mm/backing-dev.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2
   3 #include <linux/wait.h>
   4 #include <linux/rbtree.h>
   5 #include <linux/backing-dev.h>
   6 #include <linux/kthread.h>
   7 #include <linux/freezer.h>
   8 #include <linux/fs.h>
   9 #include <linux/pagemap.h>
  10 #include <linux/mm.h>
  11 #include <linux/sched.h>
  12 #include <linux/module.h>
  13 #include <linux/writeback.h>
  14 #include <linux/device.h>
  15 #include <trace/events/writeback.h>
  16
  17 struct backing_dev_info noop_backing_dev_info = {
  18         .name           = "noop",
  19         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
  20 };
  21 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  22
  23 static struct class *bdi_class;
  24
  25 /*
  26  * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU
  27  * reader side locking.
  28  */
  29 DEFINE_SPINLOCK(bdi_lock);
  30 static u64 bdi_id_cursor;
  31 static struct rb_root bdi_tree = RB_ROOT;
  32 LIST_HEAD(bdi_list);
  33
  34 /* bdi_wq serves all asynchronous writeback tasks */
  35 struct workqueue_struct *bdi_wq;
  36
  37 #ifdef CONFIG_DEBUG_FS
  38 #include <linux/debugfs.h>
  39 #include <linux/seq_file.h>
  40
  41 static struct dentry *bdi_debug_root;
  42
  43 static void bdi_debug_init(void)
  44 {
  45         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  46 }
  47
  48 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  49 {
  50         struct backing_dev_info *bdi = m->private;
  51         struct bdi_writeback *wb = &bdi->wb;
  52         unsigned long background_thresh;
  53         unsigned long dirty_thresh;
  54         unsigned long wb_thresh;
  55         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  56         struct inode *inode;
  57
  58         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  59         spin_lock(&wb->list_lock);
  60         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  61                 nr_dirty++;
  62         list_for_each_entry(inode, &wb->b_io, i_io_list)
  63                 nr_io++;
  64         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  65                 nr_more_io++;
  66         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  67                 if (inode->i_state & I_DIRTY_TIME)
  68                         nr_dirty_time++;
  69         spin_unlock(&wb->list_lock);
  70
  71         global_dirty_limits(&background_thresh, &dirty_thresh);
  72         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  73
  74 #define K(x) ((x) << (PAGE_SHIFT - 10))
  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 #undef K
 104
 105         return 0;
 106 }
 107 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 108
 109 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 110 {
 111         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 112
 113         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 114                             &bdi_debug_stats_fops);
 115 }
 116
 117 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 118 {
 119         debugfs_remove_recursive(bdi->debug_dir);
 120 }
 121 #else
 122 static inline void bdi_debug_init(void)
 123 {
 124 }
 125 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 126                                       const char *name)
 127 {
 128 }
 129 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 130 {
 131 }
 132 #endif
 133
 134 static ssize_t read_ahead_kb_store(struct device *dev,
 135                                   struct device_attribute *attr,
 136                                   const char *buf, size_t count)
 137 {
 138         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 139         unsigned long read_ahead_kb;
 140         ssize_t ret;
 141
 142         ret = kstrtoul(buf, 10, &read_ahead_kb);
 143         if (ret < 0)
 144                 return ret;
 145
 146         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 147
 148         return count;
 149 }
 150
 151 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 152
 153 #define BDI_SHOW(name, expr)                                            \
 154 static ssize_t name##_show(struct device *dev,                          \
 155                            struct device_attribute *attr, char *page)   \
 156 {                                                                       \
 157         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 158                                                                         \
 159         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 160 }                                                                       \
 161 static DEVICE_ATTR_RW(name);
 162
 163 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 164
 165 static ssize_t min_ratio_store(struct device *dev,
 166                 struct device_attribute *attr, const char *buf, size_t count)
 167 {
 168         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 169         unsigned int ratio;
 170         ssize_t ret;
 171
 172         ret = kstrtouint(buf, 10, &ratio);
 173         if (ret < 0)
 174                 return ret;
 175
 176         ret = bdi_set_min_ratio(bdi, ratio);
 177         if (!ret)
 178                 ret = count;
 179
 180         return ret;
 181 }
 182 BDI_SHOW(min_ratio, bdi->min_ratio)
 183
 184 static ssize_t max_ratio_store(struct device *dev,
 185                 struct device_attribute *attr, const char *buf, size_t count)
 186 {
