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 / BDI_RATIO_SCALE)
 182
 183 static ssize_t min_ratio_fine_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_min_ratio_no_scale(bdi, ratio);
 195         if (!ret)
 196                 ret = count;
 197
 198         return ret;
 199 }
 200 BDI_SHOW(min_ratio_fine, bdi->min_ratio)
 201
 202 static ssize_t max_ratio_store(struct device *dev,
 203                 struct device_attribute *attr, const char *buf, size_t count)
 204 {
 205         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 206         unsigned int ratio;
 207         ssize_t ret;
 208
 209         ret = kstrtouint(buf, 10, &ratio);
 210         if (ret < 0)
 211                 return ret;
 212
 213         ret = bdi_set_max_ratio(bdi, ratio);
 214         if (!ret)
 215                 ret = count;
 216
 217         return ret;
 218 }
 219 BDI_SHOW(max_ratio, bdi->max_ratio / BDI_RATIO_SCALE)
 220
 221 static ssize_t max_ratio_fine_store(struct device *dev,
 222                 struct device_attribute *attr, const char *buf, size_t count)
 223 {
 224         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 225         unsigned int ratio;
 226         ssize_t ret;
 227
 228         ret = kstrtouint(buf, 10, &ratio);
 229         if (ret < 0)
 230                 return ret;
 231
 232         ret = bdi_set_max_ratio_no_scale(bdi, ratio);
 233         if (!ret)
 234                 ret = count;
 235
 236         return ret;
 237 }
 238 BDI_SHOW(max_ratio_fine, bdi->max_ratio)
 239
 240 static ssize_t min_bytes_show(struct device *dev,
 241                               struct device_attribute *attr,
 242                               char *buf)
 243 {
 244         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 245
 246         return sysfs_emit(buf, "%llu\n", bdi_get_min_bytes(bdi));
 247 }
 248
 249 static ssize_t min_bytes_store(struct device *dev,
 250                 struct device_attribute *attr, const char *buf, size_t count)
 251 {
 252         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 253         u64 bytes;
 254         ssize_t ret;
 255
 256         ret = kstrtoull(buf, 10, &bytes);
 257         if (ret < 0)
 258                 return ret;
 259
 260         ret = bdi_set_min_bytes(bdi, bytes);
 261         if (!ret)
 262                 ret = count;
 263
 264         return ret;
 265 }
 266 static DEVICE_ATTR_RW(min_bytes);
 267
 268 static ssize_t max_bytes_show(struct device *dev,
 269                               struct device_attribute *attr,
 270                               char *buf)
 271 {
 272         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 273
 274         return sysfs_emit(buf, "%llu\n", bdi_get_max_bytes(bdi));
 275 }
 276
 277 static ssize_t max_bytes_store(struct device *dev,
 278                 struct device_attribute *attr, const char *buf, size_t count)
 279 {
 280         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 281         u64 bytes;
 282         ssize_t ret;
 283
 284         ret = kstrtoull(buf, 10, &bytes);
 285         if (ret < 0)
 286                 return ret;
 287
 288         ret = bdi_set_max_bytes(bdi, bytes);
 289         if (!ret)
 290                 ret = count;
 291
 292         return ret;
 293 }
 294 static DEVICE_ATTR_RW(max_bytes);
 295
 296 static ssize_t stable_pages_required_show(struct device *dev,
 297                                           struct device_attribute *attr,
 298                                           char *buf)
 299 {
 300         dev_warn_once(dev,
 301                 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
 302         return sysfs_emit(buf, "%d\n", 0);
 303 }
 304 static DEVICE_ATTR_RO(stable_pages_required);
 305
 306 static ssize_t strict_limit_store(struct device *dev,
 307                 struct device_attribute *attr, const char *buf, size_t count)
 308 {
 309         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 310         unsigned int strict_limit;
 311         ssize_t ret;
 312
 313         ret = kstrtouint(buf, 10, &strict_limit);
 314         if (ret < 0)
 315                 return ret;
 316
 317         ret = bdi_set_strict_limit(bdi, strict_limit);
 318         if (!ret)
 319                 ret = count;
 320
 321         return ret;
 322 }
 323
 324 static ssize_t strict_limit_show(struct device *dev,
 325                 struct device_attribute *attr, char *buf)
 326 {
 327         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 328
 329         return sysfs_emit(buf, "%d\n",
 330                         !!(bdi->capabilities & BDI_CAP_STRICTLIMIT));
 331 }
 332 static DEVICE_ATTR_RW(strict_limit);
 333
 334 static struct attribute *bdi_dev_attrs[] = {
 335         &dev_attr_read_ahead_kb.attr,
 336         &dev_attr_min_ratio.attr,
 337         &dev_attr_min_ratio_fine.attr,
 338         &dev_attr_max_ratio.attr,
