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, <[email protected]>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <[email protected]>
  10  */
  11
  12 #include <linux/module.h>
  13
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/slab.h>
  17 #include <linux/mempool.h>
  18 #include <linux/smp.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/mutex.h>
  21 #include <linux/freezer.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #include "sunrpc.h"
  26
  27 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  28 #define RPCDBG_FACILITY         RPCDBG_SCHED
  29 #endif
  30
  31 #define CREATE_TRACE_POINTS
  32 #include <trace/events/sunrpc.h>
  33
  34 /*
  35  * RPC slabs and memory pools
  36  */
  37 #define RPC_BUFFER_MAXSIZE      (2048)
  38 #define RPC_BUFFER_POOLSIZE     (8)
  39 #define RPC_TASK_POOLSIZE       (8)
  40 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  41 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  42 static mempool_t        *rpc_task_mempool __read_mostly;
  43 static mempool_t        *rpc_buffer_mempool __read_mostly;
  44
  45 static void                     rpc_async_schedule(struct work_struct *);
  46 static void                      rpc_release_task(struct rpc_task *task);
  47 static void __rpc_queue_timer_fn(struct timer_list *t);
  48
  49 /*
  50  * RPC tasks sit here while waiting for conditions to improve.
  51  */
  52 static struct rpc_wait_queue delay_queue;
  53
  54 /*
  55  * rpciod-related stuff
  56  */
  57 struct workqueue_struct *rpciod_workqueue __read_mostly;
  58 struct workqueue_struct *xprtiod_workqueue __read_mostly;
  59
  60 /*
  61  * Disable the timer for a given RPC task. Should be called with
  62  * queue->lock and bh_disabled in order to avoid races within
  63  * rpc_run_timer().
  64  */
  65 static void
  66 __rpc_disable_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  67 {
  68         if (task->tk_timeout == 0)
  69                 return;
  70         dprintk("RPC: %5u disabling timer\n", task->tk_pid);
  71         task->tk_timeout = 0;
  72         list_del(&task->u.tk_wait.timer_list);
  73         if (list_empty(&queue->timer_list.list))
  74                 del_timer(&queue->timer_list.timer);
  75 }
  76
  77 static void
  78 rpc_set_queue_timer(struct rpc_wait_queue *queue, unsigned long expires)
  79 {
  80         queue->timer_list.expires = expires;
  81         mod_timer(&queue->timer_list.timer, expires);
  82 }
  83
  84 /*
  85  * Set up a timer for the current task.
  86  */
  87 static void
  88 __rpc_add_timer(struct rpc_wait_queue *queue, struct rpc_task *task)
  89 {
  90         if (!task->tk_timeout)
  91                 return;
  92
  93         dprintk("RPC: %5u setting alarm for %u ms\n",
  94                 task->tk_pid, jiffies_to_msecs(task->tk_timeout));
  95
  96         task->u.tk_wait.expires = jiffies + task->tk_timeout;
  97         if (list_empty(&queue->timer_list.list) || time_before(task->u.tk_wait.expires, queue->timer_list.expires))
  98                 rpc_set_queue_timer(queue, task->u.tk_wait.expires);
  99         list_add(&task->u.tk_wait.timer_list, &queue->timer_list.list);
 100 }
 101
 102 static void rpc_set_waitqueue_priority(struct rpc_wait_queue *queue, int priority)
 103 {
 104         if (queue->priority != priority) {
 105                 queue->priority = priority;
 106                 queue->nr = 1U << priority;
 107         }
 108 }
 109
 110 static void rpc_reset_waitqueue_priority(struct rpc_wait_queue *queue)
 111 {
 112         rpc_set_waitqueue_priority(queue, queue->maxpriority);
 113 }
 114
 115 /*
 116  * Add a request to a queue list
 117  */
 118 static void
 119 __rpc_list_enqueue_task(struct list_head *q, struct rpc_task *task)
 120 {
 121         struct rpc_task *t;
 122
