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 const char *bdi_unknown_name = "(unknown)";
  25
  26 /*
  27  * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
  28  * locking.
  29  */
  30 DEFINE_SPINLOCK(bdi_lock);
  31 LIST_HEAD(bdi_list);
  32
  33 /* bdi_wq serves all asynchronous writeback tasks */
  34 struct workqueue_struct *bdi_wq;
  35
  36 #ifdef CONFIG_DEBUG_FS
  37 #include <linux/debugfs.h>
  38 #include <linux/seq_file.h>
  39
  40 static struct dentry *bdi_debug_root;
  41
  42 static void bdi_debug_init(void)
  43 {
  44         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  45 }
  46
  47 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  48 {
  49         struct backing_dev_info *bdi = m->private;
  50         struct bdi_writeback *wb = &bdi->wb;
  51         unsigned long background_thresh;
  52         unsigned long dirty_thresh;
  53         unsigned long wb_thresh;
  54         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  55         struct inode *inode;
  56
  57         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  58         spin_lock(&wb->list_lock);
  59         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  60                 nr_dirty++;
  61         list_for_each_entry(inode, &wb->b_io, i_io_list)
  62                 nr_io++;
  63         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  64                 nr_more_io++;
  65         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  66                 if (inode->i_state & I_DIRTY_TIME)
  67                         nr_dirty_time++;
  68         spin_unlock(&wb->list_lock);
  69
  70         global_dirty_limits(&background_thresh, &dirty_thresh);
  71         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  72
  73 #define K(x) ((x) << (PAGE_SHIFT - 10))
  74         seq_printf(m,
  75                    "BdiWriteback:       %10lu kB\n"
  76                    "BdiReclaimable:     %10lu kB\n"
  77                    "BdiDirtyThresh:     %10lu kB\n"
  78                    "DirtyThresh:        %10lu kB\n"
  79                    "BackgroundThresh:   %10lu kB\n"
  80                    "BdiDirtied:         %10lu kB\n"
  81                    "BdiWritten:         %10lu kB\n"
  82                    "BdiWriteBandwidth:  %10lu kBps\n"
  83                    "b_dirty:            %10lu\n"
  84                    "b_io:               %10lu\n"
  85                    "b_more_io:          %10lu\n"
  86                    "b_dirty_time:       %10lu\n"
  87                    "bdi_list:           %10u\n"
  88                    "state:              %10lx\n",
  89                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  90                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  91                    K(wb_thresh),
  92                    K(dirty_thresh),
  93                    K(background_thresh),
  94                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  95                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  96                    (unsigned long) K(wb->write_bandwidth),
  97                    nr_dirty,
  98                    nr_io,
  99                    nr_more_io,
 100                    nr_dirty_time,
 101                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 102 #undef K
 103
 104         return 0;
 105 }
 106
 107 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 108 {
 109         return single_open(file, bdi_debug_stats_show, inode->i_private);
 110 }
 111
 112 static const struct file_operations bdi_debug_stats_fops = {
 113         .open           = bdi_debug_stats_open,
 114         .read           = seq_read,
 115         .llseek         = seq_lseek,
 116         .release        = single_release,
 117 };
 118
 119 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 120 {
 121         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 122         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 123                                                bdi, &bdi_debug_stats_fops);
 124 }
 125
 126 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 127 {
 128         debugfs_remove(bdi->debug_stats);
 129         debugfs_remove(bdi->debug_dir);
 130 }
 131 #else
 132 static inline void bdi_debug_init(void)
 133 {
 134 }
 135 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 136                                       const char *name)
 137 {
 138 }
 139 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 140 {
 141 }
 142 #endif
 143
 144 static ssize_t read_ahead_kb_store(struct device *dev,
