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 #include <linux/fs_struct.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 struct io_wqe_acct {
  73         unsigned nr_workers;
  74         unsigned max_workers;
  75         atomic_t nr_running;
  76 };
  77
  78 enum {
  79         IO_WQ_ACCT_BOUND,
  80         IO_WQ_ACCT_UNBOUND,
  81 };
  82
  83 /*
  84  * Per-node worker thread pool
  85  */
  86 struct io_wqe {
  87         struct {
  88                 spinlock_t lock;
  89                 struct io_wq_work_list work_list;
  90                 unsigned long hash_map;
  91                 unsigned flags;
  92         } ____cacheline_aligned_in_smp;
  93
  94         int node;
  95         struct io_wqe_acct acct[2];
  96
  97         struct hlist_nulls_head free_list;
  98         struct list_head all_list;
  99
 100         struct io_wq *wq;
 101 };
 102
 103 /*
 104  * Per io_wq state
 105   */
 106 struct io_wq {
 107         struct io_wqe **wqes;
 108         unsigned long state;
 109
 110         get_work_fn *get_work;
 111         put_work_fn *put_work;
 112
 113         struct task_struct *manager;
 114         struct user_struct *user;
 115         refcount_t refs;
 116         struct completion done;
 117
 118         refcount_t use_refs;
 119 };
 120
 121 static bool io_worker_get(struct io_worker *worker)
 122 {
 123         return refcount_inc_not_zero(&worker->ref);
 124 }
 125
 126 static void io_worker_release(struct io_worker *worker)
 127 {
 128         if (refcount_dec_and_test(&worker->ref))
 129                 wake_up_process(worker->task);
 130 }
 131
 132 /*
 133  * Note: drops the wqe->lock if returning true! The caller must re-acquire
 134  * the lock in that case. Some callers need to restart handling if this
 135  * happens, so we can't just re-acquire the lock on behalf of the caller.
 136  */
 137 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
 138 {
 139         bool dropped_lock = false;
 140
 141         if (worker->saved_creds) {
 142                 revert_creds(worker->saved_creds);
 143                 worker->cur_creds = worker->saved_creds = NULL;
 144         }
 145
 146         if (current->files != worker->restore_files) {
 147                 __acquire(&wqe->lock);
 148                 spin_unlock_irq(&wqe->lock);
 149                 dropped_lock = true;
 150
 151                 task_lock(current);
 152                 current->files = worker->restore_files;
 153                 task_unlock(current);
 154         }
 155
 156         if (current->fs != worker->restore_fs)
 157                 current->fs = worker->restore_fs;
 158
 159         /*
 160          * If we have an active mm, we need to drop the wq lock before unusing
 161          * it. If we do, return true and let the caller retry the idle loop.
 162          */
 163         if (worker->mm) {
 164                 if (!dropped_lock) {
 165                         __acquire(&wqe->lock);
 166                         spin_unlock_irq(&wqe->lock);
 167                         dropped_lock = true;
 168                 }
 169                 __set_current_state(TASK_RUNNING);
 170                 set_fs(KERNEL_DS);
 171                 unuse_mm(worker->mm);
 172                 mmput(worker->mm);
 173                 worker->mm = NULL;
 174         }
 175
 176         return dropped_lock;
 177 }
 178
 179 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 180                                                    struct io_wq_work *work)
 181 {
 182         if (work->flags & IO_WQ_WORK_UNBOUND)
 183                 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 184
 185         return &wqe->acct[IO_WQ_ACCT_BOUND];
 186 }
 187
 188 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
 189                                                   struct io_worker *worker)
 190 {
 191         if (worker->flags & IO_WORKER_F_BOUND)
 192                 return &wqe->acct[IO_WQ_ACCT_BOUND];
 193
 194         return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 195 }
 196
 197 static void io_worker_exit(struct io_worker *worker)
 198 {
 199         struct io_wqe *wqe = worker->wqe;
 200         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 201         unsigned nr_workers;
 202
 203         /*
 204          * If we're not at zero, someone else is holding a brief reference
 205          * to the worker. Wait for that to go away.
