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 39bc31b86aSWu Fengguang /* 40c4a77a6cSJens Axboe * Passed into wb_writeback(), essentially a subset of writeback_control 41c4a77a6cSJens Axboe */ 4283ba7b07SChristoph Hellwig struct wb_writeback_work { 43c4a77a6cSJens Axboe long nr_pages; 44c4a77a6cSJens Axboe struct super_block *sb; 45c4a77a6cSJens Axboe enum writeback_sync_modes sync_mode; 466e6938b6SWu Fengguang unsigned int tagged_writepages:1; 4752957fe1SH Hartley Sweeten unsigned int for_kupdate:1; 4852957fe1SH Hartley Sweeten unsigned int range_cyclic:1; 4952957fe1SH Hartley Sweeten unsigned int for_background:1; 507747bd4bSDave Chinner unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */ 51ac7b19a3STejun Heo unsigned int auto_free:1; /* free on completion */ 520e175a18SCurt Wohlgemuth enum wb_reason reason; /* why was writeback initiated? */ 53c4a77a6cSJens Axboe 548010c3b6SJens Axboe struct list_head list; /* pending work list */ 55cc395d7fSTejun Heo struct wb_completion *done; /* set if the caller waits */ 5603ba3782SJens Axboe }; 5703ba3782SJens Axboe 58a2f48706STheodore Ts'o /* 59a2f48706STheodore Ts'o * If an inode is constantly having its pages dirtied, but then the 60a2f48706STheodore Ts'o * updates stop dirtytime_expire_interval seconds in the past, it's 61a2f48706STheodore Ts'o * possible for the worst case time between when an inode has its 62a2f48706STheodore Ts'o * timestamps updated and when they finally get written out to be two 63a2f48706STheodore Ts'o * dirtytime_expire_intervals. We set the default to 12 hours (in 64a2f48706STheodore Ts'o * seconds), which means most of the time inodes will have their 65a2f48706STheodore Ts'o * timestamps written to disk after 12 hours, but in the worst case a 66a2f48706STheodore Ts'o * few inodes might not their timestamps updated for 24 hours. 67a2f48706STheodore Ts'o */ 68a2f48706STheodore Ts'o unsigned int dirtytime_expire_interval = 12 * 60 * 60; 69a2f48706STheodore Ts'o 707ccf19a8SNick Piggin static inline struct inode *wb_inode(struct list_head *head) 717ccf19a8SNick Piggin { 72c7f54084SDave Chinner return list_entry(head, struct inode, i_io_list); 737ccf19a8SNick Piggin } 747ccf19a8SNick Piggin 7515eb77a0SWu Fengguang /* 7615eb77a0SWu Fengguang * Include the creation of the trace points after defining the 7715eb77a0SWu Fengguang * wb_writeback_work structure and inline functions so that the definition 7815eb77a0SWu Fengguang * remains local to this file. 7915eb77a0SWu Fengguang */ 8015eb77a0SWu Fengguang #define CREATE_TRACE_POINTS 8115eb77a0SWu Fengguang #include <trace/events/writeback.h> 8215eb77a0SWu Fengguang 83774016b2SSteven Whitehouse EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage); 84774016b2SSteven Whitehouse 85d6c10f1fSTejun Heo static bool wb_io_lists_populated(struct bdi_writeback *wb) 86d6c10f1fSTejun Heo { 87d6c10f1fSTejun Heo if (wb_has_dirty_io(wb)) { 88d6c10f1fSTejun Heo return false; 89d6c10f1fSTejun Heo } else { 90d6c10f1fSTejun Heo set_bit(WB_has_dirty_io, &wb->state); 9195a46c65STejun Heo WARN_ON_ONCE(!wb->avg_write_bandwidth); 92766a9d6eSTejun Heo atomic_long_add(wb->avg_write_bandwidth, 93766a9d6eSTejun Heo &wb->bdi->tot_write_bandwidth); 94d6c10f1fSTejun Heo return true; 95d6c10f1fSTejun Heo } 96d6c10f1fSTejun Heo } 97d6c10f1fSTejun Heo 98d6c10f1fSTejun Heo static void wb_io_lists_depopulated(struct bdi_writeback *wb) 99d6c10f1fSTejun Heo { 100d6c10f1fSTejun Heo if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) && 101766a9d6eSTejun Heo list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) { 102d6c10f1fSTejun Heo clear_bit(WB_has_dirty_io, &wb->state); 10395a46c65STejun Heo WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth, 10495a46c65STejun Heo &wb->bdi->tot_write_bandwidth) < 0); 105766a9d6eSTejun Heo } 106d6c10f1fSTejun Heo } 107d6c10f1fSTejun Heo 108d6c10f1fSTejun Heo /** 109c7f54084SDave Chinner * inode_io_list_move_locked - move an inode onto a bdi_writeback IO list 110d6c10f1fSTejun Heo * @inode: inode to be moved 111d6c10f1fSTejun Heo * @wb: target bdi_writeback 112bbbc3c1cSWang Long * @head: one of @wb->b_{dirty|io|more_io|dirty_time} 113d6c10f1fSTejun Heo * 114c7f54084SDave Chinner * Move @inode->i_io_list to @list of @wb and set %WB_has_dirty_io. 115d6c10f1fSTejun Heo * Returns %true if @inode is the first occupant of the !dirty_time IO 116d6c10f1fSTejun Heo * lists; otherwise, %false. 117d6c10f1fSTejun Heo */ 118c7f54084SDave Chinner static bool inode_io_list_move_locked(struct inode *inode, 119d6c10f1fSTejun Heo struct bdi_writeback *wb, 120d6c10f1fSTejun Heo struct list_head *head) 121d6c10f1fSTejun Heo { 122d6c10f1fSTejun Heo assert_spin_locked(&wb->list_lock); 123d6c10f1fSTejun Heo 124c7f54084SDave Chinner list_move(&inode->i_io_list, head); 125d6c10f1fSTejun Heo 126d6c10f1fSTejun Heo /* dirty_time doesn't count as dirty_io until expiration */ 127d6c10f1fSTejun Heo if (head != &wb->b_dirty_time) 128d6c10f1fSTejun Heo return wb_io_lists_populated(wb); 129d6c10f1fSTejun Heo 130d6c10f1fSTejun Heo wb_io_lists_depopulated(wb); 131d6c10f1fSTejun Heo return false; 132d6c10f1fSTejun Heo } 133d6c10f1fSTejun Heo 134f0054bb1STejun Heo static void wb_wakeup(struct bdi_writeback *wb) 1355acda9d1SJan Kara { 136f0054bb1STejun Heo spin_lock_bh(&wb->work_lock); 137f0054bb1STejun Heo if (test_bit(WB_registered, &wb->state)) 138f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 139f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 1405acda9d1SJan Kara } 1415acda9d1SJan Kara 1424a3a485bSTahsin Erdogan static void finish_writeback_work(struct bdi_writeback *wb, 1434a3a485bSTahsin Erdogan struct wb_writeback_work *work) 1444a3a485bSTahsin Erdogan { 1454a3a485bSTahsin Erdogan struct wb_completion *done = work->done; 1464a3a485bSTahsin Erdogan 1474a3a485bSTahsin Erdogan if (work->auto_free) 1484a3a485bSTahsin Erdogan kfree(work); 1498e00c4e9STejun Heo if (done) { 1508e00c4e9STejun Heo wait_queue_head_t *waitq = done->waitq; 1518e00c4e9STejun Heo 1528e00c4e9STejun Heo /* @done can't be accessed after the following dec */ 1538e00c4e9STejun Heo if (atomic_dec_and_test(&done->cnt)) 1548e00c4e9STejun Heo wake_up_all(waitq); 1558e00c4e9STejun Heo } 1564a3a485bSTahsin Erdogan } 1574a3a485bSTahsin Erdogan 158f0054bb1STejun Heo static void wb_queue_work(struct bdi_writeback *wb, 1596585027aSJan Kara struct wb_writeback_work *work) 1606585027aSJan Kara { 1615634cc2aSTejun Heo trace_writeback_queue(wb, work); 1626585027aSJan Kara 163cc395d7fSTejun Heo if (work->done) 164cc395d7fSTejun Heo atomic_inc(&work->done->cnt); 1654a3a485bSTahsin Erdogan 1664a3a485bSTahsin Erdogan spin_lock_bh(&wb->work_lock); 1674a3a485bSTahsin Erdogan 1684a3a485bSTahsin Erdogan if (test_bit(WB_registered, &wb->state)) { 169f0054bb1STejun Heo list_add_tail(&work->list, &wb->work_list); 170f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 1714a3a485bSTahsin Erdogan } else 1724a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 1734a3a485bSTahsin Erdogan 174f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 17503ba3782SJens Axboe } 1761da177e4SLinus Torvalds 177cc395d7fSTejun Heo /** 178cc395d7fSTejun Heo * wb_wait_for_completion - wait for completion of bdi_writeback_works 179cc395d7fSTejun Heo * @done: target wb_completion 180cc395d7fSTejun Heo * 181cc395d7fSTejun Heo * Wait for one or more work items issued to @bdi with their ->done field 1825b9cce4cSTejun Heo * set to @done, which should have been initialized with 1835b9cce4cSTejun Heo * DEFINE_WB_COMPLETION(). This function returns after all such work items 1845b9cce4cSTejun Heo * are completed. Work items which are waited upon aren't freed 185cc395d7fSTejun Heo * automatically on completion. 186cc395d7fSTejun Heo */ 1875b9cce4cSTejun Heo void wb_wait_for_completion(struct wb_completion *done) 188cc395d7fSTejun Heo { 189cc395d7fSTejun Heo atomic_dec(&done->cnt); /* put down the initial count */ 1905b9cce4cSTejun Heo wait_event(*done->waitq, !atomic_read(&done->cnt)); 191cc395d7fSTejun Heo } 192cc395d7fSTejun Heo 193703c2708STejun Heo #ifdef CONFIG_CGROUP_WRITEBACK 194703c2708STejun Heo 19555a694dfSTejun Heo /* 19655a694dfSTejun Heo * Parameters for foreign inode detection, see wbc_detach_inode() to see 19755a694dfSTejun Heo * how they're used. 19855a694dfSTejun Heo * 19955a694dfSTejun Heo * These paramters are inherently heuristical as the detection target 20055a694dfSTejun Heo * itself is fuzzy. All we want to do is detaching an inode from the 20155a694dfSTejun Heo * current owner if it's being written to by some other cgroups too much. 20255a694dfSTejun Heo * 20355a694dfSTejun Heo * The current cgroup writeback is built on the assumption that multiple 20455a694dfSTejun Heo * cgroups writing to the same inode concurrently is very rare and a mode 20555a694dfSTejun Heo * of operation which isn't well supported. As such, the goal is not 20655a694dfSTejun Heo * taking too long when a different cgroup takes over an inode while 20755a694dfSTejun Heo * avoiding too aggressive flip-flops from occasional foreign writes. 20855a694dfSTejun Heo * 20955a694dfSTejun Heo * We record, very roughly, 2s worth of IO time history and if more than 21055a694dfSTejun Heo * half of that is foreign, trigger the switch. The recording is quantized 21155a694dfSTejun Heo * to 16 slots. To avoid tiny writes from swinging the decision too much, 21255a694dfSTejun Heo * writes smaller than 1/8 of avg size are ignored. 21355a694dfSTejun Heo */ 2142a814908STejun Heo #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */ 2152a814908STejun Heo #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */ 21655a694dfSTejun Heo #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */ 2172a814908STejun Heo #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */ 2182a814908STejun Heo 2192a814908STejun Heo #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */ 2202a814908STejun Heo #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS) 2212a814908STejun Heo /* each slot's duration is 2s / 16 */ 2222a814908STejun Heo #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2) 2232a814908STejun Heo /* if foreign slots >= 8, switch */ 2242a814908STejun Heo #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1) 2252a814908STejun Heo /* one round can affect upto 5 slots */ 2266444f47eSTejun Heo #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */ 2272a814908STejun Heo 228c22d70a1SRoman Gushchin /* 229c22d70a1SRoman Gushchin * Maximum inodes per isw. A specific value has been chosen to make 230c22d70a1SRoman Gushchin * struct inode_switch_wbs_context fit into 1024 bytes kmalloc. 231c22d70a1SRoman Gushchin */ 232c22d70a1SRoman Gushchin #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \ 233c22d70a1SRoman Gushchin / sizeof(struct inode *)) 234c22d70a1SRoman Gushchin 235a1a0e23eSTejun Heo static atomic_t isw_nr_in_flight = ATOMIC_INIT(0); 236a1a0e23eSTejun Heo static struct workqueue_struct *isw_wq; 237a1a0e23eSTejun Heo 23821c6321fSTejun Heo void __inode_attach_wb(struct inode *inode, struct page *page) 23921c6321fSTejun Heo { 24021c6321fSTejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 24121c6321fSTejun Heo struct bdi_writeback *wb = NULL; 24221c6321fSTejun Heo 24321c6321fSTejun Heo if (inode_cgwb_enabled(inode)) { 24421c6321fSTejun Heo struct cgroup_subsys_state *memcg_css; 24521c6321fSTejun Heo 24621c6321fSTejun Heo if (page) { 24721c6321fSTejun Heo memcg_css = mem_cgroup_css_from_page(page); 24821c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 24921c6321fSTejun Heo } else { 25021c6321fSTejun Heo /* must pin memcg_css, see wb_get_create() */ 25121c6321fSTejun Heo memcg_css = task_get_css(current, memory_cgrp_id); 25221c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 25321c6321fSTejun Heo css_put(memcg_css); 25421c6321fSTejun Heo } 25521c6321fSTejun Heo } 25621c6321fSTejun Heo 25721c6321fSTejun Heo if (!wb) 25821c6321fSTejun Heo wb = &bdi->wb; 25921c6321fSTejun Heo 26021c6321fSTejun Heo /* 26121c6321fSTejun Heo * There may be multiple instances of this function racing to 26221c6321fSTejun Heo * update the same inode. Use cmpxchg() to tell the winner. 26321c6321fSTejun Heo */ 26421c6321fSTejun Heo if (unlikely(cmpxchg(&inode->i_wb, NULL, wb))) 26521c6321fSTejun Heo wb_put(wb); 26621c6321fSTejun Heo } 2679b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(__inode_attach_wb); 26821c6321fSTejun Heo 269703c2708STejun Heo /** 270f3b6a6dfSRoman Gushchin * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list 271f3b6a6dfSRoman Gushchin * @inode: inode of interest with i_lock held 272f3b6a6dfSRoman Gushchin * @wb: target bdi_writeback 273f3b6a6dfSRoman Gushchin * 274f3b6a6dfSRoman Gushchin * Remove the inode from wb's io lists and if necessarily put onto b_attached 275f3b6a6dfSRoman Gushchin * list. Only inodes attached to cgwb's are kept on this list. 276f3b6a6dfSRoman Gushchin */ 277f3b6a6dfSRoman Gushchin static void inode_cgwb_move_to_attached(struct inode *inode, 278f3b6a6dfSRoman Gushchin struct bdi_writeback *wb) 279f3b6a6dfSRoman Gushchin { 280f3b6a6dfSRoman Gushchin assert_spin_locked(&wb->list_lock); 281f3b6a6dfSRoman Gushchin assert_spin_locked(&inode->i_lock); 282f3b6a6dfSRoman Gushchin 283f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 284f3b6a6dfSRoman Gushchin if (wb != &wb->bdi->wb) 285f3b6a6dfSRoman Gushchin list_move(&inode->i_io_list, &wb->b_attached); 286f3b6a6dfSRoman Gushchin else 287f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 288f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 289f3b6a6dfSRoman Gushchin } 290f3b6a6dfSRoman Gushchin 291f3b6a6dfSRoman Gushchin /** 29287e1d789STejun Heo * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it 29387e1d789STejun Heo * @inode: inode of interest with i_lock held 29487e1d789STejun Heo * 29587e1d789STejun Heo * Returns @inode's wb with its list_lock held. @inode->i_lock must be 29687e1d789STejun Heo * held on entry and is released on return. The returned wb is guaranteed 29787e1d789STejun Heo * to stay @inode's associated wb until its list_lock is released. 29887e1d789STejun Heo */ 29987e1d789STejun Heo static struct bdi_writeback * 30087e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 30187e1d789STejun Heo __releases(&inode->i_lock) 30287e1d789STejun Heo __acquires(&wb->list_lock) 30387e1d789STejun Heo { 30487e1d789STejun Heo while (true) { 30587e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 30687e1d789STejun Heo 30787e1d789STejun Heo /* 30887e1d789STejun Heo * inode_to_wb() association is protected by both 30987e1d789STejun Heo * @inode->i_lock and @wb->list_lock but list_lock nests 31087e1d789STejun Heo * outside i_lock. Drop i_lock and verify that the 31187e1d789STejun Heo * association hasn't changed after acquiring list_lock. 31287e1d789STejun Heo */ 31387e1d789STejun Heo wb_get(wb); 31487e1d789STejun Heo spin_unlock(&inode->i_lock); 31587e1d789STejun Heo spin_lock(&wb->list_lock); 31687e1d789STejun Heo 317aaa2cacfSTejun Heo /* i_wb may have changed inbetween, can't use inode_to_wb() */ 318614a4e37STejun Heo if (likely(wb == inode->i_wb)) { 319614a4e37STejun Heo wb_put(wb); /* @inode already has ref */ 320614a4e37STejun Heo return wb; 321614a4e37STejun Heo } 32287e1d789STejun Heo 32387e1d789STejun Heo spin_unlock(&wb->list_lock); 324614a4e37STejun Heo wb_put(wb); 32587e1d789STejun Heo cpu_relax(); 32687e1d789STejun Heo spin_lock(&inode->i_lock); 32787e1d789STejun Heo } 32887e1d789STejun Heo } 32987e1d789STejun Heo 33087e1d789STejun Heo /** 33187e1d789STejun Heo * inode_to_wb_and_lock_list - determine an inode's wb and lock it 33287e1d789STejun Heo * @inode: inode of interest 33387e1d789STejun Heo * 33487e1d789STejun Heo * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held 33587e1d789STejun Heo * on entry. 33687e1d789STejun Heo */ 33787e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 33887e1d789STejun Heo __acquires(&wb->list_lock) 33987e1d789STejun Heo { 34087e1d789STejun Heo spin_lock(&inode->i_lock); 34187e1d789STejun Heo return locked_inode_to_wb_and_lock_list(inode); 34287e1d789STejun Heo } 34387e1d789STejun Heo 344682aa8e1STejun Heo struct inode_switch_wbs_context { 34529264d92SRoman Gushchin struct rcu_work work; 346f5fbe6b7SRoman Gushchin 347f5fbe6b7SRoman Gushchin /* 348f5fbe6b7SRoman Gushchin * Multiple inodes can be switched at once. The switching procedure 349f5fbe6b7SRoman Gushchin * consists of two parts, separated by a RCU grace period. To make 350f5fbe6b7SRoman Gushchin * sure that the second part is executed for each inode gone through 351f5fbe6b7SRoman Gushchin * the first part, all inode pointers are placed into a NULL-terminated 352f5fbe6b7SRoman Gushchin * array embedded into struct inode_switch_wbs_context. Otherwise 353f5fbe6b7SRoman Gushchin * an inode could be left in a non-consistent state. 