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 struct backing_dev_info noop_backing_dev_info = {
  16         .name           = "noop",
  17         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
  18 };
  19 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  20
  21 static struct class *bdi_class;
  22 const char *bdi_unknown_name = "(unknown)";
  23
  24 /*
  25  * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
  26  * locking.
  27  */
  28 DEFINE_SPINLOCK(bdi_lock);
  29 LIST_HEAD(bdi_list);
  30
  31 /* bdi_wq serves all asynchronous writeback tasks */
  32 struct workqueue_struct *bdi_wq;
  33
  34 #ifdef CONFIG_DEBUG_FS
  35 #include <linux/debugfs.h>
  36 #include <linux/seq_file.h>
  37
  38 static struct dentry *bdi_debug_root;
  39
  40 static void bdi_debug_init(void)
  41 {
  42         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  43 }
  44
  45 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  46 {
  47         struct backing_dev_info *bdi = m->private;
  48         struct bdi_writeback *wb = &bdi->wb;
  49         unsigned long background_thresh;
  50         unsigned long dirty_thresh;
  51         unsigned long wb_thresh;
  52         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  53         struct inode *inode;
  54
  55         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  56         spin_lock(&wb->list_lock);
  57         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  58                 nr_dirty++;
  59         list_for_each_entry(inode, &wb->b_io, i_io_list)
  60                 nr_io++;
  61         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  62                 nr_more_io++;
  63         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  64                 if (inode->i_state & I_DIRTY_TIME)
  65                         nr_dirty_time++;
  66         spin_unlock(&wb->list_lock);
  67
  68         global_dirty_limits(&background_thresh, &dirty_thresh);
  69         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  70
  71 #define K(x) ((x) << (PAGE_SHIFT - 10))
  72         seq_printf(m,
  73                    "BdiWriteback:       %10lu kB\n"
  74                    "BdiReclaimable:     %10lu kB\n"
  75                    "BdiDirtyThresh:     %10lu kB\n"
  76                    "DirtyThresh:        %10lu kB\n"
  77                    "BackgroundThresh:   %10lu kB\n"
  78                    "BdiDirtied:         %10lu kB\n"
  79                    "BdiWritten:         %10lu kB\n"
  80                    "BdiWriteBandwidth:  %10lu kBps\n"
  81                    "b_dirty:            %10lu\n"
  82                    "b_io:               %10lu\n"
  83                    "b_more_io:          %10lu\n"
  84                    "b_dirty_time:       %10lu\n"
  85                    "bdi_list:           %10u\n"
  86                    "state:              %10lx\n",
  87                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  88                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  89                    K(wb_thresh),
  90                    K(dirty_thresh),
  91                    K(background_thresh),
  92                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  93                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  94                    (unsigned long) K(wb->write_bandwidth),
  95                    nr_dirty,
  96                    nr_io,
  97                    nr_more_io,
  98                    nr_dirty_time,
  99                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 100 #undef K
 101
 102         return 0;
 103 }
 104
 105 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 106 {
 107         return single_open(file, bdi_debug_stats_show, inode->i_private);
 108 }
 109
 110 static const struct file_operations bdi_debug_stats_fops = {
 111         .open           = bdi_debug_stats_open,
 112         .read           = seq_read,
 113         .llseek         = seq_lseek,
 114         .release        = single_release,
 115 };
 116
 117 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 118 {
 119         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 120         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 121                                                bdi, &bdi_debug_stats_fops);
 122 }
 123
 124 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 125 {
 126         debugfs_remove(bdi->debug_stats);
 127         debugfs_remove(bdi->debug_dir);
 128 }
 129 #else
 130 static inline void bdi_debug_init(void)
 131 {
 132 }
 133 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 134                                       const char *name)
 135 {
 136 }
 137 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 138 {
 139 }