 339         &dev_attr_max_ratio_fine.attr,
 340         &dev_attr_min_bytes.attr,
 341         &dev_attr_max_bytes.attr,
 342         &dev_attr_stable_pages_required.attr,
 343         &dev_attr_strict_limit.attr,
 344         NULL,
 345 };
 346 ATTRIBUTE_GROUPS(bdi_dev);
 347
 348 static __init int bdi_class_init(void)
 349 {
 350         bdi_class = class_create("bdi");
 351         if (IS_ERR(bdi_class))
 352                 return PTR_ERR(bdi_class);
 353
 354         bdi_class->dev_groups = bdi_dev_groups;
 355         bdi_debug_init();
 356
 357         return 0;
 358 }
 359 postcore_initcall(bdi_class_init);
 360
 361 static int __init default_bdi_init(void)
 362 {
 363         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 364                                  WQ_SYSFS, 0);
 365         if (!bdi_wq)
 366                 return -ENOMEM;
 367         return 0;
 368 }
 369 subsys_initcall(default_bdi_init);
 370
 371 /*
 372  * This function is used when the first inode for this wb is marked dirty. It
 373  * wakes-up the corresponding bdi thread which should then take care of the
 374  * periodic background write-out of dirty inodes. Since the write-out would
 375  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 376  * set up a timer which wakes the bdi thread up later.
 377  *
 378  * Note, we wouldn't bother setting up the timer, but this function is on the
 379  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 380  * by delaying the wake-up.
 381  *
 382  * We have to be careful not to postpone flush work if it is scheduled for
 383  * earlier. Thus we use queue_delayed_work().
 384  */
 385 void wb_wakeup_delayed(struct bdi_writeback *wb)
 386 {
 387         unsigned long timeout;
 388
 389         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 390         spin_lock_irq(&wb->work_lock);
 391         if (test_bit(WB_registered, &wb->state))
 392                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 393         spin_unlock_irq(&wb->work_lock);
 394 }
 395
 396 static void wb_update_bandwidth_workfn(struct work_struct *work)
 397 {
 398         struct bdi_writeback *wb = container_of(to_delayed_work(work),
 399                                                 struct bdi_writeback, bw_dwork);
 400
 401         wb_update_bandwidth(wb);
 402 }
 403
 404 /*
 405  * Initial write bandwidth: 100 MB/s
 406  */
 407 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 408
 409 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 410                    gfp_t gfp)
 411 {
 412         int i, err;
 413
 414         memset(wb, 0, sizeof(*wb));
 415
 416         wb->bdi = bdi;
 417         wb->last_old_flush = jiffies;
 418         INIT_LIST_HEAD(&wb->b_dirty);
 419         INIT_LIST_HEAD(&wb->b_io);
 420         INIT_LIST_HEAD(&wb->b_more_io);
 421         INIT_LIST_HEAD(&wb->b_dirty_time);
 422         spin_lock_init(&wb->list_lock);
 423
 424         atomic_set(&wb->writeback_inodes, 0);
 425         wb->bw_time_stamp = jiffies;
 426         wb->balanced_dirty_ratelimit = INIT_BW;
 427         wb->dirty_ratelimit = INIT_BW;
 428         wb->write_bandwidth = INIT_BW;
 429         wb->avg_write_bandwidth = INIT_BW;
 430
 431         spin_lock_init(&wb->work_lock);
 432         INIT_LIST_HEAD(&wb->work_list);
 433         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 434         INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
 435         wb->dirty_sleep = jiffies;
 436
 437         err = fprop_local_init_percpu(&wb->completions, gfp);
 438         if (err)
 439                 return err;
 440
 441         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 442                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 443                 if (err)
 444                         goto out_destroy_stat;
 445         }
 446
 447         return 0;
 448
 449 out_destroy_stat:
 450         while (i--)
 451                 percpu_counter_destroy(&wb->stat[i]);
 452         fprop_local_destroy_percpu(&wb->completions);
 453         return err;
 454 }
 455
 456 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 457
 458 /*
 459  * Remove bdi from the global list and shutdown any threads we have running
 460  */
 461 static void wb_shutdown(struct bdi_writeback *wb)
 462 {
 463         /* Make sure nobody queues further work */
 464         spin_lock_irq(&wb->work_lock);
 465         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 466                 spin_unlock_irq(&wb->work_lock);
 467                 return;
 468         }
 469         spin_unlock_irq(&wb->work_lock);
 470
 471         cgwb_remove_from_bdi_list(wb);
 472         /*
 473          * Drain work list and shutdown the delayed_work.  !WB_registered
 474          * tells wb_workfn() that @wb is dying and its work_list needs to
 475          * be drained no matter what.
