mm/backing-dev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include <linux/wait.h>
   4 #include <linux/backing-dev.h>
   5 #include <linux/kthread.h>
   6 #include <linux/freezer.h>
   7 #include <linux/fs.h>
   8 #include <linux/pagemap.h>
   9 #include <linux/mm.h>
  10 #include <linux/sched.h>
  11 #include <linux/module.h>
  12 #include <linux/writeback.h>
  13 #include <linux/device.h>
  14 #include <trace/events/writeback.h>
  15
  16 struct backing_dev_info noop_backing_dev_info = {
  17         .name           = "noop",
  18         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
  19 };
  20 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  21
  22 static struct class *bdi_class;
  23
  24 /*
  25  * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
  26  * locking.
  27  */
  28 DEFINE_SPINLOCK(bdi_lock);
  29 LIST_HEAD(bdi_list);
  30
  31 /* bdi_wq serves all asynchronous writeback tasks */
  32 struct workqueue_struct *bdi_wq;
  33
  34 #ifdef CONFIG_DEBUG_FS
  35 #include <linux/debugfs.h>
  36 #include <linux/seq_file.h>
  37
  38 static struct dentry *bdi_debug_root;
  39
  40 static void bdi_debug_init(void)
  41 {
  42         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  43 }
  44
  45 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  46 {
  47         struct backing_dev_info *bdi = m->private;
  48         struct bdi_writeback *wb = &bdi->wb;
  49         unsigned long background_thresh;
  50         unsigned long dirty_thresh;
  51         unsigned long wb_thresh;
  52         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  53         struct inode *inode;
  54
  55         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  56         spin_lock(&wb->list_lock);
  57         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  58                 nr_dirty++;
  59         list_for_each_entry(inode, &wb->b_io, i_io_list)
  60                 nr_io++;
  61         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  62                 nr_more_io++;
  63         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  64                 if (inode->i_state & I_DIRTY_TIME)
  65                         nr_dirty_time++;
  66         spin_unlock(&wb->list_lock);
  67
  68         global_dirty_limits(&background_thresh, &dirty_thresh);
  69         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  70
  71 #define K(x) ((x) << (PAGE_SHIFT - 10))
  72         seq_printf(m,
  73                    "BdiWriteback:       %10lu kB\n"
  74                    "BdiReclaimable:     %10lu kB\n"
  75                    "BdiDirtyThresh:     %10lu kB\n"
  76                    "DirtyThresh:        %10lu kB\n"
  77                    "BackgroundThresh:   %10lu kB\n"
  78                    "BdiDirtied:         %10lu kB\n"
  79                    "BdiWritten:         %10lu kB\n"
  80                    "BdiWriteBandwidth:  %10lu kBps\n"
  81                    "b_dirty:            %10lu\n"
  82                    "b_io:               %10lu\n"
  83                    "b_more_io:          %10lu\n"
  84                    "b_dirty_time:       %10lu\n"
  85                    "bdi_list:           %10u\n"
  86                    "state:              %10lx\n",
  87                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  88                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  89                    K(wb_thresh),
  90                    K(dirty_thresh),
  91                    K(background_thresh),
  92                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  93                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  94                    (unsigned long) K(wb->write_bandwidth),
  95                    nr_dirty,
  96                    nr_io,
  97                    nr_more_io,
  98                    nr_dirty_time,
  99                    !list_empty(&bdi->bdi_list), bdi->wb.state);
 100 #undef K
 101
 102         return 0;
 103 }
 104 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 105
 106 static int bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 107 {
 108         if (!bdi_debug_root)
 109                 return -ENOMEM;
 110
 111         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 112         if (!bdi->debug_dir)
 113                 return -ENOMEM;
 114
 115         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 116                                                bdi, &bdi_debug_stats_fops);
 117         if (!bdi->debug_stats) {
 118                 debugfs_remove(bdi->debug_dir);
 119                 bdi->debug_dir = NULL;
 120                 return -ENOMEM;
 121         }
 122
 123         return 0;
 124 }
 125
 126 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 127 {
 128         debugfs_remove(bdi->debug_stats);