354f5fbe6b7SRoman Gushchin */ 355f5fbe6b7SRoman Gushchin struct bdi_writeback *new_wb; 356f5fbe6b7SRoman Gushchin struct inode *inodes[]; 357682aa8e1STejun Heo }; 358682aa8e1STejun Heo 3597fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) 3607fc5854fSTejun Heo { 3617fc5854fSTejun Heo down_write(&bdi->wb_switch_rwsem); 3627fc5854fSTejun Heo } 3637fc5854fSTejun Heo 3647fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) 3657fc5854fSTejun Heo { 3667fc5854fSTejun Heo up_write(&bdi->wb_switch_rwsem); 3677fc5854fSTejun Heo } 3687fc5854fSTejun Heo 369f5fbe6b7SRoman Gushchin static bool inode_do_switch_wbs(struct inode *inode, 370f5fbe6b7SRoman Gushchin struct bdi_writeback *old_wb, 37172d4512eSRoman Gushchin struct bdi_writeback *new_wb) 372682aa8e1STejun Heo { 373d10c8095STejun Heo struct address_space *mapping = inode->i_mapping; 37404edf02cSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, 0); 37522b3c8d6SMatthew Wilcox (Oracle) struct folio *folio; 376d10c8095STejun Heo bool switched = false; 377682aa8e1STejun Heo 378682aa8e1STejun Heo spin_lock(&inode->i_lock); 379b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 380682aa8e1STejun Heo 381d10c8095STejun Heo /* 3824ade5867SRoman Gushchin * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction 3834ade5867SRoman Gushchin * path owns the inode and we shouldn't modify ->i_io_list. 384d10c8095STejun Heo */ 3854ade5867SRoman Gushchin if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE))) 386d10c8095STejun Heo goto skip_switch; 387d10c8095STejun Heo 3883a8e9ac8STejun Heo trace_inode_switch_wbs(inode, old_wb, new_wb); 3893a8e9ac8STejun Heo 390d10c8095STejun Heo /* 391d10c8095STejun Heo * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points 39222b3c8d6SMatthew Wilcox (Oracle) * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to 39322b3c8d6SMatthew Wilcox (Oracle) * folios actually under writeback. 394d10c8095STejun Heo */ 39522b3c8d6SMatthew Wilcox (Oracle) xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) { 39622b3c8d6SMatthew Wilcox (Oracle) if (folio_test_dirty(folio)) { 39722b3c8d6SMatthew Wilcox (Oracle) long nr = folio_nr_pages(folio); 39822b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr); 39922b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(new_wb, WB_RECLAIMABLE, nr); 400d10c8095STejun Heo } 401d10c8095STejun Heo } 402d10c8095STejun Heo 40304edf02cSMatthew Wilcox xas_set(&xas, 0); 40422b3c8d6SMatthew Wilcox (Oracle) xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) { 40522b3c8d6SMatthew Wilcox (Oracle) long nr = folio_nr_pages(folio); 40622b3c8d6SMatthew Wilcox (Oracle) WARN_ON_ONCE(!folio_test_writeback(folio)); 40722b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(old_wb, WB_WRITEBACK, -nr); 40822b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(new_wb, WB_WRITEBACK, nr); 409d10c8095STejun Heo } 410d10c8095STejun Heo 411633a2abbSJan Kara if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) { 412633a2abbSJan Kara atomic_dec(&old_wb->writeback_inodes); 413633a2abbSJan Kara atomic_inc(&new_wb->writeback_inodes); 414633a2abbSJan Kara } 415633a2abbSJan Kara 416d10c8095STejun Heo wb_get(new_wb); 417d10c8095STejun Heo 418d10c8095STejun Heo /* 419f3b6a6dfSRoman Gushchin * Transfer to @new_wb's IO list if necessary. If the @inode is dirty, 420f3b6a6dfSRoman Gushchin * the specific list @inode was on is ignored and the @inode is put on 421f3b6a6dfSRoman Gushchin * ->b_dirty which is always correct including from ->b_dirty_time. 422f3b6a6dfSRoman Gushchin * The transfer preserves @inode->dirtied_when ordering. If the @inode 423f3b6a6dfSRoman Gushchin * was clean, it means it was on the b_attached list, so move it onto 424f3b6a6dfSRoman Gushchin * the b_attached list of @new_wb. 425d10c8095STejun Heo */ 426c7f54084SDave Chinner if (!list_empty(&inode->i_io_list)) { 427f3b6a6dfSRoman Gushchin inode->i_wb = new_wb; 428f3b6a6dfSRoman Gushchin 429f3b6a6dfSRoman Gushchin if (inode->i_state & I_DIRTY_ALL) { 430d10c8095STejun Heo struct inode *pos; 431d10c8095STejun Heo 432c7f54084SDave Chinner list_for_each_entry(pos, &new_wb->b_dirty, i_io_list) 433d10c8095STejun Heo if (time_after_eq(inode->dirtied_when, 434d10c8095STejun Heo pos->dirtied_when)) 435d10c8095STejun Heo break; 436f3b6a6dfSRoman Gushchin inode_io_list_move_locked(inode, new_wb, 437f3b6a6dfSRoman Gushchin pos->i_io_list.prev); 438f3b6a6dfSRoman Gushchin } else { 439f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, new_wb); 440f3b6a6dfSRoman Gushchin } 441d10c8095STejun Heo } else { 442d10c8095STejun Heo inode->i_wb = new_wb; 443d10c8095STejun Heo } 444d10c8095STejun Heo 445d10c8095STejun Heo /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */ 446682aa8e1STejun Heo inode->i_wb_frn_winner = 0; 447682aa8e1STejun Heo inode->i_wb_frn_avg_time = 0; 448682aa8e1STejun Heo inode->i_wb_frn_history = 0; 449d10c8095STejun Heo switched = true; 450d10c8095STejun Heo skip_switch: 451682aa8e1STejun Heo /* 452682aa8e1STejun Heo * Paired with load_acquire in unlocked_inode_to_wb_begin() and 453682aa8e1STejun Heo * ensures that the new wb is visible if they see !I_WB_SWITCH. 454682aa8e1STejun Heo */ 455682aa8e1STejun Heo smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); 456682aa8e1STejun Heo 457b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 458682aa8e1STejun Heo spin_unlock(&inode->i_lock); 459d10c8095STejun Heo 460f5fbe6b7SRoman Gushchin return switched; 46172d4512eSRoman Gushchin } 462682aa8e1STejun Heo 46372d4512eSRoman Gushchin static void inode_switch_wbs_work_fn(struct work_struct *work) 46472d4512eSRoman Gushchin { 46572d4512eSRoman Gushchin struct inode_switch_wbs_context *isw = 46672d4512eSRoman Gushchin container_of(to_rcu_work(work), struct inode_switch_wbs_context, work); 467f5fbe6b7SRoman Gushchin struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]); 468f5fbe6b7SRoman Gushchin struct bdi_writeback *old_wb = isw->inodes[0]->i_wb; 469f5fbe6b7SRoman Gushchin struct bdi_writeback *new_wb = isw->new_wb; 470f5fbe6b7SRoman Gushchin unsigned long nr_switched = 0; 471f5fbe6b7SRoman Gushchin struct inode **inodep; 47272d4512eSRoman Gushchin 473f5fbe6b7SRoman Gushchin /* 474f5fbe6b7SRoman Gushchin * If @inode switches cgwb membership while sync_inodes_sb() is 475f5fbe6b7SRoman Gushchin * being issued, sync_inodes_sb() might miss it. Synchronize. 476f5fbe6b7SRoman Gushchin */ 477f5fbe6b7SRoman Gushchin down_read(&bdi->wb_switch_rwsem); 478f5fbe6b7SRoman Gushchin 479f5fbe6b7SRoman Gushchin /* 480f5fbe6b7SRoman Gushchin * By the time control reaches here, RCU grace period has passed 481f5fbe6b7SRoman Gushchin * since I_WB_SWITCH assertion and all wb stat update transactions 482f5fbe6b7SRoman Gushchin * between unlocked_inode_to_wb_begin/end() are guaranteed to be 483f5fbe6b7SRoman Gushchin * synchronizing against the i_pages lock. 484f5fbe6b7SRoman Gushchin * 485f5fbe6b7SRoman Gushchin * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock 486f5fbe6b7SRoman Gushchin * gives us exclusion against all wb related operations on @inode 487f5fbe6b7SRoman Gushchin * including IO list manipulations and stat updates. 488f5fbe6b7SRoman Gushchin */ 489f5fbe6b7SRoman Gushchin if (old_wb < new_wb) { 490f5fbe6b7SRoman Gushchin spin_lock(&old_wb->list_lock); 491f5fbe6b7SRoman Gushchin spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING); 492f5fbe6b7SRoman Gushchin } else { 493f5fbe6b7SRoman Gushchin spin_lock(&new_wb->list_lock); 494f5fbe6b7SRoman Gushchin spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING); 495f5fbe6b7SRoman Gushchin } 496f5fbe6b7SRoman Gushchin 497f5fbe6b7SRoman Gushchin for (inodep = isw->inodes; *inodep; inodep++) { 498f5fbe6b7SRoman Gushchin WARN_ON_ONCE((*inodep)->i_wb != old_wb); 499f5fbe6b7SRoman Gushchin if (inode_do_switch_wbs(*inodep, old_wb, new_wb)) 500f5fbe6b7SRoman Gushchin nr_switched++; 501f5fbe6b7SRoman Gushchin } 502f5fbe6b7SRoman Gushchin 503f5fbe6b7SRoman Gushchin spin_unlock(&new_wb->list_lock); 504f5fbe6b7SRoman Gushchin spin_unlock(&old_wb->list_lock); 505f5fbe6b7SRoman Gushchin 506f5fbe6b7SRoman Gushchin up_read(&bdi->wb_switch_rwsem); 507f5fbe6b7SRoman Gushchin 508f5fbe6b7SRoman Gushchin if (nr_switched) { 509f5fbe6b7SRoman Gushchin wb_wakeup(new_wb); 510f5fbe6b7SRoman Gushchin wb_put_many(old_wb, nr_switched); 511f5fbe6b7SRoman Gushchin } 512f5fbe6b7SRoman Gushchin 513f5fbe6b7SRoman Gushchin for (inodep = isw->inodes; *inodep; inodep++) 514f5fbe6b7SRoman Gushchin iput(*inodep); 515f5fbe6b7SRoman Gushchin wb_put(new_wb); 516682aa8e1STejun Heo kfree(isw); 517a1a0e23eSTejun Heo atomic_dec(&isw_nr_in_flight); 518682aa8e1STejun Heo } 519682aa8e1STejun Heo 520c22d70a1SRoman Gushchin static bool inode_prepare_wbs_switch(struct inode *inode, 521c22d70a1SRoman Gushchin struct bdi_writeback *new_wb) 522c22d70a1SRoman Gushchin { 523c22d70a1SRoman Gushchin /* 524c22d70a1SRoman Gushchin * Paired with smp_mb() in cgroup_writeback_umount(). 525c22d70a1SRoman Gushchin * isw_nr_in_flight must be increased before checking SB_ACTIVE and 526c22d70a1SRoman Gushchin * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0 527c22d70a1SRoman Gushchin * in cgroup_writeback_umount() and the isw_wq will be not flushed. 528c22d70a1SRoman Gushchin */ 529c22d70a1SRoman Gushchin smp_mb(); 530c22d70a1SRoman Gushchin 531593311e8SRoman Gushchin if (IS_DAX(inode)) 532593311e8SRoman Gushchin return false; 533593311e8SRoman Gushchin 534c22d70a1SRoman Gushchin /* while holding I_WB_SWITCH, no one else can update the association */ 535c22d70a1SRoman Gushchin spin_lock(&inode->i_lock); 536c22d70a1SRoman Gushchin if (!(inode->i_sb->s_flags & SB_ACTIVE) || 537c22d70a1SRoman Gushchin inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) || 538c22d70a1SRoman Gushchin inode_to_wb(inode) == new_wb) { 539c22d70a1SRoman Gushchin spin_unlock(&inode->i_lock); 540c22d70a1SRoman Gushchin return false; 541c22d70a1SRoman Gushchin } 542c22d70a1SRoman Gushchin inode->i_state |= I_WB_SWITCH; 543c22d70a1SRoman Gushchin __iget(inode); 544c22d70a1SRoman Gushchin spin_unlock(&inode->i_lock); 545c22d70a1SRoman Gushchin 546c22d70a1SRoman Gushchin return true; 547c22d70a1SRoman Gushchin } 548c22d70a1SRoman Gushchin 549682aa8e1STejun Heo /** 550682aa8e1STejun Heo * inode_switch_wbs - change the wb association of an inode 551682aa8e1STejun Heo * @inode: target inode 552682aa8e1STejun Heo * @new_wb_id: ID of the new wb 553682aa8e1STejun Heo * 554682aa8e1STejun Heo * Switch @inode's wb association to the wb identified by @new_wb_id. The 555682aa8e1STejun Heo * switching is performed asynchronously and may fail silently. 556682aa8e1STejun Heo */ 557682aa8e1STejun Heo static void inode_switch_wbs(struct inode *inode, int new_wb_id) 558682aa8e1STejun Heo { 559682aa8e1STejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 560682aa8e1STejun Heo struct cgroup_subsys_state *memcg_css; 561682aa8e1STejun Heo struct inode_switch_wbs_context *isw; 562682aa8e1STejun Heo 563682aa8e1STejun Heo /* noop if seems to be already in progress */ 564682aa8e1STejun Heo if (inode->i_state & I_WB_SWITCH) 565682aa8e1STejun Heo return; 566682aa8e1STejun Heo 5676444f47eSTejun Heo /* avoid queueing a new switch if too many are already in flight */ 5686444f47eSTejun Heo if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT) 5697fc5854fSTejun Heo return; 5707fc5854fSTejun Heo 57198b160c8SLen Baker isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC); 572682aa8e1STejun Heo if (!isw) 5736444f47eSTejun Heo return; 574682aa8e1STejun Heo 5758826ee4fSRoman Gushchin atomic_inc(&isw_nr_in_flight); 5768826ee4fSRoman Gushchin 577682aa8e1STejun Heo /* find and pin the new wb */ 578682aa8e1STejun Heo rcu_read_lock(); 579682aa8e1STejun Heo memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); 5808b0ed844SMuchun Song if (memcg_css && !css_tryget(memcg_css)) 5818b0ed844SMuchun Song memcg_css = NULL; 582682aa8e1STejun Heo rcu_read_unlock(); 5838b0ed844SMuchun Song if (!memcg_css) 5848b0ed844SMuchun Song goto out_free; 5858b0ed844SMuchun Song 5868b0ed844SMuchun Song isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 5878b0ed844SMuchun Song css_put(memcg_css); 588682aa8e1STejun Heo if (!isw->new_wb) 589682aa8e1STejun Heo goto out_free; 590682aa8e1STejun Heo 591c22d70a1SRoman Gushchin if (!inode_prepare_wbs_switch(inode, isw->new_wb)) 592a1a0e23eSTejun Heo goto out_free; 593682aa8e1STejun Heo 594f5fbe6b7SRoman Gushchin isw->inodes[0] = inode; 595682aa8e1STejun Heo 596682aa8e1STejun Heo /* 597682aa8e1STejun Heo * In addition to synchronizing among switchers, I_WB_SWITCH tells 598b93b0163SMatthew Wilcox * the RCU protected stat update paths to grab the i_page 599b93b0163SMatthew Wilcox * lock so that stat transfer can synchronize against them. 600682aa8e1STejun Heo * Let's continue after I_WB_SWITCH is guaranteed to be visible. 601682aa8e1STejun Heo */ 60229264d92SRoman Gushchin INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); 60329264d92SRoman Gushchin queue_rcu_work(isw_wq, &isw->work); 6046444f47eSTejun Heo return; 605682aa8e1STejun Heo 606682aa8e1STejun Heo out_free: 6078826ee4fSRoman Gushchin atomic_dec(&isw_nr_in_flight); 608682aa8e1STejun Heo if (isw->new_wb) 609682aa8e1STejun Heo wb_put(isw->new_wb); 610682aa8e1STejun Heo kfree(isw); 611682aa8e1STejun Heo } 612682aa8e1STejun Heo 61387e1d789STejun Heo /** 614c22d70a1SRoman Gushchin * cleanup_offline_cgwb - detach associated inodes 615c22d70a1SRoman Gushchin * @wb: target wb 616c22d70a1SRoman Gushchin * 617c22d70a1SRoman Gushchin * Switch all inodes attached to @wb to a nearest living ancestor's wb in order 618c22d70a1SRoman Gushchin * to eventually release the dying @wb. Returns %true if not all inodes were 619c22d70a1SRoman Gushchin * switched and the function has to be restarted. 620c22d70a1SRoman Gushchin */ 621c22d70a1SRoman Gushchin bool cleanup_offline_cgwb(struct bdi_writeback *wb) 622c22d70a1SRoman Gushchin { 623c22d70a1SRoman Gushchin struct cgroup_subsys_state *memcg_css; 624c22d70a1SRoman Gushchin struct inode_switch_wbs_context *isw; 625c22d70a1SRoman Gushchin struct inode *inode; 626c22d70a1SRoman Gushchin int nr; 627c22d70a1SRoman Gushchin bool restart = false; 628c22d70a1SRoman Gushchin 62998b160c8SLen Baker isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW), 63098b160c8SLen Baker GFP_KERNEL); 631c22d70a1SRoman Gushchin if (!isw) 632c22d70a1SRoman Gushchin return restart; 633c22d70a1SRoman Gushchin 634c22d70a1SRoman Gushchin atomic_inc(&isw_nr_in_flight); 635c22d70a1SRoman Gushchin 636c22d70a1SRoman Gushchin for (memcg_css = wb->memcg_css->parent; memcg_css; 637c22d70a1SRoman Gushchin memcg_css = memcg_css->parent) { 638c22d70a1SRoman Gushchin isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL); 639c22d70a1SRoman Gushchin if (isw->new_wb) 640c22d70a1SRoman Gushchin break; 641c22d70a1SRoman Gushchin } 642c22d70a1SRoman Gushchin if (unlikely(!isw->new_wb)) 643c22d70a1SRoman Gushchin isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */ 644c22d70a1SRoman Gushchin 645c22d70a1SRoman Gushchin nr = 0; 646c22d70a1SRoman Gushchin spin_lock(&wb->list_lock); 647c22d70a1SRoman Gushchin list_for_each_entry(inode, &wb->b_attached, i_io_list) { 648c22d70a1SRoman Gushchin if (!inode_prepare_wbs_switch(inode, isw->new_wb)) 649c22d70a1SRoman Gushchin continue; 650c22d70a1SRoman Gushchin 651c22d70a1SRoman Gushchin isw->inodes[nr++] = inode; 652c22d70a1SRoman Gushchin 653c22d70a1SRoman Gushchin if (nr >= WB_MAX_INODES_PER_ISW - 1) { 654c22d70a1SRoman Gushchin restart = true; 655c22d70a1SRoman Gushchin break; 656c22d70a1SRoman Gushchin } 657c22d70a1SRoman Gushchin } 658c22d70a1SRoman Gushchin spin_unlock(&wb->list_lock); 659c22d70a1SRoman Gushchin 660c22d70a1SRoman Gushchin /* no attached inodes? bail out */ 661c22d70a1SRoman Gushchin if (nr == 0) { 662c22d70a1SRoman Gushchin atomic_dec(&isw_nr_in_flight); 663c22d70a1SRoman Gushchin wb_put(isw->new_wb); 664c22d70a1SRoman Gushchin kfree(isw); 665c22d70a1SRoman Gushchin return restart; 666c22d70a1SRoman Gushchin } 667c22d70a1SRoman Gushchin 668c22d70a1SRoman Gushchin /* 669c22d70a1SRoman Gushchin * In addition to synchronizing among switchers, I_WB_SWITCH tells 670c22d70a1SRoman Gushchin * the RCU protected stat update paths to grab the i_page 671c22d70a1SRoman Gushchin * lock so that stat transfer can synchronize against them. 672c22d70a1SRoman Gushchin * Let's continue after I_WB_SWITCH is guaranteed to be visible. 673c22d70a1SRoman Gushchin */ 674c22d70a1SRoman Gushchin INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); 675c22d70a1SRoman Gushchin queue_rcu_work(isw_wq, &isw->work); 676c22d70a1SRoman Gushchin 677c22d70a1SRoman Gushchin return restart; 678c22d70a1SRoman Gushchin } 679c22d70a1SRoman Gushchin 680c22d70a1SRoman Gushchin /** 681b16b1debSTejun Heo * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it 682b16b1debSTejun Heo * @wbc: writeback_control of interest 683b16b1debSTejun Heo * @inode: target inode 684b16b1debSTejun Heo * 685b16b1debSTejun Heo * @inode is locked and about to be written back under the control of @wbc. 686b16b1debSTejun Heo * Record @inode's writeback context into @wbc and unlock the i_lock. On 687b16b1debSTejun Heo * writeback completion, wbc_detach_inode() should be called. This is used 688b16b1debSTejun Heo * to track the cgroup writeback context. 