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
  23 /*
  24  * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
  25  * locking.
  26  */
  27 DEFINE_SPINLOCK(bdi_lock);
  28 LIST_HEAD(bdi_list);
  29
  30 /* bdi_wq serves all asynchronous writeback tasks */
  31 struct workqueue_struct *bdi_wq;
  32
  33 #ifdef CONFIG_DEBUG_FS
  34 #include <linux/debugfs.h>
  35 #include <linux/seq_file.h>
  36
  37 static struct dentry *bdi_debug_root;
  38
  39 static void bdi_debug_init(void)
  40 {
  41         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  42 }
  43
  44 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  45 {
  46         struct backing_dev_info *bdi = m->private;
  47         struct bdi_writeback *wb = &bdi->wb;
  48         unsigned long background_thresh;
  49         unsigned long dirty_thresh;
  50         unsigned long wb_thresh;
  51         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  52         struct inode *inode;
  53
  54         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  55         spin_lock(&wb->list_lock);
  56         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  57                 nr_dirty++;
  58         list_for_each_entry(inode, &wb->b_io, i_io_list)
  59                 nr_io++;
  60         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  61                 nr_more_io++;
  62         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  63                 if (inode->i_state & I_DIRTY_TIME)
  64                         nr_dirty_time++;
  65         spin_unlock(&wb->list_lock);
  66
  67         global_dirty_limits(&background_thresh, &dirty_thresh);
  68         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  69
  70 #define K(x) ((x) << (PAGE_SHIFT - 10))
  71         seq_printf(m,
  72                    "BdiWriteback:       %10lu kB\n"
  73                    "BdiReclaimable:     %10lu kB\n"
  74                    "BdiDirtyThresh:     %10lu kB\n"
  75                    "DirtyThresh:        %10lu kB\n"
  76                    "BackgroundThresh:   %10lu kB\n"
  77                    "BdiDirtied:         %10lu kB\n"
  78                    "BdiWritten:         %10lu kB\n"
  79                    "BdiWriteBandwidth:  %10lu kBps\n"
  80                    "b_dirty:            %10lu\n"
  81                    "b_io:               %10lu\n"
  82                    "b_more_io:          %10lu\n"
  83                    "b_dirty_time:       %10lu\n"
  84                    "bdi_list:           %10u\n"
  85                    "state:              %10lx\n",
  86                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  87                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  88                    K(wb_thresh),
  89                    K(dirty_thresh),
  90                    K(background_thresh),
  91                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  92                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  93                    (unsigned long) K(wb->write_bandwidth),
  94                    nr_dirty,
  95                    nr_io,
  96                    nr_more_io,
  97                    nr_dirty_time,
  98                    !list_empty(&bdi->bdi_list), bdi->wb.state);
  99 #undef K
 100
 101         return 0;
 102 }
 103
 104 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 105 {
 106         return single_open(file, bdi_debug_stats_show, inode->i_private);
 107 }
 108
 109 static const struct file_operations bdi_debug_stats_fops = {
 110         .open           = bdi_debug_stats_open,
 111         .read           = seq_read,
 112         .llseek         = seq_lseek,
 113         .release        = single_release,
 114 };
 115
 116 static int bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 117 {
 118         if (!bdi_debug_root)
 119                 return -ENOMEM;
 120
 121         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 122         if (!bdi->debug_dir)
 123                 return -ENOMEM;
 124
 125         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 126                                                bdi, &bdi_debug_stats_fops);
 127         if (!bdi->debug_stats) {
 128                 debugfs_remove(bdi->debug_dir);
 129                 return -ENOMEM;
 130         }
 131
 132         return 0;
 133 }
 134
 135 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 136 {
 137         debugfs_remove(bdi->debug_stats);
 138         debugfs_remove(bdi->debug_dir);
 139 }
 140 #else
 141 static inline void bdi_debug_init(void)
 142 {
 143 }
 144 static inline int bdi_debug_register(struct backing_dev_info *bdi,
 145                                       const char *name)
 146 {
 147         return 0;
 148 }
 149 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 150 {
