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