 476          */
 477         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 478         flush_delayed_work(&wb->dwork);
 479         WARN_ON(!list_empty(&wb->work_list));
 480         flush_delayed_work(&wb->bw_dwork);
 481 }
 482
 483 static void wb_exit(struct bdi_writeback *wb)
 484 {
 485         int i;
 486
 487         WARN_ON(delayed_work_pending(&wb->dwork));
 488
 489         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 490                 percpu_counter_destroy(&wb->stat[i]);
 491
 492         fprop_local_destroy_percpu(&wb->completions);
 493 }
 494
 495 #ifdef CONFIG_CGROUP_WRITEBACK
 496
 497 #include <linux/memcontrol.h>
 498
 499 /*
 500  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
 501  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
 502  */
 503 static DEFINE_SPINLOCK(cgwb_lock);
 504 static struct workqueue_struct *cgwb_release_wq;
 505
 506 static LIST_HEAD(offline_cgwbs);
 507 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
 508 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
 509
 510 static void cgwb_free_rcu(struct rcu_head *rcu_head)
 511 {
 512         struct bdi_writeback *wb = container_of(rcu_head,
 513                         struct bdi_writeback, rcu);
 514
 515         percpu_ref_exit(&wb->refcnt);
 516         kfree(wb);
 517 }
 518
 519 static void cgwb_release_workfn(struct work_struct *work)
 520 {
 521         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 522                                                 release_work);
 523         struct backing_dev_info *bdi = wb->bdi;
 524
 525         mutex_lock(&wb->bdi->cgwb_release_mutex);
 526         wb_shutdown(wb);
 527
 528         css_put(wb->memcg_css);
 529         css_put(wb->blkcg_css);
 530         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 531
 532         /* triggers blkg destruction if no online users left */
 533         blkcg_unpin_online(wb->blkcg_css);
 534
 535         fprop_local_destroy_percpu(&wb->memcg_completions);
 536
 537         spin_lock_irq(&cgwb_lock);
 538         list_del(&wb->offline_node);
 539         spin_unlock_irq(&cgwb_lock);
 540
 541         wb_exit(wb);
 542         bdi_put(bdi);
 543         WARN_ON_ONCE(!list_empty(&wb->b_attached));
 544         call_rcu(&wb->rcu, cgwb_free_rcu);
 545 }
 546
 547 static void cgwb_release(struct percpu_ref *refcnt)
 548 {
 549         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 550                                                 refcnt);
 551         queue_work(cgwb_release_wq, &wb->release_work);
 552 }
 553
 554 static void cgwb_kill(struct bdi_writeback *wb)
 555 {
 556         lockdep_assert_held(&cgwb_lock);
 557
 558         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 559         list_del(&wb->memcg_node);
 560         list_del(&wb->blkcg_node);
 561         list_add(&wb->offline_node, &offline_cgwbs);
 562         percpu_ref_kill(&wb->refcnt);
 563 }
 564
 565 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 566 {
 567         spin_lock_irq(&cgwb_lock);
 568         list_del_rcu(&wb->bdi_node);
 569         spin_unlock_irq(&cgwb_lock);
 570 }
 571
 572 static int cgwb_create(struct backing_dev_info *bdi,
 573                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 574 {
 575         struct mem_cgroup *memcg;
 576         struct cgroup_subsys_state *blkcg_css;
 577         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 578         struct bdi_writeback *wb;
 579         unsigned long flags;
 580         int ret = 0;
 581
 582         memcg = mem_cgroup_from_css(memcg_css);
 583         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 584         memcg_cgwb_list = &memcg->cgwb_list;
 585         blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
 586
 587         /* look up again under lock and discard on blkcg mismatch */
 588         spin_lock_irqsave(&cgwb_lock, flags);
 589         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 590         if (wb && wb->blkcg_css != blkcg_css) {
 591                 cgwb_kill(wb);
 592                 wb = NULL;
 593         }
 594         spin_unlock_irqrestore(&cgwb_lock, flags);
 595         if (wb)
 596                 goto out_put;
 597
 598         /* need to create a new one */
 599         wb = kmalloc(sizeof(*wb), gfp);
 600         if (!wb) {
 601                 ret = -ENOMEM;
 602                 goto out_put;
 603         }
 604
 605         ret = wb_init(wb, bdi, gfp);
 606         if (ret)