 151 }
 152 #endif
 153
 154 static ssize_t read_ahead_kb_store(struct device *dev,
 155                                   struct device_attribute *attr,
 156                                   const char *buf, size_t count)
 157 {
 158         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 159         unsigned long read_ahead_kb;
 160         ssize_t ret;
 161
 162         ret = kstrtoul(buf, 10, &read_ahead_kb);
 163         if (ret < 0)
 164                 return ret;
 165
 166         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 167
 168         return count;
 169 }
 170
 171 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 172
 173 #define BDI_SHOW(name, expr)                                            \
 174 static ssize_t name##_show(struct device *dev,                          \
 175                            struct device_attribute *attr, char *page)   \
 176 {                                                                       \
 177         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 178                                                                         \
 179         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 180 }                                                                       \
 181 static DEVICE_ATTR_RW(name);
 182
 183 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 184
 185 static ssize_t min_ratio_store(struct device *dev,
 186                 struct device_attribute *attr, const char *buf, size_t count)
 187 {
 188         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 189         unsigned int ratio;
 190         ssize_t ret;
 191
 192         ret = kstrtouint(buf, 10, &ratio);
 193         if (ret < 0)
 194                 return ret;
 195
 196         ret = bdi_set_min_ratio(bdi, ratio);
 197         if (!ret)
 198                 ret = count;
 199
 200         return ret;
 201 }
 202 BDI_SHOW(min_ratio, bdi->min_ratio)
 203
 204 static ssize_t max_ratio_store(struct device *dev,
 205                 struct device_attribute *attr, const char *buf, size_t count)
 206 {
 207         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 208         unsigned int ratio;
 209         ssize_t ret;
 210
 211         ret = kstrtouint(buf, 10, &ratio);
 212         if (ret < 0)
 213                 return ret;
 214
 215         ret = bdi_set_max_ratio(bdi, ratio);
 216         if (!ret)
 217                 ret = count;
 218
 219         return ret;
 220 }
 221 BDI_SHOW(max_ratio, bdi->max_ratio)
 222
 223 static ssize_t stable_pages_required_show(struct device *dev,
 224                                           struct device_attribute *attr,
 225                                           char *page)
 226 {
 227         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 228
 229         return snprintf(page, PAGE_SIZE-1, "%d\n",
 230                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 231 }
 232 static DEVICE_ATTR_RO(stable_pages_required);
 233
 234 static struct attribute *bdi_dev_attrs[] = {
 235         &dev_attr_read_ahead_kb.attr,
 236         &dev_attr_min_ratio.attr,
 237         &dev_attr_max_ratio.attr,
 238         &dev_attr_stable_pages_required.attr,
 239         NULL,
 240 };
 241 ATTRIBUTE_GROUPS(bdi_dev);
 242
 243 static __init int bdi_class_init(void)
 244 {
 245         bdi_class = class_create(THIS_MODULE, "bdi");
 246         if (IS_ERR(bdi_class))
 247                 return PTR_ERR(bdi_class);
 248
 249         bdi_class->dev_groups = bdi_dev_groups;
 250         bdi_debug_init();
 251
 252         return 0;
 253 }
 254 postcore_initcall(bdi_class_init);
 255
 256 static int bdi_init(struct backing_dev_info *bdi);
 257
 258 static int __init default_bdi_init(void)
 259 {
 260         int err;
 261
 262         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
 263                                               WQ_UNBOUND | WQ_SYSFS, 0);
 264         if (!bdi_wq)
 265                 return -ENOMEM;
 266
 267         err = bdi_init(&noop_backing_dev_info);
 268
 269         return err;
 270 }
 271 subsys_initcall(default_bdi_init);
 272
 273 /*
 274  * This function is used when the first inode for this wb is marked dirty. It
 275  * wakes-up the corresponding bdi thread which should then take care of the
 276  * periodic background write-out of dirty inodes. Since the write-out would
 277  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 278  * set up a timer which wakes the bdi thread up later.
 279  *
 280  * Note, we wouldn't bother setting up the timer, but this function is on the
 281  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 282  * by delaying the wake-up.
 283  *
 284  * We have to be careful not to postpone flush work if it is scheduled for
 285  * earlier. Thus we use queue_delayed_work().