 206          */
 207         set_current_state(TASK_INTERRUPTIBLE);
 208         if (!refcount_dec_and_test(&worker->ref))
 209                 schedule();
 210         __set_current_state(TASK_RUNNING);
 211
 212         preempt_disable();
 213         current->flags &= ~PF_IO_WORKER;
 214         if (worker->flags & IO_WORKER_F_RUNNING)
 215                 atomic_dec(&acct->nr_running);
 216         if (!(worker->flags & IO_WORKER_F_BOUND))
 217                 atomic_dec(&wqe->wq->user->processes);
 218         worker->flags = 0;
 219         preempt_enable();
 220
 221         spin_lock_irq(&wqe->lock);
 222         hlist_nulls_del_rcu(&worker->nulls_node);
 223         list_del_rcu(&worker->all_list);
 224         if (__io_worker_unuse(wqe, worker)) {
 225                 __release(&wqe->lock);
 226                 spin_lock_irq(&wqe->lock);
 227         }
 228         acct->nr_workers--;
 229         nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers +
 230                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers;
 231         spin_unlock_irq(&wqe->lock);
 232
 233         /* all workers gone, wq exit can proceed */
 234         if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs))
 235                 complete(&wqe->wq->done);
 236
 237         kfree_rcu(worker, rcu);
 238 }
 239
 240 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
 241         __must_hold(wqe->lock)
 242 {
 243         if (!wq_list_empty(&wqe->work_list) &&
 244             !(wqe->flags & IO_WQE_FLAG_STALLED))
 245                 return true;
 246         return false;
 247 }
 248
 249 /*
 250  * Check head of free list for an available worker. If one isn't available,
 251  * caller must wake up the wq manager to create one.
 252  */
 253 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
 254         __must_hold(RCU)
 255 {
 256         struct hlist_nulls_node *n;
 257         struct io_worker *worker;
 258
 259         n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
 260         if (is_a_nulls(n))
 261                 return false;
 262
 263         worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
 264         if (io_worker_get(worker)) {
 265                 wake_up_process(worker->task);
 266                 io_worker_release(worker);
 267                 return true;
 268         }
 269
 270         return false;
 271 }
 272
 273 /*
 274  * We need a worker. If we find a free one, we're good. If not, and we're
 275  * below the max number of workers, wake up the manager to create one.
 276  */
 277 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 278 {
 279         bool ret;
 280
 281         /*
 282          * Most likely an attempt to queue unbounded work on an io_wq that
 283          * wasn't setup with any unbounded workers.
 284          */
 285         WARN_ON_ONCE(!acct->max_workers);
 286
 287         rcu_read_lock();
 288         ret = io_wqe_activate_free_worker(wqe);
 289         rcu_read_unlock();
 290
 291         if (!ret && acct->nr_workers < acct->max_workers)
 292                 wake_up_process(wqe->wq->manager);
 293 }
 294
 295 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
 296 {
 297         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 298
 299         atomic_inc(&acct->nr_running);
 300 }
 301
 302 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
 303         __must_hold(wqe->lock)
 304 {
 305         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 306
 307         if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
 308                 io_wqe_wake_worker(wqe, acct);
 309 }
 310
 311 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
 312 {
 313         allow_kernel_signal(SIGINT);
 314
 315         current->flags |= PF_IO_WORKER;
 316
 317         worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 318         worker->restore_files = current->files;
 319         worker->restore_fs = current->fs;
 320         io_wqe_inc_running(wqe, worker);
 321 }
 322
 323 /*
 324  * Worker will start processing some work. Move it to the busy list, if
 325  * it's currently on the freelist
 326  */
 327 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
 328                              struct io_wq_work *work)
 329         __must_hold(wqe->lock)
 330 {
 331         bool worker_bound, work_bound;
 332
 333         if (worker->flags & IO_WORKER_F_FREE) {
 334                 worker->flags &= ~IO_WORKER_F_FREE;
 335                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 336         }
 337
 338         /*
 339          * If worker is moving from bound to unbound (or vice versa), then
 340          * ensure we update the running accounting.
 341          */
 342         worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
 343         work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
 344         if (worker_bound != work_bound) {
 345                 io_wqe_dec_running(wqe, worker);
 346                 if (work_bound) {
 347                         worker->flags |= IO_WORKER_F_BOUND;
 348                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
 349                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
 350                         atomic_dec(&wqe->wq->user->processes);
 351                 } else {
 352                         worker->flags &= ~IO_WORKER_F_BOUND;
 353                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
 354                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
 355                         atomic_inc(&wqe->wq->user->processes);
 356                 }
 357                 io_wqe_inc_running(wqe, worker);
 358          }
 359 }
 360
 361 /*
 362  * No work, worker going to sleep. Move to freelist, and unuse mm if we
 363  * have one attached. Dropping the mm may potentially sleep, so we drop
 364  * the lock in that case and return success. Since the caller has to
 365  * retry the loop in that case (we changed task state), we don't regrab
 366  * the lock if we return success.
