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/mm.h>
  12 #include <linux/sched.h>
  13 #include <linux/module.h>
  14 #include <linux/writeback.h>
  15 #include <linux/device.h>
  16 #include <trace/events/writeback.h>
  17
  18 struct backing_dev_info noop_backing_dev_info;
  19 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  20
  21 static struct class *bdi_class;
  22 static const char *bdi_unknown_name = "(unknown)";
  23
  24 /*
  25  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
  26  * reader side locking.
  27  */
  28 DEFINE_SPINLOCK(bdi_lock);
  29 static u64 bdi_id_cursor;
  30 static struct rb_root bdi_tree = RB_ROOT;
  31 LIST_HEAD(bdi_list);
  32
  33 /* bdi_wq serves all asynchronous writeback tasks */
  34 struct workqueue_struct *bdi_wq;
  35
  36 #define K(x) ((x) << (PAGE_SHIFT - 10))
  37
  38 #ifdef CONFIG_DEBUG_FS
  39 #include <linux/debugfs.h>
  40 #include <linux/seq_file.h>
  41
  42 static struct dentry *bdi_debug_root;
  43
  44 static void bdi_debug_init(void)
  45 {
  46         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  47 }
  48
  49 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  50 {
  51         struct backing_dev_info *bdi = m->private;
  52         struct bdi_writeback *wb = &bdi->wb;
  53         unsigned long background_thresh;
  54         unsigned long dirty_thresh;
  55         unsigned long wb_thresh;
  56         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  57         struct inode *inode;
  58
  59         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  60         spin_lock(&wb->list_lock);
  61         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  62                 nr_dirty++;
  63         list_for_each_entry(inode, &wb->b_io, i_io_list)
  64                 nr_io++;
  65         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  66                 nr_more_io++;
  67         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  68                 if (inode->i_state & I_DIRTY_TIME)
  69                         nr_dirty_time++;
  70         spin_unlock(&wb->list_lock);
  71
  72         global_dirty_limits(&background_thresh, &dirty_thresh);
  73         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  74
  75         seq_printf(m,
  76                    "BdiWriteback:       %10lu kB\n"
  77                    "BdiReclaimable:     %10lu kB\n"
  78                    "BdiDirtyThresh:     %10lu kB\n"
  79                    "DirtyThresh:        %10lu kB\n"
  80                    "BackgroundThresh:   %10lu kB\n"
  81                    "BdiDirtied:         %10lu kB\n"
  82                    "BdiWritten:         %10lu kB\n"
  83                    "BdiWriteBandwidth:  %10lu kBps\n"
  84                    "b_dirty:            %10lu\n"
  85                    "b_io:               %10lu\n"
  86                    "b_more_io:          %10lu\n"
  87                    "b_dirty_time:       %10lu\n"
  88                    "bdi_list:           %10u\n"
  89                    "state:              %10lx\n",
  90                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  91                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  92                    K(wb_thresh),
  93                    K(dirty_thresh),
  94                    K(background_thresh),
  95                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  96                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  97                    (unsigned long) K(wb->write_bandwidth),
  98                    nr_dirty,
  99                    nr_io,
 100                    nr_more_io,
 101                    nr_dirty_time,
 102                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 103
 104         return 0;
 105 }
 106 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 107
 108 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 109 {
 110         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 111
 112         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 113                             &bdi_debug_stats_fops);
 114 }
 115
 116 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 117 {
 118         debugfs_remove_recursive(bdi->debug_dir);
 119 }
 120 #else
 121 static inline void bdi_debug_init(void)
 122 {
 123 }
 124 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 125                                       const char *name)
 126 {
 127 }
 128 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 129 {
 130 }
