mm/backing-dev.c

   1
   2 #include <linux/wait.h>
   3 #include <linux/backing-dev.h>
   4 #include <linux/kthread.h>
   5 #include <linux/freezer.h>
   6 #include <linux/fs.h>
   7 #include <linux/pagemap.h>
   8 #include <linux/mm.h>
   9 #include <linux/sched.h>
  10 #include <linux/module.h>
  11 #include <linux/writeback.h>
  12 #include <linux/device.h>
  13 #include <trace/events/writeback.h>
  14
  15 struct backing_dev_info noop_backing_dev_info = {
  16         .name           = "noop",
  17         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
  18 };
  19 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  20
  21 static struct class *bdi_class;
  22
  23 /*
  24  * bdi_lock protects updates to bdi_list. bdi_list has RCU reader side
  25  * locking.
  26  */
  27 DEFINE_SPINLOCK(bdi_lock);
  28 LIST_HEAD(bdi_list);
  29
  30 /* bdi_wq serves all asynchronous writeback tasks */
  31 struct workqueue_struct *bdi_wq;
  32
  33 #ifdef CONFIG_DEBUG_FS
  34 #include <linux/debugfs.h>
  35 #include <linux/seq_file.h>
  36
  37 static struct dentry *bdi_debug_root;
  38
  39 static void bdi_debug_init(void)
  40 {
  41         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  42 }
  43
  44 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  45 {
  46         struct backing_dev_info *bdi = m->private;
  47         struct bdi_writeback *wb = &bdi->wb;
  48         unsigned long background_thresh;
  49         unsigned long dirty_thresh;
  50         unsigned long wb_thresh;
  51         unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time;
  52         struct inode *inode;
  53
  54         nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0;
  55         spin_lock(&wb->list_lock);
  56         list_for_each_entry(inode, &wb->b_dirty, i_io_list)
  57                 nr_dirty++;
  58         list_for_each_entry(inode, &wb->b_io, i_io_list)
  59                 nr_io++;
  60         list_for_each_entry(inode, &wb->b_more_io, i_io_list)
  61                 nr_more_io++;
  62         list_for_each_entry(inode, &wb->b_dirty_time, i_io_list)
  63                 if (inode->i_state & I_DIRTY_TIME)
  64                         nr_dirty_time++;
  65         spin_unlock(&wb->list_lock);
  66
  67         global_dirty_limits(&background_thresh, &dirty_thresh);
  68         wb_thresh = wb_calc_thresh(wb, dirty_thresh);
  69
  70 #define K(x) ((x) << (PAGE_SHIFT - 10))
  71         seq_printf(m,
  72                    "BdiWriteback:       %10lu kB\n"
  73                    "BdiReclaimable:     %10lu kB\n"
  74                    "BdiDirtyThresh:     %10lu kB\n"
  75                    "DirtyThresh:        %10lu kB\n"
  76                    "BackgroundThresh:   %10lu kB\n"
  77                    "BdiDirtied:         %10lu kB\n"
  78                    "BdiWritten:         %10lu kB\n"
  79                    "BdiWriteBandwidth:  %10lu kBps\n"
  80                    "b_dirty:            %10lu\n"
  81                    "b_io:               %10lu\n"
  82                    "b_more_io:          %10lu\n"
  83                    "b_dirty_time:       %10lu\n"
  84                    "bdi_list:           %10u\n"
  85                    "state:              %10lx\n",
  86                    (unsigned long) K(wb_stat(wb, WB_WRITEBACK)),
  87                    (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)),
  88                    K(wb_thresh),
  89                    K(dirty_thresh),
  90                    K(background_thresh),
  91                    (unsigned long) K(wb_stat(wb, WB_DIRTIED)),
  92                    (unsigned long) K(wb_stat(wb, WB_WRITTEN)),
  93                    (unsigned long) K(wb->write_bandwidth),
  94                    nr_dirty,
  95                    nr_io,
  96                    nr_more_io,
  97                    nr_dirty_time,
  98                    !list_empty(&bdi->bdi_list), bdi->wb.state);
  99 #undef K
 100
 101         return 0;
 102 }
 103 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats);
 104
 105 static int bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 106 {
 107         if (!bdi_debug_root)
 108                 return -ENOMEM;
 109
 110         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 111         if (!bdi->debug_dir)
 112                 return -ENOMEM;
 113
 114         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 115                                                bdi, &bdi_debug_stats_fops);
 116         if (!bdi->debug_stats) {
 117                 debugfs_remove(bdi->debug_dir);
 118                 bdi->debug_dir = NULL;
 119                 return -ENOMEM;
 120         }
 121
 122         return 0;
 123 }
 124
 125 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 126 {
 127         debugfs_remove(bdi->debug_stats);
