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); 37504edf02cSMatthew Wilcox struct page *page; 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 392d10c8095STejun Heo * to possibly dirty pages while PAGECACHE_TAG_WRITEBACK points to 393b93b0163SMatthew Wilcox * pages actually under writeback. 394d10c8095STejun Heo */ 39504edf02cSMatthew Wilcox xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_DIRTY) { 39604edf02cSMatthew Wilcox if (PageDirty(page)) { 3973e8f399dSNikolay Borisov dec_wb_stat(old_wb, WB_RECLAIMABLE); 3983e8f399dSNikolay Borisov inc_wb_stat(new_wb, WB_RECLAIMABLE); 399d10c8095STejun Heo } 400d10c8095STejun Heo } 401d10c8095STejun Heo 40204edf02cSMatthew Wilcox xas_set(&xas, 0); 40304edf02cSMatthew Wilcox xas_for_each_marked(&xas, page, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) { 404d10c8095STejun Heo WARN_ON_ONCE(!PageWriteback(page)); 4053e8f399dSNikolay Borisov dec_wb_stat(old_wb, WB_WRITEBACK); 4063e8f399dSNikolay Borisov inc_wb_stat(new_wb, WB_WRITEBACK); 407d10c8095STejun Heo } 408d10c8095STejun Heo 409633a2abbSJan Kara if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) { 410633a2abbSJan Kara atomic_dec(&old_wb->writeback_inodes); 411633a2abbSJan Kara atomic_inc(&new_wb->writeback_inodes); 412633a2abbSJan Kara } 413633a2abbSJan Kara 414d10c8095STejun Heo wb_get(new_wb); 415d10c8095STejun Heo 416d10c8095STejun Heo /* 417f3b6a6dfSRoman Gushchin * Transfer to @new_wb's IO list if necessary. If the @inode is dirty, 418f3b6a6dfSRoman Gushchin * the specific list @inode was on is ignored and the @inode is put on 419f3b6a6dfSRoman Gushchin * ->b_dirty which is always correct including from ->b_dirty_time. 420f3b6a6dfSRoman Gushchin * The transfer preserves @inode->dirtied_when ordering. If the @inode 421f3b6a6dfSRoman Gushchin * was clean, it means it was on the b_attached list, so move it onto 422f3b6a6dfSRoman Gushchin * the b_attached list of @new_wb. 423d10c8095STejun Heo */ 424c7f54084SDave Chinner if (!list_empty(&inode->i_io_list)) { 425f3b6a6dfSRoman Gushchin inode->i_wb = new_wb; 426f3b6a6dfSRoman Gushchin 427f3b6a6dfSRoman Gushchin if (inode->i_state & I_DIRTY_ALL) { 428d10c8095STejun Heo struct inode *pos; 429d10c8095STejun Heo 430c7f54084SDave Chinner list_for_each_entry(pos, &new_wb->b_dirty, i_io_list) 431d10c8095STejun Heo if (time_after_eq(inode->dirtied_when, 432d10c8095STejun Heo pos->dirtied_when)) 433d10c8095STejun Heo break; 434f3b6a6dfSRoman Gushchin inode_io_list_move_locked(inode, new_wb, 435f3b6a6dfSRoman Gushchin pos->i_io_list.prev); 436f3b6a6dfSRoman Gushchin } else { 437f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, new_wb); 438f3b6a6dfSRoman Gushchin } 439d10c8095STejun Heo } else { 440d10c8095STejun Heo inode->i_wb = new_wb; 441d10c8095STejun Heo } 442d10c8095STejun Heo 443d10c8095STejun Heo /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */ 444682aa8e1STejun Heo inode->i_wb_frn_winner = 0; 445682aa8e1STejun Heo inode->i_wb_frn_avg_time = 0; 446682aa8e1STejun Heo inode->i_wb_frn_history = 0; 447d10c8095STejun Heo switched = true; 448d10c8095STejun Heo skip_switch: 449682aa8e1STejun Heo /* 450682aa8e1STejun Heo * Paired with load_acquire in unlocked_inode_to_wb_begin() and 451682aa8e1STejun Heo * ensures that the new wb is visible if they see !I_WB_SWITCH. 452682aa8e1STejun Heo */ 453682aa8e1STejun Heo smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); 454682aa8e1STejun Heo 455b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 456682aa8e1STejun Heo spin_unlock(&inode->i_lock); 457d10c8095STejun Heo 458f5fbe6b7SRoman Gushchin return switched; 45972d4512eSRoman Gushchin } 460682aa8e1STejun Heo 46172d4512eSRoman Gushchin static void inode_switch_wbs_work_fn(struct work_struct *work) 46272d4512eSRoman Gushchin { 46372d4512eSRoman Gushchin struct inode_switch_wbs_context *isw = 46472d4512eSRoman Gushchin container_of(to_rcu_work(work), struct inode_switch_wbs_context, work); 465f5fbe6b7SRoman Gushchin struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]); 466f5fbe6b7SRoman Gushchin struct bdi_writeback *old_wb = isw->inodes[0]->i_wb; 467f5fbe6b7SRoman Gushchin struct bdi_writeback *new_wb = isw->new_wb; 468f5fbe6b7SRoman Gushchin unsigned long nr_switched = 0; 469f5fbe6b7SRoman Gushchin struct inode **inodep; 47072d4512eSRoman Gushchin 471f5fbe6b7SRoman Gushchin /* 472f5fbe6b7SRoman Gushchin * If @inode switches cgwb membership while sync_inodes_sb() is 473f5fbe6b7SRoman Gushchin * being issued, sync_inodes_sb() might miss it. Synchronize. 474f5fbe6b7SRoman Gushchin */ 475f5fbe6b7SRoman Gushchin down_read(&bdi->wb_switch_rwsem); 476f5fbe6b7SRoman Gushchin 477f5fbe6b7SRoman Gushchin /* 478f5fbe6b7SRoman Gushchin * By the time control reaches here, RCU grace period has passed 479f5fbe6b7SRoman Gushchin * since I_WB_SWITCH assertion and all wb stat update transactions 480f5fbe6b7SRoman Gushchin * between unlocked_inode_to_wb_begin/end() are guaranteed to be 481f5fbe6b7SRoman Gushchin * synchronizing against the i_pages lock. 482f5fbe6b7SRoman Gushchin * 483f5fbe6b7SRoman Gushchin * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock 484f5fbe6b7SRoman Gushchin * gives us exclusion against all wb related operations on @inode 485f5fbe6b7SRoman Gushchin * including IO list manipulations and stat updates. 486f5fbe6b7SRoman Gushchin */ 487f5fbe6b7SRoman Gushchin if (old_wb < new_wb) { 488f5fbe6b7SRoman Gushchin spin_lock(&old_wb->list_lock); 489f5fbe6b7SRoman Gushchin spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING); 490f5fbe6b7SRoman Gushchin } else { 491f5fbe6b7SRoman Gushchin spin_lock(&new_wb->list_lock); 492f5fbe6b7SRoman Gushchin spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING); 493f5fbe6b7SRoman Gushchin } 494f5fbe6b7SRoman Gushchin 495f5fbe6b7SRoman Gushchin for (inodep = isw->inodes; *inodep; inodep++) { 496f5fbe6b7SRoman Gushchin WARN_ON_ONCE((*inodep)->i_wb != old_wb); 497f5fbe6b7SRoman Gushchin if (inode_do_switch_wbs(*inodep, old_wb, new_wb)) 498f5fbe6b7SRoman Gushchin nr_switched++; 499f5fbe6b7SRoman Gushchin } 500f5fbe6b7SRoman Gushchin 501f5fbe6b7SRoman Gushchin spin_unlock(&new_wb->list_lock); 502f5fbe6b7SRoman Gushchin spin_unlock(&old_wb->list_lock); 503f5fbe6b7SRoman Gushchin 504f5fbe6b7SRoman Gushchin up_read(&bdi->wb_switch_rwsem); 505f5fbe6b7SRoman Gushchin 506f5fbe6b7SRoman Gushchin if (nr_switched) { 507f5fbe6b7SRoman Gushchin wb_wakeup(new_wb); 508f5fbe6b7SRoman Gushchin wb_put_many(old_wb, nr_switched); 509f5fbe6b7SRoman Gushchin } 510f5fbe6b7SRoman Gushchin 511f5fbe6b7SRoman Gushchin for (inodep = isw->inodes; *inodep; inodep++) 512f5fbe6b7SRoman Gushchin iput(*inodep); 513f5fbe6b7SRoman Gushchin wb_put(new_wb); 514682aa8e1STejun Heo kfree(isw); 515a1a0e23eSTejun Heo atomic_dec(&isw_nr_in_flight); 516682aa8e1STejun Heo } 517682aa8e1STejun Heo 518c22d70a1SRoman Gushchin static bool inode_prepare_wbs_switch(struct inode *inode, 519c22d70a1SRoman Gushchin struct bdi_writeback *new_wb) 520c22d70a1SRoman Gushchin { 521c22d70a1SRoman Gushchin /* 522c22d70a1SRoman Gushchin * Paired with smp_mb() in cgroup_writeback_umount(). 523c22d70a1SRoman Gushchin * isw_nr_in_flight must be increased before checking SB_ACTIVE and 524c22d70a1SRoman Gushchin * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0 525c22d70a1SRoman Gushchin * in cgroup_writeback_umount() and the isw_wq will be not flushed. 526c22d70a1SRoman Gushchin */ 527c22d70a1SRoman Gushchin smp_mb(); 528c22d70a1SRoman Gushchin 529593311e8SRoman Gushchin if (IS_DAX(inode)) 530593311e8SRoman Gushchin return false; 531593311e8SRoman Gushchin 532c22d70a1SRoman Gushchin /* while holding I_WB_SWITCH, no one else can update the association */ 533c22d70a1SRoman Gushchin spin_lock(&inode->i_lock); 534c22d70a1SRoman Gushchin if (!(inode->i_sb->s_flags & SB_ACTIVE) || 535c22d70a1SRoman Gushchin inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) || 536c22d70a1SRoman Gushchin inode_to_wb(inode) == new_wb) { 537c22d70a1SRoman Gushchin spin_unlock(&inode->i_lock); 538c22d70a1SRoman Gushchin return false; 539c22d70a1SRoman Gushchin } 540c22d70a1SRoman Gushchin inode->i_state |= I_WB_SWITCH; 541c22d70a1SRoman Gushchin __iget(inode); 542c22d70a1SRoman Gushchin spin_unlock(&inode->i_lock); 543c22d70a1SRoman Gushchin 544c22d70a1SRoman Gushchin return true; 545c22d70a1SRoman Gushchin } 546c22d70a1SRoman Gushchin 547682aa8e1STejun Heo /** 548682aa8e1STejun Heo * inode_switch_wbs - change the wb association of an inode 549682aa8e1STejun Heo * @inode: target inode 550682aa8e1STejun Heo * @new_wb_id: ID of the new wb 551682aa8e1STejun Heo * 552682aa8e1STejun Heo * Switch @inode's wb association to the wb identified by @new_wb_id. The 553682aa8e1STejun Heo * switching is performed asynchronously and may fail silently. 554682aa8e1STejun Heo */ 555682aa8e1STejun Heo static void inode_switch_wbs(struct inode *inode, int new_wb_id) 556682aa8e1STejun Heo { 557682aa8e1STejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 558682aa8e1STejun Heo struct cgroup_subsys_state *memcg_css; 559682aa8e1STejun Heo struct inode_switch_wbs_context *isw; 560682aa8e1STejun Heo 561682aa8e1STejun Heo /* noop if seems to be already in progress */ 562682aa8e1STejun Heo if (inode->i_state & I_WB_SWITCH) 563682aa8e1STejun Heo return; 564682aa8e1STejun Heo 5656444f47eSTejun Heo /* avoid queueing a new switch if too many are already in flight */ 5666444f47eSTejun Heo if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT) 5677fc5854fSTejun Heo return; 5687fc5854fSTejun Heo 569*98b160c8SLen Baker isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC); 570682aa8e1STejun Heo if (!isw) 5716444f47eSTejun Heo return; 572682aa8e1STejun Heo 5738826ee4fSRoman Gushchin atomic_inc(&isw_nr_in_flight); 5748826ee4fSRoman Gushchin 575682aa8e1STejun Heo /* find and pin the new wb */ 576682aa8e1STejun Heo rcu_read_lock(); 577682aa8e1STejun Heo memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); 5788b0ed844SMuchun Song if (memcg_css && !css_tryget(memcg_css)) 5798b0ed844SMuchun Song memcg_css = NULL; 580682aa8e1STejun Heo rcu_read_unlock(); 5818b0ed844SMuchun Song if (!memcg_css) 5828b0ed844SMuchun Song goto out_free; 5838b0ed844SMuchun Song 5848b0ed844SMuchun Song isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 5858b0ed844SMuchun Song css_put(memcg_css); 586682aa8e1STejun Heo if (!isw->new_wb) 587682aa8e1STejun Heo goto out_free; 588682aa8e1STejun Heo 589c22d70a1SRoman Gushchin if (!inode_prepare_wbs_switch(inode, isw->new_wb)) 590a1a0e23eSTejun Heo goto out_free; 591682aa8e1STejun Heo 592f5fbe6b7SRoman Gushchin isw->inodes[0] = inode; 593682aa8e1STejun Heo 594682aa8e1STejun Heo /* 595682aa8e1STejun Heo * In addition to synchronizing among switchers, I_WB_SWITCH tells 596b93b0163SMatthew Wilcox * the RCU protected stat update paths to grab the i_page 597b93b0163SMatthew Wilcox * lock so that stat transfer can synchronize against them. 598682aa8e1STejun Heo * Let's continue after I_WB_SWITCH is guaranteed to be visible. 599682aa8e1STejun Heo */ 60029264d92SRoman Gushchin INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); 60129264d92SRoman Gushchin queue_rcu_work(isw_wq, &isw->work); 6026444f47eSTejun Heo return; 603682aa8e1STejun Heo 604682aa8e1STejun Heo out_free: 6058826ee4fSRoman Gushchin atomic_dec(&isw_nr_in_flight); 606682aa8e1STejun Heo if (isw->new_wb) 607682aa8e1STejun Heo wb_put(isw->new_wb); 608682aa8e1STejun Heo kfree(isw); 609682aa8e1STejun Heo } 610682aa8e1STejun Heo 61187e1d789STejun Heo /** 612c22d70a1SRoman Gushchin * cleanup_offline_cgwb - detach associated inodes 613c22d70a1SRoman Gushchin * @wb: target wb 614c22d70a1SRoman Gushchin * 615c22d70a1SRoman Gushchin * Switch all inodes attached to @wb to a nearest living ancestor's wb in order 616c22d70a1SRoman Gushchin * to eventually release the dying @wb. Returns %true if not all inodes were 617c22d70a1SRoman Gushchin * switched and the function has to be restarted. 618c22d70a1SRoman Gushchin */ 619c22d70a1SRoman Gushchin bool cleanup_offline_cgwb(struct bdi_writeback *wb) 620c22d70a1SRoman Gushchin { 621c22d70a1SRoman Gushchin struct cgroup_subsys_state *memcg_css; 622c22d70a1SRoman Gushchin struct inode_switch_wbs_context *isw; 623c22d70a1SRoman Gushchin struct inode *inode; 624c22d70a1SRoman Gushchin int nr; 625c22d70a1SRoman Gushchin bool restart = false; 626c22d70a1SRoman Gushchin 627*98b160c8SLen Baker isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW), 628*98b160c8SLen Baker GFP_KERNEL); 629c22d70a1SRoman Gushchin if (!isw) 630c22d70a1SRoman Gushchin return restart; 631c22d70a1SRoman Gushchin 632c22d70a1SRoman Gushchin atomic_inc(&isw_nr_in_flight); 633c22d70a1SRoman Gushchin 634c22d70a1SRoman Gushchin for (memcg_css = wb->memcg_css->parent; memcg_css; 635c22d70a1SRoman Gushchin memcg_css = memcg_css->parent) { 636c22d70a1SRoman Gushchin isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL); 637c22d70a1SRoman Gushchin if (isw->new_wb) 638c22d70a1SRoman Gushchin break; 639c22d70a1SRoman Gushchin } 640c22d70a1SRoman Gushchin if (unlikely(!isw->new_wb)) 641c22d70a1SRoman Gushchin isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */ 642c22d70a1SRoman Gushchin 643c22d70a1SRoman Gushchin nr = 0; 644c22d70a1SRoman Gushchin spin_lock(&wb->list_lock); 645c22d70a1SRoman Gushchin list_for_each_entry(inode, &wb->b_attached, i_io_list) { 646c22d70a1SRoman Gushchin if (!inode_prepare_wbs_switch(inode, isw->new_wb)) 647c22d70a1SRoman Gushchin continue; 648c22d70a1SRoman Gushchin 649c22d70a1SRoman Gushchin isw->inodes[nr++] = inode; 650c22d70a1SRoman Gushchin 651c22d70a1SRoman Gushchin if (nr >= WB_MAX_INODES_PER_ISW - 1) { 652c22d70a1SRoman Gushchin restart = true; 653c22d70a1SRoman Gushchin break; 654c22d70a1SRoman Gushchin } 655c22d70a1SRoman Gushchin } 656c22d70a1SRoman Gushchin spin_unlock(&wb->list_lock); 657c22d70a1SRoman Gushchin 658c22d70a1SRoman Gushchin /* no attached inodes? bail out */ 659c22d70a1SRoman Gushchin if (nr == 0) { 660c22d70a1SRoman Gushchin atomic_dec(&isw_nr_in_flight); 661c22d70a1SRoman Gushchin wb_put(isw->new_wb); 662c22d70a1SRoman Gushchin kfree(isw); 663c22d70a1SRoman Gushchin return restart; 664c22d70a1SRoman Gushchin } 665c22d70a1SRoman Gushchin 666c22d70a1SRoman Gushchin /* 667c22d70a1SRoman Gushchin * In addition to synchronizing among switchers, I_WB_SWITCH tells 668c22d70a1SRoman Gushchin * the RCU protected stat update paths to grab the i_page 669c22d70a1SRoman Gushchin * lock so that stat transfer can synchronize against them. 670c22d70a1SRoman Gushchin * Let's continue after I_WB_SWITCH is guaranteed to be visible. 671c22d70a1SRoman Gushchin */ 672c22d70a1SRoman Gushchin INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); 673c22d70a1SRoman Gushchin queue_rcu_work(isw_wq, &isw->work); 674c22d70a1SRoman Gushchin 675c22d70a1SRoman Gushchin return restart; 676c22d70a1SRoman Gushchin } 677c22d70a1SRoman Gushchin 678c22d70a1SRoman Gushchin /** 679b16b1debSTejun Heo * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it 680b16b1debSTejun Heo * @wbc: writeback_control of interest 681b16b1debSTejun Heo * @inode: target inode 682b16b1debSTejun Heo * 683b16b1debSTejun Heo * @inode is locked and about to be written back under the control of @wbc. 684b16b1debSTejun Heo * Record @inode's writeback context into @wbc and unlock the i_lock. On 685b16b1debSTejun Heo * writeback completion, wbc_detach_inode() should be called. This is used 686b16b1debSTejun Heo * to track the cgroup writeback context. 