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
 104 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 105 {
 106         return single_open(file, bdi_debug_stats_show, inode->i_private);
 107 }
 108
 109 static const struct file_operations bdi_debug_stats_fops = {
 110         .open           = bdi_debug_stats_open,
 111         .read           = seq_read,
 112         .llseek         = seq_lseek,
 113         .release        = single_release,
 114 };
 115
 116 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 117 {
 118         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 119         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 120                                                bdi, &bdi_debug_stats_fops);
 121 }
 122
 123 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 124 {
 125         debugfs_remove(bdi->debug_stats);
 126         debugfs_remove(bdi->debug_dir);
 127 }
 128 #else
 129 static inline void bdi_debug_init(void)
 130 {
 131 }
 132 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 133                                       const char *name)
 134 {
 135 }
 136 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 137 {
 138 }
 139 #endif
 140
 141 static ssize_t read_ahead_kb_store(struct device *dev,
 142                                   struct device_attribute *attr,
 143                                   const char *buf, size_t count)
 144 {
 145         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 146         unsigned long read_ahead_kb;
 147         ssize_t ret;
 148
 149         ret = kstrtoul(buf, 10, &read_ahead_kb);
 150         if (ret < 0)
 151                 return ret;
 152
 153         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 154
 155         return count;
 156 }
 157
 158 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 159
 160 #define BDI_SHOW(name, expr)                                            \
 161 static ssize_t name##_show(struct device *dev,                          \
 162                            struct device_attribute *attr, char *page)   \
 163 {                                                                       \
 164         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 165                                                                         \
 166         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 167 }                                                                       \
 168 static DEVICE_ATTR_RW(name);
 169
 170 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 171
 172 static ssize_t min_ratio_store(struct device *dev,
 173                 struct device_attribute *attr, const char *buf, size_t count)
 174 {
 175         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 176         unsigned int ratio;
 177         ssize_t ret;
 178
 179         ret = kstrtouint(buf, 10, &ratio);
 180         if (ret < 0)
 181                 return ret;
 182
 183         ret = bdi_set_min_ratio(bdi, ratio);
 184         if (!ret)
 185                 ret = count;
 186
 187         return ret;
 188 }
 189 BDI_SHOW(min_ratio, bdi->min_ratio)
 190
 191 static ssize_t max_ratio_store(struct device *dev,
 192                 struct device_attribute *attr, const char *buf, size_t count)
 193 {
 194         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 195         unsigned int ratio;
 196         ssize_t ret;
 197
 198         ret = kstrtouint(buf, 10, &ratio);
 199         if (ret < 0)
 200                 return ret;
 201
 202         ret = bdi_set_max_ratio(bdi, ratio);
 203         if (!ret)
 204                 ret = count;
 205
 206         return ret;
 207 }
 208 BDI_SHOW(max_ratio, bdi->max_ratio)
 209
 210 static ssize_t stable_pages_required_show(struct device *dev,
 211                                           struct device_attribute *attr,
 212                                           char *page)
 213 {
 214         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 215
 216         return snprintf(page, PAGE_SIZE-1, "%d\n",
 217                         bdi_cap_stable_pages_required(bdi) ? 1 : 0);
 218 }
 219 static DEVICE_ATTR_RO(stable_pages_required);
 220
 221 static struct attribute *bdi_dev_attrs[] = {
 222         &dev_attr_read_ahead_kb.attr,
 223         &dev_attr_min_ratio.attr,
 224         &dev_attr_max_ratio.attr,
 225         &dev_attr_stable_pages_required.attr,
 226         NULL,
 227 };
 228 ATTRIBUTE_GROUPS(bdi_dev);
 229
 230 static __init int bdi_class_init(void)
 231 {
 232         bdi_class = class_create(THIS_MODULE, "bdi");
 233         if (IS_ERR(bdi_class))
 234                 return PTR_ERR(bdi_class);
 235
 236         bdi_class->dev_groups = bdi_dev_groups;
 237         bdi_debug_init();
 238
 239         return 0;
 240 }
 241 postcore_initcall(bdi_class_init);
 242
 243 static int bdi_init(struct backing_dev_info *bdi);
 244
 245 static int __init default_bdi_init(void)
 246 {
 247         int err;
 248
 249         bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_FREEZABLE |
 250                                               WQ_UNBOUND | WQ_SYSFS, 0);
 251         if (!bdi_wq)
 252                 return -ENOMEM;
 253
 254         err = bdi_init(&noop_backing_dev_info);
 255
 256         return err;
 257 }
 258 subsys_initcall(default_bdi_init);
 259
 260 /*
 261  * This function is used when the first inode for this wb is marked dirty. It
 262  * wakes-up the corresponding bdi thread which should then take care of the
 263  * periodic background write-out of dirty inodes. Since the write-out would
 264  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 265  * set up a timer which wakes the bdi thread up later.