 187         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 188         unsigned int ratio;
 189         ssize_t ret;
 190
 191         ret = kstrtouint(buf, 10, &ratio);
 192         if (ret < 0)
 193                 return ret;
 194
 195         ret = bdi_set_max_ratio(bdi, ratio);
 196         if (!ret)
 197                 ret = count;
 198
 199         return ret;
 200 }
 201 BDI_SHOW(max_ratio, bdi->max_ratio)
 202
 203 static ssize_t stable_pages_required_show(struct device *dev,
 204                                           struct device_attribute *attr,
 205                                           char *page)
 206 {
 207         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 208
 209         return snprintf(page, PAGE_SIZE-1, "%d\n",
 210                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 211 }
 212 static DEVICE_ATTR_RO(stable_pages_required);
 213
 214 static struct attribute *bdi_dev_attrs[] = {
 215         &dev_attr_read_ahead_kb.attr,
 216         &dev_attr_min_ratio.attr,
 217         &dev_attr_max_ratio.attr,
 218         &dev_attr_stable_pages_required.attr,
 219         NULL,
 220 };
 221 ATTRIBUTE_GROUPS(bdi_dev);
 222
 223 static __init int bdi_class_init(void)
 224 {
 225         bdi_class = class_create(THIS_MODULE, "bdi");
 226         if (IS_ERR(bdi_class))
 227                 return PTR_ERR(bdi_class);
 228
 229         bdi_class->dev_groups = bdi_dev_groups;
 230         bdi_debug_init();
 231
 232         return 0;
 233 }
 234 postcore_initcall(bdi_class_init);
 235
 236 static int bdi_init(struct backing_dev_info *bdi);
 237
 238 static int __init default_bdi_init(void)
 239 {
 240         int err;
 241
 242         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND |
 243                                  WQ_SYSFS, 0);
 244         if (!bdi_wq)
 245                 return -ENOMEM;
 246
 247         err = bdi_init(&noop_backing_dev_info);
 248
 249         return err;
 250 }
 251 subsys_initcall(default_bdi_init);
 252
 253 /*
 254  * This function is used when the first inode for this wb is marked dirty. It
 255  * wakes-up the corresponding bdi thread which should then take care of the
 256  * periodic background write-out of dirty inodes. Since the write-out would
 257  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 258  * set up a timer which wakes the bdi thread up later.
 259  *
 260  * Note, we wouldn't bother setting up the timer, but this function is on the
 261  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 262  * by delaying the wake-up.
 263  *
 264  * We have to be careful not to postpone flush work if it is scheduled for
 265  * earlier. Thus we use queue_delayed_work().
 266  */
 267 void wb_wakeup_delayed(struct bdi_writeback *wb)
 268 {
 269         unsigned long timeout;
 270
 271         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 272         spin_lock_bh(&wb->work_lock);
 273         if (test_bit(WB_registered, &wb->state))
 274                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 275         spin_unlock_bh(&wb->work_lock);
 276 }
 277
 278 /*
 279  * Initial write bandwidth: 100 MB/s
 280  */
 281 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 282
 283 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 284                    int blkcg_id, gfp_t gfp)
 285 {
 286         int i, err;
 287
 288         memset(wb, 0, sizeof(*wb));
 289
 290         if (wb != &bdi->wb)
 291                 bdi_get(bdi);
 292         wb->bdi = bdi;
 293         wb->last_old_flush = jiffies;
 294         INIT_LIST_HEAD(&wb->b_dirty);
 295         INIT_LIST_HEAD(&wb->b_io);
 296         INIT_LIST_HEAD(&wb->b_more_io);
 297         INIT_LIST_HEAD(&wb->b_dirty_time);
 298         spin_lock_init(&wb->list_lock);
 299
 300         wb->bw_time_stamp = jiffies;
 301         wb->balanced_dirty_ratelimit = INIT_BW;
 302         wb->dirty_ratelimit = INIT_BW;
 303         wb->write_bandwidth = INIT_BW;
 304         wb->avg_write_bandwidth = INIT_BW;
 305
 306         spin_lock_init(&wb->work_lock);
 307         INIT_LIST_HEAD(&wb->work_list);
 308         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 309         wb->dirty_sleep = jiffies;
 310
 311         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 312         if (!wb->congested) {
 313                 err = -ENOMEM;
 314                 goto out_put_bdi;
 315         }
 316
 317         err = fprop_local_init_percpu(&wb->completions, gfp);
 318         if (err)
 319                 goto out_put_cong;
 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_cong:
 334         wb_congested_put(wb->congested);
 335 out_put_bdi:
 336         if (wb != &bdi->wb)
 337                 bdi_put(bdi);
 338         return err;
 339 }
 340
 341 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 342
 343 /*
 344  * Remove bdi from the global list and shutdown any threads we have running
 345  */
 346 static void wb_shutdown(struct bdi_writeback *wb)
 347 {
 348         /* Make sure nobody queues further work */
 349         spin_lock_bh(&wb->work_lock);
 350         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 351                 spin_unlock_bh(&wb->work_lock);
 352                 return;
 353         }
 354         spin_unlock_bh(&wb->work_lock);
 355
 356         cgwb_remove_from_bdi_list(wb);
 357         /*
 358          * Drain work list and shutdown the delayed_work.  !WB_registered
 359          * tells wb_workfn() that @wb is dying and its work_list needs to
 360          * be drained no matter what.
