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                 return;
 360         }
 361         spin_unlock_bh(&wb->work_lock);
 362
 363         cgwb_remove_from_bdi_list(wb);
 364         /*
 365          * Drain work list and shutdown the delayed_work.  !WB_registered
 366          * tells wb_workfn() that @wb is dying and its work_list needs to
 367          * be drained no matter what.
 368          */
 369         mod_delayed_work(bdi_wq, &wb->dwork, 0);
 370         flush_delayed_work(&wb->dwork);
 371         WARN_ON(!list_empty(&wb->work_list));
 372 }
 373
 374 static void wb_exit(struct bdi_writeback *wb)
 375 {
 376         int i;
 377
 378         WARN_ON(delayed_work_pending(&wb->dwork));
 379
 380         for (i = 0; i < NR_WB_STAT_ITEMS; i++)
 381                 percpu_counter_destroy(&wb->stat[i]);
 382
 383         fprop_local_destroy_percpu(&wb->completions);
 384         wb_congested_put(wb->congested);
 385         if (wb != &wb->bdi->wb)
 386                 bdi_put(wb->bdi);
 387 }
 388
 389 #ifdef CONFIG_CGROUP_WRITEBACK
 390
 391 #include <linux/memcontrol.h>
 392
 393 /*
 394  * cgwb_lock protects bdi->cgwb_tree, bdi->cgwb_congested_tree,
 395  * blkcg->cgwb_list, and memcg->cgwb_list.  bdi->cgwb_tree is also RCU
 396  * protected.
 397  */
 398 static DEFINE_SPINLOCK(cgwb_lock);
 399 static struct workqueue_struct *cgwb_release_wq;
 400
 401 /**
 402  * wb_congested_get_create - get or create a wb_congested
 403  * @bdi: associated bdi
 404  * @blkcg_id: ID of the associated blkcg
 405  * @gfp: allocation mask
 406  *
 407  * Look up the wb_congested for @blkcg_id on @bdi.  If missing, create one.
 408  * The returned wb_congested has its reference count incremented.  Returns
 409  * NULL on failure.
 410  */
 411 struct bdi_writeback_congested *
 412 wb_congested_get_create(struct backing_dev_info *bdi, int blkcg_id, gfp_t gfp)
 413 {
 414         struct bdi_writeback_congested *new_congested = NULL, *congested;
 415         struct rb_node **node, *parent;
 416         unsigned long flags;
 417 retry:
 418         spin_lock_irqsave(&cgwb_lock, flags);
 419
 420         node = &bdi->cgwb_congested_tree.rb_node;
 421         parent = NULL;
 422
 423         while (*node != NULL) {
 424                 parent = *node;
 425                 congested = rb_entry(parent, struct bdi_writeback_congested,
 426                                      rb_node);
 427                 if (congested->blkcg_id < blkcg_id)
 428                         node = &parent->rb_left;
 429                 else if (congested->blkcg_id > blkcg_id)
 430                         node = &parent->rb_right;
 431                 else
 432                         goto found;
 433         }
 434
 435         if (new_congested) {
 436                 /* !found and storage for new one already allocated, insert */
 437                 congested = new_congested;
 438                 new_congested = NULL;
 439                 rb_link_node(&congested->rb_node, parent, node);
 440                 rb_insert_color(&congested->rb_node, &bdi->cgwb_congested_tree);
 441                 goto found;
 442         }
 443
 444         spin_unlock_irqrestore(&cgwb_lock, flags);
 445
 446         /* allocate storage for new one and retry */
 447         new_congested = kzalloc(sizeof(*new_congested), gfp);
 448         if (!new_congested)
 449                 return NULL;
 450
 451         atomic_set(&new_congested->refcnt, 0);
 452         new_congested->__bdi = bdi;
 453         new_congested->blkcg_id = blkcg_id;
 454         goto retry;
 455
 456 found:
 457         atomic_inc(&congested->refcnt);
 458         spin_unlock_irqrestore(&cgwb_lock, flags);
 459         kfree(new_congested);
 460         return congested;
 461 }
 462
 463 /**
 464  * wb_congested_put - put a wb_congested
 465  * @congested: wb_congested to put
 466  *
 467  * Put @congested and destroy it if the refcnt reaches zero.