 128         debugfs_remove(bdi->debug_dir);
 129 }
 130 #else
 131 static inline void bdi_debug_init(void)
 132 {
 133 }
 134 static inline int bdi_debug_register(struct backing_dev_info *bdi,
 135                                       const char *name)
 136 {
 137         return 0;
 138 }
 139 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 140 {
 141 }
 142 #endif
 143
 144 static ssize_t read_ahead_kb_store(struct device *dev,
 145                                   struct device_attribute *attr,
 146                                   const char *buf, size_t count)
 147 {
 148         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 149         unsigned long read_ahead_kb;
 150         ssize_t ret;
 151
 152         ret = kstrtoul(buf, 10, &read_ahead_kb);
 153         if (ret < 0)
 154                 return ret;
 155
 156         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 157
 158         return count;
 159 }
 160
 161 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 162
 163 #define BDI_SHOW(name, expr)                                            \
 164 static ssize_t name##_show(struct device *dev,                          \
 165                            struct device_attribute *attr, char *page)   \
 166 {                                                                       \
 167         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 168                                                                         \
 169         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 170 }                                                                       \
 171 static DEVICE_ATTR_RW(name);
 172
 173 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 174
 175 static ssize_t min_ratio_store(struct device *dev,
 176                 struct device_attribute *attr, const char *buf, size_t count)
 177 {
 178         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 179         unsigned int ratio;
 180         ssize_t ret;
 181
 182         ret = kstrtouint(buf, 10, &ratio);
 183         if (ret < 0)
 184                 return ret;
 185
 186         ret = bdi_set_min_ratio(bdi, ratio);
 187         if (!ret)
 188                 ret = count;
 189
 190         return ret;
 191 }
 192 BDI_SHOW(min_ratio, bdi->min_ratio)
 193
 194 static ssize_t max_ratio_store(struct device *dev,
 195                 struct device_attribute *attr, const char *buf, size_t count)
 196 {
 197         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 198         unsigned int ratio;
 199         ssize_t ret;
 200
 201         ret = kstrtouint(buf, 10, &ratio);
 202         if (ret < 0)
 203                 return ret;
 204
 205         ret = bdi_set_max_ratio(bdi, ratio);
 206         if (!ret)
 207                 ret = count;
 208
 209         return ret;
 210 }
 211 BDI_SHOW(max_ratio, bdi->max_ratio)
 212
 213 static ssize_t stable_pages_required_show(struct device *dev,
 214                                           struct device_attribute *attr,
 215                                           char *page)
 216 {
 217         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 218
 219         return snprintf(page, PAGE_SIZE-1, "%d\n",
 220                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 221 }
 222 static DEVICE_ATTR_RO(stable_pages_required);
 223
 224 static struct attribute *bdi_dev_attrs[] = {
 225         &dev_attr_read_ahead_kb.attr,
 226         &dev_attr_min_ratio.attr,
 227         &dev_attr_max_ratio.attr,
 228         &dev_attr_stable_pages_required.attr,
 229         NULL,
 230 };
 231 ATTRIBUTE_GROUPS(bdi_dev);
 232
 233 static __init int bdi_class_init(void)
 234 {
 235         bdi_class = class_create(THIS_MODULE, "bdi");
 236         if (IS_ERR(bdi_class))
 237                 return PTR_ERR(bdi_class);
 238
 239         bdi_class->dev_groups = bdi_dev_groups;
 240         bdi_debug_init();
 241
 242         return 0;
 243 }
 244 postcore_initcall(bdi_class_init);
 245
 246 static int bdi_init(struct backing_dev_info *bdi);
 247
 248 static int __init default_bdi_init(void)
 249 {
 250         int err;
 251
 252         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
 253                                               WQ_UNBOUND | WQ_SYSFS, 0);
 254         if (!bdi_wq)
 255                 return -ENOMEM;
 256
 257         err = bdi_init(&noop_backing_dev_info);
 258
 259         return err;
 260 }
 261 subsys_initcall(default_bdi_init);
 262
 263 /*
 264  * This function is used when the first inode for this wb is marked dirty. It
 265  * wakes-up the corresponding bdi thread which should then take care of the
 266  * periodic background write-out of dirty inodes. Since the write-out would
 267  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 268  * set up a timer which wakes the bdi thread up later.