 367  */
 368 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
 369         __must_hold(wqe->lock)
 370 {
 371         if (!(worker->flags & IO_WORKER_F_FREE)) {
 372                 worker->flags |= IO_WORKER_F_FREE;
 373                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 374         }
 375
 376         return __io_worker_unuse(wqe, worker);
 377 }
 378
 379 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe, unsigned *hash)
 380         __must_hold(wqe->lock)
 381 {
 382         struct io_wq_work_node *node, *prev;
 383         struct io_wq_work *work;
 384
 385         wq_list_for_each(node, prev, &wqe->work_list) {
 386                 work = container_of(node, struct io_wq_work, list);
 387
 388                 /* not hashed, can run anytime */
 389                 if (!(work->flags & IO_WQ_WORK_HASHED)) {
 390                         wq_node_del(&wqe->work_list, node, prev);
 391                         return work;
 392                 }
 393
 394                 /* hashed, can run if not already running */
 395                 *hash = work->flags >> IO_WQ_HASH_SHIFT;
 396                 if (!(wqe->hash_map & BIT_ULL(*hash))) {
 397                         wqe->hash_map |= BIT_ULL(*hash);
 398                         wq_node_del(&wqe->work_list, node, prev);
 399                         return work;
 400                 }
 401         }
 402
 403         return NULL;
 404 }
 405
 406 static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
 407 {
 408         if (worker->mm) {
 409                 unuse_mm(worker->mm);
 410                 mmput(worker->mm);
 411                 worker->mm = NULL;
 412         }
 413         if (!work->mm) {
 414                 set_fs(KERNEL_DS);
 415                 return;
 416         }
 417         if (mmget_not_zero(work->mm)) {
 418                 use_mm(work->mm);
 419                 if (!worker->mm)
 420                         set_fs(USER_DS);
 421                 worker->mm = work->mm;
 422                 /* hang on to this mm */
 423                 work->mm = NULL;
 424                 return;
 425         }
 426
 427         /* failed grabbing mm, ensure work gets cancelled */
 428         work->flags |= IO_WQ_WORK_CANCEL;
 429 }
 430
 431 static void io_wq_switch_creds(struct io_worker *worker,
 432                                struct io_wq_work *work)
 433 {
 434         const struct cred *old_creds = override_creds(work->creds);
 435
 436         worker->cur_creds = work->creds;
 437         if (worker->saved_creds)
 438                 put_cred(old_creds); /* creds set by previous switch */
 439         else
 440                 worker->saved_creds = old_creds;
 441 }
 442
 443 static void io_worker_handle_work(struct io_worker *worker)
 444         __releases(wqe->lock)
 445 {
 446         struct io_wq_work *work, *old_work = NULL, *put_work = NULL;
 447         struct io_wqe *wqe = worker->wqe;
 448         struct io_wq *wq = wqe->wq;
 449
 450         do {
 451                 unsigned hash = -1U;
 452
 453                 /*
 454                  * If we got some work, mark us as busy. If we didn't, but
 455                  * the list isn't empty, it means we stalled on hashed work.
 456                  * Mark us stalled so we don't keep looking for work when we
 457                  * can't make progress, any work completion or insertion will
 458                  * clear the stalled flag.
 459                  */
 460                 work = io_get_next_work(wqe, &hash);
 461                 if (work)
 462                         __io_worker_busy(wqe, worker, work);
 463                 else if (!wq_list_empty(&wqe->work_list))
 464                         wqe->flags |= IO_WQE_FLAG_STALLED;
 465
 466                 spin_unlock_irq(&wqe->lock);
 467                 if (put_work && wq->put_work)
 468                         wq->put_work(old_work);
 469                 if (!work)
 470                         break;
 471 next:
 472                 /* flush any pending signals before assigning new work */
 473                 if (signal_pending(current))
 474                         flush_signals(current);
 475
 476                 cond_resched();
 477
 478                 spin_lock_irq(&worker->lock);
 479                 worker->cur_work = work;
 480                 spin_unlock_irq(&worker->lock);
 481
 482                 if (work->flags & IO_WQ_WORK_CB)
 483                         work->func(&work);
 484
 485                 if (work->files && current->files != work->files) {
 486                         task_lock(current);
 487                         current->files = work->files;
 488                         task_unlock(current);
 489                 }
 490                 if (work->fs && current->fs != work->fs)
 491                         current->fs = work->fs;
 492                 if (work->mm != worker->mm)
 493                         io_wq_switch_mm(worker, work);
 494                 if (worker->cur_creds != work->creds)
 495                         io_wq_switch_creds(worker, work);
 496                 /*
 497                  * OK to set IO_WQ_WORK_CANCEL even for uncancellable work,
 498                  * the worker function will do the right thing.
