mm/backing-dev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include <linux/blkdev.h>
   4 #include <linux/wait.h>
   5 #include <linux/rbtree.h>
   6 #include <linux/kthread.h>
   7 #include <linux/backing-dev.h>
   8 #include <linux/blk-cgroup.h>
   9 #include <linux/freezer.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/mm.h>
  13 #include <linux/sched/mm.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/writeback.h>
  17 #include <linux/device.h>
  18 #include <trace/events/writeback.h>
  19
  20 struct backing_dev_info noop_backing_dev_info;
  21 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  22
  23 static struct class *bdi_class;
  24 static const char *bdi_unknown_name = "(unknown)";
  25
  26 /*
  27  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
  28  * reader side locking.
  29  */
  30 DEFINE_SPINLOCK(bdi_lock);
  31 static u64 bdi_id_cursor;
  32 static struct rb_root bdi_tree = RB_ROOT;
  33 LIST_HEAD(bdi_list);
  34
  35 /* bdi_wq serves all asynchronous writeback tasks */
  36 struct workqueue_struct *bdi_wq;
  37
  38 #define K(x) ((x) << (PAGE_SHIFT - 10))
  39
  40 #ifdef CONFIG_DEBUG_FS
  41 #include <linux/debugfs.h>
  42 #include <linux/seq_file.h>
  43
  44 static struct dentry *bdi_debug_root;
  45
  46 static void bdi_debug_init(void)
  47 {
  48         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  49 }
  50
  51 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  52 {
  53         struct backing_dev_info *bdi = m->private;
  54         struct bdi_writeback *wb = &bdi->wb;
  55         unsigned long background_thresh;
  56         unsigned long dirty_thresh;
  57         unsigned long wb_thresh;
  58         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  59         struct inode *inode;
  60
  61         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  62         spin_lock(&wb->list_lock);
  63         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  64                 nr_dirty++;
  65         list_for_each_entry(inode, &wb->b_io, i_io_list)
  66                 nr_io++;
  67         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  68                 nr_more_io++;
  69         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  70                 if (inode->i_state & I_DIRTY_TIME)
  71                         nr_dirty_time++;
  72         spin_unlock(&wb->list_lock);
  73
  74         global_dirty_limits(&background_thresh, &dirty_thresh);
  75         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  76
  77         seq_printf(m,
  78                    "BdiWriteback:       %10lu kB\n"
  79                    "BdiReclaimable:     %10lu kB\n"
  80                    "BdiDirtyThresh:     %10lu kB\n"
  81                    "DirtyThresh:        %10lu kB\n"
  82                    "BackgroundThresh:   %10lu kB\n"
  83                    "BdiDirtied:         %10lu kB\n"
  84                    "BdiWritten:         %10lu kB\n"
  85                    "BdiWriteBandwidth:  %10lu kBps\n"
  86                    "b_dirty:            %10lu\n"
  87                    "b_io:               %10lu\n"
  88                    "b_more_io:          %10lu\n"
  89                    "b_dirty_time:       %10lu\n"
  90                    "bdi_list:           %10u\n"
  91                    "state:              %10lx\n",
  92                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  93                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  94                    K(wb_thresh),
  95                    K(dirty_thresh),
  96                    K(background_thresh),
  97                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  98                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  99                    (unsigned long) K(wb->write_bandwidth),
 100                    nr_dirty,
 101                    nr_io,
 102                    nr_more_io,
 103                    nr_dirty_time,
 104                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 105
 106         return 0;
 107 }
 108 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 109
 110 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 111 {
 112         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 113
 114         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 115                             &bdi_debug_stats_fops);
 116 }
 117
 118 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 119 {
 120         debugfs_remove_recursive(bdi->debug_dir);
 121 }
 122 #else
 123 static inline void bdi_debug_init(void)
 124 {
 125 }
 126 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 127                                       const char *name)
