fs/io-wq.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Basic worker thread pool for io_uring
   4  *
   5  * Copyright (C) 2019 Jens Axboe
   6  *
   7  */
   8 #include <linux/kernel.h>
   9 #include <linux/init.h>
  10 #include <linux/errno.h>
  11 #include <linux/sched/signal.h>
  12 #include <linux/mm.h>
  13 #include <linux/sched/mm.h>
  14 #include <linux/percpu.h>
  15 #include <linux/slab.h>
  16 #include <linux/rculist_nulls.h>
  17 #include <linux/cpu.h>
  18 #include <linux/tracehook.h>
  19
  20 #include "../kernel/sched/sched.h"
  21 #include "io-wq.h"
  22
  23 #define WORKER_IDLE_TIMEOUT     (5 * HZ)
  24
  25 enum {
  26         IO_WORKER_F_UP          = 1,    /* up and active */
  27         IO_WORKER_F_RUNNING     = 2,    /* account as running */
  28         IO_WORKER_F_FREE        = 4,    /* worker on free list */
  29         IO_WORKER_F_FIXED       = 8,    /* static idle worker */
  30         IO_WORKER_F_BOUND       = 16,   /* is doing bounded work */
  31 };
  32
  33 enum {
  34         IO_WQ_BIT_EXIT          = 0,    /* wq exiting */
  35 };
  36
  37 enum {
  38         IO_WQE_FLAG_STALLED     = 1,    /* stalled on hash */
  39 };
  40
  41 /*
  42  * One for each thread in a wqe pool
  43  */
  44 struct io_worker {
  45         refcount_t ref;
  46         unsigned flags;
  47         struct hlist_nulls_node nulls_node;
  48         struct list_head all_list;
  49         struct task_struct *task;
  50         struct io_wqe *wqe;
  51
  52         struct io_wq_work *cur_work;
  53         spinlock_t lock;
  54
  55         struct completion ref_done;
  56
  57         struct rcu_head rcu;
  58 };
  59
  60 #if BITS_PER_LONG == 64
  61 #define IO_WQ_HASH_ORDER        6
  62 #else
  63 #define IO_WQ_HASH_ORDER        5
  64 #endif
  65
  66 #define IO_WQ_NR_HASH_BUCKETS   (1u << IO_WQ_HASH_ORDER)
  67
  68 struct io_wqe_acct {
  69         unsigned nr_workers;
  70         unsigned max_workers;
  71         atomic_t nr_running;
  72 };
  73
  74 enum {
  75         IO_WQ_ACCT_BOUND,
  76         IO_WQ_ACCT_UNBOUND,
  77 };
  78
  79 /*
  80  * Per-node worker thread pool
  81  */
  82 struct io_wqe {
  83         struct {
  84                 raw_spinlock_t lock;
  85                 struct io_wq_work_list work_list;
  86                 unsigned flags;
  87         } ____cacheline_aligned_in_smp;
  88
  89         int node;
  90         struct io_wqe_acct acct[2];
  91
  92         struct hlist_nulls_head free_list;
  93         struct list_head all_list;
  94
  95         struct wait_queue_entry wait;
  96
  97         struct io_wq *wq;
  98         struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
  99 };
 100
 101 /*
 102  * Per io_wq state
 103   */
 104 struct io_wq {
 105         struct io_wqe **wqes;
 106         unsigned long state;
 107
 108         free_work_fn *free_work;
 109         io_wq_work_fn *do_work;
 110
 111         struct task_struct *manager;
 112
 113         struct io_wq_hash *hash;
 114
 115         refcount_t refs;
 116         struct completion exited;
 117
 118         atomic_t worker_refs;
 119         struct completion worker_done;
 120
 121         struct hlist_node cpuhp_node;
 122
 123         pid_t task_pid;
 124 };
 125
 126 static enum cpuhp_state io_wq_online;
 127
 128 struct io_cb_cancel_data {
 129         work_cancel_fn *fn;
 130         void *data;
 131         int nr_running;
 132         int nr_pending;
 133         bool cancel_all;
 134 };
 135
 136 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
 137                                        struct io_cb_cancel_data *match);
 138
 139 static bool io_worker_get(struct io_worker *worker)
 140 {
 141         return refcount_inc_not_zero(&worker->ref);
 142 }
 143
 144 static void io_worker_release(struct io_worker *worker)
 145 {
 146         if (refcount_dec_and_test(&worker->ref))
 147                 complete(&worker->ref_done);
 148 }
 149
 150 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 151                                                    struct io_wq_work *work)
 152 {
 153         if (work->flags & IO_WQ_WORK_UNBOUND)
 154                 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 155
 156         return &wqe->acct[IO_WQ_ACCT_BOUND];
 157 }
 158
 159 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_worker *worker)
 160 {
 161         struct io_wqe *wqe = worker->wqe;