 140 #endif
 141
 142 static ssize_t read_ahead_kb_store(struct device *dev,
 143                                   struct device_attribute *attr,
 144                                   const char *buf, size_t count)
 145 {
 146         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 147         unsigned long read_ahead_kb;
 148         ssize_t ret;
 149
 150         ret = kstrtoul(buf, 10, &read_ahead_kb);
 151         if (ret < 0)
 152                 return ret;
 153
 154         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 155
 156         return count;
 157 }
 158
 159 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 160
 161 #define BDI_SHOW(name, expr)                                            \
 162 static ssize_t name##_show(struct device *dev,                          \
 163                            struct device_attribute *attr, char *page)   \
 164 {                                                                       \
 165         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 166                                                                         \
 167         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 168 }                                                                       \
 169 static DEVICE_ATTR_RW(name);
 170
 171 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 172
 173 static ssize_t min_ratio_store(struct device *dev,
 174                 struct device_attribute *attr, const char *buf, size_t count)
 175 {
 176         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 177         unsigned int ratio;
 178         ssize_t ret;
 179
 180         ret = kstrtouint(buf, 10, &ratio);
 181         if (ret < 0)
 182                 return ret;
 183
 184         ret = bdi_set_min_ratio(bdi, ratio);
 185         if (!ret)
 186                 ret = count;
 187
 188         return ret;
 189 }
 190 BDI_SHOW(min_ratio, bdi->min_ratio)
 191
 192 static ssize_t max_ratio_store(struct device *dev,
 193                 struct device_attribute *attr, const char *buf, size_t count)
 194 {
 195         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 196         unsigned int ratio;
 197         ssize_t ret;
 198
 199         ret = kstrtouint(buf, 10, &ratio);
 200         if (ret < 0)
 201                 return ret;
 202
 203         ret = bdi_set_max_ratio(bdi, ratio);
 204         if (!ret)
 205                 ret = count;
 206
 207         return ret;
 208 }
 209 BDI_SHOW(max_ratio, bdi->max_ratio)
 210
 211 static ssize_t stable_pages_required_show(struct device *dev,
 212                                           struct device_attribute *attr,
 213                                           char *page)
 214 {
 215         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 216
 217         return snprintf(page, PAGE_SIZE-1, "%d\n",
 218                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 219 }
 220 static DEVICE_ATTR_RO(stable_pages_required);
 221
 222 static struct attribute *bdi_dev_attrs[] = {
 223         &dev_attr_read_ahead_kb.attr,
 224         &dev_attr_min_ratio.attr,
 225         &dev_attr_max_ratio.attr,
 226         &dev_attr_stable_pages_required.attr,
 227         NULL,
 228 };
 229 ATTRIBUTE_GROUPS(bdi_dev);
 230
 231 static __init int bdi_class_init(void)
 232 {
 233         bdi_class = class_create(THIS_MODULE, "bdi");
 234         if (IS_ERR(bdi_class))
 235                 return PTR_ERR(bdi_class);
 236
 237         bdi_class->dev_groups = bdi_dev_groups;
 238         bdi_debug_init();
 239
 240         return 0;
 241 }
 242 postcore_initcall(bdi_class_init);
 243
 244 static int bdi_init(struct backing_dev_info *bdi);
 245
 246 static int __init default_bdi_init(void)
 247 {
 248         int err;
 249
 250         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 251                                  WQ_SYSFS, 0);
 252         if (!bdi_wq)
 253                 return -ENOMEM;
 254
 255         err = bdi_init(&noop_backing_dev_info);
 256
 257         return err;
 258 }
 259 subsys_initcall(default_bdi_init);
 260
 261 /*
 262  * This function is used when the first inode for this wb is marked dirty. It
 263  * wakes-up the corresponding bdi thread which should then take care of the
 264  * periodic background write-out of dirty inodes. Since the write-out would
 265  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 266  * set up a timer which wakes the bdi thread up later.
 267  *
 268  * Note, we wouldn't bother setting up the timer, but this function is on the
 269  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 270  * by delaying the wake-up.
 271  *
 272  * We have to be careful not to postpone flush work if it is scheduled for
 273  * earlier. Thus we use queue_delayed_work().
