mm/backing-dev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include <linux/wait.h>
   4 #include <linux/rbtree.h>
   5 #include <linux/backing-dev.h>
   6 #include <linux/kthread.h>
   7 #include <linux/freezer.h>
   8 #include <linux/fs.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/mm.h>
  11 #include <linux/sched/mm.h>
  12 #include <linux/sched.h>
  13 #include <linux/module.h>
  14 #include <linux/writeback.h>
  15 #include <linux/device.h>
  16 #include <trace/events/writeback.h>
  17
  18 struct backing_dev_info noop_backing_dev_info;
  19 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  20
  21 static struct class *bdi_class;
  22 static const char *bdi_unknown_name = "(unknown)";
  23
  24 /*
  25  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
  26  * reader side locking.
  27  */
  28 DEFINE_SPINLOCK(bdi_lock);
  29 static u64 bdi_id_cursor;
  30 static struct rb_root bdi_tree = RB_ROOT;
  31 LIST_HEAD(bdi_list);
  32
  33 /* bdi_wq serves all asynchronous writeback tasks */
  34 struct workqueue_struct *bdi_wq;
  35
  36 #define K(x) ((x) << (PAGE_SHIFT - 10))
  37
  38 #ifdef CONFIG_DEBUG_FS
  39 #include <linux/debugfs.h>
  40 #include <linux/seq_file.h>
  41
  42 static struct dentry *bdi_debug_root;
  43
  44 static void bdi_debug_init(void)
  45 {
  46         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  47 }
  48
  49 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  50 {
  51         struct backing_dev_info *bdi = m->private;
  52         struct bdi_writeback *wb = &bdi->wb;
  53         unsigned long background_thresh;
  54         unsigned long dirty_thresh;
  55         unsigned long wb_thresh;
  56         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  57         struct inode *inode;
  58
  59         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  60         spin_lock(&wb->list_lock);
  61         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  62                 nr_dirty++;
  63         list_for_each_entry(inode, &wb->b_io, i_io_list)
  64                 nr_io++;
  65         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  66                 nr_more_io++;
  67         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  68                 if (inode->i_state & I_DIRTY_TIME)
  69                         nr_dirty_time++;
  70         spin_unlock(&wb->list_lock);
  71
  72         global_dirty_limits(&background_thresh, &dirty_thresh);
  73         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  74
  75         seq_printf(m,
  76                    "BdiWriteback:       %10lu kB\n"
  77                    "BdiReclaimable:     %10lu kB\n"
  78                    "BdiDirtyThresh:     %10lu kB\n"
  79                    "DirtyThresh:        %10lu kB\n"
  80                    "BackgroundThresh:   %10lu kB\n"
  81                    "BdiDirtied:         %10lu kB\n"
  82                    "BdiWritten:         %10lu kB\n"
  83                    "BdiWriteBandwidth:  %10lu kBps\n"
  84                    "b_dirty:            %10lu\n"
  85                    "b_io:               %10lu\n"
  86                    "b_more_io:          %10lu\n"
  87                    "b_dirty_time:       %10lu\n"
  88                    "bdi_list:           %10u\n"
  89                    "state:              %10lx\n",
  90                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  91                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  92                    K(wb_thresh),
  93                    K(dirty_thresh),
  94                    K(background_thresh),
  95                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  96                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  97                    (unsigned long) K(wb->write_bandwidth),
  98                    nr_dirty,
  99                    nr_io,
 100                    nr_more_io,
 101                    nr_dirty_time,
 102                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 103
 104         return 0;
 105 }
 106 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 107
 108 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 109 {
 110         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 111
 112         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 113                             &bdi_debug_stats_fops);
 114 }
 115
 116 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 117 {
 118         debugfs_remove_recursive(bdi->debug_dir);
 119 }
 120 #else
 121 static inline void bdi_debug_init(void)
 122 {
 123 }
 124 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 125                                       const char *name)
