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/percpu.h>
  13 #include <linux/slab.h>
  14 #include <linux/rculist_nulls.h>
  15 #include <linux/cpu.h>
  16 #include <linux/tracehook.h>
  17 #include <uapi/linux/io_uring.h>
  18
  19 #include "io-wq.h"
  20
  21 #define WORKER_IDLE_TIMEOUT     (5 * HZ)
  22
  23 enum {
  24         IO_WORKER_F_UP          = 1,    /* up and active */
  25         IO_WORKER_F_RUNNING     = 2,    /* account as running */
  26         IO_WORKER_F_FREE        = 4,    /* worker on free list */
  27         IO_WORKER_F_BOUND       = 8,    /* is doing bounded work */
  28 };
  29
  30 enum {
  31         IO_WQ_BIT_EXIT          = 0,    /* wq exiting */
  32 };
  33
  34 enum {
  35         IO_ACCT_STALLED_BIT     = 0,    /* stalled on hash */
  36 };
  37
  38 /*
  39  * One for each thread in a wqe pool
  40  */
  41 struct io_worker {
  42         refcount_t ref;
  43         unsigned flags;
  44         struct hlist_nulls_node nulls_node;
  45         struct list_head all_list;
  46         struct task_struct *task;
  47         struct io_wqe *wqe;
  48
  49         struct io_wq_work *cur_work;
  50         spinlock_t lock;
  51
  52         struct completion ref_done;
  53
  54         unsigned long create_state;
  55         struct callback_head create_work;
  56         int create_index;
  57
  58         union {
  59                 struct rcu_head rcu;
  60                 struct work_struct work;
  61         };
  62 };
  63
  64 #if BITS_PER_LONG == 64
  65 #define IO_WQ_HASH_ORDER        6
  66 #else
  67 #define IO_WQ_HASH_ORDER        5
  68 #endif
  69
  70 #define IO_WQ_NR_HASH_BUCKETS   (1u << IO_WQ_HASH_ORDER)
  71
  72 struct io_wqe_acct {
  73         unsigned nr_workers;
  74         unsigned max_workers;
  75         int index;
  76         atomic_t nr_running;
  77         struct io_wq_work_list work_list;
  78         unsigned long flags;
  79 };
  80
  81 enum {
  82         IO_WQ_ACCT_BOUND,
  83         IO_WQ_ACCT_UNBOUND,
  84         IO_WQ_ACCT_NR,
  85 };
  86
  87 /*
  88  * Per-node worker thread pool
  89  */
  90 struct io_wqe {
  91         raw_spinlock_t lock;
  92         struct io_wqe_acct acct[2];
  93
  94         int node;
  95
  96         struct hlist_nulls_head free_list;
  97         struct list_head all_list;
  98
  99         struct wait_queue_entry wait;
 100
 101         struct io_wq *wq;
 102         struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
 103
 104         cpumask_var_t cpu_mask;
 105 };
 106
 107 /*
 108  * Per io_wq state
 109   */
 110 struct io_wq {
 111         unsigned long state;
 112
 113         free_work_fn *free_work;
 114         io_wq_work_fn *do_work;
 115
 116         struct io_wq_hash *hash;
 117
 118         atomic_t worker_refs;
 119         struct completion worker_done;
 120
 121         struct hlist_node cpuhp_node;
 122
 123         struct task_struct *task;
 124
 125         struct io_wqe *wqes[];
 126 };
 127
 128 static enum cpuhp_state io_wq_online;
 129
 130 struct io_cb_cancel_data {
 131         work_cancel_fn *fn;
 132         void *data;
 133         int nr_running;
 134         int nr_pending;
 135         bool cancel_all;
 136 };
 137
 138 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index);
 139 static void io_wqe_dec_running(struct io_worker *worker);
 140 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
 141                                         struct io_wqe_acct *acct,
 142                                         struct io_cb_cancel_data *match);
 143 static void create_worker_cb(struct callback_head *cb);
 144
 145 static bool io_worker_get(struct io_worker *worker)
 146 {
 147         return refcount_inc_not_zero(&worker->ref);
 148 }
 149
 150 static void io_worker_release(struct io_worker *worker)
 151 {
 152         if (refcount_dec_and_test(&worker->ref))
 153                 complete(&worker->ref_done);
 154 }
 155
 156 static inline struct io_wqe_acct *io_get_acct(struct io_wqe *wqe, bool bound)
 157 {
 158         return &wqe->acct[bound ? IO_WQ_ACCT_BOUND : IO_WQ_ACCT_UNBOUND];
 159 }
 160
 161 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 162                                                    struct io_wq_work *work)
 163 {
 164         return io_get_acct(wqe, !(work->flags & IO_WQ_WORK_UNBOUND));
 165 }
 166
 167 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
 168 {
 169         return io_get_acct(worker->wqe, worker->flags & IO_WORKER_F_BOUND);
 170 }
 171
 172 static void io_worker_ref_put(struct io_wq *wq)
 173 {
 174         if (atomic_dec_and_test(&wq->worker_refs))
 175                 complete(&wq->worker_done);
 176 }
 177
 178 static void io_worker_cancel_cb(struct io_worker *worker)
 179 {
 180         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 181         struct io_wqe *wqe = worker->wqe;
 182         struct io_wq *wq = wqe->wq;
 183
 184         atomic_dec(&acct->nr_running);
 185         raw_spin_lock(&worker->wqe->lock);
 186         acct->nr_workers--;
 187         raw_spin_unlock(&worker->wqe->lock);
 188         io_worker_ref_put(wq);
 189         clear_bit_unlock(0, &worker->create_state);
 190         io_worker_release(worker);
 191 }
 192
 193 static bool io_task_worker_match(struct callback_head *cb, void *data)
 194 {
 195         struct io_worker *worker;
 196
 197         if (cb->func != create_worker_cb)
 198                 return false;
 199         worker = container_of(cb, struct io_worker, create_work);
 200         return worker == data;
 201 }
 202