 269  *
 270  * Note, we wouldn't bother setting up the timer, but this function is on the
 271  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 272  * by delaying the wake-up.
 273  *
 274  * We have to be careful not to postpone flush work if it is scheduled for
 275  * earlier. Thus we use queue_delayed_work().
 276  */
 277 void wb_wakeup_delayed(struct bdi_writeback *wb)
 278 {
 279         unsigned long timeout;
 280
 281         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 282         spin_lock_bh(&wb->work_lock);
 283         if (test_bit(WB_registered, &wb->state))
 284                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 285         spin_unlock_bh(&wb->work_lock);
 286 }
 287
 288 /*
 289  * Initial write bandwidth: 100 MB/s
 290  */
 291 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 292
 293 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 294                    int blkcg_id, gfp_t gfp)
 295 {
 296         int i, err;
 297
 298         memset(wb, 0, sizeof(*wb));
 299
 300         if (wb != &bdi->wb)
 301                 bdi_get(bdi);
 302         wb->bdi = bdi;
 303         wb->last_old_flush = jiffies;
 304         INIT_LIST_HEAD(&wb->b_dirty);
 305         INIT_LIST_HEAD(&wb->b_io);
 306         INIT_LIST_HEAD(&wb->b_more_io);
 307         INIT_LIST_HEAD(&wb->b_dirty_time);
 308         spin_lock_init(&wb->list_lock);
 309
 310         wb->bw_time_stamp = jiffies;
 311         wb->balanced_dirty_ratelimit = INIT_BW;
 312         wb->dirty_ratelimit = INIT_BW;
 313         wb->write_bandwidth = INIT_BW;
 314         wb->avg_write_bandwidth = INIT_BW;
 315
 316         spin_lock_init(&wb->work_lock);
 317         INIT_LIST_HEAD(&wb->work_list);
 318         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 319         wb->dirty_sleep = jiffies;
 320
 321         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 322         if (!wb->congested) {
 323                 err = -ENOMEM;
 324                 goto out_put_bdi;
 325         }
 326
 327         err = fprop_local_init_percpu(&wb->completions, gfp);
 328         if (err)
 329                 goto out_put_cong;
 330
 331         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 332                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 333                 if (err)
 334                         goto out_destroy_stat;
 335         }
 336
 337         return 0;
 338
 339 out_destroy_stat:
 340         while (i--)
 341                 percpu_counter_destroy(&wb->stat[i]);
 342         fprop_local_destroy_percpu(&wb->completions);
 343 out_put_cong:
 344         wb_congested_put(wb->congested);
 345 out_put_bdi:
 346         if (wb != &bdi->wb)
 347                 bdi_put(bdi);
 348         return err;
 349 }
 350
 351 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 352
 353 /*
 354  * Remove bdi from the global list and shutdown any threads we have running
 355  */
 356 static void wb_shutdown(struct bdi_writeback *wb)
 357 {
 358         /* Make sure nobody queues further work */
 359         spin_lock_bh(&wb->work_lock);
 360         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 361                 spin_unlock_bh(&wb->work_lock);
 362                 /*
 363                  * Wait for wb shutdown to finish if someone else is just
 364                  * running wb_shutdown(). Otherwise we could proceed to wb /
 365                  * bdi destruction before wb_shutdown() is finished.
 366                  */
 367                 wait_on_bit(&wb->state, WB_shutting_down, TASK_UNINTERRUPTIBLE);
 368                 return;
 369         }
 370         set_bit(WB_shutting_down, &wb->state);
 371         spin_unlock_bh(&wb->work_lock);
 372
 373         cgwb_remove_from_bdi_list(wb);
 374         /*
 375          * Drain work list and shutdown the delayed_work.  !WB_registered
 376          * tells wb_workfn() that @wb is dying and its work_list needs to
 377          * be drained no matter what.