 128 {
 129 }
 130 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 131 {
 132 }
 133 #endif
 134
 135 static ssize_t read_ahead_kb_store(struct device *dev,
 136                                   struct device_attribute *attr,
 137                                   const char *buf, size_t count)
 138 {
 139         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 140         unsigned long read_ahead_kb;
 141         ssize_t ret;
 142
 143         ret = kstrtoul(buf, 10, &read_ahead_kb);
 144         if (ret < 0)
 145                 return ret;
 146
 147         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 148
 149         return count;
 150 }
 151
 152 #define BDI_SHOW(name, expr)                                            \
 153 static ssize_t name##_show(struct device *dev,                          \
 154                            struct device_attribute *attr, char *buf)    \
 155 {                                                                       \
 156         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 157                                                                         \
 158         return sysfs_emit(buf, "%lld\n", (long long)expr);              \
 159 }                                                                       \
 160 static DEVICE_ATTR_RW(name);
 161
 162 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 163
 164 static ssize_t min_ratio_store(struct device *dev,
 165                 struct device_attribute *attr, const char *buf, size_t count)
 166 {
 167         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 168         unsigned int ratio;
 169         ssize_t ret;
 170
 171         ret = kstrtouint(buf, 10, &ratio);
 172         if (ret < 0)
 173                 return ret;
 174
 175         ret = bdi_set_min_ratio(bdi, ratio);
 176         if (!ret)
 177                 ret = count;
 178
 179         return ret;
 180 }
 181 BDI_SHOW(min_ratio, bdi->min_ratio)
 182
 183 static ssize_t max_ratio_store(struct device *dev,
 184                 struct device_attribute *attr, const char *buf, size_t count)
 185 {
 186         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 187         unsigned int ratio;
 188         ssize_t ret;
 189
 190         ret = kstrtouint(buf, 10, &ratio);
 191         if (ret < 0)
 192                 return ret;
 193
 194         ret = bdi_set_max_ratio(bdi, ratio);
 195         if (!ret)
 196                 ret = count;
 197
 198         return ret;
 199 }
 200 BDI_SHOW(max_ratio, bdi->max_ratio)
 201
 202 static ssize_t stable_pages_required_show(struct device *dev,
 203                                           struct device_attribute *attr,
 204                                           char *buf)
 205 {
 206         dev_warn_once(dev,
 207                 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
 208         return sysfs_emit(buf, "%d\n", 0);
 209 }
 210 static DEVICE_ATTR_RO(stable_pages_required);
 211
 212 static struct attribute *bdi_dev_attrs[] = {
 213         &dev_attr_read_ahead_kb.attr,
 214         &dev_attr_min_ratio.attr,
 215         &dev_attr_max_ratio.attr,
 216         &dev_attr_stable_pages_required.attr,
 217         NULL,
 218 };
 219 ATTRIBUTE_GROUPS(bdi_dev);
 220
 221 static __init int bdi_class_init(void)
 222 {
 223         bdi_class = class_create(THIS_MODULE, "bdi");
 224         if (IS_ERR(bdi_class))
 225                 return PTR_ERR(bdi_class);
 226
 227         bdi_class->dev_groups = bdi_dev_groups;
 228         bdi_debug_init();
 229
 230         return 0;
 231 }
 232 postcore_initcall(bdi_class_init);
 233
 234 static int __init default_bdi_init(void)
 235 {
 236         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 237                                  WQ_SYSFS, 0);
 238         if (!bdi_wq)
 239                 return -ENOMEM;
 240         return 0;
 241 }
 242 subsys_initcall(default_bdi_init);
 243
 244 /*
 245  * This function is used when the first inode for this wb is marked dirty. It
 246  * wakes-up the corresponding bdi thread which should then take care of the
 247  * periodic background write-out of dirty inodes. Since the write-out would
 248  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 249  * set up a timer which wakes the bdi thread up later.
 250  *
 251  * Note, we wouldn't bother setting up the timer, but this function is on the
 252  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 253  * by delaying the wake-up.
 254  *
 255  * We have to be careful not to postpone flush work if it is scheduled for
 256  * earlier. Thus we use queue_delayed_work().