 162
 163         if (worker->flags & IO_WORKER_F_BOUND)
 164                 return &wqe->acct[IO_WQ_ACCT_BOUND];
 165
 166         return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 167 }
 168
 169 static void io_worker_exit(struct io_worker *worker)
 170 {
 171         struct io_wqe *wqe = worker->wqe;
 172         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 173         unsigned flags;
 174
 175         if (refcount_dec_and_test(&worker->ref))
 176                 complete(&worker->ref_done);
 177         wait_for_completion(&worker->ref_done);
 178
 179         preempt_disable();
 180         current->flags &= ~PF_IO_WORKER;
 181         flags = worker->flags;
 182         worker->flags = 0;
 183         if (flags & IO_WORKER_F_RUNNING)
 184                 atomic_dec(&acct->nr_running);
 185         worker->flags = 0;
 186         preempt_enable();
 187
 188         raw_spin_lock_irq(&wqe->lock);
 189         if (flags & IO_WORKER_F_FREE)
 190                 hlist_nulls_del_rcu(&worker->nulls_node);
 191         list_del_rcu(&worker->all_list);
 192         acct->nr_workers--;
 193         raw_spin_unlock_irq(&wqe->lock);
 194
 195         kfree_rcu(worker, rcu);
 196         if (atomic_dec_and_test(&wqe->wq->worker_refs))
 197                 complete(&wqe->wq->worker_done);
 198         do_exit(0);
 199 }
 200
 201 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
 202         __must_hold(wqe->lock)
 203 {
 204         if (!wq_list_empty(&wqe->work_list) &&
 205             !(wqe->flags & IO_WQE_FLAG_STALLED))
 206                 return true;
 207         return false;
 208 }
 209
 210 /*
 211  * Check head of free list for an available worker. If one isn't available,
 212  * caller must wake up the wq manager to create one.
 213  */
 214 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
 215         __must_hold(RCU)
 216 {
 217         struct hlist_nulls_node *n;
 218         struct io_worker *worker;
 219
 220         n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
 221         if (is_a_nulls(n))
 222                 return false;
 223
 224         worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
 225         if (io_worker_get(worker)) {
 226                 wake_up_process(worker->task);
 227                 io_worker_release(worker);
 228                 return true;
 229         }
 230
 231         return false;
 232 }
 233
 234 /*
 235  * We need a worker. If we find a free one, we're good. If not, and we're
 236  * below the max number of workers, wake up the manager to create one.
 237  */
 238 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 239 {
 240         bool ret;
 241
 242         /*
 243          * Most likely an attempt to queue unbounded work on an io_wq that
 244          * wasn't setup with any unbounded workers.
 245          */
 246         WARN_ON_ONCE(!acct->max_workers);
 247
 248         rcu_read_lock();
 249         ret = io_wqe_activate_free_worker(wqe);
 250         rcu_read_unlock();
 251
 252         if (!ret && acct->nr_workers < acct->max_workers)
 253                 wake_up_process(wqe->wq->manager);
 254 }
 255
 256 static void io_wqe_inc_running(struct io_worker *worker)
 257 {
 258         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 259
 260         atomic_inc(&acct->nr_running);
 261 }
 262
 263 static void io_wqe_dec_running(struct io_worker *worker)
 264         __must_hold(wqe->lock)
 265 {
 266         struct io_wqe_acct *acct = io_wqe_get_acct(worker);
 267         struct io_wqe *wqe = worker->wqe;
 268
 269         if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
 270                 io_wqe_wake_worker(wqe, acct);
 271 }
 272
 273 /*
 274  * Worker will start processing some work. Move it to the busy list, if
 275  * it's currently on the freelist
 276  */
 277 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
 278                              struct io_wq_work *work)
 279         __must_hold(wqe->lock)
 280 {
 281         bool worker_bound, work_bound;
 282
 283         if (worker->flags & IO_WORKER_F_FREE) {
 284                 worker->flags &= ~IO_WORKER_F_FREE;
 285                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 286         }
 287
 288         /*
 289          * If worker is moving from bound to unbound (or vice versa), then
 290          * ensure we update the running accounting.
