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 #include <linux/task_work.h>
  21
  22 #include "io-wq.h"
  23
  24 #define WORKER_IDLE_TIMEOUT     (5 * HZ)
  25
  26 enum {
  27         IO_WORKER_F_UP          = 1,    /* up and active */
  28         IO_WORKER_F_RUNNING     = 2,    /* account as running */
  29         IO_WORKER_F_FREE        = 4,    /* worker on free list */
  30         IO_WORKER_F_EXITING     = 8,    /* worker exiting */
  31         IO_WORKER_F_FIXED       = 16,   /* static idle worker */
  32         IO_WORKER_F_BOUND       = 32,   /* is doing bounded work */
  33 };
  34
  35 enum {
  36         IO_WQ_BIT_EXIT          = 0,    /* wq exiting */
  37         IO_WQ_BIT_CANCEL        = 1,    /* cancel work on list */
  38         IO_WQ_BIT_ERROR         = 2,    /* error on setup */
  39 };
  40
  41 enum {
  42         IO_WQE_FLAG_STALLED     = 1,    /* stalled on hash */
  43 };
  44
  45 /*
  46  * One for each thread in a wqe pool
  47  */
  48 struct io_worker {
  49         refcount_t ref;
  50         unsigned flags;
  51         struct hlist_nulls_node nulls_node;
  52         struct list_head all_list;
  53         struct task_struct *task;
  54         struct io_wqe *wqe;
  55
  56         struct io_wq_work *cur_work;
  57         spinlock_t lock;
  58
  59         struct rcu_head rcu;
  60         struct mm_struct *mm;
  61         const struct cred *cur_creds;
  62         const struct cred *saved_creds;
  63         struct files_struct *restore_files;
  64         struct fs_struct *restore_fs;
  65 };
  66
  67 #if BITS_PER_LONG == 64
  68 #define IO_WQ_HASH_ORDER        6
  69 #else
  70 #define IO_WQ_HASH_ORDER        5
  71 #endif
  72
  73 #define IO_WQ_NR_HASH_BUCKETS   (1u << IO_WQ_HASH_ORDER)
  74
  75 struct io_wqe_acct {
  76         unsigned nr_workers;
  77         unsigned max_workers;
  78         atomic_t nr_running;
  79 };
  80
  81 enum {
  82         IO_WQ_ACCT_BOUND,
  83         IO_WQ_ACCT_UNBOUND,
  84 };
  85
  86 /*
  87  * Per-node worker thread pool
  88  */
  89 struct io_wqe {
  90         struct {
  91                 spinlock_t lock;
  92                 struct io_wq_work_list work_list;
  93                 unsigned long hash_map;
  94                 unsigned flags;
  95         } ____cacheline_aligned_in_smp;
  96
  97         int node;
  98         struct io_wqe_acct acct[2];
  99
 100         struct hlist_nulls_head free_list;
 101         struct list_head all_list;
 102
 103         struct io_wq *wq;
 104         struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
 105 };
 106
 107 /*
 108  * Per io_wq state
 109   */
 110 struct io_wq {
 111         struct io_wqe **wqes;
 112         unsigned long state;
 113
 114         free_work_fn *free_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                 set_fs(KERNEL_DS);
 174                 unuse_mm(worker->mm);
 175                 mmput(worker->mm);
 176                 worker->mm = NULL;
 177         }
 178
 179         return dropped_lock;
 180 }
 181
 182 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 183                                                    struct io_wq_work *work)
 184 {
 185         if (work->flags & IO_WQ_WORK_UNBOUND)
 186                 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 187
 188         return &wqe->acct[IO_WQ_ACCT_BOUND];
 189 }
 190
 191 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
 192                                                   struct io_worker *worker)
 193 {
 194         if (worker->flags & IO_WORKER_F_BOUND)
 195                 return &wqe->acct[IO_WQ_ACCT_BOUND];
 196
 197         return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 198 }
 199
 200 static void io_worker_exit(struct io_worker *worker)
 201 {
 202         struct io_wqe *wqe = worker->wqe;
 203         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 204         unsigned nr_workers;
 205
 206         /*
 207          * If we're not at zero, someone else is holding a brief reference
 208          * to the worker. Wait for that to go away.
