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