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 void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 107 {
 108         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 109
 110         debugfs_create_file("stats", 0444, bdi->debug_dir, bdi,
 111                             &bdi_debug_stats_fops);
 112 }
 113
 114 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 115 {
 116         debugfs_remove_recursive(bdi->debug_dir);
 117 }
 118 #else
 119 static inline void bdi_debug_init(void)
 120 {
 121 }
 122 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 123                                       const char *name)
 124 {
 125 }
 126 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 127 {
 128 }
 129 #endif
 130
 131 static ssize_t read_ahead_kb_store(struct device *dev,
 132                                   struct device_attribute *attr,
 133                                   const char *buf, size_t count)
 134 {
 135         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 136         unsigned long read_ahead_kb;
 137         ssize_t ret;
 138
 139         ret = kstrtoul(buf, 10, &read_ahead_kb);
 140         if (ret < 0)
 141                 return ret;
 142
 143         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 144
 145         return count;
 146 }
 147
 148 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 149
 150 #define BDI_SHOW(name, expr)                                            \
 151 static ssize_t name##_show(struct device *dev,                          \
 152                            struct device_attribute *attr, char *page)   \
 153 {                                                                       \
 154         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 155                                                                         \
 156         return snprintf(page, PAGE_SIZE-1, "%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 *page)
 203 {
 204         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 205
 206         return snprintf(page, PAGE_SIZE-1, "%d\n",
 207                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 208 }
 209 static DEVICE_ATTR_RO(stable_pages_required);
 210
 211 static struct attribute *bdi_dev_attrs[] = {
 212         &dev_attr_read_ahead_kb.attr,
 213         &dev_attr_min_ratio.attr,
 214         &dev_attr_max_ratio.attr,
 215         &dev_attr_stable_pages_required.attr,
 216         NULL,
 217 };
 218 ATTRIBUTE_GROUPS(bdi_dev);
 219
 220 static __init int bdi_class_init(void)
 221 {
 222         bdi_class = class_create(THIS_MODULE, "bdi");
 223         if (IS_ERR(bdi_class))
 224                 return PTR_ERR(bdi_class);
 225
 226         bdi_class->dev_groups = bdi_dev_groups;
 227         bdi_debug_init();
 228
 229         return 0;
 230 }
 231 postcore_initcall(bdi_class_init);
 232
 233 static int bdi_init(struct backing_dev_info *bdi);
 234
 235 static int __init default_bdi_init(void)
 236 {
 237         int err;
 238
 239         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
 240                                               WQ_UNBOUND | WQ_SYSFS, 0);
 241         if (!bdi_wq)
 242                 return -ENOMEM;
 243
 244         err = bdi_init(&noop_backing_dev_info);
 245
 246         return err;
 247 }
 248 subsys_initcall(default_bdi_init);
 249
 250 /*
 251  * This function is used when the first inode for this wb is marked dirty. It
 252  * wakes-up the corresponding bdi thread which should then take care of the
 253  * periodic background write-out of dirty inodes. Since the write-out would
 254  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 255  * set up a timer which wakes the bdi thread up later.
 256  *
 257  * Note, we wouldn't bother setting up the timer, but this function is on the
 258  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 259  * by delaying the wake-up.
 260  *
 261  * We have to be careful not to postpone flush work if it is scheduled for
 262  * earlier. Thus we use queue_delayed_work().
