1457c8996SThomas Gleixner // SPDX-License-Identifier: GPL-2.0-only 21da177e4SLinus Torvalds /* 31da177e4SLinus Torvalds * fs/fs-writeback.c 41da177e4SLinus Torvalds * 51da177e4SLinus Torvalds * Copyright (C) 2002, Linus Torvalds. 61da177e4SLinus Torvalds * 71da177e4SLinus Torvalds * Contains all the functions related to writing back and waiting 81da177e4SLinus Torvalds * upon dirty inodes against superblocks, and writing back dirty 91da177e4SLinus Torvalds * pages against inodes. ie: data writeback. Writeout of the 101da177e4SLinus Torvalds * inode itself is not handled here. 111da177e4SLinus Torvalds * 12e1f8e874SFrancois Cami * 10Apr2002 Andrew Morton 131da177e4SLinus Torvalds * Split out of fs/inode.c 141da177e4SLinus Torvalds * Additions for address_space-based writeback 151da177e4SLinus Torvalds */ 161da177e4SLinus Torvalds 171da177e4SLinus Torvalds #include <linux/kernel.h> 18630d9c47SPaul Gortmaker #include <linux/export.h> 191da177e4SLinus Torvalds #include <linux/spinlock.h> 205a0e3ad6STejun Heo #include <linux/slab.h> 211da177e4SLinus Torvalds #include <linux/sched.h> 221da177e4SLinus Torvalds #include <linux/fs.h> 231da177e4SLinus Torvalds #include <linux/mm.h> 24bc31b86aSWu Fengguang #include <linux/pagemap.h> 2503ba3782SJens Axboe #include <linux/kthread.h> 261da177e4SLinus Torvalds #include <linux/writeback.h> 271da177e4SLinus Torvalds #include <linux/blkdev.h> 281da177e4SLinus Torvalds #include <linux/backing-dev.h> 29455b2864SDave Chinner #include <linux/tracepoint.h> 30719ea2fbSAl Viro #include <linux/device.h> 3121c6321fSTejun Heo #include <linux/memcontrol.h> 3207f3f05cSDavid Howells #include "internal.h" 331da177e4SLinus Torvalds 34d0bceac7SJens Axboe /* 35bc31b86aSWu Fengguang * 4MB minimal write chunk size 36bc31b86aSWu Fengguang */ 3709cbfeafSKirill A. Shutemov #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_SHIFT - 10)) 38bc31b86aSWu Fengguang 39cc395d7fSTejun Heo struct wb_completion { 40cc395d7fSTejun Heo atomic_t cnt; 41cc395d7fSTejun Heo }; 42cc395d7fSTejun Heo 43bc31b86aSWu Fengguang /* 44c4a77a6cSJens Axboe * Passed into wb_writeback(), essentially a subset of writeback_control 45c4a77a6cSJens Axboe */ 4683ba7b07SChristoph Hellwig struct wb_writeback_work { 47c4a77a6cSJens Axboe long nr_pages; 48c4a77a6cSJens Axboe struct super_block *sb; 490dc83bd3SJan Kara unsigned long *older_than_this; 50c4a77a6cSJens Axboe enum writeback_sync_modes sync_mode; 516e6938b6SWu Fengguang unsigned int tagged_writepages:1; 5252957fe1SH Hartley Sweeten unsigned int for_kupdate:1; 5352957fe1SH Hartley Sweeten unsigned int range_cyclic:1; 5452957fe1SH Hartley Sweeten unsigned int for_background:1; 557747bd4bSDave Chinner unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */ 56ac7b19a3STejun Heo unsigned int auto_free:1; /* free on completion */ 570e175a18SCurt Wohlgemuth enum wb_reason reason; /* why was writeback initiated? */ 58c4a77a6cSJens Axboe 598010c3b6SJens Axboe struct list_head list; /* pending work list */ 60cc395d7fSTejun Heo struct wb_completion *done; /* set if the caller waits */ 6103ba3782SJens Axboe }; 6203ba3782SJens Axboe 63a2f48706STheodore Ts'o /* 64cc395d7fSTejun Heo * If one wants to wait for one or more wb_writeback_works, each work's 65cc395d7fSTejun Heo * ->done should be set to a wb_completion defined using the following 66cc395d7fSTejun Heo * macro. Once all work items are issued with wb_queue_work(), the caller 67cc395d7fSTejun Heo * can wait for the completion of all using wb_wait_for_completion(). Work 68cc395d7fSTejun Heo * items which are waited upon aren't freed automatically on completion. 69cc395d7fSTejun Heo */ 70cc395d7fSTejun Heo #define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \ 71cc395d7fSTejun Heo struct wb_completion cmpl = { \ 72cc395d7fSTejun Heo .cnt = ATOMIC_INIT(1), \ 73cc395d7fSTejun Heo } 74cc395d7fSTejun Heo 75cc395d7fSTejun Heo 76cc395d7fSTejun Heo /* 77a2f48706STheodore Ts'o * If an inode is constantly having its pages dirtied, but then the 78a2f48706STheodore Ts'o * updates stop dirtytime_expire_interval seconds in the past, it's 79a2f48706STheodore Ts'o * possible for the worst case time between when an inode has its 80a2f48706STheodore Ts'o * timestamps updated and when they finally get written out to be two 81a2f48706STheodore Ts'o * dirtytime_expire_intervals. We set the default to 12 hours (in 82a2f48706STheodore Ts'o * seconds), which means most of the time inodes will have their 83a2f48706STheodore Ts'o * timestamps written to disk after 12 hours, but in the worst case a 84a2f48706STheodore Ts'o * few inodes might not their timestamps updated for 24 hours. 85a2f48706STheodore Ts'o */ 86a2f48706STheodore Ts'o unsigned int dirtytime_expire_interval = 12 * 60 * 60; 87a2f48706STheodore Ts'o 887ccf19a8SNick Piggin static inline struct inode *wb_inode(struct list_head *head) 897ccf19a8SNick Piggin { 90c7f54084SDave Chinner return list_entry(head, struct inode, i_io_list); 917ccf19a8SNick Piggin } 927ccf19a8SNick Piggin 9315eb77a0SWu Fengguang /* 9415eb77a0SWu Fengguang * Include the creation of the trace points after defining the 9515eb77a0SWu Fengguang * wb_writeback_work structure and inline functions so that the definition 9615eb77a0SWu Fengguang * remains local to this file. 9715eb77a0SWu Fengguang */ 9815eb77a0SWu Fengguang #define CREATE_TRACE_POINTS 9915eb77a0SWu Fengguang #include <trace/events/writeback.h> 10015eb77a0SWu Fengguang 101774016b2SSteven Whitehouse EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage); 102774016b2SSteven Whitehouse 103d6c10f1fSTejun Heo static bool wb_io_lists_populated(struct bdi_writeback *wb) 104d6c10f1fSTejun Heo { 105d6c10f1fSTejun Heo if (wb_has_dirty_io(wb)) { 106d6c10f1fSTejun Heo return false; 107d6c10f1fSTejun Heo } else { 108d6c10f1fSTejun Heo set_bit(WB_has_dirty_io, &wb->state); 10995a46c65STejun Heo WARN_ON_ONCE(!wb->avg_write_bandwidth); 110766a9d6eSTejun Heo atomic_long_add(wb->avg_write_bandwidth, 111766a9d6eSTejun Heo &wb->bdi->tot_write_bandwidth); 112d6c10f1fSTejun Heo return true; 113d6c10f1fSTejun Heo } 114d6c10f1fSTejun Heo } 115d6c10f1fSTejun Heo 116d6c10f1fSTejun Heo static void wb_io_lists_depopulated(struct bdi_writeback *wb) 117d6c10f1fSTejun Heo { 118d6c10f1fSTejun Heo if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) && 119766a9d6eSTejun Heo list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) { 120d6c10f1fSTejun Heo clear_bit(WB_has_dirty_io, &wb->state); 12195a46c65STejun Heo WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth, 12295a46c65STejun Heo &wb->bdi->tot_write_bandwidth) < 0); 123766a9d6eSTejun Heo } 124d6c10f1fSTejun Heo } 125d6c10f1fSTejun Heo 126d6c10f1fSTejun Heo /** 127c7f54084SDave Chinner * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list 128d6c10f1fSTejun Heo * @inode: inode to be moved 129d6c10f1fSTejun Heo * @wb: target bdi_writeback 130bbbc3c1cSWang Long * @head: one of @wb->b_{dirty|io|more_io|dirty_time} 131d6c10f1fSTejun Heo * 132c7f54084SDave Chinner * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io. 133d6c10f1fSTejun Heo * Returns %true if @inode is the first occupant of the !dirty_time IO 134d6c10f1fSTejun Heo * lists; otherwise, %false. 135d6c10f1fSTejun Heo */ 136c7f54084SDave Chinner static bool inode_io_list_move_locked(struct inode *inode, 137d6c10f1fSTejun Heo struct bdi_writeback *wb, 138d6c10f1fSTejun Heo struct list_head *head) 139d6c10f1fSTejun Heo { 140d6c10f1fSTejun Heo assert_spin_locked(&wb->list_lock); 141d6c10f1fSTejun Heo 142c7f54084SDave Chinner list_move(&inode->i_io_list, head); 143d6c10f1fSTejun Heo 144d6c10f1fSTejun Heo /* dirty_time doesn't count as dirty_io until expiration */ 145d6c10f1fSTejun Heo if (head != &wb->b_dirty_time) 146d6c10f1fSTejun Heo return wb_io_lists_populated(wb); 147d6c10f1fSTejun Heo 148d6c10f1fSTejun Heo wb_io_lists_depopulated(wb); 149d6c10f1fSTejun Heo return false; 150d6c10f1fSTejun Heo } 151d6c10f1fSTejun Heo 152d6c10f1fSTejun Heo /** 153c7f54084SDave Chinner * inode_io_list_del_locked - remove an inode from its bdi_writeback IO list 154d6c10f1fSTejun Heo * @inode: inode to be removed 155d6c10f1fSTejun Heo * @wb: bdi_writeback @inode is being removed from 156d6c10f1fSTejun Heo * 157d6c10f1fSTejun Heo * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and 158d6c10f1fSTejun Heo * clear %WB_has_dirty_io if all are empty afterwards. 159d6c10f1fSTejun Heo */ 160c7f54084SDave Chinner static void inode_io_list_del_locked(struct inode *inode, 161d6c10f1fSTejun Heo struct bdi_writeback *wb) 162d6c10f1fSTejun Heo { 163d6c10f1fSTejun Heo assert_spin_locked(&wb->list_lock); 164d6c10f1fSTejun Heo 165c7f54084SDave Chinner list_del_init(&inode->i_io_list); 166d6c10f1fSTejun Heo wb_io_lists_depopulated(wb); 167d6c10f1fSTejun Heo } 168d6c10f1fSTejun Heo 169f0054bb1STejun Heo static void wb_wakeup(struct bdi_writeback *wb) 1705acda9d1SJan Kara { 171f0054bb1STejun Heo spin_lock_bh(&wb->work_lock); 172f0054bb1STejun Heo if (test_bit(WB_registered, &wb->state)) 173f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 174f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 1755acda9d1SJan Kara } 1765acda9d1SJan Kara 1774a3a485bSTahsin Erdogan static void finish_writeback_work(struct bdi_writeback *wb, 1784a3a485bSTahsin Erdogan struct wb_writeback_work *work) 1794a3a485bSTahsin Erdogan { 1804a3a485bSTahsin Erdogan struct wb_completion *done = work->done; 1814a3a485bSTahsin Erdogan 1824a3a485bSTahsin Erdogan if (work->auto_free) 1834a3a485bSTahsin Erdogan kfree(work); 1844a3a485bSTahsin Erdogan if (done && atomic_dec_and_test(&done->cnt)) 1854a3a485bSTahsin Erdogan wake_up_all(&wb->bdi->wb_waitq); 1864a3a485bSTahsin Erdogan } 1874a3a485bSTahsin Erdogan 188f0054bb1STejun Heo static void wb_queue_work(struct bdi_writeback *wb, 1896585027aSJan Kara struct wb_writeback_work *work) 1906585027aSJan Kara { 1915634cc2aSTejun Heo trace_writeback_queue(wb, work); 1926585027aSJan Kara 193cc395d7fSTejun Heo if (work->done) 194cc395d7fSTejun Heo atomic_inc(&work->done->cnt); 1954a3a485bSTahsin Erdogan 1964a3a485bSTahsin Erdogan spin_lock_bh(&wb->work_lock); 1974a3a485bSTahsin Erdogan 1984a3a485bSTahsin Erdogan if (test_bit(WB_registered, &wb->state)) { 199f0054bb1STejun Heo list_add_tail(&work->list, &wb->work_list); 200f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 2014a3a485bSTahsin Erdogan } else 2024a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 2034a3a485bSTahsin Erdogan 204f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 20503ba3782SJens Axboe } 2061da177e4SLinus Torvalds 207cc395d7fSTejun Heo /** 208cc395d7fSTejun Heo * wb_wait_for_completion - wait for completion of bdi_writeback_works 209cc395d7fSTejun Heo * @bdi: bdi work items were issued to 210cc395d7fSTejun Heo * @done: target wb_completion 211cc395d7fSTejun Heo * 212cc395d7fSTejun Heo * Wait for one or more work items issued to @bdi with their ->done field 213cc395d7fSTejun Heo * set to @done, which should have been defined with 214cc395d7fSTejun Heo * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such 215cc395d7fSTejun Heo * work items are completed. Work items which are waited upon aren't freed 216cc395d7fSTejun Heo * automatically on completion. 217cc395d7fSTejun Heo */ 218cc395d7fSTejun Heo static void wb_wait_for_completion(struct backing_dev_info *bdi, 219cc395d7fSTejun Heo struct wb_completion *done) 220cc395d7fSTejun Heo { 221cc395d7fSTejun Heo atomic_dec(&done->cnt); /* put down the initial count */ 222cc395d7fSTejun Heo wait_event(bdi->wb_waitq, !atomic_read(&done->cnt)); 223cc395d7fSTejun Heo } 224cc395d7fSTejun Heo 225703c2708STejun Heo #ifdef CONFIG_CGROUP_WRITEBACK 226703c2708STejun Heo 22755a694dfSTejun Heo /* 22855a694dfSTejun Heo * Parameters for foreign inode detection, see wbc_detach_inode() to see 22955a694dfSTejun Heo * how they're used. 23055a694dfSTejun Heo * 23155a694dfSTejun Heo * These paramters are inherently heuristical as the detection target 23255a694dfSTejun Heo * itself is fuzzy. All we want to do is detaching an inode from the 23355a694dfSTejun Heo * current owner if it's being written to by some other cgroups too much. 23455a694dfSTejun Heo * 23555a694dfSTejun Heo * The current cgroup writeback is built on the assumption that multiple 23655a694dfSTejun Heo * cgroups writing to the same inode concurrently is very rare and a mode 23755a694dfSTejun Heo * of operation which isn't well supported. As such, the goal is not 23855a694dfSTejun Heo * taking too long when a different cgroup takes over an inode while 23955a694dfSTejun Heo * avoiding too aggressive flip-flops from occasional foreign writes. 24055a694dfSTejun Heo * 24155a694dfSTejun Heo * We record, very roughly, 2s worth of IO time history and if more than 24255a694dfSTejun Heo * half of that is foreign, trigger the switch. The recording is quantized 24355a694dfSTejun Heo * to 16 slots. To avoid tiny writes from swinging the decision too much, 24455a694dfSTejun Heo * writes smaller than 1/8 of avg size are ignored. 24555a694dfSTejun Heo */ 2462a814908STejun Heo #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */ 2472a814908STejun Heo #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */ 24855a694dfSTejun Heo #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */ 2492a814908STejun Heo #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */ 2502a814908STejun Heo 2512a814908STejun Heo #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */ 2522a814908STejun Heo #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS) 2532a814908STejun Heo /* each slot's duration is 2s / 16 */ 2542a814908STejun Heo #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2) 2552a814908STejun Heo /* if foreign slots >= 8, switch */ 2562a814908STejun Heo #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1) 2572a814908STejun Heo /* one round can affect upto 5 slots */ 2582a814908STejun Heo 259a1a0e23eSTejun Heo static atomic_t isw_nr_in_flight = ATOMIC_INIT(0); 260a1a0e23eSTejun Heo static struct workqueue_struct *isw_wq; 261a1a0e23eSTejun Heo 26221c6321fSTejun Heo void __inode_attach_wb(struct inode *inode, struct page *page) 26321c6321fSTejun Heo { 26421c6321fSTejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 26521c6321fSTejun Heo struct bdi_writeback *wb = NULL; 26621c6321fSTejun Heo 26721c6321fSTejun Heo if (inode_cgwb_enabled(inode)) { 26821c6321fSTejun Heo struct cgroup_subsys_state *memcg_css; 26921c6321fSTejun Heo 27021c6321fSTejun Heo if (page) { 27121c6321fSTejun Heo memcg_css = mem_cgroup_css_from_page(page); 27221c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 27321c6321fSTejun Heo } else { 27421c6321fSTejun Heo /* must pin memcg_css, see wb_get_create() */ 27521c6321fSTejun Heo memcg_css = task_get_css(current, memory_cgrp_id); 27621c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 27721c6321fSTejun Heo css_put(memcg_css); 27821c6321fSTejun Heo } 27921c6321fSTejun Heo } 28021c6321fSTejun Heo 28121c6321fSTejun Heo if (!wb) 28221c6321fSTejun Heo wb = &bdi->wb; 28321c6321fSTejun Heo 28421c6321fSTejun Heo /* 28521c6321fSTejun Heo * There may be multiple instances of this function racing to 28621c6321fSTejun Heo * update the same inode. Use cmpxchg() to tell the winner. 28721c6321fSTejun Heo */ 28821c6321fSTejun Heo if (unlikely(cmpxchg(&inode->i_wb, NULL, wb))) 28921c6321fSTejun Heo wb_put(wb); 29021c6321fSTejun Heo } 2919b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(__inode_attach_wb); 29221c6321fSTejun Heo 293703c2708STejun Heo /** 29487e1d789STejun Heo * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it 29587e1d789STejun Heo * @inode: inode of interest with i_lock held 29687e1d789STejun Heo * 29787e1d789STejun Heo * Returns @inode's wb with its list_lock held. @inode->i_lock must be 29887e1d789STejun Heo * held on entry and is released on return. The returned wb is guaranteed 29987e1d789STejun Heo * to stay @inode's associated wb until its list_lock is released. 