 499                  */
 500                 if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
 501                         work->flags |= IO_WQ_WORK_CANCEL;
 502                 if (worker->mm)
 503                         work->flags |= IO_WQ_WORK_HAS_MM;
 504
 505                 if (wq->get_work && !(work->flags & IO_WQ_WORK_INTERNAL)) {
 506                         put_work = work;
 507                         wq->get_work(work);
 508                 }
 509
 510                 old_work = work;
 511                 work->func(&work);
 512
 513                 spin_lock_irq(&worker->lock);
 514                 worker->cur_work = NULL;
 515                 spin_unlock_irq(&worker->lock);
 516
 517                 spin_lock_irq(&wqe->lock);
 518
 519                 if (hash != -1U) {
 520                         wqe->hash_map &= ~BIT_ULL(hash);
 521                         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 522                 }
 523                 if (work && work != old_work) {
 524                         spin_unlock_irq(&wqe->lock);
 525
 526                         if (put_work && wq->put_work) {
 527                                 wq->put_work(put_work);
 528                                 put_work = NULL;
 529                         }
 530
 531                         /* dependent work not hashed */
 532                         hash = -1U;
 533                         goto next;
 534                 }
 535         } while (1);
 536 }
 537
 538 static int io_wqe_worker(void *data)
 539 {
 540         struct io_worker *worker = data;
 541         struct io_wqe *wqe = worker->wqe;
 542         struct io_wq *wq = wqe->wq;
 543
 544         io_worker_start(wqe, worker);
 545
 546         while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 547                 set_current_state(TASK_INTERRUPTIBLE);
 548 loop:
 549                 spin_lock_irq(&wqe->lock);
 550                 if (io_wqe_run_queue(wqe)) {
 551                         __set_current_state(TASK_RUNNING);
 552                         io_worker_handle_work(worker);
 553                         goto loop;
 554                 }
 555                 /* drops the lock on success, retry */
 556                 if (__io_worker_idle(wqe, worker)) {
 557                         __release(&wqe->lock);
 558                         goto loop;
 559                 }
 560                 spin_unlock_irq(&wqe->lock);
 561                 if (signal_pending(current))
 562                         flush_signals(current);
 563                 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
 564                         continue;
 565                 /* timed out, exit unless we're the fixed worker */
 566                 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 567                     !(worker->flags & IO_WORKER_F_FIXED))
 568                         break;
 569         }
 570
 571         if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 572                 spin_lock_irq(&wqe->lock);
 573                 if (!wq_list_empty(&wqe->work_list))
 574                         io_worker_handle_work(worker);
 575                 else
 576                         spin_unlock_irq(&wqe->lock);
 577         }
 578
 579         io_worker_exit(worker);
 580         return 0;
 581 }
 582
 583 /*
 584  * Called when a worker is scheduled in. Mark us as currently running.
 585  */
 586 void io_wq_worker_running(struct task_struct *tsk)
 587 {
 588         struct io_worker *worker = kthread_data(tsk);
 589         struct io_wqe *wqe = worker->wqe;
 590
 591         if (!(worker->flags & IO_WORKER_F_UP))
 592                 return;
 593         if (worker->flags & IO_WORKER_F_RUNNING)
 594                 return;
 595         worker->flags |= IO_WORKER_F_RUNNING;
 596         io_wqe_inc_running(wqe, worker);
 597 }
 598
 599 /*
 600  * Called when worker is going to sleep. If there are no workers currently
 601  * running and we have work pending, wake up a free one or have the manager
 602  * set one up.
