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); 124a9438b44SJan Kara WARN_ON_ONCE(inode->i_state & I_FREEING); 125d6c10f1fSTejun Heo 126c7f54084SDave Chinner list_move(&inode->i_io_list, head); 127d6c10f1fSTejun Heo 128d6c10f1fSTejun Heo /* dirty_time doesn't count as dirty_io until expiration */ 129d6c10f1fSTejun Heo if (head != &wb->b_dirty_time) 130d6c10f1fSTejun Heo return wb_io_lists_populated(wb); 131d6c10f1fSTejun Heo 132d6c10f1fSTejun Heo wb_io_lists_depopulated(wb); 133d6c10f1fSTejun Heo return false; 134d6c10f1fSTejun Heo } 135d6c10f1fSTejun Heo 136f0054bb1STejun Heo static void wb_wakeup(struct bdi_writeback *wb) 1375acda9d1SJan Kara { 138f87904c0SKhazhismel Kumykov spin_lock_irq(&wb->work_lock); 139f0054bb1STejun Heo if (test_bit(WB_registered, &wb->state)) 140f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 141f87904c0SKhazhismel Kumykov spin_unlock_irq(&wb->work_lock); 1425acda9d1SJan Kara } 1435acda9d1SJan Kara 1444a3a485bSTahsin Erdogan static void finish_writeback_work(struct bdi_writeback *wb, 1454a3a485bSTahsin Erdogan struct wb_writeback_work *work) 1464a3a485bSTahsin Erdogan { 1474a3a485bSTahsin Erdogan struct wb_completion *done = work->done; 1484a3a485bSTahsin Erdogan 1494a3a485bSTahsin Erdogan if (work->auto_free) 1504a3a485bSTahsin Erdogan kfree(work); 1518e00c4e9STejun Heo if (done) { 1528e00c4e9STejun Heo wait_queue_head_t *waitq = done->waitq; 1538e00c4e9STejun Heo 1548e00c4e9STejun Heo /* @done can't be accessed after the following dec */ 1558e00c4e9STejun Heo if (atomic_dec_and_test(&done->cnt)) 1568e00c4e9STejun Heo wake_up_all(waitq); 1578e00c4e9STejun Heo } 1584a3a485bSTahsin Erdogan } 1594a3a485bSTahsin Erdogan 160f0054bb1STejun Heo static void wb_queue_work(struct bdi_writeback *wb, 1616585027aSJan Kara struct wb_writeback_work *work) 1626585027aSJan Kara { 1635634cc2aSTejun Heo trace_writeback_queue(wb, work); 1646585027aSJan Kara 165cc395d7fSTejun Heo if (work->done) 166cc395d7fSTejun Heo atomic_inc(&work->done->cnt); 1674a3a485bSTahsin Erdogan 168f87904c0SKhazhismel Kumykov spin_lock_irq(&wb->work_lock); 1694a3a485bSTahsin Erdogan 1704a3a485bSTahsin Erdogan if (test_bit(WB_registered, &wb->state)) { 171f0054bb1STejun Heo list_add_tail(&work->list, &wb->work_list); 172f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 1734a3a485bSTahsin Erdogan } else 1744a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 1754a3a485bSTahsin Erdogan 176f87904c0SKhazhismel Kumykov spin_unlock_irq(&wb->work_lock); 17703ba3782SJens Axboe } 1781da177e4SLinus Torvalds 179cc395d7fSTejun Heo /** 180cc395d7fSTejun Heo * wb_wait_for_completion - wait for completion of bdi_writeback_works 181cc395d7fSTejun Heo * @done: target wb_completion 182cc395d7fSTejun Heo * 183cc395d7fSTejun Heo * Wait for one or more work items issued to @bdi with their ->done field 1845b9cce4cSTejun Heo * set to @done, which should have been initialized with 1855b9cce4cSTejun Heo * DEFINE_WB_COMPLETION(). This function returns after all such work items 1865b9cce4cSTejun Heo * are completed. Work items which are waited upon aren't freed 187cc395d7fSTejun Heo * automatically on completion. 188cc395d7fSTejun Heo */ 1895b9cce4cSTejun Heo void wb_wait_for_completion(struct wb_completion *done) 190cc395d7fSTejun Heo { 191cc395d7fSTejun Heo atomic_dec(&done->cnt); /* put down the initial count */ 1925b9cce4cSTejun Heo wait_event(*done->waitq, !atomic_read(&done->cnt)); 193cc395d7fSTejun Heo } 194cc395d7fSTejun Heo 195703c2708STejun Heo #ifdef CONFIG_CGROUP_WRITEBACK 196703c2708STejun Heo 19755a694dfSTejun Heo /* 19855a694dfSTejun Heo * Parameters for foreign inode detection, see wbc_detach_inode() to see 19955a694dfSTejun Heo * how they're used. 20055a694dfSTejun Heo * 20155a694dfSTejun Heo * These paramters are inherently heuristical as the detection target 20255a694dfSTejun Heo * itself is fuzzy. All we want to do is detaching an inode from the 20355a694dfSTejun Heo * current owner if it's being written to by some other cgroups too much. 20455a694dfSTejun Heo * 20555a694dfSTejun Heo * The current cgroup writeback is built on the assumption that multiple 20655a694dfSTejun Heo * cgroups writing to the same inode concurrently is very rare and a mode 20755a694dfSTejun Heo * of operation which isn't well supported. As such, the goal is not 20855a694dfSTejun Heo * taking too long when a different cgroup takes over an inode while 20955a694dfSTejun Heo * avoiding too aggressive flip-flops from occasional foreign writes. 21055a694dfSTejun Heo * 21155a694dfSTejun Heo * We record, very roughly, 2s worth of IO time history and if more than 21255a694dfSTejun Heo * half of that is foreign, trigger the switch. The recording is quantized 21355a694dfSTejun Heo * to 16 slots. To avoid tiny writes from swinging the decision too much, 21455a694dfSTejun Heo * writes smaller than 1/8 of avg size are ignored. 21555a694dfSTejun Heo */ 2162a814908STejun Heo #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */ 2172a814908STejun Heo #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */ 21855a694dfSTejun Heo #define WB_FRN_TIME_CUT_DIV 8 /* ignore rounds < avg / 8 */ 2192a814908STejun Heo #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */ 2202a814908STejun Heo 2212a814908STejun Heo #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */ 2222a814908STejun Heo #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS) 2232a814908STejun Heo /* each slot's duration is 2s / 16 */ 2242a814908STejun Heo #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2) 2252a814908STejun Heo /* if foreign slots >= 8, switch */ 2262a814908STejun Heo #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1) 2272a814908STejun Heo /* one round can affect upto 5 slots */ 2286444f47eSTejun Heo #define WB_FRN_MAX_IN_FLIGHT 1024 /* don't queue too many concurrently */ 2292a814908STejun Heo 230c22d70a1SRoman Gushchin /* 231c22d70a1SRoman Gushchin * Maximum inodes per isw. A specific value has been chosen to make 232c22d70a1SRoman Gushchin * struct inode_switch_wbs_context fit into 1024 bytes kmalloc. 233c22d70a1SRoman Gushchin */ 234c22d70a1SRoman Gushchin #define WB_MAX_INODES_PER_ISW ((1024UL - sizeof(struct inode_switch_wbs_context)) \ 235c22d70a1SRoman Gushchin / sizeof(struct inode *)) 236c22d70a1SRoman Gushchin 237a1a0e23eSTejun Heo static atomic_t isw_nr_in_flight = ATOMIC_INIT(0); 238a1a0e23eSTejun Heo static struct workqueue_struct *isw_wq; 239a1a0e23eSTejun Heo 240*9cfb816bSMatthew Wilcox (Oracle) void __inode_attach_wb(struct inode *inode, struct folio *folio) 24121c6321fSTejun Heo { 24221c6321fSTejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 24321c6321fSTejun Heo struct bdi_writeback *wb = NULL; 24421c6321fSTejun Heo 24521c6321fSTejun Heo if (inode_cgwb_enabled(inode)) { 24621c6321fSTejun Heo struct cgroup_subsys_state *memcg_css; 24721c6321fSTejun Heo 248*9cfb816bSMatthew Wilcox (Oracle) if (folio) { 249*9cfb816bSMatthew Wilcox (Oracle) memcg_css = mem_cgroup_css_from_page(&folio->page); 25021c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 25121c6321fSTejun Heo } else { 25221c6321fSTejun Heo /* must pin memcg_css, see wb_get_create() */ 25321c6321fSTejun Heo memcg_css = task_get_css(current, memory_cgrp_id); 25421c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 25521c6321fSTejun Heo css_put(memcg_css); 25621c6321fSTejun Heo } 25721c6321fSTejun Heo } 25821c6321fSTejun Heo 25921c6321fSTejun Heo if (!wb) 26021c6321fSTejun Heo wb = &bdi->wb; 26121c6321fSTejun Heo 26221c6321fSTejun Heo /* 26321c6321fSTejun Heo * There may be multiple instances of this function racing to 26421c6321fSTejun Heo * update the same inode. Use cmpxchg() to tell the winner. 26521c6321fSTejun Heo */ 26621c6321fSTejun Heo if (unlikely(cmpxchg(&inode->i_wb, NULL, wb))) 26721c6321fSTejun Heo wb_put(wb); 26821c6321fSTejun Heo } 2699b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(__inode_attach_wb); 27021c6321fSTejun Heo 271703c2708STejun Heo /** 272f3b6a6dfSRoman Gushchin * inode_cgwb_move_to_attached - put the inode onto wb->b_attached list 273f3b6a6dfSRoman Gushchin * @inode: inode of interest with i_lock held 274f3b6a6dfSRoman Gushchin * @wb: target bdi_writeback 275f3b6a6dfSRoman Gushchin * 276f3b6a6dfSRoman Gushchin * Remove the inode from wb's io lists and if necessarily put onto b_attached 277f3b6a6dfSRoman Gushchin * list. Only inodes attached to cgwb's are kept on this list. 278f3b6a6dfSRoman Gushchin */ 279f3b6a6dfSRoman Gushchin static void inode_cgwb_move_to_attached(struct inode *inode, 280f3b6a6dfSRoman Gushchin struct bdi_writeback *wb) 281f3b6a6dfSRoman Gushchin { 282f3b6a6dfSRoman Gushchin assert_spin_locked(&wb->list_lock); 283f3b6a6dfSRoman Gushchin assert_spin_locked(&inode->i_lock); 284a9438b44SJan Kara WARN_ON_ONCE(inode->i_state & I_FREEING); 285f3b6a6dfSRoman Gushchin 286f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 287f3b6a6dfSRoman Gushchin if (wb != &wb->bdi->wb) 288f3b6a6dfSRoman Gushchin list_move(&inode->i_io_list, &wb->b_attached); 289f3b6a6dfSRoman Gushchin else 290f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 291f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 292f3b6a6dfSRoman Gushchin } 293f3b6a6dfSRoman Gushchin 294f3b6a6dfSRoman Gushchin /** 29587e1d789STejun Heo * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it 29687e1d789STejun Heo * @inode: inode of interest with i_lock held 29787e1d789STejun Heo * 29887e1d789STejun Heo * Returns @inode's wb with its list_lock held. @inode->i_lock must be 29987e1d789STejun Heo * held on entry and is released on return. The returned wb is guaranteed 30087e1d789STejun Heo * to stay @inode's associated wb until its list_lock is released. 