fs/io-wq.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Basic worker thread pool for io_uring
   4  *
   5  * Copyright (C) 2019 Jens Axboe
   6  *
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/errno.h>
  11 #include <linux/sched/signal.h>
  12 #include <linux/mm.h>
  13 #include <linux/mmu_context.h>
  14 #include <linux/sched/mm.h>
  15 #include <linux/percpu.h>
  16 #include <linux/slab.h>
  17 #include <linux/kthread.h>
  18 #include <linux/rculist_nulls.h>
  19
  20 #include "io-wq.h"
  21
  22 #define WORKER_IDLE_TIMEOUT     (5 * HZ)
  23
  24 enum {
  25         IO_WORKER_F_UP          = 1,    /* up and active */
  26         IO_WORKER_F_RUNNING     = 2,    /* account as running */
  27         IO_WORKER_F_FREE        = 4,    /* worker on free list */
  28         IO_WORKER_F_EXITING     = 8,    /* worker exiting */
  29         IO_WORKER_F_FIXED       = 16,   /* static idle worker */
  30         IO_WORKER_F_BOUND       = 32,   /* is doing bounded work */
  31 };
  32
  33 enum {
  34         IO_WQ_BIT_EXIT          = 0,    /* wq exiting */
  35         IO_WQ_BIT_CANCEL        = 1,    /* cancel work on list */
  36         IO_WQ_BIT_ERROR         = 2,    /* error on setup */
  37 };
  38
  39 enum {
  40         IO_WQE_FLAG_STALLED     = 1,    /* stalled on hash */
  41 };
  42
  43 /*
  44  * One for each thread in a wqe pool
  45  */
  46 struct io_worker {
  47         refcount_t ref;
  48         unsigned flags;
  49         struct hlist_nulls_node nulls_node;
  50         struct list_head all_list;
  51         struct task_struct *task;
  52         wait_queue_head_t wait;
  53         struct io_wqe *wqe;
  54
  55         struct io_wq_work *cur_work;
  56         spinlock_t lock;
  57
  58         struct rcu_head rcu;
  59         struct mm_struct *mm;
  60         const struct cred *creds;
  61         struct files_struct *restore_files;
  62 };
  63
  64 #if BITS_PER_LONG == 64
  65 #define IO_WQ_HASH_ORDER        6
  66 #else
  67 #define IO_WQ_HASH_ORDER        5
  68 #endif
  69
  70 struct io_wqe_acct {
  71         unsigned nr_workers;
  72         unsigned max_workers;
  73         atomic_t nr_running;
  74 };
  75
  76 enum {
  77         IO_WQ_ACCT_BOUND,
  78         IO_WQ_ACCT_UNBOUND,
  79 };
  80
  81 /*
  82  * Per-node worker thread pool
  83  */
  84 struct io_wqe {
  85         struct {
  86                 spinlock_t lock;
  87                 struct io_wq_work_list work_list;
  88                 unsigned long hash_map;
  89                 unsigned flags;
  90         } ____cacheline_aligned_in_smp;
  91
  92         int node;
  93         struct io_wqe_acct acct[2];
  94
  95         struct hlist_nulls_head free_list;
  96         struct hlist_nulls_head busy_list;
  97         struct list_head all_list;
  98
  99         struct io_wq *wq;
 100 };
 101
 102 /*
 103  * Per io_wq state
 104   */
 105 struct io_wq {
 106         struct io_wqe **wqes;
 107         unsigned long state;
 108
 109         get_work_fn *get_work;
 110         put_work_fn *put_work;
 111
 112         struct task_struct *manager;
 113         struct user_struct *user;
 114         const struct cred *creds;
 115         struct mm_struct *mm;
 116         refcount_t refs;
 117         struct completion done;
 118 };
 119
 120 static bool io_worker_get(struct io_worker *worker)
 121 {
 122         return refcount_inc_not_zero(&worker->ref);
 123 }
 124
 125 static void io_worker_release(struct io_worker *worker)
 126 {
 127         if (refcount_dec_and_test(&worker->ref))
 128                 wake_up_process(worker->task);
 129 }
 130
 131 /*
 132  * Note: drops the wqe->lock if returning true! The caller must re-acquire
 133  * the lock in that case. Some callers need to restart handling if this
 134  * happens, so we can't just re-acquire the lock on behalf of the caller.
