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 bdi_init(struct backing_dev_info *bdi);
 233
 234 static int __init default_bdi_init(void)
 235 {
 236         int err;
 237
 238         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 239                                  WQ_SYSFS, 0);
 240         if (!bdi_wq)
 241                 return -ENOMEM;
 242
 243         err = bdi_init(&noop_backing_dev_info);
 244
 245         return err;
 246 }
 247 subsys_initcall(default_bdi_init);
 248
 249 /*
 250  * This function is used when the first inode for this wb is marked dirty. It
 251  * wakes-up the corresponding bdi thread which should then take care of the
 252  * periodic background write-out of dirty inodes. Since the write-out would
 253  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 254  * set up a timer which wakes the bdi thread up later.
 255  *
 256  * Note, we wouldn't bother setting up the timer, but this function is on the
 257  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 258  * by delaying the wake-up.
 259  *
 260  * We have to be careful not to postpone flush work if it is scheduled for
 261  * earlier. Thus we use queue_delayed_work().
 262  */
 263 void wb_wakeup_delayed(struct bdi_writeback *wb)
 264 {
 265         unsigned long timeout;
 266
 267         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 268         spin_lock_bh(&wb->work_lock);
 269         if (test_bit(WB_registered, &wb->state))
 270                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 271         spin_unlock_bh(&wb->work_lock);
 272 }
 273
 274 static void wb_update_bandwidth_workfn(struct work_struct *work)
 275 {
 276         struct bdi_writeback *wb = container_of(to_delayed_work(work),
 277                                                 struct bdi_writeback, bw_dwork);
 278
 279         wb_update_bandwidth(wb);
 280 }
 281
 282 /*
 283  * Initial write bandwidth: 100 MB/s
 284  */
 285 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 286
 287 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 288                    gfp_t gfp)
 289 {
 290         int i, err;
 291
 292         memset(wb, 0, sizeof(*wb));
 293
 294         wb->bdi = bdi;
 295         wb->last_old_flush = jiffies;
 296         INIT_LIST_HEAD(&wb->b_dirty);
 297         INIT_LIST_HEAD(&wb->b_io);
 298         INIT_LIST_HEAD(&wb->b_more_io);
 299         INIT_LIST_HEAD(&wb->b_dirty_time);
 300         spin_lock_init(&wb->list_lock);
 301
 302         atomic_set(&wb->writeback_inodes, 0);
 303         wb->bw_time_stamp = jiffies;
 304         wb->balanced_dirty_ratelimit = INIT_BW;
 305         wb->dirty_ratelimit = INIT_BW;
 306         wb->write_bandwidth = INIT_BW;
 307         wb->avg_write_bandwidth = INIT_BW;
 308
 309         spin_lock_init(&wb->work_lock);
 310         INIT_LIST_HEAD(&wb->work_list);
 311         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 312         INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
 313         wb->dirty_sleep = jiffies;
 314
 315         err = fprop_local_init_percpu(&wb->completions, gfp);
 316         if (err)
 317                 return err;
 318
 319         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 320                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 321                 if (err)
 322                         goto out_destroy_stat;
 323         }
 324
 325         return 0;
 326
 327 out_destroy_stat:
 328         while (i--)
 329                 percpu_counter_destroy(&wb->stat[i]);
 330         fprop_local_destroy_percpu(&wb->completions);
 331         return err;
 332 }
 333
 334 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 335
 336 /*
 337  * Remove bdi from the global list and shutdown any threads we have running
 338  */
 339 static void wb_shutdown(struct bdi_writeback *wb)
 340 {
 341         /* Make sure nobody queues further work */
 342         spin_lock_bh(&wb->work_lock);
 343         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 344                 spin_unlock_bh(&wb->work_lock);
 345                 return;
 346         }
 347         spin_unlock_bh(&wb->work_lock);
 348
 349         cgwb_remove_from_bdi_list(wb);
 350         /*
 351          * Drain work list and shutdown the delayed_work.  !WB_registered
 352          * tells wb_workfn() that @wb is dying and its work_list needs to
 353          * be drained no matter what.