 209          */
 210         set_current_state(TASK_INTERRUPTIBLE);
 211         if (!refcount_dec_and_test(&worker->ref))
 212                 schedule();
 213         __set_current_state(TASK_RUNNING);
 214
 215         preempt_disable();
 216         current->flags &= ~PF_IO_WORKER;
 217         if (worker->flags & IO_WORKER_F_RUNNING)
 218                 atomic_dec(&acct->nr_running);
 219         if (!(worker->flags & IO_WORKER_F_BOUND))
 220                 atomic_dec(&wqe->wq->user->processes);
 221         worker->flags = 0;
 222         preempt_enable();
 223
 224         spin_lock_irq(&wqe->lock);
 225         hlist_nulls_del_rcu(&worker->nulls_node);
 226         list_del_rcu(&worker->all_list);
 227         if (__io_worker_unuse(wqe, worker)) {
 228                 __release(&wqe->lock);
 229                 spin_lock_irq(&wqe->lock);
 230         }
 231         acct->nr_workers--;
 232         nr_workers = wqe->acct[IO_WQ_ACCT_BOUND].nr_workers +
 233                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers;
 234         spin_unlock_irq(&wqe->lock);
 235
 236         /* all workers gone, wq exit can proceed */
 237         if (!nr_workers && refcount_dec_and_test(&wqe->wq->refs))
 238                 complete(&wqe->wq->done);
 239
 240         kfree_rcu(worker, rcu);
 241 }
 242
 243 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
 244         __must_hold(wqe->lock)
 245 {
 246         if (!wq_list_empty(&wqe->work_list) &&
 247             !(wqe->flags & IO_WQE_FLAG_STALLED))
 248                 return true;
 249         return false;
 250 }
 251
 252 /*
 253  * Check head of free list for an available worker. If one isn't available,
 254  * caller must wake up the wq manager to create one.
 255  */
 256 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
 257         __must_hold(RCU)
 258 {
 259         struct hlist_nulls_node *n;
 260         struct io_worker *worker;
 261
 262         n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
 263         if (is_a_nulls(n))
 264                 return false;
 265
 266         worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
 267         if (io_worker_get(worker)) {
 268                 wake_up_process(worker->task);
 269                 io_worker_release(worker);
 270                 return true;
 271         }
 272
 273         return false;
 274 }
 275
 276 /*
 277  * We need a worker. If we find a free one, we're good. If not, and we're
 278  * below the max number of workers, wake up the manager to create one.
 279  */
 280 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 281 {
 282         bool ret;
 283
 284         /*
 285          * Most likely an attempt to queue unbounded work on an io_wq that
 286          * wasn't setup with any unbounded workers.
 287          */
 288         WARN_ON_ONCE(!acct->max_workers);
 289
 290         rcu_read_lock();
 291         ret = io_wqe_activate_free_worker(wqe);
 292         rcu_read_unlock();
 293
 294         if (!ret && acct->nr_workers < acct->max_workers)
 295                 wake_up_process(wqe->wq->manager);
 296 }
 297
 298 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
 299 {
 300         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 301
 302         atomic_inc(&acct->nr_running);
 303 }
 304
 305 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
 306         __must_hold(wqe->lock)
 307 {
 308         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 309
 310         if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
 311                 io_wqe_wake_worker(wqe, acct);
 312 }
 313
 314 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
 315 {
 316         allow_kernel_signal(SIGINT);
 317
 318         current->flags |= PF_IO_WORKER;
 319
 320         worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 321         worker->restore_files = current->files;
 322         worker->restore_fs = current->fs;
 323         io_wqe_inc_running(wqe, worker);
 324 }
 325
 326 /*
 327  * Worker will start processing some work. Move it to the busy list, if
 328  * it's currently on the freelist
 329  */
 330 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
 331                              struct io_wq_work *work)
 332         __must_hold(wqe->lock)
 333 {
 334         bool worker_bound, work_bound;
 335
 336         if (worker->flags & IO_WORKER_F_FREE) {
 337                 worker->flags &= ~IO_WORKER_F_FREE;
 338                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 339         }
 340
 341         /*
 342          * If worker is moving from bound to unbound (or vice versa), then
 343          * ensure we update the running accounting.