 203 static void io_worker_exit(struct io_worker *worker)
 204 {
 205         struct io_wqe *wqe = worker->wqe;
 206         struct io_wq *wq = wqe->wq;
 207
 208         while (1) {
 209                 struct callback_head *cb = task_work_cancel_match(wq->task,
 210                                                 io_task_worker_match, worker);
 211
 212                 if (!cb)
 213                         break;
 214                 io_worker_cancel_cb(worker);
 215         }
 216
 217         if (refcount_dec_and_test(&worker->ref))
 218                 complete(&worker->ref_done);
 219         wait_for_completion(&worker->ref_done);
 220
 221         raw_spin_lock(&wqe->lock);
 222         if (worker->flags & IO_WORKER_F_FREE)
 223                 hlist_nulls_del_rcu(&worker->nulls_node);
 224         list_del_rcu(&worker->all_list);
 225         preempt_disable();
 226         io_wqe_dec_running(worker);
 227         worker->flags = 0;
 228         current->flags &= ~PF_IO_WORKER;
 229         preempt_enable();
 230         raw_spin_unlock(&wqe->lock);
 231
 232         kfree_rcu(worker, rcu);
 233         io_worker_ref_put(wqe->wq);
 234         do_exit(0);
 235 }
 236
 237 static inline bool io_acct_run_queue(struct io_wqe_acct *acct)
 238 {
 239         if (!wq_list_empty(&acct->work_list) &&
 240             !test_bit(IO_ACCT_STALLED_BIT, &acct->flags))
 241                 return true;
 242         return false;
 243 }
 244
 245 /*
 246  * Check head of free list for an available worker. If one isn't available,
 247  * caller must create one.
 248  */
 249 static bool io_wqe_activate_free_worker(struct io_wqe *wqe,
 250                                         struct io_wqe_acct *acct)
 251         __must_hold(RCU)
 252 {
 253         struct hlist_nulls_node *n;
 254         struct io_worker *worker;
 255
 256         /*
 257          * Iterate free_list and see if we can find an idle worker to
 258          * activate. If a given worker is on the free_list but in the process
 259          * of exiting, keep trying.
 260          */
 261         hlist_nulls_for_each_entry_rcu(worker, n, &wqe->free_list, nulls_node) {
 262                 if (!io_worker_get(worker))
 263                         continue;
 264                 if (io_wqe_get_acct(worker) != acct) {
 265                         io_worker_release(worker);
 266                         continue;
 267                 }
 268                 if (wake_up_process(worker->task)) {
 269                         io_worker_release(worker);
 270                         return true;
 271                 }
 272                 io_worker_release(worker);
 273         }
 274
 275         return false;
 276 }
 277
 278 /*
 279  * We need a worker. If we find a free one, we're good. If not, and we're
 280  * below the max number of workers, create one.
 281  */
 282 static bool io_wqe_create_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 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         if (unlikely(!acct->max_workers))
 289                 pr_warn_once("io-wq is not configured for unbound workers");
 290
 291         raw_spin_lock(&wqe->lock);
 292         if (acct->nr_workers >= acct->max_workers) {
 293                 raw_spin_unlock(&wqe->lock);
 294                 return true;
 295         }
 296         acct->nr_workers++;
 297         raw_spin_unlock(&wqe->lock);
 298         atomic_inc(&acct->nr_running);
 299         atomic_inc(&wqe->wq->worker_refs);
 300         return create_io_worker(wqe->wq, wqe, acct->index);
 301 }
 302
 303 static void io_wqe_inc_running(struct io_worker *worker)
 304 {
 305         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 306
 307         atomic_inc(&acct->nr_running);
 308 }
 309
 310 static void create_worker_cb(struct callback_head *cb)
 311 {
 312         struct io_worker *worker;
 313         struct io_wq *wq;
 314         struct io_wqe *wqe;
 315         struct io_wqe_acct *acct;
 316         bool do_create = false;
 317
 318         worker = container_of(cb, struct io_worker, create_work);
 319         wqe = worker->wqe;
 320         wq = wqe->wq;
 321         acct = &wqe->acct[worker->create_index];
 322         raw_spin_lock(&wqe->lock);
 323         if (acct->nr_workers < acct->max_workers) {
 324                 acct->nr_workers++;
 325                 do_create = true;
 326         }
 327         raw_spin_unlock(&wqe->lock);
 328         if (do_create) {
 329                 create_io_worker(wq, wqe, worker->create_index);
 330         } else {
 331                 atomic_dec(&acct->nr_running);
 332                 io_worker_ref_put(wq);
 333         }
 334         clear_bit_unlock(0, &worker->create_state);
 335         io_worker_release(worker);
 336 }
 337
 338 static bool io_queue_worker_create(struct io_worker *worker,
 339                                    struct io_wqe_acct *acct,
 340                                    task_work_func_t func)
 341 {
 342         struct io_wqe *wqe = worker->wqe;
 343         struct io_wq *wq = wqe->wq;
 344
 345         /* raced with exit, just ignore create call */
 346         if (test_bit(IO_WQ_BIT_EXIT, &wq->state))
 347                 goto fail;
 348         if (!io_worker_get(worker))
 349                 goto fail;
 350         /*
 351          * create_state manages ownership of create_work/index. We should
 352          * only need one entry per worker, as the worker going to sleep
 353          * will trigger the condition, and waking will clear it once it
 354          * runs the task_work.