30187e1d789STejun Heo */ 30287e1d789STejun Heo static struct bdi_writeback * 30387e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 30487e1d789STejun Heo __releases(&inode->i_lock) 30587e1d789STejun Heo __acquires(&wb->list_lock) 30687e1d789STejun Heo { 30787e1d789STejun Heo while (true) { 30887e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 30987e1d789STejun Heo 31087e1d789STejun Heo /* 31187e1d789STejun Heo * inode_to_wb() association is protected by both 31287e1d789STejun Heo * @inode->i_lock and @wb->list_lock but list_lock nests 31387e1d789STejun Heo * outside i_lock. Drop i_lock and verify that the 31487e1d789STejun Heo * association hasn't changed after acquiring list_lock. 31587e1d789STejun Heo */ 31687e1d789STejun Heo wb_get(wb); 31787e1d789STejun Heo spin_unlock(&inode->i_lock); 31887e1d789STejun Heo spin_lock(&wb->list_lock); 31987e1d789STejun Heo 320aaa2cacfSTejun Heo /* i_wb may have changed inbetween, can't use inode_to_wb() */ 321614a4e37STejun Heo if (likely(wb == inode->i_wb)) { 322614a4e37STejun Heo wb_put(wb); /* @inode already has ref */ 323614a4e37STejun Heo return wb; 324614a4e37STejun Heo } 32587e1d789STejun Heo 32687e1d789STejun Heo spin_unlock(&wb->list_lock); 327614a4e37STejun Heo wb_put(wb); 32887e1d789STejun Heo cpu_relax(); 32987e1d789STejun Heo spin_lock(&inode->i_lock); 33087e1d789STejun Heo } 33187e1d789STejun Heo } 33287e1d789STejun Heo 33387e1d789STejun Heo /** 33487e1d789STejun Heo * inode_to_wb_and_lock_list - determine an inode's wb and lock it 33587e1d789STejun Heo * @inode: inode of interest 33687e1d789STejun Heo * 33787e1d789STejun Heo * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held 33887e1d789STejun Heo * on entry. 33987e1d789STejun Heo */ 34087e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 34187e1d789STejun Heo __acquires(&wb->list_lock) 34287e1d789STejun Heo { 34387e1d789STejun Heo spin_lock(&inode->i_lock); 34487e1d789STejun Heo return locked_inode_to_wb_and_lock_list(inode); 34587e1d789STejun Heo } 34687e1d789STejun Heo 347682aa8e1STejun Heo struct inode_switch_wbs_context { 34829264d92SRoman Gushchin struct rcu_work work; 349f5fbe6b7SRoman Gushchin 350f5fbe6b7SRoman Gushchin /* 351f5fbe6b7SRoman Gushchin * Multiple inodes can be switched at once. The switching procedure 352f5fbe6b7SRoman Gushchin * consists of two parts, separated by a RCU grace period. To make 353f5fbe6b7SRoman Gushchin * sure that the second part is executed for each inode gone through 354f5fbe6b7SRoman Gushchin * the first part, all inode pointers are placed into a NULL-terminated 355f5fbe6b7SRoman Gushchin * array embedded into struct inode_switch_wbs_context. Otherwise 356f5fbe6b7SRoman Gushchin * an inode could be left in a non-consistent state. 357f5fbe6b7SRoman Gushchin */ 358f5fbe6b7SRoman Gushchin struct bdi_writeback *new_wb; 359f5fbe6b7SRoman Gushchin struct inode *inodes[]; 360682aa8e1STejun Heo }; 361682aa8e1STejun Heo 3627fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) 3637fc5854fSTejun Heo { 3647fc5854fSTejun Heo down_write(&bdi->wb_switch_rwsem); 3657fc5854fSTejun Heo } 3667fc5854fSTejun Heo 3677fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) 3687fc5854fSTejun Heo { 3697fc5854fSTejun Heo up_write(&bdi->wb_switch_rwsem); 3707fc5854fSTejun Heo } 3717fc5854fSTejun Heo 372f5fbe6b7SRoman Gushchin static bool inode_do_switch_wbs(struct inode *inode, 373f5fbe6b7SRoman Gushchin struct bdi_writeback *old_wb, 37472d4512eSRoman Gushchin struct bdi_writeback *new_wb) 375682aa8e1STejun Heo { 376d10c8095STejun Heo struct address_space *mapping = inode->i_mapping; 37704edf02cSMatthew Wilcox XA_STATE(xas, &mapping->i_pages, 0); 37822b3c8d6SMatthew Wilcox (Oracle) struct folio *folio; 379d10c8095STejun Heo bool switched = false; 380682aa8e1STejun Heo 381682aa8e1STejun Heo spin_lock(&inode->i_lock); 382b93b0163SMatthew Wilcox xa_lock_irq(&mapping->i_pages); 383682aa8e1STejun Heo 384d10c8095STejun Heo /* 3854ade5867SRoman Gushchin * Once I_FREEING or I_WILL_FREE are visible under i_lock, the eviction 3864ade5867SRoman Gushchin * path owns the inode and we shouldn't modify ->i_io_list. 387d10c8095STejun Heo */ 3884ade5867SRoman Gushchin if (unlikely(inode->i_state & (I_FREEING | I_WILL_FREE))) 389d10c8095STejun Heo goto skip_switch; 390d10c8095STejun Heo 3913a8e9ac8STejun Heo trace_inode_switch_wbs(inode, old_wb, new_wb); 3923a8e9ac8STejun Heo 393d10c8095STejun Heo /* 394d10c8095STejun Heo * Count and transfer stats. Note that PAGECACHE_TAG_DIRTY points 39522b3c8d6SMatthew Wilcox (Oracle) * to possibly dirty folios while PAGECACHE_TAG_WRITEBACK points to 39622b3c8d6SMatthew Wilcox (Oracle) * folios actually under writeback. 397d10c8095STejun Heo */ 39822b3c8d6SMatthew Wilcox (Oracle) xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_DIRTY) { 39922b3c8d6SMatthew Wilcox (Oracle) if (folio_test_dirty(folio)) { 40022b3c8d6SMatthew Wilcox (Oracle) long nr = folio_nr_pages(folio); 40122b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(old_wb, WB_RECLAIMABLE, -nr); 40222b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(new_wb, WB_RECLAIMABLE, nr); 403d10c8095STejun Heo } 404d10c8095STejun Heo } 405d10c8095STejun Heo 40604edf02cSMatthew Wilcox xas_set(&xas, 0); 40722b3c8d6SMatthew Wilcox (Oracle) xas_for_each_marked(&xas, folio, ULONG_MAX, PAGECACHE_TAG_WRITEBACK) { 40822b3c8d6SMatthew Wilcox (Oracle) long nr = folio_nr_pages(folio); 40922b3c8d6SMatthew Wilcox (Oracle) WARN_ON_ONCE(!folio_test_writeback(folio)); 41022b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(old_wb, WB_WRITEBACK, -nr); 41122b3c8d6SMatthew Wilcox (Oracle) wb_stat_mod(new_wb, WB_WRITEBACK, nr); 412d10c8095STejun Heo } 413d10c8095STejun Heo 414633a2abbSJan Kara if (mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) { 415633a2abbSJan Kara atomic_dec(&old_wb->writeback_inodes); 416633a2abbSJan Kara atomic_inc(&new_wb->writeback_inodes); 417633a2abbSJan Kara } 418633a2abbSJan Kara 419d10c8095STejun Heo wb_get(new_wb); 420d10c8095STejun Heo 421d10c8095STejun Heo /* 422f3b6a6dfSRoman Gushchin * Transfer to @new_wb's IO list if necessary. If the @inode is dirty, 423f3b6a6dfSRoman Gushchin * the specific list @inode was on is ignored and the @inode is put on 424f3b6a6dfSRoman Gushchin * ->b_dirty which is always correct including from ->b_dirty_time. 425f3b6a6dfSRoman Gushchin * The transfer preserves @inode->dirtied_when ordering. If the @inode 426f3b6a6dfSRoman Gushchin * was clean, it means it was on the b_attached list, so move it onto 427f3b6a6dfSRoman Gushchin * the b_attached list of @new_wb. 428d10c8095STejun Heo */ 429c7f54084SDave Chinner if (!list_empty(&inode->i_io_list)) { 430f3b6a6dfSRoman Gushchin inode->i_wb = new_wb; 431f3b6a6dfSRoman Gushchin 432f3b6a6dfSRoman Gushchin if (inode->i_state & I_DIRTY_ALL) { 433d10c8095STejun Heo struct inode *pos; 434d10c8095STejun Heo 435c7f54084SDave Chinner list_for_each_entry(pos, &new_wb->b_dirty, i_io_list) 436d10c8095STejun Heo if (time_after_eq(inode->dirtied_when, 437d10c8095STejun Heo pos->dirtied_when)) 438d10c8095STejun Heo break; 439f3b6a6dfSRoman Gushchin inode_io_list_move_locked(inode, new_wb, 440f3b6a6dfSRoman Gushchin pos->i_io_list.prev); 441f3b6a6dfSRoman Gushchin } else { 442f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, new_wb); 443f3b6a6dfSRoman Gushchin } 444d10c8095STejun Heo } else { 445d10c8095STejun Heo inode->i_wb = new_wb; 446d10c8095STejun Heo } 447d10c8095STejun Heo 448d10c8095STejun Heo /* ->i_wb_frn updates may race wbc_detach_inode() but doesn't matter */ 449682aa8e1STejun Heo inode->i_wb_frn_winner = 0; 450682aa8e1STejun Heo inode->i_wb_frn_avg_time = 0; 451682aa8e1STejun Heo inode->i_wb_frn_history = 0; 452d10c8095STejun Heo switched = true; 453d10c8095STejun Heo skip_switch: 454682aa8e1STejun Heo /* 455682aa8e1STejun Heo * Paired with load_acquire in unlocked_inode_to_wb_begin() and 456682aa8e1STejun Heo * ensures that the new wb is visible if they see !I_WB_SWITCH. 