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); 12310e14073SJchao Sun assert_spin_locked(&inode->i_lock); 124d6c10f1fSTejun Heo 125c7f54084SDave Chinner list_move(&inode->i_io_list, head); 126d6c10f1fSTejun Heo 127d6c10f1fSTejun Heo /* dirty_time doesn't count as dirty_io until expiration */ 128d6c10f1fSTejun Heo if (head != &wb->b_dirty_time) 129d6c10f1fSTejun Heo return wb_io_lists_populated(wb); 130d6c10f1fSTejun Heo 131d6c10f1fSTejun Heo wb_io_lists_depopulated(wb); 132d6c10f1fSTejun Heo return false; 133d6c10f1fSTejun Heo } 134d6c10f1fSTejun Heo 135f0054bb1STejun Heo static void wb_wakeup(struct bdi_writeback *wb) 1365acda9d1SJan Kara { 137f87904c0SKhazhismel Kumykov spin_lock_irq(&wb->work_lock); 138f0054bb1STejun Heo if (test_bit(WB_registered, &wb->state)) 139f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 140f87904c0SKhazhismel Kumykov spin_unlock_irq(&wb->work_lock); 1415acda9d1SJan Kara } 1425acda9d1SJan Kara 1434a3a485bSTahsin Erdogan static void finish_writeback_work(struct bdi_writeback *wb, 1444a3a485bSTahsin Erdogan struct wb_writeback_work *work) 1454a3a485bSTahsin Erdogan { 1464a3a485bSTahsin Erdogan struct wb_completion *done = work->done; 1474a3a485bSTahsin Erdogan 1484a3a485bSTahsin Erdogan if (work->auto_free) 1494a3a485bSTahsin Erdogan kfree(work); 1508e00c4e9STejun Heo if (done) { 1518e00c4e9STejun Heo wait_queue_head_t *waitq = done->waitq; 1528e00c4e9STejun Heo 1538e00c4e9STejun Heo /* @done can't be accessed after the following dec */ 1548e00c4e9STejun Heo if (atomic_dec_and_test(&done->cnt)) 1558e00c4e9STejun Heo wake_up_all(waitq); 1568e00c4e9STejun Heo } 1574a3a485bSTahsin Erdogan } 1584a3a485bSTahsin Erdogan 159f0054bb1STejun Heo static void wb_queue_work(struct bdi_writeback *wb, 1606585027aSJan Kara struct wb_writeback_work *work) 1616585027aSJan Kara { 1625634cc2aSTejun Heo trace_writeback_queue(wb, work); 1636585027aSJan Kara 164cc395d7fSTejun Heo if (work->done) 165cc395d7fSTejun Heo atomic_inc(&work->done->cnt); 1664a3a485bSTahsin Erdogan 167f87904c0SKhazhismel Kumykov spin_lock_irq(&wb->work_lock); 1684a3a485bSTahsin Erdogan 1694a3a485bSTahsin Erdogan if (test_bit(WB_registered, &wb->state)) { 170f0054bb1STejun Heo list_add_tail(&work->list, &wb->work_list); 171f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 1724a3a485bSTahsin Erdogan } else 1734a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 1744a3a485bSTahsin Erdogan 175f87904c0SKhazhismel Kumykov spin_unlock_irq(&wb->work_lock); 17603ba3782SJens Axboe } 1771da177e4SLinus Torvalds 178cc395d7fSTejun Heo /** 179cc395d7fSTejun Heo * wb_wait_for_completion - wait for completion of bdi_writeback_works 180cc395d7fSTejun Heo * @done: target wb_completion 181cc395d7fSTejun Heo * 182cc395d7fSTejun Heo * Wait for one or more work items issued to @bdi with their ->done field 1835b9cce4cSTejun Heo * set to @done, which should have been initialized with 1845b9cce4cSTejun Heo * DEFINE_WB_COMPLETION(). This function returns after all such work items 1855b9cce4cSTejun Heo * are completed. Work items which are waited upon aren't freed 186cc395d7fSTejun Heo * automatically on completion. 187cc395d7fSTejun Heo */ 1885b9cce4cSTejun Heo void wb_wait_for_completion(struct wb_completion *done) 189cc395d7fSTejun Heo { 190cc395d7fSTejun Heo atomic_dec(&done->cnt); /* put down the initial count */ 1915b9cce4cSTejun Heo wait_event(*done->waitq, !atomic_read(&done->cnt)); 192cc395d7fSTejun Heo } 193cc395d7fSTejun Heo 194703c2708STejun Heo #ifdef CONFIG_CGROUP_WRITEBACK 195703c2708STejun Heo 19655a694dfSTejun Heo /* 19755a694dfSTejun Heo * Parameters for foreign inode detection, see wbc_detach_inode() to see 19855a694dfSTejun Heo * how they're used. 19955a694dfSTejun Heo * 20055a694dfSTejun Heo * These paramters are inherently heuristical as the detection target 20155a694dfSTejun Heo * itself is fuzzy. All we want to do is detaching an inode from the 20255a694dfSTejun Heo * current owner if it's being written to by some other cgroups too much. 20355a694dfSTejun Heo * 20455a694dfSTejun Heo * The current cgroup writeback is built on the assumption that multiple 20555a694dfSTejun Heo * cgroups writing to the same inode concurrently is very rare and a mode 20655a694dfSTejun Heo * of operation which isn't well supported. As such, the goal is not 20755a694dfSTejun Heo * taking too long when a different cgroup takes over an inode while 20855a694dfSTejun Heo * avoiding too aggressive flip-flops from occasional foreign writes. 20955a694dfSTejun Heo * 21055a694dfSTejun Heo * We record, very roughly, 2s worth of IO time history and if more than 21155a694dfSTejun Heo * half of that is foreign, trigger the switch. The recording is quantized 21255a694dfSTejun Heo * to 16 slots. To avoid tiny writes from swinging the decision too much, 21355a694dfSTejun Heo * writes smaller than 1/8 of avg size are ignored. 21455a694dfSTejun Heo */ 2152a814908STejun Heo #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */ 2162a814908STejun Heo #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */ 21755a694dfSTejun Heo #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */ 2182a814908STejun Heo #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */ 2192a814908STejun Heo 2202a814908STejun Heo #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */ 2212a814908STejun Heo #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS) 2222a814908STejun Heo /* each slot's duration is 2s / 16 */ 2232a814908STejun Heo #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2) 2242a814908STejun Heo /* if foreign slots >= 8, switch */ 2252a814908STejun Heo #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1) 2262a814908STejun Heo /* one round can affect upto 5 slots */ 2276444f47eSTejun Heo #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */ 2282a814908STejun Heo 229c22d70a1SRoman Gushchin /* 230c22d70a1SRoman Gushchin * Maximum inodes per isw. A specific value has been chosen to make 231c22d70a1SRoman Gushchin * struct inode_switch_wbs_context fit into 1024 bytes kmalloc. 232c22d70a1SRoman Gushchin */ 233c22d70a1SRoman Gushchin #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \ 234c22d70a1SRoman Gushchin / sizeof(struct inode *)) 235c22d70a1SRoman Gushchin 236a1a0e23eSTejun Heo static atomic_t isw_nr_in_flight = ATOMIC_INIT(0); 237a1a0e23eSTejun Heo static struct workqueue_struct *isw_wq; 238a1a0e23eSTejun Heo 23921c6321fSTejun Heo void __inode_attach_wb(struct inode *inode, struct page *page) 24021c6321fSTejun Heo { 24121c6321fSTejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 24221c6321fSTejun Heo struct bdi_writeback *wb = NULL; 24321c6321fSTejun Heo 24421c6321fSTejun Heo if (inode_cgwb_enabled(inode)) { 24521c6321fSTejun Heo struct cgroup_subsys_state *memcg_css; 24621c6321fSTejun Heo 24721c6321fSTejun Heo if (page) { 24821c6321fSTejun Heo memcg_css = mem_cgroup_css_from_page(page); 24921c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 25021c6321fSTejun Heo } else { 25121c6321fSTejun Heo /* must pin memcg_css, see wb_get_create() */ 25221c6321fSTejun Heo memcg_css = task_get_css(current, memory_cgrp_id); 25321c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 25421c6321fSTejun Heo css_put(memcg_css); 25521c6321fSTejun Heo } 25621c6321fSTejun Heo } 25721c6321fSTejun Heo 25821c6321fSTejun Heo if (!wb) 25921c6321fSTejun Heo wb = &bdi->wb; 26021c6321fSTejun Heo 26121c6321fSTejun Heo /* 26221c6321fSTejun Heo * There may be multiple instances of this function racing to 26321c6321fSTejun Heo * update the same inode. Use cmpxchg() to tell the winner. 26421c6321fSTejun Heo */ 26521c6321fSTejun Heo if (unlikely(cmpxchg(&inode->i_wb, NULL, wb))) 26621c6321fSTejun Heo wb_put(wb); 26721c6321fSTejun Heo } 2689b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(__inode_attach_wb); 26921c6321fSTejun Heo 270703c2708STejun Heo /** 271f3b6a6dfSRoman Gushchin * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list 272f3b6a6dfSRoman Gushchin * @inode: inode of interest with i_lock held 273f3b6a6dfSRoman Gushchin * @wb: target bdi_writeback 274f3b6a6dfSRoman Gushchin * 275f3b6a6dfSRoman Gushchin * Remove the inode from wb's io lists and if necessarily put onto b_attached 276f3b6a6dfSRoman Gushchin * list. Only inodes attached to cgwb's are kept on this list. 277f3b6a6dfSRoman Gushchin */ 278f3b6a6dfSRoman Gushchin static void inode_cgwb_move_to_attached(struct inode *inode, 279f3b6a6dfSRoman Gushchin struct bdi_writeback *wb) 280f3b6a6dfSRoman Gushchin { 281f3b6a6dfSRoman Gushchin assert_spin_locked(&wb->list_lock); 282f3b6a6dfSRoman Gushchin assert_spin_locked(&inode->i_lock); 283f3b6a6dfSRoman Gushchin 284f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 285f3b6a6dfSRoman Gushchin if (wb != &wb->bdi->wb) 286f3b6a6dfSRoman Gushchin list_move(&inode->i_io_list, &wb->b_attached); 287f3b6a6dfSRoman Gushchin else 288f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 289f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 290f3b6a6dfSRoman Gushchin } 291f3b6a6dfSRoman Gushchin 292f3b6a6dfSRoman Gushchin /** 29387e1d789STejun Heo * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it 29487e1d789STejun Heo * @inode: inode of interest with i_lock held 29587e1d789STejun Heo * 29687e1d789STejun Heo * Returns @inode's wb with its list_lock held. @inode->i_lock must be 29787e1d789STejun Heo * held on entry and is released on return. The returned wb is guaranteed 29887e1d789STejun Heo * to stay @inode's associated wb until its list_lock is released. 29987e1d789STejun Heo */ 30087e1d789STejun Heo static struct bdi_writeback * 30187e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 30287e1d789STejun Heo __releases(&inode->i_lock) 30387e1d789STejun Heo __acquires(&wb->list_lock) 30487e1d789STejun Heo { 30587e1d789STejun Heo while (true) { 30687e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 30787e1d789STejun Heo 30887e1d789STejun Heo /* 30987e1d789STejun Heo * inode_to_wb() association is protected by both 31087e1d789STejun Heo * @inode->i_lock and @wb->list_lock but list_lock nests 31187e1d789STejun Heo * outside i_lock. Drop i_lock and verify that the 31287e1d789STejun Heo * association hasn't changed after acquiring list_lock. 31387e1d789STejun Heo */ 31487e1d789STejun Heo wb_get(wb); 31587e1d789STejun Heo spin_unlock(&inode->i_lock); 31687e1d789STejun Heo spin_lock(&wb->list_lock); 31787e1d789STejun Heo 318aaa2cacfSTejun Heo /* i_wb may have changed inbetween, can't use inode_to_wb() */ 319614a4e37STejun Heo if (likely(wb == inode->i_wb)) { 320614a4e37STejun Heo wb_put(wb); /* @inode already has ref */ 321614a4e37STejun Heo return wb; 322614a4e37STejun Heo } 32387e1d789STejun Heo 32487e1d789STejun Heo spin_unlock(&wb->list_lock); 325614a4e37STejun Heo wb_put(wb); 32687e1d789STejun Heo cpu_relax(); 32787e1d789STejun Heo spin_lock(&inode->i_lock); 32887e1d789STejun Heo } 32987e1d789STejun Heo } 33087e1d789STejun Heo 33187e1d789STejun Heo /** 33287e1d789STejun Heo * inode_to_wb_and_lock_list - determine an inode's wb and lock it 33387e1d789STejun Heo * @inode: inode of interest 33487e1d789STejun Heo * 33587e1d789STejun Heo * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held 33687e1d789STejun Heo * on entry. 33787e1d789STejun Heo */ 33887e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 33987e1d789STejun Heo __acquires(&wb->list_lock) 34087e1d789STejun Heo { 34187e1d789STejun Heo spin_lock(&inode->i_lock); 34287e1d789STejun Heo return locked_inode_to_wb_and_lock_list(inode); 34387e1d789STejun Heo } 34487e1d789STejun Heo 345682aa8e1STejun Heo struct inode_switch_wbs_context { 34629264d92SRoman Gushchin struct rcu_work work; 347f5fbe6b7SRoman Gushchin 348f5fbe6b7SRoman Gushchin /* 349f5fbe6b7SRoman Gushchin * Multiple inodes can be switched at once. The switching procedure 350f5fbe6b7SRoman Gushchin * consists of two parts, separated by a RCU grace period. To make 351f5fbe6b7SRoman Gushchin * sure that the second part is executed for each inode gone through 352f5fbe6b7SRoman Gushchin * the first part, all inode pointers are placed into a NULL-terminated 353f5fbe6b7SRoman Gushchin * array embedded into struct inode_switch_wbs_context. Otherwise 354f5fbe6b7SRoman Gushchin * an inode could be left in a non-consistent state. 355f5fbe6b7SRoman Gushchin */ 356f5fbe6b7SRoman Gushchin struct bdi_writeback *new_wb; 357f5fbe6b7SRoman Gushchin struct inode *inodes[]; 358682aa8e1STejun Heo }; 359682aa8e1STejun Heo 3607fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) 3617fc5854fSTejun Heo { 3627fc5854fSTejun Heo down_write(&bdi->wb_switch_rwsem); 3637fc5854fSTejun Heo } 3647fc5854fSTejun Heo 3657fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) 3667fc5854fSTejun Heo { 3677fc5854fSTejun Heo up_write(&bdi->wb_switch_rwsem); 3687fc5854fSTejun Heo } 3697fc5854fSTejun Heo 370f5fbe6b7SRoman Gushchin static bool inode_do_switch_wbs(struct inode *inode, 371f5fbe6b7SRoman Gushchin struct bdi_writeback *old_wb, 37272d4512eSRoman Gushchin struct bdi_writeback *new_wb) 373682aa8e1STejun Heo { 374d10c8095STejun Heo struct address_space *mapping = inode->i_mapping; 37504edf02cSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, 0); 37622b3c8d6SMatthew Wilcox (Oracle) struct folio *folio; 377d10c8095STejun Heo bool switched = false; 378682aa8e1STejun Heo 379682aa8e1STejun Heo spin_lock(&inode->i_lock); 380b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 381682aa8e1STejun Heo 382d10c8095STejun Heo /* 3834ade5867SRoman Gushchin * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction 3844ade5867SRoman Gushchin * path owns the inode and we shouldn't modify ->i_io_list. 385d10c8095STejun Heo */ 3864ade5867SRoman Gushchin if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE))) 387d10c8095STejun Heo goto skip_switch; 388d10c8095STejun Heo 3893a8e9ac8STejun Heo trace_inode_switch_wbs(inode, old_wb, new_wb); 3903a8e9ac8STejun Heo 391d10c8095STejun Heo /* 392d10c8095STejun Heo * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points 39322b3c8d6SMatthew Wilcox (Oracle) * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to 39422b3c8d6SMatthew Wilcox (Oracle) * folios actually under writeback. 395d10c8095STejun Heo */ 39622b3c8d6SMatthew Wilcox (Oracle) xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) { 39722b3c8d6SMatthew Wilcox (Oracle) if (folio_test_dirty(folio)) { 39822b3c8d6SMatthew Wilcox (Oracle) long nr = folio_nr_pages(folio); 39922b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr); 40022b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(new_wb, WB_RECLAIMABLE, nr); 401d10c8095STejun Heo } 402d10c8095STejun Heo } 403d10c8095STejun Heo 40404edf02cSMatthew Wilcox xas_set(&xas, 0); 40522b3c8d6SMatthew Wilcox (Oracle) xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) { 40622b3c8d6SMatthew Wilcox (Oracle) long nr = folio_nr_pages(folio); 40722b3c8d6SMatthew Wilcox (Oracle) WARN_ON_ONCE(!folio_test_writeback(folio)); 40822b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(old_wb, WB_WRITEBACK, -nr); 40922b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(new_wb, WB_WRITEBACK, nr); 410d10c8095STejun Heo } 411d10c8095STejun Heo 412633a2abbSJan Kara if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) { 413633a2abbSJan Kara atomic_dec(&old_wb->writeback_inodes); 414633a2abbSJan Kara atomic_inc(&new_wb->writeback_inodes); 415633a2abbSJan Kara } 416633a2abbSJan Kara 417d10c8095STejun Heo wb_get(new_wb); 418d10c8095STejun Heo 419d10c8095STejun Heo /* 420f3b6a6dfSRoman Gushchin * Transfer to @new_wb's IO list if necessary. If the @inode is dirty, 421f3b6a6dfSRoman Gushchin * the specific list @inode was on is ignored and the @inode is put on 422f3b6a6dfSRoman Gushchin * ->b_dirty which is always correct including from ->b_dirty_time. 423f3b6a6dfSRoman Gushchin * The transfer preserves @inode->dirtied_when ordering. If the @inode 424f3b6a6dfSRoman Gushchin * was clean, it means it was on the b_attached list, so move it onto 425f3b6a6dfSRoman Gushchin * the b_attached list of @new_wb. 426d10c8095STejun Heo */ 427c7f54084SDave Chinner if (!list_empty(&inode->i_io_list)) { 428f3b6a6dfSRoman Gushchin inode->i_wb = new_wb; 429f3b6a6dfSRoman Gushchin 430f3b6a6dfSRoman Gushchin if (inode->i_state & I_DIRTY_ALL) { 431d10c8095STejun Heo struct inode *pos; 432d10c8095STejun Heo 433c7f54084SDave Chinner list_for_each_entry(pos, &new_wb->b_dirty, i_io_list) 434d10c8095STejun Heo if (time_after_eq(inode->dirtied_when, 435d10c8095STejun Heo pos->dirtied_when)) 436d10c8095STejun Heo break; 437f3b6a6dfSRoman Gushchin inode_io_list_move_locked(inode, new_wb, 438f3b6a6dfSRoman Gushchin pos->i_io_list.prev); 439f3b6a6dfSRoman Gushchin } else { 440f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, new_wb); 441f3b6a6dfSRoman Gushchin } 442d10c8095STejun Heo } else { 443d10c8095STejun Heo inode->i_wb = new_wb; 444d10c8095STejun Heo } 445d10c8095STejun Heo 446d10c8095STejun Heo /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */ 447682aa8e1STejun Heo inode->i_wb_frn_winner = 0; 448682aa8e1STejun Heo inode->i_wb_frn_avg_time = 0; 449682aa8e1STejun Heo inode->i_wb_frn_history = 0; 450d10c8095STejun Heo switched = true; 451d10c8095STejun Heo skip_switch: 452682aa8e1STejun Heo /* 453682aa8e1STejun Heo * Paired with load_acquire in unlocked_inode_to_wb_begin() and 454682aa8e1STejun Heo * ensures that the new wb is visible if they see !I_WB_SWITCH. 