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