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 bdi_init(struct backing_dev_info *bdi);
 233
 234 static int __init default_bdi_init(void)
 235 {
 236         int err;
 237
 238         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 239                                  WQ_SYSFS, 0);
 240         if (!bdi_wq)
 241                 return -ENOMEM;
 242
 243         err = bdi_init(&noop_backing_dev_info);
 244
 245         return err;
 246 }
 247 subsys_initcall(default_bdi_init);
 248
 249 /*
 250  * This function is used when the first inode for this wb is marked dirty. It
 251  * wakes-up the corresponding bdi thread which should then take care of the
 252  * periodic background write-out of dirty inodes. Since the write-out would
 253  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 254  * set up a timer which wakes the bdi thread up later.
 255  *
 256  * Note, we wouldn't bother setting up the timer, but this function is on the
 257  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 258  * by delaying the wake-up.
 259  *
 260  * We have to be careful not to postpone flush work if it is scheduled for
 261  * earlier. Thus we use queue_delayed_work().
 262  */
 263 void wb_wakeup_delayed(struct bdi_writeback *wb)
 264 {
 265         unsigned long timeout;
 266
 267         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 268         spin_lock_bh(&wb->work_lock);
 269         if (test_bit(WB_registered, &wb->state))
 270                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 271         spin_unlock_bh(&wb->work_lock);
 272 }
 273
 274 static void wb_update_bandwidth_workfn(struct work_struct *work)
 275 {
 276         struct bdi_writeback *wb = container_of(to_delayed_work(work),
 277                                                 struct bdi_writeback, bw_dwork);
 278
 279         wb_update_bandwidth(wb);
 280 }
 281
 282 /*
 283  * Initial write bandwidth: 100 MB/s
 284  */
 285 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 286
 287 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 288                    gfp_t gfp)
 289 {
 290         int i, err;
 291
 292         memset(wb, 0, sizeof(*wb));
 293
 294         if (wb != &bdi->wb)
 295                 bdi_get(bdi);
 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                 goto out_put_bdi;
 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 out_put_bdi:
 334         if (wb != &bdi->wb)
 335                 bdi_put(bdi);
 336         return err;
 337 }
 338
 339 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 340
 341 /*
 342  * Remove bdi from the global list and shutdown any threads we have running
 343  */
 344 static void wb_shutdown(struct bdi_writeback *wb)
 345 {
 346         /* Make sure nobody queues further work */
 347         spin_lock_bh(&wb->work_lock);
 348         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 349                 spin_unlock_bh(&wb->work_lock);
 350                 return;
 351         }
 352         spin_unlock_bh(&wb->work_lock);
 353
 354         cgwb_remove_from_bdi_list(wb);
 355         /*
 356          * Drain work list and shutdown the delayed_work.  !WB_registered
 357          * tells wb_workfn() that @wb is dying and its work_list needs to
 358          * be drained no matter what.
 359          */
 360         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 361         flush_delayed_work(&wb->dwork);
 362         WARN_ON(!list_empty(&wb->work_list));
 363         flush_delayed_work(&wb->bw_dwork);
 364 }
 365
 366 static void wb_exit(struct bdi_writeback *wb)
 367 {
 368         int i;
 369
 370         WARN_ON(delayed_work_pending(&wb->dwork));
 371
 372         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 373                 percpu_counter_destroy(&wb->stat[i]);
 374
 375         fprop_local_destroy_percpu(&wb->completions);
 376         if (wb != &wb->bdi->wb)
 377                 bdi_put(wb->bdi);
 378 }
 379
 380 #ifdef CONFIG_CGROUP_WRITEBACK
 381
 382 #include <linux/memcontrol.h>
 383
 384 /*
 385  * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and
 386  * memcg->cgwb_list.  bdi->cgwb_tree is also RCU protected.
