mm/backing-dev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include <linux/blkdev.h>
   4 #include <linux/wait.h>
   5 #include <linux/rbtree.h>
   6 #include <linux/kthread.h>
   7 #include <linux/backing-dev.h>
   8 #include <linux/blk-cgroup.h>
   9 #include <linux/freezer.h>
  10 #include <linux/fs.h>
  11 #include <linux/pagemap.h>
  12 #include <linux/mm.h>
  13 #include <linux/sched/mm.h>
  14 #include <linux/sched.h>
  15 #include <linux/module.h>
  16 #include <linux/writeback.h>
  17 #include <linux/device.h>
  18 #include <trace/events/writeback.h>
  19
  20 struct backing_dev_info noop_backing_dev_info;
  21 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  22
  23 static struct class *bdi_class;
  24 static const char *bdi_unknown_name = "(unknown)";
  25
  26 /*
  27  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
  28  * reader side locking.
  29  */
  30 DEFINE_SPINLOCK(bdi_lock);
  31 static u64 bdi_id_cursor;
  32 static struct rb_root bdi_tree = RB_ROOT;
  33 LIST_HEAD(bdi_list);
  34
  35 /* bdi_wq serves all asynchronous writeback tasks */
  36 struct workqueue_struct *bdi_wq;
  37
  38 #define K(x) ((x) << (PAGE_SHIFT - 10))
  39
  40 #ifdef CONFIG_DEBUG_FS
  41 #include <linux/debugfs.h>
  42 #include <linux/seq_file.h>
  43
  44 static struct dentry *bdi_debug_root;
  45
  46 static void bdi_debug_init(void)
  47 {
  48         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  49 }
  50
  51 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  52 {
  53         struct backing_dev_info *bdi = m->private;
  54         struct bdi_writeback *wb = &bdi->wb;
  55         unsigned long background_thresh;
  56         unsigned long dirty_thresh;
  57         unsigned long wb_thresh;
  58         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  59         struct inode *inode;
  60
  61         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  62         spin_lock(&wb->list_lock);
  63         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  64                 nr_dirty++;
  65         list_for_each_entry(inode, &wb->b_io, i_io_list)
  66                 nr_io++;
  67         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  68                 nr_more_io++;
  69         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  70                 if (inode->i_state & I_DIRTY_TIME)
  71                         nr_dirty_time++;
  72         spin_unlock(&wb->list_lock);
  73
  74         global_dirty_limits(&background_thresh, &dirty_thresh);
  75         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  76
  77         seq_printf(m,
  78                    "BdiWriteback:       %10lu kB\n"
  79                    "BdiReclaimable:     %10lu kB\n"
  80                    "BdiDirtyThresh:     %10lu kB\n"
  81                    "DirtyThresh:        %10lu kB\n"
  82                    "BackgroundThresh:   %10lu kB\n"
  83                    "BdiDirtied:         %10lu kB\n"
  84                    "BdiWritten:         %10lu kB\n"
  85                    "BdiWriteBandwidth:  %10lu kBps\n"
  86                    "b_dirty:            %10lu\n"
  87                    "b_io:               %10lu\n"
  88                    "b_more_io:          %10lu\n"
  89                    "b_dirty_time:       %10lu\n"
  90                    "bdi_list:           %10u\n"
  91                    "state:              %10lx\n",
  92                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  93                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  94                    K(wb_thresh),
  95                    K(dirty_thresh),
  96                    K(background_thresh),
  97                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  98                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  99                    (unsigned long) K(wb->write_bandwidth),
 100                    nr_dirty,
 101                    nr_io,
 102                    nr_more_io,
 103                    nr_dirty_time,
 104                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 105
 106         return 0;
 107 }
 108 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 109
 110 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 111 {
 112         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 113
 114         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 115                             &bdi_debug_stats_fops);
 116 }
 117
 118 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 119 {
 120         debugfs_remove_recursive(bdi->debug_dir);
 121 }
 122 #else