 135  */
 136 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
 137 {
 138         bool dropped_lock = false;
 139
 140         if (worker->creds) {
 141                 revert_creds(worker->creds);
 142                 worker->creds = NULL;
 143         }
 144
 145         if (current->files != worker->restore_files) {
 146                 __acquire(&wqe->lock);
 147                 spin_unlock_irq(&wqe->lock);
 148                 dropped_lock = true;
 149
 150                 task_lock(current);
 151                 current->files = worker->restore_files;
 152                 task_unlock(current);
 153         }
 154
 155         /*
 156          * If we have an active mm, we need to drop the wq lock before unusing
 157          * it. If we do, return true and let the caller retry the idle loop.
 158          */
 159         if (worker->mm) {
 160                 if (!dropped_lock) {
 161                         __acquire(&wqe->lock);
 162                         spin_unlock_irq(&wqe->lock);
 163                         dropped_lock = true;
 164                 }
 165                 __set_current_state(TASK_RUNNING);
 166                 set_fs(KERNEL_DS);
 167                 unuse_mm(worker->mm);
 168                 mmput(worker->mm);
 169                 worker->mm = NULL;
 170         }
 171
 172         return dropped_lock;
 173 }
 174
 175 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 176                                                    struct io_wq_work *work)
 177 {
 178         if (work->flags & IO_WQ_WORK_UNBOUND)
 179                 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 180
 181         return &wqe->acct[IO_WQ_ACCT_BOUND];
 182 }
 183
 184 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
 185                                                   struct io_worker *worker)
 186 {
 187         if (worker->flags & IO_WORKER_F_BOUND)
 188                 return &wqe->acct[IO_WQ_ACCT_BOUND];
 189
 190         return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 191 }
 192
 193 static void io_worker_exit(struct io_worker *worker)
 194 {
 195         struct io_wqe *wqe = worker->wqe;
 196         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 197         unsigned nr_workers;
 198
 199         /*
 200          * If we're not at zero, someone else is holding a brief reference
 201          * to the worker. Wait for that to go away.
 202          */
 203         set_current_state(TASK_INTERRUPTIBLE);
 204         if (!refcount_dec_and_test(&worker->ref))
 205                 schedule();
 206         __set_current_state(TASK_RUNNING);
 207
 208         preempt_disable();
 209         current->flags &= ~PF_IO_WORKER;
 210         if (worker->flags & IO_WORKER_F_RUNNING)
 211                 atomic_dec(&acct->nr_running);
 212         if (!(worker->flags & IO_WORKER_F_BOUND))
 213                 atomic_dec(&wqe->wq->user->processes);
 214         worker->flags = 0;
 215         preempt_enable();
 216
 217         spin_lock_irq(&wqe->lock);
 218         hlist_nulls_del_rcu(&worker->nulls_node);
 219         list_del_rcu(&worker->all_list);
 220         if (__io_worker_unuse(wqe, worker)) {
 221                 __release(&wqe->lock);
 222                 spin_lock_irq(&wqe->lock);
 223         }
 224         acct->nr_workers--;
 225         nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers +
 226                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers;
 227         spin_unlock_irq(&wqe->lock);
 228
 229         /* all workers gone, wq exit can proceed */
 230         if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs))
 231                 complete(&wqe->wq->done);
 232
 233         kfree_rcu(worker, rcu);
 234 }
 235
 236 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
 237         __must_hold(wqe->lock)
 238 {
 239         if (!wq_list_empty(&wqe->work_list) &&
 240             !(wqe->flags & IO_WQE_FLAG_STALLED))
 241                 return true;
 242         return false;
 243 }
 244
 245 /*
 246  * Check head of free list for an available worker. If one isn't available,
 247  * caller must wake up the wq manager to create one.
 248  */
 249 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
 250         __must_hold(RCU)
 251 {
 252         struct hlist_nulls_node *n;
 253         struct io_worker *worker;
 254
 255         n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
 256         if (is_a_nulls(n))
 257                 return false;
 258
 259         worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
 260         if (io_worker_get(worker)) {
 261                 wake_up(&worker->wait);
 262                 io_worker_release(worker);
 263                 return true;
 264         }
 265
 266         return false;
 267 }
 268
 269 /*
 270  * We need a worker. If we find a free one, we're good. If not, and we're
 271  * below the max number of workers, wake up the manager to create one.