455682aa8e1STejun Heo */ 456682aa8e1STejun Heo smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); 457682aa8e1STejun Heo 458b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 459682aa8e1STejun Heo spin_unlock(&inode->i_lock); 460d10c8095STejun Heo 461f5fbe6b7SRoman Gushchin return switched; 46272d4512eSRoman Gushchin } 463682aa8e1STejun Heo 46472d4512eSRoman Gushchin static void inode_switch_wbs_work_fn(struct work_struct *work) 46572d4512eSRoman Gushchin { 46672d4512eSRoman Gushchin struct inode_switch_wbs_context *isw = 46772d4512eSRoman Gushchin container_of(to_rcu_work(work), struct inode_switch_wbs_context, work); 468f5fbe6b7SRoman Gushchin struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]); 469f5fbe6b7SRoman Gushchin struct bdi_writeback *old_wb = isw->inodes[0]->i_wb; 470f5fbe6b7SRoman Gushchin struct bdi_writeback *new_wb = isw->new_wb; 471f5fbe6b7SRoman Gushchin unsigned long nr_switched = 0; 472f5fbe6b7SRoman Gushchin struct inode **inodep; 47372d4512eSRoman Gushchin 474f5fbe6b7SRoman Gushchin /* 475f5fbe6b7SRoman Gushchin * If @inode switches cgwb membership while sync_inodes_sb() is 476f5fbe6b7SRoman Gushchin * being issued, sync_inodes_sb() might miss it. Synchronize. 477f5fbe6b7SRoman Gushchin */ 478f5fbe6b7SRoman Gushchin down_read(&bdi->wb_switch_rwsem); 479f5fbe6b7SRoman Gushchin 480f5fbe6b7SRoman Gushchin /* 481f5fbe6b7SRoman Gushchin * By the time control reaches here, RCU grace period has passed 482f5fbe6b7SRoman Gushchin * since I_WB_SWITCH assertion and all wb stat update transactions 483f5fbe6b7SRoman Gushchin * between unlocked_inode_to_wb_begin/end() are guaranteed to be 484f5fbe6b7SRoman Gushchin * synchronizing against the i_pages lock. 485f5fbe6b7SRoman Gushchin * 486f5fbe6b7SRoman Gushchin * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock 487f5fbe6b7SRoman Gushchin * gives us exclusion against all wb related operations on @inode 488f5fbe6b7SRoman Gushchin * including IO list manipulations and stat updates. 489f5fbe6b7SRoman Gushchin */ 490f5fbe6b7SRoman Gushchin if (old_wb < new_wb) { 491f5fbe6b7SRoman Gushchin spin_lock(&old_wb->list_lock); 492f5fbe6b7SRoman Gushchin spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING); 493f5fbe6b7SRoman Gushchin } else { 494f5fbe6b7SRoman Gushchin spin_lock(&new_wb->list_lock); 495f5fbe6b7SRoman Gushchin spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING); 496f5fbe6b7SRoman Gushchin } 497f5fbe6b7SRoman Gushchin 498f5fbe6b7SRoman Gushchin for (inodep = isw->inodes; *inodep; inodep++) { 499f5fbe6b7SRoman Gushchin WARN_ON_ONCE((*inodep)->i_wb != old_wb); 500f5fbe6b7SRoman Gushchin if (inode_do_switch_wbs(*inodep, old_wb, new_wb)) 501f5fbe6b7SRoman Gushchin nr_switched++; 502f5fbe6b7SRoman Gushchin } 503f5fbe6b7SRoman Gushchin 504f5fbe6b7SRoman Gushchin spin_unlock(&new_wb->list_lock); 505f5fbe6b7SRoman Gushchin spin_unlock(&old_wb->list_lock); 506f5fbe6b7SRoman Gushchin 507f5fbe6b7SRoman Gushchin up_read(&bdi->wb_switch_rwsem); 508f5fbe6b7SRoman Gushchin 509f5fbe6b7SRoman Gushchin if (nr_switched) { 510f5fbe6b7SRoman Gushchin wb_wakeup(new_wb); 511f5fbe6b7SRoman Gushchin wb_put_many(old_wb, nr_switched); 512f5fbe6b7SRoman Gushchin } 513f5fbe6b7SRoman Gushchin 514f5fbe6b7SRoman Gushchin for (inodep = isw->inodes; *inodep; inodep++) 515f5fbe6b7SRoman Gushchin iput(*inodep); 516f5fbe6b7SRoman Gushchin wb_put(new_wb); 517682aa8e1STejun Heo kfree(isw); 518a1a0e23eSTejun Heo atomic_dec(&isw_nr_in_flight); 519682aa8e1STejun Heo } 520682aa8e1STejun Heo 521c22d70a1SRoman Gushchin static bool inode_prepare_wbs_switch(struct inode *inode, 522c22d70a1SRoman Gushchin struct bdi_writeback *new_wb) 523c22d70a1SRoman Gushchin { 524c22d70a1SRoman Gushchin /* 525c22d70a1SRoman Gushchin * Paired with smp_mb() in cgroup_writeback_umount(). 526c22d70a1SRoman Gushchin * isw_nr_in_flight must be increased before checking SB_ACTIVE and 527c22d70a1SRoman Gushchin * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0 528c22d70a1SRoman Gushchin * in cgroup_writeback_umount() and the isw_wq will be not flushed. 529c22d70a1SRoman Gushchin */ 530c22d70a1SRoman Gushchin smp_mb(); 531c22d70a1SRoman Gushchin 532593311e8SRoman Gushchin if (IS_DAX(inode)) 533593311e8SRoman Gushchin return false; 534593311e8SRoman Gushchin 535c22d70a1SRoman Gushchin /* while holding I_WB_SWITCH, no one else can update the association */ 536c22d70a1SRoman Gushchin spin_lock(&inode->i_lock); 537c22d70a1SRoman Gushchin if (!(inode->i_sb->s_flags & SB_ACTIVE) || 538c22d70a1SRoman Gushchin inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) || 539c22d70a1SRoman Gushchin inode_to_wb(inode) == new_wb) { 540c22d70a1SRoman Gushchin spin_unlock(&inode->i_lock); 541c22d70a1SRoman Gushchin return false; 542c22d70a1SRoman Gushchin } 543c22d70a1SRoman Gushchin inode->i_state |= I_WB_SWITCH; 544c22d70a1SRoman Gushchin __iget(inode); 545c22d70a1SRoman Gushchin spin_unlock(&inode->i_lock); 546c22d70a1SRoman Gushchin 547c22d70a1SRoman Gushchin return true; 548c22d70a1SRoman Gushchin } 549c22d70a1SRoman Gushchin 550682aa8e1STejun Heo /** 551682aa8e1STejun Heo * inode_switch_wbs - change the wb association of an inode 552682aa8e1STejun Heo * @inode: target inode 553682aa8e1STejun Heo * @new_wb_id: ID of the new wb 554682aa8e1STejun Heo * 555682aa8e1STejun Heo * Switch @inode's wb association to the wb identified by @new_wb_id. The 556682aa8e1STejun Heo * switching is performed asynchronously and may fail silently. 557682aa8e1STejun Heo */ 558682aa8e1STejun Heo static void inode_switch_wbs(struct inode *inode, int new_wb_id) 559682aa8e1STejun Heo { 560682aa8e1STejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 561682aa8e1STejun Heo struct cgroup_subsys_state *memcg_css; 562682aa8e1STejun Heo struct inode_switch_wbs_context *isw; 563682aa8e1STejun Heo 564682aa8e1STejun Heo /* noop if seems to be already in progress */ 565682aa8e1STejun Heo if (inode->i_state & I_WB_SWITCH) 566682aa8e1STejun Heo return; 567682aa8e1STejun Heo 5686444f47eSTejun Heo /* avoid queueing a new switch if too many are already in flight */ 5696444f47eSTejun Heo if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT) 5707fc5854fSTejun Heo return; 5717fc5854fSTejun Heo 57298b160c8SLen Baker isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC); 573682aa8e1STejun Heo if (!isw) 5746444f47eSTejun Heo return; 575682aa8e1STejun Heo 5768826ee4fSRoman Gushchin atomic_inc(&isw_nr_in_flight); 5778826ee4fSRoman Gushchin 578682aa8e1STejun Heo /* find and pin the new wb */ 579682aa8e1STejun Heo rcu_read_lock(); 580682aa8e1STejun Heo memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); 5818b0ed844SMuchun Song if (memcg_css && !css_tryget(memcg_css)) 5828b0ed844SMuchun Song memcg_css = NULL; 583682aa8e1STejun Heo rcu_read_unlock(); 5848b0ed844SMuchun Song if (!memcg_css) 5858b0ed844SMuchun Song goto out_free; 5868b0ed844SMuchun Song 5878b0ed844SMuchun Song isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 5888b0ed844SMuchun Song css_put(memcg_css); 589682aa8e1STejun Heo if (!isw->new_wb) 590682aa8e1STejun Heo goto out_free; 591682aa8e1STejun Heo 592c22d70a1SRoman Gushchin if (!inode_prepare_wbs_switch(inode, isw->new_wb)) 593a1a0e23eSTejun Heo goto out_free; 594682aa8e1STejun Heo 595f5fbe6b7SRoman Gushchin isw->inodes[0] = inode; 596682aa8e1STejun Heo 597682aa8e1STejun Heo /* 598682aa8e1STejun Heo * In addition to synchronizing among switchers, I_WB_SWITCH tells 599b93b0163SMatthew Wilcox * the RCU protected stat update paths to grab the i_page 600b93b0163SMatthew Wilcox * lock so that stat transfer can synchronize against them. 601682aa8e1STejun Heo * Let's continue after I_WB_SWITCH is guaranteed to be visible. 602682aa8e1STejun Heo */ 60329264d92SRoman Gushchin INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); 60429264d92SRoman Gushchin queue_rcu_work(isw_wq, &isw->work); 6056444f47eSTejun Heo return; 606682aa8e1STejun Heo 607682aa8e1STejun Heo out_free: 6088826ee4fSRoman Gushchin atomic_dec(&isw_nr_in_flight); 609682aa8e1STejun Heo if (isw->new_wb) 610682aa8e1STejun Heo wb_put(isw->new_wb); 611682aa8e1STejun Heo kfree(isw); 612682aa8e1STejun Heo } 613682aa8e1STejun Heo 61487e1d789STejun Heo /** 615c22d70a1SRoman Gushchin * cleanup_offline_cgwb - detach associated inodes 616c22d70a1SRoman Gushchin * @wb: target wb 617c22d70a1SRoman Gushchin * 618c22d70a1SRoman Gushchin * Switch all inodes attached to @wb to a nearest living ancestor's wb in order 619c22d70a1SRoman Gushchin * to eventually release the dying @wb. Returns %true if not all inodes were 620c22d70a1SRoman Gushchin * switched and the function has to be restarted. 621c22d70a1SRoman Gushchin */ 622c22d70a1SRoman Gushchin bool cleanup_offline_cgwb(struct bdi_writeback *wb) 623c22d70a1SRoman Gushchin { 624c22d70a1SRoman Gushchin struct cgroup_subsys_state *memcg_css; 625c22d70a1SRoman Gushchin struct inode_switch_wbs_context *isw; 626c22d70a1SRoman Gushchin struct inode *inode; 627c22d70a1SRoman Gushchin int nr; 628c22d70a1SRoman Gushchin bool restart = false; 629c22d70a1SRoman Gushchin 63098b160c8SLen Baker isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW), 63198b160c8SLen Baker GFP_KERNEL); 632c22d70a1SRoman Gushchin if (!isw) 633c22d70a1SRoman Gushchin return restart; 634c22d70a1SRoman Gushchin 635c22d70a1SRoman Gushchin atomic_inc(&isw_nr_in_flight); 636c22d70a1SRoman Gushchin 637c22d70a1SRoman Gushchin for (memcg_css = wb->memcg_css->parent; memcg_css; 638c22d70a1SRoman Gushchin memcg_css = memcg_css->parent) { 639c22d70a1SRoman Gushchin isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL); 640c22d70a1SRoman Gushchin if (isw->new_wb) 641c22d70a1SRoman Gushchin break; 642c22d70a1SRoman Gushchin } 643c22d70a1SRoman Gushchin if (unlikely(!isw->new_wb)) 644c22d70a1SRoman Gushchin isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */ 645c22d70a1SRoman Gushchin 646c22d70a1SRoman Gushchin nr = 0; 647c22d70a1SRoman Gushchin spin_lock(&wb->list_lock); 648c22d70a1SRoman Gushchin list_for_each_entry(inode, &wb->b_attached, i_io_list) { 649c22d70a1SRoman Gushchin if (!inode_prepare_wbs_switch(inode, isw->new_wb)) 650c22d70a1SRoman Gushchin continue; 651c22d70a1SRoman Gushchin 652c22d70a1SRoman Gushchin isw->inodes[nr++] = inode; 653c22d70a1SRoman Gushchin 654c22d70a1SRoman Gushchin if (nr >= WB_MAX_INODES_PER_ISW - 1) { 655c22d70a1SRoman Gushchin restart = true; 656c22d70a1SRoman Gushchin break; 657c22d70a1SRoman Gushchin } 658c22d70a1SRoman Gushchin } 659c22d70a1SRoman Gushchin spin_unlock(&wb->list_lock); 660c22d70a1SRoman Gushchin 661c22d70a1SRoman Gushchin /* no attached inodes? bail out */ 662c22d70a1SRoman Gushchin if (nr == 0) { 663c22d70a1SRoman Gushchin atomic_dec(&isw_nr_in_flight); 664c22d70a1SRoman Gushchin wb_put(isw->new_wb); 665c22d70a1SRoman Gushchin kfree(isw); 666c22d70a1SRoman Gushchin return restart; 667c22d70a1SRoman Gushchin } 668c22d70a1SRoman Gushchin 669c22d70a1SRoman Gushchin /* 670c22d70a1SRoman Gushchin * In addition to synchronizing among switchers, I_WB_SWITCH tells 671c22d70a1SRoman Gushchin * the RCU protected stat update paths to grab the i_page 672c22d70a1SRoman Gushchin * lock so that stat transfer can synchronize against them. 673c22d70a1SRoman Gushchin * Let's continue after I_WB_SWITCH is guaranteed to be visible. 674c22d70a1SRoman Gushchin */ 675c22d70a1SRoman Gushchin INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); 676c22d70a1SRoman Gushchin queue_rcu_work(isw_wq, &isw->work); 677c22d70a1SRoman Gushchin 678c22d70a1SRoman Gushchin return restart; 679c22d70a1SRoman Gushchin } 680c22d70a1SRoman Gushchin 681c22d70a1SRoman Gushchin /** 682b16b1debSTejun Heo * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it 683b16b1debSTejun Heo * @wbc: writeback_control of interest 684b16b1debSTejun Heo * @inode: target inode 685b16b1debSTejun Heo * 686b16b1debSTejun Heo * @inode is locked and about to be written back under the control of @wbc. 687b16b1debSTejun Heo * Record @inode's writeback context into @wbc and unlock the i_lock. On 688b16b1debSTejun Heo * writeback completion, wbc_detach_inode() should be called. This is used 689b16b1debSTejun Heo * to track the cgroup writeback context. 690b16b1debSTejun Heo */ 691b16b1debSTejun Heo void wbc_attach_and_unlock_inode(struct writeback_control *wbc, 692b16b1debSTejun Heo struct inode *inode) 693b16b1debSTejun Heo { 694dd73e4b7STejun Heo if (!inode_cgwb_enabled(inode)) { 695dd73e4b7STejun Heo spin_unlock(&inode->i_lock); 696dd73e4b7STejun Heo return; 697dd73e4b7STejun Heo } 698dd73e4b7STejun Heo 699b16b1debSTejun Heo wbc->wb = inode_to_wb(inode); 7002a814908STejun Heo wbc->inode = inode; 7012a814908STejun Heo 7022a814908STejun Heo wbc->wb_id = wbc->wb->memcg_css->id; 7032a814908STejun Heo wbc->wb_lcand_id = inode->i_wb_frn_winner; 7042a814908STejun Heo wbc->wb_tcand_id = 0; 7052a814908STejun Heo wbc->wb_bytes = 0; 7062a814908STejun Heo wbc->wb_lcand_bytes = 0; 7072a814908STejun Heo wbc->wb_tcand_bytes = 0; 7082a814908STejun Heo 709b16b1debSTejun Heo wb_get(wbc->wb); 710b16b1debSTejun Heo spin_unlock(&inode->i_lock); 711e8a7abf5STejun Heo 712e8a7abf5STejun Heo /* 71365de03e2STejun Heo * A dying wb indicates that either the blkcg associated with the 71465de03e2STejun Heo * memcg changed or the associated memcg is dying. In the first 71565de03e2STejun Heo * case, a replacement wb should already be available and we should 71665de03e2STejun Heo * refresh the wb immediately. In the second case, trying to 71765de03e2STejun Heo * refresh will keep failing. 718e8a7abf5STejun Heo */ 71965de03e2STejun Heo if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css))) 720e8a7abf5STejun Heo inode_switch_wbs(inode, wbc->wb_id); 721b16b1debSTejun Heo } 7229b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode); 723b16b1debSTejun Heo 724b16b1debSTejun Heo /** 7252a814908STejun Heo * wbc_detach_inode - disassociate wbc from inode and perform foreign detection 7262a814908STejun Heo * @wbc: writeback_control of the just finished writeback 727b16b1debSTejun Heo * 728b16b1debSTejun Heo * To be called after a writeback attempt of an inode finishes and undoes 729b16b1debSTejun Heo * wbc_attach_and_unlock_inode(). Can be called under any context. 7302a814908STejun Heo * 7312a814908STejun Heo * As concurrent write sharing of an inode is expected to be very rare and 7322a814908STejun Heo * memcg only tracks page ownership on first-use basis severely confining 7332a814908STejun Heo * the usefulness of such sharing, cgroup writeback tracks ownership 7342a814908STejun Heo * per-inode. While the support for concurrent write sharing of an inode 7352a814908STejun Heo * is deemed unnecessary, an inode being written to by different cgroups at 7362a814908STejun Heo * different points in time is a lot more common, and, more importantly, 7372a814908STejun Heo * charging only by first-use can too readily lead to grossly incorrect 7382a814908STejun Heo * behaviors (single foreign page can lead to gigabytes of writeback to be 7392a814908STejun Heo * incorrectly attributed). 7402a814908STejun Heo * 7412a814908STejun Heo * To resolve this issue, cgroup writeback detects the majority dirtier of 7422999e1e3SJulia Lawall * an inode and transfers the ownership to it. To avoid unnecessary 7432a814908STejun Heo * oscillation, the detection mechanism keeps track of history and gives 7442a814908STejun Heo * out the switch verdict only if the foreign usage pattern is stable over 7452a814908STejun Heo * a certain amount of time and/or writeback attempts. 7462a814908STejun Heo * 7472a814908STejun Heo * On each writeback attempt, @wbc tries to detect the majority writer 7482a814908STejun Heo * using Boyer-Moore majority vote algorithm. In addition to the byte 7492a814908STejun Heo * count from the majority voting, it also counts the bytes written for the 7502a814908STejun Heo * current wb and the last round's winner wb (max of last round's current 7512a814908STejun Heo * wb, the winner from two rounds ago, and the last round's majority 7522a814908STejun Heo * candidate). Keeping track of the historical winner helps the algorithm 7532a814908STejun Heo * to semi-reliably detect the most active writer even when it's not the 7542a814908STejun Heo * absolute majority. 7552a814908STejun Heo * 7562a814908STejun Heo * Once the winner of the round is determined, whether the winner is 7572a814908STejun Heo * foreign or not and how much IO time the round consumed is recorded in 7582a814908STejun Heo * inode->i_wb_frn_history. If the amount of recorded foreign IO time is 7592a814908STejun Heo * over a certain threshold, the switch verdict is given. 