 378          */
 379         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 380         flush_delayed_work(&wb->dwork);
 381         WARN_ON(!list_empty(&wb->work_list));
 382         /*
 383          * Make sure bit gets cleared after shutdown is finished. Matches with
 384          * the barrier provided by test_and_clear_bit() above.
 385          */
 386         smp_wmb();
 387         clear_and_wake_up_bit(WB_shutting_down, &wb->state);
 388 }
 389
 390 static void wb_exit(struct bdi_writeback *wb)
 391 {
 392         int i;
 393
 394         WARN_ON(delayed_work_pending(&wb->dwork));
 395
 396         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 397                 percpu_counter_destroy(&wb->stat[i]);
 398
 399         fprop_local_destroy_percpu(&wb->completions);
 400         wb_congested_put(wb->congested);
 401         if (wb != &wb->bdi->wb)
 402                 bdi_put(wb->bdi);
 403 }
 404
 405 #ifdef CONFIG_CGROUP_WRITEBACK
 406
 407 #include <linux/memcontrol.h>
 408
 409 /*
 410  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 411  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 412  * protected.
 413  */
 414 static DEFINE_SPINLOCK(cgwb_lock);
 415 static struct workqueue_struct *cgwb_release_wq;
 416
 417 /**
 418  * wb_congested_get_create - get or create a wb_congested
 419  * @bdi: associated bdi
 420  * @blkcg_id: ID of the associated blkcg
 421  * @gfp: allocation mask
 422  *
 423  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 424  * The returned wb_congested has its reference count incremented.  Returns
 425  * NULL on failure.
 426  */
 427 struct bdi_writeback_congested *
 428 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 429 {
 430         struct bdi_writeback_congested *new_congested = NULL, *congested;
 431         struct rb_node **node, *parent;
 432         unsigned long flags;
 433 retry:
 434         spin_lock_irqsave(&cgwb_lock, flags);
 435
 436         node = &bdi->cgwb_congested_tree.rb_node;
 437         parent = NULL;
 438
 439         while (*node != NULL) {
 440                 parent = *node;
 441                 congested = rb_entry(parent, struct bdi_writeback_congested,
 442                                      rb_node);
 443                 if (congested->blkcg_id < blkcg_id)
 444                         node = &parent->rb_left;
 445                 else if (congested->blkcg_id > blkcg_id)
 446                         node = &parent->rb_right;
 447                 else
 448                         goto found;
 449         }
 450
 451         if (new_congested) {
 452                 /* !found and storage for new one already allocated, insert */
 453                 congested = new_congested;
 454                 new_congested = NULL;
 455                 rb_link_node(&congested->rb_node, parent, node);
 456                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 457                 goto found;
 458         }
 459
 460         spin_unlock_irqrestore(&cgwb_lock, flags);
 461
 462         /* allocate storage for new one and retry */
 463         new_congested = kzalloc(sizeof(*new_congested), gfp);
 464         if (!new_congested)
 465                 return NULL;
 466
 467         atomic_set(&new_congested->refcnt, 0);
 468         new_congested->__bdi = bdi;
 469         new_congested->blkcg_id = blkcg_id;
 470         goto retry;
 471
 472 found:
 473         atomic_inc(&congested->refcnt);
 474         spin_unlock_irqrestore(&cgwb_lock, flags);
 475         kfree(new_congested);
 476         return congested;
 477 }
 478
 479 /**
 480  * wb_congested_put - put a wb_congested
 481  * @congested: wb_congested to put
 482  *
 483  * Put @congested and destroy it if the refcnt reaches zero.