689b16b1debSTejun Heo */ 690b16b1debSTejun Heo void wbc_attach_and_unlock_inode(struct writeback_control *wbc, 691b16b1debSTejun Heo struct inode *inode) 692b16b1debSTejun Heo { 693dd73e4b7STejun Heo if (!inode_cgwb_enabled(inode)) { 694dd73e4b7STejun Heo spin_unlock(&inode->i_lock); 695dd73e4b7STejun Heo return; 696dd73e4b7STejun Heo } 697dd73e4b7STejun Heo 698b16b1debSTejun Heo wbc->wb = inode_to_wb(inode); 6992a814908STejun Heo wbc->inode = inode; 7002a814908STejun Heo 7012a814908STejun Heo wbc->wb_id = wbc->wb->memcg_css->id; 7022a814908STejun Heo wbc->wb_lcand_id = inode->i_wb_frn_winner; 7032a814908STejun Heo wbc->wb_tcand_id = 0; 7042a814908STejun Heo wbc->wb_bytes = 0; 7052a814908STejun Heo wbc->wb_lcand_bytes = 0; 7062a814908STejun Heo wbc->wb_tcand_bytes = 0; 7072a814908STejun Heo 708b16b1debSTejun Heo wb_get(wbc->wb); 709b16b1debSTejun Heo spin_unlock(&inode->i_lock); 710e8a7abf5STejun Heo 711e8a7abf5STejun Heo /* 71265de03e2STejun Heo * A dying wb indicates that either the blkcg associated with the 71365de03e2STejun Heo * memcg changed or the associated memcg is dying. In the first 71465de03e2STejun Heo * case, a replacement wb should already be available and we should 71565de03e2STejun Heo * refresh the wb immediately. In the second case, trying to 71665de03e2STejun Heo * refresh will keep failing. 717e8a7abf5STejun Heo */ 71865de03e2STejun Heo if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css))) 719e8a7abf5STejun Heo inode_switch_wbs(inode, wbc->wb_id); 720b16b1debSTejun Heo } 7219b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode); 722b16b1debSTejun Heo 723b16b1debSTejun Heo /** 7242a814908STejun Heo * wbc_detach_inode - disassociate wbc from inode and perform foreign detection 7252a814908STejun Heo * @wbc: writeback_control of the just finished writeback 726b16b1debSTejun Heo * 727b16b1debSTejun Heo * To be called after a writeback attempt of an inode finishes and undoes 728b16b1debSTejun Heo * wbc_attach_and_unlock_inode(). Can be called under any context. 7292a814908STejun Heo * 7302a814908STejun Heo * As concurrent write sharing of an inode is expected to be very rare and 7312a814908STejun Heo * memcg only tracks page ownership on first-use basis severely confining 7322a814908STejun Heo * the usefulness of such sharing, cgroup writeback tracks ownership 7332a814908STejun Heo * per-inode. While the support for concurrent write sharing of an inode 7342a814908STejun Heo * is deemed unnecessary, an inode being written to by different cgroups at 7352a814908STejun Heo * different points in time is a lot more common, and, more importantly, 7362a814908STejun Heo * charging only by first-use can too readily lead to grossly incorrect 7372a814908STejun Heo * behaviors (single foreign page can lead to gigabytes of writeback to be 7382a814908STejun Heo * incorrectly attributed). 7392a814908STejun Heo * 7402a814908STejun Heo * To resolve this issue, cgroup writeback detects the majority dirtier of 7412a814908STejun Heo * an inode and transfers the ownership to it. To avoid unnnecessary 7422a814908STejun Heo * oscillation, the detection mechanism keeps track of history and gives 7432a814908STejun Heo * out the switch verdict only if the foreign usage pattern is stable over 7442a814908STejun Heo * a certain amount of time and/or writeback attempts. 7452a814908STejun Heo * 7462a814908STejun Heo * On each writeback attempt, @wbc tries to detect the majority writer 7472a814908STejun Heo * using Boyer-Moore majority vote algorithm. In addition to the byte 7482a814908STejun Heo * count from the majority voting, it also counts the bytes written for the 7492a814908STejun Heo * current wb and the last round's winner wb (max of last round's current 7502a814908STejun Heo * wb, the winner from two rounds ago, and the last round's majority 7512a814908STejun Heo * candidate). Keeping track of the historical winner helps the algorithm 7522a814908STejun Heo * to semi-reliably detect the most active writer even when it's not the 7532a814908STejun Heo * absolute majority. 7542a814908STejun Heo * 7552a814908STejun Heo * Once the winner of the round is determined, whether the winner is 7562a814908STejun Heo * foreign or not and how much IO time the round consumed is recorded in 7572a814908STejun Heo * inode->i_wb_frn_history. If the amount of recorded foreign IO time is 7582a814908STejun Heo * over a certain threshold, the switch verdict is given. 759b16b1debSTejun Heo */ 760b16b1debSTejun Heo void wbc_detach_inode(struct writeback_control *wbc) 761b16b1debSTejun Heo { 7622a814908STejun Heo struct bdi_writeback *wb = wbc->wb; 7632a814908STejun Heo struct inode *inode = wbc->inode; 764dd73e4b7STejun Heo unsigned long avg_time, max_bytes, max_time; 765dd73e4b7STejun Heo u16 history; 7662a814908STejun Heo int max_id; 7672a814908STejun Heo 768dd73e4b7STejun Heo if (!wb) 769dd73e4b7STejun Heo return; 770dd73e4b7STejun Heo 771dd73e4b7STejun Heo history = inode->i_wb_frn_history; 772dd73e4b7STejun Heo avg_time = inode->i_wb_frn_avg_time; 773dd73e4b7STejun Heo 7742a814908STejun Heo /* pick the winner of this round */ 7752a814908STejun Heo if (wbc->wb_bytes >= wbc->wb_lcand_bytes && 7762a814908STejun Heo wbc->wb_bytes >= wbc->wb_tcand_bytes) { 7772a814908STejun Heo max_id = wbc->wb_id; 7782a814908STejun Heo max_bytes = wbc->wb_bytes; 7792a814908STejun Heo } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { 7802a814908STejun Heo max_id = wbc->wb_lcand_id; 7812a814908STejun Heo max_bytes = wbc->wb_lcand_bytes; 7822a814908STejun Heo } else { 7832a814908STejun Heo max_id = wbc->wb_tcand_id; 7842a814908STejun Heo max_bytes = wbc->wb_tcand_bytes; 7852a814908STejun Heo } 7862a814908STejun Heo 7872a814908STejun Heo /* 7882a814908STejun Heo * Calculate the amount of IO time the winner consumed and fold it 7892a814908STejun Heo * into the running average kept per inode. If the consumed IO 7902a814908STejun Heo * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for 7912a814908STejun Heo * deciding whether to switch or not. This is to prevent one-off 7922a814908STejun Heo * small dirtiers from skewing the verdict. 7932a814908STejun Heo */ 7942a814908STejun Heo max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, 7952a814908STejun Heo wb->avg_write_bandwidth); 7962a814908STejun Heo if (avg_time) 7972a814908STejun Heo avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - 7982a814908STejun Heo (avg_time >> WB_FRN_TIME_AVG_SHIFT); 7992a814908STejun Heo else 8002a814908STejun Heo avg_time = max_time; /* immediate catch up on first run */ 8012a814908STejun Heo 8022a814908STejun Heo if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { 8032a814908STejun Heo int slots; 8042a814908STejun Heo 8052a814908STejun Heo /* 8062a814908STejun Heo * The switch verdict is reached if foreign wb's consume 8072a814908STejun Heo * more than a certain proportion of IO time in a 8082a814908STejun Heo * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot 8092a814908STejun Heo * history mask where each bit represents one sixteenth of 8102a814908STejun Heo * the period. Determine the number of slots to shift into 8112a814908STejun Heo * history from @max_time. 8122a814908STejun Heo */ 8132a814908STejun Heo slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), 8142a814908STejun Heo (unsigned long)WB_FRN_HIST_MAX_SLOTS); 8152a814908STejun Heo history <<= slots; 8162a814908STejun Heo if (wbc->wb_id != max_id) 8172a814908STejun Heo history |= (1U << slots) - 1; 8182a814908STejun Heo 8193a8e9ac8STejun Heo if (history) 8203a8e9ac8STejun Heo trace_inode_foreign_history(inode, wbc, history); 8213a8e9ac8STejun Heo 8222a814908STejun Heo /* 8232a814908STejun Heo * Switch if the current wb isn't the consistent winner. 8242a814908STejun Heo * If there are multiple closely competing dirtiers, the 8252a814908STejun Heo * inode may switch across them repeatedly over time, which 8262a814908STejun Heo * is okay. The main goal is avoiding keeping an inode on 8272a814908STejun Heo * the wrong wb for an extended period of time. 8282a814908STejun Heo */ 829682aa8e1STejun Heo if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) 830682aa8e1STejun Heo inode_switch_wbs(inode, max_id); 8312a814908STejun Heo } 8322a814908STejun Heo 8332a814908STejun Heo /* 8342a814908STejun Heo * Multiple instances of this function may race to update the 8352a814908STejun Heo * following fields but we don't mind occassional inaccuracies. 8362a814908STejun Heo */ 8372a814908STejun Heo inode->i_wb_frn_winner = max_id; 8382a814908STejun Heo inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); 8392a814908STejun Heo inode->i_wb_frn_history = history; 8402a814908STejun Heo 841b16b1debSTejun Heo wb_put(wbc->wb); 842b16b1debSTejun Heo wbc->wb = NULL; 843b16b1debSTejun Heo } 8449b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_detach_inode); 845b16b1debSTejun Heo 846b16b1debSTejun Heo /** 84734e51a5eSTejun Heo * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership 8482a814908STejun Heo * @wbc: writeback_control of the writeback in progress 8492a814908STejun Heo * @page: page being written out 8502a814908STejun Heo * @bytes: number of bytes being written out 8512a814908STejun Heo * 8522a814908STejun Heo * @bytes from @page are about to written out during the writeback 8532a814908STejun Heo * controlled by @wbc. Keep the book for foreign inode detection. See 8542a814908STejun Heo * wbc_detach_inode(). 8552a814908STejun Heo */ 85634e51a5eSTejun Heo void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page, 8572a814908STejun Heo size_t bytes) 8582a814908STejun Heo { 85966311422STejun Heo struct cgroup_subsys_state *css; 8602a814908STejun Heo int id; 8612a814908STejun Heo 8622a814908STejun Heo /* 8632a814908STejun Heo * pageout() path doesn't attach @wbc to the inode being written 8642a814908STejun Heo * out. This is intentional as we don't want the function to block 8652a814908STejun Heo * behind a slow cgroup. Ultimately, we want pageout() to kick off 8662a814908STejun Heo * regular writeback instead of writing things out itself. 8672a814908STejun Heo */ 86827b36d8fSTejun Heo if (!wbc->wb || wbc->no_cgroup_owner) 8692a814908STejun Heo return; 8702a814908STejun Heo 87166311422STejun Heo css = mem_cgroup_css_from_page(page); 87266311422STejun Heo /* dead cgroups shouldn't contribute to inode ownership arbitration */ 87366311422STejun Heo if (!(css->flags & CSS_ONLINE)) 87466311422STejun Heo return; 87566311422STejun Heo 87666311422STejun Heo id = css->id; 8772a814908STejun Heo 8782a814908STejun Heo if (id == wbc->wb_id) { 8792a814908STejun Heo wbc->wb_bytes += bytes; 8802a814908STejun Heo return; 8812a814908STejun Heo } 8822a814908STejun Heo 8832a814908STejun Heo if (id == wbc->wb_lcand_id) 8842a814908STejun Heo wbc->wb_lcand_bytes += bytes; 8852a814908STejun Heo 8862a814908STejun Heo /* Boyer-Moore majority vote algorithm */ 8872a814908STejun Heo if (!wbc->wb_tcand_bytes) 8882a814908STejun Heo wbc->wb_tcand_id = id; 8892a814908STejun Heo if (id == wbc->wb_tcand_id) 8902a814908STejun Heo wbc->wb_tcand_bytes += bytes; 8912a814908STejun Heo else 8922a814908STejun Heo wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); 8932a814908STejun Heo } 89434e51a5eSTejun Heo EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner); 8952a814908STejun Heo 8962a814908STejun Heo /** 897f2b65121STejun Heo * wb_split_bdi_pages - split nr_pages to write according to bandwidth 898f2b65121STejun Heo * @wb: target bdi_writeback to split @nr_pages to 899f2b65121STejun Heo * @nr_pages: number of pages to write for the whole bdi 900f2b65121STejun Heo * 901f2b65121STejun Heo * Split @wb's portion of @nr_pages according to @wb's write bandwidth in 902f2b65121STejun Heo * relation to the total write bandwidth of all wb's w/ dirty inodes on 903f2b65121STejun Heo * @wb->bdi. 904f2b65121STejun Heo */ 905f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 906f2b65121STejun Heo { 907f2b65121STejun Heo unsigned long this_bw = wb->avg_write_bandwidth; 908f2b65121STejun Heo unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); 909f2b65121STejun Heo 910f2b65121STejun Heo if (nr_pages == LONG_MAX) 911f2b65121STejun Heo return LONG_MAX; 912f2b65121STejun Heo 913f2b65121STejun Heo /* 914f2b65121STejun Heo * This may be called on clean wb's and proportional distribution 915f2b65121STejun Heo * may not make sense, just use the original @nr_pages in those 916f2b65121STejun Heo * cases. In general, we wanna err on the side of writing more. 917f2b65121STejun Heo */ 918f2b65121STejun Heo if (!tot_bw || this_bw >= tot_bw) 919f2b65121STejun Heo return nr_pages; 920f2b65121STejun Heo else 921f2b65121STejun Heo return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); 922f2b65121STejun Heo } 923f2b65121STejun Heo 924db125360STejun Heo /** 925db125360STejun Heo * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi 926db125360STejun Heo * @bdi: target backing_dev_info 927db125360STejun Heo * @base_work: wb_writeback_work to issue 928db125360STejun Heo * @skip_if_busy: skip wb's which already have writeback in progress 929db125360STejun Heo * 930db125360STejun Heo * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which 931db125360STejun Heo * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's 932db125360STejun Heo * distributed to the busy wbs according to each wb's proportion in the 933db125360STejun Heo * total active write bandwidth of @bdi. 934db125360STejun Heo */ 935db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 936db125360STejun Heo struct wb_writeback_work *base_work, 937db125360STejun Heo bool skip_if_busy) 938db125360STejun Heo { 939b817525aSTejun Heo struct bdi_writeback *last_wb = NULL; 940b33e18f6STejun Heo struct bdi_writeback *wb = list_entry(&bdi->wb_list, 941b817525aSTejun Heo struct bdi_writeback, bdi_node); 942db125360STejun Heo 943db125360STejun Heo might_sleep(); 944db125360STejun Heo restart: 945db125360STejun Heo rcu_read_lock(); 946b817525aSTejun Heo list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) { 9475b9cce4cSTejun Heo DEFINE_WB_COMPLETION(fallback_work_done, bdi); 9488a1270cdSTejun Heo struct wb_writeback_work fallback_work; 9498a1270cdSTejun Heo struct wb_writeback_work *work; 9508a1270cdSTejun Heo long nr_pages; 9518a1270cdSTejun Heo 952b817525aSTejun Heo if (last_wb) { 953b817525aSTejun Heo wb_put(last_wb); 954b817525aSTejun Heo last_wb = NULL; 955b817525aSTejun Heo } 956b817525aSTejun Heo 957006a0973STejun Heo /* SYNC_ALL writes out I_DIRTY_TIME too */ 958006a0973STejun Heo if (!wb_has_dirty_io(wb) && 959006a0973STejun Heo (base_work->sync_mode == WB_SYNC_NONE || 960006a0973STejun Heo list_empty(&wb->b_dirty_time))) 961006a0973STejun Heo continue; 962006a0973STejun Heo if (skip_if_busy && writeback_in_progress(wb)) 963db125360STejun Heo continue; 964db125360STejun Heo 9658a1270cdSTejun Heo nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages); 9668a1270cdSTejun Heo 9678a1270cdSTejun Heo work = kmalloc(sizeof(*work), GFP_ATOMIC); 9688a1270cdSTejun Heo if (work) { 9698a1270cdSTejun Heo *work = *base_work; 9708a1270cdSTejun Heo work->nr_pages = nr_pages; 9718a1270cdSTejun Heo work->auto_free = 1; 9728a1270cdSTejun Heo wb_queue_work(wb, work); 9738a1270cdSTejun Heo continue; 974db125360STejun Heo } 9758a1270cdSTejun Heo 9768a1270cdSTejun Heo /* alloc failed, execute synchronously using on-stack fallback */ 9778a1270cdSTejun Heo work = &fallback_work; 9788a1270cdSTejun Heo *work = *base_work; 9798a1270cdSTejun Heo work->nr_pages = nr_pages; 9808a1270cdSTejun Heo work->auto_free = 0; 9818a1270cdSTejun Heo work->done = &fallback_work_done; 9828a1270cdSTejun Heo 9838a1270cdSTejun Heo wb_queue_work(wb, work); 9848a1270cdSTejun Heo 985b817525aSTejun Heo /* 986b817525aSTejun Heo * Pin @wb so that it stays on @bdi->wb_list. This allows 987b817525aSTejun Heo * continuing iteration from @wb after dropping and 988b817525aSTejun Heo * regrabbing rcu read lock. 989b817525aSTejun Heo */ 990b817525aSTejun Heo wb_get(wb); 991b817525aSTejun Heo last_wb = wb; 992b817525aSTejun Heo 993db125360STejun Heo rcu_read_unlock(); 9945b9cce4cSTejun Heo wb_wait_for_completion(&fallback_work_done); 995db125360STejun Heo goto restart; 996db125360STejun Heo } 997db125360STejun Heo rcu_read_unlock(); 998b817525aSTejun Heo 999b817525aSTejun Heo if (last_wb) 1000b817525aSTejun Heo wb_put(last_wb); 1001db125360STejun Heo } 1002db125360STejun Heo 1003a1a0e23eSTejun Heo /** 1004d62241c7STejun Heo * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs 1005d62241c7STejun Heo * @bdi_id: target bdi id 1006d62241c7STejun Heo * @memcg_id: target memcg css id 1007d62241c7STejun Heo * @reason: reason why some writeback work initiated 1008d62241c7STejun Heo * @done: target wb_completion 1009d62241c7STejun Heo * 1010d62241c7STejun Heo * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id 1011d62241c7STejun Heo * with the specified parameters. 1012d62241c7STejun Heo */ 10137490a2d2SShakeel Butt int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, 1014d62241c7STejun Heo enum wb_reason reason, struct wb_completion *done) 1015d62241c7STejun Heo { 1016d62241c7STejun Heo struct backing_dev_info *bdi; 1017d62241c7STejun Heo struct cgroup_subsys_state *memcg_css; 1018d62241c7STejun Heo struct bdi_writeback *wb; 1019d62241c7STejun Heo struct wb_writeback_work *work; 10207490a2d2SShakeel Butt unsigned long dirty; 1021d62241c7STejun Heo int ret; 1022d62241c7STejun Heo 1023d62241c7STejun Heo /* lookup bdi and memcg */ 1024d62241c7STejun Heo bdi = bdi_get_by_id(bdi_id); 1025d62241c7STejun Heo if (!bdi) 1026d62241c7STejun Heo return -ENOENT; 1027d62241c7STejun Heo 1028d62241c7STejun Heo rcu_read_lock(); 1029d62241c7STejun Heo memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys); 1030d62241c7STejun Heo if (memcg_css && !css_tryget(memcg_css)) 1031d62241c7STejun Heo memcg_css = NULL; 1032d62241c7STejun Heo rcu_read_unlock(); 1033d62241c7STejun Heo if (!memcg_css) { 1034d62241c7STejun Heo ret = -ENOENT; 1035d62241c7STejun Heo goto out_bdi_put; 1036d62241c7STejun Heo } 1037d62241c7STejun Heo 1038d62241c7STejun Heo /* 1039d62241c7STejun Heo * And find the associated wb. If the wb isn't there already 1040d62241c7STejun Heo * there's nothing to flush, don't create one. 1041d62241c7STejun Heo */ 1042d62241c7STejun Heo wb = wb_get_lookup(bdi, memcg_css); 1043d62241c7STejun Heo if (!wb) { 1044d62241c7STejun Heo ret = -ENOENT; 1045d62241c7STejun Heo goto out_css_put; 1046d62241c7STejun Heo } 1047d62241c7STejun Heo 1048d62241c7STejun Heo /* 10497490a2d2SShakeel Butt * The caller is attempting to write out most of 1050d62241c7STejun Heo * the currently dirty pages. Let's take the current dirty page 1051d62241c7STejun Heo * count and inflate it by 25% which should be large enough to 1052d62241c7STejun Heo * flush out most dirty pages while avoiding getting livelocked by 1053d62241c7STejun Heo * concurrent dirtiers. 