net/sunrpc/sched.c

   1 /*
   2  * linux/net/sunrpc/sched.c
   3  *
   4  * Scheduling for synchronous and asynchronous RPC requests.
   5  *
   6  * Copyright (C) 1996 Olaf Kirch, <okir@monad.swb.de>
   7  *
   8  * TCP NFS related read + write fixes
   9  * (C) 1999 Dave Airlie, University of Limerick, Ireland <airlied@linux.ie>
  10  */
  11
  12 #include <linux/module.h>
  13
  14 #include <linux/sched.h>
  15 #include <linux/interrupt.h>
  16 #include <linux/slab.h>
  17 #include <linux/mempool.h>
  18 #include <linux/smp.h>
  19 #include <linux/spinlock.h>
  20 #include <linux/mutex.h>
  21 #include <linux/freezer.h>
  22
  23 #include <linux/sunrpc/clnt.h>
  24
  25 #include "sunrpc.h"
  26
  27 #if IS_ENABLED(CONFIG_SUNRPC_DEBUG)
  28 #define RPCDBG_FACILITY         RPCDBG_SCHED
  29 #endif
  30
  31 #define CREATE_TRACE_POINTS
  32 #include <trace/events/sunrpc.h>
  33
  34 /*
  35  * RPC slabs and memory pools
  36  */
  37 #define RPC_BUFFER_MAXSIZE      (2048)
  38 #define RPC_BUFFER_POOLSIZE     (8)
  39 #define RPC_TASK_POOLSIZE       (8)
  40 static struct kmem_cache        *rpc_task_slabp __read_mostly;
  41 static struct kmem_cache        *rpc_buffer_slabp __read_mostly;
  42 static mempool_t        *rpc_task_mempool __read_mostly;
  43 static mempool_t        *rpc_buffer_mempool __read_mostly;
  44
  45 static void                     rpc_async_schedule(struct work_struct *);
  46 static void                      rpc_release_task(struct rpc_task *task);
  47 static void __rpc_queue_timer_fn(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 void rpc_wake_up_task_on_wq_queue_locked(struct workqueue_struct *wq,
 453                 struct rpc_wait_queue *queue, struct rpc_task *task)
 454 {
 455         if (RPC_IS_QUEUED(task)) {
 456                 smp_rmb();
 457                 if (task->tk_waitqueue == queue)
 458                         __rpc_do_wake_up_task_on_wq(wq, queue, task);
 459         }
 460 }
 461
 462 /*
 463  * Wake up a queued task while the queue lock is being held
 464  */
 465 static void rpc_wake_up_task_queue_locked(struct rpc_wait_queue *queue, struct rpc_task *task)
 466 {
 467         rpc_wake_up_task_on_wq_queue_locked(rpciod_workqueue, queue, task);
 468 }
 469
 470 /*
 471  * Wake up a task on a specific queue
 472  */
 473 void rpc_wake_up_queued_task_on_wq(struct workqueue_struct *wq,
 474                 struct rpc_wait_queue *queue,
 475                 struct rpc_task *task)
 476 {
 477         spin_lock_bh(&queue->lock);
 478         rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
 479         spin_unlock_bh(&queue->lock);
 480 }
 481
 482 /*
 483  * Wake up a task on a specific queue
 484  */
 485 void rpc_wake_up_queued_task(struct rpc_wait_queue *queue, struct rpc_task *task)
 486 {
 487         spin_lock_bh(&queue->lock);
 488         rpc_wake_up_task_queue_locked(queue, task);
 489         spin_unlock_bh(&queue->lock);
 490 }
 491 EXPORT_SYMBOL_GPL(rpc_wake_up_queued_task);
 492
 493 /*
 494  * Wake up the next task on a priority queue.
 495  */
 496 static struct rpc_task *__rpc_find_next_queued_priority(struct rpc_wait_queue *queue)
 497 {
 498         struct list_head *q;
 499         struct rpc_task *task;
 500
 501         /*
 502          * Service the privileged queue.
 503          */
 504         q = &queue->tasks[RPC_NR_PRIORITY - 1];
 505         if (queue->maxpriority > RPC_PRIORITY_PRIVILEGED && !list_empty(q)) {
 506                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 507                 goto out;
 508         }
 509
 510         /*
 511          * Service a batch of tasks from a single owner.