 344          */
 345         worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
 346         work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
 347         if (worker_bound != work_bound) {
 348                 io_wqe_dec_running(wqe, worker);
 349                 if (work_bound) {
 350                         worker->flags |= IO_WORKER_F_BOUND;
 351                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
 352                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
 353                         atomic_dec(&wqe->wq->user->processes);
 354                 } else {
 355                         worker->flags &= ~IO_WORKER_F_BOUND;
 356                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
 357                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
 358                         atomic_inc(&wqe->wq->user->processes);
 359                 }
 360                 io_wqe_inc_running(wqe, worker);
 361          }
 362 }
 363
 364 /*
 365  * No work, worker going to sleep. Move to freelist, and unuse mm if we
 366  * have one attached. Dropping the mm may potentially sleep, so we drop
 367  * the lock in that case and return success. Since the caller has to
 368  * retry the loop in that case (we changed task state), we don't regrab
 369  * the lock if we return success.
 370  */
 371 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
 372         __must_hold(wqe->lock)
 373 {
 374         if (!(worker->flags & IO_WORKER_F_FREE)) {
 375                 worker->flags |= IO_WORKER_F_FREE;
 376                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 377         }
 378
 379         return __io_worker_unuse(wqe, worker);
 380 }
 381
 382 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
 383 {
 384         return work->flags >> IO_WQ_HASH_SHIFT;
 385 }
 386
 387 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
 388         __must_hold(wqe->lock)
 389 {
 390         struct io_wq_work_node *node, *prev;
 391         struct io_wq_work *work, *tail;
 392         unsigned int hash;
 393
 394         wq_list_for_each(node, prev, &wqe->work_list) {
 395                 work = container_of(node, struct io_wq_work, list);
 396
 397                 /* not hashed, can run anytime */
 398                 if (!io_wq_is_hashed(work)) {
 399                         wq_list_del(&wqe->work_list, node, prev);
 400                         return work;
 401                 }
 402
 403                 /* hashed, can run if not already running */
 404                 hash = io_get_work_hash(work);
 405                 if (!(wqe->hash_map & BIT(hash))) {
 406                         wqe->hash_map |= BIT(hash);
 407                         /* all items with this hash lie in [work, tail] */
 408                         tail = wqe->hash_tail[hash];
 409                         wqe->hash_tail[hash] = NULL;
 410                         wq_list_cut(&wqe->work_list, &tail->list, prev);
 411                         return work;
 412                 }
 413         }
 414
 415         return NULL;
 416 }
 417
 418 static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
 419 {
 420         if (worker->mm) {
 421                 unuse_mm(worker->mm);
 422                 mmput(worker->mm);
 423                 worker->mm = NULL;
 424         }
 425         if (!work->mm) {
 426                 set_fs(KERNEL_DS);
 427                 return;
 428         }
 429         if (mmget_not_zero(work->mm)) {
 430                 use_mm(work->mm);
 431                 if (!worker->mm)
 432                         set_fs(USER_DS);
 433                 worker->mm = work->mm;
 434                 /* hang on to this mm */
 435                 work->mm = NULL;
 436                 return;
 437         }
 438
 439         /* failed grabbing mm, ensure work gets cancelled */
 440         work->flags |= IO_WQ_WORK_CANCEL;
 441 }
 442
 443 static void io_wq_switch_creds(struct io_worker *worker,
 444                                struct io_wq_work *work)
 445 {
 446         const struct cred *old_creds = override_creds(work->creds);
 447
 448         worker->cur_creds = work->creds;
 449         if (worker->saved_creds)
 450                 put_cred(old_creds); /* creds set by previous switch */
 451         else
 452                 worker->saved_creds = old_creds;
 453 }
 454
 455 static void io_impersonate_work(struct io_worker *worker,
 456                                 struct io_wq_work *work)
 457 {
 458         if (work->files && current->files != work->files) {
 459                 task_lock(current);
 460                 current->files = work->files;
 461                 task_unlock(current);
 462         }
 463         if (work->fs && current->fs != work->fs)
 464                 current->fs = work->fs;
 465         if (work->mm != worker->mm)
 466                 io_wq_switch_mm(worker, work);
 467         if (worker->cur_creds != work->creds)
 468                 io_wq_switch_creds(worker, work);
 469 }
 470
 471 static void io_assign_current_work(struct io_worker *worker,
 472                                    struct io_wq_work *work)
 473 {
 474         if (work) {
 475                 /* flush pending signals before assigning new work */
 476                 if (signal_pending(current))
 477                         flush_signals(current);
 478                 cond_resched();
 479         }
 480
 481         spin_lock_irq(&worker->lock);
 482         worker->cur_work = work;
 483         spin_unlock_irq(&worker->lock);
 484 }
 485
 486 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
 487
 488 static void io_worker_handle_work(struct io_worker *worker)
 489         __releases(wqe->lock)
 490 {
 491         struct io_wqe *wqe = worker->wqe;
 492         struct io_wq *wq = wqe->wq;
 493
 494         do {
 495                 struct io_wq_work *work;
 496                 unsigned int hash;
 497 get_next:
 498                 /*
 499                  * If we got some work, mark us as busy. If we didn't, but
 500                  * the list isn't empty, it means we stalled on hashed work.