 123         list_for_each_entry(t, q, u.tk_wait.list) {
 124                 if (t->tk_owner == task->tk_owner) {
 125                         list_add_tail(&task->u.tk_wait.links,
 126                                         &t->u.tk_wait.links);
 127                         /* Cache the queue head in task->u.tk_wait.list */
 128                         task->u.tk_wait.list.next = q;
 129                         task->u.tk_wait.list.prev = NULL;
 130                         return;
 131                 }
 132         }
 133         INIT_LIST_HEAD(&task->u.tk_wait.links);
 134         list_add_tail(&task->u.tk_wait.list, q);
 135 }
 136
 137 /*
 138  * Remove request from a queue list
 139  */
 140 static void
 141 __rpc_list_dequeue_task(struct rpc_task *task)
 142 {
 143         struct list_head *q;
 144         struct rpc_task *t;
 145
 146         if (task->u.tk_wait.list.prev == NULL) {
 147                 list_del(&task->u.tk_wait.links);
 148                 return;
 149         }
 150         if (!list_empty(&task->u.tk_wait.links)) {
 151                 t = list_first_entry(&task->u.tk_wait.links,
 152                                 struct rpc_task,
 153                                 u.tk_wait.links);
 154                 /* Assume __rpc_list_enqueue_task() cached the queue head */
 155                 q = t->u.tk_wait.list.next;
 156                 list_add_tail(&t->u.tk_wait.list, q);
 157                 list_del(&task->u.tk_wait.links);
 158         }
 159         list_del(&task->u.tk_wait.list);
 160 }
 161
 162 /*
 163  * Add new request to a priority queue.
 164  */
 165 static void __rpc_add_wait_queue_priority(struct rpc_wait_queue *queue,
 166                 struct rpc_task *task,
 167                 unsigned char queue_priority)
 168 {
 169         if (unlikely(queue_priority > queue->maxpriority))
 170                 queue_priority = queue->maxpriority;
 171         __rpc_list_enqueue_task(&queue->tasks[queue_priority], task);
 172 }
 173
 174 /*
 175  * Add new request to wait queue.
 176  *
 177  * Swapper tasks always get inserted at the head of the queue.
 178  * This should avoid many nasty memory deadlocks and hopefully
 179  * improve overall performance.
 180  * Everyone else gets appended to the queue to ensure proper FIFO behavior.
 181  */
 182 static void __rpc_add_wait_queue(struct rpc_wait_queue *queue,
 183                 struct rpc_task *task,
 184                 unsigned char queue_priority)
 185 {
 186         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 187         if (RPC_IS_QUEUED(task))
 188                 return;
 189
 190         if (RPC_IS_PRIORITY(queue))
 191                 __rpc_add_wait_queue_priority(queue, task, queue_priority);
 192         else if (RPC_IS_SWAPPER(task))
 193                 list_add(&task->u.tk_wait.list, &queue->tasks[0]);
 194         else
 195                 list_add_tail(&task->u.tk_wait.list, &queue->tasks[0]);
 196         task->tk_waitqueue = queue;
 197         queue->qlen++;
 198         /* barrier matches the read in rpc_wake_up_task_queue_locked() */
 199         smp_wmb();
 200         rpc_set_queued(task);
 201
 202         dprintk("RPC: %5u added to queue %p \"%s\"\n",
 203                         task->tk_pid, queue, rpc_qname(queue));
 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         dprintk("RPC: %5u removed from queue %p \"%s\"\n",
 227                         task->tk_pid, queue, rpc_qname(queue));
 228 }
 229
 230 static void __rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname, unsigned char nr_queues)
 231 {
 232         int i;
 233
 234         spin_lock_init(&queue->lock);
 235         for (i = 0; i < ARRAY_SIZE(queue->tasks); i++)
 236                 INIT_LIST_HEAD(&queue->tasks[i]);
 237         queue->maxpriority = nr_queues - 1;