 272  */
 273 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 274 {
 275         bool ret;
 276
 277         /*
 278          * Most likely an attempt to queue unbounded work on an io_wq that
 279          * wasn't setup with any unbounded workers.
 280          */
 281         WARN_ON_ONCE(!acct->max_workers);
 282
 283         rcu_read_lock();
 284         ret = io_wqe_activate_free_worker(wqe);
 285         rcu_read_unlock();
 286
 287         if (!ret && acct->nr_workers < acct->max_workers)
 288                 wake_up_process(wqe->wq->manager);
 289 }
 290
 291 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
 292 {
 293         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 294
 295         atomic_inc(&acct->nr_running);
 296 }
 297
 298 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
 299         __must_hold(wqe->lock)
 300 {
 301         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 302
 303         if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
 304                 io_wqe_wake_worker(wqe, acct);
 305 }
 306
 307 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
 308 {
 309         allow_kernel_signal(SIGINT);
 310
 311         current->flags |= PF_IO_WORKER;
 312
 313         worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 314         worker->restore_files = current->files;
 315         io_wqe_inc_running(wqe, worker);
 316 }
 317
 318 /*
 319  * Worker will start processing some work. Move it to the busy list, if
 320  * it's currently on the freelist
 321  */
 322 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
 323                              struct io_wq_work *work)
 324         __must_hold(wqe->lock)
 325 {
 326         bool worker_bound, work_bound;
 327
 328         if (worker->flags & IO_WORKER_F_FREE) {
 329                 worker->flags &= ~IO_WORKER_F_FREE;
 330                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 331                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->busy_list);
 332         }
 333
 334         /*
 335          * If worker is moving from bound to unbound (or vice versa), then
 336          * ensure we update the running accounting.
 337          */
 338         worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
 339         work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
 340         if (worker_bound != work_bound) {
 341                 io_wqe_dec_running(wqe, worker);
 342                 if (work_bound) {
 343                         worker->flags |= IO_WORKER_F_BOUND;
 344                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
 345                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
 346                         atomic_dec(&wqe->wq->user->processes);
 347                 } else {
 348                         worker->flags &= ~IO_WORKER_F_BOUND;
 349                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
 350                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
 351                         atomic_inc(&wqe->wq->user->processes);
 352                 }
 353                 io_wqe_inc_running(wqe, worker);
 354          }
 355 }
 356
 357 /*
 358  * No work, worker going to sleep. Move to freelist, and unuse mm if we
 359  * have one attached. Dropping the mm may potentially sleep, so we drop
 360  * the lock in that case and return success. Since the caller has to
 361  * retry the loop in that case (we changed task state), we don't regrab
 362  * the lock if we return success.
 363  */
 364 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
 365         __must_hold(wqe->lock)
 366 {
 367         if (!(worker->flags & IO_WORKER_F_FREE)) {
 368                 worker->flags |= IO_WORKER_F_FREE;
 369                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 370                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 371         }
 372
 373         return __io_worker_unuse(wqe, worker);
 374 }
 375
 376 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe, unsigned *hash)
 377         __must_hold(wqe->lock)
 378 {
 379         struct io_wq_work_node *node, *prev;
 380         struct io_wq_work *work;
 381
 382         wq_list_for_each(node, prev, &wqe->work_list) {
 383                 work = container_of(node, struct io_wq_work, list);
 384
 385                 /* not hashed, can run anytime */
 386                 if (!(work->flags & IO_WQ_WORK_HASHED)) {
 387                         wq_node_del(&wqe->work_list, node, prev);
 388                         return work;
 389                 }
 390
 391                 /* hashed, can run if not already running */
 392                 *hash = work->flags >> IO_WQ_HASH_SHIFT;
 393                 if (!(wqe->hash_map & BIT_ULL(*hash))) {
 394                         wqe->hash_map |= BIT_ULL(*hash);
 395                         wq_node_del(&wqe->work_list, node, prev);
 396                         return work;
 397                 }
 398         }
 399
 400         return NULL;
 401 }
 402
 403 static void io_worker_handle_work(struct io_worker *worker)
 404         __releases(wqe->lock)
 405 {
 406         struct io_wq_work *work, *old_work = NULL, *put_work = NULL;
 407         struct io_wqe *wqe = worker->wqe;
 408         struct io_wq *wq = wqe->wq;
 409
 410         do {
 411                 unsigned hash = -1U;
 412
 413                 /*
 414                  * If we got some work, mark us as busy. If we didn't, but
 415                  * the list isn't empty, it means we stalled on hashed work.