 129         debugfs_remove(bdi->debug_dir);
 130 }
 131 #else
 132 static inline void bdi_debug_init(void)
 133 {
 134 }
 135 static inline int bdi_debug_register(struct backing_dev_info *bdi,
 136                                       const char *name)
 137 {
 138         return 0;
 139 }
 140 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 141 {
 142 }
 143 #endif
 144
 145 static ssize_t read_ahead_kb_store(struct device *dev,
 146                                   struct device_attribute *attr,
 147                                   const char *buf, size_t count)
 148 {
 149         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 150         unsigned long read_ahead_kb;
 151         ssize_t ret;
 152
 153         ret = kstrtoul(buf, 10, &read_ahead_kb);
 154         if (ret < 0)
 155                 return ret;
 156
 157         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 158
 159         return count;
 160 }
 161
 162 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 163
 164 #define BDI_SHOW(name, expr)                                            \
 165 static ssize_t name##_show(struct device *dev,                          \
 166                            struct device_attribute *attr, char *page)   \
 167 {                                                                       \
 168         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 169                                                                         \
 170         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 171 }                                                                       \
 172 static DEVICE_ATTR_RW(name);
 173
 174 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 175
 176 static ssize_t min_ratio_store(struct device *dev,
 177                 struct device_attribute *attr, const char *buf, size_t count)
 178 {
 179         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 180         unsigned int ratio;
 181         ssize_t ret;
 182
 183         ret = kstrtouint(buf, 10, &ratio);
 184         if (ret < 0)
 185                 return ret;
 186
 187         ret = bdi_set_min_ratio(bdi, ratio);
 188         if (!ret)
 189                 ret = count;
 190
 191         return ret;
 192 }
 193 BDI_SHOW(min_ratio, bdi->min_ratio)
 194
 195 static ssize_t max_ratio_store(struct device *dev,
 196                 struct device_attribute *attr, const char *buf, size_t count)
 197 {
 198         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 199         unsigned int ratio;
 200         ssize_t ret;
 201
 202         ret = kstrtouint(buf, 10, &ratio);
 203         if (ret < 0)
 204                 return ret;
 205
 206         ret = bdi_set_max_ratio(bdi, ratio);
 207         if (!ret)
 208                 ret = count;
 209
 210         return ret;
 211 }
 212 BDI_SHOW(max_ratio, bdi->max_ratio)
 213
 214 static ssize_t stable_pages_required_show(struct device *dev,
 215                                           struct device_attribute *attr,
 216                                           char *page)
 217 {
 218         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 219
 220         return snprintf(page, PAGE_SIZE-1, "%d\n",
 221                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 222 }
 223 static DEVICE_ATTR_RO(stable_pages_required);
 224
 225 static struct attribute *bdi_dev_attrs[] = {
 226         &dev_attr_read_ahead_kb.attr,
 227         &dev_attr_min_ratio.attr,
 228         &dev_attr_max_ratio.attr,
 229         &dev_attr_stable_pages_required.attr,
 230         NULL,
 231 };
 232 ATTRIBUTE_GROUPS(bdi_dev);
 233
 234 static __init int bdi_class_init(void)
 235 {
 236         bdi_class = class_create(THIS_MODULE, "bdi");
 237         if (IS_ERR(bdi_class))
 238                 return PTR_ERR(bdi_class);
 239
 240         bdi_class->dev_groups = bdi_dev_groups;
 241         bdi_debug_init();
 242
 243         return 0;
 244 }
 245 postcore_initcall(bdi_class_init);
 246
 247 static int bdi_init(struct backing_dev_info *bdi);
 248
 249 static int __init default_bdi_init(void)
 250 {
 251         int err;
 252
 253         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
 254                                               WQ_UNBOUND | WQ_SYSFS, 0);
 255         if (!bdi_wq)
 256                 return -ENOMEM;
 257
 258         err = bdi_init(&noop_backing_dev_info);
 259
 260         return err;
 261 }
 262 subsys_initcall(default_bdi_init);
 263
 264 /*
 265  * This function is used when the first inode for this wb is marked dirty. It
 266  * wakes-up the corresponding bdi thread which should then take care of the
 267  * periodic background write-out of dirty inodes. Since the write-out would
 268  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 269  * set up a timer which wakes the bdi thread up later.