760b16b1debSTejun Heo */ 761b16b1debSTejun Heo void wbc_detach_inode(struct writeback_control *wbc) 762b16b1debSTejun Heo { 7632a814908STejun Heo struct bdi_writeback *wb = wbc->wb; 7642a814908STejun Heo struct inode *inode = wbc->inode; 765dd73e4b7STejun Heo unsigned long avg_time, max_bytes, max_time; 766dd73e4b7STejun Heo u16 history; 7672a814908STejun Heo int max_id; 7682a814908STejun Heo 769dd73e4b7STejun Heo if (!wb) 770dd73e4b7STejun Heo return; 771dd73e4b7STejun Heo 772dd73e4b7STejun Heo history = inode->i_wb_frn_history; 773dd73e4b7STejun Heo avg_time = inode->i_wb_frn_avg_time; 774dd73e4b7STejun Heo 7752a814908STejun Heo /* pick the winner of this round */ 7762a814908STejun Heo if (wbc->wb_bytes >= wbc->wb_lcand_bytes && 7772a814908STejun Heo wbc->wb_bytes >= wbc->wb_tcand_bytes) { 7782a814908STejun Heo max_id = wbc->wb_id; 7792a814908STejun Heo max_bytes = wbc->wb_bytes; 7802a814908STejun Heo } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { 7812a814908STejun Heo max_id = wbc->wb_lcand_id; 7822a814908STejun Heo max_bytes = wbc->wb_lcand_bytes; 7832a814908STejun Heo } else { 7842a814908STejun Heo max_id = wbc->wb_tcand_id; 7852a814908STejun Heo max_bytes = wbc->wb_tcand_bytes; 7862a814908STejun Heo } 7872a814908STejun Heo 7882a814908STejun Heo /* 7892a814908STejun Heo * Calculate the amount of IO time the winner consumed and fold it 7902a814908STejun Heo * into the running average kept per inode. If the consumed IO 7912a814908STejun Heo * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for 7922a814908STejun Heo * deciding whether to switch or not. This is to prevent one-off 7932a814908STejun Heo * small dirtiers from skewing the verdict. 7942a814908STejun Heo */ 7952a814908STejun Heo max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, 7962a814908STejun Heo wb->avg_write_bandwidth); 7972a814908STejun Heo if (avg_time) 7982a814908STejun Heo avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - 7992a814908STejun Heo (avg_time >> WB_FRN_TIME_AVG_SHIFT); 8002a814908STejun Heo else 8012a814908STejun Heo avg_time = max_time; /* immediate catch up on first run */ 8022a814908STejun Heo 8032a814908STejun Heo if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { 8042a814908STejun Heo int slots; 8052a814908STejun Heo 8062a814908STejun Heo /* 8072a814908STejun Heo * The switch verdict is reached if foreign wb's consume 8082a814908STejun Heo * more than a certain proportion of IO time in a 8092a814908STejun Heo * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot 8102a814908STejun Heo * history mask where each bit represents one sixteenth of 8112a814908STejun Heo * the period. Determine the number of slots to shift into 8122a814908STejun Heo * history from @max_time. 8132a814908STejun Heo */ 8142a814908STejun Heo slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), 8152a814908STejun Heo (unsigned long)WB_FRN_HIST_MAX_SLOTS); 8162a814908STejun Heo history <<= slots; 8172a814908STejun Heo if (wbc->wb_id != max_id) 8182a814908STejun Heo history |= (1U << slots) - 1; 8192a814908STejun Heo 8203a8e9ac8STejun Heo if (history) 8213a8e9ac8STejun Heo trace_inode_foreign_history(inode, wbc, history); 8223a8e9ac8STejun Heo 8232a814908STejun Heo /* 8242a814908STejun Heo * Switch if the current wb isn't the consistent winner. 8252a814908STejun Heo * If there are multiple closely competing dirtiers, the 8262a814908STejun Heo * inode may switch across them repeatedly over time, which 8272a814908STejun Heo * is okay. The main goal is avoiding keeping an inode on 8282a814908STejun Heo * the wrong wb for an extended period of time. 8292a814908STejun Heo */ 830682aa8e1STejun Heo if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) 831682aa8e1STejun Heo inode_switch_wbs(inode, max_id); 8322a814908STejun Heo } 8332a814908STejun Heo 8342a814908STejun Heo /* 8352a814908STejun Heo * Multiple instances of this function may race to update the 8362a814908STejun Heo * following fields but we don't mind occassional inaccuracies. 8372a814908STejun Heo */ 8382a814908STejun Heo inode->i_wb_frn_winner = max_id; 8392a814908STejun Heo inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); 8402a814908STejun Heo inode->i_wb_frn_history = history; 8412a814908STejun Heo 842b16b1debSTejun Heo wb_put(wbc->wb); 843b16b1debSTejun Heo wbc->wb = NULL; 844b16b1debSTejun Heo } 8459b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_detach_inode); 846b16b1debSTejun Heo 847b16b1debSTejun Heo /** 84834e51a5eSTejun Heo * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership 8492a814908STejun Heo * @wbc: writeback_control of the writeback in progress 8502a814908STejun Heo * @page: page being written out 8512a814908STejun Heo * @bytes: number of bytes being written out 8522a814908STejun Heo * 8532a814908STejun Heo * @bytes from @page are about to written out during the writeback 8542a814908STejun Heo * controlled by @wbc. Keep the book for foreign inode detection. See 8552a814908STejun Heo * wbc_detach_inode(). 8562a814908STejun Heo */ 85734e51a5eSTejun Heo void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page, 8582a814908STejun Heo size_t bytes) 8592a814908STejun Heo { 86066311422STejun Heo struct cgroup_subsys_state *css; 8612a814908STejun Heo int id; 8622a814908STejun Heo 8632a814908STejun Heo /* 8642a814908STejun Heo * pageout() path doesn't attach @wbc to the inode being written 8652a814908STejun Heo * out. This is intentional as we don't want the function to block 8662a814908STejun Heo * behind a slow cgroup. Ultimately, we want pageout() to kick off 8672a814908STejun Heo * regular writeback instead of writing things out itself. 8682a814908STejun Heo */ 86927b36d8fSTejun Heo if (!wbc->wb || wbc->no_cgroup_owner) 8702a814908STejun Heo return; 8712a814908STejun Heo 87266311422STejun Heo css = mem_cgroup_css_from_page(page); 87366311422STejun Heo /* dead cgroups shouldn't contribute to inode ownership arbitration */ 87466311422STejun Heo if (!(css->flags & CSS_ONLINE)) 87566311422STejun Heo return; 87666311422STejun Heo 87766311422STejun Heo id = css->id; 8782a814908STejun Heo 8792a814908STejun Heo if (id == wbc->wb_id) { 8802a814908STejun Heo wbc->wb_bytes += bytes; 8812a814908STejun Heo return; 8822a814908STejun Heo } 8832a814908STejun Heo 8842a814908STejun Heo if (id == wbc->wb_lcand_id) 8852a814908STejun Heo wbc->wb_lcand_bytes += bytes; 8862a814908STejun Heo 8872a814908STejun Heo /* Boyer-Moore majority vote algorithm */ 8882a814908STejun Heo if (!wbc->wb_tcand_bytes) 8892a814908STejun Heo wbc->wb_tcand_id = id; 8902a814908STejun Heo if (id == wbc->wb_tcand_id) 8912a814908STejun Heo wbc->wb_tcand_bytes += bytes; 8922a814908STejun Heo else 8932a814908STejun Heo wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); 8942a814908STejun Heo } 89534e51a5eSTejun Heo EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner); 8962a814908STejun Heo 8972a814908STejun Heo /** 898f2b65121STejun Heo * wb_split_bdi_pages - split nr_pages to write according to bandwidth 899f2b65121STejun Heo * @wb: target bdi_writeback to split @nr_pages to 900f2b65121STejun Heo * @nr_pages: number of pages to write for the whole bdi 901f2b65121STejun Heo * 902f2b65121STejun Heo * Split @wb's portion of @nr_pages according to @wb's write bandwidth in 903f2b65121STejun Heo * relation to the total write bandwidth of all wb's w/ dirty inodes on 904f2b65121STejun Heo * @wb->bdi. 905f2b65121STejun Heo */ 906f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 907f2b65121STejun Heo { 908f2b65121STejun Heo unsigned long this_bw = wb->avg_write_bandwidth; 909f2b65121STejun Heo unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); 910f2b65121STejun Heo 911f2b65121STejun Heo if (nr_pages == LONG_MAX) 912f2b65121STejun Heo return LONG_MAX; 913f2b65121STejun Heo 914f2b65121STejun Heo /* 915f2b65121STejun Heo * This may be called on clean wb's and proportional distribution 916f2b65121STejun Heo * may not make sense, just use the original @nr_pages in those 917f2b65121STejun Heo * cases. In general, we wanna err on the side of writing more. 918f2b65121STejun Heo */ 919f2b65121STejun Heo if (!tot_bw || this_bw >= tot_bw) 920f2b65121STejun Heo return nr_pages; 921f2b65121STejun Heo else 922f2b65121STejun Heo return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); 923f2b65121STejun Heo } 924f2b65121STejun Heo 925db125360STejun Heo /** 926db125360STejun Heo * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi 927db125360STejun Heo * @bdi: target backing_dev_info 928db125360STejun Heo * @base_work: wb_writeback_work to issue 929db125360STejun Heo * @skip_if_busy: skip wb's which already have writeback in progress 930db125360STejun Heo * 931db125360STejun Heo * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which 932db125360STejun Heo * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's 933db125360STejun Heo * distributed to the busy wbs according to each wb's proportion in the 934db125360STejun Heo * total active write bandwidth of @bdi. 935db125360STejun Heo */ 936db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 937db125360STejun Heo struct wb_writeback_work *base_work, 938db125360STejun Heo bool skip_if_busy) 939db125360STejun Heo { 940b817525aSTejun Heo struct bdi_writeback *last_wb = NULL; 941b33e18f6STejun Heo struct bdi_writeback *wb = list_entry(&bdi->wb_list, 942b817525aSTejun Heo struct bdi_writeback, bdi_node); 943db125360STejun Heo 944db125360STejun Heo might_sleep(); 945db125360STejun Heo restart: 946db125360STejun Heo rcu_read_lock(); 947b817525aSTejun Heo list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) { 9485b9cce4cSTejun Heo DEFINE_WB_COMPLETION(fallback_work_done, bdi); 9498a1270cdSTejun Heo struct wb_writeback_work fallback_work; 9508a1270cdSTejun Heo struct wb_writeback_work *work; 9518a1270cdSTejun Heo long nr_pages; 9528a1270cdSTejun Heo 953b817525aSTejun Heo if (last_wb) { 954b817525aSTejun Heo wb_put(last_wb); 955b817525aSTejun Heo last_wb = NULL; 956b817525aSTejun Heo } 957b817525aSTejun Heo 958006a0973STejun Heo /* SYNC_ALL writes out I_DIRTY_TIME too */ 959006a0973STejun Heo if (!wb_has_dirty_io(wb) && 960006a0973STejun Heo (base_work->sync_mode == WB_SYNC_NONE || 961006a0973STejun Heo list_empty(&wb->b_dirty_time))) 962006a0973STejun Heo continue; 963006a0973STejun Heo if (skip_if_busy && writeback_in_progress(wb)) 964db125360STejun Heo continue; 965db125360STejun Heo 9668a1270cdSTejun Heo nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages); 9678a1270cdSTejun Heo 9688a1270cdSTejun Heo work = kmalloc(sizeof(*work), GFP_ATOMIC); 9698a1270cdSTejun Heo if (work) { 9708a1270cdSTejun Heo *work = *base_work; 9718a1270cdSTejun Heo work->nr_pages = nr_pages; 9728a1270cdSTejun Heo work->auto_free = 1; 9738a1270cdSTejun Heo wb_queue_work(wb, work); 9748a1270cdSTejun Heo continue; 975db125360STejun Heo } 9768a1270cdSTejun Heo 9778a1270cdSTejun Heo /* alloc failed, execute synchronously using on-stack fallback */ 9788a1270cdSTejun Heo work = &fallback_work; 9798a1270cdSTejun Heo *work = *base_work; 9808a1270cdSTejun Heo work->nr_pages = nr_pages; 9818a1270cdSTejun Heo work->auto_free = 0; 9828a1270cdSTejun Heo work->done = &fallback_work_done; 9838a1270cdSTejun Heo 9848a1270cdSTejun Heo wb_queue_work(wb, work); 9858a1270cdSTejun Heo 986b817525aSTejun Heo /* 987b817525aSTejun Heo * Pin @wb so that it stays on @bdi->wb_list. This allows 988b817525aSTejun Heo * continuing iteration from @wb after dropping and 989b817525aSTejun Heo * regrabbing rcu read lock. 990b817525aSTejun Heo */ 991b817525aSTejun Heo wb_get(wb); 992b817525aSTejun Heo last_wb = wb; 993b817525aSTejun Heo 994db125360STejun Heo rcu_read_unlock(); 9955b9cce4cSTejun Heo wb_wait_for_completion(&fallback_work_done); 996db125360STejun Heo goto restart; 997db125360STejun Heo } 998db125360STejun Heo rcu_read_unlock(); 999b817525aSTejun Heo 1000b817525aSTejun Heo if (last_wb) 1001b817525aSTejun Heo wb_put(last_wb); 1002db125360STejun Heo } 1003db125360STejun Heo 1004a1a0e23eSTejun Heo /** 1005d62241c7STejun Heo * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs 1006d62241c7STejun Heo * @bdi_id: target bdi id 1007d62241c7STejun Heo * @memcg_id: target memcg css id 1008d62241c7STejun Heo * @reason: reason why some writeback work initiated 1009d62241c7STejun Heo * @done: target wb_completion 1010d62241c7STejun Heo * 1011d62241c7STejun Heo * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id 1012d62241c7STejun Heo * with the specified parameters. 1013d62241c7STejun Heo */ 10147490a2d2SShakeel Butt int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, 1015d62241c7STejun Heo enum wb_reason reason, struct wb_completion *done) 1016d62241c7STejun Heo { 1017d62241c7STejun Heo struct backing_dev_info *bdi; 1018d62241c7STejun Heo struct cgroup_subsys_state *memcg_css; 1019d62241c7STejun Heo struct bdi_writeback *wb; 1020d62241c7STejun Heo struct wb_writeback_work *work; 10217490a2d2SShakeel Butt unsigned long dirty; 1022d62241c7STejun Heo int ret; 1023d62241c7STejun Heo 1024d62241c7STejun Heo /* lookup bdi and memcg */ 1025d62241c7STejun Heo bdi = bdi_get_by_id(bdi_id); 1026d62241c7STejun Heo if (!bdi) 1027d62241c7STejun Heo return -ENOENT; 1028d62241c7STejun Heo 1029d62241c7STejun Heo rcu_read_lock(); 1030d62241c7STejun Heo memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys); 1031d62241c7STejun Heo if (memcg_css && !css_tryget(memcg_css)) 1032d62241c7STejun Heo memcg_css = NULL; 1033d62241c7STejun Heo rcu_read_unlock(); 1034d62241c7STejun Heo if (!memcg_css) { 1035d62241c7STejun Heo ret = -ENOENT; 1036d62241c7STejun Heo goto out_bdi_put; 1037d62241c7STejun Heo } 1038d62241c7STejun Heo 1039d62241c7STejun Heo /* 1040d62241c7STejun Heo * And find the associated wb. If the wb isn't there already 1041d62241c7STejun Heo * there's nothing to flush, don't create one. 1042d62241c7STejun Heo */ 1043d62241c7STejun Heo wb = wb_get_lookup(bdi, memcg_css); 1044d62241c7STejun Heo if (!wb) { 1045d62241c7STejun Heo ret = -ENOENT; 1046d62241c7STejun Heo goto out_css_put; 1047d62241c7STejun Heo } 1048d62241c7STejun Heo 1049d62241c7STejun Heo /* 10507490a2d2SShakeel Butt * The caller is attempting to write out most of 1051d62241c7STejun Heo * the currently dirty pages. Let's take the current dirty page 1052d62241c7STejun Heo * count and inflate it by 25% which should be large enough to 1053d62241c7STejun Heo * flush out most dirty pages while avoiding getting livelocked by 1054d62241c7STejun Heo * concurrent dirtiers. 10557490a2d2SShakeel Butt * 10567490a2d2SShakeel Butt * BTW the memcg stats are flushed periodically and this is best-effort 10577490a2d2SShakeel Butt * estimation, so some potential error is ok. 1058d62241c7STejun Heo */ 10597490a2d2SShakeel Butt dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY); 10607490a2d2SShakeel Butt dirty = dirty * 10 / 8; 1061d62241c7STejun Heo 1062d62241c7STejun Heo /* issue the writeback work */ 1063d62241c7STejun Heo work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN); 1064d62241c7STejun Heo if (work) { 10657490a2d2SShakeel Butt work->nr_pages = dirty; 1066d62241c7STejun Heo work->sync_mode = WB_SYNC_NONE; 1067d62241c7STejun Heo work->range_cyclic = 1; 1068d62241c7STejun Heo work->reason = reason; 1069d62241c7STejun Heo work->done = done; 1070d62241c7STejun Heo work->auto_free = 1; 1071d62241c7STejun Heo wb_queue_work(wb, work); 1072d62241c7STejun Heo ret = 0; 1073d62241c7STejun Heo } else { 1074d62241c7STejun Heo ret = -ENOMEM; 1075d62241c7STejun Heo } 1076d62241c7STejun Heo 1077d62241c7STejun Heo wb_put(wb); 1078d62241c7STejun Heo out_css_put: 1079d62241c7STejun Heo css_put(memcg_css); 1080d62241c7STejun Heo out_bdi_put: 1081d62241c7STejun Heo bdi_put(bdi); 1082d62241c7STejun Heo return ret; 1083d62241c7STejun Heo } 1084d62241c7STejun Heo 1085d62241c7STejun Heo /** 1086a1a0e23eSTejun Heo * cgroup_writeback_umount - flush inode wb switches for umount 1087a1a0e23eSTejun Heo * 1088a1a0e23eSTejun Heo * This function is called when a super_block is about to be destroyed and 1089a1a0e23eSTejun Heo * flushes in-flight inode wb switches. An inode wb switch goes through 1090a1a0e23eSTejun Heo * RCU and then workqueue, so the two need to be flushed in order to ensure 1091a1a0e23eSTejun Heo * that all previously scheduled switches are finished. As wb switches are 1092a1a0e23eSTejun Heo * rare occurrences and synchronize_rcu() can take a while, perform 1093a1a0e23eSTejun Heo * flushing iff wb switches are in flight. 1094a1a0e23eSTejun Heo */ 1095a1a0e23eSTejun Heo void cgroup_writeback_umount(void) 1096a1a0e23eSTejun Heo { 1097592fa002SRoman Gushchin /* 1098592fa002SRoman Gushchin * SB_ACTIVE should be reliably cleared before checking 1099592fa002SRoman Gushchin * isw_nr_in_flight, see generic_shutdown_super(). 1100592fa002SRoman Gushchin */ 1101592fa002SRoman Gushchin smp_mb(); 1102592fa002SRoman Gushchin 1103a1a0e23eSTejun Heo if (atomic_read(&isw_nr_in_flight)) { 1104ec084de9SJiufei Xue /* 1105ec084de9SJiufei Xue * Use rcu_barrier() to wait for all pending callbacks to 1106ec084de9SJiufei Xue * ensure that all in-flight wb switches are in the workqueue. 