mm/backing-dev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include <linux/wait.h>
   4 #include <linux/rbtree.h>
   5 #include <linux/backing-dev.h>
   6 #include <linux/kthread.h>
   7 #include <linux/freezer.h>
   8 #include <linux/fs.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/module.h>
  13 #include <linux/writeback.h>
  14 #include <linux/device.h>
  15 #include <trace/events/writeback.h>
  16
  17 struct backing_dev_info noop_backing_dev_info;
  18 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  19
  20 static struct class *bdi_class;
  21 static const char *bdi_unknown_name = "(unknown)";
  22
  23 /*
  24  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
  25  * reader side locking.
  26  */
  27 DEFINE_SPINLOCK(bdi_lock);
  28 static u64 bdi_id_cursor;
  29 static struct rb_root bdi_tree = RB_ROOT;
  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 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 106
 107 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 108 {
 109         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 110
 111         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 112                             &bdi_debug_stats_fops);
 113 }
 114
 115 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 116 {
 117         debugfs_remove_recursive(bdi->debug_dir);
 118 }
 119 #else
 120 static inline void bdi_debug_init(void)
 121 {
 122 }
 123 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 124                                       const char *name)
 125 {
 126 }
 127 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 128 {
 129 }
 130 #endif
 131
 132 static ssize_t read_ahead_kb_store(struct device *dev,
 133                                   struct device_attribute *attr,
 134                                   const char *buf, size_t count)
 135 {
 136         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 137         unsigned long read_ahead_kb;
 138         ssize_t ret;
 139
 140         ret = kstrtoul(buf, 10, &read_ahead_kb);
 141         if (ret < 0)
 142                 return ret;
 143
 144         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 145
 146         return count;
 147 }
 148
 149 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 150
 151 #define BDI_SHOW(name, expr)                                            \
 152 static ssize_t name##_show(struct device *dev,                          \
 153                            struct device_attribute *attr, char *page)   \
 154 {                                                                       \
 155         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 156                                                                         \
 157         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 158 }                                                                       \
 159 static DEVICE_ATTR_RW(name);
 160
 161 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 162
 163 static ssize_t min_ratio_store(struct device *dev,
 164                 struct device_attribute *attr, const char *buf, size_t count)
 165 {
 166         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 167         unsigned int ratio;
 168         ssize_t ret;
 169
 170         ret = kstrtouint(buf, 10, &ratio);
 171         if (ret < 0)
 172                 return ret;
 173
 174         ret = bdi_set_min_ratio(bdi, ratio);
 175         if (!ret)
 176                 ret = count;
 177
 178         return ret;
 179 }
 180 BDI_SHOW(min_ratio, bdi->min_ratio)
 181
 182 static ssize_t max_ratio_store(struct device *dev,
 183                 struct device_attribute *attr, const char *buf, size_t count)
 184 {
 185         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 186         unsigned int ratio;
 187         ssize_t ret;
 188
 189         ret = kstrtouint(buf, 10, &ratio);
 190         if (ret < 0)
 191                 return ret;
 192
 193         ret = bdi_set_max_ratio(bdi, ratio);
 194         if (!ret)
 195                 ret = count;
 196
 197         return ret;
 198 }
 199 BDI_SHOW(max_ratio, bdi->max_ratio)
 200
 201 static ssize_t stable_pages_required_show(struct device *dev,
 202                                           struct device_attribute *attr,
 203                                           char *page)
 204 {
 205         dev_warn_once(dev,
 206                 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
 207         return snprintf(page, PAGE_SIZE-1, "%d\n", 0);
 208 }
 209 static DEVICE_ATTR_RO(stable_pages_required);
 210
 211 static struct attribute *bdi_dev_attrs[] = {
 212         &dev_attr_read_ahead_kb.attr,
 213         &dev_attr_min_ratio.attr,
 214         &dev_attr_max_ratio.attr,
 215         &dev_attr_stable_pages_required.attr,
 216         NULL,
 217 };
 218 ATTRIBUTE_GROUPS(bdi_dev);
 219
 220 static __init int bdi_class_init(void)
 221 {
 222         bdi_class = class_create(THIS_MODULE, "bdi");
 223         if (IS_ERR(bdi_class))
 224                 return PTR_ERR(bdi_class);
 225
 226         bdi_class->dev_groups = bdi_dev_groups;
 227         bdi_debug_init();
 228
 229         return 0;
 230 }
 231 postcore_initcall(bdi_class_init);
 232
 233 static int bdi_init(struct backing_dev_info *bdi);
 234
 235 static int __init default_bdi_init(void)
 236 {
 237         int err;
 238
 239         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 240                                  WQ_SYSFS, 0);
 241         if (!bdi_wq)
 242                 return -ENOMEM;
 243
 244         err = bdi_init(&noop_backing_dev_info);
 245
 246         return err;
 247 }
 248 subsys_initcall(default_bdi_init);
 249
 250 /*
 251  * This function is used when the first inode for this wb is marked dirty. It
 252  * wakes-up the corresponding bdi thread which should then take care of the
 253  * periodic background write-out of dirty inodes. Since the write-out would
 254  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 255  * set up a timer which wakes the bdi thread up later.