 416                  * Mark us stalled so we don't keep looking for work when we
 417                  * can't make progress, any work completion or insertion will
 418                  * clear the stalled flag.
 419                  */
 420                 work = io_get_next_work(wqe, &hash);
 421                 if (work)
 422                         __io_worker_busy(wqe, worker, work);
 423                 else if (!wq_list_empty(&wqe->work_list))
 424                         wqe->flags |= IO_WQE_FLAG_STALLED;
 425
 426                 spin_unlock_irq(&wqe->lock);
 427                 if (put_work && wq->put_work)
 428                         wq->put_work(old_work);
 429                 if (!work)
 430                         break;
 431 next:
 432                 /* flush any pending signals before assigning new work */
 433                 if (signal_pending(current))
 434                         flush_signals(current);
 435
 436                 spin_lock_irq(&worker->lock);
 437                 worker->cur_work = work;
 438                 spin_unlock_irq(&worker->lock);
 439
 440                 if (work->flags & IO_WQ_WORK_CB)
 441                         work->func(&work);
 442
 443                 if ((work->flags & IO_WQ_WORK_NEEDS_FILES) &&
 444                     current->files != work->files) {
 445                         task_lock(current);
 446                         current->files = work->files;
 447                         task_unlock(current);
 448                 }
 449                 if ((work->flags & IO_WQ_WORK_NEEDS_USER) && !worker->mm &&
 450                     wq->mm && mmget_not_zero(wq->mm)) {
 451                         use_mm(wq->mm);
 452                         set_fs(USER_DS);
 453                         worker->mm = wq->mm;
 454                 }
 455                 if (!worker->creds)
 456                         worker->creds = override_creds(wq->creds);
 457                 if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
 458                         work->flags |= IO_WQ_WORK_CANCEL;
 459                 if (worker->mm)
 460                         work->flags |= IO_WQ_WORK_HAS_MM;
 461
 462                 if (wq->get_work && !(work->flags & IO_WQ_WORK_INTERNAL)) {
 463                         put_work = work;
 464                         wq->get_work(work);
 465                 }
 466
 467                 old_work = work;
 468                 work->func(&work);
 469
 470                 spin_lock_irq(&worker->lock);
 471                 worker->cur_work = NULL;
 472                 spin_unlock_irq(&worker->lock);
 473
 474                 spin_lock_irq(&wqe->lock);
 475
 476                 if (hash != -1U) {
 477                         wqe->hash_map &= ~BIT_ULL(hash);
 478                         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 479                 }
 480                 if (work && work != old_work) {
 481                         spin_unlock_irq(&wqe->lock);
 482
 483                         if (put_work && wq->put_work) {
 484                                 wq->put_work(put_work);
 485                                 put_work = NULL;
 486                         }
 487
 488                         /* dependent work not hashed */
 489                         hash = -1U;
 490                         goto next;
 491                 }
 492         } while (1);
 493 }
 494
 495 static int io_wqe_worker(void *data)
 496 {
 497         struct io_worker *worker = data;
 498         struct io_wqe *wqe = worker->wqe;
 499         struct io_wq *wq = wqe->wq;
 500         DEFINE_WAIT(wait);
 501
 502         io_worker_start(wqe, worker);
 503
 504         while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 505                 prepare_to_wait(&worker->wait, &wait, TASK_INTERRUPTIBLE);
 506
 507                 spin_lock_irq(&wqe->lock);
 508                 if (io_wqe_run_queue(wqe)) {
 509                         __set_current_state(TASK_RUNNING);
 510                         io_worker_handle_work(worker);
 511                         continue;
 512                 }
 513                 /* drops the lock on success, retry */
 514                 if (__io_worker_idle(wqe, worker)) {
 515                         __release(&wqe->lock);
 516                         continue;
 517                 }
 518                 spin_unlock_irq(&wqe->lock);
 519                 if (signal_pending(current))
 520                         flush_signals(current);
 521                 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
 522                         continue;
 523                 /* timed out, exit unless we're the fixed worker */
 524                 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 525                     !(worker->flags & IO_WORKER_F_FIXED))
 526                         break;
 527         }
 528
 529         finish_wait(&worker->wait, &wait);
 530
 531         if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 532                 spin_lock_irq(&wqe->lock);
 533                 if (!wq_list_empty(&wqe->work_list))
 534                         io_worker_handle_work(worker);
 535                 else
 536                         spin_unlock_irq(&wqe->lock);
 537         }
 538
 539         io_worker_exit(worker);
 540         return 0;
 541 }
 542
 543 /*
 544  * Called when a worker is scheduled in. Mark us as currently running.