 484  */
 485 void wb_congested_put(struct bdi_writeback_congested *congested)
 486 {
 487         unsigned long flags;
 488
 489         local_irq_save(flags);
 490         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 491                 local_irq_restore(flags);
 492                 return;
 493         }
 494
 495         /* bdi might already have been destroyed leaving @congested unlinked */
 496         if (congested->__bdi) {
 497                 rb_erase(&congested->rb_node,
 498                          &congested->__bdi->cgwb_congested_tree);
 499                 congested->__bdi = NULL;
 500         }
 501
 502         spin_unlock_irqrestore(&cgwb_lock, flags);
 503         kfree(congested);
 504 }
 505
 506 static void cgwb_release_workfn(struct work_struct *work)
 507 {
 508         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 509                                                 release_work);
 510
 511         wb_shutdown(wb);
 512
 513         css_put(wb->memcg_css);
 514         css_put(wb->blkcg_css);
 515
 516         fprop_local_destroy_percpu(&wb->memcg_completions);
 517         percpu_ref_exit(&wb->refcnt);
 518         wb_exit(wb);
 519         kfree_rcu(wb, rcu);
 520 }
 521
 522 static void cgwb_release(struct percpu_ref *refcnt)
 523 {
 524         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 525                                                 refcnt);
 526         queue_work(cgwb_release_wq, &wb->release_work);
 527 }
 528
 529 static void cgwb_kill(struct bdi_writeback *wb)
 530 {
 531         lockdep_assert_held(&cgwb_lock);
 532
 533         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 534         list_del(&wb->memcg_node);
 535         list_del(&wb->blkcg_node);
 536         percpu_ref_kill(&wb->refcnt);
 537 }
 538
 539 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 540 {
 541         spin_lock_irq(&cgwb_lock);
 542         list_del_rcu(&wb->bdi_node);
 543         spin_unlock_irq(&cgwb_lock);
 544 }
 545
 546 static int cgwb_create(struct backing_dev_info *bdi,
 547                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 548 {
 549         struct mem_cgroup *memcg;
 550         struct cgroup_subsys_state *blkcg_css;
 551         struct blkcg *blkcg;
 552         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 553         struct bdi_writeback *wb;
 554         unsigned long flags;
 555         int ret = 0;
 556
 557         memcg = mem_cgroup_from_css(memcg_css);
 558         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 559         blkcg = css_to_blkcg(blkcg_css);
 560         memcg_cgwb_list = &memcg->cgwb_list;
 561         blkcg_cgwb_list = &blkcg->cgwb_list;
 562
 563         /* look up again under lock and discard on blkcg mismatch */
 564         spin_lock_irqsave(&cgwb_lock, flags);
 565         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 566         if (wb && wb->blkcg_css != blkcg_css) {
 567                 cgwb_kill(wb);
 568                 wb = NULL;
 569         }
 570         spin_unlock_irqrestore(&cgwb_lock, flags);
 571         if (wb)
 572                 goto out_put;
 573
 574         /* need to create a new one */
 575         wb = kmalloc(sizeof(*wb), gfp);
 576         if (!wb) {
 577                 ret = -ENOMEM;
 578                 goto out_put;
 579         }
 580
 581         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 582         if (ret)
 583                 goto err_free;
 584
 585         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 586         if (ret)
 587                 goto err_wb_exit;
 588
 589         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 590         if (ret)
 591                 goto err_ref_exit;
 592
 593         wb->memcg_css = memcg_css;
 594         wb->blkcg_css = blkcg_css;
 595         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 596         set_bit(WB_registered, &wb->state);
 597
 598         /*
 599          * The root wb determines the registered state of the whole bdi and
 600          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 601          * whether they're still online.  Don't link @wb if any is dead.
 602          * See wb_memcg_offline() and wb_blkcg_offline().
 603          */
 604         ret = -ENODEV;
 605         spin_lock_irqsave(&cgwb_lock, flags);
 606         if (test_bit(WB_registered, &bdi->wb.state) &&
 607             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 608                 /* we might have raced another instance of this function */
 609                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 610                 if (!ret) {
 611                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 612                         list_add(&wb->memcg_node, memcg_cgwb_list);
 613                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 614                         css_get(memcg_css);
 615                         css_get(blkcg_css);
 616                 }
 617         }
 618         spin_unlock_irqrestore(&cgwb_lock, flags);
 619         if (ret) {
 620                 if (ret == -EEXIST)
 621                         ret = 0;
 622                 goto err_fprop_exit;
 623         }
 624         goto out_put;
 625
 626 err_fprop_exit:
 627         fprop_local_destroy_percpu(&wb->memcg_completions);
 628 err_ref_exit:
 629         percpu_ref_exit(&wb->refcnt);
 630 err_wb_exit:
 631         wb_exit(wb);
 632 err_free:
 633         kfree(wb);
 634 out_put:
 635         css_put(blkcg_css);
 636         return ret;
 637 }
 638
 639 /**
 640  * wb_get_create - get wb for a given memcg, create if necessary
 641  * @bdi: target bdi
 642  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 643  * @gfp: allocation mask to use
 644  *
 645  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 646  * create one.  The returned wb has its refcount incremented.