30087e1d789STejun Heo */ 30187e1d789STejun Heo static struct bdi_writeback * 30287e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 30387e1d789STejun Heo __releases(&inode->i_lock) 30487e1d789STejun Heo __acquires(&wb->list_lock) 30587e1d789STejun Heo { 30687e1d789STejun Heo while (true) { 30787e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 30887e1d789STejun Heo 30987e1d789STejun Heo /* 31087e1d789STejun Heo * inode_to_wb() association is protected by both 31187e1d789STejun Heo * @inode->i_lock and @wb->list_lock but list_lock nests 31287e1d789STejun Heo * outside i_lock. Drop i_lock and verify that the 31387e1d789STejun Heo * association hasn't changed after acquiring list_lock. 31487e1d789STejun Heo */ 31587e1d789STejun Heo wb_get(wb); 31687e1d789STejun Heo spin_unlock(&inode->i_lock); 31787e1d789STejun Heo spin_lock(&wb->list_lock); 31887e1d789STejun Heo 319aaa2cacfSTejun Heo /* i_wb may have changed inbetween, can't use inode_to_wb() */ 320614a4e37STejun Heo if (likely(wb == inode->i_wb)) { 321614a4e37STejun Heo wb_put(wb); /* @inode already has ref */ 322614a4e37STejun Heo return wb; 323614a4e37STejun Heo } 32487e1d789STejun Heo 32587e1d789STejun Heo spin_unlock(&wb->list_lock); 326614a4e37STejun Heo wb_put(wb); 32787e1d789STejun Heo cpu_relax(); 32887e1d789STejun Heo spin_lock(&inode->i_lock); 32987e1d789STejun Heo } 33087e1d789STejun Heo } 33187e1d789STejun Heo 33287e1d789STejun Heo /** 33387e1d789STejun Heo * inode_to_wb_and_lock_list - determine an inode's wb and lock it 33487e1d789STejun Heo * @inode: inode of interest 33587e1d789STejun Heo * 33687e1d789STejun Heo * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held 33787e1d789STejun Heo * on entry. 33887e1d789STejun Heo */ 33987e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 34087e1d789STejun Heo __acquires(&wb->list_lock) 34187e1d789STejun Heo { 34287e1d789STejun Heo spin_lock(&inode->i_lock); 34387e1d789STejun Heo return locked_inode_to_wb_and_lock_list(inode); 34487e1d789STejun Heo } 34587e1d789STejun Heo 346682aa8e1STejun Heo struct inode_switch_wbs_context { 347682aa8e1STejun Heo struct inode *inode; 348682aa8e1STejun Heo struct bdi_writeback *new_wb; 349682aa8e1STejun Heo 350682aa8e1STejun Heo struct rcu_head rcu_head; 351682aa8e1STejun Heo struct work_struct work; 352682aa8e1STejun Heo }; 353682aa8e1STejun Heo 3547fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) 3557fc5854fSTejun Heo { 3567fc5854fSTejun Heo down_write(&bdi->wb_switch_rwsem); 3577fc5854fSTejun Heo } 3587fc5854fSTejun Heo 3597fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) 3607fc5854fSTejun Heo { 3617fc5854fSTejun Heo up_write(&bdi->wb_switch_rwsem); 3627fc5854fSTejun Heo } 3637fc5854fSTejun Heo 364682aa8e1STejun Heo static void inode_switch_wbs_work_fn(struct work_struct *work) 365682aa8e1STejun Heo { 366682aa8e1STejun Heo struct inode_switch_wbs_context *isw = 367682aa8e1STejun Heo container_of(work, struct inode_switch_wbs_context, work); 368682aa8e1STejun Heo struct inode *inode = isw->inode; 3697fc5854fSTejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 370d10c8095STejun Heo struct address_space *mapping = inode->i_mapping; 371d10c8095STejun Heo struct bdi_writeback *old_wb = inode->i_wb; 372682aa8e1STejun Heo struct bdi_writeback *new_wb = isw->new_wb; 37304edf02cSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, 0); 37404edf02cSMatthew Wilcox struct page *page; 375d10c8095STejun Heo bool switched = false; 376682aa8e1STejun Heo 377682aa8e1STejun Heo /* 3787fc5854fSTejun Heo * If @inode switches cgwb membership while sync_inodes_sb() is 3797fc5854fSTejun Heo * being issued, sync_inodes_sb() might miss it. Synchronize. 3807fc5854fSTejun Heo */ 3817fc5854fSTejun Heo down_read(&bdi->wb_switch_rwsem); 3827fc5854fSTejun Heo 3837fc5854fSTejun Heo /* 384682aa8e1STejun Heo * By the time control reaches here, RCU grace period has passed 385682aa8e1STejun Heo * since I_WB_SWITCH assertion and all wb stat update transactions 386682aa8e1STejun Heo * between unlocked_inode_to_wb_begin/end() are guaranteed to be 387b93b0163SMatthew Wilcox * synchronizing against the i_pages lock. 388d10c8095STejun Heo * 389b93b0163SMatthew Wilcox * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock 390d10c8095STejun Heo * gives us exclusion against all wb related operations on @inode 391d10c8095STejun Heo * including IO list manipulations and stat updates. 392682aa8e1STejun Heo */ 393d10c8095STejun Heo if (old_wb < new_wb) { 394d10c8095STejun Heo spin_lock(&old_wb->list_lock); 395d10c8095STejun Heo spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING); 396d10c8095STejun Heo } else { 397d10c8095STejun Heo spin_lock(&new_wb->list_lock); 398d10c8095STejun Heo spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING); 399d10c8095STejun Heo } 400682aa8e1STejun Heo spin_lock(&inode->i_lock); 401b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 402682aa8e1STejun Heo 403d10c8095STejun Heo /* 404d10c8095STejun Heo * Once I_FREEING is visible under i_lock, the eviction path owns 405c7f54084SDave Chinner * the inode and we shouldn't modify ->i_io_list. 406d10c8095STejun Heo */ 407d10c8095STejun Heo if (unlikely(inode->i_state & I_FREEING)) 408d10c8095STejun Heo goto skip_switch; 409d10c8095STejun Heo 410d10c8095STejun Heo /* 411d10c8095STejun Heo * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points 412d10c8095STejun Heo * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to 413b93b0163SMatthew Wilcox * pages actually under writeback. 414d10c8095STejun Heo */ 41504edf02cSMatthew Wilcox xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) { 41604edf02cSMatthew Wilcox if (PageDirty(page)) { 4173e8f399dSNikolay Borisov dec_wb_stat(old_wb, WB_RECLAIMABLE); 4183e8f399dSNikolay Borisov inc_wb_stat(new_wb, WB_RECLAIMABLE); 419d10c8095STejun Heo } 420d10c8095STejun Heo } 421d10c8095STejun Heo 42204edf02cSMatthew Wilcox xas_set(&xas, 0); 42304edf02cSMatthew Wilcox xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) { 424d10c8095STejun Heo WARN_ON_ONCE(!PageWriteback(page)); 4253e8f399dSNikolay Borisov dec_wb_stat(old_wb, WB_WRITEBACK); 4263e8f399dSNikolay Borisov inc_wb_stat(new_wb, WB_WRITEBACK); 427d10c8095STejun Heo } 428d10c8095STejun Heo 429d10c8095STejun Heo wb_get(new_wb); 430d10c8095STejun Heo 431d10c8095STejun Heo /* 432d10c8095STejun Heo * Transfer to @new_wb's IO list if necessary. The specific list 433d10c8095STejun Heo * @inode was on is ignored and the inode is put on ->b_dirty which 434d10c8095STejun Heo * is always correct including from ->b_dirty_time. The transfer 435d10c8095STejun Heo * preserves @inode->dirtied_when ordering. 436d10c8095STejun Heo */ 437c7f54084SDave Chinner if (!list_empty(&inode->i_io_list)) { 438d10c8095STejun Heo struct inode *pos; 439d10c8095STejun Heo 440c7f54084SDave Chinner inode_io_list_del_locked(inode, old_wb); 441d10c8095STejun Heo inode->i_wb = new_wb; 442c7f54084SDave Chinner list_for_each_entry(pos, &new_wb->b_dirty, i_io_list) 443d10c8095STejun Heo if (time_after_eq(inode->dirtied_when, 444d10c8095STejun Heo pos->dirtied_when)) 445d10c8095STejun Heo break; 446c7f54084SDave Chinner inode_io_list_move_locked(inode, new_wb, pos->i_io_list.prev); 447d10c8095STejun Heo } else { 448d10c8095STejun Heo inode->i_wb = new_wb; 449d10c8095STejun Heo } 450d10c8095STejun Heo 451d10c8095STejun Heo /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */ 452682aa8e1STejun Heo inode->i_wb_frn_winner = 0; 453682aa8e1STejun Heo inode->i_wb_frn_avg_time = 0; 454682aa8e1STejun Heo inode->i_wb_frn_history = 0; 455d10c8095STejun Heo switched = true; 456d10c8095STejun Heo skip_switch: 457682aa8e1STejun Heo /* 458682aa8e1STejun Heo * Paired with load_acquire in unlocked_inode_to_wb_begin() and 459682aa8e1STejun Heo * ensures that the new wb is visible if they see !I_WB_SWITCH. 460682aa8e1STejun Heo */ 461682aa8e1STejun Heo smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); 462682aa8e1STejun Heo 463b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 464682aa8e1STejun Heo spin_unlock(&inode->i_lock); 465d10c8095STejun Heo spin_unlock(&new_wb->list_lock); 466d10c8095STejun Heo spin_unlock(&old_wb->list_lock); 467d10c8095STejun Heo 4687fc5854fSTejun Heo up_read(&bdi->wb_switch_rwsem); 4697fc5854fSTejun Heo 470d10c8095STejun Heo if (switched) { 471d10c8095STejun Heo wb_wakeup(new_wb); 472d10c8095STejun Heo wb_put(old_wb); 473d10c8095STejun Heo } 474d10c8095STejun Heo wb_put(new_wb); 475682aa8e1STejun Heo 476682aa8e1STejun Heo iput(inode); 477682aa8e1STejun Heo kfree(isw); 478a1a0e23eSTejun Heo 479a1a0e23eSTejun Heo atomic_dec(&isw_nr_in_flight); 480682aa8e1STejun Heo } 481682aa8e1STejun Heo 482682aa8e1STejun Heo static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head) 483682aa8e1STejun Heo { 484682aa8e1STejun Heo struct inode_switch_wbs_context *isw = container_of(rcu_head, 485682aa8e1STejun Heo struct inode_switch_wbs_context, rcu_head); 486682aa8e1STejun Heo 487682aa8e1STejun Heo /* needs to grab bh-unsafe locks, bounce to work item */ 488682aa8e1STejun Heo INIT_WORK(&isw->work, inode_switch_wbs_work_fn); 489a1a0e23eSTejun Heo queue_work(isw_wq, &isw->work); 490682aa8e1STejun Heo } 491682aa8e1STejun Heo 492682aa8e1STejun Heo /** 493682aa8e1STejun Heo * inode_switch_wbs - change the wb association of an inode 494682aa8e1STejun Heo * @inode: target inode 495682aa8e1STejun Heo * @new_wb_id: ID of the new wb 496682aa8e1STejun Heo * 497682aa8e1STejun Heo * Switch @inode's wb association to the wb identified by @new_wb_id. The 498682aa8e1STejun Heo * switching is performed asynchronously and may fail silently. 499682aa8e1STejun Heo */ 500682aa8e1STejun Heo static void inode_switch_wbs(struct inode *inode, int new_wb_id) 501682aa8e1STejun Heo { 502682aa8e1STejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 503682aa8e1STejun Heo struct cgroup_subsys_state *memcg_css; 504682aa8e1STejun Heo struct inode_switch_wbs_context *isw; 505682aa8e1STejun Heo 506682aa8e1STejun Heo /* noop if seems to be already in progress */ 507682aa8e1STejun Heo if (inode->i_state & I_WB_SWITCH) 508682aa8e1STejun Heo return; 509682aa8e1STejun Heo 5107fc5854fSTejun Heo /* 5117fc5854fSTejun Heo * Avoid starting new switches while sync_inodes_sb() is in 5127fc5854fSTejun Heo * progress. Otherwise, if the down_write protected issue path 5137fc5854fSTejun Heo * blocks heavily, we might end up starting a large number of 5147fc5854fSTejun Heo * switches which will block on the rwsem. 5157fc5854fSTejun Heo */ 5167fc5854fSTejun Heo if (!down_read_trylock(&bdi->wb_switch_rwsem)) 5177fc5854fSTejun Heo return; 5187fc5854fSTejun Heo 519682aa8e1STejun Heo isw = kzalloc(sizeof(*isw), GFP_ATOMIC); 520682aa8e1STejun Heo if (!isw) 5217fc5854fSTejun Heo goto out_unlock; 522682aa8e1STejun Heo 523682aa8e1STejun Heo /* find and pin the new wb */ 524682aa8e1STejun Heo rcu_read_lock(); 525682aa8e1STejun Heo memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); 526682aa8e1STejun Heo if (memcg_css) 527682aa8e1STejun Heo isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 528682aa8e1STejun Heo rcu_read_unlock(); 529682aa8e1STejun Heo if (!isw->new_wb) 530682aa8e1STejun Heo goto out_free; 531682aa8e1STejun Heo 532682aa8e1STejun Heo /* while holding I_WB_SWITCH, no one else can update the association */ 533682aa8e1STejun Heo spin_lock(&inode->i_lock); 5341751e8a6SLinus Torvalds if (!(inode->i_sb->s_flags & SB_ACTIVE) || 535a1a0e23eSTejun Heo inode->i_state & (I_WB_SWITCH | I_FREEING) || 536a1a0e23eSTejun Heo inode_to_wb(inode) == isw->new_wb) { 537a1a0e23eSTejun Heo spin_unlock(&inode->i_lock); 538a1a0e23eSTejun Heo goto out_free; 539a1a0e23eSTejun Heo } 540682aa8e1STejun Heo inode->i_state |= I_WB_SWITCH; 54174524955STahsin Erdogan __iget(inode); 542682aa8e1STejun Heo spin_unlock(&inode->i_lock); 543682aa8e1STejun Heo 544682aa8e1STejun Heo isw->inode = inode; 545682aa8e1STejun Heo 546682aa8e1STejun Heo /* 547682aa8e1STejun Heo * In addition to synchronizing among switchers, I_WB_SWITCH tells 548b93b0163SMatthew Wilcox * the RCU protected stat update paths to grab the i_page 549b93b0163SMatthew Wilcox * lock so that stat transfer can synchronize against them. 550682aa8e1STejun Heo * Let's continue after I_WB_SWITCH is guaranteed to be visible. 551682aa8e1STejun Heo */ 552682aa8e1STejun Heo call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn); 553ec084de9SJiufei Xue 554ec084de9SJiufei Xue atomic_inc(&isw_nr_in_flight); 555ec084de9SJiufei Xue 5567fc5854fSTejun Heo goto out_unlock; 557682aa8e1STejun Heo 558682aa8e1STejun Heo out_free: 559682aa8e1STejun Heo if (isw->new_wb) 560682aa8e1STejun Heo wb_put(isw->new_wb); 561682aa8e1STejun Heo kfree(isw); 5627fc5854fSTejun Heo out_unlock: 5637fc5854fSTejun Heo up_read(&bdi->wb_switch_rwsem); 564682aa8e1STejun Heo } 565682aa8e1STejun Heo 56687e1d789STejun Heo /** 567b16b1debSTejun Heo * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it 568b16b1debSTejun Heo * @wbc: writeback_control of interest 569b16b1debSTejun Heo * @inode: target inode 570b16b1debSTejun Heo * 571b16b1debSTejun Heo * @inode is locked and about to be written back under the control of @wbc. 572b16b1debSTejun Heo * Record @inode's writeback context into @wbc and unlock the i_lock. On 573b16b1debSTejun Heo * writeback completion, wbc_detach_inode() should be called. This is used 574b16b1debSTejun Heo * to track the cgroup writeback context. 575b16b1debSTejun Heo */ 576b16b1debSTejun Heo void wbc_attach_and_unlock_inode(struct writeback_control *wbc, 577b16b1debSTejun Heo struct inode *inode) 578b16b1debSTejun Heo { 579dd73e4b7STejun Heo if (!inode_cgwb_enabled(inode)) { 580dd73e4b7STejun Heo spin_unlock(&inode->i_lock); 581dd73e4b7STejun Heo return; 582dd73e4b7STejun Heo } 583dd73e4b7STejun Heo 584b16b1debSTejun Heo wbc->wb = inode_to_wb(inode); 5852a814908STejun Heo wbc->inode = inode; 5862a814908STejun Heo 5872a814908STejun Heo wbc->wb_id = wbc->wb->memcg_css->id; 5882a814908STejun Heo wbc->wb_lcand_id = inode->i_wb_frn_winner; 5892a814908STejun Heo wbc->wb_tcand_id = 0; 5902a814908STejun Heo wbc->wb_bytes = 0; 5912a814908STejun Heo wbc->wb_lcand_bytes = 0; 5922a814908STejun Heo wbc->wb_tcand_bytes = 0; 5932a814908STejun Heo 594b16b1debSTejun Heo wb_get(wbc->wb); 595b16b1debSTejun Heo spin_unlock(&inode->i_lock); 596e8a7abf5STejun Heo 597e8a7abf5STejun Heo /* 598e8a7abf5STejun Heo * A dying wb indicates that the memcg-blkcg mapping has changed 599e8a7abf5STejun Heo * and a new wb is already serving the memcg. Switch immediately. 600e8a7abf5STejun Heo */ 601e8a7abf5STejun Heo if (unlikely(wb_dying(wbc->wb))) 602e8a7abf5STejun Heo inode_switch_wbs(inode, wbc->wb_id); 603b16b1debSTejun Heo } 6049b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode); 605b16b1debSTejun Heo 606b16b1debSTejun Heo /** 6072a814908STejun Heo * wbc_detach_inode - disassociate wbc from inode and perform foreign detection 6082a814908STejun Heo * @wbc: writeback_control of the just finished writeback 609b16b1debSTejun Heo * 610b16b1debSTejun Heo * To be called after a writeback attempt of an inode finishes and undoes 611b16b1debSTejun Heo * wbc_attach_and_unlock_inode(). Can be called under any context. 6122a814908STejun Heo * 6132a814908STejun Heo * As concurrent write sharing of an inode is expected to be very rare and 6142a814908STejun Heo * memcg only tracks page ownership on first-use basis severely confining 6152a814908STejun Heo * the usefulness of such sharing, cgroup writeback tracks ownership 6162a814908STejun Heo * per-inode. While the support for concurrent write sharing of an inode 6172a814908STejun Heo * is deemed unnecessary, an inode being written to by different cgroups at 6182a814908STejun Heo * different points in time is a lot more common, and, more importantly, 6192a814908STejun Heo * charging only by first-use can too readily lead to grossly incorrect 6202a814908STejun Heo * behaviors (single foreign page can lead to gigabytes of writeback to be 6212a814908STejun Heo * incorrectly attributed). 6222a814908STejun Heo * 6232a814908STejun Heo * To resolve this issue, cgroup writeback detects the majority dirtier of 6242a814908STejun Heo * an inode and transfers the ownership to it. To avoid unnnecessary 6252a814908STejun Heo * oscillation, the detection mechanism keeps track of history and gives 6262a814908STejun Heo * out the switch verdict only if the foreign usage pattern is stable over 6272a814908STejun Heo * a certain amount of time and/or writeback attempts. 