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_CANCEL        = 1,    /* cancel work on list */
  40         IO_WQ_BIT_ERROR         = 2,    /* error on setup */
  41 };
  42
  43 enum {
  44         IO_WQE_FLAG_STALLED     = 1,    /* stalled on hash */
  45 };
  46
  47 /*
  48  * One for each thread in a wqe pool
  49  */
  50 struct io_worker {
  51         refcount_t ref;
  52         unsigned flags;
  53         struct hlist_nulls_node nulls_node;
  54         struct list_head all_list;
  55         struct task_struct *task;
  56         struct io_wqe *wqe;
  57
  58         struct io_wq_work *cur_work;
  59         spinlock_t lock;
  60
  61         struct rcu_head rcu;
  62         struct mm_struct *mm;
  63 #ifdef CONFIG_BLK_CGROUP
  64         struct cgroup_subsys_state *blkcg_css;
  65 #endif
  66         const struct cred *cur_creds;
  67         const struct cred *saved_creds;
  68         struct files_struct *restore_files;
  69         struct nsproxy *restore_nsproxy;
  70         struct fs_struct *restore_fs;
  71 };
  72
  73 #if BITS_PER_LONG == 64
  74 #define IO_WQ_HASH_ORDER        6
  75 #else
  76 #define IO_WQ_HASH_ORDER        5
  77 #endif
  78
  79 #define IO_WQ_NR_HASH_BUCKETS   (1u << IO_WQ_HASH_ORDER)
  80
  81 struct io_wqe_acct {
  82         unsigned nr_workers;
  83         unsigned max_workers;
  84         atomic_t nr_running;
  85 };
  86
  87 enum {
  88         IO_WQ_ACCT_BOUND,
  89         IO_WQ_ACCT_UNBOUND,
  90 };
  91
  92 /*
  93  * Per-node worker thread pool
  94  */
  95 struct io_wqe {
  96         struct {
  97                 raw_spinlock_t lock;
  98                 struct io_wq_work_list work_list;
  99                 unsigned long hash_map;
 100                 unsigned flags;
 101         } ____cacheline_aligned_in_smp;
 102
 103         int node;
 104         struct io_wqe_acct acct[2];
 105
 106         struct hlist_nulls_head free_list;
 107         struct list_head all_list;
 108
 109         struct io_wq *wq;
 110         struct io_wq_work *hash_tail[IO_WQ_NR_HASH_BUCKETS];
 111 };
 112
 113 /*
 114  * Per io_wq state
 115   */
 116 struct io_wq {
 117         struct io_wqe **wqes;
 118         unsigned long state;
 119
 120         free_work_fn *free_work;
 121         io_wq_work_fn *do_work;
 122
 123         struct task_struct *manager;
 124         struct user_struct *user;
 125         refcount_t refs;
 126         struct completion done;
 127
 128         struct hlist_node cpuhp_node;
 129
 130         refcount_t use_refs;
 131 };
 132
 133 static enum cpuhp_state io_wq_online;
 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                 wake_up_process(worker->task);
 144 }
 145
 146 /*
 147  * Note: drops the wqe->lock if returning true! The caller must re-acquire
 148  * the lock in that case. Some callers need to restart handling if this
 149  * happens, so we can't just re-acquire the lock on behalf of the caller.
 150  */
 151 static bool __io_worker_unuse(struct io_wqe *wqe, struct io_worker *worker)
 152 {
 153         bool dropped_lock = false;
 154
 155         if (worker->saved_creds) {
 156                 revert_creds(worker->saved_creds);
 157                 worker->cur_creds = worker->saved_creds = NULL;
 158         }
 159
 160         if (current->files != worker->restore_files) {
 161                 __acquire(&wqe->lock);
 162                 raw_spin_unlock_irq(&wqe->lock);
 163                 dropped_lock = true;
 164
 165                 task_lock(current);
 166                 current->files = worker->restore_files;
 167                 current->nsproxy = worker->restore_nsproxy;
 168                 task_unlock(current);
 169         }
 170
 171         if (current->fs != worker->restore_fs)
 172                 current->fs = worker->restore_fs;
 173
 174         /*
 175          * If we have an active mm, we need to drop the wq lock before unusing
 176          * it. If we do, return true and let the caller retry the idle loop.
 177          */
 178         if (worker->mm) {
 179                 if (!dropped_lock) {
 180                         __acquire(&wqe->lock);
 181                         raw_spin_unlock_irq(&wqe->lock);
 182                         dropped_lock = true;
 183                 }
 184                 __set_current_state(TASK_RUNNING);
 185                 kthread_unuse_mm(worker->mm);
 186                 mmput(worker->mm);
 187                 worker->mm = NULL;
 188         }
 189
 190 #ifdef CONFIG_BLK_CGROUP
 191         if (worker->blkcg_css) {
 192                 kthread_associate_blkcg(NULL);
 193                 worker->blkcg_css = NULL;
 194         }
 195 #endif
 196         if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
 197                 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
 198         return dropped_lock;
 199 }
 200
 201 static inline struct io_wqe_acct *io_work_get_acct(struct io_wqe *wqe,
 202                                                    struct io_wq_work *work)
 203 {
 204         if (work->flags & IO_WQ_WORK_UNBOUND)
 205                 return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 206
 207         return &wqe->acct[IO_WQ_ACCT_BOUND];
 208 }
 209
 210 static inline struct io_wqe_acct *io_wqe_get_acct(struct io_wqe *wqe,
 211                                                   struct io_worker *worker)
 212 {
 213         if (worker->flags & IO_WORKER_F_BOUND)
 214                 return &wqe->acct[IO_WQ_ACCT_BOUND];
 215
 216         return &wqe->acct[IO_WQ_ACCT_UNBOUND];
 217 }
 218
 219 static void io_worker_exit(struct io_worker *worker)
 220 {
 221         struct io_wqe *wqe = worker->wqe;
 222         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 223
 224         /*
 225          * If we're not at zero, someone else is holding a brief reference
 226          * to the worker. Wait for that to go away.