 387  */
 388 static DEFINE_SPINLOCK(cgwb_lock);
 389 static struct workqueue_struct *cgwb_release_wq;
 390
 391 static LIST_HEAD(offline_cgwbs);
 392 static void cleanup_offline_cgwbs_workfn(struct work_struct *work);
 393 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn);
 394
 395 static void cgwb_release_workfn(struct work_struct *work)
 396 {
 397         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 398                                                 release_work);
 399         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 400
 401         mutex_lock(&wb->bdi->cgwb_release_mutex);
 402         wb_shutdown(wb);
 403
 404         css_put(wb->memcg_css);
 405         css_put(wb->blkcg_css);
 406         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 407
 408         /* triggers blkg destruction if no online users left */
 409         blkcg_unpin_online(blkcg);
 410
 411         fprop_local_destroy_percpu(&wb->memcg_completions);
 412
 413         spin_lock_irq(&cgwb_lock);
 414         list_del(&wb->offline_node);
 415         spin_unlock_irq(&cgwb_lock);
 416
 417         percpu_ref_exit(&wb->refcnt);
 418         wb_exit(wb);
 419         WARN_ON_ONCE(!list_empty(&wb->b_attached));
 420         kfree_rcu(wb, rcu);
 421 }
 422
 423 static void cgwb_release(struct percpu_ref *refcnt)
 424 {
 425         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 426                                                 refcnt);
 427         queue_work(cgwb_release_wq, &wb->release_work);
 428 }
 429
 430 static void cgwb_kill(struct bdi_writeback *wb)
 431 {
 432         lockdep_assert_held(&cgwb_lock);
 433
 434         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 435         list_del(&wb->memcg_node);
 436         list_del(&wb->blkcg_node);
 437         list_add(&wb->offline_node, &offline_cgwbs);
 438         percpu_ref_kill(&wb->refcnt);
 439 }
 440
 441 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 442 {
 443         spin_lock_irq(&cgwb_lock);
 444         list_del_rcu(&wb->bdi_node);
 445         spin_unlock_irq(&cgwb_lock);
 446 }
 447
 448 static int cgwb_create(struct backing_dev_info *bdi,
 449                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 450 {
 451         struct mem_cgroup *memcg;
 452         struct cgroup_subsys_state *blkcg_css;
 453         struct blkcg *blkcg;
 454         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 455         struct bdi_writeback *wb;
 456         unsigned long flags;
 457         int ret = 0;
 458
 459         memcg = mem_cgroup_from_css(memcg_css);
 460         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 461         blkcg = css_to_blkcg(blkcg_css);
 462         memcg_cgwb_list = &memcg->cgwb_list;
 463         blkcg_cgwb_list = &blkcg->cgwb_list;
 464
 465         /* look up again under lock and discard on blkcg mismatch */
 466         spin_lock_irqsave(&cgwb_lock, flags);
 467         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 468         if (wb && wb->blkcg_css != blkcg_css) {
 469                 cgwb_kill(wb);
 470                 wb = NULL;
 471         }
 472         spin_unlock_irqrestore(&cgwb_lock, flags);
 473         if (wb)
 474                 goto out_put;
 475
 476         /* need to create a new one */
 477         wb = kmalloc(sizeof(*wb), gfp);
 478         if (!wb) {
 479                 ret = -ENOMEM;
 480                 goto out_put;
 481         }
 482
 483         ret = wb_init(wb, bdi, gfp);
 484         if (ret)
 485                 goto err_free;
 486
 487         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 488         if (ret)
 489                 goto err_wb_exit;
 490
 491         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 492         if (ret)
 493                 goto err_ref_exit;
 494
 495         wb->memcg_css = memcg_css;
 496         wb->blkcg_css = blkcg_css;
 497         INIT_LIST_HEAD(&wb->b_attached);
 498         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 499         set_bit(WB_registered, &wb->state);
 500
 501         /*
 502          * The root wb determines the registered state of the whole bdi and
 503          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 504          * whether they're still online.  Don't link @wb if any is dead.
 505          * See wb_memcg_offline() and wb_blkcg_offline().