 355          */
 356         if (test_bit(0, &worker->create_state) ||
 357             test_and_set_bit_lock(0, &worker->create_state))
 358                 goto fail_release;
 359
 360         init_task_work(&worker->create_work, func);
 361         worker->create_index = acct->index;
 362         if (!task_work_add(wq->task, &worker->create_work, TWA_SIGNAL)) {
 363                 clear_bit_unlock(0, &worker->create_state);
 364                 return true;
 365         }
 366         clear_bit_unlock(0, &worker->create_state);
 367 fail_release:
 368         io_worker_release(worker);
 369 fail:
 370         atomic_dec(&acct->nr_running);
 371         io_worker_ref_put(wq);
 372         return false;
 373 }
 374
 375 static void io_wqe_dec_running(struct io_worker *worker)
 376         __must_hold(wqe->lock)
 377 {
 378         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 379         struct io_wqe *wqe = worker->wqe;
 380
 381         if (!(worker->flags & IO_WORKER_F_UP))
 382                 return;
 383
 384         if (atomic_dec_and_test(&acct->nr_running) && io_acct_run_queue(acct)) {
 385                 atomic_inc(&acct->nr_running);
 386                 atomic_inc(&wqe->wq->worker_refs);
 387                 io_queue_worker_create(worker, acct, create_worker_cb);
 388         }
 389 }
 390
 391 /*
 392  * Worker will start processing some work. Move it to the busy list, if
 393  * it's currently on the freelist
 394  */
 395 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
 396                              struct io_wq_work *work)
 397         __must_hold(wqe->lock)
 398 {
 399         if (worker->flags & IO_WORKER_F_FREE) {
 400                 worker->flags &= ~IO_WORKER_F_FREE;
 401                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 402         }
 403 }
 404
 405 /*
 406  * No work, worker going to sleep. Move to freelist, and unuse mm if we
 407  * have one attached. Dropping the mm may potentially sleep, so we drop
 408  * the lock in that case and return success. Since the caller has to
 409  * retry the loop in that case (we changed task state), we don't regrab
 410  * the lock if we return success.
 411  */
 412 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
 413         __must_hold(wqe->lock)
 414 {
 415         if (!(worker->flags & IO_WORKER_F_FREE)) {
 416                 worker->flags |= IO_WORKER_F_FREE;
 417                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 418         }
 419 }
 420
 421 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
 422 {
 423         return work->flags >> IO_WQ_HASH_SHIFT;
 424 }
 425
 426 static bool io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
 427 {
 428         struct io_wq *wq = wqe->wq;
 429         bool ret = false;
 430
 431         spin_lock_irq(&wq->hash->wait.lock);
 432         if (list_empty(&wqe->wait.entry)) {
 433                 __add_wait_queue(&wq->hash->wait, &wqe->wait);
 434                 if (!test_bit(hash, &wq->hash->map)) {
 435                         __set_current_state(TASK_RUNNING);
 436                         list_del_init(&wqe->wait.entry);
 437                         ret = true;
 438                 }
 439         }
 440         spin_unlock_irq(&wq->hash->wait.lock);
 441         return ret;
 442 }
 443
 444 static struct io_wq_work *io_get_next_work(struct io_wqe_acct *acct,
 445                                            struct io_worker *worker)
 446         __must_hold(wqe->lock)
 447 {
 448         struct io_wq_work_node *node, *prev;
 449         struct io_wq_work *work, *tail;
 450         unsigned int stall_hash = -1U;
 451         struct io_wqe *wqe = worker->wqe;
 452
 453         wq_list_for_each(node, prev, &acct->work_list) {
 454                 unsigned int hash;
 455
 456                 work = container_of(node, struct io_wq_work, list);
 457
 458                 /* not hashed, can run anytime */
 459                 if (!io_wq_is_hashed(work)) {
 460                         wq_list_del(&acct->work_list, node, prev);
 461                         return work;
 462                 }
 463
 464                 hash = io_get_work_hash(work);
 465                 /* all items with this hash lie in [work, tail] */
 466                 tail = wqe->hash_tail[hash];
 467
 468                 /* hashed, can run if not already running */
 469                 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
 470                         wqe->hash_tail[hash] = NULL;
 471                         wq_list_cut(&acct->work_list, &tail->list, prev);
 472                         return work;
 473                 }
 474                 if (stall_hash == -1U)
 475                         stall_hash = hash;
 476                 /* fast forward to a next hash, for-each will fix up @prev */
 477                 node = &tail->list;
 478         }
 479
 480         if (stall_hash != -1U) {
 481                 bool unstalled;
 482
 483                 /*
 484                  * Set this before dropping the lock to avoid racing with new
 485                  * work being added and clearing the stalled bit.
 486                  */
 487                 set_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 488                 raw_spin_unlock(&wqe->lock);
 489                 unstalled = io_wait_on_hash(wqe, stall_hash);
 490                 raw_spin_lock(&wqe->lock);
 491                 if (unstalled) {
 492                         clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 493                         if (wq_has_sleeper(&wqe->wq->hash->wait))
 494                                 wake_up(&wqe->wq->hash->wait);
 495                 }
 496         }
 497
 498         return NULL;
 499 }
 500
 501 static bool io_flush_signals(void)
 502 {
 503         if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
 504                 __set_current_state(TASK_RUNNING);
 505                 tracehook_notify_signal();
 506                 return true;
 507         }
 508         return false;
 509 }
 510
 511 static void io_assign_current_work(struct io_worker *worker,
 512                                    struct io_wq_work *work)
 513 {
 514         if (work) {
 515                 io_flush_signals();
 516                 cond_resched();
 517         }
 518
 519         spin_lock(&worker->lock);
 520         worker->cur_work = work;
 521         spin_unlock(&worker->lock);
 522 }
 523
 524 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
 525
 526 static void io_worker_handle_work(struct io_worker *worker)
 527         __releases(wqe->lock)
 528 {
 529         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 530         struct io_wqe *wqe = worker->wqe;
 531         struct io_wq *wq = wqe->wq;
 532         bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
 533
 534         do {
 535                 struct io_wq_work *work;
 536 get_next:
 537                 /*
 538                  * If we got some work, mark us as busy. If we didn't, but
 539                  * the list isn't empty, it means we stalled on hashed work.
