Pileus Git - ~andy/linux/blob - 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 <linux/slab.h>
  14 #include <trace/events/writeback.h>
  15
  16 static atomic_long_t bdi_seq = ATOMIC_LONG_INIT(0);
  17
  18 struct backing_dev_info default_backing_dev_info = {
  19         .name           = "default",
  20         .ra_pages       = VM_MAX_READAHEAD * 1024 / PAGE_CACHE_SIZE,
  21         .state          = 0,
  22         .capabilities   = BDI_CAP_MAP_COPY,
  23 };
  24 EXPORT_SYMBOL_GPL(default_backing_dev_info);
  25
  26 struct backing_dev_info noop_backing_dev_info = {
  27         .name           = "noop",
  28         .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
  29 };
  30 EXPORT_SYMBOL_GPL(noop_backing_dev_info);
  31
  32 static struct class *bdi_class;
  33
  34 /*
  35  * bdi_lock protects updates to bdi_list and bdi_pending_list, as well as
  36  * reader side protection for bdi_pending_list. bdi_list has RCU reader side
  37  * locking.
  38  */
  39 DEFINE_SPINLOCK(bdi_lock);
  40 LIST_HEAD(bdi_list);
  41 LIST_HEAD(bdi_pending_list);
  42
  43 void bdi_lock_two(struct bdi_writeback *wb1, struct bdi_writeback *wb2)
  44 {
  45         if (wb1 < wb2) {
  46                 spin_lock(&wb1->list_lock);
  47                 spin_lock_nested(&wb2->list_lock, 1);
  48         } else {
  49                 spin_lock(&wb2->list_lock);
  50                 spin_lock_nested(&wb1->list_lock, 1);
  51         }
  52 }
  53
  54 #ifdef CONFIG_DEBUG_FS
  55 #include <linux/debugfs.h>
  56 #include <linux/seq_file.h>
  57
  58 static struct dentry *bdi_debug_root;
  59
  60 static void bdi_debug_init(void)
  61 {
  62         bdi_debug_root = debugfs_create_dir("bdi", NULL);
  63 }
  64
  65 static int bdi_debug_stats_show(struct seq_file *m, void *v)
  66 {
  67         struct backing_dev_info *bdi = m->private;
  68         struct bdi_writeback *wb = &bdi->wb;
  69         unsigned long background_thresh;
  70         unsigned long dirty_thresh;
  71         unsigned long bdi_thresh;
  72         unsigned long nr_dirty, nr_io, nr_more_io;
  73         struct inode *inode;
  74
  75         nr_dirty = nr_io = nr_more_io = 0;
  76         spin_lock(&wb->list_lock);
  77         list_for_each_entry(inode, &wb->b_dirty, i_wb_list)
  78                 nr_dirty++;
  79         list_for_each_entry(inode, &wb->b_io, i_wb_list)
  80                 nr_io++;
  81         list_for_each_entry(inode, &wb->b_more_io, i_wb_list)
  82                 nr_more_io++;
  83         spin_unlock(&wb->list_lock);
  84
  85         global_dirty_limits(&background_thresh, &dirty_thresh);
  86         bdi_thresh = bdi_dirty_limit(bdi, dirty_thresh);
  87
  88 #define K(x) ((x) << (PAGE_SHIFT - 10))
  89         seq_printf(m,
  90                    "BdiWriteback:       %10lu kB\n"
  91                    "BdiReclaimable:     %10lu kB\n"
  92                    "BdiDirtyThresh:     %10lu kB\n"
  93                    "DirtyThresh:        %10lu kB\n"
  94                    "BackgroundThresh:   %10lu kB\n"
  95                    "BdiDirtied:         %10lu kB\n"
  96                    "BdiWritten:         %10lu kB\n"
  97                    "BdiWriteBandwidth:  %10lu kBps\n"
  98                    "b_dirty:            %10lu\n"
  99                    "b_io:               %10lu\n"
 100                    "b_more_io:          %10lu\n"
 101                    "bdi_list:           %10u\n"
 102                    "state:              %10lx\n",
 103                    (unsigned long) K(bdi_stat(bdi, BDI_WRITEBACK)),
 104                    (unsigned long) K(bdi_stat(bdi, BDI_RECLAIMABLE)),
 105                    K(bdi_thresh),
 106                    K(dirty_thresh),
 107                    K(background_thresh),
 108                    (unsigned long) K(bdi_stat(bdi, BDI_DIRTIED)),
 109                    (unsigned long) K(bdi_stat(bdi, BDI_WRITTEN)),
 110                    (unsigned long) K(bdi->write_bandwidth),
 111                    nr_dirty,
 112                    nr_io,
 113                    nr_more_io,
 114                    !list_empty(&bdi->bdi_list), bdi->state);
 115 #undef K
 116
 117         return 0;
 118 }
 119