457682aa8e1STejun Heo */ 458682aa8e1STejun Heo smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); 459682aa8e1STejun Heo 460b93b0163SMatthew Wilcox xa_unlock_irq(&mapping->i_pages); 461682aa8e1STejun Heo spin_unlock(&inode->i_lock); 462d10c8095STejun Heo 463f5fbe6b7SRoman Gushchin return switched; 46472d4512eSRoman Gushchin } 465682aa8e1STejun Heo 46672d4512eSRoman Gushchin static void inode_switch_wbs_work_fn(struct work_struct *work) 46772d4512eSRoman Gushchin { 46872d4512eSRoman Gushchin struct inode_switch_wbs_context *isw = 46972d4512eSRoman Gushchin container_of(to_rcu_work(work), struct inode_switch_wbs_context, work); 470f5fbe6b7SRoman Gushchin struct backing_dev_info *bdi = inode_to_bdi(isw->inodes[0]); 471f5fbe6b7SRoman Gushchin struct bdi_writeback *old_wb = isw->inodes[0]->i_wb; 472f5fbe6b7SRoman Gushchin struct bdi_writeback *new_wb = isw->new_wb; 473f5fbe6b7SRoman Gushchin unsigned long nr_switched = 0; 474f5fbe6b7SRoman Gushchin struct inode **inodep; 47572d4512eSRoman Gushchin 476f5fbe6b7SRoman Gushchin /* 477f5fbe6b7SRoman Gushchin * If @inode switches cgwb membership while sync_inodes_sb() is 478f5fbe6b7SRoman Gushchin * being issued, sync_inodes_sb() might miss it. Synchronize. 479f5fbe6b7SRoman Gushchin */ 480f5fbe6b7SRoman Gushchin down_read(&bdi->wb_switch_rwsem); 481f5fbe6b7SRoman Gushchin 482f5fbe6b7SRoman Gushchin /* 483f5fbe6b7SRoman Gushchin * By the time control reaches here, RCU grace period has passed 484f5fbe6b7SRoman Gushchin * since I_WB_SWITCH assertion and all wb stat update transactions 485f5fbe6b7SRoman Gushchin * between unlocked_inode_to_wb_begin/end() are guaranteed to be 486f5fbe6b7SRoman Gushchin * synchronizing against the i_pages lock. 487f5fbe6b7SRoman Gushchin * 488f5fbe6b7SRoman Gushchin * Grabbing old_wb->list_lock, inode->i_lock and the i_pages lock 489f5fbe6b7SRoman Gushchin * gives us exclusion against all wb related operations on @inode 490f5fbe6b7SRoman Gushchin * including IO list manipulations and stat updates. 491f5fbe6b7SRoman Gushchin */ 492f5fbe6b7SRoman Gushchin if (old_wb < new_wb) { 493f5fbe6b7SRoman Gushchin spin_lock(&old_wb->list_lock); 494f5fbe6b7SRoman Gushchin spin_lock_nested(&new_wb->list_lock, SINGLE_DEPTH_NESTING); 495f5fbe6b7SRoman Gushchin } else { 496f5fbe6b7SRoman Gushchin spin_lock(&new_wb->list_lock); 497f5fbe6b7SRoman Gushchin spin_lock_nested(&old_wb->list_lock, SINGLE_DEPTH_NESTING); 498f5fbe6b7SRoman Gushchin } 499f5fbe6b7SRoman Gushchin 500f5fbe6b7SRoman Gushchin for (inodep = isw->inodes; *inodep; inodep++) { 501f5fbe6b7SRoman Gushchin WARN_ON_ONCE((*inodep)->i_wb != old_wb); 502f5fbe6b7SRoman Gushchin if (inode_do_switch_wbs(*inodep, old_wb, new_wb)) 503f5fbe6b7SRoman Gushchin nr_switched++; 504f5fbe6b7SRoman Gushchin } 505f5fbe6b7SRoman Gushchin 506f5fbe6b7SRoman Gushchin spin_unlock(&new_wb->list_lock); 507f5fbe6b7SRoman Gushchin spin_unlock(&old_wb->list_lock); 508f5fbe6b7SRoman Gushchin 509f5fbe6b7SRoman Gushchin up_read(&bdi->wb_switch_rwsem); 510f5fbe6b7SRoman Gushchin 511f5fbe6b7SRoman Gushchin if (nr_switched) { 512f5fbe6b7SRoman Gushchin wb_wakeup(new_wb); 513f5fbe6b7SRoman Gushchin wb_put_many(old_wb, nr_switched); 514f5fbe6b7SRoman Gushchin } 515f5fbe6b7SRoman Gushchin 516f5fbe6b7SRoman Gushchin for (inodep = isw->inodes; *inodep; inodep++) 517f5fbe6b7SRoman Gushchin iput(*inodep); 518f5fbe6b7SRoman Gushchin wb_put(new_wb); 519682aa8e1STejun Heo kfree(isw); 520a1a0e23eSTejun Heo atomic_dec(&isw_nr_in_flight); 521682aa8e1STejun Heo } 522682aa8e1STejun Heo 523c22d70a1SRoman Gushchin static bool inode_prepare_wbs_switch(struct inode *inode, 524c22d70a1SRoman Gushchin struct bdi_writeback *new_wb) 525c22d70a1SRoman Gushchin { 526c22d70a1SRoman Gushchin /* 527c22d70a1SRoman Gushchin * Paired with smp_mb() in cgroup_writeback_umount(). 528c22d70a1SRoman Gushchin * isw_nr_in_flight must be increased before checking SB_ACTIVE and 529c22d70a1SRoman Gushchin * grabbing an inode, otherwise isw_nr_in_flight can be observed as 0 530c22d70a1SRoman Gushchin * in cgroup_writeback_umount() and the isw_wq will be not flushed. 531c22d70a1SRoman Gushchin */ 532c22d70a1SRoman Gushchin smp_mb(); 533c22d70a1SRoman Gushchin 534593311e8SRoman Gushchin if (IS_DAX(inode)) 535593311e8SRoman Gushchin return false; 536593311e8SRoman Gushchin 537c22d70a1SRoman Gushchin /* while holding I_WB_SWITCH, no one else can update the association */ 538c22d70a1SRoman Gushchin spin_lock(&inode->i_lock); 539c22d70a1SRoman Gushchin if (!(inode->i_sb->s_flags & SB_ACTIVE) || 540c22d70a1SRoman Gushchin inode->i_state & (I_WB_SWITCH | I_FREEING | I_WILL_FREE) || 541c22d70a1SRoman Gushchin inode_to_wb(inode) == new_wb) { 542c22d70a1SRoman Gushchin spin_unlock(&inode->i_lock); 543c22d70a1SRoman Gushchin return false; 544c22d70a1SRoman Gushchin } 545c22d70a1SRoman Gushchin inode->i_state |= I_WB_SWITCH; 546c22d70a1SRoman Gushchin __iget(inode); 547c22d70a1SRoman Gushchin spin_unlock(&inode->i_lock); 548c22d70a1SRoman Gushchin 549c22d70a1SRoman Gushchin return true; 550c22d70a1SRoman Gushchin } 551c22d70a1SRoman Gushchin 552682aa8e1STejun Heo /** 553682aa8e1STejun Heo * inode_switch_wbs - change the wb association of an inode 554682aa8e1STejun Heo * @inode: target inode 555682aa8e1STejun Heo * @new_wb_id: ID of the new wb 556682aa8e1STejun Heo * 557682aa8e1STejun Heo * Switch @inode's wb association to the wb identified by @new_wb_id. The 558682aa8e1STejun Heo * switching is performed asynchronously and may fail silently. 559682aa8e1STejun Heo */ 560682aa8e1STejun Heo static void inode_switch_wbs(struct inode *inode, int new_wb_id) 561682aa8e1STejun Heo { 562682aa8e1STejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 563682aa8e1STejun Heo struct cgroup_subsys_state *memcg_css; 564682aa8e1STejun Heo struct inode_switch_wbs_context *isw; 565682aa8e1STejun Heo 566682aa8e1STejun Heo /* noop if seems to be already in progress */ 567682aa8e1STejun Heo if (inode->i_state & I_WB_SWITCH) 568682aa8e1STejun Heo return; 569682aa8e1STejun Heo 5706444f47eSTejun Heo /* avoid queueing a new switch if too many are already in flight */ 5716444f47eSTejun Heo if (atomic_read(&isw_nr_in_flight) > WB_FRN_MAX_IN_FLIGHT) 5727fc5854fSTejun Heo return; 5737fc5854fSTejun Heo 57498b160c8SLen Baker isw = kzalloc(struct_size(isw, inodes, 2), GFP_ATOMIC); 575682aa8e1STejun Heo if (!isw) 5766444f47eSTejun Heo return; 577682aa8e1STejun Heo 5788826ee4fSRoman Gushchin atomic_inc(&isw_nr_in_flight); 5798826ee4fSRoman Gushchin 580682aa8e1STejun Heo /* find and pin the new wb */ 581682aa8e1STejun Heo rcu_read_lock(); 582682aa8e1STejun Heo memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); 5838b0ed844SMuchun Song if (memcg_css && !css_tryget(memcg_css)) 5848b0ed844SMuchun Song memcg_css = NULL; 585682aa8e1STejun Heo rcu_read_unlock(); 5868b0ed844SMuchun Song if (!memcg_css) 5878b0ed844SMuchun Song goto out_free; 5888b0ed844SMuchun Song 5898b0ed844SMuchun Song isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 5908b0ed844SMuchun Song css_put(memcg_css); 591682aa8e1STejun Heo if (!isw->new_wb) 592682aa8e1STejun Heo goto out_free; 593682aa8e1STejun Heo 594c22d70a1SRoman Gushchin if (!inode_prepare_wbs_switch(inode, isw->new_wb)) 595a1a0e23eSTejun Heo goto out_free; 596682aa8e1STejun Heo 597f5fbe6b7SRoman Gushchin isw->inodes[0] = inode; 598682aa8e1STejun Heo 599682aa8e1STejun Heo /* 600682aa8e1STejun Heo * In addition to synchronizing among switchers, I_WB_SWITCH tells 601b93b0163SMatthew Wilcox * the RCU protected stat update paths to grab the i_page 602b93b0163SMatthew Wilcox * lock so that stat transfer can synchronize against them. 603682aa8e1STejun Heo * Let's continue after I_WB_SWITCH is guaranteed to be visible. 604682aa8e1STejun Heo */ 60529264d92SRoman Gushchin INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); 60629264d92SRoman Gushchin queue_rcu_work(isw_wq, &isw->work); 6076444f47eSTejun Heo return; 608682aa8e1STejun Heo 609682aa8e1STejun Heo out_free: 6108826ee4fSRoman Gushchin atomic_dec(&isw_nr_in_flight); 611682aa8e1STejun Heo if (isw->new_wb) 612682aa8e1STejun Heo wb_put(isw->new_wb); 613682aa8e1STejun Heo kfree(isw); 614682aa8e1STejun Heo } 615682aa8e1STejun Heo 61687e1d789STejun Heo /** 617c22d70a1SRoman Gushchin * cleanup_offline_cgwb - detach associated inodes 618c22d70a1SRoman Gushchin * @wb: target wb 619c22d70a1SRoman Gushchin * 620c22d70a1SRoman Gushchin * Switch all inodes attached to @wb to a nearest living ancestor's wb in order 621c22d70a1SRoman Gushchin * to eventually release the dying @wb. Returns %true if not all inodes were 622c22d70a1SRoman Gushchin * switched and the function has to be restarted. 