 270  *
 271  * Note, we wouldn't bother setting up the timer, but this function is on the
 272  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 273  * by delaying the wake-up.
 274  *
 275  * We have to be careful not to postpone flush work if it is scheduled for
 276  * earlier. Thus we use queue_delayed_work().
 277  */
 278 void wb_wakeup_delayed(struct bdi_writeback *wb)
 279 {
 280         unsigned long timeout;
 281
 282         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 283         spin_lock_bh(&wb->work_lock);
 284         if (test_bit(WB_registered, &wb->state))
 285                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 286         spin_unlock_bh(&wb->work_lock);
 287 }
 288
 289 /*
 290  * Initial write bandwidth: 100 MB/s
 291  */
 292 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 293
 294 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 295                    int blkcg_id, gfp_t gfp)
 296 {
 297         int i, err;
 298
 299         memset(wb, 0, sizeof(*wb));
 300
 301         if (wb != &bdi->wb)
 302                 bdi_get(bdi);
 303         wb->bdi = bdi;
 304         wb->last_old_flush = jiffies;
 305         INIT_LIST_HEAD(&wb->b_dirty);
 306         INIT_LIST_HEAD(&wb->b_io);
 307         INIT_LIST_HEAD(&wb->b_more_io);
 308         INIT_LIST_HEAD(&wb->b_dirty_time);
 309         spin_lock_init(&wb->list_lock);
 310
 311         wb->bw_time_stamp = jiffies;
 312         wb->balanced_dirty_ratelimit = INIT_BW;
 313         wb->dirty_ratelimit = INIT_BW;
 314         wb->write_bandwidth = INIT_BW;
 315         wb->avg_write_bandwidth = INIT_BW;
 316
 317         spin_lock_init(&wb->work_lock);
 318         INIT_LIST_HEAD(&wb->work_list);
 319         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 320         wb->dirty_sleep = jiffies;
 321
 322         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 323         if (!wb->congested) {
 324                 err = -ENOMEM;
 325                 goto out_put_bdi;
 326         }
 327
 328         err = fprop_local_init_percpu(&wb->completions, gfp);
 329         if (err)
 330                 goto out_put_cong;
 331
 332         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 333                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 334                 if (err)
 335                         goto out_destroy_stat;
 336         }
 337
 338         return 0;
 339
 340 out_destroy_stat:
 341         while (i--)
 342                 percpu_counter_destroy(&wb->stat[i]);
 343         fprop_local_destroy_percpu(&wb->completions);
 344 out_put_cong:
 345         wb_congested_put(wb->congested);
 346 out_put_bdi:
 347         if (wb != &bdi->wb)
 348                 bdi_put(bdi);
 349         return err;
 350 }
 351
 352 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 353
 354 /*
 355  * Remove bdi from the global list and shutdown any threads we have running
 356  */
 357 static void wb_shutdown(struct bdi_writeback *wb)
 358 {
 359         /* Make sure nobody queues further work */
 360         spin_lock_bh(&wb->work_lock);
 361         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 362                 spin_unlock_bh(&wb->work_lock);
 363                 return;
 364         }
 365         spin_unlock_bh(&wb->work_lock);
 366
 367         cgwb_remove_from_bdi_list(wb);
 368         /*
 369          * Drain work list and shutdown the delayed_work.  !WB_registered
 370          * tells wb_workfn() that @wb is dying and its work_list needs to
 371          * be drained no matter what.
 372          */
 373         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 374         flush_delayed_work(&wb->dwork);
 375         WARN_ON(!list_empty(&wb->work_list));
 376 }
 377
 378 static void wb_exit(struct bdi_writeback *wb)
 379 {
 380         int i;
 381
 382         WARN_ON(delayed_work_pending(&wb->dwork));
 383
 384         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 385                 percpu_counter_destroy(&wb->stat[i]);
 386
 387         fprop_local_destroy_percpu(&wb->completions);
 388         wb_congested_put(wb->congested);
 389         if (wb != &wb->bdi->wb)
 390                 bdi_put(wb->bdi);
 391 }
 392
 393 #ifdef CONFIG_CGROUP_WRITEBACK
 394
 395 #include <linux/memcontrol.h>
 396
 397 /*
 398  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 399  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 400  * protected.