 256  *
 257  * Note, we wouldn't bother setting up the timer, but this function is on the
 258  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 259  * by delaying the wake-up.
 260  *
 261  * We have to be careful not to postpone flush work if it is scheduled for
 262  * earlier. Thus we use queue_delayed_work().
 263  */
 264 void wb_wakeup_delayed(struct bdi_writeback *wb)
 265 {
 266         unsigned long timeout;
 267
 268         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 269         spin_lock_bh(&wb->work_lock);
 270         if (test_bit(WB_registered, &wb->state))
 271                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 272         spin_unlock_bh(&wb->work_lock);
 273 }
 274
 275 /*
 276  * Initial write bandwidth: 100 MB/s
 277  */
 278 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 279
 280 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 281                    gfp_t gfp)
 282 {
 283         int i, err;
 284
 285         memset(wb, 0, sizeof(*wb));
 286
 287         if (wb != &bdi->wb)
 288                 bdi_get(bdi);
 289         wb->bdi = bdi;
 290         wb->last_old_flush = jiffies;
 291         INIT_LIST_HEAD(&wb->b_dirty);
 292         INIT_LIST_HEAD(&wb->b_io);
 293         INIT_LIST_HEAD(&wb->b_more_io);
 294         INIT_LIST_HEAD(&wb->b_dirty_time);
 295         spin_lock_init(&wb->list_lock);
 296
 297         wb->bw_time_stamp = jiffies;
 298         wb->balanced_dirty_ratelimit = INIT_BW;
 299         wb->dirty_ratelimit = INIT_BW;
 300         wb->write_bandwidth = INIT_BW;
 301         wb->avg_write_bandwidth = INIT_BW;
 302
 303         spin_lock_init(&wb->work_lock);
 304         INIT_LIST_HEAD(&wb->work_list);
 305         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 306         wb->dirty_sleep = jiffies;
 307
 308         err = fprop_local_init_percpu(&wb->completions, gfp);
 309         if (err)
 310                 goto out_put_bdi;
 311
 312         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 313                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 314                 if (err)
 315                         goto out_destroy_stat;
 316         }
 317
 318         return 0;
 319
 320 out_destroy_stat:
 321         while (i--)
 322                 percpu_counter_destroy(&wb->stat[i]);
 323         fprop_local_destroy_percpu(&wb->completions);
 324 out_put_bdi:
 325         if (wb != &bdi->wb)
 326                 bdi_put(bdi);
 327         return err;
 328 }
 329
 330 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 331
 332 /*
 333  * Remove bdi from the global list and shutdown any threads we have running
 334  */
 335 static void wb_shutdown(struct bdi_writeback *wb)
 336 {
 337         /* Make sure nobody queues further work */
 338         spin_lock_bh(&wb->work_lock);
 339         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 340                 spin_unlock_bh(&wb->work_lock);
 341                 return;
 342         }
 343         spin_unlock_bh(&wb->work_lock);
 344
 345         cgwb_remove_from_bdi_list(wb);
 346         /*
 347          * Drain work list and shutdown the delayed_work.  !WB_registered
 348          * tells wb_workfn() that @wb is dying and its work_list needs to
 349          * be drained no matter what.