 263  */
 264 void wb_wakeup_delayed(struct bdi_writeback *wb)
 265 {
 266         unsigned long timeout;
 267
 268         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 269         spin_lock_bh(&wb->work_lock);
 270         if (test_bit(WB_registered, &wb->state))
 271                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 272         spin_unlock_bh(&wb->work_lock);
 273 }
 274
 275 /*
 276  * Initial write bandwidth: 100 MB/s
 277  */
 278 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 279
 280 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 281                    int blkcg_id, gfp_t gfp)
 282 {
 283         int i, err;
 284
 285         memset(wb, 0, sizeof(*wb));
 286
 287         if (wb != &bdi->wb)
 288                 bdi_get(bdi);
 289         wb->bdi = bdi;
 290         wb->last_old_flush = jiffies;
 291         INIT_LIST_HEAD(&wb->b_dirty);
 292         INIT_LIST_HEAD(&wb->b_io);
 293         INIT_LIST_HEAD(&wb->b_more_io);
 294         INIT_LIST_HEAD(&wb->b_dirty_time);
 295         spin_lock_init(&wb->list_lock);
 296
 297         wb->bw_time_stamp = jiffies;
 298         wb->balanced_dirty_ratelimit = INIT_BW;
 299         wb->dirty_ratelimit = INIT_BW;
 300         wb->write_bandwidth = INIT_BW;
 301         wb->avg_write_bandwidth = INIT_BW;
 302
 303         spin_lock_init(&wb->work_lock);
 304         INIT_LIST_HEAD(&wb->work_list);
 305         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 306         wb->dirty_sleep = jiffies;
 307
 308         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 309         if (!wb->congested) {
 310                 err = -ENOMEM;
 311                 goto out_put_bdi;
 312         }
 313
 314         err = fprop_local_init_percpu(&wb->completions, gfp);
 315         if (err)
 316                 goto out_put_cong;
 317
 318         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 319                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 320                 if (err)
 321                         goto out_destroy_stat;
 322         }
 323
 324         return 0;
 325
 326 out_destroy_stat:
 327         while (i--)
 328                 percpu_counter_destroy(&wb->stat[i]);
 329         fprop_local_destroy_percpu(&wb->completions);
 330 out_put_cong:
 331         wb_congested_put(wb->congested);
 332 out_put_bdi:
 333         if (wb != &bdi->wb)
 334                 bdi_put(bdi);
 335         return err;
 336 }
 337
 338 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 339
 340 /*
 341  * Remove bdi from the global list and shutdown any threads we have running
 342  */
 343 static void wb_shutdown(struct bdi_writeback *wb)
 344 {
 345         /* Make sure nobody queues further work */
 346         spin_lock_bh(&wb->work_lock);
 347         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 348                 spin_unlock_bh(&wb->work_lock);
 349                 return;
 350         }
 351         spin_unlock_bh(&wb->work_lock);
 352
 353         cgwb_remove_from_bdi_list(wb);
 354         /*
 355          * Drain work list and shutdown the delayed_work.  !WB_registered
 356          * tells wb_workfn() that @wb is dying and its work_list needs to
 357          * be drained no matter what.
 358          */
 359         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 360         flush_delayed_work(&wb->dwork);
 361         WARN_ON(!list_empty(&wb->work_list));
 362 }
 363
 364 static void wb_exit(struct bdi_writeback *wb)
 365 {
 366         int i;
 367
 368         WARN_ON(delayed_work_pending(&wb->dwork));
 369
 370         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 371                 percpu_counter_destroy(&wb->stat[i]);
 372
 373         fprop_local_destroy_percpu(&wb->completions);
 374         wb_congested_put(wb->congested);
 375         if (wb != &wb->bdi->wb)
 376                 bdi_put(wb->bdi);
 377 }
 378
 379 #ifdef CONFIG_CGROUP_WRITEBACK
 380
 381 #include <linux/memcontrol.h>
 382
 383 /*
 384  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 385  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 386  * protected.
 387  */
 388 static DEFINE_SPINLOCK(cgwb_lock);
 389 static struct workqueue_struct *cgwb_release_wq;
 390
 391 /**
 392  * wb_congested_get_create - get or create a wb_congested
 393  * @bdi: associated bdi
 394  * @blkcg_id: ID of the associated blkcg
 395  * @gfp: allocation mask
 396  *
 397  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 398  * The returned wb_congested has its reference count incremented.  Returns
 399  * NULL on failure.