 401  */
 402 static DEFINE_SPINLOCK(cgwb_lock);
 403 static struct workqueue_struct *cgwb_release_wq;
 404
 405 /**
 406  * wb_congested_get_create - get or create a wb_congested
 407  * @bdi: associated bdi
 408  * @blkcg_id: ID of the associated blkcg
 409  * @gfp: allocation mask
 410  *
 411  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 412  * The returned wb_congested has its reference count incremented.  Returns
 413  * NULL on failure.
 414  */
 415 struct bdi_writeback_congested *
 416 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 417 {
 418         struct bdi_writeback_congested *new_congested = NULL, *congested;
 419         struct rb_node **node, *parent;
 420         unsigned long flags;
 421 retry:
 422         spin_lock_irqsave(&cgwb_lock, flags);
 423
 424         node = &bdi->cgwb_congested_tree.rb_node;
 425         parent = NULL;
 426
 427         while (*node != NULL) {
 428                 parent = *node;
 429                 congested = rb_entry(parent, struct bdi_writeback_congested,
 430                                      rb_node);
 431                 if (congested->blkcg_id < blkcg_id)
 432                         node = &parent->rb_left;
 433                 else if (congested->blkcg_id > blkcg_id)
 434                         node = &parent->rb_right;
 435                 else
 436                         goto found;
 437         }
 438
 439         if (new_congested) {
 440                 /* !found and storage for new one already allocated, insert */
 441                 congested = new_congested;
 442                 rb_link_node(&congested->rb_node, parent, node);
 443                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 444                 spin_unlock_irqrestore(&cgwb_lock, flags);
 445                 return congested;
 446         }
 447
 448         spin_unlock_irqrestore(&cgwb_lock, flags);
 449
 450         /* allocate storage for new one and retry */
 451         new_congested = kzalloc(sizeof(*new_congested), gfp);
 452         if (!new_congested)
 453                 return NULL;
 454
 455         refcount_set(&new_congested->refcnt, 1);
 456         new_congested->__bdi = bdi;
 457         new_congested->blkcg_id = blkcg_id;
 458         goto retry;
 459
 460 found:
 461         refcount_inc(&congested->refcnt);
 462         spin_unlock_irqrestore(&cgwb_lock, flags);
 463         kfree(new_congested);
 464         return congested;
 465 }
 466
 467 /**
 468  * wb_congested_put - put a wb_congested
 469  * @congested: wb_congested to put
 470  *
 471  * Put @congested and destroy it if the refcnt reaches zero.
 472  */
 473 void wb_congested_put(struct bdi_writeback_congested *congested)
 474 {
 475         unsigned long flags;
 476
 477         if (!refcount_dec_and_lock_irqsave(&congested->refcnt, &cgwb_lock, &flags))
 478                 return;
 479
 480         /* bdi might already have been destroyed leaving @congested unlinked */
 481         if (congested->__bdi) {
 482                 rb_erase(&congested->rb_node,
 483                          &congested->__bdi->cgwb_congested_tree);
 484                 congested->__bdi = NULL;
 485         }
 486
 487         spin_unlock_irqrestore(&cgwb_lock, flags);
 488         kfree(congested);
 489 }
 490
 491 static void cgwb_release_workfn(struct work_struct *work)
 492 {
 493         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 494                                                 release_work);
 495         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 496
 497         mutex_lock(&wb->bdi->cgwb_release_mutex);
 498         wb_shutdown(wb);
 499
 500         css_put(wb->memcg_css);
 501         css_put(wb->blkcg_css);
 502         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 503
 504         /* triggers blkg destruction if cgwb_refcnt becomes zero */
 505         blkcg_cgwb_put(blkcg);
 506
 507         fprop_local_destroy_percpu(&wb->memcg_completions);
 508         percpu_ref_exit(&wb->refcnt);
 509         wb_exit(wb);
 510         kfree_rcu(wb, rcu);
 511 }
 512
 513 static void cgwb_release(struct percpu_ref *refcnt)
 514 {
 515         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 516                                                 refcnt);
 517         queue_work(cgwb_release_wq, &wb->release_work);
 518 }
 519
 520 static void cgwb_kill(struct bdi_writeback *wb)
 521 {