 227          */
 228         set_current_state(TASK_INTERRUPTIBLE);
 229         if (!refcount_dec_and_test(&worker->ref))
 230                 schedule();
 231         __set_current_state(TASK_RUNNING);
 232
 233         preempt_disable();
 234         current->flags &= ~PF_IO_WORKER;
 235         if (worker->flags & IO_WORKER_F_RUNNING)
 236                 atomic_dec(&acct->nr_running);
 237         if (!(worker->flags & IO_WORKER_F_BOUND))
 238                 atomic_dec(&wqe->wq->user->processes);
 239         worker->flags = 0;
 240         preempt_enable();
 241
 242         raw_spin_lock_irq(&wqe->lock);
 243         hlist_nulls_del_rcu(&worker->nulls_node);
 244         list_del_rcu(&worker->all_list);
 245         if (__io_worker_unuse(wqe, worker)) {
 246                 __release(&wqe->lock);
 247                 raw_spin_lock_irq(&wqe->lock);
 248         }
 249         acct->nr_workers--;
 250         raw_spin_unlock_irq(&wqe->lock);
 251
 252         kfree_rcu(worker, rcu);
 253         if (refcount_dec_and_test(&wqe->wq->refs))
 254                 complete(&wqe->wq->done);
 255 }
 256
 257 static inline bool io_wqe_run_queue(struct io_wqe *wqe)
 258         __must_hold(wqe->lock)
 259 {
 260         if (!wq_list_empty(&wqe->work_list) &&
 261             !(wqe->flags & IO_WQE_FLAG_STALLED))
 262                 return true;
 263         return false;
 264 }
 265
 266 /*
 267  * Check head of free list for an available worker. If one isn't available,
 268  * caller must wake up the wq manager to create one.
 269  */
 270 static bool io_wqe_activate_free_worker(struct io_wqe *wqe)
 271         __must_hold(RCU)
 272 {
 273         struct hlist_nulls_node *n;
 274         struct io_worker *worker;
 275
 276         n = rcu_dereference(hlist_nulls_first_rcu(&wqe->free_list));
 277         if (is_a_nulls(n))
 278                 return false;
 279
 280         worker = hlist_nulls_entry(n, struct io_worker, nulls_node);
 281         if (io_worker_get(worker)) {
 282                 wake_up_process(worker->task);
 283                 io_worker_release(worker);
 284                 return true;
 285         }
 286
 287         return false;
 288 }
 289
 290 /*
 291  * We need a worker. If we find a free one, we're good. If not, and we're
 292  * below the max number of workers, wake up the manager to create one.
 293  */
 294 static void io_wqe_wake_worker(struct io_wqe *wqe, struct io_wqe_acct *acct)
 295 {
 296         bool ret;
 297
 298         /*
 299          * Most likely an attempt to queue unbounded work on an io_wq that
 300          * wasn't setup with any unbounded workers.
 301          */
 302         WARN_ON_ONCE(!acct->max_workers);
 303
 304         rcu_read_lock();
 305         ret = io_wqe_activate_free_worker(wqe);
 306         rcu_read_unlock();
 307
 308         if (!ret && acct->nr_workers < acct->max_workers)
 309                 wake_up_process(wqe->wq->manager);
 310 }
 311
 312 static void io_wqe_inc_running(struct io_wqe *wqe, struct io_worker *worker)
 313 {
 314         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 315
 316         atomic_inc(&acct->nr_running);
 317 }
 318
 319 static void io_wqe_dec_running(struct io_wqe *wqe, struct io_worker *worker)
 320         __must_hold(wqe->lock)
 321 {
 322         struct io_wqe_acct *acct = io_wqe_get_acct(wqe, worker);
 323
 324         if (atomic_dec_and_test(&acct->nr_running) && io_wqe_run_queue(wqe))
 325                 io_wqe_wake_worker(wqe, acct);
 326 }
 327
 328 static void io_worker_start(struct io_wqe *wqe, struct io_worker *worker)
 329 {
 330         allow_kernel_signal(SIGINT);
 331
 332         current->flags |= PF_IO_WORKER;
 333
 334         worker->flags |= (IO_WORKER_F_UP | IO_WORKER_F_RUNNING);
 335         worker->restore_files = current->files;
 336         worker->restore_nsproxy = current->nsproxy;
 337         worker->restore_fs = current->fs;
 338         io_wqe_inc_running(wqe, worker);
 339 }
 340
 341 /*
 342  * Worker will start processing some work. Move it to the busy list, if
 343  * it's currently on the freelist
 344  */
 345 static void __io_worker_busy(struct io_wqe *wqe, struct io_worker *worker,
 346                              struct io_wq_work *work)
 347         __must_hold(wqe->lock)
 348 {
 349         bool worker_bound, work_bound;
 350
 351         if (worker->flags & IO_WORKER_F_FREE) {
 352                 worker->flags &= ~IO_WORKER_F_FREE;
 353                 hlist_nulls_del_init_rcu(&worker->nulls_node);
 354         }
 355
 356         /*
 357          * If worker is moving from bound to unbound (or vice versa), then
 358          * ensure we update the running accounting.
