mm/backing-dev.c

   1
   2 #include <linux/wait.h>
   3 #include <linux/backing-dev.h>
   4 #include <linux/kthread.h>
   5 #include <linux/freezer.h>
   6 #include <linux/fs.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/module.h>
  11 #include <linux/writeback.h>
  12 #include <linux/device.h>
  13 #include <trace/events/writeback.h>
  14
  15 static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
  16
  17 struct backing_dev_info noop_backing_dev_info = {
  18         .name           = "noop",
  19         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
  20 };
  21 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  22
  23 static struct class *bdi_class;
  24
  25 /*
  26  * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
  27  * locking.
  28  */
  29 DEFINE_SPINLOCK(bdi_lock);
  30 LIST_HEAD(bdi_list);
  31
  32 /* bdi_wq serves all asynchronous writeback tasks */
  33 struct workqueue_struct *bdi_wq;
  34
  35 #ifdef CONFIG_DEBUG_FS
  36 #include <linux/debugfs.h>
  37 #include <linux/seq_file.h>
  38
  39 static struct dentry *bdi_debug_root;
  40
  41 static void bdi_debug_init(void)
  42 {
  43         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  44 }
  45
  46 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  47 {
  48         struct backing_dev_info *bdi = m->private;
  49         struct bdi_writeback *wb = &bdi->wb;
  50         unsigned long background_thresh;
  51         unsigned long dirty_thresh;
  52         unsigned long wb_thresh;
  53         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  54         struct inode *inode;
  55
  56         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  57         spin_lock(&wb->list_lock);
  58         list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
  59                 nr_dirty++;
  60         list_for_each_entry(inode, &wb->b_io, i_wb_list)
  61                 nr_io++;
  62         list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
  63                 nr_more_io++;
  64         list_for_each_entry(inode, &wb->b_dirty_time, i_wb_list)
  65                 if (inode->i_state & I_DIRTY_TIME)
  66                         nr_dirty_time++;
  67         spin_unlock(&wb->list_lock);
  68
  69         global_dirty_limits(&background_thresh, &dirty_thresh);
  70         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  71
  72 #define K(x) ((x) << (PAGE_SHIFT - 10))
  73         seq_printf(m,
  74                    "BdiWriteback:       %10lu kB\n"
  75                    "BdiReclaimable:     %10lu kB\n"
  76                    "BdiDirtyThresh:     %10lu kB\n"
  77                    "DirtyThresh:        %10lu kB\n"
  78                    "BackgroundThresh:   %10lu kB\n"
  79                    "BdiDirtied:         %10lu kB\n"
  80                    "BdiWritten:         %10lu kB\n"
  81                    "BdiWriteBandwidth:  %10lu kBps\n"
  82                    "b_dirty:            %10lu\n"
  83                    "b_io:               %10lu\n"
  84                    "b_more_io:          %10lu\n"
  85                    "b_dirty_time:       %10lu\n"
  86                    "bdi_list:           %10u\n"
  87                    "state:              %10lx\n",
  88                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  89                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  90                    K(wb_thresh),
  91                    K(dirty_thresh),
  92                    K(background_thresh),
  93                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  94                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  95                    (unsigned long) K(wb->write_bandwidth),
  96                    nr_dirty,
  97                    nr_io,
  98                    nr_more_io,
  99                    nr_dirty_time,
 100                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 101 #undef K
 102
 103         return 0;
 104 }
 105
 106 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 107 {
 108         return single_open(file, bdi_debug_stats_show, inode->i_private);
 109 }
 110
 111 static const struct file_operations bdi_debug_stats_fops = {
 112         .open           = bdi_debug_stats_open,
 113         .read           = seq_read,
 114         .llseek         = seq_lseek,
 115         .release        = single_release,
 116 };
 117
 118 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 119 {
 120         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 121         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 122                                                bdi, &bdi_debug_stats_fops);
 123 }
 124
 125 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 126 {