 238         rpc_reset_waitqueue_priority(queue);
 239         queue->qlen = 0;
 240         timer_setup(&queue->timer_list.timer, __rpc_queue_timer_fn, 0);
 241         INIT_LIST_HEAD(&queue->timer_list.list);
 242         rpc_assign_waitqueue_name(queue, qname);
 243 }
 244
 245 void rpc_init_priority_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 246 {
 247         __rpc_init_priority_wait_queue(queue, qname, RPC_NR_PRIORITY);
 248 }
 249 EXPORT_SYMBOL_GPL(rpc_init_priority_wait_queue);
 250
 251 void rpc_init_wait_queue(struct rpc_wait_queue *queue, const char *qname)
 252 {
 253         __rpc_init_priority_wait_queue(queue, qname, 1);
 254 }
 255 EXPORT_SYMBOL_GPL(rpc_init_wait_queue);
 256
 257 void rpc_destroy_wait_queue(struct rpc_wait_queue *queue)
 258 {
 259         del_timer_sync(&queue->timer_list.timer);
 260 }
 261 EXPORT_SYMBOL_GPL(rpc_destroy_wait_queue);
 262
 263 static int rpc_wait_bit_killable(struct wait_bit_key *key, int mode)
 264 {
 265         freezable_schedule_unsafe();
 266         if (signal_pending_state(mode, current))
 267                 return -ERESTARTSYS;
 268         return 0;
 269 }
 270
 271 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG) || IS_ENABLED(CONFIG_TRACEPOINTS)
 272 static void rpc_task_set_debuginfo(struct rpc_task *task)
 273 {
 274         static atomic_t rpc_pid;
 275
 276         task->tk_pid = atomic_inc_return(&rpc_pid);
 277 }
 278 #else
 279 static inline void rpc_task_set_debuginfo(struct rpc_task *task)
 280 {
 281 }
 282 #endif
 283
 284 static void rpc_set_active(struct rpc_task *task)
 285 {
 286         rpc_task_set_debuginfo(task);
 287         set_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 288         trace_rpc_task_begin(task, NULL);
 289 }
 290
 291 /*
 292  * Mark an RPC call as having completed by clearing the 'active' bit
 293  * and then waking up all tasks that were sleeping.
 294  */
 295 static int rpc_complete_task(struct rpc_task *task)
 296 {
 297         void *m = &task->tk_runstate;
 298         wait_queue_head_t *wq = bit_waitqueue(m, RPC_TASK_ACTIVE);
 299         struct wait_bit_key k = __WAIT_BIT_KEY_INITIALIZER(m, RPC_TASK_ACTIVE);
 300         unsigned long flags;
 301         int ret;
 302
 303         trace_rpc_task_complete(task, NULL);
 304
 305         spin_lock_irqsave(&wq->lock, flags);
 306         clear_bit(RPC_TASK_ACTIVE, &task->tk_runstate);
 307         ret = atomic_dec_and_test(&task->tk_count);
 308         if (waitqueue_active(wq))
 309                 __wake_up_locked_key(wq, TASK_NORMAL, &k);
 310         spin_unlock_irqrestore(&wq->lock, flags);
 311         return ret;
 312 }
 313
 314 /*
 315  * Allow callers to wait for completion of an RPC call
 316  *
 317  * Note the use of out_of_line_wait_on_bit() rather than wait_on_bit()
 318  * to enforce taking of the wq->lock and hence avoid races with
 319  * rpc_complete_task().
 320  */
 321 int __rpc_wait_for_completion_task(struct rpc_task *task, wait_bit_action_f *action)
 322 {
 323         if (action == NULL)
 324                 action = rpc_wait_bit_killable;
 325         return out_of_line_wait_on_bit(&task->tk_runstate, RPC_TASK_ACTIVE,
 326                         action, TASK_KILLABLE);
 327 }
 328 EXPORT_SYMBOL_GPL(__rpc_wait_for_completion_task);
 329
 330 /*
 331  * Make an RPC task runnable.
 332  *
 333  * Note: If the task is ASYNC, and is being made runnable after sitting on an
 334  * rpc_wait_queue, this must be called with the queue spinlock held to protect
 335  * the wait queue operation.
 336  * Note the ordering of rpc_test_and_set_running() and rpc_clear_queued(),
 337  * which is needed to ensure that __rpc_execute() doesn't loop (due to the
 338  * lockless RPC_IS_QUEUED() test) before we've had a chance to test
 339  * the RPC_TASK_RUNNING flag.