 266  *
 267  * Note, we wouldn't bother setting up the timer, but this function is on the
 268  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 269  * by delaying the wake-up.
 270  *
 271  * We have to be careful not to postpone flush work if it is scheduled for
 272  * earlier. Thus we use queue_delayed_work().
 273  */
 274 void wb_wakeup_delayed(struct bdi_writeback *wb)
 275 {
 276         unsigned long timeout;
 277
 278         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 279         spin_lock_bh(&wb->work_lock);
 280         if (test_bit(WB_registered, &wb->state))
 281                 queue_delayed_work(bdi_wq, &wb->dwork, timeout);
 282         spin_unlock_bh(&wb->work_lock);
 283 }
 284
 285 /*
 286  * Initial write bandwidth: 100 MB/s
 287  */
 288 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 289
 290 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi,
 291                    int blkcg_id, gfp_t gfp)
 292 {
 293         int i, err;
 294
 295         memset(wb, 0, sizeof(*wb));
 296
 297         if (wb != &bdi->wb)
 298                 bdi_get(bdi);
 299         wb->bdi = bdi;
 300         wb->last_old_flush = jiffies;
 301         INIT_LIST_HEAD(&wb->b_dirty);
 302         INIT_LIST_HEAD(&wb->b_io);
 303         INIT_LIST_HEAD(&wb->b_more_io);
 304         INIT_LIST_HEAD(&wb->b_dirty_time);
 305         spin_lock_init(&wb->list_lock);
 306
 307         wb->bw_time_stamp = jiffies;
 308         wb->balanced_dirty_ratelimit = INIT_BW;
 309         wb->dirty_ratelimit = INIT_BW;
 310         wb->write_bandwidth = INIT_BW;
 311         wb->avg_write_bandwidth = INIT_BW;
 312
 313         spin_lock_init(&wb->work_lock);
 314         INIT_LIST_HEAD(&wb->work_list);
 315         INIT_DELAYED_WORK(&wb->dwork, wb_workfn);
 316         wb->dirty_sleep = jiffies;
 317
 318         wb->congested = wb_congested_get_create(bdi, blkcg_id, gfp);
 319         if (!wb->congested) {
 320                 err = -ENOMEM;
 321                 goto out_put_bdi;
 322         }
 323
 324         err = fprop_local_init_percpu(&wb->completions, gfp);
 325         if (err)
 326                 goto out_put_cong;
 327
 328         for (i = 0; i < NR_WB_STAT_ITEMS; i++) {
 329                 err = percpu_counter_init(&wb->stat[i], 0, gfp);
 330                 if (err)
 331                         goto out_destroy_stat;
 332         }
 333
 334         return 0;
 335
 336 out_destroy_stat:
 337         while (i--)
 338                 percpu_counter_destroy(&wb->stat[i]);
 339         fprop_local_destroy_percpu(&wb->completions);
 340 out_put_cong:
 341         wb_congested_put(wb->congested);
 342 out_put_bdi:
 343         if (wb != &bdi->wb)
 344                 bdi_put(bdi);
 345         return err;
 346 }
 347
 348 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb);
 349
 350 /*
 351  * Remove bdi from the global list and shutdown any threads we have running
 352  */
 353 static void wb_shutdown(struct bdi_writeback *wb)
 354 {
 355         /* Make sure nobody queues further work */
 356         spin_lock_bh(&wb->work_lock);
 357         if (!test_and_clear_bit(WB_registered, &wb->state)) {
 358                 spin_unlock_bh(&wb->work_lock);
 359                 /*
 360                  * Wait for wb shutdown to finish if someone else is just
 361                  * running wb_shutdown(). Otherwise we could proceed to wb /
 362                  * bdi destruction before wb_shutdown() is finished.