 127         debugfs_remove(bdi->debug_stats);
 128         debugfs_remove(bdi->debug_dir);
 129 }
 130 #else
 131 static inline void bdi_debug_init(void)
 132 {
 133 }
 134 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 135                                       const char *name)
 136 {
 137 }
 138 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 139 {
 140 }
 141 #endif
 142
 143 static ssize_t read_ahead_kb_store(struct device *dev,
 144                                   struct device_attribute *attr,
 145                                   const char *buf, size_t count)
 146 {
 147         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 148         unsigned long read_ahead_kb;
 149         ssize_t ret;
 150
 151         ret = kstrtoul(buf, 10, &read_ahead_kb);
 152         if (ret < 0)
 153                 return ret;
 154
 155         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 156
 157         return count;
 158 }
 159
 160 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 161
 162 #define BDI_SHOW(name, expr)                                            \
 163 static ssize_t name##_show(struct device *dev,                          \
 164                            struct device_attribute *attr, char *page)   \
 165 {                                                                       \
 166         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 167                                                                         \
 168         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 169 }                                                                       \
 170 static DEVICE_ATTR_RW(name);
 171
 172 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 173
 174 static ssize_t min_ratio_store(struct device *dev,
 175                 struct device_attribute *attr, const char *buf, size_t count)
 176 {
 177         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 178         unsigned int ratio;
 179         ssize_t ret;
 180
 181         ret = kstrtouint(buf, 10, &ratio);
 182         if (ret < 0)
 183                 return ret;
 184
 185         ret = bdi_set_min_ratio(bdi, ratio);
 186         if (!ret)
 187                 ret = count;
 188
 189         return ret;
 190 }
 191 BDI_SHOW(min_ratio, bdi->min_ratio)
 192
 193 static ssize_t max_ratio_store(struct device *dev,
 194                 struct device_attribute *attr, const char *buf, size_t count)
 195 {
 196         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 197         unsigned int ratio;
 198         ssize_t ret;
 199
 200         ret = kstrtouint(buf, 10, &ratio);
 201         if (ret < 0)
 202                 return ret;
 203
 204         ret = bdi_set_max_ratio(bdi, ratio);
 205         if (!ret)
 206                 ret = count;
 207
 208         return ret;
 209 }
 210 BDI_SHOW(max_ratio, bdi->max_ratio)
 211
 212 static ssize_t stable_pages_required_show(struct device *dev,
 213                                           struct device_attribute *attr,
 214                                           char *page)
 215 {
 216         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 217
 218         return snprintf(page, PAGE_SIZE-1, "%d\n",
 219                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 220 }
 221 static DEVICE_ATTR_RO(stable_pages_required);
 222
 223 static struct attribute *bdi_dev_attrs[] = {
 224         &dev_attr_read_ahead_kb.attr,
 225         &dev_attr_min_ratio.attr,
 226         &dev_attr_max_ratio.attr,
 227         &dev_attr_stable_pages_required.attr,
 228         NULL,
 229 };
 230 ATTRIBUTE_GROUPS(bdi_dev);
 231
 232 static __init int bdi_class_init(void)
 233 {
 234         bdi_class = class_create(THIS_MODULE, "bdi");
 235         if (IS_ERR(bdi_class))
 236                 return PTR_ERR(bdi_class);
 237
 238         bdi_class->dev_groups = bdi_dev_groups;
 239         bdi_debug_init();
 240         return 0;
 241 }
 242 postcore_initcall(bdi_class_init);
 243
 244 static int __init default_bdi_init(void)
 245 {
 246         int err;
 247
 248         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
 249                                               WQ_UNBOUND | WQ_SYSFS, 0);
 250         if (!bdi_wq)
 251                 return -ENOMEM;
 252
 253         err = bdi_init(&noop_backing_dev_info);
 254
 255         return err;
 256 }
 257 subsys_initcall(default_bdi_init);
 258
 259 /*
 260  * This function is used when the first inode for this wb is marked dirty. It
 261  * wakes-up the corresponding bdi thread which should then take care of the
 262  * periodic background write-out of dirty inodes. Since the write-out would
 263  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 264  * set up a timer which wakes the bdi thread up later.
