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/sched/mm.h>
  14 #include <linux/percpu.h>
  15 #include <linux/slab.h>
  16 #include <linux/kthread.h>
  17 #include <linux/rculist_nulls.h>
  18 #include <linux/fs_struct.h>
  19 #include <linux/task_work.h>
  20
  21 #include "io-wq.h"
  22
  23 #define WORKER_IDLE_TIMEOUT     (5 * HZ)
  24
  25 enum {
  26         IO_WORKER_F_UP          = 1,    /* up and active */
  27         IO_WORKER_F_RUNNING     = 2,    /* account as running */
  28         IO_WORKER_F_FREE        = 4,    /* worker on free list */
  29         IO_WORKER_F_EXITING     = 8,    /* worker exiting */
  30         IO_WORKER_F_FIXED       = 16,   /* static idle worker */
  31         IO_WORKER_F_BOUND       = 32,   /* is doing bounded work */
  32 };
  33
  34 enum {
  35         IO_WQ_BIT_EXIT          = 0,    /* wq exiting */
  36         IO_WQ_BIT_CANCEL        = 1,    /* cancel work on list */
  37         IO_WQ_BIT_ERROR         = 2,    /* error on setup */
  38 };
  39
  40 enum {
  41         IO_WQE_FLAG_STALLED     = 1,    /* stalled on hash */
  42 };
  43
  44 /*
  45  * One for each thread in a wqe pool
  46  */
  47 struct io_worker {
  48         refcount_t ref;
  49         unsigned flags;
  50         struct hlist_nulls_node nulls_node;
  51         struct list_head all_list;
  52         struct task_struct *task;
  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 *cur_creds;
  61         const struct cred *saved_creds;
  62         struct files_struct *restore_files;
  63         struct fs_struct *restore_fs;
  64 };
  65
  66 #if BITS_PER_LONG == 64
  67 #define IO_WQ_HASH_ORDER        6
  68 #else
  69 #define IO_WQ_HASH_ORDER        5
  70 #endif
  71
  72 #define IO_WQ_NR_HASH_BUCKETS   (1u << IO_WQ_HASH_ORDER)
  73
  74 struct io_wqe_acct {
  75         unsigned nr_workers;
  76         unsigned max_workers;
  77         atomic_t nr_running;
  78 };
  79
  80 enum {
  81         IO_WQ_ACCT_BOUND,
  82         IO_WQ_ACCT_UNBOUND,
  83 };
  84
  85 /*
  86  * Per-node worker thread pool
  87  */
  88 struct io_wqe {
  89         struct {
  90                 spinlock_t lock;
  91                 struct io_wq_work_list work_list;
  92                 unsigned long hash_map;
  93                 unsigned flags;
  94         } ____cacheline_aligned_in_smp;
  95
  96         int node;
  97         struct io_wqe_acct acct[2];
  98
  99         struct hlist_nulls_head free_list;
 100         struct list_head all_list;
 101
 102         struct io_wq *wq;
 103         struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
 104 };
 105
 106 /*
 107  * Per io_wq state
 108   */
 109 struct io_wq {
 110         struct io_wqe **wqes;
 111         unsigned long state;
 112
 113         free_work_fn *free_work;
 114         io_wq_work_fn *do_work;
 115
 116         struct task_struct *manager;
 117         struct user_struct *user;
 118         refcount_t refs;
 119         struct completion done;
 120
 121         refcount_t use_refs;
 122 };
 123
 124 static bool io_worker_get(struct io_worker *worker)
 125 {
 126         return refcount_inc_not_zero(&worker->ref);
 127 }
 128
 129 static void io_worker_release(struct io_worker *worker)
 130 {
 131         if (refcount_dec_and_test(&worker->ref))
 132                 wake_up_process(worker->task);
 133 }
 134
 135 /*
 136  * Note: drops the wqe->lock if returning true! The caller must re-acquire
 137  * the lock in that case. Some callers need to restart handling if this
 138  * happens, so we can't just re-acquire the lock on behalf of the caller.
 139  */
 140 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
 141 {
 142         bool dropped_lock = false;
 143
 144         if (worker->saved_creds) {
 145                 revert_creds(worker->saved_creds);
 146                 worker->cur_creds = worker->saved_creds = NULL;
 147         }
 148
 149         if (current->files != worker->restore_files) {
 150                 __acquire(&wqe->lock);
 151                 spin_unlock_irq(&wqe->lock);
 152                 dropped_lock = true;
 153
 154                 task_lock(current);
 155                 current->files = worker->restore_files;
 156                 task_unlock(current);
 157         }
 158
 159         if (current->fs != worker->restore_fs)
 160                 current->fs = worker->restore_fs;
 161
 162         /*
 163          * If we have an active mm, we need to drop the wq lock before unusing
 164          * it. If we do, return true and let the caller retry the idle loop.