 120 static int bdi_debug_stats_open(struct inode *inode, struct file *file)
 121 {
 122         return single_open(file, bdi_debug_stats_show, inode->i_private);
 123 }
 124
 125 static const struct file_operations bdi_debug_stats_fops = {
 126         .open           = bdi_debug_stats_open,
 127         .read           = seq_read,
 128         .llseek         = seq_lseek,
 129         .release        = single_release,
 130 };
 131
 132 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name)
 133 {
 134         bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root);
 135         bdi->debug_stats = debugfs_create_file("stats", 0444, bdi->debug_dir,
 136                                                bdi, &bdi_debug_stats_fops);
 137 }
 138
 139 static void bdi_debug_unregister(struct backing_dev_info *bdi)
 140 {
 141         debugfs_remove(bdi->debug_stats);
 142         debugfs_remove(bdi->debug_dir);
 143 }
 144 #else
 145 static inline void bdi_debug_init(void)
 146 {
 147 }
 148 static inline void bdi_debug_register(struct backing_dev_info *bdi,
 149                                       const char *name)
 150 {
 151 }
 152 static inline void bdi_debug_unregister(struct backing_dev_info *bdi)
 153 {
 154 }
 155 #endif
 156
 157 static ssize_t read_ahead_kb_store(struct device *dev,
 158                                   struct device_attribute *attr,
 159                                   const char *buf, size_t count)
 160 {
 161         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 162         unsigned long read_ahead_kb;
 163         ssize_t ret;
 164
 165         ret = kstrtoul(buf, 10, &read_ahead_kb);
 166         if (ret < 0)
 167                 return ret;
 168
 169         bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10);
 170
 171         return count;
 172 }
 173
 174 #define K(pages) ((pages) << (PAGE_SHIFT - 10))
 175
 176 #define BDI_SHOW(name, expr)                                            \
 177 static ssize_t name##_show(struct device *dev,                          \
 178                            struct device_attribute *attr, char *page)   \
 179 {                                                                       \
 180         struct backing_dev_info *bdi = dev_get_drvdata(dev);            \
 181                                                                         \
 182         return snprintf(page, PAGE_SIZE-1, "%lld\n", (long long)expr);  \
 183 }
 184
 185 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages))
 186
 187 static ssize_t min_ratio_store(struct device *dev,
 188                 struct device_attribute *attr, const char *buf, size_t count)
 189 {
 190         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 191         unsigned int ratio;
 192         ssize_t ret;
 193
 194         ret = kstrtouint(buf, 10, &ratio);
 195         if (ret < 0)
 196                 return ret;
 197
 198         ret = bdi_set_min_ratio(bdi, ratio);
 199         if (!ret)
 200                 ret = count;
 201
 202         return ret;
 203 }
 204 BDI_SHOW(min_ratio, bdi->min_ratio)
 205
 206 static ssize_t max_ratio_store(struct device *dev,
 207                 struct device_attribute *attr, const char *buf, size_t count)
 208 {
 209         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 210         unsigned int ratio;
 211         ssize_t ret;
 212
 213         ret = kstrtouint(buf, 10, &ratio);
 214         if (ret < 0)
 215                 return ret;
 216
 217         ret = bdi_set_max_ratio(bdi, ratio);
 218         if (!ret)
 219                 ret = count;
 220
 221         return ret;
 222 }
 223 BDI_SHOW(max_ratio, bdi->max_ratio)
 224
 225 static ssize_t cpu_list_store(struct device *dev,
 226                 struct device_attribute *attr, const char *buf, size_t count)
 227 {
 228         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 229         struct bdi_writeback *wb = &bdi->wb;
 230         cpumask_var_t newmask;
 231         ssize_t ret;
 232         struct task_struct *task;
 233
 234         if (!alloc_cpumask_var(&newmask, GFP_KERNEL))
 235                 return -ENOMEM;
 236
 237         ret = cpulist_parse(buf, newmask);
 238         if (!ret) {
 239                 spin_lock_bh(&bdi->wb_lock);