 274  */
 275 void wb_wakeup_delayed(struct bdi_writeback *wb)
 276 {
 277         unsigned long timeout;
 278
 279         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 280         spin_lock_bh(&wb->work_lock);
 281         if (test_bit(WB_registered, &wb->state))
 282                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 283         spin_unlock_bh(&wb->work_lock);
 284 }
 285
 286 /*
 287  * Initial write bandwidth: 100 MB/s
 288  */
 289 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 290
 291 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 292                    int blkcg_id, gfp_t gfp)
 293 {
 294         int i, err;
 295
 296         memset(wb, 0, sizeof(*wb));
 297
 298         if (wb != &bdi->wb)
 299                 bdi_get(bdi);
 300         wb->bdi = bdi;
 301         wb->last_old_flush = jiffies;
 302         INIT_LIST_HEAD(&wb->b_dirty);
 303         INIT_LIST_HEAD(&wb->b_io);
 304         INIT_LIST_HEAD(&wb->b_more_io);
 305         INIT_LIST_HEAD(&wb->b_dirty_time);
 306         spin_lock_init(&wb->list_lock);
 307
 308         wb->bw_time_stamp = jiffies;
 309         wb->balanced_dirty_ratelimit = INIT_BW;
 310         wb->dirty_ratelimit = INIT_BW;
 311         wb->write_bandwidth = INIT_BW;
 312         wb->avg_write_bandwidth = INIT_BW;
 313
 314         spin_lock_init(&wb->work_lock);
 315         INIT_LIST_HEAD(&wb->work_list);
 316         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 317         wb->dirty_sleep = jiffies;
 318
 319         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 320         if (!wb->congested) {
 321                 err = -ENOMEM;
 322                 goto out_put_bdi;
 323         }
 324
 325         err = fprop_local_init_percpu(&wb->completions, gfp);
 326         if (err)
 327                 goto out_put_cong;
 328
 329         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 330                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 331                 if (err)
 332                         goto out_destroy_stat;
 333         }
 334
 335         return 0;
 336
 337 out_destroy_stat:
 338         while (i--)
 339                 percpu_counter_destroy(&wb->stat[i]);
 340         fprop_local_destroy_percpu(&wb->completions);
 341 out_put_cong:
 342         wb_congested_put(wb->congested);
 343 out_put_bdi:
 344         if (wb != &bdi->wb)
 345                 bdi_put(bdi);
 346         return err;
 347 }
 348
 349 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 350
 351 /*
 352  * Remove bdi from the global list and shutdown any threads we have running
 353  */
 354 static void wb_shutdown(struct bdi_writeback *wb)
 355 {
 356         /* Make sure nobody queues further work */
 357         spin_lock_bh(&wb->work_lock);
 358         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 359                 spin_unlock_bh(&wb->work_lock);
 360                 return;
 361         }
 362         spin_unlock_bh(&wb->work_lock);
 363
 364         cgwb_remove_from_bdi_list(wb);
 365         /*
 366          * Drain work list and shutdown the delayed_work.  !WB_registered
 367          * tells wb_workfn() that @wb is dying and its work_list needs to
 368          * be drained no matter what.
 369          */
 370         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 371         flush_delayed_work(&wb->dwork);
 372         WARN_ON(!list_empty(&wb->work_list));
 373 }
 374
 375 static void wb_exit(struct bdi_writeback *wb)
 376 {
 377         int i;
 378
 379         WARN_ON(delayed_work_pending(&wb->dwork));
 380
 381         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 382                 percpu_counter_destroy(&wb->stat[i]);
 383
 384         fprop_local_destroy_percpu(&wb->completions);
 385         wb_congested_put(wb->congested);
 386         if (wb != &wb->bdi->wb)
 387                 bdi_put(wb->bdi);
 388 }
 389
 390 #ifdef CONFIG_CGROUP_WRITEBACK
 391
 392 #include <linux/memcontrol.h>
 393
 394 /*
 395  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 396  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 397  * protected.
