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