 512          */
 513         q = &queue->tasks[queue->priority];
 514         if (!list_empty(q) && queue->nr) {
 515                 queue->nr--;
 516                 task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 517                 goto out;
 518         }
 519
 520         /*
 521          * Service the next queue.
 522          */
 523         do {
 524                 if (q == &queue->tasks[0])
 525                         q = &queue->tasks[queue->maxpriority];
 526                 else
 527                         q = q - 1;
 528                 if (!list_empty(q)) {
 529                         task = list_first_entry(q, struct rpc_task, u.tk_wait.list);
 530                         goto new_queue;
 531                 }
 532         } while (q != &queue->tasks[queue->priority]);
 533
 534         rpc_reset_waitqueue_priority(queue);
 535         return NULL;
 536
 537 new_queue:
 538         rpc_set_waitqueue_priority(queue, (unsigned int)(q - &queue->tasks[0]));
 539 out:
 540         return task;
 541 }
 542
 543 static struct rpc_task *__rpc_find_next_queued(struct rpc_wait_queue *queue)
 544 {
 545         if (RPC_IS_PRIORITY(queue))
 546                 return __rpc_find_next_queued_priority(queue);
 547         if (!list_empty(&queue->tasks[0]))
 548                 return list_first_entry(&queue->tasks[0], struct rpc_task, u.tk_wait.list);
 549         return NULL;
 550 }
 551
 552 /*
 553  * Wake up the first task on the wait queue.
 554  */
 555 struct rpc_task *rpc_wake_up_first_on_wq(struct workqueue_struct *wq,
 556                 struct rpc_wait_queue *queue,
 557                 bool (*func)(struct rpc_task *, void *), void *data)
 558 {
 559         struct rpc_task *task = NULL;
 560
 561         dprintk("RPC:       wake_up_first(%p \"%s\")\n",
 562                         queue, rpc_qname(queue));
 563         spin_lock_bh(&queue->lock);
 564         task = __rpc_find_next_queued(queue);
 565         if (task != NULL) {
 566                 if (func(task, data))
 567                         rpc_wake_up_task_on_wq_queue_locked(wq, queue, task);
 568                 else
 569                         task = NULL;
 570         }
 571         spin_unlock_bh(&queue->lock);
 572
 573         return task;
 574 }
 575
 576 /*
 577  * Wake up the first task on the wait queue.
 578  */
 579 struct rpc_task *rpc_wake_up_first(struct rpc_wait_queue *queue,
 580                 bool (*func)(struct rpc_task *, void *), void *data)
 581 {
 582         return rpc_wake_up_first_on_wq(rpciod_workqueue, queue, func, data);
 583 }
 584 EXPORT_SYMBOL_GPL(rpc_wake_up_first);
 585
 586 static bool rpc_wake_up_next_func(struct rpc_task *task, void *data)
 587 {
 588         return true;
 589 }
 590
 591 /*
 592  * Wake up the next task on the wait queue.
 593 */
 594 struct rpc_task *rpc_wake_up_next(struct rpc_wait_queue *queue)
 595 {
 596         return rpc_wake_up_first(queue, rpc_wake_up_next_func, NULL);
 597 }
 598 EXPORT_SYMBOL_GPL(rpc_wake_up_next);
 599
 600 /**
 601  * rpc_wake_up - wake up all rpc_tasks
 602  * @queue: rpc_wait_queue on which the tasks are sleeping
 603  *
 604  * Grabs queue->lock
 605  */
 606 void rpc_wake_up(struct rpc_wait_queue *queue)
 607 {
 608         struct list_head *head;
 609
 610         spin_lock_bh(&queue->lock);
 611         head = &queue->tasks[queue->maxpriority];
 612         for (;;) {
 613                 while (!list_empty(head)) {
 614                         struct rpc_task *task;
 615                         task = list_first_entry(head,
 616                                         struct rpc_task,
 617                                         u.tk_wait.list);
 618                         rpc_wake_up_task_queue_locked(queue, task);
 619                 }
 620                 if (head == &queue->tasks[0])
 621                         break;
 622                 head--;
 623         }
 624         spin_unlock_bh(&queue->lock);
 625 }
 626 EXPORT_SYMBOL_GPL(rpc_wake_up);
 627
 628 /**
 629  * rpc_wake_up_status - wake up all rpc_tasks and set their status value.