 501                  * Mark us stalled so we don't keep looking for work when we
 502                  * can't make progress, any work completion or insertion will
 503                  * clear the stalled flag.
 504                  */
 505                 work = io_get_next_work(wqe);
 506                 if (work)
 507                         __io_worker_busy(wqe, worker, work);
 508                 else if (!wq_list_empty(&wqe->work_list))
 509                         wqe->flags |= IO_WQE_FLAG_STALLED;
 510
 511                 spin_unlock_irq(&wqe->lock);
 512                 if (!work)
 513                         break;
 514                 io_assign_current_work(worker, work);
 515
 516                 /* handle a whole dependent link */
 517                 do {
 518                         struct io_wq_work *old_work, *next_hashed, *linked;
 519
 520                         next_hashed = wq_next_work(work);
 521                         io_impersonate_work(worker, work);
 522                         /*
 523                          * OK to set IO_WQ_WORK_CANCEL even for uncancellable
 524                          * work, the worker function will do the right thing.
 525                          */
 526                         if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
 527                                 work->flags |= IO_WQ_WORK_CANCEL;
 528
 529                         hash = io_get_work_hash(work);
 530                         linked = old_work = work;
 531                         linked->func(&linked);
 532                         linked = (old_work == linked) ? NULL : linked;
 533
 534                         work = next_hashed;
 535                         if (!work && linked && !io_wq_is_hashed(linked)) {
 536                                 work = linked;
 537                                 linked = NULL;
 538                         }
 539                         io_assign_current_work(worker, work);
 540                         wq->free_work(old_work);
 541
 542                         if (linked)
 543                                 io_wqe_enqueue(wqe, linked);
 544
 545                         if (hash != -1U && !next_hashed) {
 546                                 spin_lock_irq(&wqe->lock);
 547                                 wqe->hash_map &= ~BIT_ULL(hash);
 548                                 wqe->flags &= ~IO_WQE_FLAG_STALLED;
 549                                 /* dependent work is not hashed */
 550                                 hash = -1U;
 551                                 /* skip unnecessary unlock-lock wqe->lock */
 552                                 if (!work)
 553                                         goto get_next;
 554                                 spin_unlock_irq(&wqe->lock);
 555                         }
 556                 } while (work);
 557
 558                 spin_lock_irq(&wqe->lock);
 559         } while (1);
 560 }
 561
 562 static int io_wqe_worker(void *data)
 563 {
 564         struct io_worker *worker = data;
 565         struct io_wqe *wqe = worker->wqe;
 566         struct io_wq *wq = wqe->wq;
 567
 568         io_worker_start(wqe, worker);
 569
 570         while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 571                 set_current_state(TASK_INTERRUPTIBLE);
 572 loop:
 573                 spin_lock_irq(&wqe->lock);
 574                 if (io_wqe_run_queue(wqe)) {
 575                         __set_current_state(TASK_RUNNING);
 576                         io_worker_handle_work(worker);
 577                         goto loop;
 578                 }
 579                 /* drops the lock on success, retry */
 580                 if (__io_worker_idle(wqe, worker)) {
 581                         __release(&wqe->lock);
 582                         goto loop;
 583                 }
 584                 spin_unlock_irq(&wqe->lock);
 585                 if (signal_pending(current))
 586                         flush_signals(current);
 587                 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
 588                         continue;
 589                 /* timed out, exit unless we're the fixed worker */
 590                 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 591                     !(worker->flags & IO_WORKER_F_FIXED))
 592                         break;
 593         }
 594
 595         if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 596                 spin_lock_irq(&wqe->lock);
 597                 if (!wq_list_empty(&wqe->work_list))
 598                         io_worker_handle_work(worker);
 599                 else
 600                         spin_unlock_irq(&wqe->lock);
 601         }
 602
 603         io_worker_exit(worker);
 604         return 0;
 605 }
 606
 607 /*
 608  * Called when a worker is scheduled in. Mark us as currently running.