10547490a2d2SShakeel Butt * 10557490a2d2SShakeel Butt * BTW the memcg stats are flushed periodically and this is best-effort 10567490a2d2SShakeel Butt * estimation, so some potential error is ok. 1057d62241c7STejun Heo */ 10587490a2d2SShakeel Butt dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY); 10597490a2d2SShakeel Butt dirty = dirty * 10 / 8; 1060d62241c7STejun Heo 1061d62241c7STejun Heo /* issue the writeback work */ 1062d62241c7STejun Heo work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN); 1063d62241c7STejun Heo if (work) { 10647490a2d2SShakeel Butt work->nr_pages = dirty; 1065d62241c7STejun Heo work->sync_mode = WB_SYNC_NONE; 1066d62241c7STejun Heo work->range_cyclic = 1; 1067d62241c7STejun Heo work->reason = reason; 1068d62241c7STejun Heo work->done = done; 1069d62241c7STejun Heo work->auto_free = 1; 1070d62241c7STejun Heo wb_queue_work(wb, work); 1071d62241c7STejun Heo ret = 0; 1072d62241c7STejun Heo } else { 1073d62241c7STejun Heo ret = -ENOMEM; 1074d62241c7STejun Heo } 1075d62241c7STejun Heo 1076d62241c7STejun Heo wb_put(wb); 1077d62241c7STejun Heo out_css_put: 1078d62241c7STejun Heo css_put(memcg_css); 1079d62241c7STejun Heo out_bdi_put: 1080d62241c7STejun Heo bdi_put(bdi); 1081d62241c7STejun Heo return ret; 1082d62241c7STejun Heo } 1083d62241c7STejun Heo 1084d62241c7STejun Heo /** 1085a1a0e23eSTejun Heo * cgroup_writeback_umount - flush inode wb switches for umount 1086a1a0e23eSTejun Heo * 1087a1a0e23eSTejun Heo * This function is called when a super_block is about to be destroyed and 1088a1a0e23eSTejun Heo * flushes in-flight inode wb switches. An inode wb switch goes through 1089a1a0e23eSTejun Heo * RCU and then workqueue, so the two need to be flushed in order to ensure 1090a1a0e23eSTejun Heo * that all previously scheduled switches are finished. As wb switches are 1091a1a0e23eSTejun Heo * rare occurrences and synchronize_rcu() can take a while, perform 1092a1a0e23eSTejun Heo * flushing iff wb switches are in flight. 1093a1a0e23eSTejun Heo */ 1094a1a0e23eSTejun Heo void cgroup_writeback_umount(void) 1095a1a0e23eSTejun Heo { 1096592fa002SRoman Gushchin /* 1097592fa002SRoman Gushchin * SB_ACTIVE should be reliably cleared before checking 1098592fa002SRoman Gushchin * isw_nr_in_flight, see generic_shutdown_super(). 1099592fa002SRoman Gushchin */ 1100592fa002SRoman Gushchin smp_mb(); 1101592fa002SRoman Gushchin 1102a1a0e23eSTejun Heo if (atomic_read(&isw_nr_in_flight)) { 1103ec084de9SJiufei Xue /* 1104ec084de9SJiufei Xue * Use rcu_barrier() to wait for all pending callbacks to 1105ec084de9SJiufei Xue * ensure that all in-flight wb switches are in the workqueue. 1106ec084de9SJiufei Xue */ 1107ec084de9SJiufei Xue rcu_barrier(); 1108a1a0e23eSTejun Heo flush_workqueue(isw_wq); 1109a1a0e23eSTejun Heo } 1110a1a0e23eSTejun Heo } 1111a1a0e23eSTejun Heo 1112a1a0e23eSTejun Heo static int __init cgroup_writeback_init(void) 1113a1a0e23eSTejun Heo { 1114a1a0e23eSTejun Heo isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0); 1115a1a0e23eSTejun Heo if (!isw_wq) 1116a1a0e23eSTejun Heo return -ENOMEM; 1117a1a0e23eSTejun Heo return 0; 1118a1a0e23eSTejun Heo } 1119a1a0e23eSTejun Heo fs_initcall(cgroup_writeback_init); 1120a1a0e23eSTejun Heo 1121f2b65121STejun Heo #else /* CONFIG_CGROUP_WRITEBACK */ 1122f2b65121STejun Heo 11237fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 11247fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 11257fc5854fSTejun Heo 1126f3b6a6dfSRoman Gushchin static void inode_cgwb_move_to_attached(struct inode *inode, 1127f3b6a6dfSRoman Gushchin struct bdi_writeback *wb) 1128f3b6a6dfSRoman Gushchin { 1129f3b6a6dfSRoman Gushchin assert_spin_locked(&wb->list_lock); 1130f3b6a6dfSRoman Gushchin assert_spin_locked(&inode->i_lock); 1131f3b6a6dfSRoman Gushchin 1132f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 1133f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 1134f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 1135f3b6a6dfSRoman Gushchin } 1136f3b6a6dfSRoman Gushchin 113787e1d789STejun Heo static struct bdi_writeback * 113887e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 113987e1d789STejun Heo __releases(&inode->i_lock) 114087e1d789STejun Heo __acquires(&wb->list_lock) 114187e1d789STejun Heo { 114287e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 114387e1d789STejun Heo 114487e1d789STejun Heo spin_unlock(&inode->i_lock); 114587e1d789STejun Heo spin_lock(&wb->list_lock); 114687e1d789STejun Heo return wb; 114787e1d789STejun Heo } 114887e1d789STejun Heo 114987e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 115087e1d789STejun Heo __acquires(&wb->list_lock) 115187e1d789STejun Heo { 115287e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 115387e1d789STejun Heo 115487e1d789STejun Heo spin_lock(&wb->list_lock); 115587e1d789STejun Heo return wb; 115687e1d789STejun Heo } 115787e1d789STejun Heo 1158f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 1159f2b65121STejun Heo { 1160f2b65121STejun Heo return nr_pages; 1161f2b65121STejun Heo } 1162f2b65121STejun Heo 1163db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 1164db125360STejun Heo struct wb_writeback_work *base_work, 1165db125360STejun Heo bool skip_if_busy) 1166db125360STejun Heo { 1167db125360STejun Heo might_sleep(); 1168db125360STejun Heo 1169006a0973STejun Heo if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) { 1170db125360STejun Heo base_work->auto_free = 0; 1171db125360STejun Heo wb_queue_work(&bdi->wb, base_work); 1172db125360STejun Heo } 1173db125360STejun Heo } 1174db125360STejun Heo 1175703c2708STejun Heo #endif /* CONFIG_CGROUP_WRITEBACK */ 1176703c2708STejun Heo 1177e8e8a0c6SJens Axboe /* 1178e8e8a0c6SJens Axboe * Add in the number of potentially dirty inodes, because each inode 1179e8e8a0c6SJens Axboe * write can dirty pagecache in the underlying blockdev. 1180e8e8a0c6SJens Axboe */ 1181e8e8a0c6SJens Axboe static unsigned long get_nr_dirty_pages(void) 1182e8e8a0c6SJens Axboe { 1183e8e8a0c6SJens Axboe return global_node_page_state(NR_FILE_DIRTY) + 1184e8e8a0c6SJens Axboe get_nr_dirty_inodes(); 1185e8e8a0c6SJens Axboe } 1186e8e8a0c6SJens Axboe 1187e8e8a0c6SJens Axboe static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason) 1188b6e51316SJens Axboe { 1189c00ddad3STejun Heo if (!wb_has_dirty_io(wb)) 1190c00ddad3STejun Heo return; 1191c00ddad3STejun Heo 1192c00ddad3STejun Heo /* 1193aac8d41cSJens Axboe * All callers of this function want to start writeback of all 1194aac8d41cSJens Axboe * dirty pages. Places like vmscan can call this at a very 1195aac8d41cSJens Axboe * high frequency, causing pointless allocations of tons of 1196aac8d41cSJens Axboe * work items and keeping the flusher threads busy retrieving 1197aac8d41cSJens Axboe * that work. Ensure that we only allow one of them pending and 119885009b4fSJens Axboe * inflight at the time. 1199aac8d41cSJens Axboe */ 120085009b4fSJens Axboe if (test_bit(WB_start_all, &wb->state) || 120185009b4fSJens Axboe test_and_set_bit(WB_start_all, &wb->state)) 1202aac8d41cSJens Axboe return; 1203aac8d41cSJens Axboe 120485009b4fSJens Axboe wb->start_all_reason = reason; 1205c00ddad3STejun Heo wb_wakeup(wb); 1206d3ddec76SWu Fengguang } 1207d3ddec76SWu Fengguang 1208c5444198SChristoph Hellwig /** 12099ecf4866STejun Heo * wb_start_background_writeback - start background writeback 12109ecf4866STejun Heo * @wb: bdi_writback to write from 1211c5444198SChristoph Hellwig * 1212c5444198SChristoph Hellwig * Description: 12136585027aSJan Kara * This makes sure WB_SYNC_NONE background writeback happens. When 12149ecf4866STejun Heo * this function returns, it is only guaranteed that for given wb 12156585027aSJan Kara * some IO is happening if we are over background dirty threshold. 12166585027aSJan Kara * Caller need not hold sb s_umount semaphore. 1217c5444198SChristoph Hellwig */ 12189ecf4866STejun Heo void wb_start_background_writeback(struct bdi_writeback *wb) 1219c5444198SChristoph Hellwig { 12206585027aSJan Kara /* 12216585027aSJan Kara * We just wake up the flusher thread. It will perform background 12226585027aSJan Kara * writeback as soon as there is no other work to do. 12236585027aSJan Kara */ 12245634cc2aSTejun Heo trace_writeback_wake_background(wb); 12259ecf4866STejun Heo wb_wakeup(wb); 12261da177e4SLinus Torvalds } 12271da177e4SLinus Torvalds 12281da177e4SLinus Torvalds /* 1229a66979abSDave Chinner * Remove the inode from the writeback list it is on. 1230a66979abSDave Chinner */ 1231c7f54084SDave Chinner void inode_io_list_del(struct inode *inode) 1232a66979abSDave Chinner { 123387e1d789STejun Heo struct bdi_writeback *wb; 1234a66979abSDave Chinner 123587e1d789STejun Heo wb = inode_to_wb_and_lock_list(inode); 1236b35250c0SJan Kara spin_lock(&inode->i_lock); 1237f3b6a6dfSRoman Gushchin 1238f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 1239f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 1240f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 1241f3b6a6dfSRoman Gushchin 1242b35250c0SJan Kara spin_unlock(&inode->i_lock); 124352ebea74STejun Heo spin_unlock(&wb->list_lock); 1244f758eeabSChristoph Hellwig } 12454301efa4SJan Kara EXPORT_SYMBOL(inode_io_list_del); 1246a66979abSDave Chinner 1247a66979abSDave Chinner /* 12486c60d2b5SDave Chinner * mark an inode as under writeback on the sb 12496c60d2b5SDave Chinner */ 12506c60d2b5SDave Chinner void sb_mark_inode_writeback(struct inode *inode) 12516c60d2b5SDave Chinner { 12526c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 12536c60d2b5SDave Chinner unsigned long flags; 12546c60d2b5SDave Chinner 12556c60d2b5SDave Chinner if (list_empty(&inode->i_wb_list)) { 12566c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 12579a46b04fSBrian Foster if (list_empty(&inode->i_wb_list)) { 12586c60d2b5SDave Chinner list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); 12599a46b04fSBrian Foster trace_sb_mark_inode_writeback(inode); 12609a46b04fSBrian Foster } 12616c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 12626c60d2b5SDave Chinner } 12636c60d2b5SDave Chinner } 12646c60d2b5SDave Chinner 12656c60d2b5SDave Chinner /* 12666c60d2b5SDave Chinner * clear an inode as under writeback on the sb 12676c60d2b5SDave Chinner */ 12686c60d2b5SDave Chinner void sb_clear_inode_writeback(struct inode *inode) 12696c60d2b5SDave Chinner { 12706c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 12716c60d2b5SDave Chinner unsigned long flags; 12726c60d2b5SDave Chinner 12736c60d2b5SDave Chinner if (!list_empty(&inode->i_wb_list)) { 12746c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 12759a46b04fSBrian Foster if (!list_empty(&inode->i_wb_list)) { 12766c60d2b5SDave Chinner list_del_init(&inode->i_wb_list); 12779a46b04fSBrian Foster trace_sb_clear_inode_writeback(inode); 12789a46b04fSBrian Foster } 12796c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 12806c60d2b5SDave Chinner } 12816c60d2b5SDave Chinner } 12826c60d2b5SDave Chinner 12836c60d2b5SDave Chinner /* 12846610a0bcSAndrew Morton * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 12856610a0bcSAndrew Morton * furthest end of its superblock's dirty-inode list. 12866610a0bcSAndrew Morton * 12876610a0bcSAndrew Morton * Before stamping the inode's ->dirtied_when, we check to see whether it is 128866f3b8e2SJens Axboe * already the most-recently-dirtied inode on the b_dirty list. If that is 12896610a0bcSAndrew Morton * the case then the inode must have been redirtied while it was being written 12906610a0bcSAndrew Morton * out and we don't reset its dirtied_when. 12916610a0bcSAndrew Morton */ 1292b35250c0SJan Kara static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb) 12936610a0bcSAndrew Morton { 1294b35250c0SJan Kara assert_spin_locked(&inode->i_lock); 1295b35250c0SJan Kara 129603ba3782SJens Axboe if (!list_empty(&wb->b_dirty)) { 129766f3b8e2SJens Axboe struct inode *tail; 12986610a0bcSAndrew Morton 12997ccf19a8SNick Piggin tail = wb_inode(wb->b_dirty.next); 130066f3b8e2SJens Axboe if (time_before(inode->dirtied_when, tail->dirtied_when)) 13016610a0bcSAndrew Morton inode->dirtied_when = jiffies; 13026610a0bcSAndrew Morton } 1303c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty); 13045afced3bSJan Kara inode->i_state &= ~I_SYNC_QUEUED; 13056610a0bcSAndrew Morton } 13066610a0bcSAndrew Morton 1307b35250c0SJan Kara static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) 1308b35250c0SJan Kara { 1309b35250c0SJan Kara spin_lock(&inode->i_lock); 1310b35250c0SJan Kara redirty_tail_locked(inode, wb); 1311b35250c0SJan Kara spin_unlock(&inode->i_lock); 1312b35250c0SJan Kara } 1313b35250c0SJan Kara 13146610a0bcSAndrew Morton /* 131566f3b8e2SJens Axboe * requeue inode for re-scanning after bdi->b_io list is exhausted. 1316c986d1e2SAndrew Morton */ 1317f758eeabSChristoph Hellwig static void requeue_io(struct inode *inode, struct bdi_writeback *wb) 1318c986d1e2SAndrew Morton { 1319c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_more_io); 1320c986d1e2SAndrew Morton } 1321c986d1e2SAndrew Morton 13221c0eeaf5SJoern Engel static void inode_sync_complete(struct inode *inode) 13231c0eeaf5SJoern Engel { 1324365b94aeSJan Kara inode->i_state &= ~I_SYNC; 13254eff96ddSJan Kara /* If inode is clean an unused, put it into LRU now... */ 13264eff96ddSJan Kara inode_add_lru(inode); 1327365b94aeSJan Kara /* Waiters must see I_SYNC cleared before being woken up */ 13281c0eeaf5SJoern Engel smp_mb(); 13291c0eeaf5SJoern Engel wake_up_bit(&inode->i_state, __I_SYNC); 13301c0eeaf5SJoern Engel } 13311c0eeaf5SJoern Engel 1332d2caa3c5SJeff Layton static bool inode_dirtied_after(struct inode *inode, unsigned long t) 1333d2caa3c5SJeff Layton { 1334d2caa3c5SJeff Layton bool ret = time_after(inode->dirtied_when, t); 1335d2caa3c5SJeff Layton #ifndef CONFIG_64BIT 1336d2caa3c5SJeff Layton /* 1337d2caa3c5SJeff Layton * For inodes being constantly redirtied, dirtied_when can get stuck. 1338d2caa3c5SJeff Layton * It _appears_ to be in the future, but is actually in distant past. 1339d2caa3c5SJeff Layton * This test is necessary to prevent such wrapped-around relative times 13405b0830cbSJens Axboe * from permanently stopping the whole bdi writeback. 1341d2caa3c5SJeff Layton */ 1342d2caa3c5SJeff Layton ret = ret && time_before_eq(inode->dirtied_when, jiffies); 1343d2caa3c5SJeff Layton #endif 1344d2caa3c5SJeff Layton return ret; 1345d2caa3c5SJeff Layton } 1346d2caa3c5SJeff Layton 13470ae45f63STheodore Ts'o #define EXPIRE_DIRTY_ATIME 0x0001 13480ae45f63STheodore Ts'o 1349c986d1e2SAndrew Morton /* 1350f9cae926SJan Kara * Move expired (dirtied before dirtied_before) dirty inodes from 1351697e6fedSJan Kara * @delaying_queue to @dispatch_queue. 13522c136579SFengguang Wu */ 1353e84d0a4fSWu Fengguang static int move_expired_inodes(struct list_head *delaying_queue, 13542c136579SFengguang Wu struct list_head *dispatch_queue, 13555fcd5750SJan Kara unsigned long dirtied_before) 13562c136579SFengguang Wu { 13575c03449dSShaohua Li LIST_HEAD(tmp); 13585c03449dSShaohua Li struct list_head *pos, *node; 1359cf137307SJens Axboe struct super_block *sb = NULL; 13605c03449dSShaohua Li struct inode *inode; 1361cf137307SJens Axboe int do_sb_sort = 0; 1362e84d0a4fSWu Fengguang int moved = 0; 13635c03449dSShaohua Li 13642c136579SFengguang Wu while (!list_empty(delaying_queue)) { 13657ccf19a8SNick Piggin inode = wb_inode(delaying_queue->prev); 1366f9cae926SJan Kara if (inode_dirtied_after(inode, dirtied_before)) 13672c136579SFengguang Wu break; 1368c7f54084SDave Chinner list_move(&inode->i_io_list, &tmp); 1369a8855990SJan Kara moved++; 13705afced3bSJan Kara spin_lock(&inode->i_lock); 13715afced3bSJan Kara inode->i_state |= I_SYNC_QUEUED; 13725afced3bSJan Kara spin_unlock(&inode->i_lock); 1373a8855990SJan Kara if (sb_is_blkdev_sb(inode->i_sb)) 1374a8855990SJan Kara continue; 1375cf137307SJens Axboe if (sb && sb != inode->i_sb) 1376cf137307SJens Axboe do_sb_sort = 1; 1377cf137307SJens Axboe sb = inode->i_sb; 13785c03449dSShaohua Li } 13795c03449dSShaohua Li 1380cf137307SJens Axboe /* just one sb in list, splice to dispatch_queue and we're done */ 1381cf137307SJens Axboe if (!do_sb_sort) { 1382cf137307SJens Axboe list_splice(&tmp, dispatch_queue); 1383e84d0a4fSWu Fengguang goto out; 1384cf137307SJens Axboe } 1385cf137307SJens Axboe 13865c03449dSShaohua Li /* Move inodes from one superblock together */ 13875c03449dSShaohua Li while (!list_empty(&tmp)) { 13887ccf19a8SNick Piggin sb = wb_inode(tmp.prev)->i_sb; 13895c03449dSShaohua Li list_for_each_prev_safe(pos, node, &tmp) { 13907ccf19a8SNick Piggin inode = wb_inode(pos); 13915c03449dSShaohua Li if (inode->i_sb == sb) 1392c7f54084SDave Chinner list_move(&inode->i_io_list, dispatch_queue); 13932c136579SFengguang Wu } 13942c136579SFengguang Wu } 1395e84d0a4fSWu Fengguang out: 1396e84d0a4fSWu Fengguang return moved; 13975c03449dSShaohua Li } 13982c136579SFengguang Wu 13992c136579SFengguang Wu /* 14002c136579SFengguang Wu * Queue all expired dirty inodes for io, eldest first. 