 340  */
 341 static void rpc_make_runnable(struct workqueue_struct *wq,
 342                 struct rpc_task *task)
 343 {
 344         bool need_wakeup = !rpc_test_and_set_running(task);
 345
 346         rpc_clear_queued(task);
 347         if (!need_wakeup)
 348                 return;
 349         if (RPC_IS_ASYNC(task)) {
 350                 INIT_WORK(&task->u.tk_work, rpc_async_schedule);
 351                 queue_work(wq, &task->u.tk_work);
 352         } else
 353                 wake_up_bit(&task->tk_runstate, RPC_TASK_QUEUED);
 354 }
 355
 356 /*
 357  * Prepare for sleeping on a wait queue.
 358  * By always appending tasks to the list we ensure FIFO behavior.
 359  * NB: An RPC task will only receive interrupt-driven events as long
 360  * as it's on a wait queue.
 361  */
 362 static void __rpc_sleep_on_priority(struct rpc_wait_queue *q,
 363                 struct rpc_task *task,
 364                 rpc_action action,
 365                 unsigned char queue_priority)
 366 {
 367         dprintk("RPC: %5u sleep_on(queue \"%s\" time %lu)\n",
 368                         task->tk_pid, rpc_qname(q), jiffies);
 369
 370         trace_rpc_task_sleep(task, q);
 371
 372         __rpc_add_wait_queue(q, task, queue_priority);
 373
 374         WARN_ON_ONCE(task->tk_callback != NULL);
 375         task->tk_callback = action;
 376         __rpc_add_timer(q, task);
 377 }
 378
 379 void rpc_sleep_on(struct rpc_wait_queue *q, struct rpc_task *task,
 380                                 rpc_action action)
 381 {
 382         /* We shouldn't ever put an inactive task to sleep */
 383         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 384         if (!RPC_IS_ACTIVATED(task)) {
 385                 task->tk_status = -EIO;
 386                 rpc_put_task_async(task);
 387                 return;
 388         }
 389
 390         /*
 391          * Protect the queue operations.
 392          */
 393         spin_lock_bh(&q->lock);
 394         __rpc_sleep_on_priority(q, task, action, task->tk_priority);
 395         spin_unlock_bh(&q->lock);
 396 }
 397 EXPORT_SYMBOL_GPL(rpc_sleep_on);
 398
 399 void rpc_sleep_on_priority(struct rpc_wait_queue *q, struct rpc_task *task,
 400                 rpc_action action, int priority)
 401 {
 402         /* We shouldn't ever put an inactive task to sleep */
 403         WARN_ON_ONCE(!RPC_IS_ACTIVATED(task));
 404         if (!RPC_IS_ACTIVATED(task)) {
 405                 task->tk_status = -EIO;
 406                 rpc_put_task_async(task);
 407                 return;
 408         }
 409
 410         /*
 411          * Protect the queue operations.
 412          */
 413         spin_lock_bh(&q->lock);
 414         __rpc_sleep_on_priority(q, task, action, priority - RPC_PRIORITY_LOW);
 415         spin_unlock_bh(&q->lock);
 416 }
 417 EXPORT_SYMBOL_GPL(rpc_sleep_on_priority);
 418
 419 /**
 420  * __rpc_do_wake_up_task_on_wq - wake up a single rpc_task
 421  * @wq: workqueue on which to run task
 422  * @queue: wait queue
 423  * @task: task to be woken up
 424  *
 425  * Caller must hold queue->lock, and have cleared the task queued flag.