 286  */
 287 void wb_wakeup_delayed(struct bdi_writeback *wb)
 288 {
 289         unsigned long timeout;
 290
 291         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 292         spin_lock_bh(&wb->work_lock);
 293         if (test_bit(WB_registered, &wb->state))
 294                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 295         spin_unlock_bh(&wb->work_lock);
 296 }
 297
 298 /*
 299  * Initial write bandwidth: 100 MB/s
 300  */
 301 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 302
 303 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 304                    int blkcg_id, gfp_t gfp)
 305 {
 306         int i, err;
 307
 308         memset(wb, 0, sizeof(*wb));
 309
 310         if (wb != &bdi->wb)
 311                 bdi_get(bdi);
 312         wb->bdi = bdi;
 313         wb->last_old_flush = jiffies;
 314         INIT_LIST_HEAD(&wb->b_dirty);
 315         INIT_LIST_HEAD(&wb->b_io);
 316         INIT_LIST_HEAD(&wb->b_more_io);
 317         INIT_LIST_HEAD(&wb->b_dirty_time);
 318         spin_lock_init(&wb->list_lock);
 319
 320         wb->bw_time_stamp = jiffies;
 321         wb->balanced_dirty_ratelimit = INIT_BW;
 322         wb->dirty_ratelimit = INIT_BW;
 323         wb->write_bandwidth = INIT_BW;
 324         wb->avg_write_bandwidth = INIT_BW;
 325
 326         spin_lock_init(&wb->work_lock);
 327         INIT_LIST_HEAD(&wb->work_list);
 328         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 329         wb->dirty_sleep = jiffies;
 330
 331         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 332         if (!wb->congested) {
 333                 err = -ENOMEM;
 334                 goto out_put_bdi;
 335         }
 336
 337         err = fprop_local_init_percpu(&wb->completions, gfp);
 338         if (err)
 339                 goto out_put_cong;
 340
 341         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 342                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 343                 if (err)
 344                         goto out_destroy_stat;
 345         }
 346
 347         return 0;
 348
 349 out_destroy_stat:
 350         while (i--)
 351                 percpu_counter_destroy(&wb->stat[i]);
 352         fprop_local_destroy_percpu(&wb->completions);
 353 out_put_cong:
 354         wb_congested_put(wb->congested);
 355 out_put_bdi:
 356         if (wb != &bdi->wb)
 357                 bdi_put(bdi);
 358         return err;
 359 }
 360
 361 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 362
 363 /*
 364  * Remove bdi from the global list and shutdown any threads we have running
 365  */
 366 static void wb_shutdown(struct bdi_writeback *wb)
 367 {
 368         /* Make sure nobody queues further work */
 369         spin_lock_bh(&wb->work_lock);
 370         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 371                 spin_unlock_bh(&wb->work_lock);
 372                 /*
 373                  * Wait for wb shutdown to finish if someone else is just
 374                  * running wb_shutdown(). Otherwise we could proceed to wb /
 375                  * bdi destruction before wb_shutdown() is finished.
 376                  */
 377                 wait_on_bit(&wb->state, WB_shutting_down, TASK_UNINTERRUPTIBLE);
 378                 return;
 379         }
 380         set_bit(WB_shutting_down, &wb->state);
 381         spin_unlock_bh(&wb->work_lock);
 382
 383         cgwb_remove_from_bdi_list(wb);
 384         /*
 385          * Drain work list and shutdown the delayed_work.  !WB_registered
 386          * tells wb_workfn() that @wb is dying and its work_list needs to
 387          * be drained no matter what.
 388          */
 389         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 390         flush_delayed_work(&wb->dwork);
 391         WARN_ON(!list_empty(&wb->work_list));
 392         /*
 393          * Make sure bit gets cleared after shutdown is finished. Matches with
 394          * the barrier provided by test_and_clear_bit() above.