 630  * @queue: rpc_wait_queue on which the tasks are sleeping
 631  * @status: status value to set
 632  *
 633  * Grabs queue->lock
 634  */
 635 void rpc_wake_up_status(struct rpc_wait_queue *queue, int status)
 636 {
 637         struct list_head *head;
 638
 639         spin_lock_bh(&queue->lock);
 640         head = &queue->tasks[queue->maxpriority];
 641         for (;;) {
 642                 while (!list_empty(head)) {
 643                         struct rpc_task *task;
 644                         task = list_first_entry(head,
 645                                         struct rpc_task,
 646                                         u.tk_wait.list);
 647                         task->tk_status = status;
 648                         rpc_wake_up_task_queue_locked(queue, task);
 649                 }
 650                 if (head == &queue->tasks[0])
 651                         break;
 652                 head--;
 653         }
 654         spin_unlock_bh(&queue->lock);
 655 }
 656 EXPORT_SYMBOL_GPL(rpc_wake_up_status);
 657
 658 static void __rpc_queue_timer_fn(struct timer_list *t)
 659 {
 660         struct rpc_wait_queue *queue = from_timer(queue, t, timer_list.timer);
 661         struct rpc_task *task, *n;
 662         unsigned long expires, now, timeo;
 663
 664         spin_lock(&queue->lock);
 665         expires = now = jiffies;
 666         list_for_each_entry_safe(task, n, &queue->timer_list.list, u.tk_wait.timer_list) {
 667                 timeo = task->u.tk_wait.expires;
 668                 if (time_after_eq(now, timeo)) {
 669                         dprintk("RPC: %5u timeout\n", task->tk_pid);
 670                         task->tk_status = -ETIMEDOUT;
 671                         rpc_wake_up_task_queue_locked(queue, task);
 672                         continue;
 673                 }
 674                 if (expires == now || time_after(expires, timeo))
 675                         expires = timeo;
 676         }
 677         if (!list_empty(&queue->timer_list.list))
 678                 rpc_set_queue_timer(queue, expires);
 679         spin_unlock(&queue->lock);
 680 }
 681
 682 static void __rpc_atrun(struct rpc_task *task)
 683 {
 684         if (task->tk_status == -ETIMEDOUT)
 685                 task->tk_status = 0;
 686 }
 687
 688 /*
 689  * Run a task at a later time
 690  */
 691 void rpc_delay(struct rpc_task *task, unsigned long delay)
 692 {
 693         task->tk_timeout = delay;
 694         rpc_sleep_on(&delay_queue, task, __rpc_atrun);
 695 }
 696 EXPORT_SYMBOL_GPL(rpc_delay);
 697
 698 /*
 699  * Helper to call task->tk_ops->rpc_call_prepare
 700  */
 701 void rpc_prepare_task(struct rpc_task *task)
 702 {
 703         task->tk_ops->rpc_call_prepare(task, task->tk_calldata);
 704 }
 705
 706 static void
 707 rpc_init_task_statistics(struct rpc_task *task)
 708 {
 709         /* Initialize retry counters */
 710         task->tk_garb_retry = 2;
 711         task->tk_cred_retry = 2;
 712
 713         /* starting timestamp */
 714         task->tk_start = ktime_get();
 715 }
 716
 717 static void
 718 rpc_reset_task_statistics(struct rpc_task *task)
 719 {
 720         task->tk_timeouts = 0;
 721         task->tk_flags &= ~(RPC_CALL_MAJORSEEN|RPC_TASK_KILLED|RPC_TASK_SENT);
 722
 723         rpc_init_task_statistics(task);
 724 }
 725
 726 /*
 727  * Helper that calls task->tk_ops->rpc_call_done if it exists
 728  */
 729 void rpc_exit_task(struct rpc_task *task)
 730 {
 731         task->tk_action = NULL;
 732         if (task->tk_ops->rpc_call_done != NULL) {
 733                 task->tk_ops->rpc_call_done(task, task->tk_calldata);
 734                 if (task->tk_action != NULL) {
 735                         WARN_ON(RPC_ASSASSINATED(task));
 736                         /* Always release the RPC slot and buffer memory */
 737                         xprt_release(task);
 738                         rpc_reset_task_statistics(task);
 739                 }
 740         }
 741 }
 742
 743 void rpc_exit(struct rpc_task *task, int status)
 744 {
 745         task->tk_status = status;
 746         task->tk_action = rpc_exit_task;
 747         if (RPC_IS_QUEUED(task))
 748                 rpc_wake_up_queued_task(task->tk_waitqueue, task);
 749 }
 750 EXPORT_SYMBOL_GPL(rpc_exit);
 751
 752 void rpc_release_calldata(const struct rpc_call_ops *ops, void *calldata)
 753 {
 754         if (ops->rpc_release != NULL)
 755                 ops->rpc_release(calldata);
 756 }
 757
 758 /*
 759  * This is the RPC `scheduler' (or rather, the finite state machine).