 123 static inline void bdi_debug_init(void)
 124 {
 125 }
 126 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 127                                       const char *name)
 128 {
 129 }
 130 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 131 {
 132 }
 133 #endif
 134
 135 static ssize_t read_ahead_kb_store(struct device *dev,
 136                                   struct device_attribute *attr,
 137                                   const char *buf, size_t count)
 138 {
 139         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 140         unsigned long read_ahead_kb;
 141         ssize_t ret;
 142
 143         ret = kstrtoul(buf, 10, &read_ahead_kb);
 144         if (ret < 0)
 145                 return ret;
 146
 147         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 148
 149         return count;
 150 }
 151
 152 #define BDI_SHOW(name, expr)                                            \
 153 static ssize_t name##_show(struct device *dev,                          \
 154                            struct device_attribute *attr, char *buf)    \
 155 {                                                                       \
 156         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 157                                                                         \
 158         return sysfs_emit(buf, "%lld\n", (long long)expr);              \
 159 }                                                                       \
 160 static DEVICE_ATTR_RW(name);
 161
 162 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 163
 164 static ssize_t min_ratio_store(struct device *dev,
 165                 struct device_attribute *attr, const char *buf, size_t count)
 166 {
 167         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 168         unsigned int ratio;
 169         ssize_t ret;
 170
 171         ret = kstrtouint(buf, 10, &ratio);
 172         if (ret < 0)
 173                 return ret;
 174
 175         ret = bdi_set_min_ratio(bdi, ratio);
 176         if (!ret)
 177                 ret = count;
 178
 179         return ret;
 180 }
 181 BDI_SHOW(min_ratio, bdi->min_ratio)
 182
 183 static ssize_t max_ratio_store(struct device *dev,
 184                 struct device_attribute *attr, const char *buf, size_t count)
 185 {
 186         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 187         unsigned int ratio;
 188         ssize_t ret;
 189
 190         ret = kstrtouint(buf, 10, &ratio);
 191         if (ret < 0)
 192                 return ret;
 193
 194         ret = bdi_set_max_ratio(bdi, ratio);
 195         if (!ret)
 196                 ret = count;
 197
 198         return ret;
 199 }
 200 BDI_SHOW(max_ratio, bdi->max_ratio)
 201
 202 static ssize_t stable_pages_required_show(struct device *dev,
 203                                           struct device_attribute *attr,
 204                                           char *buf)
 205 {
 206         dev_warn_once(dev,
 207                 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
 208         return sysfs_emit(buf, "%d\n", 0);
 209 }
 210 static DEVICE_ATTR_RO(stable_pages_required);
 211
 212 static struct attribute *bdi_dev_attrs[] = {
 213         &dev_attr_read_ahead_kb.attr,
 214         &dev_attr_min_ratio.attr,
 215         &dev_attr_max_ratio.attr,
 216         &dev_attr_stable_pages_required.attr,
 217         NULL,
 218 };
 219 ATTRIBUTE_GROUPS(bdi_dev);
 220
 221 static __init int bdi_class_init(void)
 222 {
 223         bdi_class = class_create(THIS_MODULE, "bdi");
 224         if (IS_ERR(bdi_class))
 225                 return PTR_ERR(bdi_class);
 226
 227         bdi_class->dev_groups = bdi_dev_groups;
 228         bdi_debug_init();
 229
 230         return 0;
 231 }
 232 postcore_initcall(bdi_class_init);
 233
 234 static int bdi_init(struct backing_dev_info *bdi);
 235
 236 static int __init default_bdi_init(void)
 237 {
 238         int err;
 239
 240         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 241                                  WQ_SYSFS, 0);
 242         if (!bdi_wq)
 243                 return -ENOMEM;
 244
 245         err = bdi_init(&noop_backing_dev_info);
 246
 247         return err;
 248 }
 249 subsys_initcall(default_bdi_init);
 250
 251 /*
 252  * This function is used when the first inode for this wb is marked dirty. It
 253  * wakes-up the corresponding bdi thread which should then take care of the
 254  * periodic background write-out of dirty inodes. Since the write-out would
 255  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 256  * set up a timer which wakes the bdi thread up later.