 398  */
 399 static DEFINE_SPINLOCK(cgwb_lock);
 400 static struct workqueue_struct *cgwb_release_wq;
 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
 496         mutex_lock(&wb->bdi->cgwb_release_mutex);
 497         wb_shutdown(wb);
 498
 499         css_put(wb->memcg_css);
 500         css_put(wb->blkcg_css);
 501         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 502
 503         fprop_local_destroy_percpu(&wb->memcg_completions);
 504         percpu_ref_exit(&wb->refcnt);
 505         wb_exit(wb);
 506         kfree_rcu(wb, rcu);
 507 }
 508
 509 static void cgwb_release(struct percpu_ref *refcnt)
 510 {
 511         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 512                                                 refcnt);
 513         queue_work(cgwb_release_wq, &wb->release_work);
 514 }
 515
 516 static void cgwb_kill(struct bdi_writeback *wb)
 517 {
 518         lockdep_assert_held(&cgwb_lock);
 519
 520         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 521         list_del(&wb->memcg_node);
 522         list_del(&wb->blkcg_node);
 523         percpu_ref_kill(&wb->refcnt);
 524 }
 525
 526 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 527 {
 528         spin_lock_irq(&cgwb_lock);
 529         list_del_rcu(&wb->bdi_node);
 530         spin_unlock_irq(&cgwb_lock);
 531 }
 532
 533 static int cgwb_create(struct backing_dev_info *bdi,
 534                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 535 {
 536         struct mem_cgroup *memcg;
 537         struct cgroup_subsys_state *blkcg_css;
 538         struct blkcg *blkcg;
 539         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 540         struct bdi_writeback *wb;
 541         unsigned long flags;
 542         int ret = 0;
 543
 544         memcg = mem_cgroup_from_css(memcg_css);
 545         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 546         blkcg = css_to_blkcg(blkcg_css);
 547         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 548         blkcg_cgwb_list = &blkcg->cgwb_list;
 549
 550         /* look up again under lock and discard on blkcg mismatch */
 551         spin_lock_irqsave(&cgwb_lock, flags);
 552         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 553         if (wb && wb->blkcg_css != blkcg_css) {
 554                 cgwb_kill(wb);
 555                 wb = NULL;
 556         }
 557         spin_unlock_irqrestore(&cgwb_lock, flags);
 558         if (wb)
 559                 goto out_put;
 560
 561         /* need to create a new one */
 562         wb = kmalloc(sizeof(*wb), gfp);
 563         if (!wb) {
 564                 ret = -ENOMEM;
 565                 goto out_put;
 566         }
 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                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 599                         list_add(&wb->memcg_node, memcg_cgwb_list);
 600                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 601                         css_get(memcg_css);
 602                         css_get(blkcg_css);
 603                 }
 604         }
 605         spin_unlock_irqrestore(&cgwb_lock, flags);
 606         if (ret) {
 607                 if (ret == -EEXIST)
 608                         ret = 0;
 609                 goto err_fprop_exit;
 610         }
 611         goto out_put;
 612
 613 err_fprop_exit:
 614         fprop_local_destroy_percpu(&wb->memcg_completions);
 615 err_ref_exit:
 616         percpu_ref_exit(&wb->refcnt);
 617 err_wb_exit:
 618         wb_exit(wb);
 619 err_free:
 620         kfree(wb);
 621 out_put:
 622         css_put(blkcg_css);
 623         return ret;
 624 }
 625
 626 /**
 627  * wb_get_create - get wb for a given memcg, create if necessary
 628  * @bdi: target bdi
 629  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 630  * @gfp: allocation mask to use
 631  *
 632  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 633  * create one.  The returned wb has its refcount incremented.
 634  *
 635  * This function uses css_get() on @memcg_css and thus expects its refcnt
 636  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 637  * @memcg_css isn't enough.  try_get it before calling this function.
 638  *
 639  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 640  * memcg on the default hierarchy, memcg association is guaranteed to be
 641  * more specific (equal or descendant to the associated blkcg) and thus can
 642  * identify both the memcg and blkcg associations.
 643  *
 644  * Because the blkcg associated with a memcg may change as blkcg is enabled
 645  * and disabled closer to root in the hierarchy, each wb keeps track of
 646  * both the memcg and blkcg associated with it and verifies the blkcg on
 647  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 648  * created.