 607                 goto err_free;
 608
 609         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 610         if (ret)
 611                 goto err_wb_exit;
 612
 613         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 614         if (ret)
 615                 goto err_ref_exit;
 616
 617         wb->memcg_css = memcg_css;
 618         wb->blkcg_css = blkcg_css;
 619         INIT_LIST_HEAD(&wb->b_attached);
 620         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 621         set_bit(WB_registered, &wb->state);
 622         bdi_get(bdi);
 623
 624         /*
 625          * The root wb determines the registered state of the whole bdi and
 626          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 627          * whether they're still online.  Don't link @wb if any is dead.
 628          * See wb_memcg_offline() and wb_blkcg_offline().
 629          */
 630         ret = -ENODEV;
 631         spin_lock_irqsave(&cgwb_lock, flags);
 632         if (test_bit(WB_registered, &bdi->wb.state) &&
 633             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 634                 /* we might have raced another instance of this function */
 635                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 636                 if (!ret) {
 637                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 638                         list_add(&wb->memcg_node, memcg_cgwb_list);
 639                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 640                         blkcg_pin_online(blkcg_css);
 641                         css_get(memcg_css);
 642                         css_get(blkcg_css);
 643                 }
 644         }
 645         spin_unlock_irqrestore(&cgwb_lock, flags);
 646         if (ret) {
 647                 if (ret == -EEXIST)
 648                         ret = 0;
 649                 goto err_fprop_exit;
 650         }
 651         goto out_put;
 652
 653 err_fprop_exit:
 654         bdi_put(bdi);
 655         fprop_local_destroy_percpu(&wb->memcg_completions);
 656 err_ref_exit:
 657         percpu_ref_exit(&wb->refcnt);
 658 err_wb_exit:
 659         wb_exit(wb);
 660 err_free:
 661         kfree(wb);
 662 out_put:
 663         css_put(blkcg_css);
 664         return ret;
 665 }
 666
 667 /**
 668  * wb_get_lookup - get wb for a given memcg
 669  * @bdi: target bdi
 670  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 671  *
 672  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 673  * refcount incremented.
 674  *
 675  * This function uses css_get() on @memcg_css and thus expects its refcnt
 676  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 677  * @memcg_css isn't enough.  try_get it before calling this function.
 678  *
 679  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 680  * memcg on the default hierarchy, memcg association is guaranteed to be
 681  * more specific (equal or descendant to the associated blkcg) and thus can
 682  * identify both the memcg and blkcg associations.
 683  *
 684  * Because the blkcg associated with a memcg may change as blkcg is enabled
 685  * and disabled closer to root in the hierarchy, each wb keeps track of
 686  * both the memcg and blkcg associated with it and verifies the blkcg on
 687  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 688  * created.
 689  */
 690 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 691                                     struct cgroup_subsys_state *memcg_css)
 692 {
 693         struct bdi_writeback *wb;
 694
 695         if (!memcg_css->parent)
 696                 return &bdi->wb;
 697
 698         rcu_read_lock();
 699         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 700         if (wb) {
 701                 struct cgroup_subsys_state *blkcg_css;
 702
 703                 /* see whether the blkcg association has changed */
 704                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 705                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 706                         wb = NULL;
 707                 css_put(blkcg_css);
 708         }
 709         rcu_read_unlock();
 710
 711         return wb;
 712 }
 713
 714 /**
 715  * wb_get_create - get wb for a given memcg, create if necessary
 716  * @bdi: target bdi
 717  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 718  * @gfp: allocation mask to use
 719  *
 720  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 721  * create one.  See wb_get_lookup() for more details.