 468  */
 469 void wb_congested_put(struct bdi_writeback_congested *congested)
 470 {
 471         unsigned long flags;
 472
 473         local_irq_save(flags);
 474         if (!atomic_dec_and_lock(&congested->refcnt, &cgwb_lock)) {
 475                 local_irq_restore(flags);
 476                 return;
 477         }
 478
 479         /* bdi might already have been destroyed leaving @congested unlinked */
 480         if (congested->__bdi) {
 481                 rb_erase(&congested->rb_node,
 482                          &congested->__bdi->cgwb_congested_tree);
 483                 congested->__bdi = NULL;
 484         }
 485
 486         spin_unlock_irqrestore(&cgwb_lock, flags);
 487         kfree(congested);
 488 }
 489
 490 static void cgwb_release_workfn(struct work_struct *work)
 491 {
 492         struct bdi_writeback *wb = container_of(work, struct bdi_writeback,
 493                                                 release_work);
 494
 495         mutex_lock(&wb->bdi->cgwb_release_mutex);
 496         wb_shutdown(wb);
 497
 498         css_put(wb->memcg_css);
 499         css_put(wb->blkcg_css);
 500         mutex_unlock(&wb->bdi->cgwb_release_mutex);
 501
 502         fprop_local_destroy_percpu(&wb->memcg_completions);
 503         percpu_ref_exit(&wb->refcnt);
 504         wb_exit(wb);
 505         kfree_rcu(wb, rcu);
 506 }
 507
 508 static void cgwb_release(struct percpu_ref *refcnt)
 509 {
 510         struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback,
 511                                                 refcnt);
 512         queue_work(cgwb_release_wq, &wb->release_work);
 513 }
 514
 515 static void cgwb_kill(struct bdi_writeback *wb)
 516 {
 517         lockdep_assert_held(&cgwb_lock);
 518
 519         WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id));
 520         list_del(&wb->memcg_node);
 521         list_del(&wb->blkcg_node);
 522         percpu_ref_kill(&wb->refcnt);
 523 }
 524
 525 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 526 {
 527         spin_lock_irq(&cgwb_lock);
 528         list_del_rcu(&wb->bdi_node);
 529         spin_unlock_irq(&cgwb_lock);
 530 }
 531
 532 static int cgwb_create(struct backing_dev_info *bdi,
 533                        struct cgroup_subsys_state *memcg_css, gfp_t gfp)
 534 {
 535         struct mem_cgroup *memcg;
 536         struct cgroup_subsys_state *blkcg_css;
 537         struct blkcg *blkcg;
 538         struct list_head *memcg_cgwb_list, *blkcg_cgwb_list;
 539         struct bdi_writeback *wb;
 540         unsigned long flags;
 541         int ret = 0;
 542
 543         memcg = mem_cgroup_from_css(memcg_css);
 544         blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys);
 545         blkcg = css_to_blkcg(blkcg_css);
 546         memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 547         blkcg_cgwb_list = &blkcg->cgwb_list;
 548
 549         /* look up again under lock and discard on blkcg mismatch */
 550         spin_lock_irqsave(&cgwb_lock, flags);
 551         wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 552         if (wb && wb->blkcg_css != blkcg_css) {
 553                 cgwb_kill(wb);
 554                 wb = NULL;
 555         }
 556         spin_unlock_irqrestore(&cgwb_lock, flags);
 557         if (wb)
 558                 goto out_put;
 559
 560         /* need to create a new one */
 561         wb = kmalloc(sizeof(*wb), gfp);
 562         if (!wb) {
 563                 ret = -ENOMEM;
 564                 goto out_put;
 565         }
 566
 567         ret = wb_init(wb, bdi, blkcg_css->id, gfp);
 568         if (ret)
 569                 goto err_free;
 570
 571         ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp);
 572         if (ret)
 573                 goto err_wb_exit;
 574
 575         ret = fprop_local_init_percpu(&wb->memcg_completions, gfp);
 576         if (ret)
 577                 goto err_ref_exit;
 578
 579         wb->memcg_css = memcg_css;
 580         wb->blkcg_css = blkcg_css;
 581         INIT_WORK(&wb->release_work, cgwb_release_workfn);
 582         set_bit(WB_registered, &wb->state);
 583
 584         /*
 585          * The root wb determines the registered state of the whole bdi and
 586          * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate
 587          * whether they're still online.  Don't link @wb if any is dead.