 395          */
 396         smp_wmb();
 397         clear_bit(WB_shutting_down, &wb->state);
 398 }
 399
 400 static void wb_exit(struct bdi_writeback *wb)
 401 {
 402         int i;
 403
 404         WARN_ON(delayed_work_pending(&wb->dwork));
 405
 406         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 407                 percpu_counter_destroy(&wb->stat[i]);
 408
 409         fprop_local_destroy_percpu(&wb->completions);
 410         wb_congested_put(wb->congested);
 411         if (wb != &wb->bdi->wb)
 412                 bdi_put(wb->bdi);
 413 }
 414
 415 #ifdef CONFIG_CGROUP_WRITEBACK
 416
 417 #include <linux/memcontrol.h>
 418
 419 /*
 420  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 421  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 422  * protected.
 423  */
 424 static DEFINE_SPINLOCK(cgwb_lock);
 425
 426 /**
 427  * wb_congested_get_create - get or create a wb_congested
 428  * @bdi: associated bdi
 429  * @blkcg_id: ID of the associated blkcg
 430  * @gfp: allocation mask
 431  *
 432  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 433  * The returned wb_congested has its reference count incremented.  Returns
 434  * NULL on failure.
 435  */
 436 struct bdi_writeback_congested *
 437 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 438 {
 439         struct bdi_writeback_congested *new_congested = NULL, *congested;
 440         struct rb_node **node, *parent;
 441         unsigned long flags;
 442 retry:
 443         spin_lock_irqsave(&cgwb_lock, flags);
 444
 445         node = &bdi->cgwb_congested_tree.rb_node;
 446         parent = NULL;
 447
 448         while (*node != NULL) {
 449                 parent = *node;
 450                 congested = rb_entry(parent, struct bdi_writeback_congested,
 451                                      rb_node);
 452                 if (congested->blkcg_id < blkcg_id)
 453                         node = &parent->rb_left;
 454                 else if (congested->blkcg_id > blkcg_id)
 455                         node = &parent->rb_right;
 456                 else
 457                         goto found;
 458         }
 459
 460         if (new_congested) {
 461                 /* !found and storage for new one already allocated, insert */
 462                 congested = new_congested;
 463                 new_congested = NULL;
 464                 rb_link_node(&congested->rb_node, parent, node);
 465                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 466                 goto found;
 467         }
 468
 469         spin_unlock_irqrestore(&cgwb_lock, flags);
 470
 471         /* allocate storage for new one and retry */
 472         new_congested = kzalloc(sizeof(*new_congested), gfp);
 473         if (!new_congested)
 474                 return NULL;
 475
 476         atomic_set(&new_congested->refcnt, 0);
 477         new_congested->__bdi = bdi;
 478         new_congested->blkcg_id = blkcg_id;
 479         goto retry;
 480
 481 found:
 482         atomic_inc(&congested->refcnt);
 483         spin_unlock_irqrestore(&cgwb_lock, flags);
 484         kfree(new_congested);
 485         return congested;
 486 }
 487
 488 /**
 489  * wb_congested_put - put a wb_congested
 490  * @congested: wb_congested to put
 491  *
 492  * Put @congested and destroy it if the refcnt reaches zero.
 493  */
 494 void wb_congested_put(struct bdi_writeback_congested *congested)
 495 {
 496         unsigned long flags;
 497
 498         local_irq_save(flags);
 499         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 500                 local_irq_restore(flags);
 501                 return;
 502         }
 503
 504         /* bdi might already have been destroyed leaving @congested unlinked */
 505         if (congested->__bdi) {
 506                 rb_erase(&congested->rb_node,
 507                          &congested->__bdi->cgwb_congested_tree);
 508                 congested->__bdi = NULL;
 509         }
 510
 511         spin_unlock_irqrestore(&cgwb_lock, flags);
 512         kfree(congested);
 513 }
 514
 515 static void cgwb_release_workfn(struct work_struct *work)
 516 {
 517         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 518                                                 release_work);
 519
 520         wb_shutdown(wb);
 521
 522         css_put(wb->memcg_css);
 523         css_put(wb->blkcg_css);
 524
 525         fprop_local_destroy_percpu(&wb->memcg_completions);
 526         percpu_ref_exit(&wb->refcnt);
 527         wb_exit(wb);
 528         kfree_rcu(wb, rcu);
 529 }
 530
 531 static void cgwb_release(struct percpu_ref *refcnt)
 532 {
 533         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 534                                                 refcnt);
 535         schedule_work(&wb->release_work);
 536 }
 537
 538 static void cgwb_kill(struct bdi_writeback *wb)
 539 {