 400  */
 401 struct bdi_writeback_congested *
 402 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 403 {
 404         struct bdi_writeback_congested *new_congested = NULL, *congested;
 405         struct rb_node **node, *parent;
 406         unsigned long flags;
 407 retry:
 408         spin_lock_irqsave(&cgwb_lock, flags);
 409
 410         node = &bdi->cgwb_congested_tree.rb_node;
 411         parent = NULL;
 412
 413         while (*node != NULL) {
 414                 parent = *node;
 415                 congested = rb_entry(parent, struct bdi_writeback_congested,
 416                                      rb_node);
 417                 if (congested->blkcg_id < blkcg_id)
 418                         node = &parent->rb_left;
 419                 else if (congested->blkcg_id > blkcg_id)
 420                         node = &parent->rb_right;
 421                 else
 422                         goto found;
 423         }
 424
 425         if (new_congested) {
 426                 /* !found and storage for new one already allocated, insert */
 427                 congested = new_congested;
 428                 rb_link_node(&congested->rb_node, parent, node);
 429                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 430                 spin_unlock_irqrestore(&cgwb_lock, flags);
 431                 return congested;
 432         }
 433
 434         spin_unlock_irqrestore(&cgwb_lock, flags);
 435
 436         /* allocate storage for new one and retry */
 437         new_congested = kzalloc(sizeof(*new_congested), gfp);
 438         if (!new_congested)
 439                 return NULL;
 440
 441         refcount_set(&new_congested->refcnt, 1);
 442         new_congested->__bdi = bdi;
 443         new_congested->blkcg_id = blkcg_id;
 444         goto retry;
 445
 446 found:
 447         refcount_inc(&congested->refcnt);
 448         spin_unlock_irqrestore(&cgwb_lock, flags);
 449         kfree(new_congested);
 450         return congested;
 451 }
 452
 453 /**
 454  * wb_congested_put - put a wb_congested
 455  * @congested: wb_congested to put
 456  *
 457  * Put @congested and destroy it if the refcnt reaches zero.
 458  */
 459 void wb_congested_put(struct bdi_writeback_congested *congested)
 460 {
 461         unsigned long flags;
 462
 463         if (!refcount_dec_and_lock_irqsave(&congested->refcnt, &cgwb_lock, &flags))
 464                 return;
 465
 466         /* bdi might already have been destroyed leaving @congested unlinked */
 467         if (congested->__bdi) {
 468                 rb_erase(&congested->rb_node,
 469                          &congested->__bdi->cgwb_congested_tree);
 470                 congested->__bdi = NULL;
 471         }
 472
 473         spin_unlock_irqrestore(&cgwb_lock, flags);
 474         kfree(congested);
 475 }
 476
 477 static void cgwb_release_workfn(struct work_struct *work)
 478 {
 479         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 480                                                 release_work);
 481         struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css);
 482
 483         mutex_lock(&wb->bdi->cgwb_release_mutex);
 484         wb_shutdown(wb);
 485
 486         css_put(wb->memcg_css);
 487         css_put(wb->blkcg_css);
 488         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 489
 490         /* triggers blkg destruction if cgwb_refcnt becomes zero */
 491         blkcg_cgwb_put(blkcg);
 492
 493         fprop_local_destroy_percpu(&wb->memcg_completions);
 494         percpu_ref_exit(&wb->refcnt);
 495         wb_exit(wb);
 496         kfree_rcu(wb, rcu);
 497 }
 498
 499 static void cgwb_release(struct percpu_ref *refcnt)
 500 {
 501         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 502                                                 refcnt);
 503         queue_work(cgwb_release_wq, &wb->release_work);
 504 }
 505
 506 static void cgwb_kill(struct bdi_writeback *wb)
 507 {
 508         lockdep_assert_held(&cgwb_lock);
 509
 510         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 511         list_del(&wb->memcg_node);
 512         list_del(&wb->blkcg_node);
 513         percpu_ref_kill(&wb->refcnt);
 514 }
 515
 516 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 517 {
 518         spin_lock_irq(&cgwb_lock);
 519         list_del_rcu(&wb->bdi_node);
 520         spin_unlock_irq(&cgwb_lock);
 521 }
 522
 523 static int cgwb_create(struct backing_dev_info *bdi,
 524                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 525 {
 526         struct mem_cgroup *memcg;
