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/kthread.h>
  17 #include <linux/rculist_nulls.h>
  18 #include <linux/fs_struct.h>
  19 #include <linux/task_work.h>
  20 #include <linux/blk-cgroup.h>
  21 #include <linux/audit.h>
  22 #include <linux/cpu.h>
  23
  24 #include "../kernel/sched/sched.h"
  25 #include "io-wq.h"
  26
  27 #define WORKER_IDLE_TIMEOUT     (5 * HZ)
  28
  29 enum {
  30         IO_WORKER_F_UP          = 1,    /* up and active */
  31         IO_WORKER_F_RUNNING     = 2,    /* account as running */
  32         IO_WORKER_F_FREE        = 4,    /* worker on free list */
  33         IO_WORKER_F_FIXED       = 8,    /* static idle worker */
  34         IO_WORKER_F_BOUND       = 16,   /* is doing bounded work */
  35 };
  36
  37 enum {
  38         IO_WQ_BIT_EXIT          = 0,    /* wq exiting */
  39         IO_WQ_BIT_ERROR         = 1,    /* error on setup */
  40 };
  41
  42 enum {
  43         IO_WQE_FLAG_STALLED     = 1,    /* stalled on hash */
  44 };
  45
  46 /*
  47  * One for each thread in a wqe pool
  48  */
  49 struct io_worker {
  50         refcount_t ref;
  51         unsigned flags;
  52         struct hlist_nulls_node nulls_node;
  53         struct list_head all_list;
  54         struct task_struct *task;
  55         struct io_wqe *wqe;
  56
  57         struct io_wq_work *cur_work;
  58         spinlock_t lock;
  59
  60         struct rcu_head rcu;
  61         struct mm_struct *mm;
  62 #ifdef CONFIG_BLK_CGROUP
  63         struct cgroup_subsys_state *blkcg_css;
  64 #endif
  65         const struct cred *cur_creds;
  66         const struct cred *saved_creds;
  67         struct files_struct *restore_files;
  68         struct nsproxy *restore_nsproxy;
  69         struct fs_struct *restore_fs;
  70 };
  71
  72 #if BITS_PER_LONG == 64
  73 #define IO_WQ_HASH_ORDER        6
  74 #else
  75 #define IO_WQ_HASH_ORDER        5
  76 #endif
  77
  78 #define IO_WQ_NR_HASH_BUCKETS   (1u << IO_WQ_HASH_ORDER)
  79
  80 struct io_wqe_acct {
  81         unsigned nr_workers;
  82         unsigned max_workers;
  83         atomic_t nr_running;
  84 };
  85
  86 enum {
  87         IO_WQ_ACCT_BOUND,
  88         IO_WQ_ACCT_UNBOUND,
  89 };
  90
  91 /*
  92  * Per-node worker thread pool
  93  */
  94 struct io_wqe {
  95         struct {
  96                 raw_spinlock_t lock;
  97                 struct io_wq_work_list work_list;
  98                 unsigned long hash_map;
  99                 unsigned flags;
 100         } ____cacheline_aligned_in_smp;
 101
 102         int node;
 103         struct io_wqe_acct acct[2];
 104
 105         struct hlist_nulls_head free_list;
 106         struct list_head all_list;
 107
 108         struct io_wq *wq;
 109         struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
 110 };
 111
 112 /*
 113  * Per io_wq state
 114   */
 115 struct io_wq {
 116         struct io_wqe **wqes;
 117         unsigned long state;
 118
 119         free_work_fn *free_work;
 120         io_wq_work_fn *do_work;
 121
 122         struct task_struct *manager;
 123         struct user_struct *user;
 124         refcount_t refs;
 125         struct completion done;
 126
 127         struct hlist_node cpuhp_node;
 128
 129         refcount_t use_refs;
 130 };
 131
 132 static enum cpuhp_state io_wq_online;
 133
 134 static bool io_worker_get(struct io_worker *worker)
 135 {
 136         return refcount_inc_not_zero(&worker->ref);
 137 }
 138
 139 static void io_worker_release(struct io_worker *worker)
 140 {
 141         if (refcount_dec_and_test(&worker->ref))
 142                 wake_up_process(worker->task);
 143 }
 144
 145 /*
 146  * Note: drops the wqe->lock if returning true! The caller must re-acquire
 147  * the lock in that case. Some callers need to restart handling if this
 148  * happens, so we can't just re-acquire the lock on behalf of the caller.