 240                 task = wb->task;
 241                 if (task)
 242                         get_task_struct(task);
 243                 spin_unlock_bh(&bdi->wb_lock);
 244
 245                 mutex_lock(&bdi->flusher_cpumask_lock);
 246                 if (task) {
 247                         ret = set_cpus_allowed_ptr(task, newmask);
 248                         put_task_struct(task);
 249                 }
 250                 if (ret == 0) {
 251                         cpumask_copy(bdi->flusher_cpumask, newmask);
 252                         ret = count;
 253                 }
 254                 mutex_unlock(&bdi->flusher_cpumask_lock);
 255
 256         }
 257         free_cpumask_var(newmask);
 258
 259         return ret;
 260 }
 261
 262 static ssize_t cpu_list_show(struct device *dev,
 263                 struct device_attribute *attr, char *page)
 264 {
 265         struct backing_dev_info *bdi = dev_get_drvdata(dev);
 266         ssize_t ret;
 267
 268         mutex_lock(&bdi->flusher_cpumask_lock);
 269         ret = cpulist_scnprintf(page, PAGE_SIZE-1, bdi->flusher_cpumask);
 270         mutex_unlock(&bdi->flusher_cpumask_lock);
 271
 272         return ret;
 273 }
 274
 275 #define __ATTR_RW(attr) __ATTR(attr, 0644, attr##_show, attr##_store)
 276
 277 static struct device_attribute bdi_dev_attrs[] = {
 278         __ATTR_RW(read_ahead_kb),
 279         __ATTR_RW(min_ratio),
 280         __ATTR_RW(max_ratio),
 281         __ATTR_RW(cpu_list),
 282         __ATTR_NULL,
 283 };
 284
 285 static __init int bdi_class_init(void)
 286 {
 287         bdi_class = class_create(THIS_MODULE, "bdi");
 288         if (IS_ERR(bdi_class))
 289                 return PTR_ERR(bdi_class);
 290
 291         bdi_class->dev_attrs = bdi_dev_attrs;
 292         bdi_debug_init();
 293         return 0;
 294 }
 295 postcore_initcall(bdi_class_init);
 296
 297 static int __init default_bdi_init(void)
 298 {
 299         int err;
 300
 301         err = bdi_init(&default_backing_dev_info);
 302         if (!err)
 303                 bdi_register(&default_backing_dev_info, NULL, "default");
 304         err = bdi_init(&noop_backing_dev_info);
 305
 306         return err;
 307 }
 308 subsys_initcall(default_bdi_init);
 309
 310 int bdi_has_dirty_io(struct backing_dev_info *bdi)
 311 {
 312         return wb_has_dirty_io(&bdi->wb);
 313 }
 314
 315 static void wakeup_timer_fn(unsigned long data)
 316 {
 317         struct backing_dev_info *bdi = (struct backing_dev_info *)data;
 318
 319         spin_lock_bh(&bdi->wb_lock);
 320         if (bdi->wb.task) {
 321                 trace_writeback_wake_thread(bdi);
 322                 wake_up_process(bdi->wb.task);
 323         } else if (bdi->dev) {
 324                 /*
 325                  * When bdi tasks are inactive for long time, they are killed.
 326                  * In this case we have to wake-up the forker thread which
 327                  * should create and run the bdi thread.
 328                  */
 329                 trace_writeback_wake_forker_thread(bdi);
 330                 wake_up_process(default_backing_dev_info.wb.task);
 331         }
 332         spin_unlock_bh(&bdi->wb_lock);
 333 }
 334
 335 /*
 336  * This function is used when the first inode for this bdi is marked dirty. It
 337  * wakes-up the corresponding bdi thread which should then take care of the
 338  * periodic background write-out of dirty inodes. Since the write-out would
 339  * starts only 'dirty_writeback_interval' centisecs from now anyway, we just
 340  * set up a timer which wakes the bdi thread up later.
 341  *
 342  * Note, we wouldn't bother setting up the timer, but this function is on the
 343  * fast-path (used by '__mark_inode_dirty()'), so we save few context switches
 344  * by delaying the wake-up.
 345  */
 346 void bdi_wakeup_thread_delayed(struct backing_dev_info *bdi)
 347 {
 348         unsigned long timeout;
 349
 350         timeout = msecs_to_jiffies(dirty_writeback_interval * 10);
 351         mod_timer(&bdi->wb.wakeup_timer, jiffies + timeout);
 352 }
 353
 354 /*
 355  * Calculate the longest interval (jiffies) bdi threads are allowed to be
 356  * inactive.