623c22d70a1SRoman Gushchin */ 624c22d70a1SRoman Gushchin bool cleanup_offline_cgwb(struct bdi_writeback *wb) 625c22d70a1SRoman Gushchin { 626c22d70a1SRoman Gushchin struct cgroup_subsys_state *memcg_css; 627c22d70a1SRoman Gushchin struct inode_switch_wbs_context *isw; 628c22d70a1SRoman Gushchin struct inode *inode; 629c22d70a1SRoman Gushchin int nr; 630c22d70a1SRoman Gushchin bool restart = false; 631c22d70a1SRoman Gushchin 63298b160c8SLen Baker isw = kzalloc(struct_size(isw, inodes, WB_MAX_INODES_PER_ISW), 63398b160c8SLen Baker GFP_KERNEL); 634c22d70a1SRoman Gushchin if (!isw) 635c22d70a1SRoman Gushchin return restart; 636c22d70a1SRoman Gushchin 637c22d70a1SRoman Gushchin atomic_inc(&isw_nr_in_flight); 638c22d70a1SRoman Gushchin 639c22d70a1SRoman Gushchin for (memcg_css = wb->memcg_css->parent; memcg_css; 640c22d70a1SRoman Gushchin memcg_css = memcg_css->parent) { 641c22d70a1SRoman Gushchin isw->new_wb = wb_get_create(wb->bdi, memcg_css, GFP_KERNEL); 642c22d70a1SRoman Gushchin if (isw->new_wb) 643c22d70a1SRoman Gushchin break; 644c22d70a1SRoman Gushchin } 645c22d70a1SRoman Gushchin if (unlikely(!isw->new_wb)) 646c22d70a1SRoman Gushchin isw->new_wb = &wb->bdi->wb; /* wb_get() is noop for bdi's wb */ 647c22d70a1SRoman Gushchin 648c22d70a1SRoman Gushchin nr = 0; 649c22d70a1SRoman Gushchin spin_lock(&wb->list_lock); 650c22d70a1SRoman Gushchin list_for_each_entry(inode, &wb->b_attached, i_io_list) { 651c22d70a1SRoman Gushchin if (!inode_prepare_wbs_switch(inode, isw->new_wb)) 652c22d70a1SRoman Gushchin continue; 653c22d70a1SRoman Gushchin 654c22d70a1SRoman Gushchin isw->inodes[nr++] = inode; 655c22d70a1SRoman Gushchin 656c22d70a1SRoman Gushchin if (nr >= WB_MAX_INODES_PER_ISW - 1) { 657c22d70a1SRoman Gushchin restart = true; 658c22d70a1SRoman Gushchin break; 659c22d70a1SRoman Gushchin } 660c22d70a1SRoman Gushchin } 661c22d70a1SRoman Gushchin spin_unlock(&wb->list_lock); 662c22d70a1SRoman Gushchin 663c22d70a1SRoman Gushchin /* no attached inodes? bail out */ 664c22d70a1SRoman Gushchin if (nr == 0) { 665c22d70a1SRoman Gushchin atomic_dec(&isw_nr_in_flight); 666c22d70a1SRoman Gushchin wb_put(isw->new_wb); 667c22d70a1SRoman Gushchin kfree(isw); 668c22d70a1SRoman Gushchin return restart; 669c22d70a1SRoman Gushchin } 670c22d70a1SRoman Gushchin 671c22d70a1SRoman Gushchin /* 672c22d70a1SRoman Gushchin * In addition to synchronizing among switchers, I_WB_SWITCH tells 673c22d70a1SRoman Gushchin * the RCU protected stat update paths to grab the i_page 674c22d70a1SRoman Gushchin * lock so that stat transfer can synchronize against them. 675c22d70a1SRoman Gushchin * Let's continue after I_WB_SWITCH is guaranteed to be visible. 676c22d70a1SRoman Gushchin */ 677c22d70a1SRoman Gushchin INIT_RCU_WORK(&isw->work, inode_switch_wbs_work_fn); 678c22d70a1SRoman Gushchin queue_rcu_work(isw_wq, &isw->work); 679c22d70a1SRoman Gushchin 680c22d70a1SRoman Gushchin return restart; 681c22d70a1SRoman Gushchin } 682c22d70a1SRoman Gushchin 683c22d70a1SRoman Gushchin /** 684b16b1debSTejun Heo * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it 685b16b1debSTejun Heo * @wbc: writeback_control of interest 686b16b1debSTejun Heo * @inode: target inode 687b16b1debSTejun Heo * 688b16b1debSTejun Heo * @inode is locked and about to be written back under the control of @wbc. 689b16b1debSTejun Heo * Record @inode's writeback context into @wbc and unlock the i_lock. On 690b16b1debSTejun Heo * writeback completion, wbc_detach_inode() should be called. This is used 691b16b1debSTejun Heo * to track the cgroup writeback context. 692b16b1debSTejun Heo */ 693b16b1debSTejun Heo void wbc_attach_and_unlock_inode(struct writeback_control *wbc, 694b16b1debSTejun Heo struct inode *inode) 695b16b1debSTejun Heo { 696dd73e4b7STejun Heo if (!inode_cgwb_enabled(inode)) { 697dd73e4b7STejun Heo spin_unlock(&inode->i_lock); 698dd73e4b7STejun Heo return; 699dd73e4b7STejun Heo } 700dd73e4b7STejun Heo 701b16b1debSTejun Heo wbc->wb = inode_to_wb(inode); 7022a814908STejun Heo wbc->inode = inode; 7032a814908STejun Heo 7042a814908STejun Heo wbc->wb_id = wbc->wb->memcg_css->id; 7052a814908STejun Heo wbc->wb_lcand_id = inode->i_wb_frn_winner; 7062a814908STejun Heo wbc->wb_tcand_id = 0; 7072a814908STejun Heo wbc->wb_bytes = 0; 7082a814908STejun Heo wbc->wb_lcand_bytes = 0; 7092a814908STejun Heo wbc->wb_tcand_bytes = 0; 7102a814908STejun Heo 711b16b1debSTejun Heo wb_get(wbc->wb); 712b16b1debSTejun Heo spin_unlock(&inode->i_lock); 713e8a7abf5STejun Heo 714e8a7abf5STejun Heo /* 71565de03e2STejun Heo * A dying wb indicates that either the blkcg associated with the 71665de03e2STejun Heo * memcg changed or the associated memcg is dying. In the first 71765de03e2STejun Heo * case, a replacement wb should already be available and we should 71865de03e2STejun Heo * refresh the wb immediately. In the second case, trying to 71965de03e2STejun Heo * refresh will keep failing. 720e8a7abf5STejun Heo */ 72165de03e2STejun Heo if (unlikely(wb_dying(wbc->wb) && !css_is_dying(wbc->wb->memcg_css))) 722e8a7abf5STejun Heo inode_switch_wbs(inode, wbc->wb_id); 723b16b1debSTejun Heo } 7249b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_attach_and_unlock_inode); 725b16b1debSTejun Heo 726b16b1debSTejun Heo /** 7272a814908STejun Heo * wbc_detach_inode - disassociate wbc from inode and perform foreign detection 7282a814908STejun Heo * @wbc: writeback_control of the just finished writeback 729b16b1debSTejun Heo * 730b16b1debSTejun Heo * To be called after a writeback attempt of an inode finishes and undoes 731b16b1debSTejun Heo * wbc_attach_and_unlock_inode(). Can be called under any context. 7322a814908STejun Heo * 7332a814908STejun Heo * As concurrent write sharing of an inode is expected to be very rare and 7342a814908STejun Heo * memcg only tracks page ownership on first-use basis severely confining 7352a814908STejun Heo * the usefulness of such sharing, cgroup writeback tracks ownership 7362a814908STejun Heo * per-inode. While the support for concurrent write sharing of an inode 7372a814908STejun Heo * is deemed unnecessary, an inode being written to by different cgroups at 7382a814908STejun Heo * different points in time is a lot more common, and, more importantly, 7392a814908STejun Heo * charging only by first-use can too readily lead to grossly incorrect 7402a814908STejun Heo * behaviors (single foreign page can lead to gigabytes of writeback to be 7412a814908STejun Heo * incorrectly attributed). 7422a814908STejun Heo * 7432a814908STejun Heo * To resolve this issue, cgroup writeback detects the majority dirtier of 7442999e1e3SJulia Lawall * an inode and transfers the ownership to it. To avoid unnecessary 7452a814908STejun Heo * oscillation, the detection mechanism keeps track of history and gives 7462a814908STejun Heo * out the switch verdict only if the foreign usage pattern is stable over 7472a814908STejun Heo * a certain amount of time and/or writeback attempts. 7482a814908STejun Heo * 7492a814908STejun Heo * On each writeback attempt, @wbc tries to detect the majority writer 7502a814908STejun Heo * using Boyer-Moore majority vote algorithm. In addition to the byte 7512a814908STejun Heo * count from the majority voting, it also counts the bytes written for the 7522a814908STejun Heo * current wb and the last round's winner wb (max of last round's current 7532a814908STejun Heo * wb, the winner from two rounds ago, and the last round's majority 7542a814908STejun Heo * candidate). Keeping track of the historical winner helps the algorithm 7552a814908STejun Heo * to semi-reliably detect the most active writer even when it's not the 7562a814908STejun Heo * absolute majority. 7572a814908STejun Heo * 7582a814908STejun Heo * Once the winner of the round is determined, whether the winner is 7592a814908STejun Heo * foreign or not and how much IO time the round consumed is recorded in 7602a814908STejun Heo * inode->i_wb_frn_history. If the amount of recorded foreign IO time is 7612a814908STejun Heo * over a certain threshold, the switch verdict is given. 762b16b1debSTejun Heo */ 763b16b1debSTejun Heo void wbc_detach_inode(struct writeback_control *wbc) 764b16b1debSTejun Heo { 7652a814908STejun Heo struct bdi_writeback *wb = wbc->wb; 7662a814908STejun Heo struct inode *inode = wbc->inode; 767dd73e4b7STejun Heo unsigned long avg_time, max_bytes, max_time; 768dd73e4b7STejun Heo u16 history; 7692a814908STejun Heo int max_id; 7702a814908STejun Heo 771dd73e4b7STejun Heo if (!wb) 772dd73e4b7STejun Heo return; 773dd73e4b7STejun Heo 774dd73e4b7STejun Heo history = inode->i_wb_frn_history; 775dd73e4b7STejun Heo avg_time = inode->i_wb_frn_avg_time; 776dd73e4b7STejun Heo 7772a814908STejun Heo /* pick the winner of this round */ 7782a814908STejun Heo if (wbc->wb_bytes >= wbc->wb_lcand_bytes && 7792a814908STejun Heo wbc->wb_bytes >= wbc->wb_tcand_bytes) { 7802a814908STejun Heo max_id = wbc->wb_id; 7812a814908STejun Heo max_bytes = wbc->wb_bytes; 7822a814908STejun Heo } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { 7832a814908STejun Heo max_id = wbc->wb_lcand_id; 7842a814908STejun Heo max_bytes = wbc->wb_lcand_bytes; 7852a814908STejun Heo } else { 7862a814908STejun Heo max_id = wbc->wb_tcand_id; 7872a814908STejun Heo max_bytes = wbc->wb_tcand_bytes; 7882a814908STejun Heo } 7892a814908STejun Heo 7902a814908STejun Heo /* 7912a814908STejun Heo * Calculate the amount of IO time the winner consumed and fold it 7922a814908STejun Heo * into the running average kept per inode. If the consumed IO 7932a814908STejun Heo * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for 7942a814908STejun Heo * deciding whether to switch or not. This is to prevent one-off 7952a814908STejun Heo * small dirtiers from skewing the verdict. 