 131 #endif
 132
 133 static ssize_t read_ahead_kb_store(struct device *dev,
 134                                   struct device_attribute *attr,
 135                                   const char *buf, size_t count)
 136 {
 137         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 138         unsigned long read_ahead_kb;
 139         ssize_t ret;
 140
 141         ret = kstrtoul(buf, 10, &read_ahead_kb);
 142         if (ret < 0)
 143                 return ret;
 144
 145         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 146
 147         return count;
 148 }
 149
 150 #define BDI_SHOW(name, expr)                                            \
 151 static ssize_t name##_show(struct device *dev,                          \
 152                            struct device_attribute *attr, char *buf)    \
 153 {                                                                       \
 154         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 155                                                                         \
 156         return sysfs_emit(buf, "%lld\n", (long long)expr);              \
 157 }                                                                       \
 158 static DEVICE_ATTR_RW(name);
 159
 160 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 161
 162 static ssize_t min_ratio_store(struct device *dev,
 163                 struct device_attribute *attr, const char *buf, size_t count)
 164 {
 165         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 166         unsigned int ratio;
 167         ssize_t ret;
 168
 169         ret = kstrtouint(buf, 10, &ratio);
 170         if (ret < 0)
 171                 return ret;
 172
 173         ret = bdi_set_min_ratio(bdi, ratio);
 174         if (!ret)
 175                 ret = count;
 176
 177         return ret;
 178 }
 179 BDI_SHOW(min_ratio, bdi->min_ratio)
 180
 181 static ssize_t max_ratio_store(struct device *dev,
 182                 struct device_attribute *attr, const char *buf, size_t count)
 183 {
 184         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 185         unsigned int ratio;
 186         ssize_t ret;
 187
 188         ret = kstrtouint(buf, 10, &ratio);
 189         if (ret < 0)
 190                 return ret;
 191
 192         ret = bdi_set_max_ratio(bdi, ratio);
 193         if (!ret)
 194                 ret = count;
 195
 196         return ret;
 197 }
 198 BDI_SHOW(max_ratio, bdi->max_ratio)
 199
 200 static ssize_t stable_pages_required_show(struct device *dev,
 201                                           struct device_attribute *attr,
 202                                           char *buf)
 203 {
 204         dev_warn_once(dev,
 205                 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n");
 206         return sysfs_emit(buf, "%d\n", 0);
 207 }
 208 static DEVICE_ATTR_RO(stable_pages_required);
 209
 210 static struct attribute *bdi_dev_attrs[] = {
 211         &dev_attr_read_ahead_kb.attr,
 212         &dev_attr_min_ratio.attr,
 213         &dev_attr_max_ratio.attr,
 214         &dev_attr_stable_pages_required.attr,
 215         NULL,
 216 };
 217 ATTRIBUTE_GROUPS(bdi_dev);
 218
 219 static __init int bdi_class_init(void)
 220 {
 221         bdi_class = class_create(THIS_MODULE, "bdi");
 222         if (IS_ERR(bdi_class))
 223                 return PTR_ERR(bdi_class);
 224
 225         bdi_class->dev_groups = bdi_dev_groups;
 226         bdi_debug_init();
 227
 228         return 0;
 229 }
 230 postcore_initcall(bdi_class_init);
 231
 232 static int __init default_bdi_init(void)
 233 {
 234         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 235                                  WQ_SYSFS, 0);
 236         if (!bdi_wq)
 237                 return -ENOMEM;
 238         return 0;
 239 }
 240 subsys_initcall(default_bdi_init);
 241
 242 /*
 243  * This function is used when the first inode for this wb is marked dirty. It
 244  * wakes-up the corresponding bdi thread which should then take care of the
 245  * periodic background write-out of dirty inodes. Since the write-out would
 246  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 247  * set up a timer which wakes the bdi thread up later.
 248  *
 249  * Note, we wouldn't bother setting up the timer, but this function is on the
 250  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 251  * by delaying the wake-up.
 252  *
 253  * We have to be careful not to postpone flush work if it is scheduled for
 254  * earlier. Thus we use queue_delayed_work().