 265  *
 266  * Note, we wouldn't bother setting up the timer, but this function is on the
 267  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 268  * by delaying the wake-up.
 269  *
 270  * We have to be careful not to postpone flush work if it is scheduled for
 271  * earlier. Thus we use queue_delayed_work().
 272  */
 273 void wb_wakeup_delayed(struct bdi_writeback *wb)
 274 {
 275         unsigned long timeout;
 276
 277         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 278         spin_lock_bh(&wb->work_lock);
 279         if (test_bit(WB_registered, &wb->state))
 280                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 281         spin_unlock_bh(&wb->work_lock);
 282 }
 283
 284 /*
 285  * Initial write bandwidth: 100 MB/s
 286  */
 287 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 288
 289 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 290                    gfp_t gfp)
 291 {
 292         int i, err;
 293
 294         memset(wb, 0, sizeof(*wb));
 295
 296         wb->bdi = bdi;
 297         wb->last_old_flush = jiffies;
 298         INIT_LIST_HEAD(&wb->b_dirty);
 299         INIT_LIST_HEAD(&wb->b_io);
 300         INIT_LIST_HEAD(&wb->b_more_io);
 301         INIT_LIST_HEAD(&wb->b_dirty_time);
 302         spin_lock_init(&wb->list_lock);
 303
 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
 314         err = fprop_local_init_percpu(&wb->completions, gfp);
 315         if (err)
 316                 return err;
 317
 318         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 319                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 320                 if (err) {
 321                         while (--i)
 322                                 percpu_counter_destroy(&wb->stat[i]);
 323                         fprop_local_destroy_percpu(&wb->completions);
 324                         return err;
 325                 }
 326         }
 327
 328         return 0;
 329 }
 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_bh(&wb->work_lock);
 338         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 339                 spin_unlock_bh(&wb->work_lock);
 340                 return;
 341         }
 342         spin_unlock_bh(&wb->work_lock);
 343
 344         /*
 345          * Drain work list and shutdown the delayed_work.  !WB_registered
 346          * tells wb_workfn() that @wb is dying and its work_list needs to
 347          * be drained no matter what.
 348          */
 349         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 350         flush_delayed_work(&wb->dwork);
 351         WARN_ON(!list_empty(&wb->work_list));
 352 }
 353
 354 static void wb_exit(struct bdi_writeback *wb)
 355 {
 356         int i;
 357
 358         WARN_ON(delayed_work_pending(&wb->dwork));
 359
 360         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 361                 percpu_counter_destroy(&wb->stat[i]);
 362
 363         fprop_local_destroy_percpu(&wb->completions);
 364 }
 365
 366 #ifdef CONFIG_CGROUP_WRITEBACK
 367
 368 #include <linux/memcontrol.h>
 369
 370 /*
 371  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 372  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 373  * protected.  cgwb_release_wait is used to wait for the completion of cgwb
 374  * releases from bdi destruction path.
 375  */
 376 static DEFINE_SPINLOCK(cgwb_lock);
 377 static DECLARE_WAIT_QUEUE_HEAD(cgwb_release_wait);
 378
 379 /**
 380  * wb_congested_get_create - get or create a wb_congested
 381  * @bdi: associated bdi
 382  * @blkcg_id: ID of the associated blkcg
 383  * @gfp: allocation mask
 384  *
 385  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 386  * The returned wb_congested has its reference count incremented.  Returns
 387  * NULL on failure.