 145                                   struct device_attribute *attr,
 146                                   const char *buf, size_t count)
 147 {
 148         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 149         unsigned long read_ahead_kb;
 150         ssize_t ret;
 151
 152         ret = kstrtoul(buf, 10, &read_ahead_kb);
 153         if (ret < 0)
 154                 return ret;
 155
 156         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 157
 158         return count;
 159 }
 160
 161 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 162
 163 #define BDI_SHOW(name, expr)                                            \
 164 static ssize_t name##_show(struct device *dev,                          \
 165                            struct device_attribute *attr, char *page)   \
 166 {                                                                       \
 167         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 168                                                                         \
 169         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 170 }                                                                       \
 171 static DEVICE_ATTR_RW(name);
 172
 173 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 174
 175 static ssize_t min_ratio_store(struct device *dev,
 176                 struct device_attribute *attr, const char *buf, size_t count)
 177 {
 178         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 179         unsigned int ratio;
 180         ssize_t ret;
 181
 182         ret = kstrtouint(buf, 10, &ratio);
 183         if (ret < 0)
 184                 return ret;
 185
 186         ret = bdi_set_min_ratio(bdi, ratio);
 187         if (!ret)
 188                 ret = count;
 189
 190         return ret;
 191 }
 192 BDI_SHOW(min_ratio, bdi->min_ratio)
 193
 194 static ssize_t max_ratio_store(struct device *dev,
 195                 struct device_attribute *attr, const char *buf, size_t count)
 196 {
 197         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 198         unsigned int ratio;
 199         ssize_t ret;
 200
 201         ret = kstrtouint(buf, 10, &ratio);
 202         if (ret < 0)
 203                 return ret;
 204
 205         ret = bdi_set_max_ratio(bdi, ratio);
 206         if (!ret)
 207                 ret = count;
 208
 209         return ret;
 210 }
 211 BDI_SHOW(max_ratio, bdi->max_ratio)
 212
 213 static ssize_t stable_pages_required_show(struct device *dev,
 214                                           struct device_attribute *attr,
 215                                           char *page)
 216 {
 217         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 218
 219         return snprintf(page, PAGE_SIZE-1, "%d\n",
 220                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 221 }
 222 static DEVICE_ATTR_RO(stable_pages_required);
 223
 224 static struct attribute *bdi_dev_attrs[] = {
 225         &dev_attr_read_ahead_kb.attr,
 226         &dev_attr_min_ratio.attr,
 227         &dev_attr_max_ratio.attr,
 228         &dev_attr_stable_pages_required.attr,
 229         NULL,
 230 };
 231 ATTRIBUTE_GROUPS(bdi_dev);
 232
 233 static __init int bdi_class_init(void)
 234 {
 235         bdi_class = class_create(THIS_MODULE, "bdi");
 236         if (IS_ERR(bdi_class))
 237                 return PTR_ERR(bdi_class);
 238
 239         bdi_class->dev_groups = bdi_dev_groups;
 240         bdi_debug_init();
 241         return 0;
 242 }
 243 postcore_initcall(bdi_class_init);
 244
 245 static int __init default_bdi_init(void)
 246 {
 247         int err;
 248
 249         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
 250                                               WQ_UNBOUND | WQ_SYSFS, 0);
 251         if (!bdi_wq)
 252                 return -ENOMEM;
 253
 254         err = bdi_init(&noop_backing_dev_info);
 255
 256         return err;
 257 }
 258 subsys_initcall(default_bdi_init);
 259
 260 /*
 261  * This function is used when the first inode for this wb is marked dirty. It
 262  * wakes-up the corresponding bdi thread which should then take care of the
 263  * periodic background write-out of dirty inodes. Since the write-out would
 264  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 265  * set up a timer which wakes the bdi thread up later.
 266  *
 267  * Note, we wouldn't bother setting up the timer, but this function is on the
 268  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 269  * by delaying the wake-up.
 270  *
 271  * We have to be careful not to postpone flush work if it is scheduled for
 272  * earlier. Thus we use queue_delayed_work().