 257  *
 258  * Note, we wouldn't bother setting up the timer, but this function is on the
 259  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 260  * by delaying the wake-up.
 261  *
 262  * We have to be careful not to postpone flush work if it is scheduled for
 263  * earlier. Thus we use queue_delayed_work().
 264  */
 265 void wb_wakeup_delayed(struct bdi_writeback *wb)
 266 {
 267         unsigned long timeout;
 268
 269         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 270         spin_lock_bh(&wb->work_lock);
 271         if (test_bit(WB_registered, &wb->state))
 272                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 273         spin_unlock_bh(&wb->work_lock);
 274 }
 275
 276 static void wb_update_bandwidth_workfn(struct work_struct *work)
 277 {
 278         struct bdi_writeback *wb = container_of(to_delayed_work(work),
 279                                                 struct bdi_writeback, bw_dwork);
 280
 281         wb_update_bandwidth(wb);
 282 }
 283
 284 /*
 285  * Initial write bandwidth: 100 MB/s
 286  */
 287 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 288
 289 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 290                    gfp_t gfp)
 291 {
 292         int i, err;
 293
 294         memset(wb, 0, sizeof(*wb));
 295
 296         wb->bdi = bdi;
 297         wb->last_old_flush = jiffies;
 298         INIT_LIST_HEAD(&wb->b_dirty);
 299         INIT_LIST_HEAD(&wb->b_io);
 300         INIT_LIST_HEAD(&wb->b_more_io);
 301         INIT_LIST_HEAD(&wb->b_dirty_time);
 302         spin_lock_init(&wb->list_lock);
 303
 304         atomic_set(&wb->writeback_inodes, 0);
 305         wb->bw_time_stamp = jiffies;
 306         wb->balanced_dirty_ratelimit = INIT_BW;
 307         wb->dirty_ratelimit = INIT_BW;
 308         wb->write_bandwidth = INIT_BW;
 309         wb->avg_write_bandwidth = INIT_BW;
 310
 311         spin_lock_init(&wb->work_lock);
 312         INIT_LIST_HEAD(&wb->work_list);
 313         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 314         INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
 315         wb->dirty_sleep = jiffies;
 316
 317         err = fprop_local_init_percpu(&wb->completions, gfp);
 318         if (err)
 319                 return err;
 320
 321         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 322                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 323                 if (err)
 324                         goto out_destroy_stat;
 325         }
 326
 327         return 0;
 328
 329 out_destroy_stat:
 330         while (i--)
 331                 percpu_counter_destroy(&wb->stat[i]);
 332         fprop_local_destroy_percpu(&wb->completions);
 333         return err;
 334 }
 335
 336 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 337
 338 /*
 339  * Remove bdi from the global list and shutdown any threads we have running
 340  */
 341 static void wb_shutdown(struct bdi_writeback *wb)
 342 {
 343         /* Make sure nobody queues further work */
 344         spin_lock_bh(&wb->work_lock);
 345         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 346                 spin_unlock_bh(&wb->work_lock);
 347                 return;
 348         }
 349         spin_unlock_bh(&wb->work_lock);
 350
 351         cgwb_remove_from_bdi_list(wb);
 352         /*
 353          * Drain work list and shutdown the delayed_work.  !WB_registered
 354          * tells wb_workfn() that @wb is dying and its work_list needs to
 355          * be drained no matter what.