 647  *
 648  * This function uses css_get() on @memcg_css and thus expects its refcnt
 649  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 650  * @memcg_css isn't enough.  try_get it before calling this function.
 651  *
 652  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 653  * memcg on the default hierarchy, memcg association is guaranteed to be
 654  * more specific (equal or descendant to the associated blkcg) and thus can
 655  * identify both the memcg and blkcg associations.
 656  *
 657  * Because the blkcg associated with a memcg may change as blkcg is enabled
 658  * and disabled closer to root in the hierarchy, each wb keeps track of
 659  * both the memcg and blkcg associated with it and verifies the blkcg on
 660  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 661  * created.
 662  */
 663 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 664                                     struct cgroup_subsys_state *memcg_css,
 665                                     gfp_t gfp)
 666 {
 667         struct bdi_writeback *wb;
 668
 669         might_sleep_if(gfpflags_allow_blocking(gfp));
 670
 671         if (!memcg_css->parent)
 672                 return &bdi->wb;
 673
 674         do {
 675                 rcu_read_lock();
 676                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 677                 if (wb) {
 678                         struct cgroup_subsys_state *blkcg_css;
 679
 680                         /* see whether the blkcg association has changed */
 681                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 682                                                      &io_cgrp_subsys);
 683                         if (unlikely(wb->blkcg_css != blkcg_css ||
 684                                      !wb_tryget(wb)))
 685                                 wb = NULL;
 686                         css_put(blkcg_css);
 687                 }
 688                 rcu_read_unlock();
 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
 701         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 702         if (!ret) {
 703                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 704                 bdi->wb.blkcg_css = blkcg_root_css;
 705         }
 706         return ret;
 707 }
 708
 709 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 710 {
 711         struct radix_tree_iter iter;
 712         void **slot;
 713         struct bdi_writeback *wb;
 714
 715         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 716
 717         spin_lock_irq(&cgwb_lock);
 718         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 719                 cgwb_kill(*slot);
 720
 721         while (!list_empty(&bdi->wb_list)) {
 722                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 723                                       bdi_node);
 724                 spin_unlock_irq(&cgwb_lock);
 725                 wb_shutdown(wb);
 726                 spin_lock_irq(&cgwb_lock);
 727         }
 728         spin_unlock_irq(&cgwb_lock);
 729 }
 730
 731 /**
 732  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 733  * @memcg: memcg being offlined
 734  *
 735  * Also prevents creation of any new wb's associated with @memcg.
 736  */
 737 void wb_memcg_offline(struct mem_cgroup *memcg)
 738 {
 739         struct list_head *memcg_cgwb_list = &memcg->cgwb_list;
 740         struct bdi_writeback *wb, *next;
 741
 742         spin_lock_irq(&cgwb_lock);
 743         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 744                 cgwb_kill(wb);
 745         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 746         spin_unlock_irq(&cgwb_lock);
 747 }
 748
 749 /**
 750  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 751  * @blkcg: blkcg being offlined
 752  *
 753  * Also prevents creation of any new wb's associated with @blkcg.
 754  */
 755 void wb_blkcg_offline(struct blkcg *blkcg)
 756 {
 757         struct bdi_writeback *wb, *next;
 758
 759         spin_lock_irq(&cgwb_lock);
 760         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 761                 cgwb_kill(wb);
 762         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 763         spin_unlock_irq(&cgwb_lock);
 764 }
 765
 766 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 767 {
 768         struct rb_node *rbn;
 769
 770         spin_lock_irq(&cgwb_lock);
 771         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 772                 struct bdi_writeback_congested *congested =
 773                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 774
 775                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 776                 congested->__bdi = NULL;        /* mark @congested unlinked */
 777         }
 778         spin_unlock_irq(&cgwb_lock);
 779 }
 780
 781 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 782 {
 783         spin_lock_irq(&cgwb_lock);
 784         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 785         spin_unlock_irq(&cgwb_lock);
 786 }
 787
 788 static int __init cgwb_init(void)
 789 {
 790         /*
 791          * There can be many concurrent release work items overwhelming
 792          * system_wq.  Put them in a separate wq and limit concurrency.
 793          * There's no point in executing many of these in parallel.