 522         lockdep_assert_held(&cgwb_lock);
 523
 524         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 525         list_del(&wb->memcg_node);
 526         list_del(&wb->blkcg_node);
 527         percpu_ref_kill(&wb->refcnt);
 528 }
 529
 530 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 531 {
 532         spin_lock_irq(&cgwb_lock);
 533         list_del_rcu(&wb->bdi_node);
 534         spin_unlock_irq(&cgwb_lock);
 535 }
 536
 537 static int cgwb_create(struct backing_dev_info *bdi,
 538                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 539 {
 540         struct mem_cgroup *memcg;
 541         struct cgroup_subsys_state *blkcg_css;
 542         struct blkcg *blkcg;
 543         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 544         struct bdi_writeback *wb;
 545         unsigned long flags;
 546         int ret = 0;
 547
 548         memcg = mem_cgroup_from_css(memcg_css);
 549         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 550         blkcg = css_to_blkcg(blkcg_css);
 551         memcg_cgwb_list = &memcg->cgwb_list;
 552         blkcg_cgwb_list = &blkcg->cgwb_list;
 553
 554         /* look up again under lock and discard on blkcg mismatch */
 555         spin_lock_irqsave(&cgwb_lock, flags);
 556         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 557         if (wb && wb->blkcg_css != blkcg_css) {
 558                 cgwb_kill(wb);
 559                 wb = NULL;
 560         }
 561         spin_unlock_irqrestore(&cgwb_lock, flags);
 562         if (wb)
 563                 goto out_put;
 564
 565         /* need to create a new one */
 566         wb = kmalloc(sizeof(*wb), gfp);
 567         if (!wb) {
 568                 ret = -ENOMEM;
 569                 goto out_put;
 570         }
 571
 572         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 573         if (ret)
 574                 goto err_free;
 575
 576         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 577         if (ret)
 578                 goto err_wb_exit;
 579
 580         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 581         if (ret)
 582                 goto err_ref_exit;
 583
 584         wb->memcg_css = memcg_css;
 585         wb->blkcg_css = blkcg_css;
 586         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 587         set_bit(WB_registered, &wb->state);
 588
 589         /*
 590          * The root wb determines the registered state of the whole bdi and
 591          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 592          * whether they're still online.  Don't link @wb if any is dead.
 593          * See wb_memcg_offline() and wb_blkcg_offline().
 594          */
 595         ret = -ENODEV;
 596         spin_lock_irqsave(&cgwb_lock, flags);
 597         if (test_bit(WB_registered, &bdi->wb.state) &&
 598             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 599                 /* we might have raced another instance of this function */
 600                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 601                 if (!ret) {
 602                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 603                         list_add(&wb->memcg_node, memcg_cgwb_list);
 604                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 605                         blkcg_cgwb_get(blkcg);
 606                         css_get(memcg_css);
 607                         css_get(blkcg_css);
 608                 }
 609         }
 610         spin_unlock_irqrestore(&cgwb_lock, flags);
 611         if (ret) {
 612                 if (ret == -EEXIST)
 613                         ret = 0;
 614                 goto err_fprop_exit;
 615         }
 616         goto out_put;
 617
 618 err_fprop_exit:
 619         fprop_local_destroy_percpu(&wb->memcg_completions);
 620 err_ref_exit:
 621         percpu_ref_exit(&wb->refcnt);
 622 err_wb_exit:
 623         wb_exit(wb);
 624 err_free:
 625         kfree(wb);
 626 out_put:
 627         css_put(blkcg_css);
 628         return ret;
 629 }
 630
 631 /**
 632  * wb_get_create - get wb for a given memcg, create if necessary
 633  * @bdi: target bdi
 634  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 635  * @gfp: allocation mask to use
 636  *
 637  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 638  * create one.  The returned wb has its refcount incremented.