 359          */
 360         worker_bound = (worker->flags & IO_WORKER_F_BOUND) != 0;
 361         work_bound = (work->flags & IO_WQ_WORK_UNBOUND) == 0;
 362         if (worker_bound != work_bound) {
 363                 io_wqe_dec_running(wqe, worker);
 364                 if (work_bound) {
 365                         worker->flags |= IO_WORKER_F_BOUND;
 366                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers--;
 367                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers++;
 368                         atomic_dec(&wqe->wq->user->processes);
 369                 } else {
 370                         worker->flags &= ~IO_WORKER_F_BOUND;
 371                         wqe->acct[IO_WQ_ACCT_UNBOUND].nr_workers++;
 372                         wqe->acct[IO_WQ_ACCT_BOUND].nr_workers--;
 373                         atomic_inc(&wqe->wq->user->processes);
 374                 }
 375                 io_wqe_inc_running(wqe, worker);
 376          }
 377 }
 378
 379 /*
 380  * No work, worker going to sleep. Move to freelist, and unuse mm if we
 381  * have one attached. Dropping the mm may potentially sleep, so we drop
 382  * the lock in that case and return success. Since the caller has to
 383  * retry the loop in that case (we changed task state), we don't regrab
 384  * the lock if we return success.
 385  */
 386 static bool __io_worker_idle(struct io_wqe *wqe, struct io_worker *worker)
 387         __must_hold(wqe->lock)
 388 {
 389         if (!(worker->flags & IO_WORKER_F_FREE)) {
 390                 worker->flags |= IO_WORKER_F_FREE;
 391                 hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 392         }
 393
 394         return __io_worker_unuse(wqe, worker);
 395 }
 396
 397 static inline unsigned int io_get_work_hash(struct io_wq_work *work)
 398 {
 399         return work->flags >> IO_WQ_HASH_SHIFT;
 400 }
 401
 402 static struct io_wq_work *io_get_next_work(struct io_wqe *wqe)
 403         __must_hold(wqe->lock)
 404 {
 405         struct io_wq_work_node *node, *prev;
 406         struct io_wq_work *work, *tail;
 407         unsigned int hash;
 408
 409         wq_list_for_each(node, prev, &wqe->work_list) {
 410                 work = container_of(node, struct io_wq_work, list);
 411
 412                 /* not hashed, can run anytime */
 413                 if (!io_wq_is_hashed(work)) {
 414                         wq_list_del(&wqe->work_list, node, prev);
 415                         return work;
 416                 }
 417
 418                 /* hashed, can run if not already running */
 419                 hash = io_get_work_hash(work);
 420                 if (!(wqe->hash_map & BIT(hash))) {
 421                         wqe->hash_map |= BIT(hash);
 422                         /* all items with this hash lie in [work, tail] */
 423                         tail = wqe->hash_tail[hash];
 424                         wqe->hash_tail[hash] = NULL;
 425                         wq_list_cut(&wqe->work_list, &tail->list, prev);
 426                         return work;
 427                 }
 428         }
 429
 430         return NULL;
 431 }
 432
 433 static void io_wq_switch_mm(struct io_worker *worker, struct io_wq_work *work)
 434 {
 435         if (worker->mm) {
 436                 kthread_unuse_mm(worker->mm);
 437                 mmput(worker->mm);
 438                 worker->mm = NULL;
 439         }
 440
 441         if (mmget_not_zero(work->identity->mm)) {
 442                 kthread_use_mm(work->identity->mm);
 443                 worker->mm = work->identity->mm;
 444                 return;
 445         }
 446
 447         /* failed grabbing mm, ensure work gets cancelled */
 448         work->flags |= IO_WQ_WORK_CANCEL;
 449 }
 450
 451 static inline void io_wq_switch_blkcg(struct io_worker *worker,
 452                                       struct io_wq_work *work)
 453 {
 454 #ifdef CONFIG_BLK_CGROUP
 455         if (!(work->flags & IO_WQ_WORK_BLKCG))
 456                 return;
 457         if (work->identity->blkcg_css != worker->blkcg_css) {
 458                 kthread_associate_blkcg(work->identity->blkcg_css);
 459                 worker->blkcg_css = work->identity->blkcg_css;
 460         }
 461 #endif
 462 }
 463
 464 static void io_wq_switch_creds(struct io_worker *worker,
 465                                struct io_wq_work *work)
 466 {
 467         const struct cred *old_creds = override_creds(work->identity->creds);
 468
 469         worker->cur_creds = work->identity->creds;
 470         if (worker->saved_creds)
 471                 put_cred(old_creds); /* creds set by previous switch */
 472         else
 473                 worker->saved_creds = old_creds;
 474 }
 475
 476 static void io_impersonate_work(struct io_worker *worker,
 477                                 struct io_wq_work *work)
 478 {
 479         if ((work->flags & IO_WQ_WORK_FILES) &&
 480             current->files != work->identity->files) {
 481                 task_lock(current);
 482                 current->files = work->identity->files;
 483                 current->nsproxy = work->identity->nsproxy;
 484                 task_unlock(current);
 485         }
 486         if ((work->flags & IO_WQ_WORK_FS) && current->fs != work->identity->fs)
 487                 current->fs = work->identity->fs;
 488         if ((work->flags & IO_WQ_WORK_MM) && work->identity->mm != worker->mm)
 489                 io_wq_switch_mm(worker, work);
 490         if ((work->flags & IO_WQ_WORK_CREDS) &&
 491             worker->cur_creds != work->identity->creds)
 492                 io_wq_switch_creds(worker, work);
 493         if (work->flags & IO_WQ_WORK_FSIZE)
 494                 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = work->identity->fsize;
 495         else if (current->signal->rlim[RLIMIT_FSIZE].rlim_cur != RLIM_INFINITY)
 496                 current->signal->rlim[RLIMIT_FSIZE].rlim_cur = RLIM_INFINITY;
 497         io_wq_switch_blkcg(worker, work);
 498 #ifdef CONFIG_AUDIT
 499         current->loginuid = work->identity->loginuid;
 500         current->sessionid = work->identity->sessionid;
 501 #endif
 502 }
 503
 504 static void io_assign_current_work(struct io_worker *worker,
 505                                    struct io_wq_work *work)
 506 {
 507         if (work) {
 508                 /* flush pending signals before assigning new work */
 509                 if (signal_pending(current))
 510                         flush_signals(current);
 511                 cond_resched();
 512         }
 513
 514 #ifdef CONFIG_AUDIT
 515         current->loginuid = KUIDT_INIT(AUDIT_UID_UNSET);
 516         current->sessionid = AUDIT_SID_UNSET;
 517 #endif
 518
 519         spin_lock_irq(&worker->lock);
 520         worker->cur_work = work;
 521         spin_unlock_irq(&worker->lock);
 522 }
 523
 524 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work);
 525
 526 static void io_worker_handle_work(struct io_worker *worker)
 527         __releases(wqe->lock)
 528 {
 529         struct io_wqe *wqe = worker->wqe;
 530         struct io_wq *wq = wqe->wq;
 531
 532         do {
 533                 struct io_wq_work *work;
 534 get_next:
 535                 /*
 536                  * If we got some work, mark us as busy. If we didn't, but
 537                  * the list isn't empty, it means we stalled on hashed work.