1107ec084de9SJiufei Xue */ 1108ec084de9SJiufei Xue rcu_barrier(); 1109a1a0e23eSTejun Heo flush_workqueue(isw_wq); 1110a1a0e23eSTejun Heo } 1111a1a0e23eSTejun Heo } 1112a1a0e23eSTejun Heo 1113a1a0e23eSTejun Heo static int __init cgroup_writeback_init(void) 1114a1a0e23eSTejun Heo { 1115a1a0e23eSTejun Heo isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0); 1116a1a0e23eSTejun Heo if (!isw_wq) 1117a1a0e23eSTejun Heo return -ENOMEM; 1118a1a0e23eSTejun Heo return 0; 1119a1a0e23eSTejun Heo } 1120a1a0e23eSTejun Heo fs_initcall(cgroup_writeback_init); 1121a1a0e23eSTejun Heo 1122f2b65121STejun Heo #else /* CONFIG_CGROUP_WRITEBACK */ 1123f2b65121STejun Heo 11247fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 11257fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 11267fc5854fSTejun Heo 1127f3b6a6dfSRoman Gushchin static void inode_cgwb_move_to_attached(struct inode *inode, 1128f3b6a6dfSRoman Gushchin struct bdi_writeback *wb) 1129f3b6a6dfSRoman Gushchin { 1130f3b6a6dfSRoman Gushchin assert_spin_locked(&wb->list_lock); 1131f3b6a6dfSRoman Gushchin assert_spin_locked(&inode->i_lock); 1132f3b6a6dfSRoman Gushchin 1133f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 1134f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 1135f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 1136f3b6a6dfSRoman Gushchin } 1137f3b6a6dfSRoman Gushchin 113887e1d789STejun Heo static struct bdi_writeback * 113987e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 114087e1d789STejun Heo __releases(&inode->i_lock) 114187e1d789STejun Heo __acquires(&wb->list_lock) 114287e1d789STejun Heo { 114387e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 114487e1d789STejun Heo 114587e1d789STejun Heo spin_unlock(&inode->i_lock); 114687e1d789STejun Heo spin_lock(&wb->list_lock); 114787e1d789STejun Heo return wb; 114887e1d789STejun Heo } 114987e1d789STejun Heo 115087e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 115187e1d789STejun Heo __acquires(&wb->list_lock) 115287e1d789STejun Heo { 115387e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 115487e1d789STejun Heo 115587e1d789STejun Heo spin_lock(&wb->list_lock); 115687e1d789STejun Heo return wb; 115787e1d789STejun Heo } 115887e1d789STejun Heo 1159f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 1160f2b65121STejun Heo { 1161f2b65121STejun Heo return nr_pages; 1162f2b65121STejun Heo } 1163f2b65121STejun Heo 1164db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 1165db125360STejun Heo struct wb_writeback_work *base_work, 1166db125360STejun Heo bool skip_if_busy) 1167db125360STejun Heo { 1168db125360STejun Heo might_sleep(); 1169db125360STejun Heo 1170006a0973STejun Heo if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) { 1171db125360STejun Heo base_work->auto_free = 0; 1172db125360STejun Heo wb_queue_work(&bdi->wb, base_work); 1173db125360STejun Heo } 1174db125360STejun Heo } 1175db125360STejun Heo 1176703c2708STejun Heo #endif /* CONFIG_CGROUP_WRITEBACK */ 1177703c2708STejun Heo 1178e8e8a0c6SJens Axboe /* 1179e8e8a0c6SJens Axboe * Add in the number of potentially dirty inodes, because each inode 1180e8e8a0c6SJens Axboe * write can dirty pagecache in the underlying blockdev. 1181e8e8a0c6SJens Axboe */ 1182e8e8a0c6SJens Axboe static unsigned long get_nr_dirty_pages(void) 1183e8e8a0c6SJens Axboe { 1184e8e8a0c6SJens Axboe return global_node_page_state(NR_FILE_DIRTY) + 1185e8e8a0c6SJens Axboe get_nr_dirty_inodes(); 1186e8e8a0c6SJens Axboe } 1187e8e8a0c6SJens Axboe 1188e8e8a0c6SJens Axboe static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason) 1189b6e51316SJens Axboe { 1190c00ddad3STejun Heo if (!wb_has_dirty_io(wb)) 1191c00ddad3STejun Heo return; 1192c00ddad3STejun Heo 1193c00ddad3STejun Heo /* 1194aac8d41cSJens Axboe * All callers of this function want to start writeback of all 1195aac8d41cSJens Axboe * dirty pages. Places like vmscan can call this at a very 1196aac8d41cSJens Axboe * high frequency, causing pointless allocations of tons of 1197aac8d41cSJens Axboe * work items and keeping the flusher threads busy retrieving 1198aac8d41cSJens Axboe * that work. Ensure that we only allow one of them pending and 119985009b4fSJens Axboe * inflight at the time. 1200aac8d41cSJens Axboe */ 120185009b4fSJens Axboe if (test_bit(WB_start_all, &wb->state) || 120285009b4fSJens Axboe test_and_set_bit(WB_start_all, &wb->state)) 1203aac8d41cSJens Axboe return; 1204aac8d41cSJens Axboe 120585009b4fSJens Axboe wb->start_all_reason = reason; 1206c00ddad3STejun Heo wb_wakeup(wb); 1207d3ddec76SWu Fengguang } 1208d3ddec76SWu Fengguang 1209c5444198SChristoph Hellwig /** 12109ecf4866STejun Heo * wb_start_background_writeback - start background writeback 12119ecf4866STejun Heo * @wb: bdi_writback to write from 1212c5444198SChristoph Hellwig * 1213c5444198SChristoph Hellwig * Description: 12146585027aSJan Kara * This makes sure WB_SYNC_NONE background writeback happens. When 12159ecf4866STejun Heo * this function returns, it is only guaranteed that for given wb 12166585027aSJan Kara * some IO is happening if we are over background dirty threshold. 12176585027aSJan Kara * Caller need not hold sb s_umount semaphore. 1218c5444198SChristoph Hellwig */ 12199ecf4866STejun Heo void wb_start_background_writeback(struct bdi_writeback *wb) 1220c5444198SChristoph Hellwig { 12216585027aSJan Kara /* 12226585027aSJan Kara * We just wake up the flusher thread. It will perform background 12236585027aSJan Kara * writeback as soon as there is no other work to do. 12246585027aSJan Kara */ 12255634cc2aSTejun Heo trace_writeback_wake_background(wb); 12269ecf4866STejun Heo wb_wakeup(wb); 12271da177e4SLinus Torvalds } 12281da177e4SLinus Torvalds 12291da177e4SLinus Torvalds /* 1230a66979abSDave Chinner * Remove the inode from the writeback list it is on. 1231a66979abSDave Chinner */ 1232c7f54084SDave Chinner void inode_io_list_del(struct inode *inode) 1233a66979abSDave Chinner { 123487e1d789STejun Heo struct bdi_writeback *wb; 1235a66979abSDave Chinner 123687e1d789STejun Heo wb = inode_to_wb_and_lock_list(inode); 1237b35250c0SJan Kara spin_lock(&inode->i_lock); 1238f3b6a6dfSRoman Gushchin 1239f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 1240f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 1241f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 1242f3b6a6dfSRoman Gushchin 1243b35250c0SJan Kara spin_unlock(&inode->i_lock); 124452ebea74STejun Heo spin_unlock(&wb->list_lock); 1245f758eeabSChristoph Hellwig } 12464301efa4SJan Kara EXPORT_SYMBOL(inode_io_list_del); 1247a66979abSDave Chinner 1248a66979abSDave Chinner /* 12496c60d2b5SDave Chinner * mark an inode as under writeback on the sb 12506c60d2b5SDave Chinner */ 12516c60d2b5SDave Chinner void sb_mark_inode_writeback(struct inode *inode) 12526c60d2b5SDave Chinner { 12536c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 12546c60d2b5SDave Chinner unsigned long flags; 12556c60d2b5SDave Chinner 12566c60d2b5SDave Chinner if (list_empty(&inode->i_wb_list)) { 12576c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 12589a46b04fSBrian Foster if (list_empty(&inode->i_wb_list)) { 12596c60d2b5SDave Chinner list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); 12609a46b04fSBrian Foster trace_sb_mark_inode_writeback(inode); 12619a46b04fSBrian Foster } 12626c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 12636c60d2b5SDave Chinner } 12646c60d2b5SDave Chinner } 12656c60d2b5SDave Chinner 12666c60d2b5SDave Chinner /* 12676c60d2b5SDave Chinner * clear an inode as under writeback on the sb 12686c60d2b5SDave Chinner */ 12696c60d2b5SDave Chinner void sb_clear_inode_writeback(struct inode *inode) 12706c60d2b5SDave Chinner { 12716c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 12726c60d2b5SDave Chinner unsigned long flags; 12736c60d2b5SDave Chinner 12746c60d2b5SDave Chinner if (!list_empty(&inode->i_wb_list)) { 12756c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 12769a46b04fSBrian Foster if (!list_empty(&inode->i_wb_list)) { 12776c60d2b5SDave Chinner list_del_init(&inode->i_wb_list); 12789a46b04fSBrian Foster trace_sb_clear_inode_writeback(inode); 12799a46b04fSBrian Foster } 12806c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 12816c60d2b5SDave Chinner } 12826c60d2b5SDave Chinner } 12836c60d2b5SDave Chinner 12846c60d2b5SDave Chinner /* 12856610a0bcSAndrew Morton * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 12866610a0bcSAndrew Morton * furthest end of its superblock's dirty-inode list. 12876610a0bcSAndrew Morton * 12886610a0bcSAndrew Morton * Before stamping the inode's ->dirtied_when, we check to see whether it is 128966f3b8e2SJens Axboe * already the most-recently-dirtied inode on the b_dirty list. If that is 12906610a0bcSAndrew Morton * the case then the inode must have been redirtied while it was being written 12916610a0bcSAndrew Morton * out and we don't reset its dirtied_when. 12926610a0bcSAndrew Morton */ 1293b35250c0SJan Kara static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb) 12946610a0bcSAndrew Morton { 1295b35250c0SJan Kara assert_spin_locked(&inode->i_lock); 1296b35250c0SJan Kara 129703ba3782SJens Axboe if (!list_empty(&wb->b_dirty)) { 129866f3b8e2SJens Axboe struct inode *tail; 12996610a0bcSAndrew Morton 13007ccf19a8SNick Piggin tail = wb_inode(wb->b_dirty.next); 130166f3b8e2SJens Axboe if (time_before(inode->dirtied_when, tail->dirtied_when)) 13026610a0bcSAndrew Morton inode->dirtied_when = jiffies; 13036610a0bcSAndrew Morton } 1304c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty); 13055afced3bSJan Kara inode->i_state &= ~I_SYNC_QUEUED; 13066610a0bcSAndrew Morton } 13076610a0bcSAndrew Morton 1308b35250c0SJan Kara static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) 1309b35250c0SJan Kara { 1310b35250c0SJan Kara spin_lock(&inode->i_lock); 1311b35250c0SJan Kara redirty_tail_locked(inode, wb); 1312b35250c0SJan Kara spin_unlock(&inode->i_lock); 1313b35250c0SJan Kara } 1314b35250c0SJan Kara 13156610a0bcSAndrew Morton /* 131666f3b8e2SJens Axboe * requeue inode for re-scanning after bdi->b_io list is exhausted. 1317c986d1e2SAndrew Morton */ 1318f758eeabSChristoph Hellwig static void requeue_io(struct inode *inode, struct bdi_writeback *wb) 1319c986d1e2SAndrew Morton { 1320c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_more_io); 1321c986d1e2SAndrew Morton } 1322c986d1e2SAndrew Morton 13231c0eeaf5SJoern Engel static void inode_sync_complete(struct inode *inode) 13241c0eeaf5SJoern Engel { 1325365b94aeSJan Kara inode->i_state &= ~I_SYNC; 13264eff96ddSJan Kara /* If inode is clean an unused, put it into LRU now... */ 13274eff96ddSJan Kara inode_add_lru(inode); 1328365b94aeSJan Kara /* Waiters must see I_SYNC cleared before being woken up */ 13291c0eeaf5SJoern Engel smp_mb(); 13301c0eeaf5SJoern Engel wake_up_bit(&inode->i_state, __I_SYNC); 13311c0eeaf5SJoern Engel } 13321c0eeaf5SJoern Engel 1333d2caa3c5SJeff Layton static bool inode_dirtied_after(struct inode *inode, unsigned long t) 1334d2caa3c5SJeff Layton { 1335d2caa3c5SJeff Layton bool ret = time_after(inode->dirtied_when, t); 1336d2caa3c5SJeff Layton #ifndef CONFIG_64BIT 1337d2caa3c5SJeff Layton /* 1338d2caa3c5SJeff Layton * For inodes being constantly redirtied, dirtied_when can get stuck. 1339d2caa3c5SJeff Layton * It _appears_ to be in the future, but is actually in distant past. 1340d2caa3c5SJeff Layton * This test is necessary to prevent such wrapped-around relative times 13415b0830cbSJens Axboe * from permanently stopping the whole bdi writeback. 1342d2caa3c5SJeff Layton */ 1343d2caa3c5SJeff Layton ret = ret && time_before_eq(inode->dirtied_when, jiffies); 1344d2caa3c5SJeff Layton #endif 1345d2caa3c5SJeff Layton return ret; 1346d2caa3c5SJeff Layton } 1347d2caa3c5SJeff Layton 13480ae45f63STheodore Ts'o #define EXPIRE_DIRTY_ATIME 0x0001 13490ae45f63STheodore Ts'o 1350c986d1e2SAndrew Morton /* 1351f9cae926SJan Kara * Move expired (dirtied before dirtied_before) dirty inodes from 1352697e6fedSJan Kara * @delaying_queue to @dispatch_queue. 13532c136579SFengguang Wu */ 1354e84d0a4fSWu Fengguang static int move_expired_inodes(struct list_head *delaying_queue, 13552c136579SFengguang Wu struct list_head *dispatch_queue, 13565fcd5750SJan Kara unsigned long dirtied_before) 13572c136579SFengguang Wu { 13585c03449dSShaohua Li LIST_HEAD(tmp); 13595c03449dSShaohua Li struct list_head *pos, *node; 1360cf137307SJens Axboe struct super_block *sb = NULL; 13615c03449dSShaohua Li struct inode *inode; 1362cf137307SJens Axboe int do_sb_sort = 0; 1363e84d0a4fSWu Fengguang int moved = 0; 13645c03449dSShaohua Li 13652c136579SFengguang Wu while (!list_empty(delaying_queue)) { 13667ccf19a8SNick Piggin inode = wb_inode(delaying_queue->prev); 1367f9cae926SJan Kara if (inode_dirtied_after(inode, dirtied_before)) 13682c136579SFengguang Wu break; 136910e14073SJchao Sun spin_lock(&inode->i_lock); 1370c7f54084SDave Chinner list_move(&inode->i_io_list, &tmp); 1371a8855990SJan Kara moved++; 13725afced3bSJan Kara inode->i_state |= I_SYNC_QUEUED; 13735afced3bSJan Kara spin_unlock(&inode->i_lock); 1374a8855990SJan Kara if (sb_is_blkdev_sb(inode->i_sb)) 1375a8855990SJan Kara continue; 1376cf137307SJens Axboe if (sb && sb != inode->i_sb) 1377cf137307SJens Axboe do_sb_sort = 1; 1378cf137307SJens Axboe sb = inode->i_sb; 13795c03449dSShaohua Li } 13805c03449dSShaohua Li 1381cf137307SJens Axboe /* just one sb in list, splice to dispatch_queue and we're done */ 1382cf137307SJens Axboe if (!do_sb_sort) { 1383cf137307SJens Axboe list_splice(&tmp, dispatch_queue); 1384e84d0a4fSWu Fengguang goto out; 1385cf137307SJens Axboe } 1386cf137307SJens Axboe 138710e14073SJchao Sun /* 138810e14073SJchao Sun * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue', 138910e14073SJchao Sun * we don't take inode->i_lock here because it is just a pointless overhead. 139010e14073SJchao Sun * Inode is already marked as I_SYNC_QUEUED so writeback list handling is 139110e14073SJchao Sun * fully under our control. 139210e14073SJchao Sun */ 13935c03449dSShaohua Li while (!list_empty(&tmp)) { 13947ccf19a8SNick Piggin sb = wb_inode(tmp.prev)->i_sb; 13955c03449dSShaohua Li list_for_each_prev_safe(pos, node, &tmp) { 13967ccf19a8SNick Piggin inode = wb_inode(pos); 13975c03449dSShaohua Li if (inode->i_sb == sb) 1398c7f54084SDave Chinner list_move(&inode->i_io_list, dispatch_queue); 13992c136579SFengguang Wu } 14002c136579SFengguang Wu } 1401e84d0a4fSWu Fengguang out: 1402e84d0a4fSWu Fengguang return moved; 14035c03449dSShaohua Li } 14042c136579SFengguang Wu 14052c136579SFengguang Wu /* 14062c136579SFengguang Wu * Queue all expired dirty inodes for io, eldest first. 14074ea879b9SWu Fengguang * Before 14084ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 14094ea879b9SWu Fengguang * =============> gf edc BA 14104ea879b9SWu Fengguang * After 14114ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 14124ea879b9SWu Fengguang * =============> g fBAedc 14134ea879b9SWu Fengguang * | 14144ea879b9SWu Fengguang * +--> dequeue for IO 14152c136579SFengguang Wu */ 1416f9cae926SJan Kara static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work, 1417f9cae926SJan Kara unsigned long dirtied_before) 14182c136579SFengguang Wu { 1419e84d0a4fSWu Fengguang int moved; 1420f9cae926SJan Kara unsigned long time_expire_jif = dirtied_before; 14210ae45f63STheodore Ts'o 1422f758eeabSChristoph Hellwig assert_spin_locked(&wb->list_lock); 14234ea879b9SWu Fengguang list_splice_init(&wb->b_more_io, &wb->b_io); 14245fcd5750SJan Kara moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before); 1425f9cae926SJan Kara if (!work->for_sync) 1426f9cae926SJan Kara time_expire_jif = jiffies - dirtytime_expire_interval * HZ; 14270ae45f63STheodore Ts'o moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, 14285fcd5750SJan Kara time_expire_jif); 1429d6c10f1fSTejun Heo if (moved) 1430d6c10f1fSTejun Heo wb_io_lists_populated(wb); 1431f9cae926SJan Kara trace_writeback_queue_io(wb, work, dirtied_before, moved); 143266f3b8e2SJens Axboe } 143366f3b8e2SJens Axboe 1434a9185b41SChristoph Hellwig static int write_inode(struct inode *inode, struct writeback_control *wbc) 143566f3b8e2SJens Axboe { 14369fb0a7daSTejun Heo int ret; 14379fb0a7daSTejun Heo 14389fb0a7daSTejun Heo if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { 14399fb0a7daSTejun Heo trace_writeback_write_inode_start(inode, wbc); 14409fb0a7daSTejun Heo ret = inode->i_sb->s_op->write_inode(inode, wbc); 14419fb0a7daSTejun Heo trace_writeback_write_inode(inode, wbc); 14429fb0a7daSTejun Heo return ret; 14439fb0a7daSTejun Heo } 144403ba3782SJens Axboe return 0; 144566f3b8e2SJens Axboe } 144608d8e974SFengguang Wu 14472c136579SFengguang Wu /* 1448169ebd90SJan Kara * Wait for writeback on an inode to complete. Called with i_lock held. 1449169ebd90SJan Kara * Caller must make sure inode cannot go away when we drop i_lock. 145001c03194SChristoph Hellwig */ 1451169ebd90SJan Kara static void __inode_wait_for_writeback(struct inode *inode) 1452169ebd90SJan Kara __releases(inode->i_lock) 1453169ebd90SJan Kara __acquires(inode->i_lock) 145401c03194SChristoph Hellwig { 145501c03194SChristoph Hellwig DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); 145601c03194SChristoph Hellwig wait_queue_head_t *wqh; 145701c03194SChristoph Hellwig 145801c03194SChristoph Hellwig wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 145958a9d3d8SRichard Kennedy while (inode->i_state & I_SYNC) { 1460250df6edSDave Chinner spin_unlock(&inode->i_lock); 146174316201SNeilBrown __wait_on_bit(wqh, &wq, bit_wait, 146274316201SNeilBrown TASK_UNINTERRUPTIBLE); 1463250df6edSDave Chinner spin_lock(&inode->i_lock); 146458a9d3d8SRichard Kennedy } 146501c03194SChristoph Hellwig } 146601c03194SChristoph Hellwig 146701c03194SChristoph Hellwig /* 1468169ebd90SJan Kara * Wait for writeback on an inode to complete. Caller must have inode pinned. 