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