6282a814908STejun Heo * 6292a814908STejun Heo * On each writeback attempt, @wbc tries to detect the majority writer 6302a814908STejun Heo * using Boyer-Moore majority vote algorithm. In addition to the byte 6312a814908STejun Heo * count from the majority voting, it also counts the bytes written for the 6322a814908STejun Heo * current wb and the last round's winner wb (max of last round's current 6332a814908STejun Heo * wb, the winner from two rounds ago, and the last round's majority 6342a814908STejun Heo * candidate). Keeping track of the historical winner helps the algorithm 6352a814908STejun Heo * to semi-reliably detect the most active writer even when it's not the 6362a814908STejun Heo * absolute majority. 6372a814908STejun Heo * 6382a814908STejun Heo * Once the winner of the round is determined, whether the winner is 6392a814908STejun Heo * foreign or not and how much IO time the round consumed is recorded in 6402a814908STejun Heo * inode->i_wb_frn_history. If the amount of recorded foreign IO time is 6412a814908STejun Heo * over a certain threshold, the switch verdict is given. 642b16b1debSTejun Heo */ 643b16b1debSTejun Heo void wbc_detach_inode(struct writeback_control *wbc) 644b16b1debSTejun Heo { 6452a814908STejun Heo struct bdi_writeback *wb = wbc->wb; 6462a814908STejun Heo struct inode *inode = wbc->inode; 647dd73e4b7STejun Heo unsigned long avg_time, max_bytes, max_time; 648dd73e4b7STejun Heo u16 history; 6492a814908STejun Heo int max_id; 6502a814908STejun Heo 651dd73e4b7STejun Heo if (!wb) 652dd73e4b7STejun Heo return; 653dd73e4b7STejun Heo 654dd73e4b7STejun Heo history = inode->i_wb_frn_history; 655dd73e4b7STejun Heo avg_time = inode->i_wb_frn_avg_time; 656dd73e4b7STejun Heo 6572a814908STejun Heo /* pick the winner of this round */ 6582a814908STejun Heo if (wbc->wb_bytes >= wbc->wb_lcand_bytes && 6592a814908STejun Heo wbc->wb_bytes >= wbc->wb_tcand_bytes) { 6602a814908STejun Heo max_id = wbc->wb_id; 6612a814908STejun Heo max_bytes = wbc->wb_bytes; 6622a814908STejun Heo } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { 6632a814908STejun Heo max_id = wbc->wb_lcand_id; 6642a814908STejun Heo max_bytes = wbc->wb_lcand_bytes; 6652a814908STejun Heo } else { 6662a814908STejun Heo max_id = wbc->wb_tcand_id; 6672a814908STejun Heo max_bytes = wbc->wb_tcand_bytes; 6682a814908STejun Heo } 6692a814908STejun Heo 6702a814908STejun Heo /* 6712a814908STejun Heo * Calculate the amount of IO time the winner consumed and fold it 6722a814908STejun Heo * into the running average kept per inode. If the consumed IO 6732a814908STejun Heo * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for 6742a814908STejun Heo * deciding whether to switch or not. This is to prevent one-off 6752a814908STejun Heo * small dirtiers from skewing the verdict. 6762a814908STejun Heo */ 6772a814908STejun Heo max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, 6782a814908STejun Heo wb->avg_write_bandwidth); 6792a814908STejun Heo if (avg_time) 6802a814908STejun Heo avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - 6812a814908STejun Heo (avg_time >> WB_FRN_TIME_AVG_SHIFT); 6822a814908STejun Heo else 6832a814908STejun Heo avg_time = max_time; /* immediate catch up on first run */ 6842a814908STejun Heo 6852a814908STejun Heo if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { 6862a814908STejun Heo int slots; 6872a814908STejun Heo 6882a814908STejun Heo /* 6892a814908STejun Heo * The switch verdict is reached if foreign wb's consume 6902a814908STejun Heo * more than a certain proportion of IO time in a 6912a814908STejun Heo * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot 6922a814908STejun Heo * history mask where each bit represents one sixteenth of 6932a814908STejun Heo * the period. Determine the number of slots to shift into 6942a814908STejun Heo * history from @max_time. 6952a814908STejun Heo */ 6962a814908STejun Heo slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), 6972a814908STejun Heo (unsigned long)WB_FRN_HIST_MAX_SLOTS); 6982a814908STejun Heo history <<= slots; 6992a814908STejun Heo if (wbc->wb_id != max_id) 7002a814908STejun Heo history |= (1U << slots) - 1; 7012a814908STejun Heo 7022a814908STejun Heo /* 7032a814908STejun Heo * Switch if the current wb isn't the consistent winner. 7042a814908STejun Heo * If there are multiple closely competing dirtiers, the 7052a814908STejun Heo * inode may switch across them repeatedly over time, which 7062a814908STejun Heo * is okay. The main goal is avoiding keeping an inode on 7072a814908STejun Heo * the wrong wb for an extended period of time. 7082a814908STejun Heo */ 709682aa8e1STejun Heo if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) 710682aa8e1STejun Heo inode_switch_wbs(inode, max_id); 7112a814908STejun Heo } 7122a814908STejun Heo 7132a814908STejun Heo /* 7142a814908STejun Heo * Multiple instances of this function may race to update the 7152a814908STejun Heo * following fields but we don't mind occassional inaccuracies. 7162a814908STejun Heo */ 7172a814908STejun Heo inode->i_wb_frn_winner = max_id; 7182a814908STejun Heo inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); 7192a814908STejun Heo inode->i_wb_frn_history = history; 7202a814908STejun Heo 721b16b1debSTejun Heo wb_put(wbc->wb); 722b16b1debSTejun Heo wbc->wb = NULL; 723b16b1debSTejun Heo } 7249b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_detach_inode); 725b16b1debSTejun Heo 726b16b1debSTejun Heo /** 72734e51a5eSTejun Heo * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership 7282a814908STejun Heo * @wbc: writeback_control of the writeback in progress 7292a814908STejun Heo * @page: page being written out 7302a814908STejun Heo * @bytes: number of bytes being written out 7312a814908STejun Heo * 7322a814908STejun Heo * @bytes from @page are about to written out during the writeback 7332a814908STejun Heo * controlled by @wbc. Keep the book for foreign inode detection. See 7342a814908STejun Heo * wbc_detach_inode(). 7352a814908STejun Heo */ 73634e51a5eSTejun Heo void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page, 7372a814908STejun Heo size_t bytes) 7382a814908STejun Heo { 73966311422STejun Heo struct cgroup_subsys_state *css; 7402a814908STejun Heo int id; 7412a814908STejun Heo 7422a814908STejun Heo /* 7432a814908STejun Heo * pageout() path doesn't attach @wbc to the inode being written 7442a814908STejun Heo * out. This is intentional as we don't want the function to block 7452a814908STejun Heo * behind a slow cgroup. Ultimately, we want pageout() to kick off 7462a814908STejun Heo * regular writeback instead of writing things out itself. 7472a814908STejun Heo */ 74827b36d8fSTejun Heo if (!wbc->wb || wbc->no_cgroup_owner) 7492a814908STejun Heo return; 7502a814908STejun Heo 75166311422STejun Heo css = mem_cgroup_css_from_page(page); 75266311422STejun Heo /* dead cgroups shouldn't contribute to inode ownership arbitration */ 75366311422STejun Heo if (!(css->flags & CSS_ONLINE)) 75466311422STejun Heo return; 75566311422STejun Heo 75666311422STejun Heo id = css->id; 7572a814908STejun Heo 7582a814908STejun Heo if (id == wbc->wb_id) { 7592a814908STejun Heo wbc->wb_bytes += bytes; 7602a814908STejun Heo return; 7612a814908STejun Heo } 7622a814908STejun Heo 7632a814908STejun Heo if (id == wbc->wb_lcand_id) 7642a814908STejun Heo wbc->wb_lcand_bytes += bytes; 7652a814908STejun Heo 7662a814908STejun Heo /* Boyer-Moore majority vote algorithm */ 7672a814908STejun Heo if (!wbc->wb_tcand_bytes) 7682a814908STejun Heo wbc->wb_tcand_id = id; 7692a814908STejun Heo if (id == wbc->wb_tcand_id) 7702a814908STejun Heo wbc->wb_tcand_bytes += bytes; 7712a814908STejun Heo else 7722a814908STejun Heo wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); 7732a814908STejun Heo } 77434e51a5eSTejun Heo EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner); 7752a814908STejun Heo 7762a814908STejun Heo /** 777703c2708STejun Heo * inode_congested - test whether an inode is congested 77860292bccSTejun Heo * @inode: inode to test for congestion (may be NULL) 779703c2708STejun Heo * @cong_bits: mask of WB_[a]sync_congested bits to test 780703c2708STejun Heo * 781703c2708STejun Heo * Tests whether @inode is congested. @cong_bits is the mask of congestion 782703c2708STejun Heo * bits to test and the return value is the mask of set bits. 783703c2708STejun Heo * 784703c2708STejun Heo * If cgroup writeback is enabled for @inode, the congestion state is 785703c2708STejun Heo * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg 786703c2708STejun Heo * associated with @inode is congested; otherwise, the root wb's congestion 787703c2708STejun Heo * state is used. 78860292bccSTejun Heo * 78960292bccSTejun Heo * @inode is allowed to be NULL as this function is often called on 79060292bccSTejun Heo * mapping->host which is NULL for the swapper space. 791703c2708STejun Heo */ 792703c2708STejun Heo int inode_congested(struct inode *inode, int cong_bits) 793703c2708STejun Heo { 7945cb8b824STejun Heo /* 7955cb8b824STejun Heo * Once set, ->i_wb never becomes NULL while the inode is alive. 7965cb8b824STejun Heo * Start transaction iff ->i_wb is visible. 7975cb8b824STejun Heo */ 798aaa2cacfSTejun Heo if (inode && inode_to_wb_is_valid(inode)) { 7995cb8b824STejun Heo struct bdi_writeback *wb; 8002e898e4cSGreg Thelen struct wb_lock_cookie lock_cookie = {}; 8012e898e4cSGreg Thelen bool congested; 8025cb8b824STejun Heo 8032e898e4cSGreg Thelen wb = unlocked_inode_to_wb_begin(inode, &lock_cookie); 8045cb8b824STejun Heo congested = wb_congested(wb, cong_bits); 8052e898e4cSGreg Thelen unlocked_inode_to_wb_end(inode, &lock_cookie); 8065cb8b824STejun Heo return congested; 807703c2708STejun Heo } 808703c2708STejun Heo 809703c2708STejun Heo return wb_congested(&inode_to_bdi(inode)->wb, cong_bits); 810703c2708STejun Heo } 811703c2708STejun Heo EXPORT_SYMBOL_GPL(inode_congested); 812703c2708STejun Heo 813f2b65121STejun Heo /** 814f2b65121STejun Heo * wb_split_bdi_pages - split nr_pages to write according to bandwidth 815f2b65121STejun Heo * @wb: target bdi_writeback to split @nr_pages to 816f2b65121STejun Heo * @nr_pages: number of pages to write for the whole bdi 817f2b65121STejun Heo * 818f2b65121STejun Heo * Split @wb's portion of @nr_pages according to @wb's write bandwidth in 819f2b65121STejun Heo * relation to the total write bandwidth of all wb's w/ dirty inodes on 820f2b65121STejun Heo * @wb->bdi. 821f2b65121STejun Heo */ 822f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 823f2b65121STejun Heo { 824f2b65121STejun Heo unsigned long this_bw = wb->avg_write_bandwidth; 825f2b65121STejun Heo unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); 826f2b65121STejun Heo 827f2b65121STejun Heo if (nr_pages == LONG_MAX) 828f2b65121STejun Heo return LONG_MAX; 829f2b65121STejun Heo 830f2b65121STejun Heo /* 831f2b65121STejun Heo * This may be called on clean wb's and proportional distribution 832f2b65121STejun Heo * may not make sense, just use the original @nr_pages in those 833f2b65121STejun Heo * cases. In general, we wanna err on the side of writing more. 834f2b65121STejun Heo */ 835f2b65121STejun Heo if (!tot_bw || this_bw >= tot_bw) 836f2b65121STejun Heo return nr_pages; 837f2b65121STejun Heo else 838f2b65121STejun Heo return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); 839f2b65121STejun Heo } 840f2b65121STejun Heo 841db125360STejun Heo /** 842db125360STejun Heo * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi 843db125360STejun Heo * @bdi: target backing_dev_info 844db125360STejun Heo * @base_work: wb_writeback_work to issue 845db125360STejun Heo * @skip_if_busy: skip wb's which already have writeback in progress 846db125360STejun Heo * 847db125360STejun Heo * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which 848db125360STejun Heo * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's 849db125360STejun Heo * distributed to the busy wbs according to each wb's proportion in the 850db125360STejun Heo * total active write bandwidth of @bdi. 851db125360STejun Heo */ 852db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 853db125360STejun Heo struct wb_writeback_work *base_work, 854db125360STejun Heo bool skip_if_busy) 855db125360STejun Heo { 856b817525aSTejun Heo struct bdi_writeback *last_wb = NULL; 857b33e18f6STejun Heo struct bdi_writeback *wb = list_entry(&bdi->wb_list, 858b817525aSTejun Heo struct bdi_writeback, bdi_node); 859db125360STejun Heo 860db125360STejun Heo might_sleep(); 861db125360STejun Heo restart: 862db125360STejun Heo rcu_read_lock(); 863b817525aSTejun Heo list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) { 8648a1270cdSTejun Heo DEFINE_WB_COMPLETION_ONSTACK(fallback_work_done); 8658a1270cdSTejun Heo struct wb_writeback_work fallback_work; 8668a1270cdSTejun Heo struct wb_writeback_work *work; 8678a1270cdSTejun Heo long nr_pages; 8688a1270cdSTejun Heo 869b817525aSTejun Heo if (last_wb) { 870b817525aSTejun Heo wb_put(last_wb); 871b817525aSTejun Heo last_wb = NULL; 872b817525aSTejun Heo } 873b817525aSTejun Heo 874006a0973STejun Heo /* SYNC_ALL writes out I_DIRTY_TIME too */ 875006a0973STejun Heo if (!wb_has_dirty_io(wb) && 876006a0973STejun Heo (base_work->sync_mode == WB_SYNC_NONE || 877006a0973STejun Heo list_empty(&wb->b_dirty_time))) 878006a0973STejun Heo continue; 879006a0973STejun Heo if (skip_if_busy && writeback_in_progress(wb)) 880db125360STejun Heo continue; 881db125360STejun Heo 8828a1270cdSTejun Heo nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages); 8838a1270cdSTejun Heo 8848a1270cdSTejun Heo work = kmalloc(sizeof(*work), GFP_ATOMIC); 8858a1270cdSTejun Heo if (work) { 8868a1270cdSTejun Heo *work = *base_work; 8878a1270cdSTejun Heo work->nr_pages = nr_pages; 8888a1270cdSTejun Heo work->auto_free = 1; 8898a1270cdSTejun Heo wb_queue_work(wb, work); 8908a1270cdSTejun Heo continue; 891db125360STejun Heo } 8928a1270cdSTejun Heo 8938a1270cdSTejun Heo /* alloc failed, execute synchronously using on-stack fallback */ 8948a1270cdSTejun Heo work = &fallback_work; 8958a1270cdSTejun Heo *work = *base_work; 8968a1270cdSTejun Heo work->nr_pages = nr_pages; 8978a1270cdSTejun Heo work->auto_free = 0; 8988a1270cdSTejun Heo work->done = &fallback_work_done; 8998a1270cdSTejun Heo 9008a1270cdSTejun Heo wb_queue_work(wb, work); 9018a1270cdSTejun Heo 902b817525aSTejun Heo /* 903b817525aSTejun Heo * Pin @wb so that it stays on @bdi->wb_list. This allows 904b817525aSTejun Heo * continuing iteration from @wb after dropping and 905b817525aSTejun Heo * regrabbing rcu read lock. 906b817525aSTejun Heo */ 907b817525aSTejun Heo wb_get(wb); 908b817525aSTejun Heo last_wb = wb; 909b817525aSTejun Heo 910db125360STejun Heo rcu_read_unlock(); 9118a1270cdSTejun Heo wb_wait_for_completion(bdi, &fallback_work_done); 912db125360STejun Heo goto restart; 913db125360STejun Heo } 914db125360STejun Heo rcu_read_unlock(); 915b817525aSTejun Heo 916b817525aSTejun Heo if (last_wb) 917b817525aSTejun Heo wb_put(last_wb); 918db125360STejun Heo } 919db125360STejun Heo 920a1a0e23eSTejun Heo /** 921a1a0e23eSTejun Heo * cgroup_writeback_umount - flush inode wb switches for umount 922a1a0e23eSTejun Heo * 923a1a0e23eSTejun Heo * This function is called when a super_block is about to be destroyed and 924a1a0e23eSTejun Heo * flushes in-flight inode wb switches. An inode wb switch goes through 925a1a0e23eSTejun Heo * RCU and then workqueue, so the two need to be flushed in order to ensure 926a1a0e23eSTejun Heo * that all previously scheduled switches are finished. As wb switches are 927a1a0e23eSTejun Heo * rare occurrences and synchronize_rcu() can take a while, perform 928a1a0e23eSTejun Heo * flushing iff wb switches are in flight. 929a1a0e23eSTejun Heo */ 930a1a0e23eSTejun Heo void cgroup_writeback_umount(void) 931a1a0e23eSTejun Heo { 932a1a0e23eSTejun Heo if (atomic_read(&isw_nr_in_flight)) { 933ec084de9SJiufei Xue /* 934ec084de9SJiufei Xue * Use rcu_barrier() to wait for all pending callbacks to 935ec084de9SJiufei Xue * ensure that all in-flight wb switches are in the workqueue. 