 722  */
 723 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 724                                     struct cgroup_subsys_state *memcg_css,
 725                                     gfp_t gfp)
 726 {
 727         struct bdi_writeback *wb;
 728
 729         might_alloc(gfp);
 730
 731         if (!memcg_css->parent)
 732                 return &bdi->wb;
 733
 734         do {
 735                 wb = wb_get_lookup(bdi, memcg_css);
 736         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 737
 738         return wb;
 739 }
 740
 741 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 742 {
 743         int ret;
 744
 745         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 746         mutex_init(&bdi->cgwb_release_mutex);
 747         init_rwsem(&bdi->wb_switch_rwsem);
 748
 749         ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 750         if (!ret) {
 751                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 752                 bdi->wb.blkcg_css = blkcg_root_css;
 753         }
 754         return ret;
 755 }
 756
 757 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 758 {
 759         struct radix_tree_iter iter;
 760         void **slot;
 761         struct bdi_writeback *wb;
 762
 763         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 764
 765         spin_lock_irq(&cgwb_lock);
 766         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 767                 cgwb_kill(*slot);
 768         spin_unlock_irq(&cgwb_lock);
 769
 770         mutex_lock(&bdi->cgwb_release_mutex);
 771         spin_lock_irq(&cgwb_lock);
 772         while (!list_empty(&bdi->wb_list)) {
 773                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 774                                       bdi_node);
 775                 spin_unlock_irq(&cgwb_lock);
 776                 wb_shutdown(wb);
 777                 spin_lock_irq(&cgwb_lock);
 778         }
 779         spin_unlock_irq(&cgwb_lock);
 780         mutex_unlock(&bdi->cgwb_release_mutex);
 781 }
 782
 783 /*
 784  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
 785  *
 786  * Try to release dying cgwbs by switching attached inodes to the nearest
 787  * living ancestor's writeback. Processed wbs are placed at the end
 788  * of the list to guarantee the forward progress.
 789  */
 790 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
 791 {
 792         struct bdi_writeback *wb;
 793         LIST_HEAD(processed);
 794
 795         spin_lock_irq(&cgwb_lock);
 796
 797         while (!list_empty(&offline_cgwbs)) {
 798                 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
 799                                       offline_node);
 800                 list_move(&wb->offline_node, &processed);
 801
 802                 /*
 803                  * If wb is dirty, cleaning up the writeback by switching
 804                  * attached inodes will result in an effective removal of any
 805                  * bandwidth restrictions, which isn't the goal.  Instead,
 806                  * it can be postponed until the next time, when all io
 807                  * will be likely completed.  If in the meantime some inodes
 808                  * will get re-dirtied, they should be eventually switched to
 809                  * a new cgwb.
 810                  */
 811                 if (wb_has_dirty_io(wb))
 812                         continue;
 813
 814                 if (!wb_tryget(wb))
 815                         continue;
 816
 817                 spin_unlock_irq(&cgwb_lock);
 818                 while (cleanup_offline_cgwb(wb))
 819                         cond_resched();
 820                 spin_lock_irq(&cgwb_lock);
 821
 822                 wb_put(wb);
 823         }
 824
 825         if (!list_empty(&processed))
 826                 list_splice_tail(&processed, &offline_cgwbs);
 827
 828         spin_unlock_irq(&cgwb_lock);
 829 }
 830
 831 /**
 832  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 833  * @memcg: memcg being offlined
 834  *
 835  * Also prevents creation of any new wb's associated with @memcg.
 836  */
 837 void wb_memcg_offline(struct mem_cgroup *memcg)
 838 {
 839         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 840         struct bdi_writeback *wb, *next;
 841
 842         spin_lock_irq(&cgwb_lock);
 843         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 844                 cgwb_kill(wb);
 845         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 846         spin_unlock_irq(&cgwb_lock);
 847
 848         queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
 849 }
 850
 851 /**
 852  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 853  * @css: blkcg being offlined
 854  *
 855  * Also prevents creation of any new wb's associated with @blkcg.
