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