936ec084de9SJiufei Xue */ 937ec084de9SJiufei Xue rcu_barrier(); 938a1a0e23eSTejun Heo flush_workqueue(isw_wq); 939a1a0e23eSTejun Heo } 940a1a0e23eSTejun Heo } 941a1a0e23eSTejun Heo 942a1a0e23eSTejun Heo static int __init cgroup_writeback_init(void) 943a1a0e23eSTejun Heo { 944a1a0e23eSTejun Heo isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0); 945a1a0e23eSTejun Heo if (!isw_wq) 946a1a0e23eSTejun Heo return -ENOMEM; 947a1a0e23eSTejun Heo return 0; 948a1a0e23eSTejun Heo } 949a1a0e23eSTejun Heo fs_initcall(cgroup_writeback_init); 950a1a0e23eSTejun Heo 951f2b65121STejun Heo #else /* CONFIG_CGROUP_WRITEBACK */ 952f2b65121STejun Heo 9537fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 9547fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 9557fc5854fSTejun Heo 95687e1d789STejun Heo static struct bdi_writeback * 95787e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 95887e1d789STejun Heo __releases(&inode->i_lock) 95987e1d789STejun Heo __acquires(&wb->list_lock) 96087e1d789STejun Heo { 96187e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 96287e1d789STejun Heo 96387e1d789STejun Heo spin_unlock(&inode->i_lock); 96487e1d789STejun Heo spin_lock(&wb->list_lock); 96587e1d789STejun Heo return wb; 96687e1d789STejun Heo } 96787e1d789STejun Heo 96887e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 96987e1d789STejun Heo __acquires(&wb->list_lock) 97087e1d789STejun Heo { 97187e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 97287e1d789STejun Heo 97387e1d789STejun Heo spin_lock(&wb->list_lock); 97487e1d789STejun Heo return wb; 97587e1d789STejun Heo } 97687e1d789STejun Heo 977f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 978f2b65121STejun Heo { 979f2b65121STejun Heo return nr_pages; 980f2b65121STejun Heo } 981f2b65121STejun Heo 982db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 983db125360STejun Heo struct wb_writeback_work *base_work, 984db125360STejun Heo bool skip_if_busy) 985db125360STejun Heo { 986db125360STejun Heo might_sleep(); 987db125360STejun Heo 988006a0973STejun Heo if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) { 989db125360STejun Heo base_work->auto_free = 0; 990db125360STejun Heo wb_queue_work(&bdi->wb, base_work); 991db125360STejun Heo } 992db125360STejun Heo } 993db125360STejun Heo 994703c2708STejun Heo #endif /* CONFIG_CGROUP_WRITEBACK */ 995703c2708STejun Heo 996e8e8a0c6SJens Axboe /* 997e8e8a0c6SJens Axboe * Add in the number of potentially dirty inodes, because each inode 998e8e8a0c6SJens Axboe * write can dirty pagecache in the underlying blockdev. 999e8e8a0c6SJens Axboe */ 1000e8e8a0c6SJens Axboe static unsigned long get_nr_dirty_pages(void) 1001e8e8a0c6SJens Axboe { 1002e8e8a0c6SJens Axboe return global_node_page_state(NR_FILE_DIRTY) + 1003e8e8a0c6SJens Axboe global_node_page_state(NR_UNSTABLE_NFS) + 1004e8e8a0c6SJens Axboe get_nr_dirty_inodes(); 1005e8e8a0c6SJens Axboe } 1006e8e8a0c6SJens Axboe 1007e8e8a0c6SJens Axboe static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason) 1008b6e51316SJens Axboe { 1009c00ddad3STejun Heo if (!wb_has_dirty_io(wb)) 1010c00ddad3STejun Heo return; 1011c00ddad3STejun Heo 1012c00ddad3STejun Heo /* 1013aac8d41cSJens Axboe * All callers of this function want to start writeback of all 1014aac8d41cSJens Axboe * dirty pages. Places like vmscan can call this at a very 1015aac8d41cSJens Axboe * high frequency, causing pointless allocations of tons of 1016aac8d41cSJens Axboe * work items and keeping the flusher threads busy retrieving 1017aac8d41cSJens Axboe * that work. Ensure that we only allow one of them pending and 101885009b4fSJens Axboe * inflight at the time. 1019aac8d41cSJens Axboe */ 102085009b4fSJens Axboe if (test_bit(WB_start_all, &wb->state) || 102185009b4fSJens Axboe test_and_set_bit(WB_start_all, &wb->state)) 1022aac8d41cSJens Axboe return; 1023aac8d41cSJens Axboe 102485009b4fSJens Axboe wb->start_all_reason = reason; 1025c00ddad3STejun Heo wb_wakeup(wb); 1026d3ddec76SWu Fengguang } 1027d3ddec76SWu Fengguang 1028c5444198SChristoph Hellwig /** 10299ecf4866STejun Heo * wb_start_background_writeback - start background writeback 10309ecf4866STejun Heo * @wb: bdi_writback to write from 1031c5444198SChristoph Hellwig * 1032c5444198SChristoph Hellwig * Description: 10336585027aSJan Kara * This makes sure WB_SYNC_NONE background writeback happens. When 10349ecf4866STejun Heo * this function returns, it is only guaranteed that for given wb 10356585027aSJan Kara * some IO is happening if we are over background dirty threshold. 10366585027aSJan Kara * Caller need not hold sb s_umount semaphore. 1037c5444198SChristoph Hellwig */ 10389ecf4866STejun Heo void wb_start_background_writeback(struct bdi_writeback *wb) 1039c5444198SChristoph Hellwig { 10406585027aSJan Kara /* 10416585027aSJan Kara * We just wake up the flusher thread. It will perform background 10426585027aSJan Kara * writeback as soon as there is no other work to do. 10436585027aSJan Kara */ 10445634cc2aSTejun Heo trace_writeback_wake_background(wb); 10459ecf4866STejun Heo wb_wakeup(wb); 10461da177e4SLinus Torvalds } 10471da177e4SLinus Torvalds 10481da177e4SLinus Torvalds /* 1049a66979abSDave Chinner * Remove the inode from the writeback list it is on. 1050a66979abSDave Chinner */ 1051c7f54084SDave Chinner void inode_io_list_del(struct inode *inode) 1052a66979abSDave Chinner { 105387e1d789STejun Heo struct bdi_writeback *wb; 1054a66979abSDave Chinner 105587e1d789STejun Heo wb = inode_to_wb_and_lock_list(inode); 1056c7f54084SDave Chinner inode_io_list_del_locked(inode, wb); 105752ebea74STejun Heo spin_unlock(&wb->list_lock); 1058f758eeabSChristoph Hellwig } 1059a66979abSDave Chinner 1060a66979abSDave Chinner /* 10616c60d2b5SDave Chinner * mark an inode as under writeback on the sb 10626c60d2b5SDave Chinner */ 10636c60d2b5SDave Chinner void sb_mark_inode_writeback(struct inode *inode) 10646c60d2b5SDave Chinner { 10656c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 10666c60d2b5SDave Chinner unsigned long flags; 10676c60d2b5SDave Chinner 10686c60d2b5SDave Chinner if (list_empty(&inode->i_wb_list)) { 10696c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 10709a46b04fSBrian Foster if (list_empty(&inode->i_wb_list)) { 10716c60d2b5SDave Chinner list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); 10729a46b04fSBrian Foster trace_sb_mark_inode_writeback(inode); 10739a46b04fSBrian Foster } 10746c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 10756c60d2b5SDave Chinner } 10766c60d2b5SDave Chinner } 10776c60d2b5SDave Chinner 10786c60d2b5SDave Chinner /* 10796c60d2b5SDave Chinner * clear an inode as under writeback on the sb 10806c60d2b5SDave Chinner */ 10816c60d2b5SDave Chinner void sb_clear_inode_writeback(struct inode *inode) 10826c60d2b5SDave Chinner { 10836c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 10846c60d2b5SDave Chinner unsigned long flags; 10856c60d2b5SDave Chinner 10866c60d2b5SDave Chinner if (!list_empty(&inode->i_wb_list)) { 10876c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 10889a46b04fSBrian Foster if (!list_empty(&inode->i_wb_list)) { 10896c60d2b5SDave Chinner list_del_init(&inode->i_wb_list); 10909a46b04fSBrian Foster trace_sb_clear_inode_writeback(inode); 10919a46b04fSBrian Foster } 10926c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 10936c60d2b5SDave Chinner } 10946c60d2b5SDave Chinner } 10956c60d2b5SDave Chinner 10966c60d2b5SDave Chinner /* 10976610a0bcSAndrew Morton * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 10986610a0bcSAndrew Morton * furthest end of its superblock's dirty-inode list. 10996610a0bcSAndrew Morton * 11006610a0bcSAndrew Morton * Before stamping the inode's ->dirtied_when, we check to see whether it is 110166f3b8e2SJens Axboe * already the most-recently-dirtied inode on the b_dirty list. If that is 11026610a0bcSAndrew Morton * the case then the inode must have been redirtied while it was being written 11036610a0bcSAndrew Morton * out and we don't reset its dirtied_when. 11046610a0bcSAndrew Morton */ 1105f758eeabSChristoph Hellwig static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) 11066610a0bcSAndrew Morton { 110703ba3782SJens Axboe if (!list_empty(&wb->b_dirty)) { 110866f3b8e2SJens Axboe struct inode *tail; 11096610a0bcSAndrew Morton 11107ccf19a8SNick Piggin tail = wb_inode(wb->b_dirty.next); 111166f3b8e2SJens Axboe if (time_before(inode->dirtied_when, tail->dirtied_when)) 11126610a0bcSAndrew Morton inode->dirtied_when = jiffies; 11136610a0bcSAndrew Morton } 1114c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty); 11156610a0bcSAndrew Morton } 11166610a0bcSAndrew Morton 11176610a0bcSAndrew Morton /* 111866f3b8e2SJens Axboe * requeue inode for re-scanning after bdi->b_io list is exhausted. 1119c986d1e2SAndrew Morton */ 1120f758eeabSChristoph Hellwig static void requeue_io(struct inode *inode, struct bdi_writeback *wb) 1121c986d1e2SAndrew Morton { 1122c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_more_io); 1123c986d1e2SAndrew Morton } 1124c986d1e2SAndrew Morton 11251c0eeaf5SJoern Engel static void inode_sync_complete(struct inode *inode) 11261c0eeaf5SJoern Engel { 1127365b94aeSJan Kara inode->i_state &= ~I_SYNC; 11284eff96ddSJan Kara /* If inode is clean an unused, put it into LRU now... */ 11294eff96ddSJan Kara inode_add_lru(inode); 1130365b94aeSJan Kara /* Waiters must see I_SYNC cleared before being woken up */ 11311c0eeaf5SJoern Engel smp_mb(); 11321c0eeaf5SJoern Engel wake_up_bit(&inode->i_state, __I_SYNC); 11331c0eeaf5SJoern Engel } 11341c0eeaf5SJoern Engel 1135d2caa3c5SJeff Layton static bool inode_dirtied_after(struct inode *inode, unsigned long t) 1136d2caa3c5SJeff Layton { 1137d2caa3c5SJeff Layton bool ret = time_after(inode->dirtied_when, t); 1138d2caa3c5SJeff Layton #ifndef CONFIG_64BIT 1139d2caa3c5SJeff Layton /* 1140d2caa3c5SJeff Layton * For inodes being constantly redirtied, dirtied_when can get stuck. 1141d2caa3c5SJeff Layton * It _appears_ to be in the future, but is actually in distant past. 1142d2caa3c5SJeff Layton * This test is necessary to prevent such wrapped-around relative times 11435b0830cbSJens Axboe * from permanently stopping the whole bdi writeback. 1144d2caa3c5SJeff Layton */ 1145d2caa3c5SJeff Layton ret = ret && time_before_eq(inode->dirtied_when, jiffies); 1146d2caa3c5SJeff Layton #endif 1147d2caa3c5SJeff Layton return ret; 1148d2caa3c5SJeff Layton } 1149d2caa3c5SJeff Layton 11500ae45f63STheodore Ts'o #define EXPIRE_DIRTY_ATIME 0x0001 11510ae45f63STheodore Ts'o 1152c986d1e2SAndrew Morton /* 11530e2f2b23SWang Sheng-Hui * Move expired (dirtied before work->older_than_this) dirty inodes from 1154697e6fedSJan Kara * @delaying_queue to @dispatch_queue. 11552c136579SFengguang Wu */ 1156e84d0a4fSWu Fengguang static int move_expired_inodes(struct list_head *delaying_queue, 11572c136579SFengguang Wu struct list_head *dispatch_queue, 11580ae45f63STheodore Ts'o int flags, 1159ad4e38ddSCurt Wohlgemuth struct wb_writeback_work *work) 11602c136579SFengguang Wu { 11610ae45f63STheodore Ts'o unsigned long *older_than_this = NULL; 11620ae45f63STheodore Ts'o unsigned long expire_time; 11635c03449dSShaohua Li LIST_HEAD(tmp); 11645c03449dSShaohua Li struct list_head *pos, *node; 1165cf137307SJens Axboe struct super_block *sb = NULL; 11665c03449dSShaohua Li struct inode *inode; 1167cf137307SJens Axboe int do_sb_sort = 0; 1168e84d0a4fSWu Fengguang int moved = 0; 11695c03449dSShaohua Li 11700ae45f63STheodore Ts'o if ((flags & EXPIRE_DIRTY_ATIME) == 0) 11710ae45f63STheodore Ts'o older_than_this = work->older_than_this; 1172a2f48706STheodore Ts'o else if (!work->for_sync) { 1173a2f48706STheodore Ts'o expire_time = jiffies - (dirtytime_expire_interval * HZ); 11740ae45f63STheodore Ts'o older_than_this = &expire_time; 11750ae45f63STheodore Ts'o } 11762c136579SFengguang Wu while (!list_empty(delaying_queue)) { 11777ccf19a8SNick Piggin inode = wb_inode(delaying_queue->prev); 11780ae45f63STheodore Ts'o if (older_than_this && 11790ae45f63STheodore Ts'o inode_dirtied_after(inode, *older_than_this)) 11802c136579SFengguang Wu break; 1181c7f54084SDave Chinner list_move(&inode->i_io_list, &tmp); 1182a8855990SJan Kara moved++; 11830ae45f63STheodore Ts'o if (flags & EXPIRE_DIRTY_ATIME) 11840ae45f63STheodore Ts'o set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state); 1185a8855990SJan Kara if (sb_is_blkdev_sb(inode->i_sb)) 1186a8855990SJan Kara continue; 1187cf137307SJens Axboe if (sb && sb != inode->i_sb) 1188cf137307SJens Axboe do_sb_sort = 1; 1189cf137307SJens Axboe sb = inode->i_sb; 11905c03449dSShaohua Li } 11915c03449dSShaohua Li 1192cf137307SJens Axboe /* just one sb in list, splice to dispatch_queue and we're done */ 1193cf137307SJens Axboe if (!do_sb_sort) { 1194cf137307SJens Axboe list_splice(&tmp, dispatch_queue); 1195e84d0a4fSWu Fengguang goto out; 1196cf137307SJens Axboe } 1197cf137307SJens Axboe 11985c03449dSShaohua Li /* Move inodes from one superblock together */ 11995c03449dSShaohua Li while (!list_empty(&tmp)) { 12007ccf19a8SNick Piggin sb = wb_inode(tmp.prev)->i_sb; 12015c03449dSShaohua Li list_for_each_prev_safe(pos, node, &tmp) { 12027ccf19a8SNick Piggin inode = wb_inode(pos); 12035c03449dSShaohua Li if (inode->i_sb == sb) 1204c7f54084SDave Chinner list_move(&inode->i_io_list, dispatch_queue); 12052c136579SFengguang Wu } 12062c136579SFengguang Wu } 1207e84d0a4fSWu Fengguang out: 1208e84d0a4fSWu Fengguang return moved; 12095c03449dSShaohua Li } 12102c136579SFengguang Wu 12112c136579SFengguang Wu /* 12122c136579SFengguang Wu * Queue all expired dirty inodes for io, eldest first. 12134ea879b9SWu Fengguang * Before 12144ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 12154ea879b9SWu Fengguang * =============> gf edc BA 12164ea879b9SWu Fengguang * After 12174ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 12184ea879b9SWu Fengguang * =============> g fBAedc 12194ea879b9SWu Fengguang * | 12204ea879b9SWu Fengguang * +--> dequeue for IO 12212c136579SFengguang Wu */ 1222ad4e38ddSCurt Wohlgemuth static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work) 12232c136579SFengguang Wu { 1224e84d0a4fSWu Fengguang int moved; 12250ae45f63STheodore Ts'o 1226f758eeabSChristoph Hellwig assert_spin_locked(&wb->list_lock); 12274ea879b9SWu Fengguang list_splice_init(&wb->b_more_io, &wb->b_io); 12280ae45f63STheodore Ts'o moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work); 12290ae45f63STheodore Ts'o moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, 12300ae45f63STheodore Ts'o EXPIRE_DIRTY_ATIME, work); 1231d6c10f1fSTejun Heo if (moved) 1232d6c10f1fSTejun Heo wb_io_lists_populated(wb); 1233ad4e38ddSCurt Wohlgemuth trace_writeback_queue_io(wb, work, moved); 123466f3b8e2SJens Axboe } 123566f3b8e2SJens Axboe 1236a9185b41SChristoph Hellwig static int write_inode(struct inode *inode, struct writeback_control *wbc) 123766f3b8e2SJens Axboe { 12389fb0a7daSTejun Heo int ret; 12399fb0a7daSTejun Heo 12409fb0a7daSTejun Heo if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { 12419fb0a7daSTejun Heo trace_writeback_write_inode_start(inode, wbc); 12429fb0a7daSTejun Heo ret = inode->i_sb->s_op->write_inode(inode, wbc); 12439fb0a7daSTejun Heo trace_writeback_write_inode(inode, wbc); 12449fb0a7daSTejun Heo return ret; 12459fb0a7daSTejun Heo } 124603ba3782SJens Axboe return 0; 124766f3b8e2SJens Axboe } 124808d8e974SFengguang Wu 12492c136579SFengguang Wu /* 1250169ebd90SJan Kara * Wait for writeback on an inode to complete. Called with i_lock held. 1251169ebd90SJan Kara * Caller must make sure inode cannot go away when we drop i_lock. 125201c03194SChristoph Hellwig */ 1253169ebd90SJan Kara static void __inode_wait_for_writeback(struct inode *inode) 1254169ebd90SJan Kara __releases(inode->i_lock) 1255169ebd90SJan Kara __acquires(inode->i_lock) 125601c03194SChristoph Hellwig { 125701c03194SChristoph Hellwig DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); 125801c03194SChristoph Hellwig wait_queue_head_t *wqh; 125901c03194SChristoph Hellwig 126001c03194SChristoph Hellwig wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 126158a9d3d8SRichard Kennedy while (inode->i_state & I_SYNC) { 1262250df6edSDave Chinner spin_unlock(&inode->i_lock); 126374316201SNeilBrown __wait_on_bit(wqh, &wq, bit_wait, 126474316201SNeilBrown TASK_UNINTERRUPTIBLE); 1265250df6edSDave Chinner spin_lock(&inode->i_lock); 126658a9d3d8SRichard Kennedy } 126701c03194SChristoph Hellwig } 126801c03194SChristoph Hellwig 126901c03194SChristoph Hellwig /* 1270169ebd90SJan Kara * Wait for writeback on an inode to complete. Caller must have inode pinned. 