1469169ebd90SJan Kara */ 1470169ebd90SJan Kara void inode_wait_for_writeback(struct inode *inode) 1471169ebd90SJan Kara { 1472169ebd90SJan Kara spin_lock(&inode->i_lock); 1473169ebd90SJan Kara __inode_wait_for_writeback(inode); 1474169ebd90SJan Kara spin_unlock(&inode->i_lock); 1475169ebd90SJan Kara } 1476169ebd90SJan Kara 1477169ebd90SJan Kara /* 1478169ebd90SJan Kara * Sleep until I_SYNC is cleared. This function must be called with i_lock 1479169ebd90SJan Kara * held and drops it. It is aimed for callers not holding any inode reference 1480169ebd90SJan Kara * so once i_lock is dropped, inode can go away. 1481169ebd90SJan Kara */ 1482169ebd90SJan Kara static void inode_sleep_on_writeback(struct inode *inode) 1483169ebd90SJan Kara __releases(inode->i_lock) 1484169ebd90SJan Kara { 1485169ebd90SJan Kara DEFINE_WAIT(wait); 1486169ebd90SJan Kara wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1487169ebd90SJan Kara int sleep; 1488169ebd90SJan Kara 1489169ebd90SJan Kara prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1490169ebd90SJan Kara sleep = inode->i_state & I_SYNC; 1491169ebd90SJan Kara spin_unlock(&inode->i_lock); 1492169ebd90SJan Kara if (sleep) 1493169ebd90SJan Kara schedule(); 1494169ebd90SJan Kara finish_wait(wqh, &wait); 1495169ebd90SJan Kara } 1496169ebd90SJan Kara 1497169ebd90SJan Kara /* 1498ccb26b5aSJan Kara * Find proper writeback list for the inode depending on its current state and 1499ccb26b5aSJan Kara * possibly also change of its state while we were doing writeback. Here we 1500ccb26b5aSJan Kara * handle things such as livelock prevention or fairness of writeback among 1501ccb26b5aSJan Kara * inodes. This function can be called only by flusher thread - noone else 1502ccb26b5aSJan Kara * processes all inodes in writeback lists and requeueing inodes behind flusher 1503ccb26b5aSJan Kara * thread's back can have unexpected consequences. 1504ccb26b5aSJan Kara */ 1505ccb26b5aSJan Kara static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, 1506ccb26b5aSJan Kara struct writeback_control *wbc) 1507ccb26b5aSJan Kara { 1508ccb26b5aSJan Kara if (inode->i_state & I_FREEING) 1509ccb26b5aSJan Kara return; 1510ccb26b5aSJan Kara 1511ccb26b5aSJan Kara /* 1512ccb26b5aSJan Kara * Sync livelock prevention. Each inode is tagged and synced in one 1513ccb26b5aSJan Kara * shot. If still dirty, it will be redirty_tail()'ed below. Update 1514ccb26b5aSJan Kara * the dirty time to prevent enqueue and sync it again. 1515ccb26b5aSJan Kara */ 1516ccb26b5aSJan Kara if ((inode->i_state & I_DIRTY) && 1517ccb26b5aSJan Kara (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) 1518ccb26b5aSJan Kara inode->dirtied_when = jiffies; 1519ccb26b5aSJan Kara 15204f8ad655SJan Kara if (wbc->pages_skipped) { 15214f8ad655SJan Kara /* 15224f8ad655SJan Kara * writeback is not making progress due to locked 15234f8ad655SJan Kara * buffers. Skip this inode for now. 15244f8ad655SJan Kara */ 1525b35250c0SJan Kara redirty_tail_locked(inode, wb); 15264f8ad655SJan Kara return; 15274f8ad655SJan Kara } 15284f8ad655SJan Kara 1529ccb26b5aSJan Kara if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { 1530ccb26b5aSJan Kara /* 1531ccb26b5aSJan Kara * We didn't write back all the pages. nfs_writepages() 1532ccb26b5aSJan Kara * sometimes bales out without doing anything. 1533ccb26b5aSJan Kara */ 1534ccb26b5aSJan Kara if (wbc->nr_to_write <= 0) { 1535ccb26b5aSJan Kara /* Slice used up. Queue for next turn. */ 1536ccb26b5aSJan Kara requeue_io(inode, wb); 1537ccb26b5aSJan Kara } else { 1538ccb26b5aSJan Kara /* 1539ccb26b5aSJan Kara * Writeback blocked by something other than 1540ccb26b5aSJan Kara * congestion. Delay the inode for some time to 1541ccb26b5aSJan Kara * avoid spinning on the CPU (100% iowait) 1542ccb26b5aSJan Kara * retrying writeback of the dirty page/inode 1543ccb26b5aSJan Kara * that cannot be performed immediately. 1544ccb26b5aSJan Kara */ 1545b35250c0SJan Kara redirty_tail_locked(inode, wb); 1546ccb26b5aSJan Kara } 1547ccb26b5aSJan Kara } else if (inode->i_state & I_DIRTY) { 1548ccb26b5aSJan Kara /* 1549ccb26b5aSJan Kara * Filesystems can dirty the inode during writeback operations, 1550ccb26b5aSJan Kara * such as delayed allocation during submission or metadata 1551ccb26b5aSJan Kara * updates after data IO completion. 1552ccb26b5aSJan Kara */ 1553b35250c0SJan Kara redirty_tail_locked(inode, wb); 15540ae45f63STheodore Ts'o } else if (inode->i_state & I_DIRTY_TIME) { 1555a2f48706STheodore Ts'o inode->dirtied_when = jiffies; 1556c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty_time); 15575afced3bSJan Kara inode->i_state &= ~I_SYNC_QUEUED; 1558ccb26b5aSJan Kara } else { 1559ccb26b5aSJan Kara /* The inode is clean. Remove from writeback lists. */ 1560f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, wb); 1561ccb26b5aSJan Kara } 1562ccb26b5aSJan Kara } 1563ccb26b5aSJan Kara 1564ccb26b5aSJan Kara /* 1565da0c4c60SEric Biggers * Write out an inode and its dirty pages (or some of its dirty pages, depending 1566da0c4c60SEric Biggers * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state. 1567da0c4c60SEric Biggers * 1568da0c4c60SEric Biggers * This doesn't remove the inode from the writeback list it is on, except 1569da0c4c60SEric Biggers * potentially to move it from b_dirty_time to b_dirty due to timestamp 1570da0c4c60SEric Biggers * expiration. The caller is otherwise responsible for writeback list handling. 1571da0c4c60SEric Biggers * 1572da0c4c60SEric Biggers * The caller is also responsible for setting the I_SYNC flag beforehand and 1573da0c4c60SEric Biggers * calling inode_sync_complete() to clear it afterwards. 15741da177e4SLinus Torvalds */ 15751da177e4SLinus Torvalds static int 1576cd8ed2a4SYan Hong __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 15771da177e4SLinus Torvalds { 15781da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 1579251d6a47SWu Fengguang long nr_to_write = wbc->nr_to_write; 158001c03194SChristoph Hellwig unsigned dirty; 15811da177e4SLinus Torvalds int ret; 15821da177e4SLinus Torvalds 15834f8ad655SJan Kara WARN_ON(!(inode->i_state & I_SYNC)); 15841da177e4SLinus Torvalds 15859fb0a7daSTejun Heo trace_writeback_single_inode_start(inode, wbc, nr_to_write); 15869fb0a7daSTejun Heo 15871da177e4SLinus Torvalds ret = do_writepages(mapping, wbc); 15881da177e4SLinus Torvalds 158926821ed4SChristoph Hellwig /* 159026821ed4SChristoph Hellwig * Make sure to wait on the data before writing out the metadata. 159126821ed4SChristoph Hellwig * This is important for filesystems that modify metadata on data 15927747bd4bSDave Chinner * I/O completion. We don't do it for sync(2) writeback because it has a 15937747bd4bSDave Chinner * separate, external IO completion path and ->sync_fs for guaranteeing 15947747bd4bSDave Chinner * inode metadata is written back correctly. 159526821ed4SChristoph Hellwig */ 15967747bd4bSDave Chinner if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { 159726821ed4SChristoph Hellwig int err = filemap_fdatawait(mapping); 15981da177e4SLinus Torvalds if (ret == 0) 15991da177e4SLinus Torvalds ret = err; 16001da177e4SLinus Torvalds } 16011da177e4SLinus Torvalds 16025547e8aaSDmitry Monakhov /* 16031e249cb5SEric Biggers * If the inode has dirty timestamps and we need to write them, call 16041e249cb5SEric Biggers * mark_inode_dirty_sync() to notify the filesystem about it and to 16051e249cb5SEric Biggers * change I_DIRTY_TIME into I_DIRTY_SYNC. 16061e249cb5SEric Biggers */ 16071e249cb5SEric Biggers if ((inode->i_state & I_DIRTY_TIME) && 160883dc881dSEric Biggers (wbc->sync_mode == WB_SYNC_ALL || 16091e249cb5SEric Biggers time_after(jiffies, inode->dirtied_time_when + 16101e249cb5SEric Biggers dirtytime_expire_interval * HZ))) { 16111e249cb5SEric Biggers trace_writeback_lazytime(inode); 16121e249cb5SEric Biggers mark_inode_dirty_sync(inode); 16131e249cb5SEric Biggers } 16141e249cb5SEric Biggers 16151e249cb5SEric Biggers /* 1616da0c4c60SEric Biggers * Get and clear the dirty flags from i_state. This needs to be done 1617da0c4c60SEric Biggers * after calling writepages because some filesystems may redirty the 1618da0c4c60SEric Biggers * inode during writepages due to delalloc. It also needs to be done 1619da0c4c60SEric Biggers * after handling timestamp expiration, as that may dirty the inode too. 16205547e8aaSDmitry Monakhov */ 1621250df6edSDave Chinner spin_lock(&inode->i_lock); 16225547e8aaSDmitry Monakhov dirty = inode->i_state & I_DIRTY; 16230ae45f63STheodore Ts'o inode->i_state &= ~dirty; 16249c6ac78eSTejun Heo 16259c6ac78eSTejun Heo /* 16269c6ac78eSTejun Heo * Paired with smp_mb() in __mark_inode_dirty(). This allows 16279c6ac78eSTejun Heo * __mark_inode_dirty() to test i_state without grabbing i_lock - 16289c6ac78eSTejun Heo * either they see the I_DIRTY bits cleared or we see the dirtied 16299c6ac78eSTejun Heo * inode. 16309c6ac78eSTejun Heo * 16319c6ac78eSTejun Heo * I_DIRTY_PAGES is always cleared together above even if @mapping 16329c6ac78eSTejun Heo * still has dirty pages. The flag is reinstated after smp_mb() if 16339c6ac78eSTejun Heo * necessary. This guarantees that either __mark_inode_dirty() 16349c6ac78eSTejun Heo * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. 16359c6ac78eSTejun Heo */ 16369c6ac78eSTejun Heo smp_mb(); 16379c6ac78eSTejun Heo 16389c6ac78eSTejun Heo if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 16399c6ac78eSTejun Heo inode->i_state |= I_DIRTY_PAGES; 164008276bdaSDavid Howells else if (unlikely(inode->i_state & I_PINNING_FSCACHE_WB)) { 164108276bdaSDavid Howells if (!(inode->i_state & I_DIRTY_PAGES)) { 164208276bdaSDavid Howells inode->i_state &= ~I_PINNING_FSCACHE_WB; 164308276bdaSDavid Howells wbc->unpinned_fscache_wb = true; 164408276bdaSDavid Howells dirty |= I_PINNING_FSCACHE_WB; /* Cause write_inode */ 164508276bdaSDavid Howells } 164608276bdaSDavid Howells } 16479c6ac78eSTejun Heo 1648250df6edSDave Chinner spin_unlock(&inode->i_lock); 16499c6ac78eSTejun Heo 165026821ed4SChristoph Hellwig /* Don't write the inode if only I_DIRTY_PAGES was set */ 16510ae45f63STheodore Ts'o if (dirty & ~I_DIRTY_PAGES) { 1652a9185b41SChristoph Hellwig int err = write_inode(inode, wbc); 16531da177e4SLinus Torvalds if (ret == 0) 16541da177e4SLinus Torvalds ret = err; 16551da177e4SLinus Torvalds } 165608276bdaSDavid Howells wbc->unpinned_fscache_wb = false; 16574f8ad655SJan Kara trace_writeback_single_inode(inode, wbc, nr_to_write); 16584f8ad655SJan Kara return ret; 16594f8ad655SJan Kara } 16604f8ad655SJan Kara 16614f8ad655SJan Kara /* 1662da0c4c60SEric Biggers * Write out an inode's dirty data and metadata on-demand, i.e. separately from 1663da0c4c60SEric Biggers * the regular batched writeback done by the flusher threads in 1664da0c4c60SEric Biggers * writeback_sb_inodes(). @wbc controls various aspects of the write, such as 1665da0c4c60SEric Biggers * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE). 16664f8ad655SJan Kara * 1667da0c4c60SEric Biggers * To prevent the inode from going away, either the caller must have a reference 1668da0c4c60SEric Biggers * to the inode, or the inode must have I_WILL_FREE or I_FREEING set. 16694f8ad655SJan Kara */ 1670aaf25593STejun Heo static int writeback_single_inode(struct inode *inode, 16714f8ad655SJan Kara struct writeback_control *wbc) 16724f8ad655SJan Kara { 1673aaf25593STejun Heo struct bdi_writeback *wb; 16744f8ad655SJan Kara int ret = 0; 16754f8ad655SJan Kara 16764f8ad655SJan Kara spin_lock(&inode->i_lock); 16774f8ad655SJan Kara if (!atomic_read(&inode->i_count)) 16784f8ad655SJan Kara WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 16794f8ad655SJan Kara else 16804f8ad655SJan Kara WARN_ON(inode->i_state & I_WILL_FREE); 16814f8ad655SJan Kara 16824f8ad655SJan Kara if (inode->i_state & I_SYNC) { 1683da0c4c60SEric Biggers /* 1684da0c4c60SEric Biggers * Writeback is already running on the inode. For WB_SYNC_NONE, 1685da0c4c60SEric Biggers * that's enough and we can just return. For WB_SYNC_ALL, we 1686da0c4c60SEric Biggers * must wait for the existing writeback to complete, then do 1687da0c4c60SEric Biggers * writeback again if there's anything left. 1688da0c4c60SEric Biggers */ 16894f8ad655SJan Kara if (wbc->sync_mode != WB_SYNC_ALL) 16904f8ad655SJan Kara goto out; 1691169ebd90SJan Kara __inode_wait_for_writeback(inode); 16924f8ad655SJan Kara } 16934f8ad655SJan Kara WARN_ON(inode->i_state & I_SYNC); 16944f8ad655SJan Kara /* 1695da0c4c60SEric Biggers * If the inode is already fully clean, then there's nothing to do. 1696da0c4c60SEric Biggers * 1697da0c4c60SEric Biggers * For data-integrity syncs we also need to check whether any pages are 1698da0c4c60SEric Biggers * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If 1699da0c4c60SEric Biggers * there are any such pages, we'll need to wait for them. 17004f8ad655SJan Kara */ 17010ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL) && 1702f9b0e058SJan Kara (wbc->sync_mode != WB_SYNC_ALL || 1703f9b0e058SJan Kara !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) 17044f8ad655SJan Kara goto out; 17054f8ad655SJan Kara inode->i_state |= I_SYNC; 1706b16b1debSTejun Heo wbc_attach_and_unlock_inode(wbc, inode); 17074f8ad655SJan Kara 1708cd8ed2a4SYan Hong ret = __writeback_single_inode(inode, wbc); 17091da177e4SLinus Torvalds 1710b16b1debSTejun Heo wbc_detach_inode(wbc); 1711aaf25593STejun Heo 1712aaf25593STejun Heo wb = inode_to_wb_and_lock_list(inode); 1713250df6edSDave Chinner spin_lock(&inode->i_lock); 17144f8ad655SJan Kara /* 1715*4e3c51f4SSvyatoslav Feldsherov * If the inode is freeing, its i_io_list shoudn't be updated 1716*4e3c51f4SSvyatoslav Feldsherov * as it can be finally deleted at this moment. 1717*4e3c51f4SSvyatoslav Feldsherov */ 1718*4e3c51f4SSvyatoslav Feldsherov if (!(inode->i_state & I_FREEING)) { 1719*4e3c51f4SSvyatoslav Feldsherov /* 1720*4e3c51f4SSvyatoslav Feldsherov * If the inode is now fully clean, then it can be safely 1721*4e3c51f4SSvyatoslav Feldsherov * removed from its writeback list (if any). Otherwise the 1722*4e3c51f4SSvyatoslav Feldsherov * flusher threads are responsible for the writeback lists. 17234f8ad655SJan Kara */ 17240ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1725f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, wb); 1726cbfecb92SLukas Czerner else if (!(inode->i_state & I_SYNC_QUEUED)) { 1727cbfecb92SLukas Czerner if ((inode->i_state & I_DIRTY)) 1728846a3351SJing Xia redirty_tail_locked(inode, wb); 1729cbfecb92SLukas Czerner else if (inode->i_state & I_DIRTY_TIME) { 1730cbfecb92SLukas Czerner inode->dirtied_when = jiffies; 1731*4e3c51f4SSvyatoslav Feldsherov inode_io_list_move_locked(inode, 1732*4e3c51f4SSvyatoslav Feldsherov wb, 1733*4e3c51f4SSvyatoslav Feldsherov &wb->b_dirty_time); 1734*4e3c51f4SSvyatoslav Feldsherov } 1735cbfecb92SLukas Czerner } 1736cbfecb92SLukas Czerner } 1737846a3351SJing Xia 17384f8ad655SJan Kara spin_unlock(&wb->list_lock); 17391c0eeaf5SJoern Engel inode_sync_complete(inode); 17404f8ad655SJan Kara out: 17414f8ad655SJan Kara spin_unlock(&inode->i_lock); 17421da177e4SLinus Torvalds return ret; 17431da177e4SLinus Torvalds } 17441da177e4SLinus Torvalds 1745a88a341aSTejun Heo static long writeback_chunk_size(struct bdi_writeback *wb, 17461a12d8bdSWu Fengguang struct wb_writeback_work *work) 1747d46db3d5SWu Fengguang { 1748d46db3d5SWu Fengguang long pages; 1749d46db3d5SWu Fengguang 1750d46db3d5SWu Fengguang /* 1751d46db3d5SWu Fengguang * WB_SYNC_ALL mode does livelock avoidance by syncing dirty 1752d46db3d5SWu Fengguang * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX 1753d46db3d5SWu Fengguang * here avoids calling into writeback_inodes_wb() more than once. 1754d46db3d5SWu Fengguang * 1755d46db3d5SWu Fengguang * The intended call sequence for WB_SYNC_ALL writeback is: 1756d46db3d5SWu Fengguang * 1757d46db3d5SWu Fengguang * wb_writeback() 1758d46db3d5SWu Fengguang * writeback_sb_inodes() <== called only once 1759d46db3d5SWu Fengguang * write_cache_pages() <== called once for each inode 1760d46db3d5SWu Fengguang * (quickly) tag currently dirty pages 1761d46db3d5SWu Fengguang * (maybe slowly) sync all tagged pages 1762d46db3d5SWu Fengguang */ 1763d46db3d5SWu Fengguang if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) 1764d46db3d5SWu Fengguang pages = LONG_MAX; 17651a12d8bdSWu Fengguang else { 1766a88a341aSTejun Heo pages = min(wb->avg_write_bandwidth / 2, 1767dcc25ae7STejun Heo global_wb_domain.dirty_limit / DIRTY_SCOPE); 17681a12d8bdSWu Fengguang pages = min(pages, work->nr_pages); 17691a12d8bdSWu Fengguang pages = round_down(pages + MIN_WRITEBACK_PAGES, 17701a12d8bdSWu Fengguang MIN_WRITEBACK_PAGES); 17711a12d8bdSWu Fengguang } 1772d46db3d5SWu Fengguang 1773d46db3d5SWu Fengguang return pages; 1774d46db3d5SWu Fengguang } 1775d46db3d5SWu Fengguang 177603ba3782SJens Axboe /* 1777f11c9c5cSEdward Shishkin * Write a portion of b_io inodes which belong to @sb. 1778edadfb10SChristoph Hellwig * 1779d46db3d5SWu Fengguang * Return the number of pages and/or inodes written. 