687b16b1debSTejun Heo */ 688b16b1debSTejun Heo void wbc_attach_and_unlock_inode(struct writeback_control *wbc, 689b16b1debSTejun Heo struct inode *inode) 690b16b1debSTejun Heo { 691dd73e4b7STejun Heo if (!inode_cgwb_enabled(inode)) { 692dd73e4b7STejun Heo spin_unlock(&inode->i_lock); 693dd73e4b7STejun Heo return; 694dd73e4b7STejun Heo } 695dd73e4b7STejun Heo 696b16b1debSTejun Heo wbc->wb = inode_to_wb(inode); 6972a814908STejun Heo wbc->inode = inode; 6982a814908STejun Heo 6992a814908STejun Heo wbc->wb_id = wbc->wb->memcg_css->id; 7002a814908STejun Heo wbc->wb_lcand_id = inode->i_wb_frn_winner; 7012a814908STejun Heo wbc->wb_tcand_id = 0; 7022a814908STejun Heo wbc->wb_bytes = 0; 7032a814908STejun Heo wbc->wb_lcand_bytes = 0; 7042a814908STejun Heo wbc->wb_tcand_bytes = 0; 7052a814908STejun Heo 706b16b1debSTejun Heo wb_get(wbc->wb); 707b16b1debSTejun Heo spin_unlock(&inode->i_lock); 708e8a7abf5STejun Heo 709e8a7abf5STejun Heo /* 71065de03e2STejun Heo * A dying wb indicates that either the blkcg associated with the 71165de03e2STejun Heo * memcg changed or the associated memcg is dying. In the first 71265de03e2STejun Heo * case, a replacement wb should already be available and we should 71365de03e2STejun Heo * refresh the wb immediately. In the second case, trying to 71465de03e2STejun Heo * refresh will keep failing. 715e8a7abf5STejun Heo */ 71665de03e2STejun Heo if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css))) 717e8a7abf5STejun Heo inode_switch_wbs(inode, wbc->wb_id); 718b16b1debSTejun Heo } 7199b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode); 720b16b1debSTejun Heo 721b16b1debSTejun Heo /** 7222a814908STejun Heo * wbc_detach_inode - disassociate wbc from inode and perform foreign detection 7232a814908STejun Heo * @wbc: writeback_control of the just finished writeback 724b16b1debSTejun Heo * 725b16b1debSTejun Heo * To be called after a writeback attempt of an inode finishes and undoes 726b16b1debSTejun Heo * wbc_attach_and_unlock_inode(). Can be called under any context. 7272a814908STejun Heo * 7282a814908STejun Heo * As concurrent write sharing of an inode is expected to be very rare and 7292a814908STejun Heo * memcg only tracks page ownership on first-use basis severely confining 7302a814908STejun Heo * the usefulness of such sharing, cgroup writeback tracks ownership 7312a814908STejun Heo * per-inode. While the support for concurrent write sharing of an inode 7322a814908STejun Heo * is deemed unnecessary, an inode being written to by different cgroups at 7332a814908STejun Heo * different points in time is a lot more common, and, more importantly, 7342a814908STejun Heo * charging only by first-use can too readily lead to grossly incorrect 7352a814908STejun Heo * behaviors (single foreign page can lead to gigabytes of writeback to be 7362a814908STejun Heo * incorrectly attributed). 7372a814908STejun Heo * 7382a814908STejun Heo * To resolve this issue, cgroup writeback detects the majority dirtier of 7392a814908STejun Heo * an inode and transfers the ownership to it. To avoid unnnecessary 7402a814908STejun Heo * oscillation, the detection mechanism keeps track of history and gives 7412a814908STejun Heo * out the switch verdict only if the foreign usage pattern is stable over 7422a814908STejun Heo * a certain amount of time and/or writeback attempts. 7432a814908STejun Heo * 7442a814908STejun Heo * On each writeback attempt, @wbc tries to detect the majority writer 7452a814908STejun Heo * using Boyer-Moore majority vote algorithm. In addition to the byte 7462a814908STejun Heo * count from the majority voting, it also counts the bytes written for the 7472a814908STejun Heo * current wb and the last round's winner wb (max of last round's current 7482a814908STejun Heo * wb, the winner from two rounds ago, and the last round's majority 7492a814908STejun Heo * candidate). Keeping track of the historical winner helps the algorithm 7502a814908STejun Heo * to semi-reliably detect the most active writer even when it's not the 7512a814908STejun Heo * absolute majority. 7522a814908STejun Heo * 7532a814908STejun Heo * Once the winner of the round is determined, whether the winner is 7542a814908STejun Heo * foreign or not and how much IO time the round consumed is recorded in 7552a814908STejun Heo * inode->i_wb_frn_history. If the amount of recorded foreign IO time is 7562a814908STejun Heo * over a certain threshold, the switch verdict is given. 757b16b1debSTejun Heo */ 758b16b1debSTejun Heo void wbc_detach_inode(struct writeback_control *wbc) 759b16b1debSTejun Heo { 7602a814908STejun Heo struct bdi_writeback *wb = wbc->wb; 7612a814908STejun Heo struct inode *inode = wbc->inode; 762dd73e4b7STejun Heo unsigned long avg_time, max_bytes, max_time; 763dd73e4b7STejun Heo u16 history; 7642a814908STejun Heo int max_id; 7652a814908STejun Heo 766dd73e4b7STejun Heo if (!wb) 767dd73e4b7STejun Heo return; 768dd73e4b7STejun Heo 769dd73e4b7STejun Heo history = inode->i_wb_frn_history; 770dd73e4b7STejun Heo avg_time = inode->i_wb_frn_avg_time; 771dd73e4b7STejun Heo 7722a814908STejun Heo /* pick the winner of this round */ 7732a814908STejun Heo if (wbc->wb_bytes >= wbc->wb_lcand_bytes && 7742a814908STejun Heo wbc->wb_bytes >= wbc->wb_tcand_bytes) { 7752a814908STejun Heo max_id = wbc->wb_id; 7762a814908STejun Heo max_bytes = wbc->wb_bytes; 7772a814908STejun Heo } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { 7782a814908STejun Heo max_id = wbc->wb_lcand_id; 7792a814908STejun Heo max_bytes = wbc->wb_lcand_bytes; 7802a814908STejun Heo } else { 7812a814908STejun Heo max_id = wbc->wb_tcand_id; 7822a814908STejun Heo max_bytes = wbc->wb_tcand_bytes; 7832a814908STejun Heo } 7842a814908STejun Heo 7852a814908STejun Heo /* 7862a814908STejun Heo * Calculate the amount of IO time the winner consumed and fold it 7872a814908STejun Heo * into the running average kept per inode. If the consumed IO 7882a814908STejun Heo * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for 7892a814908STejun Heo * deciding whether to switch or not. This is to prevent one-off 7902a814908STejun Heo * small dirtiers from skewing the verdict. 7912a814908STejun Heo */ 7922a814908STejun Heo max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, 7932a814908STejun Heo wb->avg_write_bandwidth); 7942a814908STejun Heo if (avg_time) 7952a814908STejun Heo avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - 7962a814908STejun Heo (avg_time >> WB_FRN_TIME_AVG_SHIFT); 7972a814908STejun Heo else 7982a814908STejun Heo avg_time = max_time; /* immediate catch up on first run */ 7992a814908STejun Heo 8002a814908STejun Heo if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { 8012a814908STejun Heo int slots; 8022a814908STejun Heo 8032a814908STejun Heo /* 8042a814908STejun Heo * The switch verdict is reached if foreign wb's consume 8052a814908STejun Heo * more than a certain proportion of IO time in a 8062a814908STejun Heo * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot 8072a814908STejun Heo * history mask where each bit represents one sixteenth of 8082a814908STejun Heo * the period. Determine the number of slots to shift into 8092a814908STejun Heo * history from @max_time. 8102a814908STejun Heo */ 8112a814908STejun Heo slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), 8122a814908STejun Heo (unsigned long)WB_FRN_HIST_MAX_SLOTS); 8132a814908STejun Heo history <<= slots; 8142a814908STejun Heo if (wbc->wb_id != max_id) 8152a814908STejun Heo history |= (1U << slots) - 1; 8162a814908STejun Heo 8173a8e9ac8STejun Heo if (history) 8183a8e9ac8STejun Heo trace_inode_foreign_history(inode, wbc, history); 8193a8e9ac8STejun Heo 8202a814908STejun Heo /* 8212a814908STejun Heo * Switch if the current wb isn't the consistent winner. 8222a814908STejun Heo * If there are multiple closely competing dirtiers, the 8232a814908STejun Heo * inode may switch across them repeatedly over time, which 8242a814908STejun Heo * is okay. The main goal is avoiding keeping an inode on 8252a814908STejun Heo * the wrong wb for an extended period of time. 8262a814908STejun Heo */ 827682aa8e1STejun Heo if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) 828682aa8e1STejun Heo inode_switch_wbs(inode, max_id); 8292a814908STejun Heo } 8302a814908STejun Heo 8312a814908STejun Heo /* 8322a814908STejun Heo * Multiple instances of this function may race to update the 8332a814908STejun Heo * following fields but we don't mind occassional inaccuracies. 8342a814908STejun Heo */ 8352a814908STejun Heo inode->i_wb_frn_winner = max_id; 8362a814908STejun Heo inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); 8372a814908STejun Heo inode->i_wb_frn_history = history; 8382a814908STejun Heo 839b16b1debSTejun Heo wb_put(wbc->wb); 840b16b1debSTejun Heo wbc->wb = NULL; 841b16b1debSTejun Heo } 8429b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_detach_inode); 843b16b1debSTejun Heo 844b16b1debSTejun Heo /** 84534e51a5eSTejun Heo * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership 8462a814908STejun Heo * @wbc: writeback_control of the writeback in progress 8472a814908STejun Heo * @page: page being written out 8482a814908STejun Heo * @bytes: number of bytes being written out 8492a814908STejun Heo * 8502a814908STejun Heo * @bytes from @page are about to written out during the writeback 8512a814908STejun Heo * controlled by @wbc. Keep the book for foreign inode detection. See 8522a814908STejun Heo * wbc_detach_inode(). 8532a814908STejun Heo */ 85434e51a5eSTejun Heo void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page, 8552a814908STejun Heo size_t bytes) 8562a814908STejun Heo { 85766311422STejun Heo struct cgroup_subsys_state *css; 8582a814908STejun Heo int id; 8592a814908STejun Heo 8602a814908STejun Heo /* 8612a814908STejun Heo * pageout() path doesn't attach @wbc to the inode being written 8622a814908STejun Heo * out. This is intentional as we don't want the function to block 8632a814908STejun Heo * behind a slow cgroup. Ultimately, we want pageout() to kick off 8642a814908STejun Heo * regular writeback instead of writing things out itself. 8652a814908STejun Heo */ 86627b36d8fSTejun Heo if (!wbc->wb || wbc->no_cgroup_owner) 8672a814908STejun Heo return; 8682a814908STejun Heo 86966311422STejun Heo css = mem_cgroup_css_from_page(page); 87066311422STejun Heo /* dead cgroups shouldn't contribute to inode ownership arbitration */ 87166311422STejun Heo if (!(css->flags & CSS_ONLINE)) 87266311422STejun Heo return; 87366311422STejun Heo 87466311422STejun Heo id = css->id; 8752a814908STejun Heo 8762a814908STejun Heo if (id == wbc->wb_id) { 8772a814908STejun Heo wbc->wb_bytes += bytes; 8782a814908STejun Heo return; 8792a814908STejun Heo } 8802a814908STejun Heo 8812a814908STejun Heo if (id == wbc->wb_lcand_id) 8822a814908STejun Heo wbc->wb_lcand_bytes += bytes; 8832a814908STejun Heo 8842a814908STejun Heo /* Boyer-Moore majority vote algorithm */ 8852a814908STejun Heo if (!wbc->wb_tcand_bytes) 8862a814908STejun Heo wbc->wb_tcand_id = id; 8872a814908STejun Heo if (id == wbc->wb_tcand_id) 8882a814908STejun Heo wbc->wb_tcand_bytes += bytes; 8892a814908STejun Heo else 8902a814908STejun Heo wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); 8912a814908STejun Heo } 89234e51a5eSTejun Heo EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner); 8932a814908STejun Heo 8942a814908STejun Heo /** 895703c2708STejun Heo * inode_congested - test whether an inode is congested 89660292bccSTejun Heo * @inode: inode to test for congestion (may be NULL) 897703c2708STejun Heo * @cong_bits: mask of WB_[a]sync_congested bits to test 898703c2708STejun Heo * 899703c2708STejun Heo * Tests whether @inode is congested. @cong_bits is the mask of congestion 900703c2708STejun Heo * bits to test and the return value is the mask of set bits. 901703c2708STejun Heo * 902703c2708STejun Heo * If cgroup writeback is enabled for @inode, the congestion state is 903703c2708STejun Heo * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg 904703c2708STejun Heo * associated with @inode is congested; otherwise, the root wb's congestion 905703c2708STejun Heo * state is used. 90660292bccSTejun Heo * 90760292bccSTejun Heo * @inode is allowed to be NULL as this function is often called on 90860292bccSTejun Heo * mapping->host which is NULL for the swapper space. 909703c2708STejun Heo */ 910703c2708STejun Heo int inode_congested(struct inode *inode, int cong_bits) 911703c2708STejun Heo { 9125cb8b824STejun Heo /* 9135cb8b824STejun Heo * Once set, ->i_wb never becomes NULL while the inode is alive. 9145cb8b824STejun Heo * Start transaction iff ->i_wb is visible. 9155cb8b824STejun Heo */ 916aaa2cacfSTejun Heo if (inode && inode_to_wb_is_valid(inode)) { 9175cb8b824STejun Heo struct bdi_writeback *wb; 9182e898e4cSGreg Thelen struct wb_lock_cookie lock_cookie = {}; 9192e898e4cSGreg Thelen bool congested; 9205cb8b824STejun Heo 9212e898e4cSGreg Thelen wb = unlocked_inode_to_wb_begin(inode, &lock_cookie); 9225cb8b824STejun Heo congested = wb_congested(wb, cong_bits); 9232e898e4cSGreg Thelen unlocked_inode_to_wb_end(inode, &lock_cookie); 9245cb8b824STejun Heo return congested; 925703c2708STejun Heo } 926703c2708STejun Heo 927703c2708STejun Heo return wb_congested(&inode_to_bdi(inode)->wb, cong_bits); 928703c2708STejun Heo } 929703c2708STejun Heo EXPORT_SYMBOL_GPL(inode_congested); 930703c2708STejun Heo 931f2b65121STejun Heo /** 932f2b65121STejun Heo * wb_split_bdi_pages - split nr_pages to write according to bandwidth 933f2b65121STejun Heo * @wb: target bdi_writeback to split @nr_pages to 934f2b65121STejun Heo * @nr_pages: number of pages to write for the whole bdi 935f2b65121STejun Heo * 936f2b65121STejun Heo * Split @wb's portion of @nr_pages according to @wb's write bandwidth in 937f2b65121STejun Heo * relation to the total write bandwidth of all wb's w/ dirty inodes on 938f2b65121STejun Heo * @wb->bdi. 939f2b65121STejun Heo */ 940f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 941f2b65121STejun Heo { 942f2b65121STejun Heo unsigned long this_bw = wb->avg_write_bandwidth; 943f2b65121STejun Heo unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); 944f2b65121STejun Heo 945f2b65121STejun Heo if (nr_pages == LONG_MAX) 946f2b65121STejun Heo return LONG_MAX; 947f2b65121STejun Heo 948f2b65121STejun Heo /* 949f2b65121STejun Heo * This may be called on clean wb's and proportional distribution 950f2b65121STejun Heo * may not make sense, just use the original @nr_pages in those 951f2b65121STejun Heo * cases. In general, we wanna err on the side of writing more. 952f2b65121STejun Heo */ 953f2b65121STejun Heo if (!tot_bw || this_bw >= tot_bw) 954f2b65121STejun Heo return nr_pages; 955f2b65121STejun Heo else 956f2b65121STejun Heo return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); 957f2b65121STejun Heo } 958f2b65121STejun Heo 959db125360STejun Heo /** 960db125360STejun Heo * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi 961db125360STejun Heo * @bdi: target backing_dev_info 962db125360STejun