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_io_list)
  59                 nr_dirty++;
  60         list_for_each_entry(inode, &wb->b_io, i_io_list)
  61                 nr_io++;
  62         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  63                 nr_more_io++;
  64         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  65                 if (inode->i_state & I_DIRTY_TIME)
  66                         nr_dirty_time++;
  67         spin_unlock(&wb->list_lock);
  68
  69         global_dirty_limits(&background_thresh, &dirty_thresh);
  70         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  71
  72 #define K(x) ((x) << (PAGE_SHIFT - 10))
  73         seq_printf(m,
  74                    "BdiWriteback:       %10lu kB\n"
  75                    "BdiReclaimable:     %10lu kB\n"
  76                    "BdiDirtyThresh:     %10lu kB\n"
  77                    "DirtyThresh:        %10lu kB\n"
  78                    "BackgroundThresh:   %10lu kB\n"
  79                    "BdiDirtied:         %10lu kB\n"
  80                    "BdiWritten:         %10lu kB\n"
  81                    "BdiWriteBandwidth:  %10lu kBps\n"
  82                    "b_dirty:            %10lu\n"
  83                    "b_io:               %10lu\n"
  84                    "b_more_io:          %10lu\n"
  85                    "b_dirty_time:       %10lu\n"
  86                    "bdi_list:           %10u\n"
  87                    "state:              %10lx\n",
  88                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  89                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  90                    K(wb_thresh),
  91                    K(dirty_thresh),
  92                    K(background_thresh),
  93                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  94                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  95                    (unsigned long) K(wb->write_bandwidth),
  96                    nr_dirty,
  97                    nr_io,
  98                    nr_more_io,
  99                    nr_dirty_time,
 100                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 101 #undef K
 102
 103         return 0;
 104 }
 105
 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                    int blkcg_id, 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         wb->dirty_sleep = jiffies;
 314
 315         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 316         if (!wb->congested)
 317                 return -ENOMEM;
 318
 319         err = fprop_local_init_percpu(&wb->completions, gfp);
 320         if (err)
 321                 goto out_put_cong;
 322
 323         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 324                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 325                 if (err)
 326                         goto out_destroy_stat;
 327         }
 328
 329         return 0;
 330
 331 out_destroy_stat:
 332         while (i--)
 333                 percpu_counter_destroy(&wb->stat[i]);
 334         fprop_local_destroy_percpu(&wb->completions);
 335 out_put_cong:
 336         wb_congested_put(wb->congested);
 337         return err;
 338 }
 339
 340 /*
 341  * Remove bdi from the global list and shutdown any threads we have running
 342  */
 343 static void wb_shutdown(struct bdi_writeback *wb)
 344 {
 345         /* Make sure nobody queues further work */
 346         spin_lock_bh(&wb->work_lock);
 347         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 348                 spin_unlock_bh(&wb->work_lock);
 349                 return;
 350         }
 351         spin_unlock_bh(&wb->work_lock);
 352
 353         /*
 354          * Drain work list and shutdown the delayed_work.  !WB_registered
 355          * tells wb_workfn() that @wb is dying and its work_list needs to
 356          * be drained no matter what.
 357          */
 358         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 359         flush_delayed_work(&wb->dwork);
 360         WARN_ON(!list_empty(&wb->work_list));
 361 }
 362
 363 static void wb_exit(struct bdi_writeback *wb)
 364 {
 365         int i;
 366
 367         WARN_ON(delayed_work_pending(&wb->dwork));
 368
 369         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 370                 percpu_counter_destroy(&wb->stat[i]);
 371
 372         fprop_local_destroy_percpu(&wb->completions);
 373         wb_congested_put(wb->congested);
 374 }
 375
 376 #ifdef CONFIG_CGROUP_WRITEBACK
 377
 378 #include <linux/memcontrol.h>
 379
 380 /*
 381  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 382  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 383  * protected.  cgwb_release_wait is used to wait for the completion of cgwb
 384  * releases from bdi destruction path.