 273  */
 274 void wb_wakeup_delayed(struct bdi_writeback *wb)
 275 {
 276         unsigned long timeout;
 277
 278         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 279         spin_lock_bh(&wb->work_lock);
 280         if (test_bit(WB_registered, &wb->state))
 281                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 282         spin_unlock_bh(&wb->work_lock);
 283 }
 284
 285 /*
 286  * Initial write bandwidth: 100 MB/s
 287  */
 288 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 289
 290 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 291                    int blkcg_id, gfp_t gfp)
 292 {
 293         int i, err;
 294
 295         memset(wb, 0, sizeof(*wb));
 296
 297         wb->bdi = bdi;
 298         wb->last_old_flush = jiffies;
 299         INIT_LIST_HEAD(&wb->b_dirty);
 300         INIT_LIST_HEAD(&wb->b_io);
 301         INIT_LIST_HEAD(&wb->b_more_io);
 302         INIT_LIST_HEAD(&wb->b_dirty_time);
 303         spin_lock_init(&wb->list_lock);
 304
 305         wb->bw_time_stamp = jiffies;
 306         wb->balanced_dirty_ratelimit = INIT_BW;
 307         wb->dirty_ratelimit = INIT_BW;
 308         wb->write_bandwidth = INIT_BW;
 309         wb->avg_write_bandwidth = INIT_BW;
 310
 311         spin_lock_init(&wb->work_lock);
 312         INIT_LIST_HEAD(&wb->work_list);
 313         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 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         init_rwsem(&bdi->wb_switch_rwsem);
 674
 675         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 676         if (!ret) {
 677                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 678                 bdi->wb.blkcg_css = blkcg_root_css;
 679         }
 680         return ret;
 681 }
 682
 683 static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
 684 {
 685         struct radix_tree_iter iter;
 686         struct rb_node *rbn;
 687         void **slot;
 688
 689         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 690
 691         spin_lock_irq(&cgwb_lock);
 692
 693         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 694                 cgwb_kill(*slot);
 695
 696         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 697                 struct bdi_writeback_congested *congested =
 698                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 699
 700                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 701                 congested->bdi = NULL;  /* mark @congested unlinked */
 702         }
 703
 704         spin_unlock_irq(&cgwb_lock);
 705
 706         /*
 707          * All cgwb's and their congested states must be shutdown and
 708          * released before returning.  Drain the usage counter to wait for
 709          * all cgwb's and cgwb_congested's ever created on @bdi.
 710          */
 711         atomic_dec(&bdi->usage_cnt);
 712         wait_event(cgwb_release_wait, !atomic_read(&bdi->usage_cnt));
 713 }
 714
 715 /**
 716  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 717  * @memcg: memcg being offlined
 718  *
 719  * Also prevents creation of any new wb's associated with @memcg.
 720  */
 721 void wb_memcg_offline(struct mem_cgroup *memcg)
 722 {
 723         LIST_HEAD(to_destroy);
 724         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 725         struct bdi_writeback *wb, *next;
 726
 727         spin_lock_irq(&cgwb_lock);
 728         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 729                 cgwb_kill(wb);
 730         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 731         spin_unlock_irq(&cgwb_lock);
 732 }
 733
 734 /**
 735  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 736  * @blkcg: blkcg being offlined
 737  *
 738  * Also prevents creation of any new wb's associated with @blkcg.
 739  */
 740 void wb_blkcg_offline(struct blkcg *blkcg)
 741 {
 742         LIST_HEAD(to_destroy);
 743         struct bdi_writeback *wb, *next;
 744
 745         spin_lock_irq(&cgwb_lock);
 746         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 747                 cgwb_kill(wb);
 748         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 749         spin_unlock_irq(&cgwb_lock);
 750 }
 751
 752 #else   /* CONFIG_CGROUP_WRITEBACK */
 753
 754 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 755 {
 756         int err;
 757
 758         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 759         if (!bdi->wb_congested)
 760                 return -ENOMEM;
 761
 762         atomic_set(&bdi->wb_congested->refcnt, 1);
 763
 764         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 765         if (err) {
 766                 wb_congested_put(bdi->wb_congested);
 767                 return err;
 768         }
 769         return 0;
 770 }
 771
 772 static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
 773 {
 774         wb_congested_put(bdi->wb_congested);
 775 }
 776
 777 #endif  /* CONFIG_CGROUP_WRITEBACK */
 778
 779 int bdi_init(struct backing_dev_info *bdi)
 780 {
 781         int ret;
 782
 783         bdi->dev = NULL;
 784
 785         bdi->min_ratio = 0;
 786         bdi->max_ratio = 100;
 787         bdi->max_prop_frac = FPROP_FRAC_BASE;
 788         INIT_LIST_HEAD(&bdi->bdi_list);
 789         INIT_LIST_HEAD(&bdi->wb_list);
 790         init_waitqueue_head(&bdi->wb_waitq);
 791
 792         ret = cgwb_bdi_init(bdi);
 793
 794         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 795
 796         return ret;
 797 }
 798 EXPORT_SYMBOL(bdi_init);
 799
 800 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 801                 const char *fmt, ...)