 255  */
 256 void wb_wakeup_delayed(struct bdi_writeback *wb)
 257 {
 258         unsigned long timeout;
 259
 260         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 261         spin_lock_irq(&wb->work_lock);
 262         if (test_bit(WB_registered, &wb->state))
 263                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 264         spin_unlock_irq(&wb->work_lock);
 265 }
 266
 267 static void wb_update_bandwidth_workfn(struct work_struct *work)
 268 {
 269         struct bdi_writeback *wb = container_of(to_delayed_work(work),
 270                                                 struct bdi_writeback, bw_dwork);
 271
 272         wb_update_bandwidth(wb);
 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         atomic_set(&wb->writeback_inodes, 0);
 298         wb->bw_time_stamp = jiffies;
 299         wb->balanced_dirty_ratelimit = INIT_BW;
 300         wb->dirty_ratelimit = INIT_BW;
 301         wb->write_bandwidth = INIT_BW;
 302         wb->avg_write_bandwidth = INIT_BW;
 303
 304         spin_lock_init(&wb->work_lock);
 305         INIT_LIST_HEAD(&wb->work_list);
 306         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 307         INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn);
 308         wb->dirty_sleep = jiffies;
 309
 310         err = fprop_local_init_percpu(&wb->completions, gfp);
 311         if (err)
 312                 goto out_put_bdi;
 313
 314         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 315                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 316                 if (err)
 317                         goto out_destroy_stat;
 318         }
 319
 320         return 0;
 321
 322 out_destroy_stat:
 323         while (i--)
 324                 percpu_counter_destroy(&wb->stat[i]);
 325         fprop_local_destroy_percpu(&wb->completions);
 326 out_put_bdi:
 327         if (wb != &bdi->wb)
 328                 bdi_put(bdi);
 329         return err;
 330 }
 331
 332 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 333
 334 /*
 335  * Remove bdi from the global list and shutdown any threads we have running
 336  */
 337 static void wb_shutdown(struct bdi_writeback *wb)
 338 {
 339         /* Make sure nobody queues further work */
 340         spin_lock_irq(&wb->work_lock);
 341         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 342                 spin_unlock_irq(&wb->work_lock);
 343                 return;
 344         }
 345         spin_unlock_irq(&wb->work_lock);
 346
 347         cgwb_remove_from_bdi_list(wb);
 348         /*
 349          * Drain work list and shutdown the delayed_work.  !WB_registered
 350          * tells wb_workfn() that @wb is dying and its work_list needs to
 351          * be drained no matter what.
 352          */
 353         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 354         flush_delayed_work(&wb->dwork);
 355         WARN_ON(!list_empty(&wb->work_list));
 356         flush_delayed_work(&wb->bw_dwork);
 357 }
 358
 359 static void wb_exit(struct bdi_writeback *wb)
 360 {
 361         int i;
 362
 363         WARN_ON(delayed_work_pending(&wb->dwork));
 364
 365         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 366                 percpu_counter_destroy(&wb->stat[i]);
 367
 368         fprop_local_destroy_percpu(&wb->completions);
 369         if (wb != &wb->bdi->wb)
 370                 bdi_put(wb->bdi);
 371 }
 372
 373 #ifdef CONFIG_CGROUP_WRITEBACK
 374
 375 #include <linux/memcontrol.h>
 376
 377 /*
 378  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
 379  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
 380  */
 381 static DEFINE_SPINLOCK(cgwb_lock);
 382 static struct workqueue_struct *cgwb_release_wq;
 383
 384 static LIST_HEAD(offline_cgwbs);
 385 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
 386 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
 387
 388 static void cgwb_release_workfn(struct work_struct *work)
 389 {
 390         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 391                                                 release_work);
 392         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 393
 394         mutex_lock(&wb->bdi->cgwb_release_mutex);
 395         wb_shutdown(wb);
 396
 397         css_put(wb->memcg_css);
 398         css_put(wb->blkcg_css);
 399         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 400
 401         /* triggers blkg destruction if no online users left */
 402         blkcg_unpin_online(blkcg);
 403
 404         fprop_local_destroy_percpu(&wb->memcg_completions);
 405
 406         spin_lock_irq(&cgwb_lock);
 407         list_del(&wb->offline_node);
 408         spin_unlock_irq(&cgwb_lock);
 409
 410         percpu_ref_exit(&wb->refcnt);
 411         wb_exit(wb);
 412         WARN_ON_ONCE(!list_empty(&wb->b_attached));
 413         kfree_rcu(wb, rcu);
 414 }
 415