 363                  */
 364                 wait_on_bit(&wb->state, WB_shutting_down, TASK_UNINTERRUPTIBLE);
 365                 return;
 366         }
 367         set_bit(WB_shutting_down, &wb->state);
 368         spin_unlock_bh(&wb->work_lock);
 369
 370         cgwb_remove_from_bdi_list(wb);
 371         /*
 372          * Drain work list and shutdown the delayed_work.  !WB_registered
 373          * tells wb_workfn() that @wb is dying and its work_list needs to
 374          * be drained no matter what.
 375          */
 376         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 377         flush_delayed_work(&wb->dwork);
 378         WARN_ON(!list_empty(&wb->work_list));
 379         /*
 380          * Make sure bit gets cleared after shutdown is finished. Matches with
 381          * the barrier provided by test_and_clear_bit() above.
 382          */
 383         smp_wmb();
 384         clear_bit(WB_shutting_down, &wb->state);
 385 }
 386
 387 static void wb_exit(struct bdi_writeback *wb)
 388 {
 389         int i;
 390
 391         WARN_ON(delayed_work_pending(&wb->dwork));
 392
 393         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 394                 percpu_counter_destroy(&wb->stat[i]);
 395
 396         fprop_local_destroy_percpu(&wb->completions);
 397         wb_congested_put(wb->congested);
 398         if (wb != &wb->bdi->wb)
 399                 bdi_put(wb->bdi);
 400 }
 401
 402 #ifdef CONFIG_CGROUP_WRITEBACK
 403
 404 #include <linux/memcontrol.h>
 405
 406 /*
 407  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 408  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 409  * protected.
 410  */
 411 static DEFINE_SPINLOCK(cgwb_lock);
 412
 413 /**
 414  * wb_congested_get_create - get or create a wb_congested
 415  * @bdi: associated bdi
 416  * @blkcg_id: ID of the associated blkcg
 417  * @gfp: allocation mask
 418  *
 419  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 420  * The returned wb_congested has its reference count incremented.  Returns
 421  * NULL on failure.
 422  */
 423 struct bdi_writeback_congested *
 424 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 425 {
 426         struct bdi_writeback_congested *new_congested = NULL, *congested;
 427         struct rb_node **node, *parent;
 428         unsigned long flags;
 429 retry:
 430         spin_lock_irqsave(&cgwb_lock, flags);
 431
 432         node = &bdi->cgwb_congested_tree.rb_node;
 433         parent = NULL;
 434
 435         while (*node != NULL) {
 436                 parent = *node;
 437                 congested = rb_entry(parent, struct bdi_writeback_congested,
 438                                      rb_node);
 439                 if (congested->blkcg_id < blkcg_id)
 440                         node = &parent->rb_left;
 441                 else if (congested->blkcg_id > blkcg_id)
 442                         node = &parent->rb_right;
 443                 else
 444                         goto found;
 445         }
 446
 447         if (new_congested) {
 448                 /* !found and storage for new one already allocated, insert */
 449                 congested = new_congested;
 450                 new_congested = NULL;
 451                 rb_link_node(&congested->rb_node, parent, node);
 452                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 453                 goto found;
 454         }
 455
 456         spin_unlock_irqrestore(&cgwb_lock, flags);
 457
 458         /* allocate storage for new one and retry */
 459         new_congested = kzalloc(sizeof(*new_congested), gfp);
 460         if (!new_congested)
 461                 return NULL;
 462
 463         atomic_set(&new_congested->refcnt, 0);
 464         new_congested->__bdi = bdi;
 465         new_congested->blkcg_id = blkcg_id;
 466         goto retry;
 467
 468 found:
 469         atomic_inc(&congested->refcnt);
 470         spin_unlock_irqrestore(&cgwb_lock, flags);
 471         kfree(new_congested);
 472         return congested;
 473 }
 474
 475 /**
 476  * wb_congested_put - put a wb_congested
 477  * @congested: wb_congested to put
 478  *
 479  * Put @congested and destroy it if the refcnt reaches zero.