 649  */
 650 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 651                                     struct cgroup_subsys_state *memcg_css,
 652                                     gfp_t gfp)
 653 {
 654         struct bdi_writeback *wb;
 655
 656         might_sleep_if(gfpflags_allow_blocking(gfp));
 657
 658         if (!memcg_css->parent)
 659                 return &bdi->wb;
 660
 661         do {
 662                 rcu_read_lock();
 663                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 664                 if (wb) {
 665                         struct cgroup_subsys_state *blkcg_css;
 666
 667                         /* see whether the blkcg association has changed */
 668                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 669                                                      &io_cgrp_subsys);
 670                         if (unlikely(wb->blkcg_css != blkcg_css ||
 671                                      !wb_tryget(wb)))
 672                                 wb = NULL;
 673                         css_put(blkcg_css);
 674                 }
 675                 rcu_read_unlock();
 676         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 677
 678         return wb;
 679 }
 680
 681 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 682 {
 683         int ret;
 684
 685         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 686         bdi->cgwb_congested_tree = RB_ROOT;
 687         mutex_init(&bdi->cgwb_release_mutex);
 688         init_rwsem(&bdi->wb_switch_rwsem);
 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_unregister(struct backing_dev_info *bdi)
 699 {
 700         struct radix_tree_iter iter;
 701         void **slot;
 702         struct bdi_writeback *wb;
 703
 704         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 705
 706         spin_lock_irq(&cgwb_lock);
 707         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 708                 cgwb_kill(*slot);
 709         spin_unlock_irq(&cgwb_lock);
 710
 711         mutex_lock(&bdi->cgwb_release_mutex);
 712         spin_lock_irq(&cgwb_lock);
 713         while (!list_empty(&bdi->wb_list)) {
 714                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 715                                       bdi_node);
 716                 spin_unlock_irq(&cgwb_lock);
 717                 wb_shutdown(wb);
 718                 spin_lock_irq(&cgwb_lock);
 719         }
 720         spin_unlock_irq(&cgwb_lock);
 721         mutex_unlock(&bdi->cgwb_release_mutex);
 722 }
 723
 724 /**
 725  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 726  * @memcg: memcg being offlined
 727  *
 728  * Also prevents creation of any new wb's associated with @memcg.
 729  */
 730 void wb_memcg_offline(struct mem_cgroup *memcg)
 731 {
 732         LIST_HEAD(to_destroy);
 733         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 734         struct bdi_writeback *wb, *next;
 735
 736         spin_lock_irq(&cgwb_lock);
 737         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 738                 cgwb_kill(wb);
 739         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 740         spin_unlock_irq(&cgwb_lock);
 741 }
 742
 743 /**
 744  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 745  * @blkcg: blkcg being offlined
 746  *
 747  * Also prevents creation of any new wb's associated with @blkcg.
 748  */
 749 void wb_blkcg_offline(struct blkcg *blkcg)
 750 {
 751         LIST_HEAD(to_destroy);
 752         struct bdi_writeback *wb, *next;
 753
 754         spin_lock_irq(&cgwb_lock);
 755         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 756                 cgwb_kill(wb);
 757         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 758         spin_unlock_irq(&cgwb_lock);
 759 }
 760
 761 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 762 {
 763         struct rb_node *rbn;
 764
 765         spin_lock_irq(&cgwb_lock);
 766         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 767                 struct bdi_writeback_congested *congested =
 768                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 769
 770                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 771                 congested->__bdi = NULL;        /* mark @congested unlinked */
 772         }
 773         spin_unlock_irq(&cgwb_lock);
 774 }
 775
 776 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 777 {
 778         spin_lock_irq(&cgwb_lock);
 779         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 780         spin_unlock_irq(&cgwb_lock);
 781 }
 782
 783 static int __init cgwb_init(void)
 784 {
 785         /*
 786          * There can be many concurrent release work items overwhelming
 787          * system_wq.  Put them in a separate wq and limit concurrency.
 788          * There's no point in executing many of these in parallel.