 165          */
 166         if (worker->mm) {
 167                 if (!dropped_lock) {
 168                         __acquire(&wqe->lock);
 169                         spin_unlock_irq(&wqe->lock);
 170                         dropped_lock = true;
 171                 }
 172                 __set_current_state(TASK_RUNNING);
 173                 kthread_unuse_mm(worker->mm);
 174                 mmput(worker->mm);
 175                 worker->mm = NULL;
 176         }
 177
 178         return dropped_lock;
 179 }
 180
 181 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 182                                                    struct io_wq_work *work)
 183 {
 184         if (work->flags & IO_WQ_WORK_UNBOUND)
 185                 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 186
 187         return &wqe->acct[IO_WQ_ACCT_BOUND];
 188 }
 189
 190 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
 191                                                   struct io_worker *worker)
 192 {
 193         if (worker->flags & IO_WORKER_F_BOUND)
 194                 return &wqe->acct[IO_WQ_ACCT_BOUND];
 195
 196         return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 197 }
 198
 199 static void io_worker_exit(struct io_worker *worker)
 200 {
 201         struct io_wqe *wqe = worker->wqe;
 202         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 203         unsigned nr_workers;
 204
 205         /*
 206          * If we're not at zero, someone else is holding a brief reference
 207          * to the worker. Wait for that to go away.
 208          */
 209         set_current_state(TASK_INTERRUPTIBLE);
 210         if (!refcount_dec_and_test(&worker->ref))
 211                 schedule();
 212         __set_current_state(TASK_RUNNING);
 213
 214         preempt_disable();
 215         current->flags &= ~PF_IO_WORKER;
 216         if (worker->flags & IO_WORKER_F_RUNNING)
 217                 atomic_dec(&acct->nr_running);
 218         if (!(worker->flags & IO_WORKER_F_BOUND))
 219                 atomic_dec(&wqe->wq->user->processes);
 220         worker->flags = 0;
 221         preempt_enable();
 222
 223         spin_lock_irq(&wqe->lock);
 224         hlist_nulls_del_rcu(&worker->nulls_node);
 225         list_del_rcu(&worker->all_list);
 226         if (__io_worker_unuse(wqe, worker)) {
 227                 __release(&wqe->lock);
 228                 spin_lock_irq(&wqe->lock);
 229         }
 230         acct->nr_workers--;
 231         nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers +
 232                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers;
 233         spin_unlock_irq(&wqe->lock);
 234
 235         /* all workers gone, wq exit can proceed */
 236         if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs))
 237                 complete(&wqe->wq->done);
 238
 239         kfree_rcu(worker, rcu);
 240 }
 241
 242 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
 243         __must_hold(wqe->lock)
 244 {
 245         if (!wq_list_empty(&wqe->work_list) &&
 246             !(wqe->flags & IO_WQE_FLAG_STALLED))
 247                 return true;
 248         return false;
 249 }
 250
 251 /*
 252  * Check head of free list for an available worker. If one isn't available,
 253  * caller must wake up the wq manager to create one.
 254  */
 255 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
 256         __must_hold(RCU)
 257 {
 258         struct hlist_nulls_node *n;
 259         struct io_worker *worker;
 260
 261         n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
 262         if (is_a_nulls(n))
 263                 return false;
 264
 265         worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
 266         if (io_worker_get(worker)) {
 267                 wake_up_process(worker->task);
 268                 io_worker_release(worker);
 269                 return true;
 270         }
 271
 272         return false;
 273 }
 274
 275 /*
 276  * We need a worker. If we find a free one, we're good. If not, and we're
 277  * below the max number of workers, wake up the manager to create one.
 278  */
 279 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 280 {
 281         bool ret;
 282
 283         /*
 284          * Most likely an attempt to queue unbounded work on an io_wq that
 285          * wasn't setup with any unbounded workers.