 361          */
 362         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 363         flush_delayed_work(&wb->dwork);
 364         WARN_ON(!list_empty(&wb->work_list));
 365 }
 366
 367 static void wb_exit(struct bdi_writeback *wb)
 368 {
 369         int i;
 370
 371         WARN_ON(delayed_work_pending(&wb->dwork));
 372
 373         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 374                 percpu_counter_destroy(&wb->stat[i]);
 375
 376         fprop_local_destroy_percpu(&wb->completions);
 377         wb_congested_put(wb->congested);
 378         if (wb != &wb->bdi->wb)
 379                 bdi_put(wb->bdi);
 380 }
 381
 382 #ifdef CONFIG_CGROUP_WRITEBACK
 383
 384 #include <linux/memcontrol.h>
 385
 386 /*
 387  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 388  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 389  * protected.
 390  */
 391 static DEFINE_SPINLOCK(cgwb_lock);
 392 static struct workqueue_struct *cgwb_release_wq;
 393
 394 /**
 395  * wb_congested_get_create - get or create a wb_congested
 396  * @bdi: associated bdi
 397  * @blkcg_id: ID of the associated blkcg
 398  * @gfp: allocation mask
 399  *
 400  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 401  * The returned wb_congested has its reference count incremented.  Returns
 402  * NULL on failure.
 403  */
 404 struct bdi_writeback_congested *
 405 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 406 {
 407         struct bdi_writeback_congested *new_congested = NULL, *congested;
 408         struct rb_node **node, *parent;
 409         unsigned long flags;
 410 retry:
 411         spin_lock_irqsave(&cgwb_lock, flags);
 412
 413         node = &bdi->cgwb_congested_tree.rb_node;
 414         parent = NULL;
 415
 416         while (*node != NULL) {
 417                 parent = *node;
 418                 congested = rb_entry(parent, struct bdi_writeback_congested,
 419                                      rb_node);
 420                 if (congested->blkcg_id < blkcg_id)
 421                         node = &parent->rb_left;
 422                 else if (congested->blkcg_id > blkcg_id)
 423                         node = &parent->rb_right;
 424                 else
 425                         goto found;
 426         }
 427
 428         if (new_congested) {
 429                 /* !found and storage for new one already allocated, insert */
 430                 congested = new_congested;
 431                 rb_link_node(&congested->rb_node, parent, node);
 432                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 433                 spin_unlock_irqrestore(&cgwb_lock, flags);
 434                 return congested;
 435         }
 436
 437         spin_unlock_irqrestore(&cgwb_lock, flags);
 438
 439         /* allocate storage for new one and retry */
 440         new_congested = kzalloc(sizeof(*new_congested), gfp);
 441         if (!new_congested)
 442                 return NULL;
 443
 444         refcount_set(&new_congested->refcnt, 1);
 445         new_congested->__bdi = bdi;
 446         new_congested->blkcg_id = blkcg_id;
 447         goto retry;
 448
 449 found:
 450         refcount_inc(&congested->refcnt);
 451         spin_unlock_irqrestore(&cgwb_lock, flags);
 452         kfree(new_congested);
 453         return congested;
 454 }
 455
 456 /**
 457  * wb_congested_put - put a wb_congested
 458  * @congested: wb_congested to put
 459  *
 460  * Put @congested and destroy it if the refcnt reaches zero.