 540         lockdep_assert_held(&cgwb_lock);
 541
 542         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 543         list_del(&wb->memcg_node);
 544         list_del(&wb->blkcg_node);
 545         percpu_ref_kill(&wb->refcnt);
 546 }
 547
 548 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 549 {
 550         spin_lock_irq(&cgwb_lock);
 551         list_del_rcu(&wb->bdi_node);
 552         spin_unlock_irq(&cgwb_lock);
 553 }
 554
 555 static int cgwb_create(struct backing_dev_info *bdi,
 556                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 557 {
 558         struct mem_cgroup *memcg;
 559         struct cgroup_subsys_state *blkcg_css;
 560         struct blkcg *blkcg;
 561         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 562         struct bdi_writeback *wb;
 563         unsigned long flags;
 564         int ret = 0;
 565
 566         memcg = mem_cgroup_from_css(memcg_css);
 567         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 568         blkcg = css_to_blkcg(blkcg_css);
 569         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 570         blkcg_cgwb_list = &blkcg->cgwb_list;
 571
 572         /* look up again under lock and discard on blkcg mismatch */
 573         spin_lock_irqsave(&cgwb_lock, flags);
 574         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 575         if (wb && wb->blkcg_css != blkcg_css) {
 576                 cgwb_kill(wb);
 577                 wb = NULL;
 578         }
 579         spin_unlock_irqrestore(&cgwb_lock, flags);
 580         if (wb)
 581                 goto out_put;
 582
 583         /* need to create a new one */
 584         wb = kmalloc(sizeof(*wb), gfp);
 585         if (!wb) {
 586                 ret = -ENOMEM;
 587                 goto out_put;
 588         }
 589
 590         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 591         if (ret)
 592                 goto err_free;
 593
 594         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 595         if (ret)
 596                 goto err_wb_exit;
 597
 598         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 599         if (ret)
 600                 goto err_ref_exit;
 601
 602         wb->memcg_css = memcg_css;
 603         wb->blkcg_css = blkcg_css;
 604         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 605         set_bit(WB_registered, &wb->state);
 606
 607         /*
 608          * The root wb determines the registered state of the whole bdi and
 609          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 610          * whether they're still online.  Don't link @wb if any is dead.
 611          * See wb_memcg_offline() and wb_blkcg_offline().
 612          */
 613         ret = -ENODEV;
 614         spin_lock_irqsave(&cgwb_lock, flags);
 615         if (test_bit(WB_registered, &bdi->wb.state) &&
 616             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 617                 /* we might have raced another instance of this function */
 618                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 619                 if (!ret) {
 620                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 621                         list_add(&wb->memcg_node, memcg_cgwb_list);
 622                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 623                         css_get(memcg_css);
 624                         css_get(blkcg_css);
 625                 }
 626         }
 627         spin_unlock_irqrestore(&cgwb_lock, flags);
 628         if (ret) {
 629                 if (ret == -EEXIST)
 630                         ret = 0;
 631                 goto err_fprop_exit;
 632         }
 633         goto out_put;
 634
 635 err_fprop_exit:
 636         fprop_local_destroy_percpu(&wb->memcg_completions);
 637 err_ref_exit:
 638         percpu_ref_exit(&wb->refcnt);
 639 err_wb_exit:
 640         wb_exit(wb);
 641 err_free:
 642         kfree(wb);
 643 out_put:
 644         css_put(blkcg_css);
 645         return ret;
 646 }
 647
 648 /**
 649  * wb_get_create - get wb for a given memcg, create if necessary
 650  * @bdi: target bdi
 651  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 652  * @gfp: allocation mask to use
 653  *
 654  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 655  * create one.  The returned wb has its refcount incremented.
 656  *
 657  * This function uses css_get() on @memcg_css and thus expects its refcnt
 658  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 659  * @memcg_css isn't enough.  try_get it before calling this function.
 660  *
 661  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 662  * memcg on the default hierarchy, memcg association is guaranteed to be
 663  * more specific (equal or descendant to the associated blkcg) and thus can
 664  * identify both the memcg and blkcg associations.