 416 static void cgwb_release(struct percpu_ref *refcnt)
 417 {
 418         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 419                                                 refcnt);
 420         queue_work(cgwb_release_wq, &wb->release_work);
 421 }
 422
 423 static void cgwb_kill(struct bdi_writeback *wb)
 424 {
 425         lockdep_assert_held(&cgwb_lock);
 426
 427         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 428         list_del(&wb->memcg_node);
 429         list_del(&wb->blkcg_node);
 430         list_add(&wb->offline_node, &offline_cgwbs);
 431         percpu_ref_kill(&wb->refcnt);
 432 }
 433
 434 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 435 {
 436         spin_lock_irq(&cgwb_lock);
 437         list_del_rcu(&wb->bdi_node);
 438         spin_unlock_irq(&cgwb_lock);
 439 }
 440
 441 static int cgwb_create(struct backing_dev_info *bdi,
 442                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 443 {
 444         struct mem_cgroup *memcg;
 445         struct cgroup_subsys_state *blkcg_css;
 446         struct blkcg *blkcg;
 447         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 448         struct bdi_writeback *wb;
 449         unsigned long flags;
 450         int ret = 0;
 451
 452         memcg = mem_cgroup_from_css(memcg_css);
 453         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 454         blkcg = css_to_blkcg(blkcg_css);
 455         memcg_cgwb_list = &memcg->cgwb_list;
 456         blkcg_cgwb_list = &blkcg->cgwb_list;
 457
 458         /* look up again under lock and discard on blkcg mismatch */
 459         spin_lock_irqsave(&cgwb_lock, flags);
 460         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 461         if (wb && wb->blkcg_css != blkcg_css) {
 462                 cgwb_kill(wb);
 463                 wb = NULL;
 464         }
 465         spin_unlock_irqrestore(&cgwb_lock, flags);
 466         if (wb)
 467                 goto out_put;
 468
 469         /* need to create a new one */
 470         wb = kmalloc(sizeof(*wb), gfp);
 471         if (!wb) {
 472                 ret = -ENOMEM;
 473                 goto out_put;
 474         }
 475
 476         ret = wb_init(wb, bdi, gfp);
 477         if (ret)
 478                 goto err_free;
 479
 480         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 481         if (ret)
 482                 goto err_wb_exit;
 483
 484         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 485         if (ret)
 486                 goto err_ref_exit;
 487
 488         wb->memcg_css = memcg_css;
 489         wb->blkcg_css = blkcg_css;
 490         INIT_LIST_HEAD(&wb->b_attached);
 491         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 492         set_bit(WB_registered, &wb->state);
 493
 494         /*
 495          * The root wb determines the registered state of the whole bdi and
 496          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 497          * whether they're still online.  Don't link @wb if any is dead.
 498          * See wb_memcg_offline() and wb_blkcg_offline().
 499          */
 500         ret = -ENODEV;
 501         spin_lock_irqsave(&cgwb_lock, flags);
 502         if (test_bit(WB_registered, &bdi->wb.state) &&
 503             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 504                 /* we might have raced another instance of this function */
 505                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 506                 if (!ret) {
 507                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 508                         list_add(&wb->memcg_node, memcg_cgwb_list);
 509                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 510                         blkcg_pin_online(blkcg);
 511                         css_get(memcg_css);
 512                         css_get(blkcg_css);
 513                 }
 514         }
 515         spin_unlock_irqrestore(&cgwb_lock, flags);
 516         if (ret) {
 517                 if (ret == -EEXIST)
 518                         ret = 0;
 519                 goto err_fprop_exit;
 520         }
 521         goto out_put;
 522
 523 err_fprop_exit:
 524         fprop_local_destroy_percpu(&wb->memcg_completions);
 525 err_ref_exit:
 526         percpu_ref_exit(&wb->refcnt);
 527 err_wb_exit:
 528         wb_exit(wb);
 529 err_free:
 530         kfree(wb);
 531 out_put:
 532         css_put(blkcg_css);
 533         return ret;
 534 }
 535
 536 /**
 537  * wb_get_lookup - get wb for a given memcg
 538  * @bdi: target bdi
 539  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 540  *
 541  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 542  * refcount incremented.
 543  *
 544  * This function uses css_get() on @memcg_css and thus expects its refcnt
 545  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 546  * @memcg_css isn't enough.  try_get it before calling this function.