 760  */
 761 static void __rpc_execute(struct rpc_task *task)
 762 {
 763         struct rpc_wait_queue *queue;
 764         int task_is_async = RPC_IS_ASYNC(task);
 765         int status = 0;
 766
 767         dprintk("RPC: %5u __rpc_execute flags=0x%x\n",
 768                         task->tk_pid, task->tk_flags);
 769
 770         WARN_ON_ONCE(RPC_IS_QUEUED(task));
 771         if (RPC_IS_QUEUED(task))
 772                 return;
 773
 774         for (;;) {
 775                 void (*do_action)(struct rpc_task *);
 776
 777                 /*
 778                  * Perform the next FSM step or a pending callback.
 779                  *
 780                  * tk_action may be NULL if the task has been killed.
 781                  * In particular, note that rpc_killall_tasks may
 782                  * do this at any time, so beware when dereferencing.
 783                  */
 784                 do_action = task->tk_action;
 785                 if (task->tk_callback) {
 786                         do_action = task->tk_callback;
 787                         task->tk_callback = NULL;
 788                 }
 789                 if (!do_action)
 790                         break;
 791                 trace_rpc_task_run_action(task, do_action);
 792                 do_action(task);
 793
 794                 /*
 795                  * Lockless check for whether task is sleeping or not.
 796                  */
 797                 if (!RPC_IS_QUEUED(task))
 798                         continue;
 799                 /*
 800                  * The queue->lock protects against races with
 801                  * rpc_make_runnable().
 802                  *
 803                  * Note that once we clear RPC_TASK_RUNNING on an asynchronous
 804                  * rpc_task, rpc_make_runnable() can assign it to a
 805                  * different workqueue. We therefore cannot assume that the
 806                  * rpc_task pointer may still be dereferenced.
 807                  */
 808                 queue = task->tk_waitqueue;
 809                 spin_lock_bh(&queue->lock);
 810                 if (!RPC_IS_QUEUED(task)) {
 811                         spin_unlock_bh(&queue->lock);
 812                         continue;
 813                 }
 814                 rpc_clear_running(task);
 815                 spin_unlock_bh(&queue->lock);
 816                 if (task_is_async)
 817                         return;
 818
 819                 /* sync task: sleep here */
 820                 dprintk("RPC: %5u sync task going to sleep\n", task->tk_pid);
 821                 status = out_of_line_wait_on_bit(&task->tk_runstate,
 822                                 RPC_TASK_QUEUED, rpc_wait_bit_killable,
 823                                 TASK_KILLABLE);
 824                 if (status == -ERESTARTSYS) {
 825                         /*
 826                          * When a sync task receives a signal, it exits with
 827                          * -ERESTARTSYS. In order to catch any callbacks that
 828                          * clean up after sleeping on some queue, we don't
 829                          * break the loop here, but go around once more.
 830                          */
 831                         dprintk("RPC: %5u got signal\n", task->tk_pid);
 832                         task->tk_flags |= RPC_TASK_KILLED;
 833                         rpc_exit(task, -ERESTARTSYS);
 834                 }
 835                 dprintk("RPC: %5u sync task resuming\n", task->tk_pid);
 836         }
 837
 838         dprintk("RPC: %5u return %d, status %d\n", task->tk_pid, status,
 839                         task->tk_status);
 840         /* Release all resources associated with the task */
 841         rpc_release_task(task);
 842 }
 843
 844 /*
 845  * User-visible entry point to the scheduler.
 846  *
 847  * This may be called recursively if e.g. an async NFS task updates
 848  * the attributes and finds that dirty pages must be flushed.
 849  * NOTE: Upon exit of this function the task is guaranteed to be
 850  *       released. In particular note that tk_release() will have
 851  *       been called, so your task memory may have been freed.