 609  */
 610 void io_wq_worker_running(struct task_struct *tsk)
 611 {
 612         struct io_worker *worker = kthread_data(tsk);
 613         struct io_wqe *wqe = worker->wqe;
 614
 615         if (!(worker->flags & IO_WORKER_F_UP))
 616                 return;
 617         if (worker->flags & IO_WORKER_F_RUNNING)
 618                 return;
 619         worker->flags |= IO_WORKER_F_RUNNING;
 620         io_wqe_inc_running(wqe, worker);
 621 }
 622
 623 /*
 624  * Called when worker is going to sleep. If there are no workers currently
 625  * running and we have work pending, wake up a free one or have the manager
 626  * set one up.
 627  */
 628 void io_wq_worker_sleeping(struct task_struct *tsk)
 629 {
 630         struct io_worker *worker = kthread_data(tsk);
 631         struct io_wqe *wqe = worker->wqe;
 632
 633         if (!(worker->flags & IO_WORKER_F_UP))
 634                 return;
 635         if (!(worker->flags & IO_WORKER_F_RUNNING))
 636                 return;
 637
 638         worker->flags &= ~IO_WORKER_F_RUNNING;
 639
 640         spin_lock_irq(&wqe->lock);
 641         io_wqe_dec_running(wqe, worker);
 642         spin_unlock_irq(&wqe->lock);
 643 }
 644
 645 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
 646 {
 647         struct io_wqe_acct *acct =&wqe->acct[index];
 648         struct io_worker *worker;
 649
 650         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
 651         if (!worker)
 652                 return false;
 653
 654         refcount_set(&worker->ref, 1);
 655         worker->nulls_node.pprev = NULL;
 656         worker->wqe = wqe;
 657         spin_lock_init(&worker->lock);
 658
 659         worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
 660                                 "io_wqe_worker-%d/%d", index, wqe->node);
 661         if (IS_ERR(worker->task)) {
 662                 kfree(worker);
 663                 return false;
 664         }
 665
 666         spin_lock_irq(&wqe->lock);
 667         hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 668         list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 669         worker->flags |= IO_WORKER_F_FREE;
 670         if (index == IO_WQ_ACCT_BOUND)
 671                 worker->flags |= IO_WORKER_F_BOUND;
 672         if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
 673                 worker->flags |= IO_WORKER_F_FIXED;
 674         acct->nr_workers++;
 675         spin_unlock_irq(&wqe->lock);
 676
 677         if (index == IO_WQ_ACCT_UNBOUND)
 678                 atomic_inc(&wq->user->processes);
 679
 680         wake_up_process(worker->task);
 681         return true;
 682 }
 683
 684 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
 685         __must_hold(wqe->lock)
 686 {
 687         struct io_wqe_acct *acct = &wqe->acct[index];
 688
 689         /* if we have available workers or no work, no need */
 690         if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
 691                 return false;
 692         return acct->nr_workers < acct->max_workers;
 693 }
 694
 695 /*
 696  * Manager thread. Tasked with creating new workers, if we need them.