 603  */
 604 void io_wq_worker_sleeping(struct task_struct *tsk)
 605 {
 606         struct io_worker *worker = kthread_data(tsk);
 607         struct io_wqe *wqe = worker->wqe;
 608
 609         if (!(worker->flags & IO_WORKER_F_UP))
 610                 return;
 611         if (!(worker->flags & IO_WORKER_F_RUNNING))
 612                 return;
 613
 614         worker->flags &= ~IO_WORKER_F_RUNNING;
 615
 616         spin_lock_irq(&wqe->lock);
 617         io_wqe_dec_running(wqe, worker);
 618         spin_unlock_irq(&wqe->lock);
 619 }
 620
 621 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
 622 {
 623         struct io_wqe_acct *acct =&wqe->acct[index];
 624         struct io_worker *worker;
 625
 626         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
 627         if (!worker)
 628                 return false;
 629
 630         refcount_set(&worker->ref, 1);
 631         worker->nulls_node.pprev = NULL;
 632         worker->wqe = wqe;
 633         spin_lock_init(&worker->lock);
 634
 635         worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
 636                                 "io_wqe_worker-%d/%d", index, wqe->node);
 637         if (IS_ERR(worker->task)) {
 638                 kfree(worker);
 639                 return false;
 640         }
 641
 642         spin_lock_irq(&wqe->lock);
 643         hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 644         list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 645         worker->flags |= IO_WORKER_F_FREE;
 646         if (index == IO_WQ_ACCT_BOUND)
 647                 worker->flags |= IO_WORKER_F_BOUND;
 648         if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
 649                 worker->flags |= IO_WORKER_F_FIXED;
 650         acct->nr_workers++;
 651         spin_unlock_irq(&wqe->lock);
 652
 653         if (index == IO_WQ_ACCT_UNBOUND)
 654                 atomic_inc(&wq->user->processes);
 655
 656         wake_up_process(worker->task);
 657         return true;
 658 }
 659
 660 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
 661         __must_hold(wqe->lock)
 662 {
 663         struct io_wqe_acct *acct = &wqe->acct[index];
 664
 665         /* if we have available workers or no work, no need */
 666         if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
 667                 return false;
 668         return acct->nr_workers < acct->max_workers;
 669 }
 670
 671 /*
 672  * Manager thread. Tasked with creating new workers, if we need them.
 673  */
 674 static int io_wq_manager(void *data)
 675 {
 676         struct io_wq *wq = data;
 677         int workers_to_create = num_possible_nodes();
 678         int node;
 679
 680         /* create fixed workers */
 681         refcount_set(&wq->refs, workers_to_create);
 682         for_each_node(node) {
 683                 if (!node_online(node))
 684                         continue;
 685                 if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
 686                         goto err;
 687                 workers_to_create--;
 688         }
 689
 690         while (workers_to_create--)
 691                 refcount_dec(&wq->refs);
 692
 693         complete(&wq->done);
 694
 695         while (!kthread_should_stop()) {
 696                 for_each_node(node) {
 697                         struct io_wqe *wqe = wq->wqes[node];
 698                         bool fork_worker[2] = { false, false };
 699
 700                         if (!node_online(node))
 701                                 continue;
 702
 703                         spin_lock_irq(&wqe->lock);
 704                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
 705                                 fork_worker[IO_WQ_ACCT_BOUND] = true;
 706                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
 707                                 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
 708                         spin_unlock_irq(&wqe->lock);
 709                         if (fork_worker[IO_WQ_ACCT_BOUND])
 710                                 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
 711                         if (fork_worker[IO_WQ_ACCT_UNBOUND])
 712                                 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
 713                 }
 714                 set_current_state(TASK_INTERRUPTIBLE);
 715                 schedule_timeout(HZ);
 716         }
 717
 718         return 0;
 719 err:
 720         set_bit(IO_WQ_BIT_ERROR, &wq->state);
 721         set_bit(IO_WQ_BIT_EXIT, &wq->state);
 722         if (refcount_sub_and_test(workers_to_create, &wq->refs))
 723                 complete(&wq->done);
 724         return 0;
 725 }
 726
 727 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
 728                             struct io_wq_work *work)
 729 {
 730         bool free_worker;
 731
 732         if (!(work->flags & IO_WQ_WORK_UNBOUND))
 733                 return true;
 734         if (atomic_read(&acct->nr_running))
 735                 return true;
 736
 737         rcu_read_lock();
 738         free_worker = !hlist_nulls_empty(&wqe->free_list);
 739         rcu_read_unlock();
 740         if (free_worker)
 741                 return true;
 742
 743         if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
 744             !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
 745                 return false;
 746
 747         return true;
 748 }
 749
 750 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 751 {
 752         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 753         int work_flags;
 754         unsigned long flags;
 755
 756         /*
 757          * Do early check to see if we need a new unbound worker, and if we do,
 758          * if we're allowed to do so. This isn't 100% accurate as there's a
 759          * gap between this check and incrementing the value, but that's OK.