 357  */
 358 static unsigned long bdi_longest_inactive(void)
 359 {
 360         unsigned long interval;
 361
 362         interval = msecs_to_jiffies(dirty_writeback_interval * 10);
 363         return max(5UL * 60 * HZ, interval);
 364 }
 365
 366 /*
 367  * Clear pending bit and wakeup anybody waiting for flusher thread creation or
 368  * shutdown
 369  */
 370 static void bdi_clear_pending(struct backing_dev_info *bdi)
 371 {
 372         clear_bit(BDI_pending, &bdi->state);
 373         smp_mb__after_clear_bit();
 374         wake_up_bit(&bdi->state, BDI_pending);
 375 }
 376
 377 static int bdi_forker_thread(void *ptr)
 378 {
 379         struct bdi_writeback *me = ptr;
 380
 381         current->flags |= PF_SWAPWRITE;
 382         set_freezable();
 383
 384         /*
 385          * Our parent may run at a different priority, just set us to normal
 386          */
 387         set_user_nice(current, 0);
 388
 389         for (;;) {
 390                 struct task_struct *task = NULL;
 391                 struct backing_dev_info *bdi;
 392                 enum {
 393                         NO_ACTION,   /* Nothing to do */
 394                         FORK_THREAD, /* Fork bdi thread */
 395                         KILL_THREAD, /* Kill inactive bdi thread */
 396                 } action = NO_ACTION;
 397
 398                 /*
 399                  * Temporary measure, we want to make sure we don't see
 400                  * dirty data on the default backing_dev_info
 401                  */
 402                 if (wb_has_dirty_io(me) || !list_empty(&me->bdi->work_list)) {
 403                         del_timer(&me->wakeup_timer);
 404                         wb_do_writeback(me, 0);
 405                 }
 406
 407                 spin_lock_bh(&bdi_lock);
 408                 /*
 409                  * In the following loop we are going to check whether we have
 410                  * some work to do without any synchronization with tasks
 411                  * waking us up to do work for them. Set the task state here
 412                  * so that we don't miss wakeups after verifying conditions.
 413                  */
 414                 set_current_state(TASK_INTERRUPTIBLE);
 415
 416                 list_for_each_entry(bdi, &bdi_list, bdi_list) {
 417                         bool have_dirty_io;
 418
 419                         if (!bdi_cap_writeback_dirty(bdi) ||
 420                              bdi_cap_flush_forker(bdi))
 421                                 continue;
 422
 423                         WARN(!test_bit(BDI_registered, &bdi->state),
 424                              "bdi %p/%s is not registered!\n", bdi, bdi->name);
 425
 426                         have_dirty_io = !list_empty(&bdi->work_list) ||
 427                                         wb_has_dirty_io(&bdi->wb);
 428
 429                         /*
 430                          * If the bdi has work to do, but the thread does not
 431                          * exist - create it.
 432                          */
 433                         if (!bdi->wb.task && have_dirty_io) {
 434                                 /*
 435                                  * Set the pending bit - if someone will try to
 436                                  * unregister this bdi - it'll wait on this bit.
 437                                  */
 438                                 set_bit(BDI_pending, &bdi->state);
 439                                 action = FORK_THREAD;
 440                                 break;
 441                         }
 442
 443                         spin_lock(&bdi->wb_lock);
 444
 445                         /*
 446                          * If there is no work to do and the bdi thread was
 447                          * inactive long enough - kill it. The wb_lock is taken
 448                          * to make sure no-one adds more work to this bdi and
 449                          * wakes the bdi thread up.
 450                          */
 451                         if (bdi->wb.task && !have_dirty_io &&
 452                             time_after(jiffies, bdi->wb.last_active +
 453                                                 bdi_longest_inactive())) {
 454                                 task = bdi->wb.task;
 455                                 bdi->wb.task = NULL;
 456                                 spin_unlock(&bdi->wb_lock);
 457                                 set_bit(BDI_pending, &bdi->state);
 458                                 action = KILL_THREAD;
 459                                 break;
 460                         }
 461                         spin_unlock(&bdi->wb_lock);
 462                 }
 463                 spin_unlock_bh(&bdi_lock);
 464
 465                 /* Keep working if default bdi still has things to do */
 466                 if (!list_empty(&me->bdi->work_list))
 467                         __set_current_state(TASK_RUNNING);
 468
 469                 switch (action) {
 470                 case FORK_THREAD:
 471                         __set_current_state(TASK_RUNNING);
 472                         task = kthread_create(bdi_writeback_thread, &bdi->wb,
 473                                               "flush-%s", dev_name(bdi->dev));
 474                         if (IS_ERR(task)) {
 475                                 /*
 476                                  * If thread creation fails, force writeout of
 477                                  * the bdi from the thread. Hopefully 1024 is
 478                                  * large enough for efficient IO.