 697  */
 698 static int io_wq_manager(void *data)
 699 {
 700         struct io_wq *wq = data;
 701         int workers_to_create = num_possible_nodes();
 702         int node;
 703
 704         /* create fixed workers */
 705         refcount_set(&wq->refs, workers_to_create);
 706         for_each_node(node) {
 707                 if (!node_online(node))
 708                         continue;
 709                 if (!create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
 710                         goto err;
 711                 workers_to_create--;
 712         }
 713
 714         while (workers_to_create--)
 715                 refcount_dec(&wq->refs);
 716
 717         complete(&wq->done);
 718
 719         while (!kthread_should_stop()) {
 720                 if (current->task_works)
 721                         task_work_run();
 722
 723                 for_each_node(node) {
 724                         struct io_wqe *wqe = wq->wqes[node];
 725                         bool fork_worker[2] = { false, false };
 726
 727                         if (!node_online(node))
 728                                 continue;
 729
 730                         spin_lock_irq(&wqe->lock);
 731                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
 732                                 fork_worker[IO_WQ_ACCT_BOUND] = true;
 733                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
 734                                 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
 735                         spin_unlock_irq(&wqe->lock);
 736                         if (fork_worker[IO_WQ_ACCT_BOUND])
 737                                 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
 738                         if (fork_worker[IO_WQ_ACCT_UNBOUND])
 739                                 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
 740                 }
 741                 set_current_state(TASK_INTERRUPTIBLE);
 742                 schedule_timeout(HZ);
 743         }
 744
 745         if (current->task_works)
 746                 task_work_run();
 747
 748         return 0;
 749 err:
 750         set_bit(IO_WQ_BIT_ERROR, &wq->state);
 751         set_bit(IO_WQ_BIT_EXIT, &wq->state);
 752         if (refcount_sub_and_test(workers_to_create, &wq->refs))
 753                 complete(&wq->done);
 754         return 0;
 755 }
 756
 757 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
 758                             struct io_wq_work *work)
 759 {
 760         bool free_worker;
 761
 762         if (!(work->flags & IO_WQ_WORK_UNBOUND))
 763                 return true;
 764         if (atomic_read(&acct->nr_running))
 765                 return true;
 766
 767         rcu_read_lock();
 768         free_worker = !hlist_nulls_empty(&wqe->free_list);
 769         rcu_read_unlock();
 770         if (free_worker)
 771                 return true;
 772
 773         if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
 774             !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
 775                 return false;
 776
 777         return true;
 778 }
 779
 780 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
 781 {
 782         struct io_wq *wq = wqe->wq;
 783
 784         do {
 785                 struct io_wq_work *old_work = work;
 786
 787                 work->flags |= IO_WQ_WORK_CANCEL;
 788                 work->func(&work);
 789                 work = (work == old_work) ? NULL : work;
 790                 wq->free_work(old_work);
 791         } while (work);
 792 }
 793
 794 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
 795 {
 796         unsigned int hash;
 797         struct io_wq_work *tail;
 798
 799         if (!io_wq_is_hashed(work)) {
 800 append:
 801                 wq_list_add_tail(&work->list, &wqe->work_list);
 802                 return;
 803         }
 804
 805         hash = io_get_work_hash(work);
 806         tail = wqe->hash_tail[hash];
 807         wqe->hash_tail[hash] = work;
 808         if (!tail)
 809                 goto append;
 810
 811         wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
 812 }
 813
 814 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 815 {
 816         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 817         int work_flags;
 818         unsigned long flags;
 819
 820         /*
 821          * Do early check to see if we need a new unbound worker, and if we do,
 822          * if we're allowed to do so. This isn't 100% accurate as there's a
 823          * gap between this check and incrementing the value, but that's OK.
 824          * It's close enough to not be an issue, fork() has the same delay.
 825          */
 826         if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
 827                 io_run_cancel(work, wqe);
 828                 return;
 829         }
 830
 831         work_flags = work->flags;
 832         spin_lock_irqsave(&wqe->lock, flags);
 833         io_wqe_insert_work(wqe, work);
 834         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 835         spin_unlock_irqrestore(&wqe->lock, flags);
 836
 837         if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 838             !atomic_read(&acct->nr_running))
 839                 io_wqe_wake_worker(wqe, acct);
 840 }
 841
 842 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 843 {
 844         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 845
 846         io_wqe_enqueue(wqe, work);
 847 }
 848
 849 /*
 850  * Work items that hash to the same value will not be done in parallel.
 851  * Used to limit concurrent writes, generally hashed by inode.