 354          */
 355         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 356         flush_delayed_work(&wb->dwork);
 357         WARN_ON(!list_empty(&wb->work_list));
 358         flush_delayed_work(&wb->bw_dwork);
 359 }
 360
 361 static void wb_exit(struct bdi_writeback *wb)
 362 {
 363         int i;
 364
 365         WARN_ON(delayed_work_pending(&wb->dwork));
 366
 367         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 368                 percpu_counter_destroy(&wb->stat[i]);
 369
 370         fprop_local_destroy_percpu(&wb->completions);
 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_release_workfn(struct work_struct *work)
 389 {
 390         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 391                                                 release_work);
 392         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 393         struct backing_dev_info *bdi = wb->bdi;
 394
 395         mutex_lock(&wb->bdi->cgwb_release_mutex);
 396         wb_shutdown(wb);
 397
 398         css_put(wb->memcg_css);
 399         css_put(wb->blkcg_css);
 400         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 401
 402         /* triggers blkg destruction if no online users left */
 403         blkcg_unpin_online(blkcg);
 404
 405         fprop_local_destroy_percpu(&wb->memcg_completions);
 406
 407         spin_lock_irq(&cgwb_lock);
 408         list_del(&wb->offline_node);
 409         spin_unlock_irq(&cgwb_lock);
 410
 411         percpu_ref_exit(&wb->refcnt);
 412         wb_exit(wb);
 413         bdi_put(bdi);
 414         WARN_ON_ONCE(!list_empty(&wb->b_attached));
 415         kfree_rcu(wb, rcu);
 416 }
 417
 418 static void cgwb_release(struct percpu_ref *refcnt)
 419 {
 420         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 421                                                 refcnt);
 422         queue_work(cgwb_release_wq, &wb->release_work);
 423 }
 424
 425 static void cgwb_kill(struct bdi_writeback *wb)
 426 {
 427         lockdep_assert_held(&cgwb_lock);
 428
 429         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 430         list_del(&wb->memcg_node);
 431         list_del(&wb->blkcg_node);
 432         list_add(&wb->offline_node, &offline_cgwbs);
 433         percpu_ref_kill(&wb->refcnt);
 434 }
 435
 436 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 437 {
 438         spin_lock_irq(&cgwb_lock);
 439         list_del_rcu(&wb->bdi_node);
 440         spin_unlock_irq(&cgwb_lock);
 441 }
 442
 443 static int cgwb_create(struct backing_dev_info *bdi,
 444                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 445 {
 446         struct mem_cgroup *memcg;
 447         struct cgroup_subsys_state *blkcg_css;
 448         struct blkcg *blkcg;
 449         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 450         struct bdi_writeback *wb;
 451         unsigned long flags;
 452         int ret = 0;
 453
 454         memcg = mem_cgroup_from_css(memcg_css);
 455         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 456         blkcg = css_to_blkcg(blkcg_css);
 457         memcg_cgwb_list = &memcg->cgwb_list;
 458         blkcg_cgwb_list = &blkcg->cgwb_list;
 459
 460         /* look up again under lock and discard on blkcg mismatch */
 461         spin_lock_irqsave(&cgwb_lock, flags);
 462         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 463         if (wb && wb->blkcg_css != blkcg_css) {
 464                 cgwb_kill(wb);
 465                 wb = NULL;
 466         }
 467         spin_unlock_irqrestore(&cgwb_lock, flags);
 468         if (wb)
 469                 goto out_put;
 470
 471         /* need to create a new one */
 472         wb = kmalloc(sizeof(*wb), gfp);
 473         if (!wb) {
 474                 ret = -ENOMEM;
 475                 goto out_put;
 476         }
 477
 478         ret = wb_init(wb, bdi, gfp);
 479         if (ret)
 480                 goto err_free;
 481
 482         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 483         if (ret)
 484                 goto err_wb_exit;
 485
 486         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 487         if (ret)
 488                 goto err_ref_exit;
 489
 490         wb->memcg_css = memcg_css;
 491         wb->blkcg_css = blkcg_css;
 492         INIT_LIST_HEAD(&wb->b_attached);
 493         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 494         set_bit(WB_registered, &wb->state);
 495         bdi_get(bdi);
 496
 497         /*
 498          * The root wb determines the registered state of the whole bdi and
 499          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 500          * whether they're still online.  Don't link @wb if any is dead.