 480  */
 481 void wb_congested_put(struct bdi_writeback_congested *congested)
 482 {
 483         unsigned long flags;
 484
 485         local_irq_save(flags);
 486         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 487                 local_irq_restore(flags);
 488                 return;
 489         }
 490
 491         /* bdi might already have been destroyed leaving @congested unlinked */
 492         if (congested->__bdi) {
 493                 rb_erase(&congested->rb_node,
 494                          &congested->__bdi->cgwb_congested_tree);
 495                 congested->__bdi = NULL;
 496         }
 497
 498         spin_unlock_irqrestore(&cgwb_lock, flags);
 499         kfree(congested);
 500 }
 501
 502 static void cgwb_release_workfn(struct work_struct *work)
 503 {
 504         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 505                                                 release_work);
 506
 507         wb_shutdown(wb);
 508
 509         css_put(wb->memcg_css);
 510         css_put(wb->blkcg_css);
 511
 512         fprop_local_destroy_percpu(&wb->memcg_completions);
 513         percpu_ref_exit(&wb->refcnt);
 514         wb_exit(wb);
 515         kfree_rcu(wb, rcu);
 516 }
 517
 518 static void cgwb_release(struct percpu_ref *refcnt)
 519 {
 520         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 521                                                 refcnt);
 522         schedule_work(&wb->release_work);
 523 }
 524
 525 static void cgwb_kill(struct bdi_writeback *wb)
 526 {
 527         lockdep_assert_held(&cgwb_lock);
 528
 529         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 530         list_del(&wb->memcg_node);
 531         list_del(&wb->blkcg_node);
 532         percpu_ref_kill(&wb->refcnt);
 533 }
 534
 535 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 536 {
 537         spin_lock_irq(&cgwb_lock);
 538         list_del_rcu(&wb->bdi_node);
 539         spin_unlock_irq(&cgwb_lock);
 540 }
 541
 542 static int cgwb_create(struct backing_dev_info *bdi,
 543                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 544 {
 545         struct mem_cgroup *memcg;
 546         struct cgroup_subsys_state *blkcg_css;
 547         struct blkcg *blkcg;
 548         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 549         struct bdi_writeback *wb;
 550         unsigned long flags;
 551         int ret = 0;
 552
 553         memcg = mem_cgroup_from_css(memcg_css);
 554         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 555         blkcg = css_to_blkcg(blkcg_css);
 556         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 557         blkcg_cgwb_list = &blkcg->cgwb_list;
 558
 559         /* look up again under lock and discard on blkcg mismatch */
 560         spin_lock_irqsave(&cgwb_lock, flags);
 561         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 562         if (wb && wb->blkcg_css != blkcg_css) {
 563                 cgwb_kill(wb);
 564                 wb = NULL;
 565         }
 566         spin_unlock_irqrestore(&cgwb_lock, flags);
 567         if (wb)
 568                 goto out_put;
 569
 570         /* need to create a new one */
 571         wb = kmalloc(sizeof(*wb), gfp);
 572         if (!wb)
 573                 return -ENOMEM;
 574
 575         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 576         if (ret)
 577                 goto err_free;
 578
 579         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 580         if (ret)
 581                 goto err_wb_exit;
 582
 583         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 584         if (ret)
 585                 goto err_ref_exit;
 586
 587         wb->memcg_css = memcg_css;
 588         wb->blkcg_css = blkcg_css;
 589         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 590         set_bit(WB_registered, &wb->state);
 591
 592         /*
 593          * The root wb determines the registered state of the whole bdi and
 594          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 595          * whether they're still online.  Don't link @wb if any is dead.