 479                                  */
 480                                 writeback_inodes_wb(&bdi->wb, 1024,
 481                                                     WB_REASON_FORKER_THREAD);
 482                         } else {
 483                                 int ret;
 484                                 /*
 485                                  * The spinlock makes sure we do not lose
 486                                  * wake-ups when racing with 'bdi_queue_work()'.
 487                                  * And as soon as the bdi thread is visible, we
 488                                  * can start it.
 489                                  */
 490                                 spin_lock_bh(&bdi->wb_lock);
 491                                 bdi->wb.task = task;
 492                                 spin_unlock_bh(&bdi->wb_lock);
 493                                 mutex_lock(&bdi->flusher_cpumask_lock);
 494                                 ret = set_cpus_allowed_ptr(task,
 495                                                         bdi->flusher_cpumask);
 496                                 mutex_unlock(&bdi->flusher_cpumask_lock);
 497                                 if (ret)
 498                                         printk_once("%s: failed to bind flusher"
 499                                                     " thread %s, error %d\n",
 500                                                     __func__, task->comm, ret);
 501                                 wake_up_process(task);
 502                         }
 503                         bdi_clear_pending(bdi);
 504                         break;
 505
 506                 case KILL_THREAD:
 507                         __set_current_state(TASK_RUNNING);
 508                         kthread_stop(task);
 509                         bdi_clear_pending(bdi);
 510                         break;
 511
 512                 case NO_ACTION:
 513                         if (!wb_has_dirty_io(me) || !dirty_writeback_interval)
 514                                 /*
 515                                  * There are no dirty data. The only thing we
 516                                  * should now care about is checking for
 517                                  * inactive bdi threads and killing them. Thus,
 518                                  * let's sleep for longer time, save energy and
 519                                  * be friendly for battery-driven devices.
 520                                  */
 521                                 schedule_timeout(bdi_longest_inactive());
 522                         else
 523                                 schedule_timeout(msecs_to_jiffies(dirty_writeback_interval * 10));
 524                         try_to_freeze();
 525                         break;
 526                 }
 527         }
 528
 529         return 0;
 530 }
 531
 532 /*
 533  * Remove bdi from bdi_list, and ensure that it is no longer visible
 534  */
 535 static void bdi_remove_from_list(struct backing_dev_info *bdi)
 536 {
 537         spin_lock_bh(&bdi_lock);
 538         list_del_rcu(&bdi->bdi_list);
 539         spin_unlock_bh(&bdi_lock);
 540
 541         synchronize_rcu_expedited();
 542 }
 543
 544 int bdi_register(struct backing_dev_info *bdi, struct device *parent,
 545                 const char *fmt, ...)
 546 {
 547         va_list args;
 548         struct device *dev;
 549
 550         if (bdi->dev)   /* The driver needs to use separate queues per device */
 551                 return 0;
 552
 553         va_start(args, fmt);
 554         dev = device_create_vargs(bdi_class, parent, MKDEV(0, 0), bdi, fmt, args);
 555         va_end(args);
 556         if (IS_ERR(dev))
 557                 return PTR_ERR(dev);
 558
 559         bdi->dev = dev;
 560
 561         /*
 562          * Just start the forker thread for our default backing_dev_info,
 563          * and add other bdi's to the list. They will get a thread created
 564          * on-demand when they need it.
 565          */
 566         if (bdi_cap_flush_forker(bdi)) {
 567                 struct bdi_writeback *wb = &bdi->wb;
 568
 569                 wb->task = kthread_run(bdi_forker_thread, wb, "bdi-%s",
 570                                                 dev_name(dev));
 571                 if (IS_ERR(wb->task))
 572                         return PTR_ERR(wb->task);
 573         } else {
 574                 int node;
 575                 /*
 576                  * Set up a default cpumask for the flusher threads that
 577                  * includes all cpus on the same numa node as the device.
 578                  * The mask may be overridden via sysfs.