Heo * @base_work: wb_writeback_work to issue 963db125360STejun Heo * @skip_if_busy: skip wb's which already have writeback in progress 964db125360STejun Heo * 965db125360STejun Heo * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which 966db125360STejun Heo * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's 967db125360STejun Heo * distributed to the busy wbs according to each wb's proportion in the 968db125360STejun Heo * total active write bandwidth of @bdi. 969db125360STejun Heo */ 970db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 971db125360STejun Heo struct wb_writeback_work *base_work, 972db125360STejun Heo bool skip_if_busy) 973db125360STejun Heo { 974b817525aSTejun Heo struct bdi_writeback *last_wb = NULL; 975b33e18f6STejun Heo struct bdi_writeback *wb = list_entry(&bdi->wb_list, 976b817525aSTejun Heo struct bdi_writeback, bdi_node); 977db125360STejun Heo 978db125360STejun Heo might_sleep(); 979db125360STejun Heo restart: 980db125360STejun Heo rcu_read_lock(); 981b817525aSTejun Heo list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) { 9825b9cce4cSTejun Heo DEFINE_WB_COMPLETION(fallback_work_done, bdi); 9838a1270cdSTejun Heo struct wb_writeback_work fallback_work; 9848a1270cdSTejun Heo struct wb_writeback_work *work; 9858a1270cdSTejun Heo long nr_pages; 9868a1270cdSTejun Heo 987b817525aSTejun Heo if (last_wb) { 988b817525aSTejun Heo wb_put(last_wb); 989b817525aSTejun Heo last_wb = NULL; 990b817525aSTejun Heo } 991b817525aSTejun Heo 992006a0973STejun Heo /* SYNC_ALL writes out I_DIRTY_TIME too */ 993006a0973STejun Heo if (!wb_has_dirty_io(wb) && 994006a0973STejun Heo (base_work->sync_mode == WB_SYNC_NONE || 995006a0973STejun Heo list_empty(&wb->b_dirty_time))) 996006a0973STejun Heo continue; 997006a0973STejun Heo if (skip_if_busy && writeback_in_progress(wb)) 998db125360STejun Heo continue; 999db125360STejun Heo 10008a1270cdSTejun Heo nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages); 10018a1270cdSTejun Heo 10028a1270cdSTejun Heo work = kmalloc(sizeof(*work), GFP_ATOMIC); 10038a1270cdSTejun Heo if (work) { 10048a1270cdSTejun Heo *work = *base_work; 10058a1270cdSTejun Heo work->nr_pages = nr_pages; 10068a1270cdSTejun Heo work->auto_free = 1; 10078a1270cdSTejun Heo wb_queue_work(wb, work); 10088a1270cdSTejun Heo continue; 1009db125360STejun Heo } 10108a1270cdSTejun Heo 10118a1270cdSTejun Heo /* alloc failed, execute synchronously using on-stack fallback */ 10128a1270cdSTejun Heo work = &fallback_work; 10138a1270cdSTejun Heo *work = *base_work; 10148a1270cdSTejun Heo work->nr_pages = nr_pages; 10158a1270cdSTejun Heo work->auto_free = 0; 10168a1270cdSTejun Heo work->done = &fallback_work_done; 10178a1270cdSTejun Heo 10188a1270cdSTejun Heo wb_queue_work(wb, work); 10198a1270cdSTejun Heo 1020b817525aSTejun Heo /* 1021b817525aSTejun Heo * Pin @wb so that it stays on @bdi->wb_list. This allows 1022b817525aSTejun Heo * continuing iteration from @wb after dropping and 1023b817525aSTejun Heo * regrabbing rcu read lock. 1024b817525aSTejun Heo */ 1025b817525aSTejun Heo wb_get(wb); 1026b817525aSTejun Heo last_wb = wb; 1027b817525aSTejun Heo 1028db125360STejun Heo rcu_read_unlock(); 10295b9cce4cSTejun Heo wb_wait_for_completion(&fallback_work_done); 1030db125360STejun Heo goto restart; 1031db125360STejun Heo } 1032db125360STejun Heo rcu_read_unlock(); 1033b817525aSTejun Heo 1034b817525aSTejun Heo if (last_wb) 1035b817525aSTejun Heo wb_put(last_wb); 1036db125360STejun Heo } 1037db125360STejun Heo 1038a1a0e23eSTejun Heo /** 1039d62241c7STejun Heo * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs 1040d62241c7STejun Heo * @bdi_id: target bdi id 1041d62241c7STejun Heo * @memcg_id: target memcg css id 1042d62241c7STejun Heo * @reason: reason why some writeback work initiated 1043d62241c7STejun Heo * @done: target wb_completion 1044d62241c7STejun Heo * 1045d62241c7STejun Heo * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id 1046d62241c7STejun Heo * with the specified parameters. 1047d62241c7STejun Heo */ 10487490a2d2SShakeel Butt int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, 1049d62241c7STejun Heo enum wb_reason reason, struct wb_completion *done) 1050d62241c7STejun Heo { 1051d62241c7STejun Heo struct backing_dev_info *bdi; 1052d62241c7STejun Heo struct cgroup_subsys_state *memcg_css; 1053d62241c7STejun Heo struct bdi_writeback *wb; 1054d62241c7STejun Heo struct wb_writeback_work *work; 10557490a2d2SShakeel Butt unsigned long dirty; 1056d62241c7STejun Heo int ret; 1057d62241c7STejun Heo 1058d62241c7STejun Heo /* lookup bdi and memcg */ 1059d62241c7STejun Heo bdi = bdi_get_by_id(bdi_id); 1060d62241c7STejun Heo if (!bdi) 1061d62241c7STejun Heo return -ENOENT; 1062d62241c7STejun Heo 1063d62241c7STejun Heo rcu_read_lock(); 1064d62241c7STejun Heo memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys); 1065d62241c7STejun Heo if (memcg_css && !css_tryget(memcg_css)) 1066d62241c7STejun Heo memcg_css = NULL; 1067d62241c7STejun Heo rcu_read_unlock(); 1068d62241c7STejun Heo if (!memcg_css) { 1069d62241c7STejun Heo ret = -ENOENT; 1070d62241c7STejun Heo goto out_bdi_put; 1071d62241c7STejun Heo } 1072d62241c7STejun Heo 1073d62241c7STejun Heo /* 1074d62241c7STejun Heo * And find the associated wb. If the wb isn't there already 1075d62241c7STejun Heo * there's nothing to flush, don't create one. 1076d62241c7STejun Heo */ 1077d62241c7STejun Heo wb = wb_get_lookup(bdi, memcg_css); 1078d62241c7STejun Heo if (!wb) { 1079d62241c7STejun Heo ret = -ENOENT; 1080d62241c7STejun Heo goto out_css_put; 1081d62241c7STejun Heo } 1082d62241c7STejun Heo 1083d62241c7STejun Heo /* 10847490a2d2SShakeel Butt * The caller is attempting to write out most of 1085d62241c7STejun Heo * the currently dirty pages. Let's take the current dirty page 1086d62241c7STejun Heo * count and inflate it by 25% which should be large enough to 1087d62241c7STejun Heo * flush out most dirty pages while avoiding getting livelocked by 1088d62241c7STejun Heo * concurrent dirtiers. 10897490a2d2SShakeel Butt * 10907490a2d2SShakeel Butt * BTW the memcg stats are flushed periodically and this is best-effort 10917490a2d2SShakeel Butt * estimation, so some potential error is ok. 1092d62241c7STejun Heo */ 10937490a2d2SShakeel Butt dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY); 10947490a2d2SShakeel Butt dirty = dirty * 10 / 8; 1095d62241c7STejun Heo 1096d62241c7STejun Heo /* issue the writeback work */ 1097d62241c7STejun Heo work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN); 1098d62241c7STejun Heo if (work) { 10997490a2d2SShakeel Butt work->nr_pages = dirty; 1100d62241c7STejun Heo work->sync_mode = WB_SYNC_NONE; 1101d62241c7STejun Heo work->range_cyclic = 1; 1102d62241c7STejun Heo work->reason = reason; 1103d62241c7STejun Heo work->done = done; 1104d62241c7STejun Heo work->auto_free = 1; 1105d62241c7STejun Heo wb_queue_work(wb, work); 1106d62241c7STejun Heo ret = 0; 1107d62241c7STejun Heo } else { 1108d62241c7STejun Heo ret = -ENOMEM; 1109d62241c7STejun Heo } 1110d62241c7STejun Heo 1111d62241c7STejun Heo wb_put(wb); 1112d62241c7STejun Heo out_css_put: 1113d62241c7STejun Heo css_put(memcg_css); 1114d62241c7STejun Heo out_bdi_put: 1115d62241c7STejun Heo bdi_put(bdi); 1116d62241c7STejun Heo return ret; 1117d62241c7STejun Heo } 1118d62241c7STejun Heo 1119d62241c7STejun Heo /** 1120a1a0e23eSTejun Heo * cgroup_writeback_umount - flush inode wb switches for umount 1121a1a0e23eSTejun Heo * 1122a1a0e23eSTejun Heo * This function is called when a super_block is about to be destroyed and 1123a1a0e23eSTejun Heo * flushes in-flight inode wb switches. An inode wb switch goes through 1124a1a0e23eSTejun Heo * RCU and then workqueue, so the two need to be flushed in order to ensure 1125a1a0e23eSTejun Heo * that all previously scheduled switches are finished. As wb switches are 1126a1a0e23eSTejun Heo * rare occurrences and synchronize_rcu() can take a while, perform 1127a1a0e23eSTejun Heo * flushing iff wb switches are in flight. 1128a1a0e23eSTejun Heo */ 1129a1a0e23eSTejun Heo void cgroup_writeback_umount(void) 1130a1a0e23eSTejun Heo { 1131592fa002SRoman Gushchin /* 1132592fa002SRoman Gushchin * SB_ACTIVE should be reliably cleared before checking 1133592fa002SRoman Gushchin * isw_nr_in_flight, see generic_shutdown_super(). 1134592fa002SRoman Gushchin */ 1135592fa002SRoman Gushchin smp_mb(); 1136592fa002SRoman Gushchin 1137a1a0e23eSTejun Heo if (atomic_read(&isw_nr_in_flight)) { 1138ec084de9SJiufei Xue /* 1139ec084de9SJiufei Xue * Use rcu_barrier() to wait for all pending callbacks to 1140ec084de9SJiufei Xue * ensure that all in-flight wb switches are in the workqueue. 1141ec084de9SJiufei Xue */ 1142ec084de9SJiufei Xue rcu_barrier(); 1143a1a0e23eSTejun Heo flush_workqueue(isw_wq); 1144a1a0e23eSTejun Heo } 1145a1a0e23eSTejun Heo } 1146a1a0e23eSTejun Heo 1147a1a0e23eSTejun Heo static int __init cgroup_writeback_init(void) 1148a1a0e23eSTejun Heo { 1149a1a0e23eSTejun Heo isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0); 1150a1a0e23eSTejun Heo if (!isw_wq) 1151a1a0e23eSTejun Heo return -ENOMEM; 1152a1a0e23eSTejun Heo return 0; 1153a1a0e23eSTejun Heo } 1154a1a0e23eSTejun Heo fs_initcall(cgroup_writeback_init); 1155a1a0e23eSTejun Heo 1156f2b65121STejun Heo #else /* CONFIG_CGROUP_WRITEBACK */ 1157f2b65121STejun Heo 11587fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 11597fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 11607fc5854fSTejun Heo 1161f3b6a6dfSRoman Gushchin static void inode_cgwb_move_to_attached(struct inode *inode, 1162f3b6a6dfSRoman Gushchin struct bdi_writeback *wb) 1163f3b6a6dfSRoman Gushchin { 1164f3b6a6dfSRoman Gushchin assert_spin_locked(&wb->list_lock); 1165f3b6a6dfSRoman Gushchin assert_spin_locked(&inode->i_lock); 1166f3b6a6dfSRoman Gushchin 1167f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 1168f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 1169f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 1170f3b6a6dfSRoman Gushchin } 1171f3b6a6dfSRoman Gushchin 117287e1d789STejun Heo static struct bdi_writeback * 117387e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 117487e1d789STejun Heo __releases(&inode->i_lock) 117587e1d789STejun Heo __acquires(&wb->list_lock) 117687e1d789STejun Heo { 117787e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 117887e1d789STejun Heo 117987e1d789STejun Heo spin_unlock(&inode->i_lock); 118087e1d789STejun Heo spin_lock(&wb->list_lock); 118187e1d789STejun Heo return wb; 118287e1d789STejun Heo } 118387e1d789STejun Heo 118487e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 118587e1d789STejun Heo __acquires(&wb->list_lock) 118687e1d789STejun Heo { 118787e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 118887e1d789STejun Heo 118987e1d789STejun Heo spin_lock(&wb->list_lock); 119087e1d789STejun Heo return wb; 119187e1d789STejun Heo } 119287e1d789STejun Heo 1193f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 1194f2b65121STejun Heo { 1195f2b65121STejun Heo return nr_pages; 1196f2b65121STejun Heo } 1197f2b65121STejun Heo 1198db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 1199db125360STejun Heo struct wb_writeback_work *base_work, 1200db125360STejun Heo bool skip_if_busy) 1201db125360STejun Heo { 1202db125360STejun Heo might_sleep(); 1203db125360STejun Heo 1204006a0973STejun Heo if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) { 1205db125360STejun Heo base_work->auto_free = 0; 1206db125360STejun Heo wb_queue_work(&bdi->wb, base_work); 1207db125360STejun Heo } 1208db125360STejun Heo } 1209db125360STejun Heo 1210703c2708STejun Heo #endif /* CONFIG_CGROUP_WRITEBACK */ 1211703c2708STejun Heo 1212e8e8a0c6SJens Axboe /* 1213e8e8a0c6SJens Axboe * Add in the number of potentially dirty inodes, because each inode 1214e8e8a0c6SJens Axboe * write can dirty pagecache in the underlying blockdev. 1215e8e8a0c6SJens Axboe */ 1216e8e8a0c6SJens Axboe static unsigned long get_nr_dirty_pages(void) 1217e8e8a0c6SJens Axboe { 1218e8e8a0c6SJens Axboe return global_node_page_state(NR_FILE_DIRTY) + 1219e8e8a0c6SJens Axboe get_nr_dirty_inodes(); 1220e8e8a0c6SJens Axboe } 1221e8e8a0c6SJens Axboe 1222e8e8a0c6SJens Axboe static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason) 1223b6e51316SJens Axboe { 1224c00ddad3STejun Heo if (!wb_has_dirty_io(wb)) 1225c00ddad3STejun Heo return; 1226c00ddad3STejun Heo 1227c00ddad3STejun Heo /* 1228aac8d41cSJens Axboe * All callers of this function want to start writeback of all 1229aac8d41cSJens Axboe * dirty pages. Places like vmscan can call this at a very 1230aac8d41cSJens Axboe * high frequency, causing pointless allocations of tons of 1231aac8d41cSJens Axboe * work items and keeping the flusher threads busy retrieving 1232aac8d41cSJens Axboe * that work. Ensure that we only allow one of them pending and 123385009b4fSJens Axboe * inflight at the time. 1234aac8d41cSJens Axboe */ 123585009b4fSJens Axboe if (test_bit(WB_start_all, &wb->state) || 123685009b4fSJens Axboe test_and_set_bit(WB_start_all, &wb->state)) 1237aac8d41cSJens Axboe return; 1238aac8d41cSJens Axboe 123985009b4fSJens Axboe wb->start_all_reason = reason; 1240c00ddad3STejun Heo wb_wakeup(wb); 1241d3ddec76SWu Fengguang } 1242d3ddec76SWu Fengguang 1243c5444198SChristoph Hellwig /** 12449ecf4866STejun Heo * wb_start_background_writeback - start background writeback 12459ecf4866STejun Heo * @wb: bdi_writback to write from 1246c5444198SChristoph Hellwig * 1247c5444198SChristoph Hellwig * Description: 12486585027aSJan Kara * This makes sure WB_SYNC_NONE background writeback happens. When 12499ecf4866STejun Heo * this function returns, it is only guaranteed that for given wb 12506585027aSJan Kara * some IO is happening if we are over background dirty threshold. 12516585027aSJan Kara * Caller need not hold sb s_umount semaphore. 