1 // SPDX-License-Identifier: GPL-2.0-only 2 3 #include <linux/wait.h> 4 #include <linux/rbtree.h> 5 #include <linux/kthread.h> 6 #include <linux/backing-dev.h> 7 #include <linux/blk-cgroup.h> 8 #include <linux/freezer.h> 9 #include <linux/fs.h> 10 #include <linux/pagemap.h> 11 #include <linux/mm.h> 12 #include <linux/sched/mm.h> 13 #include <linux/sched.h> 14 #include <linux/module.h> 15 #include <linux/writeback.h> 16 #include <linux/device.h> 17 #include <trace/events/writeback.h> 18 19 struct backing_dev_info noop_backing_dev_info; 20 EXPORT_SYMBOL_GPL(noop_backing_dev_info); 21 22 static struct class *bdi_class; 23 static const char *bdi_unknown_name = "(unknown)"; 24 25 /* 26 * bdi_lock protects bdi_tree and updates to bdi_list. bdi_list has RCU 27 * reader side locking. 28 */ 29 DEFINE_SPINLOCK(bdi_lock); 30 static u64 bdi_id_cursor; 31 static struct rb_root bdi_tree = RB_ROOT; 32 LIST_HEAD(bdi_list); 33 34 /* bdi_wq serves all asynchronous writeback tasks */ 35 struct workqueue_struct *bdi_wq; 36 37 #define K(x) ((x) << (PAGE_SHIFT - 10)) 38 39 #ifdef CONFIG_DEBUG_FS 40 #include <linux/debugfs.h> 41 #include <linux/seq_file.h> 42 43 static struct dentry *bdi_debug_root; 44 45 static void bdi_debug_init(void) 46 { 47 bdi_debug_root = debugfs_create_dir("bdi", NULL); 48 } 49 50 static int bdi_debug_stats_show(struct seq_file *m, void *v) 51 { 52 struct backing_dev_info *bdi = m->private; 53 struct bdi_writeback *wb = &bdi->wb; 54 unsigned long background_thresh; 55 unsigned long dirty_thresh; 56 unsigned long wb_thresh; 57 unsigned long nr_dirty, nr_io, nr_more_io, nr_dirty_time; 58 struct inode *inode; 59 60 nr_dirty = nr_io = nr_more_io = nr_dirty_time = 0; 61 spin_lock(&wb->list_lock); 62 list_for_each_entry(inode, &wb->b_dirty, i_io_list) 63 nr_dirty++; 64 list_for_each_entry(inode, &wb->b_io, i_io_list) 65 nr_io++; 66 list_for_each_entry(inode, &wb->b_more_io, i_io_list) 67 nr_more_io++; 68 list_for_each_entry(inode, &wb->b_dirty_time, i_io_list) 69 if (inode->i_state & I_DIRTY_TIME) 70 nr_dirty_time++; 71 spin_unlock(&wb->list_lock); 72 73 global_dirty_limits(&background_thresh, &dirty_thresh); 74 wb_thresh = wb_calc_thresh(wb, dirty_thresh); 75 76 seq_printf(m, 77 "BdiWriteback: %10lu kB\n" 78 "BdiReclaimable: %10lu kB\n" 79 "BdiDirtyThresh: %10lu kB\n" 80 "DirtyThresh: %10lu kB\n" 81 "BackgroundThresh: %10lu kB\n" 82 "BdiDirtied: %10lu kB\n" 83 "BdiWritten: %10lu kB\n" 84 "BdiWriteBandwidth: %10lu kBps\n" 85 "b_dirty: %10lu\n" 86 "b_io: %10lu\n" 87 "b_more_io: %10lu\n" 88 "b_dirty_time: %10lu\n" 89 "bdi_list: %10u\n" 90 "state: %10lx\n", 91 (unsigned long) K(wb_stat(wb, WB_WRITEBACK)), 92 (unsigned long) K(wb_stat(wb, WB_RECLAIMABLE)), 93 K(wb_thresh), 94 K(dirty_thresh), 95 K(background_thresh), 96 (unsigned long) K(wb_stat(wb, WB_DIRTIED)), 97 (unsigned long) K(wb_stat(wb, WB_WRITTEN)), 98 (unsigned long) K(wb->write_bandwidth), 99 nr_dirty, 100 nr_io, 101 nr_more_io, 102 nr_dirty_time, 103 !list_empty(&bdi->bdi_list), bdi->wb.state); 104 105 return 0; 106 } 107 DEFINE_SHOW_ATTRIBUTE(bdi_debug_stats); 108 109 static void bdi_debug_register(struct backing_dev_info *bdi, const char *name) 110 { 111 bdi->debug_dir = debugfs_create_dir(name, bdi_debug_root); 112 113 debugfs_create_file("stats", 0444, bdi->debug_dir, bdi, 114 &bdi_debug_stats_fops); 115 } 116 117 static void bdi_debug_unregister(struct backing_dev_info *bdi) 118 { 119 debugfs_remove_recursive(bdi->debug_dir); 120 } 121 #else 122 static inline void bdi_debug_init(void) 123 { 124 } 125 static inline void bdi_debug_register(struct backing_dev_info *bdi, 126 const char *name) 127 { 128 } 129 static inline void bdi_debug_unregister(struct backing_dev_info *bdi) 130 { 131 } 132 #endif 133 134 static ssize_t read_ahead_kb_store(struct device *dev, 135 struct device_attribute *attr, 136 const char *buf, size_t count) 137 { 138 struct backing_dev_info *bdi = dev_get_drvdata(dev); 