 639  *
 640  * This function uses css_get() on @memcg_css and thus expects its refcnt
 641  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 642  * @memcg_css isn't enough.  try_get it before calling this function.
 643  *
 644  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 645  * memcg on the default hierarchy, memcg association is guaranteed to be
 646  * more specific (equal or descendant to the associated blkcg) and thus can
 647  * identify both the memcg and blkcg associations.
 648  *
 649  * Because the blkcg associated with a memcg may change as blkcg is enabled
 650  * and disabled closer to root in the hierarchy, each wb keeps track of
 651  * both the memcg and blkcg associated with it and verifies the blkcg on
 652  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 653  * created.
 654  */
 655 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 656                                     struct cgroup_subsys_state *memcg_css,
 657                                     gfp_t gfp)
 658 {
 659         struct bdi_writeback *wb;
 660
 661         might_sleep_if(gfpflags_allow_blocking(gfp));
 662
 663         if (!memcg_css->parent)
 664                 return &bdi->wb;
 665
 666         do {
 667                 rcu_read_lock();
 668                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 669                 if (wb) {
 670                         struct cgroup_subsys_state *blkcg_css;
 671
 672                         /* see whether the blkcg association has changed */
 673                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 674                                                      &io_cgrp_subsys);
 675                         if (unlikely(wb->blkcg_css != blkcg_css ||
 676                                      !wb_tryget(wb)))
 677                                 wb = NULL;
 678                         css_put(blkcg_css);
 679                 }
 680                 rcu_read_unlock();
 681         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 682
 683         return wb;
 684 }
 685
 686 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 687 {
 688         int ret;
 689
 690         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 691         bdi->cgwb_congested_tree = RB_ROOT;
 692         mutex_init(&bdi->cgwb_release_mutex);
 693         init_rwsem(&bdi->wb_switch_rwsem);
 694
 695         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 696         if (!ret) {
 697                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 698                 bdi->wb.blkcg_css = blkcg_root_css;
 699         }
 700         return ret;
 701 }
 702
 703 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 704 {
 705         struct radix_tree_iter iter;
 706         void **slot;
 707         struct bdi_writeback *wb;
 708
 709         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 710
 711         spin_lock_irq(&cgwb_lock);
 712         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 713                 cgwb_kill(*slot);
 714         spin_unlock_irq(&cgwb_lock);
 715
 716         mutex_lock(&bdi->cgwb_release_mutex);
 717         spin_lock_irq(&cgwb_lock);
 718         while (!list_empty(&bdi->wb_list)) {
 719                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 720                                       bdi_node);
 721                 spin_unlock_irq(&cgwb_lock);
 722                 wb_shutdown(wb);
 723                 spin_lock_irq(&cgwb_lock);
 724         }
 725         spin_unlock_irq(&cgwb_lock);
 726         mutex_unlock(&bdi->cgwb_release_mutex);
 727 }
 728
 729 /**
 730  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 731  * @memcg: memcg being offlined
 732  *
 733  * Also prevents creation of any new wb's associated with @memcg.
 734  */
 735 void wb_memcg_offline(struct mem_cgroup *memcg)
 736 {
 737         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 738         struct bdi_writeback *wb, *next;
 739
 740         spin_lock_irq(&cgwb_lock);
 741         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 742                 cgwb_kill(wb);
 743         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 744         spin_unlock_irq(&cgwb_lock);
 745 }
 746
 747 /**
 748  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 749  * @blkcg: blkcg being offlined
 750  *
 751  * Also prevents creation of any new wb's associated with @blkcg.
 752  */
 753 void wb_blkcg_offline(struct blkcg *blkcg)
 754 {
 755         struct bdi_writeback *wb, *next;
 756
 757         spin_lock_irq(&cgwb_lock);
 758         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 759                 cgwb_kill(wb);
 760         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 761         spin_unlock_irq(&cgwb_lock);
 762 }
 763
 764 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 765 {
 766         struct rb_node *rbn;
 767
 768         spin_lock_irq(&cgwb_lock);
 769         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 770                 struct bdi_writeback_congested *congested =
 771                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 772
 773                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 774                 congested->__bdi = NULL;        /* mark @congested unlinked */
 775         }
 776         spin_unlock_irq(&cgwb_lock);
 777 }
 778
 779 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 780 {
 781         spin_lock_irq(&cgwb_lock);
 782         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 783         spin_unlock_irq(&cgwb_lock);
 784 }
 785
 786 static int __init cgwb_init(void)
 787 {
 788         /*
 789          * There can be many concurrent release work items overwhelming
 790          * system_wq.  Put them in a separate wq and limit concurrency.