 665  *
 666  * Because the blkcg associated with a memcg may change as blkcg is enabled
 667  * and disabled closer to root in the hierarchy, each wb keeps track of
 668  * both the memcg and blkcg associated with it and verifies the blkcg on
 669  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 670  * created.
 671  */
 672 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 673                                     struct cgroup_subsys_state *memcg_css,
 674                                     gfp_t gfp)
 675 {
 676         struct bdi_writeback *wb;
 677
 678         might_sleep_if(gfpflags_allow_blocking(gfp));
 679
 680         if (!memcg_css->parent)
 681                 return &bdi->wb;
 682
 683         do {
 684                 rcu_read_lock();
 685                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 686                 if (wb) {
 687                         struct cgroup_subsys_state *blkcg_css;
 688
 689                         /* see whether the blkcg association has changed */
 690                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 691                                                      &io_cgrp_subsys);
 692                         if (unlikely(wb->blkcg_css != blkcg_css ||
 693                                      !wb_tryget(wb)))
 694                                 wb = NULL;
 695                         css_put(blkcg_css);
 696                 }
 697                 rcu_read_unlock();
 698         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 699
 700         return wb;
 701 }
 702
 703 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 704 {
 705         int ret;
 706
 707         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 708         bdi->cgwb_congested_tree = RB_ROOT;
 709
 710         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 711         if (!ret) {
 712                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 713                 bdi->wb.blkcg_css = blkcg_root_css;
 714         }
 715         return ret;
 716 }
 717
 718 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 719 {
 720         struct radix_tree_iter iter;
 721         void **slot;
 722         struct bdi_writeback *wb;
 723
 724         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 725
 726         spin_lock_irq(&cgwb_lock);
 727         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 728                 cgwb_kill(*slot);
 729
 730         while (!list_empty(&bdi->wb_list)) {
 731                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 732                                       bdi_node);
 733                 spin_unlock_irq(&cgwb_lock);
 734                 wb_shutdown(wb);
 735                 spin_lock_irq(&cgwb_lock);
 736         }
 737         spin_unlock_irq(&cgwb_lock);
 738 }
 739
 740 /**
 741  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 742  * @memcg: memcg being offlined
 743  *
 744  * Also prevents creation of any new wb's associated with @memcg.
 745  */
 746 void wb_memcg_offline(struct mem_cgroup *memcg)
 747 {
 748         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 749         struct bdi_writeback *wb, *next;
 750
 751         spin_lock_irq(&cgwb_lock);
 752         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 753                 cgwb_kill(wb);
 754         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 755         spin_unlock_irq(&cgwb_lock);
 756 }
 757
 758 /**
 759  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 760  * @blkcg: blkcg being offlined
 761  *
 762  * Also prevents creation of any new wb's associated with @blkcg.
 763  */
 764 void wb_blkcg_offline(struct blkcg *blkcg)
 765 {
 766         struct bdi_writeback *wb, *next;
 767
 768         spin_lock_irq(&cgwb_lock);
 769         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 770                 cgwb_kill(wb);
 771         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 772         spin_unlock_irq(&cgwb_lock);
 773 }
 774
 775 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 776 {
 777         struct rb_node *rbn;
 778
 779         spin_lock_irq(&cgwb_lock);
 780         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 781                 struct bdi_writeback_congested *congested =
 782                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 783
 784                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 785                 congested->__bdi = NULL;        /* mark @congested unlinked */
 786         }
 787         spin_unlock_irq(&cgwb_lock);
 788 }
 789
 790 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 791 {
 792         spin_lock_irq(&cgwb_lock);