 126 {
 127 }
 128 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 129 {
 130 }
 131 #endif
 132
 133 static ssize_t read_ahead_kb_store(struct device *dev,
 134                                   struct device_attribute *attr,
 135                                   const char *buf, size_t count)
 136 {
 137         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 138         unsigned long read_ahead_kb;
 139         ssize_t ret;
 140
 141         ret = kstrtoul(buf, 10, &read_ahead_kb);
 142         if (ret < 0)
 143                 return ret;
 144
 145         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 146
 147         return count;
 148 }
 149
 150 #define BDI_SHOW(name, expr)                                            \
 151 static ssize_t name##_show(struct device *dev,                          \
 152                            struct device_attribute *attr, char *buf)    \
 153 {                                                                       \
 154         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 155                                                                         \
 156         return sysfs_emit(buf, "%lld\n", (long long)expr);              \
 157 }                                                                       \
 158 static DEVICE_ATTR_RW(name);
 159
 160 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 161
 162 static ssize_t min_ratio_store(struct device *dev,
 163                 struct device_attribute *attr, const char *buf, size_t count)
 164 {
 165         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 166         unsigned int ratio;
 167         ssize_t ret;
 168
 169         ret = kstrtouint(buf, 10, &ratio);
 170         if (ret < 0)
 171                 return ret;
 172
 173         ret = bdi_set_min_ratio(bdi, ratio);
 174         if (!ret)
 175                 ret = count;
 176
 177         return ret;
 178 }
 179 BDI_SHOW(min_ratio, bdi->min_ratio)
 180
 181 static ssize_t max_ratio_store(struct device *dev,
 182                 struct device_attribute *attr, const char *buf, size_t count)
 183 {
 184         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 185         unsigned int ratio;
 186         ssize_t ret;
 187
 188         ret = kstrtouint(buf, 10, &ratio);
 189         if (ret < 0)
 190                 return ret;
 191
 192         ret = bdi_set_max_ratio(bdi, ratio);
 193         if (!ret)
 194                 ret = count;
 195
 196         return ret;
 197 }
 198 BDI_SHOW(max_ratio, bdi->max_ratio)
 199
 200 static ssize_t stable_pages_required_show(struct device *dev,
 201                                           struct device_attribute *attr,
 202                                           char *buf)
 203 {
 204         dev_warn_once(dev,
 205                 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
 206         return sysfs_emit(buf, "%d\n", 0);
 207 }
 208 static DEVICE_ATTR_RO(stable_pages_required);
 209
 210 static struct attribute *bdi_dev_attrs[] = {
 211         &dev_attr_read_ahead_kb.attr,
 212         &dev_attr_min_ratio.attr,
 213         &dev_attr_max_ratio.attr,
 214         &dev_attr_stable_pages_required.attr,
 215         NULL,
 216 };
 217 ATTRIBUTE_GROUPS(bdi_dev);
 218
 219 static __init int bdi_class_init(void)
 220 {
 221         bdi_class = class_create(THIS_MODULE, "bdi");
 222         if (IS_ERR(bdi_class))
 223                 return PTR_ERR(bdi_class);
 224
 225         bdi_class->dev_groups = bdi_dev_groups;
 226         bdi_debug_init();
 227
 228         return 0;
 229 }
 230 postcore_initcall(bdi_class_init);
 231
 232 static int __init default_bdi_init(void)
 233 {
 234         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 235                                  WQ_SYSFS, 0);
 236         if (!bdi_wq)
 237                 return -ENOMEM;
 238         return 0;
 239 }
 240 subsys_initcall(default_bdi_init);
 241
 242 /*
 243  * This function is used when the first inode for this wb is marked dirty. It
 244  * wakes-up the corresponding bdi thread which should then take care of the
 245  * periodic background write-out of dirty inodes. Since the write-out would
 246  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 247  * set up a timer which wakes the bdi thread up later.
 248  *
 249  * Note, we wouldn't bother setting up the timer, but this function is on the
 250  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 251  * by delaying the wake-up.
 252  *
 253  * We have to be careful not to postpone flush work if it is scheduled for
 254  * earlier. Thus we use queue_delayed_work().