 588          * See wb_memcg_offline() and wb_blkcg_offline().
 589          */
 590         ret = -ENODEV;
 591         spin_lock_irqsave(&cgwb_lock, flags);
 592         if (test_bit(WB_registered, &bdi->wb.state) &&
 593             blkcg_cgwb_list->next && memcg_cgwb_list->next) {
 594                 /* we might have raced another instance of this function */
 595                 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb);
 596                 if (!ret) {
 597                         list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list);
 598                         list_add(&wb->memcg_node, memcg_cgwb_list);
 599                         list_add(&wb->blkcg_node, blkcg_cgwb_list);
 600                         css_get(memcg_css);
 601                         css_get(blkcg_css);
 602                 }
 603         }
 604         spin_unlock_irqrestore(&cgwb_lock, flags);
 605         if (ret) {
 606                 if (ret == -EEXIST)
 607                         ret = 0;
 608                 goto err_fprop_exit;
 609         }
 610         goto out_put;
 611
 612 err_fprop_exit:
 613         fprop_local_destroy_percpu(&wb->memcg_completions);
 614 err_ref_exit:
 615         percpu_ref_exit(&wb->refcnt);
 616 err_wb_exit:
 617         wb_exit(wb);
 618 err_free:
 619         kfree(wb);
 620 out_put:
 621         css_put(blkcg_css);
 622         return ret;
 623 }
 624
 625 /**
 626  * wb_get_create - get wb for a given memcg, create if necessary
 627  * @bdi: target bdi
 628  * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref)
 629  * @gfp: allocation mask to use
 630  *
 631  * Try to get the wb for @memcg_css on @bdi.  If it doesn't exist, try to
 632  * create one.  The returned wb has its refcount incremented.
 633  *
 634  * This function uses css_get() on @memcg_css and thus expects its refcnt
 635  * to be positive on invocation.  IOW, rcu_read_lock() protection on
 636  * @memcg_css isn't enough.  try_get it before calling this function.
 637  *
 638  * A wb is keyed by its associated memcg.  As blkcg implicitly enables
 639  * memcg on the default hierarchy, memcg association is guaranteed to be
 640  * more specific (equal or descendant to the associated blkcg) and thus can
 641  * identify both the memcg and blkcg associations.
 642  *
 643  * Because the blkcg associated with a memcg may change as blkcg is enabled
 644  * and disabled closer to root in the hierarchy, each wb keeps track of
 645  * both the memcg and blkcg associated with it and verifies the blkcg on
 646  * each lookup.  On mismatch, the existing wb is discarded and a new one is
 647  * created.