 350          */
 351         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 352         flush_delayed_work(&wb->dwork);
 353         WARN_ON(!list_empty(&wb->work_list));
 354 }
 355
 356 static void wb_exit(struct bdi_writeback *wb)
 357 {
 358         int i;
 359
 360         WARN_ON(delayed_work_pending(&wb->dwork));
 361
 362         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 363                 percpu_counter_destroy(&wb->stat[i]);
 364
 365         fprop_local_destroy_percpu(&wb->completions);
 366         if (wb != &wb->bdi->wb)
 367                 bdi_put(wb->bdi);
 368 }
 369
 370 #ifdef CONFIG_CGROUP_WRITEBACK
 371
 372 #include <linux/memcontrol.h>
 373
 374 /*
 375  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, and memcg->cgwb_list.
 376  * bdi->cgwb_tree is also RCU protected.
 377  */
 378 static DEFINE_SPINLOCK(cgwb_lock);
 379 static struct workqueue_struct *cgwb_release_wq;
 380
 381 static void cgwb_free_rcu(struct rcu_head *rcu_head)
 382 {
 383         struct bdi_writeback *wb = container_of(rcu_head,
 384                         struct bdi_writeback, rcu);
 385
 386         percpu_ref_exit(&wb->refcnt);
 387         kfree(wb);
 388 }
 389
 390 static void cgwb_release_workfn(struct work_struct *work)
 391 {
 392         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 393                                                 release_work);
 394         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 395
 396         mutex_lock(&wb->bdi->cgwb_release_mutex);
 397         wb_shutdown(wb);
 398
 399         css_put(wb->memcg_css);
 400         css_put(wb->blkcg_css);
 401         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 402
 403         /* triggers blkg destruction if no online users left */
 404         blkcg_unpin_online(blkcg);
 405
 406         fprop_local_destroy_percpu(&wb->memcg_completions);
 407         wb_exit(wb);
 408         call_rcu(&wb->rcu, cgwb_free_rcu);
 409 }
 410
 411 static void cgwb_release(struct percpu_ref *refcnt)
 412 {
 413         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 414                                                 refcnt);
 415         queue_work(cgwb_release_wq, &wb->release_work);
 416 }
 417
 418 static void cgwb_kill(struct bdi_writeback *wb)
 419 {
 420         lockdep_assert_held(&cgwb_lock);
 421
 422         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 423         list_del(&wb->memcg_node);
 424         list_del(&wb->blkcg_node);
 425         percpu_ref_kill(&wb->refcnt);
 426 }
 427
 428 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 429 {
 430         spin_lock_irq(&cgwb_lock);
 431         list_del_rcu(&wb->bdi_node);
 432         spin_unlock_irq(&cgwb_lock);
 433 }
 434
 435 static int cgwb_create(struct backing_dev_info *bdi,
 436                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 437 {
 438         struct mem_cgroup *memcg;
 439         struct cgroup_subsys_state *blkcg_css;
 440         struct blkcg *blkcg;
 441         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 442         struct bdi_writeback *wb;
 443         unsigned long flags;
 444         int ret = 0;
 445
 446         memcg = mem_cgroup_from_css(memcg_css);
 447         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 448         blkcg = css_to_blkcg(blkcg_css);
 449         memcg_cgwb_list = &memcg->cgwb_list;
 450         blkcg_cgwb_list = &blkcg->cgwb_list;
 451
 452         /* look up again under lock and discard on blkcg mismatch */
 453         spin_lock_irqsave(&cgwb_lock, flags);
 454         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 455         if (wb && wb->blkcg_css != blkcg_css) {
 456                 cgwb_kill(wb);
 457                 wb = NULL;
 458         }
 459         spin_unlock_irqrestore(&cgwb_lock, flags);
 460         if (wb)
 461                 goto out_put;
 462
 463         /* need to create a new one */
 464         wb = kmalloc(sizeof(*wb), gfp);
 465         if (!wb) {
 466                 ret = -ENOMEM;
 467                 goto out_put;
 468         }
 469
 470         ret = wb_init(wb, bdi, gfp);
 471         if (ret)
 472                 goto err_free;
 473
 474         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 475         if (ret)
 476                 goto err_wb_exit;
 477
 478         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 479         if (ret)
 480                 goto err_ref_exit;
 481
 482         wb->memcg_css = memcg_css;
 483         wb->blkcg_css = blkcg_css;
 484         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 485         set_bit(WB_registered, &wb->state);
 486
 487         /*
 488          * The root wb determines the registered state of the whole bdi and
 489          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 490          * whether they're still online.  Don't link @wb if any is dead.