 385  */
 386 static DEFINE_SPINLOCK(cgwb_lock);
 387 static DECLARE_WAIT_QUEUE_HEAD(cgwb_release_wait);
 388
 389 /**
 390  * wb_congested_get_create - get or create a wb_congested
 391  * @bdi: associated bdi
 392  * @blkcg_id: ID of the associated blkcg
 393  * @gfp: allocation mask
 394  *
 395  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 396  * The returned wb_congested has its reference count incremented.  Returns
 397  * NULL on failure.
 398  */
 399 struct bdi_writeback_congested *
 400 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 401 {
 402         struct bdi_writeback_congested *new_congested = NULL, *congested;
 403         struct rb_node **node, *parent;
 404         unsigned long flags;
 405 retry:
 406         spin_lock_irqsave(&cgwb_lock, flags);
 407
 408         node = &bdi->cgwb_congested_tree.rb_node;
 409         parent = NULL;
 410
 411         while (*node != NULL) {
 412                 parent = *node;
 413                 congested = container_of(parent, struct bdi_writeback_congested,
 414                                          rb_node);
 415                 if (congested->blkcg_id < blkcg_id)
 416                         node = &parent->rb_left;
 417                 else if (congested->blkcg_id > blkcg_id)
 418                         node = &parent->rb_right;
 419                 else
 420                         goto found;
 421         }
 422
 423         if (new_congested) {
 424                 /* !found and storage for new one already allocated, insert */
 425                 congested = new_congested;
 426                 new_congested = NULL;
 427                 rb_link_node(&congested->rb_node, parent, node);
 428                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 429                 goto found;
 430         }
 431
 432         spin_unlock_irqrestore(&cgwb_lock, flags);
 433
 434         /* allocate storage for new one and retry */
 435         new_congested = kzalloc(sizeof(*new_congested), gfp);
 436         if (!new_congested)
 437                 return NULL;
 438
 439         atomic_set(&new_congested->refcnt, 0);
 440         new_congested->bdi = bdi;
 441         new_congested->blkcg_id = blkcg_id;
 442         goto retry;
 443
 444 found:
 445         atomic_inc(&congested->refcnt);
 446         spin_unlock_irqrestore(&cgwb_lock, flags);
 447         kfree(new_congested);
 448         return congested;
 449 }
 450
 451 /**
 452  * wb_congested_put - put a wb_congested
 453  * @congested: wb_congested to put
 454  *
 455  * Put @congested and destroy it if the refcnt reaches zero.
 456  */
 457 void wb_congested_put(struct bdi_writeback_congested *congested)
 458 {
 459         unsigned long flags;
 460
 461         local_irq_save(flags);
 462         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 463                 local_irq_restore(flags);
 464                 return;
 465         }
 466
 467         /* bdi might already have been destroyed leaving @congested unlinked */
 468         if (congested->bdi) {
 469                 rb_erase(&congested->rb_node,
 470                          &congested->bdi->cgwb_congested_tree);
 471                 congested->bdi = NULL;
 472         }
 473
 474         spin_unlock_irqrestore(&cgwb_lock, flags);
 475         kfree(congested);
 476 }
 477
 478 static void cgwb_release_workfn(struct work_struct *work)
 479 {
 480         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 481                                                 release_work);
 482         struct backing_dev_info *bdi = wb->bdi;
 483
 484         spin_lock_irq(&cgwb_lock);
 485         list_del_rcu(&wb->bdi_node);
 486         spin_unlock_irq(&cgwb_lock);
 487
 488         wb_shutdown(wb);
 489
 490         css_put(wb->memcg_css);
 491         css_put(wb->blkcg_css);
 492
 493         fprop_local_destroy_percpu(&wb->memcg_completions);
 494         percpu_ref_exit(&wb->refcnt);
 495         wb_exit(wb);
 496         kfree_rcu(wb, rcu);
 497
 498         if (atomic_dec_and_test(&bdi->usage_cnt))
 499                 wake_up_all(&cgwb_release_wait);
 500 }
 501
 502 static void cgwb_release(struct percpu_ref *refcnt)
 503 {
 504         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 505                                                 refcnt);
 506         schedule_work(&wb->release_work);
 507 }
 508
 509 static void cgwb_kill(struct bdi_writeback *wb)
 510 {
 511         lockdep_assert_held(&cgwb_lock);
 512
 513         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 514         list_del(&wb->memcg_node);
 515         list_del(&wb->blkcg_node);
 516         percpu_ref_kill(&wb->refcnt);
 517 }
 518
 519 static int cgwb_create(struct backing_dev_info *bdi,