 540                  * Mark us stalled so we don't keep looking for work when we
 541                  * can't make progress, any work completion or insertion will
 542                  * clear the stalled flag.
 543                  */
 544                 work = io_get_next_work(acct, worker);
 545                 if (work)
 546                         __io_worker_busy(wqe, worker, work);
 547
 548                 raw_spin_unlock(&wqe->lock);
 549                 if (!work)
 550                         break;
 551                 io_assign_current_work(worker, work);
 552                 __set_current_state(TASK_RUNNING);
 553
 554                 /* handle a whole dependent link */
 555                 do {
 556                         struct io_wq_work *next_hashed, *linked;
 557                         unsigned int hash = io_get_work_hash(work);
 558
 559                         next_hashed = wq_next_work(work);
 560
 561                         if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
 562                                 work->flags |= IO_WQ_WORK_CANCEL;
 563                         wq->do_work(work);
 564                         io_assign_current_work(worker, NULL);
 565
 566                         linked = wq->free_work(work);
 567                         work = next_hashed;
 568                         if (!work && linked && !io_wq_is_hashed(linked)) {
 569                                 work = linked;
 570                                 linked = NULL;
 571                         }
 572                         io_assign_current_work(worker, work);
 573                         if (linked)
 574                                 io_wqe_enqueue(wqe, linked);
 575
 576                         if (hash != -1U && !next_hashed) {
 577                                 /* serialize hash clear with wake_up() */
 578                                 spin_lock_irq(&wq->hash->wait.lock);
 579                                 clear_bit(hash, &wq->hash->map);
 580                                 clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 581                                 spin_unlock_irq(&wq->hash->wait.lock);
 582                                 if (wq_has_sleeper(&wq->hash->wait))
 583                                         wake_up(&wq->hash->wait);
 584                                 raw_spin_lock(&wqe->lock);
 585                                 /* skip unnecessary unlock-lock wqe->lock */
 586                                 if (!work)
 587                                         goto get_next;
 588                                 raw_spin_unlock(&wqe->lock);
 589                         }
 590                 } while (work);
 591
 592                 raw_spin_lock(&wqe->lock);
 593         } while (1);
 594 }
 595
 596 static int io_wqe_worker(void *data)
 597 {
 598         struct io_worker *worker = data;
 599         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 600         struct io_wqe *wqe = worker->wqe;
 601         struct io_wq *wq = wqe->wq;
 602         bool last_timeout = false;
 603         char buf[TASK_COMM_LEN];
 604
 605         worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 606
 607         snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task->pid);
 608         set_task_comm(current, buf);
 609
 610         while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 611                 long ret;
 612
 613                 set_current_state(TASK_INTERRUPTIBLE);
 614 loop:
 615                 raw_spin_lock(&wqe->lock);
 616                 if (io_acct_run_queue(acct)) {
 617                         io_worker_handle_work(worker);
 618                         goto loop;
 619                 }
 620                 /* timed out, exit unless we're the last worker */
 621                 if (last_timeout && acct->nr_workers > 1) {
 622                         acct->nr_workers--;
 623                         raw_spin_unlock(&wqe->lock);
 624                         __set_current_state(TASK_RUNNING);
 625                         break;
 626                 }
 627                 last_timeout = false;
 628                 __io_worker_idle(wqe, worker);
 629                 raw_spin_unlock(&wqe->lock);
 630                 if (io_flush_signals())
 631                         continue;
 632                 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
 633                 if (signal_pending(current)) {
 634                         struct ksignal ksig;
 635
 636                         if (!get_signal(&ksig))
 637                                 continue;
 638                         break;
 639                 }
 640                 last_timeout = !ret;
 641         }
 642
 643         if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 644                 raw_spin_lock(&wqe->lock);
 645                 io_worker_handle_work(worker);
 646         }
 647
 648         io_worker_exit(worker);
 649         return 0;
 650 }
 651
 652 /*
 653  * Called when a worker is scheduled in. Mark us as currently running.
 654  */
 655 void io_wq_worker_running(struct task_struct *tsk)
 656 {
 657         struct io_worker *worker = tsk->pf_io_worker;
 658
 659         if (!worker)
 660                 return;
 661         if (!(worker->flags & IO_WORKER_F_UP))
 662                 return;
 663         if (worker->flags & IO_WORKER_F_RUNNING)
 664                 return;
 665         worker->flags |= IO_WORKER_F_RUNNING;
 666         io_wqe_inc_running(worker);
 667 }
 668
 669 /*
 670  * Called when worker is going to sleep. If there are no workers currently
 671  * running and we have work pending, wake up a free one or create a new one.