 255  */
 256 void wb_wakeup_delayed(struct bdi_writeback *wb)
 257 {
 258         unsigned long timeout;
 259
 260         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 261         spin_lock_irq(&wb->work_lock);
 262         if (test_bit(WB_registered, &wb->state))
 263                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 264         spin_unlock_irq(&wb->work_lock);
 265 }
 266
 267 static void wb_update_bandwidth_workfn(struct work_struct *work)
 268 {
 269         struct bdi_writeback *wb = container_of(to_delayed_work(work),
 270                                                 struct bdi_writeback, bw_dwork);
 271
 272         wb_update_bandwidth(wb);
 273 }
 274
 275 /*
 276  * Initial write bandwidth: 100 MB/s
 277  */
 278 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 279
 280 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 281                    gfp_t gfp)
 282 {
 283         int i, err;
 284
 285         memset(wb, 0, sizeof(*wb));
 286
 287         if (wb != &bdi->wb)
 288                 bdi_get(bdi);
 289         wb->bdi = bdi;
 290         wb->last_old_flush = jiffies;
 291         INIT_LIST_HEAD(&wb->b_dirty);
 292         INIT_LIST_HEAD(&wb->b_io);
 293         INIT_LIST_HEAD(&wb->b_more_io);
 294         INIT_LIST_HEAD(&wb->b_dirty_time);
 295         spin_lock_init(&wb->list_lock);
 296
 297         atomic_set(&wb->writeback_inodes, 0);
 298         wb->bw_time_stamp = jiffies;
 299         wb->balanced_dirty_ratelimit = INIT_BW;
 300         wb->dirty_ratelimit = INIT_BW;
 301         wb->write_bandwidth = INIT_BW;
 302         wb->avg_write_bandwidth = INIT_BW;
 303
 304         spin_lock_init(&wb->work_lock);
 305         INIT_LIST_HEAD(&wb->work_list);
 306         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 307         INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
 308         wb->dirty_sleep = jiffies;
 309
 310         err = fprop_local_init_percpu(&wb->completions, gfp);
 311         if (err)
 312                 goto out_put_bdi;
 313
 314         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 315                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 316                 if (err)
 317                         goto out_destroy_stat;
 318         }
 319
 320         return 0;
 321
 322 out_destroy_stat:
 323         while (i--)
 324                 percpu_counter_destroy(&wb->stat[i]);
 325         fprop_local_destroy_percpu(&wb->completions);
 326 out_put_bdi:
 327         if (wb != &bdi->wb)
 328                 bdi_put(bdi);
 329         return err;
 330 }
 331
 332 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 333
 334 /*
 335  * Remove bdi from the global list and shutdown any threads we have running
 336  */
 337 static void wb_shutdown(struct bdi_writeback *wb)
 338 {
 339         /* Make sure nobody queues further work */
 340         spin_lock_irq(&wb->work_lock);
 341         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 342                 spin_unlock_irq(&wb->work_lock);
 343                 return;
 344         }
 345         spin_unlock_irq(&wb->work_lock);
 346
 347         cgwb_remove_from_bdi_list(wb);
 348         /*
 349          * Drain work list and shutdown the delayed_work.  !WB_registered
 350          * tells wb_workfn() that @wb is dying and its work_list needs to
 351          * be drained no matter what.
 352          */
 353         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 354         flush_delayed_work(&wb->dwork);
 355         WARN_ON(!list_empty(&wb->work_list));
 356         flush_delayed_work(&wb->bw_dwork);
 357 }
 358
 359 static void wb_exit(struct bdi_writeback *wb)
 360 {
 361         int i;
 362
 363         WARN_ON(delayed_work_pending(&wb->dwork));
 364
 365         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 366                 percpu_counter_destroy(&wb->stat[i]);
 367
 368         fprop_local_destroy_percpu(&wb->completions);
 369         if (wb != &wb->bdi->wb)
 370                 bdi_put(wb->bdi);
 371 }
 372
 373 #ifdef CONFIG_CGROUP_WRITEBACK
 374
 375 #include <linux/memcontrol.h>
 376
 377 /*
 378  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
 379  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
 380  */
 381 static DEFINE_SPINLOCK(cgwb_lock);
 382 static struct workqueue_struct *cgwb_release_wq;
 383
 384 static LIST_HEAD(offline_cgwbs);
 385 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
 386 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
 387
 388 static void cgwb_free_rcu(struct rcu_head *rcu_head)
 389 {
 390         struct bdi_writeback *wb = container_of(rcu_head,
 391                         struct bdi_writeback, rcu);
 392
 393         percpu_ref_exit(&wb->refcnt);
 394         kfree(wb);
 395 }
 396
 397 static void cgwb_release_workfn(struct work_struct *work)
 398 {
 399         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 400                                                 release_work);
 401         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 402
 403         mutex_lock(&wb->bdi->cgwb_release_mutex);
 404         wb_shutdown(wb);
 405
 406         css_put(wb->memcg_css);
 407         css_put(wb->blkcg_css);