 520                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 521 {
 522         struct mem_cgroup *memcg;
 523         struct cgroup_subsys_state *blkcg_css;
 524         struct blkcg *blkcg;
 525         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 526         struct bdi_writeback *wb;
 527         unsigned long flags;
 528         int ret = 0;
 529
 530         memcg = mem_cgroup_from_css(memcg_css);
 531         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 532         blkcg = css_to_blkcg(blkcg_css);
 533         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 534         blkcg_cgwb_list = &blkcg->cgwb_list;
 535
 536         /* look up again under lock and discard on blkcg mismatch */
 537         spin_lock_irqsave(&cgwb_lock, flags);
 538         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 539         if (wb && wb->blkcg_css != blkcg_css) {
 540                 cgwb_kill(wb);
 541                 wb = NULL;
 542         }
 543         spin_unlock_irqrestore(&cgwb_lock, flags);
 544         if (wb)
 545                 goto out_put;
 546
 547         /* need to create a new one */
 548         wb = kmalloc(sizeof(*wb), gfp);
 549         if (!wb)
 550                 return -ENOMEM;
 551
 552         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 553         if (ret)
 554                 goto err_free;
 555
 556         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 557         if (ret)
 558                 goto err_wb_exit;
 559
 560         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 561         if (ret)
 562                 goto err_ref_exit;
 563
 564         wb->memcg_css = memcg_css;
 565         wb->blkcg_css = blkcg_css;
 566         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 567         set_bit(WB_registered, &wb->state);
 568
 569         /*
 570          * The root wb determines the registered state of the whole bdi and
 571          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 572          * whether they're still online.  Don't link @wb if any is dead.
 573          * See wb_memcg_offline() and wb_blkcg_offline().
 574          */
 575         ret = -ENODEV;
 576         spin_lock_irqsave(&cgwb_lock, flags);
 577         if (test_bit(WB_registered, &bdi->wb.state) &&
 578             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 579                 /* we might have raced another instance of this function */
 580                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 581                 if (!ret) {
 582                         atomic_inc(&bdi->usage_cnt);
 583                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 584                         list_add(&wb->memcg_node, memcg_cgwb_list);
 585                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 586                         css_get(memcg_css);
 587                         css_get(blkcg_css);
 588                 }
 589         }
 590         spin_unlock_irqrestore(&cgwb_lock, flags);
 591         if (ret) {
 592                 if (ret == -EEXIST)
 593                         ret = 0;
 594                 goto err_fprop_exit;
 595         }
 596         goto out_put;
 597
 598 err_fprop_exit:
 599         fprop_local_destroy_percpu(&wb->memcg_completions);
 600 err_ref_exit:
 601         percpu_ref_exit(&wb->refcnt);
 602 err_wb_exit:
 603         wb_exit(wb);
 604 err_free:
 605         kfree(wb);
 606 out_put:
 607         css_put(blkcg_css);
 608         return ret;
 609 }
 610
 611 /**
 612  * wb_get_create - get wb for a given memcg, create if necessary
 613  * @bdi: target bdi
 614  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 615  * @gfp: allocation mask to use
 616  *
 617  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 618  * create one.  The returned wb has its refcount incremented.
 619  *
 620  * This function uses css_get() on @memcg_css and thus expects its refcnt
 621  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 622  * @memcg_css isn't enough.  try_get it before calling this function.
 623  *
 624  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 625  * memcg on the default hierarchy, memcg association is guaranteed to be
 626  * more specific (equal or descendant to the associated blkcg) and thus can
 627  * identify both the memcg and blkcg associations.
 628  *
 629  * Because the blkcg associated with a memcg may change as blkcg is enabled
 630  * and disabled closer to root in the hierarchy, each wb keeps track of
 631  * both the memcg and blkcg associated with it and verifies the blkcg on
 632  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 633  * created.