1252c5444198SChristoph Hellwig */ 12539ecf4866STejun Heo void wb_start_background_writeback(struct bdi_writeback *wb) 1254c5444198SChristoph Hellwig { 12556585027aSJan Kara /* 12566585027aSJan Kara * We just wake up the flusher thread. It will perform background 12576585027aSJan Kara * writeback as soon as there is no other work to do. 12586585027aSJan Kara */ 12595634cc2aSTejun Heo trace_writeback_wake_background(wb); 12609ecf4866STejun Heo wb_wakeup(wb); 12611da177e4SLinus Torvalds } 12621da177e4SLinus Torvalds 12631da177e4SLinus Torvalds /* 1264a66979abSDave Chinner * Remove the inode from the writeback list it is on. 1265a66979abSDave Chinner */ 1266c7f54084SDave Chinner void inode_io_list_del(struct inode *inode) 1267a66979abSDave Chinner { 126887e1d789STejun Heo struct bdi_writeback *wb; 1269a66979abSDave Chinner 127087e1d789STejun Heo wb = inode_to_wb_and_lock_list(inode); 1271b35250c0SJan Kara spin_lock(&inode->i_lock); 1272f3b6a6dfSRoman Gushchin 1273f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 1274f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 1275f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 1276f3b6a6dfSRoman Gushchin 1277b35250c0SJan Kara spin_unlock(&inode->i_lock); 127852ebea74STejun Heo spin_unlock(&wb->list_lock); 1279f758eeabSChristoph Hellwig } 12804301efa4SJan Kara EXPORT_SYMBOL(inode_io_list_del); 1281a66979abSDave Chinner 1282a66979abSDave Chinner /* 12836c60d2b5SDave Chinner * mark an inode as under writeback on the sb 12846c60d2b5SDave Chinner */ 12856c60d2b5SDave Chinner void sb_mark_inode_writeback(struct inode *inode) 12866c60d2b5SDave Chinner { 12876c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 12886c60d2b5SDave Chinner unsigned long flags; 12896c60d2b5SDave Chinner 12906c60d2b5SDave Chinner if (list_empty(&inode->i_wb_list)) { 12916c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 12929a46b04fSBrian Foster if (list_empty(&inode->i_wb_list)) { 12936c60d2b5SDave Chinner list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); 12949a46b04fSBrian Foster trace_sb_mark_inode_writeback(inode); 12959a46b04fSBrian Foster } 12966c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 12976c60d2b5SDave Chinner } 12986c60d2b5SDave Chinner } 12996c60d2b5SDave Chinner 13006c60d2b5SDave Chinner /* 13016c60d2b5SDave Chinner * clear an inode as under writeback on the sb 13026c60d2b5SDave Chinner */ 13036c60d2b5SDave Chinner void sb_clear_inode_writeback(struct inode *inode) 13046c60d2b5SDave Chinner { 13056c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 13066c60d2b5SDave Chinner unsigned long flags; 13076c60d2b5SDave Chinner 13086c60d2b5SDave Chinner if (!list_empty(&inode->i_wb_list)) { 13096c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 13109a46b04fSBrian Foster if (!list_empty(&inode->i_wb_list)) { 13116c60d2b5SDave Chinner list_del_init(&inode->i_wb_list); 13129a46b04fSBrian Foster trace_sb_clear_inode_writeback(inode); 13139a46b04fSBrian Foster } 13146c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 13156c60d2b5SDave Chinner } 13166c60d2b5SDave Chinner } 13176c60d2b5SDave Chinner 13186c60d2b5SDave Chinner /* 13196610a0bcSAndrew Morton * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 13206610a0bcSAndrew Morton * furthest end of its superblock's dirty-inode list. 13216610a0bcSAndrew Morton * 13226610a0bcSAndrew Morton * Before stamping the inode's ->dirtied_when, we check to see whether it is 132366f3b8e2SJens Axboe * already the most-recently-dirtied inode on the b_dirty list. If that is 13246610a0bcSAndrew Morton * the case then the inode must have been redirtied while it was being written 13256610a0bcSAndrew Morton * out and we don't reset its dirtied_when. 13266610a0bcSAndrew Morton */ 1327b35250c0SJan Kara static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb) 13286610a0bcSAndrew Morton { 1329b35250c0SJan Kara assert_spin_locked(&inode->i_lock); 1330b35250c0SJan Kara 133103ba3782SJens Axboe if (!list_empty(&wb->b_dirty)) { 133266f3b8e2SJens Axboe struct inode *tail; 13336610a0bcSAndrew Morton 13347ccf19a8SNick Piggin tail = wb_inode(wb->b_dirty.next); 133566f3b8e2SJens Axboe if (time_before(inode->dirtied_when, tail->dirtied_when)) 13366610a0bcSAndrew Morton inode->dirtied_when = jiffies; 13376610a0bcSAndrew Morton } 1338c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty); 13395afced3bSJan Kara inode->i_state &= ~I_SYNC_QUEUED; 13406610a0bcSAndrew Morton } 13416610a0bcSAndrew Morton 1342b35250c0SJan Kara static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) 1343b35250c0SJan Kara { 1344b35250c0SJan Kara spin_lock(&inode->i_lock); 1345b35250c0SJan Kara redirty_tail_locked(inode, wb); 1346b35250c0SJan Kara spin_unlock(&inode->i_lock); 1347b35250c0SJan Kara } 1348b35250c0SJan Kara 13496610a0bcSAndrew Morton /* 135066f3b8e2SJens Axboe * requeue inode for re-scanning after bdi->b_io list is exhausted. 1351c986d1e2SAndrew Morton */ 1352f758eeabSChristoph Hellwig static void requeue_io(struct inode *inode, struct bdi_writeback *wb) 1353c986d1e2SAndrew Morton { 1354c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_more_io); 1355c986d1e2SAndrew Morton } 1356c986d1e2SAndrew Morton 13571c0eeaf5SJoern Engel static void inode_sync_complete(struct inode *inode) 13581c0eeaf5SJoern Engel { 1359365b94aeSJan Kara inode->i_state &= ~I_SYNC; 13604eff96ddSJan Kara /* If inode is clean an unused, put it into LRU now... */ 13614eff96ddSJan Kara inode_add_lru(inode); 1362365b94aeSJan Kara /* Waiters must see I_SYNC cleared before being woken up */ 13631c0eeaf5SJoern Engel smp_mb(); 13641c0eeaf5SJoern Engel wake_up_bit(&inode->i_state, __I_SYNC); 13651c0eeaf5SJoern Engel } 13661c0eeaf5SJoern Engel 1367d2caa3c5SJeff Layton static bool inode_dirtied_after(struct inode *inode, unsigned long t) 1368d2caa3c5SJeff Layton { 1369d2caa3c5SJeff Layton bool ret = time_after(inode->dirtied_when, t); 1370d2caa3c5SJeff Layton #ifndef CONFIG_64BIT 1371d2caa3c5SJeff Layton /* 1372d2caa3c5SJeff Layton * For inodes being constantly redirtied, dirtied_when can get stuck. 1373d2caa3c5SJeff Layton * It _appears_ to be in the future, but is actually in distant past. 1374d2caa3c5SJeff Layton * This test is necessary to prevent such wrapped-around relative times 13755b0830cbSJens Axboe * from permanently stopping the whole bdi writeback. 1376d2caa3c5SJeff Layton */ 1377d2caa3c5SJeff Layton ret = ret && time_before_eq(inode->dirtied_when, jiffies); 1378d2caa3c5SJeff Layton #endif 1379d2caa3c5SJeff Layton return ret; 1380d2caa3c5SJeff Layton } 1381d2caa3c5SJeff Layton 13820ae45f63STheodore Ts'o #define EXPIRE_DIRTY_ATIME 0x0001 13830ae45f63STheodore Ts'o 1384c986d1e2SAndrew Morton /* 1385f9cae926SJan Kara * Move expired (dirtied before dirtied_before) dirty inodes from 1386697e6fedSJan Kara * @delaying_queue to @dispatch_queue. 13872c136579SFengguang Wu */ 1388e84d0a4fSWu Fengguang static int move_expired_inodes(struct list_head *delaying_queue, 13892c136579SFengguang Wu struct list_head *dispatch_queue, 13905fcd5750SJan Kara unsigned long dirtied_before) 13912c136579SFengguang Wu { 13925c03449dSShaohua Li LIST_HEAD(tmp); 13935c03449dSShaohua Li struct list_head *pos, *node; 1394cf137307SJens Axboe struct super_block *sb = NULL; 13955c03449dSShaohua Li struct inode *inode; 1396cf137307SJens Axboe int do_sb_sort = 0; 1397e84d0a4fSWu Fengguang int moved = 0; 13985c03449dSShaohua Li 13992c136579SFengguang Wu while (!list_empty(delaying_queue)) { 14007ccf19a8SNick Piggin inode = wb_inode(delaying_queue->prev); 1401f9cae926SJan Kara if (inode_dirtied_after(inode, dirtied_before)) 14022c136579SFengguang Wu break; 1403c7f54084SDave Chinner list_move(&inode->i_io_list, &tmp); 1404a8855990SJan Kara moved++; 14055afced3bSJan Kara spin_lock(&inode->i_lock); 14065afced3bSJan Kara inode->i_state |= I_SYNC_QUEUED; 14075afced3bSJan Kara spin_unlock(&inode->i_lock); 1408a8855990SJan Kara if (sb_is_blkdev_sb(inode->i_sb)) 1409a8855990SJan Kara continue; 1410cf137307SJens Axboe if (sb && sb != inode->i_sb) 1411cf137307SJens Axboe do_sb_sort = 1; 1412cf137307SJens Axboe sb = inode->i_sb; 14135c03449dSShaohua Li } 14145c03449dSShaohua Li 1415cf137307SJens Axboe /* just one sb in list, splice to dispatch_queue and we're done */ 1416cf137307SJens Axboe if (!do_sb_sort) { 1417cf137307SJens Axboe list_splice(&tmp, dispatch_queue); 1418e84d0a4fSWu Fengguang goto out; 1419cf137307SJens Axboe } 1420cf137307SJens Axboe 14215c03449dSShaohua Li /* Move inodes from one superblock together */ 14225c03449dSShaohua Li while (!list_empty(&tmp)) { 14237ccf19a8SNick Piggin sb = wb_inode(tmp.prev)->i_sb; 14245c03449dSShaohua Li list_for_each_prev_safe(pos, node, &tmp) { 14257ccf19a8SNick Piggin inode = wb_inode(pos); 14265c03449dSShaohua Li if (inode->i_sb == sb) 1427c7f54084SDave Chinner list_move(&inode->i_io_list, dispatch_queue); 14282c136579SFengguang Wu } 14292c136579SFengguang Wu } 1430e84d0a4fSWu Fengguang out: 1431e84d0a4fSWu Fengguang return moved; 14325c03449dSShaohua Li } 14332c136579SFengguang Wu 14342c136579SFengguang Wu /* 14352c136579SFengguang Wu * Queue all expired dirty inodes for io, eldest first. 14364ea879b9SWu Fengguang * Before 14374ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 14384ea879b9SWu Fengguang * =============> gf edc BA 14394ea879b9SWu Fengguang * After 14404ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 14414ea879b9SWu Fengguang * =============> g fBAedc 14424ea879b9SWu Fengguang * | 14434ea879b9SWu Fengguang * +--> dequeue for IO 14442c136579SFengguang Wu */ 1445f9cae926SJan Kara static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work, 1446f9cae926SJan Kara unsigned long dirtied_before) 14472c136579SFengguang Wu { 1448e84d0a4fSWu Fengguang int moved; 1449f9cae926SJan Kara unsigned long time_expire_jif = dirtied_before; 14500ae45f63STheodore Ts'o 1451f758eeabSChristoph Hellwig assert_spin_locked(&wb->list_lock); 14524ea879b9SWu Fengguang list_splice_init(&wb->b_more_io, &wb->b_io); 14535fcd5750SJan Kara moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before); 1454f9cae926SJan Kara if (!work->for_sync) 1455f9cae926SJan Kara time_expire_jif = jiffies - dirtytime_expire_interval * HZ; 14560ae45f63STheodore Ts'o moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, 14575fcd5750SJan Kara time_expire_jif); 1458d6c10f1fSTejun Heo if (moved) 1459d6c10f1fSTejun Heo wb_io_lists_populated(wb); 1460f9cae926SJan Kara trace_writeback_queue_io(wb, work, dirtied_before, moved); 146166f3b8e2SJens Axboe } 146266f3b8e2SJens Axboe 1463a9185b41SChristoph Hellwig static int write_inode(struct inode *inode, struct writeback_control *wbc) 146466f3b8e2SJens Axboe { 14659fb0a7daSTejun Heo int ret; 14669fb0a7daSTejun Heo 14679fb0a7daSTejun Heo if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { 14689fb0a7daSTejun Heo trace_writeback_write_inode_start(inode, wbc); 14699fb0a7daSTejun Heo ret = inode->i_sb->s_op->write_inode(inode, wbc); 14709fb0a7daSTejun Heo trace_writeback_write_inode(inode, wbc); 14719fb0a7daSTejun Heo return ret; 14729fb0a7daSTejun Heo } 147303ba3782SJens Axboe return 0; 147466f3b8e2SJens Axboe } 147508d8e974SFengguang Wu 14762c136579SFengguang Wu /* 1477169ebd90SJan Kara * Wait for writeback on an inode to complete. Called with i_lock held. 1478169ebd90SJan Kara * Caller must make sure inode cannot go away when we drop i_lock. 147901c03194SChristoph Hellwig */ 1480169ebd90SJan Kara static void __inode_wait_for_writeback(struct inode *inode) 1481169ebd90SJan Kara __releases(inode->i_lock) 1482169ebd90SJan Kara __acquires(inode->i_lock) 148301c03194SChristoph Hellwig { 148401c03194SChristoph Hellwig DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); 148501c03194SChristoph Hellwig wait_queue_head_t *wqh; 148601c03194SChristoph Hellwig 148701c03194SChristoph Hellwig wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 148858a9d3d8SRichard Kennedy while (inode->i_state & I_SYNC) { 1489250df6edSDave Chinner spin_unlock(&inode->i_lock); 149074316201SNeilBrown __wait_on_bit(wqh, &wq, bit_wait, 149174316201SNeilBrown TASK_UNINTERRUPTIBLE); 1492250df6edSDave Chinner spin_lock(&inode->i_lock); 149358a9d3d8SRichard Kennedy } 149401c03194SChristoph Hellwig } 149501c03194SChristoph Hellwig 149601c03194SChristoph Hellwig /* 1497169ebd90SJan Kara * Wait for writeback on an inode to complete. Caller must have inode pinned. 1498169ebd90SJan Kara */ 1499169ebd90SJan Kara void inode_wait_for_writeback(struct inode *inode) 1500169ebd90SJan Kara { 1501169ebd90SJan Kara spin_lock(&inode->i_lock); 1502169ebd90SJan Kara __inode_wait_for_writeback(inode); 1503169ebd90SJan Kara spin_unlock(&inode->i_lock); 1504169ebd90SJan Kara } 1505169ebd90SJan Kara 1506169ebd90SJan Kara /* 1507169ebd90SJan Kara * Sleep until I_SYNC is cleared. This function must be called with i_lock 1508169ebd90SJan Kara * held and drops it. It is aimed for callers not holding any inode reference 1509169ebd90SJan Kara * so once i_lock is dropped, inode can go away. 1510169ebd90SJan Kara */ 1511169ebd90SJan Kara static void inode_sleep_on_writeback(struct inode *inode) 1512169ebd90SJan Kara __releases(inode->i_lock) 1513169ebd90SJan Kara { 1514169ebd90SJan Kara DEFINE_WAIT(wait); 1515169ebd90SJan Kara wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1516169ebd90SJan Kara int sleep; 1517169ebd90SJan Kara 1518169ebd90SJan Kara prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1519169ebd90SJan Kara sleep = inode->i_state & I_SYNC; 1520169ebd90SJan Kara spin_unlock(&inode->i_lock); 1521169ebd90SJan Kara if (sleep) 1522169ebd90SJan Kara schedule(); 1523169ebd90SJan Kara finish_wait(wqh, &wait); 1524169ebd90SJan Kara } 1525169ebd90SJan Kara 1526169ebd90SJan Kara /* 1527ccb26b5aSJan Kara * Find proper writeback list for the inode depending on its current state and 1528ccb26b5aSJan Kara * possibly also change of its state while we were doing writeback. Here we 1529ccb26b5aSJan Kara * handle things such as livelock prevention or fairness of writeback among 1530ccb26b5aSJan Kara * inodes. This function can be called only by flusher thread - noone else 1531ccb26b5aSJan Kara * processes all inodes in writeback lists and requeueing inodes behind flusher 1532ccb26b5aSJan Kara * thread's back can have unexpected consequences. 1533ccb26b5aSJan Kara */ 1534ccb26b5aSJan Kara static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, 1535ccb26b5aSJan Kara struct writeback_control *wbc) 1536ccb26b5aSJan Kara { 1537ccb26b5aSJan Kara if (inode->i_state & I_FREEING) 1538ccb26b5aSJan Kara return; 1539ccb26b5aSJan Kara 1540ccb26b5aSJan Kara /* 1541ccb26b5aSJan Kara * Sync livelock prevention. Each inode is tagged and synced in one 1542ccb26b5aSJan Kara * shot. If still dirty, it will be redirty_tail()'ed below. Update 1543ccb26b5aSJan Kara * the dirty time to prevent enqueue and sync it again. 1544ccb26b5aSJan Kara */ 1545ccb26b5aSJan Kara if ((inode->i_state & I_DIRTY) && 1546ccb26b5aSJan Kara (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) 1547ccb26b5aSJan Kara inode->dirtied_when = jiffies; 1548ccb26b5aSJan Kara 15494f8ad655SJan Kara if (wbc->pages_skipped) { 15504f8ad655SJan Kara /* 15514f8ad655SJan Kara * writeback is not making progress due to locked 15524f8ad655SJan Kara * buffers. Skip this inode for now. 15534f8ad655SJan Kara */ 1554b35250c0SJan Kara redirty_tail_locked(inode, wb); 15554f8ad655SJan Kara return; 15564f8ad655SJan Kara } 15574f8ad655SJan Kara 1558ccb26b5aSJan Kara if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { 1559ccb26b5aSJan Kara /* 1560ccb26b5aSJan Kara * We didn't write back all the pages. nfs_writepages() 1561ccb26b5aSJan Kara * sometimes bales out without doing anything. 1562ccb26b5aSJan Kara */ 1563ccb26b5aSJan Kara if (wbc->nr_to_write <= 0) { 1564ccb26b5aSJan Kara /* Slice used up. Queue for next turn. */ 1565ccb26b5aSJan Kara requeue_io(inode, wb); 1566ccb26b5aSJan Kara } else { 1567ccb26b5aSJan Kara /* 1568ccb26b5aSJan Kara * Writeback blocked by something other than 1569ccb26b5aSJan Kara * congestion. Delay the inode for some time to 1570ccb26b5aSJan Kara * avoid spinning on the CPU (100% iowait) 1571ccb26b5aSJan Kara * retrying writeback of the dirty page/inode 1572ccb26b5aSJan Kara * that cannot be performed immediately. 1573ccb26b5aSJan Kara */ 1574b35250c0SJan Kara redirty_tail_locked(inode, wb); 1575ccb26b5aSJan Kara } 1576ccb26b5aSJan Kara } else if (inode->i_state & I_DIRTY) { 1577ccb26b5aSJan Kara /* 1578ccb26b5aSJan Kara * Filesystems can dirty the inode during writeback operations, 1579ccb26b5aSJan Kara * such as delayed allocation during submission or metadata 1580ccb26b5aSJan Kara * updates after data IO completion. 