 789          */
 790         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 791         if (!cgwb_release_wq)
 792                 return -ENOMEM;
 793
 794         return 0;
 795 }
 796 subsys_initcall(cgwb_init);
 797
 798 #else   /* CONFIG_CGROUP_WRITEBACK */
 799
 800 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 801 {
 802         int err;
 803
 804         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 805         if (!bdi->wb_congested)
 806                 return -ENOMEM;
 807
 808         atomic_set(&bdi->wb_congested->refcnt, 1);
 809
 810         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 811         if (err) {
 812                 wb_congested_put(bdi->wb_congested);
 813                 return err;
 814         }
 815         return 0;
 816 }
 817
 818 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 819
 820 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 821 {
 822         wb_congested_put(bdi->wb_congested);
 823 }
 824
 825 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 826 {
 827         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 828 }
 829
 830 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 831 {
 832         list_del_rcu(&wb->bdi_node);
 833 }
 834
 835 #endif  /* CONFIG_CGROUP_WRITEBACK */
 836
 837 static int bdi_init(struct backing_dev_info *bdi)
 838 {
 839         int ret;
 840
 841         bdi->dev = NULL;
 842
 843         kref_init(&bdi->refcnt);
 844         bdi->min_ratio = 0;
 845         bdi->max_ratio = 100;
 846         bdi->max_prop_frac = FPROP_FRAC_BASE;
 847         INIT_LIST_HEAD(&bdi->bdi_list);
 848         INIT_LIST_HEAD(&bdi->wb_list);
 849         init_waitqueue_head(&bdi->wb_waitq);
 850
 851         ret = cgwb_bdi_init(bdi);
 852
 853         return ret;
 854 }
 855
 856 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 857 {
 858         struct backing_dev_info *bdi;
 859
 860         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 861                            gfp_mask | __GFP_ZERO, node_id);
 862         if (!bdi)
 863                 return NULL;
 864
 865         if (bdi_init(bdi)) {
 866                 kfree(bdi);
 867                 return NULL;
 868         }
 869         return bdi;
 870 }
 871 EXPORT_SYMBOL(bdi_alloc_node);
 872
 873 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 874 {
 875         struct device *dev;
 876
 877         if (bdi->dev)   /* The driver needs to use separate queues per device */
 878                 return 0;
 879
 880         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 881         if (IS_ERR(dev))
 882                 return PTR_ERR(dev);
 883
 884         cgwb_bdi_register(bdi);
 885         bdi->dev = dev;
 886
 887         bdi_debug_register(bdi, dev_name(dev));
 888         set_bit(WB_registered, &bdi->wb.state);
 889
 890         spin_lock_bh(&bdi_lock);
 891         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 892         spin_unlock_bh(&bdi_lock);
 893
 894         trace_writeback_bdi_register(bdi);
 895         return 0;
 896 }
 897 EXPORT_SYMBOL(bdi_register_va);
 898
 899 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 900 {
 901         va_list args;
 902         int ret;
 903
 904         va_start(args, fmt);
 905         ret = bdi_register_va(bdi, fmt, args);
 906         va_end(args);
 907         return ret;
 908 }
 909 EXPORT_SYMBOL(bdi_register);
 910
 911 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 912 {
 913         int rc;
 914
 915         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 916         if (rc)
 917                 return rc;
 918         /* Leaking owner reference... */
 919         WARN_ON(bdi->owner);
 920         bdi->owner = owner;
 921         get_device(owner);
 922         return 0;
 923 }
 924 EXPORT_SYMBOL(bdi_register_owner);
 925
 926 /*
 927  * Remove bdi from bdi_list, and ensure that it is no longer visible
 928  */
 929 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 930 {
 931         spin_lock_bh(&bdi_lock);
 932         list_del_rcu(&bdi->bdi_list);
 933         spin_unlock_bh(&bdi_lock);
 934
 935         synchronize_rcu_expedited();
 936 }
 937
 938 void bdi_unregister(struct backing_dev_info *bdi)
 939 {
 940         /* make sure nobody finds us on the bdi_list anymore */
 941         bdi_remove_from_list(bdi);
 942         wb_shutdown(&bdi->wb);
 943         cgwb_bdi_unregister(bdi);
 944
 945         /*
 946          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
 947          * update the global bdi_min_ratio.