 634  */
 635 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 636                                     struct cgroup_subsys_state *memcg_css,
 637                                     gfp_t gfp)
 638 {
 639         struct bdi_writeback *wb;
 640
 641         might_sleep_if(gfpflags_allow_blocking(gfp));
 642
 643         if (!memcg_css->parent)
 644                 return &bdi->wb;
 645
 646         do {
 647                 rcu_read_lock();
 648                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 649                 if (wb) {
 650                         struct cgroup_subsys_state *blkcg_css;
 651
 652                         /* see whether the blkcg association has changed */
 653                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 654                                                      &io_cgrp_subsys);
 655                         if (unlikely(wb->blkcg_css != blkcg_css ||
 656                                      !wb_tryget(wb)))
 657                                 wb = NULL;
 658                         css_put(blkcg_css);
 659                 }
 660                 rcu_read_unlock();
 661         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 662
 663         return wb;
 664 }
 665
 666 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 667 {
 668         int ret;
 669
 670         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 671         bdi->cgwb_congested_tree = RB_ROOT;
 672         atomic_set(&bdi->usage_cnt, 1);
 673
 674         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 675         if (!ret) {
 676                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 677                 bdi->wb.blkcg_css = blkcg_root_css;
 678         }
 679         return ret;
 680 }
 681
 682 static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
 683 {
 684         struct radix_tree_iter iter;
 685         struct rb_node *rbn;
 686         void **slot;
 687
 688         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 689
 690         spin_lock_irq(&cgwb_lock);
 691
 692         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 693                 cgwb_kill(*slot);
 694
 695         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 696                 struct bdi_writeback_congested *congested =
 697                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 698
 699                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 700                 congested->bdi = NULL;  /* mark @congested unlinked */
 701         }
 702
 703         spin_unlock_irq(&cgwb_lock);
 704
 705         /*
 706          * All cgwb's and their congested states must be shutdown and
 707          * released before returning.  Drain the usage counter to wait for
 708          * all cgwb's and cgwb_congested's ever created on @bdi.
 709          */
 710         atomic_dec(&bdi->usage_cnt);
 711         wait_event(cgwb_release_wait, !atomic_read(&bdi->usage_cnt));
 712 }
 713
 714 /**
 715  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 716  * @memcg: memcg being offlined
 717  *
 718  * Also prevents creation of any new wb's associated with @memcg.
 719  */
 720 void wb_memcg_offline(struct mem_cgroup *memcg)
 721 {
 722         LIST_HEAD(to_destroy);
 723         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 724         struct bdi_writeback *wb, *next;
 725
 726         spin_lock_irq(&cgwb_lock);
 727         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 728                 cgwb_kill(wb);
 729         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 730         spin_unlock_irq(&cgwb_lock);
 731 }
 732
 733 /**
 734  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 735  * @blkcg: blkcg being offlined
 736  *
 737  * Also prevents creation of any new wb's associated with @blkcg.
 738  */
 739 void wb_blkcg_offline(struct blkcg *blkcg)
 740 {
 741         LIST_HEAD(to_destroy);
 742         struct bdi_writeback *wb, *next;
 743
 744         spin_lock_irq(&cgwb_lock);
 745         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 746                 cgwb_kill(wb);
 747         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 748         spin_unlock_irq(&cgwb_lock);
 749 }
 750
 751 #else   /* CONFIG_CGROUP_WRITEBACK */
 752
 753 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 754 {
 755         int err;
 756
 757         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 758         if (!bdi->wb_congested)
 759                 return -ENOMEM;
 760
 761         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 762         if (err) {
 763                 kfree(bdi->wb_congested);
 764                 return err;
 765         }
 766         return 0;
 767 }
 768
 769 static void cgwb_bdi_destroy(struct backing_dev_info *bdi) { }
 770
 771 #endif  /* CONFIG_CGROUP_WRITEBACK */
 772
 773 int bdi_init(struct backing_dev_info *bdi)
 774 {
 775         int ret;
 776
 777         bdi->dev = NULL;
 778
 779         bdi->min_ratio = 0;
 780         bdi->max_ratio = 100;
 781         bdi->max_prop_frac = FPROP_FRAC_BASE;
 782         INIT_LIST_HEAD(&bdi->bdi_list);
 783         INIT_LIST_HEAD(&bdi->wb_list);
 784         init_waitqueue_head(&bdi->wb_waitq);
 785
 786         ret = cgwb_bdi_init(bdi);
 787
 788         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 789
 790         return ret;
 791 }
 792 EXPORT_SYMBOL(bdi_init);
 793
 794 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 795                 const char *fmt, ...)
