615                         task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 616                         goto new_queue;
 617                 }
 618         } while (q != &queue->tasks[queue->priority]);
 619
 620         rpc_reset_waitqueue_priority(queue);
 621         return NULL;
 622
 623 new_queue:
 624         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 625 out:
 626         return task;
 627 }
 628
 629 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 630 {
 631         if (RPC_IS_PRIORITY(queue))
 632                 return __rpc_find_next_queued_priority(queue);
 633         if (!list_empty(&queue->tasks[0]))
 634                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 635         return NULL;
 636 }
 637
 638 /*
 639  * Wake up the first task on the wait queue.
 640  */
 641 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 642                 struct rpc_wait_queue *queue,
 643                 bool (*func)(struct rpc_task *, void *), void *data)
 644 {
 645         struct rpc_task *task = NULL;
 646
 647         spin_lock(&queue->lock);
 648         task = __rpc_find_next_queued(queue);
 649         if (task != NULL)
 650                 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
 651                                 task, func, data);
 652         spin_unlock(&queue->lock);
 653
 654         return task;
 655 }
 656
 657 /*
 658  * Wake up the first task on the wait queue.
 659  */
 660 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 661                 bool (*func)(struct rpc_task *, void *), void *data)
 662 {
 663         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 664 }
 665 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 666
 667 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 668 {
 669         return true;
 670 }
 671
 672 /*
 673  * Wake up the next task on the wait queue.
 674 */
 675 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 676 {
 677         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 678 }
 679 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 680
 681 /**
 682  * rpc_wake_up_locked - wake up all rpc_tasks
 683  * @queue: rpc_wait_queue on which the tasks are sleeping
 684  *
 685  */
 686 static void rpc_wake_up_locked(struct rpc_wait_queue *queue)
 687 {
 688         struct rpc_task *task;
 689
 690         for (;;) {
 691                 task = __rpc_find_next_queued(queue);
 692                 if (task == NULL)
 693                         break;
 694                 rpc_wake_up_task_queue_locked(queue, task);
 695         }
 696 }
 697
 698 /**
 699  * rpc_wake_up - wake up all rpc_tasks
 700  * @queue: rpc_wait_queue on which the tasks are sleeping
 701  *
 702  * Grabs queue->lock
 703  */
 704 void rpc_wake_up(struct rpc_wait_queue *queue)
 705 {
 706         spin_lock(&queue->lock);
 707         rpc_wake_up_locked(queue);
 708         spin_unlock(&queue->lock);
 709 }
 710 EXPORT_SYMBOL_GPL(rpc_wake_up);
 711
 712 /**
 713  * rpc_wake_up_status_locked - wake up all rpc_tasks and set their status value.
 714  * @queue: rpc_wait_queue on which the tasks are sleeping
 715  * @status: status value to set
 716  */
 717 static void rpc_wake_up_status_locked(struct rpc_wait_queue *queue, int status)
 718 {
 719         struct rpc_task *task;
 720
 721         for (;;) {
 722                 task = __rpc_find_next_queued(queue);
 723                 if (task == NULL)
 724                         break;
 725                 rpc_wake_up_task_queue_set_status_locked(queue, task, status);
 726         }
 727 }
 728
 729 /**
 730  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 731  * @queue: rpc_wait_queue on which the tasks are sleeping
 732  * @status: status value to set
 733  *
 734  * Grabs queue->lock
 735  */
 736 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 737 {
 738         spin_lock(&queue->lock);
 739         rpc_wake_up_status_locked(queue, status);
 740         spin_unlock(&queue->lock);
 741 }
 742 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 743
 744 static void __rpc_queue_timer_fn(struct work_struct *work)
 745 {
 746         struct rpc_wait_queue *queue = container_of(work,
 747                         struct rpc_wait_queue,
 748                         timer_list.dwork.work);
 749         struct rpc_task *task, *n;
 750         unsigned long expires, now, timeo;
 751
 752         spin_lock(&queue->lock);
 753         expires = now = jiffies;
 754         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 755                 timeo = task->tk_timeout;
 756                 if (time_after_eq(now, timeo)) {
 757                         trace_rpc_task_timeout(task, task->tk_action);
 758                         task->tk_status = -ETIMEDOUT;
 759                         rpc_wake_up_task_queue_locked(queue, task);
 760                         continue;