14014ea879b9SWu Fengguang * Before 14024ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 14034ea879b9SWu Fengguang * =============> gf edc BA 14044ea879b9SWu Fengguang * After 14054ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 14064ea879b9SWu Fengguang * =============> g fBAedc 14074ea879b9SWu Fengguang * | 14084ea879b9SWu Fengguang * +--> dequeue for IO 14092c136579SFengguang Wu */ 1410f9cae926SJan Kara static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work, 1411f9cae926SJan Kara unsigned long dirtied_before) 14122c136579SFengguang Wu { 1413e84d0a4fSWu Fengguang int moved; 1414f9cae926SJan Kara unsigned long time_expire_jif = dirtied_before; 14150ae45f63STheodore Ts'o 1416f758eeabSChristoph Hellwig assert_spin_locked(&wb->list_lock); 14174ea879b9SWu Fengguang list_splice_init(&wb->b_more_io, &wb->b_io); 14185fcd5750SJan Kara moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before); 1419f9cae926SJan Kara if (!work->for_sync) 1420f9cae926SJan Kara time_expire_jif = jiffies - dirtytime_expire_interval * HZ; 14210ae45f63STheodore Ts'o moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, 14225fcd5750SJan Kara time_expire_jif); 1423d6c10f1fSTejun Heo if (moved) 1424d6c10f1fSTejun Heo wb_io_lists_populated(wb); 1425f9cae926SJan Kara trace_writeback_queue_io(wb, work, dirtied_before, moved); 142666f3b8e2SJens Axboe } 142766f3b8e2SJens Axboe 1428a9185b41SChristoph Hellwig static int write_inode(struct inode *inode, struct writeback_control *wbc) 142966f3b8e2SJens Axboe { 14309fb0a7daSTejun Heo int ret; 14319fb0a7daSTejun Heo 14329fb0a7daSTejun Heo if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { 14339fb0a7daSTejun Heo trace_writeback_write_inode_start(inode, wbc); 14349fb0a7daSTejun Heo ret = inode->i_sb->s_op->write_inode(inode, wbc); 14359fb0a7daSTejun Heo trace_writeback_write_inode(inode, wbc); 14369fb0a7daSTejun Heo return ret; 14379fb0a7daSTejun Heo } 143803ba3782SJens Axboe return 0; 143966f3b8e2SJens Axboe } 144008d8e974SFengguang Wu 14412c136579SFengguang Wu /* 1442169ebd90SJan Kara * Wait for writeback on an inode to complete. Called with i_lock held. 1443169ebd90SJan Kara * Caller must make sure inode cannot go away when we drop i_lock. 144401c03194SChristoph Hellwig */ 1445169ebd90SJan Kara static void __inode_wait_for_writeback(struct inode *inode) 1446169ebd90SJan Kara __releases(inode->i_lock) 1447169ebd90SJan Kara __acquires(inode->i_lock) 144801c03194SChristoph Hellwig { 144901c03194SChristoph Hellwig DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); 145001c03194SChristoph Hellwig wait_queue_head_t *wqh; 145101c03194SChristoph Hellwig 145201c03194SChristoph Hellwig wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 145358a9d3d8SRichard Kennedy while (inode->i_state & I_SYNC) { 1454250df6edSDave Chinner spin_unlock(&inode->i_lock); 145574316201SNeilBrown __wait_on_bit(wqh, &wq, bit_wait, 145674316201SNeilBrown TASK_UNINTERRUPTIBLE); 1457250df6edSDave Chinner spin_lock(&inode->i_lock); 145858a9d3d8SRichard Kennedy } 145901c03194SChristoph Hellwig } 146001c03194SChristoph Hellwig 146101c03194SChristoph Hellwig /* 1462169ebd90SJan Kara * Wait for writeback on an inode to complete. Caller must have inode pinned. 1463169ebd90SJan Kara */ 1464169ebd90SJan Kara void inode_wait_for_writeback(struct inode *inode) 1465169ebd90SJan Kara { 1466169ebd90SJan Kara spin_lock(&inode->i_lock); 1467169ebd90SJan Kara __inode_wait_for_writeback(inode); 1468169ebd90SJan Kara spin_unlock(&inode->i_lock); 1469169ebd90SJan Kara } 1470169ebd90SJan Kara 1471169ebd90SJan Kara /* 1472169ebd90SJan Kara * Sleep until I_SYNC is cleared. This function must be called with i_lock 1473169ebd90SJan Kara * held and drops it. It is aimed for callers not holding any inode reference 1474169ebd90SJan Kara * so once i_lock is dropped, inode can go away. 1475169ebd90SJan Kara */ 1476169ebd90SJan Kara static void inode_sleep_on_writeback(struct inode *inode) 1477169ebd90SJan Kara __releases(inode->i_lock) 1478169ebd90SJan Kara { 1479169ebd90SJan Kara DEFINE_WAIT(wait); 1480169ebd90SJan Kara wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1481169ebd90SJan Kara int sleep; 1482169ebd90SJan Kara 1483169ebd90SJan Kara prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1484169ebd90SJan Kara sleep = inode->i_state & I_SYNC; 1485169ebd90SJan Kara spin_unlock(&inode->i_lock); 1486169ebd90SJan Kara if (sleep) 1487169ebd90SJan Kara schedule(); 1488169ebd90SJan Kara finish_wait(wqh, &wait); 1489169ebd90SJan Kara } 1490169ebd90SJan Kara 1491169ebd90SJan Kara /* 1492ccb26b5aSJan Kara * Find proper writeback list for the inode depending on its current state and 1493ccb26b5aSJan Kara * possibly also change of its state while we were doing writeback. Here we 1494ccb26b5aSJan Kara * handle things such as livelock prevention or fairness of writeback among 1495ccb26b5aSJan Kara * inodes. This function can be called only by flusher thread - noone else 1496ccb26b5aSJan Kara * processes all inodes in writeback lists and requeueing inodes behind flusher 1497ccb26b5aSJan Kara * thread's back can have unexpected consequences. 1498ccb26b5aSJan Kara */ 1499ccb26b5aSJan Kara static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, 1500ccb26b5aSJan Kara struct writeback_control *wbc) 1501ccb26b5aSJan Kara { 1502ccb26b5aSJan Kara if (inode->i_state & I_FREEING) 1503ccb26b5aSJan Kara return; 1504ccb26b5aSJan Kara 1505ccb26b5aSJan Kara /* 1506ccb26b5aSJan Kara * Sync livelock prevention. Each inode is tagged and synced in one 1507ccb26b5aSJan Kara * shot. If still dirty, it will be redirty_tail()'ed below. Update 1508ccb26b5aSJan Kara * the dirty time to prevent enqueue and sync it again. 1509ccb26b5aSJan Kara */ 1510ccb26b5aSJan Kara if ((inode->i_state & I_DIRTY) && 1511ccb26b5aSJan Kara (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) 1512ccb26b5aSJan Kara inode->dirtied_when = jiffies; 1513ccb26b5aSJan Kara 15144f8ad655SJan Kara if (wbc->pages_skipped) { 15154f8ad655SJan Kara /* 15164f8ad655SJan Kara * writeback is not making progress due to locked 15174f8ad655SJan Kara * buffers. Skip this inode for now. 15184f8ad655SJan Kara */ 1519b35250c0SJan Kara redirty_tail_locked(inode, wb); 15204f8ad655SJan Kara return; 15214f8ad655SJan Kara } 15224f8ad655SJan Kara 1523ccb26b5aSJan Kara if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { 1524ccb26b5aSJan Kara /* 1525ccb26b5aSJan Kara * We didn't write back all the pages. nfs_writepages() 1526ccb26b5aSJan Kara * sometimes bales out without doing anything. 1527ccb26b5aSJan Kara */ 1528ccb26b5aSJan Kara if (wbc->nr_to_write <= 0) { 1529ccb26b5aSJan Kara /* Slice used up. Queue for next turn. */ 1530ccb26b5aSJan Kara requeue_io(inode, wb); 1531ccb26b5aSJan Kara } else { 1532ccb26b5aSJan Kara /* 1533ccb26b5aSJan Kara * Writeback blocked by something other than 1534ccb26b5aSJan Kara * congestion. Delay the inode for some time to 1535ccb26b5aSJan Kara * avoid spinning on the CPU (100% iowait) 1536ccb26b5aSJan Kara * retrying writeback of the dirty page/inode 1537ccb26b5aSJan Kara * that cannot be performed immediately. 1538ccb26b5aSJan Kara */ 1539b35250c0SJan Kara redirty_tail_locked(inode, wb); 1540ccb26b5aSJan Kara } 1541ccb26b5aSJan Kara } else if (inode->i_state & I_DIRTY) { 1542ccb26b5aSJan Kara /* 1543ccb26b5aSJan Kara * Filesystems can dirty the inode during writeback operations, 1544ccb26b5aSJan Kara * such as delayed allocation during submission or metadata 1545ccb26b5aSJan Kara * updates after data IO completion. 1546ccb26b5aSJan Kara */ 1547b35250c0SJan Kara redirty_tail_locked(inode, wb); 15480ae45f63STheodore Ts'o } else if (inode->i_state & I_DIRTY_TIME) { 1549a2f48706STheodore Ts'o inode->dirtied_when = jiffies; 1550c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty_time); 15515afced3bSJan Kara inode->i_state &= ~I_SYNC_QUEUED; 1552ccb26b5aSJan Kara } else { 1553ccb26b5aSJan Kara /* The inode is clean. Remove from writeback lists. */ 1554f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, wb); 1555ccb26b5aSJan Kara } 1556ccb26b5aSJan Kara } 1557ccb26b5aSJan Kara 1558ccb26b5aSJan Kara /* 1559da0c4c60SEric Biggers * Write out an inode and its dirty pages (or some of its dirty pages, depending 1560da0c4c60SEric Biggers * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state. 1561da0c4c60SEric Biggers * 1562da0c4c60SEric Biggers * This doesn't remove the inode from the writeback list it is on, except 1563da0c4c60SEric Biggers * potentially to move it from b_dirty_time to b_dirty due to timestamp 1564da0c4c60SEric Biggers * expiration. The caller is otherwise responsible for writeback list handling. 1565da0c4c60SEric Biggers * 1566da0c4c60SEric Biggers * The caller is also responsible for setting the I_SYNC flag beforehand and 1567da0c4c60SEric Biggers * calling inode_sync_complete() to clear it afterwards. 15681da177e4SLinus Torvalds */ 15691da177e4SLinus Torvalds static int 1570cd8ed2a4SYan Hong __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 15711da177e4SLinus Torvalds { 15721da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 1573251d6a47SWu Fengguang long nr_to_write = wbc->nr_to_write; 157401c03194SChristoph Hellwig unsigned dirty; 15751da177e4SLinus Torvalds int ret; 15761da177e4SLinus Torvalds 15774f8ad655SJan Kara WARN_ON(!(inode->i_state & I_SYNC)); 15781da177e4SLinus Torvalds 15799fb0a7daSTejun Heo trace_writeback_single_inode_start(inode, wbc, nr_to_write); 15809fb0a7daSTejun Heo 15811da177e4SLinus Torvalds ret = do_writepages(mapping, wbc); 15821da177e4SLinus Torvalds 158326821ed4SChristoph Hellwig /* 158426821ed4SChristoph Hellwig * Make sure to wait on the data before writing out the metadata. 158526821ed4SChristoph Hellwig * This is important for filesystems that modify metadata on data 15867747bd4bSDave Chinner * I/O completion. We don't do it for sync(2) writeback because it has a 15877747bd4bSDave Chinner * separate, external IO completion path and ->sync_fs for guaranteeing 15887747bd4bSDave Chinner * inode metadata is written back correctly. 158926821ed4SChristoph Hellwig */ 15907747bd4bSDave Chinner if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { 159126821ed4SChristoph Hellwig int err = filemap_fdatawait(mapping); 15921da177e4SLinus Torvalds if (ret == 0) 15931da177e4SLinus Torvalds ret = err; 15941da177e4SLinus Torvalds } 15951da177e4SLinus Torvalds 15965547e8aaSDmitry Monakhov /* 15971e249cb5SEric Biggers * If the inode has dirty timestamps and we need to write them, call 15981e249cb5SEric Biggers * mark_inode_dirty_sync() to notify the filesystem about it and to 15991e249cb5SEric Biggers * change I_DIRTY_TIME into I_DIRTY_SYNC. 16001e249cb5SEric Biggers */ 16011e249cb5SEric Biggers if ((inode->i_state & I_DIRTY_TIME) && 160283dc881dSEric Biggers (wbc->sync_mode == WB_SYNC_ALL || 16031e249cb5SEric Biggers time_after(jiffies, inode->dirtied_time_when + 16041e249cb5SEric Biggers dirtytime_expire_interval * HZ))) { 16051e249cb5SEric Biggers trace_writeback_lazytime(inode); 16061e249cb5SEric Biggers mark_inode_dirty_sync(inode); 16071e249cb5SEric Biggers } 16081e249cb5SEric Biggers 16091e249cb5SEric Biggers /* 1610da0c4c60SEric Biggers * Get and clear the dirty flags from i_state. This needs to be done 1611da0c4c60SEric Biggers * after calling writepages because some filesystems may redirty the 1612da0c4c60SEric Biggers * inode during writepages due to delalloc. It also needs to be done 1613da0c4c60SEric Biggers * after handling timestamp expiration, as that may dirty the inode too. 16145547e8aaSDmitry Monakhov */ 1615250df6edSDave Chinner spin_lock(&inode->i_lock); 16165547e8aaSDmitry Monakhov dirty = inode->i_state & I_DIRTY; 16170ae45f63STheodore Ts'o inode->i_state &= ~dirty; 16189c6ac78eSTejun Heo 16199c6ac78eSTejun Heo /* 16209c6ac78eSTejun Heo * Paired with smp_mb() in __mark_inode_dirty(). This allows 16219c6ac78eSTejun Heo * __mark_inode_dirty() to test i_state without grabbing i_lock - 16229c6ac78eSTejun Heo * either they see the I_DIRTY bits cleared or we see the dirtied 16239c6ac78eSTejun Heo * inode. 16249c6ac78eSTejun Heo * 16259c6ac78eSTejun Heo * I_DIRTY_PAGES is always cleared together above even if @mapping 16269c6ac78eSTejun Heo * still has dirty pages. The flag is reinstated after smp_mb() if 16279c6ac78eSTejun Heo * necessary. This guarantees that either __mark_inode_dirty() 16289c6ac78eSTejun Heo * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. 16299c6ac78eSTejun Heo */ 16309c6ac78eSTejun Heo smp_mb(); 16319c6ac78eSTejun Heo 16329c6ac78eSTejun Heo if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 16339c6ac78eSTejun Heo inode->i_state |= I_DIRTY_PAGES; 163408276bdaSDavid Howells else if (unlikely(inode->i_state & I_PINNING_FSCACHE_WB)) { 163508276bdaSDavid Howells if (!(inode->i_state & I_DIRTY_PAGES)) { 163608276bdaSDavid Howells inode->i_state &= ~I_PINNING_FSCACHE_WB; 163708276bdaSDavid Howells wbc->unpinned_fscache_wb = true; 163808276bdaSDavid Howells dirty |= I_PINNING_FSCACHE_WB; /* Cause write_inode */ 163908276bdaSDavid Howells } 164008276bdaSDavid Howells } 16419c6ac78eSTejun Heo 1642250df6edSDave Chinner spin_unlock(&inode->i_lock); 16439c6ac78eSTejun Heo 164426821ed4SChristoph Hellwig /* Don't write the inode if only I_DIRTY_PAGES was set */ 16450ae45f63STheodore Ts'o if (dirty & ~I_DIRTY_PAGES) { 1646a9185b41SChristoph Hellwig int err = write_inode(inode, wbc); 16471da177e4SLinus Torvalds if (ret == 0) 16481da177e4SLinus Torvalds ret = err; 16491da177e4SLinus Torvalds } 165008276bdaSDavid Howells wbc->unpinned_fscache_wb = false; 16514f8ad655SJan Kara trace_writeback_single_inode(inode, wbc, nr_to_write); 16524f8ad655SJan Kara return ret; 16534f8ad655SJan Kara } 16544f8ad655SJan Kara 16554f8ad655SJan Kara /* 1656da0c4c60SEric Biggers * Write out an inode's dirty data and metadata on-demand, i.e. separately from 1657da0c4c60SEric Biggers * the regular batched writeback done by the flusher threads in 1658da0c4c60SEric Biggers * writeback_sb_inodes(). @wbc controls various aspects of the write, such as 1659da0c4c60SEric Biggers * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE). 16604f8ad655SJan Kara * 1661da0c4c60SEric Biggers * To prevent the inode from going away, either the caller must have a reference 1662da0c4c60SEric Biggers * to the inode, or the inode must have I_WILL_FREE or I_FREEING set. 16634f8ad655SJan Kara */ 1664aaf25593STejun Heo static int writeback_single_inode(struct inode *inode, 16654f8ad655SJan Kara struct writeback_control *wbc) 16664f8ad655SJan Kara { 1667aaf25593STejun Heo struct bdi_writeback *wb; 16684f8ad655SJan Kara int ret = 0; 16694f8ad655SJan Kara 16704f8ad655SJan Kara spin_lock(&inode->i_lock); 16714f8ad655SJan Kara if (!atomic_read(&inode->i_count)) 16724f8ad655SJan Kara WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 16734f8ad655SJan Kara else 16744f8ad655SJan Kara WARN_ON(inode->i_state & I_WILL_FREE); 16754f8ad655SJan Kara 16764f8ad655SJan Kara if (inode->i_state & I_SYNC) { 1677da0c4c60SEric Biggers /* 1678da0c4c60SEric Biggers * Writeback is already running on the inode. For WB_SYNC_NONE, 1679da0c4c60SEric Biggers * that's enough and we can just return. For WB_SYNC_ALL, we 1680da0c4c60SEric Biggers * must wait for the existing writeback to complete, then do 1681da0c4c60SEric Biggers * writeback again if there's anything left. 1682da0c4c60SEric Biggers */ 16834f8ad655SJan Kara if (wbc->sync_mode != WB_SYNC_ALL) 16844f8ad655SJan Kara goto out; 1685169ebd90SJan Kara __inode_wait_for_writeback(inode); 16864f8ad655SJan Kara } 16874f8ad655SJan Kara WARN_ON(inode->i_state & I_SYNC); 16884f8ad655SJan Kara /* 1689da0c4c60SEric Biggers * If the inode is already fully clean, then there's nothing to do. 1690da0c4c60SEric Biggers * 1691da0c4c60SEric Biggers * For data-integrity syncs we also need to check whether any pages are 1692da0c4c60SEric Biggers * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If 1693da0c4c60SEric Biggers * there are any such pages, we'll need to wait for them. 16944f8ad655SJan Kara */ 16950ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL) && 1696f9b0e058SJan Kara (wbc->sync_mode != WB_SYNC_ALL || 1697f9b0e058SJan Kara !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) 16984f8ad655SJan Kara goto out; 16994f8ad655SJan Kara inode->i_state |= I_SYNC; 1700b16b1debSTejun Heo wbc_attach_and_unlock_inode(wbc, inode); 17014f8ad655SJan Kara 1702cd8ed2a4SYan Hong ret = __writeback_single_inode(inode, wbc); 17031da177e4SLinus Torvalds 1704b16b1debSTejun Heo wbc_detach_inode(wbc); 1705aaf25593STejun Heo 1706aaf25593STejun Heo wb = inode_to_wb_and_lock_list(inode); 1707250df6edSDave Chinner spin_lock(&inode->i_lock); 17084f8ad655SJan Kara /* 1709da0c4c60SEric Biggers * If the inode is now fully clean, then it can be safely removed from 1710da0c4c60SEric Biggers * its writeback list (if any). Otherwise the flusher threads are 1711da0c4c60SEric Biggers * responsible for the writeback lists. 