17800ba13fd1SLinus Torvalds * 17810ba13fd1SLinus Torvalds * NOTE! This is called with wb->list_lock held, and will 17820ba13fd1SLinus Torvalds * unlock and relock that for each inode it ends up doing 17830ba13fd1SLinus Torvalds * IO for. 1784f11c9c5cSEdward Shishkin */ 1785d46db3d5SWu Fengguang static long writeback_sb_inodes(struct super_block *sb, 1786d46db3d5SWu Fengguang struct bdi_writeback *wb, 1787d46db3d5SWu Fengguang struct wb_writeback_work *work) 178803ba3782SJens Axboe { 1789d46db3d5SWu Fengguang struct writeback_control wbc = { 1790d46db3d5SWu Fengguang .sync_mode = work->sync_mode, 1791d46db3d5SWu Fengguang .tagged_writepages = work->tagged_writepages, 1792d46db3d5SWu Fengguang .for_kupdate = work->for_kupdate, 1793d46db3d5SWu Fengguang .for_background = work->for_background, 17947747bd4bSDave Chinner .for_sync = work->for_sync, 1795d46db3d5SWu Fengguang .range_cyclic = work->range_cyclic, 1796d46db3d5SWu Fengguang .range_start = 0, 1797d46db3d5SWu Fengguang .range_end = LLONG_MAX, 1798d46db3d5SWu Fengguang }; 1799d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1800d46db3d5SWu Fengguang long write_chunk; 180168f4c6ebSZhihao Cheng long total_wrote = 0; /* count both pages and inodes */ 1802d46db3d5SWu Fengguang 180303ba3782SJens Axboe while (!list_empty(&wb->b_io)) { 18047ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1805aaf25593STejun Heo struct bdi_writeback *tmp_wb; 180668f4c6ebSZhihao Cheng long wrote; 1807edadfb10SChristoph Hellwig 1808edadfb10SChristoph Hellwig if (inode->i_sb != sb) { 1809d46db3d5SWu Fengguang if (work->sb) { 1810edadfb10SChristoph Hellwig /* 1811edadfb10SChristoph Hellwig * We only want to write back data for this 1812edadfb10SChristoph Hellwig * superblock, move all inodes not belonging 1813edadfb10SChristoph Hellwig * to it back onto the dirty list. 1814edadfb10SChristoph Hellwig */ 1815f758eeabSChristoph Hellwig redirty_tail(inode, wb); 181666f3b8e2SJens Axboe continue; 181766f3b8e2SJens Axboe } 1818edadfb10SChristoph Hellwig 1819edadfb10SChristoph Hellwig /* 1820edadfb10SChristoph Hellwig * The inode belongs to a different superblock. 1821edadfb10SChristoph Hellwig * Bounce back to the caller to unpin this and 1822edadfb10SChristoph Hellwig * pin the next superblock. 1823edadfb10SChristoph Hellwig */ 1824d46db3d5SWu Fengguang break; 1825edadfb10SChristoph Hellwig } 1826edadfb10SChristoph Hellwig 18279843b76aSChristoph Hellwig /* 1828331cbdeeSWanpeng Li * Don't bother with new inodes or inodes being freed, first 1829331cbdeeSWanpeng Li * kind does not need periodic writeout yet, and for the latter 18309843b76aSChristoph Hellwig * kind writeout is handled by the freer. 18319843b76aSChristoph Hellwig */ 1832250df6edSDave Chinner spin_lock(&inode->i_lock); 18339843b76aSChristoph Hellwig if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { 1834b35250c0SJan Kara redirty_tail_locked(inode, wb); 1835250df6edSDave Chinner spin_unlock(&inode->i_lock); 18367ef0d737SNick Piggin continue; 18377ef0d737SNick Piggin } 1838cc1676d9SJan Kara if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { 1839cc1676d9SJan Kara /* 1840cc1676d9SJan Kara * If this inode is locked for writeback and we are not 1841cc1676d9SJan Kara * doing writeback-for-data-integrity, move it to 1842cc1676d9SJan Kara * b_more_io so that writeback can proceed with the 1843cc1676d9SJan Kara * other inodes on s_io. 1844cc1676d9SJan Kara * 1845cc1676d9SJan Kara * We'll have another go at writing back this inode 1846cc1676d9SJan Kara * when we completed a full scan of b_io. 1847cc1676d9SJan Kara */ 1848cc1676d9SJan Kara requeue_io(inode, wb); 184910e14073SJchao Sun spin_unlock(&inode->i_lock); 1850cc1676d9SJan Kara trace_writeback_sb_inodes_requeue(inode); 1851cc1676d9SJan Kara continue; 1852cc1676d9SJan Kara } 1853f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 1854f0d07b7fSJan Kara 18554f8ad655SJan Kara /* 18564f8ad655SJan Kara * We already requeued the inode if it had I_SYNC set and we 18574f8ad655SJan Kara * are doing WB_SYNC_NONE writeback. So this catches only the 18584f8ad655SJan Kara * WB_SYNC_ALL case. 18594f8ad655SJan Kara */ 1860169ebd90SJan Kara if (inode->i_state & I_SYNC) { 1861169ebd90SJan Kara /* Wait for I_SYNC. This function drops i_lock... */ 1862169ebd90SJan Kara inode_sleep_on_writeback(inode); 1863169ebd90SJan Kara /* Inode may be gone, start again */ 1864ead188f9SJan Kara spin_lock(&wb->list_lock); 1865169ebd90SJan Kara continue; 1866169ebd90SJan Kara } 18674f8ad655SJan Kara inode->i_state |= I_SYNC; 1868b16b1debSTejun Heo wbc_attach_and_unlock_inode(&wbc, inode); 1869169ebd90SJan Kara 1870a88a341aSTejun Heo write_chunk = writeback_chunk_size(wb, work); 1871d46db3d5SWu Fengguang wbc.nr_to_write = write_chunk; 1872d46db3d5SWu Fengguang wbc.pages_skipped = 0; 1873250df6edSDave Chinner 1874169ebd90SJan Kara /* 1875169ebd90SJan Kara * We use I_SYNC to pin the inode in memory. While it is set 1876169ebd90SJan Kara * evict_inode() will wait so the inode cannot be freed. 1877169ebd90SJan Kara */ 1878cd8ed2a4SYan Hong __writeback_single_inode(inode, &wbc); 1879d46db3d5SWu Fengguang 1880b16b1debSTejun Heo wbc_detach_inode(&wbc); 1881d46db3d5SWu Fengguang work->nr_pages -= write_chunk - wbc.nr_to_write; 188268f4c6ebSZhihao Cheng wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped; 188368f4c6ebSZhihao Cheng wrote = wrote < 0 ? 0 : wrote; 188468f4c6ebSZhihao Cheng total_wrote += wrote; 1885590dca3aSChris Mason 1886590dca3aSChris Mason if (need_resched()) { 1887590dca3aSChris Mason /* 1888590dca3aSChris Mason * We're trying to balance between building up a nice 1889590dca3aSChris Mason * long list of IOs to improve our merge rate, and 1890590dca3aSChris Mason * getting those IOs out quickly for anyone throttling 1891590dca3aSChris Mason * in balance_dirty_pages(). cond_resched() doesn't 1892590dca3aSChris Mason * unplug, so get our IOs out the door before we 1893590dca3aSChris Mason * give up the CPU. 1894590dca3aSChris Mason */ 1895008f75a2SChristoph Hellwig blk_flush_plug(current->plug, false); 1896590dca3aSChris Mason cond_resched(); 1897590dca3aSChris Mason } 1898590dca3aSChris Mason 1899aaf25593STejun Heo /* 1900aaf25593STejun Heo * Requeue @inode if still dirty. Be careful as @inode may 1901aaf25593STejun Heo * have been switched to another wb in the meantime. 1902aaf25593STejun Heo */ 1903aaf25593STejun Heo tmp_wb = inode_to_wb_and_lock_list(inode); 19044f8ad655SJan Kara spin_lock(&inode->i_lock); 19050ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 190668f4c6ebSZhihao Cheng total_wrote++; 1907aaf25593STejun Heo requeue_inode(inode, tmp_wb, &wbc); 19084f8ad655SJan Kara inode_sync_complete(inode); 19090f1b1fd8SDave Chinner spin_unlock(&inode->i_lock); 1910590dca3aSChris Mason 1911aaf25593STejun Heo if (unlikely(tmp_wb != wb)) { 1912aaf25593STejun Heo spin_unlock(&tmp_wb->list_lock); 1913aaf25593STejun Heo spin_lock(&wb->list_lock); 1914aaf25593STejun Heo } 1915aaf25593STejun Heo 1916d46db3d5SWu Fengguang /* 1917d46db3d5SWu Fengguang * bail out to wb_writeback() often enough to check 1918d46db3d5SWu Fengguang * background threshold and other termination conditions. 1919d46db3d5SWu Fengguang */ 192068f4c6ebSZhihao Cheng if (total_wrote) { 1921d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1922d46db3d5SWu Fengguang break; 1923d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1924d46db3d5SWu Fengguang break; 19251da177e4SLinus Torvalds } 19268bc3be27SFengguang Wu } 192768f4c6ebSZhihao Cheng return total_wrote; 1928f11c9c5cSEdward Shishkin } 192938f21977SNick Piggin 1930d46db3d5SWu Fengguang static long __writeback_inodes_wb(struct bdi_writeback *wb, 1931d46db3d5SWu Fengguang struct wb_writeback_work *work) 1932f11c9c5cSEdward Shishkin { 1933d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1934d46db3d5SWu Fengguang long wrote = 0; 1935f11c9c5cSEdward Shishkin 1936f11c9c5cSEdward Shishkin while (!list_empty(&wb->b_io)) { 19377ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1938f11c9c5cSEdward Shishkin struct super_block *sb = inode->i_sb; 1939f11c9c5cSEdward Shishkin 1940eb6ef3dfSKonstantin Khlebnikov if (!trylock_super(sb)) { 19410e995816SWu Fengguang /* 1942eb6ef3dfSKonstantin Khlebnikov * trylock_super() may fail consistently due to 19430e995816SWu Fengguang * s_umount being grabbed by someone else. Don't use 19440e995816SWu Fengguang * requeue_io() to avoid busy retrying the inode/sb. 19450e995816SWu Fengguang */ 19460e995816SWu Fengguang redirty_tail(inode, wb); 1947d19de7edSChristoph Hellwig continue; 1948334132aeSChristoph Hellwig } 1949d46db3d5SWu Fengguang wrote += writeback_sb_inodes(sb, wb, work); 1950eb6ef3dfSKonstantin Khlebnikov up_read(&sb->s_umount); 1951f11c9c5cSEdward Shishkin 1952d46db3d5SWu Fengguang /* refer to the same tests at the end of writeback_sb_inodes */ 1953d46db3d5SWu Fengguang if (wrote) { 1954d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1955d46db3d5SWu Fengguang break; 1956d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1957f11c9c5cSEdward Shishkin break; 1958f11c9c5cSEdward Shishkin } 1959d46db3d5SWu Fengguang } 196066f3b8e2SJens Axboe /* Leave any unwritten inodes on b_io */ 1961d46db3d5SWu Fengguang return wrote; 196266f3b8e2SJens Axboe } 196366f3b8e2SJens Axboe 19647d9f073bSWanpeng Li static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, 19650e175a18SCurt Wohlgemuth enum wb_reason reason) 1966edadfb10SChristoph Hellwig { 1967d46db3d5SWu Fengguang struct wb_writeback_work work = { 1968d46db3d5SWu Fengguang .nr_pages = nr_pages, 1969d46db3d5SWu Fengguang .sync_mode = WB_SYNC_NONE, 1970d46db3d5SWu Fengguang .range_cyclic = 1, 19710e175a18SCurt Wohlgemuth .reason = reason, 1972d46db3d5SWu Fengguang }; 1973505a666eSLinus Torvalds struct blk_plug plug; 1974edadfb10SChristoph Hellwig 1975505a666eSLinus Torvalds blk_start_plug(&plug); 1976f758eeabSChristoph Hellwig spin_lock(&wb->list_lock); 1977424b351fSWu Fengguang if (list_empty(&wb->b_io)) 1978f9cae926SJan Kara queue_io(wb, &work, jiffies); 1979d46db3d5SWu Fengguang __writeback_inodes_wb(wb, &work); 1980f758eeabSChristoph Hellwig spin_unlock(&wb->list_lock); 1981505a666eSLinus Torvalds blk_finish_plug(&plug); 1982edadfb10SChristoph Hellwig 1983d46db3d5SWu Fengguang return nr_pages - work.nr_pages; 198466f3b8e2SJens Axboe } 198566f3b8e2SJens Axboe 198603ba3782SJens Axboe /* 198703ba3782SJens Axboe * Explicit flushing or periodic writeback of "old" data. 198803ba3782SJens Axboe * 198903ba3782SJens Axboe * Define "old": the first time one of an inode's pages is dirtied, we mark the 199003ba3782SJens Axboe * dirtying-time in the inode's address_space. So this periodic writeback code 199103ba3782SJens Axboe * just walks the superblock inode list, writing back any inodes which are 199203ba3782SJens Axboe * older than a specific point in time. 199303ba3782SJens Axboe * 199403ba3782SJens Axboe * Try to run once per dirty_writeback_interval. But if a writeback event 199503ba3782SJens Axboe * takes longer than a dirty_writeback_interval interval, then leave a 199603ba3782SJens Axboe * one-second gap. 199703ba3782SJens Axboe * 1998f9cae926SJan Kara * dirtied_before takes precedence over nr_to_write. So we'll only write back 199903ba3782SJens Axboe * all dirty pages if they are all attached to "old" mappings. 200003ba3782SJens Axboe */ 2001c4a77a6cSJens Axboe static long wb_writeback(struct bdi_writeback *wb, 200283ba7b07SChristoph Hellwig struct wb_writeback_work *work) 200303ba3782SJens Axboe { 2004d46db3d5SWu Fengguang long nr_pages = work->nr_pages; 2005f9cae926SJan Kara unsigned long dirtied_before = jiffies; 2006a5989bdcSJan Kara struct inode *inode; 2007d46db3d5SWu Fengguang long progress; 2008505a666eSLinus Torvalds struct blk_plug plug; 200903ba3782SJens Axboe 2010505a666eSLinus Torvalds blk_start_plug(&plug); 2011e8dfc305SWu Fengguang spin_lock(&wb->list_lock); 201203ba3782SJens Axboe for (;;) { 201303ba3782SJens Axboe /* 2014d3ddec76SWu Fengguang * Stop writeback when nr_pages has been consumed 201503ba3782SJens Axboe */ 201683ba7b07SChristoph Hellwig if (work->nr_pages <= 0) 201703ba3782SJens Axboe break; 201803ba3782SJens Axboe 201903ba3782SJens Axboe /* 2020aa373cf5SJan Kara * Background writeout and kupdate-style writeback may 2021aa373cf5SJan Kara * run forever. Stop them if there is other work to do 2022aa373cf5SJan Kara * so that e.g. sync can proceed. They'll be restarted 2023aa373cf5SJan Kara * after the other works are all done. 2024aa373cf5SJan Kara */ 2025aa373cf5SJan Kara if ((work->for_background || work->for_kupdate) && 2026f0054bb1STejun Heo !list_empty(&wb->work_list)) 2027aa373cf5SJan Kara break; 2028aa373cf5SJan Kara 2029aa373cf5SJan Kara /* 2030d3ddec76SWu Fengguang * For background writeout, stop when we are below the 2031d3ddec76SWu Fengguang * background dirty threshold 203203ba3782SJens Axboe */ 2033aa661bbeSTejun Heo if (work->for_background && !wb_over_bg_thresh(wb)) 203403ba3782SJens Axboe break; 203503ba3782SJens Axboe 20361bc36b64SJan Kara /* 20371bc36b64SJan Kara * Kupdate and background works are special and we want to 20381bc36b64SJan Kara * include all inodes that need writing. Livelock avoidance is 20391bc36b64SJan Kara * handled by these works yielding to any other work so we are 20401bc36b64SJan Kara * safe. 20411bc36b64SJan Kara */ 2042ba9aa839SWu Fengguang if (work->for_kupdate) { 2043f9cae926SJan Kara dirtied_before = jiffies - 2044ba9aa839SWu Fengguang msecs_to_jiffies(dirty_expire_interval * 10); 20451bc36b64SJan Kara } else if (work->for_background) 2046f9cae926SJan Kara dirtied_before = jiffies; 2047028c2dd1SDave Chinner 20485634cc2aSTejun Heo trace_writeback_start(wb, work); 2049e8dfc305SWu Fengguang if (list_empty(&wb->b_io)) 2050f9cae926SJan Kara queue_io(wb, work, dirtied_before); 205183ba7b07SChristoph Hellwig if (work->sb) 2052d46db3d5SWu Fengguang progress = writeback_sb_inodes(work->sb, wb, work); 2053edadfb10SChristoph Hellwig else 2054d46db3d5SWu Fengguang progress = __writeback_inodes_wb(wb, work); 20555634cc2aSTejun Heo trace_writeback_written(wb, work); 2056028c2dd1SDave Chinner 205703ba3782SJens Axboe /* 205871fd05a8SJens Axboe * Did we write something? Try for more 2059e6fb6da2SWu Fengguang * 2060e6fb6da2SWu Fengguang * Dirty inodes are moved to b_io for writeback in batches. 2061e6fb6da2SWu Fengguang * The completion of the current batch does not necessarily 2062e6fb6da2SWu Fengguang * mean the overall work is done. So we keep looping as long 2063e6fb6da2SWu Fengguang * as made some progress on cleaning pages or inodes. 206471fd05a8SJens Axboe */ 2065d46db3d5SWu Fengguang if (progress) 206603ba3782SJens Axboe continue; 2067a5989bdcSJan Kara /* 2068e6fb6da2SWu Fengguang * No more inodes for IO, bail 2069a5989bdcSJan Kara */ 2070b7a2441fSWu Fengguang if (list_empty(&wb->b_more_io)) 207103ba3782SJens Axboe break; 207203ba3782SJens Axboe /* 20738010c3b6SJens Axboe * Nothing written. Wait for some inode to 20748010c3b6SJens Axboe * become available for writeback. Otherwise 20758010c3b6SJens Axboe * we'll just busyloop. 