 794          */
 795         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 796         if (!cgwb_release_wq)
 797                 return -ENOMEM;
 798
 799         return 0;
 800 }
 801 subsys_initcall(cgwb_init);
 802
 803 #else   /* CONFIG_CGROUP_WRITEBACK */
 804
 805 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 806 {
 807         int err;
 808
 809         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 810         if (!bdi->wb_congested)
 811                 return -ENOMEM;
 812
 813         atomic_set(&bdi->wb_congested->refcnt, 1);
 814
 815         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 816         if (err) {
 817                 wb_congested_put(bdi->wb_congested);
 818                 return err;
 819         }
 820         return 0;
 821 }
 822
 823 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 824
 825 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 826 {
 827         wb_congested_put(bdi->wb_congested);
 828 }
 829
 830 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 831 {
 832         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 833 }
 834
 835 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 836 {
 837         list_del_rcu(&wb->bdi_node);
 838 }
 839
 840 #endif  /* CONFIG_CGROUP_WRITEBACK */
 841
 842 static int bdi_init(struct backing_dev_info *bdi)
 843 {
 844         int ret;
 845
 846         bdi->dev = NULL;
 847
 848         kref_init(&bdi->refcnt);
 849         bdi->min_ratio = 0;
 850         bdi->max_ratio = 100;
 851         bdi->max_prop_frac = FPROP_FRAC_BASE;
 852         INIT_LIST_HEAD(&bdi->bdi_list);
 853         INIT_LIST_HEAD(&bdi->wb_list);
 854         init_waitqueue_head(&bdi->wb_waitq);
 855
 856         ret = cgwb_bdi_init(bdi);
 857
 858         return ret;
 859 }
 860
 861 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 862 {
 863         struct backing_dev_info *bdi;
 864
 865         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 866                            gfp_mask | __GFP_ZERO, node_id);
 867         if (!bdi)
 868                 return NULL;
 869
 870         if (bdi_init(bdi)) {
 871                 kfree(bdi);
 872                 return NULL;
 873         }
 874         return bdi;
 875 }
 876 EXPORT_SYMBOL(bdi_alloc_node);
 877
 878 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 879 {
 880         struct device *dev;
 881
 882         if (bdi->dev)   /* The driver needs to use separate queues per device */
 883                 return 0;
 884
 885         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 886         if (IS_ERR(dev))
 887                 return PTR_ERR(dev);
 888
 889         cgwb_bdi_register(bdi);
 890         bdi->dev = dev;
 891
 892         bdi_debug_register(bdi, dev_name(dev));
 893         set_bit(WB_registered, &bdi->wb.state);
 894
 895         spin_lock_bh(&bdi_lock);
 896         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 897         spin_unlock_bh(&bdi_lock);
 898
 899         trace_writeback_bdi_register(bdi);
 900         return 0;
 901 }
 902 EXPORT_SYMBOL(bdi_register_va);
 903
 904 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 905 {
 906         va_list args;
 907         int ret;
 908
 909         va_start(args, fmt);
 910         ret = bdi_register_va(bdi, fmt, args);
 911         va_end(args);
 912         return ret;
 913 }
 914 EXPORT_SYMBOL(bdi_register);
 915
 916 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 917 {
 918         int rc;
 919
 920         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 921         if (rc)
 922                 return rc;
 923         /* Leaking owner reference... */
 924         WARN_ON(bdi->owner);
 925         bdi->owner = owner;
 926         get_device(owner);
 927         return 0;
 928 }
 929 EXPORT_SYMBOL(bdi_register_owner);
 930
 931 /*
 932  * Remove bdi from bdi_list, and ensure that it is no longer visible
 933  */
 934 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 935 {
 936         spin_lock_bh(&bdi_lock);
 937         list_del_rcu(&bdi->bdi_list);
 938         spin_unlock_bh(&bdi_lock);
 939
 940         synchronize_rcu_expedited();
 941 }
 942
 943 void bdi_unregister(struct backing_dev_info *bdi)
 944 {
 945         /* make sure nobody finds us on the bdi_list anymore */
 946         bdi_remove_from_list(bdi);
 947         wb_shutdown(&bdi->wb);
 948         cgwb_bdi_unregister(bdi);
 949
 950         if (bdi->dev) {
 951                 bdi_debug_unregister(bdi);
 952                 device_unregister(bdi->dev);
 953                 bdi->dev = NULL;