1271169ebd90SJan Kara */ 1272169ebd90SJan Kara void inode_wait_for_writeback(struct inode *inode) 1273169ebd90SJan Kara { 1274169ebd90SJan Kara spin_lock(&inode->i_lock); 1275169ebd90SJan Kara __inode_wait_for_writeback(inode); 1276169ebd90SJan Kara spin_unlock(&inode->i_lock); 1277169ebd90SJan Kara } 1278169ebd90SJan Kara 1279169ebd90SJan Kara /* 1280169ebd90SJan Kara * Sleep until I_SYNC is cleared. This function must be called with i_lock 1281169ebd90SJan Kara * held and drops it. It is aimed for callers not holding any inode reference 1282169ebd90SJan Kara * so once i_lock is dropped, inode can go away. 1283169ebd90SJan Kara */ 1284169ebd90SJan Kara static void inode_sleep_on_writeback(struct inode *inode) 1285169ebd90SJan Kara __releases(inode->i_lock) 1286169ebd90SJan Kara { 1287169ebd90SJan Kara DEFINE_WAIT(wait); 1288169ebd90SJan Kara wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1289169ebd90SJan Kara int sleep; 1290169ebd90SJan Kara 1291169ebd90SJan Kara prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1292169ebd90SJan Kara sleep = inode->i_state & I_SYNC; 1293169ebd90SJan Kara spin_unlock(&inode->i_lock); 1294169ebd90SJan Kara if (sleep) 1295169ebd90SJan Kara schedule(); 1296169ebd90SJan Kara finish_wait(wqh, &wait); 1297169ebd90SJan Kara } 1298169ebd90SJan Kara 1299169ebd90SJan Kara /* 1300ccb26b5aSJan Kara * Find proper writeback list for the inode depending on its current state and 1301ccb26b5aSJan Kara * possibly also change of its state while we were doing writeback. Here we 1302ccb26b5aSJan Kara * handle things such as livelock prevention or fairness of writeback among 1303ccb26b5aSJan Kara * inodes. This function can be called only by flusher thread - noone else 1304ccb26b5aSJan Kara * processes all inodes in writeback lists and requeueing inodes behind flusher 1305ccb26b5aSJan Kara * thread's back can have unexpected consequences. 1306ccb26b5aSJan Kara */ 1307ccb26b5aSJan Kara static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, 1308ccb26b5aSJan Kara struct writeback_control *wbc) 1309ccb26b5aSJan Kara { 1310ccb26b5aSJan Kara if (inode->i_state & I_FREEING) 1311ccb26b5aSJan Kara return; 1312ccb26b5aSJan Kara 1313ccb26b5aSJan Kara /* 1314ccb26b5aSJan Kara * Sync livelock prevention. Each inode is tagged and synced in one 1315ccb26b5aSJan Kara * shot. If still dirty, it will be redirty_tail()'ed below. Update 1316ccb26b5aSJan Kara * the dirty time to prevent enqueue and sync it again. 1317ccb26b5aSJan Kara */ 1318ccb26b5aSJan Kara if ((inode->i_state & I_DIRTY) && 1319ccb26b5aSJan Kara (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) 1320ccb26b5aSJan Kara inode->dirtied_when = jiffies; 1321ccb26b5aSJan Kara 13224f8ad655SJan Kara if (wbc->pages_skipped) { 13234f8ad655SJan Kara /* 13244f8ad655SJan Kara * writeback is not making progress due to locked 13254f8ad655SJan Kara * buffers. Skip this inode for now. 13264f8ad655SJan Kara */ 13274f8ad655SJan Kara redirty_tail(inode, wb); 13284f8ad655SJan Kara return; 13294f8ad655SJan Kara } 13304f8ad655SJan Kara 1331ccb26b5aSJan Kara if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { 1332ccb26b5aSJan Kara /* 1333ccb26b5aSJan Kara * We didn't write back all the pages. nfs_writepages() 1334ccb26b5aSJan Kara * sometimes bales out without doing anything. 1335ccb26b5aSJan Kara */ 1336ccb26b5aSJan Kara if (wbc->nr_to_write <= 0) { 1337ccb26b5aSJan Kara /* Slice used up. Queue for next turn. */ 1338ccb26b5aSJan Kara requeue_io(inode, wb); 1339ccb26b5aSJan Kara } else { 1340ccb26b5aSJan Kara /* 1341ccb26b5aSJan Kara * Writeback blocked by something other than 1342ccb26b5aSJan Kara * congestion. Delay the inode for some time to 1343ccb26b5aSJan Kara * avoid spinning on the CPU (100% iowait) 1344ccb26b5aSJan Kara * retrying writeback of the dirty page/inode 1345ccb26b5aSJan Kara * that cannot be performed immediately. 1346ccb26b5aSJan Kara */ 1347ccb26b5aSJan Kara redirty_tail(inode, wb); 1348ccb26b5aSJan Kara } 1349ccb26b5aSJan Kara } else if (inode->i_state & I_DIRTY) { 1350ccb26b5aSJan Kara /* 1351ccb26b5aSJan Kara * Filesystems can dirty the inode during writeback operations, 1352ccb26b5aSJan Kara * such as delayed allocation during submission or metadata 1353ccb26b5aSJan Kara * updates after data IO completion. 1354ccb26b5aSJan Kara */ 1355ccb26b5aSJan Kara redirty_tail(inode, wb); 13560ae45f63STheodore Ts'o } else if (inode->i_state & I_DIRTY_TIME) { 1357a2f48706STheodore Ts'o inode->dirtied_when = jiffies; 1358c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty_time); 1359ccb26b5aSJan Kara } else { 1360ccb26b5aSJan Kara /* The inode is clean. Remove from writeback lists. */ 1361c7f54084SDave Chinner inode_io_list_del_locked(inode, wb); 1362ccb26b5aSJan Kara } 1363ccb26b5aSJan Kara } 1364ccb26b5aSJan Kara 1365ccb26b5aSJan Kara /* 13664f8ad655SJan Kara * Write out an inode and its dirty pages. Do not update the writeback list 13674f8ad655SJan Kara * linkage. That is left to the caller. The caller is also responsible for 13684f8ad655SJan Kara * setting I_SYNC flag and calling inode_sync_complete() to clear it. 13691da177e4SLinus Torvalds */ 13701da177e4SLinus Torvalds static int 1371cd8ed2a4SYan Hong __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 13721da177e4SLinus Torvalds { 13731da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 1374251d6a47SWu Fengguang long nr_to_write = wbc->nr_to_write; 137501c03194SChristoph Hellwig unsigned dirty; 13761da177e4SLinus Torvalds int ret; 13771da177e4SLinus Torvalds 13784f8ad655SJan Kara WARN_ON(!(inode->i_state & I_SYNC)); 13791da177e4SLinus Torvalds 13809fb0a7daSTejun Heo trace_writeback_single_inode_start(inode, wbc, nr_to_write); 13819fb0a7daSTejun Heo 13821da177e4SLinus Torvalds ret = do_writepages(mapping, wbc); 13831da177e4SLinus Torvalds 138426821ed4SChristoph Hellwig /* 138526821ed4SChristoph Hellwig * Make sure to wait on the data before writing out the metadata. 138626821ed4SChristoph Hellwig * This is important for filesystems that modify metadata on data 13877747bd4bSDave Chinner * I/O completion. We don't do it for sync(2) writeback because it has a 13887747bd4bSDave Chinner * separate, external IO completion path and ->sync_fs for guaranteeing 13897747bd4bSDave Chinner * inode metadata is written back correctly. 139026821ed4SChristoph Hellwig */ 13917747bd4bSDave Chinner if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { 139226821ed4SChristoph Hellwig int err = filemap_fdatawait(mapping); 13931da177e4SLinus Torvalds if (ret == 0) 13941da177e4SLinus Torvalds ret = err; 13951da177e4SLinus Torvalds } 13961da177e4SLinus Torvalds 13975547e8aaSDmitry Monakhov /* 13985547e8aaSDmitry Monakhov * Some filesystems may redirty the inode during the writeback 13995547e8aaSDmitry Monakhov * due to delalloc, clear dirty metadata flags right before 14005547e8aaSDmitry Monakhov * write_inode() 14015547e8aaSDmitry Monakhov */ 1402250df6edSDave Chinner spin_lock(&inode->i_lock); 14039c6ac78eSTejun Heo 14045547e8aaSDmitry Monakhov dirty = inode->i_state & I_DIRTY; 1405a2f48706STheodore Ts'o if (inode->i_state & I_DIRTY_TIME) { 14060e11f644SChristoph Hellwig if ((dirty & I_DIRTY_INODE) || 1407dc5ff2b1SJan Kara wbc->sync_mode == WB_SYNC_ALL || 1408a2f48706STheodore Ts'o unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) || 1409a2f48706STheodore Ts'o unlikely(time_after(jiffies, 1410a2f48706STheodore Ts'o (inode->dirtied_time_when + 1411a2f48706STheodore Ts'o dirtytime_expire_interval * HZ)))) { 14120ae45f63STheodore Ts'o dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED; 14130ae45f63STheodore Ts'o trace_writeback_lazytime(inode); 14140ae45f63STheodore Ts'o } 1415a2f48706STheodore Ts'o } else 1416a2f48706STheodore Ts'o inode->i_state &= ~I_DIRTY_TIME_EXPIRED; 14170ae45f63STheodore Ts'o inode->i_state &= ~dirty; 14189c6ac78eSTejun Heo 14199c6ac78eSTejun Heo /* 14209c6ac78eSTejun Heo * Paired with smp_mb() in __mark_inode_dirty(). This allows 14219c6ac78eSTejun Heo * __mark_inode_dirty() to test i_state without grabbing i_lock - 14229c6ac78eSTejun Heo * either they see the I_DIRTY bits cleared or we see the dirtied 14239c6ac78eSTejun Heo * inode. 14249c6ac78eSTejun Heo * 14259c6ac78eSTejun Heo * I_DIRTY_PAGES is always cleared together above even if @mapping 14269c6ac78eSTejun Heo * still has dirty pages. The flag is reinstated after smp_mb() if 14279c6ac78eSTejun Heo * necessary. This guarantees that either __mark_inode_dirty() 14289c6ac78eSTejun Heo * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. 14299c6ac78eSTejun Heo */ 14309c6ac78eSTejun Heo smp_mb(); 14319c6ac78eSTejun Heo 14329c6ac78eSTejun Heo if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 14339c6ac78eSTejun Heo inode->i_state |= I_DIRTY_PAGES; 14349c6ac78eSTejun Heo 1435250df6edSDave Chinner spin_unlock(&inode->i_lock); 14369c6ac78eSTejun Heo 14370ae45f63STheodore Ts'o if (dirty & I_DIRTY_TIME) 14380ae45f63STheodore Ts'o mark_inode_dirty_sync(inode); 143926821ed4SChristoph Hellwig /* Don't write the inode if only I_DIRTY_PAGES was set */ 14400ae45f63STheodore Ts'o if (dirty & ~I_DIRTY_PAGES) { 1441a9185b41SChristoph Hellwig int err = write_inode(inode, wbc); 14421da177e4SLinus Torvalds if (ret == 0) 14431da177e4SLinus Torvalds ret = err; 14441da177e4SLinus Torvalds } 14454f8ad655SJan Kara trace_writeback_single_inode(inode, wbc, nr_to_write); 14464f8ad655SJan Kara return ret; 14474f8ad655SJan Kara } 14484f8ad655SJan Kara 14494f8ad655SJan Kara /* 14504f8ad655SJan Kara * Write out an inode's dirty pages. Either the caller has an active reference 14514f8ad655SJan Kara * on the inode or the inode has I_WILL_FREE set. 14524f8ad655SJan Kara * 14534f8ad655SJan Kara * This function is designed to be called for writing back one inode which 14544f8ad655SJan Kara * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode() 14554f8ad655SJan Kara * and does more profound writeback list handling in writeback_sb_inodes(). 14564f8ad655SJan Kara */ 1457aaf25593STejun Heo static int writeback_single_inode(struct inode *inode, 14584f8ad655SJan Kara struct writeback_control *wbc) 14594f8ad655SJan Kara { 1460aaf25593STejun Heo struct bdi_writeback *wb; 14614f8ad655SJan Kara int ret = 0; 14624f8ad655SJan Kara 14634f8ad655SJan Kara spin_lock(&inode->i_lock); 14644f8ad655SJan Kara if (!atomic_read(&inode->i_count)) 14654f8ad655SJan Kara WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 14664f8ad655SJan Kara else 14674f8ad655SJan Kara WARN_ON(inode->i_state & I_WILL_FREE); 14684f8ad655SJan Kara 14694f8ad655SJan Kara if (inode->i_state & I_SYNC) { 14704f8ad655SJan Kara if (wbc->sync_mode != WB_SYNC_ALL) 14714f8ad655SJan Kara goto out; 14724f8ad655SJan Kara /* 1473169ebd90SJan Kara * It's a data-integrity sync. We must wait. Since callers hold 1474169ebd90SJan Kara * inode reference or inode has I_WILL_FREE set, it cannot go 1475169ebd90SJan Kara * away under us. 14764f8ad655SJan Kara */ 1477169ebd90SJan Kara __inode_wait_for_writeback(inode); 14784f8ad655SJan Kara } 14794f8ad655SJan Kara WARN_ON(inode->i_state & I_SYNC); 14804f8ad655SJan Kara /* 1481f9b0e058SJan Kara * Skip inode if it is clean and we have no outstanding writeback in 1482f9b0e058SJan Kara * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this 1483f9b0e058SJan Kara * function since flusher thread may be doing for example sync in 1484f9b0e058SJan Kara * parallel and if we move the inode, it could get skipped. So here we 1485f9b0e058SJan Kara * make sure inode is on some writeback list and leave it there unless 1486f9b0e058SJan Kara * we have completely cleaned the inode. 14874f8ad655SJan Kara */ 14880ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL) && 1489f9b0e058SJan Kara (wbc->sync_mode != WB_SYNC_ALL || 1490f9b0e058SJan Kara !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) 14914f8ad655SJan Kara goto out; 14924f8ad655SJan Kara inode->i_state |= I_SYNC; 1493b16b1debSTejun Heo wbc_attach_and_unlock_inode(wbc, inode); 14944f8ad655SJan Kara 1495cd8ed2a4SYan Hong ret = __writeback_single_inode(inode, wbc); 14961da177e4SLinus Torvalds 1497b16b1debSTejun Heo wbc_detach_inode(wbc); 1498aaf25593STejun Heo 1499aaf25593STejun Heo wb = inode_to_wb_and_lock_list(inode); 1500250df6edSDave Chinner spin_lock(&inode->i_lock); 15014f8ad655SJan Kara /* 15024f8ad655SJan Kara * If inode is clean, remove it from writeback lists. Otherwise don't 15034f8ad655SJan Kara * touch it. See comment above for explanation. 15044f8ad655SJan Kara */ 15050ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1506c7f54084SDave Chinner inode_io_list_del_locked(inode, wb); 15074f8ad655SJan Kara spin_unlock(&wb->list_lock); 15081c0eeaf5SJoern Engel inode_sync_complete(inode); 15094f8ad655SJan Kara out: 15104f8ad655SJan Kara spin_unlock(&inode->i_lock); 15111da177e4SLinus Torvalds return ret; 15121da177e4SLinus Torvalds } 15131da177e4SLinus Torvalds 1514a88a341aSTejun Heo static long writeback_chunk_size(struct bdi_writeback *wb, 15151a12d8bdSWu Fengguang struct wb_writeback_work *work) 1516d46db3d5SWu Fengguang { 1517d46db3d5SWu Fengguang long pages; 1518d46db3d5SWu Fengguang 1519d46db3d5SWu Fengguang /* 1520d46db3d5SWu Fengguang * WB_SYNC_ALL mode does livelock avoidance by syncing dirty 1521d46db3d5SWu Fengguang * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX 1522d46db3d5SWu Fengguang * here avoids calling into writeback_inodes_wb() more than once. 1523d46db3d5SWu Fengguang * 1524d46db3d5SWu Fengguang * The intended call sequence for WB_SYNC_ALL writeback is: 1525d46db3d5SWu Fengguang * 1526d46db3d5SWu Fengguang * wb_writeback() 1527d46db3d5SWu Fengguang * writeback_sb_inodes() <== called only once 1528d46db3d5SWu Fengguang * write_cache_pages() <== called once for each inode 1529d46db3d5SWu Fengguang * (quickly) tag currently dirty pages 1530d46db3d5SWu Fengguang * (maybe slowly) sync all tagged pages 1531d46db3d5SWu Fengguang */ 1532d46db3d5SWu Fengguang if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) 1533d46db3d5SWu Fengguang pages = LONG_MAX; 15341a12d8bdSWu Fengguang else { 1535a88a341aSTejun Heo pages = min(wb->avg_write_bandwidth / 2, 1536dcc25ae7STejun Heo global_wb_domain.dirty_limit / DIRTY_SCOPE); 15371a12d8bdSWu Fengguang pages = min(pages, work->nr_pages); 15381a12d8bdSWu Fengguang pages = round_down(pages + MIN_WRITEBACK_PAGES, 15391a12d8bdSWu Fengguang MIN_WRITEBACK_PAGES); 15401a12d8bdSWu Fengguang } 1541d46db3d5SWu Fengguang 1542d46db3d5SWu Fengguang return pages; 1543d46db3d5SWu Fengguang } 1544d46db3d5SWu Fengguang 154503ba3782SJens Axboe /* 1546f11c9c5cSEdward Shishkin * Write a portion of b_io inodes which belong to @sb. 1547edadfb10SChristoph Hellwig * 1548d46db3d5SWu Fengguang * Return the number of pages and/or inodes written. 15490ba13fd1SLinus Torvalds * 15500ba13fd1SLinus Torvalds * NOTE! This is called with wb->list_lock held, and will 15510ba13fd1SLinus Torvalds * unlock and relock that for each inode it ends up doing 15520ba13fd1SLinus Torvalds * IO for. 1553f11c9c5cSEdward Shishkin */ 1554d46db3d5SWu Fengguang static long writeback_sb_inodes(struct super_block *sb, 1555d46db3d5SWu Fengguang struct bdi_writeback *wb, 1556d46db3d5SWu Fengguang struct wb_writeback_work *work) 155703ba3782SJens Axboe { 1558d46db3d5SWu Fengguang struct writeback_control wbc = { 1559d46db3d5SWu Fengguang .sync_mode = work->sync_mode, 1560d46db3d5SWu Fengguang .tagged_writepages = work->tagged_writepages, 1561d46db3d5SWu Fengguang .for_kupdate = work->for_kupdate, 1562d46db3d5SWu Fengguang .for_background = work->for_background, 15637747bd4bSDave Chinner .for_sync = work->for_sync, 1564d46db3d5SWu Fengguang .range_cyclic = work->range_cyclic, 1565d46db3d5SWu Fengguang .range_start = 0, 1566d46db3d5SWu Fengguang .range_end = LLONG_MAX, 1567d46db3d5SWu Fengguang }; 1568d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1569d46db3d5SWu Fengguang long write_chunk; 1570d46db3d5SWu Fengguang long wrote = 0; /* count both pages and inodes */ 1571d46db3d5SWu Fengguang 157203ba3782SJens Axboe while (!list_empty(&wb->b_io)) { 15737ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1574aaf25593STejun Heo struct bdi_writeback *tmp_wb; 1575edadfb10SChristoph Hellwig 1576edadfb10SChristoph Hellwig if (inode->i_sb != sb) { 1577d46db3d5SWu Fengguang if (work->sb) { 1578edadfb10SChristoph Hellwig /* 1579edadfb10SChristoph Hellwig * We only want to write back data for this 1580edadfb10SChristoph Hellwig * superblock, move all inodes not belonging 1581edadfb10SChristoph Hellwig * to it back onto the dirty list. 1582edadfb10SChristoph Hellwig */ 1583f758eeabSChristoph Hellwig redirty_tail(inode, wb); 158466f3b8e2SJens Axboe continue; 158566f3b8e2SJens Axboe } 1586edadfb10SChristoph Hellwig 1587edadfb10SChristoph Hellwig /* 1588edadfb10SChristoph Hellwig * The inode belongs to a different superblock. 