 856  */
 857 void wb_blkcg_offline(struct cgroup_subsys_state *css)
 858 {
 859         struct bdi_writeback *wb, *next;
 860         struct list_head *list = blkcg_get_cgwb_list(css);
 861
 862         spin_lock_irq(&cgwb_lock);
 863         list_for_each_entry_safe(wb, next, list, blkcg_node)
 864                 cgwb_kill(wb);
 865         list->next = NULL;      /* prevent new wb's */
 866         spin_unlock_irq(&cgwb_lock);
 867 }
 868
 869 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 870 {
 871         spin_lock_irq(&cgwb_lock);
 872         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 873         spin_unlock_irq(&cgwb_lock);
 874 }
 875
 876 static int __init cgwb_init(void)
 877 {
 878         /*
 879          * There can be many concurrent release work items overwhelming
 880          * system_wq.  Put them in a separate wq and limit concurrency.
 881          * There's no point in executing many of these in parallel.
 882          */
 883         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 884         if (!cgwb_release_wq)
 885                 return -ENOMEM;
 886
 887         return 0;
 888 }
 889 subsys_initcall(cgwb_init);
 890
 891 #else   /* CONFIG_CGROUP_WRITEBACK */
 892
 893 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 894 {
 895         return wb_init(&bdi->wb, bdi, GFP_KERNEL);
 896 }
 897
 898 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 899
 900 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 901 {
 902         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 903 }
 904
 905 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 906 {
 907         list_del_rcu(&wb->bdi_node);
 908 }
 909
 910 #endif  /* CONFIG_CGROUP_WRITEBACK */
 911
 912 int bdi_init(struct backing_dev_info *bdi)
 913 {
 914         bdi->dev = NULL;
 915
 916         kref_init(&bdi->refcnt);
 917         bdi->min_ratio = 0;
 918         bdi->max_ratio = 100 * BDI_RATIO_SCALE;
 919         bdi->max_prop_frac = FPROP_FRAC_BASE;
 920         INIT_LIST_HEAD(&bdi->bdi_list);
 921         INIT_LIST_HEAD(&bdi->wb_list);
 922         init_waitqueue_head(&bdi->wb_waitq);
 923
 924         return cgwb_bdi_init(bdi);
 925 }
 926
 927 struct backing_dev_info *bdi_alloc(int node_id)
 928 {
 929         struct backing_dev_info *bdi;
 930
 931         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 932         if (!bdi)
 933                 return NULL;
 934
 935         if (bdi_init(bdi)) {
 936                 kfree(bdi);
 937                 return NULL;
 938         }
 939         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 940         bdi->ra_pages = VM_READAHEAD_PAGES;
 941         bdi->io_pages = VM_READAHEAD_PAGES;
 942         timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
 943         return bdi;
 944 }
 945 EXPORT_SYMBOL(bdi_alloc);
 946
 947 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 948 {
 949         struct rb_node **p = &bdi_tree.rb_node;
 950         struct rb_node *parent = NULL;
 951         struct backing_dev_info *bdi;
 952
 953         lockdep_assert_held(&bdi_lock);
 954
 955         while (*p) {
 956                 parent = *p;
 957                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 958
 959                 if (bdi->id > id)
 960                         p = &(*p)->rb_left;
 961                 else if (bdi->id < id)
 962                         p = &(*p)->rb_right;
 963                 else
 964                         break;
 965         }
 966
 967         if (parentp)
 968                 *parentp = parent;
 969         return p;
 970 }
 971
 972 /**
 973  * bdi_get_by_id - lookup and get bdi from its id
 974  * @id: bdi id to lookup
 975  *
 976  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 977  * doesn't exist or is already unregistered.