 547  *
 548  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 549  * memcg on the default hierarchy, memcg association is guaranteed to be
 550  * more specific (equal or descendant to the associated blkcg) and thus can
 551  * identify both the memcg and blkcg associations.
 552  *
 553  * Because the blkcg associated with a memcg may change as blkcg is enabled
 554  * and disabled closer to root in the hierarchy, each wb keeps track of
 555  * both the memcg and blkcg associated with it and verifies the blkcg on
 556  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 557  * created.
 558  */
 559 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 560                                     struct cgroup_subsys_state *memcg_css)
 561 {
 562         struct bdi_writeback *wb;
 563
 564         if (!memcg_css->parent)
 565                 return &bdi->wb;
 566
 567         rcu_read_lock();
 568         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 569         if (wb) {
 570                 struct cgroup_subsys_state *blkcg_css;
 571
 572                 /* see whether the blkcg association has changed */
 573                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 574                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 575                         wb = NULL;
 576                 css_put(blkcg_css);
 577         }
 578         rcu_read_unlock();
 579
 580         return wb;
 581 }
 582
 583 /**
 584  * wb_get_create - get wb for a given memcg, create if necessary
 585  * @bdi: target bdi
 586  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 587  * @gfp: allocation mask to use
 588  *
 589  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 590  * create one.  See wb_get_lookup() for more details.
 591  */
 592 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 593                                     struct cgroup_subsys_state *memcg_css,
 594                                     gfp_t gfp)
 595 {
 596         struct bdi_writeback *wb;
 597
 598         might_alloc(gfp);
 599
 600         if (!memcg_css->parent)
 601                 return &bdi->wb;
 602
 603         do {
 604                 wb = wb_get_lookup(bdi, memcg_css);
 605         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 606
 607         return wb;
 608 }
 609
 610 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 611 {
 612         int ret;
 613
 614         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 615         mutex_init(&bdi->cgwb_release_mutex);
 616         init_rwsem(&bdi->wb_switch_rwsem);
 617
 618         ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 619         if (!ret) {
 620                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 621                 bdi->wb.blkcg_css = blkcg_root_css;
 622         }
 623         return ret;
 624 }
 625
 626 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 627 {
 628         struct radix_tree_iter iter;
 629         void **slot;
 630         struct bdi_writeback *wb;
 631
 632         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 633
 634         spin_lock_irq(&cgwb_lock);
 635         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 636                 cgwb_kill(*slot);
 637         spin_unlock_irq(&cgwb_lock);
 638
 639         mutex_lock(&bdi->cgwb_release_mutex);
 640         spin_lock_irq(&cgwb_lock);
 641         while (!list_empty(&bdi->wb_list)) {
 642                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 643                                       bdi_node);
 644                 spin_unlock_irq(&cgwb_lock);
 645                 wb_shutdown(wb);
 646                 spin_lock_irq(&cgwb_lock);
 647         }
 648         spin_unlock_irq(&cgwb_lock);
 649         mutex_unlock(&bdi->cgwb_release_mutex);
 650 }
 651
 652 /*
 653  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
 654  *
 655  * Try to release dying cgwbs by switching attached inodes to the nearest
 656  * living ancestor's writeback. Processed wbs are placed at the end
 657  * of the list to guarantee the forward progress.
 658  */
 659 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
 660 {
 661         struct bdi_writeback *wb;
 662         LIST_HEAD(processed);
 663
 664         spin_lock_irq(&cgwb_lock);
 665
 666         while (!list_empty(&offline_cgwbs)) {
 667                 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
 668                                       offline_node);
 669                 list_move(&wb->offline_node, &processed);
 670
 671                 /*
 672                  * If wb is dirty, cleaning up the writeback by switching
 673                  * attached inodes will result in an effective removal of any
 674                  * bandwidth restrictions, which isn't the goal.  Instead,
 675                  * it can be postponed until the next time, when all io
 676                  * will be likely completed.  If in the meantime some inodes
 677                  * will get re-dirtied, they should be eventually switched to
 678                  * a new cgwb.