 761                 }
 762                 if (expires == now || time_after(expires, timeo))
 763                         expires = timeo;
 764         }
 765         if (!list_empty(&queue->timer_list.list))
 766                 rpc_set_queue_timer(queue, expires);
 767         spin_unlock(&queue->lock);
 768 }
 769
 770 static void __rpc_atrun(struct rpc_task *task)
 771 {
 772         if (task->tk_status == -ETIMEDOUT)
 773                 task->tk_status = 0;
 774 }
 775
 776 /*
 777  * Run a task at a later time
 778  */
 779 void rpc_delay(struct rpc_task *task, unsigned long delay)
 780 {
 781         rpc_sleep_on_timeout(&delay_queue, task, __rpc_atrun, jiffies + delay);
 782 }
 783 EXPORT_SYMBOL_GPL(rpc_delay);
 784
 785 /*
 786  * Helper to call task->tk_ops->rpc_call_prepare
 787  */
 788 void rpc_prepare_task(struct rpc_task *task)
 789 {
 790         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 791 }
 792
 793 static void
 794 rpc_init_task_statistics(struct rpc_task *task)
 795 {
 796         /* Initialize retry counters */
 797         task->tk_garb_retry = 2;
 798         task->tk_cred_retry = 2;
 799
 800         /* starting timestamp */
 801         task->tk_start = ktime_get();
 802 }
 803
 804 static void
 805 rpc_reset_task_statistics(struct rpc_task *task)
 806 {
 807         task->tk_timeouts = 0;
 808         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_SENT);
 809         rpc_init_task_statistics(task);
 810 }
 811
 812 /*
 813  * Helper that calls task->tk_ops->rpc_call_done if it exists
 814  */
 815 void rpc_exit_task(struct rpc_task *task)
 816 {
 817         trace_rpc_task_end(task, task->tk_action);
 818         task->tk_action = NULL;
 819         if (task->tk_ops->rpc_count_stats)
 820                 task->tk_ops->rpc_count_stats(task, task->tk_calldata);
 821         else if (task->tk_client)
 822                 rpc_count_iostats(task, task->tk_client->cl_metrics);
 823         if (task->tk_ops->rpc_call_done != NULL) {
 824                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 825                 if (task->tk_action != NULL) {
 826                         /* Always release the RPC slot and buffer memory */
 827                         xprt_release(task);
 828                         rpc_reset_task_statistics(task);
 829                 }
 830         }
 831 }
 832
 833 void rpc_signal_task(struct rpc_task *task)
 834 {
 835         struct rpc_wait_queue *queue;
 836
 837         if (!RPC_IS_ACTIVATED(task))
 838                 return;
 839
 840         trace_rpc_task_signalled(task, task->tk_action);
 841         set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
 842         smp_mb__after_atomic();
 843         queue = READ_ONCE(task->tk_waitqueue);
 844         if (queue)
 845                 rpc_wake_up_queued_task_set_status(queue, task, -ERESTARTSYS);
 846 }
 847
 848 void rpc_exit(struct rpc_task *task, int status)
 849 {
 850         task->tk_status = status;
 851         task->tk_action = rpc_exit_task;
 852         rpc_wake_up_queued_task(task->tk_waitqueue, task);
 853 }
 854 EXPORT_SYMBOL_GPL(rpc_exit);
 855
 856 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 857 {
 858         if (ops->rpc_release != NULL)
 859                 ops->rpc_release(calldata);
 860 }
 861
 862 /*
 863  * This is the RPC `scheduler' (or rather, the finite state machine).
 864  */
 865 static void __rpc_execute(struct rpc_task *task)
 866 {
 867         struct rpc_wait_queue *queue;
 868         int task_is_async = RPC_IS_ASYNC(task);
 869         int status = 0;
 870
 871         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 872         if (RPC_IS_QUEUED(task))
 873                 return;
 874
 875         for (;;) {
 876                 void (*do_action)(struct rpc_task *);
 877
 878                 /*
 879                  * Perform the next FSM step or a pending callback.
 880                  *
 881                  * tk_action may be NULL if the task has been killed.
 882                  * In particular, note that rpc_killall_tasks may
 883                  * do this at any time, so beware when dereferencing.
 884                  */
 885                 do_action = task->tk_action;
 886                 if (task->tk_callback) {
 887                         do_action = task->tk_callback;
 888                         task->tk_callback = NULL;
 889                 }
 890                 if (!do_action)
 891                         break;
 892                 trace_rpc_task_run_action(task, do_action);
 893                 do_action(task);
 894
 895                 /*
 896                  * Lockless check for whether task is sleeping or not.
 897                  */
 898                 if (!RPC_IS_QUEUED(task))
 899                         continue;
 900
 901                 /*
 902                  * Signalled tasks should exit rather than sleep.