 461  */
 462 void wb_congested_put(struct bdi_writeback_congested *congested)
 463 {
 464         unsigned long flags;
 465
 466         if (!refcount_dec_and_lock_irqsave(&congested->refcnt, &cgwb_lock, &flags))
 467                 return;
 468
 469         /* bdi might already have been destroyed leaving @congested unlinked */
 470         if (congested->__bdi) {
 471                 rb_erase(&congested->rb_node,
 472                          &congested->__bdi->cgwb_congested_tree);
 473                 congested->__bdi = NULL;
 474         }
 475
 476         spin_unlock_irqrestore(&cgwb_lock, flags);
 477         kfree(congested);
 478 }
 479
 480 static void cgwb_release_workfn(struct work_struct *work)
 481 {
 482         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 483                                                 release_work);
 484         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 485
 486         mutex_lock(&wb->bdi->cgwb_release_mutex);
 487         wb_shutdown(wb);
 488
 489         css_put(wb->memcg_css);
 490         css_put(wb->blkcg_css);
 491         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 492
 493         /* triggers blkg destruction if cgwb_refcnt becomes zero */
 494         blkcg_cgwb_put(blkcg);
 495
 496         fprop_local_destroy_percpu(&wb->memcg_completions);
 497         percpu_ref_exit(&wb->refcnt);
 498         wb_exit(wb);
 499         kfree_rcu(wb, rcu);
 500 }
 501
 502 static void cgwb_release(struct percpu_ref *refcnt)
 503 {
 504         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 505                                                 refcnt);
 506         queue_work(cgwb_release_wq, &wb->release_work);
 507 }
 508
 509 static void cgwb_kill(struct bdi_writeback *wb)
 510 {
 511         lockdep_assert_held(&cgwb_lock);
 512
 513         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 514         list_del(&wb->memcg_node);
 515         list_del(&wb->blkcg_node);
 516         percpu_ref_kill(&wb->refcnt);
 517 }
 518
 519 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 520 {
 521         spin_lock_irq(&cgwb_lock);
 522         list_del_rcu(&wb->bdi_node);
 523         spin_unlock_irq(&cgwb_lock);
 524 }
 525
 526 static int cgwb_create(struct backing_dev_info *bdi,
 527                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 528 {
 529         struct mem_cgroup *memcg;
 530         struct cgroup_subsys_state *blkcg_css;
 531         struct blkcg *blkcg;
 532         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 533         struct bdi_writeback *wb;
 534         unsigned long flags;
 535         int ret = 0;
 536
 537         memcg = mem_cgroup_from_css(memcg_css);
 538         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 539         blkcg = css_to_blkcg(blkcg_css);
 540         memcg_cgwb_list = &memcg->cgwb_list;
 541         blkcg_cgwb_list = &blkcg->cgwb_list;
 542
 543         /* look up again under lock and discard on blkcg mismatch */
 544         spin_lock_irqsave(&cgwb_lock, flags);
 545         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 546         if (wb && wb->blkcg_css != blkcg_css) {
 547                 cgwb_kill(wb);
 548                 wb = NULL;
 549         }
 550         spin_unlock_irqrestore(&cgwb_lock, flags);
 551         if (wb)
 552                 goto out_put;
 553
 554         /* need to create a new one */
 555         wb = kmalloc(sizeof(*wb), gfp);
 556         if (!wb) {
 557                 ret = -ENOMEM;
 558                 goto out_put;
 559         }
 560
 561         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 562         if (ret)
 563                 goto err_free;
 564
 565         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 566         if (ret)
 567                 goto err_wb_exit;
 568
 569         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 570         if (ret)
 571                 goto err_ref_exit;
 572
 573         wb->memcg_css = memcg_css;
 574         wb->blkcg_css = blkcg_css;
 575         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 576         set_bit(WB_registered, &wb->state);
 577
 578         /*
 579          * The root wb determines the registered state of the whole bdi and
 580          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 581          * whether they're still online.  Don't link @wb if any is dead.
 582          * See wb_memcg_offline() and wb_blkcg_offline().