 579                  */
 580                 node = dev_to_node(bdi->dev);
 581                 if (node != NUMA_NO_NODE)
 582                         cpumask_copy(bdi->flusher_cpumask,
 583                                      cpumask_of_node(node));
 584         }
 585
 586         bdi_debug_register(bdi, dev_name(dev));
 587         set_bit(BDI_registered, &bdi->state);
 588
 589         spin_lock_bh(&bdi_lock);
 590         list_add_tail_rcu(&bdi->bdi_list, &bdi_list);
 591         spin_unlock_bh(&bdi_lock);
 592
 593         trace_writeback_bdi_register(bdi);
 594         return 0;
 595 }
 596 EXPORT_SYMBOL(bdi_register);
 597
 598 int bdi_register_dev(struct backing_dev_info *bdi, dev_t dev)
 599 {
 600         return bdi_register(bdi, NULL, "%u:%u", MAJOR(dev), MINOR(dev));
 601 }
 602 EXPORT_SYMBOL(bdi_register_dev);
 603
 604 /*
 605  * Remove bdi from the global list and shutdown any threads we have running
 606  */
 607 static void bdi_wb_shutdown(struct backing_dev_info *bdi)
 608 {
 609         struct task_struct *task;
 610
 611         if (!bdi_cap_writeback_dirty(bdi))
 612                 return;
 613
 614         /*
 615          * Make sure nobody finds us on the bdi_list anymore
 616          */
 617         bdi_remove_from_list(bdi);
 618
 619         /*
 620          * If setup is pending, wait for that to complete first
 621          */
 622         wait_on_bit(&bdi->state, BDI_pending, bdi_sched_wait,
 623                         TASK_UNINTERRUPTIBLE);
 624
 625         /*
 626          * Finally, kill the kernel thread. We don't need to be RCU
 627          * safe anymore, since the bdi is gone from visibility.
 628          */
 629         spin_lock_bh(&bdi->wb_lock);
 630         task = bdi->wb.task;
 631         bdi->wb.task = NULL;
 632         spin_unlock_bh(&bdi->wb_lock);
 633
 634         if (task)
 635                 kthread_stop(task);
 636 }
 637
 638 /*
 639  * This bdi is going away now, make sure that no super_blocks point to it
 640  */
 641 static void bdi_prune_sb(struct backing_dev_info *bdi)
 642 {
 643         struct super_block *sb;
 644
 645         spin_lock(&sb_lock);
 646         list_for_each_entry(sb, &super_blocks, s_list) {
 647                 if (sb->s_bdi == bdi)
 648                         sb->s_bdi = &default_backing_dev_info;
 649         }
 650         spin_unlock(&sb_lock);
 651 }
 652
 653 void bdi_unregister(struct backing_dev_info *bdi)
 654 {
 655         struct device *dev = bdi->dev;
 656
 657         if (dev) {
 658                 bdi_set_min_ratio(bdi, 0);
 659                 trace_writeback_bdi_unregister(bdi);
 660                 bdi_prune_sb(bdi);
 661                 del_timer_sync(&bdi->wb.wakeup_timer);
 662
 663                 if (!bdi_cap_flush_forker(bdi))
 664                         bdi_wb_shutdown(bdi);
 665                 bdi_debug_unregister(bdi);
 666
 667                 spin_lock_bh(&bdi->wb_lock);
 668                 bdi->dev = NULL;
 669                 spin_unlock_bh(&bdi->wb_lock);
 670
 671                 device_unregister(dev);
 672         }
 673 }
 674 EXPORT_SYMBOL(bdi_unregister);
 675
 676 static void bdi_wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi)
 677 {
 678         memset(wb, 0, sizeof(*wb));
 679
 680         wb->bdi = bdi;
 681         wb->last_old_flush = jiffies;
 682         INIT_LIST_HEAD(&wb->b_dirty);
 683         INIT_LIST_HEAD(&wb->b_io);
 684         INIT_LIST_HEAD(&wb->b_more_io);
 685         spin_lock_init(&wb->list_lock);
 686         setup_timer(&wb->wakeup_timer, wakeup_timer_fn, (unsigned long)bdi);
 687 }
 688
 689 /*
 690  * Initial write bandwidth: 100 MB/s
 691  */
 692 #define INIT_BW         (100 << (20 - PAGE_SHIFT))
 693
 694 int bdi_init(struct backing_dev_info *bdi)
 695 {
 696         int i, err;
 697
 698         bdi->dev = NULL;
 699
 700         bdi->min_ratio = 0;
 701         bdi->max_ratio = 100;
 702         bdi->max_prop_frac = FPROP_FRAC_BASE;
 703         spin_lock_init(&bdi->wb_lock);
 704         INIT_LIST_HEAD(&bdi->bdi_list);
 705         INIT_LIST_HEAD(&bdi->work_list);
 706
 707         bdi_wb_init(&bdi->wb, bdi);
 708
 709         if (!bdi_cap_flush_forker(bdi)) {
 710                 bdi->flusher_cpumask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
 711                 if (!bdi->flusher_cpumask)
 712                         return -ENOMEM;
 713                 cpumask_setall(bdi->flusher_cpumask);
 714                 mutex_init(&bdi->flusher_cpumask_lock);
 715         } else
 716                 bdi->flusher_cpumask = NULL;
