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