 672  */
 673 void io_wq_worker_sleeping(struct task_struct *tsk)
 674 {
 675         struct io_worker *worker = tsk->pf_io_worker;
 676
 677         if (!worker)
 678                 return;
 679         if (!(worker->flags & IO_WORKER_F_UP))
 680                 return;
 681         if (!(worker->flags & IO_WORKER_F_RUNNING))
 682                 return;
 683
 684         worker->flags &= ~IO_WORKER_F_RUNNING;
 685
 686         raw_spin_lock(&worker->wqe->lock);
 687         io_wqe_dec_running(worker);
 688         raw_spin_unlock(&worker->wqe->lock);
 689 }
 690
 691 static void io_init_new_worker(struct io_wqe *wqe, struct io_worker *worker,
 692                                struct task_struct *tsk)
 693 {
 694         tsk->pf_io_worker = worker;
 695         worker->task = tsk;
 696         set_cpus_allowed_ptr(tsk, wqe->cpu_mask);
 697         tsk->flags |= PF_NO_SETAFFINITY;
 698
 699         raw_spin_lock(&wqe->lock);
 700         hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 701         list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 702         worker->flags |= IO_WORKER_F_FREE;
 703         raw_spin_unlock(&wqe->lock);
 704         wake_up_new_task(tsk);
 705 }
 706
 707 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
 708 {
 709         return true;
 710 }
 711
 712 static inline bool io_should_retry_thread(long err)
 713 {
 714         /*
 715          * Prevent perpetual task_work retry, if the task (or its group) is
 716          * exiting.
 717          */
 718         if (fatal_signal_pending(current))
 719                 return false;
 720
 721         switch (err) {
 722         case -EAGAIN:
 723         case -ERESTARTSYS:
 724         case -ERESTARTNOINTR:
 725         case -ERESTARTNOHAND:
 726                 return true;
 727         default:
 728                 return false;
 729         }
 730 }
 731
 732 static void create_worker_cont(struct callback_head *cb)
 733 {
 734         struct io_worker *worker;
 735         struct task_struct *tsk;
 736         struct io_wqe *wqe;
 737
 738         worker = container_of(cb, struct io_worker, create_work);
 739         clear_bit_unlock(0, &worker->create_state);
 740         wqe = worker->wqe;
 741         tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
 742         if (!IS_ERR(tsk)) {
 743                 io_init_new_worker(wqe, worker, tsk);
 744                 io_worker_release(worker);
 745                 return;
 746         } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
 747                 struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 748
 749                 atomic_dec(&acct->nr_running);
 750                 raw_spin_lock(&wqe->lock);
 751                 acct->nr_workers--;
 752                 if (!acct->nr_workers) {
 753                         struct io_cb_cancel_data match = {
 754                                 .fn             = io_wq_work_match_all,
 755                                 .cancel_all     = true,
 756                         };
 757
 758                         while (io_acct_cancel_pending_work(wqe, acct, &match))
 759                                 raw_spin_lock(&wqe->lock);
 760                 }
 761                 raw_spin_unlock(&wqe->lock);
 762                 io_worker_ref_put(wqe->wq);
 763                 kfree(worker);
 764                 return;
 765         }
 766
 767         /* re-create attempts grab a new worker ref, drop the existing one */
 768         io_worker_release(worker);
 769         schedule_work(&worker->work);
 770 }
 771
 772 static void io_workqueue_create(struct work_struct *work)
 773 {
 774         struct io_worker *worker = container_of(work, struct io_worker, work);
 775         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 776
 777         if (!io_queue_worker_create(worker, acct, create_worker_cont))
 778                 kfree(worker);
 779 }
 780
 781 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
 782 {
 783         struct io_wqe_acct *acct = &wqe->acct[index];
 784         struct io_worker *worker;
 785         struct task_struct *tsk;
 786
 787         __set_current_state(TASK_RUNNING);
 788
 789         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
 790         if (!worker) {
 791 fail:
 792                 atomic_dec(&acct->nr_running);
 793                 raw_spin_lock(&wqe->lock);
 794                 acct->nr_workers--;
 795                 raw_spin_unlock(&wqe->lock);
 796                 io_worker_ref_put(wq);
 797                 return false;
 798         }
 799
 800         refcount_set(&worker->ref, 1);
 801         worker->wqe = wqe;
 802         spin_lock_init(&worker->lock);
 803         init_completion(&worker->ref_done);
 804
 805         if (index == IO_WQ_ACCT_BOUND)
 806                 worker->flags |= IO_WORKER_F_BOUND;
 807
 808         tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
 809         if (!IS_ERR(tsk)) {
 810                 io_init_new_worker(wqe, worker, tsk);
 811         } else if (!io_should_retry_thread(PTR_ERR(tsk))) {
 812                 kfree(worker);
 813                 goto fail;
 814         } else {
 815                 INIT_WORK(&worker->work, io_workqueue_create);
 816                 schedule_work(&worker->work);
 817         }
 818
 819         return true;
 820 }
 821
 822 /*
 823  * Iterate the passed in list and call the specific function for each
 824  * worker that isn't exiting
 825  */
 826 static bool io_wq_for_each_worker(struct io_wqe *wqe,
 827                                   bool (*func)(struct io_worker *, void *),
 828                                   void *data)
 829 {
 830         struct io_worker *worker;
 831         bool ret = false;
 832
 833         list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
 834                 if (io_worker_get(worker)) {
 835                         /* no task if node is/was offline */
 836                         if (worker->task)
 837                                 ret = func(worker, data);
 838                         io_worker_release(worker);
 839                         if (ret)
 840                                 break;
 841                 }
 842         }
 843
 844         return ret;
 845 }
 846
 847 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
 848 {
 849         set_notify_signal(worker->task);
 850         wake_up_process(worker->task);
 851         return false;
 852 }
 853
 854 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
 855 {
 856         struct io_wq *wq = wqe->wq;
 857
 858         do {
 859                 work->flags |= IO_WQ_WORK_CANCEL;
 860                 wq->do_work(work);
 861                 work = wq->free_work(work);
 862         } while (work);
 863 }
 864
 865 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
 866 {
 867         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 868         unsigned int hash;
 869         struct io_wq_work *tail;
 870
 871         if (!io_wq_is_hashed(work)) {
 872 append:
 873                 wq_list_add_tail(&work->list, &acct->work_list);
 874                 return;
 875         }
 876
 877         hash = io_get_work_hash(work);
 878         tail = wqe->hash_tail[hash];
 879         wqe->hash_tail[hash] = work;
 880         if (!tail)
 881                 goto append;
 882
 883         wq_list_add_after(&work->list, &tail->list, &acct->work_list);
 884 }
 885
 886 static bool io_wq_work_match_item(struct io_wq_work *work, void *data)
 887 {
 888         return work == data;
 889 }
 890
 891 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 892 {
 893         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 894         unsigned work_flags = work->flags;
 895         bool do_create;
 896
 897         /*
 898          * If io-wq is exiting for this task, or if the request has explicitly
 899          * been marked as one that should not get executed, cancel it here.