 648  */
 649 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi,
 650                                     struct cgroup_subsys_state *memcg_css,
 651                                     gfp_t gfp)
 652 {
 653         struct bdi_writeback *wb;
 654
 655         might_sleep_if(gfpflags_allow_blocking(gfp));
 656
 657         if (!memcg_css->parent)
 658                 return &bdi->wb;
 659
 660         do {
 661                 rcu_read_lock();
 662                 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id);
 663                 if (wb) {
 664                         struct cgroup_subsys_state *blkcg_css;
 665
 666                         /* see whether the blkcg association has changed */
 667                         blkcg_css = cgroup_get_e_css(memcg_css->cgroup,
 668                                                      &io_cgrp_subsys);
 669                         if (unlikely(wb->blkcg_css != blkcg_css ||
 670                                      !wb_tryget(wb)))
 671                                 wb = NULL;
 672                         css_put(blkcg_css);
 673                 }
 674                 rcu_read_unlock();
 675         } while (!wb && !cgwb_create(bdi, memcg_css, gfp));
 676
 677         return wb;
 678 }
 679
 680 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 681 {
 682         int ret;
 683
 684         INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC);
 685         bdi->cgwb_congested_tree = RB_ROOT;
 686         mutex_init(&bdi->cgwb_release_mutex);
 687
 688         ret = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 689         if (!ret) {
 690                 bdi->wb.memcg_css = &root_mem_cgroup->css;
 691                 bdi->wb.blkcg_css = blkcg_root_css;
 692         }
 693         return ret;
 694 }
 695
 696 static void cgwb_bdi_unregister(struct backing_dev_info *bdi)
 697 {
 698         struct radix_tree_iter iter;
 699         void **slot;
 700         struct bdi_writeback *wb;
 701
 702         WARN_ON(test_bit(WB_registered, &bdi->wb.state));
 703
 704         spin_lock_irq(&cgwb_lock);
 705         radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0)
 706                 cgwb_kill(*slot);
 707         spin_unlock_irq(&cgwb_lock);
 708
 709         mutex_lock(&bdi->cgwb_release_mutex);
 710         spin_lock_irq(&cgwb_lock);
 711         while (!list_empty(&bdi->wb_list)) {
 712                 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback,
 713                                       bdi_node);
 714                 spin_unlock_irq(&cgwb_lock);
 715                 wb_shutdown(wb);
 716                 spin_lock_irq(&cgwb_lock);
 717         }
 718         spin_unlock_irq(&cgwb_lock);
 719         mutex_unlock(&bdi->cgwb_release_mutex);
 720 }
 721
 722 /**
 723  * wb_memcg_offline - kill all wb's associated with a memcg being offlined
 724  * @memcg: memcg being offlined
 725  *
 726  * Also prevents creation of any new wb's associated with @memcg.
 727  */
 728 void wb_memcg_offline(struct mem_cgroup *memcg)
 729 {
 730         LIST_HEAD(to_destroy);
 731         struct list_head *memcg_cgwb_list = mem_cgroup_cgwb_list(memcg);
 732         struct bdi_writeback *wb, *next;
 733
 734         spin_lock_irq(&cgwb_lock);
 735         list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node)
 736                 cgwb_kill(wb);
 737         memcg_cgwb_list->next = NULL;   /* prevent new wb's */
 738         spin_unlock_irq(&cgwb_lock);
 739 }
 740
 741 /**
 742  * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined
 743  * @blkcg: blkcg being offlined
 744  *
 745  * Also prevents creation of any new wb's associated with @blkcg.
 746  */
 747 void wb_blkcg_offline(struct blkcg *blkcg)
 748 {
 749         LIST_HEAD(to_destroy);
 750         struct bdi_writeback *wb, *next;
 751
 752         spin_lock_irq(&cgwb_lock);
 753         list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node)
 754                 cgwb_kill(wb);
 755         blkcg->cgwb_list.next = NULL;   /* prevent new wb's */
 756         spin_unlock_irq(&cgwb_lock);
 757 }
 758
 759 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 760 {
 761         struct rb_node *rbn;
 762
 763         spin_lock_irq(&cgwb_lock);
 764         while ((rbn = rb_first(&bdi->cgwb_congested_tree))) {
 765                 struct bdi_writeback_congested *congested =
 766                         rb_entry(rbn, struct bdi_writeback_congested, rb_node);
 767
 768                 rb_erase(rbn, &bdi->cgwb_congested_tree);
 769                 congested->__bdi = NULL;        /* mark @congested unlinked */
 770         }
 771         spin_unlock_irq(&cgwb_lock);
 772 }
 773
 774 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 775 {
 776         spin_lock_irq(&cgwb_lock);
 777         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 778         spin_unlock_irq(&cgwb_lock);
 779 }
 780
 781 static int __init cgwb_init(void)
 782 {
 783         /*
 784          * There can be many concurrent release work items overwhelming
 785          * system_wq.  Put them in a separate wq and limit concurrency.