 545  */
 546 void io_wq_worker_running(struct task_struct *tsk)
 547 {
 548         struct io_worker *worker = kthread_data(tsk);
 549         struct io_wqe *wqe = worker->wqe;
 550
 551         if (!(worker->flags & IO_WORKER_F_UP))
 552                 return;
 553         if (worker->flags & IO_WORKER_F_RUNNING)
 554                 return;
 555         worker->flags |= IO_WORKER_F_RUNNING;
 556         io_wqe_inc_running(wqe, worker);
 557 }
 558
 559 /*
 560  * Called when worker is going to sleep. If there are no workers currently
 561  * running and we have work pending, wake up a free one or have the manager
 562  * set one up.
 563  */
 564 void io_wq_worker_sleeping(struct task_struct *tsk)
 565 {
 566         struct io_worker *worker = kthread_data(tsk);
 567         struct io_wqe *wqe = worker->wqe;
 568
 569         if (!(worker->flags & IO_WORKER_F_UP))
 570                 return;
 571         if (!(worker->flags & IO_WORKER_F_RUNNING))
 572                 return;
 573
 574         worker->flags &= ~IO_WORKER_F_RUNNING;
 575
 576         spin_lock_irq(&wqe->lock);
 577         io_wqe_dec_running(wqe, worker);
 578         spin_unlock_irq(&wqe->lock);
 579 }
 580
 581 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
 582 {
 583         struct io_wqe_acct *acct =&wqe->acct[index];
 584         struct io_worker *worker;
 585
 586         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
 587         if (!worker)
 588                 return false;
 589
 590         refcount_set(&worker->ref, 1);
 591         worker->nulls_node.pprev = NULL;
 592         init_waitqueue_head(&worker->wait);
 593         worker->wqe = wqe;
 594         spin_lock_init(&worker->lock);
 595
 596         worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
 597                                 "io_wqe_worker-%d/%d", index, wqe->node);
 598         if (IS_ERR(worker->task)) {
 599                 kfree(worker);
 600                 return false;
 601         }
 602
 603         spin_lock_irq(&wqe->lock);
 604         hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 605         list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 606         worker->flags |= IO_WORKER_F_FREE;
 607         if (index == IO_WQ_ACCT_BOUND)
 608                 worker->flags |= IO_WORKER_F_BOUND;
 609         if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
 610                 worker->flags |= IO_WORKER_F_FIXED;
 611         acct->nr_workers++;
 612         spin_unlock_irq(&wqe->lock);
 613
 614         if (index == IO_WQ_ACCT_UNBOUND)
 615                 atomic_inc(&wq->user->processes);
 616
 617         wake_up_process(worker->task);
 618         return true;
 619 }
 620
 621 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
 622         __must_hold(wqe->lock)
 623 {
 624         struct io_wqe_acct *acct = &wqe->acct[index];
 625
 626         /* if we have available workers or no work, no need */
 627         if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
 628                 return false;
 629         return acct->nr_workers < acct->max_workers;
 630 }
 631
 632 /*
 633  * Manager thread. Tasked with creating new workers, if we need them.