 491          * See wb_memcg_offline() and wb_blkcg_offline().
 492          */
 493         ret = -ENODEV;
 494         spin_lock_irqsave(&cgwb_lock, flags);
 495         if (test_bit(WB_registered, &bdi->wb.state) &&
 496             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 497                 /* we might have raced another instance of this function */
 498                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 499                 if (!ret) {
 500                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 501                         list_add(&wb->memcg_node, memcg_cgwb_list);
 502                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 503                         blkcg_pin_online(blkcg);
 504                         css_get(memcg_css);
 505                         css_get(blkcg_css);
 506                 }
 507         }
 508         spin_unlock_irqrestore(&cgwb_lock, flags);
 509         if (ret) {
 510                 if (ret == -EEXIST)
 511                         ret = 0;
 512                 goto err_fprop_exit;
 513         }
 514         goto out_put;
 515
 516 err_fprop_exit:
 517         fprop_local_destroy_percpu(&wb->memcg_completions);
 518 err_ref_exit:
 519         percpu_ref_exit(&wb->refcnt);
 520 err_wb_exit:
 521         wb_exit(wb);
 522 err_free:
 523         kfree(wb);
 524 out_put:
 525         css_put(blkcg_css);
 526         return ret;
 527 }
 528
 529 /**
 530  * wb_get_lookup - get wb for a given memcg
 531  * @bdi: target bdi
 532  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 533  *
 534  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 535  * refcount incremented.
 536  *
 537  * This function uses css_get() on @memcg_css and thus expects its refcnt
 538  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 539  * @memcg_css isn't enough.  try_get it before calling this function.
 540  *
 541  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 542  * memcg on the default hierarchy, memcg association is guaranteed to be
 543  * more specific (equal or descendant to the associated blkcg) and thus can
 544  * identify both the memcg and blkcg associations.
 545  *
 546  * Because the blkcg associated with a memcg may change as blkcg is enabled
 547  * and disabled closer to root in the hierarchy, each wb keeps track of
 548  * both the memcg and blkcg associated with it and verifies the blkcg on
 549  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 550  * created.
 551  */
 552 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 553                                     struct cgroup_subsys_state *memcg_css)
 554 {
 555         struct bdi_writeback *wb;
 556
 557         if (!memcg_css->parent)
 558                 return &bdi->wb;
 559
 560         rcu_read_lock();
 561         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 562         if (wb) {
 563                 struct cgroup_subsys_state *blkcg_css;
 564
 565                 /* see whether the blkcg association has changed */
 566                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 567                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 568                         wb = NULL;
 569                 css_put(blkcg_css);
 570         }
 571         rcu_read_unlock();
 572
 573         return wb;
 574 }
 575
 576 /**
 577  * wb_get_create - get wb for a given memcg, create if necessary
 578  * @bdi: target bdi
 579  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 580  * @gfp: allocation mask to use
 581  *
 582  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 583  * create one.  See wb_get_lookup() for more details.
 584  */
 585 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 586                                     struct cgroup_subsys_state *memcg_css,
 587                                     gfp_t gfp)
 588 {
 589         struct bdi_writeback *wb;
 590
 591         might_sleep_if(gfpflags_allow_blocking(gfp));
 592
 593         if (!memcg_css->parent)
 594                 return &bdi->wb;
 595
 596         do {
 597                 wb = wb_get_lookup(bdi, memcg_css);
 598         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 599
 600         return wb;
 601 }
 602
 603 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 604 {
 605         int ret;
 606
 607         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 608         mutex_init(&bdi->cgwb_release_mutex);
 609         init_rwsem(&bdi->wb_switch_rwsem);
 610
 611         ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 612         if (!ret) {
 613                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 614                 bdi->wb.blkcg_css = blkcg_root_css;
 615         }
 616         return ret;
 617 }
 618
 619 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 620 {
 621         struct radix_tree_iter iter;
 622         void **slot;
 623         struct bdi_writeback *wb;
 624
 625         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 626
 627         spin_lock_irq(&cgwb_lock);
 628         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 629                 cgwb_kill(*slot);
 630         spin_unlock_irq(&cgwb_lock);
 631
 632         mutex_lock(&bdi->cgwb_release_mutex);
 633         spin_lock_irq(&cgwb_lock);
 634         while (!list_empty(&bdi->wb_list)) {
 635                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 636                                       bdi_node);
 637                 spin_unlock_irq(&cgwb_lock);
 638                 wb_shutdown(wb);
 639                 spin_lock_irq(&cgwb_lock);
 640         }
 641         spin_unlock_irq(&cgwb_lock);
 642         mutex_unlock(&bdi->cgwb_release_mutex);
 643 }
 644
 645 /**
 646  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 647  * @memcg: memcg being offlined
 648  *
 649  * Also prevents creation of any new wb's associated with @memcg.