 388  */
 389 struct bdi_writeback_congested *
 390 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 391 {
 392         struct bdi_writeback_congested *new_congested = NULL, *congested;
 393         struct rb_node **node, *parent;
 394         unsigned long flags;
 395
 396         if (blkcg_id == 1)
 397                 return &bdi->wb_congested;
 398 retry:
 399         spin_lock_irqsave(&cgwb_lock, flags);
 400
 401         node = &bdi->cgwb_congested_tree.rb_node;
 402         parent = NULL;
 403
 404         while (*node != NULL) {
 405                 parent = *node;
 406                 congested = container_of(parent, struct bdi_writeback_congested,
 407                                          rb_node);
 408                 if (congested->blkcg_id < blkcg_id)
 409                         node = &parent->rb_left;
 410                 else if (congested->blkcg_id > blkcg_id)
 411                         node = &parent->rb_right;
 412                 else
 413                         goto found;
 414         }
 415
 416         if (new_congested) {
 417                 /* !found and storage for new one already allocated, insert */
 418                 congested = new_congested;
 419                 new_congested = NULL;
 420                 rb_link_node(&congested->rb_node, parent, node);
 421                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 422                 atomic_inc(&bdi->usage_cnt);
 423                 goto found;
 424         }
 425
 426         spin_unlock_irqrestore(&cgwb_lock, flags);
 427
 428         /* allocate storage for new one and retry */
 429         new_congested = kzalloc(sizeof(*new_congested), gfp);
 430         if (!new_congested)
 431                 return NULL;
 432
 433         atomic_set(&new_congested->refcnt, 0);
 434         new_congested->bdi = bdi;
 435         new_congested->blkcg_id = blkcg_id;
 436         goto retry;
 437
 438 found:
 439         atomic_inc(&congested->refcnt);
 440         spin_unlock_irqrestore(&cgwb_lock, flags);
 441         kfree(new_congested);
 442         return congested;
 443 }
 444
 445 /**
 446  * wb_congested_put - put a wb_congested
 447  * @congested: wb_congested to put
 448  *
 449  * Put @congested and destroy it if the refcnt reaches zero.
 450  */
 451 void wb_congested_put(struct bdi_writeback_congested *congested)
 452 {
 453         struct backing_dev_info *bdi = congested->bdi;
 454         unsigned long flags;
 455
 456         if (congested->blkcg_id == 1)
 457                 return;
 458
 459         local_irq_save(flags);
 460         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 461                 local_irq_restore(flags);
 462                 return;
 463         }
 464
 465         rb_erase(&congested->rb_node, &congested->bdi->cgwb_congested_tree);
 466         spin_unlock_irqrestore(&cgwb_lock, flags);
 467         kfree(congested);
 468
 469         if (atomic_dec_and_test(&bdi->usage_cnt))
 470                 wake_up_all(&cgwb_release_wait);
 471 }
 472
 473 static void cgwb_release_workfn(struct work_struct *work)
 474 {
 475         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 476                                                 release_work);
 477         struct backing_dev_info *bdi = wb->bdi;
 478
 479         wb_shutdown(wb);
 480
 481         css_put(wb->memcg_css);
 482         css_put(wb->blkcg_css);
 483         wb_congested_put(wb->congested);
 484
 485         fprop_local_destroy_percpu(&wb->memcg_completions);
 486         percpu_ref_exit(&wb->refcnt);
 487         wb_exit(wb);
 488         kfree_rcu(wb, rcu);
 489
 490         if (atomic_dec_and_test(&bdi->usage_cnt))
 491                 wake_up_all(&cgwb_release_wait);
 492 }
 493
 494 static void cgwb_release(struct percpu_ref *refcnt)
 495 {
 496         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 497                                                 refcnt);
 498         schedule_work(&wb->release_work);
 499 }
 500
 501 static void cgwb_kill(struct bdi_writeback *wb)
 502 {
 503         lockdep_assert_held(&cgwb_lock);
 504
 505         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 506         list_del(&wb->memcg_node);
 507         list_del(&wb->blkcg_node);
 508         percpu_ref_kill(&wb->refcnt);
 509 }
 510
 511 static int cgwb_create(struct backing_dev_info *bdi,
 512                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 513 {
 514         struct mem_cgroup *memcg;
 515         struct cgroup_subsys_state *blkcg_css;
 516         struct blkcg *blkcg;