 978  */
 979 struct backing_dev_info *bdi_get_by_id(u64 id)
 980 {
 981         struct backing_dev_info *bdi = NULL;
 982         struct rb_node **p;
 983
 984         spin_lock_bh(&bdi_lock);
 985         p = bdi_lookup_rb_node(id, NULL);
 986         if (*p) {
 987                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 988                 bdi_get(bdi);
 989         }
 990         spin_unlock_bh(&bdi_lock);
 991
 992         return bdi;
 993 }
 994
 995 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 996 {
 997         struct device *dev;
 998         struct rb_node *parent, **p;
 999
1000         if (bdi->dev)   /* The driver needs to use separate queues per device */
1001                 return 0;
1002
1003         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
1004         dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
1005         if (IS_ERR(dev))
1006                 return PTR_ERR(dev);
1007
1008         cgwb_bdi_register(bdi);
1009         bdi->dev = dev;
1010
1011         bdi_debug_register(bdi, dev_name(dev));
1012         set_bit(WB_registered, &bdi->wb.state);
1013
1014         spin_lock_bh(&bdi_lock);
1015
1016         bdi->id = ++bdi_id_cursor;
1017
1018         p = bdi_lookup_rb_node(bdi->id, &parent);
1019         rb_link_node(&bdi->rb_node, parent, p);
1020         rb_insert_color(&bdi->rb_node, &bdi_tree);
1021
1022         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
1023
1024         spin_unlock_bh(&bdi_lock);
1025
1026         trace_writeback_bdi_register(bdi);
1027         return 0;
1028 }
1029
1030 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
1031 {
1032         va_list args;
1033         int ret;
1034
1035         va_start(args, fmt);
1036         ret = bdi_register_va(bdi, fmt, args);
1037         va_end(args);
1038         return ret;
1039 }
1040 EXPORT_SYMBOL(bdi_register);
1041
1042 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
1043 {
1044         WARN_ON_ONCE(bdi->owner);
1045         bdi->owner = owner;
1046         get_device(owner);
1047 }
1048
1049 /*
1050  * Remove bdi from bdi_list, and ensure that it is no longer visible
1051  */
1052 static void bdi_remove_from_list(struct backing_dev_info *bdi)
1053 {
1054         spin_lock_bh(&bdi_lock);
1055         rb_erase(&bdi->rb_node, &bdi_tree);
1056         list_del_rcu(&bdi->bdi_list);
1057         spin_unlock_bh(&bdi_lock);
1058
1059         synchronize_rcu_expedited();
1060 }
1061
1062 void bdi_unregister(struct backing_dev_info *bdi)
1063 {
1064         del_timer_sync(&bdi->laptop_mode_wb_timer);
1065
1066         /* make sure nobody finds us on the bdi_list anymore */
1067         bdi_remove_from_list(bdi);
1068         wb_shutdown(&bdi->wb);
1069         cgwb_bdi_unregister(bdi);
1070
1071         /*
1072          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
1073          * update the global bdi_min_ratio.
1074          */
1075         if (bdi->min_ratio)
1076                 bdi_set_min_ratio(bdi, 0);
1077
1078         if (bdi->dev) {
1079                 bdi_debug_unregister(bdi);
1080                 device_unregister(bdi->dev);
1081                 bdi->dev = NULL;
1082         }
1083
1084         if (bdi->owner) {
1085                 put_device(bdi->owner);
1086                 bdi->owner = NULL;
1087         }
1088 }
1089 EXPORT_SYMBOL(bdi_unregister);
1090
1091 static void release_bdi(struct kref *ref)
1092 {
1093         struct backing_dev_info *bdi =
1094                         container_of(ref, struct backing_dev_info, refcnt);
1095
1096         WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
1097         WARN_ON_ONCE(bdi->dev);
1098         wb_exit(&bdi->wb);
1099         kfree(bdi);
1100 }
1101
1102 void bdi_put(struct backing_dev_info *bdi)
1103 {
1104         kref_put(&bdi->refcnt, release_bdi);
1105 }
1106 EXPORT_SYMBOL(bdi_put);
1107
1108 struct backing_dev_info *inode_to_bdi(struct inode *inode)
1109 {
1110         struct super_block *sb;
1111
1112         if (!inode)
1113                 return &noop_backing_dev_info;
1114
1115         sb = inode->i_sb;
1116 #ifdef CONFIG_BLOCK
1117         if (sb_is_blkdev_sb(sb))
1118                 return I_BDEV(inode)->bd_disk->bdi;
1119 #endif
1120         return sb->s_bdi;
1121 }
1122 EXPORT_SYMBOL(inode_to_bdi);
1123
1124 const char *bdi_dev_name(struct backing_dev_info *bdi)
1125 {
1126         if (!bdi || !bdi->dev)
1127                 return bdi_unknown_name;
1128         return bdi->dev_name;
1129 }
1130 EXPORT_SYMBOL_GPL(bdi_dev_name);