 538                  * Mark us stalled so we don't keep looking for work when we
 539                  * can't make progress, any work completion or insertion will
 540                  * clear the stalled flag.
 541                  */
 542                 work = io_get_next_work(wqe);
 543                 if (work)
 544                         __io_worker_busy(wqe, worker, work);
 545                 else if (!wq_list_empty(&wqe->work_list))
 546                         wqe->flags |= IO_WQE_FLAG_STALLED;
 547
 548                 raw_spin_unlock_irq(&wqe->lock);
 549                 if (!work)
 550                         break;
 551                 io_assign_current_work(worker, work);
 552
 553                 /* handle a whole dependent link */
 554                 do {
 555                         struct io_wq_work *old_work, *next_hashed, *linked;
 556                         unsigned int hash = io_get_work_hash(work);
 557
 558                         next_hashed = wq_next_work(work);
 559                         io_impersonate_work(worker, work);
 560                         /*
 561                          * OK to set IO_WQ_WORK_CANCEL even for uncancellable
 562                          * work, the worker function will do the right thing.
 563                          */
 564                         if (test_bit(IO_WQ_BIT_CANCEL, &wq->state))
 565                                 work->flags |= IO_WQ_WORK_CANCEL;
 566
 567                         old_work = work;
 568                         linked = wq->do_work(work);
 569
 570                         work = next_hashed;
 571                         if (!work && linked && !io_wq_is_hashed(linked)) {
 572                                 work = linked;
 573                                 linked = NULL;
 574                         }
 575                         io_assign_current_work(worker, work);
 576                         wq->free_work(old_work);
 577
 578                         if (linked)
 579                                 io_wqe_enqueue(wqe, linked);
 580
 581                         if (hash != -1U && !next_hashed) {
 582                                 raw_spin_lock_irq(&wqe->lock);
 583                                 wqe->hash_map &= ~BIT_ULL(hash);
 584                                 wqe->flags &= ~IO_WQE_FLAG_STALLED;
 585                                 /* skip unnecessary unlock-lock wqe->lock */
 586                                 if (!work)
 587                                         goto get_next;
 588                                 raw_spin_unlock_irq(&wqe->lock);
 589                         }
 590                 } while (work);
 591
 592                 raw_spin_lock_irq(&wqe->lock);
 593         } while (1);
 594 }
 595
 596 static int io_wqe_worker(void *data)
 597 {
 598         struct io_worker *worker = data;
 599         struct io_wqe *wqe = worker->wqe;
 600         struct io_wq *wq = wqe->wq;
 601
 602         io_worker_start(wqe, worker);
 603
 604         while (!test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 605                 set_current_state(TASK_INTERRUPTIBLE);
 606 loop:
 607                 raw_spin_lock_irq(&wqe->lock);
 608                 if (io_wqe_run_queue(wqe)) {
 609                         __set_current_state(TASK_RUNNING);
 610                         io_worker_handle_work(worker);
 611                         goto loop;
 612                 }
 613                 /* drops the lock on success, retry */
 614                 if (__io_worker_idle(wqe, worker)) {
 615                         __release(&wqe->lock);
 616                         goto loop;
 617                 }
 618                 raw_spin_unlock_irq(&wqe->lock);
 619                 if (signal_pending(current))
 620                         flush_signals(current);
 621                 if (schedule_timeout(WORKER_IDLE_TIMEOUT))
 622                         continue;
 623                 /* timed out, exit unless we're the fixed worker */
 624                 if (test_bit(IO_WQ_BIT_EXIT, &wq->state) ||
 625                     !(worker->flags & IO_WORKER_F_FIXED))
 626                         break;
 627         }
 628
 629         if (test_bit(IO_WQ_BIT_EXIT, &wq->state)) {
 630                 raw_spin_lock_irq(&wqe->lock);
 631                 if (!wq_list_empty(&wqe->work_list))
 632                         io_worker_handle_work(worker);
 633                 else
 634                         raw_spin_unlock_irq(&wqe->lock);
 635         }
 636
 637         io_worker_exit(worker);
 638         return 0;
 639 }
 640
 641 /*
 642  * Called when a worker is scheduled in. Mark us as currently running.