 501          * See wb_memcg_offline() and wb_blkcg_offline().
 502          */
 503         ret = -ENODEV;
 504         spin_lock_irqsave(&cgwb_lock, flags);
 505         if (test_bit(WB_registered, &bdi->wb.state) &&
 506             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 507                 /* we might have raced another instance of this function */
 508                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 509                 if (!ret) {
 510                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 511                         list_add(&wb->memcg_node, memcg_cgwb_list);
 512                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 513                         blkcg_pin_online(blkcg);
 514                         css_get(memcg_css);
 515                         css_get(blkcg_css);
 516                 }
 517         }
 518         spin_unlock_irqrestore(&cgwb_lock, flags);
 519         if (ret) {
 520                 if (ret == -EEXIST)
 521                         ret = 0;
 522                 goto err_fprop_exit;
 523         }
 524         goto out_put;
 525
 526 err_fprop_exit:
 527         bdi_put(bdi);
 528         fprop_local_destroy_percpu(&wb->memcg_completions);
 529 err_ref_exit:
 530         percpu_ref_exit(&wb->refcnt);
 531 err_wb_exit:
 532         wb_exit(wb);
 533 err_free:
 534         kfree(wb);
 535 out_put:
 536         css_put(blkcg_css);
 537         return ret;
 538 }
 539
 540 /**
 541  * wb_get_lookup - get wb for a given memcg
 542  * @bdi: target bdi
 543  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 544  *
 545  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 546  * refcount incremented.
 547  *
 548  * This function uses css_get() on @memcg_css and thus expects its refcnt
 549  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 550  * @memcg_css isn't enough.  try_get it before calling this function.
 551  *
 552  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 553  * memcg on the default hierarchy, memcg association is guaranteed to be
 554  * more specific (equal or descendant to the associated blkcg) and thus can
 555  * identify both the memcg and blkcg associations.
 556  *
 557  * Because the blkcg associated with a memcg may change as blkcg is enabled
 558  * and disabled closer to root in the hierarchy, each wb keeps track of
 559  * both the memcg and blkcg associated with it and verifies the blkcg on
 560  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 561  * created.
 562  */
 563 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 564                                     struct cgroup_subsys_state *memcg_css)
 565 {
 566         struct bdi_writeback *wb;
 567
 568         if (!memcg_css->parent)
 569                 return &bdi->wb;
 570
 571         rcu_read_lock();
 572         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 573         if (wb) {
 574                 struct cgroup_subsys_state *blkcg_css;
 575
 576                 /* see whether the blkcg association has changed */
 577                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 578                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 579                         wb = NULL;
 580                 css_put(blkcg_css);
 581         }
 582         rcu_read_unlock();
 583
 584         return wb;
 585 }
 586
 587 /**
 588  * wb_get_create - get wb for a given memcg, create if necessary
 589  * @bdi: target bdi
 590  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 591  * @gfp: allocation mask to use
 592  *
 593  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 594  * create one.  See wb_get_lookup() for more details.
 595  */
 596 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 597                                     struct cgroup_subsys_state *memcg_css,
 598                                     gfp_t gfp)
 599 {
 600         struct bdi_writeback *wb;
 601
 602         might_alloc(gfp);
 603
 604         if (!memcg_css->parent)
 605                 return &bdi->wb;
 606
 607         do {
 608                 wb = wb_get_lookup(bdi, memcg_css);
 609         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 610
 611         return wb;
 612 }
 613
 614 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 615 {
 616         int ret;
 617
 618         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 619         mutex_init(&bdi->cgwb_release_mutex);
 620         init_rwsem(&bdi->wb_switch_rwsem);
 621
 622         ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 623         if (!ret) {
 624                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 625                 bdi->wb.blkcg_css = blkcg_root_css;
 626         }
 627         return ret;
 628 }
 629
 630 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 631 {
 632         struct radix_tree_iter iter;
 633         void **slot;
 634         struct bdi_writeback *wb;
 635
 636         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 637
 638         spin_lock_irq(&cgwb_lock);
 639         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 640                 cgwb_kill(*slot);
 641         spin_unlock_irq(&cgwb_lock);
 642
 643         mutex_lock(&bdi->cgwb_release_mutex);
 644         spin_lock_irq(&cgwb_lock);
 645         while (!list_empty(&bdi->wb_list)) {
 646                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 647                                       bdi_node);
 648                 spin_unlock_irq(&cgwb_lock);
 649                 wb_shutdown(wb);
 650                 spin_lock_irq(&cgwb_lock);
 651         }
 652         spin_unlock_irq(&cgwb_lock);
 653         mutex_unlock(&bdi->cgwb_release_mutex);
 654 }
 655
 656 /*
 657  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
 658  *
 659  * Try to release dying cgwbs by switching attached inodes to the nearest
 660  * living ancestor's writeback. Processed wbs are placed at the end
 661  * of the list to guarantee the forward progress.