 852  */
 853 void rpc_execute(struct rpc_task *task)
 854 {
 855         bool is_async = RPC_IS_ASYNC(task);
 856
 857         rpc_set_active(task);
 858         rpc_make_runnable(rpciod_workqueue, task);
 859         if (!is_async)
 860                 __rpc_execute(task);
 861 }
 862
 863 static void rpc_async_schedule(struct work_struct *work)
 864 {
 865         __rpc_execute(container_of(work, struct rpc_task, u.tk_work));
 866 }
 867
 868 /**
 869  * rpc_malloc - allocate RPC buffer resources
 870  * @task: RPC task
 871  *
 872  * A single memory region is allocated, which is split between the
 873  * RPC call and RPC reply that this task is being used for. When
 874  * this RPC is retired, the memory is released by calling rpc_free.
 875  *
 876  * To prevent rpciod from hanging, this allocator never sleeps,
 877  * returning -ENOMEM and suppressing warning if the request cannot
 878  * be serviced immediately. The caller can arrange to sleep in a
 879  * way that is safe for rpciod.
 880  *
 881  * Most requests are 'small' (under 2KiB) and can be serviced from a
 882  * mempool, ensuring that NFS reads and writes can always proceed,
 883  * and that there is good locality of reference for these buffers.
 884  *
 885  * In order to avoid memory starvation triggering more writebacks of
 886  * NFS requests, we avoid using GFP_KERNEL.
 887  */
 888 int rpc_malloc(struct rpc_task *task)
 889 {
 890         struct rpc_rqst *rqst = task->tk_rqstp;
 891         size_t size = rqst->rq_callsize + rqst->rq_rcvsize;
 892         struct rpc_buffer *buf;
 893         gfp_t gfp = GFP_NOIO | __GFP_NOWARN;
 894
 895         if (RPC_IS_ASYNC(task))
 896                 gfp = GFP_NOWAIT | __GFP_NOWARN;
 897         if (RPC_IS_SWAPPER(task))
 898                 gfp |= __GFP_MEMALLOC;
 899
 900         size += sizeof(struct rpc_buffer);
 901         if (size <= RPC_BUFFER_MAXSIZE)
 902                 buf = mempool_alloc(rpc_buffer_mempool, gfp);
 903         else
 904                 buf = kmalloc(size, gfp);
 905
 906         if (!buf)
 907                 return -ENOMEM;
 908
 909         buf->len = size;
 910         dprintk("RPC: %5u allocated buffer of size %zu at %p\n",
 911                         task->tk_pid, size, buf);
 912         rqst->rq_buffer = buf->data;
 913         rqst->rq_rbuffer = (char *)rqst->rq_buffer + rqst->rq_callsize;
 914         return 0;
 915 }
 916 EXPORT_SYMBOL_GPL(rpc_malloc);
 917
 918 /**
 919  * rpc_free - free RPC buffer resources allocated via rpc_malloc
 920  * @task: RPC task
 921  *
 922  */
 923 void rpc_free(struct rpc_task *task)
 924 {
 925         void *buffer = task->tk_rqstp->rq_buffer;
 926         size_t size;
 927         struct rpc_buffer *buf;
 928
 929         buf = container_of(buffer, struct rpc_buffer, data);
 930         size = buf->len;
 931
 932         dprintk("RPC:       freeing buffer of size %zu at %p\n",
 933                         size, buf);
 934
 935         if (size <= RPC_BUFFER_MAXSIZE)
 936                 mempool_free(buf, rpc_buffer_mempool);
 937         else
 938                 kfree(buf);
 939 }
 940 EXPORT_SYMBOL_GPL(rpc_free);
 941
 942 /*
 943  * Creation and deletion of RPC task structures
 944  */
 945 static void rpc_init_task(struct rpc_task *task, const struct rpc_task_setup *task_setup_data)
 946 {
 947         memset(task, 0, sizeof(*task));
 948         atomic_set(&task->tk_count, 1);
 949         task->tk_flags  = task_setup_data->flags;
 950         task->tk_ops = task_setup_data->callback_ops;
 951         task->tk_calldata = task_setup_data->callback_data;
 952         INIT_LIST_HEAD(&task->tk_task);
 953
 954         task->tk_priority = task_setup_data->priority - RPC_PRIORITY_LOW;
 955         task->tk_owner = current->tgid;
 956
 957         /* Initialize workqueue for async tasks */
 958         task->tk_workqueue = task_setup_data->workqueue;
 959
 960         task->tk_xprt = xprt_get(task_setup_data->rpc_xprt);
 961
 962         if (task->tk_ops->rpc_call_prepare != NULL)
 963                 task->tk_action = rpc_prepare_task;
 964
 965         rpc_init_task_statistics(task);
 966
 967         dprintk("RPC:       new task initialized, procpid %u\n",
 968                                 task_pid_nr(current));
 969 }
 970
 971 static struct rpc_task *
 972 rpc_alloc_task(void)
 973 {
 974         return (struct rpc_task *)mempool_alloc(rpc_task_mempool, GFP_NOIO);
 975 }
 976
 977 /*
 978  * Create a new task for the specified client.