7962a814908STejun Heo */ 7972a814908STejun Heo max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, 7982a814908STejun Heo wb->avg_write_bandwidth); 7992a814908STejun Heo if (avg_time) 8002a814908STejun Heo avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - 8012a814908STejun Heo (avg_time >> WB_FRN_TIME_AVG_SHIFT); 8022a814908STejun Heo else 8032a814908STejun Heo avg_time = max_time; /* immediate catch up on first run */ 8042a814908STejun Heo 8052a814908STejun Heo if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { 8062a814908STejun Heo int slots; 8072a814908STejun Heo 8082a814908STejun Heo /* 8092a814908STejun Heo * The switch verdict is reached if foreign wb's consume 8102a814908STejun Heo * more than a certain proportion of IO time in a 8112a814908STejun Heo * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot 8122a814908STejun Heo * history mask where each bit represents one sixteenth of 8132a814908STejun Heo * the period. Determine the number of slots to shift into 8142a814908STejun Heo * history from @max_time. 8152a814908STejun Heo */ 8162a814908STejun Heo slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), 8172a814908STejun Heo (unsigned long)WB_FRN_HIST_MAX_SLOTS); 8182a814908STejun Heo history <<= slots; 8192a814908STejun Heo if (wbc->wb_id != max_id) 8202a814908STejun Heo history |= (1U << slots) - 1; 8212a814908STejun Heo 8223a8e9ac8STejun Heo if (history) 8233a8e9ac8STejun Heo trace_inode_foreign_history(inode, wbc, history); 8243a8e9ac8STejun Heo 8252a814908STejun Heo /* 8262a814908STejun Heo * Switch if the current wb isn't the consistent winner. 8272a814908STejun Heo * If there are multiple closely competing dirtiers, the 8282a814908STejun Heo * inode may switch across them repeatedly over time, which 8292a814908STejun Heo * is okay. The main goal is avoiding keeping an inode on 8302a814908STejun Heo * the wrong wb for an extended period of time. 8312a814908STejun Heo */ 832682aa8e1STejun Heo if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) 833682aa8e1STejun Heo inode_switch_wbs(inode, max_id); 8342a814908STejun Heo } 8352a814908STejun Heo 8362a814908STejun Heo /* 8372a814908STejun Heo * Multiple instances of this function may race to update the 8382a814908STejun Heo * following fields but we don't mind occassional inaccuracies. 8392a814908STejun Heo */ 8402a814908STejun Heo inode->i_wb_frn_winner = max_id; 8412a814908STejun Heo inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); 8422a814908STejun Heo inode->i_wb_frn_history = history; 8432a814908STejun Heo 844b16b1debSTejun Heo wb_put(wbc->wb); 845b16b1debSTejun Heo wbc->wb = NULL; 846b16b1debSTejun Heo } 8479b0eb69bSTejun Heo EXPORT_SYMBOL_GPL(wbc_detach_inode); 848b16b1debSTejun Heo 849b16b1debSTejun Heo /** 85034e51a5eSTejun Heo * wbc_account_cgroup_owner - account writeback to update inode cgroup ownership 8512a814908STejun Heo * @wbc: writeback_control of the writeback in progress 8522a814908STejun Heo * @page: page being written out 8532a814908STejun Heo * @bytes: number of bytes being written out 8542a814908STejun Heo * 8552a814908STejun Heo * @bytes from @page are about to written out during the writeback 8562a814908STejun Heo * controlled by @wbc. Keep the book for foreign inode detection. See 8572a814908STejun Heo * wbc_detach_inode(). 8582a814908STejun Heo */ 85934e51a5eSTejun Heo void wbc_account_cgroup_owner(struct writeback_control *wbc, struct page *page, 8602a814908STejun Heo size_t bytes) 8612a814908STejun Heo { 86266311422STejun Heo struct cgroup_subsys_state *css; 8632a814908STejun Heo int id; 8642a814908STejun Heo 8652a814908STejun Heo /* 8662a814908STejun Heo * pageout() path doesn't attach @wbc to the inode being written 8672a814908STejun Heo * out. This is intentional as we don't want the function to block 8682a814908STejun Heo * behind a slow cgroup. Ultimately, we want pageout() to kick off 8692a814908STejun Heo * regular writeback instead of writing things out itself. 8702a814908STejun Heo */ 87127b36d8fSTejun Heo if (!wbc->wb || wbc->no_cgroup_owner) 8722a814908STejun Heo return; 8732a814908STejun Heo 87466311422STejun Heo css = mem_cgroup_css_from_page(page); 87566311422STejun Heo /* dead cgroups shouldn't contribute to inode ownership arbitration */ 87666311422STejun Heo if (!(css->flags & CSS_ONLINE)) 87766311422STejun Heo return; 87866311422STejun Heo 87966311422STejun Heo id = css->id; 8802a814908STejun Heo 8812a814908STejun Heo if (id == wbc->wb_id) { 8822a814908STejun Heo wbc->wb_bytes += bytes; 8832a814908STejun Heo return; 8842a814908STejun Heo } 8852a814908STejun Heo 8862a814908STejun Heo if (id == wbc->wb_lcand_id) 8872a814908STejun Heo wbc->wb_lcand_bytes += bytes; 8882a814908STejun Heo 8892a814908STejun Heo /* Boyer-Moore majority vote algorithm */ 8902a814908STejun Heo if (!wbc->wb_tcand_bytes) 8912a814908STejun Heo wbc->wb_tcand_id = id; 8922a814908STejun Heo if (id == wbc->wb_tcand_id) 8932a814908STejun Heo wbc->wb_tcand_bytes += bytes; 8942a814908STejun Heo else 8952a814908STejun Heo wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); 8962a814908STejun Heo } 89734e51a5eSTejun Heo EXPORT_SYMBOL_GPL(wbc_account_cgroup_owner); 8982a814908STejun Heo 8992a814908STejun Heo /** 900f2b65121STejun Heo * wb_split_bdi_pages - split nr_pages to write according to bandwidth 901f2b65121STejun Heo * @wb: target bdi_writeback to split @nr_pages to 902f2b65121STejun Heo * @nr_pages: number of pages to write for the whole bdi 903f2b65121STejun Heo * 904f2b65121STejun Heo * Split @wb's portion of @nr_pages according to @wb's write bandwidth in 905f2b65121STejun Heo * relation to the total write bandwidth of all wb's w/ dirty inodes on 906f2b65121STejun Heo * @wb->bdi. 907f2b65121STejun Heo */ 908f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 909f2b65121STejun Heo { 910f2b65121STejun Heo unsigned long this_bw = wb->avg_write_bandwidth; 911f2b65121STejun Heo unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); 912f2b65121STejun Heo 913f2b65121STejun Heo if (nr_pages == LONG_MAX) 914f2b65121STejun Heo return LONG_MAX; 915f2b65121STejun Heo 916f2b65121STejun Heo /* 917f2b65121STejun Heo * This may be called on clean wb's and proportional distribution 918f2b65121STejun Heo * may not make sense, just use the original @nr_pages in those 919f2b65121STejun Heo * cases. In general, we wanna err on the side of writing more. 920f2b65121STejun Heo */ 921f2b65121STejun Heo if (!tot_bw || this_bw >= tot_bw) 922f2b65121STejun Heo return nr_pages; 923f2b65121STejun Heo else 924f2b65121STejun Heo return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); 925f2b65121STejun Heo } 926f2b65121STejun Heo 927db125360STejun Heo /** 928db125360STejun Heo * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi 929db125360STejun Heo * @bdi: target backing_dev_info 930db125360STejun Heo * @base_work: wb_writeback_work to issue 931db125360STejun Heo * @skip_if_busy: skip wb's which already have writeback in progress 932db125360STejun Heo * 933db125360STejun Heo * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which 934db125360STejun Heo * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's 935db125360STejun Heo * distributed to the busy wbs according to each wb's proportion in the 936db125360STejun Heo * total active write bandwidth of @bdi. 937db125360STejun Heo */ 938db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 939db125360STejun Heo struct wb_writeback_work *base_work, 940db125360STejun Heo bool skip_if_busy) 941db125360STejun Heo { 942b817525aSTejun Heo struct bdi_writeback *last_wb = NULL; 943b33e18f6STejun Heo struct bdi_writeback *wb = list_entry(&bdi->wb_list, 944b817525aSTejun Heo struct bdi_writeback, bdi_node); 945db125360STejun Heo 946db125360STejun Heo might_sleep(); 947db125360STejun Heo restart: 948db125360STejun Heo rcu_read_lock(); 949b817525aSTejun Heo list_for_each_entry_continue_rcu(wb, &bdi->wb_list, bdi_node) { 9505b9cce4cSTejun Heo DEFINE_WB_COMPLETION(fallback_work_done, bdi); 9518a1270cdSTejun Heo struct wb_writeback_work fallback_work; 9528a1270cdSTejun Heo struct wb_writeback_work *work; 9538a1270cdSTejun Heo long nr_pages; 9548a1270cdSTejun Heo 955b817525aSTejun Heo if (last_wb) { 956b817525aSTejun Heo wb_put(last_wb); 957b817525aSTejun Heo last_wb = NULL; 958b817525aSTejun Heo } 959b817525aSTejun Heo 960006a0973STejun Heo /* SYNC_ALL writes out I_DIRTY_TIME too */ 961006a0973STejun Heo if (!wb_has_dirty_io(wb) && 962006a0973STejun Heo (base_work->sync_mode == WB_SYNC_NONE || 963006a0973STejun Heo list_empty(&wb->b_dirty_time))) 964006a0973STejun Heo continue; 965006a0973STejun Heo if (skip_if_busy && writeback_in_progress(wb)) 966db125360STejun Heo continue; 967db125360STejun Heo 9688a1270cdSTejun Heo nr_pages = wb_split_bdi_pages(wb, base_work->nr_pages); 9698a1270cdSTejun Heo 9708a1270cdSTejun Heo work = kmalloc(sizeof(*work), GFP_ATOMIC); 9718a1270cdSTejun Heo if (work) { 9728a1270cdSTejun Heo *work = *base_work; 9738a1270cdSTejun Heo work->nr_pages = nr_pages; 9748a1270cdSTejun Heo work->auto_free = 1; 9758a1270cdSTejun Heo wb_queue_work(wb, work); 9768a1270cdSTejun Heo continue; 977db125360STejun Heo } 9788a1270cdSTejun Heo 9798a1270cdSTejun Heo /* alloc failed, execute synchronously using on-stack fallback */ 9808a1270cdSTejun Heo work = &fallback_work; 9818a1270cdSTejun Heo *work = *base_work; 9828a1270cdSTejun Heo work->nr_pages = nr_pages; 9838a1270cdSTejun Heo work->auto_free = 0; 9848a1270cdSTejun Heo work->done = &fallback_work_done; 9858a1270cdSTejun Heo 9868a1270cdSTejun Heo wb_queue_work(wb, work); 9878a1270cdSTejun Heo 988b817525aSTejun Heo /* 989b817525aSTejun Heo * Pin @wb so that it stays on @bdi->wb_list. This allows 990b817525aSTejun Heo * continuing iteration from @wb after dropping and 991b817525aSTejun Heo * regrabbing rcu read lock. 992b817525aSTejun Heo */ 993b817525aSTejun Heo wb_get(wb); 994b817525aSTejun Heo last_wb = wb; 995b817525aSTejun Heo 996db125360STejun Heo rcu_read_unlock(); 9975b9cce4cSTejun Heo wb_wait_for_completion(&fallback_work_done); 998db125360STejun Heo goto restart; 999db125360STejun Heo } 1000db125360STejun Heo rcu_read_unlock(); 1001b817525aSTejun Heo 1002b817525aSTejun Heo if (last_wb) 1003b817525aSTejun Heo wb_put(last_wb); 1004db125360STejun Heo } 1005db125360STejun Heo 1006a1a0e23eSTejun Heo /** 1007d62241c7STejun Heo * cgroup_writeback_by_id - initiate cgroup writeback from bdi and memcg IDs 1008d62241c7STejun Heo * @bdi_id: target bdi id 1009d62241c7STejun Heo * @memcg_id: target memcg css id 1010d62241c7STejun Heo * @reason: reason why some writeback work initiated 1011d62241c7STejun Heo * @done: target wb_completion 1012d62241c7STejun Heo * 1013d62241c7STejun Heo * Initiate flush of the bdi_writeback identified by @bdi_id and @memcg_id 1014d62241c7STejun Heo * with the specified parameters. 1015d62241c7STejun Heo */ 10167490a2d2SShakeel Butt int cgroup_writeback_by_id(u64 bdi_id, int memcg_id, 1017d62241c7STejun Heo enum wb_reason reason, struct wb_completion *done) 1018d62241c7STejun Heo { 1019d62241c7STejun Heo struct backing_dev_info *bdi; 1020d62241c7STejun Heo struct cgroup_subsys_state *memcg_css; 1021d62241c7STejun Heo struct bdi_writeback *wb; 1022d62241c7STejun Heo struct wb_writeback_work *work; 10237490a2d2SShakeel Butt unsigned long dirty; 1024d62241c7STejun Heo int ret; 1025d62241c7STejun Heo 1026d62241c7STejun Heo /* lookup bdi and memcg */ 1027d62241c7STejun Heo bdi = bdi_get_by_id(bdi_id); 1028d62241c7STejun Heo if (!bdi) 1029d62241c7STejun Heo return -ENOENT; 1030d62241c7STejun Heo 1031d62241c7STejun Heo rcu_read_lock(); 1032d62241c7STejun Heo memcg_css = css_from_id(memcg_id, &memory_cgrp_subsys); 1033d62241c7STejun Heo if (memcg_css && !css_tryget(memcg_css)) 1034d62241c7STejun Heo memcg_css = NULL; 1035d62241c7STejun Heo rcu_read_unlock(); 1036d62241c7STejun Heo if (!memcg_css) { 1037d62241c7STejun Heo ret = -ENOENT; 1038d62241c7STejun Heo goto out_bdi_put; 1039d62241c7STejun Heo } 1040d62241c7STejun Heo 1041d62241c7STejun Heo /* 1042d62241c7STejun Heo * And find the associated wb. If the wb isn't there already 1043d62241c7STejun Heo * there's nothing to flush, don't create one. 1044d62241c7STejun Heo */ 1045d62241c7STejun Heo wb = wb_get_lookup(bdi, memcg_css); 1046d62241c7STejun Heo if (!wb) { 1047d62241c7STejun Heo ret = -ENOENT; 1048d62241c7STejun Heo goto out_css_put; 1049d62241c7STejun Heo } 1050d62241c7STejun Heo 1051d62241c7STejun Heo /* 10527490a2d2SShakeel Butt * The caller is attempting to write out most of 1053d62241c7STejun Heo * the currently dirty pages. Let's take the current dirty page 1054d62241c7STejun Heo * count and inflate it by 25% which should be large enough to 1055d62241c7STejun Heo * flush out most dirty pages while avoiding getting livelocked by 1056d62241c7STejun Heo * concurrent dirtiers. 10577490a2d2SShakeel Butt * 10587490a2d2SShakeel Butt * BTW the memcg stats are flushed periodically and this is best-effort 10597490a2d2SShakeel Butt * estimation, so some potential error is ok. 1060d62241c7STejun Heo */ 10617490a2d2SShakeel Butt dirty = memcg_page_state(mem_cgroup_from_css(memcg_css), NR_FILE_DIRTY); 10627490a2d2SShakeel Butt dirty = dirty * 10 / 8; 1063d62241c7STejun Heo 1064d62241c7STejun Heo /* issue the writeback work */ 1065d62241c7STejun Heo work = kzalloc(sizeof(*work), GFP_NOWAIT | __GFP_NOWARN); 1066d62241c7STejun Heo if (work) { 10677490a2d2SShakeel Butt work->nr_pages = dirty; 1068d62241c7STejun Heo work->sync_mode = WB_SYNC_NONE; 1069d62241c7STejun Heo work->range_cyclic = 1; 1070d62241c7STejun Heo work->reason = reason; 1071d62241c7STejun Heo work->done = done; 1072d62241c7STejun Heo work->auto_free = 1; 1073d62241c7STejun Heo wb_queue_work(wb, work); 1074d62241c7STejun Heo ret = 0; 1075d62241c7STejun Heo } else { 1076d62241c7STejun Heo ret = -ENOMEM; 1077d62241c7STejun Heo } 1078d62241c7STejun Heo 1079d62241c7STejun Heo wb_put(wb); 1080d62241c7STejun Heo out_css_put: 1081d62241c7STejun Heo css_put(memcg_css); 1082d62241c7STejun Heo out_bdi_put: 1083d62241c7STejun Heo bdi_put(bdi); 1084d62241c7STejun Heo return ret; 1085d62241c7STejun Heo } 1086d62241c7STejun Heo 1087d62241c7STejun Heo /** 1088a1a0e23eSTejun Heo * cgroup_writeback_umount - flush inode wb switches for umount 1089a1a0e23eSTejun Heo * 1090a1a0e23eSTejun Heo * This function is called when a super_block is about to be destroyed and 1091a1a0e23eSTejun Heo * flushes in-flight inode wb switches. An inode wb switch goes through 1092a1a0e23eSTejun Heo * RCU and then workqueue, so the two need to be flushed in order to ensure 1093a1a0e23eSTejun Heo * that all previously scheduled switches are finished. As wb switches are 1094a1a0e23eSTejun Heo * rare occurrences and synchronize_rcu() can take a while, perform 1095a1a0e23eSTejun Heo * flushing iff wb switches are in flight. 1096a1a0e23eSTejun Heo */ 1097a1a0e23eSTejun Heo void cgroup_writeback_umount(void) 1098a1a0e23eSTejun Heo { 1099592fa002SRoman Gushchin /* 1100592fa002SRoman Gushchin * SB_ACTIVE should be reliably cleared before checking 1101592fa002SRoman Gushchin * isw_nr_in_flight, see generic_shutdown_super(). 1102592fa002SRoman Gushchin */ 1103592fa002SRoman Gushchin smp_mb(); 1104592fa002SRoman Gushchin 1105a1a0e23eSTejun Heo if (atomic_read(&isw_nr_in_flight)) { 1106ec084de9SJiufei Xue /* 1107ec084de9SJiufei Xue * Use rcu_barrier() to wait for all pending callbacks to 1108ec084de9SJiufei Xue * ensure that all in-flight wb switches are in the workqueue. 1109ec084de9SJiufei Xue */ 1110ec084de9SJiufei Xue rcu_barrier(); 1111a1a0e23eSTejun Heo flush_workqueue(isw_wq); 1112a1a0e23eSTejun Heo } 1113a1a0e23eSTejun Heo } 1114a1a0e23eSTejun Heo 1115a1a0e23eSTejun Heo static int __init cgroup_writeback_init(void) 1116a1a0e23eSTejun Heo { 1117a1a0e23eSTejun Heo isw_wq = alloc_workqueue("inode_switch_wbs", 0, 0); 1118a1a0e23eSTejun Heo if (!isw_wq) 1119a1a0e23eSTejun Heo return -ENOMEM; 1120a1a0e23eSTejun Heo return 0; 1121a1a0e23eSTejun Heo } 1122a1a0e23eSTejun Heo fs_initcall(cgroup_writeback_init); 1123a1a0e23eSTejun Heo 1124f2b65121STejun Heo #else /* CONFIG_CGROUP_WRITEBACK */ 1125f2b65121STejun Heo 11267fc5854fSTejun Heo static void bdi_down_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 11277fc5854fSTejun Heo static void bdi_up_write_wb_switch_rwsem(struct backing_dev_info *bdi) { } 11287fc5854fSTejun Heo 1129f3b6a6dfSRoman Gushchin static void inode_cgwb_move_to_attached(struct inode *inode, 1130f3b6a6dfSRoman Gushchin struct bdi_writeback *wb) 1131f3b6a6dfSRoman Gushchin { 1132f3b6a6dfSRoman Gushchin assert_spin_locked(&wb->list_lock); 1133f3b6a6dfSRoman Gushchin assert_spin_locked(&inode->i_lock); 1134a9438b44SJan Kara WARN_ON_ONCE(inode->i_state & I_FREEING); 1135f3b6a6dfSRoman Gushchin 1136f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 1137f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 1138f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 1139f3b6a6dfSRoman Gushchin } 1140f3b6a6dfSRoman Gushchin 114187e1d789STejun Heo static struct bdi_writeback * 114287e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 114387e1d789STejun Heo __releases(&inode->i_lock) 114487e1d789STejun Heo __acquires(&wb->list_lock) 114587e1d789STejun Heo { 114687e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 114787e1d789STejun Heo 114887e1d789STejun Heo spin_unlock(&inode->i_lock); 114987e1d789STejun Heo spin_lock(&wb->list_lock); 115087e1d789STejun Heo return wb; 115187e1d789STejun Heo } 115287e1d789STejun Heo 115387e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 115487e1d789STejun Heo __acquires(&wb->list_lock) 115587e1d789STejun Heo { 115687e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 115787e1d789STejun Heo 115887e1d789STejun Heo spin_lock(&wb->list_lock); 115987e1d789STejun Heo return wb; 116087e1d789STejun Heo } 116187e1d789STejun Heo 1162f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 1163f2b65121STejun Heo { 1164f2b65121STejun Heo return nr_pages; 1165f2b65121STejun Heo } 1166f2b65121STejun Heo 1167db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 1168db125360STejun Heo struct wb_writeback_work *base_work, 1169db125360STejun Heo bool skip_if_busy) 1170db125360STejun Heo { 1171db125360STejun Heo might_sleep(); 1172db125360STejun Heo 1173006a0973STejun Heo if (!skip_if_busy || !writeback_in_progress(&bdi->wb)) { 1174db125360STejun Heo base_work->auto_free = 0; 1175db125360STejun Heo wb_queue_work(&bdi->wb, base_work); 1176db125360STejun Heo } 1177db125360STejun Heo } 1178db125360STejun Heo 1179703c2708STejun Heo #endif /* CONFIG_CGROUP_WRITEBACK */ 1180703c2708STejun Heo 1181e8e8a0c6SJens Axboe /* 1182e8e8a0c6SJens Axboe * Add in the number of potentially dirty inodes, because each inode 1183e8e8a0c6SJens Axboe * write can dirty pagecache in the underlying blockdev. 