 426  */
 427 static void __rpc_do_wake_up_task_on_wq(struct workqueue_struct *wq,
 428                 struct rpc_wait_queue *queue,
 429                 struct rpc_task *task)
 430 {
 431         dprintk("RPC: %5u __rpc_wake_up_task (now %lu)\n",
 432                         task->tk_pid, jiffies);
 433
 434         /* Has the task been executed yet? If not, we cannot wake it up! */
 435         if (!RPC_IS_ACTIVATED(task)) {
 436                 printk(KERN_ERR "RPC: Inactive task (%p) being woken up!\n", task);
 437                 return;
 438         }
 439
 440         trace_rpc_task_wakeup(task, queue);
 441
 442         __rpc_remove_wait_queue(queue, task);
 443
 444         rpc_make_runnable(wq, task);
 445
 446         dprintk("RPC:       __rpc_wake_up_task done\n");
 447 }
 448
 449 /*
 450  * Wake up a queued task while the queue lock is being held
 451  */
 452 static struct rpc_task *
 453 rpc_wake_up_task_on_wq_queue_action_locked(struct workqueue_struct *wq,
 454                 struct rpc_wait_queue *queue, struct rpc_task *task,
 455                 bool (*action)(struct rpc_task *, void *), void *data)
 456 {
 457         if (RPC_IS_QUEUED(task)) {
 458                 smp_rmb();
 459                 if (task->tk_waitqueue == queue) {
 460                         if (action == NULL || action(task, data)) {
 461                                 __rpc_do_wake_up_task_on_wq(wq, queue, task);
 462                                 return task;
 463                         }
 464                 }
 465         }
 466         return NULL;
 467 }
 468
 469 static void
 470 rpc_wake_up_task_on_wq_queue_locked(struct workqueue_struct *wq,
 471                 struct rpc_wait_queue *queue, struct rpc_task *task)
 472 {
 473         rpc_wake_up_task_on_wq_queue_action_locked(wq, queue, task, NULL, NULL);
 474 }
 475
 476 /*
 477  * Wake up a queued task while the queue lock is being held
 478  */
 479 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 480 {
 481         rpc_wake_up_task_on_wq_queue_locked(rpciod_workqueue, queue, task);
 482 }
 483
 484 /*
 485  * Wake up a task on a specific queue
 486  */
 487 void rpc_wake_up_queued_task_on_wq(struct workqueue_struct *wq,
 488                 struct rpc_wait_queue *queue,
 489                 struct rpc_task *task)
 490 {
 491         if (!RPC_IS_QUEUED(task))
 492                 return;
 493         spin_lock_bh(&queue->lock);
 494         rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
 495         spin_unlock_bh(&queue->lock);
 496 }
 497
 498 /*
 499  * Wake up a task on a specific queue
 500  */
 501 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 502 {
 503         if (!RPC_IS_QUEUED(task))
 504                 return;
 505         spin_lock_bh(&queue->lock);
 506         rpc_wake_up_task_queue_locked(queue, task);
 507         spin_unlock_bh(&queue->lock);
 508 }
 509 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 510
 511 static bool rpc_task_action_set_status(struct rpc_task *task, void *status)
 512 {
 513         task->tk_status = *(int *)status;
 514         return true;
 515 }
 516
 517 static void
 518 rpc_wake_up_task_queue_set_status_locked(struct rpc_wait_queue *queue,
 519                 struct rpc_task *task, int status)
 520 {
 521         rpc_wake_up_task_on_wq_queue_action_locked(rpciod_workqueue, queue,
 522                         task, rpc_task_action_set_status, &status);
 523 }
 524
 525 /**
 526  * rpc_wake_up_queued_task_set_status - wake up a task and set task->tk_status
 527  * @queue: pointer to rpc_wait_queue
 528  * @task: pointer to rpc_task
 529  * @status: integer error value
 530  *
 531  * If @task is queued on @queue, then it is woken up, and @task->tk_status is
 532  * set to the value of @status.
 533  */
 534 void
 535 rpc_wake_up_queued_task_set_status(struct rpc_wait_queue *queue,
 536                 struct rpc_task *task, int status)
 537 {
 538         if (!RPC_IS_QUEUED(task))
 539                 return;
 540         spin_lock_bh(&queue->lock);
 541         rpc_wake_up_task_queue_set_status_locked(queue, task, status);
 542         spin_unlock_bh(&queue->lock);
 543 }
 544
 545 /*
 546  * Wake up the next task on a priority queue.
 547  */
 548 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 549 {
 550         struct list_head *q;
 551         struct rpc_task *task;
 552
 553         /*
 554          * Service a batch of tasks from a single owner.
 555          */
 556         q = &queue->tasks[queue->priority];
 557         if (!list_empty(q) && --queue->nr) {
 558                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 559                 goto out;
 560         }
 561
 562         /*
 563          * Service the next queue.