 291          */
 292         worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
 293         work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
 294         if (worker_bound != work_bound) {
 295                 io_wqe_dec_running(worker);
 296                 if (work_bound) {
 297                         worker->flags |= IO_WORKER_F_BOUND;
 298                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
 299                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
 300                 } else {
 301                         worker->flags &= ~IO_WORKER_F_BOUND;
 302                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
 303                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
 304                 }
 305                 io_wqe_inc_running(worker);
 306          }
 307 }
 308
 309 /*
 310  * No work, worker going to sleep. Move to freelist, and unuse mm if we
 311  * have one attached. Dropping the mm may potentially sleep, so we drop
 312  * the lock in that case and return success. Since the caller has to
 313  * retry the loop in that case (we changed task state), we don't regrab
 314  * the lock if we return success.
 315  */
 316 static void __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
 317         __must_hold(wqe->lock)
 318 {
 319         if (!(worker->flags & IO_WORKER_F_FREE)) {
 320                 worker->flags |= IO_WORKER_F_FREE;
 321                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 322         }
 323 }
 324
 325 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
 326 {
 327         return work->flags >> IO_WQ_HASH_SHIFT;
 328 }
 329
 330 static void io_wait_on_hash(struct io_wqe *wqe, unsigned int hash)
 331 {
 332         struct io_wq *wq = wqe->wq;
 333
 334         spin_lock(&wq->hash->wait.lock);
 335         if (list_empty(&wqe->wait.entry)) {
 336                 __add_wait_queue(&wq->hash->wait, &wqe->wait);
 337                 if (!test_bit(hash, &wq->hash->map)) {
 338                         __set_current_state(TASK_RUNNING);
 339                         list_del_init(&wqe->wait.entry);
 340                 }
 341         }
 342         spin_unlock(&wq->hash->wait.lock);
 343 }
 344
 345 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
 346         __must_hold(wqe->lock)
 347 {
 348         struct io_wq_work_node *node, *prev;
 349         struct io_wq_work *work, *tail;
 350         unsigned int stall_hash = -1U;
 351
 352         wq_list_for_each(node, prev, &wqe->work_list) {
 353                 unsigned int hash;
 354
 355                 work = container_of(node, struct io_wq_work, list);
 356
 357                 /* not hashed, can run anytime */
 358                 if (!io_wq_is_hashed(work)) {
 359                         wq_list_del(&wqe->work_list, node, prev);
 360                         return work;
 361                 }
 362
 363                 hash = io_get_work_hash(work);
 364                 /* all items with this hash lie in [work, tail] */
 365                 tail = wqe->hash_tail[hash];
 366
 367                 /* hashed, can run if not already running */
 368                 if (!test_and_set_bit(hash, &wqe->wq->hash->map)) {
 369                         wqe->hash_tail[hash] = NULL;
 370                         wq_list_cut(&wqe->work_list, &tail->list, prev);
 371                         return work;
 372                 }
 373                 if (stall_hash == -1U)
 374                         stall_hash = hash;
 375                 /* fast forward to a next hash, for-each will fix up @prev */
 376                 node = &tail->list;
 377         }
 378
 379         if (stall_hash != -1U) {
 380                 raw_spin_unlock(&wqe->lock);
 381                 io_wait_on_hash(wqe, stall_hash);
 382                 raw_spin_lock(&wqe->lock);
 383         }
 384
 385         return NULL;
 386 }
 387
 388 static bool io_flush_signals(void)
 389 {
 390         if (unlikely(test_thread_flag(TIF_NOTIFY_SIGNAL))) {
 391                 __set_current_state(TASK_RUNNING);
 392                 tracehook_notify_signal();
 393                 return true;
 394         }
 395         return false;
 396 }
 397
 398 static void io_assign_current_work(struct io_worker *worker,
 399                                    struct io_wq_work *work)
 400 {
 401         if (work) {
 402                 io_flush_signals();
 403                 cond_resched();
 404         }
 405
 406         spin_lock_irq(&worker->lock);
 407         worker->cur_work = work;
 408         spin_unlock_irq(&worker->lock);
 409 }
 410
 411 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
 412
 413 static void io_worker_handle_work(struct io_worker *worker)
 414         __releases(wqe->lock)
 415 {
 416         struct io_wqe *wqe = worker->wqe;
 417         struct io_wq *wq = wqe->wq;
 418         bool do_kill = test_bit(IO_WQ_BIT_EXIT, &wq->state);
 419
 420         do {
 421                 struct io_wq_work *work;
 422 get_next:
 423                 /*
 424                  * If we got some work, mark us as busy. If we didn't, but
 425                  * the list isn't empty, it means we stalled on hashed work.
 426                  * Mark us stalled so we don't keep looking for work when we
 427                  * can't make progress, any work completion or insertion will
 428                  * clear the stalled flag.