 948          */
 949         if (bdi->min_ratio)
 950                 bdi_set_min_ratio(bdi, 0);
 951
 952         if (bdi->dev) {
 953                 bdi_debug_unregister(bdi);
 954                 device_unregister(bdi->dev);
 955                 bdi->dev = NULL;
 956         }
 957
 958         if (bdi->owner) {
 959                 put_device(bdi->owner);
 960                 bdi->owner = NULL;
 961         }
 962 }
 963
 964 static void release_bdi(struct kref *ref)
 965 {
 966         struct backing_dev_info *bdi =
 967                         container_of(ref, struct backing_dev_info, refcnt);
 968
 969         if (test_bit(WB_registered, &bdi->wb.state))
 970                 bdi_unregister(bdi);
 971         WARN_ON_ONCE(bdi->dev);
 972         wb_exit(&bdi->wb);
 973         cgwb_bdi_exit(bdi);
 974         kfree(bdi);
 975 }
 976
 977 void bdi_put(struct backing_dev_info *bdi)
 978 {
 979         kref_put(&bdi->refcnt, release_bdi);
 980 }
 981 EXPORT_SYMBOL(bdi_put);
 982
 983 static wait_queue_head_t congestion_wqh[2] = {
 984                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 985                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 986         };
 987 static atomic_t nr_wb_congested[2];
 988
 989 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 990 {
 991         wait_queue_head_t *wqh = &congestion_wqh[sync];
 992         enum wb_congested_state bit;
 993
 994         bit = sync ? WB_sync_congested : WB_async_congested;
 995         if (test_and_clear_bit(bit, &congested->state))
 996                 atomic_dec(&nr_wb_congested[sync]);
 997         smp_mb__after_atomic();
 998         if (waitqueue_active(wqh))
 999                 wake_up(wqh);
1000 }
1001 EXPORT_SYMBOL(clear_wb_congested);
1002
1003 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
1004 {
1005         enum wb_congested_state bit;
1006
1007         bit = sync ? WB_sync_congested : WB_async_congested;
1008         if (!test_and_set_bit(bit, &congested->state))
1009                 atomic_inc(&nr_wb_congested[sync]);
1010 }
1011 EXPORT_SYMBOL(set_wb_congested);
1012
1013 /**
1014  * congestion_wait - wait for a backing_dev to become uncongested
1015  * @sync: SYNC or ASYNC IO
1016  * @timeout: timeout in jiffies
1017  *
1018  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1019  * write congestion.  If no backing_devs are congested then just wait for the
1020  * next write to be completed.
1021  */
1022 long congestion_wait(int sync, long timeout)
1023 {
1024         long ret;
1025         unsigned long start = jiffies;
1026         DEFINE_WAIT(wait);
1027         wait_queue_head_t *wqh = &congestion_wqh[sync];
1028
1029         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1030         ret = io_schedule_timeout(timeout);
1031         finish_wait(wqh, &wait);
1032
1033         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1034                                         jiffies_to_usecs(jiffies - start));
1035
1036         return ret;
1037 }
1038 EXPORT_SYMBOL(congestion_wait);
1039
1040 /**
1041  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1042  * @pgdat: A pgdat to check if it is heavily congested
1043  * @sync: SYNC or ASYNC IO
1044  * @timeout: timeout in jiffies
1045  *
1046  * In the event of a congested backing_dev (any backing_dev) and the given
1047  * @pgdat has experienced recent congestion, this waits for up to @timeout
1048  * jiffies for either a BDI to exit congestion of the given @sync queue
1049  * or a write to complete.
1050  *
1051  * In the absence of pgdat congestion, cond_resched() is called to yield
1052  * the processor if necessary but otherwise does not sleep.
1053  *
1054  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1055  * it is the number of jiffies that were still remaining when the function
1056  * returned. return_value == timeout implies the function did not sleep.
1057  */
1058 long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
1059 {
1060         long ret;
1061         unsigned long start = jiffies;
1062         DEFINE_WAIT(wait);
1063         wait_queue_head_t *wqh = &congestion_wqh[sync];
1064
1065         /*
1066          * If there is no congestion, or heavy congestion is not being
1067          * encountered in the current pgdat, yield if necessary instead
1068          * of sleeping on the congestion queue
1069          */
1070         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1071             !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1072                 cond_resched();
1073
1074                 /* In case we scheduled, work out time remaining */
1075                 ret = timeout - (jiffies - start);
1076                 if (ret < 0)
1077                         ret = 0;
1078
1079                 goto out;
1080         }
1081
1082         /* Sleep until uncongested or a write happens */
1083         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1084         ret = io_schedule_timeout(timeout);
1085         finish_wait(wqh, &wait);
1086
1087 out:
1088         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1089                                         jiffies_to_usecs(jiffies - start));
1090
1091         return ret;
1092 }
1093 EXPORT_SYMBOL(wait_iff_congested);
1094
1095 int pdflush_proc_obsolete(struct ctl_table *table, int write,
1096                         void __user *buffer, size_t *lenp, loff_t *ppos)
1097 {
1098         char kbuf[] = "0\n";
1099
1100         if (*ppos || *lenp < sizeof(kbuf)) {
1101                 *lenp = 0;
1102                 return 0;
1103         }
1104
1105         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1106                 return -EFAULT;
1107         pr_warn_once("%s exported in /proc is scheduled for removal\n",
1108                      table->procname);
1109
1110         *lenp = 2;
1111         *ppos += *lenp;
1112         return 2;
1113 }