 564          */
 565         do {
 566                 if (q == &queue->tasks[0])
 567                         q = &queue->tasks[queue->maxpriority];
 568                 else
 569                         q = q - 1;
 570                 if (!list_empty(q)) {
 571                         task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 572                         goto new_queue;
 573                 }
 574         } while (q != &queue->tasks[queue->priority]);
 575
 576         rpc_reset_waitqueue_priority(queue);
 577         return NULL;
 578
 579 new_queue:
 580         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 581 out:
 582         return task;
 583 }
 584
 585 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 586 {
 587         if (RPC_IS_PRIORITY(queue))
 588                 return __rpc_find_next_queued_priority(queue);
 589         if (!list_empty(&queue->tasks[0]))
 590                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 591         return NULL;
 592 }
 593
 594 /*
 595  * Wake up the first task on the wait queue.
 596  */
 597 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 598                 struct rpc_wait_queue *queue,
 599                 bool (*func)(struct rpc_task *, void *), void *data)
 600 {
 601         struct rpc_task *task = NULL;
 602
 603         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 604                         queue, rpc_qname(queue));
 605         spin_lock_bh(&queue->lock);
 606         task = __rpc_find_next_queued(queue);
 607         if (task != NULL)
 608                 task = rpc_wake_up_task_on_wq_queue_action_locked(wq, queue,
 609                                 task, func, data);
 610         spin_unlock_bh(&queue->lock);
 611
 612         return task;
 613 }
 614
 615 /*
 616  * Wake up the first task on the wait queue.
 617  */
 618 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 619                 bool (*func)(struct rpc_task *, void *), void *data)
 620 {
 621         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 622 }
 623 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 624
 625 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 626 {
 627         return true;
 628 }
 629
 630 /*
 631  * Wake up the next task on the wait queue.
 632 */
 633 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 634 {
 635         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 636 }
 637 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 638
 639 /**
 640  * rpc_wake_up - wake up all rpc_tasks
 641  * @queue: rpc_wait_queue on which the tasks are sleeping
 642  *
 643  * Grabs queue->lock
 644  */
 645 void rpc_wake_up(struct rpc_wait_queue *queue)
 646 {
 647         struct list_head *head;
 648
 649         spin_lock_bh(&queue->lock);
 650         head = &queue->tasks[queue->maxpriority];
 651         for (;;) {
 652                 while (!list_empty(head)) {
 653                         struct rpc_task *task;
 654                         task = list_first_entry(head,
 655                                         struct rpc_task,
 656                                         u.tk_wait.list);
 657                         rpc_wake_up_task_queue_locked(queue, task);
 658                 }
 659                 if (head == &queue->tasks[0])
 660                         break;
 661                 head--;
 662         }
 663         spin_unlock_bh(&queue->lock);
 664 }
 665 EXPORT_SYMBOL_GPL(rpc_wake_up);
 666
 667 /**
 668  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 669  * @queue: rpc_wait_queue on which the tasks are sleeping
 670  * @status: status value to set
 671  *
 672  * Grabs queue->lock
 673  */
 674 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 675 {
 676         struct list_head *head;
 677
 678         spin_lock_bh(&queue->lock);
 679         head = &queue->tasks[queue->maxpriority];
 680         for (;;) {
 681                 while (!list_empty(head)) {
 682                         struct rpc_task *task;
 683                         task = list_first_entry(head,
 684                                         struct rpc_task,
 685                                         u.tk_wait.list);
 686                         task->tk_status = status;
 687                         rpc_wake_up_task_queue_locked(queue, task);
 688                 }
 689                 if (head == &queue->tasks[0])
 690                         break;
 691                 head--;
 692         }
 693         spin_unlock_bh(&queue->lock);
 694 }
 695 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 696
 697 static void __rpc_queue_timer_fn(struct timer_list *t)
 698 {
 699         struct rpc_wait_queue *queue = from_timer(queue, t, timer_list.timer);
 700         struct rpc_task *task, *n;
 701         unsigned long expires, now, timeo;
 702
 703         spin_lock(&queue->lock);
 704         expires = now = jiffies;
 705         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 706                 timeo = task->u.tk_wait.expires;
 707                 if (time_after_eq(now, timeo)) {
 708                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 709                         task->tk_status = -ETIMEDOUT;
 710                         rpc_wake_up_task_queue_locked(queue, task);
 711                         continue;
 712                 }