 900          */
 901         if (test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state) ||
 902             (work->flags & IO_WQ_WORK_CANCEL)) {
 903                 io_run_cancel(work, wqe);
 904                 return;
 905         }
 906
 907         raw_spin_lock(&wqe->lock);
 908         io_wqe_insert_work(wqe, work);
 909         clear_bit(IO_ACCT_STALLED_BIT, &acct->flags);
 910
 911         rcu_read_lock();
 912         do_create = !io_wqe_activate_free_worker(wqe, acct);
 913         rcu_read_unlock();
 914
 915         raw_spin_unlock(&wqe->lock);
 916
 917         if (do_create && ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 918             !atomic_read(&acct->nr_running))) {
 919                 bool did_create;
 920
 921                 did_create = io_wqe_create_worker(wqe, acct);
 922                 if (likely(did_create))
 923                         return;
 924
 925                 raw_spin_lock(&wqe->lock);
 926                 /* fatal condition, failed to create the first worker */
 927                 if (!acct->nr_workers) {
 928                         struct io_cb_cancel_data match = {
 929                                 .fn             = io_wq_work_match_item,
 930                                 .data           = work,
 931                                 .cancel_all     = false,
 932                         };
 933
 934                         if (io_acct_cancel_pending_work(wqe, acct, &match))
 935                                 raw_spin_lock(&wqe->lock);
 936                 }
 937                 raw_spin_unlock(&wqe->lock);
 938         }
 939 }
 940
 941 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 942 {
 943         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 944
 945         io_wqe_enqueue(wqe, work);
 946 }
 947
 948 /*
 949  * Work items that hash to the same value will not be done in parallel.
 950  * Used to limit concurrent writes, generally hashed by inode.
 951  */
 952 void io_wq_hash_work(struct io_wq_work *work, void *val)
 953 {
 954         unsigned int bit;
 955
 956         bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 957         work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 958 }
 959
 960 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 961 {
 962         struct io_cb_cancel_data *match = data;
 963
 964         /*
 965          * Hold the lock to avoid ->cur_work going out of scope, caller
 966          * may dereference the passed in work.
 967          */
 968         spin_lock(&worker->lock);
 969         if (worker->cur_work &&
 970             match->fn(worker->cur_work, match->data)) {
 971                 set_notify_signal(worker->task);
 972                 match->nr_running++;
 973         }
 974         spin_unlock(&worker->lock);
 975
 976         return match->nr_running && !match->cancel_all;
 977 }
 978
 979 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
 980                                          struct io_wq_work *work,
 981                                          struct io_wq_work_node *prev)
 982 {
 983         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 984         unsigned int hash = io_get_work_hash(work);
 985         struct io_wq_work *prev_work = NULL;
 986
 987         if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
 988                 if (prev)
 989                         prev_work = container_of(prev, struct io_wq_work, list);
 990                 if (prev_work && io_get_work_hash(prev_work) == hash)
 991                         wqe->hash_tail[hash] = prev_work;
 992                 else
 993                         wqe->hash_tail[hash] = NULL;
 994         }
 995         wq_list_del(&acct->work_list, &work->list, prev);
 996 }
 997
 998 static bool io_acct_cancel_pending_work(struct io_wqe *wqe,
 999                                         struct io_wqe_acct *acct,
1000                                         struct io_cb_cancel_data *match)
1001         __releases(wqe->lock)
1002 {
1003         struct io_wq_work_node *node, *prev;
1004         struct io_wq_work *work;
1005
1006         wq_list_for_each(node, prev, &acct->work_list) {
1007                 work = container_of(node, struct io_wq_work, list);
1008                 if (!match->fn(work, match->data))
1009                         continue;
1010                 io_wqe_remove_pending(wqe, work, prev);
1011                 raw_spin_unlock(&wqe->lock);
1012                 io_run_cancel(work, wqe);
1013                 match->nr_pending++;
1014                 /* not safe to continue after unlock */
1015                 return true;
1016         }
1017
1018         return false;
1019 }
1020
1021 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1022                                        struct io_cb_cancel_data *match)
1023 {
1024         int i;
1025 retry:
1026         raw_spin_lock(&wqe->lock);
1027         for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1028                 struct io_wqe_acct *acct = io_get_acct(wqe, i == 0);
1029
1030                 if (io_acct_cancel_pending_work(wqe, acct, match)) {
1031                         if (match->cancel_all)
1032                                 goto retry;
1033                         return;
1034                 }
1035         }
1036         raw_spin_unlock(&wqe->lock);
1037 }
1038
1039 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1040                                        struct io_cb_cancel_data *match)
1041 {
1042         rcu_read_lock();
1043         io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1044         rcu_read_unlock();
1045 }
1046
1047 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1048                                   void *data, bool cancel_all)
1049 {
1050         struct io_cb_cancel_data match = {
1051                 .fn             = cancel,
1052                 .data           = data,
1053                 .cancel_all     = cancel_all,
1054         };
1055         int node;
1056
1057         /*
1058          * First check pending list, if we're lucky we can just remove it
1059          * from there. CANCEL_OK means that the work is returned as-new,
1060          * no completion will be posted for it.