 793         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 794         spin_unlock_irq(&cgwb_lock);
 795 }
 796
 797 #else   /* CONFIG_CGROUP_WRITEBACK */
 798
 799 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 800 {
 801         int err;
 802
 803         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 804         if (!bdi->wb_congested)
 805                 return -ENOMEM;
 806
 807         atomic_set(&bdi->wb_congested->refcnt, 1);
 808
 809         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 810         if (err) {
 811                 wb_congested_put(bdi->wb_congested);
 812                 return err;
 813         }
 814         return 0;
 815 }
 816
 817 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 818
 819 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 820 {
 821         wb_congested_put(bdi->wb_congested);
 822 }
 823
 824 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 825 {
 826         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 827 }
 828
 829 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 830 {
 831         list_del_rcu(&wb->bdi_node);
 832 }
 833
 834 #endif  /* CONFIG_CGROUP_WRITEBACK */
 835
 836 static int bdi_init(struct backing_dev_info *bdi)
 837 {
 838         int ret;
 839
 840         bdi->dev = NULL;
 841
 842         kref_init(&bdi->refcnt);
 843         bdi->min_ratio = 0;
 844         bdi->max_ratio = 100;
 845         bdi->max_prop_frac = FPROP_FRAC_BASE;
 846         INIT_LIST_HEAD(&bdi->bdi_list);
 847         INIT_LIST_HEAD(&bdi->wb_list);
 848         init_waitqueue_head(&bdi->wb_waitq);
 849
 850         ret = cgwb_bdi_init(bdi);
 851
 852         return ret;
 853 }
 854
 855 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 856 {
 857         struct backing_dev_info *bdi;
 858
 859         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 860                            gfp_mask | __GFP_ZERO, node_id);
 861         if (!bdi)
 862                 return NULL;
 863
 864         if (bdi_init(bdi)) {
 865                 kfree(bdi);
 866                 return NULL;
 867         }
 868         return bdi;
 869 }
 870 EXPORT_SYMBOL(bdi_alloc_node);
 871
 872 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 873 {
 874         struct device *dev;
 875
 876         if (bdi->dev)   /* The driver needs to use separate queues per device */
 877                 return 0;
 878
 879         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 880         if (IS_ERR(dev))
 881                 return PTR_ERR(dev);
 882
 883         cgwb_bdi_register(bdi);
 884         bdi->dev = dev;
 885
 886         bdi_debug_register(bdi, dev_name(dev));
 887         set_bit(WB_registered, &bdi->wb.state);
 888
 889         spin_lock_bh(&bdi_lock);
 890         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 891         spin_unlock_bh(&bdi_lock);
 892
 893         trace_writeback_bdi_register(bdi);
 894         return 0;
 895 }
 896 EXPORT_SYMBOL(bdi_register_va);
 897
 898 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 899 {
 900         va_list args;
 901         int ret;
 902
 903         va_start(args, fmt);
 904         ret = bdi_register_va(bdi, fmt, args);
 905         va_end(args);
 906         return ret;
 907 }
 908 EXPORT_SYMBOL(bdi_register);
 909
 910 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 911 {
 912         int rc;
 913
 914         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 915         if (rc)
 916                 return rc;
 917         /* Leaking owner reference... */
 918         WARN_ON(bdi->owner);
 919         bdi->owner = owner;
 920         get_device(owner);
 921         return 0;
 922 }
 923 EXPORT_SYMBOL(bdi_register_owner);
 924
 925 /*
 926  * Remove bdi from bdi_list, and ensure that it is no longer visible
 927  */
 928 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 929 {
 930         spin_lock_bh(&bdi_lock);
 931         list_del_rcu(&bdi->bdi_list);
 932         spin_unlock_bh(&bdi_lock);
 933
 934         synchronize_rcu_expedited();
 935 }
 936
 937 void bdi_unregister(struct backing_dev_info *bdi)
 938 {
 939         /* make sure nobody finds us on the bdi_list anymore */
 940         bdi_remove_from_list(bdi);
 941         wb_shutdown(&bdi->wb);
 942         cgwb_bdi_unregister(bdi);
 943
 944         if (bdi->dev) {
 945                 bdi_debug_unregister(bdi);
 946                 device_unregister(bdi->dev);
 947                 bdi->dev = NULL;
 948         }
 949
 950         if (bdi->owner) {
 951                 put_device(bdi->owner);
 952                 bdi->owner = NULL;
 953         }
 954 }
 955
 956 static void release_bdi(struct kref *ref)
 957 {
 958         struct backing_dev_info *bdi =