 257  */
 258 void wb_wakeup_delayed(struct bdi_writeback *wb)
 259 {
 260         unsigned long timeout;
 261
 262         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 263         spin_lock_irq(&wb->work_lock);
 264         if (test_bit(WB_registered, &wb->state))
 265                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 266         spin_unlock_irq(&wb->work_lock);
 267 }
 268
 269 static void wb_update_bandwidth_workfn(struct work_struct *work)
 270 {
 271         struct bdi_writeback *wb = container_of(to_delayed_work(work),
 272                                                 struct bdi_writeback, bw_dwork);
 273
 274         wb_update_bandwidth(wb);
 275 }
 276
 277 /*
 278  * Initial write bandwidth: 100 MB/s
 279  */
 280 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 281
 282 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 283                    gfp_t gfp)
 284 {
 285         int i, err;
 286
 287         memset(wb, 0, sizeof(*wb));
 288
 289         wb->bdi = bdi;
 290         wb->last_old_flush = jiffies;
 291         INIT_LIST_HEAD(&wb->b_dirty);
 292         INIT_LIST_HEAD(&wb->b_io);
 293         INIT_LIST_HEAD(&wb->b_more_io);
 294         INIT_LIST_HEAD(&wb->b_dirty_time);
 295         spin_lock_init(&wb->list_lock);
 296
 297         atomic_set(&wb->writeback_inodes, 0);
 298         wb->bw_time_stamp = jiffies;
 299         wb->balanced_dirty_ratelimit = INIT_BW;
 300         wb->dirty_ratelimit = INIT_BW;
 301         wb->write_bandwidth = INIT_BW;
 302         wb->avg_write_bandwidth = INIT_BW;
 303
 304         spin_lock_init(&wb->work_lock);
 305         INIT_LIST_HEAD(&wb->work_list);
 306         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 307         INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
 308         wb->dirty_sleep = jiffies;
 309
 310         err = fprop_local_init_percpu(&wb->completions, gfp);
 311         if (err)
 312                 return err;
 313
 314         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 315                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 316                 if (err)
 317                         goto out_destroy_stat;
 318         }
 319
 320         return 0;
 321
 322 out_destroy_stat:
 323         while (i--)
 324                 percpu_counter_destroy(&wb->stat[i]);
 325         fprop_local_destroy_percpu(&wb->completions);
 326         return err;
 327 }
 328
 329 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 330
 331 /*
 332  * Remove bdi from the global list and shutdown any threads we have running
 333  */
 334 static void wb_shutdown(struct bdi_writeback *wb)
 335 {
 336         /* Make sure nobody queues further work */
 337         spin_lock_irq(&wb->work_lock);
 338         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 339                 spin_unlock_irq(&wb->work_lock);
 340                 return;
 341         }
 342         spin_unlock_irq(&wb->work_lock);
 343
 344         cgwb_remove_from_bdi_list(wb);
 345         /*
 346          * Drain work list and shutdown the delayed_work.  !WB_registered
 347          * tells wb_workfn() that @wb is dying and its work_list needs to
 348          * be drained no matter what.
 349          */
 350         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 351         flush_delayed_work(&wb->dwork);
 352         WARN_ON(!list_empty(&wb->work_list));
 353         flush_delayed_work(&wb->bw_dwork);
 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 }
 367
 368 #ifdef CONFIG_CGROUP_WRITEBACK
 369
 370 #include <linux/memcontrol.h>
 371
 372 /*
 373  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
 374  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
 375  */
 376 static DEFINE_SPINLOCK(cgwb_lock);
 377 static struct workqueue_struct *cgwb_release_wq;
 378
 379 static LIST_HEAD(offline_cgwbs);
 380 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
 381 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
 382
 383 static void cgwb_free_rcu(struct rcu_head *rcu_head)
 384 {
 385         struct bdi_writeback *wb = container_of(rcu_head,
 386                         struct bdi_writeback, rcu);
 387
 388         percpu_ref_exit(&wb->refcnt);
 389         kfree(wb);
 390 }
 391
 392 static void cgwb_release_workfn(struct work_struct *work)
 393 {
 394         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 395                                                 release_work);
 396         struct backing_dev_info *bdi = wb->bdi;
 397
 398         mutex_lock(&wb->bdi->cgwb_release_mutex);