 356          */
 357         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 358         flush_delayed_work(&wb->dwork);
 359         WARN_ON(!list_empty(&wb->work_list));
 360         flush_delayed_work(&wb->bw_dwork);
 361 }
 362
 363 static void wb_exit(struct bdi_writeback *wb)
 364 {
 365         int i;
 366
 367         WARN_ON(delayed_work_pending(&wb->dwork));
 368
 369         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 370                 percpu_counter_destroy(&wb->stat[i]);
 371
 372         fprop_local_destroy_percpu(&wb->completions);
 373 }
 374
 375 #ifdef CONFIG_CGROUP_WRITEBACK
 376
 377 #include <linux/memcontrol.h>
 378
 379 /*
 380  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
 381  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
 382  */
 383 static DEFINE_SPINLOCK(cgwb_lock);
 384 static struct workqueue_struct *cgwb_release_wq;
 385
 386 static LIST_HEAD(offline_cgwbs);
 387 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
 388 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
 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 backing_dev_info *bdi = wb->bdi;
 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(wb->blkcg_css);
 405
 406         fprop_local_destroy_percpu(&wb->memcg_completions);
 407
 408         spin_lock_irq(&cgwb_lock);
 409         list_del(&wb->offline_node);
 410         spin_unlock_irq(&cgwb_lock);
 411
 412         percpu_ref_exit(&wb->refcnt);
 413         wb_exit(wb);
 414         bdi_put(bdi);
 415         WARN_ON_ONCE(!list_empty(&wb->b_attached));
 416         kfree_rcu(wb, rcu);
 417 }
 418
 419 static void cgwb_release(struct percpu_ref *refcnt)
 420 {
 421         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 422                                                 refcnt);
 423         queue_work(cgwb_release_wq, &wb->release_work);
 424 }
 425
 426 static void cgwb_kill(struct bdi_writeback *wb)
 427 {
 428         lockdep_assert_held(&cgwb_lock);
 429
 430         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 431         list_del(&wb->memcg_node);
 432         list_del(&wb->blkcg_node);
 433         list_add(&wb->offline_node, &offline_cgwbs);
 434         percpu_ref_kill(&wb->refcnt);
 435 }
 436
 437 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 438 {
 439         spin_lock_irq(&cgwb_lock);
 440         list_del_rcu(&wb->bdi_node);
 441         spin_unlock_irq(&cgwb_lock);
 442 }
 443
 444 static int cgwb_create(struct backing_dev_info *bdi,
 445                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 446 {
 447         struct mem_cgroup *memcg;
 448         struct cgroup_subsys_state *blkcg_css;
 449         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 450         struct bdi_writeback *wb;
 451         unsigned long flags;
 452         int ret = 0;
 453
 454         memcg = mem_cgroup_from_css(memcg_css);
 455         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 456         memcg_cgwb_list = &memcg->cgwb_list;
 457         blkcg_cgwb_list = blkcg_get_cgwb_list(blkcg_css);
 458
 459         /* look up again under lock and discard on blkcg mismatch */
 460         spin_lock_irqsave(&cgwb_lock, flags);
 461         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 462         if (wb && wb->blkcg_css != blkcg_css) {
 463                 cgwb_kill(wb);
 464                 wb = NULL;
 465         }
 466         spin_unlock_irqrestore(&cgwb_lock, flags);
 467         if (wb)
 468                 goto out_put;
 469
 470         /* need to create a new one */
 471         wb = kmalloc(sizeof(*wb), gfp);
 472         if (!wb) {
 473                 ret = -ENOMEM;
 474                 goto out_put;
 475         }
 476
 477         ret = wb_init(wb, bdi, gfp);
 478         if (ret)
 479                 goto err_free;
 480
 481         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 482         if (ret)
 483                 goto err_wb_exit;
 484
 485         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 486         if (ret)
 487                 goto err_ref_exit;
 488
 489         wb->memcg_css = memcg_css;
 490         wb->blkcg_css = blkcg_css;
 491         INIT_LIST_HEAD(&wb->b_attached);
 492         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 493         set_bit(WB_registered, &wb->state);
 494         bdi_get(bdi);
 495
 496         /*
 497          * The root wb determines the registered state of the whole bdi and
 498          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 499          * whether they're still online.  Don't link @wb if any is dead.