 596          * See wb_memcg_offline() and wb_blkcg_offline().
 597          */
 598         ret = -ENODEV;
 599         spin_lock_irqsave(&cgwb_lock, flags);
 600         if (test_bit(WB_registered, &bdi->wb.state) &&
 601             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 602                 /* we might have raced another instance of this function */
 603                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 604                 if (!ret) {
 605                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 606                         list_add(&wb->memcg_node, memcg_cgwb_list);
 607                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 608                         css_get(memcg_css);
 609                         css_get(blkcg_css);
 610                 }
 611         }
 612         spin_unlock_irqrestore(&cgwb_lock, flags);
 613         if (ret) {
 614                 if (ret == -EEXIST)
 615                         ret = 0;
 616                 goto err_fprop_exit;
 617         }
 618         goto out_put;
 619
 620 err_fprop_exit:
 621         fprop_local_destroy_percpu(&wb->memcg_completions);
 622 err_ref_exit:
 623         percpu_ref_exit(&wb->refcnt);
 624 err_wb_exit:
 625         wb_exit(wb);
 626 err_free:
 627         kfree(wb);
 628 out_put:
 629         css_put(blkcg_css);
 630         return ret;
 631 }
 632
 633 /**
 634  * wb_get_create - get wb for a given memcg, create if necessary
 635  * @bdi: target bdi
 636  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 637  * @gfp: allocation mask to use
 638  *
 639  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 640  * create one.  The returned wb has its refcount incremented.
 641  *
 642  * This function uses css_get() on @memcg_css and thus expects its refcnt
 643  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 644  * @memcg_css isn't enough.  try_get it before calling this function.
 645  *
 646  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 647  * memcg on the default hierarchy, memcg association is guaranteed to be
 648  * more specific (equal or descendant to the associated blkcg) and thus can
 649  * identify both the memcg and blkcg associations.
 650  *
 651  * Because the blkcg associated with a memcg may change as blkcg is enabled
 652  * and disabled closer to root in the hierarchy, each wb keeps track of
 653  * both the memcg and blkcg associated with it and verifies the blkcg on
 654  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 655  * created.
 656  */
 657 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 658                                     struct cgroup_subsys_state *memcg_css,
 659                                     gfp_t gfp)
 660 {
 661         struct bdi_writeback *wb;
 662
 663         might_sleep_if(gfpflags_allow_blocking(gfp));
 664
 665         if (!memcg_css->parent)
 666                 return &bdi->wb;
 667
 668         do {
 669                 rcu_read_lock();
 670                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 671                 if (wb) {
 672                         struct cgroup_subsys_state *blkcg_css;
 673
 674                         /* see whether the blkcg association has changed */
 675                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 676                                                      &io_cgrp_subsys);
 677                         if (unlikely(wb->blkcg_css != blkcg_css ||
 678                                      !wb_tryget(wb)))
 679                                 wb = NULL;
 680                         css_put(blkcg_css);
 681                 }
 682                 rcu_read_unlock();
 683         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 684
 685         return wb;
 686 }
 687
 688 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 689 {
 690         int ret;
 691
 692         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 693         bdi->cgwb_congested_tree = RB_ROOT;
 694
 695         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 696         if (!ret) {
 697                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 698                 bdi->wb.blkcg_css = blkcg_root_css;
 699         }
 700         return ret;
 701 }
 702
 703 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 704 {
 705         struct radix_tree_iter iter;
 706         void **slot;
 707         struct bdi_writeback *wb;
 708
 709         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 710
 711         spin_lock_irq(&cgwb_lock);
 712         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 713                 cgwb_kill(*slot);
 714
 715         while (!list_empty(&bdi->wb_list)) {
 716                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 717                                       bdi_node);
 718                 spin_unlock_irq(&cgwb_lock);
 719                 wb_shutdown(wb);
 720                 spin_lock_irq(&cgwb_lock);
 721         }
 722         spin_unlock_irq(&cgwb_lock);
 723 }
 724
 725 /**
 726  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 727  * @memcg: memcg being offlined
 728  *
 729  * Also prevents creation of any new wb's associated with @memcg.