1581ccb26b5aSJan Kara */ 1582b35250c0SJan Kara redirty_tail_locked(inode, wb); 15830ae45f63STheodore Ts'o } else if (inode->i_state & I_DIRTY_TIME) { 1584a2f48706STheodore Ts'o inode->dirtied_when = jiffies; 1585c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty_time); 15865afced3bSJan Kara inode->i_state &= ~I_SYNC_QUEUED; 1587ccb26b5aSJan Kara } else { 1588ccb26b5aSJan Kara /* The inode is clean. Remove from writeback lists. */ 1589f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, wb); 1590ccb26b5aSJan Kara } 1591ccb26b5aSJan Kara } 1592ccb26b5aSJan Kara 1593ccb26b5aSJan Kara /* 1594da0c4c60SEric Biggers * Write out an inode and its dirty pages (or some of its dirty pages, depending 1595da0c4c60SEric Biggers * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state. 1596da0c4c60SEric Biggers * 1597da0c4c60SEric Biggers * This doesn't remove the inode from the writeback list it is on, except 1598da0c4c60SEric Biggers * potentially to move it from b_dirty_time to b_dirty due to timestamp 1599da0c4c60SEric Biggers * expiration. The caller is otherwise responsible for writeback list handling. 1600da0c4c60SEric Biggers * 1601da0c4c60SEric Biggers * The caller is also responsible for setting the I_SYNC flag beforehand and 1602da0c4c60SEric Biggers * calling inode_sync_complete() to clear it afterwards. 16031da177e4SLinus Torvalds */ 16041da177e4SLinus Torvalds static int 1605cd8ed2a4SYan Hong __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 16061da177e4SLinus Torvalds { 16071da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 1608251d6a47SWu Fengguang long nr_to_write = wbc->nr_to_write; 160901c03194SChristoph Hellwig unsigned dirty; 16101da177e4SLinus Torvalds int ret; 16111da177e4SLinus Torvalds 16124f8ad655SJan Kara WARN_ON(!(inode->i_state & I_SYNC)); 16131da177e4SLinus Torvalds 16149fb0a7daSTejun Heo trace_writeback_single_inode_start(inode, wbc, nr_to_write); 16159fb0a7daSTejun Heo 16161da177e4SLinus Torvalds ret = do_writepages(mapping, wbc); 16171da177e4SLinus Torvalds 161826821ed4SChristoph Hellwig /* 161926821ed4SChristoph Hellwig * Make sure to wait on the data before writing out the metadata. 162026821ed4SChristoph Hellwig * This is important for filesystems that modify metadata on data 16217747bd4bSDave Chinner * I/O completion. We don't do it for sync(2) writeback because it has a 16227747bd4bSDave Chinner * separate, external IO completion path and ->sync_fs for guaranteeing 16237747bd4bSDave Chinner * inode metadata is written back correctly. 162426821ed4SChristoph Hellwig */ 16257747bd4bSDave Chinner if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { 162626821ed4SChristoph Hellwig int err = filemap_fdatawait(mapping); 16271da177e4SLinus Torvalds if (ret == 0) 16281da177e4SLinus Torvalds ret = err; 16291da177e4SLinus Torvalds } 16301da177e4SLinus Torvalds 16315547e8aaSDmitry Monakhov /* 16321e249cb5SEric Biggers * If the inode has dirty timestamps and we need to write them, call 16331e249cb5SEric Biggers * mark_inode_dirty_sync() to notify the filesystem about it and to 16341e249cb5SEric Biggers * change I_DIRTY_TIME into I_DIRTY_SYNC. 16351e249cb5SEric Biggers */ 16361e249cb5SEric Biggers if ((inode->i_state & I_DIRTY_TIME) && 163783dc881dSEric Biggers (wbc->sync_mode == WB_SYNC_ALL || 16381e249cb5SEric Biggers time_after(jiffies, inode->dirtied_time_when + 16391e249cb5SEric Biggers dirtytime_expire_interval * HZ))) { 16401e249cb5SEric Biggers trace_writeback_lazytime(inode); 16411e249cb5SEric Biggers mark_inode_dirty_sync(inode); 16421e249cb5SEric Biggers } 16431e249cb5SEric Biggers 16441e249cb5SEric Biggers /* 1645da0c4c60SEric Biggers * Get and clear the dirty flags from i_state. This needs to be done 1646da0c4c60SEric Biggers * after calling writepages because some filesystems may redirty the 1647da0c4c60SEric Biggers * inode during writepages due to delalloc. It also needs to be done 1648da0c4c60SEric Biggers * after handling timestamp expiration, as that may dirty the inode too. 16495547e8aaSDmitry Monakhov */ 1650250df6edSDave Chinner spin_lock(&inode->i_lock); 16515547e8aaSDmitry Monakhov dirty = inode->i_state & I_DIRTY; 16520ae45f63STheodore Ts'o inode->i_state &= ~dirty; 16539c6ac78eSTejun Heo 16549c6ac78eSTejun Heo /* 16559c6ac78eSTejun Heo * Paired with smp_mb() in __mark_inode_dirty(). This allows 16569c6ac78eSTejun Heo * __mark_inode_dirty() to test i_state without grabbing i_lock - 16579c6ac78eSTejun Heo * either they see the I_DIRTY bits cleared or we see the dirtied 16589c6ac78eSTejun Heo * inode. 16599c6ac78eSTejun Heo * 16609c6ac78eSTejun Heo * I_DIRTY_PAGES is always cleared together above even if @mapping 16619c6ac78eSTejun Heo * still has dirty pages. The flag is reinstated after smp_mb() if 16629c6ac78eSTejun Heo * necessary. This guarantees that either __mark_inode_dirty() 16639c6ac78eSTejun Heo * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. 16649c6ac78eSTejun Heo */ 16659c6ac78eSTejun Heo smp_mb(); 16669c6ac78eSTejun Heo 16679c6ac78eSTejun Heo if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 16689c6ac78eSTejun Heo inode->i_state |= I_DIRTY_PAGES; 16699c6ac78eSTejun Heo 1670250df6edSDave Chinner spin_unlock(&inode->i_lock); 16719c6ac78eSTejun Heo 167226821ed4SChristoph Hellwig /* Don't write the inode if only I_DIRTY_PAGES was set */ 16730ae45f63STheodore Ts'o if (dirty & ~I_DIRTY_PAGES) { 1674a9185b41SChristoph Hellwig int err = write_inode(inode, wbc); 16751da177e4SLinus Torvalds if (ret == 0) 16761da177e4SLinus Torvalds ret = err; 16771da177e4SLinus Torvalds } 16784f8ad655SJan Kara trace_writeback_single_inode(inode, wbc, nr_to_write); 16794f8ad655SJan Kara return ret; 16804f8ad655SJan Kara } 16814f8ad655SJan Kara 16824f8ad655SJan Kara /* 1683da0c4c60SEric Biggers * Write out an inode's dirty data and metadata on-demand, i.e. separately from 1684da0c4c60SEric Biggers * the regular batched writeback done by the flusher threads in 1685da0c4c60SEric Biggers * writeback_sb_inodes(). @wbc controls various aspects of the write, such as 1686da0c4c60SEric Biggers * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE). 16874f8ad655SJan Kara * 1688da0c4c60SEric Biggers * To prevent the inode from going away, either the caller must have a reference 1689da0c4c60SEric Biggers * to the inode, or the inode must have I_WILL_FREE or I_FREEING set. 16904f8ad655SJan Kara */ 1691aaf25593STejun Heo static int writeback_single_inode(struct inode *inode, 16924f8ad655SJan Kara struct writeback_control *wbc) 16934f8ad655SJan Kara { 1694aaf25593STejun Heo struct bdi_writeback *wb; 16954f8ad655SJan Kara int ret = 0; 16964f8ad655SJan Kara 16974f8ad655SJan Kara spin_lock(&inode->i_lock); 16984f8ad655SJan Kara if (!atomic_read(&inode->i_count)) 16994f8ad655SJan Kara WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 17004f8ad655SJan Kara else 17014f8ad655SJan Kara WARN_ON(inode->i_state & I_WILL_FREE); 17024f8ad655SJan Kara 17034f8ad655SJan Kara if (inode->i_state & I_SYNC) { 1704da0c4c60SEric Biggers /* 1705da0c4c60SEric Biggers * Writeback is already running on the inode. For WB_SYNC_NONE, 1706da0c4c60SEric Biggers * that's enough and we can just return. For WB_SYNC_ALL, we 1707da0c4c60SEric Biggers * must wait for the existing writeback to complete, then do 1708da0c4c60SEric Biggers * writeback again if there's anything left. 1709da0c4c60SEric Biggers */ 17104f8ad655SJan Kara if (wbc->sync_mode != WB_SYNC_ALL) 17114f8ad655SJan Kara goto out; 1712169ebd90SJan Kara __inode_wait_for_writeback(inode); 17134f8ad655SJan Kara } 17144f8ad655SJan Kara WARN_ON(inode->i_state & I_SYNC); 17154f8ad655SJan Kara /* 1716da0c4c60SEric Biggers * If the inode is already fully clean, then there's nothing to do. 1717da0c4c60SEric Biggers * 1718da0c4c60SEric Biggers * For data-integrity syncs we also need to check whether any pages are 1719da0c4c60SEric Biggers * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If 1720da0c4c60SEric Biggers * there are any such pages, we'll need to wait for them. 17214f8ad655SJan Kara */ 17220ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL) && 1723f9b0e058SJan Kara (wbc->sync_mode != WB_SYNC_ALL || 1724f9b0e058SJan Kara !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) 17254f8ad655SJan Kara goto out; 17264f8ad655SJan Kara inode->i_state |= I_SYNC; 1727b16b1debSTejun Heo wbc_attach_and_unlock_inode(wbc, inode); 17284f8ad655SJan Kara 1729cd8ed2a4SYan Hong ret = __writeback_single_inode(inode, wbc); 17301da177e4SLinus Torvalds 1731b16b1debSTejun Heo wbc_detach_inode(wbc); 1732aaf25593STejun Heo 1733aaf25593STejun Heo wb = inode_to_wb_and_lock_list(inode); 1734250df6edSDave Chinner spin_lock(&inode->i_lock); 17354f8ad655SJan Kara /* 1736da0c4c60SEric Biggers * If the inode is now fully clean, then it can be safely removed from 1737da0c4c60SEric Biggers * its writeback list (if any). Otherwise the flusher threads are 1738da0c4c60SEric Biggers * responsible for the writeback lists. 17394f8ad655SJan Kara */ 17400ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1741f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, wb); 17424f8ad655SJan Kara spin_unlock(&wb->list_lock); 17431c0eeaf5SJoern Engel inode_sync_complete(inode); 17444f8ad655SJan Kara out: 17454f8ad655SJan Kara spin_unlock(&inode->i_lock); 17461da177e4SLinus Torvalds return ret; 17471da177e4SLinus Torvalds } 17481da177e4SLinus Torvalds 1749a88a341aSTejun Heo static long writeback_chunk_size(struct bdi_writeback *wb, 17501a12d8bdSWu Fengguang struct wb_writeback_work *work) 1751d46db3d5SWu Fengguang { 1752d46db3d5SWu Fengguang long pages; 1753d46db3d5SWu Fengguang 1754d46db3d5SWu Fengguang /* 1755d46db3d5SWu Fengguang * WB_SYNC_ALL mode does livelock avoidance by syncing dirty 1756d46db3d5SWu Fengguang * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX 1757d46db3d5SWu Fengguang * here avoids calling into writeback_inodes_wb() more than once. 1758d46db3d5SWu Fengguang * 1759d46db3d5SWu Fengguang * The intended call sequence for WB_SYNC_ALL writeback is: 1760d46db3d5SWu Fengguang * 1761d46db3d5SWu Fengguang * wb_writeback() 1762d46db3d5SWu Fengguang * writeback_sb_inodes() <== called only once 1763d46db3d5SWu Fengguang * write_cache_pages() <== called once for each inode 1764d46db3d5SWu Fengguang * (quickly) tag currently dirty pages 1765d46db3d5SWu Fengguang * (maybe slowly) sync all tagged pages 1766d46db3d5SWu Fengguang */ 1767d46db3d5SWu Fengguang if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) 1768d46db3d5SWu Fengguang pages = LONG_MAX; 17691a12d8bdSWu Fengguang else { 1770a88a341aSTejun Heo pages = min(wb->avg_write_bandwidth / 2, 1771dcc25ae7STejun Heo global_wb_domain.dirty_limit / DIRTY_SCOPE); 17721a12d8bdSWu Fengguang pages = min(pages, work->nr_pages); 17731a12d8bdSWu Fengguang pages = round_down(pages + MIN_WRITEBACK_PAGES, 17741a12d8bdSWu Fengguang MIN_WRITEBACK_PAGES); 17751a12d8bdSWu Fengguang } 1776d46db3d5SWu Fengguang 1777d46db3d5SWu Fengguang return pages; 1778d46db3d5SWu Fengguang } 1779d46db3d5SWu Fengguang 178003ba3782SJens Axboe /* 1781f11c9c5cSEdward Shishkin * Write a portion of b_io inodes which belong to @sb. 1782edadfb10SChristoph Hellwig * 1783d46db3d5SWu Fengguang * Return the number of pages and/or inodes written. 17840ba13fd1SLinus Torvalds * 17850ba13fd1SLinus Torvalds * NOTE! This is called with wb->list_lock held, and will 17860ba13fd1SLinus Torvalds * unlock and relock that for each inode it ends up doing 17870ba13fd1SLinus Torvalds * IO for. 1788f11c9c5cSEdward Shishkin */ 1789d46db3d5SWu Fengguang static long writeback_sb_inodes(struct super_block *sb, 1790d46db3d5SWu Fengguang struct bdi_writeback *wb, 1791d46db3d5SWu Fengguang struct wb_writeback_work *work) 179203ba3782SJens Axboe { 1793d46db3d5SWu Fengguang struct writeback_control wbc = { 1794d46db3d5SWu Fengguang .sync_mode = work->sync_mode, 1795d46db3d5SWu Fengguang .tagged_writepages = work->tagged_writepages, 1796d46db3d5SWu Fengguang .for_kupdate = work->for_kupdate, 1797d46db3d5SWu Fengguang .for_background = work->for_background, 17987747bd4bSDave Chinner .for_sync = work->for_sync, 1799d46db3d5SWu Fengguang .range_cyclic = work->range_cyclic, 1800d46db3d5SWu Fengguang .range_start = 0, 1801d46db3d5SWu Fengguang .range_end = LLONG_MAX, 1802d46db3d5SWu Fengguang }; 1803d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1804d46db3d5SWu Fengguang long write_chunk; 1805d46db3d5SWu Fengguang long wrote = 0; /* count both pages and inodes */ 1806d46db3d5SWu Fengguang 180703ba3782SJens Axboe while (!list_empty(&wb->b_io)) { 18087ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1809aaf25593STejun Heo struct bdi_writeback *tmp_wb; 1810edadfb10SChristoph Hellwig 1811edadfb10SChristoph Hellwig if (inode->i_sb != sb) { 1812d46db3d5SWu Fengguang if (work->sb) { 1813edadfb10SChristoph Hellwig /* 1814edadfb10SChristoph Hellwig * We only want to write back data for this 1815edadfb10SChristoph Hellwig * superblock, move all inodes not belonging 1816edadfb10SChristoph Hellwig * to it back onto the dirty list. 1817edadfb10SChristoph Hellwig */ 1818f758eeabSChristoph Hellwig redirty_tail(inode, wb); 181966f3b8e2SJens Axboe continue; 182066f3b8e2SJens Axboe } 1821edadfb10SChristoph Hellwig 1822edadfb10SChristoph Hellwig /* 1823edadfb10SChristoph Hellwig * The inode belongs to a different superblock. 1824edadfb10SChristoph Hellwig * Bounce back to the caller to unpin this and 1825edadfb10SChristoph Hellwig * pin the next superblock. 1826edadfb10SChristoph Hellwig */ 1827d46db3d5SWu Fengguang break; 1828edadfb10SChristoph Hellwig } 1829edadfb10SChristoph Hellwig 18309843b76aSChristoph Hellwig /* 1831331cbdeeSWanpeng Li * Don't bother with new inodes or inodes being freed, first 1832331cbdeeSWanpeng Li * kind does not need periodic writeout yet, and for the latter 18339843b76aSChristoph Hellwig * kind writeout is handled by the freer. 18349843b76aSChristoph Hellwig */ 1835250df6edSDave Chinner spin_lock(&inode->i_lock); 18369843b76aSChristoph Hellwig if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { 1837b35250c0SJan Kara redirty_tail_locked(inode, wb); 1838250df6edSDave Chinner spin_unlock(&inode->i_lock); 18397ef0d737SNick Piggin continue; 18407ef0d737SNick Piggin } 1841cc1676d9SJan Kara if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { 1842cc1676d9SJan Kara /* 1843cc1676d9SJan Kara * If this inode is locked for writeback and we are not 1844cc1676d9SJan Kara * doing writeback-for-data-integrity, move it to 1845cc1676d9SJan Kara * b_more_io so that writeback can proceed with the 1846cc1676d9SJan Kara * other inodes on s_io. 1847cc1676d9SJan Kara * 1848cc1676d9SJan Kara * We'll have another go at writing back this inode 1849cc1676d9SJan Kara * when we completed a full scan of b_io. 1850cc1676d9SJan Kara */ 1851cc1676d9SJan Kara spin_unlock(&inode->i_lock); 1852cc1676d9SJan Kara requeue_io(inode, wb); 1853cc1676d9SJan Kara trace_writeback_sb_inodes_requeue(inode); 1854cc1676d9SJan Kara continue; 1855cc1676d9SJan Kara } 1856f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 1857f0d07b7fSJan Kara 18584f8ad655SJan Kara /* 18594f8ad655SJan Kara * We already requeued the inode if it had I_SYNC set and we 18604f8ad655SJan Kara * are doing WB_SYNC_NONE writeback. So this catches only the 18614f8ad655SJan Kara * WB_SYNC_ALL case. 18624f8ad655SJan Kara */ 1863169ebd90SJan Kara if (inode->i_state & I_SYNC) { 1864169ebd90SJan Kara /* Wait for I_SYNC. This function drops i_lock... */ 1865169ebd90SJan Kara inode_sleep_on_writeback(inode); 1866169ebd90SJan Kara /* Inode may be gone, start again */ 1867ead188f9SJan Kara spin_lock(&wb->list_lock); 1868169ebd90SJan Kara continue; 1869169ebd90SJan Kara } 18704f8ad655SJan Kara inode->i_state |= I_SYNC; 1871b16b1debSTejun Heo wbc_attach_and_unlock_inode(&wbc, inode); 1872169ebd90SJan Kara 1873a88a341aSTejun Heo write_chunk = writeback_chunk_size(wb, work); 1874d46db3d5SWu Fengguang wbc.nr_to_write = write_chunk; 1875d46db3d5SWu Fengguang wbc.pages_skipped = 0; 1876250df6edSDave Chinner 1877169ebd90SJan Kara /* 1878169ebd90SJan Kara * We use I_SYNC to pin the inode in memory. While it is set 1879169ebd90SJan Kara * evict_inode() will wait so the inode cannot be freed. 