17124f8ad655SJan Kara */ 17130ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1714f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, wb); 1715*846a3351SJing Xia else if (!(inode->i_state & I_SYNC_QUEUED) && 1716*846a3351SJing Xia (inode->i_state & I_DIRTY)) 1717*846a3351SJing Xia redirty_tail_locked(inode, wb); 1718*846a3351SJing Xia 17194f8ad655SJan Kara spin_unlock(&wb->list_lock); 17201c0eeaf5SJoern Engel inode_sync_complete(inode); 17214f8ad655SJan Kara out: 17224f8ad655SJan Kara spin_unlock(&inode->i_lock); 17231da177e4SLinus Torvalds return ret; 17241da177e4SLinus Torvalds } 17251da177e4SLinus Torvalds 1726a88a341aSTejun Heo static long writeback_chunk_size(struct bdi_writeback *wb, 17271a12d8bdSWu Fengguang struct wb_writeback_work *work) 1728d46db3d5SWu Fengguang { 1729d46db3d5SWu Fengguang long pages; 1730d46db3d5SWu Fengguang 1731d46db3d5SWu Fengguang /* 1732d46db3d5SWu Fengguang * WB_SYNC_ALL mode does livelock avoidance by syncing dirty 1733d46db3d5SWu Fengguang * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX 1734d46db3d5SWu Fengguang * here avoids calling into writeback_inodes_wb() more than once. 1735d46db3d5SWu Fengguang * 1736d46db3d5SWu Fengguang * The intended call sequence for WB_SYNC_ALL writeback is: 1737d46db3d5SWu Fengguang * 1738d46db3d5SWu Fengguang * wb_writeback() 1739d46db3d5SWu Fengguang * writeback_sb_inodes() <== called only once 1740d46db3d5SWu Fengguang * write_cache_pages() <== called once for each inode 1741d46db3d5SWu Fengguang * (quickly) tag currently dirty pages 1742d46db3d5SWu Fengguang * (maybe slowly) sync all tagged pages 1743d46db3d5SWu Fengguang */ 1744d46db3d5SWu Fengguang if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) 1745d46db3d5SWu Fengguang pages = LONG_MAX; 17461a12d8bdSWu Fengguang else { 1747a88a341aSTejun Heo pages = min(wb->avg_write_bandwidth / 2, 1748dcc25ae7STejun Heo global_wb_domain.dirty_limit / DIRTY_SCOPE); 17491a12d8bdSWu Fengguang pages = min(pages, work->nr_pages); 17501a12d8bdSWu Fengguang pages = round_down(pages + MIN_WRITEBACK_PAGES, 17511a12d8bdSWu Fengguang MIN_WRITEBACK_PAGES); 17521a12d8bdSWu Fengguang } 1753d46db3d5SWu Fengguang 1754d46db3d5SWu Fengguang return pages; 1755d46db3d5SWu Fengguang } 1756d46db3d5SWu Fengguang 175703ba3782SJens Axboe /* 1758f11c9c5cSEdward Shishkin * Write a portion of b_io inodes which belong to @sb. 1759edadfb10SChristoph Hellwig * 1760d46db3d5SWu Fengguang * Return the number of pages and/or inodes written. 17610ba13fd1SLinus Torvalds * 17620ba13fd1SLinus Torvalds * NOTE! This is called with wb->list_lock held, and will 17630ba13fd1SLinus Torvalds * unlock and relock that for each inode it ends up doing 17640ba13fd1SLinus Torvalds * IO for. 1765f11c9c5cSEdward Shishkin */ 1766d46db3d5SWu Fengguang static long writeback_sb_inodes(struct super_block *sb, 1767d46db3d5SWu Fengguang struct bdi_writeback *wb, 1768d46db3d5SWu Fengguang struct wb_writeback_work *work) 176903ba3782SJens Axboe { 1770d46db3d5SWu Fengguang struct writeback_control wbc = { 1771d46db3d5SWu Fengguang .sync_mode = work->sync_mode, 1772d46db3d5SWu Fengguang .tagged_writepages = work->tagged_writepages, 1773d46db3d5SWu Fengguang .for_kupdate = work->for_kupdate, 1774d46db3d5SWu Fengguang .for_background = work->for_background, 17757747bd4bSDave Chinner .for_sync = work->for_sync, 1776d46db3d5SWu Fengguang .range_cyclic = work->range_cyclic, 1777d46db3d5SWu Fengguang .range_start = 0, 1778d46db3d5SWu Fengguang .range_end = LLONG_MAX, 1779d46db3d5SWu Fengguang }; 1780d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1781d46db3d5SWu Fengguang long write_chunk; 1782d46db3d5SWu Fengguang long wrote = 0; /* count both pages and inodes */ 1783d46db3d5SWu Fengguang 178403ba3782SJens Axboe while (!list_empty(&wb->b_io)) { 17857ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1786aaf25593STejun Heo struct bdi_writeback *tmp_wb; 1787edadfb10SChristoph Hellwig 1788edadfb10SChristoph Hellwig if (inode->i_sb != sb) { 1789d46db3d5SWu Fengguang if (work->sb) { 1790edadfb10SChristoph Hellwig /* 1791edadfb10SChristoph Hellwig * We only want to write back data for this 1792edadfb10SChristoph Hellwig * superblock, move all inodes not belonging 1793edadfb10SChristoph Hellwig * to it back onto the dirty list. 1794edadfb10SChristoph Hellwig */ 1795f758eeabSChristoph Hellwig redirty_tail(inode, wb); 179666f3b8e2SJens Axboe continue; 179766f3b8e2SJens Axboe } 1798edadfb10SChristoph Hellwig 1799edadfb10SChristoph Hellwig /* 1800edadfb10SChristoph Hellwig * The inode belongs to a different superblock. 1801edadfb10SChristoph Hellwig * Bounce back to the caller to unpin this and 1802edadfb10SChristoph Hellwig * pin the next superblock. 1803edadfb10SChristoph Hellwig */ 1804d46db3d5SWu Fengguang break; 1805edadfb10SChristoph Hellwig } 1806edadfb10SChristoph Hellwig 18079843b76aSChristoph Hellwig /* 1808331cbdeeSWanpeng Li * Don't bother with new inodes or inodes being freed, first 1809331cbdeeSWanpeng Li * kind does not need periodic writeout yet, and for the latter 18109843b76aSChristoph Hellwig * kind writeout is handled by the freer. 18119843b76aSChristoph Hellwig */ 1812250df6edSDave Chinner spin_lock(&inode->i_lock); 18139843b76aSChristoph Hellwig if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { 1814b35250c0SJan Kara redirty_tail_locked(inode, wb); 1815250df6edSDave Chinner spin_unlock(&inode->i_lock); 18167ef0d737SNick Piggin continue; 18177ef0d737SNick Piggin } 1818cc1676d9SJan Kara if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { 1819cc1676d9SJan Kara /* 1820cc1676d9SJan Kara * If this inode is locked for writeback and we are not 1821cc1676d9SJan Kara * doing writeback-for-data-integrity, move it to 1822cc1676d9SJan Kara * b_more_io so that writeback can proceed with the 1823cc1676d9SJan Kara * other inodes on s_io. 1824cc1676d9SJan Kara * 1825cc1676d9SJan Kara * We'll have another go at writing back this inode 1826cc1676d9SJan Kara * when we completed a full scan of b_io. 1827cc1676d9SJan Kara */ 1828cc1676d9SJan Kara spin_unlock(&inode->i_lock); 1829cc1676d9SJan Kara requeue_io(inode, wb); 1830cc1676d9SJan Kara trace_writeback_sb_inodes_requeue(inode); 1831cc1676d9SJan Kara continue; 1832cc1676d9SJan Kara } 1833f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 1834f0d07b7fSJan Kara 18354f8ad655SJan Kara /* 18364f8ad655SJan Kara * We already requeued the inode if it had I_SYNC set and we 18374f8ad655SJan Kara * are doing WB_SYNC_NONE writeback. So this catches only the 18384f8ad655SJan Kara * WB_SYNC_ALL case. 18394f8ad655SJan Kara */ 1840169ebd90SJan Kara if (inode->i_state & I_SYNC) { 1841169ebd90SJan Kara /* Wait for I_SYNC. This function drops i_lock... */ 1842169ebd90SJan Kara inode_sleep_on_writeback(inode); 1843169ebd90SJan Kara /* Inode may be gone, start again */ 1844ead188f9SJan Kara spin_lock(&wb->list_lock); 1845169ebd90SJan Kara continue; 1846169ebd90SJan Kara } 18474f8ad655SJan Kara inode->i_state |= I_SYNC; 1848b16b1debSTejun Heo wbc_attach_and_unlock_inode(&wbc, inode); 1849169ebd90SJan Kara 1850a88a341aSTejun Heo write_chunk = writeback_chunk_size(wb, work); 1851d46db3d5SWu Fengguang wbc.nr_to_write = write_chunk; 1852d46db3d5SWu Fengguang wbc.pages_skipped = 0; 1853250df6edSDave Chinner 1854169ebd90SJan Kara /* 1855169ebd90SJan Kara * We use I_SYNC to pin the inode in memory. While it is set 1856169ebd90SJan Kara * evict_inode() will wait so the inode cannot be freed. 1857169ebd90SJan Kara */ 1858cd8ed2a4SYan Hong __writeback_single_inode(inode, &wbc); 1859d46db3d5SWu Fengguang 1860b16b1debSTejun Heo wbc_detach_inode(&wbc); 1861d46db3d5SWu Fengguang work->nr_pages -= write_chunk - wbc.nr_to_write; 1862d46db3d5SWu Fengguang wrote += write_chunk - wbc.nr_to_write; 1863590dca3aSChris Mason 1864590dca3aSChris Mason if (need_resched()) { 1865590dca3aSChris Mason /* 1866590dca3aSChris Mason * We're trying to balance between building up a nice 1867590dca3aSChris Mason * long list of IOs to improve our merge rate, and 1868590dca3aSChris Mason * getting those IOs out quickly for anyone throttling 1869590dca3aSChris Mason * in balance_dirty_pages(). cond_resched() doesn't 1870590dca3aSChris Mason * unplug, so get our IOs out the door before we 1871590dca3aSChris Mason * give up the CPU. 1872590dca3aSChris Mason */ 1873008f75a2SChristoph Hellwig blk_flush_plug(current->plug, false); 1874590dca3aSChris Mason cond_resched(); 1875590dca3aSChris Mason } 1876590dca3aSChris Mason 1877aaf25593STejun Heo /* 1878aaf25593STejun Heo * Requeue @inode if still dirty. Be careful as @inode may 1879aaf25593STejun Heo * have been switched to another wb in the meantime. 1880aaf25593STejun Heo */ 1881aaf25593STejun Heo tmp_wb = inode_to_wb_and_lock_list(inode); 18824f8ad655SJan Kara spin_lock(&inode->i_lock); 18830ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1884d46db3d5SWu Fengguang wrote++; 1885aaf25593STejun Heo requeue_inode(inode, tmp_wb, &wbc); 18864f8ad655SJan Kara inode_sync_complete(inode); 18870f1b1fd8SDave Chinner spin_unlock(&inode->i_lock); 1888590dca3aSChris Mason 1889aaf25593STejun Heo if (unlikely(tmp_wb != wb)) { 1890aaf25593STejun Heo spin_unlock(&tmp_wb->list_lock); 1891aaf25593STejun Heo spin_lock(&wb->list_lock); 1892aaf25593STejun Heo } 1893aaf25593STejun Heo 1894d46db3d5SWu Fengguang /* 1895d46db3d5SWu Fengguang * bail out to wb_writeback() often enough to check 1896d46db3d5SWu Fengguang * background threshold and other termination conditions. 1897d46db3d5SWu Fengguang */ 1898d46db3d5SWu Fengguang if (wrote) { 1899d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1900d46db3d5SWu Fengguang break; 1901d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1902d46db3d5SWu Fengguang break; 19031da177e4SLinus Torvalds } 19048bc3be27SFengguang Wu } 1905d46db3d5SWu Fengguang return wrote; 1906f11c9c5cSEdward Shishkin } 190738f21977SNick Piggin 1908d46db3d5SWu Fengguang static long __writeback_inodes_wb(struct bdi_writeback *wb, 1909d46db3d5SWu Fengguang struct wb_writeback_work *work) 1910f11c9c5cSEdward Shishkin { 1911d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1912d46db3d5SWu Fengguang long wrote = 0; 1913f11c9c5cSEdward Shishkin 1914f11c9c5cSEdward Shishkin while (!list_empty(&wb->b_io)) { 19157ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1916f11c9c5cSEdward Shishkin struct super_block *sb = inode->i_sb; 1917f11c9c5cSEdward Shishkin 1918eb6ef3dfSKonstantin Khlebnikov if (!trylock_super(sb)) { 19190e995816SWu Fengguang /* 1920eb6ef3dfSKonstantin Khlebnikov * trylock_super() may fail consistently due to 19210e995816SWu Fengguang * s_umount being grabbed by someone else. Don't use 19220e995816SWu Fengguang * requeue_io() to avoid busy retrying the inode/sb. 19230e995816SWu Fengguang */ 19240e995816SWu Fengguang redirty_tail(inode, wb); 1925d19de7edSChristoph Hellwig continue; 1926334132aeSChristoph Hellwig } 1927d46db3d5SWu Fengguang wrote += writeback_sb_inodes(sb, wb, work); 1928eb6ef3dfSKonstantin Khlebnikov up_read(&sb->s_umount); 1929f11c9c5cSEdward Shishkin 1930d46db3d5SWu Fengguang /* refer to the same tests at the end of writeback_sb_inodes */ 1931d46db3d5SWu Fengguang if (wrote) { 1932d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1933d46db3d5SWu Fengguang break; 1934d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1935f11c9c5cSEdward Shishkin break; 1936f11c9c5cSEdward Shishkin } 1937d46db3d5SWu Fengguang } 193866f3b8e2SJens Axboe /* Leave any unwritten inodes on b_io */ 1939d46db3d5SWu Fengguang return wrote; 194066f3b8e2SJens Axboe } 194166f3b8e2SJens Axboe 19427d9f073bSWanpeng Li static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, 19430e175a18SCurt Wohlgemuth enum wb_reason reason) 1944edadfb10SChristoph Hellwig { 1945d46db3d5SWu Fengguang struct wb_writeback_work work = { 1946d46db3d5SWu Fengguang .nr_pages = nr_pages, 1947d46db3d5SWu Fengguang .sync_mode = WB_SYNC_NONE, 1948d46db3d5SWu Fengguang .range_cyclic = 1, 19490e175a18SCurt Wohlgemuth .reason = reason, 1950d46db3d5SWu Fengguang }; 1951505a666eSLinus Torvalds struct blk_plug plug; 1952edadfb10SChristoph Hellwig 1953505a666eSLinus Torvalds blk_start_plug(&plug); 1954f758eeabSChristoph Hellwig spin_lock(&wb->list_lock); 1955424b351fSWu Fengguang if (list_empty(&wb->b_io)) 1956f9cae926SJan Kara queue_io(wb, &work, jiffies); 1957d46db3d5SWu Fengguang __writeback_inodes_wb(wb, &work); 1958f758eeabSChristoph Hellwig spin_unlock(&wb->list_lock); 1959505a666eSLinus Torvalds blk_finish_plug(&plug); 1960edadfb10SChristoph Hellwig 1961d46db3d5SWu Fengguang return nr_pages - work.nr_pages; 196266f3b8e2SJens Axboe } 196366f3b8e2SJens Axboe 196403ba3782SJens Axboe /* 196503ba3782SJens Axboe * Explicit flushing or periodic writeback of "old" data. 196603ba3782SJens Axboe * 196703ba3782SJens Axboe * Define "old": the first time one of an inode's pages is dirtied, we mark the 196803ba3782SJens Axboe * dirtying-time in the inode's address_space. So this periodic writeback code 196903ba3782SJens Axboe * just walks the superblock inode list, writing back any inodes which are 197003ba3782SJens Axboe * older than a specific point in time. 197103ba3782SJens Axboe * 197203ba3782SJens Axboe * Try to run once per dirty_writeback_interval. But if a writeback event 197303ba3782SJens Axboe * takes longer than a dirty_writeback_interval interval, then leave a 197403ba3782SJens Axboe * one-second gap. 197503ba3782SJens Axboe * 1976f9cae926SJan Kara * dirtied_before takes precedence over nr_to_write. So we'll only write back 197703ba3782SJens Axboe * all dirty pages if they are all attached to "old" mappings. 197803ba3782SJens Axboe */ 1979c4a77a6cSJens Axboe static long wb_writeback(struct bdi_writeback *wb, 198083ba7b07SChristoph Hellwig struct wb_writeback_work *work) 198103ba3782SJens Axboe { 1982d46db3d5SWu Fengguang long nr_pages = work->nr_pages; 1983f9cae926SJan Kara unsigned long dirtied_before = jiffies; 1984a5989bdcSJan Kara struct inode *inode; 1985d46db3d5SWu Fengguang long progress; 1986505a666eSLinus Torvalds struct blk_plug plug; 198703ba3782SJens Axboe 1988505a666eSLinus Torvalds blk_start_plug(&plug); 1989e8dfc305SWu Fengguang spin_lock(&wb->list_lock); 199003ba3782SJens Axboe for (;;) { 199103ba3782SJens Axboe /* 1992d3ddec76SWu Fengguang * Stop writeback when nr_pages has been consumed 199303ba3782SJens Axboe */ 199483ba7b07SChristoph Hellwig if (work->nr_pages <= 0) 199503ba3782SJens Axboe break; 199603ba3782SJens Axboe 199703ba3782SJens Axboe /* 1998aa373cf5SJan Kara * Background writeout and kupdate-style writeback may 1999aa373cf5SJan Kara * run forever. Stop them if there is other work to do 2000aa373cf5SJan Kara * so that e.g. sync can proceed. They'll be restarted 2001aa373cf5SJan Kara * after the other works are all done. 2002aa373cf5SJan Kara */ 2003aa373cf5SJan Kara if ((work->for_background || work->for_kupdate) && 2004f0054bb1STejun Heo !list_empty(&wb->work_list)) 2005aa373cf5SJan Kara break; 2006aa373cf5SJan Kara 2007aa373cf5SJan Kara /* 2008d3ddec76SWu Fengguang * For background writeout, stop when we are below the 2009d3ddec76SWu Fengguang * background dirty threshold 201003ba3782SJens Axboe */ 2011aa661bbeSTejun Heo if (work->for_background && !wb_over_bg_thresh(wb)) 201203ba3782SJens Axboe break; 201303ba3782SJens Axboe 20141bc36b64SJan Kara /* 20151bc36b64SJan Kara * Kupdate and background works are special and we want to 20161bc36b64SJan Kara * include all inodes that need writing. Livelock avoidance is 20171bc36b64SJan Kara * handled by these works yielding to any other work so we are 20181bc36b64SJan Kara * safe. 20191bc36b64SJan Kara */ 2020ba9aa839SWu Fengguang if (work->for_kupdate) { 2021f9cae926SJan Kara dirtied_before = jiffies - 2022ba9aa839SWu Fengguang msecs_to_jiffies(dirty_expire_interval * 10); 20231bc36b64SJan Kara } else if (work->for_background) 2024f9cae926SJan Kara dirtied_before = jiffies; 2025028c2dd1SDave Chinner 20265634cc2aSTejun Heo trace_writeback_start(wb, work); 2027e8dfc305SWu Fengguang if (list_empty(&wb->b_io)) 2028f9cae926SJan Kara queue_io(wb, work, dirtied_before); 202983ba7b07SChristoph Hellwig if (work->sb) 2030d46db3d5SWu Fengguang progress = writeback_sb_inodes(work->sb, wb, work); 2031edadfb10SChristoph Hellwig else 2032d46db3d5SWu Fengguang progress = __writeback_inodes_wb(wb, work); 20335634cc2aSTejun Heo trace_writeback_written(wb, work); 2034028c2dd1SDave Chinner 203503ba3782SJens Axboe /* 203671fd05a8SJens Axboe * Did we write something? Try for more 2037e6fb6da2SWu Fengguang * 2038e6fb6da2SWu Fengguang * Dirty inodes are moved to b_io for writeback in batches. 2039e6fb6da2SWu Fengguang * The completion of the current batch does not necessarily 2040e6fb6da2SWu Fengguang * mean the overall work is done. So we keep looping as long 2041e6fb6da2SWu Fengguang * as made some progress on cleaning pages or inodes. 204271fd05a8SJens Axboe */ 2043d46db3d5SWu Fengguang if (progress) 204403ba3782SJens Axboe continue; 2045a5989bdcSJan Kara /* 2046e6fb6da2SWu Fengguang * No more inodes for IO, bail 2047a5989bdcSJan Kara */ 2048b7a2441fSWu Fengguang if (list_empty(&wb->b_more_io)) 204903ba3782SJens Axboe break; 205003ba3782SJens Axboe /* 20518010c3b6SJens Axboe * Nothing written. Wait for some inode to 20528010c3b6SJens Axboe * become available for writeback. Otherwise 20538010c3b6SJens Axboe * we'll just busyloop. 