 527         struct cgroup_subsys_state *blkcg_css;
 528         struct blkcg *blkcg;
 529         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 530         struct bdi_writeback *wb;
 531         unsigned long flags;
 532         int ret = 0;
 533
 534         memcg = mem_cgroup_from_css(memcg_css);
 535         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 536         blkcg = css_to_blkcg(blkcg_css);
 537         memcg_cgwb_list = &memcg->cgwb_list;
 538         blkcg_cgwb_list = &blkcg->cgwb_list;
 539
 540         /* look up again under lock and discard on blkcg mismatch */
 541         spin_lock_irqsave(&cgwb_lock, flags);
 542         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 543         if (wb && wb->blkcg_css != blkcg_css) {
 544                 cgwb_kill(wb);
 545                 wb = NULL;
 546         }
 547         spin_unlock_irqrestore(&cgwb_lock, flags);
 548         if (wb)
 549                 goto out_put;
 550
 551         /* need to create a new one */
 552         wb = kmalloc(sizeof(*wb), gfp);
 553         if (!wb) {
 554                 ret = -ENOMEM;
 555                 goto out_put;
 556         }
 557
 558         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 559         if (ret)
 560                 goto err_free;
 561
 562         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 563         if (ret)
 564                 goto err_wb_exit;
 565
 566         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 567         if (ret)
 568                 goto err_ref_exit;
 569
 570         wb->memcg_css = memcg_css;
 571         wb->blkcg_css = blkcg_css;
 572         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 573         set_bit(WB_registered, &wb->state);
 574
 575         /*
 576          * The root wb determines the registered state of the whole bdi and
 577          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 578          * whether they're still online.  Don't link @wb if any is dead.
 579          * See wb_memcg_offline() and wb_blkcg_offline().
 580          */
 581         ret = -ENODEV;
 582         spin_lock_irqsave(&cgwb_lock, flags);
 583         if (test_bit(WB_registered, &bdi->wb.state) &&
 584             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 585                 /* we might have raced another instance of this function */
 586                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 587                 if (!ret) {
 588                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 589                         list_add(&wb->memcg_node, memcg_cgwb_list);
 590                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 591                         blkcg_cgwb_get(blkcg);
 592                         css_get(memcg_css);
 593                         css_get(blkcg_css);
 594                 }
 595         }
 596         spin_unlock_irqrestore(&cgwb_lock, flags);
 597         if (ret) {
 598                 if (ret == -EEXIST)
 599                         ret = 0;
 600                 goto err_fprop_exit;
 601         }
 602         goto out_put;
 603
 604 err_fprop_exit:
 605         fprop_local_destroy_percpu(&wb->memcg_completions);
 606 err_ref_exit:
 607         percpu_ref_exit(&wb->refcnt);
 608 err_wb_exit:
 609         wb_exit(wb);
 610 err_free:
 611         kfree(wb);
 612 out_put:
 613         css_put(blkcg_css);
 614         return ret;
 615 }
 616
 617 /**
 618  * wb_get_create - get wb for a given memcg, create if necessary
 619  * @bdi: target bdi
 620  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 621  * @gfp: allocation mask to use
 622  *
 623  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 624  * create one.  The returned wb has its refcount incremented.
 625  *
 626  * This function uses css_get() on @memcg_css and thus expects its refcnt
 627  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 628  * @memcg_css isn't enough.  try_get it before calling this function.
 629  *
 630  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 631  * memcg on the default hierarchy, memcg association is guaranteed to be
 632  * more specific (equal or descendant to the associated blkcg) and thus can
 633  * identify both the memcg and blkcg associations.
 634  *
 635  * Because the blkcg associated with a memcg may change as blkcg is enabled
 636  * and disabled closer to root in the hierarchy, each wb keeps track of
 637  * both the memcg and blkcg associated with it and verifies the blkcg on
 638  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 639  * created.