 429                  */
 430                 work = io_get_next_work(wqe);
 431                 if (work)
 432                         __io_worker_busy(wqe, worker, work);
 433                 else if (!wq_list_empty(&wqe->work_list))
 434                         wqe->flags |= IO_WQE_FLAG_STALLED;
 435
 436                 raw_spin_unlock_irq(&wqe->lock);
 437                 if (!work)
 438                         break;
 439                 io_assign_current_work(worker, work);
 440                 __set_current_state(TASK_RUNNING);
 441
 442                 /* handle a whole dependent link */
 443                 do {
 444                         struct io_wq_work *next_hashed, *linked;
 445                         unsigned int hash = io_get_work_hash(work);
 446
 447                         next_hashed = wq_next_work(work);
 448
 449                         if (unlikely(do_kill) && (work->flags & IO_WQ_WORK_UNBOUND))
 450                                 work->flags |= IO_WQ_WORK_CANCEL;
 451                         wq->do_work(work);
 452                         io_assign_current_work(worker, NULL);
 453
 454                         linked = wq->free_work(work);
 455                         work = next_hashed;
 456                         if (!work && linked && !io_wq_is_hashed(linked)) {
 457                                 work = linked;
 458                                 linked = NULL;
 459                         }
 460                         io_assign_current_work(worker, work);
 461                         if (linked)
 462                                 io_wqe_enqueue(wqe, linked);
 463
 464                         if (hash != -1U && !next_hashed) {
 465                                 clear_bit(hash, &wq->hash->map);
 466                                 if (wq_has_sleeper(&wq->hash->wait))
 467                                         wake_up(&wq->hash->wait);
 468                                 raw_spin_lock_irq(&wqe->lock);
 469                                 wqe->flags &= ~IO_WQE_FLAG_STALLED;
 470                                 /* skip unnecessary unlock-lock wqe->lock */
 471                                 if (!work)
 472                                         goto get_next;
 473                                 raw_spin_unlock_irq(&wqe->lock);
 474                         }
 475                 } while (work);
 476
 477                 raw_spin_lock_irq(&wqe->lock);
 478         } while (1);
 479 }
 480
 481 static int io_wqe_worker(void *data)
 482 {
 483         struct io_worker *worker = data;
 484         struct io_wqe *wqe = worker->wqe;
 485         struct io_wq *wq = wqe->wq;
 486         char buf[TASK_COMM_LEN];
 487
 488         worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 489         io_wqe_inc_running(worker);
 490
 491         snprintf(buf, sizeof(buf), "iou-wrk-%d", wq->task_pid);
 492         set_task_comm(current, buf);
 493
 494         while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 495                 long ret;
 496
 497                 set_current_state(TASK_INTERRUPTIBLE);
 498 loop:
 499                 raw_spin_lock_irq(&wqe->lock);
 500                 if (io_wqe_run_queue(wqe)) {
 501                         io_worker_handle_work(worker);
 502                         goto loop;
 503                 }
 504                 __io_worker_idle(wqe, worker);
 505                 raw_spin_unlock_irq(&wqe->lock);
 506                 if (io_flush_signals())
 507                         continue;
 508                 ret = schedule_timeout(WORKER_IDLE_TIMEOUT);
 509                 if (signal_pending(current)) {
 510                         struct ksignal ksig;
 511
 512                         if (!get_signal(&ksig))
 513                                 continue;
 514                         break;
 515                 }
 516                 if (ret)
 517                         continue;
 518                 /* timed out, exit unless we're the fixed worker */
 519                 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 520                     !(worker->flags & IO_WORKER_F_FIXED))
 521                         break;
 522         }
 523
 524         if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 525                 raw_spin_lock_irq(&wqe->lock);
 526                 if (!wq_list_empty(&wqe->work_list))
 527                         io_worker_handle_work(worker);
 528                 else
 529                         raw_spin_unlock_irq(&wqe->lock);
 530         }
 531
 532         io_worker_exit(worker);
 533         return 0;
 534 }
 535
 536 /*
 537  * Called when a worker is scheduled in. Mark us as currently running.
 538  */
 539 void io_wq_worker_running(struct task_struct *tsk)
 540 {
 541         struct io_worker *worker = tsk->pf_io_worker;
 542
 543         if (!worker)
 544                 return;
 545         if (!(worker->flags & IO_WORKER_F_UP))
 546                 return;
 547         if (worker->flags & IO_WORKER_F_RUNNING)
 548                 return;
 549         worker->flags |= IO_WORKER_F_RUNNING;
 550         io_wqe_inc_running(worker);
 551 }
 552
 553 /*
 554  * Called when worker is going to sleep. If there are no workers currently
 555  * running and we have work pending, wake up a free one or have the manager
 556  * set one up.