 408         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 409
 410         /* triggers blkg destruction if no online users left */
 411         blkcg_unpin_online(blkcg);
 412
 413         fprop_local_destroy_percpu(&wb->memcg_completions);
 414
 415         spin_lock_irq(&cgwb_lock);
 416         list_del(&wb->offline_node);
 417         spin_unlock_irq(&cgwb_lock);
 418
 419         wb_exit(wb);
 420         WARN_ON_ONCE(!list_empty(&wb->b_attached));
 421         call_rcu(&wb->rcu, cgwb_free_rcu);
 422 }
 423
 424 static void cgwb_release(struct percpu_ref *refcnt)
 425 {
 426         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 427                                                 refcnt);
 428         queue_work(cgwb_release_wq, &wb->release_work);
 429 }
 430
 431 static void cgwb_kill(struct bdi_writeback *wb)
 432 {
 433         lockdep_assert_held(&cgwb_lock);
 434
 435         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 436         list_del(&wb->memcg_node);
 437         list_del(&wb->blkcg_node);
 438         list_add(&wb->offline_node, &offline_cgwbs);
 439         percpu_ref_kill(&wb->refcnt);
 440 }
 441
 442 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 443 {
 444         spin_lock_irq(&cgwb_lock);
 445         list_del_rcu(&wb->bdi_node);
 446         spin_unlock_irq(&cgwb_lock);
 447 }
 448
 449 static int cgwb_create(struct backing_dev_info *bdi,
 450                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 451 {
 452         struct mem_cgroup *memcg;
 453         struct cgroup_subsys_state *blkcg_css;
 454         struct blkcg *blkcg;
 455         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 456         struct bdi_writeback *wb;
 457         unsigned long flags;
 458         int ret = 0;
 459
 460         memcg = mem_cgroup_from_css(memcg_css);
 461         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 462         blkcg = css_to_blkcg(blkcg_css);
 463         memcg_cgwb_list = &memcg->cgwb_list;
 464         blkcg_cgwb_list = &blkcg->cgwb_list;
 465
 466         /* look up again under lock and discard on blkcg mismatch */
 467         spin_lock_irqsave(&cgwb_lock, flags);
 468         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 469         if (wb && wb->blkcg_css != blkcg_css) {
 470                 cgwb_kill(wb);
 471                 wb = NULL;
 472         }
 473         spin_unlock_irqrestore(&cgwb_lock, flags);
 474         if (wb)
 475                 goto out_put;
 476
 477         /* need to create a new one */
 478         wb = kmalloc(sizeof(*wb), gfp);
 479         if (!wb) {
 480                 ret = -ENOMEM;
 481                 goto out_put;
 482         }
 483
 484         ret = wb_init(wb, bdi, gfp);
 485         if (ret)
 486                 goto err_free;
 487
 488         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 489         if (ret)
 490                 goto err_wb_exit;
 491
 492         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 493         if (ret)
 494                 goto err_ref_exit;
 495
 496         wb->memcg_css = memcg_css;
 497         wb->blkcg_css = blkcg_css;
 498         INIT_LIST_HEAD(&wb->b_attached);
 499         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 500         set_bit(WB_registered, &wb->state);
 501
 502         /*
 503          * The root wb determines the registered state of the whole bdi and
 504          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 505          * whether they're still online.  Don't link @wb if any is dead.
 506          * See wb_memcg_offline() and wb_blkcg_offline().
 507          */
 508         ret = -ENODEV;
 509         spin_lock_irqsave(&cgwb_lock, flags);
 510         if (test_bit(WB_registered, &bdi->wb.state) &&
 511             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 512                 /* we might have raced another instance of this function */
 513                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 514                 if (!ret) {
 515                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 516                         list_add(&wb->memcg_node, memcg_cgwb_list);
 517                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 518                         blkcg_pin_online(blkcg);
 519                         css_get(memcg_css);
 520                         css_get(blkcg_css);
 521                 }
 522         }
 523         spin_unlock_irqrestore(&cgwb_lock, flags);
 524         if (ret) {
 525                 if (ret == -EEXIST)
 526                         ret = 0;
 527                 goto err_fprop_exit;
 528         }
 529         goto out_put;
 530
 531 err_fprop_exit:
 532         fprop_local_destroy_percpu(&wb->memcg_completions);
 533 err_ref_exit:
 534         percpu_ref_exit(&wb->refcnt);
 535 err_wb_exit:
 536         wb_exit(wb);
 537 err_free:
 538         kfree(wb);
 539 out_put:
 540         css_put(blkcg_css);
 541         return ret;
 542 }
 543
 544 /**
 545  * wb_get_lookup - get wb for a given memcg
 546  * @bdi: target bdi
 547  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 548  *
 549  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 550  * refcount incremented.