 760          * It's close enough to not be an issue, fork() has the same delay.
 761          */
 762         if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
 763                 work->flags |= IO_WQ_WORK_CANCEL;
 764                 work->func(&work);
 765                 return;
 766         }
 767
 768         work_flags = work->flags;
 769         spin_lock_irqsave(&wqe->lock, flags);
 770         wq_list_add_tail(&work->list, &wqe->work_list);
 771         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 772         spin_unlock_irqrestore(&wqe->lock, flags);
 773
 774         if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 775             !atomic_read(&acct->nr_running))
 776                 io_wqe_wake_worker(wqe, acct);
 777 }
 778
 779 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 780 {
 781         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 782
 783         io_wqe_enqueue(wqe, work);
 784 }
 785
 786 /*
 787  * Enqueue work, hashed by some key. Work items that hash to the same value
 788  * will not be done in parallel. Used to limit concurrent writes, generally
 789  * hashed by inode.
 790  */
 791 void io_wq_enqueue_hashed(struct io_wq *wq, struct io_wq_work *work, void *val)
 792 {
 793         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 794         unsigned bit;
 795
 796
 797         bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 798         work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 799         io_wqe_enqueue(wqe, work);
 800 }
 801
 802 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
 803 {
 804         send_sig(SIGINT, worker->task, 1);
 805         return false;
 806 }
 807
 808 /*
 809  * Iterate the passed in list and call the specific function for each
 810  * worker that isn't exiting
 811  */
 812 static bool io_wq_for_each_worker(struct io_wqe *wqe,
 813                                   bool (*func)(struct io_worker *, void *),
 814                                   void *data)
 815 {
 816         struct io_worker *worker;
 817         bool ret = false;
 818
 819         list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
 820                 if (io_worker_get(worker)) {
 821                         /* no task if node is/was offline */
 822                         if (worker->task)
 823                                 ret = func(worker, data);
 824                         io_worker_release(worker);
 825                         if (ret)
 826                                 break;
 827                 }
 828         }
 829
 830         return ret;
 831 }
 832
 833 void io_wq_cancel_all(struct io_wq *wq)
 834 {
 835         int node;
 836
 837         set_bit(IO_WQ_BIT_CANCEL, &wq->state);
 838
 839         rcu_read_lock();
 840         for_each_node(node) {
 841                 struct io_wqe *wqe = wq->wqes[node];
 842
 843                 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
 844         }
 845         rcu_read_unlock();
 846 }
 847
 848 struct io_cb_cancel_data {
 849         struct io_wqe *wqe;
 850         work_cancel_fn *cancel;
 851         void *caller_data;
 852 };
 853
 854 static bool io_work_cancel(struct io_worker *worker, void *cancel_data)
 855 {
 856         struct io_cb_cancel_data *data = cancel_data;
 857         unsigned long flags;
 858         bool ret = false;
 859
 860         /*
 861          * Hold the lock to avoid ->cur_work going out of scope, caller
 862          * may dereference the passed in work.
 863          */
 864         spin_lock_irqsave(&worker->lock, flags);
 865         if (worker->cur_work &&
 866             !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) &&
 867             data->cancel(worker->cur_work, data->caller_data)) {
 868                 send_sig(SIGINT, worker->task, 1);
 869                 ret = true;
 870         }
 871         spin_unlock_irqrestore(&worker->lock, flags);
 872
 873         return ret;
 874 }
 875
 876 static enum io_wq_cancel io_wqe_cancel_cb_work(struct io_wqe *wqe,
 877                                                work_cancel_fn *cancel,
 878                                                void *cancel_data)
 879 {
 880         struct io_cb_cancel_data data = {
 881                 .wqe = wqe,
 882                 .cancel = cancel,
 883                 .caller_data = cancel_data,
 884         };
 885         struct io_wq_work_node *node, *prev;
 886         struct io_wq_work *work;
 887         unsigned long flags;
 888         bool found = false;
 889
 890         spin_lock_irqsave(&wqe->lock, flags);
 891         wq_list_for_each(node, prev, &wqe->work_list) {
 892                 work = container_of(node, struct io_wq_work, list);
 893
 894                 if (cancel(work, cancel_data)) {
 895                         wq_node_del(&wqe->work_list, node, prev);
 896                         found = true;
 897                         break;
 898                 }
 899         }
 900         spin_unlock_irqrestore(&wqe->lock, flags);
 901
 902         if (found) {
 903                 work->flags |= IO_WQ_WORK_CANCEL;
 904                 work->func(&work);
 905                 return IO_WQ_CANCEL_OK;
 906         }
 907
 908         rcu_read_lock();
 909         found = io_wq_for_each_worker(wqe, io_work_cancel, &data);
 910         rcu_read_unlock();
