net/sunrpc/sched.c

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