 634  */
 635 static int io_wq_manager(void *data)
 636 {
 637         struct io_wq *wq = data;
 638         int workers_to_create = num_possible_nodes();
 639         int node;
 640
 641         /* create fixed workers */
 642         refcount_set(&wq->refs, workers_to_create);
 643         for_each_node(node) {
 644                 if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
 645                         goto err;
 646                 workers_to_create--;
 647         }
 648
 649         complete(&wq->done);
 650
 651         while (!kthread_should_stop()) {
 652                 for_each_node(node) {
 653                         struct io_wqe *wqe = wq->wqes[node];
 654                         bool fork_worker[2] = { false, false };
 655
 656                         spin_lock_irq(&wqe->lock);
 657                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
 658                                 fork_worker[IO_WQ_ACCT_BOUND] = true;
 659                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
 660                                 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
 661                         spin_unlock_irq(&wqe->lock);
 662                         if (fork_worker[IO_WQ_ACCT_BOUND])
 663                                 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
 664                         if (fork_worker[IO_WQ_ACCT_UNBOUND])
 665                                 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
 666                 }
 667                 set_current_state(TASK_INTERRUPTIBLE);
 668                 schedule_timeout(HZ);
 669         }
 670
 671         return 0;
 672 err:
 673         set_bit(IO_WQ_BIT_ERROR, &wq->state);
 674         set_bit(IO_WQ_BIT_EXIT, &wq->state);
 675         if (refcount_sub_and_test(workers_to_create, &wq->refs))
 676                 complete(&wq->done);
 677         return 0;
 678 }
 679
 680 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
 681                             struct io_wq_work *work)
 682 {
 683         bool free_worker;
 684
 685         if (!(work->flags & IO_WQ_WORK_UNBOUND))
 686                 return true;
 687         if (atomic_read(&acct->nr_running))
 688                 return true;
 689
 690         rcu_read_lock();
 691         free_worker = !hlist_nulls_empty(&wqe->free_list);
 692         rcu_read_unlock();
 693         if (free_worker)
 694                 return true;
 695
 696         if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
 697             !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
 698                 return false;
 699
 700         return true;
 701 }
 702
 703 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 704 {
 705         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 706         unsigned long flags;
 707
 708         /*
 709          * Do early check to see if we need a new unbound worker, and if we do,
 710          * if we're allowed to do so. This isn't 100% accurate as there's a
 711          * gap between this check and incrementing the value, but that's OK.
 712          * It's close enough to not be an issue, fork() has the same delay.
 713          */
 714         if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
 715                 work->flags |= IO_WQ_WORK_CANCEL;
 716                 work->func(&work);
 717                 return;
 718         }
 719
 720         spin_lock_irqsave(&wqe->lock, flags);
 721         wq_list_add_tail(&work->list, &wqe->work_list);
 722         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 723         spin_unlock_irqrestore(&wqe->lock, flags);
 724
 725         if (!atomic_read(&acct->nr_running))
 726                 io_wqe_wake_worker(wqe, acct);
 727 }
 728
 729 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 730 {
 731         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 732
 733         io_wqe_enqueue(wqe, work);
 734 }
 735
 736 /*
 737  * Enqueue work, hashed by some key. Work items that hash to the same value
 738  * will not be done in parallel. Used to limit concurrent writes, generally
 739  * hashed by inode.
 740  */
 741 void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val)
 742 {
 743         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 744         unsigned bit;
 745
 746
 747         bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 748         work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 749         io_wqe_enqueue(wqe, work);
 750 }
 751
 752 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
 753 {
 754         send_sig(SIGINT, worker->task, 1);
 755         return false;
 756 }
 757
 758 /*
 759  * Iterate the passed in list and call the specific function for each
 760  * worker that isn't exiting
 761  */
 762 static bool io_wq_for_each_worker(struct io_wqe *wqe,
 763                                   bool (*func)(struct io_worker *, void *),
 764                                   void *data)
 765 {
 766         struct io_worker *worker;
 767         bool ret = false;
 768
 769         list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
 770                 if (io_worker_get(worker)) {
 771                         ret = func(worker, data);
 772                         io_worker_release(worker);
 773                         if (ret)
 774                                 break;
 775                 }
 776         }
 777
 778         return ret;
 779 }
 780
 781 void io_wq_cancel_all(struct io_wq *wq)
 782 {
 783         int node;
 784
 785         set_bit(IO_WQ_BIT_CANCEL, &wq->state);
 786
 787         /*
 788          * Browse both lists, as there's a gap between handing work off
 789          * to a worker and the worker putting itself on the busy_list
 790          */
 791         rcu_read_lock();
 792         for_each_node(node) {
 793                 struct io_wqe *wqe = wq->wqes[node];
 794
 795                 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
 796         }
 797         rcu_read_unlock();
 798 }
 799
 800 struct io_cb_cancel_data {
 801         struct io_wqe *wqe;
 802         work_cancel_fn *cancel;
 803         void *caller_data;
 804 };
 805
 806 static bool io_work_cancel(struct io_worker *worker, void *cancel_data)
 807 {
 808         struct io_cb_cancel_data *data = cancel_data;
 809         unsigned long flags;
 810         bool ret = false;
 811
 812         /*
 813          * Hold the lock to avoid ->cur_work going out of scope, caller
 814          * may dereference the passed in work.