 286          */
 287         WARN_ON_ONCE(!acct->max_workers);
 288
 289         rcu_read_lock();
 290         ret = io_wqe_activate_free_worker(wqe);
 291         rcu_read_unlock();
 292
 293         if (!ret && acct->nr_workers < acct->max_workers)
 294                 wake_up_process(wqe->wq->manager);
 295 }
 296
 297 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
 298 {
 299         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 300
 301         atomic_inc(&acct->nr_running);
 302 }
 303
 304 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
 305         __must_hold(wqe->lock)
 306 {
 307         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 308
 309         if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
 310                 io_wqe_wake_worker(wqe, acct);
 311 }
 312
 313 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
 314 {
 315         allow_kernel_signal(SIGINT);
 316
 317         current->flags |= PF_IO_WORKER;
 318
 319         worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 320         worker->restore_files = current->files;
 321         worker->restore_fs = current->fs;
 322         io_wqe_inc_running(wqe, worker);
 323 }
 324
 325 /*
 326  * Worker will start processing some work. Move it to the busy list, if
 327  * it's currently on the freelist
 328  */
 329 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
 330                              struct io_wq_work *work)
 331         __must_hold(wqe->lock)
 332 {
 333         bool worker_bound, work_bound;
 334
 335         if (worker->flags & IO_WORKER_F_FREE) {
 336                 worker->flags &= ~IO_WORKER_F_FREE;
 337                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 338         }
 339
 340         /*
 341          * If worker is moving from bound to unbound (or vice versa), then
 342          * ensure we update the running accounting.
 343          */
 344         worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
 345         work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
 346         if (worker_bound != work_bound) {
 347                 io_wqe_dec_running(wqe, worker);
 348                 if (work_bound) {
 349                         worker->flags |= IO_WORKER_F_BOUND;
 350                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
 351                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
 352                         atomic_dec(&wqe->wq->user->processes);
 353                 } else {
 354                         worker->flags &= ~IO_WORKER_F_BOUND;
 355                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
 356                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
 357                         atomic_inc(&wqe->wq->user->processes);
 358                 }
 359                 io_wqe_inc_running(wqe, worker);
 360          }
 361 }
 362
 363 /*
 364  * No work, worker going to sleep. Move to freelist, and unuse mm if we
 365  * have one attached. Dropping the mm may potentially sleep, so we drop
 366  * the lock in that case and return success. Since the caller has to
 367  * retry the loop in that case (we changed task state), we don't regrab
 368  * the lock if we return success.
 369  */
 370 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
 371         __must_hold(wqe->lock)
 372 {
 373         if (!(worker->flags & IO_WORKER_F_FREE)) {
 374                 worker->flags |= IO_WORKER_F_FREE;
 375                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 376         }
 377
 378         return __io_worker_unuse(wqe, worker);
 379 }
 380
 381 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
 382 {
 383         return work->flags >> IO_WQ_HASH_SHIFT;
 384 }
 385
 386 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
 387         __must_hold(wqe->lock)
 388 {
 389         struct io_wq_work_node *node, *prev;
 390         struct io_wq_work *work, *tail;
 391         unsigned int hash;
 392
 393         wq_list_for_each(node, prev, &wqe->work_list) {
 394                 work = container_of(node, struct io_wq_work, list);
 395
 396                 /* not hashed, can run anytime */
 397                 if (!io_wq_is_hashed(work)) {
 398                         wq_list_del(&wqe->work_list, node, prev);
 399                         return work;
 400                 }
 401
 402                 /* hashed, can run if not already running */
 403                 hash = io_get_work_hash(work);
 404                 if (!(wqe->hash_map & BIT(hash))) {
 405                         wqe->hash_map |= BIT(hash);
 406                         /* all items with this hash lie in [work, tail] */
 407                         tail = wqe->hash_tail[hash];
 408                         wqe->hash_tail[hash] = NULL;
 409                         wq_list_cut(&wqe->work_list, &tail->list, prev);
 410                         return work;
 411                 }
 412         }
 413
 414         return NULL;
 415 }
 416
 417 static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
 418 {
 419         if (worker->mm) {
 420                 kthread_unuse_mm(worker->mm);
 421                 mmput(worker->mm);
 422                 worker->mm = NULL;
 423         }
 424         if (!work->mm)
 425                 return;
 426
 427         if (mmget_not_zero(work->mm)) {
 428                 kthread_use_mm(work->mm);
 429                 worker->mm = work->mm;
 430                 /* hang on to this mm */
 431                 work->mm = NULL;
 432                 return;
 433         }
 434
 435         /* failed grabbing mm, ensure work gets cancelled */
 436         work->flags |= IO_WQ_WORK_CANCEL;
 437 }
 438
 439 static void io_wq_switch_creds(struct io_worker *worker,
 440                                struct io_wq_work *work)
 441 {
 442         const struct cred *old_creds = override_creds(work->creds);
 443
 444         worker->cur_creds = work->creds;
 445         if (worker->saved_creds)
 446                 put_cred(old_creds); /* creds set by previous switch */
 447         else
 448                 worker->saved_creds = old_creds;
 449 }
 450
 451 static void io_impersonate_work(struct io_worker *worker,
 452                                 struct io_wq_work *work)
 453 {