 500          * See wb_memcg_offline() and wb_blkcg_offline().
 501          */
 502         ret = -ENODEV;
 503         spin_lock_irqsave(&cgwb_lock, flags);
 504         if (test_bit(WB_registered, &bdi->wb.state) &&
 505             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 506                 /* we might have raced another instance of this function */
 507                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 508                 if (!ret) {
 509                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 510                         list_add(&wb->memcg_node, memcg_cgwb_list);
 511                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 512                         blkcg_pin_online(blkcg_css);
 513                         css_get(memcg_css);
 514                         css_get(blkcg_css);
 515                 }
 516         }
 517         spin_unlock_irqrestore(&cgwb_lock, flags);
 518         if (ret) {
 519                 if (ret == -EEXIST)
 520                         ret = 0;
 521                 goto err_fprop_exit;
 522         }
 523         goto out_put;
 524
 525 err_fprop_exit:
 526         bdi_put(bdi);
 527         fprop_local_destroy_percpu(&wb->memcg_completions);
 528 err_ref_exit:
 529         percpu_ref_exit(&wb->refcnt);
 530 err_wb_exit:
 531         wb_exit(wb);
 532 err_free:
 533         kfree(wb);
 534 out_put:
 535         css_put(blkcg_css);
 536         return ret;
 537 }
 538
 539 /**
 540  * wb_get_lookup - get wb for a given memcg
 541  * @bdi: target bdi
 542  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 543  *
 544  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 545  * refcount incremented.
 546  *
 547  * This function uses css_get() on @memcg_css and thus expects its refcnt
 548  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 549  * @memcg_css isn't enough.  try_get it before calling this function.
 550  *
 551  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 552  * memcg on the default hierarchy, memcg association is guaranteed to be
 553  * more specific (equal or descendant to the associated blkcg) and thus can
 554  * identify both the memcg and blkcg associations.
 555  *
 556  * Because the blkcg associated with a memcg may change as blkcg is enabled
 557  * and disabled closer to root in the hierarchy, each wb keeps track of
 558  * both the memcg and blkcg associated with it and verifies the blkcg on
 559  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 560  * created.
 561  */
 562 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 563                                     struct cgroup_subsys_state *memcg_css)
 564 {
 565         struct bdi_writeback *wb;
 566
 567         if (!memcg_css->parent)
 568                 return &bdi->wb;
 569
 570         rcu_read_lock();
 571         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 572         if (wb) {
 573                 struct cgroup_subsys_state *blkcg_css;
 574
 575                 /* see whether the blkcg association has changed */
 576                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 577                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 578                         wb = NULL;
 579                 css_put(blkcg_css);
 580         }
 581         rcu_read_unlock();
 582
 583         return wb;
 584 }
 585
 586 /**
 587  * wb_get_create - get wb for a given memcg, create if necessary
 588  * @bdi: target bdi
 589  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 590  * @gfp: allocation mask to use
 591  *
 592  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 593  * create one.  See wb_get_lookup() for more details.