 506          */
 507         ret = -ENODEV;
 508         spin_lock_irqsave(&cgwb_lock, flags);
 509         if (test_bit(WB_registered, &bdi->wb.state) &&
 510             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 511                 /* we might have raced another instance of this function */
 512                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 513                 if (!ret) {
 514                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 515                         list_add(&wb->memcg_node, memcg_cgwb_list);
 516                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 517                         blkcg_pin_online(blkcg);
 518                         css_get(memcg_css);
 519                         css_get(blkcg_css);
 520                 }
 521         }
 522         spin_unlock_irqrestore(&cgwb_lock, flags);
 523         if (ret) {
 524                 if (ret == -EEXIST)
 525                         ret = 0;
 526                 goto err_fprop_exit;
 527         }
 528         goto out_put;
 529
 530 err_fprop_exit:
 531         fprop_local_destroy_percpu(&wb->memcg_completions);
 532 err_ref_exit:
 533         percpu_ref_exit(&wb->refcnt);
 534 err_wb_exit:
 535         wb_exit(wb);
 536 err_free:
 537         kfree(wb);
 538 out_put:
 539         css_put(blkcg_css);
 540         return ret;
 541 }
 542
 543 /**
 544  * wb_get_lookup - get wb for a given memcg
 545  * @bdi: target bdi
 546  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 547  *
 548  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 549  * refcount incremented.
 550  *
 551  * This function uses css_get() on @memcg_css and thus expects its refcnt
 552  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 553  * @memcg_css isn't enough.  try_get it before calling this function.
 554  *
 555  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 556  * memcg on the default hierarchy, memcg association is guaranteed to be
 557  * more specific (equal or descendant to the associated blkcg) and thus can
 558  * identify both the memcg and blkcg associations.
 559  *
 560  * Because the blkcg associated with a memcg may change as blkcg is enabled
 561  * and disabled closer to root in the hierarchy, each wb keeps track of
 562  * both the memcg and blkcg associated with it and verifies the blkcg on
 563  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 564  * created.
 565  */
 566 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 567                                     struct cgroup_subsys_state *memcg_css)
 568 {
 569         struct bdi_writeback *wb;
 570
 571         if (!memcg_css->parent)
 572                 return &bdi->wb;
 573
 574         rcu_read_lock();
 575         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 576         if (wb) {
 577                 struct cgroup_subsys_state *blkcg_css;
 578
 579                 /* see whether the blkcg association has changed */
 580                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 581                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 582                         wb = NULL;
 583                 css_put(blkcg_css);
 584         }
 585         rcu_read_unlock();
 586
 587         return wb;
 588 }
 589
 590 /**
 591  * wb_get_create - get wb for a given memcg, create if necessary
 592  * @bdi: target bdi
 593  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 594  * @gfp: allocation mask to use
 595  *
 596  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 597  * create one.  See wb_get_lookup() for more details.
 598  */
 599 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 600                                     struct cgroup_subsys_state *memcg_css,
 601                                     gfp_t gfp)
 602 {
 603         struct bdi_writeback *wb;
 604
 605         might_alloc(gfp);
 606
 607         if (!memcg_css->parent)
 608                 return &bdi->wb;
 609
 610         do {
 611                 wb = wb_get_lookup(bdi, memcg_css);
 612         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 613
 614         return wb;
 615 }
 616
 617 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 618 {
 619         int ret;
 620
 621         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 622         mutex_init(&bdi->cgwb_release_mutex);
 623         init_rwsem(&bdi->wb_switch_rwsem);
 624
 625         ret = wb_init(&bdi->wb, bdi, GFP_KERNEL);
 626         if (!ret) {
 627                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 628                 bdi->wb.blkcg_css = blkcg_root_css;
 629         }
 630         return ret;
 631 }
 632
 633 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 634 {
 635         struct radix_tree_iter iter;
 636         void **slot;
 637         struct bdi_writeback *wb;
 638
 639         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 640
 641         spin_lock_irq(&cgwb_lock);
 642         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 643                 cgwb_kill(*slot);
 644         spin_unlock_irq(&cgwb_lock);
 645
 646         mutex_lock(&bdi->cgwb_release_mutex);
 647         spin_lock_irq(&cgwb_lock);
 648         while (!list_empty(&bdi->wb_list)) {
 649                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 650                                       bdi_node);
 651                 spin_unlock_irq(&cgwb_lock);
 652                 wb_shutdown(wb);
 653                 spin_lock_irq(&cgwb_lock);
 654         }
 655         spin_unlock_irq(&cgwb_lock);
 656         mutex_unlock(&bdi->cgwb_release_mutex);
 657 }
 658
 659 /*
 660  * cleanup_offline_cgwbs_workfn - try to release dying cgwbs
 661  *
 662  * Try to release dying cgwbs by switching attached inodes to the nearest
 663  * living ancestor's writeback. Processed wbs are placed at the end
 664  * of the list to guarantee the forward progress.