 454         if (work->files && current->files != work->files) {
 455                 task_lock(current);
 456                 current->files = work->files;
 457                 task_unlock(current);
 458         }
 459         if (work->fs && current->fs != work->fs)
 460                 current->fs = work->fs;
 461         if (work->mm != worker->mm)
 462                 io_wq_switch_mm(worker, work);
 463         if (worker->cur_creds != work->creds)
 464                 io_wq_switch_creds(worker, work);
 465         current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->fsize;
 466 }
 467
 468 static void io_assign_current_work(struct io_worker *worker,
 469                                    struct io_wq_work *work)
 470 {
 471         if (work) {
 472                 /* flush pending signals before assigning new work */
 473                 if (signal_pending(current))
 474                         flush_signals(current);
 475                 cond_resched();
 476         }
 477
 478         spin_lock_irq(&worker->lock);
 479         worker->cur_work = work;
 480         spin_unlock_irq(&worker->lock);
 481 }
 482
 483 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
 484
 485 static void io_worker_handle_work(struct io_worker *worker)
 486         __releases(wqe->lock)
 487 {
 488         struct io_wqe *wqe = worker->wqe;
 489         struct io_wq *wq = wqe->wq;
 490
 491         do {
 492                 struct io_wq_work *work;
 493 get_next:
 494                 /*
 495                  * If we got some work, mark us as busy. If we didn't, but
 496                  * the list isn't empty, it means we stalled on hashed work.
 497                  * Mark us stalled so we don't keep looking for work when we
 498                  * can't make progress, any work completion or insertion will
 499                  * clear the stalled flag.
 500                  */
 501                 work = io_get_next_work(wqe);
 502                 if (work)
 503                         __io_worker_busy(wqe, worker, work);
 504                 else if (!wq_list_empty(&wqe->work_list))
 505                         wqe->flags |= IO_WQE_FLAG_STALLED;
 506
 507                 spin_unlock_irq(&wqe->lock);
 508                 if (!work)
 509                         break;
 510                 io_assign_current_work(worker, work);
 511
 512                 /* handle a whole dependent link */
 513                 do {
 514                         struct io_wq_work *old_work, *next_hashed, *linked;
 515                         unsigned int hash = io_get_work_hash(work);
 516
 517                         next_hashed = wq_next_work(work);
 518                         io_impersonate_work(worker, work);
 519                         /*
 520                          * OK to set IO_WQ_WORK_CANCEL even for uncancellable
 521                          * work, the worker function will do the right thing.
 522                          */
 523                         if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
 524                                 work->flags |= IO_WQ_WORK_CANCEL;
 525
 526                         old_work = work;
 527                         linked = wq->do_work(work);
 528
 529                         work = next_hashed;
 530                         if (!work && linked && !io_wq_is_hashed(linked)) {
 531                                 work = linked;
 532                                 linked = NULL;
 533                         }
 534                         io_assign_current_work(worker, work);
 535                         wq->free_work(old_work);
 536
 537                         if (linked)
 538                                 io_wqe_enqueue(wqe, linked);
 539
 540                         if (hash != -1U && !next_hashed) {
 541                                 spin_lock_irq(&wqe->lock);
 542                                 wqe->hash_map &= ~BIT_ULL(hash);
 543                                 wqe->flags &= ~IO_WQE_FLAG_STALLED;
 544                                 /* skip unnecessary unlock-lock wqe->lock */
 545                                 if (!work)
 546                                         goto get_next;
 547                                 spin_unlock_irq(&wqe->lock);
 548                         }
 549                 } while (work);
 550
 551                 spin_lock_irq(&wqe->lock);
 552         } while (1);
 553 }
 554
 555 static int io_wqe_worker(void *data)
 556 {
 557         struct io_worker *worker = data;
 558         struct io_wqe *wqe = worker->wqe;
 559         struct io_wq *wq = wqe->wq;
 560
 561         io_worker_start(wqe, worker);
 562
 563         while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 564                 set_current_state(TASK_INTERRUPTIBLE);
 565 loop:
 566                 spin_lock_irq(&wqe->lock);
 567                 if (io_wqe_run_queue(wqe)) {
 568                         __set_current_state(TASK_RUNNING);
 569                         io_worker_handle_work(worker);
 570                         goto loop;
 571                 }
 572                 /* drops the lock on success, retry */
 573                 if (__io_worker_idle(wqe, worker)) {
 574                         __release(&wqe->lock);
 575                         goto loop;
 576                 }
 577                 spin_unlock_irq(&wqe->lock);
 578                 if (signal_pending(current))
 579                         flush_signals(current);
 580                 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
 581                         continue;
 582                 /* timed out, exit unless we're the fixed worker */
 583                 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 584                     !(worker->flags & IO_WORKER_F_FIXED))
 585                         break;
 586         }
 587
 588         if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 589                 spin_lock_irq(&wqe->lock);
 590                 if (!wq_list_empty(&wqe->work_list))
 591                         io_worker_handle_work(worker);
 592                 else
 593                         spin_unlock_irq(&wqe->lock);
 594         }
 595
 596         io_worker_exit(worker);
 597         return 0;
 598 }
 599
 600 /*
 601  * Called when a worker is scheduled in. Mark us as currently running.