 583          */
 584         ret = -ENODEV;
 585         spin_lock_irqsave(&cgwb_lock, flags);
 586         if (test_bit(WB_registered, &bdi->wb.state) &&
 587             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 588                 /* we might have raced another instance of this function */
 589                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 590                 if (!ret) {
 591                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 592                         list_add(&wb->memcg_node, memcg_cgwb_list);
 593                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 594                         blkcg_cgwb_get(blkcg);
 595                         css_get(memcg_css);
 596                         css_get(blkcg_css);
 597                 }
 598         }
 599         spin_unlock_irqrestore(&cgwb_lock, flags);
 600         if (ret) {
 601                 if (ret == -EEXIST)
 602                         ret = 0;
 603                 goto err_fprop_exit;
 604         }
 605         goto out_put;
 606
 607 err_fprop_exit:
 608         fprop_local_destroy_percpu(&wb->memcg_completions);
 609 err_ref_exit:
 610         percpu_ref_exit(&wb->refcnt);
 611 err_wb_exit:
 612         wb_exit(wb);
 613 err_free:
 614         kfree(wb);
 615 out_put:
 616         css_put(blkcg_css);
 617         return ret;
 618 }
 619
 620 /**
 621  * wb_get_lookup - get wb for a given memcg
 622  * @bdi: target bdi
 623  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 624  *
 625  * Try to get the wb for @memcg_css on @bdi.  The returned wb has its
 626  * refcount incremented.
 627  *
 628  * This function uses css_get() on @memcg_css and thus expects its refcnt
 629  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 630  * @memcg_css isn't enough.  try_get it before calling this function.
 631  *
 632  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 633  * memcg on the default hierarchy, memcg association is guaranteed to be
 634  * more specific (equal or descendant to the associated blkcg) and thus can
 635  * identify both the memcg and blkcg associations.
 636  *
 637  * Because the blkcg associated with a memcg may change as blkcg is enabled
 638  * and disabled closer to root in the hierarchy, each wb keeps track of
 639  * both the memcg and blkcg associated with it and verifies the blkcg on
 640  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 641  * created.
 642  */
 643 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi,
 644                                     struct cgroup_subsys_state *memcg_css)
 645 {
 646         struct bdi_writeback *wb;
 647
 648         if (!memcg_css->parent)
 649                 return &bdi->wb;
 650
 651         rcu_read_lock();
 652         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 653         if (wb) {
 654                 struct cgroup_subsys_state *blkcg_css;
 655
 656                 /* see whether the blkcg association has changed */
 657                 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 658                 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb)))
 659                         wb = NULL;
 660                 css_put(blkcg_css);
 661         }
 662         rcu_read_unlock();
 663
 664         return wb;
 665 }
 666
 667 /**
 668  * wb_get_create - get wb for a given memcg, create if necessary
 669  * @bdi: target bdi
 670  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 671  * @gfp: allocation mask to use
 672  *
 673  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 674  * create one.  See wb_get_lookup() for more details.
 675  */
 676 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 677                                     struct cgroup_subsys_state *memcg_css,
 678                                     gfp_t gfp)
 679 {
 680         struct bdi_writeback *wb;
 681
 682         might_sleep_if(gfpflags_allow_blocking(gfp));
 683
 684         if (!memcg_css->parent)
 685                 return &bdi->wb;
 686
 687         do {
 688                 wb = wb_get_lookup(bdi, memcg_css);
 689         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 690
 691         return wb;
 692 }
 693
 694 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 695 {
 696         int ret;
 697
 698         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 699         bdi->cgwb_congested_tree = RB_ROOT;
 700         mutex_init(&bdi->cgwb_release_mutex);
 701         init_rwsem(&bdi->wb_switch_rwsem);
 702
 703         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 704         if (!ret) {
 705                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 706                 bdi->wb.blkcg_css = blkcg_root_css;
 707         }
 708         return ret;
 709 }
 710
 711 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 712 {
 713         struct radix_tree_iter iter;
 714         void **slot;
 715         struct bdi_writeback *wb;
 716
 717         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 718
 719         spin_lock_irq(&cgwb_lock);
 720         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 721                 cgwb_kill(*slot);
 722         spin_unlock_irq(&cgwb_lock);
 723
 724         mutex_lock(&bdi->cgwb_release_mutex);
 725         spin_lock_irq(&cgwb_lock);
 726         while (!list_empty(&bdi->wb_list)) {
 727                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 728                                       bdi_node);
 729                 spin_unlock_irq(&cgwb_lock);
 730                 wb_shutdown(wb);
 731                 spin_lock_irq(&cgwb_lock);
 732         }
 733         spin_unlock_irq(&cgwb_lock);
 734         mutex_unlock(&bdi->cgwb_release_mutex);
 735 }
 736
 737 /**
 738  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 739  * @memcg: memcg being offlined
 740  *
 741  * Also prevents creation of any new wb's associated with @memcg.