1589edadfb10SChristoph Hellwig * Bounce back to the caller to unpin this and 1590edadfb10SChristoph Hellwig * pin the next superblock. 1591edadfb10SChristoph Hellwig */ 1592d46db3d5SWu Fengguang break; 1593edadfb10SChristoph Hellwig } 1594edadfb10SChristoph Hellwig 15959843b76aSChristoph Hellwig /* 1596331cbdeeSWanpeng Li * Don't bother with new inodes or inodes being freed, first 1597331cbdeeSWanpeng Li * kind does not need periodic writeout yet, and for the latter 15989843b76aSChristoph Hellwig * kind writeout is handled by the freer. 15999843b76aSChristoph Hellwig */ 1600250df6edSDave Chinner spin_lock(&inode->i_lock); 16019843b76aSChristoph Hellwig if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { 1602250df6edSDave Chinner spin_unlock(&inode->i_lock); 1603fcc5c222SWu Fengguang redirty_tail(inode, wb); 16047ef0d737SNick Piggin continue; 16057ef0d737SNick Piggin } 1606cc1676d9SJan Kara if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { 1607cc1676d9SJan Kara /* 1608cc1676d9SJan Kara * If this inode is locked for writeback and we are not 1609cc1676d9SJan Kara * doing writeback-for-data-integrity, move it to 1610cc1676d9SJan Kara * b_more_io so that writeback can proceed with the 1611cc1676d9SJan Kara * other inodes on s_io. 1612cc1676d9SJan Kara * 1613cc1676d9SJan Kara * We'll have another go at writing back this inode 1614cc1676d9SJan Kara * when we completed a full scan of b_io. 1615cc1676d9SJan Kara */ 1616cc1676d9SJan Kara spin_unlock(&inode->i_lock); 1617cc1676d9SJan Kara requeue_io(inode, wb); 1618cc1676d9SJan Kara trace_writeback_sb_inodes_requeue(inode); 1619cc1676d9SJan Kara continue; 1620cc1676d9SJan Kara } 1621f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 1622f0d07b7fSJan Kara 16234f8ad655SJan Kara /* 16244f8ad655SJan Kara * We already requeued the inode if it had I_SYNC set and we 16254f8ad655SJan Kara * are doing WB_SYNC_NONE writeback. So this catches only the 16264f8ad655SJan Kara * WB_SYNC_ALL case. 16274f8ad655SJan Kara */ 1628169ebd90SJan Kara if (inode->i_state & I_SYNC) { 1629169ebd90SJan Kara /* Wait for I_SYNC. This function drops i_lock... */ 1630169ebd90SJan Kara inode_sleep_on_writeback(inode); 1631169ebd90SJan Kara /* Inode may be gone, start again */ 1632ead188f9SJan Kara spin_lock(&wb->list_lock); 1633169ebd90SJan Kara continue; 1634169ebd90SJan Kara } 16354f8ad655SJan Kara inode->i_state |= I_SYNC; 1636b16b1debSTejun Heo wbc_attach_and_unlock_inode(&wbc, inode); 1637169ebd90SJan Kara 1638a88a341aSTejun Heo write_chunk = writeback_chunk_size(wb, work); 1639d46db3d5SWu Fengguang wbc.nr_to_write = write_chunk; 1640d46db3d5SWu Fengguang wbc.pages_skipped = 0; 1641250df6edSDave Chinner 1642169ebd90SJan Kara /* 1643169ebd90SJan Kara * We use I_SYNC to pin the inode in memory. While it is set 1644169ebd90SJan Kara * evict_inode() will wait so the inode cannot be freed. 1645169ebd90SJan Kara */ 1646cd8ed2a4SYan Hong __writeback_single_inode(inode, &wbc); 1647d46db3d5SWu Fengguang 1648b16b1debSTejun Heo wbc_detach_inode(&wbc); 1649d46db3d5SWu Fengguang work->nr_pages -= write_chunk - wbc.nr_to_write; 1650d46db3d5SWu Fengguang wrote += write_chunk - wbc.nr_to_write; 1651590dca3aSChris Mason 1652590dca3aSChris Mason if (need_resched()) { 1653590dca3aSChris Mason /* 1654590dca3aSChris Mason * We're trying to balance between building up a nice 1655590dca3aSChris Mason * long list of IOs to improve our merge rate, and 1656590dca3aSChris Mason * getting those IOs out quickly for anyone throttling 1657590dca3aSChris Mason * in balance_dirty_pages(). cond_resched() doesn't 1658590dca3aSChris Mason * unplug, so get our IOs out the door before we 1659590dca3aSChris Mason * give up the CPU. 1660590dca3aSChris Mason */ 1661590dca3aSChris Mason blk_flush_plug(current); 1662590dca3aSChris Mason cond_resched(); 1663590dca3aSChris Mason } 1664590dca3aSChris Mason 1665aaf25593STejun Heo /* 1666aaf25593STejun Heo * Requeue @inode if still dirty. Be careful as @inode may 1667aaf25593STejun Heo * have been switched to another wb in the meantime. 1668aaf25593STejun Heo */ 1669aaf25593STejun Heo tmp_wb = inode_to_wb_and_lock_list(inode); 16704f8ad655SJan Kara spin_lock(&inode->i_lock); 16710ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1672d46db3d5SWu Fengguang wrote++; 1673aaf25593STejun Heo requeue_inode(inode, tmp_wb, &wbc); 16744f8ad655SJan Kara inode_sync_complete(inode); 16750f1b1fd8SDave Chinner spin_unlock(&inode->i_lock); 1676590dca3aSChris Mason 1677aaf25593STejun Heo if (unlikely(tmp_wb != wb)) { 1678aaf25593STejun Heo spin_unlock(&tmp_wb->list_lock); 1679aaf25593STejun Heo spin_lock(&wb->list_lock); 1680aaf25593STejun Heo } 1681aaf25593STejun Heo 1682d46db3d5SWu Fengguang /* 1683d46db3d5SWu Fengguang * bail out to wb_writeback() often enough to check 1684d46db3d5SWu Fengguang * background threshold and other termination conditions. 1685d46db3d5SWu Fengguang */ 1686d46db3d5SWu Fengguang if (wrote) { 1687d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1688d46db3d5SWu Fengguang break; 1689d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1690d46db3d5SWu Fengguang break; 16911da177e4SLinus Torvalds } 16928bc3be27SFengguang Wu } 1693d46db3d5SWu Fengguang return wrote; 1694f11c9c5cSEdward Shishkin } 169538f21977SNick Piggin 1696d46db3d5SWu Fengguang static long __writeback_inodes_wb(struct bdi_writeback *wb, 1697d46db3d5SWu Fengguang struct wb_writeback_work *work) 1698f11c9c5cSEdward Shishkin { 1699d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1700d46db3d5SWu Fengguang long wrote = 0; 1701f11c9c5cSEdward Shishkin 1702f11c9c5cSEdward Shishkin while (!list_empty(&wb->b_io)) { 17037ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1704f11c9c5cSEdward Shishkin struct super_block *sb = inode->i_sb; 1705f11c9c5cSEdward Shishkin 1706eb6ef3dfSKonstantin Khlebnikov if (!trylock_super(sb)) { 17070e995816SWu Fengguang /* 1708eb6ef3dfSKonstantin Khlebnikov * trylock_super() may fail consistently due to 17090e995816SWu Fengguang * s_umount being grabbed by someone else. Don't use 17100e995816SWu Fengguang * requeue_io() to avoid busy retrying the inode/sb. 17110e995816SWu Fengguang */ 17120e995816SWu Fengguang redirty_tail(inode, wb); 1713d19de7edSChristoph Hellwig continue; 1714334132aeSChristoph Hellwig } 1715d46db3d5SWu Fengguang wrote += writeback_sb_inodes(sb, wb, work); 1716eb6ef3dfSKonstantin Khlebnikov up_read(&sb->s_umount); 1717f11c9c5cSEdward Shishkin 1718d46db3d5SWu Fengguang /* refer to the same tests at the end of writeback_sb_inodes */ 1719d46db3d5SWu Fengguang if (wrote) { 1720d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1721d46db3d5SWu Fengguang break; 1722d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1723f11c9c5cSEdward Shishkin break; 1724f11c9c5cSEdward Shishkin } 1725d46db3d5SWu Fengguang } 172666f3b8e2SJens Axboe /* Leave any unwritten inodes on b_io */ 1727d46db3d5SWu Fengguang return wrote; 172866f3b8e2SJens Axboe } 172966f3b8e2SJens Axboe 17307d9f073bSWanpeng Li static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, 17310e175a18SCurt Wohlgemuth enum wb_reason reason) 1732edadfb10SChristoph Hellwig { 1733d46db3d5SWu Fengguang struct wb_writeback_work work = { 1734d46db3d5SWu Fengguang .nr_pages = nr_pages, 1735d46db3d5SWu Fengguang .sync_mode = WB_SYNC_NONE, 1736d46db3d5SWu Fengguang .range_cyclic = 1, 17370e175a18SCurt Wohlgemuth .reason = reason, 1738d46db3d5SWu Fengguang }; 1739505a666eSLinus Torvalds struct blk_plug plug; 1740edadfb10SChristoph Hellwig 1741505a666eSLinus Torvalds blk_start_plug(&plug); 1742f758eeabSChristoph Hellwig spin_lock(&wb->list_lock); 1743424b351fSWu Fengguang if (list_empty(&wb->b_io)) 1744ad4e38ddSCurt Wohlgemuth queue_io(wb, &work); 1745d46db3d5SWu Fengguang __writeback_inodes_wb(wb, &work); 1746f758eeabSChristoph Hellwig spin_unlock(&wb->list_lock); 1747505a666eSLinus Torvalds blk_finish_plug(&plug); 1748edadfb10SChristoph Hellwig 1749d46db3d5SWu Fengguang return nr_pages - work.nr_pages; 175066f3b8e2SJens Axboe } 175166f3b8e2SJens Axboe 175203ba3782SJens Axboe /* 175303ba3782SJens Axboe * Explicit flushing or periodic writeback of "old" data. 175403ba3782SJens Axboe * 175503ba3782SJens Axboe * Define "old": the first time one of an inode's pages is dirtied, we mark the 175603ba3782SJens Axboe * dirtying-time in the inode's address_space. So this periodic writeback code 175703ba3782SJens Axboe * just walks the superblock inode list, writing back any inodes which are 175803ba3782SJens Axboe * older than a specific point in time. 175903ba3782SJens Axboe * 176003ba3782SJens Axboe * Try to run once per dirty_writeback_interval. But if a writeback event 176103ba3782SJens Axboe * takes longer than a dirty_writeback_interval interval, then leave a 176203ba3782SJens Axboe * one-second gap. 176303ba3782SJens Axboe * 176403ba3782SJens Axboe * older_than_this takes precedence over nr_to_write. So we'll only write back 176503ba3782SJens Axboe * all dirty pages if they are all attached to "old" mappings. 176603ba3782SJens Axboe */ 1767c4a77a6cSJens Axboe static long wb_writeback(struct bdi_writeback *wb, 176883ba7b07SChristoph Hellwig struct wb_writeback_work *work) 176903ba3782SJens Axboe { 1770e98be2d5SWu Fengguang unsigned long wb_start = jiffies; 1771d46db3d5SWu Fengguang long nr_pages = work->nr_pages; 17720dc83bd3SJan Kara unsigned long oldest_jif; 1773a5989bdcSJan Kara struct inode *inode; 1774d46db3d5SWu Fengguang long progress; 1775505a666eSLinus Torvalds struct blk_plug plug; 177603ba3782SJens Axboe 17770dc83bd3SJan Kara oldest_jif = jiffies; 17780dc83bd3SJan Kara work->older_than_this = &oldest_jif; 177903ba3782SJens Axboe 1780505a666eSLinus Torvalds blk_start_plug(&plug); 1781e8dfc305SWu Fengguang spin_lock(&wb->list_lock); 178203ba3782SJens Axboe for (;;) { 178303ba3782SJens Axboe /* 1784d3ddec76SWu Fengguang * Stop writeback when nr_pages has been consumed 178503ba3782SJens Axboe */ 178683ba7b07SChristoph Hellwig if (work->nr_pages <= 0) 178703ba3782SJens Axboe break; 178803ba3782SJens Axboe 178903ba3782SJens Axboe /* 1790aa373cf5SJan Kara * Background writeout and kupdate-style writeback may 1791aa373cf5SJan Kara * run forever. Stop them if there is other work to do 1792aa373cf5SJan Kara * so that e.g. sync can proceed. They'll be restarted 1793aa373cf5SJan Kara * after the other works are all done. 1794aa373cf5SJan Kara */ 1795aa373cf5SJan Kara if ((work->for_background || work->for_kupdate) && 1796f0054bb1STejun Heo !list_empty(&wb->work_list)) 1797aa373cf5SJan Kara break; 1798aa373cf5SJan Kara 1799aa373cf5SJan Kara /* 1800d3ddec76SWu Fengguang * For background writeout, stop when we are below the 1801d3ddec76SWu Fengguang * background dirty threshold 180203ba3782SJens Axboe */ 1803aa661bbeSTejun Heo if (work->for_background && !wb_over_bg_thresh(wb)) 180403ba3782SJens Axboe break; 180503ba3782SJens Axboe 18061bc36b64SJan Kara /* 18071bc36b64SJan Kara * Kupdate and background works are special and we want to 18081bc36b64SJan Kara * include all inodes that need writing. Livelock avoidance is 18091bc36b64SJan Kara * handled by these works yielding to any other work so we are 18101bc36b64SJan Kara * safe. 18111bc36b64SJan Kara */ 1812ba9aa839SWu Fengguang if (work->for_kupdate) { 18130dc83bd3SJan Kara oldest_jif = jiffies - 1814ba9aa839SWu Fengguang msecs_to_jiffies(dirty_expire_interval * 10); 18151bc36b64SJan Kara } else if (work->for_background) 18160dc83bd3SJan Kara oldest_jif = jiffies; 1817028c2dd1SDave Chinner 18185634cc2aSTejun Heo trace_writeback_start(wb, work); 1819e8dfc305SWu Fengguang if (list_empty(&wb->b_io)) 1820ad4e38ddSCurt Wohlgemuth queue_io(wb, work); 182183ba7b07SChristoph Hellwig if (work->sb) 1822d46db3d5SWu Fengguang progress = writeback_sb_inodes(work->sb, wb, work); 1823edadfb10SChristoph Hellwig else 1824d46db3d5SWu Fengguang progress = __writeback_inodes_wb(wb, work); 18255634cc2aSTejun Heo trace_writeback_written(wb, work); 1826028c2dd1SDave Chinner 1827e98be2d5SWu Fengguang wb_update_bandwidth(wb, wb_start); 182803ba3782SJens Axboe 182903ba3782SJens Axboe /* 183071fd05a8SJens Axboe * Did we write something? Try for more 1831e6fb6da2SWu Fengguang * 1832e6fb6da2SWu Fengguang * Dirty inodes are moved to b_io for writeback in batches. 1833e6fb6da2SWu Fengguang * The completion of the current batch does not necessarily 1834e6fb6da2SWu Fengguang * mean the overall work is done. So we keep looping as long 1835e6fb6da2SWu Fengguang * as made some progress on cleaning pages or inodes. 183671fd05a8SJens Axboe */ 1837d46db3d5SWu Fengguang if (progress) 183803ba3782SJens Axboe continue; 1839a5989bdcSJan Kara /* 1840e6fb6da2SWu Fengguang * No more inodes for IO, bail 1841a5989bdcSJan Kara */ 1842b7a2441fSWu Fengguang if (list_empty(&wb->b_more_io)) 184303ba3782SJens Axboe break; 184403ba3782SJens Axboe /* 18458010c3b6SJens Axboe * Nothing written. Wait for some inode to 18468010c3b6SJens Axboe * become available for writeback. Otherwise 18478010c3b6SJens Axboe * we'll just busyloop. 184803ba3782SJens Axboe */ 18495634cc2aSTejun Heo trace_writeback_wait(wb, work); 185003ba3782SJens Axboe inode = wb_inode(wb->b_more_io.prev); 1851250df6edSDave Chinner spin_lock(&inode->i_lock); 1852f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 1853169ebd90SJan Kara /* This function drops i_lock... */ 1854169ebd90SJan Kara inode_sleep_on_writeback(inode); 1855f0d07b7fSJan Kara spin_lock(&wb->list_lock); 185603ba3782SJens Axboe } 1857e8dfc305SWu Fengguang spin_unlock(&wb->list_lock); 1858505a666eSLinus Torvalds blk_finish_plug(&plug); 185903ba3782SJens Axboe 1860d46db3d5SWu Fengguang return nr_pages - work->nr_pages; 186103ba3782SJens Axboe } 186203ba3782SJens Axboe 186303ba3782SJens Axboe /* 186483ba7b07SChristoph Hellwig * Return the next wb_writeback_work struct that hasn't been processed yet. 186503ba3782SJens Axboe */ 1866f0054bb1STejun Heo static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) 186703ba3782SJens Axboe { 186883ba7b07SChristoph Hellwig struct wb_writeback_work *work = NULL; 186903ba3782SJens Axboe 1870f0054bb1STejun Heo spin_lock_bh(&wb->work_lock); 1871f0054bb1STejun Heo if (!list_empty(&wb->work_list)) { 1872f0054bb1STejun Heo work = list_entry(wb->work_list.next, 187383ba7b07SChristoph Hellwig struct wb_writeback_work, list); 187483ba7b07SChristoph Hellwig list_del_init(&work->list); 187503ba3782SJens Axboe } 1876f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 187783ba7b07SChristoph Hellwig return work; 187803ba3782SJens Axboe } 187903ba3782SJens Axboe 18806585027aSJan Kara static long wb_check_background_flush(struct bdi_writeback *wb) 18816585027aSJan Kara { 1882aa661bbeSTejun Heo if (wb_over_bg_thresh(wb)) { 18836585027aSJan Kara 18846585027aSJan Kara struct wb_writeback_work work = { 18856585027aSJan Kara .nr_pages = LONG_MAX, 18866585027aSJan Kara .sync_mode = WB_SYNC_NONE, 18876585027aSJan Kara .for_background = 1, 18886585027aSJan Kara .range_cyclic = 1, 18890e175a18SCurt Wohlgemuth .reason = WB_REASON_BACKGROUND, 18906585027aSJan Kara }; 18916585027aSJan Kara 18926585027aSJan Kara return wb_writeback(wb, &work); 18936585027aSJan Kara } 18946585027aSJan Kara 18956585027aSJan Kara return 0; 18966585027aSJan Kara } 18976585027aSJan Kara 189803ba3782SJens Axboe static long wb_check_old_data_flush(struct bdi_writeback *wb) 189903ba3782SJens Axboe { 190003ba3782SJens Axboe unsigned long expired; 190103ba3782SJens Axboe long nr_pages; 190203ba3782SJens Axboe 190369b62d01SJens Axboe /* 190469b62d01SJens Axboe * When set to zero, disable periodic writeback 190569b62d01SJens Axboe */ 190669b62d01SJens Axboe if (!dirty_writeback_interval) 190769b62d01SJens Axboe return 0; 190869b62d01SJens Axboe 190903ba3782SJens Axboe expired = wb->last_old_flush + 191003ba3782SJens Axboe msecs_to_jiffies(dirty_writeback_interval * 10); 191103ba3782SJens Axboe if (time_before(jiffies, expired)) 191203ba3782SJens Axboe return 0; 191303ba3782SJens Axboe 191403ba3782SJens Axboe wb->last_old_flush = jiffies; 1915cdf01dd5SLinus Torvalds nr_pages = get_nr_dirty_pages(); 191603ba3782SJens Axboe 1917c4a77a6cSJens Axboe if (nr_pages) { 191883ba7b07SChristoph Hellwig struct wb_writeback_work work = { 1919c4a77a6cSJens Axboe .nr_pages = nr_pages, 1920c4a77a6cSJens Axboe .sync_mode = WB_SYNC_NONE, 1921c4a77a6cSJens Axboe .for_kupdate = 1, 1922c4a77a6cSJens Axboe .range_cyclic = 1, 19230e175a18SCurt