 557  */
 558 void io_wq_worker_sleeping(struct task_struct *tsk)
 559 {
 560         struct io_worker *worker = tsk->pf_io_worker;
 561
 562         if (!worker)
 563                 return;
 564         if (!(worker->flags & IO_WORKER_F_UP))
 565                 return;
 566         if (!(worker->flags & IO_WORKER_F_RUNNING))
 567                 return;
 568
 569         worker->flags &= ~IO_WORKER_F_RUNNING;
 570
 571         raw_spin_lock_irq(&worker->wqe->lock);
 572         io_wqe_dec_running(worker);
 573         raw_spin_unlock_irq(&worker->wqe->lock);
 574 }
 575
 576 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
 577 {
 578         struct io_wqe_acct *acct = &wqe->acct[index];
 579         struct io_worker *worker;
 580         struct task_struct *tsk;
 581
 582         __set_current_state(TASK_RUNNING);
 583
 584         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
 585         if (!worker)
 586                 return false;
 587
 588         refcount_set(&worker->ref, 1);
 589         worker->nulls_node.pprev = NULL;
 590         worker->wqe = wqe;
 591         spin_lock_init(&worker->lock);
 592         init_completion(&worker->ref_done);
 593
 594         atomic_inc(&wq->worker_refs);
 595
 596         tsk = create_io_thread(io_wqe_worker, worker, wqe->node);
 597         if (IS_ERR(tsk)) {
 598                 if (atomic_dec_and_test(&wq->worker_refs))
 599                         complete(&wq->worker_done);
 600                 kfree(worker);
 601                 return false;
 602         }
 603
 604         tsk->pf_io_worker = worker;
 605         worker->task = tsk;
 606         set_cpus_allowed_ptr(tsk, cpumask_of_node(wqe->node));
 607         tsk->flags |= PF_NO_SETAFFINITY;
 608
 609         raw_spin_lock_irq(&wqe->lock);
 610         hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 611         list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 612         worker->flags |= IO_WORKER_F_FREE;
 613         if (index == IO_WQ_ACCT_BOUND)
 614                 worker->flags |= IO_WORKER_F_BOUND;
 615         if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
 616                 worker->flags |= IO_WORKER_F_FIXED;
 617         acct->nr_workers++;
 618         raw_spin_unlock_irq(&wqe->lock);
 619         wake_up_new_task(tsk);
 620         return true;
 621 }
 622
 623 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
 624         __must_hold(wqe->lock)
 625 {
 626         struct io_wqe_acct *acct = &wqe->acct[index];
 627
 628         if (acct->nr_workers && test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state))
 629                 return false;
 630         /* if we have available workers or no work, no need */
 631         if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
 632                 return false;
 633         return acct->nr_workers < acct->max_workers;
 634 }
 635
 636 /*
 637  * Iterate the passed in list and call the specific function for each
 638  * worker that isn't exiting
 639  */
 640 static bool io_wq_for_each_worker(struct io_wqe *wqe,
 641                                   bool (*func)(struct io_worker *, void *),
 642                                   void *data)
 643 {
 644         struct io_worker *worker;
 645         bool ret = false;
 646
 647         list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
 648                 if (io_worker_get(worker)) {
 649                         /* no task if node is/was offline */
 650                         if (worker->task)
 651                                 ret = func(worker, data);
 652                         io_worker_release(worker);
 653                         if (ret)
 654                                 break;
 655                 }
 656         }
 657
 658         return ret;
 659 }
 660
 661 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
 662 {
 663         set_notify_signal(worker->task);
 664         wake_up_process(worker->task);
 665         return false;
 666 }
 667
 668 static void io_wq_check_workers(struct io_wq *wq)
 669 {
 670         int node;
 671
 672         for_each_node(node) {
 673                 struct io_wqe *wqe = wq->wqes[node];
 674                 bool fork_worker[2] = { false, false };
 675
 676                 if (!node_online(node))
 677                         continue;
 678
 679                 raw_spin_lock_irq(&wqe->lock);
 680                 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
 681                         fork_worker[IO_WQ_ACCT_BOUND] = true;
 682                 if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
 683                         fork_worker[IO_WQ_ACCT_UNBOUND] = true;
 684                 raw_spin_unlock_irq(&wqe->lock);
 685                 if (fork_worker[IO_WQ_ACCT_BOUND])
 686                         create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
 687                 if (fork_worker[IO_WQ_ACCT_UNBOUND])
 688                         create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
 689         }
 690 }
 691
 692 static bool io_wq_work_match_all(struct io_wq_work *work, void *data)
 693 {
 694         return true;
 695 }
 696
 697 static void io_wq_cancel_pending(struct io_wq *wq)
 698 {
 699         struct io_cb_cancel_data match = {
 700                 .fn             = io_wq_work_match_all,
 701                 .cancel_all     = true,
 702         };
 703         int node;
 704
 705         for_each_node(node)
 706                 io_wqe_cancel_pending_work(wq->wqes[node], &match);
 707 }
 708
 709 /*
 710  * Manager thread. Tasked with creating new workers, if we need them.