 149  */
 150 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
 151 {
 152         bool dropped_lock = false;
 153
 154         if (worker->saved_creds) {
 155                 revert_creds(worker->saved_creds);
 156                 worker->cur_creds = worker->saved_creds = NULL;
 157         }
 158
 159         if (current->files != worker->restore_files) {
 160                 __acquire(&wqe->lock);
 161                 raw_spin_unlock_irq(&wqe->lock);
 162                 dropped_lock = true;
 163
 164                 task_lock(current);
 165                 current->files = worker->restore_files;
 166                 current->nsproxy = worker->restore_nsproxy;
 167                 task_unlock(current);
 168         }
 169
 170         if (current->fs != worker->restore_fs)
 171                 current->fs = worker->restore_fs;
 172
 173         /*
 174          * If we have an active mm, we need to drop the wq lock before unusing
 175          * it. If we do, return true and let the caller retry the idle loop.
 176          */
 177         if (worker->mm) {
 178                 if (!dropped_lock) {
 179                         __acquire(&wqe->lock);
 180                         raw_spin_unlock_irq(&wqe->lock);
 181                         dropped_lock = true;
 182                 }
 183                 __set_current_state(TASK_RUNNING);
 184                 kthread_unuse_mm(worker->mm);
 185                 mmput(worker->mm);
 186                 worker->mm = NULL;
 187         }
 188
 189 #ifdef CONFIG_BLK_CGROUP
 190         if (worker->blkcg_css) {
 191                 kthread_associate_blkcg(NULL);
 192                 worker->blkcg_css = NULL;
 193         }
 194 #endif
 195         if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
 196                 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
 197         return dropped_lock;
 198 }
 199
 200 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 201                                                    struct io_wq_work *work)
 202 {
 203         if (work->flags & IO_WQ_WORK_UNBOUND)
 204                 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 205
 206         return &wqe->acct[IO_WQ_ACCT_BOUND];
 207 }
 208
 209 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
 210                                                   struct io_worker *worker)
 211 {
 212         if (worker->flags & IO_WORKER_F_BOUND)
 213                 return &wqe->acct[IO_WQ_ACCT_BOUND];
 214
 215         return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 216 }
 217
 218 static void io_worker_exit(struct io_worker *worker)
 219 {
 220         struct io_wqe *wqe = worker->wqe;
 221         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 222
 223         /*
 224          * If we're not at zero, someone else is holding a brief reference
 225          * to the worker. Wait for that to go away.
 226          */
 227         set_current_state(TASK_INTERRUPTIBLE);
 228         if (!refcount_dec_and_test(&worker->ref))
 229                 schedule();
 230         __set_current_state(TASK_RUNNING);
 231
 232         preempt_disable();
 233         current->flags &= ~PF_IO_WORKER;
 234         if (worker->flags & IO_WORKER_F_RUNNING)
 235                 atomic_dec(&acct->nr_running);
 236         if (!(worker->flags & IO_WORKER_F_BOUND))
 237                 atomic_dec(&wqe->wq->user->processes);
 238         worker->flags = 0;
 239         preempt_enable();
 240
 241         raw_spin_lock_irq(&wqe->lock);
 242         hlist_nulls_del_rcu(&worker->nulls_node);
 243         list_del_rcu(&worker->all_list);
 244         if (__io_worker_unuse(wqe, worker)) {
 245                 __release(&wqe->lock);
 246                 raw_spin_lock_irq(&wqe->lock);
 247         }
 248         acct->nr_workers--;
 249         raw_spin_unlock_irq(&wqe->lock);
 250
 251         kfree_rcu(worker, rcu);
 252         if (refcount_dec_and_test(&wqe->wq->refs))
 253                 complete(&wqe->wq->done);
 254 }
 255
 256 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
 257         __must_hold(wqe->lock)
 258 {
 259         if (!wq_list_empty(&wqe->work_list) &&
 260             !(wqe->flags & IO_WQE_FLAG_STALLED))
 261                 return true;
 262         return false;
 263 }
 264
 265 /*
 266  * Check head of free list for an available worker. If one isn't available,
 267  * caller must wake up the wq manager to create one.
 268  */
 269 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
 270         __must_hold(RCU)
 271 {
 272         struct hlist_nulls_node *n;
 273         struct io_worker *worker;
 274
 275         n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
 276         if (is_a_nulls(n))
 277                 return false;
 278
 279         worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
 280         if (io_worker_get(worker)) {
 281                 wake_up_process(worker->task);
 282                 io_worker_release(worker);
 283                 return true;
 284         }
 285
 286         return false;
 287 }
 288
 289 /*
 290  * We need a worker. If we find a free one, we're good. If not, and we're
 291  * below the max number of workers, wake up the manager to create one.
 292  */
 293 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 294 {
 295         bool ret;
 296
 297         /*
 298          * Most likely an attempt to queue unbounded work on an io_wq that
 299          * wasn't setup with any unbounded workers.