 551  *
 552  * This function uses css_get() on @memcg_css and thus expects its refcnt
 553  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 554  * @memcg_css isn't enough.  try_get it before calling this function.
 555  *
 556  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 557  * memcg on the default hierarchy, memcg association is guaranteed to be
 558  * more specific (equal or descendant to the associated blkcg) and thus can
 559  * identify both the memcg and blkcg associations.
 560  *
 561  * Because the blkcg associated with a memcg may change as blkcg is enabled
 562  * and disabled closer to root in the hierarchy, each wb keeps track of
 563  * both the memcg and blkcg associated with it and verifies the blkcg on
 564  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 565  * created.
 566  */
 567 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 568                                     struct cgroup_subsys_state *memcg_css)
 569 {
 570         struct bdi_writeback *wb;
 571
 572         if (!memcg_css->parent)
 573                 return &bdi->wb;
 574
 575         rcu_read_lock();
 576         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 577         if (wb) {
 578                 struct cgroup_subsys_state *blkcg_css;
 579
 580                 /* see whether the blkcg association has changed */
 581                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 582                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 583                         wb = NULL;
 584                 css_put(blkcg_css);
 585         }
 586         rcu_read_unlock();
 587
 588         return wb;
 589 }
 590
 591 /**
 592  * wb_get_create - get wb for a given memcg, create if necessary
 593  * @bdi: target bdi
 594  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 595  * @gfp: allocation mask to use
 596  *
 597  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 598  * create one.  See wb_get_lookup() for more details.
 599  */
 600 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 601                                     struct cgroup_subsys_state *memcg_css,
 602                                     gfp_t gfp)
 603 {
 604         struct bdi_writeback *wb;
 605
 606         might_alloc(gfp);
 607
 608         if (!memcg_css->parent)
 609                 return &bdi->wb;
 610
 611         do {
 612                 wb = wb_get_lookup(bdi, memcg_css);
 613         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 614
 615         return wb;
 616 }
 617
 618 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 619 {
 620         int ret;
 621
 622         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 623         mutex_init(&bdi->cgwb_release_mutex);
 624         init_rwsem(&bdi->wb_switch_rwsem);
 625
 626         ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 627         if (!ret) {
 628                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 629                 bdi->wb.blkcg_css = blkcg_root_css;
 630         }
 631         return ret;
 632 }
 633
 634 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 635 {
 636         struct radix_tree_iter iter;
 637         void **slot;
 638         struct bdi_writeback *wb;
 639
 640         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 641
 642         spin_lock_irq(&cgwb_lock);
 643         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 644                 cgwb_kill(*slot);
 645         spin_unlock_irq(&cgwb_lock);
 646
 647         mutex_lock(&bdi->cgwb_release_mutex);
 648         spin_lock_irq(&cgwb_lock);
 649         while (!list_empty(&bdi->wb_list)) {
 650                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 651                                       bdi_node);
 652                 spin_unlock_irq(&cgwb_lock);
 653                 wb_shutdown(wb);
 654                 spin_lock_irq(&cgwb_lock);
 655         }
 656         spin_unlock_irq(&cgwb_lock);
 657         mutex_unlock(&bdi->cgwb_release_mutex);
 658 }
 659
 660 /*
 661  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
 662  *
 663  * Try to release dying cgwbs by switching attached inodes to the nearest
 664  * living ancestor's writeback. Processed wbs are placed at the end
 665  * of the list to guarantee the forward progress.
 666  */
 667 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
 668 {
 669         struct bdi_writeback *wb;
 670         LIST_HEAD(processed);
 671
 672         spin_lock_irq(&cgwb_lock);
 673
 674         while (!list_empty(&offline_cgwbs)) {
 675                 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
 676                                       offline_node);
 677                 list_move(&wb->offline_node, &processed);
 678
 679                 /*
 680                  * If wb is dirty, cleaning up the writeback by switching
 681                  * attached inodes will result in an effective removal of any
 682                  * bandwidth restrictions, which isn't the goal.  Instead,
 683                  * it can be postponed until the next time, when all io
 684                  * will be likely completed.  If in the meantime some inodes
 685                  * will get re-dirtied, they should be eventually switched to
 686                  * a new cgwb.