 665  */
 666 static void cleanup_offline_cgwbs_workfn(struct work_struct *work)
 667 {
 668         struct bdi_writeback *wb;
 669         LIST_HEAD(processed);
 670
 671         spin_lock_irq(&cgwb_lock);
 672
 673         while (!list_empty(&offline_cgwbs)) {
 674                 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback,
 675                                       offline_node);
 676                 list_move(&wb->offline_node, &processed);
 677
 678                 /*
 679                  * If wb is dirty, cleaning up the writeback by switching
 680                  * attached inodes will result in an effective removal of any
 681                  * bandwidth restrictions, which isn't the goal.  Instead,
 682                  * it can be postponed until the next time, when all io
 683                  * will be likely completed.  If in the meantime some inodes
 684                  * will get re-dirtied, they should be eventually switched to
 685                  * a new cgwb.
 686                  */
 687                 if (wb_has_dirty_io(wb))
 688                         continue;
 689
 690                 if (!wb_tryget(wb))
 691                         continue;
 692
 693                 spin_unlock_irq(&cgwb_lock);
 694                 while (cleanup_offline_cgwb(wb))
 695                         cond_resched();
 696                 spin_lock_irq(&cgwb_lock);
 697
 698                 wb_put(wb);
 699         }
 700
 701         if (!list_empty(&processed))
 702                 list_splice_tail(&processed, &offline_cgwbs);
 703
 704         spin_unlock_irq(&cgwb_lock);
 705 }
 706
 707 /**
 708  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 709  * @memcg: memcg being offlined
 710  *
 711  * Also prevents creation of any new wb's associated with @memcg.
 712  */
 713 void wb_memcg_offline(struct mem_cgroup *memcg)
 714 {
 715         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 716         struct bdi_writeback *wb, *next;
 717
 718         spin_lock_irq(&cgwb_lock);
 719         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 720                 cgwb_kill(wb);
 721         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 722         spin_unlock_irq(&cgwb_lock);
 723
 724         queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work);
 725 }
 726
 727 /**
 728  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 729  * @blkcg: blkcg being offlined
 730  *
 731  * Also prevents creation of any new wb's associated with @blkcg.
 732  */
 733 void wb_blkcg_offline(struct blkcg *blkcg)
 734 {
 735         struct bdi_writeback *wb, *next;
 736
 737         spin_lock_irq(&cgwb_lock);
 738         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 739                 cgwb_kill(wb);
 740         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 741         spin_unlock_irq(&cgwb_lock);
 742 }
 743
 744 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 745 {
 746         spin_lock_irq(&cgwb_lock);
 747         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 748         spin_unlock_irq(&cgwb_lock);
 749 }
 750
 751 static int __init cgwb_init(void)
 752 {
 753         /*
 754          * There can be many concurrent release work items overwhelming
 755          * system_wq.  Put them in a separate wq and limit concurrency.
 756          * There's no point in executing many of these in parallel.
 757          */
 758         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 759         if (!cgwb_release_wq)
 760                 return -ENOMEM;
 761
 762         return 0;
 763 }
 764 subsys_initcall(cgwb_init);
 765
 766 #else   /* CONFIG_CGROUP_WRITEBACK */
 767
 768 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 769 {
 770         return wb_init(&bdi->wb, bdi, GFP_KERNEL);
 771 }
 772
 773 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 774
 775 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 776 {
 777         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 778 }
 779
 780 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 781 {
 782         list_del_rcu(&wb->bdi_node);
 783 }
 784
 785 #endif  /* CONFIG_CGROUP_WRITEBACK */
 786
 787 static int bdi_init(struct backing_dev_info *bdi)
 788 {
 789         int ret;
 790
 791         bdi->dev = NULL;
 792
 793         kref_init(&bdi->refcnt);
 794         bdi->min_ratio = 0;
 795         bdi->max_ratio = 100;
 796         bdi->max_prop_frac = FPROP_FRAC_BASE;
 797         INIT_LIST_HEAD(&bdi->bdi_list);
 798         INIT_LIST_HEAD(&bdi->wb_list);