 796 {
 797         va_list args;
 798         struct device *dev;
 799
 800         if (bdi->dev)   /* The driver needs to use separate queues per device */
 801                 return 0;
 802
 803         va_start(args, fmt);
 804         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 805         va_end(args);
 806         if (IS_ERR(dev))
 807                 return PTR_ERR(dev);
 808
 809         bdi->dev = dev;
 810
 811         bdi_debug_register(bdi, dev_name(dev));
 812         set_bit(WB_registered, &bdi->wb.state);
 813
 814         spin_lock_bh(&bdi_lock);
 815         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 816         spin_unlock_bh(&bdi_lock);
 817
 818         trace_writeback_bdi_register(bdi);
 819         return 0;
 820 }
 821 EXPORT_SYMBOL(bdi_register);
 822
 823 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 824 {
 825         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 826 }
 827 EXPORT_SYMBOL(bdi_register_dev);
 828
 829 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 830 {
 831         int rc;
 832
 833         rc = bdi_register(bdi, NULL, "%u:%u", MAJOR(owner->devt),
 834                         MINOR(owner->devt));
 835         if (rc)
 836                 return rc;
 837         bdi->owner = owner;
 838         get_device(owner);
 839         return 0;
 840 }
 841 EXPORT_SYMBOL(bdi_register_owner);
 842
 843 /*
 844  * Remove bdi from bdi_list, and ensure that it is no longer visible
 845  */
 846 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 847 {
 848         spin_lock_bh(&bdi_lock);
 849         list_del_rcu(&bdi->bdi_list);
 850         spin_unlock_bh(&bdi_lock);
 851
 852         synchronize_rcu_expedited();
 853 }
 854
 855 void bdi_unregister(struct backing_dev_info *bdi)
 856 {
 857         /* make sure nobody finds us on the bdi_list anymore */
 858         bdi_remove_from_list(bdi);
 859         wb_shutdown(&bdi->wb);
 860         cgwb_bdi_destroy(bdi);
 861
 862         if (bdi->dev) {
 863                 bdi_debug_unregister(bdi);
 864                 device_unregister(bdi->dev);
 865                 bdi->dev = NULL;
 866         }
 867
 868         if (bdi->owner) {
 869                 put_device(bdi->owner);
 870                 bdi->owner = NULL;
 871         }
 872 }
 873
 874 void bdi_exit(struct backing_dev_info *bdi)
 875 {
 876         WARN_ON_ONCE(bdi->dev);
 877         wb_exit(&bdi->wb);
 878 }
 879
 880 void bdi_destroy(struct backing_dev_info *bdi)
 881 {
 882         bdi_unregister(bdi);
 883         bdi_exit(bdi);
 884 }
 885 EXPORT_SYMBOL(bdi_destroy);
 886
 887 /*
 888  * For use from filesystems to quickly init and register a bdi associated
 889  * with dirty writeback
 890  */
 891 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
 892 {
 893         int err;
 894
 895         bdi->name = name;
 896         bdi->capabilities = 0;
 897         err = bdi_init(bdi);
 898         if (err)
 899                 return err;
 900
 901         err = bdi_register(bdi, NULL, "%.28s-%ld", name,
 902                            atomic_long_inc_return(&bdi_seq));
 903         if (err) {
 904                 bdi_destroy(bdi);
 905                 return err;
 906         }
 907
 908         return 0;
 909 }
 910 EXPORT_SYMBOL(bdi_setup_and_register);
 911
 912 static wait_queue_head_t congestion_wqh[2] = {
 913                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 914                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 915         };
 916 static atomic_t nr_wb_congested[2];
 917
 918 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 919 {
 920         wait_queue_head_t *wqh = &congestion_wqh[sync];
 921         enum wb_congested_state bit;
 922
 923         bit = sync ? WB_sync_congested : WB_async_congested;
 924         if (test_and_clear_bit(bit, &congested->state))
 925                 atomic_dec(&nr_wb_congested[sync]);
 926         smp_mb__after_atomic();
 927         if (waitqueue_active(wqh))
 928                 wake_up(wqh);
 929 }
 930 EXPORT_SYMBOL(clear_wb_congested);
 931
 932 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
 933 {
 934         enum wb_congested_state bit;
 935
 936         bit = sync ? WB_sync_congested : WB_async_congested;
 937         if (!test_and_set_bit(bit, &congested->state))
 938                 atomic_inc(&nr_wb_congested[sync]);
 939 }
 940 EXPORT_SYMBOL(set_wb_congested);
 941
 942 /**
 943  * congestion_wait - wait for a backing_dev to become uncongested
 944  * @sync: SYNC or ASYNC IO
 945  * @timeout: timeout in jiffies
 946  *
 947  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 948  * write congestion.  If no backing_devs are congested then just wait for the
 949  * next write to be completed.