1880169ebd90SJan Kara */ 1881cd8ed2a4SYan Hong __writeback_single_inode(inode, &wbc); 1882d46db3d5SWu Fengguang 1883b16b1debSTejun Heo wbc_detach_inode(&wbc); 1884d46db3d5SWu Fengguang work->nr_pages -= write_chunk - wbc.nr_to_write; 1885d46db3d5SWu Fengguang wrote += write_chunk - wbc.nr_to_write; 1886590dca3aSChris Mason 1887590dca3aSChris Mason if (need_resched()) { 1888590dca3aSChris Mason /* 1889590dca3aSChris Mason * We're trying to balance between building up a nice 1890590dca3aSChris Mason * long list of IOs to improve our merge rate, and 1891590dca3aSChris Mason * getting those IOs out quickly for anyone throttling 1892590dca3aSChris Mason * in balance_dirty_pages(). cond_resched() doesn't 1893590dca3aSChris Mason * unplug, so get our IOs out the door before we 1894590dca3aSChris Mason * give up the CPU. 1895590dca3aSChris Mason */ 1896590dca3aSChris Mason blk_flush_plug(current); 1897590dca3aSChris Mason cond_resched(); 1898590dca3aSChris Mason } 1899590dca3aSChris Mason 1900aaf25593STejun Heo /* 1901aaf25593STejun Heo * Requeue @inode if still dirty. Be careful as @inode may 1902aaf25593STejun Heo * have been switched to another wb in the meantime. 1903aaf25593STejun Heo */ 1904aaf25593STejun Heo tmp_wb = inode_to_wb_and_lock_list(inode); 19054f8ad655SJan Kara spin_lock(&inode->i_lock); 19060ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1907d46db3d5SWu Fengguang wrote++; 1908aaf25593STejun Heo requeue_inode(inode, tmp_wb, &wbc); 19094f8ad655SJan Kara inode_sync_complete(inode); 19100f1b1fd8SDave Chinner spin_unlock(&inode->i_lock); 1911590dca3aSChris Mason 1912aaf25593STejun Heo if (unlikely(tmp_wb != wb)) { 1913aaf25593STejun Heo spin_unlock(&tmp_wb->list_lock); 1914aaf25593STejun Heo spin_lock(&wb->list_lock); 1915aaf25593STejun Heo } 1916aaf25593STejun Heo 1917d46db3d5SWu Fengguang /* 1918d46db3d5SWu Fengguang * bail out to wb_writeback() often enough to check 1919d46db3d5SWu Fengguang * background threshold and other termination conditions. 1920d46db3d5SWu Fengguang */ 1921d46db3d5SWu Fengguang if (wrote) { 1922d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1923d46db3d5SWu Fengguang break; 1924d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1925d46db3d5SWu Fengguang break; 19261da177e4SLinus Torvalds } 19278bc3be27SFengguang Wu } 1928d46db3d5SWu Fengguang return wrote; 1929f11c9c5cSEdward Shishkin } 193038f21977SNick Piggin 1931d46db3d5SWu Fengguang static long __writeback_inodes_wb(struct bdi_writeback *wb, 1932d46db3d5SWu Fengguang struct wb_writeback_work *work) 1933f11c9c5cSEdward Shishkin { 1934d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1935d46db3d5SWu Fengguang long wrote = 0; 1936f11c9c5cSEdward Shishkin 1937f11c9c5cSEdward Shishkin while (!list_empty(&wb->b_io)) { 19387ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1939f11c9c5cSEdward Shishkin struct super_block *sb = inode->i_sb; 1940f11c9c5cSEdward Shishkin 1941eb6ef3dfSKonstantin Khlebnikov if (!trylock_super(sb)) { 19420e995816SWu Fengguang /* 1943eb6ef3dfSKonstantin Khlebnikov * trylock_super() may fail consistently due to 19440e995816SWu Fengguang * s_umount being grabbed by someone else. Don't use 19450e995816SWu Fengguang * requeue_io() to avoid busy retrying the inode/sb. 19460e995816SWu Fengguang */ 19470e995816SWu Fengguang redirty_tail(inode, wb); 1948d19de7edSChristoph Hellwig continue; 1949334132aeSChristoph Hellwig } 1950d46db3d5SWu Fengguang wrote += writeback_sb_inodes(sb, wb, work); 1951eb6ef3dfSKonstantin Khlebnikov up_read(&sb->s_umount); 1952f11c9c5cSEdward Shishkin 1953d46db3d5SWu Fengguang /* refer to the same tests at the end of writeback_sb_inodes */ 1954d46db3d5SWu Fengguang if (wrote) { 1955d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1956d46db3d5SWu Fengguang break; 1957d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1958f11c9c5cSEdward Shishkin break; 1959f11c9c5cSEdward Shishkin } 1960d46db3d5SWu Fengguang } 196166f3b8e2SJens Axboe /* Leave any unwritten inodes on b_io */ 1962d46db3d5SWu Fengguang return wrote; 196366f3b8e2SJens Axboe } 196466f3b8e2SJens Axboe 19657d9f073bSWanpeng Li static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, 19660e175a18SCurt Wohlgemuth enum wb_reason reason) 1967edadfb10SChristoph Hellwig { 1968d46db3d5SWu Fengguang struct wb_writeback_work work = { 1969d46db3d5SWu Fengguang .nr_pages = nr_pages, 1970d46db3d5SWu Fengguang .sync_mode = WB_SYNC_NONE, 1971d46db3d5SWu Fengguang .range_cyclic = 1, 19720e175a18SCurt Wohlgemuth .reason = reason, 1973d46db3d5SWu Fengguang }; 1974505a666eSLinus Torvalds struct blk_plug plug; 1975edadfb10SChristoph Hellwig 1976505a666eSLinus Torvalds blk_start_plug(&plug); 1977f758eeabSChristoph Hellwig spin_lock(&wb->list_lock); 1978424b351fSWu Fengguang if (list_empty(&wb->b_io)) 1979f9cae926SJan Kara queue_io(wb, &work, jiffies); 1980d46db3d5SWu Fengguang __writeback_inodes_wb(wb, &work); 1981f758eeabSChristoph Hellwig spin_unlock(&wb->list_lock); 1982505a666eSLinus Torvalds blk_finish_plug(&plug); 1983edadfb10SChristoph Hellwig 1984d46db3d5SWu Fengguang return nr_pages - work.nr_pages; 198566f3b8e2SJens Axboe } 198666f3b8e2SJens Axboe 198703ba3782SJens Axboe /* 198803ba3782SJens Axboe * Explicit flushing or periodic writeback of "old" data. 198903ba3782SJens Axboe * 199003ba3782SJens Axboe * Define "old": the first time one of an inode's pages is dirtied, we mark the 199103ba3782SJens Axboe * dirtying-time in the inode's address_space. So this periodic writeback code 199203ba3782SJens Axboe * just walks the superblock inode list, writing back any inodes which are 199303ba3782SJens Axboe * older than a specific point in time. 199403ba3782SJens Axboe * 199503ba3782SJens Axboe * Try to run once per dirty_writeback_interval. But if a writeback event 199603ba3782SJens Axboe * takes longer than a dirty_writeback_interval interval, then leave a 199703ba3782SJens Axboe * one-second gap. 199803ba3782SJens Axboe * 1999f9cae926SJan Kara * dirtied_before takes precedence over nr_to_write. So we'll only write back 200003ba3782SJens Axboe * all dirty pages if they are all attached to "old" mappings. 200103ba3782SJens Axboe */ 2002c4a77a6cSJens Axboe static long wb_writeback(struct bdi_writeback *wb, 200383ba7b07SChristoph Hellwig struct wb_writeback_work *work) 200403ba3782SJens Axboe { 2005d46db3d5SWu Fengguang long nr_pages = work->nr_pages; 2006f9cae926SJan Kara unsigned long dirtied_before = jiffies; 2007a5989bdcSJan Kara struct inode *inode; 2008d46db3d5SWu Fengguang long progress; 2009505a666eSLinus Torvalds struct blk_plug plug; 201003ba3782SJens Axboe 2011505a666eSLinus Torvalds blk_start_plug(&plug); 2012e8dfc305SWu Fengguang spin_lock(&wb->list_lock); 201303ba3782SJens Axboe for (;;) { 201403ba3782SJens Axboe /* 2015d3ddec76SWu Fengguang * Stop writeback when nr_pages has been consumed 201603ba3782SJens Axboe */ 201783ba7b07SChristoph Hellwig if (work->nr_pages <= 0) 201803ba3782SJens Axboe break; 201903ba3782SJens Axboe 202003ba3782SJens Axboe /* 2021aa373cf5SJan Kara * Background writeout and kupdate-style writeback may 2022aa373cf5SJan Kara * run forever. Stop them if there is other work to do 2023aa373cf5SJan Kara * so that e.g. sync can proceed. They'll be restarted 2024aa373cf5SJan Kara * after the other works are all done. 2025aa373cf5SJan Kara */ 2026aa373cf5SJan Kara if ((work->for_background || work->for_kupdate) && 2027f0054bb1STejun Heo !list_empty(&wb->work_list)) 2028aa373cf5SJan Kara break; 2029aa373cf5SJan Kara 2030aa373cf5SJan Kara /* 2031d3ddec76SWu Fengguang * For background writeout, stop when we are below the 2032d3ddec76SWu Fengguang * background dirty threshold 203303ba3782SJens Axboe */ 2034aa661bbeSTejun Heo if (work->for_background && !wb_over_bg_thresh(wb)) 203503ba3782SJens Axboe break; 203603ba3782SJens Axboe 20371bc36b64SJan Kara /* 20381bc36b64SJan Kara * Kupdate and background works are special and we want to 20391bc36b64SJan Kara * include all inodes that need writing. Livelock avoidance is 20401bc36b64SJan Kara * handled by these works yielding to any other work so we are 20411bc36b64SJan Kara * safe. 20421bc36b64SJan Kara */ 2043ba9aa839SWu Fengguang if (work->for_kupdate) { 2044f9cae926SJan Kara dirtied_before = jiffies - 2045ba9aa839SWu Fengguang msecs_to_jiffies(dirty_expire_interval * 10); 20461bc36b64SJan Kara } else if (work->for_background) 2047f9cae926SJan Kara dirtied_before = jiffies; 2048028c2dd1SDave Chinner 20495634cc2aSTejun Heo trace_writeback_start(wb, work); 2050e8dfc305SWu Fengguang if (list_empty(&wb->b_io)) 2051f9cae926SJan Kara queue_io(wb, work, dirtied_before); 205283ba7b07SChristoph Hellwig if (work->sb) 2053d46db3d5SWu Fengguang progress = writeback_sb_inodes(work->sb, wb, work); 2054edadfb10SChristoph Hellwig else 2055d46db3d5SWu Fengguang progress = __writeback_inodes_wb(wb, work); 20565634cc2aSTejun Heo trace_writeback_written(wb, work); 2057028c2dd1SDave Chinner 205803ba3782SJens Axboe /* 205971fd05a8SJens Axboe * Did we write something? Try for more 2060e6fb6da2SWu Fengguang * 2061e6fb6da2SWu Fengguang * Dirty inodes are moved to b_io for writeback in batches. 2062e6fb6da2SWu Fengguang * The completion of the current batch does not necessarily 2063e6fb6da2SWu Fengguang * mean the overall work is done. So we keep looping as long 2064e6fb6da2SWu Fengguang * as made some progress on cleaning pages or inodes. 206571fd05a8SJens Axboe */ 2066d46db3d5SWu Fengguang if (progress) 206703ba3782SJens Axboe continue; 2068a5989bdcSJan Kara /* 2069e6fb6da2SWu Fengguang * No more inodes for IO, bail 2070a5989bdcSJan Kara */ 2071b7a2441fSWu Fengguang if (list_empty(&wb->b_more_io)) 207203ba3782SJens Axboe break; 207303ba3782SJens Axboe /* 20748010c3b6SJens Axboe * Nothing written. Wait for some inode to 20758010c3b6SJens Axboe * become available for writeback. Otherwise 20768010c3b6SJens Axboe * we'll just busyloop. 207703ba3782SJens Axboe */ 20785634cc2aSTejun Heo trace_writeback_wait(wb, work); 207903ba3782SJens Axboe inode = wb_inode(wb->b_more_io.prev); 2080250df6edSDave Chinner spin_lock(&inode->i_lock); 2081f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 2082169ebd90SJan Kara /* This function drops i_lock... */ 2083169ebd90SJan Kara inode_sleep_on_writeback(inode); 2084f0d07b7fSJan Kara spin_lock(&wb->list_lock); 208503ba3782SJens Axboe } 2086e8dfc305SWu Fengguang spin_unlock(&wb->list_lock); 2087505a666eSLinus Torvalds blk_finish_plug(&plug); 208803ba3782SJens Axboe 2089d46db3d5SWu Fengguang return nr_pages - work->nr_pages; 209003ba3782SJens Axboe } 209103ba3782SJens Axboe 209203ba3782SJens Axboe /* 209383ba7b07SChristoph Hellwig * Return the next wb_writeback_work struct that hasn't been processed yet. 209403ba3782SJens Axboe */ 2095f0054bb1STejun Heo static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) 209603ba3782SJens Axboe { 209783ba7b07SChristoph Hellwig struct wb_writeback_work *work = NULL; 209803ba3782SJens Axboe 2099f0054bb1STejun Heo spin_lock_bh(&wb->work_lock); 2100f0054bb1STejun Heo if (!list_empty(&wb->work_list)) { 2101f0054bb1STejun Heo work = list_entry(wb->work_list.next, 210283ba7b07SChristoph Hellwig struct wb_writeback_work, list); 210383ba7b07SChristoph Hellwig list_del_init(&work->list); 210403ba3782SJens Axboe } 2105f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 210683ba7b07SChristoph Hellwig return work; 210703ba3782SJens Axboe } 210803ba3782SJens Axboe 21096585027aSJan Kara static long wb_check_background_flush(struct bdi_writeback *wb) 21106585027aSJan Kara { 2111aa661bbeSTejun Heo if (wb_over_bg_thresh(wb)) { 21126585027aSJan Kara 21136585027aSJan Kara struct wb_writeback_work work = { 21146585027aSJan Kara .nr_pages = LONG_MAX, 21156585027aSJan Kara .sync_mode = WB_SYNC_NONE, 21166585027aSJan Kara .for_background = 1, 21176585027aSJan Kara .range_cyclic = 1, 21180e175a18SCurt Wohlgemuth .reason = WB_REASON_BACKGROUND, 21196585027aSJan Kara }; 21206585027aSJan Kara 21216585027aSJan Kara return wb_writeback(wb, &work); 21226585027aSJan Kara } 21236585027aSJan Kara 21246585027aSJan Kara return 0; 21256585027aSJan Kara } 21266585027aSJan Kara 212703ba3782SJens Axboe static long wb_check_old_data_flush(struct bdi_writeback *wb) 212803ba3782SJens Axboe { 212903ba3782SJens Axboe unsigned long expired; 213003ba3782SJens Axboe long nr_pages; 213103ba3782SJens Axboe 213269b62d01SJens Axboe /* 213369b62d01SJens Axboe * When set to zero, disable periodic writeback 213469b62d01SJens Axboe */ 213569b62d01SJens Axboe if (!dirty_writeback_interval) 213669b62d01SJens Axboe return 0; 213769b62d01SJens Axboe 213803ba3782SJens Axboe expired = wb->last_old_flush + 213903ba3782SJens Axboe msecs_to_jiffies(dirty_writeback_interval * 10); 214003ba3782SJens Axboe if (time_before(jiffies, expired)) 214103ba3782SJens Axboe return 0; 214203ba3782SJens Axboe 214303ba3782SJens Axboe wb->last_old_flush = jiffies; 2144cdf01dd5SLinus Torvalds nr_pages = get_nr_dirty_pages(); 214503ba3782SJens Axboe 2146c4a77a6cSJens Axboe if (nr_pages) { 214783ba7b07SChristoph Hellwig struct wb_writeback_work work = { 2148c4a77a6cSJens Axboe .nr_pages = nr_pages, 2149c4a77a6cSJens Axboe .sync_mode = WB_SYNC_NONE, 2150c4a77a6cSJens Axboe .for_kupdate = 1, 2151c4a77a6cSJens Axboe .range_cyclic = 1, 21520e175a18SCurt Wohlgemuth .reason = WB_REASON_PERIODIC, 2153c4a77a6cSJens Axboe }; 2154c4a77a6cSJens Axboe 215583ba7b07SChristoph Hellwig return wb_writeback(wb, &work); 2156c4a77a6cSJens Axboe } 215703ba3782SJens Axboe 215803ba3782SJens Axboe return 0; 215903ba3782SJens Axboe } 216003ba3782SJens Axboe 216185009b4fSJens Axboe static long wb_check_start_all(struct bdi_writeback *wb) 216285009b4fSJens Axboe { 216385009b4fSJens Axboe long nr_pages; 216485009b4fSJens Axboe 216585009b4fSJens Axboe if (!test_bit(WB_start_all, &wb->state)) 216685009b4fSJens Axboe return 0; 216785009b4fSJens Axboe 216885009b4fSJens Axboe nr_pages = get_nr_dirty_pages(); 216985009b4fSJens Axboe if (nr_pages) { 217085009b4fSJens Axboe struct wb_writeback_work work = { 217185009b4fSJens Axboe .nr_pages = wb_split_bdi_pages(wb, nr_pages), 217285009b4fSJens Axboe .sync_mode = WB_SYNC_NONE, 217385009b4fSJens Axboe .range_cyclic = 1, 217485009b4fSJens Axboe .reason = wb->start_all_reason, 217585009b4fSJens Axboe }; 217685009b4fSJens Axboe 217785009b4fSJens Axboe nr_pages = wb_writeback(wb, &work); 217885009b4fSJens Axboe } 217985009b4fSJens Axboe 218085009b4fSJens Axboe clear_bit(WB_start_all, &wb->state); 218185009b4fSJens Axboe return nr_pages; 218285009b4fSJens Axboe } 218385009b4fSJens Axboe 218485009b4fSJens Axboe 218503ba3782SJens Axboe /* 218603ba3782SJens Axboe * Retrieve work items and do the writeback they describe 218703ba3782SJens Axboe */ 218825d130baSWanpeng Li static long wb_do_writeback(struct bdi_writeback *wb) 218903ba3782SJens Axboe { 219083ba7b07SChristoph Hellwig struct wb_writeback_work *work; 2191c4a77a6cSJens Axboe long wrote = 0; 219203ba3782SJens Axboe 21934452226eSTejun Heo set_bit(WB_writeback_running, &wb->state); 2194f0054bb1STejun Heo while ((work = get_next_work_item(wb)) != NULL) { 21955634cc2aSTejun Heo trace_writeback_exec(wb, work); 219683ba7b07SChristoph Hellwig wrote += wb_writeback(wb, work); 21974a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 219803ba3782SJens Axboe } 219903ba3782SJens Axboe 220003ba3782SJens Axboe /* 220185009b4fSJens Axboe * Check for a flush-everything request 220285009b4fSJens Axboe */ 220385009b4fSJens Axboe wrote += wb_check_start_all(wb); 220485009b4fSJens Axboe 220585009b4fSJens Axboe /* 220603ba3782SJens Axboe * Check for periodic writeback, kupdated() style 220703ba3782SJens Axboe */ 220803ba3782SJens Axboe wrote += wb_check_old_data_flush(wb); 22096585027aSJan Kara wrote += wb_check_background_flush(wb); 22104452226eSTejun Heo clear_bit(WB_writeback_running, &wb->state); 221103ba3782SJens Axboe 221203ba3782SJens Axboe return wrote; 221303ba3782SJens Axboe } 221403ba3782SJens Axboe 221503ba3782SJens Axboe /* 221603ba3782SJens Axboe * Handle writeback of dirty data for the device backed by this bdi. Also 2217839a8e86STejun Heo * reschedules periodically and does kupdated style flushing. 