 640  */
 641 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 642                                     struct cgroup_subsys_state *memcg_css,
 643                                     gfp_t gfp)
 644 {
 645         struct bdi_writeback *wb;
 646
 647         might_sleep_if(gfpflags_allow_blocking(gfp));
 648
 649         if (!memcg_css->parent)
 650                 return &bdi->wb;
 651
 652         do {
 653                 rcu_read_lock();
 654                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 655                 if (wb) {
 656                         struct cgroup_subsys_state *blkcg_css;
 657
 658                         /* see whether the blkcg association has changed */
 659                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 660                                                      &io_cgrp_subsys);
 661                         if (unlikely(wb->blkcg_css != blkcg_css ||
 662                                      !wb_tryget(wb)))
 663                                 wb = NULL;
 664                         css_put(blkcg_css);
 665                 }
 666                 rcu_read_unlock();
 667         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 668
 669         return wb;
 670 }
 671
 672 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 673 {
 674         int ret;
 675
 676         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 677         bdi->cgwb_congested_tree = RB_ROOT;
 678         mutex_init(&bdi->cgwb_release_mutex);
 679         init_rwsem(&bdi->wb_switch_rwsem);
 680
 681         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 682         if (!ret) {
 683                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 684                 bdi->wb.blkcg_css = blkcg_root_css;
 685         }
 686         return ret;
 687 }
 688
 689 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 690 {
 691         struct radix_tree_iter iter;
 692         void **slot;
 693         struct bdi_writeback *wb;
 694
 695         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 696
 697         spin_lock_irq(&cgwb_lock);
 698         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 699                 cgwb_kill(*slot);
 700         spin_unlock_irq(&cgwb_lock);
 701
 702         mutex_lock(&bdi->cgwb_release_mutex);
 703         spin_lock_irq(&cgwb_lock);
 704         while (!list_empty(&bdi->wb_list)) {
 705                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 706                                       bdi_node);
 707                 spin_unlock_irq(&cgwb_lock);
 708                 wb_shutdown(wb);
 709                 spin_lock_irq(&cgwb_lock);
 710         }
 711         spin_unlock_irq(&cgwb_lock);
 712         mutex_unlock(&bdi->cgwb_release_mutex);
 713 }
 714
 715 /**
 716  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 717  * @memcg: memcg being offlined
 718  *
 719  * Also prevents creation of any new wb's associated with @memcg.
 720  */
 721 void wb_memcg_offline(struct mem_cgroup *memcg)
 722 {
 723         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 724         struct bdi_writeback *wb, *next;
 725
 726         spin_lock_irq(&cgwb_lock);
 727         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 728                 cgwb_kill(wb);
 729         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 730         spin_unlock_irq(&cgwb_lock);
 731 }
 732
 733 /**
 734  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 735  * @blkcg: blkcg being offlined
 736  *
 737  * Also prevents creation of any new wb's associated with @blkcg.
 738  */
 739 void wb_blkcg_offline(struct blkcg *blkcg)
 740 {
 741         struct bdi_writeback *wb, *next;
 742
 743         spin_lock_irq(&cgwb_lock);
 744         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 745                 cgwb_kill(wb);
 746         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 747         spin_unlock_irq(&cgwb_lock);
 748 }
 749
 750 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 751 {
 752         struct rb_node *rbn;
 753
 754         spin_lock_irq(&cgwb_lock);
 755         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 756                 struct bdi_writeback_congested *congested =
 757                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 758
 759                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 760                 congested->__bdi = NULL;        /* mark @congested unlinked */
 761         }
 762         spin_unlock_irq(&cgwb_lock);
 763 }
 764
 765 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 766 {
 767         spin_lock_irq(&cgwb_lock);
 768         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 769         spin_unlock_irq(&cgwb_lock);
 770 }
 771
 772 static int __init cgwb_init(void)
 773 {
 774         /*
 775          * There can be many concurrent release work items overwhelming
 776          * system_wq.  Put them in a separate wq and limit concurrency.
 777          * There's no point in executing many of these in parallel.