 730  */
 731 void wb_memcg_offline(struct mem_cgroup *memcg)
 732 {
 733         LIST_HEAD(to_destroy);
 734         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 735         struct bdi_writeback *wb, *next;
 736
 737         spin_lock_irq(&cgwb_lock);
 738         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 739                 cgwb_kill(wb);
 740         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 741         spin_unlock_irq(&cgwb_lock);
 742 }
 743
 744 /**
 745  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 746  * @blkcg: blkcg being offlined
 747  *
 748  * Also prevents creation of any new wb's associated with @blkcg.
 749  */
 750 void wb_blkcg_offline(struct blkcg *blkcg)
 751 {
 752         LIST_HEAD(to_destroy);
 753         struct bdi_writeback *wb, *next;
 754
 755         spin_lock_irq(&cgwb_lock);
 756         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 757                 cgwb_kill(wb);
 758         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 759         spin_unlock_irq(&cgwb_lock);
 760 }
 761
 762 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 763 {
 764         struct rb_node *rbn;
 765
 766         spin_lock_irq(&cgwb_lock);
 767         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 768                 struct bdi_writeback_congested *congested =
 769                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 770
 771                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 772                 congested->__bdi = NULL;        /* mark @congested unlinked */
 773         }
 774         spin_unlock_irq(&cgwb_lock);
 775 }
 776
 777 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 778 {
 779         spin_lock_irq(&cgwb_lock);
 780         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 781         spin_unlock_irq(&cgwb_lock);
 782 }
 783
 784 #else   /* CONFIG_CGROUP_WRITEBACK */
 785
 786 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 787 {
 788         int err;
 789
 790         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 791         if (!bdi->wb_congested)
 792                 return -ENOMEM;
 793
 794         atomic_set(&bdi->wb_congested->refcnt, 1);
 795
 796         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 797         if (err) {
 798                 wb_congested_put(bdi->wb_congested);
 799                 return err;
 800         }
 801         return 0;
 802 }
 803
 804 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 805
 806 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 807 {
 808         wb_congested_put(bdi->wb_congested);
 809 }
 810
 811 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 812 {
 813         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 814 }
 815
 816 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 817 {
 818         list_del_rcu(&wb->bdi_node);
 819 }
 820
 821 #endif  /* CONFIG_CGROUP_WRITEBACK */
 822
 823 static int bdi_init(struct backing_dev_info *bdi)
 824 {
 825         int ret;
 826
 827         bdi->dev = NULL;
 828
 829         kref_init(&bdi->refcnt);
 830         bdi->min_ratio = 0;
 831         bdi->max_ratio = 100;
 832         bdi->max_prop_frac = FPROP_FRAC_BASE;
 833         INIT_LIST_HEAD(&bdi->bdi_list);
 834         INIT_LIST_HEAD(&bdi->wb_list);
 835         init_waitqueue_head(&bdi->wb_waitq);
 836
 837         ret = cgwb_bdi_init(bdi);
 838
 839         return ret;
 840 }
 841
 842 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 843 {
 844         struct backing_dev_info *bdi;
 845
 846         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 847                            gfp_mask | __GFP_ZERO, node_id);
 848         if (!bdi)
 849                 return NULL;
 850
 851         if (bdi_init(bdi)) {
 852                 kfree(bdi);
 853                 return NULL;
 854         }
 855         return bdi;
 856 }
 857 EXPORT_SYMBOL(bdi_alloc_node);
 858
 859 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 860 {
 861         struct device *dev;
 862
 863         if (bdi->dev)   /* The driver needs to use separate queues per device */
 864                 return 0;
 865