205403ba3782SJens Axboe */ 20555634cc2aSTejun Heo trace_writeback_wait(wb, work); 205603ba3782SJens Axboe inode = wb_inode(wb->b_more_io.prev); 2057250df6edSDave Chinner spin_lock(&inode->i_lock); 2058f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 2059169ebd90SJan Kara /* This function drops i_lock... */ 2060169ebd90SJan Kara inode_sleep_on_writeback(inode); 2061f0d07b7fSJan Kara spin_lock(&wb->list_lock); 206203ba3782SJens Axboe } 2063e8dfc305SWu Fengguang spin_unlock(&wb->list_lock); 2064505a666eSLinus Torvalds blk_finish_plug(&plug); 206503ba3782SJens Axboe 2066d46db3d5SWu Fengguang return nr_pages - work->nr_pages; 206703ba3782SJens Axboe } 206803ba3782SJens Axboe 206903ba3782SJens Axboe /* 207083ba7b07SChristoph Hellwig * Return the next wb_writeback_work struct that hasn't been processed yet. 207103ba3782SJens Axboe */ 2072f0054bb1STejun Heo static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) 207303ba3782SJens Axboe { 207483ba7b07SChristoph Hellwig struct wb_writeback_work *work = NULL; 207503ba3782SJens Axboe 2076f0054bb1STejun Heo spin_lock_bh(&wb->work_lock); 2077f0054bb1STejun Heo if (!list_empty(&wb->work_list)) { 2078f0054bb1STejun Heo work = list_entry(wb->work_list.next, 207983ba7b07SChristoph Hellwig struct wb_writeback_work, list); 208083ba7b07SChristoph Hellwig list_del_init(&work->list); 208103ba3782SJens Axboe } 2082f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 208383ba7b07SChristoph Hellwig return work; 208403ba3782SJens Axboe } 208503ba3782SJens Axboe 20866585027aSJan Kara static long wb_check_background_flush(struct bdi_writeback *wb) 20876585027aSJan Kara { 2088aa661bbeSTejun Heo if (wb_over_bg_thresh(wb)) { 20896585027aSJan Kara 20906585027aSJan Kara struct wb_writeback_work work = { 20916585027aSJan Kara .nr_pages = LONG_MAX, 20926585027aSJan Kara .sync_mode = WB_SYNC_NONE, 20936585027aSJan Kara .for_background = 1, 20946585027aSJan Kara .range_cyclic = 1, 20950e175a18SCurt Wohlgemuth .reason = WB_REASON_BACKGROUND, 20966585027aSJan Kara }; 20976585027aSJan Kara 20986585027aSJan Kara return wb_writeback(wb, &work); 20996585027aSJan Kara } 21006585027aSJan Kara 21016585027aSJan Kara return 0; 21026585027aSJan Kara } 21036585027aSJan Kara 210403ba3782SJens Axboe static long wb_check_old_data_flush(struct bdi_writeback *wb) 210503ba3782SJens Axboe { 210603ba3782SJens Axboe unsigned long expired; 210703ba3782SJens Axboe long nr_pages; 210803ba3782SJens Axboe 210969b62d01SJens Axboe /* 211069b62d01SJens Axboe * When set to zero, disable periodic writeback 211169b62d01SJens Axboe */ 211269b62d01SJens Axboe if (!dirty_writeback_interval) 211369b62d01SJens Axboe return 0; 211469b62d01SJens Axboe 211503ba3782SJens Axboe expired = wb->last_old_flush + 211603ba3782SJens Axboe msecs_to_jiffies(dirty_writeback_interval * 10); 211703ba3782SJens Axboe if (time_before(jiffies, expired)) 211803ba3782SJens Axboe return 0; 211903ba3782SJens Axboe 212003ba3782SJens Axboe wb->last_old_flush = jiffies; 2121cdf01dd5SLinus Torvalds nr_pages = get_nr_dirty_pages(); 212203ba3782SJens Axboe 2123c4a77a6cSJens Axboe if (nr_pages) { 212483ba7b07SChristoph Hellwig struct wb_writeback_work work = { 2125c4a77a6cSJens Axboe .nr_pages = nr_pages, 2126c4a77a6cSJens Axboe .sync_mode = WB_SYNC_NONE, 2127c4a77a6cSJens Axboe .for_kupdate = 1, 2128c4a77a6cSJens Axboe .range_cyclic = 1, 21290e175a18SCurt Wohlgemuth .reason = WB_REASON_PERIODIC, 2130c4a77a6cSJens Axboe }; 2131c4a77a6cSJens Axboe 213283ba7b07SChristoph Hellwig return wb_writeback(wb, &work); 2133c4a77a6cSJens Axboe } 213403ba3782SJens Axboe 213503ba3782SJens Axboe return 0; 213603ba3782SJens Axboe } 213703ba3782SJens Axboe 213885009b4fSJens Axboe static long wb_check_start_all(struct bdi_writeback *wb) 213985009b4fSJens Axboe { 214085009b4fSJens Axboe long nr_pages; 214185009b4fSJens Axboe 214285009b4fSJens Axboe if (!test_bit(WB_start_all, &wb->state)) 214385009b4fSJens Axboe return 0; 214485009b4fSJens Axboe 214585009b4fSJens Axboe nr_pages = get_nr_dirty_pages(); 214685009b4fSJens Axboe if (nr_pages) { 214785009b4fSJens Axboe struct wb_writeback_work work = { 214885009b4fSJens Axboe .nr_pages = wb_split_bdi_pages(wb, nr_pages), 214985009b4fSJens Axboe .sync_mode = WB_SYNC_NONE, 215085009b4fSJens Axboe .range_cyclic = 1, 215185009b4fSJens Axboe .reason = wb->start_all_reason, 215285009b4fSJens Axboe }; 215385009b4fSJens Axboe 215485009b4fSJens Axboe nr_pages = wb_writeback(wb, &work); 215585009b4fSJens Axboe } 215685009b4fSJens Axboe 215785009b4fSJens Axboe clear_bit(WB_start_all, &wb->state); 215885009b4fSJens Axboe return nr_pages; 215985009b4fSJens Axboe } 216085009b4fSJens Axboe 216185009b4fSJens Axboe 216203ba3782SJens Axboe /* 216303ba3782SJens Axboe * Retrieve work items and do the writeback they describe 216403ba3782SJens Axboe */ 216525d130baSWanpeng Li static long wb_do_writeback(struct bdi_writeback *wb) 216603ba3782SJens Axboe { 216783ba7b07SChristoph Hellwig struct wb_writeback_work *work; 2168c4a77a6cSJens Axboe long wrote = 0; 216903ba3782SJens Axboe 21704452226eSTejun Heo set_bit(WB_writeback_running, &wb->state); 2171f0054bb1STejun Heo while ((work = get_next_work_item(wb)) != NULL) { 21725634cc2aSTejun Heo trace_writeback_exec(wb, work); 217383ba7b07SChristoph Hellwig wrote += wb_writeback(wb, work); 21744a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 217503ba3782SJens Axboe } 217603ba3782SJens Axboe 217703ba3782SJens Axboe /* 217885009b4fSJens Axboe * Check for a flush-everything request 217985009b4fSJens Axboe */ 218085009b4fSJens Axboe wrote += wb_check_start_all(wb); 218185009b4fSJens Axboe 218285009b4fSJens Axboe /* 218303ba3782SJens Axboe * Check for periodic writeback, kupdated() style 218403ba3782SJens Axboe */ 218503ba3782SJens Axboe wrote += wb_check_old_data_flush(wb); 21866585027aSJan Kara wrote += wb_check_background_flush(wb); 21874452226eSTejun Heo clear_bit(WB_writeback_running, &wb->state); 218803ba3782SJens Axboe 218903ba3782SJens Axboe return wrote; 219003ba3782SJens Axboe } 219103ba3782SJens Axboe 219203ba3782SJens Axboe /* 219303ba3782SJens Axboe * Handle writeback of dirty data for the device backed by this bdi. Also 2194839a8e86STejun Heo * reschedules periodically and does kupdated style flushing. 219503ba3782SJens Axboe */ 2196f0054bb1STejun Heo void wb_workfn(struct work_struct *work) 219703ba3782SJens Axboe { 2198839a8e86STejun Heo struct bdi_writeback *wb = container_of(to_delayed_work(work), 2199839a8e86STejun Heo struct bdi_writeback, dwork); 220003ba3782SJens Axboe long pages_written; 220103ba3782SJens Axboe 220268f23b89STheodore Ts'o set_worker_desc("flush-%s", bdi_dev_name(wb->bdi)); 220303ba3782SJens Axboe 2204839a8e86STejun Heo if (likely(!current_is_workqueue_rescuer() || 22054452226eSTejun Heo !test_bit(WB_registered, &wb->state))) { 220603ba3782SJens Axboe /* 2207f0054bb1STejun Heo * The normal path. Keep writing back @wb until its 2208839a8e86STejun Heo * work_list is empty. Note that this path is also taken 2209f0054bb1STejun Heo * if @wb is shutting down even when we're running off the 2210839a8e86STejun Heo * rescuer as work_list needs to be drained. 221103ba3782SJens Axboe */ 2212839a8e86STejun Heo do { 221325d130baSWanpeng Li pages_written = wb_do_writeback(wb); 2214455b2864SDave Chinner trace_writeback_pages_written(pages_written); 2215f0054bb1STejun Heo } while (!list_empty(&wb->work_list)); 2216839a8e86STejun Heo } else { 2217253c34e9SArtem Bityutskiy /* 2218839a8e86STejun Heo * bdi_wq can't get enough workers and we're running off 2219839a8e86STejun Heo * the emergency worker. Don't hog it. Hopefully, 1024 is 2220839a8e86STejun Heo * enough for efficient IO. 2221253c34e9SArtem Bityutskiy */ 2222f0054bb1STejun Heo pages_written = writeback_inodes_wb(wb, 1024, 2223839a8e86STejun Heo WB_REASON_FORKER_THREAD); 2224839a8e86STejun Heo trace_writeback_pages_written(pages_written); 222503ba3782SJens Axboe } 222603ba3782SJens Axboe 2227f0054bb1STejun Heo if (!list_empty(&wb->work_list)) 2228b8b78495SJan Kara wb_wakeup(wb); 22296ca738d6SDerek Basehore else if (wb_has_dirty_io(wb) && dirty_writeback_interval) 2230f0054bb1STejun Heo wb_wakeup_delayed(wb); 223103ba3782SJens Axboe } 223203ba3782SJens Axboe 223303ba3782SJens Axboe /* 2234595043e5SJens Axboe * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero, 2235595043e5SJens Axboe * write back the whole world. 2236595043e5SJens Axboe */ 2237595043e5SJens Axboe static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2238e8e8a0c6SJens Axboe enum wb_reason reason) 2239595043e5SJens Axboe { 2240595043e5SJens Axboe struct bdi_writeback *wb; 2241595043e5SJens Axboe 2242595043e5SJens Axboe if (!bdi_has_dirty_io(bdi)) 2243595043e5SJens Axboe return; 2244595043e5SJens Axboe 2245595043e5SJens Axboe list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 2246e8e8a0c6SJens Axboe wb_start_writeback(wb, reason); 2247595043e5SJens Axboe } 2248595043e5SJens Axboe 2249595043e5SJens Axboe void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2250595043e5SJens Axboe enum wb_reason reason) 2251595043e5SJens Axboe { 2252595043e5SJens Axboe rcu_read_lock(); 2253e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2254595043e5SJens Axboe rcu_read_unlock(); 2255595043e5SJens Axboe } 2256595043e5SJens Axboe 2257595043e5SJens Axboe /* 22589ba4b2dfSJens Axboe * Wakeup the flusher threads to start writeback of all currently dirty pages 225903ba3782SJens Axboe */ 22609ba4b2dfSJens Axboe void wakeup_flusher_threads(enum wb_reason reason) 226103ba3782SJens Axboe { 2262b8c2f347SChristoph Hellwig struct backing_dev_info *bdi; 2263b8c2f347SChristoph Hellwig 226451350ea0SKonstantin Khlebnikov /* 226551350ea0SKonstantin Khlebnikov * If we are expecting writeback progress we must submit plugged IO. 226651350ea0SKonstantin Khlebnikov */ 2267008f75a2SChristoph Hellwig blk_flush_plug(current->plug, true); 226851350ea0SKonstantin Khlebnikov 2269b8c2f347SChristoph Hellwig rcu_read_lock(); 2270595043e5SJens Axboe list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) 2271e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2272b8c2f347SChristoph Hellwig rcu_read_unlock(); 227303ba3782SJens Axboe } 227403ba3782SJens Axboe 2275a2f48706STheodore Ts'o /* 2276a2f48706STheodore Ts'o * Wake up bdi's periodically to make sure dirtytime inodes gets 2277a2f48706STheodore Ts'o * written back periodically. We deliberately do *not* check the 2278a2f48706STheodore Ts'o * b_dirtytime list in wb_has_dirty_io(), since this would cause the 2279a2f48706STheodore Ts'o * kernel to be constantly waking up once there are any dirtytime 2280a2f48706STheodore Ts'o * inodes on the system. So instead we define a separate delayed work 2281a2f48706STheodore Ts'o * function which gets called much more rarely. (By default, only 2282a2f48706STheodore Ts'o * once every 12 hours.) 2283a2f48706STheodore Ts'o * 2284a2f48706STheodore Ts'o * If there is any other write activity going on in the file system, 2285a2f48706STheodore Ts'o * this function won't be necessary. But if the only thing that has 2286a2f48706STheodore Ts'o * happened on the file system is a dirtytime inode caused by an atime 2287a2f48706STheodore Ts'o * update, we need this infrastructure below to make sure that inode 2288a2f48706STheodore Ts'o * eventually gets pushed out to disk. 2289a2f48706STheodore Ts'o */ 2290a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w); 2291a2f48706STheodore Ts'o static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); 2292a2f48706STheodore Ts'o 2293a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w) 2294a2f48706STheodore Ts'o { 2295a2f48706STheodore Ts'o struct backing_dev_info *bdi; 2296a2f48706STheodore Ts'o 2297a2f48706STheodore Ts'o rcu_read_lock(); 2298a2f48706STheodore Ts'o list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 2299001fe6f6STejun Heo struct bdi_writeback *wb; 2300001fe6f6STejun Heo 2301b817525aSTejun Heo list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 23026fdf860fSTejun Heo if (!list_empty(&wb->b_dirty_time)) 23036fdf860fSTejun Heo wb_wakeup(wb); 2304a2f48706STheodore Ts'o } 2305a2f48706STheodore Ts'o rcu_read_unlock(); 2306a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2307a2f48706STheodore Ts'o } 2308a2f48706STheodore Ts'o 2309a2f48706STheodore Ts'o static int __init start_dirtytime_writeback(void) 2310a2f48706STheodore Ts'o { 2311a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2312a2f48706STheodore Ts'o return 0; 2313a2f48706STheodore Ts'o } 2314a2f48706STheodore Ts'o __initcall(start_dirtytime_writeback); 2315a2f48706STheodore Ts'o 23161efff914STheodore Ts'o int dirtytime_interval_handler(struct ctl_table *table, int write, 23179ca48e20STobias Klauser void *buffer, size_t *lenp, loff_t *ppos) 23181efff914STheodore Ts'o { 23191efff914STheodore Ts'o int ret; 23201efff914STheodore Ts'o 23211efff914STheodore Ts'o ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 23221efff914STheodore Ts'o if (ret == 0 && write) 23231efff914STheodore Ts'o mod_delayed_work(system_wq, &dirtytime_work, 0); 23241efff914STheodore Ts'o return ret; 23251efff914STheodore Ts'o } 23261efff914STheodore Ts'o 232703ba3782SJens Axboe /** 232835d14f27SEric Biggers * __mark_inode_dirty - internal function to mark an inode dirty 23290117d427SMauro Carvalho Chehab * 233003ba3782SJens Axboe * @inode: inode to mark 233135d14f27SEric Biggers * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of 233235d14f27SEric Biggers * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined 233335d14f27SEric Biggers * with I_DIRTY_PAGES. 23340117d427SMauro Carvalho Chehab * 233535d14f27SEric Biggers * Mark an inode as dirty. We notify the filesystem, then update the inode's 233635d14f27SEric Biggers * dirty flags. Then, if needed we add the inode to the appropriate dirty list. 233703ba3782SJens Axboe * 233835d14f27SEric Biggers * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync() 233935d14f27SEric Biggers * instead of calling this directly. 234003ba3782SJens Axboe * 234135d14f27SEric Biggers * CAREFUL! We only add the inode to the dirty list if it is hashed or if it 234235d14f27SEric Biggers * refers to a blockdev. Unhashed inodes will never be added to the dirty list 234335d14f27SEric Biggers * even if they are later hashed, as they will have been marked dirty already. 234403ba3782SJens Axboe * 234535d14f27SEric Biggers * In short, ensure you hash any inodes _before_ you start marking them dirty. 234603ba3782SJens Axboe * 234703ba3782SJens Axboe * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 234803ba3782SJens Axboe * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 234903ba3782SJens Axboe * the kernel-internal blockdev inode represents the dirtying time of the 235003ba3782SJens Axboe * blockdev's pages. This is why for I_DIRTY_PAGES we always use 235103ba3782SJens Axboe * page->mapping->host, so the page-dirtying time is recorded in the internal 235203ba3782SJens Axboe * blockdev inode. 235303ba3782SJens Axboe */ 235403ba3782SJens Axboe void __mark_inode_dirty(struct inode *inode, int flags) 235503ba3782SJens Axboe { 235603ba3782SJens Axboe struct super_block *sb = inode->i_sb; 235735d14f27SEric Biggers int dirtytime = 0; 23580ae45f63STheodore Ts'o 23590ae45f63STheodore Ts'o trace_writeback_mark_inode_dirty(inode, flags); 236003ba3782SJens Axboe 2361e2728c56SEric Biggers if (flags & I_DIRTY_INODE) { 236235d14f27SEric Biggers /* 236335d14f27SEric Biggers * Notify the filesystem about the inode being dirtied, so that 236435d14f27SEric Biggers * (if needed) it can update on-disk fields and journal the 236535d14f27SEric Biggers * inode. This is only needed when the inode itself is being 236635d14f27SEric Biggers * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not 236735d14f27SEric Biggers * for just I_DIRTY_PAGES or I_DIRTY_TIME. 236835d14f27SEric Biggers */ 23699fb0a7daSTejun Heo trace_writeback_dirty_inode_start(inode, flags); 237003ba3782SJens Axboe if (sb->s_op->dirty_inode) 2371a38ed483SEric Biggers sb->s_op->dirty_inode(inode, flags & I_DIRTY_INODE); 23729fb0a7daSTejun Heo trace_writeback_dirty_inode(inode, flags); 2373e2728c56SEric Biggers 237435d14f27SEric Biggers /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */ 23750ae45f63STheodore Ts'o flags &= ~I_DIRTY_TIME; 237635d14f27SEric Biggers } else { 237735d14f27SEric Biggers /* 237835d14f27SEric Biggers * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing. 237935d14f27SEric Biggers * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME 238035d14f27SEric Biggers * in one call to __mark_inode_dirty().) 238135d14f27SEric Biggers */ 23820ae45f63STheodore Ts'o dirtytime = flags & I_DIRTY_TIME; 238335d14f27SEric Biggers WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME); 238435d14f27SEric Biggers } 238503ba3782SJens Axboe 238603ba3782SJens Axboe /* 23879c6ac78eSTejun Heo * Paired with smp_mb() in __writeback_single_inode() for the 23889c6ac78eSTejun Heo * following lockless i_state test. See there for details. 