 679                  */
 680                 if (wb_has_dirty_io(wb))
 681                         continue;
 682
 683                 if (!wb_tryget(wb))
 684                         continue;
 685
 686                 spin_unlock_irq(&cgwb_lock);
 687                 while (cleanup_offline_cgwb(wb))
 688                         cond_resched();
 689                 spin_lock_irq(&cgwb_lock);
 690
 691                 wb_put(wb);
 692         }
 693
 694         if (!list_empty(&processed))
 695                 list_splice_tail(&processed, &offline_cgwbs);
 696
 697         spin_unlock_irq(&cgwb_lock);
 698 }
 699
 700 /**
 701  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 702  * @memcg: memcg being offlined
 703  *
 704  * Also prevents creation of any new wb's associated with @memcg.
 705  */
 706 void wb_memcg_offline(struct mem_cgroup *memcg)
 707 {
 708         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 709         struct bdi_writeback *wb, *next;
 710
 711         spin_lock_irq(&cgwb_lock);
 712         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 713                 cgwb_kill(wb);
 714         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 715         spin_unlock_irq(&cgwb_lock);
 716
 717         queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
 718 }
 719
 720 /**
 721  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 722  * @blkcg: blkcg being offlined
 723  *
 724  * Also prevents creation of any new wb's associated with @blkcg.
 725  */
 726 void wb_blkcg_offline(struct blkcg *blkcg)
 727 {
 728         struct bdi_writeback *wb, *next;
 729
 730         spin_lock_irq(&cgwb_lock);
 731         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 732                 cgwb_kill(wb);
 733         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 734         spin_unlock_irq(&cgwb_lock);
 735 }
 736
 737 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 738 {
 739         spin_lock_irq(&cgwb_lock);
 740         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 741         spin_unlock_irq(&cgwb_lock);
 742 }
 743
 744 static int __init cgwb_init(void)
 745 {
 746         /*
 747          * There can be many concurrent release work items overwhelming
 748          * system_wq.  Put them in a separate wq and limit concurrency.
 749          * There's no point in executing many of these in parallel.
 750          */
 751         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 752         if (!cgwb_release_wq)
 753                 return -ENOMEM;
 754
 755         return 0;
 756 }
 757 subsys_initcall(cgwb_init);
 758
 759 #else   /* CONFIG_CGROUP_WRITEBACK */
 760
 761 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 762 {
 763         return wb_init(&bdi->wb, bdi, GFP_KERNEL);
 764 }
 765
 766 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 767
 768 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 769 {
 770         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 771 }
 772
 773 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 774 {
 775         list_del_rcu(&wb->bdi_node);
 776 }
 777
 778 #endif  /* CONFIG_CGROUP_WRITEBACK */
 779
 780 int bdi_init(struct backing_dev_info *bdi)
 781 {
 782         int ret;
 783
 784         bdi->dev = NULL;
 785
 786         kref_init(&bdi->refcnt);
 787         bdi->min_ratio = 0;
 788         bdi->max_ratio = 100;
 789         bdi->max_prop_frac = FPROP_FRAC_BASE;
 790         INIT_LIST_HEAD(&bdi->bdi_list);
 791         INIT_LIST_HEAD(&bdi->wb_list);
 792         init_waitqueue_head(&bdi->wb_waitq);
 793
 794         ret = cgwb_bdi_init(bdi);
 795
 796         return ret;
 797 }
 798
 799 struct backing_dev_info *bdi_alloc(int node_id)
 800 {
 801         struct backing_dev_info *bdi;
 802
 803         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 804         if (!bdi)
 805                 return NULL;
 806
 807         if (bdi_init(bdi)) {
 808                 kfree(bdi);
 809                 return NULL;
 810         }
 811         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 812         bdi->ra_pages = VM_READAHEAD_PAGES;
 813         bdi->io_pages = VM_READAHEAD_PAGES;
 814         timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
 815         return bdi;
 816 }
 817 EXPORT_SYMBOL(bdi_alloc);
 818
 819 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 820 {
 821         struct rb_node **p = &bdi_tree.rb_node;
 822         struct rb_node *parent = NULL;
 823         struct backing_dev_info *bdi;
 824
 825         lockdep_assert_held(&bdi_lock);
 826
 827         while (*p) {
 828                 parent = *p;
 829                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 830
 831                 if (bdi->id > id)
 832                         p = &(*p)->rb_left;
 833                 else if (bdi->id < id)
 834                         p = &(*p)->rb_right;
 835                 else
 836                         break;
 837         }
 838
 839         if (parentp)
 840                 *parentp = parent;
 841         return p;
 842 }
 843
 844 /**
 845  * bdi_get_by_id - lookup and get bdi from its id
 846  * @id: bdi id to lookup
 847  *
 848  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 849  * doesn't exist or is already unregistered.