 650  */
 651 void wb_memcg_offline(struct mem_cgroup *memcg)
 652 {
 653         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 654         struct bdi_writeback *wb, *next;
 655
 656         spin_lock_irq(&cgwb_lock);
 657         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 658                 cgwb_kill(wb);
 659         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 660         spin_unlock_irq(&cgwb_lock);
 661 }
 662
 663 /**
 664  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 665  * @blkcg: blkcg being offlined
 666  *
 667  * Also prevents creation of any new wb's associated with @blkcg.
 668  */
 669 void wb_blkcg_offline(struct blkcg *blkcg)
 670 {
 671         struct bdi_writeback *wb, *next;
 672
 673         spin_lock_irq(&cgwb_lock);
 674         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 675                 cgwb_kill(wb);
 676         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 677         spin_unlock_irq(&cgwb_lock);
 678 }
 679
 680 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 681 {
 682         spin_lock_irq(&cgwb_lock);
 683         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 684         spin_unlock_irq(&cgwb_lock);
 685 }
 686
 687 static int __init cgwb_init(void)
 688 {
 689         /*
 690          * There can be many concurrent release work items overwhelming
 691          * system_wq.  Put them in a separate wq and limit concurrency.
 692          * There's no point in executing many of these in parallel.
 693          */
 694         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 695         if (!cgwb_release_wq)
 696                 return -ENOMEM;
 697
 698         return 0;
 699 }
 700 subsys_initcall(cgwb_init);
 701
 702 #else   /* CONFIG_CGROUP_WRITEBACK */
 703
 704 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 705 {
 706         return wb_init(&bdi->wb, bdi, GFP_KERNEL);
 707 }
 708
 709 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 710
 711 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 712 {
 713         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 714 }
 715
 716 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 717 {
 718         list_del_rcu(&wb->bdi_node);
 719 }
 720
 721 #endif  /* CONFIG_CGROUP_WRITEBACK */
 722
 723 static int bdi_init(struct backing_dev_info *bdi)
 724 {
 725         int ret;
 726
 727         bdi->dev = NULL;
 728
 729         kref_init(&bdi->refcnt);
 730         bdi->min_ratio = 0;
 731         bdi->max_ratio = 100;
 732         bdi->max_prop_frac = FPROP_FRAC_BASE;
 733         INIT_LIST_HEAD(&bdi->bdi_list);
 734         INIT_LIST_HEAD(&bdi->wb_list);
 735         init_waitqueue_head(&bdi->wb_waitq);
 736
 737         ret = cgwb_bdi_init(bdi);
 738
 739         return ret;
 740 }
 741
 742 struct backing_dev_info *bdi_alloc(int node_id)
 743 {
 744         struct backing_dev_info *bdi;
 745
 746         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 747         if (!bdi)
 748                 return NULL;
 749
 750         if (bdi_init(bdi)) {
 751                 kfree(bdi);
 752                 return NULL;
 753         }
 754         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 755         bdi->ra_pages = VM_READAHEAD_PAGES;
 756         bdi->io_pages = VM_READAHEAD_PAGES;
 757         return bdi;
 758 }
 759 EXPORT_SYMBOL(bdi_alloc);
 760
 761 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 762 {
 763         struct rb_node **p = &bdi_tree.rb_node;
 764         struct rb_node *parent = NULL;
 765         struct backing_dev_info *bdi;
 766
 767         lockdep_assert_held(&bdi_lock);
 768
 769         while (*p) {
 770                 parent = *p;
 771                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 772
 773                 if (bdi->id > id)
 774                         p = &(*p)->rb_left;
 775                 else if (bdi->id < id)
 776                         p = &(*p)->rb_right;
 777                 else
 778                         break;
 779         }
 780
 781         if (parentp)
 782                 *parentp = parent;
 783         return p;
 784 }
 785
 786 /**
 787  * bdi_get_by_id - lookup and get bdi from its id
 788  * @id: bdi id to lookup
 789  *
 790  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 791  * doesn't exist or is already unregistered.