 687                  */
 688                 if (wb_has_dirty_io(wb))
 689                         continue;
 690
 691                 if (!wb_tryget(wb))
 692                         continue;
 693
 694                 spin_unlock_irq(&cgwb_lock);
 695                 while (cleanup_offline_cgwb(wb))
 696                         cond_resched();
 697                 spin_lock_irq(&cgwb_lock);
 698
 699                 wb_put(wb);
 700         }
 701
 702         if (!list_empty(&processed))
 703                 list_splice_tail(&processed, &offline_cgwbs);
 704
 705         spin_unlock_irq(&cgwb_lock);
 706 }
 707
 708 /**
 709  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 710  * @memcg: memcg being offlined
 711  *
 712  * Also prevents creation of any new wb's associated with @memcg.
 713  */
 714 void wb_memcg_offline(struct mem_cgroup *memcg)
 715 {
 716         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 717         struct bdi_writeback *wb, *next;
 718
 719         spin_lock_irq(&cgwb_lock);
 720         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 721                 cgwb_kill(wb);
 722         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 723         spin_unlock_irq(&cgwb_lock);
 724
 725         queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
 726 }
 727
 728 /**
 729  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 730  * @blkcg: blkcg being offlined
 731  *
 732  * Also prevents creation of any new wb's associated with @blkcg.
 733  */
 734 void wb_blkcg_offline(struct blkcg *blkcg)
 735 {
 736         struct bdi_writeback *wb, *next;
 737
 738         spin_lock_irq(&cgwb_lock);
 739         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 740                 cgwb_kill(wb);
 741         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 742         spin_unlock_irq(&cgwb_lock);
 743 }
 744
 745 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 746 {
 747         spin_lock_irq(&cgwb_lock);
 748         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 749         spin_unlock_irq(&cgwb_lock);
 750 }
 751
 752 static int __init cgwb_init(void)
 753 {
 754         /*
 755          * There can be many concurrent release work items overwhelming
 756          * system_wq.  Put them in a separate wq and limit concurrency.
 757          * There's no point in executing many of these in parallel.
 758          */
 759         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 760         if (!cgwb_release_wq)
 761                 return -ENOMEM;
 762
 763         return 0;
 764 }
 765 subsys_initcall(cgwb_init);
 766
 767 #else   /* CONFIG_CGROUP_WRITEBACK */
 768
 769 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 770 {
 771         return wb_init(&bdi->wb, bdi, GFP_KERNEL);
 772 }
 773
 774 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 775
 776 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 777 {
 778         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 779 }
 780
 781 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 782 {
 783         list_del_rcu(&wb->bdi_node);
 784 }
 785
 786 #endif  /* CONFIG_CGROUP_WRITEBACK */
 787
 788 int bdi_init(struct backing_dev_info *bdi)
 789 {
 790         int ret;
 791
 792         bdi->dev = NULL;
 793
 794         kref_init(&bdi->refcnt);
 795         bdi->min_ratio = 0;
 796         bdi->max_ratio = 100;
 797         bdi->max_prop_frac = FPROP_FRAC_BASE;
 798         INIT_LIST_HEAD(&bdi->bdi_list);
 799         INIT_LIST_HEAD(&bdi->wb_list);
 800         init_waitqueue_head(&bdi->wb_waitq);
 801
 802         ret = cgwb_bdi_init(bdi);
 803
 804         return ret;
 805 }
 806
 807 struct backing_dev_info *bdi_alloc(int node_id)
 808 {
 809         struct backing_dev_info *bdi;
 810
 811         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 812         if (!bdi)
 813                 return NULL;
 814
 815         if (bdi_init(bdi)) {
 816                 kfree(bdi);
 817                 return NULL;
 818         }
 819         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 820         bdi->ra_pages = VM_READAHEAD_PAGES;
 821         bdi->io_pages = VM_READAHEAD_PAGES;
 822         timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
 823         return bdi;
 824 }
 825 EXPORT_SYMBOL(bdi_alloc);
 826
 827 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 828 {
 829         struct rb_node **p = &bdi_tree.rb_node;
 830         struct rb_node *parent = NULL;
 831         struct backing_dev_info *bdi;
 832
 833         lockdep_assert_held(&bdi_lock);
 834
 835         while (*p) {
 836                 parent = *p;
 837                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 838
 839                 if (bdi->id > id)
 840                         p = &(*p)->rb_left;
 841                 else if (bdi->id < id)
 842                         p = &(*p)->rb_right;
 843                 else
 844                         break;
 845         }
 846
 847         if (parentp)
 848                 *parentp = parent;
 849         return p;
 850 }
 851
 852 /**
 853  * bdi_get_by_id - lookup and get bdi from its id
 854  * @id: bdi id to lookup
 855  *
 856  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 857  * doesn't exist or is already unregistered.