 778          */
 779         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 780         if (!cgwb_release_wq)
 781                 return -ENOMEM;
 782
 783         return 0;
 784 }
 785 subsys_initcall(cgwb_init);
 786
 787 #else   /* CONFIG_CGROUP_WRITEBACK */
 788
 789 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 790 {
 791         int err;
 792
 793         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 794         if (!bdi->wb_congested)
 795                 return -ENOMEM;
 796
 797         refcount_set(&bdi->wb_congested->refcnt, 1);
 798
 799         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 800         if (err) {
 801                 wb_congested_put(bdi->wb_congested);
 802                 return err;
 803         }
 804         return 0;
 805 }
 806
 807 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 808
 809 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 810 {
 811         wb_congested_put(bdi->wb_congested);
 812 }
 813
 814 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 815 {
 816         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 817 }
 818
 819 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 820 {
 821         list_del_rcu(&wb->bdi_node);
 822 }
 823
 824 #endif  /* CONFIG_CGROUP_WRITEBACK */
 825
 826 static int bdi_init(struct backing_dev_info *bdi)
 827 {
 828         int ret;
 829
 830         bdi->dev = NULL;
 831
 832         kref_init(&bdi->refcnt);
 833         bdi->min_ratio = 0;
 834         bdi->max_ratio = 100;
 835         bdi->max_prop_frac = FPROP_FRAC_BASE;
 836         INIT_LIST_HEAD(&bdi->bdi_list);
 837         INIT_LIST_HEAD(&bdi->wb_list);
 838         init_waitqueue_head(&bdi->wb_waitq);
 839
 840         ret = cgwb_bdi_init(bdi);
 841
 842         return ret;
 843 }
 844
 845 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 846 {
 847         struct backing_dev_info *bdi;
 848
 849         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 850                            gfp_mask | __GFP_ZERO, node_id);
 851         if (!bdi)
 852                 return NULL;
 853
 854         if (bdi_init(bdi)) {
 855                 kfree(bdi);
 856                 return NULL;
 857         }
 858         return bdi;
 859 }
 860 EXPORT_SYMBOL(bdi_alloc_node);
 861
 862 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 863 {
 864         struct device *dev;
 865
 866         if (bdi->dev)   /* The driver needs to use separate queues per device */
 867                 return 0;
 868
 869         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 870         if (IS_ERR(dev))
 871                 return PTR_ERR(dev);
 872
 873         cgwb_bdi_register(bdi);
 874         bdi->dev = dev;
 875
 876         bdi_debug_register(bdi, dev_name(dev));
 877         set_bit(WB_registered, &bdi->wb.state);
 878
 879         spin_lock_bh(&bdi_lock);
 880         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 881         spin_unlock_bh(&bdi_lock);
 882
 883         trace_writeback_bdi_register(bdi);
 884         return 0;
 885 }
 886 EXPORT_SYMBOL(bdi_register_va);
 887
 888 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 889 {
 890         va_list args;
 891         int ret;
 892
 893         va_start(args, fmt);
 894         ret = bdi_register_va(bdi, fmt, args);
 895         va_end(args);
 896         return ret;
 897 }
 898 EXPORT_SYMBOL(bdi_register);
 899
 900 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 901 {
 902         int rc;
 903
 904         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 905         if (rc)
 906                 return rc;
 907         /* Leaking owner reference... */
 908         WARN_ON(bdi->owner);
 909         bdi->owner = owner;
 910         get_device(owner);
 911         return 0;
 912 }
 913 EXPORT_SYMBOL(bdi_register_owner);
 914
 915 /*
 916  * Remove bdi from bdi_list, and ensure that it is no longer visible
 917  */
 918 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 919 {
 920         spin_lock_bh(&bdi_lock);
 921         list_del_rcu(&bdi->bdi_list);
 922         spin_unlock_bh(&bdi_lock);
 923
 924         synchronize_rcu_expedited();
 925 }
 926
 927 void bdi_unregister(struct backing_dev_info *bdi)
 928 {
 929         /* make sure nobody finds us on the bdi_list anymore */
 930         bdi_remove_from_list(bdi);
 931         wb_shutdown(&bdi->wb);
 932         cgwb_bdi_unregister(bdi);
 933
 934         if (bdi->dev) {
 935                 bdi_debug_unregister(bdi);
 936                 device_unregister(bdi->dev);
 937                 bdi->dev = NULL;
