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