 866         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 867         if (IS_ERR(dev))
 868                 return PTR_ERR(dev);
 869
 870         cgwb_bdi_register(bdi);
 871         bdi->dev = dev;
 872
 873         bdi_debug_register(bdi, dev_name(dev));
 874         set_bit(WB_registered, &bdi->wb.state);
 875
 876         spin_lock_bh(&bdi_lock);
 877         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 878         spin_unlock_bh(&bdi_lock);
 879
 880         trace_writeback_bdi_register(bdi);
 881         return 0;
 882 }
 883 EXPORT_SYMBOL(bdi_register_va);
 884
 885 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 886 {
 887         va_list args;
 888         int ret;
 889
 890         va_start(args, fmt);
 891         ret = bdi_register_va(bdi, fmt, args);
 892         va_end(args);
 893         return ret;
 894 }
 895 EXPORT_SYMBOL(bdi_register);
 896
 897 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 898 {
 899         int rc;
 900
 901         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 902         if (rc)
 903                 return rc;
 904         /* Leaking owner reference... */
 905         WARN_ON(bdi->owner);
 906         bdi->owner = owner;
 907         get_device(owner);
 908         return 0;
 909 }
 910 EXPORT_SYMBOL(bdi_register_owner);
 911
 912 /*
 913  * Remove bdi from bdi_list, and ensure that it is no longer visible
 914  */
 915 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 916 {
 917         spin_lock_bh(&bdi_lock);
 918         list_del_rcu(&bdi->bdi_list);
 919         spin_unlock_bh(&bdi_lock);
 920
 921         synchronize_rcu_expedited();
 922 }
 923
 924 void bdi_unregister(struct backing_dev_info *bdi)
 925 {
 926         /* make sure nobody finds us on the bdi_list anymore */
 927         bdi_remove_from_list(bdi);
 928         wb_shutdown(&bdi->wb);
 929         cgwb_bdi_unregister(bdi);
 930
 931         if (bdi->dev) {
 932                 bdi_debug_unregister(bdi);
 933                 device_unregister(bdi->dev);
 934                 bdi->dev = NULL;
 935         }
 936
 937         if (bdi->owner) {
 938                 put_device(bdi->owner);
 939                 bdi->owner = NULL;
 940         }
 941 }
 942
 943 static void release_bdi(struct kref *ref)
 944 {
 945         struct backing_dev_info *bdi =
 946                         container_of(ref, struct backing_dev_info, refcnt);
 947
 948         if (test_bit(WB_registered, &bdi->wb.state))
 949                 bdi_unregister(bdi);
 950         WARN_ON_ONCE(bdi->dev);
 951         wb_exit(&bdi->wb);
 952         cgwb_bdi_exit(bdi);
 953         kfree(bdi);
 954 }
 955
 956 void bdi_put(struct backing_dev_info *bdi)
 957 {
 958         kref_put(&bdi->refcnt, release_bdi);
 959 }
 960 EXPORT_SYMBOL(bdi_put);
 961
 962 static wait_queue_head_t congestion_wqh[2] = {
 963                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 964                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 965         };
 966 static atomic_t nr_wb_congested[2];
 967
 968 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 969 {
 970         wait_queue_head_t *wqh = &congestion_wqh[sync];
 971         enum wb_congested_state bit;
 972
 973         bit = sync ? WB_sync_congested : WB_async_congested;
 974         if (test_and_clear_bit(bit, &congested->state))
 975                 atomic_dec(&nr_wb_congested[sync]);
 976         smp_mb__after_atomic();
 977         if (waitqueue_active(wqh))
 978                 wake_up(wqh);
 979 }
 980 EXPORT_SYMBOL(clear_wb_congested);
 981
 982 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
 983 {
 984         enum wb_congested_state bit;
 985
 986         bit = sync ? WB_sync_congested : WB_async_congested;
 987         if (!test_and_set_bit(bit, &congested->state))
 988                 atomic_inc(&nr_wb_congested[sync]);
 989 }
 990 EXPORT_SYMBOL(set_wb_congested);
 991
 992 /**
 993  * congestion_wait - wait for a backing_dev to become uncongested
 994  * @sync: SYNC or ASYNC IO
 995  * @timeout: timeout in jiffies
 996  *
 997  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 998  * write congestion.  If no backing_devs are congested then just wait for the
 999  * next write to be completed.