 858  */
 859 struct backing_dev_info *bdi_get_by_id(u64 id)
 860 {
 861         struct backing_dev_info *bdi = NULL;
 862         struct rb_node **p;
 863
 864         spin_lock_bh(&bdi_lock);
 865         p = bdi_lookup_rb_node(id, NULL);
 866         if (*p) {
 867                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 868                 bdi_get(bdi);
 869         }
 870         spin_unlock_bh(&bdi_lock);
 871
 872         return bdi;
 873 }
 874
 875 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 876 {
 877         struct device *dev;
 878         struct rb_node *parent, **p;
 879
 880         if (bdi->dev)   /* The driver needs to use separate queues per device */
 881                 return 0;
 882
 883         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
 884         dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
 885         if (IS_ERR(dev))
 886                 return PTR_ERR(dev);
 887
 888         cgwb_bdi_register(bdi);
 889         bdi->dev = dev;
 890
 891         bdi_debug_register(bdi, dev_name(dev));
 892         set_bit(WB_registered, &bdi->wb.state);
 893
 894         spin_lock_bh(&bdi_lock);
 895
 896         bdi->id = ++bdi_id_cursor;
 897
 898         p = bdi_lookup_rb_node(bdi->id, &parent);
 899         rb_link_node(&bdi->rb_node, parent, p);
 900         rb_insert_color(&bdi->rb_node, &bdi_tree);
 901
 902         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 903
 904         spin_unlock_bh(&bdi_lock);
 905
 906         trace_writeback_bdi_register(bdi);
 907         return 0;
 908 }
 909
 910 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 911 {
 912         va_list args;
 913         int ret;
 914
 915         va_start(args, fmt);
 916         ret = bdi_register_va(bdi, fmt, args);
 917         va_end(args);
 918         return ret;
 919 }
 920 EXPORT_SYMBOL(bdi_register);
 921
 922 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
 923 {
 924         WARN_ON_ONCE(bdi->owner);
 925         bdi->owner = owner;
 926         get_device(owner);
 927 }
 928
 929 /*
 930  * Remove bdi from bdi_list, and ensure that it is no longer visible
 931  */
 932 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 933 {
 934         spin_lock_bh(&bdi_lock);
 935         rb_erase(&bdi->rb_node, &bdi_tree);
 936         list_del_rcu(&bdi->bdi_list);
 937         spin_unlock_bh(&bdi_lock);
 938
 939         synchronize_rcu_expedited();
 940 }
 941
 942 void bdi_unregister(struct backing_dev_info *bdi)
 943 {
 944         del_timer_sync(&bdi->laptop_mode_wb_timer);
 945
 946         /* make sure nobody finds us on the bdi_list anymore */
 947         bdi_remove_from_list(bdi);
 948         wb_shutdown(&bdi->wb);
 949         cgwb_bdi_unregister(bdi);
 950
 951         /*
 952          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
 953          * update the global bdi_min_ratio.