 938         }
 939
 940         if (bdi->owner) {
 941                 put_device(bdi->owner);
 942                 bdi->owner = NULL;
 943         }
 944 }
 945
 946 static void release_bdi(struct kref *ref)
 947 {
 948         struct backing_dev_info *bdi =
 949                         container_of(ref, struct backing_dev_info, refcnt);
 950
 951         if (test_bit(WB_registered, &bdi->wb.state))
 952                 bdi_unregister(bdi);
 953         WARN_ON_ONCE(bdi->dev);
 954         wb_exit(&bdi->wb);
 955         cgwb_bdi_exit(bdi);
 956         kfree(bdi);
 957 }
 958
 959 void bdi_put(struct backing_dev_info *bdi)
 960 {
 961         kref_put(&bdi->refcnt, release_bdi);
 962 }
 963 EXPORT_SYMBOL(bdi_put);
 964
 965 static wait_queue_head_t congestion_wqh[2] = {
 966                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 967                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 968         };
 969 static atomic_t nr_wb_congested[2];
 970
 971 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 972 {
 973         wait_queue_head_t *wqh = &congestion_wqh[sync];
 974         enum wb_congested_state bit;
 975
 976         bit = sync ? WB_sync_congested : WB_async_congested;
 977         if (test_and_clear_bit(bit, &congested->state))
 978                 atomic_dec(&nr_wb_congested[sync]);
 979         smp_mb__after_atomic();
 980         if (waitqueue_active(wqh))
 981                 wake_up(wqh);
 982 }
 983 EXPORT_SYMBOL(clear_wb_congested);
 984
 985 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
 986 {
 987         enum wb_congested_state bit;
 988
 989         bit = sync ? WB_sync_congested : WB_async_congested;
 990         if (!test_and_set_bit(bit, &congested->state))
 991                 atomic_inc(&nr_wb_congested[sync]);
 992 }
 993 EXPORT_SYMBOL(set_wb_congested);
 994
 995 /**
 996  * congestion_wait - wait for a backing_dev to become uncongested
 997  * @sync: SYNC or ASYNC IO
 998  * @timeout: timeout in jiffies
 999  *
1000  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1001  * write congestion.  If no backing_devs are congested then just wait for the
1002  * next write to be completed.
1003  */
1004 long congestion_wait(int sync, long timeout)
1005 {
1006         long ret;
1007         unsigned long start = jiffies;
1008         DEFINE_WAIT(wait);
1009         wait_queue_head_t *wqh = &congestion_wqh[sync];
1010
1011         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1012         ret = io_schedule_timeout(timeout);
1013         finish_wait(wqh, &wait);
1014
1015         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1016                                         jiffies_to_usecs(jiffies - start));
1017
1018         return ret;
1019 }
1020 EXPORT_SYMBOL(congestion_wait);
1021
1022 /**
1023  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1024  * @sync: SYNC or ASYNC IO
1025  * @timeout: timeout in jiffies
1026  *
1027  * In the event of a congested backing_dev (any backing_dev) this waits
1028  * for up to @timeout jiffies for either a BDI to exit congestion of the
1029  * given @sync queue or a write to complete.
1030  *
1031  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1032  * it is the number of jiffies that were still remaining when the function
1033  * returned. return_value == timeout implies the function did not sleep.
1034  */
1035 long wait_iff_congested(int sync, long timeout)
1036 {
1037         long ret;
1038         unsigned long start = jiffies;
1039         DEFINE_WAIT(wait);
1040         wait_queue_head_t *wqh = &congestion_wqh[sync];
1041
1042         /*
1043          * If there is no congestion, yield if necessary instead
1044          * of sleeping on the congestion queue
1045          */
1046         if (atomic_read(&nr_wb_congested[sync]) == 0) {
1047                 cond_resched();
1048
1049                 /* In case we scheduled, work out time remaining */
1050                 ret = timeout - (jiffies - start);
1051                 if (ret < 0)
1052                         ret = 0;
1053
1054                 goto out;
1055         }
1056
1057         /* Sleep until uncongested or a write happens */
1058         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1059         ret = io_schedule_timeout(timeout);
1060         finish_wait(wqh, &wait);
1061
1062 out:
1063         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1064                                         jiffies_to_usecs(jiffies - start));
1065
1066         return ret;
1067 }
1068 EXPORT_SYMBOL(wait_iff_congested);