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
 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         if (wb != &bdi->wb)
 298                 bdi_get(bdi);
 299         wb->bdi = bdi;
 300         wb->last_old_flush = jiffies;
 301         INIT_LIST_HEAD(&wb->b_dirty);
 302         INIT_LIST_HEAD(&wb->b_io);
 303         INIT_LIST_HEAD(&wb->b_more_io);
 304         INIT_LIST_HEAD(&wb->b_dirty_time);
 305         spin_lock_init(&wb->list_lock);
 306
 307         wb->bw_time_stamp = jiffies;
 308         wb->balanced_dirty_ratelimit = INIT_BW;
 309         wb->dirty_ratelimit = INIT_BW;
 310         wb->write_bandwidth = INIT_BW;
 311         wb->avg_write_bandwidth = INIT_BW;
 312
 313         spin_lock_init(&wb->work_lock);
 314         INIT_LIST_HEAD(&wb->work_list);
 315         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 316         wb->dirty_sleep = jiffies;
 317
 318         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 319         if (!wb->congested) {
 320                 err = -ENOMEM;
 321                 goto out_put_bdi;
 322         }
 323
 324         err = fprop_local_init_percpu(&wb->completions, gfp);
 325         if (err)
 326                 goto out_put_cong;
 327
 328         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 329                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 330                 if (err)
 331                         goto out_destroy_stat;
 332         }
 333
 334         return 0;
 335
 336 out_destroy_stat:
 337         while (i--)
 338                 percpu_counter_destroy(&wb->stat[i]);
 339         fprop_local_destroy_percpu(&wb->completions);
 340 out_put_cong:
 341         wb_congested_put(wb->congested);
 342 out_put_bdi:
 343         if (wb != &bdi->wb)
 344                 bdi_put(bdi);
 345         return err;
 346 }
 347
 348 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 349
 350 /*
 351  * Remove bdi from the global list and shutdown any threads we have running
 352  */
 353 static void wb_shutdown(struct bdi_writeback *wb)
 354 {
 355         /* Make sure nobody queues further work */
 356         spin_lock_bh(&wb->work_lock);
 357         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 358                 spin_unlock_bh(&wb->work_lock);
 359                 return;
 360         }
 361         spin_unlock_bh(&wb->work_lock);
 362
 363         cgwb_remove_from_bdi_list(wb);
 364         /*
 365          * Drain work list and shutdown the delayed_work.  !WB_registered
 366          * tells wb_workfn() that @wb is dying and its work_list needs to
 367          * be drained no matter what.
 368          */
 369         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 370         flush_delayed_work(&wb->dwork);
 371         WARN_ON(!list_empty(&wb->work_list));
 372 }
 373
 374 static void wb_exit(struct bdi_writeback *wb)
 375 {
 376         int i;
 377
 378         WARN_ON(delayed_work_pending(&wb->dwork));
 379
 380         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 381                 percpu_counter_destroy(&wb->stat[i]);
 382
 383         fprop_local_destroy_percpu(&wb->completions);
 384         wb_congested_put(wb->congested);
 385         if (wb != &wb->bdi->wb)
 386                 bdi_put(wb->bdi);
 387 }
 388
 389 #ifdef CONFIG_CGROUP_WRITEBACK
 390
 391 #include <linux/memcontrol.h>
 392
 393 /*
 394  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 395  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 396  * protected.  cgwb_release_wait is used to wait for the completion of cgwb
 397  * releases from bdi destruction path.
 398  */
 399 static DEFINE_SPINLOCK(cgwb_lock);
 400 static DECLARE_WAIT_QUEUE_HEAD(cgwb_release_wait);
 401
 402 /**
 403  * wb_congested_get_create - get or create a wb_congested
 404  * @bdi: associated bdi
 405  * @blkcg_id: ID of the associated blkcg
 406  * @gfp: allocation mask
 407  *
 408  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 409  * The returned wb_congested has its reference count incremented.  Returns
 410  * NULL on failure.