 786          * There's no point in executing many of these in parallel.
 787          */
 788         cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1);
 789         if (!cgwb_release_wq)
 790                 return -ENOMEM;
 791
 792         return 0;
 793 }
 794 subsys_initcall(cgwb_init);
 795
 796 #else   /* CONFIG_CGROUP_WRITEBACK */
 797
 798 static int cgwb_bdi_init(struct backing_dev_info *bdi)
 799 {
 800         int err;
 801
 802         bdi->wb_congested = kzalloc(sizeof(*bdi->wb_congested), GFP_KERNEL);
 803         if (!bdi->wb_congested)
 804                 return -ENOMEM;
 805
 806         atomic_set(&bdi->wb_congested->refcnt, 1);
 807
 808         err = wb_init(&bdi->wb, bdi, 1, GFP_KERNEL);
 809         if (err) {
 810                 wb_congested_put(bdi->wb_congested);
 811                 return err;
 812         }
 813         return 0;
 814 }
 815
 816 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { }
 817
 818 static void cgwb_bdi_exit(struct backing_dev_info *bdi)
 819 {
 820         wb_congested_put(bdi->wb_congested);
 821 }
 822
 823 static void cgwb_bdi_register(struct backing_dev_info *bdi)
 824 {
 825         list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list);
 826 }
 827
 828 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb)
 829 {
 830         list_del_rcu(&wb->bdi_node);
 831 }
 832
 833 #endif  /* CONFIG_CGROUP_WRITEBACK */
 834
 835 static int bdi_init(struct backing_dev_info *bdi)
 836 {
 837         int ret;
 838
 839         bdi->dev = NULL;
 840
 841         kref_init(&bdi->refcnt);
 842         bdi->min_ratio = 0;
 843         bdi->max_ratio = 100;
 844         bdi->max_prop_frac = FPROP_FRAC_BASE;
 845         INIT_LIST_HEAD(&bdi->bdi_list);
 846         INIT_LIST_HEAD(&bdi->wb_list);
 847         init_waitqueue_head(&bdi->wb_waitq);
 848
 849         ret = cgwb_bdi_init(bdi);
 850
 851         return ret;
 852 }
 853
 854 struct backing_dev_info *bdi_alloc_node(gfp_t gfp_mask, int node_id)
 855 {
 856         struct backing_dev_info *bdi;
 857
 858         bdi = kmalloc_node(sizeof(struct backing_dev_info),
 859                            gfp_mask | __GFP_ZERO, node_id);
 860         if (!bdi)
 861                 return NULL;
 862
 863         if (bdi_init(bdi)) {
 864                 kfree(bdi);
 865                 return NULL;
 866         }
 867         return bdi;
 868 }
 869 EXPORT_SYMBOL(bdi_alloc_node);
 870
 871 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args)
 872 {
 873         struct device *dev;
 874
 875         if (bdi->dev)   /* The driver needs to use separate queues per device */
 876                 return 0;
 877
 878         dev = device_create_vargs(bdi_class, NULL, MKDEV(0, 0), bdi, fmt, args);
 879         if (IS_ERR(dev))
 880                 return PTR_ERR(dev);
 881
 882         cgwb_bdi_register(bdi);
 883         bdi->dev = dev;
 884
 885         bdi_debug_register(bdi, dev_name(dev));
 886         set_bit(WB_registered, &bdi->wb.state);
 887
 888         spin_lock_bh(&bdi_lock);
 889         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 890         spin_unlock_bh(&bdi_lock);
 891
 892         trace_writeback_bdi_register(bdi);
 893         return 0;
 894 }
 895 EXPORT_SYMBOL(bdi_register_va);
 896
 897 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...)