 850  */
 851 struct backing_dev_info *bdi_get_by_id(u64 id)
 852 {
 853         struct backing_dev_info *bdi = NULL;
 854         struct rb_node **p;
 855
 856         spin_lock_bh(&bdi_lock);
 857         p = bdi_lookup_rb_node(id, NULL);
 858         if (*p) {
 859                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 860                 bdi_get(bdi);
 861         }
 862         spin_unlock_bh(&bdi_lock);
 863
 864         return bdi;
 865 }
 866
 867 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 868 {
 869         struct device *dev;
 870         struct rb_node *parent, **p;
 871
 872         if (bdi->dev)   /* The driver needs to use separate queues per device */
 873                 return 0;
 874
 875         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
 876         dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
 877         if (IS_ERR(dev))
 878                 return PTR_ERR(dev);
 879
 880         cgwb_bdi_register(bdi);
 881         bdi->dev = dev;
 882
 883         bdi_debug_register(bdi, dev_name(dev));
 884         set_bit(WB_registered, &bdi->wb.state);
 885
 886         spin_lock_bh(&bdi_lock);
 887
 888         bdi->id = ++bdi_id_cursor;
 889
 890         p = bdi_lookup_rb_node(bdi->id, &parent);
 891         rb_link_node(&bdi->rb_node, parent, p);
 892         rb_insert_color(&bdi->rb_node, &bdi_tree);
 893
 894         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 895
 896         spin_unlock_bh(&bdi_lock);
 897
 898         trace_writeback_bdi_register(bdi);
 899         return 0;
 900 }
 901
 902 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 903 {
 904         va_list args;
 905         int ret;
 906
 907         va_start(args, fmt);
 908         ret = bdi_register_va(bdi, fmt, args);
 909         va_end(args);
 910         return ret;
 911 }
 912 EXPORT_SYMBOL(bdi_register);
 913
 914 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
 915 {
 916         WARN_ON_ONCE(bdi->owner);
 917         bdi->owner = owner;
 918         get_device(owner);
 919 }
 920
 921 /*
 922  * Remove bdi from bdi_list, and ensure that it is no longer visible
 923  */
 924 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 925 {
 926         spin_lock_bh(&bdi_lock);
 927         rb_erase(&bdi->rb_node, &bdi_tree);
 928         list_del_rcu(&bdi->bdi_list);
 929         spin_unlock_bh(&bdi_lock);
 930
 931         synchronize_rcu_expedited();
 932 }
 933
 934 void bdi_unregister(struct backing_dev_info *bdi)
 935 {
 936         del_timer_sync(&bdi->laptop_mode_wb_timer);
 937
 938         /* make sure nobody finds us on the bdi_list anymore */
 939         bdi_remove_from_list(bdi);
 940         wb_shutdown(&bdi->wb);
 941         cgwb_bdi_unregister(bdi);
 942
 943         /*
 944          * If this BDI's min ratio has been set, use bdi_set_min_ratio() to
 945          * update the global bdi_min_ratio.