 517         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 518         struct bdi_writeback *wb;
 519         unsigned long flags;
 520         int ret = 0;
 521
 522         memcg = mem_cgroup_from_css(memcg_css);
 523         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &blkio_cgrp_subsys);
 524         blkcg = css_to_blkcg(blkcg_css);
 525         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 526         blkcg_cgwb_list = &blkcg->cgwb_list;
 527
 528         /* look up again under lock and discard on blkcg mismatch */
 529         spin_lock_irqsave(&cgwb_lock, flags);
 530         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 531         if (wb && wb->blkcg_css != blkcg_css) {
 532                 cgwb_kill(wb);
 533                 wb = NULL;
 534         }
 535         spin_unlock_irqrestore(&cgwb_lock, flags);
 536         if (wb)
 537                 goto out_put;
 538
 539         /* need to create a new one */
 540         wb = kmalloc(sizeof(*wb), gfp);
 541         if (!wb)
 542                 return -ENOMEM;
 543
 544         ret = wb_init(wb, bdi, gfp);
 545         if (ret)
 546                 goto err_free;
 547
 548         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 549         if (ret)
 550                 goto err_wb_exit;
 551
 552         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 553         if (ret)
 554                 goto err_ref_exit;
 555
 556         wb->congested = wb_congested_get_create(bdi, blkcg_css->id, gfp);
 557         if (!wb->congested)
 558                 goto err_fprop_exit;
 559
 560         wb->memcg_css = memcg_css;
 561         wb->blkcg_css = blkcg_css;
 562         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 563         set_bit(WB_registered, &wb->state);
 564
 565         /*
 566          * The root wb determines the registered state of the whole bdi and
 567          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 568          * whether they're still online.  Don't link @wb if any is dead.
 569          * See wb_memcg_offline() and wb_blkcg_offline().
 570          */
 571         ret = -ENODEV;
 572         spin_lock_irqsave(&cgwb_lock, flags);
 573         if (test_bit(WB_registered, &bdi->wb.state) &&
 574             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 575                 /* we might have raced another instance of this function */
 576                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 577                 if (!ret) {
 578                         atomic_inc(&bdi->usage_cnt);
 579                         list_add(&wb->memcg_node, memcg_cgwb_list);
 580                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 581                         css_get(memcg_css);
 582                         css_get(blkcg_css);
 583                 }
 584         }
 585         spin_unlock_irqrestore(&cgwb_lock, flags);
 586         if (ret) {
 587                 if (ret == -EEXIST)
 588                         ret = 0;
 589                 goto err_put_congested;
 590         }
 591         goto out_put;
 592
 593 err_put_congested:
 594         wb_congested_put(wb->congested);
 595 err_fprop_exit:
 596         fprop_local_destroy_percpu(&wb->memcg_completions);
 597 err_ref_exit:
 598         percpu_ref_exit(&wb->refcnt);
 599 err_wb_exit:
 600         wb_exit(wb);
 601 err_free:
 602         kfree(wb);
 603 out_put:
 604         css_put(blkcg_css);
 605         return ret;
 606 }
 607
 608 /**
 609  * wb_get_create - get wb for a given memcg, create if necessary
 610  * @bdi: target bdi
 611  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 612  * @gfp: allocation mask to use
 613  *
 614  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 615  * create one.  The returned wb has its refcount incremented.
 616  *
 617  * This function uses css_get() on @memcg_css and thus expects its refcnt
 618  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 619  * @memcg_css isn't enough.  try_get it before calling this function.
 620  *
 621  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 622  * memcg on the default hierarchy, memcg association is guaranteed to be
 623  * more specific (equal or descendant to the associated blkcg) and thus can
 624  * identify both the memcg and blkcg associations.
 625  *
 626  * Because the blkcg associated with a memcg may change as blkcg is enabled
 627  * and disabled closer to root in the hierarchy, each wb keeps track of
 628  * both the memcg and blkcg associated with it and verifies the blkcg on
 629  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 630  * created.