 662  */
 663 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
 664 {
 665         struct bdi_writeback *wb;
 666         LIST_HEAD(processed);
 667
 668         spin_lock_irq(&cgwb_lock);
 669
 670         while (!list_empty(&offline_cgwbs)) {
 671                 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
 672                                       offline_node);
 673                 list_move(&wb->offline_node, &processed);
 674
 675                 /*
 676                  * If wb is dirty, cleaning up the writeback by switching
 677                  * attached inodes will result in an effective removal of any
 678                  * bandwidth restrictions, which isn't the goal.  Instead,
 679                  * it can be postponed until the next time, when all io
 680                  * will be likely completed.  If in the meantime some inodes
 681                  * will get re-dirtied, they should be eventually switched to
 682                  * a new cgwb.
 683                  */
 684                 if (wb_has_dirty_io(wb))
 685                         continue;
 686
 687                 if (!wb_tryget(wb))
 688                         continue;
 689
 690                 spin_unlock_irq(&cgwb_lock);
 691                 while (cleanup_offline_cgwb(wb))
 692                         cond_resched();
 693                 spin_lock_irq(&cgwb_lock);
 694
 695                 wb_put(wb);
 696         }
 697
 698         if (!list_empty(&processed))
 699                 list_splice_tail(&processed, &offline_cgwbs);
 700
 701         spin_unlock_irq(&cgwb_lock);
 702 }
 703
 704 /**
 705  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 706  * @memcg: memcg being offlined
 707  *
 708  * Also prevents creation of any new wb's associated with @memcg.
 709  */
 710 void wb_memcg_offline(struct mem_cgroup *memcg)
 711 {
 712         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 713         struct bdi_writeback *wb, *next;
 714
 715         spin_lock_irq(&cgwb_lock);
 716         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 717                 cgwb_kill(wb);
 718         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 719         spin_unlock_irq(&cgwb_lock);
 720
 721         queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
 722 }
 723
 724 /**
 725  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 726  * @blkcg: blkcg being offlined
 727  *
 728  * Also prevents creation of any new wb's associated with @blkcg.
 729  */
 730 void wb_blkcg_offline(struct blkcg *blkcg)
 731 {
 732         struct bdi_writeback *wb, *next;
 733
 734         spin_lock_irq(&cgwb_lock);
 735         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 736                 cgwb_kill(wb);
 737         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 738         spin_unlock_irq(&cgwb_lock);
 739 }
 740
 741 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 742 {
 743         spin_lock_irq(&cgwb_lock);
 744         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 745         spin_unlock_irq(&cgwb_lock);
 746 }
 747
 748 static int __init cgwb_init(void)
 749 {
 750         /*
 751          * There can be many concurrent release work items overwhelming
 752          * system_wq.  Put them in a separate wq and limit concurrency.
 753          * There's no point in executing many of these in parallel.