 717
 718         for (i = 0; i < NR_BDI_STAT_ITEMS; i++) {
 719                 err = percpu_counter_init(&bdi->bdi_stat[i], 0);
 720                 if (err)
 721                         goto err;
 722         }
 723
 724         bdi->dirty_exceeded = 0;
 725
 726         bdi->bw_time_stamp = jiffies;
 727         bdi->written_stamp = 0;
 728
 729         bdi->balanced_dirty_ratelimit = INIT_BW;
 730         bdi->dirty_ratelimit = INIT_BW;
 731         bdi->write_bandwidth = INIT_BW;
 732         bdi->avg_write_bandwidth = INIT_BW;
 733
 734         err = fprop_local_init_percpu(&bdi->completions);
 735
 736         if (err) {
 737 err:
 738                 while (i--)
 739                         percpu_counter_destroy(&bdi->bdi_stat[i]);
 740                 kfree(bdi->flusher_cpumask);
 741         }
 742
 743         return err;
 744 }
 745 EXPORT_SYMBOL(bdi_init);
 746
 747 void bdi_destroy(struct backing_dev_info *bdi)
 748 {
 749         int i;
 750
 751         /*
 752          * Splice our entries to the default_backing_dev_info, if this
 753          * bdi disappears
 754          */
 755         if (bdi_has_dirty_io(bdi)) {
 756                 struct bdi_writeback *dst = &default_backing_dev_info.wb;
 757
 758                 bdi_lock_two(&bdi->wb, dst);
 759                 list_splice(&bdi->wb.b_dirty, &dst->b_dirty);
 760                 list_splice(&bdi->wb.b_io, &dst->b_io);
 761                 list_splice(&bdi->wb.b_more_io, &dst->b_more_io);
 762                 spin_unlock(&bdi->wb.list_lock);
 763                 spin_unlock(&dst->list_lock);
 764         }
 765
 766         bdi_unregister(bdi);
 767
 768         kfree(bdi->flusher_cpumask);
 769
 770         /*
 771          * If bdi_unregister() had already been called earlier, the
 772          * wakeup_timer could still be armed because bdi_prune_sb()
 773          * can race with the bdi_wakeup_thread_delayed() calls from
 774          * __mark_inode_dirty().
 775          */
 776         del_timer_sync(&bdi->wb.wakeup_timer);
 777
 778         for (i = 0; i < NR_BDI_STAT_ITEMS; i++)
 779                 percpu_counter_destroy(&bdi->bdi_stat[i]);
 780
 781         fprop_local_destroy_percpu(&bdi->completions);
 782 }
 783 EXPORT_SYMBOL(bdi_destroy);
 784
 785 /*
 786  * For use from filesystems to quickly init and register a bdi associated
 787  * with dirty writeback
 788  */
 789 int bdi_setup_and_register(struct backing_dev_info *bdi, char *name,
 790                            unsigned int cap)
 791 {
 792         char tmp[32];
 793         int err;
 794
 795         bdi->name = name;
 796         bdi->capabilities = cap;
 797         err = bdi_init(bdi);
 798         if (err)
 799                 return err;
 800
 801         sprintf(tmp, "%.28s%s", name, "-%d");
 802         err = bdi_register(bdi, NULL, tmp, atomic_long_inc_return(&bdi_seq));
 803         if (err) {
 804                 bdi_destroy(bdi);
 805                 return err;
 806         }
 807
 808         return 0;
 809 }
 810 EXPORT_SYMBOL(bdi_setup_and_register);
 811
 812 static wait_queue_head_t congestion_wqh[2] = {
 813                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]),
 814                 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1])
 815         };
 816 static atomic_t nr_bdi_congested[2];
 817
 818 void clear_bdi_congested(struct backing_dev_info *bdi, int sync)
 819 {
 820         enum bdi_state bit;
 821         wait_queue_head_t *wqh = &congestion_wqh[sync];
 822
 823         bit = sync ? BDI_sync_congested : BDI_async_congested;
 824         if (test_and_clear_bit(bit, &bdi->state))
 825                 atomic_dec(&nr_bdi_congested[sync]);
 826         smp_mb__after_clear_bit();
 827         if (waitqueue_active(wqh))
 828                 wake_up(wqh);
 829 }
 830 EXPORT_SYMBOL(clear_bdi_congested);
 831
 832 void set_bdi_congested(struct backing_dev_info *bdi, int sync)
 833 {
 834         enum bdi_state bit;
 835
 836         bit = sync ? BDI_sync_congested : BDI_async_congested;
 837         if (!test_and_set_bit(bit, &bdi->state))
 838                 atomic_inc(&nr_bdi_congested[sync]);
 839 }
 840 EXPORT_SYMBOL(set_bdi_congested);
 841
 842 /**
 843  * congestion_wait - wait for a backing_dev to become uncongested
 844  * @sync: SYNC or ASYNC IO
 845  * @timeout: timeout in jiffies
 846  *
 847  * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit
 848  * write congestion.  If no backing_devs are congested then just wait for the
 849  * next write to be completed.