 399         wb_shutdown(wb);
 400
 401         css_put(wb->memcg_css);
 402         css_put(wb->blkcg_css);
 403         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 404
 405         /* triggers blkg destruction if no online users left */
 406         blkcg_unpin_online(wb->blkcg_css);
 407
 408         fprop_local_destroy_percpu(&wb->memcg_completions);
 409
 410         spin_lock_irq(&cgwb_lock);
 411         list_del(&wb->offline_node);
 412         spin_unlock_irq(&cgwb_lock);
 413
 414         wb_exit(wb);
 415         bdi_put(bdi);
 416         WARN_ON_ONCE(!list_empty(&wb->b_attached));
 417         call_rcu(&wb->rcu, cgwb_free_rcu);
 418 }
 419
 420 static void cgwb_release(struct percpu_ref *refcnt)
 421 {
 422         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 423                                                 refcnt);
 424         queue_work(cgwb_release_wq, &wb->release_work);
 425 }
 426
 427 static void cgwb_kill(struct bdi_writeback *wb)
 428 {
 429         lockdep_assert_held(&cgwb_lock);
 430
 431         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 432         list_del(&wb->memcg_node);
 433         list_del(&wb->blkcg_node);
 434         list_add(&wb->offline_node, &offline_cgwbs);
 435         percpu_ref_kill(&wb->refcnt);
 436 }
 437
 438 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 439 {
 440         spin_lock_irq(&cgwb_lock);
 441         list_del_rcu(&wb->bdi_node);
 442         spin_unlock_irq(&cgwb_lock);
 443 }
 444
 445 static int cgwb_create(struct backing_dev_info *bdi,
 446                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 447 {
 448         struct mem_cgroup *memcg;
 449         struct cgroup_subsys_state *blkcg_css;
 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         memcg_cgwb_list = &memcg->cgwb_list;
 458         blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
 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_css);
 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  * @css: blkcg being offlined
 727  *
 728  * Also prevents creation of any new wb's associated with @blkcg.
 729  */
 730 void wb_blkcg_offline(struct cgroup_subsys_state *css)
 731 {
 732         struct bdi_writeback *wb, *next;
 733         struct list_head *list = blkcg_get_cgwb_list(css);
 734
 735         spin_lock_irq(&cgwb_lock);
 736         list_for_each_entry_safe(wb, next, list, blkcg_node)
 737                 cgwb_kill(wb);
 738         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 int bdi_init(struct backing_dev_info *bdi)
 786 {
 787         bdi->dev = NULL;
 788
 789         kref_init(&bdi->refcnt);
 790         bdi->min_ratio = 0;
 791         bdi->max_ratio = 100;
 792         bdi->max_prop_frac = FPROP_FRAC_BASE;
 793         INIT_LIST_HEAD(&bdi->bdi_list);
 794         INIT_LIST_HEAD(&bdi->wb_list);
 795         init_waitqueue_head(&bdi->wb_waitq);
 796
 797         return cgwb_bdi_init(bdi);
 798 }
 799
 800 struct backing_dev_info *bdi_alloc(int node_id)
 801 {
 802         struct backing_dev_info *bdi;
 803
 804         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 805         if (!bdi)
 806                 return NULL;
 807
 808         if (bdi_init(bdi)) {
 809                 kfree(bdi);
 810                 return NULL;
 811         }
 812         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 813         bdi->ra_pages = VM_READAHEAD_PAGES;
 814         bdi->io_pages = VM_READAHEAD_PAGES;
 815         timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
 816         return bdi;
 817 }
 818 EXPORT_SYMBOL(bdi_alloc);
 819
 820 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 821 {
 822         struct rb_node **p = &bdi_tree.rb_node;
 823         struct rb_node *parent = NULL;
 824         struct backing_dev_info *bdi;
 825
 826         lockdep_assert_held(&bdi_lock);
 827
 828         while (*p) {
 829                 parent = *p;
 830                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 831
 832                 if (bdi->id > id)
 833                         p = &(*p)->rb_left;
 834                 else if (bdi->id < id)
 835                         p = &(*p)->rb_right;
 836                 else
 837                         break;
 838         }
 839
 840         if (parentp)
 841                 *parentp = parent;
 842         return p;
 843 }
 844
 845 /**
 846  * bdi_get_by_id - lookup and get bdi from its id
 847  * @id: bdi id to lookup
 848  *
 849  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 850  * doesn't exist or is already unregistered.