Wohlgemuth .reason = WB_REASON_PERIODIC, 1924c4a77a6cSJens Axboe }; 1925c4a77a6cSJens Axboe 192683ba7b07SChristoph Hellwig return wb_writeback(wb, &work); 1927c4a77a6cSJens Axboe } 192803ba3782SJens Axboe 192903ba3782SJens Axboe return 0; 193003ba3782SJens Axboe } 193103ba3782SJens Axboe 193285009b4fSJens Axboe static long wb_check_start_all(struct bdi_writeback *wb) 193385009b4fSJens Axboe { 193485009b4fSJens Axboe long nr_pages; 193585009b4fSJens Axboe 193685009b4fSJens Axboe if (!test_bit(WB_start_all, &wb->state)) 193785009b4fSJens Axboe return 0; 193885009b4fSJens Axboe 193985009b4fSJens Axboe nr_pages = get_nr_dirty_pages(); 194085009b4fSJens Axboe if (nr_pages) { 194185009b4fSJens Axboe struct wb_writeback_work work = { 194285009b4fSJens Axboe .nr_pages = wb_split_bdi_pages(wb, nr_pages), 194385009b4fSJens Axboe .sync_mode = WB_SYNC_NONE, 194485009b4fSJens Axboe .range_cyclic = 1, 194585009b4fSJens Axboe .reason = wb->start_all_reason, 194685009b4fSJens Axboe }; 194785009b4fSJens Axboe 194885009b4fSJens Axboe nr_pages = wb_writeback(wb, &work); 194985009b4fSJens Axboe } 195085009b4fSJens Axboe 195185009b4fSJens Axboe clear_bit(WB_start_all, &wb->state); 195285009b4fSJens Axboe return nr_pages; 195385009b4fSJens Axboe } 195485009b4fSJens Axboe 195585009b4fSJens Axboe 195603ba3782SJens Axboe /* 195703ba3782SJens Axboe * Retrieve work items and do the writeback they describe 195803ba3782SJens Axboe */ 195925d130baSWanpeng Li static long wb_do_writeback(struct bdi_writeback *wb) 196003ba3782SJens Axboe { 196183ba7b07SChristoph Hellwig struct wb_writeback_work *work; 1962c4a77a6cSJens Axboe long wrote = 0; 196303ba3782SJens Axboe 19644452226eSTejun Heo set_bit(WB_writeback_running, &wb->state); 1965f0054bb1STejun Heo while ((work = get_next_work_item(wb)) != NULL) { 19665634cc2aSTejun Heo trace_writeback_exec(wb, work); 196783ba7b07SChristoph Hellwig wrote += wb_writeback(wb, work); 19684a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 196903ba3782SJens Axboe } 197003ba3782SJens Axboe 197103ba3782SJens Axboe /* 197285009b4fSJens Axboe * Check for a flush-everything request 197385009b4fSJens Axboe */ 197485009b4fSJens Axboe wrote += wb_check_start_all(wb); 197585009b4fSJens Axboe 197685009b4fSJens Axboe /* 197703ba3782SJens Axboe * Check for periodic writeback, kupdated() style 197803ba3782SJens Axboe */ 197903ba3782SJens Axboe wrote += wb_check_old_data_flush(wb); 19806585027aSJan Kara wrote += wb_check_background_flush(wb); 19814452226eSTejun Heo clear_bit(WB_writeback_running, &wb->state); 198203ba3782SJens Axboe 198303ba3782SJens Axboe return wrote; 198403ba3782SJens Axboe } 198503ba3782SJens Axboe 198603ba3782SJens Axboe /* 198703ba3782SJens Axboe * Handle writeback of dirty data for the device backed by this bdi. Also 1988839a8e86STejun Heo * reschedules periodically and does kupdated style flushing. 198903ba3782SJens Axboe */ 1990f0054bb1STejun Heo void wb_workfn(struct work_struct *work) 199103ba3782SJens Axboe { 1992839a8e86STejun Heo struct bdi_writeback *wb = container_of(to_delayed_work(work), 1993839a8e86STejun Heo struct bdi_writeback, dwork); 199403ba3782SJens Axboe long pages_written; 199503ba3782SJens Axboe 1996f0054bb1STejun Heo set_worker_desc("flush-%s", dev_name(wb->bdi->dev)); 1997766f9164SPeter Zijlstra current->flags |= PF_SWAPWRITE; 199803ba3782SJens Axboe 1999839a8e86STejun Heo if (likely(!current_is_workqueue_rescuer() || 20004452226eSTejun Heo !test_bit(WB_registered, &wb->state))) { 200103ba3782SJens Axboe /* 2002f0054bb1STejun Heo * The normal path. Keep writing back @wb until its 2003839a8e86STejun Heo * work_list is empty. Note that this path is also taken 2004f0054bb1STejun Heo * if @wb is shutting down even when we're running off the 2005839a8e86STejun Heo * rescuer as work_list needs to be drained. 200603ba3782SJens Axboe */ 2007839a8e86STejun Heo do { 200825d130baSWanpeng Li pages_written = wb_do_writeback(wb); 2009455b2864SDave Chinner trace_writeback_pages_written(pages_written); 2010f0054bb1STejun Heo } while (!list_empty(&wb->work_list)); 2011839a8e86STejun Heo } else { 2012253c34e9SArtem Bityutskiy /* 2013839a8e86STejun Heo * bdi_wq can't get enough workers and we're running off 2014839a8e86STejun Heo * the emergency worker. Don't hog it. Hopefully, 1024 is 2015839a8e86STejun Heo * enough for efficient IO. 2016253c34e9SArtem Bityutskiy */ 2017f0054bb1STejun Heo pages_written = writeback_inodes_wb(wb, 1024, 2018839a8e86STejun Heo WB_REASON_FORKER_THREAD); 2019839a8e86STejun Heo trace_writeback_pages_written(pages_written); 202003ba3782SJens Axboe } 202103ba3782SJens Axboe 2022f0054bb1STejun Heo if (!list_empty(&wb->work_list)) 2023b8b78495SJan Kara wb_wakeup(wb); 20246ca738d6SDerek Basehore else if (wb_has_dirty_io(wb) && dirty_writeback_interval) 2025f0054bb1STejun Heo wb_wakeup_delayed(wb); 2026455b2864SDave Chinner 2027839a8e86STejun Heo current->flags &= ~PF_SWAPWRITE; 202803ba3782SJens Axboe } 202903ba3782SJens Axboe 203003ba3782SJens Axboe /* 2031595043e5SJens Axboe * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero, 2032595043e5SJens Axboe * write back the whole world. 2033595043e5SJens Axboe */ 2034595043e5SJens Axboe static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2035e8e8a0c6SJens Axboe enum wb_reason reason) 2036595043e5SJens Axboe { 2037595043e5SJens Axboe struct bdi_writeback *wb; 2038595043e5SJens Axboe 2039595043e5SJens Axboe if (!bdi_has_dirty_io(bdi)) 2040595043e5SJens Axboe return; 2041595043e5SJens Axboe 2042595043e5SJens Axboe list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 2043e8e8a0c6SJens Axboe wb_start_writeback(wb, reason); 2044595043e5SJens Axboe } 2045595043e5SJens Axboe 2046595043e5SJens Axboe void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2047595043e5SJens Axboe enum wb_reason reason) 2048595043e5SJens Axboe { 2049595043e5SJens Axboe rcu_read_lock(); 2050e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2051595043e5SJens Axboe rcu_read_unlock(); 2052595043e5SJens Axboe } 2053595043e5SJens Axboe 2054595043e5SJens Axboe /* 20559ba4b2dfSJens Axboe * Wakeup the flusher threads to start writeback of all currently dirty pages 205603ba3782SJens Axboe */ 20579ba4b2dfSJens Axboe void wakeup_flusher_threads(enum wb_reason reason) 205803ba3782SJens Axboe { 2059b8c2f347SChristoph Hellwig struct backing_dev_info *bdi; 2060b8c2f347SChristoph Hellwig 206151350ea0SKonstantin Khlebnikov /* 206251350ea0SKonstantin Khlebnikov * If we are expecting writeback progress we must submit plugged IO. 206351350ea0SKonstantin Khlebnikov */ 206451350ea0SKonstantin Khlebnikov if (blk_needs_flush_plug(current)) 206551350ea0SKonstantin Khlebnikov blk_schedule_flush_plug(current); 206651350ea0SKonstantin Khlebnikov 2067b8c2f347SChristoph Hellwig rcu_read_lock(); 2068595043e5SJens Axboe list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) 2069e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2070b8c2f347SChristoph Hellwig rcu_read_unlock(); 207103ba3782SJens Axboe } 207203ba3782SJens Axboe 2073a2f48706STheodore Ts'o /* 2074a2f48706STheodore Ts'o * Wake up bdi's periodically to make sure dirtytime inodes gets 2075a2f48706STheodore Ts'o * written back periodically. We deliberately do *not* check the 2076a2f48706STheodore Ts'o * b_dirtytime list in wb_has_dirty_io(), since this would cause the 2077a2f48706STheodore Ts'o * kernel to be constantly waking up once there are any dirtytime 2078a2f48706STheodore Ts'o * inodes on the system. So instead we define a separate delayed work 2079a2f48706STheodore Ts'o * function which gets called much more rarely. (By default, only 2080a2f48706STheodore Ts'o * once every 12 hours.) 2081a2f48706STheodore Ts'o * 2082a2f48706STheodore Ts'o * If there is any other write activity going on in the file system, 2083a2f48706STheodore Ts'o * this function won't be necessary. But if the only thing that has 2084a2f48706STheodore Ts'o * happened on the file system is a dirtytime inode caused by an atime 2085a2f48706STheodore Ts'o * update, we need this infrastructure below to make sure that inode 2086a2f48706STheodore Ts'o * eventually gets pushed out to disk. 2087a2f48706STheodore Ts'o */ 2088a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w); 2089a2f48706STheodore Ts'o static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); 2090a2f48706STheodore Ts'o 2091a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w) 2092a2f48706STheodore Ts'o { 2093a2f48706STheodore Ts'o struct backing_dev_info *bdi; 2094a2f48706STheodore Ts'o 2095a2f48706STheodore Ts'o rcu_read_lock(); 2096a2f48706STheodore Ts'o list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 2097001fe6f6STejun Heo struct bdi_writeback *wb; 2098001fe6f6STejun Heo 2099b817525aSTejun Heo list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 21006fdf860fSTejun Heo if (!list_empty(&wb->b_dirty_time)) 21016fdf860fSTejun Heo wb_wakeup(wb); 2102a2f48706STheodore Ts'o } 2103a2f48706STheodore Ts'o rcu_read_unlock(); 2104a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2105a2f48706STheodore Ts'o } 2106a2f48706STheodore Ts'o 2107a2f48706STheodore Ts'o static int __init start_dirtytime_writeback(void) 2108a2f48706STheodore Ts'o { 2109a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2110a2f48706STheodore Ts'o return 0; 2111a2f48706STheodore Ts'o } 2112a2f48706STheodore Ts'o __initcall(start_dirtytime_writeback); 2113a2f48706STheodore Ts'o 21141efff914STheodore Ts'o int dirtytime_interval_handler(struct ctl_table *table, int write, 21151efff914STheodore Ts'o void __user *buffer, size_t *lenp, loff_t *ppos) 21161efff914STheodore Ts'o { 21171efff914STheodore Ts'o int ret; 21181efff914STheodore Ts'o 21191efff914STheodore Ts'o ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 21201efff914STheodore Ts'o if (ret == 0 && write) 21211efff914STheodore Ts'o mod_delayed_work(system_wq, &dirtytime_work, 0); 21221efff914STheodore Ts'o return ret; 21231efff914STheodore Ts'o } 21241efff914STheodore Ts'o 212503ba3782SJens Axboe static noinline void block_dump___mark_inode_dirty(struct inode *inode) 212603ba3782SJens Axboe { 212703ba3782SJens Axboe if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) { 212803ba3782SJens Axboe struct dentry *dentry; 212903ba3782SJens Axboe const char *name = "?"; 213003ba3782SJens Axboe 213103ba3782SJens Axboe dentry = d_find_alias(inode); 213203ba3782SJens Axboe if (dentry) { 213303ba3782SJens Axboe spin_lock(&dentry->d_lock); 213403ba3782SJens Axboe name = (const char *) dentry->d_name.name; 213503ba3782SJens Axboe } 213603ba3782SJens Axboe printk(KERN_DEBUG 213703ba3782SJens Axboe "%s(%d): dirtied inode %lu (%s) on %s\n", 213803ba3782SJens Axboe current->comm, task_pid_nr(current), inode->i_ino, 213903ba3782SJens Axboe name, inode->i_sb->s_id); 214003ba3782SJens Axboe if (dentry) { 214103ba3782SJens Axboe spin_unlock(&dentry->d_lock); 214203ba3782SJens Axboe dput(dentry); 214303ba3782SJens Axboe } 214403ba3782SJens Axboe } 214503ba3782SJens Axboe } 214603ba3782SJens Axboe 214703ba3782SJens Axboe /** 214803ba3782SJens Axboe * __mark_inode_dirty - internal function 21490117d427SMauro Carvalho Chehab * 215003ba3782SJens Axboe * @inode: inode to mark 215103ba3782SJens Axboe * @flags: what kind of dirty (i.e. I_DIRTY_SYNC) 21520117d427SMauro Carvalho Chehab * 215303ba3782SJens Axboe * Mark an inode as dirty. Callers should use mark_inode_dirty or 215403ba3782SJens Axboe * mark_inode_dirty_sync. 215503ba3782SJens Axboe * 215603ba3782SJens Axboe * Put the inode on the super block's dirty list. 215703ba3782SJens Axboe * 215803ba3782SJens Axboe * CAREFUL! We mark it dirty unconditionally, but move it onto the 215903ba3782SJens Axboe * dirty list only if it is hashed or if it refers to a blockdev. 216003ba3782SJens Axboe * If it was not hashed, it will never be added to the dirty list 216103ba3782SJens Axboe * even if it is later hashed, as it will have been marked dirty already. 216203ba3782SJens Axboe * 216303ba3782SJens Axboe * In short, make sure you hash any inodes _before_ you start marking 216403ba3782SJens Axboe * them dirty. 216503ba3782SJens Axboe * 216603ba3782SJens Axboe * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 216703ba3782SJens Axboe * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 216803ba3782SJens Axboe * the kernel-internal blockdev inode represents the dirtying time of the 216903ba3782SJens Axboe * blockdev's pages. This is why for I_DIRTY_PAGES we always use 217003ba3782SJens Axboe * page->mapping->host, so the page-dirtying time is recorded in the internal 217103ba3782SJens Axboe * blockdev inode. 217203ba3782SJens Axboe */ 217303ba3782SJens Axboe void __mark_inode_dirty(struct inode *inode, int flags) 217403ba3782SJens Axboe { 217503ba3782SJens Axboe struct super_block *sb = inode->i_sb; 21760ae45f63STheodore Ts'o int dirtytime; 21770ae45f63STheodore Ts'o 21780ae45f63STheodore Ts'o trace_writeback_mark_inode_dirty(inode, flags); 217903ba3782SJens Axboe 218003ba3782SJens Axboe /* 218103ba3782SJens Axboe * Don't do this for I_DIRTY_PAGES - that doesn't actually 218203ba3782SJens Axboe * dirty the inode itself 218303ba3782SJens Axboe */ 21840e11f644SChristoph Hellwig if (flags & (I_DIRTY_INODE | I_DIRTY_TIME)) { 21859fb0a7daSTejun Heo trace_writeback_dirty_inode_start(inode, flags); 21869fb0a7daSTejun Heo 218703ba3782SJens Axboe if (sb->s_op->dirty_inode) 2188aa385729SChristoph Hellwig sb->s_op->dirty_inode(inode, flags); 21899fb0a7daSTejun Heo 21909fb0a7daSTejun Heo trace_writeback_dirty_inode(inode, flags); 219103ba3782SJens Axboe } 21920ae45f63STheodore Ts'o if (flags & I_DIRTY_INODE) 21930ae45f63STheodore Ts'o flags &= ~I_DIRTY_TIME; 21940ae45f63STheodore Ts'o dirtytime = flags & I_DIRTY_TIME; 219503ba3782SJens Axboe 219603ba3782SJens Axboe /* 21979c6ac78eSTejun Heo * Paired with smp_mb() in __writeback_single_inode() for the 21989c6ac78eSTejun Heo * following lockless i_state test. See there for details. 219903ba3782SJens Axboe */ 220003ba3782SJens Axboe smp_mb(); 220103ba3782SJens Axboe 22020ae45f63STheodore Ts'o if (((inode->i_state & flags) == flags) || 22030ae45f63STheodore Ts'o (dirtytime && (inode->i_state & I_DIRTY_INODE))) 220403ba3782SJens Axboe return; 220503ba3782SJens Axboe 220603ba3782SJens Axboe if (unlikely(block_dump)) 220703ba3782SJens Axboe block_dump___mark_inode_dirty(inode); 220803ba3782SJens Axboe 2209250df6edSDave Chinner spin_lock(&inode->i_lock); 22100ae45f63STheodore Ts'o if (dirtytime && (inode->i_state & I_DIRTY_INODE)) 22110ae45f63STheodore Ts'o goto out_unlock_inode; 221203ba3782SJens Axboe if ((inode->i_state & flags) != flags) { 221303ba3782SJens Axboe const int was_dirty = inode->i_state & I_DIRTY; 221403ba3782SJens Axboe 221552ebea74STejun Heo inode_attach_wb(inode, NULL); 221652ebea74STejun Heo 22170ae45f63STheodore Ts'o if (flags & I_DIRTY_INODE) 22180ae45f63STheodore Ts'o inode->i_state &= ~I_DIRTY_TIME; 221903ba3782SJens Axboe inode->i_state |= flags; 222003ba3782SJens Axboe 222103ba3782SJens Axboe /* 222203ba3782SJens Axboe * If the inode is being synced, just update its dirty state. 222303ba3782SJens Axboe * The unlocker will place the inode on the appropriate 222403ba3782SJens Axboe * superblock list, based upon its state. 222503ba3782SJens Axboe */ 222603ba3782SJens Axboe if (inode->i_state & I_SYNC) 2227250df6edSDave Chinner goto out_unlock_inode; 222803ba3782SJens Axboe 222903ba3782SJens Axboe /* 223003ba3782SJens Axboe * Only add valid (hashed) inodes to the superblock's 223103ba3782SJens Axboe * dirty list. Add blockdev inodes as well. 223203ba3782SJens Axboe */ 223303ba3782SJens Axboe if (!S_ISBLK(inode->i_mode)) { 22341d3382cbSAl Viro if (inode_unhashed(inode)) 2235250df6edSDave Chinner goto out_unlock_inode; 223603ba3782SJens Axboe } 2237a4ffdde6SAl Viro if (inode->i_state & I_FREEING) 2238250df6edSDave Chinner goto out_unlock_inode; 223903ba3782SJens Axboe 224003ba3782SJens Axboe /* 224103ba3782SJens Axboe * If the inode was already on b_dirty/b_io/b_more_io, don't 224203ba3782SJens Axboe * reposition it (that would break b_dirty time-ordering). 