 898 {
 899         va_list args;
 900         int ret;
 901
 902         va_start(args, fmt);
 903         ret = bdi_register_va(bdi, fmt, args);
 904         va_end(args);
 905         return ret;
 906 }
 907 EXPORT_SYMBOL(bdi_register);
 908
 909 int bdi_register_owner(struct backing_dev_info *bdi, struct device *owner)
 910 {
 911         int rc;
 912
 913         rc = bdi_register(bdi, "%u:%u", MAJOR(owner->devt), MINOR(owner->devt));
 914         if (rc)
 915                 return rc;
 916         /* Leaking owner reference... */
 917         WARN_ON(bdi->owner);
 918         bdi->owner = owner;
 919         get_device(owner);
 920         return 0;
 921 }
 922 EXPORT_SYMBOL(bdi_register_owner);
 923
 924 /*
 925  * Remove bdi from bdi_list, and ensure that it is no longer visible
 926  */
 927 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 928 {
 929         spin_lock_bh(&bdi_lock);
 930         list_del_rcu(&bdi->bdi_list);
 931         spin_unlock_bh(&bdi_lock);
 932
 933         synchronize_rcu_expedited();
 934 }
 935
 936 void bdi_unregister(struct backing_dev_info *bdi)
 937 {
 938         /* make sure nobody finds us on the bdi_list anymore */
 939         bdi_remove_from_list(bdi);
 940         wb_shutdown(&bdi->wb);
 941         cgwb_bdi_unregister(bdi);
 942
 943         if (bdi->dev) {
 944                 bdi_debug_unregister(bdi);
 945                 device_unregister(bdi->dev);
 946                 bdi->dev = NULL;
 947         }
 948
 949         if (bdi->owner) {
 950                 put_device(bdi->owner);
 951                 bdi->owner = NULL;
 952         }
 953 }
 954
 955 static void release_bdi(struct kref *ref)
 956 {
 957         struct backing_dev_info *bdi =
 958                         container_of(ref, struct backing_dev_info, refcnt);
 959
 960         if (test_bit(WB_registered, &bdi->wb.state))
 961                 bdi_unregister(bdi);
 962         WARN_ON_ONCE(bdi->dev);
 963         wb_exit(&bdi->wb);
 964         cgwb_bdi_exit(bdi);
 965         kfree(bdi);
 966 }
 967
 968 void bdi_put(struct backing_dev_info *bdi)
 969 {
 970         kref_put(&bdi->refcnt, release_bdi);
 971 }
 972 EXPORT_SYMBOL(bdi_put);
 973
 974 static wait_queue_head_t congestion_wqh[2] = {
 975                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 976                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 977         };
 978 static atomic_t nr_wb_congested[2];
 979
 980 void clear_wb_congested(struct bdi_writeback_congested *congested, int sync)
 981 {
 982         wait_queue_head_t *wqh = &congestion_wqh[sync];
 983         enum wb_congested_state bit;
 984
 985         bit = sync ? WB_sync_congested : WB_async_congested;
 986         if (test_and_clear_bit(bit, &congested->state))
 987                 atomic_dec(&nr_wb_congested[sync]);
 988         smp_mb__after_atomic();
 989         if (waitqueue_active(wqh))
 990                 wake_up(wqh);
 991 }
 992 EXPORT_SYMBOL(clear_wb_congested);
 993
 994 void set_wb_congested(struct bdi_writeback_congested *congested, int sync)
 995 {
 996         enum wb_congested_state bit;
 997
 998         bit = sync ? WB_sync_congested : WB_async_congested;
 999         if (!test_and_set_bit(bit, &congested->state))
1000                 atomic_inc(&nr_wb_congested[sync]);
1001 }
1002 EXPORT_SYMBOL(set_wb_congested);
1003
1004 /**
1005  * congestion_wait - wait for a backing_dev to become uncongested
1006  * @sync: SYNC or ASYNC IO
1007  * @timeout: timeout in jiffies
1008  *
1009  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
1010  * write congestion.  If no backing_devs are congested then just wait for the
1011  * next write to be completed.