 631  */
 632 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 633                                     struct cgroup_subsys_state *memcg_css,
 634                                     gfp_t gfp)
 635 {
 636         struct bdi_writeback *wb;
 637
 638         might_sleep_if(gfp & __GFP_WAIT);
 639
 640         if (!memcg_css->parent)
 641                 return &bdi->wb;
 642
 643         do {
 644                 rcu_read_lock();
 645                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 646                 if (wb) {
 647                         struct cgroup_subsys_state *blkcg_css;
 648
 649                         /* see whether the blkcg association has changed */
 650                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 651                                                      &blkio_cgrp_subsys);
 652                         if (unlikely(wb->blkcg_css != blkcg_css ||
 653                                      !wb_tryget(wb)))
 654                                 wb = NULL;
 655                         css_put(blkcg_css);
 656                 }
 657                 rcu_read_unlock();
 658         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 659
 660         return wb;
 661 }
 662
 663 static void cgwb_bdi_init(struct backing_dev_info *bdi)
 664 {
 665         bdi->wb.memcg_css = mem_cgroup_root_css;
 666         bdi->wb.blkcg_css = blkcg_root_css;
 667         bdi->wb_congested.blkcg_id = 1;
 668         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 669         bdi->cgwb_congested_tree = RB_ROOT;
 670         atomic_set(&bdi->usage_cnt, 1);
 671 }
 672
 673 static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
 674 {
 675         struct radix_tree_iter iter;
 676         void **slot;
 677
 678         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 679
 680         spin_lock_irq(&cgwb_lock);
 681         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 682                 cgwb_kill(*slot);
 683         spin_unlock_irq(&cgwb_lock);
 684
 685         /*
 686          * All cgwb's and their congested states must be shutdown and
 687          * released before returning.  Drain the usage counter to wait for
 688          * all cgwb's and cgwb_congested's ever created on @bdi.
 689          */
 690         atomic_dec(&bdi->usage_cnt);
 691         wait_event(cgwb_release_wait, !atomic_read(&bdi->usage_cnt));
 692 }
 693
 694 /**
 695  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 696  * @memcg: memcg being offlined
 697  *
 698  * Also prevents creation of any new wb's associated with @memcg.
 699  */
 700 void wb_memcg_offline(struct mem_cgroup *memcg)
 701 {
 702         LIST_HEAD(to_destroy);
 703         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 704         struct bdi_writeback *wb, *next;
 705
 706         spin_lock_irq(&cgwb_lock);
 707         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 708                 cgwb_kill(wb);
 709         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 710         spin_unlock_irq(&cgwb_lock);
 711 }
 712
 713 /**
 714  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 715  * @blkcg: blkcg being offlined
 716  *
 717  * Also prevents creation of any new wb's associated with @blkcg.
 718  */
 719 void wb_blkcg_offline(struct blkcg *blkcg)
 720 {
 721         LIST_HEAD(to_destroy);
 722         struct bdi_writeback *wb, *next;
 723
 724         spin_lock_irq(&cgwb_lock);
 725         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 726                 cgwb_kill(wb);
 727         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 728         spin_unlock_irq(&cgwb_lock);
 729 }
 730
 731 #else   /* CONFIG_CGROUP_WRITEBACK */
 732
 733 static void cgwb_bdi_init(struct backing_dev_info *bdi) { }
 734 static void cgwb_bdi_destroy(struct backing_dev_info *bdi) { }
 735
 736 #endif  /* CONFIG_CGROUP_WRITEBACK */
 737
 738 int bdi_init(struct backing_dev_info *bdi)
 739 {
 740         int err;
 741
 742         bdi->dev = NULL;
 743
 744         bdi->min_ratio = 0;
 745         bdi->max_ratio = 100;
 746         bdi->max_prop_frac = FPROP_FRAC_BASE;
 747         INIT_LIST_HEAD(&bdi->bdi_list);
 748         init_waitqueue_head(&bdi->wb_waitq);
 749
 750         err = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 751         if (err)
 752                 return err;
 753
 754         bdi->wb_congested.state = 0;
 755         bdi->wb.congested = &bdi->wb_congested;
 756
 757         cgwb_bdi_init(bdi);
 758         return 0;
 759 }
 760 EXPORT_SYMBOL(bdi_init);
 761
 762 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 763                 const char *fmt, ...)
