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
 416 /**
 417  * wb_congested_get_create - get or create a wb_congested
 418  * @bdi: associated bdi
 419  * @blkcg_id: ID of the associated blkcg
 420  * @gfp: allocation mask
 421  *
 422  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 423  * The returned wb_congested has its reference count incremented.  Returns
 424  * NULL on failure.
 425  */
 426 struct bdi_writeback_congested *
 427 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 428 {
 429         struct bdi_writeback_congested *new_congested = NULL, *congested;
 430         struct rb_node **node, *parent;
 431         unsigned long flags;
 432 retry:
 433         spin_lock_irqsave(&cgwb_lock, flags);
 434
 435         node = &bdi->cgwb_congested_tree.rb_node;
 436         parent = NULL;
 437
 438         while (*node != NULL) {
 439                 parent = *node;
 440                 congested = rb_entry(parent, struct bdi_writeback_congested,
 441                                      rb_node);
 442                 if (congested->blkcg_id < blkcg_id)
 443                         node = &parent->rb_left;
 444                 else if (congested->blkcg_id > blkcg_id)
 445                         node = &parent->rb_right;
 446                 else
 447                         goto found;
 448         }
 449
 450         if (new_congested) {
 451                 /* !found and storage for new one already allocated, insert */
 452                 congested = new_congested;
 453                 new_congested = NULL;
 454                 rb_link_node(&congested->rb_node, parent, node);
 455                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 456                 goto found;
 457         }
 458
 459         spin_unlock_irqrestore(&cgwb_lock, flags);
 460
 461         /* allocate storage for new one and retry */
 462         new_congested = kzalloc(sizeof(*new_congested), gfp);
 463         if (!new_congested)
 464                 return NULL;
 465
 466         atomic_set(&new_congested->refcnt, 0);
 467         new_congested->__bdi = bdi;
 468         new_congested->blkcg_id = blkcg_id;
 469         goto retry;
 470
 471 found:
 472         atomic_inc(&congested->refcnt);
 473         spin_unlock_irqrestore(&cgwb_lock, flags);
 474         kfree(new_congested);
 475         return congested;
 476 }
 477
 478 /**
 479  * wb_congested_put - put a wb_congested
 480  * @congested: wb_congested to put
 481  *
 482  * Put @congested and destroy it if the refcnt reaches zero.
 483  */
 484 void wb_congested_put(struct bdi_writeback_congested *congested)
 485 {
 486         unsigned long flags;
 487
 488         local_irq_save(flags);
 489         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 490                 local_irq_restore(flags);
 491                 return;
 492         }
 493
 494         /* bdi might already have been destroyed leaving @congested unlinked */
 495         if (congested->__bdi) {
 496                 rb_erase(&congested->rb_node,
 497                          &congested->__bdi->cgwb_congested_tree);
 498                 congested->__bdi = NULL;
 499         }
 500
 501         spin_unlock_irqrestore(&cgwb_lock, flags);
 502         kfree(congested);
 503 }
 504
 505 static void cgwb_release_workfn(struct work_struct *work)
 506 {
 507         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 508                                                 release_work);
 509
 510         wb_shutdown(wb);
 511
 512         css_put(wb->memcg_css);
 513         css_put(wb->blkcg_css);
 514
 515         fprop_local_destroy_percpu(&wb->memcg_completions);
 516         percpu_ref_exit(&wb->refcnt);
 517         wb_exit(wb);
 518         kfree_rcu(wb, rcu);
 519 }
 520
 521 static void cgwb_release(struct percpu_ref *refcnt)
 522 {
 523         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 524                                                 refcnt);
 525         schedule_work(&wb->release_work);
 526 }
 527
 528 static void cgwb_kill(struct bdi_writeback *wb)
 529 {
 530         lockdep_assert_held(&cgwb_lock);
 531
 532         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 533         list_del(&wb->memcg_node);
 534         list_del(&wb->blkcg_node);
 535         percpu_ref_kill(&wb->refcnt);
 536 }
 537
 538 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 539 {
 540         spin_lock_irq(&cgwb_lock);
 541         list_del_rcu(&wb->bdi_node);
 542         spin_unlock_irq(&cgwb_lock);
 543 }
 544
 545 static int cgwb_create(struct backing_dev_info *bdi,
 546                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 547 {
 548         struct mem_cgroup *memcg;
 549         struct cgroup_subsys_state *blkcg_css;
 550         struct blkcg *blkcg;
 551         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 552         struct bdi_writeback *wb;
 553         unsigned long flags;
 554         int ret = 0;
 555
 556         memcg = mem_cgroup_from_css(memcg_css);
 557         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 558         blkcg = css_to_blkcg(blkcg_css);
 559         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 560         blkcg_cgwb_list = &blkcg->cgwb_list;
 561
 562         /* look up again under lock and discard on blkcg mismatch */
 563         spin_lock_irqsave(&cgwb_lock, flags);
 564         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 565         if (wb && wb->blkcg_css != blkcg_css) {
 566                 cgwb_kill(wb);
 567                 wb = NULL;
 568         }
 569         spin_unlock_irqrestore(&cgwb_lock, flags);
 570         if (wb)
 571                 goto out_put;
 572
 573         /* need to create a new one */
 574         wb = kmalloc(sizeof(*wb), gfp);
 575         if (!wb) {
 576                 ret = -ENOMEM;
 577                 goto out_put;
 578         }
 579
 580         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 581         if (ret)
 582                 goto err_free;
 583
 584         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 585         if (ret)
 586                 goto err_wb_exit;
 587
 588         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 589         if (ret)
 590                 goto err_ref_exit;
 591
 592         wb->memcg_css = memcg_css;
 593         wb->blkcg_css = blkcg_css;
 594         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 595         set_bit(WB_registered, &wb->state);
 596
 597         /*
 598          * The root wb determines the registered state of the whole bdi and
 599          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 600          * whether they're still online.  Don't link @wb if any is dead.