 711  */
 712 static int io_wq_manager(void *data)
 713 {
 714         struct io_wq *wq = data;
 715         char buf[TASK_COMM_LEN];
 716         int node;
 717
 718         snprintf(buf, sizeof(buf), "iou-mgr-%d", wq->task_pid);
 719         set_task_comm(current, buf);
 720
 721         do {
 722                 set_current_state(TASK_INTERRUPTIBLE);
 723                 io_wq_check_workers(wq);
 724                 schedule_timeout(HZ);
 725                 if (signal_pending(current)) {
 726                         struct ksignal ksig;
 727
 728                         if (!get_signal(&ksig))
 729                                 continue;
 730                         set_bit(IO_WQ_BIT_EXIT, &wq->state);
 731                 }
 732         } while (!test_bit(IO_WQ_BIT_EXIT, &wq->state));
 733
 734         io_wq_check_workers(wq);
 735
 736         rcu_read_lock();
 737         for_each_node(node)
 738                 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
 739         rcu_read_unlock();
 740
 741         if (atomic_dec_and_test(&wq->worker_refs))
 742                 complete(&wq->worker_done);
 743         wait_for_completion(&wq->worker_done);
 744
 745         spin_lock_irq(&wq->hash->wait.lock);
 746         for_each_node(node)
 747                 list_del_init(&wq->wqes[node]->wait.entry);
 748         spin_unlock_irq(&wq->hash->wait.lock);
 749
 750         io_wq_cancel_pending(wq);
 751         complete(&wq->exited);
 752         do_exit(0);
 753 }
 754
 755 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
 756 {
 757         struct io_wq *wq = wqe->wq;
 758
 759         do {
 760                 work->flags |= IO_WQ_WORK_CANCEL;
 761                 wq->do_work(work);
 762                 work = wq->free_work(work);
 763         } while (work);
 764 }
 765
 766 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
 767 {
 768         unsigned int hash;
 769         struct io_wq_work *tail;
 770
 771         if (!io_wq_is_hashed(work)) {
 772 append:
 773                 wq_list_add_tail(&work->list, &wqe->work_list);
 774                 return;
 775         }
 776
 777         hash = io_get_work_hash(work);
 778         tail = wqe->hash_tail[hash];
 779         wqe->hash_tail[hash] = work;
 780         if (!tail)
 781                 goto append;
 782
 783         wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
 784 }
 785
 786 static int io_wq_fork_manager(struct io_wq *wq)
 787 {
 788         struct task_struct *tsk;
 789
 790         if (wq->manager)
 791                 return 0;
 792
 793         WARN_ON_ONCE(test_bit(IO_WQ_BIT_EXIT, &wq->state));
 794
 795         init_completion(&wq->worker_done);
 796         atomic_set(&wq->worker_refs, 1);
 797         tsk = create_io_thread(io_wq_manager, wq, NUMA_NO_NODE);
 798         if (!IS_ERR(tsk)) {
 799                 wq->manager = get_task_struct(tsk);
 800                 wake_up_new_task(tsk);
 801                 return 0;
 802         }
 803
 804         if (atomic_dec_and_test(&wq->worker_refs))
 805                 complete(&wq->worker_done);
 806
 807         return PTR_ERR(tsk);
 808 }
 809
 810 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 811 {
 812         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 813         int work_flags;
 814         unsigned long flags;
 815
 816         /* Can only happen if manager creation fails after exec */
 817         if (io_wq_fork_manager(wqe->wq) ||
 818             test_bit(IO_WQ_BIT_EXIT, &wqe->wq->state)) {
 819                 io_run_cancel(work, wqe);
 820                 return;
 821         }
 822
 823         work_flags = work->flags;
 824         raw_spin_lock_irqsave(&wqe->lock, flags);
 825         io_wqe_insert_work(wqe, work);
 826         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 827         raw_spin_unlock_irqrestore(&wqe->lock, flags);
 828
 829         if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 830             !atomic_read(&acct->nr_running))
 831                 io_wqe_wake_worker(wqe, acct);
 832 }
 833
 834 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 835 {
 836         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 837
 838         io_wqe_enqueue(wqe, work);
 839 }
 840
 841 /*
 842  * Work items that hash to the same value will not be done in parallel.
 843  * Used to limit concurrent writes, generally hashed by inode.