 594  */
 595 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 596                                     struct cgroup_subsys_state *memcg_css,
 597                                     gfp_t gfp)
 598 {
 599         struct bdi_writeback *wb;
 600
 601         might_alloc(gfp);
 602
 603         if (!memcg_css->parent)
 604                 return &bdi->wb;
 605
 606         do {
 607                 wb = wb_get_lookup(bdi, memcg_css);
 608         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 609
 610         return wb;
 611 }
 612
 613 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 614 {
 615         int ret;
 616
 617         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 618         mutex_init(&bdi->cgwb_release_mutex);
 619         init_rwsem(&bdi->wb_switch_rwsem);
 620
 621         ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 622         if (!ret) {
 623                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 624                 bdi->wb.blkcg_css = blkcg_root_css;
 625         }
 626         return ret;
 627 }
 628
 629 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 630 {
 631         struct radix_tree_iter iter;
 632         void **slot;
 633         struct bdi_writeback *wb;
 634
 635         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 636
 637         spin_lock_irq(&cgwb_lock);
 638         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 639                 cgwb_kill(*slot);
 640         spin_unlock_irq(&cgwb_lock);
 641
 642         mutex_lock(&bdi->cgwb_release_mutex);
 643         spin_lock_irq(&cgwb_lock);
 644         while (!list_empty(&bdi->wb_list)) {
 645                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 646                                       bdi_node);
 647                 spin_unlock_irq(&cgwb_lock);
 648                 wb_shutdown(wb);
 649                 spin_lock_irq(&cgwb_lock);
 650         }
 651         spin_unlock_irq(&cgwb_lock);
 652         mutex_unlock(&bdi->cgwb_release_mutex);
 653 }
 654
 655 /*
 656  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
 657  *
 658  * Try to release dying cgwbs by switching attached inodes to the nearest
 659  * living ancestor's writeback. Processed wbs are placed at the end
 660  * of the list to guarantee the forward progress.
 661  */
 662 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
 663 {
 664         struct bdi_writeback *wb;
 665         LIST_HEAD(processed);
 666
 667         spin_lock_irq(&cgwb_lock);
 668
 669         while (!list_empty(&offline_cgwbs)) {
 670                 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
 671                                       offline_node);
 672                 list_move(&wb->offline_node, &processed);
 673
 674                 /*
 675                  * If wb is dirty, cleaning up the writeback by switching
 676                  * attached inodes will result in an effective removal of any
 677                  * bandwidth restrictions, which isn't the goal.  Instead,
 678                  * it can be postponed until the next time, when all io
 679                  * will be likely completed.  If in the meantime some inodes
 680                  * will get re-dirtied, they should be eventually switched to
 681                  * a new cgwb.
 682                  */
 683                 if (wb_has_dirty_io(wb))
 684                         continue;
 685
 686                 if (!wb_tryget(wb))
 687                         continue;
 688
 689                 spin_unlock_irq(&cgwb_lock);
 690                 while (cleanup_offline_cgwb(wb))
 691                         cond_resched();
 692                 spin_lock_irq(&cgwb_lock);
 693
 694                 wb_put(wb);
 695         }
 696
 697         if (!list_empty(&processed))
 698                 list_splice_tail(&processed, &offline_cgwbs);
 699
 700         spin_unlock_irq(&cgwb_lock);
 701 }
 702
 703 /**
 704  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 705  * @memcg: memcg being offlined
 706  *
 707  * Also prevents creation of any new wb's associated with @memcg.
 708  */
 709 void wb_memcg_offline(struct mem_cgroup *memcg)
 710 {
 711         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 712         struct bdi_writeback *wb, *next;
 713
 714         spin_lock_irq(&cgwb_lock);
 715         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 716                 cgwb_kill(wb);
 717         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 718         spin_unlock_irq(&cgwb_lock);
 719
 720         queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
 721 }
 722
 723 /**
 724  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 725  * @css: blkcg being offlined
 726  *
 727  * Also prevents creation of any new wb's associated with @blkcg.
 728  */
 729 void wb_blkcg_offline(struct cgroup_subsys_state *css)
 730 {
 731         struct bdi_writeback *wb, *next;
 732         struct list_head *list = blkcg_get_cgwb_list(css);
 733
 734         spin_lock_irq(&cgwb_lock);
 735         list_for_each_entry_safe(wb, next, list, blkcg_node)
 736                 cgwb_kill(wb);
 737         list->next = NULL;      /* prevent new wb's */
 738         spin_unlock_irq(&cgwb_lock);
 739 }
 740
 741 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 742 {
 743         spin_lock_irq(&cgwb_lock);
 744         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 745         spin_unlock_irq(&cgwb_lock);
 746 }
 747
 748 static int __init cgwb_init(void)
 749 {
 750         /*
 751          * There can be many concurrent release work items overwhelming
 752          * system_wq.  Put them in a separate wq and limit concurrency.