 300          */
 301         WARN_ON_ONCE(!acct->max_workers);
 302
 303         rcu_read_lock();
 304         ret = io_wqe_activate_free_worker(wqe);
 305         rcu_read_unlock();
 306
 307         if (!ret && acct->nr_workers < acct->max_workers)
 308                 wake_up_process(wqe->wq->manager);
 309 }
 310
 311 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
 312 {
 313         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 314
 315         atomic_inc(&acct->nr_running);
 316 }
 317
 318 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
 319         __must_hold(wqe->lock)
 320 {
 321         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 322
 323         if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
 324                 io_wqe_wake_worker(wqe, acct);
 325 }
 326
 327 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
 328 {
 329         allow_kernel_signal(SIGINT);
 330
 331         current->flags |= PF_IO_WORKER;
 332
 333         worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 334         worker->restore_files = current->files;
 335         worker->restore_nsproxy = current->nsproxy;
 336         worker->restore_fs = current->fs;
 337         io_wqe_inc_running(wqe, worker);
 338 }
 339
 340 /*
 341  * Worker will start processing some work. Move it to the busy list, if
 342  * it's currently on the freelist
 343  */
 344 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
 345                              struct io_wq_work *work)
 346         __must_hold(wqe->lock)
 347 {
 348         bool worker_bound, work_bound;
 349
 350         if (worker->flags & IO_WORKER_F_FREE) {
 351                 worker->flags &= ~IO_WORKER_F_FREE;
 352                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 353         }
 354
 355         /*
 356          * If worker is moving from bound to unbound (or vice versa), then
 357          * ensure we update the running accounting.
 358          */
 359         worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
 360         work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
 361         if (worker_bound != work_bound) {
 362                 io_wqe_dec_running(wqe, worker);
 363                 if (work_bound) {
 364                         worker->flags |= IO_WORKER_F_BOUND;
 365                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
 366                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
 367                         atomic_dec(&wqe->wq->user->processes);
 368                 } else {
 369                         worker->flags &= ~IO_WORKER_F_BOUND;
 370                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
 371                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
 372                         atomic_inc(&wqe->wq->user->processes);
 373                 }
 374                 io_wqe_inc_running(wqe, worker);
 375          }
 376 }
 377
 378 /*
 379  * No work, worker going to sleep. Move to freelist, and unuse mm if we
 380  * have one attached. Dropping the mm may potentially sleep, so we drop
 381  * the lock in that case and return success. Since the caller has to
 382  * retry the loop in that case (we changed task state), we don't regrab
 383  * the lock if we return success.
 384  */
 385 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
 386         __must_hold(wqe->lock)
 387 {
 388         if (!(worker->flags & IO_WORKER_F_FREE)) {
 389                 worker->flags |= IO_WORKER_F_FREE;
 390                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 391         }
 392
 393         return __io_worker_unuse(wqe, worker);
 394 }
 395
 396 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
 397 {
 398         return work->flags >> IO_WQ_HASH_SHIFT;
 399 }
 400
 401 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
 402         __must_hold(wqe->lock)
 403 {
 404         struct io_wq_work_node *node, *prev;
 405         struct io_wq_work *work, *tail;
 406         unsigned int hash;
 407
 408         wq_list_for_each(node, prev, &wqe->work_list) {
 409                 work = container_of(node, struct io_wq_work, list);
 410
 411                 /* not hashed, can run anytime */
 412                 if (!io_wq_is_hashed(work)) {
 413                         wq_list_del(&wqe->work_list, node, prev);
 414                         return work;
 415                 }
 416
 417                 /* hashed, can run if not already running */
 418                 hash = io_get_work_hash(work);
 419                 if (!(wqe->hash_map & BIT(hash))) {
 420                         wqe->hash_map |= BIT(hash);
 421                         /* all items with this hash lie in [work, tail] */
 422                         tail = wqe->hash_tail[hash];
 423                         wqe->hash_tail[hash] = NULL;