139 unsigned long read_ahead_kb; 140 ssize_t ret; 141 142 ret = kstrtoul(buf, 10, &read_ahead_kb); 143 if (ret < 0) 144 return ret; 145 146 bdi->ra_pages = read_ahead_kb >> (PAGE_SHIFT - 10); 147 148 return count; 149 } 150 151 #define BDI_SHOW(name, expr) \ 152 static ssize_t name##_show(struct device *dev, \ 153 struct device_attribute *attr, char *buf) \ 154 { \ 155 struct backing_dev_info *bdi = dev_get_drvdata(dev); \ 156 \ 157 return sysfs_emit(buf, "%lld\n", (long long)expr); \ 158 } \ 159 static DEVICE_ATTR_RW(name); 160 161 BDI_SHOW(read_ahead_kb, K(bdi->ra_pages)) 162 163 static ssize_t min_ratio_store(struct device *dev, 164 struct device_attribute *attr, const char *buf, size_t count) 165 { 166 struct backing_dev_info *bdi = dev_get_drvdata(dev); 167 unsigned int ratio; 168 ssize_t ret; 169 170 ret = kstrtouint(buf, 10, &ratio); 171 if (ret < 0) 172 return ret; 173 174 ret = bdi_set_min_ratio(bdi, ratio); 175 if (!ret) 176 ret = count; 177 178 return ret; 179 } 180 BDI_SHOW(min_ratio, bdi->min_ratio) 181 182 static ssize_t max_ratio_store(struct device *dev, 183 struct device_attribute *attr, const char *buf, size_t count) 184 { 185 struct backing_dev_info *bdi = dev_get_drvdata(dev); 186 unsigned int ratio; 187 ssize_t ret; 188 189 ret = kstrtouint(buf, 10, &ratio); 190 if (ret < 0) 191 return ret; 192 193 ret = bdi_set_max_ratio(bdi, ratio); 194 if (!ret) 195 ret = count; 196 197 return ret; 198 } 199 BDI_SHOW(max_ratio, bdi->max_ratio) 200 201 static ssize_t stable_pages_required_show(struct device *dev, 202 struct device_attribute *attr, 203 char *buf) 204 { 205 dev_warn_once(dev, 206 "the stable_pages_required attribute has been removed. Use the stable_writes queue attribute instead.\n"); 207 return sysfs_emit(buf, "%d\n", 0); 208 } 209 static DEVICE_ATTR_RO(stable_pages_required); 210 211 static struct attribute *bdi_dev_attrs[] = { 212 &dev_attr_read_ahead_kb.attr, 213 &dev_attr_min_ratio.attr, 214 &dev_attr_max_ratio.attr, 215 &dev_attr_stable_pages_required.attr, 216 NULL, 217 }; 218 ATTRIBUTE_GROUPS(bdi_dev); 219 220 static __init int bdi_class_init(void) 221 { 222 bdi_class = class_create(THIS_MODULE, "bdi"); 223 if (IS_ERR(bdi_class)) 224 return PTR_ERR(bdi_class); 225 226 bdi_class->dev_groups = bdi_dev_groups; 227 bdi_debug_init(); 228 229 return 0; 230 } 231 postcore_initcall(bdi_class_init); 232 233 static int bdi_init(struct backing_dev_info *bdi); 234 235 static int __init default_bdi_init(void) 236 { 237 int err; 238 239 bdi_wq = alloc_workqueue("writeback", WQ_MEM_RECLAIM | WQ_UNBOUND | 240 WQ_SYSFS, 0); 241 if (!bdi_wq) 242 return -ENOMEM; 243 244 err = bdi_init(&noop_backing_dev_info); 245 246 return err; 247 } 248 subsys_initcall(default_bdi_init); 249 250 /* 251 * This function is used when the first inode for this wb is marked dirty. It 252 * wakes-up the corresponding bdi thread which should then take care of the 253 * periodic background write-out of dirty inodes. Since the write-out would 254 * starts only 'dirty_writeback_interval' centisecs from now anyway, we just 255 * set up a timer which wakes the bdi thread up later. 256 * 257 * Note, we wouldn't bother setting up the timer, but this function is on the 258 * fast-path (used by '__mark_inode_dirty()'), so we save few context switches 259 * by delaying the wake-up. 260 * 261 * We have to be careful not to postpone flush work if it is scheduled for 262 * earlier. Thus we use queue_delayed_work(). 