 950  */
 951 long congestion_wait(int sync, long timeout)
 952 {
 953         long ret;
 954         unsigned long start = jiffies;
 955         DEFINE_WAIT(wait);
 956         wait_queue_head_t *wqh = &congestion_wqh[sync];
 957
 958         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 959         ret = io_schedule_timeout(timeout);
 960         finish_wait(wqh, &wait);
 961
 962         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
 963                                         jiffies_to_usecs(jiffies - start));
 964
 965         return ret;
 966 }
 967 EXPORT_SYMBOL(congestion_wait);
 968
 969 /**
 970  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
 971  * @pgdat: A pgdat to check if it is heavily congested
 972  * @sync: SYNC or ASYNC IO
 973  * @timeout: timeout in jiffies
 974  *
 975  * In the event of a congested backing_dev (any backing_dev) and the given
 976  * @pgdat has experienced recent congestion, this waits for up to @timeout
 977  * jiffies for either a BDI to exit congestion of the given @sync queue
 978  * or a write to complete.
 979  *
 980  * In the absence of pgdat congestion, cond_resched() is called to yield
 981  * the processor if necessary but otherwise does not sleep.
 982  *
 983  * The return value is 0 if the sleep is for the full timeout. Otherwise,
 984  * it is the number of jiffies that were still remaining when the function
 985  * returned. return_value == timeout implies the function did not sleep.
 986  */
 987 long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
 988 {
 989         long ret;
 990         unsigned long start = jiffies;
 991         DEFINE_WAIT(wait);
 992         wait_queue_head_t *wqh = &congestion_wqh[sync];
 993
 994         /*
 995          * If there is no congestion, or heavy congestion is not being
 996          * encountered in the current pgdat, yield if necessary instead
 997          * of sleeping on the congestion queue
 998          */
 999         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1000             !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1001                 cond_resched();
1002
1003                 /* In case we scheduled, work out time remaining */
1004                 ret = timeout - (jiffies - start);
1005                 if (ret < 0)
1006                         ret = 0;
1007
1008                 goto out;
1009         }
1010
1011         /* Sleep until uncongested or a write happens */
1012         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1013         ret = io_schedule_timeout(timeout);
1014         finish_wait(wqh, &wait);
1015
1016 out:
1017         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1018                                         jiffies_to_usecs(jiffies - start));
1019
1020         return ret;
1021 }
1022 EXPORT_SYMBOL(wait_iff_congested);
1023
1024 int pdflush_proc_obsolete(struct ctl_table *table, int write,
1025                         void __user *buffer, size_t *lenp, loff_t *ppos)
1026 {
1027         char kbuf[] = "0\n";
1028
1029         if (*ppos || *lenp < sizeof(kbuf)) {
1030                 *lenp = 0;
1031                 return 0;
1032         }
1033
1034         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1035                 return -EFAULT;
1036         pr_warn_once("%s exported in /proc is scheduled for removal\n",
1037                      table->procname);
1038
1039         *lenp = 2;
1040         *ppos += *lenp;
1041         return 2;
1042 }