 411  */
 412 struct bdi_writeback_congested *
 413 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 414 {
 415         struct bdi_writeback_congested *new_congested = NULL, *congested;
 416         struct rb_node **node, *parent;
 417         unsigned long flags;
 418 retry:
 419         spin_lock_irqsave(&cgwb_lock, flags);
 420
 421         node = &bdi->cgwb_congested_tree.rb_node;
 422         parent = NULL;
 423
 424         while (*node != NULL) {
 425                 parent = *node;
 426                 congested = rb_entry(parent, struct bdi_writeback_congested,
 427                                      rb_node);
 428                 if (congested->blkcg_id < blkcg_id)
 429                         node = &parent->rb_left;
 430                 else if (congested->blkcg_id > blkcg_id)
 431                         node = &parent->rb_right;
 432                 else
 433                         goto found;
 434         }
 435
 436         if (new_congested) {
 437                 /* !found and storage for new one already allocated, insert */
 438                 congested = new_congested;
 439                 new_congested = NULL;
 440                 rb_link_node(&congested->rb_node, parent, node);
 441                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 442                 goto found;
 443         }
 444
 445         spin_unlock_irqrestore(&cgwb_lock, flags);
 446
 447         /* allocate storage for new one and retry */
 448         new_congested = kzalloc(sizeof(*new_congested), gfp);
 449         if (!new_congested)
 450                 return NULL;
 451
 452         atomic_set(&new_congested->refcnt, 0);
 453         new_congested->__bdi = bdi;
 454         new_congested->blkcg_id = blkcg_id;
 455         goto retry;
 456
 457 found:
 458         atomic_inc(&congested->refcnt);
 459         spin_unlock_irqrestore(&cgwb_lock, flags);
 460         kfree(new_congested);
 461         return congested;
 462 }
 463
 464 /**
 465  * wb_congested_put - put a wb_congested
 466  * @congested: wb_congested to put
 467  *
 468  * Put @congested and destroy it if the refcnt reaches zero.
 469  */
 470 void wb_congested_put(struct bdi_writeback_congested *congested)
 471 {
 472         unsigned long flags;
 473
 474         local_irq_save(flags);
 475         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 476                 local_irq_restore(flags);
 477                 return;
 478         }
 479
 480         /* bdi might already have been destroyed leaving @congested unlinked */
 481         if (congested->__bdi) {
 482                 rb_erase(&congested->rb_node,
 483                          &congested->__bdi->cgwb_congested_tree);
 484                 congested->__bdi = NULL;
 485         }
 486
 487         spin_unlock_irqrestore(&cgwb_lock, flags);
 488         kfree(congested);
 489 }
 490
 491 static void cgwb_release_workfn(struct work_struct *work)
 492 {
 493         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 494                                                 release_work);
 495         struct backing_dev_info *bdi = wb->bdi;
 496
 497         wb_shutdown(wb);
 498
 499         css_put(wb->memcg_css);
 500         css_put(wb->blkcg_css);
 501
 502         fprop_local_destroy_percpu(&wb->memcg_completions);
 503         percpu_ref_exit(&wb->refcnt);
 504         wb_exit(wb);
 505         kfree_rcu(wb, rcu);
 506
 507         if (atomic_dec_and_test(&bdi->usage_cnt))
 508                 wake_up_all(&cgwb_release_wait);
 509 }
 510
 511 static void cgwb_release(struct percpu_ref *refcnt)
 512 {
 513         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 514                                                 refcnt);
 515         schedule_work(&wb->release_work);
 516 }
 517
 518 static void cgwb_kill(struct bdi_writeback *wb)
 519 {
 520         lockdep_assert_held(&cgwb_lock);
 521
 522         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 523         list_del(&wb->memcg_node);
 524         list_del(&wb->blkcg_node);
 525         percpu_ref_kill(&wb->refcnt);
 526 }
 527
 528 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 529 {
 530         spin_lock_irq(&cgwb_lock);
 531         list_del_rcu(&wb->bdi_node);
 532         spin_unlock_irq(&cgwb_lock);
 533 }
 534
 535 static int cgwb_create(struct backing_dev_info *bdi,
 536                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 537 {
 538         struct mem_cgroup *memcg;
 539         struct cgroup_subsys_state *blkcg_css;
 540         struct blkcg *blkcg;
 541         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 542         struct bdi_writeback *wb;
 543         unsigned long flags;
 544         int ret = 0;
 545
 546         memcg = mem_cgroup_from_css(memcg_css);
 547         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 548         blkcg = css_to_blkcg(blkcg_css);
 549         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 550         blkcg_cgwb_list = &blkcg->cgwb_list;
 551
 552         /* look up again under lock and discard on blkcg mismatch */
 553         spin_lock_irqsave(&cgwb_lock, flags);
 554         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 555         if (wb && wb->blkcg_css != blkcg_css) {
 556                 cgwb_kill(wb);
 557                 wb = NULL;
 558         }
 559         spin_unlock_irqrestore(&cgwb_lock, flags);
 560         if (wb)
 561                 goto out_put;
 562
 563         /* need to create a new one */
 564         wb = kmalloc(sizeof(*wb), gfp);
 565         if (!wb)
 566                 return -ENOMEM;
 567
 568         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 569         if (ret)
 570                 goto err_free;
 571
 572         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 573         if (ret)
 574                 goto err_wb_exit;
 575
 576         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 577         if (ret)
 578                 goto err_ref_exit;
 579
 580         wb->memcg_css = memcg_css;
 581         wb->blkcg_css = blkcg_css;
 582         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 583         set_bit(WB_registered, &wb->state);
 584
 585         /*
 586          * The root wb determines the registered state of the whole bdi and
 587          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 588          * whether they're still online.  Don't link @wb if any is dead.