 792  */
 793 struct backing_dev_info *bdi_get_by_id(u64 id)
 794 {
 795         struct backing_dev_info *bdi = NULL;
 796         struct rb_node **p;
 797
 798         spin_lock_bh(&bdi_lock);
 799         p = bdi_lookup_rb_node(id, NULL);
 800         if (*p) {
 801                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 802                 bdi_get(bdi);
 803         }
 804         spin_unlock_bh(&bdi_lock);
 805
 806         return bdi;
 807 }
 808
 809 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 810 {
 811         struct device *dev;
 812         struct rb_node *parent, **p;
 813
 814         if (bdi->dev)   /* The driver needs to use separate queues per device */
 815                 return 0;
 816
 817         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
 818         dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
 819         if (IS_ERR(dev))
 820                 return PTR_ERR(dev);
 821
 822         cgwb_bdi_register(bdi);
 823         bdi->dev = dev;
 824
 825         bdi_debug_register(bdi, dev_name(dev));
 826         set_bit(WB_registered, &bdi->wb.state);
 827
 828         spin_lock_bh(&bdi_lock);
 829
 830         bdi->id = ++bdi_id_cursor;
 831
 832         p = bdi_lookup_rb_node(bdi->id, &parent);
 833         rb_link_node(&bdi->rb_node, parent, p);
 834         rb_insert_color(&bdi->rb_node, &bdi_tree);
 835
 836         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 837
 838         spin_unlock_bh(&bdi_lock);
 839
 840         trace_writeback_bdi_register(bdi);
 841         return 0;
 842 }
 843
 844 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 845 {
 846         va_list args;
 847         int ret;
 848
 849         va_start(args, fmt);
 850         ret = bdi_register_va(bdi, fmt, args);
 851         va_end(args);
 852         return ret;
 853 }
 854 EXPORT_SYMBOL(bdi_register);
 855
 856 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
 857 {
 858         WARN_ON_ONCE(bdi->owner);
 859         bdi->owner = owner;
 860         get_device(owner);
 861 }
 862
 863 /*
 864  * Remove bdi from bdi_list, and ensure that it is no longer visible
 865  */
 866 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 867 {
 868         spin_lock_bh(&bdi_lock);
 869         rb_erase(&bdi->rb_node, &bdi_tree);
 870         list_del_rcu(&bdi->bdi_list);
 871         spin_unlock_bh(&bdi_lock);
 872
 873         synchronize_rcu_expedited();
 874 }
 875
 876 void bdi_unregister(struct backing_dev_info *bdi)
 877 {
 878         /* make sure nobody finds us on the bdi_list anymore */
 879         bdi_remove_from_list(bdi);
 880         wb_shutdown(&bdi->wb);
 881         cgwb_bdi_unregister(bdi);
 882
 883         /*
 884          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
 885          * update the global bdi_min_ratio.