263 */ 264 void wb_wakeup_delayed(struct bdi_writeback *wb) 265 { 266 unsigned long timeout; 267 268 timeout = msecs_to_jiffies(dirty_writeback_interval * 10); 269 spin_lock_bh(&wb->work_lock); 270 if (test_bit(WB_registered, &wb->state)) 271 queue_delayed_work(bdi_wq, &wb->dwork, timeout); 272 spin_unlock_bh(&wb->work_lock); 273 } 274 275 static void wb_update_bandwidth_workfn(struct work_struct *work) 276 { 277 struct bdi_writeback *wb = container_of(to_delayed_work(work), 278 struct bdi_writeback, bw_dwork); 279 280 wb_update_bandwidth(wb); 281 } 282 283 /* 284 * Initial write bandwidth: 100 MB/s 285 */ 286 #define INIT_BW (100 << (20 - PAGE_SHIFT)) 287 288 static int wb_init(struct bdi_writeback *wb, struct backing_dev_info *bdi, 289 gfp_t gfp) 290 { 291 int i, err; 292 293 memset(wb, 0, sizeof(*wb)); 294 295 if (wb != &bdi->wb) 296 bdi_get(bdi); 297 wb->bdi = bdi; 298 wb->last_old_flush = jiffies; 299 INIT_LIST_HEAD(&wb->b_dirty); 300 INIT_LIST_HEAD(&wb->b_io); 301 INIT_LIST_HEAD(&wb->b_more_io); 302 INIT_LIST_HEAD(&wb->b_dirty_time); 303 spin_lock_init(&wb->list_lock); 304 305 atomic_set(&wb->writeback_inodes, 0); 306 wb->bw_time_stamp = jiffies; 307 wb->balanced_dirty_ratelimit = INIT_BW; 308 wb->dirty_ratelimit = INIT_BW; 309 wb->write_bandwidth = INIT_BW; 310 wb->avg_write_bandwidth = INIT_BW; 311 312 spin_lock_init(&wb->work_lock); 313 INIT_LIST_HEAD(&wb->work_list); 314 INIT_DELAYED_WORK(&wb->dwork, wb_workfn); 315 INIT_DELAYED_WORK(&wb->bw_dwork, wb_update_bandwidth_workfn); 316 wb->dirty_sleep = jiffies; 317 318 err = fprop_local_init_percpu(&wb->completions, gfp); 319 if (err) 320 goto out_put_bdi; 321 322 for (i = 0; i < NR_WB_STAT_ITEMS; i++) { 323 err = percpu_counter_init(&wb->stat[i], 0, gfp); 324 if (err) 325 goto out_destroy_stat; 326 } 327 328 return 0; 329 330 out_destroy_stat: 331 while (i--) 332 percpu_counter_destroy(&wb->stat[i]); 333 fprop_local_destroy_percpu(&wb->completions); 334 out_put_bdi: 335 if (wb != &bdi->wb) 336 bdi_put(bdi); 337 return err; 338 } 339 340 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb); 341 342 /* 343 * Remove bdi from the global list and shutdown any threads we have running 344 */ 345 static void wb_shutdown(struct bdi_writeback *wb) 346 { 347 /* Make sure nobody queues further work */ 348 spin_lock_bh(&wb->work_lock); 349 if (!test_and_clear_bit(WB_registered, &wb->state)) { 350 spin_unlock_bh(&wb->work_lock); 351 return; 352 } 353 spin_unlock_bh(&wb->work_lock); 354 355 cgwb_remove_from_bdi_list(wb); 356 /* 357 * Drain work list and shutdown the delayed_work. !WB_registered 358 * tells wb_workfn() that @wb is dying and its work_list needs to 359 * be drained no matter what. 360 */ 361 mod_delayed_work(bdi_wq, &wb->dwork, 0); 362 flush_delayed_work(&wb->dwork); 363 WARN_ON(!list_empty(&wb->work_list)); 364 flush_delayed_work(&wb->bw_dwork); 365 } 366 367 static void wb_exit(struct bdi_writeback *wb) 368 { 369 int i; 370 371 WARN_ON(delayed_work_pending(&wb->dwork)); 372 373 for (i = 0; i < NR_WB_STAT_ITEMS; i++) 374 percpu_counter_destroy(&wb->stat[i]); 375 376 fprop_local_destroy_percpu(&wb->completions); 377 if (wb != &wb->bdi->wb) 378 bdi_put(wb->bdi); 379 } 380 381 #ifdef CONFIG_CGROUP_WRITEBACK 382 383 #include <linux/memcontrol.h> 384 385 /* 386 * cgwb_lock protects bdi->cgwb_tree, blkcg->cgwb_list, offline_cgwbs and 387 * memcg->cgwb_list. bdi->cgwb_tree is also RCU protected. 388 */ 389 static DEFINE_SPINLOCK(cgwb_lock); 390 static struct workqueue_struct *cgwb_release_wq; 391 392 static LIST_HEAD(offline_cgwbs); 393 static void cleanup_offline_cgwbs_workfn(struct work_struct *work); 394 static DECLARE_WORK(cleanup_offline_cgwbs_work, cleanup_offline_cgwbs_workfn); 395 396 static void cgwb_release_workfn(struct work_struct *work) 397 { 398 struct bdi_writeback *wb = container_of(work, struct bdi_writeback, 399 release_work); 400 struct blkcg *blkcg = css_to_blkcg(wb->blkcg_css); 401 402 mutex_lock(&wb->bdi->cgwb_release_mutex); 403 wb_shutdown(wb); 404 405 css_put(wb->memcg_css); 406 css_put(wb->blkcg_css); 407 mutex_unlock(&wb->bdi->cgwb_release_mutex); 408 409 /* triggers blkg destruction if no online users left */ 410 blkcg_unpin_online(blkcg); 411 412 fprop_local_destroy_percpu(&wb->memcg_completions); 413 414 spin_lock_irq(&cgwb_lock); 415 list_del(&wb->offline_node); 416 spin_unlock_irq(&cgwb_lock); 417 