 589          * See wb_memcg_offline() and wb_blkcg_offline().
 590          */
 591         ret = -ENODEV;
 592         spin_lock_irqsave(&cgwb_lock, flags);
 593         if (test_bit(WB_registered, &bdi->wb.state) &&
 594             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 595                 /* we might have raced another instance of this function */
 596                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 597                 if (!ret) {
 598                         atomic_inc(&bdi->usage_cnt);
 599                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 600                         list_add(&wb->memcg_node, memcg_cgwb_list);
 601                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 602                         css_get(memcg_css);
 603                         css_get(blkcg_css);
 604                 }
 605         }
 606         spin_unlock_irqrestore(&cgwb_lock, flags);
 607         if (ret) {
 608                 if (ret == -EEXIST)
 609                         ret = 0;
 610                 goto err_fprop_exit;
 611         }
 612         goto out_put;
 613
 614 err_fprop_exit:
 615         fprop_local_destroy_percpu(&wb->memcg_completions);
 616 err_ref_exit:
 617         percpu_ref_exit(&wb->refcnt);
 618 err_wb_exit:
 619         wb_exit(wb);
 620 err_free:
 621         kfree(wb);
 622 out_put:
 623         css_put(blkcg_css);
 624         return ret;
 625 }
 626
 627 /**
 628  * wb_get_create - get wb for a given memcg, create if necessary
 629  * @bdi: target bdi
 630  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 631  * @gfp: allocation mask to use
 632  *
 633  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 634  * create one.  The returned wb has its refcount incremented.
 635  *
 636  * This function uses css_get() on @memcg_css and thus expects its refcnt
 637  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 638  * @memcg_css isn't enough.  try_get it before calling this function.
 639  *
 640  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 641  * memcg on the default hierarchy, memcg association is guaranteed to be
 642  * more specific (equal or descendant to the associated blkcg) and thus can
 643  * identify both the memcg and blkcg associations.
 644  *
 645  * Because the blkcg associated with a memcg may change as blkcg is enabled
 646  * and disabled closer to root in the hierarchy, each wb keeps track of
 647  * both the memcg and blkcg associated with it and verifies the blkcg on
 648  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 649  * created.
 650  */
 651 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 652                                     struct cgroup_subsys_state *memcg_css,
 653                                     gfp_t gfp)
 654 {
 655         struct bdi_writeback *wb;
 656
 657         might_sleep_if(gfpflags_allow_blocking(gfp));
 658
 659         if (!memcg_css->parent)
 660                 return &bdi->wb;
 661
 662         do {
 663                 rcu_read_lock();
 664                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 665                 if (wb) {
 666                         struct cgroup_subsys_state *blkcg_css;
 667
 668                         /* see whether the blkcg association has changed */
 669                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 670                                                      &io_cgrp_subsys);
 671                         if (unlikely(wb->blkcg_css != blkcg_css ||
 672                                      !wb_tryget(wb)))
 673                                 wb = NULL;
 674                         css_put(blkcg_css);
 675                 }
 676                 rcu_read_unlock();
 677         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 678
 679         return wb;
 680 }
 681
 682 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 683 {
 684         int ret;
 685
 686         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 687         bdi->cgwb_congested_tree = RB_ROOT;
 688         atomic_set(&bdi->usage_cnt, 1);
 689
 690         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 691         if (!ret) {
 692                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 693                 bdi->wb.blkcg_css = blkcg_root_css;
 694         }
 695         return ret;
 696 }
 697
 698 static void cgwb_bdi_destroy(struct backing_dev_info *bdi)
 699 {
 700         struct radix_tree_iter iter;
 701         void **slot;
 702
 703         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 704
 705         spin_lock_irq(&cgwb_lock);
 706         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 707                 cgwb_kill(*slot);
 708         spin_unlock_irq(&cgwb_lock);
 709
 710         /*
 711          * All cgwb's must be shutdown and released before returning.  Drain
 712          * the usage counter to wait for all cgwb's ever created on @bdi.