 911         return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND;
 912 }
 913
 914 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
 915                                   void *data)
 916 {
 917         enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
 918         int node;
 919
 920         for_each_node(node) {
 921                 struct io_wqe *wqe = wq->wqes[node];
 922
 923                 ret = io_wqe_cancel_cb_work(wqe, cancel, data);
 924                 if (ret != IO_WQ_CANCEL_NOTFOUND)
 925                         break;
 926         }
 927
 928         return ret;
 929 }
 930
 931 struct work_match {
 932         bool (*fn)(struct io_wq_work *, void *data);
 933         void *data;
 934 };
 935
 936 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 937 {
 938         struct work_match *match = data;
 939         unsigned long flags;
 940         bool ret = false;
 941
 942         spin_lock_irqsave(&worker->lock, flags);
 943         if (match->fn(worker->cur_work, match->data) &&
 944             !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL)) {
 945                 send_sig(SIGINT, worker->task, 1);
 946                 ret = true;
 947         }
 948         spin_unlock_irqrestore(&worker->lock, flags);
 949
 950         return ret;
 951 }
 952
 953 static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe,
 954                                             struct work_match *match)
 955 {
 956         struct io_wq_work_node *node, *prev;
 957         struct io_wq_work *work;
 958         unsigned long flags;
 959         bool found = false;
 960
 961         /*
 962          * First check pending list, if we're lucky we can just remove it
 963          * from there. CANCEL_OK means that the work is returned as-new,
 964          * no completion will be posted for it.
 965          */
 966         spin_lock_irqsave(&wqe->lock, flags);
 967         wq_list_for_each(node, prev, &wqe->work_list) {
 968                 work = container_of(node, struct io_wq_work, list);
 969
 970                 if (match->fn(work, match->data)) {
 971                         wq_node_del(&wqe->work_list, node, prev);
 972                         found = true;
 973                         break;
 974                 }
 975         }
 976         spin_unlock_irqrestore(&wqe->lock, flags);
 977
 978         if (found) {
 979                 work->flags |= IO_WQ_WORK_CANCEL;
 980                 work->func(&work);
 981                 return IO_WQ_CANCEL_OK;
 982         }
 983
 984         /*
 985          * Now check if a free (going busy) or busy worker has the work
 986          * currently running. If we find it there, we'll return CANCEL_RUNNING
 987          * as an indication that we attempt to signal cancellation. The
 988          * completion will run normally in this case.
 989          */
 990         rcu_read_lock();
 991         found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
 992         rcu_read_unlock();
 993         return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND;
 994 }
 995
 996 static bool io_wq_work_match(struct io_wq_work *work, void *data)
 997 {
 998         return work == data;
 999 }
1000
1001 enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork)
1002 {
1003         struct work_match match = {
1004                 .fn     = io_wq_work_match,
1005                 .data   = cwork
1006         };
1007         enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
1008         int node;
1009
1010         cwork->flags |= IO_WQ_WORK_CANCEL;
1011
1012         for_each_node(node) {
1013                 struct io_wqe *wqe = wq->wqes[node];
1014
1015                 ret = io_wqe_cancel_work(wqe, &match);
1016                 if (ret != IO_WQ_CANCEL_NOTFOUND)
1017                         break;
1018         }
1019
1020         return ret;
1021 }
1022
1023 static bool io_wq_pid_match(struct io_wq_work *work, void *data)
1024 {
1025         pid_t pid = (pid_t) (unsigned long) data;
1026
1027         if (work)
1028                 return work->task_pid == pid;
1029         return false;
1030 }
1031
1032 enum io_wq_cancel io_wq_cancel_pid(struct io_wq *wq, pid_t pid)
1033 {
1034         struct work_match match = {
1035                 .fn     = io_wq_pid_match,
1036                 .data   = (void *) (unsigned long) pid
1037         };
1038         enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
1039         int node;
1040
1041         for_each_node(node) {
1042                 struct io_wqe *wqe = wq->wqes[node];
1043
1044                 ret = io_wqe_cancel_work(wqe, &match);
1045                 if (ret != IO_WQ_CANCEL_NOTFOUND)
1046                         break;
1047         }
1048
1049         return ret;
1050 }
1051
1052 struct io_wq_flush_data {
1053         struct io_wq_work work;
1054         struct completion done;
1055 };
1056
1057 static void io_wq_flush_func(struct io_wq_work **workptr)
1058 {
1059         struct io_wq_work *work = *workptr;
1060         struct io_wq_flush_data *data;
1061
1062         data = container_of(work, struct io_wq_flush_data, work);
1063         complete(&data->done);
1064 }
1065
1066 /*
1067  * Doesn't wait for previously queued work to finish. When this completes,
1068  * it just means that previously queued work was started.
