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