 713          */
 714         atomic_dec(&bdi->usage_cnt);
 715         wait_event(cgwb_release_wait, !atomic_read(&bdi->usage_cnt));
 716         /*
 717          * Grab back our reference so that we hold it when @bdi gets
 718          * re-registered.
 719          */
 720         atomic_inc(&bdi->usage_cnt);
 721 }
 722
 723 /**
 724  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 725  * @memcg: memcg being offlined
 726  *
 727  * Also prevents creation of any new wb's associated with @memcg.
 728  */
 729 void wb_memcg_offline(struct mem_cgroup *memcg)
 730 {
 731         LIST_HEAD(to_destroy);
 732         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 733         struct bdi_writeback *wb, *next;
 734
 735         spin_lock_irq(&cgwb_lock);
 736         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 737                 cgwb_kill(wb);
 738         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 739         spin_unlock_irq(&cgwb_lock);
 740 }
 741
 742 /**
 743  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 744  * @blkcg: blkcg being offlined
 745  *
 746  * Also prevents creation of any new wb's associated with @blkcg.
 747  */
 748 void wb_blkcg_offline(struct blkcg *blkcg)
 749 {
 750         LIST_HEAD(to_destroy);
 751         struct bdi_writeback *wb, *next;
 752
 753         spin_lock_irq(&cgwb_lock);
 754         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 755                 cgwb_kill(wb);
 756         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 757         spin_unlock_irq(&cgwb_lock);
 758 }
 759
 760 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 761 {
 762         struct rb_node *rbn;
 763
 764         spin_lock_irq(&cgwb_lock);
 765         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 766                 struct bdi_writeback_congested *congested =
 767                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 768
 769                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 770                 congested->__bdi = NULL;        /* mark @congested unlinked */
 771         }
 772         spin_unlock_irq(&cgwb_lock);
 773 }
 774
 775 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 776 {
 777         spin_lock_irq(&cgwb_lock);
 778         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 779         spin_unlock_irq(&cgwb_lock);
 780 }
 781
 782 #else   /* CONFIG_CGROUP_WRITEBACK */
 783
 784 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 785 {
 786         int err;
 787
 788         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 789         if (!bdi->wb_congested)
 790                 return -ENOMEM;
 791
 792         atomic_set(&bdi->wb_congested->refcnt, 1);
 793
 794         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 795         if (err) {
 796                 wb_congested_put(bdi->wb_congested);
 797                 return err;
 798         }
 799         return 0;
 800 }
 801
 802 static void cgwb_bdi_destroy(struct backing_dev_info *bdi) { }
 803
 804 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 805 {
 806         wb_congested_put(bdi->wb_congested);
 807 }
 808
 809 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 810 {
 811         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 812 }
 813
 814 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 815 {
 816         list_del_rcu(&wb->bdi_node);
 817 }
 818
 819 #endif  /* CONFIG_CGROUP_WRITEBACK */
 820
 821 int bdi_init(struct backing_dev_info *bdi)
 822 {
 823         int ret;
 824
 825         bdi->dev = NULL;
 826
 827         kref_init(&bdi->refcnt);
 828         bdi->min_ratio = 0;
 829         bdi->max_ratio = 100;
 830         bdi->max_prop_frac = FPROP_FRAC_BASE;
 831         INIT_LIST_HEAD(&bdi->bdi_list);
 832         INIT_LIST_HEAD(&bdi->wb_list);
 833         init_waitqueue_head(&bdi->wb_waitq);
 834
 835         ret = cgwb_bdi_init(bdi);
 836
 837         return ret;
 838 }
 839 EXPORT_SYMBOL(bdi_init);
 840
 841 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 842 {
 843         struct backing_dev_info *bdi;
 844
 845         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 846                            gfp_mask | __GFP_ZERO, node_id);
 847         if (!bdi)
 848                 return NULL;
 849
 850         if (bdi_init(bdi)) {
 851                 kfree(bdi);
 852                 return NULL;
 853         }
 854         return bdi;
 855 }
 856
 857 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 858                 const char *fmt, ...)