 815          */
 816         spin_lock_irqsave(&worker->lock, flags);
 817         if (worker->cur_work &&
 818             data->cancel(worker->cur_work, data->caller_data)) {
 819                 send_sig(SIGINT, worker->task, 1);
 820                 ret = true;
 821         }
 822         spin_unlock_irqrestore(&worker->lock, flags);
 823
 824         return ret;
 825 }
 826
 827 static enum io_wq_cancel io_wqe_cancel_cb_work(struct io_wqe *wqe,
 828                                                work_cancel_fn *cancel,
 829                                                void *cancel_data)
 830 {
 831         struct io_cb_cancel_data data = {
 832                 .wqe = wqe,
 833                 .cancel = cancel,
 834                 .caller_data = cancel_data,
 835         };
 836         struct io_wq_work_node *node, *prev;
 837         struct io_wq_work *work;
 838         unsigned long flags;
 839         bool found = false;
 840
 841         spin_lock_irqsave(&wqe->lock, flags);
 842         wq_list_for_each(node, prev, &wqe->work_list) {
 843                 work = container_of(node, struct io_wq_work, list);
 844
 845                 if (cancel(work, cancel_data)) {
 846                         wq_node_del(&wqe->work_list, node, prev);
 847                         found = true;
 848                         break;
 849                 }
 850         }
 851         spin_unlock_irqrestore(&wqe->lock, flags);
 852
 853         if (found) {
 854                 work->flags |= IO_WQ_WORK_CANCEL;
 855                 work->func(&work);
 856                 return IO_WQ_CANCEL_OK;
 857         }
 858
 859         rcu_read_lock();
 860         found = io_wq_for_each_worker(wqe, io_work_cancel, &data);
 861         rcu_read_unlock();
 862         return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND;
 863 }
 864
 865 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
 866                                   void *data)
 867 {
 868         enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
 869         int node;
 870
 871         for_each_node(node) {
 872                 struct io_wqe *wqe = wq->wqes[node];
 873
 874                 ret = io_wqe_cancel_cb_work(wqe, cancel, data);
 875                 if (ret != IO_WQ_CANCEL_NOTFOUND)
 876                         break;
 877         }
 878
 879         return ret;
 880 }
 881
 882 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 883 {
 884         struct io_wq_work *work = data;
 885         unsigned long flags;
 886         bool ret = false;
 887
 888         if (worker->cur_work != work)
 889                 return false;
 890
 891         spin_lock_irqsave(&worker->lock, flags);
 892         if (worker->cur_work == work) {
 893                 send_sig(SIGINT, worker->task, 1);
 894                 ret = true;
 895         }
 896         spin_unlock_irqrestore(&worker->lock, flags);
 897
 898         return ret;
 899 }
 900
 901 static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe,
 902                                             struct io_wq_work *cwork)
 903 {
 904         struct io_wq_work_node *node, *prev;
 905         struct io_wq_work *work;
 906         unsigned long flags;
 907         bool found = false;
 908
 909         cwork->flags |= IO_WQ_WORK_CANCEL;
 910
 911         /*
 912          * First check pending list, if we're lucky we can just remove it
 913          * from there. CANCEL_OK means that the work is returned as-new,
 914          * no completion will be posted for it.
 915          */
 916         spin_lock_irqsave(&wqe->lock, flags);
 917         wq_list_for_each(node, prev, &wqe->work_list) {
 918                 work = container_of(node, struct io_wq_work, list);
 919
 920                 if (work == cwork) {
 921                         wq_node_del(&wqe->work_list, node, prev);
 922                         found = true;
 923                         break;
 924                 }
 925         }
 926         spin_unlock_irqrestore(&wqe->lock, flags);
 927
 928         if (found) {
 929                 work->flags |= IO_WQ_WORK_CANCEL;
 930                 work->func(&work);
 931                 return IO_WQ_CANCEL_OK;
 932         }
 933
 934         /*
 935          * Now check if a free (going busy) or busy worker has the work
 936          * currently running. If we find it there, we'll return CANCEL_RUNNING
 937          * as an indication that we attempte to signal cancellation. The
 938          * completion will run normally in this case.
