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