 754          */
 755         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 756         if (!cgwb_release_wq)
 757                 return -ENOMEM;
 758
 759         return 0;
 760 }
 761 subsys_initcall(cgwb_init);
 762
 763 #else   /* CONFIG_CGROUP_WRITEBACK */
 764
 765 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 766 {
 767         return wb_init(&bdi->wb, bdi, GFP_KERNEL);
 768 }
 769
 770 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 771
 772 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 773 {
 774         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 775 }
 776
 777 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 778 {
 779         list_del_rcu(&wb->bdi_node);
 780 }
 781
 782 #endif  /* CONFIG_CGROUP_WRITEBACK */
 783
 784 static int bdi_init(struct backing_dev_info *bdi)
 785 {
 786         int ret;
 787
 788         bdi->dev = NULL;
 789
 790         kref_init(&bdi->refcnt);
 791         bdi->min_ratio = 0;
 792         bdi->max_ratio = 100;
 793         bdi->max_prop_frac = FPROP_FRAC_BASE;
 794         INIT_LIST_HEAD(&bdi->bdi_list);
 795         INIT_LIST_HEAD(&bdi->wb_list);
 796         init_waitqueue_head(&bdi->wb_waitq);
 797
 798         ret = cgwb_bdi_init(bdi);
 799
 800         return ret;
 801 }
 802
 803 struct backing_dev_info *bdi_alloc(int node_id)
 804 {
 805         struct backing_dev_info *bdi;
 806
 807         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 808         if (!bdi)
 809                 return NULL;
 810
 811         if (bdi_init(bdi)) {
 812                 kfree(bdi);
 813                 return NULL;
 814         }
 815         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 816         bdi->ra_pages = VM_READAHEAD_PAGES;
 817         bdi->io_pages = VM_READAHEAD_PAGES;
 818         timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
 819         return bdi;
 820 }
 821 EXPORT_SYMBOL(bdi_alloc);
 822
 823 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 824 {
 825         struct rb_node **p = &bdi_tree.rb_node;
 826         struct rb_node *parent = NULL;
 827         struct backing_dev_info *bdi;
 828
 829         lockdep_assert_held(&bdi_lock);
 830
 831         while (*p) {
 832                 parent = *p;
 833                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 834
 835                 if (bdi->id > id)
 836                         p = &(*p)->rb_left;
 837                 else if (bdi->id < id)
 838                         p = &(*p)->rb_right;
 839                 else
 840                         break;
 841         }
 842
 843         if (parentp)
 844                 *parentp = parent;
 845         return p;
 846 }
 847
 848 /**
 849  * bdi_get_by_id - lookup and get bdi from its id
 850  * @id: bdi id to lookup
 851  *
 852  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 853  * doesn't exist or is already unregistered.
 854  */
 855 struct backing_dev_info *bdi_get_by_id(u64 id)
 856 {
 857         struct backing_dev_info *bdi = NULL;
 858         struct rb_node **p;
 859
 860         spin_lock_bh(&bdi_lock);
 861         p = bdi_lookup_rb_node(id, NULL);
 862         if (*p) {
 863                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 864                 bdi_get(bdi);
 865         }
 866         spin_unlock_bh(&bdi_lock);
 867
 868         return bdi;
 869 }
 870
 871 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 872 {
 873         struct device *dev;
 874         struct rb_node *parent, **p;
 875
 876         if (bdi->dev)   /* The driver needs to use separate queues per device */
 877                 return 0;
 878
 879         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
 880         dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
 881         if (IS_ERR(dev))
 882                 return PTR_ERR(dev);
 883
 884         cgwb_bdi_register(bdi);
 885         bdi->dev = dev;
 886
 887         bdi_debug_register(bdi, dev_name(dev));
 888         set_bit(WB_registered, &bdi->wb.state);
 889
 890         spin_lock_bh(&bdi_lock);
 891
 892         bdi->id = ++bdi_id_cursor;
 893
 894         p = bdi_lookup_rb_node(bdi->id, &parent);
 895         rb_link_node(&bdi->rb_node, parent, p);
 896         rb_insert_color(&bdi->rb_node, &bdi_tree);
 897
 898         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 899
 900         spin_unlock_bh(&bdi_lock);
 901