 903                  */
 904                 if (RPC_SIGNALLED(task)) {
 905                         task->tk_rpc_status = -ERESTARTSYS;
 906                         rpc_exit(task, -ERESTARTSYS);
 907                 }
 908
 909                 /*
 910                  * The queue->lock protects against races with
 911                  * rpc_make_runnable().
 912                  *
 913                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 914                  * rpc_task, rpc_make_runnable() can assign it to a
 915                  * different workqueue. We therefore cannot assume that the
 916                  * rpc_task pointer may still be dereferenced.
 917                  */
 918                 queue = task->tk_waitqueue;
 919                 spin_lock(&queue->lock);
 920                 if (!RPC_IS_QUEUED(task)) {
 921                         spin_unlock(&queue->lock);
 922                         continue;
 923                 }
 924                 rpc_clear_running(task);
 925                 spin_unlock(&queue->lock);
 926                 if (task_is_async)
 927                         return;
 928
 929                 /* sync task: sleep here */
 930                 trace_rpc_task_sync_sleep(task, task->tk_action);
 931                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 932                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 933                                 TASK_KILLABLE);
 934                 if (status < 0) {
 935                         /*
 936                          * When a sync task receives a signal, it exits with
 937                          * -ERESTARTSYS. In order to catch any callbacks that
 938                          * clean up after sleeping on some queue, we don't
 939                          * break the loop here, but go around once more.
 940                          */
 941                         trace_rpc_task_signalled(task, task->tk_action);
 942                         set_bit(RPC_TASK_SIGNALLED, &task->tk_runstate);
 943                         task->tk_rpc_status = -ERESTARTSYS;
 944                         rpc_exit(task, -ERESTARTSYS);
 945                 }
 946                 trace_rpc_task_sync_wake(task, task->tk_action);
 947         }
 948
 949         /* Release all resources associated with the task */
 950         rpc_release_task(task);
 951 }
 952
 953 /*
 954  * User-visible entry point to the scheduler.
 955  *
 956  * This may be called recursively if e.g. an async NFS task updates
 957  * the attributes and finds that dirty pages must be flushed.
 958  * NOTE: Upon exit of this function the task is guaranteed to be
 959  *       released. In particular note that tk_release() will have
 960  *       been called, so your task memory may have been freed.
 961  */
 962 void rpc_execute(struct rpc_task *task)
 963 {
 964         bool is_async = RPC_IS_ASYNC(task);
 965
 966         rpc_set_active(task);
 967         rpc_make_runnable(rpciod_workqueue, task);
 968         if (!is_async) {
 969                 unsigned int pflags = memalloc_nofs_save();
 970                 __rpc_execute(task);
 971                 memalloc_nofs_restore(pflags);
 972         }
 973 }
 974
 975 static void rpc_async_schedule(struct work_struct *work)
 976 {
 977         unsigned int pflags = memalloc_nofs_save();
 978
 979         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 980         memalloc_nofs_restore(pflags);
 981 }
 982
 983 /**
 984  * rpc_malloc - allocate RPC buffer resources
 985  * @task: RPC task
 986  *
 987  * A single memory region is allocated, which is split between the
 988  * RPC call and RPC reply that this task is being used for. When
 989  * this RPC is retired, the memory is released by calling rpc_free.
 990  *
 991  * To prevent rpciod from hanging, this allocator never sleeps,
 992  * returning -ENOMEM and suppressing warning if the request cannot
 993  * be serviced immediately. The caller can arrange to sleep in a
 994  * way that is safe for rpciod.
 995  *
 996  * Most requests are 'small' (under 2KiB) and can be serviced from a
 997  * mempool, ensuring that NFS reads and writes can always proceed,
 998  * and that there is good locality of reference for these buffers.