 764 {
 765         va_list args;
 766         struct device *dev;
 767
 768         if (bdi->dev)   /* The driver needs to use separate queues per device */
 769                 return 0;
 770
 771         va_start(args, fmt);
 772         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 773         va_end(args);
 774         if (IS_ERR(dev))
 775                 return PTR_ERR(dev);
 776
 777         bdi->dev = dev;
 778
 779         bdi_debug_register(bdi, dev_name(dev));
 780         set_bit(WB_registered, &bdi->wb.state);
 781
 782         spin_lock_bh(&bdi_lock);
 783         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 784         spin_unlock_bh(&bdi_lock);
 785
 786         trace_writeback_bdi_register(bdi);
 787         return 0;
 788 }
 789 EXPORT_SYMBOL(bdi_register);
 790
 791 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 792 {
 793         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 794 }
 795 EXPORT_SYMBOL(bdi_register_dev);
 796
 797 /*
 798  * Remove bdi from bdi_list, and ensure that it is no longer visible
 799  */
 800 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 801 {
 802         spin_lock_bh(&bdi_lock);
 803         list_del_rcu(&bdi->bdi_list);
 804         spin_unlock_bh(&bdi_lock);
 805
 806         synchronize_rcu_expedited();
 807 }
 808
 809 /*
 810  * Called when the device behind @bdi has been removed or ejected.
 811  *
 812  * We can't really do much here except for reducing the dirty ratio at
 813  * the moment.  In the future we should be able to set a flag so that
 814  * the filesystem can handle errors at mark_inode_dirty time instead
 815  * of only at writeback time.
 816  */
 817 void bdi_unregister(struct backing_dev_info *bdi)
 818 {
 819         if (WARN_ON_ONCE(!bdi->dev))
 820                 return;
 821
 822         bdi_set_min_ratio(bdi, 0);
 823 }
 824 EXPORT_SYMBOL(bdi_unregister);
 825
 826 void bdi_destroy(struct backing_dev_info *bdi)
 827 {
 828         /* make sure nobody finds us on the bdi_list anymore */
 829         bdi_remove_from_list(bdi);
 830         wb_shutdown(&bdi->wb);
 831         cgwb_bdi_destroy(bdi);
 832
 833         if (bdi->dev) {
 834                 bdi_debug_unregister(bdi);
 835                 device_unregister(bdi->dev);
 836                 bdi->dev = NULL;
 837         }
 838
 839         wb_exit(&bdi->wb);
 840 }
 841 EXPORT_SYMBOL(bdi_destroy);
 842
 843 /*
 844  * For use from filesystems to quickly init and register a bdi associated
 845  * with dirty writeback
 846  */
 847 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
 848 {
 849         int err;
 850
 851         bdi->name = name;
 852         bdi->capabilities = 0;
 853         err = bdi_init(bdi);
 854         if (err)
 855                 return err;
 856
 857         err = bdi_register(bdi, NULL, "%.28s-%ld", name,
 858                            atomic_long_inc_return(&bdi_seq));
 859         if (err) {
 860                 bdi_destroy(bdi);
 861                 return err;
 862         }
 863
 864         return 0;
 865 }
 866 EXPORT_SYMBOL(bdi_setup_and_register);
 867
 868 static wait_queue_head_t congestion_wqh[2] = {
 869                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 870                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 871         };
 872 static atomic_t nr_wb_congested[2];
 873
 874 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 875 {
 876         wait_queue_head_t *wqh = &congestion_wqh[sync];
 877         enum wb_state bit;
 878
 879         bit = sync ? WB_sync_congested : WB_async_congested;
 880         if (test_and_clear_bit(bit, &congested->state))
 881                 atomic_dec(&nr_wb_congested[sync]);
 882         smp_mb__after_atomic();
 883         if (waitqueue_active(wqh))
 884                 wake_up(wqh);
 885 }
 886 EXPORT_SYMBOL(clear_wb_congested);
 887
 888 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
 889 {
 890         enum wb_state bit;
 891
 892         bit = sync ? WB_sync_congested : WB_async_congested;
 893         if (!test_and_set_bit(bit, &congested->state))
 894                 atomic_inc(&nr_wb_congested[sync]);
 895 }
 896 EXPORT_SYMBOL(set_wb_congested);
 897
 898 /**
 899  * congestion_wait - wait for a backing_dev to become uncongested
 900  * @sync: SYNC or ASYNC IO
 901  * @timeout: timeout in jiffies
 902  *
 903  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 904  * write congestion.  If no backing_devs are congested then just wait for the
 905  * next write to be completed.