 799         init_waitqueue_head(&bdi->wb_waitq);
 800
 801         ret = cgwb_bdi_init(bdi);
 802
 803         return ret;
 804 }
 805
 806 struct backing_dev_info *bdi_alloc(int node_id)
 807 {
 808         struct backing_dev_info *bdi;
 809
 810         bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id);
 811         if (!bdi)
 812                 return NULL;
 813
 814         if (bdi_init(bdi)) {
 815                 kfree(bdi);
 816                 return NULL;
 817         }
 818         bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT;
 819         bdi->ra_pages = VM_READAHEAD_PAGES;
 820         bdi->io_pages = VM_READAHEAD_PAGES;
 821         timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0);
 822         return bdi;
 823 }
 824 EXPORT_SYMBOL(bdi_alloc);
 825
 826 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 827 {
 828         struct rb_node **p = &bdi_tree.rb_node;
 829         struct rb_node *parent = NULL;
 830         struct backing_dev_info *bdi;
 831
 832         lockdep_assert_held(&bdi_lock);
 833
 834         while (*p) {
 835                 parent = *p;
 836                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 837
 838                 if (bdi->id > id)
 839                         p = &(*p)->rb_left;
 840                 else if (bdi->id < id)
 841                         p = &(*p)->rb_right;
 842                 else
 843                         break;
 844         }
 845
 846         if (parentp)
 847                 *parentp = parent;
 848         return p;
 849 }
 850
 851 /**
 852  * bdi_get_by_id - lookup and get bdi from its id
 853  * @id: bdi id to lookup
 854  *
 855  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 856  * doesn't exist or is already unregistered.
 857  */
 858 struct backing_dev_info *bdi_get_by_id(u64 id)
 859 {
 860         struct backing_dev_info *bdi = NULL;
 861         struct rb_node **p;
 862
 863         spin_lock_bh(&bdi_lock);
 864         p = bdi_lookup_rb_node(id, NULL);
 865         if (*p) {
 866                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 867                 bdi_get(bdi);
 868         }
 869         spin_unlock_bh(&bdi_lock);
 870
 871         return bdi;
 872 }
 873
 874 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 875 {
 876         struct device *dev;
 877         struct rb_node *parent, **p;
 878
 879         if (bdi->dev)   /* The driver needs to use separate queues per device */
 880                 return 0;
 881
 882         vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args);
 883         dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name);
 884         if (IS_ERR(dev))
 885                 return PTR_ERR(dev);
 886
 887         cgwb_bdi_register(bdi);
 888         bdi->dev = dev;
 889
 890         bdi_debug_register(bdi, dev_name(dev));
 891         set_bit(WB_registered, &bdi->wb.state);
 892
 893         spin_lock_bh(&bdi_lock);
 894
 895         bdi->id = ++bdi_id_cursor;
 896
 897         p = bdi_lookup_rb_node(bdi->id, &parent);
 898         rb_link_node(&bdi->rb_node, parent, p);
 899         rb_insert_color(&bdi->rb_node, &bdi_tree);
 900
 901         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 902
 903         spin_unlock_bh(&bdi_lock);
 904
 905         trace_writeback_bdi_register(bdi);
 906         return 0;
 907 }
 908
 909 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 910 {
 911         va_list args;
 912         int ret;
 913
 914         va_start(args, fmt);
 915         ret = bdi_register_va(bdi, fmt, args);
 916         va_end(args);
 917         return ret;
 918 }
 919 EXPORT_SYMBOL(bdi_register);
 920
 921 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner)
 922 {
 923         WARN_ON_ONCE(bdi->owner);
 924         bdi->owner = owner;
 925         get_device(owner);
 926 }
 927
 928 /*
 929  * Remove bdi from bdi_list, and ensure that it is no longer visible
 930  */
 931 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 932 {
 933         spin_lock_bh(&bdi_lock);
 934         rb_erase(&bdi->rb_node, &bdi_tree);
 935         list_del_rcu(&bdi->bdi_list);
 936         spin_unlock_bh(&bdi_lock);
 937
 938         synchronize_rcu_expedited();
 939 }
 940
 941 void bdi_unregister(struct backing_dev_info *bdi)
 942 {
 943         del_timer_sync(&bdi->laptop_mode_wb_timer);
 944
 945         /* make sure nobody finds us on the bdi_list anymore */
 946         bdi_remove_from_list(bdi);
 947         wb_shutdown(&bdi->wb);
 948         cgwb_bdi_unregister(bdi);
 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);