224303ba3782SJens Axboe */ 224403ba3782SJens Axboe if (!was_dirty) { 224587e1d789STejun Heo struct bdi_writeback *wb; 2246d6c10f1fSTejun Heo struct list_head *dirty_list; 2247a66979abSDave Chinner bool wakeup_bdi = false; 2248500b067cSJens Axboe 224987e1d789STejun Heo wb = locked_inode_to_wb_and_lock_list(inode); 2250253c34e9SArtem Bityutskiy 22510747259dSTejun Heo WARN(bdi_cap_writeback_dirty(wb->bdi) && 22520747259dSTejun Heo !test_bit(WB_registered, &wb->state), 22530747259dSTejun Heo "bdi-%s not registered\n", wb->bdi->name); 225403ba3782SJens Axboe 225503ba3782SJens Axboe inode->dirtied_when = jiffies; 2256a2f48706STheodore Ts'o if (dirtytime) 2257a2f48706STheodore Ts'o inode->dirtied_time_when = jiffies; 2258d6c10f1fSTejun Heo 22590e11f644SChristoph Hellwig if (inode->i_state & I_DIRTY) 22600747259dSTejun Heo dirty_list = &wb->b_dirty; 2261a2f48706STheodore Ts'o else 22620747259dSTejun Heo dirty_list = &wb->b_dirty_time; 2263d6c10f1fSTejun Heo 2264c7f54084SDave Chinner wakeup_bdi = inode_io_list_move_locked(inode, wb, 2265d6c10f1fSTejun Heo dirty_list); 2266d6c10f1fSTejun Heo 22670747259dSTejun Heo spin_unlock(&wb->list_lock); 22680ae45f63STheodore Ts'o trace_writeback_dirty_inode_enqueue(inode); 2269253c34e9SArtem Bityutskiy 2270d6c10f1fSTejun Heo /* 2271d6c10f1fSTejun Heo * If this is the first dirty inode for this bdi, 2272d6c10f1fSTejun Heo * we have to wake-up the corresponding bdi thread 2273d6c10f1fSTejun Heo * to make sure background write-back happens 2274d6c10f1fSTejun Heo * later. 2275d6c10f1fSTejun Heo */ 22760747259dSTejun Heo if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi) 22770747259dSTejun Heo wb_wakeup_delayed(wb); 2278a66979abSDave Chinner return; 2279a66979abSDave Chinner } 2280a66979abSDave Chinner } 2281a66979abSDave Chinner out_unlock_inode: 2282a66979abSDave Chinner spin_unlock(&inode->i_lock); 228303ba3782SJens Axboe } 228403ba3782SJens Axboe EXPORT_SYMBOL(__mark_inode_dirty); 228503ba3782SJens Axboe 2286e97fedb9SDave Chinner /* 2287e97fedb9SDave Chinner * The @s_sync_lock is used to serialise concurrent sync operations 2288e97fedb9SDave Chinner * to avoid lock contention problems with concurrent wait_sb_inodes() calls. 2289e97fedb9SDave Chinner * Concurrent callers will block on the s_sync_lock rather than doing contending 2290e97fedb9SDave Chinner * walks. The queueing maintains sync(2) required behaviour as all the IO that 2291e97fedb9SDave Chinner * has been issued up to the time this function is enter is guaranteed to be 2292e97fedb9SDave Chinner * completed by the time we have gained the lock and waited for all IO that is 2293e97fedb9SDave Chinner * in progress regardless of the order callers are granted the lock. 2294e97fedb9SDave Chinner */ 2295b6e51316SJens Axboe static void wait_sb_inodes(struct super_block *sb) 229666f3b8e2SJens Axboe { 22976c60d2b5SDave Chinner LIST_HEAD(sync_list); 229838f21977SNick Piggin 229903ba3782SJens Axboe /* 230003ba3782SJens Axboe * We need to be protected against the filesystem going from 230103ba3782SJens Axboe * r/o to r/w or vice versa. 230203ba3782SJens Axboe */ 2303b6e51316SJens Axboe WARN_ON(!rwsem_is_locked(&sb->s_umount)); 230403ba3782SJens Axboe 2305e97fedb9SDave Chinner mutex_lock(&sb->s_sync_lock); 230666f3b8e2SJens Axboe 230738f21977SNick Piggin /* 23086c60d2b5SDave Chinner * Splice the writeback list onto a temporary list to avoid waiting on 23096c60d2b5SDave Chinner * inodes that have started writeback after this point. 23106c60d2b5SDave Chinner * 23116c60d2b5SDave Chinner * Use rcu_read_lock() to keep the inodes around until we have a 23126c60d2b5SDave Chinner * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as 23136c60d2b5SDave Chinner * the local list because inodes can be dropped from either by writeback 23146c60d2b5SDave Chinner * completion. 231538f21977SNick Piggin */ 23166c60d2b5SDave Chinner rcu_read_lock(); 23176c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 23186c60d2b5SDave Chinner list_splice_init(&sb->s_inodes_wb, &sync_list); 23196c60d2b5SDave Chinner 23206c60d2b5SDave Chinner /* 23216c60d2b5SDave Chinner * Data integrity sync. Must wait for all pages under writeback, because 23226c60d2b5SDave Chinner * there may have been pages dirtied before our sync call, but which had 23236c60d2b5SDave Chinner * writeout started before we write it out. In which case, the inode 23246c60d2b5SDave Chinner * may not be on the dirty list, but we still have to wait for that 23256c60d2b5SDave Chinner * writeout. 23266c60d2b5SDave Chinner */ 23276c60d2b5SDave Chinner while (!list_empty(&sync_list)) { 23286c60d2b5SDave Chinner struct inode *inode = list_first_entry(&sync_list, struct inode, 23296c60d2b5SDave Chinner i_wb_list); 2330250df6edSDave Chinner struct address_space *mapping = inode->i_mapping; 233138f21977SNick Piggin 23326c60d2b5SDave Chinner /* 23336c60d2b5SDave Chinner * Move each inode back to the wb list before we drop the lock 23346c60d2b5SDave Chinner * to preserve consistency between i_wb_list and the mapping 23356c60d2b5SDave Chinner * writeback tag. Writeback completion is responsible to remove 23366c60d2b5SDave Chinner * the inode from either list once the writeback tag is cleared. 23376c60d2b5SDave Chinner */ 23386c60d2b5SDave Chinner list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); 23396c60d2b5SDave Chinner 23406c60d2b5SDave Chinner /* 23416c60d2b5SDave Chinner * The mapping can appear untagged while still on-list since we 23426c60d2b5SDave Chinner * do not have the mapping lock. Skip it here, wb completion 23436c60d2b5SDave Chinner * will remove it. 23446c60d2b5SDave Chinner */ 23456c60d2b5SDave Chinner if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) 23466c60d2b5SDave Chinner continue; 23476c60d2b5SDave Chinner 23486c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 23496c60d2b5SDave Chinner 2350250df6edSDave Chinner spin_lock(&inode->i_lock); 23516c60d2b5SDave Chinner if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { 2352250df6edSDave Chinner spin_unlock(&inode->i_lock); 23536c60d2b5SDave Chinner 23546c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 235538f21977SNick Piggin continue; 2356250df6edSDave Chinner } 235738f21977SNick Piggin __iget(inode); 2358250df6edSDave Chinner spin_unlock(&inode->i_lock); 23596c60d2b5SDave Chinner rcu_read_unlock(); 236038f21977SNick Piggin 2361aa750fd7SJunichi Nomura /* 2362aa750fd7SJunichi Nomura * We keep the error status of individual mapping so that 2363aa750fd7SJunichi Nomura * applications can catch the writeback error using fsync(2). 2364aa750fd7SJunichi Nomura * See filemap_fdatawait_keep_errors() for details. 2365aa750fd7SJunichi Nomura */ 2366aa750fd7SJunichi Nomura filemap_fdatawait_keep_errors(mapping); 236738f21977SNick Piggin 236838f21977SNick Piggin cond_resched(); 236938f21977SNick Piggin 23706c60d2b5SDave Chinner iput(inode); 23716c60d2b5SDave Chinner 23726c60d2b5SDave Chinner rcu_read_lock(); 23736c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 237438f21977SNick Piggin } 23756c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 23766c60d2b5SDave Chinner rcu_read_unlock(); 2377e97fedb9SDave Chinner mutex_unlock(&sb->s_sync_lock); 237866f3b8e2SJens Axboe } 23791da177e4SLinus Torvalds 2380f30a7d0cSTejun Heo static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2381f30a7d0cSTejun Heo enum wb_reason reason, bool skip_if_busy) 23821da177e4SLinus Torvalds { 2383cc395d7fSTejun Heo DEFINE_WB_COMPLETION_ONSTACK(done); 238483ba7b07SChristoph Hellwig struct wb_writeback_work work = { 23853c4d7165SChristoph Hellwig .sb = sb, 23863c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_NONE, 23876e6938b6SWu Fengguang .tagged_writepages = 1, 238883ba7b07SChristoph Hellwig .done = &done, 23893259f8beSChris Mason .nr_pages = nr, 23900e175a18SCurt Wohlgemuth .reason = reason, 23913c4d7165SChristoph Hellwig }; 2392e7972912STejun Heo struct backing_dev_info *bdi = sb->s_bdi; 23930e3c9a22SJens Axboe 2394e7972912STejun Heo if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 23956eedc701SJan Kara return; 2396cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2397f30a7d0cSTejun Heo 2398db125360STejun Heo bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); 2399cc395d7fSTejun Heo wb_wait_for_completion(bdi, &done); 24001da177e4SLinus Torvalds } 2401f30a7d0cSTejun Heo 2402f30a7d0cSTejun Heo /** 2403f30a7d0cSTejun Heo * writeback_inodes_sb_nr - writeback dirty inodes from given super_block 2404f30a7d0cSTejun Heo * @sb: the superblock 2405f30a7d0cSTejun Heo * @nr: the number of pages to write 2406f30a7d0cSTejun Heo * @reason: reason why some writeback work initiated 2407f30a7d0cSTejun Heo * 2408f30a7d0cSTejun Heo * Start writeback on some inodes on this super_block. No guarantees are made 2409f30a7d0cSTejun Heo * on how many (if any) will be written, and this function does not wait 2410f30a7d0cSTejun Heo * for IO completion of submitted IO. 2411f30a7d0cSTejun Heo */ 2412f30a7d0cSTejun Heo void writeback_inodes_sb_nr(struct super_block *sb, 2413f30a7d0cSTejun Heo unsigned long nr, 2414f30a7d0cSTejun Heo enum wb_reason reason) 2415f30a7d0cSTejun Heo { 2416f30a7d0cSTejun Heo __writeback_inodes_sb_nr(sb, nr, reason, false); 2417f30a7d0cSTejun Heo } 24183259f8beSChris Mason EXPORT_SYMBOL(writeback_inodes_sb_nr); 24193259f8beSChris Mason 24203259f8beSChris Mason /** 24213259f8beSChris Mason * writeback_inodes_sb - writeback dirty inodes from given super_block 24223259f8beSChris Mason * @sb: the superblock 2423786228abSMarcos Paulo de Souza * @reason: reason why some writeback work was initiated 24243259f8beSChris Mason * 24253259f8beSChris Mason * Start writeback on some inodes on this super_block. No guarantees are made 24263259f8beSChris Mason * on how many (if any) will be written, and this function does not wait 24273259f8beSChris Mason * for IO completion of submitted IO. 24283259f8beSChris Mason */ 24290e175a18SCurt Wohlgemuth void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 24303259f8beSChris Mason { 24310e175a18SCurt Wohlgemuth return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 24323259f8beSChris Mason } 2433d8a8559cSJens Axboe EXPORT_SYMBOL(writeback_inodes_sb); 2434d8a8559cSJens Axboe 2435d8a8559cSJens Axboe /** 243610ee27a0SMiao Xie * try_to_writeback_inodes_sb - try to start writeback if none underway 243710ee27a0SMiao Xie * @sb: the superblock 243810ee27a0SMiao Xie * @reason: reason why some writeback work was initiated 243910ee27a0SMiao Xie * 24408264c321SRakesh Pandit * Invoke __writeback_inodes_sb_nr if no writeback is currently underway. 244110ee27a0SMiao Xie */ 24428264c321SRakesh Pandit void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 244310ee27a0SMiao Xie { 24448264c321SRakesh Pandit if (!down_read_trylock(&sb->s_umount)) 24458264c321SRakesh Pandit return; 24468264c321SRakesh Pandit 24478264c321SRakesh Pandit __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true); 24488264c321SRakesh Pandit up_read(&sb->s_umount); 244910ee27a0SMiao Xie } 245010ee27a0SMiao Xie EXPORT_SYMBOL(try_to_writeback_inodes_sb); 24513259f8beSChris Mason 24523259f8beSChris Mason /** 2453d8a8559cSJens Axboe * sync_inodes_sb - sync sb inode pages 2454d8a8559cSJens Axboe * @sb: the superblock 2455d8a8559cSJens Axboe * 2456d8a8559cSJens Axboe * This function writes and waits on any dirty inode belonging to this 24570dc83bd3SJan Kara * super_block. 2458d8a8559cSJens Axboe */ 24590dc83bd3SJan Kara void sync_inodes_sb(struct super_block *sb) 2460d8a8559cSJens Axboe { 2461cc395d7fSTejun Heo DEFINE_WB_COMPLETION_ONSTACK(done); 246283ba7b07SChristoph Hellwig struct wb_writeback_work work = { 24633c4d7165SChristoph Hellwig .sb = sb, 24643c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_ALL, 24653c4d7165SChristoph Hellwig .nr_pages = LONG_MAX, 24663c4d7165SChristoph Hellwig .range_cyclic = 0, 246783ba7b07SChristoph Hellwig .done = &done, 24680e175a18SCurt Wohlgemuth .reason = WB_REASON_SYNC, 24697747bd4bSDave Chinner .for_sync = 1, 24703c4d7165SChristoph Hellwig }; 2471e7972912STejun Heo struct backing_dev_info *bdi = sb->s_bdi; 24723c4d7165SChristoph Hellwig 2473006a0973STejun Heo /* 2474006a0973STejun Heo * Can't skip on !bdi_has_dirty() because we should wait for !dirty 2475006a0973STejun Heo * inodes under writeback and I_DIRTY_TIME inodes ignored by 2476006a0973STejun Heo * bdi_has_dirty() need to be written out too. 2477006a0973STejun Heo */ 2478006a0973STejun Heo if (bdi == &noop_backing_dev_info) 24796eedc701SJan Kara return; 2480cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2481cf37e972SChristoph Hellwig 24827fc5854fSTejun Heo /* protect against inode wb switch, see inode_switch_wbs_work_fn() */ 24837fc5854fSTejun Heo bdi_down_write_wb_switch_rwsem(bdi); 2484db125360STejun Heo bdi_split_work_to_wbs(bdi, &work, false); 2485cc395d7fSTejun Heo wb_wait_for_completion(bdi, &done); 24867fc5854fSTejun Heo bdi_up_write_wb_switch_rwsem(bdi); 248783ba7b07SChristoph Hellwig 2488b6e51316SJens Axboe wait_sb_inodes(sb); 2489d8a8559cSJens Axboe } 2490d8a8559cSJens Axboe EXPORT_SYMBOL(sync_inodes_sb); 24911da177e4SLinus Torvalds 24921da177e4SLinus Torvalds /** 24931da177e4SLinus Torvalds * write_inode_now - write an inode to disk 24941da177e4SLinus Torvalds * @inode: inode to write to disk 24951da177e4SLinus Torvalds * @sync: whether the write should be synchronous or not 24961da177e4SLinus Torvalds * 24977f04c26dSAndrea Arcangeli * This function commits an inode to disk immediately if it is dirty. This is 24987f04c26dSAndrea Arcangeli * primarily needed by knfsd. 24997f04c26dSAndrea Arcangeli * 25007f04c26dSAndrea Arcangeli * The caller must either have a ref on the inode or must have set I_WILL_FREE. 25011da177e4SLinus Torvalds */ 25021da177e4SLinus Torvalds int write_inode_now(struct inode *inode, int sync) 25031da177e4SLinus Torvalds { 25041da177e4SLinus Torvalds struct writeback_control wbc = { 25051da177e4SLinus Torvalds .nr_to_write = LONG_MAX, 250618914b18SMike Galbraith .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, 2507111ebb6eSOGAWA Hirofumi .range_start = 0, 2508111ebb6eSOGAWA Hirofumi .range_end = LLONG_MAX, 25091da177e4SLinus Torvalds }; 25101da177e4SLinus Torvalds 25111da177e4SLinus Torvalds if (!mapping_cap_writeback_dirty(inode->i_mapping)) 251249364ce2SAndrew Morton wbc.nr_to_write = 0; 25131da177e4SLinus Torvalds 25141da177e4SLinus Torvalds might_sleep(); 2515aaf25593STejun Heo return writeback_single_inode(inode, &wbc); 25161da177e4SLinus Torvalds } 25171da177e4SLinus Torvalds EXPORT_SYMBOL(write_inode_now); 25181da177e4SLinus Torvalds 25191da177e4SLinus Torvalds /** 25201da177e4SLinus Torvalds * sync_inode - write an inode and its pages to disk. 25211da177e4SLinus Torvalds * @inode: the inode to sync 25221da177e4SLinus Torvalds * @wbc: controls the writeback mode 25231da177e4SLinus Torvalds * 25241da177e4SLinus Torvalds * sync_inode() will write an inode and its pages to disk. It will also 25251da177e4SLinus Torvalds * correctly update the inode on its superblock's dirty inode lists and will 25261da177e4SLinus Torvalds * update inode->i_state. 25271da177e4SLinus Torvalds * 25281da177e4SLinus Torvalds * The caller must have a ref on the inode. 25291da177e4SLinus Torvalds */ 25301da177e4SLinus Torvalds int sync_inode(struct inode *inode, struct writeback_control *wbc) 25311da177e4SLinus Torvalds { 2532aaf25593STejun Heo return writeback_single_inode(inode, wbc); 25331da177e4SLinus Torvalds } 25341da177e4SLinus Torvalds EXPORT_SYMBOL(sync_inode); 2535c3765016SChristoph Hellwig 2536c3765016SChristoph Hellwig /** 2537c691b9d9SAndrew Morton * sync_inode_metadata - write an inode to disk 2538c3765016SChristoph Hellwig * @inode: the inode to sync 2539c3765016SChristoph Hellwig * @wait: wait for I/O to complete. 2540c3765016SChristoph Hellwig * 2541c691b9d9SAndrew Morton * Write an inode to disk and adjust its dirty state after completion. 2542c3765016SChristoph Hellwig * 2543c3765016SChristoph Hellwig * Note: only writes the actual inode, no associated data or other metadata. 2544c3765016SChristoph Hellwig */ 2545c3765016SChristoph Hellwig int sync_inode_metadata(struct inode *inode, int wait) 2546c3765016SChristoph Hellwig { 2547c3765016SChristoph Hellwig struct writeback_control wbc = { 2548c3765016SChristoph Hellwig .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, 2549c3765016SChristoph Hellwig .nr_to_write = 0, /* metadata-only */ 2550c3765016SChristoph Hellwig }; 2551c3765016SChristoph Hellwig 2552c3765016SChristoph Hellwig return sync_inode(inode, &wbc); 2553c3765016SChristoph Hellwig } 2554c3765016SChristoph Hellwig EXPORT_SYMBOL(sync_inode_metadata); 2555