 844  */
 845 void io_wq_hash_work(struct io_wq_work *work, void *val)
 846 {
 847         unsigned int bit;
 848
 849         bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 850         work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 851 }
 852
 853 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 854 {
 855         struct io_cb_cancel_data *match = data;
 856         unsigned long flags;
 857
 858         /*
 859          * Hold the lock to avoid ->cur_work going out of scope, caller
 860          * may dereference the passed in work.
 861          */
 862         spin_lock_irqsave(&worker->lock, flags);
 863         if (worker->cur_work &&
 864             match->fn(worker->cur_work, match->data)) {
 865                 set_notify_signal(worker->task);
 866                 match->nr_running++;
 867         }
 868         spin_unlock_irqrestore(&worker->lock, flags);
 869
 870         return match->nr_running && !match->cancel_all;
 871 }
 872
 873 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
 874                                          struct io_wq_work *work,
 875                                          struct io_wq_work_node *prev)
 876 {
 877         unsigned int hash = io_get_work_hash(work);
 878         struct io_wq_work *prev_work = NULL;
 879
 880         if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
 881                 if (prev)
 882                         prev_work = container_of(prev, struct io_wq_work, list);
 883                 if (prev_work && io_get_work_hash(prev_work) == hash)
 884                         wqe->hash_tail[hash] = prev_work;
 885                 else
 886                         wqe->hash_tail[hash] = NULL;
 887         }
 888         wq_list_del(&wqe->work_list, &work->list, prev);
 889 }
 890
 891 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
 892                                        struct io_cb_cancel_data *match)
 893 {
 894         struct io_wq_work_node *node, *prev;
 895         struct io_wq_work *work;
 896         unsigned long flags;
 897
 898 retry:
 899         raw_spin_lock_irqsave(&wqe->lock, flags);
 900         wq_list_for_each(node, prev, &wqe->work_list) {
 901                 work = container_of(node, struct io_wq_work, list);
 902                 if (!match->fn(work, match->data))
 903                         continue;
 904                 io_wqe_remove_pending(wqe, work, prev);
 905                 raw_spin_unlock_irqrestore(&wqe->lock, flags);
 906                 io_run_cancel(work, wqe);
 907                 match->nr_pending++;
 908                 if (!match->cancel_all)
 909                         return;
 910
 911                 /* not safe to continue after unlock */
 912                 goto retry;
 913         }
 914         raw_spin_unlock_irqrestore(&wqe->lock, flags);
 915 }
 916
 917 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
 918                                        struct io_cb_cancel_data *match)
 919 {
 920         rcu_read_lock();
 921         io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
 922         rcu_read_unlock();
 923 }
 924
 925 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
 926                                   void *data, bool cancel_all)
 927 {
 928         struct io_cb_cancel_data match = {
 929                 .fn             = cancel,
 930                 .data           = data,
 931                 .cancel_all     = cancel_all,
 932         };
 933         int node;
 934
 935         /*
 936          * First check pending list, if we're lucky we can just remove it
 937          * from there. CANCEL_OK means that the work is returned as-new,
 938          * no completion will be posted for it.
 939          */
 940         for_each_node(node) {
 941                 struct io_wqe *wqe = wq->wqes[node];
 942
 943                 io_wqe_cancel_pending_work(wqe, &match);
 944                 if (match.nr_pending && !match.cancel_all)
 945                         return IO_WQ_CANCEL_OK;
 946         }
 947
 948         /*
 949          * Now check if a free (going busy) or busy worker has the work
 950          * currently running. If we find it there, we'll return CANCEL_RUNNING
 951          * as an indication that we attempt to signal cancellation. The
 952          * completion will run normally in this case.