418 percpu_ref_exit(&wb->refcnt); 419 wb_exit(wb); 420 WARN_ON_ONCE(!list_empty(&wb->b_attached)); 421 kfree_rcu(wb, rcu); 422 } 423 424 static void cgwb_release(struct percpu_ref *refcnt) 425 { 426 struct bdi_writeback *wb = container_of(refcnt, struct bdi_writeback, 427 refcnt); 428 queue_work(cgwb_release_wq, &wb->release_work); 429 } 430 431 static void cgwb_kill(struct bdi_writeback *wb) 432 { 433 lockdep_assert_held(&cgwb_lock); 434 435 WARN_ON(!radix_tree_delete(&wb->bdi->cgwb_tree, wb->memcg_css->id)); 436 list_del(&wb->memcg_node); 437 list_del(&wb->blkcg_node); 438 list_add(&wb->offline_node, &offline_cgwbs); 439 percpu_ref_kill(&wb->refcnt); 440 } 441 442 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb) 443 { 444 spin_lock_irq(&cgwb_lock); 445 list_del_rcu(&wb->bdi_node); 446 spin_unlock_irq(&cgwb_lock); 447 } 448 449 static int cgwb_create(struct backing_dev_info *bdi, 450 struct cgroup_subsys_state *memcg_css, gfp_t gfp) 451 { 452 struct mem_cgroup *memcg; 453 struct cgroup_subsys_state *blkcg_css; 454 struct blkcg *blkcg; 455 struct list_head *memcg_cgwb_list, *blkcg_cgwb_list; 456 struct bdi_writeback *wb; 457 unsigned long flags; 458 int ret = 0; 459 460 memcg = mem_cgroup_from_css(memcg_css); 461 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys); 462 blkcg = css_to_blkcg(blkcg_css); 463 memcg_cgwb_list = &memcg->cgwb_list; 464 blkcg_cgwb_list = &blkcg->cgwb_list; 465 466 /* look up again under lock and discard on blkcg mismatch */ 467 spin_lock_irqsave(&cgwb_lock, flags); 468 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id); 469 if (wb && wb->blkcg_css != blkcg_css) { 470 cgwb_kill(wb); 471 wb = NULL; 472 } 473 spin_unlock_irqrestore(&cgwb_lock, flags); 474 if (wb) 475 goto out_put; 476 477 /* need to create a new one */ 478 wb = kmalloc(sizeof(*wb), gfp); 479 if (!wb) { 480 ret = -ENOMEM; 481 goto out_put; 482 } 483 484 ret = wb_init(wb, bdi, gfp); 485 if (ret) 486 goto err_free; 487 488 ret = percpu_ref_init(&wb->refcnt, cgwb_release, 0, gfp); 489 if (ret) 490 goto err_wb_exit; 491 492 ret = fprop_local_init_percpu(&wb->memcg_completions, gfp); 493 if (ret) 494 goto err_ref_exit; 495 496 wb->memcg_css = memcg_css; 497 wb->blkcg_css = blkcg_css; 498 INIT_LIST_HEAD(&wb->b_attached); 499 INIT_WORK(&wb->release_work, cgwb_release_workfn); 500 set_bit(WB_registered, &wb->state); 501 502 /* 503 * The root wb determines the registered state of the whole bdi and 504 * memcg_cgwb_list and blkcg_cgwb_list's next pointers indicate 505 * whether they're still online. Don't link @wb if any is dead. 506 * See wb_memcg_offline() and wb_blkcg_offline(). 507 */ 508 ret = -ENODEV; 509 spin_lock_irqsave(&cgwb_lock, flags); 510 if (test_bit(WB_registered, &bdi->wb.state) && 511 blkcg_cgwb_list->next && memcg_cgwb_list->next) { 512 /* we might have raced another instance of this function */ 513 ret = radix_tree_insert(&bdi->cgwb_tree, memcg_css->id, wb); 514 if (!ret) { 515 list_add_tail_rcu(&wb->bdi_node, &bdi->wb_list); 516 list_add(&wb->memcg_node, memcg_cgwb_list); 517 list_add(&wb->blkcg_node, blkcg_cgwb_list); 518 blkcg_pin_online(blkcg); 519 css_get(memcg_css); 520 css_get(blkcg_css); 521 } 522 } 523 spin_unlock_irqrestore(&cgwb_lock, flags); 524 if (ret) { 525 if (ret == -EEXIST) 526 ret = 0; 527 goto err_fprop_exit; 528 } 529 goto out_put; 530 531 err_fprop_exit: 532 fprop_local_destroy_percpu(&wb->memcg_completions); 533 err_ref_exit: 534 percpu_ref_exit(&wb->refcnt); 535 err_wb_exit: 536 wb_exit(wb); 537 err_free: 538 kfree(wb); 539 out_put: 540 css_put(blkcg_css); 541 return ret; 542 } 543 544 /** 545 * wb_get_lookup - get wb for a given memcg 546 * @bdi: target bdi 547 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref) 548 * 549 * Try to get the wb for @memcg_css on @bdi. The returned wb has its 550 * refcount incremented. 