 742  */
 743 void wb_memcg_offline(struct mem_cgroup *memcg)
 744 {
 745         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 746         struct bdi_writeback *wb, *next;
 747
 748         spin_lock_irq(&cgwb_lock);
 749         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 750                 cgwb_kill(wb);
 751         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 752         spin_unlock_irq(&cgwb_lock);
 753 }
 754
 755 /**
 756  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 757  * @blkcg: blkcg being offlined
 758  *
 759  * Also prevents creation of any new wb's associated with @blkcg.
 760  */
 761 void wb_blkcg_offline(struct blkcg *blkcg)
 762 {
 763         struct bdi_writeback *wb, *next;
 764
 765         spin_lock_irq(&cgwb_lock);
 766         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 767                 cgwb_kill(wb);
 768         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 769         spin_unlock_irq(&cgwb_lock);
 770 }
 771
 772 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 773 {
 774         struct rb_node *rbn;
 775
 776         spin_lock_irq(&cgwb_lock);
 777         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 778                 struct bdi_writeback_congested *congested =
 779                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 780
 781                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 782                 congested->__bdi = NULL;        /* mark @congested unlinked */
 783         }
 784         spin_unlock_irq(&cgwb_lock);
 785 }
 786
 787 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 788 {
 789         spin_lock_irq(&cgwb_lock);
 790         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 791         spin_unlock_irq(&cgwb_lock);
 792 }
 793
 794 static int __init cgwb_init(void)
 795 {
 796         /*
 797          * There can be many concurrent release work items overwhelming
 798          * system_wq.  Put them in a separate wq and limit concurrency.
 799          * There's no point in executing many of these in parallel.
 800          */
 801         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 802         if (!cgwb_release_wq)
 803                 return -ENOMEM;
 804
 805         return 0;
 806 }
 807 subsys_initcall(cgwb_init);
 808
 809 #else   /* CONFIG_CGROUP_WRITEBACK */
 810
 811 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 812 {
 813         int err;
 814
 815         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 816         if (!bdi->wb_congested)
 817                 return -ENOMEM;
 818
 819         refcount_set(&bdi->wb_congested->refcnt, 1);
 820
 821         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 822         if (err) {
 823                 wb_congested_put(bdi->wb_congested);
 824                 return err;
 825         }
 826         return 0;
 827 }
 828
 829 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 830
 831 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 832 {
 833         wb_congested_put(bdi->wb_congested);
 834 }
 835
 836 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 837 {
 838         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 839 }
 840
 841 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 842 {
 843         list_del_rcu(&wb->bdi_node);
 844 }
 845
 846 #endif  /* CONFIG_CGROUP_WRITEBACK */
 847
 848 static int bdi_init(struct backing_dev_info *bdi)
 849 {
 850         int ret;
 851
 852         bdi->dev = NULL;
 853
 854         kref_init(&bdi->refcnt);
 855         bdi->min_ratio = 0;
 856         bdi->max_ratio = 100;
 857         bdi->max_prop_frac = FPROP_FRAC_BASE;
 858         INIT_LIST_HEAD(&bdi->bdi_list);
 859         INIT_LIST_HEAD(&bdi->wb_list);
 860         init_waitqueue_head(&bdi->wb_waitq);
 861
 862         ret = cgwb_bdi_init(bdi);
 863
 864         return ret;
 865 }
 866
 867 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 868 {
 869         struct backing_dev_info *bdi;
 870
 871         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 872                            gfp_mask | __GFP_ZERO, node_id);
 873         if (!bdi)
 874                 return NULL;
 875
 876         if (bdi_init(bdi)) {
 877                 kfree(bdi);
 878                 return NULL;
 879         }
 880         return bdi;
 881 }
 882 EXPORT_SYMBOL(bdi_alloc_node);
 883
 884 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp)
 885 {
 886         struct rb_node **p = &bdi_tree.rb_node;
 887         struct rb_node *parent = NULL;
 888         struct backing_dev_info *bdi;
 889
 890         lockdep_assert_held(&bdi_lock);
 891
 892         while (*p) {
 893                 parent = *p;
 894                 bdi = rb_entry(parent, struct backing_dev_info, rb_node);
 895
 896                 if (bdi->id > id)
 897                         p = &(*p)->rb_left;
 898                 else if (bdi->id < id)
 899                         p = &(*p)->rb_right;
 900                 else
 901                         break;
 902         }
 903
 904         if (parentp)
 905                 *parentp = parent;
 906         return p;
 907 }
 908
 909 /**
 910  * bdi_get_by_id - lookup and get bdi from its id
 911  * @id: bdi id to lookup
 912  *
 913  * Find bdi matching @id and get it.  Returns NULL if the matching bdi
 914  * doesn't exist or is already unregistered.