 802 {
 803         va_list args;
 804         struct device *dev;
 805
 806         if (bdi->dev)   /* The driver needs to use separate queues per device */
 807                 return 0;
 808
 809         va_start(args, fmt);
 810         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 811         va_end(args);
 812         if (IS_ERR(dev))
 813                 return PTR_ERR(dev);
 814
 815         bdi->dev = dev;
 816
 817         bdi_debug_register(bdi, dev_name(dev));
 818         set_bit(WB_registered, &bdi->wb.state);
 819
 820         spin_lock_bh(&bdi_lock);
 821         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 822         spin_unlock_bh(&bdi_lock);
 823
 824         trace_writeback_bdi_register(bdi);
 825         return 0;
 826 }
 827 EXPORT_SYMBOL(bdi_register);
 828
 829 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 830 {
 831         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 832 }
 833 EXPORT_SYMBOL(bdi_register_dev);
 834
 835 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 836 {
 837         int rc;
 838
 839         rc = bdi_register(bdi, NULL, "%u:%u", MAJOR(owner->devt),
 840                         MINOR(owner->devt));
 841         if (rc)
 842                 return rc;
 843         bdi->owner = owner;
 844         get_device(owner);
 845         return 0;
 846 }
 847 EXPORT_SYMBOL(bdi_register_owner);
 848
 849 /*
 850  * Remove bdi from bdi_list, and ensure that it is no longer visible
 851  */
 852 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 853 {
 854         spin_lock_bh(&bdi_lock);
 855         list_del_rcu(&bdi->bdi_list);
 856         spin_unlock_bh(&bdi_lock);
 857
 858         synchronize_rcu_expedited();
 859 }
 860
 861 void bdi_unregister(struct backing_dev_info *bdi)
 862 {
 863         /* make sure nobody finds us on the bdi_list anymore */
 864         bdi_remove_from_list(bdi);
 865         wb_shutdown(&bdi->wb);
 866         cgwb_bdi_destroy(bdi);
 867
 868         /*
 869          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
 870          * update the global bdi_min_ratio.
 871          */
 872         if (bdi->min_ratio)
 873                 bdi_set_min_ratio(bdi, 0);
 874
 875         if (bdi->dev) {
 876                 bdi_debug_unregister(bdi);
 877                 device_unregister(bdi->dev);
 878                 bdi->dev = NULL;
 879         }
 880
 881         if (bdi->owner) {
 882                 put_device(bdi->owner);
 883                 bdi->owner = NULL;
 884         }
 885 }
 886
 887 void bdi_exit(struct backing_dev_info *bdi)
 888 {
 889         WARN_ON_ONCE(bdi->dev);
 890         wb_exit(&bdi->wb);
 891 }
 892
 893 void bdi_destroy(struct backing_dev_info *bdi)
 894 {
 895         bdi_unregister(bdi);
 896         bdi_exit(bdi);
 897 }
 898 EXPORT_SYMBOL(bdi_destroy);
 899
 900 /*
 901  * For use from filesystems to quickly init and register a bdi associated
 902  * with dirty writeback
 903  */
 904 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
 905 {
 906         int err;
 907
 908         bdi->name = name;
 909         bdi->capabilities = 0;
 910         err = bdi_init(bdi);
 911         if (err)
 912                 return err;
 913
 914         err = bdi_register(bdi, NULL, "%.28s-%ld", name,
 915                            atomic_long_inc_return(&bdi_seq));
 916         if (err) {
 917                 bdi_destroy(bdi);
 918                 return err;
 919         }
 920
 921         return 0;
 922 }
 923 EXPORT_SYMBOL(bdi_setup_and_register);
 924
 925 static wait_queue_head_t congestion_wqh[2] = {
 926                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 927                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 928         };
 929 static atomic_t nr_wb_congested[2];
 930
 931 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 932 {
 933         wait_queue_head_t *wqh = &congestion_wqh[sync];
 934         enum wb_congested_state bit;
 935
 936         bit = sync ? WB_sync_congested : WB_async_congested;
 937         if (test_and_clear_bit(bit, &congested->state))
 938                 atomic_dec(&nr_wb_congested[sync]);
 939         smp_mb__after_atomic();
 940         if (waitqueue_active(wqh))
 941                 wake_up(wqh);
 942 }
 943 EXPORT_SYMBOL(clear_wb_congested);
 944
 945 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
 946 {
 947         enum wb_congested_state bit;
 948
 949         bit = sync ? WB_sync_congested : WB_async_congested;
 950         if (!test_and_set_bit(bit, &congested->state))
 951                 atomic_inc(&nr_wb_congested[sync]);
 952 }
 953 EXPORT_SYMBOL(set_wb_congested);
 954
 955 /**
 956  * congestion_wait - wait for a backing_dev to become uncongested
 957  * @sync: SYNC or ASYNC IO
 958  * @timeout: timeout in jiffies
 959  *
 960  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 961  * write congestion.  If no backing_devs are congested then just wait for the
 962  * next write to be completed.