551 * 552 * This function uses css_get() on @memcg_css and thus expects its refcnt 553 * to be positive on invocation. IOW, rcu_read_lock() protection on 554 * @memcg_css isn't enough. try_get it before calling this function. 555 * 556 * A wb is keyed by its associated memcg. As blkcg implicitly enables 557 * memcg on the default hierarchy, memcg association is guaranteed to be 558 * more specific (equal or descendant to the associated blkcg) and thus can 559 * identify both the memcg and blkcg associations. 560 * 561 * Because the blkcg associated with a memcg may change as blkcg is enabled 562 * and disabled closer to root in the hierarchy, each wb keeps track of 563 * both the memcg and blkcg associated with it and verifies the blkcg on 564 * each lookup. On mismatch, the existing wb is discarded and a new one is 565 * created. 566 */ 567 struct bdi_writeback *wb_get_lookup(struct backing_dev_info *bdi, 568 struct cgroup_subsys_state *memcg_css) 569 { 570 struct bdi_writeback *wb; 571 572 if (!memcg_css->parent) 573 return &bdi->wb; 574 575 rcu_read_lock(); 576 wb = radix_tree_lookup(&bdi->cgwb_tree, memcg_css->id); 577 if (wb) { 578 struct cgroup_subsys_state *blkcg_css; 579 580 /* see whether the blkcg association has changed */ 581 blkcg_css = cgroup_get_e_css(memcg_css->cgroup, &io_cgrp_subsys); 582 if (unlikely(wb->blkcg_css != blkcg_css || !wb_tryget(wb))) 583 wb = NULL; 584 css_put(blkcg_css); 585 } 586 rcu_read_unlock(); 587 588 return wb; 589 } 590 591 /** 592 * wb_get_create - get wb for a given memcg, create if necessary 593 * @bdi: target bdi 594 * @memcg_css: cgroup_subsys_state of the target memcg (must have positive ref) 595 * @gfp: allocation mask to use 596 * 597 * Try to get the wb for @memcg_css on @bdi. If it doesn't exist, try to 598 * create one. See wb_get_lookup() for more details. 599 */ 600 struct bdi_writeback *wb_get_create(struct backing_dev_info *bdi, 601 struct cgroup_subsys_state *memcg_css, 602 gfp_t gfp) 603 { 604 struct bdi_writeback *wb; 605 606 might_alloc(gfp); 607 608 if (!memcg_css->parent) 609 return &bdi->wb; 610 611 do { 612 wb = wb_get_lookup(bdi, memcg_css); 613 } while (!wb && !cgwb_create(bdi, memcg_css, gfp)); 614 615 return wb; 616 } 617 618 static int cgwb_bdi_init(struct backing_dev_info *bdi) 619 { 620 int ret; 621 622 INIT_RADIX_TREE(&bdi->cgwb_tree, GFP_ATOMIC); 623 mutex_init(&bdi->cgwb_release_mutex); 624 init_rwsem(&bdi->wb_switch_rwsem); 625 626 ret = wb_init(&bdi->wb, bdi, GFP_KERNEL); 627 if (!ret) { 628 bdi->wb.memcg_css = &root_mem_cgroup->css; 629 bdi->wb.blkcg_css = blkcg_root_css; 630 } 631 return ret; 632 } 633 634 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) 635 { 636 struct radix_tree_iter iter; 637 void **slot; 638 struct bdi_writeback *wb; 639 640 WARN_ON(test_bit(WB_registered, &bdi->wb.state)); 641 642 spin_lock_irq(&cgwb_lock); 643 radix_tree_for_each_slot(slot, &bdi->cgwb_tree, &iter, 0) 644 cgwb_kill(*slot); 645 spin_unlock_irq(&cgwb_lock); 646 647 mutex_lock(&bdi->cgwb_release_mutex); 648 spin_lock_irq(&cgwb_lock); 649 while (!list_empty(&bdi->wb_list)) { 650 wb = list_first_entry(&bdi->wb_list, struct bdi_writeback, 651 bdi_node); 652 spin_unlock_irq(&cgwb_lock); 653 wb_shutdown(wb); 654 spin_lock_irq(&cgwb_lock); 655 } 656 spin_unlock_irq(&cgwb_lock); 657 mutex_unlock(&bdi->cgwb_release_mutex); 658 } 659 660 /* 661 * cleanup_offline_cgwbs_workfn - try to release dying cgwbs 662 * 663 * Try to release dying cgwbs by switching attached inodes to the nearest 664 * living ancestor's writeback. Processed wbs are placed at the end 665 * of the list to guarantee the forward progress. 