 859 {
 860         va_list args;
 861         struct device *dev;
 862
 863         if (bdi->dev)   /* The driver needs to use separate queues per device */
 864                 return 0;
 865
 866         va_start(args, fmt);
 867         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 868         va_end(args);
 869         if (IS_ERR(dev))
 870                 return PTR_ERR(dev);
 871
 872         cgwb_bdi_register(bdi);
 873         bdi->dev = dev;
 874
 875         bdi_debug_register(bdi, dev_name(dev));
 876         set_bit(WB_registered, &bdi->wb.state);
 877
 878         spin_lock_bh(&bdi_lock);
 879         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 880         spin_unlock_bh(&bdi_lock);
 881
 882         trace_writeback_bdi_register(bdi);
 883         return 0;
 884 }
 885 EXPORT_SYMBOL(bdi_register);
 886
 887 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 888 {
 889         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 890 }
 891 EXPORT_SYMBOL(bdi_register_dev);
 892
 893 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 894 {
 895         int rc;
 896
 897         rc = bdi_register(bdi, NULL, "%u:%u", MAJOR(owner->devt),
 898                         MINOR(owner->devt));
 899         if (rc)
 900                 return rc;
 901         /* Leaking owner reference... */
 902         WARN_ON(bdi->owner);
 903         bdi->owner = owner;
 904         get_device(owner);
 905         return 0;
 906 }
 907 EXPORT_SYMBOL(bdi_register_owner);
 908
 909 /*
 910  * Remove bdi from bdi_list, and ensure that it is no longer visible
 911  */
 912 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 913 {
 914         spin_lock_bh(&bdi_lock);
 915         list_del_rcu(&bdi->bdi_list);
 916         spin_unlock_bh(&bdi_lock);
 917
 918         synchronize_rcu_expedited();
 919 }
 920
 921 void bdi_unregister(struct backing_dev_info *bdi)
 922 {
 923         /* make sure nobody finds us on the bdi_list anymore */
 924         bdi_remove_from_list(bdi);
 925         wb_shutdown(&bdi->wb);
 926         cgwb_bdi_destroy(bdi);
 927
 928         if (bdi->dev) {
 929                 bdi_debug_unregister(bdi);
 930                 device_unregister(bdi->dev);
 931                 bdi->dev = NULL;
 932         }
 933
 934         if (bdi->owner) {
 935                 put_device(bdi->owner);
 936                 bdi->owner = NULL;
 937         }
 938 }
 939
 940 static void bdi_exit(struct backing_dev_info *bdi)
 941 {
 942         WARN_ON_ONCE(bdi->dev);
 943         wb_exit(&bdi->wb);
 944         cgwb_bdi_exit(bdi);
 945 }
 946
 947 static void release_bdi(struct kref *ref)
 948 {
 949         struct backing_dev_info *bdi =
 950                         container_of(ref, struct backing_dev_info, refcnt);
 951
 952         bdi_exit(bdi);
 953         kfree(bdi);
 954 }
 955
 956 void bdi_put(struct backing_dev_info *bdi)
 957 {
 958         kref_put(&bdi->refcnt, release_bdi);
 959 }
 960
 961 void bdi_destroy(struct backing_dev_info *bdi)
 962 {
 963         bdi_unregister(bdi);
 964         bdi_exit(bdi);
 965 }
 966 EXPORT_SYMBOL(bdi_destroy);
 967
 968 /*
 969  * For use from filesystems to quickly init and register a bdi associated
 970  * with dirty writeback
 971  */
 972 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name)
 973 {
 974         int err;
 975
 976         bdi->name = name;
 977         bdi->capabilities = 0;
 978         err = bdi_init(bdi);
 979         if (err)
 980                 return err;
 981
 982         err = bdi_register(bdi, NULL, "%.28s-%ld", name,
 983                            atomic_long_inc_return(&bdi_seq));
 984         if (err) {
 985                 bdi_destroy(bdi);
 986                 return err;
 987         }
 988
 989         return 0;
 990 }
 991 EXPORT_SYMBOL(bdi_setup_and_register);
 992
 993 static wait_queue_head_t congestion_wqh[2] = {
 994                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 995                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 996         };
 997 static atomic_t nr_wb_congested[2];
 998
 999 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
1000 {
1001         wait_queue_head_t *wqh = &congestion_wqh[sync];
1002         enum wb_congested_state bit;
1003
1004         bit = sync ? WB_sync_congested : WB_async_congested;
1005         if (test_and_clear_bit(bit, &congested->state))