1069  */
1070 void io_wq_flush(struct io_wq *wq)
1071 {
1072         struct io_wq_flush_data data;
1073         int node;
1074
1075         for_each_node(node) {
1076                 struct io_wqe *wqe = wq->wqes[node];
1077
1078                 if (!node_online(node))
1079                         continue;
1080                 init_completion(&data.done);
1081                 INIT_IO_WORK(&data.work, io_wq_flush_func);
1082                 data.work.flags |= IO_WQ_WORK_INTERNAL;
1083                 io_wqe_enqueue(wqe, &data.work);
1084                 wait_for_completion(&data.done);
1085         }
1086 }
1087
1088 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1089 {
1090         int ret = -ENOMEM, node;
1091         struct io_wq *wq;
1092
1093         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1094         if (!wq)
1095                 return ERR_PTR(-ENOMEM);
1096
1097         wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1098         if (!wq->wqes) {
1099                 kfree(wq);
1100                 return ERR_PTR(-ENOMEM);
1101         }
1102
1103         wq->get_work = data->get_work;
1104         wq->put_work = data->put_work;
1105
1106         /* caller must already hold a reference to this */
1107         wq->user = data->user;
1108
1109         for_each_node(node) {
1110                 struct io_wqe *wqe;
1111                 int alloc_node = node;
1112
1113                 if (!node_online(alloc_node))
1114                         alloc_node = NUMA_NO_NODE;
1115                 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1116                 if (!wqe)
1117                         goto err;
1118                 wq->wqes[node] = wqe;
1119                 wqe->node = alloc_node;
1120                 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1121                 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1122                 if (wq->user) {
1123                         wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1124                                         task_rlimit(current, RLIMIT_NPROC);
1125                 }
1126                 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1127                 wqe->wq = wq;
1128                 spin_lock_init(&wqe->lock);
1129                 INIT_WQ_LIST(&wqe->work_list);
1130                 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1131                 INIT_LIST_HEAD(&wqe->all_list);
1132         }
1133
1134         init_completion(&wq->done);
1135
1136         wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1137         if (!IS_ERR(wq->manager)) {
1138                 wake_up_process(wq->manager);
1139                 wait_for_completion(&wq->done);
1140                 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1141                         ret = -ENOMEM;
1142                         goto err;
1143                 }
1144                 refcount_set(&wq->use_refs, 1);
1145                 reinit_completion(&wq->done);
1146                 return wq;
1147         }
1148
1149         ret = PTR_ERR(wq->manager);
1150         complete(&wq->done);
1151 err:
1152         for_each_node(node)
1153                 kfree(wq->wqes[node]);
1154         kfree(wq->wqes);
1155         kfree(wq);
1156         return ERR_PTR(ret);
1157 }
1158
1159 bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1160 {
1161         if (data->get_work != wq->get_work || data->put_work != wq->put_work)
1162                 return false;
1163
1164         return refcount_inc_not_zero(&wq->use_refs);
1165 }
1166
1167 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
1168 {
1169         wake_up_process(worker->task);
1170         return false;
1171 }
1172
1173 static void __io_wq_destroy(struct io_wq *wq)
1174 {
1175         int node;
1176
1177         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1178         if (wq->manager)
1179                 kthread_stop(wq->manager);
1180
1181         rcu_read_lock();
1182         for_each_node(node)
1183                 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1184         rcu_read_unlock();
1185
1186         wait_for_completion(&wq->done);
1187
1188         for_each_node(node)
1189                 kfree(wq->wqes[node]);
1190         kfree(wq->wqes);
1191         kfree(wq);
1192 }
1193
1194 void io_wq_destroy(struct io_wq *wq)
1195 {
1196         if (refcount_dec_and_test(&wq->use_refs))
1197                 __io_wq_destroy(wq);
1198 }