 852  */
 853 void io_wq_hash_work(struct io_wq_work *work, void *val)
 854 {
 855         unsigned int bit;
 856
 857         bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 858         work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 859 }
 860
 861 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
 862 {
 863         send_sig(SIGINT, worker->task, 1);
 864         return false;
 865 }
 866
 867 /*
 868  * Iterate the passed in list and call the specific function for each
 869  * worker that isn't exiting
 870  */
 871 static bool io_wq_for_each_worker(struct io_wqe *wqe,
 872                                   bool (*func)(struct io_worker *, void *),
 873                                   void *data)
 874 {
 875         struct io_worker *worker;
 876         bool ret = false;
 877
 878         list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
 879                 if (io_worker_get(worker)) {
 880                         /* no task if node is/was offline */
 881                         if (worker->task)
 882                                 ret = func(worker, data);
 883                         io_worker_release(worker);
 884                         if (ret)
 885                                 break;
 886                 }
 887         }
 888
 889         return ret;
 890 }
 891
 892 void io_wq_cancel_all(struct io_wq *wq)
 893 {
 894         int node;
 895
 896         set_bit(IO_WQ_BIT_CANCEL, &wq->state);
 897
 898         rcu_read_lock();
 899         for_each_node(node) {
 900                 struct io_wqe *wqe = wq->wqes[node];
 901
 902                 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
 903         }
 904         rcu_read_unlock();
 905 }
 906
 907 struct io_cb_cancel_data {
 908         work_cancel_fn *fn;
 909         void *data;
 910 };
 911
 912 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 913 {
 914         struct io_cb_cancel_data *match = data;
 915         unsigned long flags;
 916         bool ret = false;
 917
 918         /*
 919          * Hold the lock to avoid ->cur_work going out of scope, caller
 920          * may dereference the passed in work.
 921          */
 922         spin_lock_irqsave(&worker->lock, flags);
 923         if (worker->cur_work &&
 924             !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) &&
 925             match->fn(worker->cur_work, match->data)) {
 926                 send_sig(SIGINT, worker->task, 1);
 927                 ret = true;
 928         }
 929         spin_unlock_irqrestore(&worker->lock, flags);
 930
 931         return ret;
 932 }
 933
 934 static enum io_wq_cancel io_wqe_cancel_work(struct io_wqe *wqe,
 935                                             struct io_cb_cancel_data *match)
 936 {
 937         struct io_wq_work_node *node, *prev;
 938         struct io_wq_work *work;
 939         unsigned long flags;
 940         bool found = false;
 941
 942         /*
 943          * First check pending list, if we're lucky we can just remove it
 944          * from there. CANCEL_OK means that the work is returned as-new,
 945          * no completion will be posted for it.
 946          */
 947         spin_lock_irqsave(&wqe->lock, flags);
 948         wq_list_for_each(node, prev, &wqe->work_list) {
 949                 work = container_of(node, struct io_wq_work, list);
 950
 951                 if (match->fn(work, match->data)) {
 952                         wq_list_del(&wqe->work_list, node, prev);
 953                         found = true;
 954                         break;
 955                 }
 956         }
 957         spin_unlock_irqrestore(&wqe->lock, flags);
 958
 959         if (found) {
 960                 io_run_cancel(work, wqe);
 961                 return IO_WQ_CANCEL_OK;
 962         }
 963
 964         /*
 965          * Now check if a free (going busy) or busy worker has the work
 966          * currently running. If we find it there, we'll return CANCEL_RUNNING
 967          * as an indication that we attempt to signal cancellation. The
 968          * completion will run normally in this case.