 791          * There's no point in executing many of these in parallel.
 792          */
 793         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 794         if (!cgwb_release_wq)
 795                 return -ENOMEM;
 796
 797         return 0;
 798 }
 799 subsys_initcall(cgwb_init);
 800
 801 #else   /* CONFIG_CGROUP_WRITEBACK */
 802
 803 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 804 {
 805         int err;
 806
 807         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 808         if (!bdi->wb_congested)
 809                 return -ENOMEM;
 810
 811         refcount_set(&bdi->wb_congested->refcnt, 1);
 812
 813         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 814         if (err) {
 815                 wb_congested_put(bdi->wb_congested);
 816                 return err;
 817         }
 818         return 0;
 819 }
 820
 821 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 822
 823 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 824 {
 825         wb_congested_put(bdi->wb_congested);
 826 }
 827
 828 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 829 {
 830         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 831 }
 832
 833 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 834 {
 835         list_del_rcu(&wb->bdi_node);
 836 }
 837
 838 #endif  /* CONFIG_CGROUP_WRITEBACK */
 839
 840 static int bdi_init(struct backing_dev_info *bdi)
 841 {
 842         int ret;
 843
 844         bdi->dev = NULL;
 845
 846         kref_init(&bdi->refcnt);
 847         bdi->min_ratio = 0;
 848         bdi->max_ratio = 100;
 849         bdi->max_prop_frac = FPROP_FRAC_BASE;
 850         INIT_LIST_HEAD(&bdi->bdi_list);
 851         INIT_LIST_HEAD(&bdi->wb_list);
 852         init_waitqueue_head(&bdi->wb_waitq);
 853
 854         ret = cgwb_bdi_init(bdi);
 855
 856         return ret;
 857 }
 858
 859 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 860 {
 861         struct backing_dev_info *bdi;
 862
 863         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 864                            gfp_mask | __GFP_ZERO, node_id);
 865         if (!bdi)
 866                 return NULL;
 867
 868         if (bdi_init(bdi)) {
 869                 kfree(bdi);
 870                 return NULL;
 871         }
 872         return bdi;
 873 }
 874 EXPORT_SYMBOL(bdi_alloc_node);
 875
 876 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 877 {
 878         struct device *dev;
 879
 880         if (bdi->dev)   /* The driver needs to use separate queues per device */
 881                 return 0;
 882
 883         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 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         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 895         spin_unlock_bh(&bdi_lock);
 896
 897         trace_writeback_bdi_register(bdi);
 898         return 0;
 899 }
 900 EXPORT_SYMBOL(bdi_register_va);
 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 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 915 {
 916         int rc;
 917
 918         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 919         if (rc)
 920                 return rc;
 921         /* Leaking owner reference... */
 922         WARN_ON(bdi->owner);
 923         bdi->owner = owner;
 924         get_device(owner);
 925         return 0;
 926 }
 927 EXPORT_SYMBOL(bdi_register_owner);
 928
 929 /*
 930  * Remove bdi from bdi_list, and ensure that it is no longer visible
 931  */
 932 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 933 {
 934         spin_lock_bh(&bdi_lock);
 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         /* make sure nobody finds us on the bdi_list anymore */
 944         bdi_remove_from_list(bdi);
 945         wb_shutdown(&bdi->wb);
 946         cgwb_bdi_unregister(bdi);
 947
 948         if (bdi->dev) {
 949                 bdi_debug_unregister(bdi);
 950                 device_unregister(bdi->dev);
 951                 bdi->dev = NULL;
 952         }
 953
 954         if (bdi->owner) {
 955                 put_device(bdi->owner);
 956                 bdi->owner = NULL;
 957         }
 958 }
 959
 960 static void release_bdi(struct kref *ref)
 961 {
 962         struct backing_dev_info *bdi =