 643  */
 644 void io_wq_worker_running(struct task_struct *tsk)
 645 {
 646         struct io_worker *worker = kthread_data(tsk);
 647         struct io_wqe *wqe = worker->wqe;
 648
 649         if (!(worker->flags & IO_WORKER_F_UP))
 650                 return;
 651         if (worker->flags & IO_WORKER_F_RUNNING)
 652                 return;
 653         worker->flags |= IO_WORKER_F_RUNNING;
 654         io_wqe_inc_running(wqe, worker);
 655 }
 656
 657 /*
 658  * Called when worker is going to sleep. If there are no workers currently
 659  * running and we have work pending, wake up a free one or have the manager
 660  * set one up.
 661  */
 662 void io_wq_worker_sleeping(struct task_struct *tsk)
 663 {
 664         struct io_worker *worker = kthread_data(tsk);
 665         struct io_wqe *wqe = worker->wqe;
 666
 667         if (!(worker->flags & IO_WORKER_F_UP))
 668                 return;
 669         if (!(worker->flags & IO_WORKER_F_RUNNING))
 670                 return;
 671
 672         worker->flags &= ~IO_WORKER_F_RUNNING;
 673
 674         raw_spin_lock_irq(&wqe->lock);
 675         io_wqe_dec_running(wqe, worker);
 676         raw_spin_unlock_irq(&wqe->lock);
 677 }
 678
 679 static bool create_io_worker(struct io_wq *wq, struct io_wqe *wqe, int index)
 680 {
 681         struct io_wqe_acct *acct = &wqe->acct[index];
 682         struct io_worker *worker;
 683
 684         worker = kzalloc_node(sizeof(*worker), GFP_KERNEL, wqe->node);
 685         if (!worker)
 686                 return false;
 687
 688         refcount_set(&worker->ref, 1);
 689         worker->nulls_node.pprev = NULL;
 690         worker->wqe = wqe;
 691         spin_lock_init(&worker->lock);
 692
 693         worker->task = kthread_create_on_node(io_wqe_worker, worker, wqe->node,
 694                                 "io_wqe_worker-%d/%d", index, wqe->node);
 695         if (IS_ERR(worker->task)) {
 696                 kfree(worker);
 697                 return false;
 698         }
 699         kthread_bind_mask(worker->task, cpumask_of_node(wqe->node));
 700
 701         raw_spin_lock_irq(&wqe->lock);
 702         hlist_nulls_add_head_rcu(&worker->nulls_node, &wqe->free_list);
 703         list_add_tail_rcu(&worker->all_list, &wqe->all_list);
 704         worker->flags |= IO_WORKER_F_FREE;
 705         if (index == IO_WQ_ACCT_BOUND)
 706                 worker->flags |= IO_WORKER_F_BOUND;
 707         if (!acct->nr_workers && (worker->flags & IO_WORKER_F_BOUND))
 708                 worker->flags |= IO_WORKER_F_FIXED;
 709         acct->nr_workers++;
 710         raw_spin_unlock_irq(&wqe->lock);
 711
 712         if (index == IO_WQ_ACCT_UNBOUND)
 713                 atomic_inc(&wq->user->processes);
 714
 715         refcount_inc(&wq->refs);
 716         wake_up_process(worker->task);
 717         return true;
 718 }
 719
 720 static inline bool io_wqe_need_worker(struct io_wqe *wqe, int index)
 721         __must_hold(wqe->lock)
 722 {
 723         struct io_wqe_acct *acct = &wqe->acct[index];
 724
 725         /* if we have available workers or no work, no need */
 726         if (!hlist_nulls_empty(&wqe->free_list) || !io_wqe_run_queue(wqe))
 727                 return false;
 728         return acct->nr_workers < acct->max_workers;
 729 }
 730
 731 static bool io_wqe_worker_send_sig(struct io_worker *worker, void *data)
 732 {
 733         send_sig(SIGINT, worker->task, 1);
 734         return false;
 735 }
 736
 737 /*
 738  * Iterate the passed in list and call the specific function for each
 739  * worker that isn't exiting
 740  */
 741 static bool io_wq_for_each_worker(struct io_wqe *wqe,
 742                                   bool (*func)(struct io_worker *, void *),
 743                                   void *data)
 744 {
 745         struct io_worker *worker;
 746         bool ret = false;
 747
 748         list_for_each_entry_rcu(worker, &wqe->all_list, all_list) {
 749                 if (io_worker_get(worker)) {
 750                         /* no task if node is/was offline */
 751                         if (worker->task)
 752                                 ret = func(worker, data);
 753                         io_worker_release(worker);
 754                         if (ret)
 755                                 break;
 756                 }
 757         }
 758
 759         return ret;
 760 }
 761
 762 static bool io_wq_worker_wake(struct io_worker *worker, void *data)
 763 {
 764         wake_up_process(worker->task);
 765         return false;
 766 }
 767
 768 /*
 769  * Manager thread. Tasked with creating new workers, if we need them.