 601          * See wb_memcg_offline() and wb_blkcg_offline().
 602          */
 603         ret = -ENODEV;
 604         spin_lock_irqsave(&cgwb_lock, flags);
 605         if (test_bit(WB_registered, &bdi->wb.state) &&
 606             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 607                 /* we might have raced another instance of this function */
 608                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 609                 if (!ret) {
 610                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 611                         list_add(&wb->memcg_node, memcg_cgwb_list);
 612                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 613                         css_get(memcg_css);
 614                         css_get(blkcg_css);
 615                 }
 616         }
 617         spin_unlock_irqrestore(&cgwb_lock, flags);
 618         if (ret) {
 619                 if (ret == -EEXIST)
 620                         ret = 0;
 621                 goto err_fprop_exit;
 622         }
 623         goto out_put;
 624
 625 err_fprop_exit:
 626         fprop_local_destroy_percpu(&wb->memcg_completions);
 627 err_ref_exit:
 628         percpu_ref_exit(&wb->refcnt);
 629 err_wb_exit:
 630         wb_exit(wb);
 631 err_free:
 632         kfree(wb);
 633 out_put:
 634         css_put(blkcg_css);
 635         return ret;
 636 }
 637
 638 /**
 639  * wb_get_create - get wb for a given memcg, create if necessary
 640  * @bdi: target bdi
 641  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 642  * @gfp: allocation mask to use
 643  *
 644  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 645  * create one.  The returned wb has its refcount incremented.
 646  *
 647  * This function uses css_get() on @memcg_css and thus expects its refcnt
 648  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 649  * @memcg_css isn't enough.  try_get it before calling this function.
 650  *
 651  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 652  * memcg on the default hierarchy, memcg association is guaranteed to be
 653  * more specific (equal or descendant to the associated blkcg) and thus can
 654  * identify both the memcg and blkcg associations.
 655  *
 656  * Because the blkcg associated with a memcg may change as blkcg is enabled
 657  * and disabled closer to root in the hierarchy, each wb keeps track of
 658  * both the memcg and blkcg associated with it and verifies the blkcg on
 659  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 660  * created.
 661  */
 662 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 663                                     struct cgroup_subsys_state *memcg_css,
 664                                     gfp_t gfp)
 665 {
 666         struct bdi_writeback *wb;
 667
 668         might_sleep_if(gfpflags_allow_blocking(gfp));
 669
 670         if (!memcg_css->parent)
 671                 return &bdi->wb;
 672
 673         do {
 674                 rcu_read_lock();
 675                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 676                 if (wb) {
 677                         struct cgroup_subsys_state *blkcg_css;
 678
 679                         /* see whether the blkcg association has changed */
 680                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 681                                                      &io_cgrp_subsys);
 682                         if (unlikely(wb->blkcg_css != blkcg_css ||
 683                                      !wb_tryget(wb)))
 684                                 wb = NULL;
 685                         css_put(blkcg_css);
 686                 }
 687                 rcu_read_unlock();
 688         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 689
 690         return wb;
 691 }
 692
 693 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 694 {
 695         int ret;
 696
 697         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 698         bdi->cgwb_congested_tree = RB_ROOT;
 699
 700         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 701         if (!ret) {
 702                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 703                 bdi->wb.blkcg_css = blkcg_root_css;
 704         }
 705         return ret;
 706 }
 707
 708 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 709 {
 710         struct radix_tree_iter iter;
 711         void **slot;
 712         struct bdi_writeback *wb;
 713
 714         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 715
 716         spin_lock_irq(&cgwb_lock);
 717         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 718                 cgwb_kill(*slot);
 719
 720         while (!list_empty(&bdi->wb_list)) {
 721                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 722                                       bdi_node);
 723                 spin_unlock_irq(&cgwb_lock);
 724                 wb_shutdown(wb);
 725                 spin_lock_irq(&cgwb_lock);
 726         }
 727         spin_unlock_irq(&cgwb_lock);
 728 }
 729
 730 /**
 731  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 732  * @memcg: memcg being offlined
 733  *
 734  * Also prevents creation of any new wb's associated with @memcg.
 735  */
 736 void wb_memcg_offline(struct mem_cgroup *memcg)
 737 {
 738         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 739         struct bdi_writeback *wb, *next;
 740
 741         spin_lock_irq(&cgwb_lock);
 742         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 743                 cgwb_kill(wb);
 744         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 745         spin_unlock_irq(&cgwb_lock);
 746 }
 747
 748 /**
 749  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 750  * @blkcg: blkcg being offlined
 751  *
 752  * Also prevents creation of any new wb's associated with @blkcg.
 753  */
 754 void wb_blkcg_offline(struct blkcg *blkcg)
 755 {
 756         struct bdi_writeback *wb, *next;
 757
 758         spin_lock_irq(&cgwb_lock);
 759         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 760                 cgwb_kill(wb);
 761         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 762         spin_unlock_irq(&cgwb_lock);
 763 }
 764
 765 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 766 {
 767         struct rb_node *rbn;
 768
 769         spin_lock_irq(&cgwb_lock);
 770         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 771                 struct bdi_writeback_congested *congested =
 772                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 773
 774                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 775                 congested->__bdi = NULL;        /* mark @congested unlinked */
 776         }
 777         spin_unlock_irq(&cgwb_lock);
 778 }
 779
 780 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 781 {
 782         spin_lock_irq(&cgwb_lock);
 783         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 784         spin_unlock_irq(&cgwb_lock);
 785 }
 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         atomic_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);