 963                         container_of(ref, struct backing_dev_info, refcnt);
 964
 965         if (test_bit(WB_registered, &bdi->wb.state))
 966                 bdi_unregister(bdi);
 967         WARN_ON_ONCE(bdi->dev);
 968         wb_exit(&bdi->wb);
 969         cgwb_bdi_exit(bdi);
 970         kfree(bdi);
 971 }
 972
 973 void bdi_put(struct backing_dev_info *bdi)
 974 {
 975         kref_put(&bdi->refcnt, release_bdi);
 976 }
 977 EXPORT_SYMBOL(bdi_put);
 978
 979 static wait_queue_head_t congestion_wqh[2] = {
 980                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 981                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 982         };
 983 static atomic_t nr_wb_congested[2];
 984
 985 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 986 {
 987         wait_queue_head_t *wqh = &congestion_wqh[sync];
 988         enum wb_congested_state bit;
 989
 990         bit = sync ? WB_sync_congested : WB_async_congested;
 991         if (test_and_clear_bit(bit, &congested->state))
 992                 atomic_dec(&nr_wb_congested[sync]);
 993         smp_mb__after_atomic();
 994         if (waitqueue_active(wqh))
 995                 wake_up(wqh);
 996 }
 997 EXPORT_SYMBOL(clear_wb_congested);
 998
 999 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
1000 {
1001         enum wb_congested_state bit;
1002
1003         bit = sync ? WB_sync_congested : WB_async_congested;
1004         if (!test_and_set_bit(bit, &congested->state))
1005                 atomic_inc(&nr_wb_congested[sync]);
1006 }
1007 EXPORT_SYMBOL(set_wb_congested);
1008
1009 /**
1010  * congestion_wait - wait for a backing_dev to become uncongested
1011  * @sync: SYNC or ASYNC IO
1012  * @timeout: timeout in jiffies
1013  *
1014  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1015  * write congestion.  If no backing_devs are congested then just wait for the
1016  * next write to be completed.
1017  */
1018 long congestion_wait(int sync, long timeout)
1019 {
1020         long ret;
1021         unsigned long start = jiffies;
1022         DEFINE_WAIT(wait);
1023         wait_queue_head_t *wqh = &congestion_wqh[sync];
1024
1025         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1026         ret = io_schedule_timeout(timeout);
1027         finish_wait(wqh, &wait);
1028
1029         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1030                                         jiffies_to_usecs(jiffies - start));
1031
1032         return ret;
1033 }
1034 EXPORT_SYMBOL(congestion_wait);
1035
1036 /**
1037  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1038  * @sync: SYNC or ASYNC IO
1039  * @timeout: timeout in jiffies
1040  *
1041  * In the event of a congested backing_dev (any backing_dev) this waits
1042  * for up to @timeout jiffies for either a BDI to exit congestion of the
1043  * given @sync queue or a write to complete.
1044  *
1045  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1046  * it is the number of jiffies that were still remaining when the function
1047  * returned. return_value == timeout implies the function did not sleep.
1048  */
1049 long wait_iff_congested(int sync, long timeout)
1050 {
1051         long ret;
1052         unsigned long start = jiffies;
1053         DEFINE_WAIT(wait);
1054         wait_queue_head_t *wqh = &congestion_wqh[sync];
1055
1056         /*
1057          * If there is no congestion, yield if necessary instead
1058          * of sleeping on the congestion queue
1059          */
1060         if (atomic_read(&nr_wb_congested[sync]) == 0) {
1061                 cond_resched();
1062
1063                 /* In case we scheduled, work out time remaining */
1064                 ret = timeout - (jiffies - start);
1065                 if (ret < 0)
1066                         ret = 0;
1067
1068                 goto out;
1069         }
1070
1071         /* Sleep until uncongested or a write happens */
1072         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1073         ret = io_schedule_timeout(timeout);
1074         finish_wait(wqh, &wait);
1075
1076 out:
1077         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1078                                         jiffies_to_usecs(jiffies - start));
1079
1080         return ret;
1081 }
1082 EXPORT_SYMBOL(wait_iff_congested);