 954         }
 955
 956         if (bdi->owner) {
 957                 put_device(bdi->owner);
 958                 bdi->owner = NULL;
 959         }
 960 }
 961
 962 static void release_bdi(struct kref *ref)
 963 {
 964         struct backing_dev_info *bdi =
 965                         container_of(ref, struct backing_dev_info, refcnt);
 966
 967         if (test_bit(WB_registered, &bdi->wb.state))
 968                 bdi_unregister(bdi);
 969         WARN_ON_ONCE(bdi->dev);
 970         wb_exit(&bdi->wb);
 971         cgwb_bdi_exit(bdi);
 972         kfree(bdi);
 973 }
 974
 975 void bdi_put(struct backing_dev_info *bdi)
 976 {
 977         kref_put(&bdi->refcnt, release_bdi);
 978 }
 979 EXPORT_SYMBOL(bdi_put);
 980
 981 static wait_queue_head_t congestion_wqh[2] = {
 982                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 983                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 984         };
 985 static atomic_t nr_wb_congested[2];
 986
 987 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 988 {
 989         wait_queue_head_t *wqh = &congestion_wqh[sync];
 990         enum wb_congested_state bit;
 991
 992         bit = sync ? WB_sync_congested : WB_async_congested;
 993         if (test_and_clear_bit(bit, &congested->state))
 994                 atomic_dec(&nr_wb_congested[sync]);
 995         smp_mb__after_atomic();
 996         if (waitqueue_active(wqh))
 997                 wake_up(wqh);
 998 }
 999 EXPORT_SYMBOL(clear_wb_congested);
1000
1001 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
1002 {
1003         enum wb_congested_state bit;
1004
1005         bit = sync ? WB_sync_congested : WB_async_congested;
1006         if (!test_and_set_bit(bit, &congested->state))
1007                 atomic_inc(&nr_wb_congested[sync]);
1008 }
1009 EXPORT_SYMBOL(set_wb_congested);
1010
1011 /**
1012  * congestion_wait - wait for a backing_dev to become uncongested
1013  * @sync: SYNC or ASYNC IO
1014  * @timeout: timeout in jiffies
1015  *
1016  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1017  * write congestion.  If no backing_devs are congested then just wait for the
1018  * next write to be completed.
1019  */
1020 long congestion_wait(int sync, long timeout)
1021 {
1022         long ret;
1023         unsigned long start = jiffies;
1024         DEFINE_WAIT(wait);
1025         wait_queue_head_t *wqh = &congestion_wqh[sync];
1026
1027         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1028         ret = io_schedule_timeout(timeout);
1029         finish_wait(wqh, &wait);
1030
1031         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1032                                         jiffies_to_usecs(jiffies - start));
1033
1034         return ret;
1035 }
1036 EXPORT_SYMBOL(congestion_wait);
1037
1038 /**
1039  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1040  * @sync: SYNC or ASYNC IO
1041  * @timeout: timeout in jiffies
1042  *
1043  * In the event of a congested backing_dev (any backing_dev) this waits
1044  * for up to @timeout jiffies for either a BDI to exit congestion of the
1045  * given @sync queue or a write to complete.
1046  *
1047  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1048  * it is the number of jiffies that were still remaining when the function
1049  * returned. return_value == timeout implies the function did not sleep.
1050  */
1051 long wait_iff_congested(int sync, long timeout)
1052 {
1053         long ret;
1054         unsigned long start = jiffies;
1055         DEFINE_WAIT(wait);
1056         wait_queue_head_t *wqh = &congestion_wqh[sync];
1057
1058         /*
1059          * If there is no congestion, yield if necessary instead
1060          * of sleeping on the congestion queue
1061          */
1062         if (atomic_read(&nr_wb_congested[sync]) == 0) {
1063                 cond_resched();
1064
1065                 /* In case we scheduled, work out time remaining */
1066                 ret = timeout - (jiffies - start);
1067                 if (ret < 0)
1068                         ret = 0;
1069
1070                 goto out;
1071         }
1072
1073         /* Sleep until uncongested or a write happens */
1074         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1075         ret = io_schedule_timeout(timeout);
1076         finish_wait(wqh, &wait);
1077
1078 out:
1079         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1080                                         jiffies_to_usecs(jiffies - start));
1081
1082         return ret;
1083 }
1084 EXPORT_SYMBOL(wait_iff_congested);