1061          */
1062         for_each_node(node) {
1063                 struct io_wqe *wqe = wq->wqes[node];
1064
1065                 io_wqe_cancel_pending_work(wqe, &match);
1066                 if (match.nr_pending && !match.cancel_all)
1067                         return IO_WQ_CANCEL_OK;
1068         }
1069
1070         /*
1071          * Now check if a free (going busy) or busy worker has the work
1072          * currently running. If we find it there, we'll return CANCEL_RUNNING
1073          * as an indication that we attempt to signal cancellation. The
1074          * completion will run normally in this case.
1075          */
1076         for_each_node(node) {
1077                 struct io_wqe *wqe = wq->wqes[node];
1078
1079                 io_wqe_cancel_running_work(wqe, &match);
1080                 if (match.nr_running && !match.cancel_all)
1081                         return IO_WQ_CANCEL_RUNNING;
1082         }
1083
1084         if (match.nr_running)
1085                 return IO_WQ_CANCEL_RUNNING;
1086         if (match.nr_pending)
1087                 return IO_WQ_CANCEL_OK;
1088         return IO_WQ_CANCEL_NOTFOUND;
1089 }
1090
1091 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
1092                             int sync, void *key)
1093 {
1094         struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
1095         int i;
1096
1097         list_del_init(&wait->entry);
1098
1099         rcu_read_lock();
1100         for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1101                 struct io_wqe_acct *acct = &wqe->acct[i];
1102
1103                 if (test_and_clear_bit(IO_ACCT_STALLED_BIT, &acct->flags))
1104                         io_wqe_activate_free_worker(wqe, acct);
1105         }
1106         rcu_read_unlock();
1107         return 1;
1108 }
1109
1110 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1111 {
1112         int ret, node, i;
1113         struct io_wq *wq;
1114
1115         if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1116                 return ERR_PTR(-EINVAL);
1117         if (WARN_ON_ONCE(!bounded))
1118                 return ERR_PTR(-EINVAL);
1119
1120         wq = kzalloc(struct_size(wq, wqes, nr_node_ids), GFP_KERNEL);
1121         if (!wq)
1122                 return ERR_PTR(-ENOMEM);
1123         ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1124         if (ret)
1125                 goto err_wq;
1126
1127         refcount_inc(&data->hash->refs);
1128         wq->hash = data->hash;
1129         wq->free_work = data->free_work;
1130         wq->do_work = data->do_work;
1131
1132         ret = -ENOMEM;
1133         for_each_node(node) {
1134                 struct io_wqe *wqe;
1135                 int alloc_node = node;
1136
1137                 if (!node_online(alloc_node))
1138                         alloc_node = NUMA_NO_NODE;
1139                 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1140                 if (!wqe)
1141                         goto err;
1142                 if (!alloc_cpumask_var(&wqe->cpu_mask, GFP_KERNEL))
1143                         goto err;
1144                 cpumask_copy(wqe->cpu_mask, cpumask_of_node(node));
1145                 wq->wqes[node] = wqe;
1146                 wqe->node = alloc_node;
1147                 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1148                 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1149                                         task_rlimit(current, RLIMIT_NPROC);
1150                 INIT_LIST_HEAD(&wqe->wait.entry);
1151                 wqe->wait.func = io_wqe_hash_wake;
1152                 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1153                         struct io_wqe_acct *acct = &wqe->acct[i];
1154
1155                         acct->index = i;
1156                         atomic_set(&acct->nr_running, 0);
1157                         INIT_WQ_LIST(&acct->work_list);
1158                 }
1159                 wqe->wq = wq;
1160                 raw_spin_lock_init(&wqe->lock);
1161                 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1162                 INIT_LIST_HEAD(&wqe->all_list);
1163         }
1164
1165         wq->task = get_task_struct(data->task);
1166         atomic_set(&wq->worker_refs, 1);
1167         init_completion(&wq->worker_done);
1168         return wq;
1169 err:
1170         io_wq_put_hash(data->hash);
1171         cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1172         for_each_node(node) {
1173                 if (!wq->wqes[node])
1174                         continue;
1175                 free_cpumask_var(wq->wqes[node]->cpu_mask);
1176                 kfree(wq->wqes[node]);
1177         }
1178 err_wq:
1179         kfree(wq);
1180         return ERR_PTR(ret);
1181 }
1182
1183 static bool io_task_work_match(struct callback_head *cb, void *data)
1184 {
1185         struct io_worker *worker;
1186
1187         if (cb->func != create_worker_cb && cb->func != create_worker_cont)
1188                 return false;
1189         worker = container_of(cb, struct io_worker, create_work);
1190         return worker->wqe->wq == data;
1191 }
1192
1193 void io_wq_exit_start(struct io_wq *wq)
1194 {
1195         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1196 }
1197
1198 static void io_wq_exit_workers(struct io_wq *wq)
1199 {
1200         struct callback_head *cb;
1201         int node;
1202
1203         if (!wq->task)
1204                 return;
1205
1206         while ((cb = task_work_cancel_match(wq->task, io_task_work_match, wq)) != NULL) {
1207                 struct io_worker *worker;