 602  */
 603 void io_wq_worker_running(struct task_struct *tsk)
 604 {
 605         struct io_worker *worker = kthread_data(tsk);
 606         struct io_wqe *wqe = worker->wqe;
 607
 608         if (!(worker->flags & IO_WORKER_F_UP))
 609                 return;
 610         if (worker->flags & IO_WORKER_F_RUNNING)
 611                 return;
 612         worker->flags |= IO_WORKER_F_RUNNING;
 613         io_wqe_inc_running(wqe, worker);
 614 }
 615
 616 /*
 617  * Called when worker is going to sleep. If there are no workers currently
 618  * running and we have work pending, wake up a free one or have the manager
 619  * set one up.
 620  */
 621 void io_wq_worker_sleeping(struct task_struct *tsk)
 622 {
 623         struct io_worker *worker = kthread_data(tsk);
 624         struct io_wqe *wqe = worker->wqe;
 625
 626         if (!(worker->flags & IO_WORKER_F_UP))
 627                 return;
 628         if (!(worker->flags & IO_WORKER_F_RUNNING))
 629                 return;
 630
 631         worker->flags &= ~IO_WORKER_F_RUNNING;
 632
 633         spin_lock_irq(&wqe->lock);
 634         io_wqe_dec_running(wqe, worker);
 635         spin_unlock_irq(&wqe->lock);
 636 }
 637
 638 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
 639 {
 640         struct io_wqe_acct *acct =&wqe->acct[index];
 641         struct io_worker *worker;
 642
 643         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
 644         if (!worker)
 645                 return false;
 646
 647         refcount_set(&worker->ref, 1);
 648         worker->nulls_node.pprev = NULL;
 649         worker->wqe = wqe;
 650         spin_lock_init(&worker->lock);
 651
 652         worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
 653                                 "io_wqe_worker-%d/%d", index, wqe->node);
 654         if (IS_ERR(worker->task)) {
 655                 kfree(worker);
 656                 return false;
 657         }
 658
 659         spin_lock_irq(&wqe->lock);
 660         hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 661         list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 662         worker->flags |= IO_WORKER_F_FREE;
 663         if (index == IO_WQ_ACCT_BOUND)
 664                 worker->flags |= IO_WORKER_F_BOUND;
 665         if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
 666                 worker->flags |= IO_WORKER_F_FIXED;
 667         acct->nr_workers++;
 668         spin_unlock_irq(&wqe->lock);
 669
 670         if (index == IO_WQ_ACCT_UNBOUND)
 671                 atomic_inc(&wq->user->processes);
 672
 673         wake_up_process(worker->task);
 674         return true;
 675 }
 676
 677 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
 678         __must_hold(wqe->lock)
 679 {
 680         struct io_wqe_acct *acct = &wqe->acct[index];
 681
 682         /* if we have available workers or no work, no need */
 683         if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
 684                 return false;
 685         return acct->nr_workers < acct->max_workers;
 686 }
 687
 688 /*
 689  * Manager thread. Tasked with creating new workers, if we need them.