 963  */
 964 long congestion_wait(int sync, long timeout)
 965 {
 966         long ret;
 967         unsigned long start = jiffies;
 968         DEFINE_WAIT(wait);
 969         wait_queue_head_t *wqh = &congestion_wqh[sync];
 970
 971         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 972         ret = io_schedule_timeout(timeout);
 973         finish_wait(wqh, &wait);
 974
 975         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
 976                                         jiffies_to_usecs(jiffies - start));
 977
 978         return ret;
 979 }
 980 EXPORT_SYMBOL(congestion_wait);
 981
 982 /**
 983  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
 984  * @pgdat: A pgdat to check if it is heavily congested
 985  * @sync: SYNC or ASYNC IO
 986  * @timeout: timeout in jiffies
 987  *
 988  * In the event of a congested backing_dev (any backing_dev) and the given
 989  * @pgdat has experienced recent congestion, this waits for up to @timeout
 990  * jiffies for either a BDI to exit congestion of the given @sync queue
 991  * or a write to complete.
 992  *
 993  * In the absence of pgdat congestion, cond_resched() is called to yield
 994  * the processor if necessary but otherwise does not sleep.
 995  *
 996  * The return value is 0 if the sleep is for the full timeout. Otherwise,
 997  * it is the number of jiffies that were still remaining when the function
 998  * returned. return_value == timeout implies the function did not sleep.
 999  */
1000 long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
1001 {
1002         long ret;
1003         unsigned long start = jiffies;
1004         DEFINE_WAIT(wait);
1005         wait_queue_head_t *wqh = &congestion_wqh[sync];
1006
1007         /*
1008          * If there is no congestion, or heavy congestion is not being
1009          * encountered in the current pgdat, yield if necessary instead
1010          * of sleeping on the congestion queue
1011          */
1012         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1013             !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1014                 cond_resched();
1015
1016                 /* In case we scheduled, work out time remaining */
1017                 ret = timeout - (jiffies - start);
1018                 if (ret < 0)
1019                         ret = 0;
1020
1021                 goto out;
1022         }
1023
1024         /* Sleep until uncongested or a write happens */
1025         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1026         ret = io_schedule_timeout(timeout);
1027         finish_wait(wqh, &wait);
1028
1029 out:
1030         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1031                                         jiffies_to_usecs(jiffies - start));
1032
1033         return ret;
1034 }
1035 EXPORT_SYMBOL(wait_iff_congested);
1036
1037 int pdflush_proc_obsolete(struct ctl_table *table, int write,
1038                         void __user *buffer, size_t *lenp, loff_t *ppos)
1039 {
1040         char kbuf[] = "0\n";
1041
1042         if (*ppos || *lenp < sizeof(kbuf)) {
1043                 *lenp = 0;
1044                 return 0;
1045         }
1046
1047         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1048                 return -EFAULT;
1049         pr_warn_once("%s exported in /proc is scheduled for removal\n",
1050                      table->procname);
1051
1052         *lenp = 2;
1053         *ppos += *lenp;
1054         return 2;
1055 }