 753          * There's no point in executing many of these in parallel.
 754          */
 755         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 756         if (!cgwb_release_wq)
 757                 return -ENOMEM;
 758
 759         return 0;
 760 }
 761 subsys_initcall(cgwb_init);
 762
 763 #else   /* CONFIG_CGROUP_WRITEBACK */
 764
 765 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 766 {
 767         return wb_init(&bdi->wb, bdi, GFP_KERNEL);
 768 }
 769
 770 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 771
 772 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 773 {
 774         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 775 }
 776
 777 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 778 {
 779         list_del_rcu(&wb->bdi_node);
 780 }
 781
 782 #endif  /* CONFIG_CGROUP_WRITEBACK */
 783
 784 static int bdi_init(struct backing_dev_info *bdi)
 785 {
 786         int ret;
 787
 788         bdi->dev = NULL;
 789
 790         kref_init(&bdi->refcnt);
 791         bdi->min_ratio = 0;
 792         bdi->max_ratio = 100;
 793         bdi->max_prop_frac = FPROP_FRAC_BASE;
 794         INIT_LIST_HEAD(&bdi->bdi_list);
 795         INIT_LIST_HEAD(&bdi->wb_list);
 796         init_waitqueue_head(&bdi->wb_waitq);
 797
 798         ret = cgwb_bdi_init(bdi);
 799
 800         return ret;
 801 }
 802
 803 struct backing_dev_info *bdi_alloc(int node_id)
 804 {
 805         struct backing_dev_info *bdi;
 806
 807         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 808         if (!bdi)
 809                 return NULL;
 810
 811         if (bdi_init(bdi)) {
 812                 kfree(bdi);
 813                 return NULL;
 814         }
 815         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 816         bdi->ra_pages = VM_READAHEAD_PAGES;
 817         bdi->io_pages = VM_READAHEAD_PAGES;
 818         timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
 819         return bdi;
 820 }
 821 EXPORT_SYMBOL(bdi_alloc);
 822
 823 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 824 {
 825         struct rb_node **p = &bdi_tree.rb_node;
 826         struct rb_node *parent = NULL;
 827         struct backing_dev_info *bdi;
 828
 829         lockdep_assert_held(&bdi_lock);
 830
 831         while (*p) {
 832                 parent = *p;
 833                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 834
 835                 if (bdi->id > id)
 836                         p = &(*p)->rb_left;
 837                 else if (bdi->id < id)
 838                         p = &(*p)->rb_right;
 839                 else
 840                         break;
 841         }
 842
 843         if (parentp)
 844                 *parentp = parent;
 845         return p;
 846 }
 847
 848 /**
 849  * bdi_get_by_id - lookup and get bdi from its id
 850  * @id: bdi id to lookup
 851  *
 852  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 853  * doesn't exist or is already unregistered.