 690  */
 691 static int io_wq_manager(void *data)
 692 {
 693         struct io_wq *wq = data;
 694         int workers_to_create = num_possible_nodes();
 695         int node;
 696
 697         /* create fixed workers */
 698         refcount_set(&wq->refs, workers_to_create);
 699         for_each_node(node) {
 700                 if (!node_online(node))
 701                         continue;
 702                 if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
 703                         goto err;
 704                 workers_to_create--;
 705         }
 706
 707         while (workers_to_create--)
 708                 refcount_dec(&wq->refs);
 709
 710         complete(&wq->done);
 711
 712         while (!kthread_should_stop()) {
 713                 if (current->task_works)
 714                         task_work_run();
 715
 716                 for_each_node(node) {
 717                         struct io_wqe *wqe = wq->wqes[node];
 718                         bool fork_worker[2] = { false, false };
 719
 720                         if (!node_online(node))
 721                                 continue;
 722
 723                         spin_lock_irq(&wqe->lock);
 724                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
 725                                 fork_worker[IO_WQ_ACCT_BOUND] = true;
 726                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
 727                                 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
 728                         spin_unlock_irq(&wqe->lock);
 729                         if (fork_worker[IO_WQ_ACCT_BOUND])
 730                                 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
 731                         if (fork_worker[IO_WQ_ACCT_UNBOUND])
 732                                 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
 733                 }
 734                 set_current_state(TASK_INTERRUPTIBLE);
 735                 schedule_timeout(HZ);
 736         }
 737
 738         if (current->task_works)
 739                 task_work_run();
 740
 741         return 0;
 742 err:
 743         set_bit(IO_WQ_BIT_ERROR, &wq->state);
 744         set_bit(IO_WQ_BIT_EXIT, &wq->state);
 745         if (refcount_sub_and_test(workers_to_create, &wq->refs))
 746                 complete(&wq->done);
 747         return 0;
 748 }
 749
 750 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
 751                             struct io_wq_work *work)
 752 {
 753         bool free_worker;
 754
 755         if (!(work->flags & IO_WQ_WORK_UNBOUND))
 756                 return true;
 757         if (atomic_read(&acct->nr_running))
 758                 return true;
 759
 760         rcu_read_lock();
 761         free_worker = !hlist_nulls_empty(&wqe->free_list);
 762         rcu_read_unlock();
 763         if (free_worker)
 764                 return true;
 765
 766         if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
 767             !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
 768                 return false;
 769
 770         return true;
 771 }
 772
 773 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
 774 {
 775         struct io_wq *wq = wqe->wq;
 776
 777         do {
 778                 struct io_wq_work *old_work = work;
 779
 780                 work->flags |= IO_WQ_WORK_CANCEL;
 781                 work = wq->do_work(work);
 782                 wq->free_work(old_work);
 783         } while (work);
 784 }
 785
 786 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
 787 {
 788         unsigned int hash;
 789         struct io_wq_work *tail;
 790
 791         if (!io_wq_is_hashed(work)) {
 792 append:
 793                 wq_list_add_tail(&work->list, &wqe->work_list);
 794                 return;
 795         }
 796
 797         hash = io_get_work_hash(work);
 798         tail = wqe->hash_tail[hash];
 799         wqe->hash_tail[hash] = work;
 800         if (!tail)
 801                 goto append;
 802
 803         wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
 804 }
 805
 806 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 807 {
 808         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 809         int work_flags;
 810         unsigned long flags;
 811
 812         /*
 813          * Do early check to see if we need a new unbound worker, and if we do,
 814          * if we're allowed to do so. This isn't 100% accurate as there's a
 815          * gap between this check and incrementing the value, but that's OK.
 816          * It's close enough to not be an issue, fork() has the same delay.
 817          */
 818         if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
 819                 io_run_cancel(work, wqe);
 820                 return;
 821         }
 822
 823         work_flags = work->flags;
 824         spin_lock_irqsave(&wqe->lock, flags);
 825         io_wqe_insert_work(wqe, work);
 826         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 827         spin_unlock_irqrestore(&wqe->lock, flags);
 828
 829         if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 830             !atomic_read(&acct->nr_running))
 831                 io_wqe_wake_worker(wqe, acct);
 832 }
 833
 834 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 835 {
 836         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 837
 838         io_wqe_enqueue(wqe, work);
 839 }
 840
 841 /*
 842  * Work items that hash to the same value will not be done in parallel.
 843  * Used to limit concurrent writes, generally hashed by inode.