 770  */
 771 static int io_wq_manager(void *data)
 772 {
 773         struct io_wq *wq = data;
 774         int node;
 775
 776         /* create fixed workers */
 777         refcount_set(&wq->refs, 1);
 778         for_each_node(node) {
 779                 if (!node_online(node))
 780                         continue;
 781                 if (create_io_worker(wq, wq->wqes[node], IO_WQ_ACCT_BOUND))
 782                         continue;
 783                 set_bit(IO_WQ_BIT_ERROR, &wq->state);
 784                 set_bit(IO_WQ_BIT_EXIT, &wq->state);
 785                 goto out;
 786         }
 787
 788         complete(&wq->done);
 789
 790         while (!kthread_should_stop()) {
 791                 if (current->task_works)
 792                         task_work_run();
 793
 794                 for_each_node(node) {
 795                         struct io_wqe *wqe = wq->wqes[node];
 796                         bool fork_worker[2] = { false, false };
 797
 798                         if (!node_online(node))
 799                                 continue;
 800
 801                         raw_spin_lock_irq(&wqe->lock);
 802                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_BOUND))
 803                                 fork_worker[IO_WQ_ACCT_BOUND] = true;
 804                         if (io_wqe_need_worker(wqe, IO_WQ_ACCT_UNBOUND))
 805                                 fork_worker[IO_WQ_ACCT_UNBOUND] = true;
 806                         raw_spin_unlock_irq(&wqe->lock);
 807                         if (fork_worker[IO_WQ_ACCT_BOUND])
 808                                 create_io_worker(wq, wqe, IO_WQ_ACCT_BOUND);
 809                         if (fork_worker[IO_WQ_ACCT_UNBOUND])
 810                                 create_io_worker(wq, wqe, IO_WQ_ACCT_UNBOUND);
 811                 }
 812                 set_current_state(TASK_INTERRUPTIBLE);
 813                 schedule_timeout(HZ);
 814         }
 815
 816         if (current->task_works)
 817                 task_work_run();
 818
 819 out:
 820         if (refcount_dec_and_test(&wq->refs)) {
 821                 complete(&wq->done);
 822                 return 0;
 823         }
 824         /* if ERROR is set and we get here, we have workers to wake */
 825         if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
 826                 rcu_read_lock();
 827                 for_each_node(node)
 828                         io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
 829                 rcu_read_unlock();
 830         }
 831         return 0;
 832 }
 833
 834 static bool io_wq_can_queue(struct io_wqe *wqe, struct io_wqe_acct *acct,
 835                             struct io_wq_work *work)
 836 {
 837         bool free_worker;
 838
 839         if (!(work->flags & IO_WQ_WORK_UNBOUND))
 840                 return true;
 841         if (atomic_read(&acct->nr_running))
 842                 return true;
 843
 844         rcu_read_lock();
 845         free_worker = !hlist_nulls_empty(&wqe->free_list);
 846         rcu_read_unlock();
 847         if (free_worker)
 848                 return true;
 849
 850         if (atomic_read(&wqe->wq->user->processes) >= acct->max_workers &&
 851             !(capable(CAP_SYS_RESOURCE) || capable(CAP_SYS_ADMIN)))
 852                 return false;
 853
 854         return true;
 855 }
 856
 857 static void io_run_cancel(struct io_wq_work *work, struct io_wqe *wqe)
 858 {
 859         struct io_wq *wq = wqe->wq;
 860
 861         do {
 862                 struct io_wq_work *old_work = work;
 863
 864                 work->flags |= IO_WQ_WORK_CANCEL;
 865                 work = wq->do_work(work);
 866                 wq->free_work(old_work);
 867         } while (work);
 868 }
 869
 870 static void io_wqe_insert_work(struct io_wqe *wqe, struct io_wq_work *work)
 871 {
 872         unsigned int hash;
 873         struct io_wq_work *tail;
 874
 875         if (!io_wq_is_hashed(work)) {
 876 append:
 877                 wq_list_add_tail(&work->list, &wqe->work_list);
 878                 return;
 879         }
 880
 881         hash = io_get_work_hash(work);
 882         tail = wqe->hash_tail[hash];
 883         wqe->hash_tail[hash] = work;
 884         if (!tail)
 885                 goto append;
 886
 887         wq_list_add_after(&work->list, &tail->list, &wqe->work_list);
 888 }
 889
 890 static void io_wqe_enqueue(struct io_wqe *wqe, struct io_wq_work *work)
 891 {
 892         struct io_wqe_acct *acct = io_work_get_acct(wqe, work);
 893         int work_flags;
 894         unsigned long flags;
 895
 896         /*
 897          * Do early check to see if we need a new unbound worker, and if we do,
 898          * if we're allowed to do so. This isn't 100% accurate as there's a
 899          * gap between this check and incrementing the value, but that's OK.
 900          * It's close enough to not be an issue, fork() has the same delay.
 901          */
 902         if (unlikely(!io_wq_can_queue(wqe, acct, work))) {
 903                 io_run_cancel(work, wqe);
 904                 return;
 905         }
 906
 907         work_flags = work->flags;
 908         raw_spin_lock_irqsave(&wqe->lock, flags);
 909         io_wqe_insert_work(wqe, work);
 910         wqe->flags &= ~IO_WQE_FLAG_STALLED;
 911         raw_spin_unlock_irqrestore(&wqe->lock, flags);
 912
 913         if ((work_flags & IO_WQ_WORK_CONCURRENT) ||
 914             !atomic_read(&acct->nr_running))
 915                 io_wqe_wake_worker(wqe, acct);
 916 }
 917
 918 void io_wq_enqueue(struct io_wq *wq, struct io_wq_work *work)
 919 {
 920         struct io_wqe *wqe = wq->wqes[numa_node_id()];
 921
 922         io_wqe_enqueue(wqe, work);
 923 }
 924
 925 /*
 926  * Work items that hash to the same value will not be done in parallel.
 927  * Used to limit concurrent writes, generally hashed by inode.
 928  */
 929 void io_wq_hash_work(struct io_wq_work *work, void *val)
 930 {
 931         unsigned int bit;
 932
 933         bit = hash_ptr(val, IO_WQ_HASH_ORDER);
 934         work->flags |= (IO_WQ_WORK_HASHED | (bit << IO_WQ_HASH_SHIFT));
 935 }
 936
 937 void io_wq_cancel_all(struct io_wq *wq)
 938 {
 939         int node;
 940
 941         set_bit(IO_WQ_BIT_CANCEL, &wq->state);
 942
 943         rcu_read_lock();
 944         for_each_node(node) {
 945                 struct io_wqe *wqe = wq->wqes[node];
 946
 947                 io_wq_for_each_worker(wqe, io_wqe_worker_send_sig, NULL);
 948         }
 949         rcu_read_unlock();
 950 }
 951
 952 struct io_cb_cancel_data {
 953         work_cancel_fn *fn;
 954         void *data;
 955         int nr_running;
 956         int nr_pending;
 957         bool cancel_all;
 958 };
 959
 960 static bool io_wq_worker_cancel(struct io_worker *worker, void *data)
 961 {
 962         struct io_cb_cancel_data *match = data;
 963         unsigned long flags;
 964
 965         /*
 966          * Hold the lock to avoid ->cur_work going out of scope, caller
 967          * may dereference the passed in work.