 851  */
 852 struct backing_dev_info *bdi_get_by_id(u64 id)
 853 {
 854         struct backing_dev_info *bdi = NULL;
 855         struct rb_node **p;
 856
 857         spin_lock_bh(&bdi_lock);
 858         p = bdi_lookup_rb_node(id, NULL);
 859         if (*p) {
 860                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 861                 bdi_get(bdi);
 862         }
 863         spin_unlock_bh(&bdi_lock);
 864
 865         return bdi;
 866 }
 867
 868 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 869 {
 870         struct device *dev;
 871         struct rb_node *parent, **p;
 872
 873         if (bdi->dev)   /* The driver needs to use separate queues per device */
 874                 return 0;
 875
 876         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
 877         dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
 878         if (IS_ERR(dev))
 879                 return PTR_ERR(dev);
 880
 881         cgwb_bdi_register(bdi);
 882         bdi->dev = dev;
 883
 884         bdi_debug_register(bdi, dev_name(dev));
 885         set_bit(WB_registered, &bdi->wb.state);
 886
 887         spin_lock_bh(&bdi_lock);
 888
 889         bdi->id = ++bdi_id_cursor;
 890
 891         p = bdi_lookup_rb_node(bdi->id, &parent);
 892         rb_link_node(&bdi->rb_node, parent, p);
 893         rb_insert_color(&bdi->rb_node, &bdi_tree);
 894
 895         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 896
 897         spin_unlock_bh(&bdi_lock);
 898
 899         trace_writeback_bdi_register(bdi);
 900         return 0;
 901 }
 902
 903 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 904 {
 905         va_list args;
 906         int ret;
 907
 908         va_start(args, fmt);
 909         ret = bdi_register_va(bdi, fmt, args);
 910         va_end(args);
 911         return ret;
 912 }
 913 EXPORT_SYMBOL(bdi_register);
 914
 915 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
 916 {
 917         WARN_ON_ONCE(bdi->owner);
 918         bdi->owner = owner;
 919         get_device(owner);
 920 }
 921
 922 /*
 923  * Remove bdi from bdi_list, and ensure that it is no longer visible
 924  */
 925 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 926 {
 927         spin_lock_bh(&bdi_lock);
 928         rb_erase(&bdi->rb_node, &bdi_tree);
 929         list_del_rcu(&bdi->bdi_list);
 930         spin_unlock_bh(&bdi_lock);
 931
 932         synchronize_rcu_expedited();
 933 }
 934
 935 void bdi_unregister(struct backing_dev_info *bdi)
 936 {
 937         del_timer_sync(&bdi->laptop_mode_wb_timer);
 938
 939         /* make sure nobody finds us on the bdi_list anymore */
 940         bdi_remove_from_list(bdi);
 941         wb_shutdown(&bdi->wb);
 942         cgwb_bdi_unregister(bdi);
 943
 944         /*
 945          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
 946          * update the global bdi_min_ratio.
 947          */
 948         if (bdi->min_ratio)
 949                 bdi_set_min_ratio(bdi, 0);
 950
 951         if (bdi->dev) {
 952                 bdi_debug_unregister(bdi);
 953                 device_unregister(bdi->dev);
 954                 bdi->dev = NULL;
 955         }
 956
 957         if (bdi->owner) {
 958                 put_device(bdi->owner);
 959                 bdi->owner = NULL;
 960         }
 961 }
 962 EXPORT_SYMBOL(bdi_unregister);
 963
 964 static void release_bdi(struct kref *ref)
 965 {
 966         struct backing_dev_info *bdi =
 967                         container_of(ref, struct backing_dev_info, refcnt);
 968
 969         WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
 970         WARN_ON_ONCE(bdi->dev);
 971         wb_exit(&bdi->wb);
 972         kfree(bdi);
 973 }
 974
 975 void bdi_put(struct backing_dev_info *bdi)
 976 {
 977         kref_put(&bdi->refcnt, release_bdi);
 978 }
 979 EXPORT_SYMBOL(bdi_put);
 980
 981 struct backing_dev_info *inode_to_bdi(struct inode *inode)
 982 {
 983         struct super_block *sb;
 984
 985         if (!inode)
 986                 return &noop_backing_dev_info;
 987
 988         sb = inode->i_sb;
 989 #ifdef CONFIG_BLOCK
 990         if (sb_is_blkdev_sb(sb))
 991                 return I_BDEV(inode)->bd_disk->bdi;
 992 #endif
 993         return sb->s_bdi;
 994 }
 995 EXPORT_SYMBOL(inode_to_bdi);
 996
 997 const char *bdi_dev_name(struct backing_dev_info *bdi)
 998 {
 999         if (!bdi || !bdi->dev)
1000                 return bdi_unknown_name;
1001         return bdi->dev_name;
1002 }
1003 EXPORT_SYMBOL_GPL(bdi_dev_name);