 844  */
 845 void io_wq_hash_work(struct io_wq_work *work, void *val)
 846 {
 847         unsigned int bit;
 848
 849         bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 850         work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 851 }
 852
 853 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
 854 {
 855         send_sig(SIGINT, worker->task, 1);
 856         return false;
 857 }
 858
 859 /*
 860  * Iterate the passed in list and call the specific function for each
 861  * worker that isn't exiting
 862  */
 863 static bool io_wq_for_each_worker(struct io_wqe *wqe,
 864                                   bool (*func)(struct io_worker *, void *),
 865                                   void *data)
 866 {
 867         struct io_worker *worker;
 868         bool ret = false;
 869
 870         list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
 871                 if (io_worker_get(worker)) {
 872                         /* no task if node is/was offline */
 873                         if (worker->task)
 874                                 ret = func(worker, data);
 875                         io_worker_release(worker);
 876                         if (ret)
 877                                 break;
 878                 }
 879         }
 880
 881         return ret;
 882 }
 883
 884 void io_wq_cancel_all(struct io_wq *wq)
 885 {
 886         int node;
 887
 888         set_bit(IO_WQ_BIT_CANCEL, &wq->state);
 889
 890         rcu_read_lock();
 891         for_each_node(node) {
 892                 struct io_wqe *wqe = wq->wqes[node];
 893
 894                 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
 895         }
 896         rcu_read_unlock();
 897 }
 898
 899 struct io_cb_cancel_data {
 900         work_cancel_fn *fn;
 901         void *data;
 902         int nr_running;
 903         int nr_pending;
 904         bool cancel_all;
 905 };
 906
 907 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 908 {
 909         struct io_cb_cancel_data *match = data;
 910         unsigned long flags;
 911
 912         /*
 913          * Hold the lock to avoid ->cur_work going out of scope, caller
 914          * may dereference the passed in work.
 915          */
 916         spin_lock_irqsave(&worker->lock, flags);
 917         if (worker->cur_work &&
 918             !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) &&
 919             match->fn(worker->cur_work, match->data)) {
 920                 send_sig(SIGINT, worker->task, 1);
 921                 match->nr_running++;
 922         }
 923         spin_unlock_irqrestore(&worker->lock, flags);
 924
 925         return match->nr_running && !match->cancel_all;
 926 }
 927
 928 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
 929                                          struct io_wq_work *work,
 930                                          struct io_wq_work_node *prev)
 931 {
 932         unsigned int hash = io_get_work_hash(work);
 933         struct io_wq_work *prev_work = NULL;
 934
 935         if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
 936                 if (prev)
 937                         prev_work = container_of(prev, struct io_wq_work, list);
 938                 if (prev_work && io_get_work_hash(prev_work) == hash)
 939                         wqe->hash_tail[hash] = prev_work;
 940                 else
 941                         wqe->hash_tail[hash] = NULL;
 942         }
 943         wq_list_del(&wqe->work_list, &work->list, prev);
 944 }
 945
 946 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
 947                                        struct io_cb_cancel_data *match)
 948 {
 949         struct io_wq_work_node *node, *prev;
 950         struct io_wq_work *work;
 951         unsigned long flags;
 952
 953 retry:
 954         spin_lock_irqsave(&wqe->lock, flags);
 955         wq_list_for_each(node, prev, &wqe->work_list) {
 956                 work = container_of(node, struct io_wq_work, list);
 957                 if (!match->fn(work, match->data))
 958                         continue;
 959                 io_wqe_remove_pending(wqe, work, prev);
 960                 spin_unlock_irqrestore(&wqe->lock, flags);
 961                 io_run_cancel(work, wqe);
 962                 match->nr_pending++;
 963                 if (!match->cancel_all)
 964                         return;
 965
 966                 /* not safe to continue after unlock */
 967                 goto retry;
 968         }
 969         spin_unlock_irqrestore(&wqe->lock, flags);
 970 }
 971
 972 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
 973                                        struct io_cb_cancel_data *match)
 974 {
 975         rcu_read_lock();
 976         io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
 977         rcu_read_unlock();
 978 }
 979
 980 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
 981                                   void *data, bool cancel_all)
 982 {
 983         struct io_cb_cancel_data match = {
 984                 .fn             = cancel,
 985                 .data           = data,
 986                 .cancel_all     = cancel_all,
 987         };
 988         int node;
 989
 990         /*
 991          * First check pending list, if we're lucky we can just remove it
 992          * from there. CANCEL_OK means that the work is returned as-new,
 993          * no completion will be posted for it.