666 */ 667 static void cleanup_offline_cgwbs_workfn(struct work_struct *work) 668 { 669 struct bdi_writeback *wb; 670 LIST_HEAD(processed); 671 672 spin_lock_irq(&cgwb_lock); 673 674 while (!list_empty(&offline_cgwbs)) { 675 wb = list_first_entry(&offline_cgwbs, struct bdi_writeback, 676 offline_node); 677 list_move(&wb->offline_node, &processed); 678 679 /* 680 * If wb is dirty, cleaning up the writeback by switching 681 * attached inodes will result in an effective removal of any 682 * bandwidth restrictions, which isn't the goal. Instead, 683 * it can be postponed until the next time, when all io 684 * will be likely completed. If in the meantime some inodes 685 * will get re-dirtied, they should be eventually switched to 686 * a new cgwb. 687 */ 688 if (wb_has_dirty_io(wb)) 689 continue; 690 691 if (!wb_tryget(wb)) 692 continue; 693 694 spin_unlock_irq(&cgwb_lock); 695 while (cleanup_offline_cgwb(wb)) 696 cond_resched(); 697 spin_lock_irq(&cgwb_lock); 698 699 wb_put(wb); 700 } 701 702 if (!list_empty(&processed)) 703 list_splice_tail(&processed, &offline_cgwbs); 704 705 spin_unlock_irq(&cgwb_lock); 706 } 707 708 /** 709 * wb_memcg_offline - kill all wb's associated with a memcg being offlined 710 * @memcg: memcg being offlined 711 * 712 * Also prevents creation of any new wb's associated with @memcg. 713 */ 714 void wb_memcg_offline(struct mem_cgroup *memcg) 715 { 716 struct list_head *memcg_cgwb_list = &memcg->cgwb_list; 717 struct bdi_writeback *wb, *next; 718 719 spin_lock_irq(&cgwb_lock); 720 list_for_each_entry_safe(wb, next, memcg_cgwb_list, memcg_node) 721 cgwb_kill(wb); 722 memcg_cgwb_list->next = NULL; /* prevent new wb's */ 723 spin_unlock_irq(&cgwb_lock); 724 725 queue_work(system_unbound_wq, &cleanup_offline_cgwbs_work); 726 } 727 728 /** 729 * wb_blkcg_offline - kill all wb's associated with a blkcg being offlined 730 * @blkcg: blkcg being offlined 731 * 732 * Also prevents creation of any new wb's associated with @blkcg. 733 */ 734 void wb_blkcg_offline(struct blkcg *blkcg) 735 { 736 struct bdi_writeback *wb, *next; 737 738 spin_lock_irq(&cgwb_lock); 739 list_for_each_entry_safe(wb, next, &blkcg->cgwb_list, blkcg_node) 740 cgwb_kill(wb); 741 blkcg->cgwb_list.next = NULL; /* prevent new wb's */ 742 spin_unlock_irq(&cgwb_lock); 743 } 744 745 static void cgwb_bdi_register(struct backing_dev_info *bdi) 746 { 747 spin_lock_irq(&cgwb_lock); 748 list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list); 749 spin_unlock_irq(&cgwb_lock); 750 } 751 752 static int __init cgwb_init(void) 753 { 754 /* 755 * There can be many concurrent release work items overwhelming 756 * system_wq. Put them in a separate wq and limit concurrency. 757 * There's no point in executing many of these in parallel. 758 */ 759 cgwb_release_wq = alloc_workqueue("cgwb_release", 0, 1); 760 if (!cgwb_release_wq) 761 return -ENOMEM; 762 763 return 0; 764 } 765 subsys_initcall(cgwb_init); 766 767 #else /* CONFIG_CGROUP_WRITEBACK */ 768 769 static int cgwb_bdi_init(struct backing_dev_info *bdi) 770 { 771 return wb_init(&bdi->wb, bdi, GFP_KERNEL); 772 } 773 774 static void cgwb_bdi_unregister(struct backing_dev_info *bdi) { } 775 776 static void cgwb_bdi_register(struct backing_dev_info *bdi) 777 { 778 list_add_tail_rcu(&bdi->wb.bdi_node, &bdi->wb_list); 779 } 780 781 static void cgwb_remove_from_bdi_list(struct bdi_writeback *wb) 782 { 783 list_del_rcu(&wb->bdi_node); 784 } 785 786 #endif /* CONFIG_CGROUP_WRITEBACK */ 787 788 static int bdi_init(struct backing_dev_info *bdi) 789 { 790 int ret; 791 792 bdi->dev = NULL; 793 794 kref_init(&bdi->refcnt); 795 bdi->min_ratio = 0; 796 bdi->max_ratio = 100; 797 bdi->max_prop_frac = FPROP_FRAC_BASE; 798 INIT_LIST_HEAD(&bdi->bdi_list); 799 INIT_LIST_HEAD(&bdi->wb_list); 800 init_waitqueue_head(&bdi->wb_waitq); 801 802 ret = cgwb_bdi_init(bdi); 803 804 return ret; 805 } 806 807 struct backing_dev_info *bdi_alloc(int node_id) 808 { 809 struct backing_dev_info *bdi; 810 811 bdi = kzalloc_node(sizeof(*bdi), GFP_KERNEL, node_id); 812 if (!bdi) 813 return NULL; 814 815 if (bdi_init(bdi)) { 816 kfree(bdi); 817 return NULL; 818 } 819 bdi->capabilities = BDI_CAP_WRITEBACK | BDI_CAP_WRITEBACK_ACCT; 820 bdi->ra_pages = VM_READAHEAD_PAGES; 821 bdi->io_pages = VM_READAHEAD_PAGES; 822 