 915  */
 916 struct backing_dev_info *bdi_get_by_id(u64 id)
 917 {
 918         struct backing_dev_info *bdi = NULL;
 919         struct rb_node **p;
 920
 921         spin_lock_bh(&bdi_lock);
 922         p = bdi_lookup_rb_node(id, NULL);
 923         if (*p) {
 924                 bdi = rb_entry(*p, struct backing_dev_info, rb_node);
 925                 bdi_get(bdi);
 926         }
 927         spin_unlock_bh(&bdi_lock);
 928
 929         return bdi;
 930 }
 931
 932 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 933 {
 934         struct device *dev;
 935         struct rb_node *parent, **p;
 936
 937         if (bdi->dev)   /* The driver needs to use separate queues per device */
 938                 return 0;
 939
 940         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 941         if (IS_ERR(dev))
 942                 return PTR_ERR(dev);
 943
 944         cgwb_bdi_register(bdi);
 945         bdi->dev = dev;
 946
 947         bdi_debug_register(bdi, dev_name(dev));
 948         set_bit(WB_registered, &bdi->wb.state);
 949
 950         spin_lock_bh(&bdi_lock);
 951
 952         bdi->id = ++bdi_id_cursor;
 953
 954         p = bdi_lookup_rb_node(bdi->id, &parent);
 955         rb_link_node(&bdi->rb_node, parent, p);
 956         rb_insert_color(&bdi->rb_node, &bdi_tree);
 957
 958         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 959
 960         spin_unlock_bh(&bdi_lock);
 961
 962         trace_writeback_bdi_register(bdi);
 963         return 0;
 964 }
 965 EXPORT_SYMBOL(bdi_register_va);
 966
 967 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 968 {
 969         va_list args;
 970         int ret;
 971
 972         va_start(args, fmt);
 973         ret = bdi_register_va(bdi, fmt, args);
 974         va_end(args);
 975         return ret;
 976 }
 977 EXPORT_SYMBOL(bdi_register);
 978
 979 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 980 {
 981         int rc;
 982
 983         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 984         if (rc)
 985                 return rc;
 986         /* Leaking owner reference... */
 987         WARN_ON(bdi->owner);
 988         bdi->owner = owner;
 989         get_device(owner);
 990         return 0;
 991 }
 992 EXPORT_SYMBOL(bdi_register_owner);
 993
 994 /*
 995  * Remove bdi from bdi_list, and ensure that it is no longer visible
 996  */
 997 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 998 {
 999         spin_lock_bh(&bdi_lock);
1000         rb_erase(&bdi->rb_node, &bdi_tree);
1001         list_del_rcu(&bdi->bdi_list);
1002         spin_unlock_bh(&bdi_lock);
1003
1004         synchronize_rcu_expedited();
1005 }
1006
1007 void bdi_unregister(struct backing_dev_info *bdi)
1008 {
1009         /* make sure nobody finds us on the bdi_list anymore */
1010         bdi_remove_from_list(bdi);
1011         wb_shutdown(&bdi->wb);
1012         cgwb_bdi_unregister(bdi);
1013
1014         if (bdi->dev) {
1015                 bdi_debug_unregister(bdi);
1016                 device_unregister(bdi->dev);
1017                 bdi->dev = NULL;
1018         }
1019
1020         if (bdi->owner) {
1021                 put_device(bdi->owner);
1022                 bdi->owner = NULL;
1023         }
1024 }
1025
1026 static void release_bdi(struct kref *ref)
1027 {
1028         struct backing_dev_info *bdi =
1029                         container_of(ref, struct backing_dev_info, refcnt);
1030
1031         if (test_bit(WB_registered, &bdi->wb.state))
1032                 bdi_unregister(bdi);
1033         WARN_ON_ONCE(bdi->dev);
1034         wb_exit(&bdi->wb);
1035         cgwb_bdi_exit(bdi);
1036         kfree(bdi);
1037 }
1038
1039 void bdi_put(struct backing_dev_info *bdi)
1040 {
1041         kref_put(&bdi->refcnt, release_bdi);
1042 }
1043 EXPORT_SYMBOL(bdi_put);
1044