1012  */
1013 long congestion_wait(int sync, long timeout)
1014 {
1015         long ret;
1016         unsigned long start = jiffies;
1017         DEFINE_WAIT(wait);
1018         wait_queue_head_t *wqh = &congestion_wqh[sync];
1019
1020         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1021         ret = io_schedule_timeout(timeout);
1022         finish_wait(wqh, &wait);
1023
1024         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
1025                                         jiffies_to_usecs(jiffies - start));
1026
1027         return ret;
1028 }
1029 EXPORT_SYMBOL(congestion_wait);
1030
1031 /**
1032  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes
1033  * @pgdat: A pgdat to check if it is heavily congested
1034  * @sync: SYNC or ASYNC IO
1035  * @timeout: timeout in jiffies
1036  *
1037  * In the event of a congested backing_dev (any backing_dev) and the given
1038  * @pgdat has experienced recent congestion, this waits for up to @timeout
1039  * jiffies for either a BDI to exit congestion of the given @sync queue
1040  * or a write to complete.
1041  *
1042  * In the absence of pgdat congestion, cond_resched() is called to yield
1043  * the processor if necessary but otherwise does not sleep.
1044  *
1045  * The return value is 0 if the sleep is for the full timeout. Otherwise,
1046  * it is the number of jiffies that were still remaining when the function
1047  * returned. return_value == timeout implies the function did not sleep.
1048  */
1049 long wait_iff_congested(struct pglist_data *pgdat, int sync, long timeout)
1050 {
1051         long ret;
1052         unsigned long start = jiffies;
1053         DEFINE_WAIT(wait);
1054         wait_queue_head_t *wqh = &congestion_wqh[sync];
1055
1056         /*
1057          * If there is no congestion, or heavy congestion is not being
1058          * encountered in the current pgdat, yield if necessary instead
1059          * of sleeping on the congestion queue
1060          */
1061         if (atomic_read(&nr_wb_congested[sync]) == 0 ||
1062             !test_bit(PGDAT_CONGESTED, &pgdat->flags)) {
1063                 cond_resched();
1064
1065                 /* In case we scheduled, work out time remaining */
1066                 ret = timeout - (jiffies - start);
1067                 if (ret < 0)
1068                         ret = 0;
1069
1070                 goto out;
1071         }
1072
1073         /* Sleep until uncongested or a write happens */
1074         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
1075         ret = io_schedule_timeout(timeout);
1076         finish_wait(wqh, &wait);
1077
1078 out:
1079         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
1080                                         jiffies_to_usecs(jiffies - start));
1081
1082         return ret;
1083 }
1084 EXPORT_SYMBOL(wait_iff_congested);
1085
1086 int pdflush_proc_obsolete(struct ctl_table *table, int write,
1087                         void __user *buffer, size_t *lenp, loff_t *ppos)
1088 {
1089         char kbuf[] = "0\n";
1090
1091         if (*ppos || *lenp < sizeof(kbuf)) {
1092                 *lenp = 0;
1093                 return 0;
1094         }
1095
1096         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
1097                 return -EFAULT;
1098         pr_warn_once("%s exported in /proc is scheduled for removal\n",
1099                      table->procname);
1100
1101         *lenp = 2;
1102         *ppos += *lenp;
1103         return 2;
1104 }