timer_setup(&bdi->laptop_mode_wb_timer, laptop_mode_timer_fn, 0); 823 return bdi; 824 } 825 EXPORT_SYMBOL(bdi_alloc); 826 827 static struct rb_node **bdi_lookup_rb_node(u64 id, struct rb_node **parentp) 828 { 829 struct rb_node **p = &bdi_tree.rb_node; 830 struct rb_node *parent = NULL; 831 struct backing_dev_info *bdi; 832 833 lockdep_assert_held(&bdi_lock); 834 835 while (*p) { 836 parent = *p; 837 bdi = rb_entry(parent, struct backing_dev_info, rb_node); 838 839 if (bdi->id > id) 840 p = &(*p)->rb_left; 841 else if (bdi->id < id) 842 p = &(*p)->rb_right; 843 else 844 break; 845 } 846 847 if (parentp) 848 *parentp = parent; 849 return p; 850 } 851 852 /** 853 * bdi_get_by_id - lookup and get bdi from its id 854 * @id: bdi id to lookup 855 * 856 * Find bdi matching @id and get it. Returns NULL if the matching bdi 857 * doesn't exist or is already unregistered. 858 */ 859 struct backing_dev_info *bdi_get_by_id(u64 id) 860 { 861 struct backing_dev_info *bdi = NULL; 862 struct rb_node **p; 863 864 spin_lock_bh(&bdi_lock); 865 p = bdi_lookup_rb_node(id, NULL); 866 if (*p) { 867 bdi = rb_entry(*p, struct backing_dev_info, rb_node); 868 bdi_get(bdi); 869 } 870 spin_unlock_bh(&bdi_lock); 871 872 return bdi; 873 } 874 875 int bdi_register_va(struct backing_dev_info *bdi, const char *fmt, va_list args) 876 { 877 struct device *dev; 878 struct rb_node *parent, **p; 879 880 if (bdi->dev) /* The driver needs to use separate queues per device */ 881 return 0; 882 883 vsnprintf(bdi->dev_name, sizeof(bdi->dev_name), fmt, args); 884 dev = device_create(bdi_class, NULL, MKDEV(0, 0), bdi, bdi->dev_name); 885 if (IS_ERR(dev)) 886 return PTR_ERR(dev); 887 888 cgwb_bdi_register(bdi); 889 bdi->dev = dev; 890 891 bdi_debug_register(bdi, dev_name(dev)); 892 set_bit(WB_registered, &bdi->wb.state); 893 894 spin_lock_bh(&bdi_lock); 895 896 bdi->id = ++bdi_id_cursor; 897 898 p = bdi_lookup_rb_node(bdi->id, &parent); 899 rb_link_node(&bdi->rb_node, parent, p); 900 rb_insert_color(&bdi->rb_node, &bdi_tree); 901 902 list_add_tail_rcu(&bdi->bdi_list, &bdi_list); 903 904 spin_unlock_bh(&bdi_lock); 905 906 trace_writeback_bdi_register(bdi); 907 return 0; 908 } 909 910 int bdi_register(struct backing_dev_info *bdi, const char *fmt, ...) 911 { 912 va_list args; 913 int ret; 914 915 va_start(args, fmt); 916 ret = bdi_register_va(bdi, fmt, args); 917 va_end(args); 918 return ret; 919 } 920 EXPORT_SYMBOL(bdi_register); 921 922 void bdi_set_owner(struct backing_dev_info *bdi, struct device *owner) 923 { 924 WARN_ON_ONCE(bdi->owner); 925 bdi->owner = owner; 926 get_device(owner); 927 } 928 929 /* 930 * Remove bdi from bdi_list, and ensure that it is no longer visible 931 */ 932 static void bdi_remove_from_list(struct backing_dev_info *bdi) 933 { 934 spin_lock_bh(&bdi_lock); 935 rb_erase(&bdi->rb_node, &bdi_tree); 936 list_del_rcu(&bdi->bdi_list); 937 spin_unlock_bh(&bdi_lock); 938 939 synchronize_rcu_expedited(); 940 } 941 942 void bdi_unregister(struct backing_dev_info *bdi) 943 { 944 del_timer_sync(&bdi->laptop_mode_wb_timer); 945 946 /* make sure nobody finds us on the bdi_list anymore */ 947 bdi_remove_from_list(bdi); 948 wb_shutdown(&bdi->wb); 949 cgwb_bdi_unregister(bdi); 950 951 if (bdi->dev) { 952 bdi_debug_unregister(bdi); 953 device_unregister(bdi->dev); 954 bdi->dev = NULL; 955 } 956 957 if (bdi->owner) { 958 put_device(bdi->owner); 959 bdi->owner = NULL; 960 } 961 } 962 963 static void release_bdi(struct kref *ref) 964 { 965 struct backing_dev_info *bdi = 966 container_of(ref, struct backing_dev_info, refcnt); 967 968 if (test_bit(WB_registered, &bdi->wb.state)) 969 bdi_unregister(bdi); 970 WARN_ON_ONCE(bdi->dev); 971 wb_exit(&bdi->wb); 972 kfree(bdi); 973 } 974 975 void bdi_put(struct backing_dev_info *bdi) 976 { 977 kref_put(&bdi->refcnt, release_bdi); 978 } 979 EXPORT_SYMBOL(bdi_put); 980 981 struct backing_dev_info *inode_to_bdi(struct inode *inode) 982 { 983 