1045 static wait_queue_head_t congestion_wqh[2] = {
1046                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
1047                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
1048         };
1049 static atomic_t nr_wb_congested[2];
1050
1051 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
1052 {
1053         wait_queue_head_t *wqh = &congestion_wqh[sync];
1054         enum wb_congested_state bit;
1055
1056         bit = sync ? WB_sync_congested : WB_async_congested;
1057         if (test_and_clear_bit(bit, &congested->state))
1058                 atomic_dec(&nr_wb_congested[sync]);
1059         smp_mb__after_atomic();
1060         if (waitqueue_active(wqh))
1061                 wake_up(wqh);
1062 }
1063 EXPORT_SYMBOL(clear_wb_congested);
1064
1065 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
1066 {
1067         enum wb_congested_state bit;
1068
1069         bit = sync ? WB_sync_congested : WB_async_congested;
1070         if (!test_and_set_bit(bit, &congested->state))
1071                 atomic_inc(&nr_wb_congested[sync]);
1072 }
1073 EXPORT_SYMBOL(set_wb_congested);
1074
1075 /**
1076  * congestion_wait - wait for a backing_dev to become uncongested
1077  * @sync: SYNC or ASYNC IO
1078  * @timeout: timeout in jiffies
1079  *
1080  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1081  * write congestion.  If no backing_devs are congested then just wait for the
1082  * next write to be completed.
1083  */
1084 long congestion_wait(int sync, long timeout)
1085 {
1086         long ret;
1087         unsigned long start = jiffies;
1088         DEFINE_WAIT(wait);
1089         wait_queue_head_t *wqh = &congestion_wqh[sync];
1090
1091         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1092         ret = io_schedule_timeout(timeout);
1093         finish_wait(wqh, &wait);
1094
1095         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1096                                         jiffies_to_usecs(jiffies - start));
1097
1098         return ret;
1099 }
1100 EXPORT_SYMBOL(congestion_wait);
1101
1102 /**
1103  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1104  * @sync: SYNC or ASYNC IO
1105  * @timeout: timeout in jiffies
1106  *
1107  * In the event of a congested backing_dev (any backing_dev) this waits
1108  * for up to @timeout jiffies for either a BDI to exit congestion of the
1109  * given @sync queue or a write to complete.
1110  *
1111  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1112  * it is the number of jiffies that were still remaining when the function
1113  * returned. return_value == timeout implies the function did not sleep.
1114  */
1115 long wait_iff_congested(int sync, long timeout)
1116 {
1117         long ret;
1118         unsigned long start = jiffies;
1119         DEFINE_WAIT(wait);
1120         wait_queue_head_t *wqh = &congestion_wqh[sync];
1121
1122         /*
1123          * If there is no congestion, yield if necessary instead
1124          * of sleeping on the congestion queue
1125          */
1126         if (atomic_read(&nr_wb_congested[sync]) == 0) {
1127                 cond_resched();
1128
1129                 /* In case we scheduled, work out time remaining */
1130                 ret = timeout - (jiffies - start);
1131                 if (ret < 0)
1132                         ret = 0;
1133
1134                 goto out;
1135         }
1136
1137         /* Sleep until uncongested or a write happens */
1138         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1139         ret = io_schedule_timeout(timeout);
1140         finish_wait(wqh, &wait);
1141
1142 out:
1143         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1144                                         jiffies_to_usecs(jiffies - start));
1145
1146         return ret;
1147 }
1148 EXPORT_SYMBOL(wait_iff_congested);