207603ba3782SJens Axboe */ 20775634cc2aSTejun Heo trace_writeback_wait(wb, work); 207803ba3782SJens Axboe inode = wb_inode(wb->b_more_io.prev); 2079250df6edSDave Chinner spin_lock(&inode->i_lock); 2080f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 2081169ebd90SJan Kara /* This function drops i_lock... */ 2082169ebd90SJan Kara inode_sleep_on_writeback(inode); 2083f0d07b7fSJan Kara spin_lock(&wb->list_lock); 208403ba3782SJens Axboe } 2085e8dfc305SWu Fengguang spin_unlock(&wb->list_lock); 2086505a666eSLinus Torvalds blk_finish_plug(&plug); 208703ba3782SJens Axboe 2088d46db3d5SWu Fengguang return nr_pages - work->nr_pages; 208903ba3782SJens Axboe } 209003ba3782SJens Axboe 209103ba3782SJens Axboe /* 209283ba7b07SChristoph Hellwig * Return the next wb_writeback_work struct that hasn't been processed yet. 209303ba3782SJens Axboe */ 2094f0054bb1STejun Heo static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) 209503ba3782SJens Axboe { 209683ba7b07SChristoph Hellwig struct wb_writeback_work *work = NULL; 209703ba3782SJens Axboe 2098f87904c0SKhazhismel Kumykov spin_lock_irq(&wb->work_lock); 2099f0054bb1STejun Heo if (!list_empty(&wb->work_list)) { 2100f0054bb1STejun Heo work = list_entry(wb->work_list.next, 210183ba7b07SChristoph Hellwig struct wb_writeback_work, list); 210283ba7b07SChristoph Hellwig list_del_init(&work->list); 210303ba3782SJens Axboe } 2104f87904c0SKhazhismel Kumykov spin_unlock_irq(&wb->work_lock); 210583ba7b07SChristoph Hellwig return work; 210603ba3782SJens Axboe } 210703ba3782SJens Axboe 21086585027aSJan Kara static long wb_check_background_flush(struct bdi_writeback *wb) 21096585027aSJan Kara { 2110aa661bbeSTejun Heo if (wb_over_bg_thresh(wb)) { 21116585027aSJan Kara 21126585027aSJan Kara struct wb_writeback_work work = { 21136585027aSJan Kara .nr_pages = LONG_MAX, 21146585027aSJan Kara .sync_mode = WB_SYNC_NONE, 21156585027aSJan Kara .for_background = 1, 21166585027aSJan Kara .range_cyclic = 1, 21170e175a18SCurt Wohlgemuth .reason = WB_REASON_BACKGROUND, 21186585027aSJan Kara }; 21196585027aSJan Kara 21206585027aSJan Kara return wb_writeback(wb, &work); 21216585027aSJan Kara } 21226585027aSJan Kara 21236585027aSJan Kara return 0; 21246585027aSJan Kara } 21256585027aSJan Kara 212603ba3782SJens Axboe static long wb_check_old_data_flush(struct bdi_writeback *wb) 212703ba3782SJens Axboe { 212803ba3782SJens Axboe unsigned long expired; 212903ba3782SJens Axboe long nr_pages; 213003ba3782SJens Axboe 213169b62d01SJens Axboe /* 213269b62d01SJens Axboe * When set to zero, disable periodic writeback 213369b62d01SJens Axboe */ 213469b62d01SJens Axboe if (!dirty_writeback_interval) 213569b62d01SJens Axboe return 0; 213669b62d01SJens Axboe 213703ba3782SJens Axboe expired = wb->last_old_flush + 213803ba3782SJens Axboe msecs_to_jiffies(dirty_writeback_interval * 10); 213903ba3782SJens Axboe if (time_before(jiffies, expired)) 214003ba3782SJens Axboe return 0; 214103ba3782SJens Axboe 214203ba3782SJens Axboe wb->last_old_flush = jiffies; 2143cdf01dd5SLinus Torvalds nr_pages = get_nr_dirty_pages(); 214403ba3782SJens Axboe 2145c4a77a6cSJens Axboe if (nr_pages) { 214683ba7b07SChristoph Hellwig struct wb_writeback_work work = { 2147c4a77a6cSJens Axboe .nr_pages = nr_pages, 2148c4a77a6cSJens Axboe .sync_mode = WB_SYNC_NONE, 2149c4a77a6cSJens Axboe .for_kupdate = 1, 2150c4a77a6cSJens Axboe .range_cyclic = 1, 21510e175a18SCurt Wohlgemuth .reason = WB_REASON_PERIODIC, 2152c4a77a6cSJens Axboe }; 2153c4a77a6cSJens Axboe 215483ba7b07SChristoph Hellwig return wb_writeback(wb, &work); 2155c4a77a6cSJens Axboe } 215603ba3782SJens Axboe 215703ba3782SJens Axboe return 0; 215803ba3782SJens Axboe } 215903ba3782SJens Axboe 216085009b4fSJens Axboe static long wb_check_start_all(struct bdi_writeback *wb) 216185009b4fSJens Axboe { 216285009b4fSJens Axboe long nr_pages; 216385009b4fSJens Axboe 216485009b4fSJens Axboe if (!test_bit(WB_start_all, &wb->state)) 216585009b4fSJens Axboe return 0; 216685009b4fSJens Axboe 216785009b4fSJens Axboe nr_pages = get_nr_dirty_pages(); 216885009b4fSJens Axboe if (nr_pages) { 216985009b4fSJens Axboe struct wb_writeback_work work = { 217085009b4fSJens Axboe .nr_pages = wb_split_bdi_pages(wb, nr_pages), 217185009b4fSJens Axboe .sync_mode = WB_SYNC_NONE, 217285009b4fSJens Axboe .range_cyclic = 1, 217385009b4fSJens Axboe .reason = wb->start_all_reason, 217485009b4fSJens Axboe }; 217585009b4fSJens Axboe 217685009b4fSJens Axboe nr_pages = wb_writeback(wb, &work); 217785009b4fSJens Axboe } 217885009b4fSJens Axboe 217985009b4fSJens Axboe clear_bit(WB_start_all, &wb->state); 218085009b4fSJens Axboe return nr_pages; 218185009b4fSJens Axboe } 218285009b4fSJens Axboe 218385009b4fSJens Axboe 218403ba3782SJens Axboe /* 218503ba3782SJens Axboe * Retrieve work items and do the writeback they describe 218603ba3782SJens Axboe */ 218725d130baSWanpeng Li static long wb_do_writeback(struct bdi_writeback *wb) 218803ba3782SJens Axboe { 218983ba7b07SChristoph Hellwig struct wb_writeback_work *work; 2190c4a77a6cSJens Axboe long wrote = 0; 219103ba3782SJens Axboe 21924452226eSTejun Heo set_bit(WB_writeback_running, &wb->state); 2193f0054bb1STejun Heo while ((work = get_next_work_item(wb)) != NULL) { 21945634cc2aSTejun Heo trace_writeback_exec(wb, work); 219583ba7b07SChristoph Hellwig wrote += wb_writeback(wb, work); 21964a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 219703ba3782SJens Axboe } 219803ba3782SJens Axboe 219903ba3782SJens Axboe /* 220085009b4fSJens Axboe * Check for a flush-everything request 220185009b4fSJens Axboe */ 220285009b4fSJens Axboe wrote += wb_check_start_all(wb); 220385009b4fSJens Axboe 220485009b4fSJens Axboe /* 220503ba3782SJens Axboe * Check for periodic writeback, kupdated() style 220603ba3782SJens Axboe */ 220703ba3782SJens Axboe wrote += wb_check_old_data_flush(wb); 22086585027aSJan Kara wrote += wb_check_background_flush(wb); 22094452226eSTejun Heo clear_bit(WB_writeback_running, &wb->state); 221003ba3782SJens Axboe 221103ba3782SJens Axboe return wrote; 221203ba3782SJens Axboe } 221303ba3782SJens Axboe 221403ba3782SJens Axboe /* 221503ba3782SJens Axboe * Handle writeback of dirty data for the device backed by this bdi. Also 2216839a8e86STejun Heo * reschedules periodically and does kupdated style flushing. 221703ba3782SJens Axboe */ 2218f0054bb1STejun Heo void wb_workfn(struct work_struct *work) 221903ba3782SJens Axboe { 2220839a8e86STejun Heo struct bdi_writeback *wb = container_of(to_delayed_work(work), 2221839a8e86STejun Heo struct bdi_writeback, dwork); 222203ba3782SJens Axboe long pages_written; 222303ba3782SJens Axboe 222468f23b89STheodore Ts'o set_worker_desc("flush-%s", bdi_dev_name(wb->bdi)); 222503ba3782SJens Axboe 2226839a8e86STejun Heo if (likely(!current_is_workqueue_rescuer() || 22274452226eSTejun Heo !test_bit(WB_registered, &wb->state))) { 222803ba3782SJens Axboe /* 2229f0054bb1STejun Heo * The normal path. Keep writing back @wb until its 2230839a8e86STejun Heo * work_list is empty. Note that this path is also taken 2231f0054bb1STejun Heo * if @wb is shutting down even when we're running off the 2232839a8e86STejun Heo * rescuer as work_list needs to be drained. 223303ba3782SJens Axboe */ 2234839a8e86STejun Heo do { 223525d130baSWanpeng Li pages_written = wb_do_writeback(wb); 2236455b2864SDave Chinner trace_writeback_pages_written(pages_written); 2237f0054bb1STejun Heo } while (!list_empty(&wb->work_list)); 2238839a8e86STejun Heo } else { 2239253c34e9SArtem Bityutskiy /* 2240839a8e86STejun Heo * bdi_wq can't get enough workers and we're running off 2241839a8e86STejun Heo * the emergency worker. Don't hog it. Hopefully, 1024 is 2242839a8e86STejun Heo * enough for efficient IO. 2243253c34e9SArtem Bityutskiy */ 2244f0054bb1STejun Heo pages_written = writeback_inodes_wb(wb, 1024, 2245839a8e86STejun Heo WB_REASON_FORKER_THREAD); 2246839a8e86STejun Heo trace_writeback_pages_written(pages_written); 224703ba3782SJens Axboe } 224803ba3782SJens Axboe 2249f0054bb1STejun Heo if (!list_empty(&wb->work_list)) 2250b8b78495SJan Kara wb_wakeup(wb); 22516ca738d6SDerek Basehore else if (wb_has_dirty_io(wb) && dirty_writeback_interval) 2252f0054bb1STejun Heo wb_wakeup_delayed(wb); 225303ba3782SJens Axboe } 225403ba3782SJens Axboe 225503ba3782SJens Axboe /* 2256595043e5SJens Axboe * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero, 2257595043e5SJens Axboe * write back the whole world. 2258595043e5SJens Axboe */ 2259595043e5SJens Axboe static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2260e8e8a0c6SJens Axboe enum wb_reason reason) 2261595043e5SJens Axboe { 2262595043e5SJens Axboe struct bdi_writeback *wb; 2263595043e5SJens Axboe 2264595043e5SJens Axboe if (!bdi_has_dirty_io(bdi)) 2265595043e5SJens Axboe return; 2266595043e5SJens Axboe 2267595043e5SJens Axboe list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 2268e8e8a0c6SJens Axboe wb_start_writeback(wb, reason); 2269595043e5SJens Axboe } 2270595043e5SJens Axboe 2271595043e5SJens Axboe void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2272595043e5SJens Axboe enum wb_reason reason) 2273595043e5SJens Axboe { 2274595043e5SJens Axboe rcu_read_lock(); 2275e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2276595043e5SJens Axboe rcu_read_unlock(); 2277595043e5SJens Axboe } 2278595043e5SJens Axboe 2279595043e5SJens Axboe /* 22809ba4b2dfSJens Axboe * Wakeup the flusher threads to start writeback of all currently dirty pages 228103ba3782SJens Axboe */ 22829ba4b2dfSJens Axboe void wakeup_flusher_threads(enum wb_reason reason) 228303ba3782SJens Axboe { 2284b8c2f347SChristoph Hellwig struct backing_dev_info *bdi; 2285b8c2f347SChristoph Hellwig 228651350ea0SKonstantin Khlebnikov /* 228751350ea0SKonstantin Khlebnikov * If we are expecting writeback progress we must submit plugged IO. 228851350ea0SKonstantin Khlebnikov */ 2289008f75a2SChristoph Hellwig blk_flush_plug(current->plug, true); 229051350ea0SKonstantin Khlebnikov 2291b8c2f347SChristoph Hellwig rcu_read_lock(); 2292595043e5SJens Axboe list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) 2293e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2294b8c2f347SChristoph Hellwig rcu_read_unlock(); 229503ba3782SJens Axboe } 229603ba3782SJens Axboe 2297a2f48706STheodore Ts'o /* 2298a2f48706STheodore Ts'o * Wake up bdi's periodically to make sure dirtytime inodes gets 2299a2f48706STheodore Ts'o * written back periodically. We deliberately do *not* check the 2300a2f48706STheodore Ts'o * b_dirtytime list in wb_has_dirty_io(), since this would cause the 2301a2f48706STheodore Ts'o * kernel to be constantly waking up once there are any dirtytime 2302a2f48706STheodore Ts'o * inodes on the system. So instead we define a separate delayed work 2303a2f48706STheodore Ts'o * function which gets called much more rarely. (By default, only 2304a2f48706STheodore Ts'o * once every 12 hours.) 2305a2f48706STheodore Ts'o * 2306a2f48706STheodore Ts'o * If there is any other write activity going on in the file system, 2307a2f48706STheodore Ts'o * this function won't be necessary. But if the only thing that has 2308a2f48706STheodore Ts'o * happened on the file system is a dirtytime inode caused by an atime 2309a2f48706STheodore Ts'o * update, we need this infrastructure below to make sure that inode 2310a2f48706STheodore Ts'o * eventually gets pushed out to disk. 2311a2f48706STheodore Ts'o */ 2312a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w); 2313a2f48706STheodore Ts'o static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); 2314a2f48706STheodore Ts'o 2315a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w) 2316a2f48706STheodore Ts'o { 2317a2f48706STheodore Ts'o struct backing_dev_info *bdi; 2318a2f48706STheodore Ts'o 2319a2f48706STheodore Ts'o rcu_read_lock(); 2320a2f48706STheodore Ts'o list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 2321001fe6f6STejun Heo struct bdi_writeback *wb; 2322001fe6f6STejun Heo 2323b817525aSTejun Heo list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 23246fdf860fSTejun Heo if (!list_empty(&wb->b_dirty_time)) 23256fdf860fSTejun Heo wb_wakeup(wb); 2326a2f48706STheodore Ts'o } 2327a2f48706STheodore Ts'o rcu_read_unlock(); 2328a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2329a2f48706STheodore Ts'o } 2330a2f48706STheodore Ts'o 2331a2f48706STheodore Ts'o static int __init start_dirtytime_writeback(void) 2332a2f48706STheodore Ts'o { 2333a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2334a2f48706STheodore Ts'o return 0; 2335a2f48706STheodore Ts'o } 2336a2f48706STheodore Ts'o __initcall(start_dirtytime_writeback); 2337a2f48706STheodore Ts'o 23381efff914STheodore Ts'o int dirtytime_interval_handler(struct ctl_table *table, int write, 23399ca48e20STobias Klauser void *buffer, size_t *lenp, loff_t *ppos) 23401efff914STheodore Ts'o { 23411efff914STheodore Ts'o int ret; 23421efff914STheodore Ts'o 23431efff914STheodore Ts'o ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 23441efff914STheodore Ts'o if (ret == 0 && write) 23451efff914STheodore Ts'o mod_delayed_work(system_wq, &dirtytime_work, 0); 23461efff914STheodore Ts'o return ret; 23471efff914STheodore Ts'o } 23481efff914STheodore Ts'o 234903ba3782SJens Axboe /** 235035d14f27SEric Biggers * __mark_inode_dirty - internal function to mark an inode dirty 23510117d427SMauro Carvalho Chehab * 235203ba3782SJens Axboe * @inode: inode to mark 235335d14f27SEric Biggers * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of 235435d14f27SEric Biggers * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined 235535d14f27SEric Biggers * with I_DIRTY_PAGES. 23560117d427SMauro Carvalho Chehab * 235735d14f27SEric Biggers * Mark an inode as dirty. We notify the filesystem, then update the inode's 235835d14f27SEric Biggers * dirty flags. Then, if needed we add the inode to the appropriate dirty list. 235903ba3782SJens Axboe * 236035d14f27SEric Biggers * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync() 236135d14f27SEric Biggers * instead of calling this directly. 236203ba3782SJens Axboe * 236335d14f27SEric Biggers * CAREFUL! We only add the inode to the dirty list if it is hashed or if it 236435d14f27SEric Biggers * refers to a blockdev. Unhashed inodes will never be added to the dirty list 236535d14f27SEric Biggers * even if they are later hashed, as they will have been marked dirty already. 236603ba3782SJens Axboe * 236735d14f27SEric Biggers * In short, ensure you hash any inodes _before_ you start marking them dirty. 236803ba3782SJens Axboe * 236903ba3782SJens Axboe * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 237003ba3782SJens Axboe * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 237103ba3782SJens Axboe * the kernel-internal blockdev inode represents the dirtying time of the 237203ba3782SJens Axboe * blockdev's pages. This is why for I_DIRTY_PAGES we always use 237303ba3782SJens Axboe * page->mapping->host, so the page-dirtying time is recorded in the internal 237403ba3782SJens Axboe * blockdev inode. 237503ba3782SJens Axboe */ 237603ba3782SJens Axboe void __mark_inode_dirty(struct inode *inode, int flags) 237703ba3782SJens Axboe { 237803ba3782SJens Axboe struct super_block *sb = inode->i_sb; 237935d14f27SEric Biggers int dirtytime = 0; 238010e14073SJchao Sun struct bdi_writeback *wb = NULL; 23810ae45f63STheodore Ts'o 23820ae45f63STheodore Ts'o trace_writeback_mark_inode_dirty(inode, flags); 238303ba3782SJens Axboe 2384e2728c56SEric Biggers if (flags & I_DIRTY_INODE) { 238535d14f27SEric Biggers /* 2386cbfecb92SLukas Czerner * Inode timestamp update will piggback on this dirtying. 2387cbfecb92SLukas Czerner * We tell ->dirty_inode callback that timestamps need to 2388cbfecb92SLukas Czerner * be updated by setting I_DIRTY_TIME in flags. 2389cbfecb92SLukas Czerner */ 2390cbfecb92SLukas Czerner if (inode->i_state & I_DIRTY_TIME) { 2391cbfecb92SLukas Czerner spin_lock(&inode->i_lock); 2392cbfecb92SLukas Czerner if (inode->i_state & I_DIRTY_TIME) { 2393cbfecb92SLukas Czerner inode->i_state &= ~I_DIRTY_TIME; 2394cbfecb92SLukas Czerner flags |= I_DIRTY_TIME; 2395cbfecb92SLukas Czerner } 2396cbfecb92SLukas Czerner spin_unlock(&inode->i_lock); 2397cbfecb92SLukas Czerner } 2398cbfecb92SLukas Czerner 2399cbfecb92SLukas Czerner /* 240035d14f27SEric Biggers * Notify the filesystem about the inode being dirtied, so that 240135d14f27SEric Biggers * (if needed) it can update on-disk fields and journal the 240235d14f27SEric Biggers * inode. This is only needed when the inode itself is being 240335d14f27SEric Biggers * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not 240435d14f27SEric Biggers * for just I_DIRTY_PAGES or I_DIRTY_TIME. 240535d14f27SEric Biggers */ 24069fb0a7daSTejun Heo trace_writeback_dirty_inode_start(inode, flags); 240703ba3782SJens Axboe if (sb->s_op->dirty_inode) 2408cbfecb92SLukas Czerner sb->s_op->dirty_inode(inode, 2409cbfecb92SLukas Czerner flags & (I_DIRTY_INODE | I_DIRTY_TIME)); 24109fb0a7daSTejun Heo trace_writeback_dirty_inode(inode, flags); 2411e2728c56SEric Biggers 241235d14f27SEric Biggers /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */ 24130ae45f63STheodore Ts'o flags &= ~I_DIRTY_TIME; 241435d14f27SEric Biggers } else { 241535d14f27SEric Biggers /* 241635d14f27SEric Biggers * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing. 241735d14f27SEric Biggers * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME 241835d14f27SEric Biggers * in one call to __mark_inode_dirty().) 241935d14f27SEric Biggers */ 24200ae45f63STheodore Ts'o dirtytime = flags & I_DIRTY_TIME; 242135d14f27SEric Biggers WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME); 242235d14f27SEric Biggers } 242303ba3782SJens Axboe 242403ba3782SJens Axboe /* 24259c6ac78eSTejun Heo * Paired with smp_mb() in __writeback_single_inode() for the 24269c6ac78eSTejun Heo * following lockless i_state test. See there for details. 242703ba3782SJens Axboe */ 242803ba3782SJens Axboe smp_mb(); 242903ba3782SJens Axboe 2430cbfecb92SLukas Czerner if ((inode->i_state & flags) == flags) 243103ba3782SJens Axboe return; 243203ba3782SJens Axboe 2433250df6edSDave Chinner spin_lock(&inode->i_lock); 243403ba3782SJens Axboe if ((inode->i_state & flags) != flags) { 243503ba3782SJens Axboe const int was_dirty = inode->i_state & I_DIRTY; 243603ba3782SJens Axboe 243752ebea74STejun Heo inode_attach_wb(inode, NULL); 243852ebea74STejun Heo 243903ba3782SJens Axboe inode->i_state |= flags; 244003ba3782SJens Axboe 244103ba3782SJens Axboe /* 244210e14073SJchao Sun * Grab inode's wb early because it requires dropping i_lock and we 244310e14073SJchao Sun * need to make sure following checks happen atomically with dirty 244410e14073SJchao Sun * list handling so that we don't move inodes under flush worker's 244510e14073SJchao Sun * hands. 