 979  */
 980 struct rpc_task *rpc_new_task(const struct rpc_task_setup *setup_data)
 981 {
 982         struct rpc_task *task = setup_data->task;
 983         unsigned short flags = 0;
 984
 985         if (task == NULL) {
 986                 task = rpc_alloc_task();
 987                 flags = RPC_TASK_DYNAMIC;
 988         }
 989
 990         rpc_init_task(task, setup_data);
 991         task->tk_flags |= flags;
 992         dprintk("RPC:       allocated task %p\n", task);
 993         return task;
 994 }
 995
 996 /*
 997  * rpc_free_task - release rpc task and perform cleanups
 998  *
 999  * Note that we free up the rpc_task _after_ rpc_release_calldata()
1000  * in order to work around a workqueue dependency issue.
1001  *
1002  * Tejun Heo states:
1003  * "Workqueue currently considers two work items to be the same if they're
1004  * on the same address and won't execute them concurrently - ie. it
1005  * makes a work item which is queued again while being executed wait
1006  * for the previous execution to complete.
1007  *
1008  * If a work function frees the work item, and then waits for an event
1009  * which should be performed by another work item and *that* work item
1010  * recycles the freed work item, it can create a false dependency loop.
1011  * There really is no reliable way to detect this short of verifying
1012  * every memory free."
1013  *
1014  */
1015 static void rpc_free_task(struct rpc_task *task)
1016 {
1017         unsigned short tk_flags = task->tk_flags;
1018
1019         rpc_release_calldata(task->tk_ops, task->tk_calldata);
1020
1021         if (tk_flags & RPC_TASK_DYNAMIC) {
1022                 dprintk("RPC: %5u freeing task\n", task->tk_pid);
1023                 mempool_free(task, rpc_task_mempool);
1024         }
1025 }
1026
1027 static void rpc_async_release(struct work_struct *work)
1028 {
1029         rpc_free_task(container_of(work, struct rpc_task, u.tk_work));
1030 }
1031
1032 static void rpc_release_resources_task(struct rpc_task *task)
1033 {
1034         xprt_release(task);
1035         if (task->tk_msg.rpc_cred) {
1036                 put_rpccred(task->tk_msg.rpc_cred);
1037                 task->tk_msg.rpc_cred = NULL;
1038         }
1039         rpc_task_release_client(task);
1040 }
1041
1042 static void rpc_final_put_task(struct rpc_task *task,
1043                 struct workqueue_struct *q)
1044 {
1045         if (q != NULL) {
1046                 INIT_WORK(&task->u.tk_work, rpc_async_release);
1047                 queue_work(q, &task->u.tk_work);
1048         } else
1049                 rpc_free_task(task);
1050 }
1051
1052 static void rpc_do_put_task(struct rpc_task *task, struct workqueue_struct *q)
1053 {
1054         if (atomic_dec_and_test(&task->tk_count)) {
1055                 rpc_release_resources_task(task);
1056                 rpc_final_put_task(task, q);
1057         }
1058 }
1059
1060 void rpc_put_task(struct rpc_task *task)
1061 {
1062         rpc_do_put_task(task, NULL);
1063 }
1064 EXPORT_SYMBOL_GPL(rpc_put_task);
1065
1066 void rpc_put_task_async(struct rpc_task *task)
1067 {
1068         rpc_do_put_task(task, task->tk_workqueue);
1069 }
1070 EXPORT_SYMBOL_GPL(rpc_put_task_async);
1071
1072 static void rpc_release_task(struct rpc_task *task)
1073 {
1074         dprintk("RPC: %5u release task\n", task->tk_pid);
1075
1076         WARN_ON_ONCE(RPC_IS_QUEUED(task));
1077
1078         rpc_release_resources_task(task);
1079
1080         /*
1081          * Note: at this point we have been removed from rpc_clnt->cl_tasks,
1082          * so it should be safe to use task->tk_count as a test for whether
1083          * or not any other processes still hold references to our rpc_task.