 954          */
 955         if (bdi->min_ratio)
 956                 bdi_set_min_ratio(bdi, 0);
 957
 958         if (bdi->dev) {
 959                 bdi_debug_unregister(bdi);
 960                 device_unregister(bdi->dev);
 961                 bdi->dev = NULL;
 962         }
 963
 964         if (bdi->owner) {
 965                 put_device(bdi->owner);
 966                 bdi->owner = NULL;
 967         }
 968 }
 969
 970 static void release_bdi(struct kref *ref)
 971 {
 972         struct backing_dev_info *bdi =
 973                         container_of(ref, struct backing_dev_info, refcnt);
 974
 975         if (test_bit(WB_registered, &bdi->wb.state))
 976                 bdi_unregister(bdi);
 977         WARN_ON_ONCE(bdi->dev);
 978         wb_exit(&bdi->wb);
 979         kfree(bdi);
 980 }
 981
 982 void bdi_put(struct backing_dev_info *bdi)
 983 {
 984         kref_put(&bdi->refcnt, release_bdi);
 985 }
 986 EXPORT_SYMBOL(bdi_put);
 987
 988 const char *bdi_dev_name(struct backing_dev_info *bdi)
 989 {
 990         if (!bdi || !bdi->dev)
 991                 return bdi_unknown_name;
 992         return bdi->dev_name;
 993 }
 994 EXPORT_SYMBOL_GPL(bdi_dev_name);
 995
 996 static wait_queue_head_t congestion_wqh[2] = {
 997                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 998                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 999         };
1000 static atomic_t nr_wb_congested[2];
1001
1002 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
1003 {
1004         wait_queue_head_t *wqh = &congestion_wqh[sync];
1005         enum wb_congested_state bit;
1006
1007         bit = sync ? WB_sync_congested : WB_async_congested;
1008         if (test_and_clear_bit(bit, &bdi->wb.congested))
1009                 atomic_dec(&nr_wb_congested[sync]);
1010         smp_mb__after_atomic();
1011         if (waitqueue_active(wqh))
1012                 wake_up(wqh);
1013 }
1014 EXPORT_SYMBOL(clear_bdi_congested);
1015
1016 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
1017 {
1018         enum wb_congested_state bit;
1019
1020         bit = sync ? WB_sync_congested : WB_async_congested;
1021         if (!test_and_set_bit(bit, &bdi->wb.congested))
1022                 atomic_inc(&nr_wb_congested[sync]);
1023 }
1024 EXPORT_SYMBOL(set_bdi_congested);
1025
1026 /**
1027  * congestion_wait - wait for a backing_dev to become uncongested
1028  * @sync: SYNC or ASYNC IO
1029  * @timeout: timeout in jiffies
1030  *
1031  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1032  * write congestion.  If no backing_devs are congested then just wait for the
1033  * next write to be completed.
1034  */
1035 long congestion_wait(int sync, long timeout)
1036 {
1037         long ret;
1038         unsigned long start = jiffies;
1039         DEFINE_WAIT(wait);
1040         wait_queue_head_t *wqh = &congestion_wqh[sync];
1041
1042         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1043         ret = io_schedule_timeout(timeout);
1044         finish_wait(wqh, &wait);
1045
1046         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1047                                         jiffies_to_usecs(jiffies - start));
1048
1049         return ret;
1050 }
1051 EXPORT_SYMBOL(congestion_wait);
1052
1053 /**
1054  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1055  * @sync: SYNC or ASYNC IO
1056  * @timeout: timeout in jiffies
1057  *
1058  * In the event of a congested backing_dev (any backing_dev) this waits
1059  * for up to @timeout jiffies for either a BDI to exit congestion of the
1060  * given @sync queue or a write to complete.
1061  *
1062  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1063  * it is the number of jiffies that were still remaining when the function
1064  * returned. return_value == timeout implies the function did not sleep.
1065  */
1066 long wait_iff_congested(int sync, long timeout)
1067 {
1068         long ret;
1069         unsigned long start = jiffies;
1070         DEFINE_WAIT(wait);
1071         wait_queue_head_t *wqh = &congestion_wqh[sync];
1072
1073         /*
1074          * If there is no congestion, yield if necessary instead
1075          * of sleeping on the congestion queue
1076          */
1077         if (atomic_read(&nr_wb_congested[sync]) == 0) {
1078                 cond_resched();
1079
1080                 /* In case we scheduled, work out time remaining */
1081                 ret = timeout - (jiffies - start);
1082                 if (ret < 0)
1083                         ret = 0;
1084
1085                 goto out;
1086         }
1087
1088         /* Sleep until uncongested or a write happens */
1089         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1090         ret = io_schedule_timeout(timeout);
1091         finish_wait(wqh, &wait);
1092
1093 out:
1094         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1095                                         jiffies_to_usecs(jiffies - start));
1096
1097         return ret;
1098 }
1099 EXPORT_SYMBOL(wait_iff_congested);