1184e8e8a0c6SJens Axboe */ 1185e8e8a0c6SJens Axboe static unsigned long get_nr_dirty_pages(void) 1186e8e8a0c6SJens Axboe { 1187e8e8a0c6SJens Axboe return global_node_page_state(NR_FILE_DIRTY) + 1188e8e8a0c6SJens Axboe get_nr_dirty_inodes(); 1189e8e8a0c6SJens Axboe } 1190e8e8a0c6SJens Axboe 1191e8e8a0c6SJens Axboe static void wb_start_writeback(struct bdi_writeback *wb, enum wb_reason reason) 1192b6e51316SJens Axboe { 1193c00ddad3STejun Heo if (!wb_has_dirty_io(wb)) 1194c00ddad3STejun Heo return; 1195c00ddad3STejun Heo 1196c00ddad3STejun Heo /* 1197aac8d41cSJens Axboe * All callers of this function want to start writeback of all 1198aac8d41cSJens Axboe * dirty pages. Places like vmscan can call this at a very 1199aac8d41cSJens Axboe * high frequency, causing pointless allocations of tons of 1200aac8d41cSJens Axboe * work items and keeping the flusher threads busy retrieving 1201aac8d41cSJens Axboe * that work. Ensure that we only allow one of them pending and 120285009b4fSJens Axboe * inflight at the time. 1203aac8d41cSJens Axboe */ 120485009b4fSJens Axboe if (test_bit(WB_start_all, &wb->state) || 120585009b4fSJens Axboe test_and_set_bit(WB_start_all, &wb->state)) 1206aac8d41cSJens Axboe return; 1207aac8d41cSJens Axboe 120885009b4fSJens Axboe wb->start_all_reason = reason; 1209c00ddad3STejun Heo wb_wakeup(wb); 1210d3ddec76SWu Fengguang } 1211d3ddec76SWu Fengguang 1212c5444198SChristoph Hellwig /** 12139ecf4866STejun Heo * wb_start_background_writeback - start background writeback 12149ecf4866STejun Heo * @wb: bdi_writback to write from 1215c5444198SChristoph Hellwig * 1216c5444198SChristoph Hellwig * Description: 12176585027aSJan Kara * This makes sure WB_SYNC_NONE background writeback happens. When 12189ecf4866STejun Heo * this function returns, it is only guaranteed that for given wb 12196585027aSJan Kara * some IO is happening if we are over background dirty threshold. 12206585027aSJan Kara * Caller need not hold sb s_umount semaphore. 1221c5444198SChristoph Hellwig */ 12229ecf4866STejun Heo void wb_start_background_writeback(struct bdi_writeback *wb) 1223c5444198SChristoph Hellwig { 12246585027aSJan Kara /* 12256585027aSJan Kara * We just wake up the flusher thread. It will perform background 12266585027aSJan Kara * writeback as soon as there is no other work to do. 12276585027aSJan Kara */ 12285634cc2aSTejun Heo trace_writeback_wake_background(wb); 12299ecf4866STejun Heo wb_wakeup(wb); 12301da177e4SLinus Torvalds } 12311da177e4SLinus Torvalds 12321da177e4SLinus Torvalds /* 1233a66979abSDave Chinner * Remove the inode from the writeback list it is on. 1234a66979abSDave Chinner */ 1235c7f54084SDave Chinner void inode_io_list_del(struct inode *inode) 1236a66979abSDave Chinner { 123787e1d789STejun Heo struct bdi_writeback *wb; 1238a66979abSDave Chinner 123987e1d789STejun Heo wb = inode_to_wb_and_lock_list(inode); 1240b35250c0SJan Kara spin_lock(&inode->i_lock); 1241f3b6a6dfSRoman Gushchin 1242f3b6a6dfSRoman Gushchin inode->i_state &= ~I_SYNC_QUEUED; 1243f3b6a6dfSRoman Gushchin list_del_init(&inode->i_io_list); 1244f3b6a6dfSRoman Gushchin wb_io_lists_depopulated(wb); 1245f3b6a6dfSRoman Gushchin 1246b35250c0SJan Kara spin_unlock(&inode->i_lock); 124752ebea74STejun Heo spin_unlock(&wb->list_lock); 1248f758eeabSChristoph Hellwig } 12494301efa4SJan Kara EXPORT_SYMBOL(inode_io_list_del); 1250a66979abSDave Chinner 1251a66979abSDave Chinner /* 12526c60d2b5SDave Chinner * mark an inode as under writeback on the sb 12536c60d2b5SDave Chinner */ 12546c60d2b5SDave Chinner void sb_mark_inode_writeback(struct inode *inode) 12556c60d2b5SDave Chinner { 12566c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 12576c60d2b5SDave Chinner unsigned long flags; 12586c60d2b5SDave Chinner 12596c60d2b5SDave Chinner if (list_empty(&inode->i_wb_list)) { 12606c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 12619a46b04fSBrian Foster if (list_empty(&inode->i_wb_list)) { 12626c60d2b5SDave Chinner list_add_tail(&inode->i_wb_list, &sb->s_inodes_wb); 12639a46b04fSBrian Foster trace_sb_mark_inode_writeback(inode); 12649a46b04fSBrian Foster } 12656c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 12666c60d2b5SDave Chinner } 12676c60d2b5SDave Chinner } 12686c60d2b5SDave Chinner 12696c60d2b5SDave Chinner /* 12706c60d2b5SDave Chinner * clear an inode as under writeback on the sb 12716c60d2b5SDave Chinner */ 12726c60d2b5SDave Chinner void sb_clear_inode_writeback(struct inode *inode) 12736c60d2b5SDave Chinner { 12746c60d2b5SDave Chinner struct super_block *sb = inode->i_sb; 12756c60d2b5SDave Chinner unsigned long flags; 12766c60d2b5SDave Chinner 12776c60d2b5SDave Chinner if (!list_empty(&inode->i_wb_list)) { 12786c60d2b5SDave Chinner spin_lock_irqsave(&sb->s_inode_wblist_lock, flags); 12799a46b04fSBrian Foster if (!list_empty(&inode->i_wb_list)) { 12806c60d2b5SDave Chinner list_del_init(&inode->i_wb_list); 12819a46b04fSBrian Foster trace_sb_clear_inode_writeback(inode); 12829a46b04fSBrian Foster } 12836c60d2b5SDave Chinner spin_unlock_irqrestore(&sb->s_inode_wblist_lock, flags); 12846c60d2b5SDave Chinner } 12856c60d2b5SDave Chinner } 12866c60d2b5SDave Chinner 12876c60d2b5SDave Chinner /* 12886610a0bcSAndrew Morton * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 12896610a0bcSAndrew Morton * furthest end of its superblock's dirty-inode list. 12906610a0bcSAndrew Morton * 12916610a0bcSAndrew Morton * Before stamping the inode's ->dirtied_when, we check to see whether it is 129266f3b8e2SJens Axboe * already the most-recently-dirtied inode on the b_dirty list. If that is 12936610a0bcSAndrew Morton * the case then the inode must have been redirtied while it was being written 12946610a0bcSAndrew Morton * out and we don't reset its dirtied_when. 12956610a0bcSAndrew Morton */ 1296b35250c0SJan Kara static void redirty_tail_locked(struct inode *inode, struct bdi_writeback *wb) 12976610a0bcSAndrew Morton { 1298b35250c0SJan Kara assert_spin_locked(&inode->i_lock); 1299b35250c0SJan Kara 1300a9438b44SJan Kara inode->i_state &= ~I_SYNC_QUEUED; 1301a9438b44SJan Kara /* 1302a9438b44SJan Kara * When the inode is being freed just don't bother with dirty list 1303a9438b44SJan Kara * tracking. Flush worker will ignore this inode anyway and it will 1304a9438b44SJan Kara * trigger assertions in inode_io_list_move_locked(). 1305a9438b44SJan Kara */ 1306a9438b44SJan Kara if (inode->i_state & I_FREEING) { 1307a9438b44SJan Kara list_del_init(&inode->i_io_list); 1308a9438b44SJan Kara wb_io_lists_depopulated(wb); 1309a9438b44SJan Kara return; 1310a9438b44SJan Kara } 131103ba3782SJens Axboe if (!list_empty(&wb->b_dirty)) { 131266f3b8e2SJens Axboe struct inode *tail; 13136610a0bcSAndrew Morton 13147ccf19a8SNick Piggin tail = wb_inode(wb->b_dirty.next); 131566f3b8e2SJens Axboe if (time_before(inode->dirtied_when, tail->dirtied_when)) 13166610a0bcSAndrew Morton inode->dirtied_when = jiffies; 13176610a0bcSAndrew Morton } 1318c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty); 13196610a0bcSAndrew Morton } 13206610a0bcSAndrew Morton 1321b35250c0SJan Kara static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) 1322b35250c0SJan Kara { 1323b35250c0SJan Kara spin_lock(&inode->i_lock); 1324b35250c0SJan Kara redirty_tail_locked(inode, wb); 1325b35250c0SJan Kara spin_unlock(&inode->i_lock); 1326b35250c0SJan Kara } 1327b35250c0SJan Kara 13286610a0bcSAndrew Morton /* 132966f3b8e2SJens Axboe * requeue inode for re-scanning after bdi->b_io list is exhausted. 1330c986d1e2SAndrew Morton */ 1331f758eeabSChristoph Hellwig static void requeue_io(struct inode *inode, struct bdi_writeback *wb) 1332c986d1e2SAndrew Morton { 1333c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_more_io); 1334c986d1e2SAndrew Morton } 1335c986d1e2SAndrew Morton 13361c0eeaf5SJoern Engel static void inode_sync_complete(struct inode *inode) 13371c0eeaf5SJoern Engel { 1338365b94aeSJan Kara inode->i_state &= ~I_SYNC; 13394eff96ddSJan Kara /* If inode is clean an unused, put it into LRU now... */ 13404eff96ddSJan Kara inode_add_lru(inode); 1341365b94aeSJan Kara /* Waiters must see I_SYNC cleared before being woken up */ 13421c0eeaf5SJoern Engel smp_mb(); 13431c0eeaf5SJoern Engel wake_up_bit(&inode->i_state, __I_SYNC); 13441c0eeaf5SJoern Engel } 13451c0eeaf5SJoern Engel 1346d2caa3c5SJeff Layton static bool inode_dirtied_after(struct inode *inode, unsigned long t) 1347d2caa3c5SJeff Layton { 1348d2caa3c5SJeff Layton bool ret = time_after(inode->dirtied_when, t); 1349d2caa3c5SJeff Layton #ifndef CONFIG_64BIT 1350d2caa3c5SJeff Layton /* 1351d2caa3c5SJeff Layton * For inodes being constantly redirtied, dirtied_when can get stuck. 