1084          */
1085         if (atomic_read(&task->tk_count) != 1 + !RPC_IS_ASYNC(task)) {
1086                 /* Wake up anyone who may be waiting for task completion */
1087                 if (!rpc_complete_task(task))
1088                         return;
1089         } else {
1090                 if (!atomic_dec_and_test(&task->tk_count))
1091                         return;
1092         }
1093         rpc_final_put_task(task, task->tk_workqueue);
1094 }
1095
1096 int rpciod_up(void)
1097 {
1098         return try_module_get(THIS_MODULE) ? 0 : -EINVAL;
1099 }
1100
1101 void rpciod_down(void)
1102 {
1103         module_put(THIS_MODULE);
1104 }
1105
1106 /*
1107  * Start up the rpciod workqueue.
1108  */
1109 static int rpciod_start(void)
1110 {
1111         struct workqueue_struct *wq;
1112
1113         /*
1114          * Create the rpciod thread and wait for it to start.
1115          */
1116         dprintk("RPC:       creating workqueue rpciod\n");
1117         wq = alloc_workqueue("rpciod", WQ_MEM_RECLAIM | WQ_UNBOUND, 0);
1118         if (!wq)
1119                 goto out_failed;
1120         rpciod_workqueue = wq;
1121         /* Note: highpri because network receive is latency sensitive */
1122         wq = alloc_workqueue("xprtiod", WQ_UNBOUND|WQ_MEM_RECLAIM|WQ_HIGHPRI, 0);
1123         if (!wq)
1124                 goto free_rpciod;
1125         xprtiod_workqueue = wq;
1126         return 1;
1127 free_rpciod:
1128         wq = rpciod_workqueue;
1129         rpciod_workqueue = NULL;
1130         destroy_workqueue(wq);
1131 out_failed:
1132         return 0;
1133 }
1134
1135 static void rpciod_stop(void)
1136 {
1137         struct workqueue_struct *wq = NULL;
1138
1139         if (rpciod_workqueue == NULL)
1140                 return;
1141         dprintk("RPC:       destroying workqueue rpciod\n");
1142
1143         wq = rpciod_workqueue;
1144         rpciod_workqueue = NULL;
1145         destroy_workqueue(wq);
1146         wq = xprtiod_workqueue;
1147         xprtiod_workqueue = NULL;
1148         destroy_workqueue(wq);
1149 }
1150
1151 void
1152 rpc_destroy_mempool(void)
1153 {
1154         rpciod_stop();
1155         mempool_destroy(rpc_buffer_mempool);
1156         mempool_destroy(rpc_task_mempool);
1157         kmem_cache_destroy(rpc_task_slabp);
1158         kmem_cache_destroy(rpc_buffer_slabp);
1159         rpc_destroy_wait_queue(&delay_queue);
1160 }
1161
1162 int
1163 rpc_init_mempool(void)
1164 {
1165         /*
1166          * The following is not strictly a mempool initialisation,
1167          * but there is no harm in doing it here
1168          */
1169         rpc_init_wait_queue(&delay_queue, "delayq");
1170         if (!rpciod_start())
1171                 goto err_nomem;
1172
1173         rpc_task_slabp = kmem_cache_create("rpc_tasks",
1174                                              sizeof(struct rpc_task),
1175                                              0, SLAB_HWCACHE_ALIGN,
1176                                              NULL);
1177         if (!rpc_task_slabp)
1178                 goto err_nomem;
1179         rpc_buffer_slabp = kmem_cache_create("rpc_buffers",
1180                                              RPC_BUFFER_MAXSIZE,
1181                                              0, SLAB_HWCACHE_ALIGN,
1182                                              NULL);
1183         if (!rpc_buffer_slabp)
1184                 goto err_nomem;
1185         rpc_task_mempool = mempool_create_slab_pool(RPC_TASK_POOLSIZE,
1186                                                     rpc_task_slabp);
1187         if (!rpc_task_mempool)
1188                 goto err_nomem;
1189         rpc_buffer_mempool = mempool_create_slab_pool(RPC_BUFFER_POOLSIZE,
1190                                                       rpc_buffer_slabp);
1191         if (!rpc_buffer_mempool)
1192                 goto err_nomem;
1193         return 0;
1194 err_nomem:
1195         rpc_destroy_mempool();
1196         return -ENOMEM;
1197 }