 953          */
 954         for_each_node(node) {
 955                 struct io_wqe *wqe = wq->wqes[node];
 956
 957                 io_wqe_cancel_running_work(wqe, &match);
 958                 if (match.nr_running && !match.cancel_all)
 959                         return IO_WQ_CANCEL_RUNNING;
 960         }
 961
 962         if (match.nr_running)
 963                 return IO_WQ_CANCEL_RUNNING;
 964         if (match.nr_pending)
 965                 return IO_WQ_CANCEL_OK;
 966         return IO_WQ_CANCEL_NOTFOUND;
 967 }
 968
 969 static int io_wqe_hash_wake(struct wait_queue_entry *wait, unsigned mode,
 970                             int sync, void *key)
 971 {
 972         struct io_wqe *wqe = container_of(wait, struct io_wqe, wait);
 973         int ret;
 974
 975         list_del_init(&wait->entry);
 976
 977         rcu_read_lock();
 978         ret = io_wqe_activate_free_worker(wqe);
 979         rcu_read_unlock();
 980
 981         if (!ret)
 982                 wake_up_process(wqe->wq->manager);
 983
 984         return 1;
 985 }
 986
 987 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
 988 {
 989         int ret = -ENOMEM, node;
 990         struct io_wq *wq;
 991
 992         if (WARN_ON_ONCE(!data->free_work || !data->do_work))
 993                 return ERR_PTR(-EINVAL);
 994
 995         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
 996         if (!wq)
 997                 return ERR_PTR(-ENOMEM);
 998
 999         wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1000         if (!wq->wqes)
1001                 goto err_wq;
1002
1003         ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1004         if (ret)
1005                 goto err_wqes;
1006
1007         refcount_inc(&data->hash->refs);
1008         wq->hash = data->hash;
1009         wq->free_work = data->free_work;
1010         wq->do_work = data->do_work;
1011
1012         ret = -ENOMEM;
1013         for_each_node(node) {
1014                 struct io_wqe *wqe;
1015                 int alloc_node = node;
1016
1017                 if (!node_online(alloc_node))
1018                         alloc_node = NUMA_NO_NODE;
1019                 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1020                 if (!wqe)
1021                         goto err;
1022                 wq->wqes[node] = wqe;
1023                 wqe->node = alloc_node;
1024                 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1025                 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1026                 wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1027                                         task_rlimit(current, RLIMIT_NPROC);
1028                 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1029                 wqe->wait.func = io_wqe_hash_wake;
1030                 INIT_LIST_HEAD(&wqe->wait.entry);
1031                 wqe->wq = wq;
1032                 raw_spin_lock_init(&wqe->lock);
1033                 INIT_WQ_LIST(&wqe->work_list);
1034                 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1035                 INIT_LIST_HEAD(&wqe->all_list);
1036         }
1037
1038         wq->task_pid = current->pid;
1039         init_completion(&wq->exited);
1040         refcount_set(&wq->refs, 1);
1041
1042         ret = io_wq_fork_manager(wq);
1043         if (!ret)
1044                 return wq;
1045 err:
1046         io_wq_put_hash(data->hash);
1047         cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1048         for_each_node(node)
1049                 kfree(wq->wqes[node]);
1050 err_wqes:
1051         kfree(wq->wqes);
1052 err_wq:
1053         kfree(wq);
1054         return ERR_PTR(ret);
1055 }
1056
1057 static void io_wq_destroy_manager(struct io_wq *wq)
1058 {
1059         if (wq->manager) {
1060                 wake_up_process(wq->manager);
1061                 wait_for_completion(&wq->exited);
1062                 put_task_struct(wq->manager);
1063                 wq->manager = NULL;
1064         }
1065 }
1066
1067 static void io_wq_destroy(struct io_wq *wq)
1068 {
1069         int node;
1070
1071         cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1072
1073         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1074         io_wq_destroy_manager(wq);
1075
1076         for_each_node(node) {
1077                 struct io_wqe *wqe = wq->wqes[node];
1078                 struct io_cb_cancel_data match = {
1079                         .fn             = io_wq_work_match_all,
1080                         .cancel_all     = true,
1081                 };
1082                 io_wqe_cancel_pending_work(wqe, &match);
1083                 kfree(wqe);
1084         }
1085         io_wq_put_hash(wq->hash);
1086         kfree(wq->wqes);
1087         kfree(wq);
1088 }
1089
1090 void io_wq_put(struct io_wq *wq)
1091 {
1092         if (refcount_dec_and_test(&wq->refs))
1093                 io_wq_destroy(wq);
1094 }
1095
1096 void io_wq_put_and_exit(struct io_wq *wq)
1097 {
1098         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1099         io_wq_destroy_manager(wq);
1100         io_wq_put(wq);
1101 }
1102
1103 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1104 {
1105         struct task_struct *task = worker->task;
1106         struct rq_flags rf;
1107         struct rq *rq;
1108
1109         rq = task_rq_lock(task, &rf);
1110         do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
1111         task->flags |= PF_NO_SETAFFINITY;
1112         task_rq_unlock(rq, task, &rf);
1113         return false;
1114 }
1115
1116 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1117 {
1118         struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1119         int i;
1120
1121         rcu_read_lock();
1122         for_each_node(i)
1123                 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
1124         rcu_read_unlock();
1125         return 0;
1126 }
1127
1128 static __init int io_wq_init(void)
1129 {
1130         int ret;
1131
1132         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1133                                         io_wq_cpu_online, NULL);
1134         if (ret < 0)
1135                 return ret;
1136         io_wq_online = ret;
1137         return 0;
1138 }
1139 subsys_initcall(io_wq_init);