 886          */
 887         if (bdi->min_ratio)
 888                 bdi_set_min_ratio(bdi, 0);
 889
 890         if (bdi->dev) {
 891                 bdi_debug_unregister(bdi);
 892                 device_unregister(bdi->dev);
 893                 bdi->dev = NULL;
 894         }
 895
 896         if (bdi->owner) {
 897                 put_device(bdi->owner);
 898                 bdi->owner = NULL;
 899         }
 900 }
 901
 902 static void release_bdi(struct kref *ref)
 903 {
 904         struct backing_dev_info *bdi =
 905                         container_of(ref, struct backing_dev_info, refcnt);
 906
 907         if (test_bit(WB_registered, &bdi->wb.state))
 908                 bdi_unregister(bdi);
 909         WARN_ON_ONCE(bdi->dev);
 910         wb_exit(&bdi->wb);
 911         kfree(bdi);
 912 }
 913
 914 void bdi_put(struct backing_dev_info *bdi)
 915 {
 916         kref_put(&bdi->refcnt, release_bdi);
 917 }
 918 EXPORT_SYMBOL(bdi_put);
 919
 920 const char *bdi_dev_name(struct backing_dev_info *bdi)
 921 {
 922         if (!bdi || !bdi->dev)
 923                 return bdi_unknown_name;
 924         return bdi->dev_name;
 925 }
 926 EXPORT_SYMBOL_GPL(bdi_dev_name);
 927
 928 static wait_queue_head_t congestion_wqh[2] = {
 929                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 930                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 931         };
 932 static atomic_t nr_wb_congested[2];
 933
 934 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 935 {
 936         wait_queue_head_t *wqh = &congestion_wqh[sync];
 937         enum wb_congested_state bit;
 938
 939         bit = sync ? WB_sync_congested : WB_async_congested;
 940         if (test_and_clear_bit(bit, &bdi->wb.congested))
 941                 atomic_dec(&nr_wb_congested[sync]);
 942         smp_mb__after_atomic();
 943         if (waitqueue_active(wqh))
 944                 wake_up(wqh);
 945 }
 946 EXPORT_SYMBOL(clear_bdi_congested);
 947
 948 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
 949 {
 950         enum wb_congested_state bit;
 951
 952         bit = sync ? WB_sync_congested : WB_async_congested;
 953         if (!test_and_set_bit(bit, &bdi->wb.congested))
 954                 atomic_inc(&nr_wb_congested[sync]);
 955 }
 956 EXPORT_SYMBOL(set_bdi_congested);
 957
 958 /**
 959  * congestion_wait - wait for a backing_dev to become uncongested
 960  * @sync: SYNC or ASYNC IO
 961  * @timeout: timeout in jiffies
 962  *
 963  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 964  * write congestion.  If no backing_devs are congested then just wait for the
 965  * next write to be completed.
 966  */
 967 long congestion_wait(int sync, long timeout)
 968 {
 969         long ret;
 970         unsigned long start = jiffies;
 971         DEFINE_WAIT(wait);
 972         wait_queue_head_t *wqh = &congestion_wqh[sync];
 973
 974         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 975         ret = io_schedule_timeout(timeout);
 976         finish_wait(wqh, &wait);
 977
 978         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
 979                                         jiffies_to_usecs(jiffies - start));
 980
 981         return ret;
 982 }
 983 EXPORT_SYMBOL(congestion_wait);
 984
 985 /**
 986  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
 987  * @sync: SYNC or ASYNC IO
 988  * @timeout: timeout in jiffies
 989  *
 990  * In the event of a congested backing_dev (any backing_dev) this waits
 991  * for up to @timeout jiffies for either a BDI to exit congestion of the
 992  * given @sync queue or a write to complete.
 993  *
 994  * The return value is 0 if the sleep is for the full timeout. Otherwise,
 995  * it is the number of jiffies that were still remaining when the function
 996  * returned. return_value == timeout implies the function did not sleep.
 997  */
 998 long wait_iff_congested(int sync, long timeout)
 999 {
1000         long ret;
1001         unsigned long start = jiffies;
1002         DEFINE_WAIT(wait);
1003         wait_queue_head_t *wqh = &congestion_wqh[sync];
1004
1005         /*
1006          * If there is no congestion, yield if necessary instead
1007          * of sleeping on the congestion queue
1008          */
1009         if (atomic_read(&nr_wb_congested[sync]) == 0) {
1010                 cond_resched();
1011
1012                 /* In case we scheduled, work out time remaining */
1013                 ret = timeout - (jiffies - start);
1014                 if (ret < 0)
1015                         ret = 0;
1016
1017                 goto out;
1018         }
1019
1020         /* Sleep until uncongested or a write happens */
1021         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1022         ret = io_schedule_timeout(timeout);
1023         finish_wait(wqh, &wait);
1024
1025 out:
1026         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1027                                         jiffies_to_usecs(jiffies - start));
1028
1029         return ret;
1030 }
1031 EXPORT_SYMBOL(wait_iff_congested);