244610e14073SJchao Sun */ 244710e14073SJchao Sun if (!was_dirty) { 244810e14073SJchao Sun wb = locked_inode_to_wb_and_lock_list(inode); 244910e14073SJchao Sun spin_lock(&inode->i_lock); 245010e14073SJchao Sun } 245110e14073SJchao Sun 245210e14073SJchao Sun /* 24535afced3bSJan Kara * If the inode is queued for writeback by flush worker, just 24545afced3bSJan Kara * update its dirty state. Once the flush worker is done with 24555afced3bSJan Kara * the inode it will place it on the appropriate superblock 24565afced3bSJan Kara * list, based upon its state. 245703ba3782SJens Axboe */ 24585afced3bSJan Kara if (inode->i_state & I_SYNC_QUEUED) 245910e14073SJchao Sun goto out_unlock; 246003ba3782SJens Axboe 246103ba3782SJens Axboe /* 246203ba3782SJens Axboe * Only add valid (hashed) inodes to the superblock's 246303ba3782SJens Axboe * dirty list. Add blockdev inodes as well. 246403ba3782SJens Axboe */ 246503ba3782SJens Axboe if (!S_ISBLK(inode->i_mode)) { 24661d3382cbSAl Viro if (inode_unhashed(inode)) 246710e14073SJchao Sun goto out_unlock; 246803ba3782SJens Axboe } 2469a4ffdde6SAl Viro if (inode->i_state & I_FREEING) 247010e14073SJchao Sun goto out_unlock; 247103ba3782SJens Axboe 247203ba3782SJens Axboe /* 247303ba3782SJens Axboe * If the inode was already on b_dirty/b_io/b_more_io, don't 247403ba3782SJens Axboe * reposition it (that would break b_dirty time-ordering). 247503ba3782SJens Axboe */ 247603ba3782SJens Axboe if (!was_dirty) { 2477d6c10f1fSTejun Heo struct list_head *dirty_list; 2478a66979abSDave Chinner bool wakeup_bdi = false; 2479500b067cSJens Axboe 248003ba3782SJens Axboe inode->dirtied_when = jiffies; 2481a2f48706STheodore Ts'o if (dirtytime) 2482a2f48706STheodore Ts'o inode->dirtied_time_when = jiffies; 2483d6c10f1fSTejun Heo 24840e11f644SChristoph Hellwig if (inode->i_state & I_DIRTY) 24850747259dSTejun Heo dirty_list = &wb->b_dirty; 2486a2f48706STheodore Ts'o else 24870747259dSTejun Heo dirty_list = &wb->b_dirty_time; 2488d6c10f1fSTejun Heo 2489c7f54084SDave Chinner wakeup_bdi = inode_io_list_move_locked(inode, wb, 2490d6c10f1fSTejun Heo dirty_list); 2491d6c10f1fSTejun Heo 24920747259dSTejun Heo spin_unlock(&wb->list_lock); 249310e14073SJchao Sun spin_unlock(&inode->i_lock); 24940ae45f63STheodore Ts'o trace_writeback_dirty_inode_enqueue(inode); 2495253c34e9SArtem Bityutskiy 2496d6c10f1fSTejun Heo /* 2497d6c10f1fSTejun Heo * If this is the first dirty inode for this bdi, 2498d6c10f1fSTejun Heo * we have to wake-up the corresponding bdi thread 2499d6c10f1fSTejun Heo * to make sure background write-back happens 2500d6c10f1fSTejun Heo * later. 2501d6c10f1fSTejun Heo */ 2502f56753acSChristoph Hellwig if (wakeup_bdi && 2503f56753acSChristoph Hellwig (wb->bdi->capabilities & BDI_CAP_WRITEBACK)) 25040747259dSTejun Heo wb_wakeup_delayed(wb); 2505a66979abSDave Chinner return; 2506a66979abSDave Chinner } 2507a66979abSDave Chinner } 250810e14073SJchao Sun out_unlock: 250910e14073SJchao Sun if (wb) 251010e14073SJchao Sun spin_unlock(&wb->list_lock); 2511a66979abSDave Chinner spin_unlock(&inode->i_lock); 251203ba3782SJens Axboe } 251303ba3782SJens Axboe EXPORT_SYMBOL(__mark_inode_dirty); 251403ba3782SJens Axboe 2515e97fedb9SDave Chinner /* 2516e97fedb9SDave Chinner * The @s_sync_lock is used to serialise concurrent sync operations 2517e97fedb9SDave Chinner * to avoid lock contention problems with concurrent wait_sb_inodes() calls. 2518e97fedb9SDave Chinner * Concurrent callers will block on the s_sync_lock rather than doing contending 2519e97fedb9SDave Chinner * walks. The queueing maintains sync(2) required behaviour as all the IO that 2520e97fedb9SDave Chinner * has been issued up to the time this function is enter is guaranteed to be 2521e97fedb9SDave Chinner * completed by the time we have gained the lock and waited for all IO that is 2522e97fedb9SDave Chinner * in progress regardless of the order callers are granted the lock. 2523e97fedb9SDave Chinner */ 2524b6e51316SJens Axboe static void wait_sb_inodes(struct super_block *sb) 252566f3b8e2SJens Axboe { 25266c60d2b5SDave Chinner LIST_HEAD(sync_list); 252738f21977SNick Piggin 252803ba3782SJens Axboe /* 252903ba3782SJens Axboe * We need to be protected against the filesystem going from 253003ba3782SJens Axboe * r/o to r/w or vice versa. 253103ba3782SJens Axboe */ 2532b6e51316SJens Axboe WARN_ON(!rwsem_is_locked(&sb->s_umount)); 253303ba3782SJens Axboe 2534e97fedb9SDave Chinner mutex_lock(&sb->s_sync_lock); 253566f3b8e2SJens Axboe 253638f21977SNick Piggin /* 25376c60d2b5SDave Chinner * Splice the writeback list onto a temporary list to avoid waiting on 25386c60d2b5SDave Chinner * inodes that have started writeback after this point. 25396c60d2b5SDave Chinner * 25406c60d2b5SDave Chinner * Use rcu_read_lock() to keep the inodes around until we have a 25416c60d2b5SDave Chinner * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as 25426c60d2b5SDave Chinner * the local list because inodes can be dropped from either by writeback 25436c60d2b5SDave Chinner * completion. 254438f21977SNick Piggin */ 25456c60d2b5SDave Chinner rcu_read_lock(); 25466c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 25476c60d2b5SDave Chinner list_splice_init(&sb->s_inodes_wb, &sync_list); 25486c60d2b5SDave Chinner 25496c60d2b5SDave Chinner /* 25506c60d2b5SDave Chinner * Data integrity sync. Must wait for all pages under writeback, because 25516c60d2b5SDave Chinner * there may have been pages dirtied before our sync call, but which had 25526c60d2b5SDave Chinner * writeout started before we write it out. In which case, the inode 25536c60d2b5SDave Chinner * may not be on the dirty list, but we still have to wait for that 25546c60d2b5SDave Chinner * writeout. 25556c60d2b5SDave Chinner */ 25566c60d2b5SDave Chinner while (!list_empty(&sync_list)) { 25576c60d2b5SDave Chinner struct inode *inode = list_first_entry(&sync_list, struct inode, 25586c60d2b5SDave Chinner i_wb_list); 2559250df6edSDave Chinner struct address_space *mapping = inode->i_mapping; 256038f21977SNick Piggin 25616c60d2b5SDave Chinner /* 25626c60d2b5SDave Chinner * Move each inode back to the wb list before we drop the lock 25636c60d2b5SDave Chinner * to preserve consistency between i_wb_list and the mapping 25646c60d2b5SDave Chinner * writeback tag. Writeback completion is responsible to remove 25656c60d2b5SDave Chinner * the inode from either list once the writeback tag is cleared. 25666c60d2b5SDave Chinner */ 25676c60d2b5SDave Chinner list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); 25686c60d2b5SDave Chinner 25696c60d2b5SDave Chinner /* 25706c60d2b5SDave Chinner * The mapping can appear untagged while still on-list since we 25716c60d2b5SDave Chinner * do not have the mapping lock. Skip it here, wb completion 25726c60d2b5SDave Chinner * will remove it. 25736c60d2b5SDave Chinner */ 25746c60d2b5SDave Chinner if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) 25756c60d2b5SDave Chinner continue; 25766c60d2b5SDave Chinner 25776c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 25786c60d2b5SDave Chinner 2579250df6edSDave Chinner spin_lock(&inode->i_lock); 25806c60d2b5SDave Chinner if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { 2581250df6edSDave Chinner spin_unlock(&inode->i_lock); 25826c60d2b5SDave Chinner 25836c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 258438f21977SNick Piggin continue; 2585250df6edSDave Chinner } 258638f21977SNick Piggin __iget(inode); 2587250df6edSDave Chinner spin_unlock(&inode->i_lock); 25886c60d2b5SDave Chinner rcu_read_unlock(); 258938f21977SNick Piggin 2590aa750fd7SJunichi Nomura /* 2591aa750fd7SJunichi Nomura * We keep the error status of individual mapping so that 2592aa750fd7SJunichi Nomura * applications can catch the writeback error using fsync(2). 2593aa750fd7SJunichi Nomura * See filemap_fdatawait_keep_errors() for details. 2594aa750fd7SJunichi Nomura */ 2595aa750fd7SJunichi Nomura filemap_fdatawait_keep_errors(mapping); 259638f21977SNick Piggin 259738f21977SNick Piggin cond_resched(); 259838f21977SNick Piggin 25996c60d2b5SDave Chinner iput(inode); 26006c60d2b5SDave Chinner 26016c60d2b5SDave Chinner rcu_read_lock(); 26026c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 260338f21977SNick Piggin } 26046c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 26056c60d2b5SDave Chinner rcu_read_unlock(); 2606e97fedb9SDave Chinner mutex_unlock(&sb->s_sync_lock); 260766f3b8e2SJens Axboe } 26081da177e4SLinus Torvalds 2609f30a7d0cSTejun Heo static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2610f30a7d0cSTejun Heo enum wb_reason reason, bool skip_if_busy) 26111da177e4SLinus Torvalds { 26125b9cce4cSTejun Heo struct backing_dev_info *bdi = sb->s_bdi; 26135b9cce4cSTejun Heo DEFINE_WB_COMPLETION(done, bdi); 261483ba7b07SChristoph Hellwig struct wb_writeback_work work = { 26153c4d7165SChristoph Hellwig .sb = sb, 26163c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_NONE, 26176e6938b6SWu Fengguang .tagged_writepages = 1, 261883ba7b07SChristoph Hellwig .done = &done, 26193259f8beSChris Mason .nr_pages = nr, 26200e175a18SCurt Wohlgemuth .reason = reason, 26213c4d7165SChristoph Hellwig }; 26220e3c9a22SJens Axboe 2623e7972912STejun Heo if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 26246eedc701SJan Kara return; 2625cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2626f30a7d0cSTejun Heo 2627db125360STejun Heo bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); 26285b9cce4cSTejun Heo wb_wait_for_completion(&done); 26291da177e4SLinus Torvalds } 2630f30a7d0cSTejun Heo 2631f30a7d0cSTejun Heo /** 2632f30a7d0cSTejun Heo * writeback_inodes_sb_nr - writeback dirty inodes from given super_block 2633f30a7d0cSTejun Heo * @sb: the superblock 2634f30a7d0cSTejun Heo * @nr: the number of pages to write 2635f30a7d0cSTejun Heo * @reason: reason why some writeback work initiated 2636f30a7d0cSTejun Heo * 2637f30a7d0cSTejun Heo * Start writeback on some inodes on this super_block. No guarantees are made 2638f30a7d0cSTejun Heo * on how many (if any) will be written, and this function does not wait 2639f30a7d0cSTejun Heo * for IO completion of submitted IO. 2640f30a7d0cSTejun Heo */ 2641f30a7d0cSTejun Heo void writeback_inodes_sb_nr(struct super_block *sb, 2642f30a7d0cSTejun Heo unsigned long nr, 2643f30a7d0cSTejun Heo enum wb_reason reason) 2644f30a7d0cSTejun Heo { 2645f30a7d0cSTejun Heo __writeback_inodes_sb_nr(sb, nr, reason, false); 2646f30a7d0cSTejun Heo } 26473259f8beSChris Mason EXPORT_SYMBOL(writeback_inodes_sb_nr); 26483259f8beSChris Mason 26493259f8beSChris Mason /** 26503259f8beSChris Mason * writeback_inodes_sb - writeback dirty inodes from given super_block 26513259f8beSChris Mason * @sb: the superblock 2652786228abSMarcos Paulo de Souza * @reason: reason why some writeback work was initiated 26533259f8beSChris Mason * 26543259f8beSChris Mason * Start writeback on some inodes on this super_block. No guarantees are made 26553259f8beSChris Mason * on how many (if any) will be written, and this function does not wait 26563259f8beSChris Mason * for IO completion of submitted IO. 26573259f8beSChris Mason */ 26580e175a18SCurt Wohlgemuth void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 26593259f8beSChris Mason { 26600e175a18SCurt Wohlgemuth return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 26613259f8beSChris Mason } 2662d8a8559cSJens Axboe EXPORT_SYMBOL(writeback_inodes_sb); 2663d8a8559cSJens Axboe 2664d8a8559cSJens Axboe /** 266510ee27a0SMiao Xie * try_to_writeback_inodes_sb - try to start writeback if none underway 266610ee27a0SMiao Xie * @sb: the superblock 266710ee27a0SMiao Xie * @reason: reason why some writeback work was initiated 266810ee27a0SMiao Xie * 26698264c321SRakesh Pandit * Invoke __writeback_inodes_sb_nr if no writeback is currently underway. 267010ee27a0SMiao Xie */ 26718264c321SRakesh Pandit void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 267210ee27a0SMiao Xie { 26738264c321SRakesh Pandit if (!down_read_trylock(&sb->s_umount)) 26748264c321SRakesh Pandit return; 26758264c321SRakesh Pandit 26768264c321SRakesh Pandit __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true); 26778264c321SRakesh Pandit up_read(&sb->s_umount); 267810ee27a0SMiao Xie } 267910ee27a0SMiao Xie EXPORT_SYMBOL(try_to_writeback_inodes_sb); 26803259f8beSChris Mason 26813259f8beSChris Mason /** 2682d8a8559cSJens Axboe * sync_inodes_sb - sync sb inode pages 2683d8a8559cSJens Axboe * @sb: the superblock 2684d8a8559cSJens Axboe * 2685d8a8559cSJens Axboe * This function writes and waits on any dirty inode belonging to this 26860dc83bd3SJan Kara * super_block. 2687d8a8559cSJens Axboe */ 26880dc83bd3SJan Kara void sync_inodes_sb(struct super_block *sb) 2689d8a8559cSJens Axboe { 26905b9cce4cSTejun Heo struct backing_dev_info *bdi = sb->s_bdi; 26915b9cce4cSTejun Heo DEFINE_WB_COMPLETION(done, bdi); 269283ba7b07SChristoph Hellwig struct wb_writeback_work work = { 26933c4d7165SChristoph Hellwig .sb = sb, 26943c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_ALL, 26953c4d7165SChristoph Hellwig .nr_pages = LONG_MAX, 26963c4d7165SChristoph Hellwig .range_cyclic = 0, 269783ba7b07SChristoph Hellwig .done = &done, 26980e175a18SCurt Wohlgemuth .reason = WB_REASON_SYNC, 26997747bd4bSDave Chinner .for_sync = 1, 27003c4d7165SChristoph Hellwig }; 27013c4d7165SChristoph Hellwig 2702006a0973STejun Heo /* 2703006a0973STejun Heo * Can't skip on !bdi_has_dirty() because we should wait for !dirty 2704006a0973STejun Heo * inodes under writeback and I_DIRTY_TIME inodes ignored by 2705006a0973STejun Heo * bdi_has_dirty() need to be written out too. 2706006a0973STejun Heo */ 2707006a0973STejun Heo if (bdi == &noop_backing_dev_info) 27086eedc701SJan Kara return; 2709cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2710cf37e972SChristoph Hellwig 27117fc5854fSTejun Heo /* protect against inode wb switch, see inode_switch_wbs_work_fn() */ 27127fc5854fSTejun Heo bdi_down_write_wb_switch_rwsem(bdi); 2713db125360STejun Heo bdi_split_work_to_wbs(bdi, &work, false); 27145b9cce4cSTejun Heo wb_wait_for_completion(&done); 27157fc5854fSTejun Heo bdi_up_write_wb_switch_rwsem(bdi); 271683ba7b07SChristoph Hellwig 2717b6e51316SJens Axboe wait_sb_inodes(sb); 2718d8a8559cSJens Axboe } 2719d8a8559cSJens Axboe EXPORT_SYMBOL(sync_inodes_sb); 27201da177e4SLinus Torvalds 27211da177e4SLinus Torvalds /** 27221da177e4SLinus Torvalds * write_inode_now - write an inode to disk 27231da177e4SLinus Torvalds * @inode: inode to write to disk 27241da177e4SLinus Torvalds * @sync: whether the write should be synchronous or not 27251da177e4SLinus Torvalds * 27267f04c26dSAndrea Arcangeli * This function commits an inode to disk immediately if it is dirty. This is 27277f04c26dSAndrea Arcangeli * primarily needed by knfsd. 27287f04c26dSAndrea Arcangeli * 27297f04c26dSAndrea Arcangeli * The caller must either have a ref on the inode or must have set I_WILL_FREE. 27301da177e4SLinus Torvalds */ 27311da177e4SLinus Torvalds int write_inode_now(struct inode *inode, int sync) 27321da177e4SLinus Torvalds { 27331da177e4SLinus Torvalds struct writeback_control wbc = { 27341da177e4SLinus Torvalds .nr_to_write = LONG_MAX, 273518914b18SMike Galbraith .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, 2736111ebb6eSOGAWA Hirofumi .range_start = 0, 2737111ebb6eSOGAWA Hirofumi .range_end = LLONG_MAX, 27381da177e4SLinus Torvalds }; 27391da177e4SLinus Torvalds 2740f56753acSChristoph Hellwig if (!mapping_can_writeback(inode->i_mapping)) 274149364ce2SAndrew Morton wbc.nr_to_write = 0; 27421da177e4SLinus Torvalds 27431da177e4SLinus Torvalds might_sleep(); 2744aaf25593STejun Heo return writeback_single_inode(inode, &wbc); 27451da177e4SLinus Torvalds } 27461da177e4SLinus Torvalds EXPORT_SYMBOL(write_inode_now); 27471da177e4SLinus Torvalds 27481da177e4SLinus Torvalds /** 2749c691b9d9SAndrew Morton * sync_inode_metadata - write an inode to disk 2750c3765016SChristoph Hellwig * @inode: the inode to sync 2751c3765016SChristoph Hellwig * @wait: wait for I/O to complete. 2752c3765016SChristoph Hellwig * 2753c691b9d9SAndrew Morton * Write an inode to disk and adjust its dirty state after completion. 2754c3765016SChristoph Hellwig * 2755c3765016SChristoph Hellwig * Note: only writes the actual inode, no associated data or other metadata. 2756c3765016SChristoph Hellwig */ 2757c3765016SChristoph Hellwig int sync_inode_metadata(struct inode *inode, int wait) 2758c3765016SChristoph Hellwig { 2759c3765016SChristoph Hellwig struct writeback_control wbc = { 2760c3765016SChristoph Hellwig .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, 2761c3765016SChristoph Hellwig .nr_to_write = 0, /* metadata-only */ 2762c3765016SChristoph Hellwig }; 2763c3765016SChristoph Hellwig 27645662c967SJosef Bacik return writeback_single_inode(inode, &wbc); 2765c3765016SChristoph Hellwig } 2766c3765016SChristoph Hellwig EXPORT_SYMBOL(sync_inode_metadata); 2767