 424                         wq_list_cut(&wqe->work_list, &tail->list, prev);
 425                         return work;
 426                 }
 427         }
 428
 429         return NULL;
 430 }
 431
 432 static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
 433 {
 434         if (worker->mm) {
 435                 kthread_unuse_mm(worker->mm);
 436                 mmput(worker->mm);
 437                 worker->mm = NULL;
 438         }
 439
 440         if (mmget_not_zero(work->identity->mm)) {
 441                 kthread_use_mm(work->identity->mm);
 442                 worker->mm = work->identity->mm;
 443                 return;
 444         }
 445
 446         /* failed grabbing mm, ensure work gets cancelled */
 447         work->flags |= IO_WQ_WORK_CANCEL;
 448 }
 449
 450 static inline void io_wq_switch_blkcg(struct io_worker *worker,
 451                                       struct io_wq_work *work)
 452 {
 453 #ifdef CONFIG_BLK_CGROUP
 454         if (!(work->flags & IO_WQ_WORK_BLKCG))
 455                 return;
 456         if (work->identity->blkcg_css != worker->blkcg_css) {
 457                 kthread_associate_blkcg(work->identity->blkcg_css);
 458                 worker->blkcg_css = work->identity->blkcg_css;
 459         }
 460 #endif
 461 }
 462
 463 static void io_wq_switch_creds(struct io_worker *worker,
 464                                struct io_wq_work *work)
 465 {
 466         const struct cred *old_creds = override_creds(work->identity->creds);
 467
 468         worker->cur_creds = work->identity->creds;
 469         if (worker->saved_creds)
 470                 put_cred(old_creds); /* creds set by previous switch */
 471         else
 472                 worker->saved_creds = old_creds;
 473 }
 474
 475 static void io_impersonate_work(struct io_worker *worker,
 476                                 struct io_wq_work *work)
 477 {
 478         if ((work->flags & IO_WQ_WORK_FILES) &&
 479             current->files != work->identity->files) {
 480                 task_lock(current);
 481                 current->files = work->identity->files;
 482                 current->nsproxy = work->identity->nsproxy;
 483                 task_unlock(current);
 484                 if (!work->identity->files) {
 485                         /* failed grabbing files, ensure work gets cancelled */
 486                         work->flags |= IO_WQ_WORK_CANCEL;
 487                 }
 488         }
 489         if ((work->flags & IO_WQ_WORK_FS) && current->fs != work->identity->fs)
 490                 current->fs = work->identity->fs;
 491         if ((work->flags & IO_WQ_WORK_MM) && work->identity->mm != worker->mm)
 492                 io_wq_switch_mm(worker, work);
 493         if ((work->flags & IO_WQ_WORK_CREDS) &&
 494             worker->cur_creds != work->identity->creds)
 495                 io_wq_switch_creds(worker, work);
 496         if (work->flags & IO_WQ_WORK_FSIZE)
 497                 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->identity->fsize;
 498         else if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
 499                 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
 500         io_wq_switch_blkcg(worker, work);
 501 #ifdef CONFIG_AUDIT
 502         current->loginuid = work->identity->loginuid;
 503         current->sessionid = work->identity->sessionid;
 504 #endif
 505 }
 506
 507 static void io_assign_current_work(struct io_worker *worker,
 508                                    struct io_wq_work *work)
 509 {
 510         if (work) {
 511                 /* flush pending signals before assigning new work */
 512                 if (signal_pending(current))
 513                         flush_signals(current);
 514                 cond_resched();
 515         }
 516
 517 #ifdef CONFIG_AUDIT
 518         current->loginuid = KUIDT_INIT(AUDIT_UID_UNSET);
 519         current->sessionid = AUDIT_SID_UNSET;
 520 #endif
 521
 522         spin_lock_irq(&worker->lock);
 523         worker->cur_work = work;
 524         spin_unlock_irq(&worker->lock);
 525 }
 526
 527 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
 528
 529 static void io_worker_handle_work(struct io_worker *worker)
 530         __releases(wqe->lock)
 531 {
 532         struct io_wqe *wqe = worker->wqe;
 533         struct io_wq *wq = wqe->wq;
 534
 535         do {
 536                 struct io_wq_work *work;
 537 get_next:
 538                 /*
 539                  * If we got some work, mark us as busy. If we didn't, but
 540                  * the list isn't empty, it means we stalled on hashed work.
 541                  * Mark us stalled so we don't keep looking for work when we
 542                  * can't make progress, any work completion or insertion will
 543                  * clear the stalled flag.