221803ba3782SJens Axboe */ 2219f0054bb1STejun Heo void wb_workfn(struct work_struct *work) 222003ba3782SJens Axboe { 2221839a8e86STejun Heo struct bdi_writeback *wb = container_of(to_delayed_work(work), 2222839a8e86STejun Heo struct bdi_writeback, dwork); 222303ba3782SJens Axboe long pages_written; 222403ba3782SJens Axboe 222568f23b89STheodore Ts'o set_worker_desc("flush-%s", bdi_dev_name(wb->bdi)); 2226766f9164SPeter Zijlstra current->flags |= PF_SWAPWRITE; 222703ba3782SJens Axboe 2228839a8e86STejun Heo if (likely(!current_is_workqueue_rescuer() || 22294452226eSTejun Heo !test_bit(WB_registered, &wb->state))) { 223003ba3782SJens Axboe /* 2231f0054bb1STejun Heo * The normal path. Keep writing back @wb until its 2232839a8e86STejun Heo * work_list is empty. Note that this path is also taken 2233f0054bb1STejun Heo * if @wb is shutting down even when we're running off the 2234839a8e86STejun Heo * rescuer as work_list needs to be drained. 223503ba3782SJens Axboe */ 2236839a8e86STejun Heo do { 223725d130baSWanpeng Li pages_written = wb_do_writeback(wb); 2238455b2864SDave Chinner trace_writeback_pages_written(pages_written); 2239f0054bb1STejun Heo } while (!list_empty(&wb->work_list)); 2240839a8e86STejun Heo } else { 2241253c34e9SArtem Bityutskiy /* 2242839a8e86STejun Heo * bdi_wq can't get enough workers and we're running off 2243839a8e86STejun Heo * the emergency worker. Don't hog it. Hopefully, 1024 is 2244839a8e86STejun Heo * enough for efficient IO. 2245253c34e9SArtem Bityutskiy */ 2246f0054bb1STejun Heo pages_written = writeback_inodes_wb(wb, 1024, 2247839a8e86STejun Heo WB_REASON_FORKER_THREAD); 2248839a8e86STejun Heo trace_writeback_pages_written(pages_written); 224903ba3782SJens Axboe } 225003ba3782SJens Axboe 2251f0054bb1STejun Heo if (!list_empty(&wb->work_list)) 2252b8b78495SJan Kara wb_wakeup(wb); 22536ca738d6SDerek Basehore else if (wb_has_dirty_io(wb) && dirty_writeback_interval) 2254f0054bb1STejun Heo wb_wakeup_delayed(wb); 2255455b2864SDave Chinner 2256839a8e86STejun Heo current->flags &= ~PF_SWAPWRITE; 225703ba3782SJens Axboe } 225803ba3782SJens Axboe 225903ba3782SJens Axboe /* 2260595043e5SJens Axboe * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero, 2261595043e5SJens Axboe * write back the whole world. 2262595043e5SJens Axboe */ 2263595043e5SJens Axboe static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2264e8e8a0c6SJens Axboe enum wb_reason reason) 2265595043e5SJens Axboe { 2266595043e5SJens Axboe struct bdi_writeback *wb; 2267595043e5SJens Axboe 2268595043e5SJens Axboe if (!bdi_has_dirty_io(bdi)) 2269595043e5SJens Axboe return; 2270595043e5SJens Axboe 2271595043e5SJens Axboe list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 2272e8e8a0c6SJens Axboe wb_start_writeback(wb, reason); 2273595043e5SJens Axboe } 2274595043e5SJens Axboe 2275595043e5SJens Axboe void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2276595043e5SJens Axboe enum wb_reason reason) 2277595043e5SJens Axboe { 2278595043e5SJens Axboe rcu_read_lock(); 2279e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2280595043e5SJens Axboe rcu_read_unlock(); 2281595043e5SJens Axboe } 2282595043e5SJens Axboe 2283595043e5SJens Axboe /* 22849ba4b2dfSJens Axboe * Wakeup the flusher threads to start writeback of all currently dirty pages 228503ba3782SJens Axboe */ 22869ba4b2dfSJens Axboe void wakeup_flusher_threads(enum wb_reason reason) 228703ba3782SJens Axboe { 2288b8c2f347SChristoph Hellwig struct backing_dev_info *bdi; 2289b8c2f347SChristoph Hellwig 229051350ea0SKonstantin Khlebnikov /* 229151350ea0SKonstantin Khlebnikov * If we are expecting writeback progress we must submit plugged IO. 229251350ea0SKonstantin Khlebnikov */ 229351350ea0SKonstantin Khlebnikov if (blk_needs_flush_plug(current)) 229451350ea0SKonstantin Khlebnikov blk_schedule_flush_plug(current); 229551350ea0SKonstantin Khlebnikov 2296b8c2f347SChristoph Hellwig rcu_read_lock(); 2297595043e5SJens Axboe list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) 2298e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2299b8c2f347SChristoph Hellwig rcu_read_unlock(); 230003ba3782SJens Axboe } 230103ba3782SJens Axboe 2302a2f48706STheodore Ts'o /* 2303a2f48706STheodore Ts'o * Wake up bdi's periodically to make sure dirtytime inodes gets 2304a2f48706STheodore Ts'o * written back periodically. We deliberately do *not* check the 2305a2f48706STheodore Ts'o * b_dirtytime list in wb_has_dirty_io(), since this would cause the 2306a2f48706STheodore Ts'o * kernel to be constantly waking up once there are any dirtytime 2307a2f48706STheodore Ts'o * inodes on the system. So instead we define a separate delayed work 2308a2f48706STheodore Ts'o * function which gets called much more rarely. (By default, only 2309a2f48706STheodore Ts'o * once every 12 hours.) 2310a2f48706STheodore Ts'o * 2311a2f48706STheodore Ts'o * If there is any other write activity going on in the file system, 2312a2f48706STheodore Ts'o * this function won't be necessary. But if the only thing that has 2313a2f48706STheodore Ts'o * happened on the file system is a dirtytime inode caused by an atime 2314a2f48706STheodore Ts'o * update, we need this infrastructure below to make sure that inode 2315a2f48706STheodore Ts'o * eventually gets pushed out to disk. 2316a2f48706STheodore Ts'o */ 2317a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w); 2318a2f48706STheodore Ts'o static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); 2319a2f48706STheodore Ts'o 2320a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w) 2321a2f48706STheodore Ts'o { 2322a2f48706STheodore Ts'o struct backing_dev_info *bdi; 2323a2f48706STheodore Ts'o 2324a2f48706STheodore Ts'o rcu_read_lock(); 2325a2f48706STheodore Ts'o list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 2326001fe6f6STejun Heo struct bdi_writeback *wb; 2327001fe6f6STejun Heo 2328b817525aSTejun Heo list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 23296fdf860fSTejun Heo if (!list_empty(&wb->b_dirty_time)) 23306fdf860fSTejun Heo wb_wakeup(wb); 2331a2f48706STheodore Ts'o } 2332a2f48706STheodore Ts'o rcu_read_unlock(); 2333a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2334a2f48706STheodore Ts'o } 2335a2f48706STheodore Ts'o 2336a2f48706STheodore Ts'o static int __init start_dirtytime_writeback(void) 2337a2f48706STheodore Ts'o { 2338a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2339a2f48706STheodore Ts'o return 0; 2340a2f48706STheodore Ts'o } 2341a2f48706STheodore Ts'o __initcall(start_dirtytime_writeback); 2342a2f48706STheodore Ts'o 23431efff914STheodore Ts'o int dirtytime_interval_handler(struct ctl_table *table, int write, 23449ca48e20STobias Klauser void *buffer, size_t *lenp, loff_t *ppos) 23451efff914STheodore Ts'o { 23461efff914STheodore Ts'o int ret; 23471efff914STheodore Ts'o 23481efff914STheodore Ts'o ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 23491efff914STheodore Ts'o if (ret == 0 && write) 23501efff914STheodore Ts'o mod_delayed_work(system_wq, &dirtytime_work, 0); 23511efff914STheodore Ts'o return ret; 23521efff914STheodore Ts'o } 23531efff914STheodore Ts'o 235403ba3782SJens Axboe /** 235535d14f27SEric Biggers * __mark_inode_dirty - internal function to mark an inode dirty 23560117d427SMauro Carvalho Chehab * 235703ba3782SJens Axboe * @inode: inode to mark 235835d14f27SEric Biggers * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of 235935d14f27SEric Biggers * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined 236035d14f27SEric Biggers * with I_DIRTY_PAGES. 23610117d427SMauro Carvalho Chehab * 236235d14f27SEric Biggers * Mark an inode as dirty. We notify the filesystem, then update the inode's 236335d14f27SEric Biggers * dirty flags. Then, if needed we add the inode to the appropriate dirty list. 236403ba3782SJens Axboe * 236535d14f27SEric Biggers * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync() 236635d14f27SEric Biggers * instead of calling this directly. 236703ba3782SJens Axboe * 236835d14f27SEric Biggers * CAREFUL! We only add the inode to the dirty list if it is hashed or if it 236935d14f27SEric Biggers * refers to a blockdev. Unhashed inodes will never be added to the dirty list 237035d14f27SEric Biggers * even if they are later hashed, as they will have been marked dirty already. 237103ba3782SJens Axboe * 237235d14f27SEric Biggers * In short, ensure you hash any inodes _before_ you start marking them dirty. 237303ba3782SJens Axboe * 237403ba3782SJens Axboe * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 237503ba3782SJens Axboe * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 237603ba3782SJens Axboe * the kernel-internal blockdev inode represents the dirtying time of the 237703ba3782SJens Axboe * blockdev's pages. This is why for I_DIRTY_PAGES we always use 237803ba3782SJens Axboe * page->mapping->host, so the page-dirtying time is recorded in the internal 237903ba3782SJens Axboe * blockdev inode. 238003ba3782SJens Axboe */ 238103ba3782SJens Axboe void __mark_inode_dirty(struct inode *inode, int flags) 238203ba3782SJens Axboe { 238303ba3782SJens Axboe struct super_block *sb = inode->i_sb; 238435d14f27SEric Biggers int dirtytime = 0; 23850ae45f63STheodore Ts'o 23860ae45f63STheodore Ts'o trace_writeback_mark_inode_dirty(inode, flags); 238703ba3782SJens Axboe 2388e2728c56SEric Biggers if (flags & I_DIRTY_INODE) { 238935d14f27SEric Biggers /* 239035d14f27SEric Biggers * Notify the filesystem about the inode being dirtied, so that 239135d14f27SEric Biggers * (if needed) it can update on-disk fields and journal the 239235d14f27SEric Biggers * inode. This is only needed when the inode itself is being 239335d14f27SEric Biggers * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not 239435d14f27SEric Biggers * for just I_DIRTY_PAGES or I_DIRTY_TIME. 239535d14f27SEric Biggers */ 23969fb0a7daSTejun Heo trace_writeback_dirty_inode_start(inode, flags); 239703ba3782SJens Axboe if (sb->s_op->dirty_inode) 2398a38ed483SEric Biggers sb->s_op->dirty_inode(inode, flags & I_DIRTY_INODE); 23999fb0a7daSTejun Heo trace_writeback_dirty_inode(inode, flags); 2400e2728c56SEric Biggers 240135d14f27SEric Biggers /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */ 24020ae45f63STheodore Ts'o flags &= ~I_DIRTY_TIME; 240335d14f27SEric Biggers } else { 240435d14f27SEric Biggers /* 240535d14f27SEric Biggers * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing. 240635d14f27SEric Biggers * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME 240735d14f27SEric Biggers * in one call to __mark_inode_dirty().) 240835d14f27SEric Biggers */ 24090ae45f63STheodore Ts'o dirtytime = flags & I_DIRTY_TIME; 241035d14f27SEric Biggers WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME); 241135d14f27SEric Biggers } 241203ba3782SJens Axboe 241303ba3782SJens Axboe /* 24149c6ac78eSTejun Heo * Paired with smp_mb() in __writeback_single_inode() for the 24159c6ac78eSTejun Heo * following lockless i_state test. See there for details. 241603ba3782SJens Axboe */ 241703ba3782SJens Axboe smp_mb(); 241803ba3782SJens Axboe 24190ae45f63STheodore Ts'o if (((inode->i_state & flags) == flags) || 24200ae45f63STheodore Ts'o (dirtytime && (inode->i_state & I_DIRTY_INODE))) 242103ba3782SJens Axboe return; 242203ba3782SJens Axboe 2423250df6edSDave Chinner spin_lock(&inode->i_lock); 24240ae45f63STheodore Ts'o if (dirtytime && (inode->i_state & I_DIRTY_INODE)) 24250ae45f63STheodore Ts'o goto out_unlock_inode; 242603ba3782SJens Axboe if ((inode->i_state & flags) != flags) { 242703ba3782SJens Axboe const int was_dirty = inode->i_state & I_DIRTY; 242803ba3782SJens Axboe 242952ebea74STejun Heo inode_attach_wb(inode, NULL); 243052ebea74STejun Heo 243135d14f27SEric Biggers /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */ 24320ae45f63STheodore Ts'o if (flags & I_DIRTY_INODE) 24330ae45f63STheodore Ts'o inode->i_state &= ~I_DIRTY_TIME; 243403ba3782SJens Axboe inode->i_state |= flags; 243503ba3782SJens Axboe 243603ba3782SJens Axboe /* 24375afced3bSJan Kara * If the inode is queued for writeback by flush worker, just 24385afced3bSJan Kara * update its dirty state. Once the flush worker is done with 24395afced3bSJan Kara * the inode it will place it on the appropriate superblock 24405afced3bSJan Kara * list, based upon its state. 244103ba3782SJens Axboe */ 24425afced3bSJan Kara if (inode->i_state & I_SYNC_QUEUED) 2443250df6edSDave Chinner goto out_unlock_inode; 244403ba3782SJens Axboe 244503ba3782SJens Axboe /* 244603ba3782SJens Axboe * Only add valid (hashed) inodes to the superblock's 244703ba3782SJens Axboe * dirty list. Add blockdev inodes as well. 244803ba3782SJens Axboe */ 244903ba3782SJens Axboe if (!S_ISBLK(inode->i_mode)) { 24501d3382cbSAl Viro if (inode_unhashed(inode)) 2451250df6edSDave Chinner goto out_unlock_inode; 245203ba3782SJens Axboe } 2453a4ffdde6SAl Viro if (inode->i_state & I_FREEING) 2454250df6edSDave Chinner goto out_unlock_inode; 245503ba3782SJens Axboe 245603ba3782SJens Axboe /* 245703ba3782SJens Axboe * If the inode was already on b_dirty/b_io/b_more_io, don't 245803ba3782SJens Axboe * reposition it (that would break b_dirty time-ordering). 