1000  */
1001 long congestion_wait(int sync, long timeout)
1002 {
1003         long ret;
1004         unsigned long start = jiffies;
1005         DEFINE_WAIT(wait);
1006         wait_queue_head_t *wqh = &congestion_wqh[sync];
1007
1008         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1009         ret = io_schedule_timeout(timeout);
1010         finish_wait(wqh, &wait);
1011
1012         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1013                                         jiffies_to_usecs(jiffies - start));
1014
1015         return ret;
1016 }
1017 EXPORT_SYMBOL(congestion_wait);
1018
1019 /**
1020  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1021  * @pgdat: A pgdat to check if it is heavily congested
1022  * @sync: SYNC or ASYNC IO
1023  * @timeout: timeout in jiffies
1024  *
1025  * In the event of a congested backing_dev (any backing_dev) and the given
1026  * @pgdat has experienced recent congestion, this waits for up to @timeout
1027  * jiffies for either a BDI to exit congestion of the given @sync queue
1028  * or a write to complete.
1029  *
1030  * In the absence of pgdat congestion, cond_resched() is called to yield
1031  * the processor if necessary but otherwise does not sleep.
1032  *
1033  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1034  * it is the number of jiffies that were still remaining when the function
1035  * returned. return_value == timeout implies the function did not sleep.
1036  */
1037 long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
1038 {
1039         long ret;
1040         unsigned long start = jiffies;
1041         DEFINE_WAIT(wait);
1042         wait_queue_head_t *wqh = &congestion_wqh[sync];
1043
1044         /*
1045          * If there is no congestion, or heavy congestion is not being
1046          * encountered in the current pgdat, yield if necessary instead
1047          * of sleeping on the congestion queue
1048          */
1049         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1050             !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1051                 cond_resched();
1052
1053                 /* In case we scheduled, work out time remaining */
1054                 ret = timeout - (jiffies - start);
1055                 if (ret < 0)
1056                         ret = 0;
1057
1058                 goto out;
1059         }
1060
1061         /* Sleep until uncongested or a write happens */
1062         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1063         ret = io_schedule_timeout(timeout);
1064         finish_wait(wqh, &wait);
1065
1066 out:
1067         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1068                                         jiffies_to_usecs(jiffies - start));
1069
1070         return ret;
1071 }
1072 EXPORT_SYMBOL(wait_iff_congested);
1073
1074 int pdflush_proc_obsolete(struct ctl_table *table, int write,
1075                         void __user *buffer, size_t *lenp, loff_t *ppos)
1076 {
1077         char kbuf[] = "0\n";
1078
1079         if (*ppos || *lenp < sizeof(kbuf)) {
1080                 *lenp = 0;
1081                 return 0;
1082         }
1083
1084         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1085                 return -EFAULT;
1086         pr_warn_once("%s exported in /proc is scheduled for removal\n",
1087                      table->procname);
1088
1089         *lenp = 2;
1090         *ppos += *lenp;
1091         return 2;
1092 }