 544                  */
 545                 work = io_get_next_work(wqe);
 546                 if (work)
 547                         __io_worker_busy(wqe, worker, work);
 548                 else if (!wq_list_empty(&wqe->work_list))
 549                         wqe->flags |= IO_WQE_FLAG_STALLED;
 550
 551                 raw_spin_unlock_irq(&wqe->lock);
 552                 if (!work)
 553                         break;
 554                 io_assign_current_work(worker, work);
 555
 556                 /* handle a whole dependent link */
 557                 do {
 558                         struct io_wq_work *next_hashed, *linked;
 559                         unsigned int hash = io_get_work_hash(work);
 560
 561                         next_hashed = wq_next_work(work);
 562                         io_impersonate_work(worker, work);
 563                         wq->do_work(work);
 564                         io_assign_current_work(worker, NULL);
 565
 566                         linked = wq->free_work(work);
 567                         work = next_hashed;
 568                         if (!work && linked && !io_wq_is_hashed(linked)) {
 569                                 work = linked;
 570                                 linked = NULL;
 571                         }
 572                         io_assign_current_work(worker, work);
 573                         if (linked)
 574                                 io_wqe_enqueue(wqe, linked);
 575
 576                         if (hash != -1U && !next_hashed) {
 577                                 raw_spin_lock_irq(&wqe->lock);
 578                                 wqe->hash_map &= ~BIT_ULL(hash);
 579                                 wqe->flags &= ~IO_WQE_FLAG_STALLED;
 580                                 /* skip unnecessary unlock-lock wqe->lock */
 581                                 if (!work)
 582                                         goto get_next;
 583                                 raw_spin_unlock_irq(&wqe->lock);
 584                         }
 585                 } while (work);
 586
 587                 raw_spin_lock_irq(&wqe->lock);
 588         } while (1);
 589 }
 590
 591 static int io_wqe_worker(void *data)
 592 {
 593         struct io_worker *worker = data;
 594         struct io_wqe *wqe = worker->wqe;
 595         struct io_wq *wq = wqe->wq;
 596
 597         io_worker_start(wqe, worker);
 598
 599         while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 600                 set_current_state(TASK_INTERRUPTIBLE);
 601 loop:
 602                 raw_spin_lock_irq(&wqe->lock);
 603                 if (io_wqe_run_queue(wqe)) {
 604                         __set_current_state(TASK_RUNNING);
 605                         io_worker_handle_work(worker);
 606                         goto loop;
 607                 }
 608                 /* drops the lock on success, retry */
 609                 if (__io_worker_idle(wqe, worker)) {
 610                         __release(&wqe->lock);
 611                         goto loop;
 612                 }
 613                 raw_spin_unlock_irq(&wqe->lock);
 614                 if (signal_pending(current))
 615                         flush_signals(current);
 616                 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
 617                         continue;
 618                 /* timed out, exit unless we're the fixed worker */
 619                 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 620                     !(worker->flags & IO_WORKER_F_FIXED))
 621                         break;
 622         }
 623
 624         if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 625                 raw_spin_lock_irq(&wqe->lock);
 626                 if (!wq_list_empty(&wqe->work_list))
 627                         io_worker_handle_work(worker);
 628                 else
 629                         raw_spin_unlock_irq(&wqe->lock);
 630         }
 631
 632         io_worker_exit(worker);
 633         return 0;
 634 }
 635
 636 /*
 637  * Called when a worker is scheduled in. Mark us as currently running.
 638  */
 639 void io_wq_worker_running(struct task_struct *tsk)
 640 {
 641         struct io_worker *worker = kthread_data(tsk);
 642         struct io_wqe *wqe = worker->wqe;
 643
 644         if (!(worker->flags & IO_WORKER_F_UP))
 645                 return;
 646         if (worker->flags & IO_WORKER_F_RUNNING)
 647                 return;
 648         worker->flags |= IO_WORKER_F_RUNNING;
 649         io_wqe_inc_running(wqe, worker);
 650 }
 651
 652 /*
 653  * Called when worker is going to sleep. If there are no workers currently
 654  * running and we have work pending, wake up a free one or have the manager
 655  * set one up.
 656  */
 657 void io_wq_worker_sleeping(struct task_struct *tsk)
 658 {
 659         struct io_worker *worker = kthread_data(tsk);
 660         struct io_wqe *wqe = worker->wqe;
 661
 662         if (!(worker->flags & IO_WORKER_F_UP))
 663                 return;
 664         if (!(worker->flags & IO_WORKER_F_RUNNING))
 665                 return;
 666
 667         worker->flags &= ~IO_WORKER_F_RUNNING;
 668
 669         raw_spin_lock_irq(&wqe->lock);
 670         io_wqe_dec_running(wqe, worker);
 671         raw_spin_unlock_irq(&wqe->lock);
 672 }
 673
 674 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
 675 {
 676         struct io_wqe_acct *acct = &wqe->acct[index];
 677         struct io_worker *worker;
 678
 679         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
 680         if (!worker)
 681                 return false;
 682
 683         refcount_set(&worker->ref, 1);
 684         worker->nulls_node.pprev = NULL;
 685         worker->wqe = wqe;
 686         spin_lock_init(&worker->lock);
 687
 688         worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
 689                                 "io_wqe_worker-%d/%d", index, wqe->node);
 690         if (IS_ERR(worker->task)) {
 691                 kfree(worker);
 692                 return false;
 693         }
 694         kthread_bind_mask(worker->task, cpumask_of_node(wqe->node));
 695
 696         raw_spin_lock_irq(&wqe->lock);
 697         hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 698         list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 699         worker->flags |= IO_WORKER_F_FREE;
 700         if (index == IO_WQ_ACCT_BOUND)
 701                 worker->flags |= IO_WORKER_F_BOUND;
 702         if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
 703                 worker->flags |= IO_WORKER_F_FIXED;
 704         acct->nr_workers++;
 705         raw_spin_unlock_irq(&wqe->lock);
 706
 707         if (index == IO_WQ_ACCT_UNBOUND)
 708                 atomic_inc(&wq->user->processes);
 709
 710         refcount_inc(&wq->refs);
 711         wake_up_process(worker->task);
 712         return true;
 713 }
 714
 715 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
 716         __must_hold(wqe->lock)
 717 {
 718         struct io_wqe_acct *acct = &wqe->acct[index];
 719
 720         /* if we have available workers or no work, no need */
 721         if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
 722                 return false;
 723         return acct->nr_workers < acct->max_workers;
 724 }
 725
 726 /*
 727  * Iterate the passed in list and call the specific function for each
 728  * worker that isn't exiting
 729  */
 730 static bool io_wq_for_each_worker(struct io_wqe *wqe,
 731                                   bool (*func)(struct io_worker *, void *),
 732                                   void *data)
 733 {
 734         struct io_worker *worker;
 735         bool ret = false;
 736
 737         list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
 738                 if (io_worker_get(worker)) {
 739                         /* no task if node is/was offline */
 740                         if (worker->task)
 741                                 ret = func(worker, data);
 742                         io_worker_release(worker);
 743                         if (ret)
 744                                 break;
 745                 }
 746         }
 747
 748         return ret;
 749 }
 750
 751 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
 752 {
 753         wake_up_process(worker->task);
 754         return false;
 755 }
 756
 757 /*
 758  * Manager thread. Tasked with creating new workers, if we need them.