 969          */
 970         rcu_read_lock();
 971         found = io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
 972         rcu_read_unlock();
 973         return found ? IO_WQ_CANCEL_RUNNING : IO_WQ_CANCEL_NOTFOUND;
 974 }
 975
 976 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
 977                                   void *data)
 978 {
 979         struct io_cb_cancel_data match = {
 980                 .fn     = cancel,
 981                 .data   = data,
 982         };
 983         enum io_wq_cancel ret = IO_WQ_CANCEL_NOTFOUND;
 984         int node;
 985
 986         for_each_node(node) {
 987                 struct io_wqe *wqe = wq->wqes[node];
 988
 989                 ret = io_wqe_cancel_work(wqe, &match);
 990                 if (ret != IO_WQ_CANCEL_NOTFOUND)
 991                         break;
 992         }
 993
 994         return ret;
 995 }
 996
 997 static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data)
 998 {
 999         return work == data;
1000 }
1001
1002 enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork)
1003 {
1004         return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork);
1005 }
1006
1007 static bool io_wq_pid_match(struct io_wq_work *work, void *data)
1008 {
1009         pid_t pid = (pid_t) (unsigned long) data;
1010
1011         return work->task_pid == pid;
1012 }
1013
1014 enum io_wq_cancel io_wq_cancel_pid(struct io_wq *wq, pid_t pid)
1015 {
1016         void *data = (void *) (unsigned long) pid;
1017
1018         return io_wq_cancel_cb(wq, io_wq_pid_match, data);
1019 }
1020
1021 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1022 {
1023         int ret = -ENOMEM, node;
1024         struct io_wq *wq;
1025
1026         if (WARN_ON_ONCE(!data->free_work))
1027                 return ERR_PTR(-EINVAL);
1028
1029         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1030         if (!wq)
1031                 return ERR_PTR(-ENOMEM);
1032
1033         wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1034         if (!wq->wqes) {
1035                 kfree(wq);
1036                 return ERR_PTR(-ENOMEM);
1037         }
1038
1039         wq->free_work = data->free_work;
1040
1041         /* caller must already hold a reference to this */
1042         wq->user = data->user;
1043
1044         for_each_node(node) {
1045                 struct io_wqe *wqe;
1046                 int alloc_node = node;
1047
1048                 if (!node_online(alloc_node))
1049                         alloc_node = NUMA_NO_NODE;
1050                 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1051                 if (!wqe)
1052                         goto err;
1053                 wq->wqes[node] = wqe;
1054                 wqe->node = alloc_node;
1055                 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1056                 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1057                 if (wq->user) {
1058                         wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1059                                         task_rlimit(current, RLIMIT_NPROC);
1060                 }
1061                 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1062                 wqe->wq = wq;
1063                 spin_lock_init(&wqe->lock);
1064                 INIT_WQ_LIST(&wqe->work_list);
1065                 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1066                 INIT_LIST_HEAD(&wqe->all_list);
1067         }
1068
1069         init_completion(&wq->done);
1070
1071         wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1072         if (!IS_ERR(wq->manager)) {
1073                 wake_up_process(wq->manager);
1074                 wait_for_completion(&wq->done);
1075                 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1076                         ret = -ENOMEM;
1077                         goto err;
1078                 }
1079                 refcount_set(&wq->use_refs, 1);
1080                 reinit_completion(&wq->done);
1081                 return wq;
1082         }
1083
1084         ret = PTR_ERR(wq->manager);
1085         complete(&wq->done);
1086 err:
1087         for_each_node(node)
1088                 kfree(wq->wqes[node]);
1089         kfree(wq->wqes);
1090         kfree(wq);
1091         return ERR_PTR(ret);
1092 }
1093
1094 bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1095 {
1096         if (data->free_work != wq->free_work)
1097                 return false;
1098
1099         return refcount_inc_not_zero(&wq->use_refs);
1100 }
1101
1102 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
1103 {
1104         wake_up_process(worker->task);
1105         return false;
1106 }
1107
1108 static void __io_wq_destroy(struct io_wq *wq)
1109 {
1110         int node;
1111
1112         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1113         if (wq->manager)
1114                 kthread_stop(wq->manager);
1115
1116         rcu_read_lock();
1117         for_each_node(node)
1118                 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1119         rcu_read_unlock();
1120
1121         wait_for_completion(&wq->done);
1122
1123         for_each_node(node)
1124                 kfree(wq->wqes[node]);
1125         kfree(wq->wqes);
1126         kfree(wq);
1127 }
1128
1129 void io_wq_destroy(struct io_wq *wq)
1130 {
1131         if (refcount_dec_and_test(&wq->use_refs))
1132                 __io_wq_destroy(wq);
1133 }
1134
1135 struct task_struct *io_wq_get_task(struct io_wq *wq)
1136 {
1137         return wq->manager;
1138 }