 713                 if (expires == now || time_after(expires, timeo))
 714                         expires = timeo;
 715         }
 716         if (!list_empty(&queue->timer_list.list))
 717                 rpc_set_queue_timer(queue, expires);
 718         spin_unlock(&queue->lock);
 719 }
 720
 721 static void __rpc_atrun(struct rpc_task *task)
 722 {
 723         if (task->tk_status == -ETIMEDOUT)
 724                 task->tk_status = 0;
 725 }
 726
 727 /*
 728  * Run a task at a later time
 729  */
 730 void rpc_delay(struct rpc_task *task, unsigned long delay)
 731 {
 732         task->tk_timeout = delay;
 733         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 734 }
 735 EXPORT_SYMBOL_GPL(rpc_delay);
 736
 737 /*
 738  * Helper to call task->tk_ops->rpc_call_prepare
 739  */
 740 void rpc_prepare_task(struct rpc_task *task)
 741 {
 742         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 743 }
 744
 745 static void
 746 rpc_init_task_statistics(struct rpc_task *task)
 747 {
 748         /* Initialize retry counters */
 749         task->tk_garb_retry = 2;
 750         task->tk_cred_retry = 2;
 751         task->tk_rebind_retry = 2;
 752
 753         /* starting timestamp */
 754         task->tk_start = ktime_get();
 755 }
 756
 757 static void
 758 rpc_reset_task_statistics(struct rpc_task *task)
 759 {
 760         task->tk_timeouts = 0;
 761         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 762
 763         rpc_init_task_statistics(task);
 764 }
 765
 766 /*
 767  * Helper that calls task->tk_ops->rpc_call_done if it exists
 768  */
 769 void rpc_exit_task(struct rpc_task *task)
 770 {
 771         task->tk_action = NULL;
 772         if (task->tk_ops->rpc_call_done != NULL) {
 773                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 774                 if (task->tk_action != NULL) {
 775                         WARN_ON(RPC_ASSASSINATED(task));
 776                         /* Always release the RPC slot and buffer memory */
 777                         xprt_release(task);
 778                         rpc_reset_task_statistics(task);
 779                 }
 780         }
 781 }
 782
 783 void rpc_exit(struct rpc_task *task, int status)
 784 {
 785         task->tk_status = status;
 786         task->tk_action = rpc_exit_task;
 787         if (RPC_IS_QUEUED(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         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 906 }
 907
 908 /**
 909  * rpc_malloc - allocate RPC buffer resources
 910  * @task: RPC task
 911  *
 912  * A single memory region is allocated, which is split between the
 913  * RPC call and RPC reply that this task is being used for. When
 914  * this RPC is retired, the memory is released by calling rpc_free.
 915  *
 916  * To prevent rpciod from hanging, this allocator never sleeps,
 917  * returning -ENOMEM and suppressing warning if the request cannot
 918  * be serviced immediately. The caller can arrange to sleep in a
 919  * way that is safe for rpciod.
 920  *
 921  * Most requests are 'small' (under 2KiB) and can be serviced from a
 922  * mempool, ensuring that NFS reads and writes can always proceed,
 923  * and that there is good locality of reference for these buffers.
 924  *
 925  * In order to avoid memory starvation triggering more writebacks of
 926  * NFS requests, we avoid using GFP_KERNEL.
 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_NOIO | __GFP_NOWARN;
 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_NOIO);
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         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1071 }
1072
1073 static void rpc_release_resources_task(struct rpc_task *task)
1074 {
1075         xprt_release(task);
1076         if (task->tk_msg.rpc_cred) {
1077                 put_cred(task->tk_msg.rpc_cred);
1078                 task->tk_msg.rpc_cred = NULL;
1079         }
1080         rpc_task_release_client(task);
1081 }
1082
1083 static void rpc_final_put_task(struct rpc_task *task,
1084                 struct workqueue_struct *q)
1085 {
1086         if (q != NULL) {
1087                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1088                 queue_work(q, &task->u.tk_work);
1089         } else
1090                 rpc_free_task(task);
1091 }
1092
1093 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1094 {
1095         if (atomic_dec_and_test(&task->tk_count)) {
1096                 rpc_release_resources_task(task);
1097                 rpc_final_put_task(task, q);
1098         }
1099 }
1100
1101 void rpc_put_task(struct rpc_task *task)
1102 {
1103         rpc_do_put_task(task, NULL);
1104 }
1105 EXPORT_SYMBOL_GPL(rpc_put_task);
1106
1107 void rpc_put_task_async(struct rpc_task *task)
1108 {
1109         rpc_do_put_task(task, task->tk_workqueue);
1110 }
1111 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1112
1113 static void rpc_release_task(struct rpc_task *task)
1114 {
1115         dprintk("RPC: %5u release task\n", task->tk_pid);
1116
1117         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1118
1119         rpc_release_resources_task(task);
1120
1121         /*
1122          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1123          * so it should be safe to use task->tk_count as a test for whether
1124          * or not any other processes still hold references to our rpc_task.