 759  */
 760 static int io_wq_manager(void *data)
 761 {
 762         struct io_wq *wq = data;
 763         int node;
 764
 765         /* create fixed workers */
 766         refcount_set(&wq->refs, 1);
 767         for_each_node(node) {
 768                 if (!node_online(node))
 769                         continue;
 770                 if (create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
 771                         continue;
 772                 set_bit(IO_WQ_BIT_ERROR, &wq->state);
 773                 set_bit(IO_WQ_BIT_EXIT, &wq->state);
 774                 goto out;
 775         }
 776
 777         complete(&wq->done);
 778
 779         while (!kthread_should_stop()) {
 780                 if (current->task_works)
 781                         task_work_run();
 782
 783                 for_each_node(node) {
 784                         struct io_wqe *wqe = wq->wqes[node];
 785                         bool fork_worker[2] = { false, false };
 786
 787                         if (!node_online(node))
 788                                 continue;
 789
 790                         raw_spin_lock_irq(&wqe->lock);
 791                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
 792                                 fork_worker[IO_WQ_ACCT_BOUND] = true;
 793                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
 794                                 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
 795                         raw_spin_unlock_irq(&wqe->lock);
 796                         if (fork_worker[IO_WQ_ACCT_BOUND])
 797                                 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
 798                         if (fork_worker[IO_WQ_ACCT_UNBOUND])
 799                                 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
 800                 }
 801                 set_current_state(TASK_INTERRUPTIBLE);
 802                 schedule_timeout(HZ);
 803         }
 804
 805         if (current->task_works)
 806                 task_work_run();
 807
 808 out:
 809         if (refcount_dec_and_test(&wq->refs)) {
 810                 complete(&wq->done);
 811                 return 0;
 812         }
 813         /* if ERROR is set and we get here, we have workers to wake */
 814         if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
 815                 rcu_read_lock();
 816                 for_each_node(node)
 817                         io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
 818                 rcu_read_unlock();
 819         }
 820         return 0;
 821 }
 822
 823 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
 824                             struct io_wq_work *work)
 825 {
 826         bool free_worker;
 827
 828         if (!(work->flags & IO_WQ_WORK_UNBOUND))
 829                 return true;
 830         if (atomic_read(&acct->nr_running))
 831                 return true;
 832
 833         rcu_read_lock();
 834         free_worker = !hlist_nulls_empty(&wqe->free_list);
 835         rcu_read_unlock();
 836         if (free_worker)
 837                 return true;
 838
 839         if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
 840             !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
 841                 return false;
 842
 843         return true;
 844 }
 845
 846 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
 847 {
 848         struct io_wq *wq = wqe->wq;
 849
 850         do {
 851                 work->flags |= IO_WQ_WORK_CANCEL;
 852                 wq->do_work(work);
 853                 work = wq->free_work(work);
 854         } while (work);
 855 }
 856
 857 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
 858 {
 859         unsigned int hash;
 860         struct io_wq_work *tail;
 861
 862         if (!io_wq_is_hashed(work)) {
 863 append:
 864                 wq_list_add_tail(&work->list, &wqe->work_list);
 865                 return;
 866         }
 867
 868         hash = io_get_work_hash(work);
 869         tail = wqe->hash_tail[hash];
 870         wqe->hash_tail[hash] = work;
 871         if (!tail)
 872                 goto append;
 873
 874         wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
 875 }
 876
 877 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 878 {
 879         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 880         int work_flags;
 881         unsigned long flags;
 882
 883         /*
 884          * Do early check to see if we need a new unbound worker, and if we do,
 885          * if we're allowed to do so. This isn't 100% accurate as there's a
 886          * gap between this check and incrementing the value, but that's OK.