 906  */
 907 long congestion_wait(int sync, long timeout)
 908 {
 909         long ret;
 910         unsigned long start = jiffies;
 911         DEFINE_WAIT(wait);
 912         wait_queue_head_t *wqh = &congestion_wqh[sync];
 913
 914         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 915         ret = io_schedule_timeout(timeout);
 916         finish_wait(wqh, &wait);
 917
 918         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
 919                                         jiffies_to_usecs(jiffies - start));
 920
 921         return ret;
 922 }
 923 EXPORT_SYMBOL(congestion_wait);
 924
 925 /**
 926  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
 927  * @zone: A zone to check if it is heavily congested
 928  * @sync: SYNC or ASYNC IO
 929  * @timeout: timeout in jiffies
 930  *
 931  * In the event of a congested backing_dev (any backing_dev) and the given
 932  * @zone has experienced recent congestion, this waits for up to @timeout
 933  * jiffies for either a BDI to exit congestion of the given @sync queue
 934  * or a write to complete.
 935  *
 936  * In the absence of zone congestion, cond_resched() is called to yield
 937  * the processor if necessary but otherwise does not sleep.
 938  *
 939  * The return value is 0 if the sleep is for the full timeout. Otherwise,
 940  * it is the number of jiffies that were still remaining when the function
 941  * returned. return_value == timeout implies the function did not sleep.
 942  */
 943 long wait_iff_congested(struct zone *zone, int sync, long timeout)
 944 {
 945         long ret;
 946         unsigned long start = jiffies;
 947         DEFINE_WAIT(wait);
 948         wait_queue_head_t *wqh = &congestion_wqh[sync];
 949
 950         /*
 951          * If there is no congestion, or heavy congestion is not being
 952          * encountered in the current zone, yield if necessary instead
 953          * of sleeping on the congestion queue
 954          */
 955         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
 956             !test_bit(ZONE_CONGESTED, &zone->flags)) {
 957                 cond_resched();
 958
 959                 /* In case we scheduled, work out time remaining */
 960                 ret = timeout - (jiffies - start);
 961                 if (ret < 0)
 962                         ret = 0;
 963
 964                 goto out;
 965         }
 966
 967         /* Sleep until uncongested or a write happens */
 968         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 969         ret = io_schedule_timeout(timeout);
 970         finish_wait(wqh, &wait);
 971
 972 out:
 973         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
 974                                         jiffies_to_usecs(jiffies - start));
 975
 976         return ret;
 977 }
 978 EXPORT_SYMBOL(wait_iff_congested);
 979
 980 int pdflush_proc_obsolete(struct ctl_table *table, int write,
 981                         void __user *buffer, size_t *lenp, loff_t *ppos)
 982 {
 983         char kbuf[] = "0\n";
 984
 985         if (*ppos || *lenp < sizeof(kbuf)) {
 986                 *lenp = 0;
 987                 return 0;
 988         }
 989
 990         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
 991                 return -EFAULT;
 992         printk_once(KERN_WARNING "%s exported in /proc is scheduled for removal\n",
 993                         table->procname);
 994
 995         *lenp = 2;
 996         *ppos += *lenp;
 997         return 2;
 998 }