 968          */
 969         spin_lock_irqsave(&worker->lock, flags);
 970         if (worker->cur_work &&
 971             !(worker->cur_work->flags & IO_WQ_WORK_NO_CANCEL) &&
 972             match->fn(worker->cur_work, match->data)) {
 973                 send_sig(SIGINT, worker->task, 1);
 974                 match->nr_running++;
 975         }
 976         spin_unlock_irqrestore(&worker->lock, flags);
 977
 978         return match->nr_running && !match->cancel_all;
 979 }
 980
 981 static inline void io_wqe_remove_pending(struct io_wqe *wqe,
 982                                          struct io_wq_work *work,
 983                                          struct io_wq_work_node *prev)
 984 {
 985         unsigned int hash = io_get_work_hash(work);
 986         struct io_wq_work *prev_work = NULL;
 987
 988         if (io_wq_is_hashed(work) && work == wqe->hash_tail[hash]) {
 989                 if (prev)
 990                         prev_work = container_of(prev, struct io_wq_work, list);
 991                 if (prev_work && io_get_work_hash(prev_work) == hash)
 992                         wqe->hash_tail[hash] = prev_work;
 993                 else
 994                         wqe->hash_tail[hash] = NULL;
 995         }
 996         wq_list_del(&wqe->work_list, &work->list, prev);
 997 }
 998
 999 static void io_wqe_cancel_pending_work(struct io_wqe *wqe,
1000                                        struct io_cb_cancel_data *match)
1001 {
1002         struct io_wq_work_node *node, *prev;
1003         struct io_wq_work *work;
1004         unsigned long flags;
1005
1006 retry:
1007         raw_spin_lock_irqsave(&wqe->lock, flags);
1008         wq_list_for_each(node, prev, &wqe->work_list) {
1009                 work = container_of(node, struct io_wq_work, list);
1010                 if (!match->fn(work, match->data))
1011                         continue;
1012                 io_wqe_remove_pending(wqe, work, prev);
1013                 raw_spin_unlock_irqrestore(&wqe->lock, flags);
1014                 io_run_cancel(work, wqe);
1015                 match->nr_pending++;
1016                 if (!match->cancel_all)
1017                         return;
1018
1019                 /* not safe to continue after unlock */
1020                 goto retry;
1021         }
1022         raw_spin_unlock_irqrestore(&wqe->lock, flags);
1023 }
1024
1025 static void io_wqe_cancel_running_work(struct io_wqe *wqe,
1026                                        struct io_cb_cancel_data *match)
1027 {
1028         rcu_read_lock();
1029         io_wq_for_each_worker(wqe, io_wq_worker_cancel, match);
1030         rcu_read_unlock();
1031 }
1032
1033 enum io_wq_cancel io_wq_cancel_cb(struct io_wq *wq, work_cancel_fn *cancel,
1034                                   void *data, bool cancel_all)
1035 {
1036         struct io_cb_cancel_data match = {
1037                 .fn             = cancel,
1038                 .data           = data,
1039                 .cancel_all     = cancel_all,
1040         };
1041         int node;
1042
1043         /*
1044          * First check pending list, if we're lucky we can just remove it
1045          * from there. CANCEL_OK means that the work is returned as-new,
1046          * no completion will be posted for it.
1047          */
1048         for_each_node(node) {
1049                 struct io_wqe *wqe = wq->wqes[node];
1050
1051                 io_wqe_cancel_pending_work(wqe, &match);
1052                 if (match.nr_pending && !match.cancel_all)
1053                         return IO_WQ_CANCEL_OK;
1054         }
1055
1056         /*
1057          * Now check if a free (going busy) or busy worker has the work
1058          * currently running. If we find it there, we'll return CANCEL_RUNNING
1059          * as an indication that we attempt to signal cancellation. The
1060          * completion will run normally in this case.