 887          * It's close enough to not be an issue, fork() has the same delay.
 888          */
 889         if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
 890                 io_run_cancel(work, wqe);
 891                 return;
 892         }
 893
 894         work_flags = work->flags;
 895         raw_spin_lock_irqsave(&wqe->lock, flags);
 896         io_wqe_insert_work(wqe, work);
 897         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 898         raw_spin_unlock_irqrestore(&wqe->lock, flags);
 899
 900         if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 901             !atomic_read(&acct->nr_running))
 902                 io_wqe_wake_worker(wqe, acct);
 903 }
 904
 905 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 906 {
 907         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 908
 909         io_wqe_enqueue(wqe, work);
 910 }
 911
 912 /*
 913  * Work items that hash to the same value will not be done in parallel.
 914  * Used to limit concurrent writes, generally hashed by inode.
 915  */
 916 void io_wq_hash_work(struct io_wq_work *work, void *val)
 917 {
 918         unsigned int bit;
 919
 920         bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 921         work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 922 }
 923
 924 struct io_cb_cancel_data {
 925         work_cancel_fn *fn;
 926         void *data;
 927         int nr_running;
 928         int nr_pending;
 929         bool cancel_all;
 930 };
 931
 932 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 933 {
 934         struct io_cb_cancel_data *match = data;
 935         unsigned long flags;
 936
 937         /*
 938          * Hold the lock to avoid ->cur_work going out of scope, caller
 939          * may dereference the passed in work.
 940          */
 941         spin_lock_irqsave(&worker->lock, flags);
 942         if (worker->cur_work &&
 943             match->fn(worker->cur_work, match->data)) {
 944                 send_sig(SIGINT, worker->task, 1);
 945                 match->nr_running++;
 946         }
 947         spin_unlock_irqrestore(&worker->lock, flags);
 948
 949         return match->nr_running && !match->cancel_all;
 950 }
 951
 952 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
 953                                          struct io_wq_work *work,
 954                                          struct io_wq_work_node *prev)
 955 {
 956         unsigned int hash = io_get_work_hash(work);
 957         struct io_wq_work *prev_work = NULL;
 958
 959         if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
 960                 if (prev)
 961                         prev_work = container_of(prev, struct io_wq_work, list);
 962                 if (prev_work && io_get_work_hash(prev_work) == hash)
 963                         wqe->hash_tail[hash] = prev_work;
 964                 else
 965                         wqe->hash_tail[hash] = NULL;
 966         }
 967         wq_list_del(&wqe->work_list, &work->list, prev);
 968 }
 969
 970 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
 971                                        struct io_cb_cancel_data *match)
 972 {
 973         struct io_wq_work_node *node, *prev;
 974         struct io_wq_work *work;
 975         unsigned long flags;
 976
 977 retry:
 978         raw_spin_lock_irqsave(&wqe->lock, flags);
 979         wq_list_for_each(node, prev, &wqe->work_list) {
 980                 work = container_of(node, struct io_wq_work, list);
 981                 if (!match->fn(work, match->data))
 982                         continue;
 983                 io_wqe_remove_pending(wqe, work, prev);
 984                 raw_spin_unlock_irqrestore(&wqe->lock, flags);
 985                 io_run_cancel(work, wqe);
 986                 match->nr_pending++;
 987                 if (!match->cancel_all)
 988                         return;
 989
 990                 /* not safe to continue after unlock */
 991                 goto retry;
 992         }
 993         raw_spin_unlock_irqrestore(&wqe->lock, flags);
 994 }
 995
 996 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
 997                                        struct io_cb_cancel_data *match)
 998 {
 999         rcu_read_lock();
1000         io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1001         rcu_read_unlock();
1002 }
1003
1004 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1005                                   void *data, bool cancel_all)
1006 {
1007         struct io_cb_cancel_data match = {
1008                 .fn             = cancel,
1009                 .data           = data,
1010                 .cancel_all     = cancel_all,
1011         };
1012         int node;
1013
1014         /*
1015          * First check pending list, if we're lucky we can just remove it
1016          * from there. CANCEL_OK means that the work is returned as-new,
1017          * no completion will be posted for it.
1018          */
1019         for_each_node(node) {
1020                 struct io_wqe *wqe = wq->wqes[node];
1021
1022                 io_wqe_cancel_pending_work(wqe, &match);
1023                 if (match.nr_pending && !match.cancel_all)
1024                         return IO_WQ_CANCEL_OK;
1025         }
1026
1027         /*
1028          * Now check if a free (going busy) or busy worker has the work
1029          * currently running. If we find it there, we'll return CANCEL_RUNNING
1030          * as an indication that we attempt to signal cancellation. The
1031          * completion will run normally in this case.