 939          */
 940         rcu_read_lock();
 941         found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, cwork);
 942         rcu_read_unlock();
 943         return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND;
 944 }
 945
 946 enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork)
 947 {
 948         enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
 949         int node;
 950
 951         for_each_node(node) {
 952                 struct io_wqe *wqe = wq->wqes[node];
 953
 954                 ret = io_wqe_cancel_work(wqe, cwork);
 955                 if (ret != IO_WQ_CANCEL_NOTFOUND)
 956                         break;
 957         }
 958
 959         return ret;
 960 }
 961
 962 struct io_wq_flush_data {
 963         struct io_wq_work work;
 964         struct completion done;
 965 };
 966
 967 static void io_wq_flush_func(struct io_wq_work **workptr)
 968 {
 969         struct io_wq_work *work = *workptr;
 970         struct io_wq_flush_data *data;
 971
 972         data = container_of(work, struct io_wq_flush_data, work);
 973         complete(&data->done);
 974 }
 975
 976 /*
 977  * Doesn't wait for previously queued work to finish. When this completes,
 978  * it just means that previously queued work was started.
 979  */
 980 void io_wq_flush(struct io_wq *wq)
 981 {
 982         struct io_wq_flush_data data;
 983         int node;
 984
 985         for_each_node(node) {
 986                 struct io_wqe *wqe = wq->wqes[node];
 987
 988                 init_completion(&data.done);
 989                 INIT_IO_WORK(&data.work, io_wq_flush_func);
 990                 data.work.flags |= IO_WQ_WORK_INTERNAL;
 991                 io_wqe_enqueue(wqe, &data.work);
 992                 wait_for_completion(&data.done);
 993         }
 994 }
 995
 996 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 997 {
 998         int ret = -ENOMEM, node;
 999         struct io_wq *wq;
1000
1001         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1002         if (!wq)
1003                 return ERR_PTR(-ENOMEM);
1004
1005         wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1006         if (!wq->wqes) {
1007                 kfree(wq);
1008                 return ERR_PTR(-ENOMEM);
1009         }
1010
1011         wq->get_work = data->get_work;
1012         wq->put_work = data->put_work;
1013
1014         /* caller must already hold a reference to this */
1015         wq->user = data->user;
1016         wq->creds = data->creds;
1017
1018         for_each_node(node) {
1019                 struct io_wqe *wqe;
1020
1021                 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, node);
1022                 if (!wqe)
1023                         goto err;
1024                 wq->wqes[node] = wqe;
1025                 wqe->node = node;
1026                 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1027                 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1028                 if (wq->user) {
1029                         wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1030                                         task_rlimit(current, RLIMIT_NPROC);
1031                 }
1032                 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1033                 wqe->node = node;
1034                 wqe->wq = wq;
1035                 spin_lock_init(&wqe->lock);
1036                 INIT_WQ_LIST(&wqe->work_list);
1037                 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1038                 INIT_HLIST_NULLS_HEAD(&wqe->busy_list, 1);
1039                 INIT_LIST_HEAD(&wqe->all_list);
1040         }
1041
1042         init_completion(&wq->done);
1043
1044         /* caller must have already done mmgrab() on this mm */
1045         wq->mm = data->mm;
1046
1047         wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1048         if (!IS_ERR(wq->manager)) {
1049                 wake_up_process(wq->manager);
1050                 wait_for_completion(&wq->done);
1051                 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1052                         ret = -ENOMEM;
1053                         goto err;
1054                 }
1055                 reinit_completion(&wq->done);
1056                 return wq;
1057         }
1058
1059         ret = PTR_ERR(wq->manager);
1060         complete(&wq->done);
1061 err:
1062         for_each_node(node)
1063                 kfree(wq->wqes[node]);
1064         kfree(wq->wqes);
1065         kfree(wq);
1066         return ERR_PTR(ret);
1067 }
1068
1069 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
1070 {
1071         wake_up_process(worker->task);
1072         return false;
1073 }
1074
1075 void io_wq_destroy(struct io_wq *wq)
1076 {
1077         int node;
1078
1079         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1080         if (wq->manager)
1081                 kthread_stop(wq->manager);
1082
1083         rcu_read_lock();
1084         for_each_node(node)
1085                 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1086         rcu_read_unlock();
1087
1088         wait_for_completion(&wq->done);
1089
1090         for_each_node(node)
1091                 kfree(wq->wqes[node]);
1092         kfree(wq->wqes);
1093         kfree(wq);
1094 }