245903ba3782SJens Axboe */ 246003ba3782SJens Axboe if (!was_dirty) { 246187e1d789STejun Heo struct bdi_writeback *wb; 2462d6c10f1fSTejun Heo struct list_head *dirty_list; 2463a66979abSDave Chinner bool wakeup_bdi = false; 2464500b067cSJens Axboe 246587e1d789STejun Heo wb = locked_inode_to_wb_and_lock_list(inode); 2466253c34e9SArtem Bityutskiy 246703ba3782SJens Axboe inode->dirtied_when = jiffies; 2468a2f48706STheodore Ts'o if (dirtytime) 2469a2f48706STheodore Ts'o inode->dirtied_time_when = jiffies; 2470d6c10f1fSTejun Heo 24710e11f644SChristoph Hellwig if (inode->i_state & I_DIRTY) 24720747259dSTejun Heo dirty_list = &wb->b_dirty; 2473a2f48706STheodore Ts'o else 24740747259dSTejun Heo dirty_list = &wb->b_dirty_time; 2475d6c10f1fSTejun Heo 2476c7f54084SDave Chinner wakeup_bdi = inode_io_list_move_locked(inode, wb, 2477d6c10f1fSTejun Heo dirty_list); 2478d6c10f1fSTejun Heo 24790747259dSTejun Heo spin_unlock(&wb->list_lock); 24800ae45f63STheodore Ts'o trace_writeback_dirty_inode_enqueue(inode); 2481253c34e9SArtem Bityutskiy 2482d6c10f1fSTejun Heo /* 2483d6c10f1fSTejun Heo * If this is the first dirty inode for this bdi, 2484d6c10f1fSTejun Heo * we have to wake-up the corresponding bdi thread 2485d6c10f1fSTejun Heo * to make sure background write-back happens 2486d6c10f1fSTejun Heo * later. 2487d6c10f1fSTejun Heo */ 2488f56753acSChristoph Hellwig if (wakeup_bdi && 2489f56753acSChristoph Hellwig (wb->bdi->capabilities & BDI_CAP_WRITEBACK)) 24900747259dSTejun Heo wb_wakeup_delayed(wb); 2491a66979abSDave Chinner return; 2492a66979abSDave Chinner } 2493a66979abSDave Chinner } 2494a66979abSDave Chinner out_unlock_inode: 2495a66979abSDave Chinner spin_unlock(&inode->i_lock); 249603ba3782SJens Axboe } 249703ba3782SJens Axboe EXPORT_SYMBOL(__mark_inode_dirty); 249803ba3782SJens Axboe 2499e97fedb9SDave Chinner /* 2500e97fedb9SDave Chinner * The @s_sync_lock is used to serialise concurrent sync operations 2501e97fedb9SDave Chinner * to avoid lock contention problems with concurrent wait_sb_inodes() calls. 2502e97fedb9SDave Chinner * Concurrent callers will block on the s_sync_lock rather than doing contending 2503e97fedb9SDave Chinner * walks. The queueing maintains sync(2) required behaviour as all the IO that 2504e97fedb9SDave Chinner * has been issued up to the time this function is enter is guaranteed to be 2505e97fedb9SDave Chinner * completed by the time we have gained the lock and waited for all IO that is 2506e97fedb9SDave Chinner * in progress regardless of the order callers are granted the lock. 2507e97fedb9SDave Chinner */ 2508b6e51316SJens Axboe static void wait_sb_inodes(struct super_block *sb) 250966f3b8e2SJens Axboe { 25106c60d2b5SDave Chinner LIST_HEAD(sync_list); 251138f21977SNick Piggin 251203ba3782SJens Axboe /* 251303ba3782SJens Axboe * We need to be protected against the filesystem going from 251403ba3782SJens Axboe * r/o to r/w or vice versa. 251503ba3782SJens Axboe */ 2516b6e51316SJens Axboe WARN_ON(!rwsem_is_locked(&sb->s_umount)); 251703ba3782SJens Axboe 2518e97fedb9SDave Chinner mutex_lock(&sb->s_sync_lock); 251966f3b8e2SJens Axboe 252038f21977SNick Piggin /* 25216c60d2b5SDave Chinner * Splice the writeback list onto a temporary list to avoid waiting on 25226c60d2b5SDave Chinner * inodes that have started writeback after this point. 25236c60d2b5SDave Chinner * 25246c60d2b5SDave Chinner * Use rcu_read_lock() to keep the inodes around until we have a 25256c60d2b5SDave Chinner * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as 25266c60d2b5SDave Chinner * the local list because inodes can be dropped from either by writeback 25276c60d2b5SDave Chinner * completion. 252838f21977SNick Piggin */ 25296c60d2b5SDave Chinner rcu_read_lock(); 25306c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 25316c60d2b5SDave Chinner list_splice_init(&sb->s_inodes_wb, &sync_list); 25326c60d2b5SDave Chinner 25336c60d2b5SDave Chinner /* 25346c60d2b5SDave Chinner * Data integrity sync. Must wait for all pages under writeback, because 25356c60d2b5SDave Chinner * there may have been pages dirtied before our sync call, but which had 25366c60d2b5SDave Chinner * writeout started before we write it out. In which case, the inode 25376c60d2b5SDave Chinner * may not be on the dirty list, but we still have to wait for that 25386c60d2b5SDave Chinner * writeout. 25396c60d2b5SDave Chinner */ 25406c60d2b5SDave Chinner while (!list_empty(&sync_list)) { 25416c60d2b5SDave Chinner struct inode *inode = list_first_entry(&sync_list, struct inode, 25426c60d2b5SDave Chinner i_wb_list); 2543250df6edSDave Chinner struct address_space *mapping = inode->i_mapping; 254438f21977SNick Piggin 25456c60d2b5SDave Chinner /* 25466c60d2b5SDave Chinner * Move each inode back to the wb list before we drop the lock 25476c60d2b5SDave Chinner * to preserve consistency between i_wb_list and the mapping 25486c60d2b5SDave Chinner * writeback tag. Writeback completion is responsible to remove 25496c60d2b5SDave Chinner * the inode from either list once the writeback tag is cleared. 25506c60d2b5SDave Chinner */ 25516c60d2b5SDave Chinner list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); 25526c60d2b5SDave Chinner 25536c60d2b5SDave Chinner /* 25546c60d2b5SDave Chinner * The mapping can appear untagged while still on-list since we 25556c60d2b5SDave Chinner * do not have the mapping lock. Skip it here, wb completion 25566c60d2b5SDave Chinner * will remove it. 25576c60d2b5SDave Chinner */ 25586c60d2b5SDave Chinner if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) 25596c60d2b5SDave Chinner continue; 25606c60d2b5SDave Chinner 25616c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 25626c60d2b5SDave Chinner 2563250df6edSDave Chinner spin_lock(&inode->i_lock); 25646c60d2b5SDave Chinner if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { 2565250df6edSDave Chinner spin_unlock(&inode->i_lock); 25666c60d2b5SDave Chinner 25676c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 256838f21977SNick Piggin continue; 2569250df6edSDave Chinner } 257038f21977SNick Piggin __iget(inode); 2571250df6edSDave Chinner spin_unlock(&inode->i_lock); 25726c60d2b5SDave Chinner rcu_read_unlock(); 257338f21977SNick Piggin 2574aa750fd7SJunichi Nomura /* 2575aa750fd7SJunichi Nomura * We keep the error status of individual mapping so that 2576aa750fd7SJunichi Nomura * applications can catch the writeback error using fsync(2). 2577aa750fd7SJunichi Nomura * See filemap_fdatawait_keep_errors() for details. 2578aa750fd7SJunichi Nomura */ 2579aa750fd7SJunichi Nomura filemap_fdatawait_keep_errors(mapping); 258038f21977SNick Piggin 258138f21977SNick Piggin cond_resched(); 258238f21977SNick Piggin 25836c60d2b5SDave Chinner iput(inode); 25846c60d2b5SDave Chinner 25856c60d2b5SDave Chinner rcu_read_lock(); 25866c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 258738f21977SNick Piggin } 25886c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 25896c60d2b5SDave Chinner rcu_read_unlock(); 2590e97fedb9SDave Chinner mutex_unlock(&sb->s_sync_lock); 259166f3b8e2SJens Axboe } 25921da177e4SLinus Torvalds 2593f30a7d0cSTejun Heo static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2594f30a7d0cSTejun Heo enum wb_reason reason, bool skip_if_busy) 25951da177e4SLinus Torvalds { 25965b9cce4cSTejun Heo struct backing_dev_info *bdi = sb->s_bdi; 25975b9cce4cSTejun Heo DEFINE_WB_COMPLETION(done, bdi); 259883ba7b07SChristoph Hellwig struct wb_writeback_work work = { 25993c4d7165SChristoph Hellwig .sb = sb, 26003c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_NONE, 26016e6938b6SWu Fengguang .tagged_writepages = 1, 260283ba7b07SChristoph Hellwig .done = &done, 26033259f8beSChris Mason .nr_pages = nr, 26040e175a18SCurt Wohlgemuth .reason = reason, 26053c4d7165SChristoph Hellwig }; 26060e3c9a22SJens Axboe 2607e7972912STejun Heo if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 26086eedc701SJan Kara return; 2609cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2610f30a7d0cSTejun Heo 2611db125360STejun Heo bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); 26125b9cce4cSTejun Heo wb_wait_for_completion(&done); 26131da177e4SLinus Torvalds } 2614f30a7d0cSTejun Heo 2615f30a7d0cSTejun Heo /** 2616f30a7d0cSTejun Heo * writeback_inodes_sb_nr - writeback dirty inodes from given super_block 2617f30a7d0cSTejun Heo * @sb: the superblock 2618f30a7d0cSTejun Heo * @nr: the number of pages to write 2619f30a7d0cSTejun Heo * @reason: reason why some writeback work initiated 2620f30a7d0cSTejun Heo * 2621f30a7d0cSTejun Heo * Start writeback on some inodes on this super_block. No guarantees are made 2622f30a7d0cSTejun Heo * on how many (if any) will be written, and this function does not wait 2623f30a7d0cSTejun Heo * for IO completion of submitted IO. 2624f30a7d0cSTejun Heo */ 2625f30a7d0cSTejun Heo void writeback_inodes_sb_nr(struct super_block *sb, 2626f30a7d0cSTejun Heo unsigned long nr, 2627f30a7d0cSTejun Heo enum wb_reason reason) 2628f30a7d0cSTejun Heo { 2629f30a7d0cSTejun Heo __writeback_inodes_sb_nr(sb, nr, reason, false); 2630f30a7d0cSTejun Heo } 26313259f8beSChris Mason EXPORT_SYMBOL(writeback_inodes_sb_nr); 26323259f8beSChris Mason 26333259f8beSChris Mason /** 26343259f8beSChris Mason * writeback_inodes_sb - writeback dirty inodes from given super_block 26353259f8beSChris Mason * @sb: the superblock 2636786228abSMarcos Paulo de Souza * @reason: reason why some writeback work was initiated 26373259f8beSChris Mason * 26383259f8beSChris Mason * Start writeback on some inodes on this super_block. No guarantees are made 26393259f8beSChris Mason * on how many (if any) will be written, and this function does not wait 26403259f8beSChris Mason * for IO completion of submitted IO. 26413259f8beSChris Mason */ 26420e175a18SCurt Wohlgemuth void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 26433259f8beSChris Mason { 26440e175a18SCurt Wohlgemuth return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 26453259f8beSChris Mason } 2646d8a8559cSJens Axboe EXPORT_SYMBOL(writeback_inodes_sb); 2647d8a8559cSJens Axboe 2648d8a8559cSJens Axboe /** 264910ee27a0SMiao Xie * try_to_writeback_inodes_sb - try to start writeback if none underway 265010ee27a0SMiao Xie * @sb: the superblock 265110ee27a0SMiao Xie * @reason: reason why some writeback work was initiated 265210ee27a0SMiao Xie * 26538264c321SRakesh Pandit * Invoke __writeback_inodes_sb_nr if no writeback is currently underway. 265410ee27a0SMiao Xie */ 26558264c321SRakesh Pandit void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 265610ee27a0SMiao Xie { 26578264c321SRakesh Pandit if (!down_read_trylock(&sb->s_umount)) 26588264c321SRakesh Pandit return; 26598264c321SRakesh Pandit 26608264c321SRakesh Pandit __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true); 26618264c321SRakesh Pandit up_read(&sb->s_umount); 266210ee27a0SMiao Xie } 266310ee27a0SMiao Xie EXPORT_SYMBOL(try_to_writeback_inodes_sb); 26643259f8beSChris Mason 26653259f8beSChris Mason /** 2666d8a8559cSJens Axboe * sync_inodes_sb - sync sb inode pages 2667d8a8559cSJens Axboe * @sb: the superblock 2668d8a8559cSJens Axboe * 2669d8a8559cSJens Axboe * This function writes and waits on any dirty inode belonging to this 26700dc83bd3SJan Kara * super_block. 2671d8a8559cSJens Axboe */ 26720dc83bd3SJan Kara void sync_inodes_sb(struct super_block *sb) 2673d8a8559cSJens Axboe { 26745b9cce4cSTejun Heo struct backing_dev_info *bdi = sb->s_bdi; 26755b9cce4cSTejun Heo DEFINE_WB_COMPLETION(done, bdi); 267683ba7b07SChristoph Hellwig struct wb_writeback_work work = { 26773c4d7165SChristoph Hellwig .sb = sb, 26783c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_ALL, 26793c4d7165SChristoph Hellwig .nr_pages = LONG_MAX, 26803c4d7165SChristoph Hellwig .range_cyclic = 0, 268183ba7b07SChristoph Hellwig .done = &done, 26820e175a18SCurt Wohlgemuth .reason = WB_REASON_SYNC, 26837747bd4bSDave Chinner .for_sync = 1, 26843c4d7165SChristoph Hellwig }; 26853c4d7165SChristoph Hellwig 2686006a0973STejun Heo /* 2687006a0973STejun Heo * Can't skip on !bdi_has_dirty() because we should wait for !dirty 2688006a0973STejun Heo * inodes under writeback and I_DIRTY_TIME inodes ignored by 2689006a0973STejun Heo * bdi_has_dirty() need to be written out too. 2690006a0973STejun Heo */ 2691006a0973STejun Heo if (bdi == &noop_backing_dev_info) 26926eedc701SJan Kara return; 2693cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2694cf37e972SChristoph Hellwig 26957fc5854fSTejun Heo /* protect against inode wb switch, see inode_switch_wbs_work_fn() */ 26967fc5854fSTejun Heo bdi_down_write_wb_switch_rwsem(bdi); 2697db125360STejun Heo bdi_split_work_to_wbs(bdi, &work, false); 26985b9cce4cSTejun Heo wb_wait_for_completion(&done); 26997fc5854fSTejun Heo bdi_up_write_wb_switch_rwsem(bdi); 270083ba7b07SChristoph Hellwig 2701b6e51316SJens Axboe wait_sb_inodes(sb); 2702d8a8559cSJens Axboe } 2703d8a8559cSJens Axboe EXPORT_SYMBOL(sync_inodes_sb); 27041da177e4SLinus Torvalds 27051da177e4SLinus Torvalds /** 27061da177e4SLinus Torvalds * write_inode_now - write an inode to disk 27071da177e4SLinus Torvalds * @inode: inode to write to disk 27081da177e4SLinus Torvalds * @sync: whether the write should be synchronous or not 27091da177e4SLinus Torvalds * 27107f04c26dSAndrea Arcangeli * This function commits an inode to disk immediately if it is dirty. This is 27117f04c26dSAndrea Arcangeli * primarily needed by knfsd. 27127f04c26dSAndrea Arcangeli * 27137f04c26dSAndrea Arcangeli * The caller must either have a ref on the inode or must have set I_WILL_FREE. 27141da177e4SLinus Torvalds */ 27151da177e4SLinus Torvalds int write_inode_now(struct inode *inode, int sync) 27161da177e4SLinus Torvalds { 27171da177e4SLinus Torvalds struct writeback_control wbc = { 27181da177e4SLinus Torvalds .nr_to_write = LONG_MAX, 271918914b18SMike Galbraith .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, 2720111ebb6eSOGAWA Hirofumi .range_start = 0, 2721111ebb6eSOGAWA Hirofumi .range_end = LLONG_MAX, 27221da177e4SLinus Torvalds }; 27231da177e4SLinus Torvalds 2724f56753acSChristoph Hellwig if (!mapping_can_writeback(inode->i_mapping)) 272549364ce2SAndrew Morton wbc.nr_to_write = 0; 27261da177e4SLinus Torvalds 27271da177e4SLinus Torvalds might_sleep(); 2728aaf25593STejun Heo return writeback_single_inode(inode, &wbc); 27291da177e4SLinus Torvalds } 27301da177e4SLinus Torvalds EXPORT_SYMBOL(write_inode_now); 27311da177e4SLinus Torvalds 27321da177e4SLinus Torvalds /** 2733c691b9d9SAndrew Morton * sync_inode_metadata - write an inode to disk 2734c3765016SChristoph Hellwig * @inode: the inode to sync 2735c3765016SChristoph Hellwig * @wait: wait for I/O to complete. 2736c3765016SChristoph Hellwig * 2737c691b9d9SAndrew Morton * Write an inode to disk and adjust its dirty state after completion. 2738c3765016SChristoph Hellwig * 2739c3765016SChristoph Hellwig * Note: only writes the actual inode, no associated data or other metadata. 2740c3765016SChristoph Hellwig */ 2741c3765016SChristoph Hellwig int sync_inode_metadata(struct inode *inode, int wait) 2742c3765016SChristoph Hellwig { 2743c3765016SChristoph Hellwig struct writeback_control wbc = { 2744c3765016SChristoph Hellwig .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, 2745c3765016SChristoph Hellwig .nr_to_write = 0, /* metadata-only */ 2746c3765016SChristoph Hellwig }; 2747c3765016SChristoph Hellwig 27485662c967SJosef Bacik return writeback_single_inode(inode, &wbc); 2749c3765016SChristoph Hellwig } 2750c3765016SChristoph Hellwig EXPORT_SYMBOL(sync_inode_metadata); 2751