 854  */
 855 struct backing_dev_info *bdi_get_by_id(u64 id)
 856 {
 857         struct backing_dev_info *bdi = NULL;
 858         struct rb_node **p;
 859
 860         spin_lock_bh(&bdi_lock);
 861         p = bdi_lookup_rb_node(id, NULL);
 862         if (*p) {
 863                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 864                 bdi_get(bdi);
 865         }
 866         spin_unlock_bh(&bdi_lock);
 867
 868         return bdi;
 869 }
 870
 871 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 872 {
 873         struct device *dev;
 874         struct rb_node *parent, **p;
 875
 876         if (bdi->dev)   /* The driver needs to use separate queues per device */
 877                 return 0;
 878
 879         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
 880         dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
 881         if (IS_ERR(dev))
 882                 return PTR_ERR(dev);
 883
 884         cgwb_bdi_register(bdi);
 885         bdi->dev = dev;
 886
 887         bdi_debug_register(bdi, dev_name(dev));
 888         set_bit(WB_registered, &bdi->wb.state);
 889
 890         spin_lock_bh(&bdi_lock);
 891
 892         bdi->id = ++bdi_id_cursor;
 893
 894         p = bdi_lookup_rb_node(bdi->id, &parent);
 895         rb_link_node(&bdi->rb_node, parent, p);
 896         rb_insert_color(&bdi->rb_node, &bdi_tree);
 897
 898         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 899
 900         spin_unlock_bh(&bdi_lock);
 901
 902         trace_writeback_bdi_register(bdi);
 903         return 0;
 904 }
 905
 906 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 907 {
 908         va_list args;
 909         int ret;
 910
 911         va_start(args, fmt);
 912         ret = bdi_register_va(bdi, fmt, args);
 913         va_end(args);
 914         return ret;
 915 }
 916 EXPORT_SYMBOL(bdi_register);
 917
 918 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
 919 {
 920         WARN_ON_ONCE(bdi->owner);
 921         bdi->owner = owner;
 922         get_device(owner);
 923 }
 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         rb_erase(&bdi->rb_node, &bdi_tree);
 932         list_del_rcu(&bdi->bdi_list);
 933         spin_unlock_bh(&bdi_lock);
 934
 935         synchronize_rcu_expedited();
 936 }
 937
 938 void bdi_unregister(struct backing_dev_info *bdi)
 939 {
 940         del_timer_sync(&bdi->laptop_mode_wb_timer);
 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         /*
 948          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
 949          * update the global bdi_min_ratio.
 950          */
 951         if (bdi->min_ratio)
 952                 bdi_set_min_ratio(bdi, 0);
 953
 954         if (bdi->dev) {
 955                 bdi_debug_unregister(bdi);
 956                 device_unregister(bdi->dev);
 957                 bdi->dev = NULL;
 958         }
 959
 960         if (bdi->owner) {
 961                 put_device(bdi->owner);
 962                 bdi->owner = NULL;
 963         }
 964 }
 965 EXPORT_SYMBOL(bdi_unregister);
 966
 967 static void release_bdi(struct kref *ref)
 968 {
 969         struct backing_dev_info *bdi =
 970                         container_of(ref, struct backing_dev_info, refcnt);
 971
 972         WARN_ON_ONCE(test_bit(WB_registered, &bdi->wb.state));
 973         WARN_ON_ONCE(bdi->dev);
 974         wb_exit(&bdi->wb);
 975         kfree(bdi);
 976 }
 977
 978 void bdi_put(struct backing_dev_info *bdi)
 979 {
 980         kref_put(&bdi->refcnt, release_bdi);
 981 }
 982 EXPORT_SYMBOL(bdi_put);
 983
 984 struct backing_dev_info *inode_to_bdi(struct inode *inode)
 985 {
 986         struct super_block *sb;
 987
 988         if (!inode)
 989                 return &noop_backing_dev_info;
 990
 991         sb = inode->i_sb;
 992 #ifdef CONFIG_BLOCK
 993         if (sb_is_blkdev_sb(sb))
 994                 return I_BDEV(inode)->bd_disk->bdi;
 995 #endif
 996         return sb->s_bdi;
 997 }
 998 EXPORT_SYMBOL(inode_to_bdi);
 999
1000 const char *bdi_dev_name(struct backing_dev_info *bdi)
1001 {
1002         if (!bdi || !bdi->dev)
1003                 return bdi_unknown_name;
1004         return bdi->dev_name;
1005 }
1006 EXPORT_SYMBOL_GPL(bdi_dev_name);