 999  */
1000 int rpc_malloc(struct rpc_task *task)
1001 {
1002         struct rpc_rqst *rqst = task->tk_rqstp;
1003         size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
1004         struct rpc_buffer *buf;
1005         gfp_t gfp = GFP_NOFS;
1006
1007         if (RPC_IS_ASYNC(task))
1008                 gfp = GFP_NOWAIT | __GFP_NOWARN;
1009         if (RPC_IS_SWAPPER(task))
1010                 gfp |= __GFP_MEMALLOC;
1011
1012         size += sizeof(struct rpc_buffer);
1013         if (size <= RPC_BUFFER_MAXSIZE)
1014                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
1015         else
1016                 buf = kmalloc(size, gfp);
1017
1018         if (!buf)
1019                 return -ENOMEM;
1020
1021         buf->len = size;
1022         rqst->rq_buffer = buf->data;
1023         rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
1024         return 0;
1025 }
1026 EXPORT_SYMBOL_GPL(rpc_malloc);
1027
1028 /**
1029  * rpc_free - free RPC buffer resources allocated via rpc_malloc
1030  * @task: RPC task
1031  *
1032  */
1033 void rpc_free(struct rpc_task *task)
1034 {
1035         void *buffer = task->tk_rqstp->rq_buffer;
1036         size_t size;
1037         struct rpc_buffer *buf;
1038
1039         buf = container_of(buffer, struct rpc_buffer, data);
1040         size = buf->len;
1041
1042         if (size <= RPC_BUFFER_MAXSIZE)
1043                 mempool_free(buf, rpc_buffer_mempool);
1044         else
1045                 kfree(buf);
1046 }
1047 EXPORT_SYMBOL_GPL(rpc_free);
1048
1049 /*
1050  * Creation and deletion of RPC task structures
1051  */
1052 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
1053 {
1054         memset(task, 0, sizeof(*task));
1055         atomic_set(&task->tk_count, 1);
1056         task->tk_flags  = task_setup_data->flags;
1057         task->tk_ops = task_setup_data->callback_ops;
1058         task->tk_calldata = task_setup_data->callback_data;
1059         INIT_LIST_HEAD(&task->tk_task);
1060
1061         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
1062         task->tk_owner = current->tgid;
1063
1064         /* Initialize workqueue for async tasks */
1065         task->tk_workqueue = task_setup_data->workqueue;
1066
1067         task->tk_xprt = rpc_task_get_xprt(task_setup_data->rpc_client,
1068                         xprt_get(task_setup_data->rpc_xprt));
1069
1070         task->tk_op_cred = get_rpccred(task_setup_data->rpc_op_cred);
1071
1072         if (task->tk_ops->rpc_call_prepare != NULL)
1073                 task->tk_action = rpc_prepare_task;
1074
1075         rpc_init_task_statistics(task);
1076 }
1077
1078 static struct rpc_task *
1079 rpc_alloc_task(void)
1080 {
1081         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOFS);
1082 }
1083
1084 /*
1085  * Create a new task for the specified client.
1086  */
1087 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
1088 {
1089         struct rpc_task *task = setup_data->task;
1090         unsigned short flags = 0;
1091
1092         if (task == NULL) {
1093                 task = rpc_alloc_task();
1094                 flags = RPC_TASK_DYNAMIC;
1095         }
1096
1097         rpc_init_task(task, setup_data);
1098         task->tk_flags |= flags;
1099         return task;
1100 }
1101
1102 /*
1103  * rpc_free_task - release rpc task and perform cleanups
1104  *
1105  * Note that we free up the rpc_task _after_ rpc_release_calldata()
1106  * in order to work around a workqueue dependency issue.
1107  *
1108  * Tejun Heo states:
1109  * "Workqueue currently considers two work items to be the same if they're
1110  * on the same address and won't execute them concurrently - ie. it
1111  * makes a work item which is queued again while being executed wait
1112  * for the previous execution to complete.
1113  *
1114  * If a work function frees the work item, and then waits for an event
1115  * which should be performed by another work item and *that* work item
1116  * recycles the freed work item, it can create a false dependency loop.
1117  * There really is no reliable way to detect this short of verifying
1118  * every memory free."
1119  *
1120  */
1121 static void rpc_free_task(struct rpc_task *task)
1122 {
1123         unsigned short tk_flags = task->tk_flags;
1124
1125         put_rpccred(task->tk_op_cred);
1126         rpc_release_calldata(task->tk_ops, task->tk_calldata);
1127
1128         if (tk_flags & RPC_TASK_DYNAMIC)
1129                 mempool_free(task, rpc_task_mempool);
1130 }
1131
1132 static void rpc_async_release(struct work_struct *work)
1133 {
1134         unsigned int pflags = memalloc_nofs_save();
1135
1136         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1137         memalloc_nofs_restore(pflags);
1138 }
1139
1140 static void rpc_release_resources_task(struct rpc_task *task)
1141 {
1142         xprt_release(task);
1143         if (task->tk_msg.rpc_cred) {
1144                 if (!(task->tk_flags & RPC_TASK_CRED_NOREF))
1145                         put_cred(task->tk_msg.rpc_cred);
1146                 task->tk_msg.rpc_cred = NULL;
1147         }
1148         rpc_task_release_client(task);
1149 }
1150
1151 static void rpc_final_put_task(struct rpc_task *task,
1152                 struct workqueue_struct *q)
1153 {
1154         if (q != NULL) {
1155                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1156                 queue_work(q, &task->u.tk_work);
1157         } else
1158                 rpc_free_task(task);
1159 }
1160
1161 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1162 {
1163         if (atomic_dec_and_test(&task->tk_count)) {
1164                 rpc_release_resources_task(task);
1165                 rpc_final_put_task(task, q);
1166         }
1167 }
1168
1169 void rpc_put_task(struct rpc_task *task)
1170 {
1171         rpc_do_put_task(task, NULL);
1172 }
1173 EXPORT_SYMBOL_GPL(rpc_put_task);
1174
1175 void rpc_put_task_async(struct rpc_task *task)
1176 {
1177         rpc_do_put_task(task, task->tk_workqueue);
1178 }
1179 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1180
1181 static void rpc_release_task(struct rpc_task *task)
1182 {
1183         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1184
1185         rpc_release_resources_task(task);
1186
1187         /*
1188          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1189          * so it should be safe to use task->tk_count as a test for whether
1190          * or not any other processes still hold references to our rpc_task.