1125          */
1126         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1127                 /* Wake up anyone who may be waiting for task completion */
1128                 if (!rpc_complete_task(task))
1129                         return;
1130         } else {
1131                 if (!atomic_dec_and_test(&task->tk_count))
1132                         return;
1133         }
1134         rpc_final_put_task(task, task->tk_workqueue);
1135 }
1136
1137 int rpciod_up(void)
1138 {
1139         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1140 }
1141
1142 void rpciod_down(void)
1143 {
1144         module_put(THIS_MODULE);
1145 }
1146
1147 /*
1148  * Start up the rpciod workqueue.
1149  */
1150 static int rpciod_start(void)
1151 {
1152         struct workqueue_struct *wq;
1153
1154         /*
1155          * Create the rpciod thread and wait for it to start.
1156          */
1157         dprintk("RPC:       creating workqueue rpciod\n");
1158         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1159         if (!wq)
1160                 goto out_failed;
1161         rpciod_workqueue = wq;
1162         /* Note: highpri because network receive is latency sensitive */
1163         wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1164         if (!wq)
1165                 goto free_rpciod;
1166         xprtiod_workqueue = wq;
1167         return 1;
1168 free_rpciod:
1169         wq = rpciod_workqueue;
1170         rpciod_workqueue = NULL;
1171         destroy_workqueue(wq);
1172 out_failed:
1173         return 0;
1174 }
1175
1176 static void rpciod_stop(void)
1177 {
1178         struct workqueue_struct *wq = NULL;
1179
1180         if (rpciod_workqueue == NULL)
1181                 return;
1182         dprintk("RPC:       destroying workqueue rpciod\n");
1183
1184         wq = rpciod_workqueue;
1185         rpciod_workqueue = NULL;
1186         destroy_workqueue(wq);
1187         wq = xprtiod_workqueue;
1188         xprtiod_workqueue = NULL;
1189         destroy_workqueue(wq);
1190 }
1191
1192 void
1193 rpc_destroy_mempool(void)
1194 {
1195         rpciod_stop();
1196         mempool_destroy(rpc_buffer_mempool);
1197         mempool_destroy(rpc_task_mempool);
1198         kmem_cache_destroy(rpc_task_slabp);
1199         kmem_cache_destroy(rpc_buffer_slabp);
1200         rpc_destroy_wait_queue(&delay_queue);
1201 }
1202
1203 int
1204 rpc_init_mempool(void)
1205 {
1206         /*
1207          * The following is not strictly a mempool initialisation,
1208          * but there is no harm in doing it here
1209          */
1210         rpc_init_wait_queue(&delay_queue, "delayq");
1211         if (!rpciod_start())
1212                 goto err_nomem;
1213
1214         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1215                                              sizeof(struct rpc_task),
1216                                              0, SLAB_HWCACHE_ALIGN,
1217                                              NULL);
1218         if (!rpc_task_slabp)
1219                 goto err_nomem;
1220         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1221                                              RPC_BUFFER_MAXSIZE,
1222                                              0, SLAB_HWCACHE_ALIGN,
1223                                              NULL);
1224         if (!rpc_buffer_slabp)
1225                 goto err_nomem;
1226         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1227                                                     rpc_task_slabp);
1228         if (!rpc_task_mempool)
1229                 goto err_nomem;
1230         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1231                                                       rpc_buffer_slabp);
1232         if (!rpc_buffer_mempool)
1233                 goto err_nomem;
1234         return 0;
1235 err_nomem:
1236         rpc_destroy_mempool();
1237         return -ENOMEM;
1238 }