1208
1209                 worker = container_of(cb, struct io_worker, create_work);
1210                 io_worker_cancel_cb(worker);
1211         }
1212
1213         rcu_read_lock();
1214         for_each_node(node) {
1215                 struct io_wqe *wqe = wq->wqes[node];
1216
1217                 io_wq_for_each_worker(wqe, io_wq_worker_wake, NULL);
1218         }
1219         rcu_read_unlock();
1220         io_worker_ref_put(wq);
1221         wait_for_completion(&wq->worker_done);
1222
1223         for_each_node(node) {
1224                 spin_lock_irq(&wq->hash->wait.lock);
1225                 list_del_init(&wq->wqes[node]->wait.entry);
1226                 spin_unlock_irq(&wq->hash->wait.lock);
1227         }
1228         put_task_struct(wq->task);
1229         wq->task = NULL;
1230 }
1231
1232 static void io_wq_destroy(struct io_wq *wq)
1233 {
1234         int node;
1235
1236         cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1237
1238         for_each_node(node) {
1239                 struct io_wqe *wqe = wq->wqes[node];
1240                 struct io_cb_cancel_data match = {
1241                         .fn             = io_wq_work_match_all,
1242                         .cancel_all     = true,
1243                 };
1244                 io_wqe_cancel_pending_work(wqe, &match);
1245                 free_cpumask_var(wqe->cpu_mask);
1246                 kfree(wqe);
1247         }
1248         io_wq_put_hash(wq->hash);
1249         kfree(wq);
1250 }
1251
1252 void io_wq_put_and_exit(struct io_wq *wq)
1253 {
1254         WARN_ON_ONCE(!test_bit(IO_WQ_BIT_EXIT, &wq->state));
1255
1256         io_wq_exit_workers(wq);
1257         io_wq_destroy(wq);
1258 }
1259
1260 struct online_data {
1261         unsigned int cpu;
1262         bool online;
1263 };
1264
1265 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1266 {
1267         struct online_data *od = data;
1268
1269         if (od->online)
1270                 cpumask_set_cpu(od->cpu, worker->wqe->cpu_mask);
1271         else
1272                 cpumask_clear_cpu(od->cpu, worker->wqe->cpu_mask);
1273         return false;
1274 }
1275
1276 static int __io_wq_cpu_online(struct io_wq *wq, unsigned int cpu, bool online)
1277 {
1278         struct online_data od = {
1279                 .cpu = cpu,
1280                 .online = online
1281         };
1282         int i;
1283
1284         rcu_read_lock();
1285         for_each_node(i)
1286                 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, &od);
1287         rcu_read_unlock();
1288         return 0;
1289 }
1290
1291 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1292 {
1293         struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1294
1295         return __io_wq_cpu_online(wq, cpu, true);
1296 }
1297
1298 static int io_wq_cpu_offline(unsigned int cpu, struct hlist_node *node)
1299 {
1300         struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1301
1302         return __io_wq_cpu_online(wq, cpu, false);
1303 }
1304
1305 int io_wq_cpu_affinity(struct io_wq *wq, cpumask_var_t mask)
1306 {
1307         int i;
1308
1309         rcu_read_lock();
1310         for_each_node(i) {
1311                 struct io_wqe *wqe = wq->wqes[i];
1312
1313                 if (mask)
1314                         cpumask_copy(wqe->cpu_mask, mask);
1315                 else
1316                         cpumask_copy(wqe->cpu_mask, cpumask_of_node(i));
1317         }
1318         rcu_read_unlock();
1319         return 0;
1320 }
1321
1322 /*
1323  * Set max number of unbounded workers, returns old value. If new_count is 0,
1324  * then just return the old value.
1325  */
1326 int io_wq_max_workers(struct io_wq *wq, int *new_count)
1327 {
1328         int prev[IO_WQ_ACCT_NR];
1329         bool first_node = true;
1330         int i, node;
1331
1332         BUILD_BUG_ON((int) IO_WQ_ACCT_BOUND   != (int) IO_WQ_BOUND);
1333         BUILD_BUG_ON((int) IO_WQ_ACCT_UNBOUND != (int) IO_WQ_UNBOUND);
1334         BUILD_BUG_ON((int) IO_WQ_ACCT_NR      != 2);
1335
1336         for (i = 0; i < 2; i++) {
1337                 if (new_count[i] > task_rlimit(current, RLIMIT_NPROC))
1338                         new_count[i] = task_rlimit(current, RLIMIT_NPROC);
1339         }
1340
1341         for (i = 0; i < IO_WQ_ACCT_NR; i++)
1342                 prev[i] = 0;
1343
1344         rcu_read_lock();
1345         for_each_node(node) {
1346                 struct io_wqe *wqe = wq->wqes[node];
1347                 struct io_wqe_acct *acct;
1348
1349                 raw_spin_lock(&wqe->lock);
1350                 for (i = 0; i < IO_WQ_ACCT_NR; i++) {
1351                         acct = &wqe->acct[i];
1352                         if (first_node)
1353                                 prev[i] = max_t(int, acct->max_workers, prev[i]);
1354                         if (new_count[i])
1355                                 acct->max_workers = new_count[i];
1356                 }
1357                 raw_spin_unlock(&wqe->lock);
1358                 first_node = false;
1359         }
1360         rcu_read_unlock();
1361
1362         for (i = 0; i < IO_WQ_ACCT_NR; i++)
1363                 new_count[i] = prev[i];
1364
1365         return 0;
1366 }
1367
1368 static __init int io_wq_init(void)
1369 {
1370         int ret;
1371
1372         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1373                                         io_wq_cpu_online, io_wq_cpu_offline);
1374         if (ret < 0)
1375                 return ret;
1376         io_wq_online = ret;
1377         return 0;
1378 }
1379 subsys_initcall(io_wq_init);