 850  */
 851 long congestion_wait(int sync, long timeout)
 852 {
 853         long ret;
 854         unsigned long start = jiffies;
 855         DEFINE_WAIT(wait);
 856         wait_queue_head_t *wqh = &congestion_wqh[sync];
 857
 858         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 859         ret = io_schedule_timeout(timeout);
 860         finish_wait(wqh, &wait);
 861
 862         trace_writeback_congestion_wait(jiffies_to_usecs(timeout),
 863                                         jiffies_to_usecs(jiffies - start));
 864
 865         return ret;
 866 }
 867 EXPORT_SYMBOL(congestion_wait);
 868
 869 /**
 870  * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a zone to complete writes
 871  * @zone: A zone to check if it is heavily congested
 872  * @sync: SYNC or ASYNC IO
 873  * @timeout: timeout in jiffies
 874  *
 875  * In the event of a congested backing_dev (any backing_dev) and the given
 876  * @zone has experienced recent congestion, this waits for up to @timeout
 877  * jiffies for either a BDI to exit congestion of the given @sync queue
 878  * or a write to complete.
 879  *
 880  * In the absence of zone congestion, cond_resched() is called to yield
 881  * the processor if necessary but otherwise does not sleep.
 882  *
 883  * The return value is 0 if the sleep is for the full timeout. Otherwise,
 884  * it is the number of jiffies that were still remaining when the function
 885  * returned. return_value == timeout implies the function did not sleep.
 886  */
 887 long wait_iff_congested(struct zone *zone, int sync, long timeout)
 888 {
 889         long ret;
 890         unsigned long start = jiffies;
 891         DEFINE_WAIT(wait);
 892         wait_queue_head_t *wqh = &congestion_wqh[sync];
 893
 894         /*
 895          * If there is no congestion, or heavy congestion is not being
 896          * encountered in the current zone, yield if necessary instead
 897          * of sleeping on the congestion queue
 898          */
 899         if (atomic_read(&nr_bdi_congested[sync]) == 0 ||
 900                         !zone_is_reclaim_congested(zone)) {
 901                 cond_resched();
 902
 903                 /* In case we scheduled, work out time remaining */
 904                 ret = timeout - (jiffies - start);
 905                 if (ret < 0)
 906                         ret = 0;
 907
 908                 goto out;
 909         }
 910
 911         /* Sleep until uncongested or a write happens */
 912         prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE);
 913         ret = io_schedule_timeout(timeout);
 914         finish_wait(wqh, &wait);
 915
 916 out:
 917         trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout),
 918                                         jiffies_to_usecs(jiffies - start));
 919
 920         return ret;
 921 }
 922 EXPORT_SYMBOL(wait_iff_congested);
 923
 924 int pdflush_proc_obsolete(struct ctl_table *table, int write,
 925                         void __user *buffer, size_t *lenp, loff_t *ppos)
 926 {
 927         char kbuf[] = "0\n";
 928
 929         if (*ppos) {
 930                 *lenp = 0;
 931                 return 0;
 932         }
 933
 934         if (copy_to_user(buffer, kbuf, sizeof(kbuf)))
 935                 return -EFAULT;
 936         printk_once(KERN_WARNING "%s exported in /proc is scheduled for removal\n",
 937                         table->procname);
 938
 939         *lenp = 2;
 940         *ppos += *lenp;
 941         return 2;
 942 }