 946          */
 947         if (bdi->min_ratio)
 948                 bdi_set_min_ratio(bdi, 0);
 949
 950         if (bdi->dev) {
 951                 bdi_debug_unregister(bdi);
 952                 device_unregister(bdi->dev);
 953                 bdi->dev = NULL;
 954         }
 955
 956         if (bdi->owner) {
 957                 put_device(bdi->owner);
 958                 bdi->owner = NULL;
 959         }
 960 }
 961
 962 static void release_bdi(struct kref *ref)
 963 {
 964         struct backing_dev_info *bdi =
 965                         container_of(ref, struct backing_dev_info, refcnt);
 966
 967         if (test_bit(WB_registered, &bdi->wb.state))
 968                 bdi_unregister(bdi);
 969         WARN_ON_ONCE(bdi->dev);
 970         wb_exit(&bdi->wb);
 971         kfree(bdi);
 972 }
 973
 974 void bdi_put(struct backing_dev_info *bdi)
 975 {
 976         kref_put(&bdi->refcnt, release_bdi);
 977 }
 978 EXPORT_SYMBOL(bdi_put);
 979
 980 const char *bdi_dev_name(struct backing_dev_info *bdi)
 981 {
 982         if (!bdi || !bdi->dev)
 983                 return bdi_unknown_name;
 984         return bdi->dev_name;
 985 }
 986 EXPORT_SYMBOL_GPL(bdi_dev_name);
 987
 988 static wait_queue_head_t congestion_wqh[2] = {
 989                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 990                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 991         };
 992 static atomic_t nr_wb_congested[2];
 993
 994 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 995 {
 996         wait_queue_head_t *wqh = &congestion_wqh[sync];
 997         enum wb_congested_state bit;
 998
 999         bit = sync ? WB_sync_congested : WB_async_congested;
1000         if (test_and_clear_bit(bit, &bdi->wb.congested))
1001                 atomic_dec(&nr_wb_congested[sync]);
1002         smp_mb__after_atomic();
1003         if (waitqueue_active(wqh))
1004                 wake_up(wqh);
1005 }
1006 EXPORT_SYMBOL(clear_bdi_congested);
1007
1008 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
1009 {
1010         enum wb_congested_state bit;
1011
1012         bit = sync ? WB_sync_congested : WB_async_congested;
1013         if (!test_and_set_bit(bit, &bdi->wb.congested))
1014                 atomic_inc(&nr_wb_congested[sync]);
1015 }
1016 EXPORT_SYMBOL(set_bdi_congested);
1017
1018 /**
1019  * congestion_wait - wait for a backing_dev to become uncongested
1020  * @sync: SYNC or ASYNC IO
1021  * @timeout: timeout in jiffies
1022  *
1023  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1024  * write congestion.  If no backing_devs are congested then just wait for the
1025  * next write to be completed.
1026  */
1027 long congestion_wait(int sync, long timeout)
1028 {
1029         long ret;
1030         unsigned long start = jiffies;
1031         DEFINE_WAIT(wait);
1032         wait_queue_head_t *wqh = &congestion_wqh[sync];
1033
1034         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1035         ret = io_schedule_timeout(timeout);
1036         finish_wait(wqh, &wait);
1037
1038         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1039                                         jiffies_to_usecs(jiffies - start));
1040
1041         return ret;
1042 }
1043 EXPORT_SYMBOL(congestion_wait);
1044
1045 /**
1046  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1047  * @sync: SYNC or ASYNC IO
1048  * @timeout: timeout in jiffies
1049  *
1050  * In the event of a congested backing_dev (any backing_dev) this waits
1051  * for up to @timeout jiffies for either a BDI to exit congestion of the
1052  * given @sync queue or a write to complete.
1053  *
1054  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1055  * it is the number of jiffies that were still remaining when the function
1056  * returned. return_value == timeout implies the function did not sleep.
1057  */
1058 long wait_iff_congested(int sync, long timeout)
1059 {
1060         long ret;
1061         unsigned long start = jiffies;
1062         DEFINE_WAIT(wait);
1063         wait_queue_head_t *wqh = &congestion_wqh[sync];
1064
1065         /*
1066          * If there is no congestion, yield if necessary instead
1067          * of sleeping on the congestion queue
1068          */
1069         if (atomic_read(&nr_wb_congested[sync]) == 0) {
1070                 cond_resched();
1071
1072                 /* In case we scheduled, work out time remaining */
1073                 ret = timeout - (jiffies - start);
1074                 if (ret < 0)
1075                         ret = 0;
1076
1077                 goto out;
1078         }
1079
1080         /* Sleep until uncongested or a write happens */
1081         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1082         ret = io_schedule_timeout(timeout);
1083         finish_wait(wqh, &wait);
1084
1085 out:
1086         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1087                                         jiffies_to_usecs(jiffies - start));
1088
1089         return ret;
1090 }
1091 EXPORT_SYMBOL(wait_iff_congested);