1191          */
1192         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1193                 /* Wake up anyone who may be waiting for task completion */
1194                 if (!rpc_complete_task(task))
1195                         return;
1196         } else {
1197                 if (!atomic_dec_and_test(&task->tk_count))
1198                         return;
1199         }
1200         rpc_final_put_task(task, task->tk_workqueue);
1201 }
1202
1203 int rpciod_up(void)
1204 {
1205         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1206 }
1207
1208 void rpciod_down(void)
1209 {
1210         module_put(THIS_MODULE);
1211 }
1212
1213 /*
1214  * Start up the rpciod workqueue.
1215  */
1216 static int rpciod_start(void)
1217 {
1218         struct workqueue_struct *wq;
1219
1220         /*
1221          * Create the rpciod thread and wait for it to start.
1222          */
1223         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1224         if (!wq)
1225                 goto out_failed;
1226         rpciod_workqueue = wq;
1227         /* Note: highpri because network receive is latency sensitive */
1228         wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1229         if (!wq)
1230                 goto free_rpciod;
1231         xprtiod_workqueue = wq;
1232         return 1;
1233 free_rpciod:
1234         wq = rpciod_workqueue;
1235         rpciod_workqueue = NULL;
1236         destroy_workqueue(wq);
1237 out_failed:
1238         return 0;
1239 }
1240
1241 static void rpciod_stop(void)
1242 {
1243         struct workqueue_struct *wq = NULL;
1244
1245         if (rpciod_workqueue == NULL)
1246                 return;
1247
1248         wq = rpciod_workqueue;
1249         rpciod_workqueue = NULL;
1250         destroy_workqueue(wq);
1251         wq = xprtiod_workqueue;
1252         xprtiod_workqueue = NULL;
1253         destroy_workqueue(wq);
1254 }
1255
1256 void
1257 rpc_destroy_mempool(void)
1258 {
1259         rpciod_stop();
1260         mempool_destroy(rpc_buffer_mempool);
1261         mempool_destroy(rpc_task_mempool);
1262         kmem_cache_destroy(rpc_task_slabp);
1263         kmem_cache_destroy(rpc_buffer_slabp);
1264         rpc_destroy_wait_queue(&delay_queue);
1265 }
1266
1267 int
1268 rpc_init_mempool(void)
1269 {
1270         /*
1271          * The following is not strictly a mempool initialisation,
1272          * but there is no harm in doing it here
1273          */
1274         rpc_init_wait_queue(&delay_queue, "delayq");
1275         if (!rpciod_start())
1276                 goto err_nomem;
1277
1278         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1279                                              sizeof(struct rpc_task),
1280                                              0, SLAB_HWCACHE_ALIGN,
1281                                              NULL);
1282         if (!rpc_task_slabp)
1283                 goto err_nomem;
1284         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1285                                              RPC_BUFFER_MAXSIZE,
1286                                              0, SLAB_HWCACHE_ALIGN,
1287                                              NULL);
1288         if (!rpc_buffer_slabp)
1289                 goto err_nomem;
1290         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1291                                                     rpc_task_slabp);
1292         if (!rpc_task_mempool)
1293                 goto err_nomem;
1294         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1295                                                       rpc_buffer_slabp);
1296         if (!rpc_buffer_mempool)
1297                 goto err_nomem;
1298         return 0;
1299 err_nomem:
1300         rpc_destroy_mempool();
1301         return -ENOMEM;
1302 }