1061          */
1062         for_each_node(node) {
1063                 struct io_wqe *wqe = wq->wqes[node];
1064
1065                 io_wqe_cancel_running_work(wqe, &match);
1066                 if (match.nr_running && !match.cancel_all)
1067                         return IO_WQ_CANCEL_RUNNING;
1068         }
1069
1070         if (match.nr_running)
1071                 return IO_WQ_CANCEL_RUNNING;
1072         if (match.nr_pending)
1073                 return IO_WQ_CANCEL_OK;
1074         return IO_WQ_CANCEL_NOTFOUND;
1075 }
1076
1077 static bool io_wq_io_cb_cancel_data(struct io_wq_work *work, void *data)
1078 {
1079         return work == data;
1080 }
1081
1082 enum io_wq_cancel io_wq_cancel_work(struct io_wq *wq, struct io_wq_work *cwork)
1083 {
1084         return io_wq_cancel_cb(wq, io_wq_io_cb_cancel_data, (void *)cwork, false);
1085 }
1086
1087 struct io_wq *io_wq_create(unsigned bounded, struct io_wq_data *data)
1088 {
1089         int ret = -ENOMEM, node;
1090         struct io_wq *wq;
1091
1092         if (WARN_ON_ONCE(!data->free_work || !data->do_work))
1093                 return ERR_PTR(-EINVAL);
1094
1095         wq = kzalloc(sizeof(*wq), GFP_KERNEL);
1096         if (!wq)
1097                 return ERR_PTR(-ENOMEM);
1098
1099         wq->wqes = kcalloc(nr_node_ids, sizeof(struct io_wqe *), GFP_KERNEL);
1100         if (!wq->wqes)
1101                 goto err_wq;
1102
1103         ret = cpuhp_state_add_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1104         if (ret)
1105                 goto err_wqes;
1106
1107         wq->free_work = data->free_work;
1108         wq->do_work = data->do_work;
1109
1110         /* caller must already hold a reference to this */
1111         wq->user = data->user;
1112
1113         ret = -ENOMEM;
1114         for_each_node(node) {
1115                 struct io_wqe *wqe;
1116                 int alloc_node = node;
1117
1118                 if (!node_online(alloc_node))
1119                         alloc_node = NUMA_NO_NODE;
1120                 wqe = kzalloc_node(sizeof(struct io_wqe), GFP_KERNEL, alloc_node);
1121                 if (!wqe)
1122                         goto err;
1123                 wq->wqes[node] = wqe;
1124                 wqe->node = alloc_node;
1125                 wqe->acct[IO_WQ_ACCT_BOUND].max_workers = bounded;
1126                 atomic_set(&wqe->acct[IO_WQ_ACCT_BOUND].nr_running, 0);
1127                 if (wq->user) {
1128                         wqe->acct[IO_WQ_ACCT_UNBOUND].max_workers =
1129                                         task_rlimit(current, RLIMIT_NPROC);
1130                 }
1131                 atomic_set(&wqe->acct[IO_WQ_ACCT_UNBOUND].nr_running, 0);
1132                 wqe->wq = wq;
1133                 raw_spin_lock_init(&wqe->lock);
1134                 INIT_WQ_LIST(&wqe->work_list);
1135                 INIT_HLIST_NULLS_HEAD(&wqe->free_list, 0);
1136                 INIT_LIST_HEAD(&wqe->all_list);
1137         }
1138
1139         init_completion(&wq->done);
1140
1141         wq->manager = kthread_create(io_wq_manager, wq, "io_wq_manager");
1142         if (!IS_ERR(wq->manager)) {
1143                 wake_up_process(wq->manager);
1144                 wait_for_completion(&wq->done);
1145                 if (test_bit(IO_WQ_BIT_ERROR, &wq->state)) {
1146                         ret = -ENOMEM;
1147                         goto err;
1148                 }
1149                 refcount_set(&wq->use_refs, 1);
1150                 reinit_completion(&wq->done);
1151                 return wq;
1152         }
1153
1154         ret = PTR_ERR(wq->manager);
1155         complete(&wq->done);
1156 err:
1157         cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1158         for_each_node(node)
1159                 kfree(wq->wqes[node]);
1160 err_wqes:
1161         kfree(wq->wqes);
1162 err_wq:
1163         kfree(wq);
1164         return ERR_PTR(ret);
1165 }
1166
1167 bool io_wq_get(struct io_wq *wq, struct io_wq_data *data)
1168 {
1169         if (data->free_work != wq->free_work || data->do_work != wq->do_work)
1170                 return false;
1171
1172         return refcount_inc_not_zero(&wq->use_refs);
1173 }
1174
1175 static void __io_wq_destroy(struct io_wq *wq)
1176 {
1177         int node;
1178
1179         cpuhp_state_remove_instance_nocalls(io_wq_online, &wq->cpuhp_node);
1180
1181         set_bit(IO_WQ_BIT_EXIT, &wq->state);
1182         if (wq->manager)
1183                 kthread_stop(wq->manager);
1184
1185         rcu_read_lock();
1186         for_each_node(node)
1187                 io_wq_for_each_worker(wq->wqes[node], io_wq_worker_wake, NULL);
1188         rcu_read_unlock();
1189
1190         wait_for_completion(&wq->done);
1191
1192         for_each_node(node)
1193                 kfree(wq->wqes[node]);
1194         kfree(wq->wqes);
1195         kfree(wq);
1196 }
1197
1198 void io_wq_destroy(struct io_wq *wq)
1199 {
1200         if (refcount_dec_and_test(&wq->use_refs))
1201                 __io_wq_destroy(wq);
1202 }
1203
1204 struct task_struct *io_wq_get_task(struct io_wq *wq)
1205 {
1206         return wq->manager;
1207 }
1208
1209 static bool io_wq_worker_affinity(struct io_worker *worker, void *data)
1210 {
1211         struct task_struct *task = worker->task;
1212         struct rq_flags rf;
1213         struct rq *rq;
1214
1215         rq = task_rq_lock(task, &rf);
1216         do_set_cpus_allowed(task, cpumask_of_node(worker->wqe->node));
1217         task->flags |= PF_NO_SETAFFINITY;
1218         task_rq_unlock(rq, task, &rf);
1219         return false;
1220 }
1221
1222 static int io_wq_cpu_online(unsigned int cpu, struct hlist_node *node)
1223 {
1224         struct io_wq *wq = hlist_entry_safe(node, struct io_wq, cpuhp_node);
1225         int i;
1226
1227         rcu_read_lock();
1228         for_each_node(i)
1229                 io_wq_for_each_worker(wq->wqes[i], io_wq_worker_affinity, NULL);
1230         rcu_read_unlock();
1231         return 0;
1232 }
1233
1234 static __init int io_wq_init(void)
1235 {
1236         int ret;
1237
1238         ret = cpuhp_setup_state_multi(CPUHP_AP_ONLINE_DYN, "io-wq/online",
1239                                         io_wq_cpu_online, NULL);
1240         if (ret < 0)
1241                 return ret;
1242         io_wq_online = ret;
1243         return 0;
1244 }
1245 subsys_initcall(io_wq_init);