 902         trace_writeback_bdi_register(bdi);
 903         return 0;
 904 }
 905
 906 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 907 {
 908         va_list args;
 909         int ret;
 910
 911         va_start(args, fmt);
 912         ret = bdi_register_va(bdi, fmt, args);
 913         va_end(args);
 914         return ret;
 915 }
 916 EXPORT_SYMBOL(bdi_register);
 917
 918 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
 919 {
 920         WARN_ON_ONCE(bdi->owner);
 921         bdi->owner = owner;
 922         get_device(owner);
 923 }
 924
 925 /*
 926  * Remove bdi from bdi_list, and ensure that it is no longer visible
 927  */
 928 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 929 {
 930         spin_lock_bh(&bdi_lock);
 931         rb_erase(&bdi->rb_node, &bdi_tree);
 932         list_del_rcu(&bdi->bdi_list);
 933         spin_unlock_bh(&bdi_lock);
 934
 935         synchronize_rcu_expedited();
 936 }
 937
 938 void bdi_unregister(struct backing_dev_info *bdi)
 939 {
 940         del_timer_sync(&bdi->laptop_mode_wb_timer);
 941
 942         /* make sure nobody finds us on the bdi_list anymore */
 943         bdi_remove_from_list(bdi);
 944         wb_shutdown(&bdi->wb);
 945         cgwb_bdi_unregister(bdi);
 946
 947         if (bdi->dev) {
 948                 bdi_debug_unregister(bdi);
 949                 device_unregister(bdi->dev);
 950                 bdi->dev = NULL;
 951         }
 952
 953         if (bdi->owner) {
 954                 put_device(bdi->owner);
 955                 bdi->owner = NULL;
 956         }
 957 }
 958 EXPORT_SYMBOL(bdi_unregister);
 959
 960 static void release_bdi(struct kref *ref)
 961 {
 962         struct backing_dev_info *bdi =
 963                         container_of(ref, struct backing_dev_info, refcnt);
 964
 965         WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
 966         WARN_ON_ONCE(bdi->dev);
 967         wb_exit(&bdi->wb);
 968         kfree(bdi);
 969 }
 970
 971 void bdi_put(struct backing_dev_info *bdi)
 972 {
 973         kref_put(&bdi->refcnt, release_bdi);
 974 }
 975 EXPORT_SYMBOL(bdi_put);
 976
 977 const char *bdi_dev_name(struct backing_dev_info *bdi)
 978 {
 979         if (!bdi || !bdi->dev)
 980                 return bdi_unknown_name;
 981         return bdi->dev_name;
 982 }
 983 EXPORT_SYMBOL_GPL(bdi_dev_name);
 984
 985 static wait_queue_head_t congestion_wqh[2] = {
 986                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 987                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 988         };
 989 static atomic_t nr_wb_congested[2];
 990
 991 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 992 {
 993         wait_queue_head_t *wqh = &congestion_wqh[sync];
 994         enum wb_congested_state bit;
 995
 996         bit = sync ? WB_sync_congested : WB_async_congested;
 997         if (test_and_clear_bit(bit, &bdi->wb.congested))
 998                 atomic_dec(&nr_wb_congested[sync]);
 999         smp_mb__after_atomic();
1000         if (waitqueue_active(wqh))
1001                 wake_up(wqh);
1002 }
1003 EXPORT_SYMBOL(clear_bdi_congested);
1004
1005 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
1006 {
1007         enum wb_congested_state bit;
1008
1009         bit = sync ? WB_sync_congested : WB_async_congested;
1010         if (!test_and_set_bit(bit, &bdi->wb.congested))
1011                 atomic_inc(&nr_wb_congested[sync]);
1012 }
1013 EXPORT_SYMBOL(set_bdi_congested);
1014
1015 /**
1016  * congestion_wait - wait for a backing_dev to become uncongested
1017  * @sync: SYNC or ASYNC IO
1018  * @timeout: timeout in jiffies
1019  *
1020  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1021  * write congestion.  If no backing_devs are congested then just wait for the
1022  * next write to be completed.
1023  */
1024 long congestion_wait(int sync, long timeout)
1025 {
1026         long ret;
1027         unsigned long start = jiffies;
1028         DEFINE_WAIT(wait);
1029         wait_queue_head_t *wqh = &congestion_wqh[sync];
1030
1031         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1032         ret = io_schedule_timeout(timeout);
1033         finish_wait(wqh, &wait);
1034
1035         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1036                                         jiffies_to_usecs(jiffies - start));
1037
1038         return ret;
1039 }
1040 EXPORT_SYMBOL(congestion_wait);