238903ba3782SJens Axboe */ 239003ba3782SJens Axboe smp_mb(); 239103ba3782SJens Axboe 23920ae45f63STheodore Ts'o if (((inode->i_state & flags) == flags) || 23930ae45f63STheodore Ts'o (dirtytime && (inode->i_state & I_DIRTY_INODE))) 239403ba3782SJens Axboe return; 239503ba3782SJens Axboe 2396250df6edSDave Chinner spin_lock(&inode->i_lock); 23970ae45f63STheodore Ts'o if (dirtytime && (inode->i_state & I_DIRTY_INODE)) 23980ae45f63STheodore Ts'o goto out_unlock_inode; 239903ba3782SJens Axboe if ((inode->i_state & flags) != flags) { 240003ba3782SJens Axboe const int was_dirty = inode->i_state & I_DIRTY; 240103ba3782SJens Axboe 240252ebea74STejun Heo inode_attach_wb(inode, NULL); 240352ebea74STejun Heo 240435d14f27SEric Biggers /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */ 24050ae45f63STheodore Ts'o if (flags & I_DIRTY_INODE) 24060ae45f63STheodore Ts'o inode->i_state &= ~I_DIRTY_TIME; 240703ba3782SJens Axboe inode->i_state |= flags; 240803ba3782SJens Axboe 240903ba3782SJens Axboe /* 24105afced3bSJan Kara * If the inode is queued for writeback by flush worker, just 24115afced3bSJan Kara * update its dirty state. Once the flush worker is done with 24125afced3bSJan Kara * the inode it will place it on the appropriate superblock 24135afced3bSJan Kara * list, based upon its state. 241403ba3782SJens Axboe */ 24155afced3bSJan Kara if (inode->i_state & I_SYNC_QUEUED) 2416250df6edSDave Chinner goto out_unlock_inode; 241703ba3782SJens Axboe 241803ba3782SJens Axboe /* 241903ba3782SJens Axboe * Only add valid (hashed) inodes to the superblock's 242003ba3782SJens Axboe * dirty list. Add blockdev inodes as well. 242103ba3782SJens Axboe */ 242203ba3782SJens Axboe if (!S_ISBLK(inode->i_mode)) { 24231d3382cbSAl Viro if (inode_unhashed(inode)) 2424250df6edSDave Chinner goto out_unlock_inode; 242503ba3782SJens Axboe } 2426a4ffdde6SAl Viro if (inode->i_state & I_FREEING) 2427250df6edSDave Chinner goto out_unlock_inode; 242803ba3782SJens Axboe 242903ba3782SJens Axboe /* 243003ba3782SJens Axboe * If the inode was already on b_dirty/b_io/b_more_io, don't 243103ba3782SJens Axboe * reposition it (that would break b_dirty time-ordering). 243203ba3782SJens Axboe */ 243303ba3782SJens Axboe if (!was_dirty) { 243487e1d789STejun Heo struct bdi_writeback *wb; 2435d6c10f1fSTejun Heo struct list_head *dirty_list; 2436a66979abSDave Chinner bool wakeup_bdi = false; 2437500b067cSJens Axboe 243887e1d789STejun Heo wb = locked_inode_to_wb_and_lock_list(inode); 2439253c34e9SArtem Bityutskiy 244003ba3782SJens Axboe inode->dirtied_when = jiffies; 2441a2f48706STheodore Ts'o if (dirtytime) 2442a2f48706STheodore Ts'o inode->dirtied_time_when = jiffies; 2443d6c10f1fSTejun Heo 24440e11f644SChristoph Hellwig if (inode->i_state & I_DIRTY) 24450747259dSTejun Heo dirty_list = &wb->b_dirty; 2446a2f48706STheodore Ts'o else 24470747259dSTejun Heo dirty_list = &wb->b_dirty_time; 2448d6c10f1fSTejun Heo 2449c7f54084SDave Chinner wakeup_bdi = inode_io_list_move_locked(inode, wb, 2450d6c10f1fSTejun Heo dirty_list); 2451d6c10f1fSTejun Heo 24520747259dSTejun Heo spin_unlock(&wb->list_lock); 24530ae45f63STheodore Ts'o trace_writeback_dirty_inode_enqueue(inode); 2454253c34e9SArtem Bityutskiy 2455d6c10f1fSTejun Heo /* 2456d6c10f1fSTejun Heo * If this is the first dirty inode for this bdi, 2457d6c10f1fSTejun Heo * we have to wake-up the corresponding bdi thread 2458d6c10f1fSTejun Heo * to make sure background write-back happens 2459d6c10f1fSTejun Heo * later. 2460d6c10f1fSTejun Heo */ 2461f56753acSChristoph Hellwig if (wakeup_bdi && 2462f56753acSChristoph Hellwig (wb->bdi->capabilities & BDI_CAP_WRITEBACK)) 24630747259dSTejun Heo wb_wakeup_delayed(wb); 2464a66979abSDave Chinner return; 2465a66979abSDave Chinner } 2466a66979abSDave Chinner } 2467a66979abSDave Chinner out_unlock_inode: 2468a66979abSDave Chinner spin_unlock(&inode->i_lock); 246903ba3782SJens Axboe } 247003ba3782SJens Axboe EXPORT_SYMBOL(__mark_inode_dirty); 247103ba3782SJens Axboe 2472e97fedb9SDave Chinner /* 2473e97fedb9SDave Chinner * The @s_sync_lock is used to serialise concurrent sync operations 2474e97fedb9SDave Chinner * to avoid lock contention problems with concurrent wait_sb_inodes() calls. 2475e97fedb9SDave Chinner * Concurrent callers will block on the s_sync_lock rather than doing contending 2476e97fedb9SDave Chinner * walks. The queueing maintains sync(2) required behaviour as all the IO that 2477e97fedb9SDave Chinner * has been issued up to the time this function is enter is guaranteed to be 2478e97fedb9SDave Chinner * completed by the time we have gained the lock and waited for all IO that is 2479e97fedb9SDave Chinner * in progress regardless of the order callers are granted the lock. 2480e97fedb9SDave Chinner */ 2481b6e51316SJens Axboe static void wait_sb_inodes(struct super_block *sb) 248266f3b8e2SJens Axboe { 24836c60d2b5SDave Chinner LIST_HEAD(sync_list); 248438f21977SNick Piggin 248503ba3782SJens Axboe /* 248603ba3782SJens Axboe * We need to be protected against the filesystem going from 248703ba3782SJens Axboe * r/o to r/w or vice versa. 248803ba3782SJens Axboe */ 2489b6e51316SJens Axboe WARN_ON(!rwsem_is_locked(&sb->s_umount)); 249003ba3782SJens Axboe 2491e97fedb9SDave Chinner mutex_lock(&sb->s_sync_lock); 249266f3b8e2SJens Axboe 249338f21977SNick Piggin /* 24946c60d2b5SDave Chinner * Splice the writeback list onto a temporary list to avoid waiting on 24956c60d2b5SDave Chinner * inodes that have started writeback after this point. 24966c60d2b5SDave Chinner * 24976c60d2b5SDave Chinner * Use rcu_read_lock() to keep the inodes around until we have a 24986c60d2b5SDave Chinner * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as 24996c60d2b5SDave Chinner * the local list because inodes can be dropped from either by writeback 25006c60d2b5SDave Chinner * completion. 250138f21977SNick Piggin */ 25026c60d2b5SDave Chinner rcu_read_lock(); 25036c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 25046c60d2b5SDave Chinner list_splice_init(&sb->s_inodes_wb, &sync_list); 25056c60d2b5SDave Chinner 25066c60d2b5SDave Chinner /* 25076c60d2b5SDave Chinner * Data integrity sync. Must wait for all pages under writeback, because 25086c60d2b5SDave Chinner * there may have been pages dirtied before our sync call, but which had 25096c60d2b5SDave Chinner * writeout started before we write it out. In which case, the inode 25106c60d2b5SDave Chinner * may not be on the dirty list, but we still have to wait for that 25116c60d2b5SDave Chinner * writeout. 25126c60d2b5SDave Chinner */ 25136c60d2b5SDave Chinner while (!list_empty(&sync_list)) { 25146c60d2b5SDave Chinner struct inode *inode = list_first_entry(&sync_list, struct inode, 25156c60d2b5SDave Chinner i_wb_list); 2516250df6edSDave Chinner struct address_space *mapping = inode->i_mapping; 251738f21977SNick Piggin 25186c60d2b5SDave Chinner /* 25196c60d2b5SDave Chinner * Move each inode back to the wb list before we drop the lock 25206c60d2b5SDave Chinner * to preserve consistency between i_wb_list and the mapping 25216c60d2b5SDave Chinner * writeback tag. Writeback completion is responsible to remove 25226c60d2b5SDave Chinner * the inode from either list once the writeback tag is cleared. 25236c60d2b5SDave Chinner */ 25246c60d2b5SDave Chinner list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); 25256c60d2b5SDave Chinner 25266c60d2b5SDave Chinner /* 25276c60d2b5SDave Chinner * The mapping can appear untagged while still on-list since we 25286c60d2b5SDave Chinner * do not have the mapping lock. Skip it here, wb completion 25296c60d2b5SDave Chinner * will remove it. 25306c60d2b5SDave Chinner */ 25316c60d2b5SDave Chinner if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) 25326c60d2b5SDave Chinner continue; 25336c60d2b5SDave Chinner 25346c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 25356c60d2b5SDave Chinner 2536250df6edSDave Chinner spin_lock(&inode->i_lock); 25376c60d2b5SDave Chinner if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { 2538250df6edSDave Chinner spin_unlock(&inode->i_lock); 25396c60d2b5SDave Chinner 25406c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 254138f21977SNick Piggin continue; 2542250df6edSDave Chinner } 254338f21977SNick Piggin __iget(inode); 2544250df6edSDave Chinner spin_unlock(&inode->i_lock); 25456c60d2b5SDave Chinner rcu_read_unlock(); 254638f21977SNick Piggin 2547aa750fd7SJunichi Nomura /* 2548aa750fd7SJunichi Nomura * We keep the error status of individual mapping so that 2549aa750fd7SJunichi Nomura * applications can catch the writeback error using fsync(2). 2550aa750fd7SJunichi Nomura * See filemap_fdatawait_keep_errors() for details. 2551aa750fd7SJunichi Nomura */ 2552aa750fd7SJunichi Nomura filemap_fdatawait_keep_errors(mapping); 255338f21977SNick Piggin 255438f21977SNick Piggin cond_resched(); 255538f21977SNick Piggin 25566c60d2b5SDave Chinner iput(inode); 25576c60d2b5SDave Chinner 25586c60d2b5SDave Chinner rcu_read_lock(); 25596c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 256038f21977SNick Piggin } 25616c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 25626c60d2b5SDave Chinner rcu_read_unlock(); 2563e97fedb9SDave Chinner mutex_unlock(&sb->s_sync_lock); 256466f3b8e2SJens Axboe } 25651da177e4SLinus Torvalds 2566f30a7d0cSTejun Heo static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2567f30a7d0cSTejun Heo enum wb_reason reason, bool skip_if_busy) 25681da177e4SLinus Torvalds { 25695b9cce4cSTejun Heo struct backing_dev_info *bdi = sb->s_bdi; 25705b9cce4cSTejun Heo DEFINE_WB_COMPLETION(done, bdi); 257183ba7b07SChristoph Hellwig struct wb_writeback_work work = { 25723c4d7165SChristoph Hellwig .sb = sb, 25733c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_NONE, 25746e6938b6SWu Fengguang .tagged_writepages = 1, 257583ba7b07SChristoph Hellwig .done = &done, 25763259f8beSChris Mason .nr_pages = nr, 25770e175a18SCurt Wohlgemuth .reason = reason, 25783c4d7165SChristoph Hellwig }; 25790e3c9a22SJens Axboe 2580e7972912STejun Heo if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 25816eedc701SJan Kara return; 2582cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2583f30a7d0cSTejun Heo 2584db125360STejun Heo bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); 25855b9cce4cSTejun Heo wb_wait_for_completion(&done); 25861da177e4SLinus Torvalds } 2587f30a7d0cSTejun Heo 2588f30a7d0cSTejun Heo /** 2589f30a7d0cSTejun Heo * writeback_inodes_sb_nr - writeback dirty inodes from given super_block 2590f30a7d0cSTejun Heo * @sb: the superblock 2591f30a7d0cSTejun Heo * @nr: the number of pages to write 2592f30a7d0cSTejun Heo * @reason: reason why some writeback work initiated 2593f30a7d0cSTejun Heo * 2594f30a7d0cSTejun Heo * Start writeback on some inodes on this super_block. No guarantees are made 2595f30a7d0cSTejun Heo * on how many (if any) will be written, and this function does not wait 2596f30a7d0cSTejun Heo * for IO completion of submitted IO. 2597f30a7d0cSTejun Heo */ 2598f30a7d0cSTejun Heo void writeback_inodes_sb_nr(struct super_block *sb, 2599f30a7d0cSTejun Heo unsigned long nr, 2600f30a7d0cSTejun Heo enum wb_reason reason) 2601f30a7d0cSTejun Heo { 2602f30a7d0cSTejun Heo __writeback_inodes_sb_nr(sb, nr, reason, false); 2603f30a7d0cSTejun Heo } 26043259f8beSChris Mason EXPORT_SYMBOL(writeback_inodes_sb_nr); 26053259f8beSChris Mason 26063259f8beSChris Mason /** 26073259f8beSChris Mason * writeback_inodes_sb - writeback dirty inodes from given super_block 26083259f8beSChris Mason * @sb: the superblock 2609786228abSMarcos Paulo de Souza * @reason: reason why some writeback work was initiated 26103259f8beSChris Mason * 26113259f8beSChris Mason * Start writeback on some inodes on this super_block. No guarantees are made 26123259f8beSChris Mason * on how many (if any) will be written, and this function does not wait 26133259f8beSChris Mason * for IO completion of submitted IO. 26143259f8beSChris Mason */ 26150e175a18SCurt Wohlgemuth void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 26163259f8beSChris Mason { 26170e175a18SCurt Wohlgemuth return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 26183259f8beSChris Mason } 2619d8a8559cSJens Axboe EXPORT_SYMBOL(writeback_inodes_sb); 2620d8a8559cSJens Axboe 2621d8a8559cSJens Axboe /** 262210ee27a0SMiao Xie * try_to_writeback_inodes_sb - try to start writeback if none underway 262310ee27a0SMiao Xie * @sb: the superblock 262410ee27a0SMiao Xie * @reason: reason why some writeback work was initiated 262510ee27a0SMiao Xie * 26268264c321SRakesh Pandit * Invoke __writeback_inodes_sb_nr if no writeback is currently underway. 262710ee27a0SMiao Xie */ 26288264c321SRakesh Pandit void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 262910ee27a0SMiao Xie { 26308264c321SRakesh Pandit if (!down_read_trylock(&sb->s_umount)) 26318264c321SRakesh Pandit return; 26328264c321SRakesh Pandit 26338264c321SRakesh Pandit __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true); 26348264c321SRakesh Pandit up_read(&sb->s_umount); 263510ee27a0SMiao Xie } 263610ee27a0SMiao Xie EXPORT_SYMBOL(try_to_writeback_inodes_sb); 26373259f8beSChris Mason 26383259f8beSChris Mason /** 2639d8a8559cSJens Axboe * sync_inodes_sb - sync sb inode pages 2640d8a8559cSJens Axboe * @sb: the superblock 2641d8a8559cSJens Axboe * 2642d8a8559cSJens Axboe * This function writes and waits on any dirty inode belonging to this 26430dc83bd3SJan Kara * super_block. 2644d8a8559cSJens Axboe */ 26450dc83bd3SJan Kara void sync_inodes_sb(struct super_block *sb) 2646d8a8559cSJens Axboe { 26475b9cce4cSTejun Heo struct backing_dev_info *bdi = sb->s_bdi; 26485b9cce4cSTejun Heo DEFINE_WB_COMPLETION(done, bdi); 264983ba7b07SChristoph Hellwig struct wb_writeback_work work = { 26503c4d7165SChristoph Hellwig .sb = sb, 26513c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_ALL, 26523c4d7165SChristoph Hellwig .nr_pages = LONG_MAX, 26533c4d7165SChristoph Hellwig .range_cyclic = 0, 265483ba7b07SChristoph Hellwig .done = &done, 26550e175a18SCurt Wohlgemuth .reason = WB_REASON_SYNC, 26567747bd4bSDave Chinner .for_sync = 1, 26573c4d7165SChristoph Hellwig }; 26583c4d7165SChristoph Hellwig 2659006a0973STejun Heo /* 2660006a0973STejun Heo * Can't skip on !bdi_has_dirty() because we should wait for !dirty 2661006a0973STejun Heo * inodes under writeback and I_DIRTY_TIME inodes ignored by 2662006a0973STejun Heo * bdi_has_dirty() need to be written out too. 2663006a0973STejun Heo */ 2664006a0973STejun Heo if (bdi == &noop_backing_dev_info) 26656eedc701SJan Kara return; 2666cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2667cf37e972SChristoph Hellwig 26687fc5854fSTejun Heo /* protect against inode wb switch, see inode_switch_wbs_work_fn() */ 26697fc5854fSTejun Heo bdi_down_write_wb_switch_rwsem(bdi); 2670db125360STejun Heo bdi_split_work_to_wbs(bdi, &work, false); 26715b9cce4cSTejun Heo wb_wait_for_completion(&done); 26727fc5854fSTejun Heo bdi_up_write_wb_switch_rwsem(bdi); 267383ba7b07SChristoph Hellwig 2674b6e51316SJens Axboe wait_sb_inodes(sb); 2675d8a8559cSJens Axboe } 2676d8a8559cSJens Axboe EXPORT_SYMBOL(sync_inodes_sb); 26771da177e4SLinus Torvalds 26781da177e4SLinus Torvalds /** 26791da177e4SLinus Torvalds * write_inode_now - write an inode to disk 26801da177e4SLinus Torvalds * @inode: inode to write to disk 26811da177e4SLinus Torvalds * @sync: whether the write should be synchronous or not 26821da177e4SLinus Torvalds * 26837f04c26dSAndrea Arcangeli * This function commits an inode to disk immediately if it is dirty. This is 26847f04c26dSAndrea Arcangeli * primarily needed by knfsd. 26857f04c26dSAndrea Arcangeli * 26867f04c26dSAndrea Arcangeli * The caller must either have a ref on the inode or must have set I_WILL_FREE. 26871da177e4SLinus Torvalds */ 26881da177e4SLinus Torvalds int write_inode_now(struct inode *inode, int sync) 26891da177e4SLinus Torvalds { 26901da177e4SLinus Torvalds struct writeback_control wbc = { 26911da177e4SLinus Torvalds .nr_to_write = LONG_MAX, 269218914b18SMike Galbraith .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, 2693111ebb6eSOGAWA Hirofumi .range_start = 0, 2694111ebb6eSOGAWA Hirofumi .range_end = LLONG_MAX, 26951da177e4SLinus Torvalds }; 26961da177e4SLinus Torvalds 2697f56753acSChristoph Hellwig if (!mapping_can_writeback(inode->i_mapping)) 269849364ce2SAndrew Morton wbc.nr_to_write = 0; 26991da177e4SLinus Torvalds 27001da177e4SLinus Torvalds might_sleep(); 2701aaf25593STejun Heo return writeback_single_inode(inode, &wbc); 27021da177e4SLinus Torvalds } 27031da177e4SLinus Torvalds EXPORT_SYMBOL(write_inode_now); 27041da177e4SLinus Torvalds 27051da177e4SLinus Torvalds /** 2706c691b9d9SAndrew Morton * sync_inode_metadata - write an inode to disk 2707c3765016SChristoph Hellwig * @inode: the inode to sync 2708c3765016SChristoph Hellwig * @wait: wait for I/O to complete. 2709c3765016SChristoph Hellwig * 2710c691b9d9SAndrew Morton * Write an inode to disk and adjust its dirty state after completion. 2711c3765016SChristoph Hellwig * 2712c3765016SChristoph Hellwig * Note: only writes the actual inode, no associated data or other metadata. 2713c3765016SChristoph Hellwig */ 2714c3765016SChristoph Hellwig int sync_inode_metadata(struct inode *inode, int wait) 2715c3765016SChristoph Hellwig { 2716c3765016SChristoph Hellwig struct writeback_control wbc = { 2717c3765016SChristoph Hellwig .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, 2718c3765016SChristoph Hellwig .nr_to_write = 0, /* metadata-only */ 2719c3765016SChristoph Hellwig }; 2720c3765016SChristoph Hellwig 27215662c967SJosef Bacik return writeback_single_inode(inode, &wbc); 2722c3765016SChristoph Hellwig } 2723c3765016SChristoph Hellwig EXPORT_SYMBOL(sync_inode_metadata); 2724