 994          */
 995         for_each_node(node) {
 996                 struct io_wqe *wqe = wq->wqes[node];
 997
 998                 io_wqe_cancel_pending_work(wqe, &match);
 999                 if (match.nr_pending && !match.cancel_all)
1000                         return IO_WQ_CANCEL_OK;
1001         }
1002
1003         /*
1004          * Now check if a free (going busy) or busy worker has the work
1005          * currently running. If we find it there, we'll return CANCEL_RUNNING
1006          * as an indication that we attempt to signal cancellation. The
1007          * completion will run normally in this case.
1008          */
1009         for_each_node(node) {
1010                 struct io_wqe *wqe = wq->wqes[node];
1011
1012                 io_wqe_cancel_running_work(wqe, &match);
1013                 if (match.nr_running && !match.cancel_all)
1014                         return IO_WQ_CANCEL_RUNNING;
1015         }
1016
1017         if (match.nr_running)
1018                 return IO_WQ_CANCEL_RUNNING;
1019         if (match.nr_pending)
1020                 return IO_WQ_CANCEL_OK;
1021         return IO_WQ_CANCEL_NOTFOUND;
1022 }
1023
1024 static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data)
1025 {
1026         return work == data;
1027 }
1028
1029 enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork)
1030 {
1031         return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork, false);
1032 }
1033
1034 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1035 {
1036         int ret = -ENOMEM, node;
1037         struct io_wq *wq;
1038
1039         if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1040                 return ERR_PTR(-EINVAL);
1041
1042         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1043         if (!wq)
1044                 return ERR_PTR(-ENOMEM);
1045
1046         wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1047         if (!wq->wqes) {
1048                 kfree(wq);
1049                 return ERR_PTR(-ENOMEM);
1050         }
1051
1052         wq->free_work = data->free_work;
1053         wq->do_work = data->do_work;
1054
1055         /* caller must already hold a reference to this */
1056         wq->user = data->user;
1057
1058         for_each_node(node) {
1059                 struct io_wqe *wqe;
1060                 int alloc_node = node;
1061
1062                 if (!node_online(alloc_node))
1063                         alloc_node = NUMA_NO_NODE;
1064                 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1065                 if (!wqe)
1066                         goto err;
1067                 wq->wqes[node] = wqe;
1068                 wqe->node = alloc_node;
1069                 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1070                 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1071                 if (wq->user) {
1072                         wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1073                                         task_rlimit(current, RLIMIT_NPROC);
1074                 }
1075                 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1076                 wqe->wq = wq;
1077                 spin_lock_init(&wqe->lock);
1078                 INIT_WQ_LIST(&wqe->work_list);
1079                 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1080                 INIT_LIST_HEAD(&wqe->all_list);
1081         }
1082
1083         init_completion(&wq->done);
1084
1085         wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1086         if (!IS_ERR(wq->manager)) {
1087                 wake_up_process(wq->manager);
1088                 wait_for_completion(&wq->done);
1089                 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1090                         ret = -ENOMEM;
1091                         goto err;
1092                 }
1093                 refcount_set(&wq->use_refs, 1);
1094                 reinit_completion(&wq->done);
1095                 return wq;
1096         }
1097
1098         ret = PTR_ERR(wq->manager);
1099         complete(&wq->done);
1100 err:
1101         for_each_node(node)
1102                 kfree(wq->wqes[node]);
1103         kfree(wq->wqes);
1104         kfree(wq);
1105         return ERR_PTR(ret);
1106 }
1107
1108 bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1109 {
1110         if (data->free_work != wq->free_work || data->do_work != wq->do_work)
1111                 return false;
1112
1113         return refcount_inc_not_zero(&wq->use_refs);
1114 }
1115
1116 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
1117 {
1118         wake_up_process(worker->task);
1119         return false;
1120 }
1121
1122 static void __io_wq_destroy(struct io_wq *wq)
1123 {
1124         int node;
1125
1126         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1127         if (wq->manager)
1128                 kthread_stop(wq->manager);
1129
1130         rcu_read_lock();
1131         for_each_node(node)
1132                 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1133         rcu_read_unlock();
1134
1135         wait_for_completion(&wq->done);
1136
1137         for_each_node(node)
1138                 kfree(wq->wqes[node]);
1139         kfree(wq->wqes);
1140         kfree(wq);
1141 }
1142
1143 void io_wq_destroy(struct io_wq *wq)
1144 {
1145         if (refcount_dec_and_test(&wq->use_refs))
1146                 __io_wq_destroy(wq);
1147 }
1148
1149 struct task_struct *io_wq_get_task(struct io_wq *wq)
1150 {
1151         return wq->manager;
1152 }