 959                         container_of(ref, struct backing_dev_info, refcnt);
 960
 961         if (test_bit(WB_registered, &bdi->wb.state))
 962                 bdi_unregister(bdi);
 963         WARN_ON_ONCE(bdi->dev);
 964         wb_exit(&bdi->wb);
 965         cgwb_bdi_exit(bdi);
 966         kfree(bdi);
 967 }
 968
 969 void bdi_put(struct backing_dev_info *bdi)
 970 {
 971         kref_put(&bdi->refcnt, release_bdi);
 972 }
 973 EXPORT_SYMBOL(bdi_put);
 974
 975 static wait_queue_head_t congestion_wqh[2] = {
 976                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 977                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 978         };
 979 static atomic_t nr_wb_congested[2];
 980
 981 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 982 {
 983         wait_queue_head_t *wqh = &congestion_wqh[sync];
 984         enum wb_congested_state bit;
 985
 986         bit = sync ? WB_sync_congested : WB_async_congested;
 987         if (test_and_clear_bit(bit, &congested->state))
 988                 atomic_dec(&nr_wb_congested[sync]);
 989         smp_mb__after_atomic();
 990         if (waitqueue_active(wqh))
 991                 wake_up(wqh);
 992 }
 993 EXPORT_SYMBOL(clear_wb_congested);
 994
 995 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
 996 {
 997         enum wb_congested_state bit;
 998
 999         bit = sync ? WB_sync_congested : WB_async_congested;
1000         if (!test_and_set_bit(bit, &congested->state))
1001                 atomic_inc(&nr_wb_congested[sync]);
1002 }
1003 EXPORT_SYMBOL(set_wb_congested);
1004
1005 /**
1006  * congestion_wait - wait for a backing_dev to become uncongested
1007  * @sync: SYNC or ASYNC IO
1008  * @timeout: timeout in jiffies
1009  *
1010  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1011  * write congestion.  If no backing_devs are congested then just wait for the
1012  * next write to be completed.
1013  */
1014 long congestion_wait(int sync, long timeout)
1015 {
1016         long ret;
1017         unsigned long start = jiffies;
1018         DEFINE_WAIT(wait);
1019         wait_queue_head_t *wqh = &congestion_wqh[sync];
1020
1021         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1022         ret = io_schedule_timeout(timeout);
1023         finish_wait(wqh, &wait);
1024
1025         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1026                                         jiffies_to_usecs(jiffies - start));
1027
1028         return ret;
1029 }
1030 EXPORT_SYMBOL(congestion_wait);
1031
1032 /**
1033  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1034  * @pgdat: A pgdat to check if it is heavily congested
1035  * @sync: SYNC or ASYNC IO
1036  * @timeout: timeout in jiffies
1037  *
1038  * In the event of a congested backing_dev (any backing_dev) and the given
1039  * @pgdat has experienced recent congestion, this waits for up to @timeout
1040  * jiffies for either a BDI to exit congestion of the given @sync queue
1041  * or a write to complete.
1042  *
1043  * In the absence of pgdat congestion, cond_resched() is called to yield
1044  * the processor if necessary but otherwise does not sleep.
1045  *
1046  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1047  * it is the number of jiffies that were still remaining when the function
1048  * returned. return_value == timeout implies the function did not sleep.
1049  */
1050 long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
1051 {
1052         long ret;
1053         unsigned long start = jiffies;
1054         DEFINE_WAIT(wait);
1055         wait_queue_head_t *wqh = &congestion_wqh[sync];
1056
1057         /*
1058          * If there is no congestion, or heavy congestion is not being
1059          * encountered in the current pgdat, yield if necessary instead
1060          * of sleeping on the congestion queue
1061          */
1062         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1063             !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1064                 cond_resched();
1065
1066                 /* In case we scheduled, work out time remaining */
1067                 ret = timeout - (jiffies - start);
1068                 if (ret < 0)
1069                         ret = 0;
1070
1071                 goto out;
1072         }
1073
1074         /* Sleep until uncongested or a write happens */
1075         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1076         ret = io_schedule_timeout(timeout);
1077         finish_wait(wqh, &wait);
1078
1079 out:
1080         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1081                                         jiffies_to_usecs(jiffies - start));
1082
1083         return ret;
1084 }
1085 EXPORT_SYMBOL(wait_iff_congested);