struct super_block *sb; 984 985 if (!inode) 986 return &noop_backing_dev_info; 987 988 sb = inode->i_sb; 989 #ifdef CONFIG_BLOCK 990 if (sb_is_blkdev_sb(sb)) 991 return I_BDEV(inode)->bd_disk->bdi; 992 #endif 993 return sb->s_bdi; 994 } 995 EXPORT_SYMBOL(inode_to_bdi); 996 997 const char *bdi_dev_name(struct backing_dev_info *bdi) 998 { 999 if (!bdi || !bdi->dev) 1000 return bdi_unknown_name; 1001 return bdi->dev_name; 1002 } 1003 EXPORT_SYMBOL_GPL(bdi_dev_name); 1004 1005 static wait_queue_head_t congestion_wqh[2] = { 1006 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[0]), 1007 __WAIT_QUEUE_HEAD_INITIALIZER(congestion_wqh[1]) 1008 }; 1009 static atomic_t nr_wb_congested[2]; 1010 1011 void clear_bdi_congested(struct backing_dev_info *bdi, int sync) 1012 { 1013 wait_queue_head_t *wqh = &congestion_wqh[sync]; 1014 enum wb_congested_state bit; 1015 1016 bit = sync ? WB_sync_congested : WB_async_congested; 1017 if (test_and_clear_bit(bit, &bdi->wb.congested)) 1018 atomic_dec(&nr_wb_congested[sync]); 1019 smp_mb__after_atomic(); 1020 if (waitqueue_active(wqh)) 1021 wake_up(wqh); 1022 } 1023 EXPORT_SYMBOL(clear_bdi_congested); 1024 1025 void set_bdi_congested(struct backing_dev_info *bdi, int sync) 1026 { 1027 enum wb_congested_state bit; 1028 1029 bit = sync ? WB_sync_congested : WB_async_congested; 1030 if (!test_and_set_bit(bit, &bdi->wb.congested)) 1031 atomic_inc(&nr_wb_congested[sync]); 1032 } 1033 EXPORT_SYMBOL(set_bdi_congested); 1034 1035 /** 1036 * congestion_wait - wait for a backing_dev to become uncongested 1037 * @sync: SYNC or ASYNC IO 1038 * @timeout: timeout in jiffies 1039 * 1040 * Waits for up to @timeout jiffies for a backing_dev (any backing_dev) to exit 1041 * write congestion. If no backing_devs are congested then just wait for the 1042 * next write to be completed. 1043 */ 1044 long congestion_wait(int sync, long timeout) 1045 { 1046 long ret; 1047 unsigned long start = jiffies; 1048 DEFINE_WAIT(wait); 1049 wait_queue_head_t *wqh = &congestion_wqh[sync]; 1050 1051 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1052 ret = io_schedule_timeout(timeout); 1053 finish_wait(wqh, &wait); 1054 1055 trace_writeback_congestion_wait(jiffies_to_usecs(timeout), 1056 jiffies_to_usecs(jiffies - start)); 1057 1058 return ret; 1059 } 1060 EXPORT_SYMBOL(congestion_wait); 1061 1062 /** 1063 * wait_iff_congested - Conditionally wait for a backing_dev to become uncongested or a pgdat to complete writes 1064 * @sync: SYNC or ASYNC IO 1065 * @timeout: timeout in jiffies 1066 * 1067 * In the event of a congested backing_dev (any backing_dev) this waits 1068 * for up to @timeout jiffies for either a BDI to exit congestion of the 1069 * given @sync queue or a write to complete. 1070 * 1071 * The return value is 0 if the sleep is for the full timeout. Otherwise, 1072 * it is the number of jiffies that were still remaining when the function 1073 * returned. return_value == timeout implies the function did not sleep. 1074 */ 1075 long wait_iff_congested(int sync, long timeout) 1076 { 1077 long ret; 1078 unsigned long start = jiffies; 1079 DEFINE_WAIT(wait); 1080 wait_queue_head_t *wqh = &congestion_wqh[sync]; 1081 1082 /* 1083 * If there is no congestion, yield if necessary instead 1084 * of sleeping on the congestion queue 1085 */ 1086 if (atomic_read(&nr_wb_congested[sync]) == 0) { 1087 cond_resched(); 1088 1089 /* In case we scheduled, work out time remaining */ 1090 ret = timeout - (jiffies - start); 1091 if (ret < 0) 1092 ret = 0; 1093 1094 goto out; 1095 } 1096 1097 /* Sleep until uncongested or a write happens */ 1098 prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1099 ret = io_schedule_timeout(timeout); 1100 finish_wait(wqh, &wait); 1101 1102 out: 1103 trace_writeback_wait_iff_congested(jiffies_to_usecs(timeout), 1104 jiffies_to_usecs(jiffies - start)); 1105 1106 return ret; 1107 } 1108 EXPORT_SYMBOL(wait_iff_congested); 1109