1006                 atomic_dec(&nr_wb_congested[sync]);
1007         smp_mb__after_atomic();
1008         if (waitqueue_active(wqh))
1009                 wake_up(wqh);
1010 }
1011 EXPORT_SYMBOL(clear_wb_congested);
1012
1013 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
1014 {
1015         enum wb_congested_state bit;
1016
1017         bit = sync ? WB_sync_congested : WB_async_congested;
1018         if (!test_and_set_bit(bit, &congested->state))
1019                 atomic_inc(&nr_wb_congested[sync]);
1020 }
1021 EXPORT_SYMBOL(set_wb_congested);
1022
1023 /**
1024  * congestion_wait - wait for a backing_dev to become uncongested
1025  * @sync: SYNC or ASYNC IO
1026  * @timeout: timeout in jiffies
1027  *
1028  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1029  * write congestion.  If no backing_devs are congested then just wait for the
1030  * next write to be completed.
1031  */
1032 long congestion_wait(int sync, long timeout)
1033 {
1034         long ret;
1035         unsigned long start = jiffies;
1036         DEFINE_WAIT(wait);
1037         wait_queue_head_t *wqh = &congestion_wqh[sync];
1038
1039         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1040         ret = io_schedule_timeout(timeout);
1041         finish_wait(wqh, &wait);
1042
1043         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1044                                         jiffies_to_usecs(jiffies - start));
1045
1046         return ret;
1047 }
1048 EXPORT_SYMBOL(congestion_wait);
1049
1050 /**
1051  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1052  * @pgdat: A pgdat to check if it is heavily congested
1053  * @sync: SYNC or ASYNC IO
1054  * @timeout: timeout in jiffies
1055  *
1056  * In the event of a congested backing_dev (any backing_dev) and the given
1057  * @pgdat has experienced recent congestion, this waits for up to @timeout
1058  * jiffies for either a BDI to exit congestion of the given @sync queue
1059  * or a write to complete.
1060  *
1061  * In the absence of pgdat congestion, cond_resched() is called to yield
1062  * the processor if necessary but otherwise does not sleep.
1063  *
1064  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1065  * it is the number of jiffies that were still remaining when the function
1066  * returned. return_value == timeout implies the function did not sleep.
1067  */
1068 long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
1069 {
1070         long ret;
1071         unsigned long start = jiffies;
1072         DEFINE_WAIT(wait);
1073         wait_queue_head_t *wqh = &congestion_wqh[sync];
1074
1075         /*
1076          * If there is no congestion, or heavy congestion is not being
1077          * encountered in the current pgdat, yield if necessary instead
1078          * of sleeping on the congestion queue
1079          */
1080         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1081             !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1082                 cond_resched();
1083
1084                 /* In case we scheduled, work out time remaining */
1085                 ret = timeout - (jiffies - start);
1086                 if (ret < 0)
1087                         ret = 0;
1088
1089                 goto out;
1090         }
1091
1092         /* Sleep until uncongested or a write happens */
1093         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1094         ret = io_schedule_timeout(timeout);
1095         finish_wait(wqh, &wait);
1096
1097 out:
1098         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1099                                         jiffies_to_usecs(jiffies - start));
1100
1101         return ret;
1102 }
1103 EXPORT_SYMBOL(wait_iff_congested);
1104
1105 int pdflush_proc_obsolete(struct ctl_table *table, int write,
1106                         void __user *buffer, size_t *lenp, loff_t *ppos)
1107 {
1108         char kbuf[] = "0\n";
1109
1110         if (*ppos || *lenp < sizeof(kbuf)) {
1111                 *lenp = 0;
1112                 return 0;
1113         }
1114
1115         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1116                 return -EFAULT;
1117         pr_warn_once("%s exported in /proc is scheduled for removal\n",
1118                      table->procname);
1119
1120         *lenp = 2;
1121         *ppos += *lenp;
1122         return 2;
1123 }