1032          */
1033         for_each_node(node) {
1034                 struct io_wqe *wqe = wq->wqes[node];
1035
1036                 io_wqe_cancel_running_work(wqe, &match);
1037                 if (match.nr_running && !match.cancel_all)
1038                         return IO_WQ_CANCEL_RUNNING;
1039         }
1040
1041         if (match.nr_running)
1042                 return IO_WQ_CANCEL_RUNNING;
1043         if (match.nr_pending)
1044                 return IO_WQ_CANCEL_OK;
1045         return IO_WQ_CANCEL_NOTFOUND;
1046 }
1047
1048 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1049 {
1050         int ret = -ENOMEM, node;
1051         struct io_wq *wq;
1052
1053         if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1054                 return ERR_PTR(-EINVAL);
1055
1056         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1057         if (!wq)
1058                 return ERR_PTR(-ENOMEM);
1059
1060         wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1061         if (!wq->wqes)
1062                 goto err_wq;
1063
1064         ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1065         if (ret)
1066                 goto err_wqes;
1067
1068         wq->free_work = data->free_work;
1069         wq->do_work = data->do_work;
1070
1071         /* caller must already hold a reference to this */
1072         wq->user = data->user;
1073
1074         ret = -ENOMEM;
1075         for_each_node(node) {
1076                 struct io_wqe *wqe;
1077                 int alloc_node = node;
1078
1079                 if (!node_online(alloc_node))
1080                         alloc_node = NUMA_NO_NODE;
1081                 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1082                 if (!wqe)
1083                         goto err;
1084                 wq->wqes[node] = wqe;
1085                 wqe->node = alloc_node;
1086                 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1087                 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1088                 if (wq->user) {
1089                         wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1090                                         task_rlimit(current, RLIMIT_NPROC);
1091                 }
1092                 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1093                 wqe->wq = wq;
1094                 raw_spin_lock_init(&wqe->lock);
1095                 INIT_WQ_LIST(&wqe->work_list);
1096                 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1097                 INIT_LIST_HEAD(&wqe->all_list);
1098         }
1099
1100         init_completion(&wq->done);
1101
1102         wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1103         if (!IS_ERR(wq->manager)) {
1104                 wake_up_process(wq->manager);
1105                 wait_for_completion(&wq->done);
1106                 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1107                         ret = -ENOMEM;
1108                         goto err;
1109                 }
1110                 refcount_set(&wq->use_refs, 1);
1111                 reinit_completion(&wq->done);
1112                 return wq;
1113         }
1114
1115         ret = PTR_ERR(wq->manager);
1116         complete(&wq->done);
1117 err:
1118         cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1119         for_each_node(node)
1120                 kfree(wq->wqes[node]);
1121 err_wqes:
1122         kfree(wq->wqes);
1123 err_wq:
1124         kfree(wq);
1125         return ERR_PTR(ret);
1126 }
1127
1128 bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1129 {
1130         if (data->free_work != wq->free_work || data->do_work != wq->do_work)
1131                 return false;
1132
1133         return refcount_inc_not_zero(&wq->use_refs);
1134 }
1135
1136 static void __io_wq_destroy(struct io_wq *wq)
1137 {
1138         int node;
1139
1140         cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1141
1142         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1143         if (wq->manager)
1144                 kthread_stop(wq->manager);
1145
1146         rcu_read_lock();
1147         for_each_node(node)
1148                 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1149         rcu_read_unlock();
1150
1151         wait_for_completion(&wq->done);
1152
1153         for_each_node(node)
1154                 kfree(wq->wqes[node]);
1155         kfree(wq->wqes);
1156         kfree(wq);
1157 }
1158
1159 void io_wq_destroy(struct io_wq *wq)
1160 {
1161         if (refcount_dec_and_test(&wq->use_refs))
1162                 __io_wq_destroy(wq);
1163 }
1164
1165 struct task_struct *io_wq_get_task(struct io_wq *wq)
1166 {
1167         return wq->manager;
1168 }
1169
1170 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1171 {
1172         struct task_struct *task = worker->task;
1173         struct rq_flags rf;
1174         struct rq *rq;
1175
1176         rq = task_rq_lock(task, &rf);
1177         do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
1178         task->flags |= PF_NO_SETAFFINITY;
1179         task_rq_unlock(rq, task, &rf);
1180         return false;
1181 }
1182
1183 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1184 {
1185         struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1186         int i;
1187
1188         rcu_read_lock();
1189         for_each_node(i)
1190                 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
1191         rcu_read_unlock();
1192         return 0;
1193 }
1194
1195 static __init int io_wq_init(void)
1196 {
1197         int ret;
1198
1199         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1200                                         io_wq_cpu_online, NULL);
1201         if (ret < 0)
1202                 return ret;
1203         io_wq_online = ret;
1204         return 0;
1205 }
1206 subsys_initcall(io_wq_init);