1352d2caa3c5SJeff Layton * It _appears_ to be in the future, but is actually in distant past. 1353d2caa3c5SJeff Layton * This test is necessary to prevent such wrapped-around relative times 13545b0830cbSJens Axboe * from permanently stopping the whole bdi writeback. 1355d2caa3c5SJeff Layton */ 1356d2caa3c5SJeff Layton ret = ret && time_before_eq(inode->dirtied_when, jiffies); 1357d2caa3c5SJeff Layton #endif 1358d2caa3c5SJeff Layton return ret; 1359d2caa3c5SJeff Layton } 1360d2caa3c5SJeff Layton 1361c986d1e2SAndrew Morton /* 1362f9cae926SJan Kara * Move expired (dirtied before dirtied_before) dirty inodes from 1363697e6fedSJan Kara * @delaying_queue to @dispatch_queue. 13642c136579SFengguang Wu */ 1365e84d0a4fSWu Fengguang static int move_expired_inodes(struct list_head *delaying_queue, 13662c136579SFengguang Wu struct list_head *dispatch_queue, 13675fcd5750SJan Kara unsigned long dirtied_before) 13682c136579SFengguang Wu { 13695c03449dSShaohua Li LIST_HEAD(tmp); 13705c03449dSShaohua Li struct list_head *pos, *node; 1371cf137307SJens Axboe struct super_block *sb = NULL; 13725c03449dSShaohua Li struct inode *inode; 1373cf137307SJens Axboe int do_sb_sort = 0; 1374e84d0a4fSWu Fengguang int moved = 0; 13755c03449dSShaohua Li 13762c136579SFengguang Wu while (!list_empty(delaying_queue)) { 13777ccf19a8SNick Piggin inode = wb_inode(delaying_queue->prev); 1378f9cae926SJan Kara if (inode_dirtied_after(inode, dirtied_before)) 13792c136579SFengguang Wu break; 138010e14073SJchao Sun spin_lock(&inode->i_lock); 1381c7f54084SDave Chinner list_move(&inode->i_io_list, &tmp); 1382a8855990SJan Kara moved++; 13835afced3bSJan Kara inode->i_state |= I_SYNC_QUEUED; 13845afced3bSJan Kara spin_unlock(&inode->i_lock); 1385a8855990SJan Kara if (sb_is_blkdev_sb(inode->i_sb)) 1386a8855990SJan Kara continue; 1387cf137307SJens Axboe if (sb && sb != inode->i_sb) 1388cf137307SJens Axboe do_sb_sort = 1; 1389cf137307SJens Axboe sb = inode->i_sb; 13905c03449dSShaohua Li } 13915c03449dSShaohua Li 1392cf137307SJens Axboe /* just one sb in list, splice to dispatch_queue and we're done */ 1393cf137307SJens Axboe if (!do_sb_sort) { 1394cf137307SJens Axboe list_splice(&tmp, dispatch_queue); 1395e84d0a4fSWu Fengguang goto out; 1396cf137307SJens Axboe } 1397cf137307SJens Axboe 139810e14073SJchao Sun /* 139910e14073SJchao Sun * Although inode's i_io_list is moved from 'tmp' to 'dispatch_queue', 140010e14073SJchao Sun * we don't take inode->i_lock here because it is just a pointless overhead. 140110e14073SJchao Sun * Inode is already marked as I_SYNC_QUEUED so writeback list handling is 140210e14073SJchao Sun * fully under our control. 140310e14073SJchao Sun */ 14045c03449dSShaohua Li while (!list_empty(&tmp)) { 14057ccf19a8SNick Piggin sb = wb_inode(tmp.prev)->i_sb; 14065c03449dSShaohua Li list_for_each_prev_safe(pos, node, &tmp) { 14077ccf19a8SNick Piggin inode = wb_inode(pos); 14085c03449dSShaohua Li if (inode->i_sb == sb) 1409c7f54084SDave Chinner list_move(&inode->i_io_list, dispatch_queue); 14102c136579SFengguang Wu } 14112c136579SFengguang Wu } 1412e84d0a4fSWu Fengguang out: 1413e84d0a4fSWu Fengguang return moved; 14145c03449dSShaohua Li } 14152c136579SFengguang Wu 14162c136579SFengguang Wu /* 14172c136579SFengguang Wu * Queue all expired dirty inodes for io, eldest first. 14184ea879b9SWu Fengguang * Before 14194ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 14204ea879b9SWu Fengguang * =============> gf edc BA 14214ea879b9SWu Fengguang * After 14224ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 14234ea879b9SWu Fengguang * =============> g fBAedc 14244ea879b9SWu Fengguang * | 14254ea879b9SWu Fengguang * +--> dequeue for IO 14262c136579SFengguang Wu */ 1427f9cae926SJan Kara static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work, 1428f9cae926SJan Kara unsigned long dirtied_before) 14292c136579SFengguang Wu { 1430e84d0a4fSWu Fengguang int moved; 1431f9cae926SJan Kara unsigned long time_expire_jif = dirtied_before; 14320ae45f63STheodore Ts'o 1433f758eeabSChristoph Hellwig assert_spin_locked(&wb->list_lock); 14344ea879b9SWu Fengguang list_splice_init(&wb->b_more_io, &wb->b_io); 14355fcd5750SJan Kara moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, dirtied_before); 1436f9cae926SJan Kara if (!work->for_sync) 1437f9cae926SJan Kara time_expire_jif = jiffies - dirtytime_expire_interval * HZ; 14380ae45f63STheodore Ts'o moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, 14395fcd5750SJan Kara time_expire_jif); 1440d6c10f1fSTejun Heo if (moved) 1441d6c10f1fSTejun Heo wb_io_lists_populated(wb); 1442f9cae926SJan Kara trace_writeback_queue_io(wb, work, dirtied_before, moved); 144366f3b8e2SJens Axboe } 144466f3b8e2SJens Axboe 1445a9185b41SChristoph Hellwig static int write_inode(struct inode *inode, struct writeback_control *wbc) 144666f3b8e2SJens Axboe { 14479fb0a7daSTejun Heo int ret; 14489fb0a7daSTejun Heo 14499fb0a7daSTejun Heo if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { 14509fb0a7daSTejun Heo trace_writeback_write_inode_start(inode, wbc); 14519fb0a7daSTejun Heo ret = inode->i_sb->s_op->write_inode(inode, wbc); 14529fb0a7daSTejun Heo trace_writeback_write_inode(inode, wbc); 14539fb0a7daSTejun Heo return ret; 14549fb0a7daSTejun Heo } 145503ba3782SJens Axboe return 0; 145666f3b8e2SJens Axboe } 145708d8e974SFengguang Wu 14582c136579SFengguang Wu /* 1459169ebd90SJan Kara * Wait for writeback on an inode to complete. Called with i_lock held. 1460169ebd90SJan Kara * Caller must make sure inode cannot go away when we drop i_lock. 146101c03194SChristoph Hellwig */ 1462169ebd90SJan Kara static void __inode_wait_for_writeback(struct inode *inode) 1463169ebd90SJan Kara __releases(inode->i_lock) 1464169ebd90SJan Kara __acquires(inode->i_lock) 146501c03194SChristoph Hellwig { 146601c03194SChristoph Hellwig DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); 146701c03194SChristoph Hellwig wait_queue_head_t *wqh; 146801c03194SChristoph Hellwig 146901c03194SChristoph Hellwig wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 147058a9d3d8SRichard Kennedy while (inode->i_state & I_SYNC) { 1471250df6edSDave Chinner spin_unlock(&inode->i_lock); 147274316201SNeilBrown __wait_on_bit(wqh, &wq, bit_wait, 147374316201SNeilBrown TASK_UNINTERRUPTIBLE); 1474250df6edSDave Chinner spin_lock(&inode->i_lock); 147558a9d3d8SRichard Kennedy } 147601c03194SChristoph Hellwig } 147701c03194SChristoph Hellwig 147801c03194SChristoph Hellwig /* 1479169ebd90SJan Kara * Wait for writeback on an inode to complete. Caller must have inode pinned. 1480169ebd90SJan Kara */ 1481169ebd90SJan Kara void inode_wait_for_writeback(struct inode *inode) 1482169ebd90SJan Kara { 1483169ebd90SJan Kara spin_lock(&inode->i_lock); 1484169ebd90SJan Kara __inode_wait_for_writeback(inode); 1485169ebd90SJan Kara spin_unlock(&inode->i_lock); 1486169ebd90SJan Kara } 1487169ebd90SJan Kara 1488169ebd90SJan Kara /* 1489169ebd90SJan Kara * Sleep until I_SYNC is cleared. This function must be called with i_lock 1490169ebd90SJan Kara * held and drops it. It is aimed for callers not holding any inode reference 1491169ebd90SJan Kara * so once i_lock is dropped, inode can go away. 1492169ebd90SJan Kara */ 1493169ebd90SJan Kara static void inode_sleep_on_writeback(struct inode *inode) 1494169ebd90SJan Kara __releases(inode->i_lock) 1495169ebd90SJan Kara { 1496169ebd90SJan Kara DEFINE_WAIT(wait); 1497169ebd90SJan Kara wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1498169ebd90SJan Kara int sleep; 1499169ebd90SJan Kara 1500169ebd90SJan Kara prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1501169ebd90SJan Kara sleep = inode->i_state & I_SYNC; 1502169ebd90SJan Kara spin_unlock(&inode->i_lock); 1503169ebd90SJan Kara if (sleep) 1504169ebd90SJan Kara schedule(); 1505169ebd90SJan Kara finish_wait(wqh, &wait); 1506169ebd90SJan Kara } 1507169ebd90SJan Kara 1508169ebd90SJan Kara /* 1509ccb26b5aSJan Kara * Find proper writeback list for the inode depending on its current state and 1510ccb26b5aSJan Kara * possibly also change of its state while we were doing writeback. Here we 1511ccb26b5aSJan Kara * handle things such as livelock prevention or fairness of writeback among 1512ccb26b5aSJan Kara * inodes. This function can be called only by flusher thread - noone else 1513ccb26b5aSJan Kara * processes all inodes in writeback lists and requeueing inodes behind flusher 1514ccb26b5aSJan Kara * thread's back can have unexpected consequences. 1515ccb26b5aSJan Kara */ 1516ccb26b5aSJan Kara static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, 1517ccb26b5aSJan Kara struct writeback_control *wbc) 1518ccb26b5aSJan Kara { 1519ccb26b5aSJan Kara if (inode->i_state & I_FREEING) 1520ccb26b5aSJan Kara return; 1521ccb26b5aSJan Kara 1522ccb26b5aSJan Kara /* 1523ccb26b5aSJan Kara * Sync livelock prevention. Each inode is tagged and synced in one 1524ccb26b5aSJan Kara * shot. If still dirty, it will be redirty_tail()'ed below. Update 1525ccb26b5aSJan Kara * the dirty time to prevent enqueue and sync it again. 1526ccb26b5aSJan Kara */ 1527ccb26b5aSJan Kara if ((inode->i_state & I_DIRTY) && 1528ccb26b5aSJan Kara (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) 1529ccb26b5aSJan Kara inode->dirtied_when = jiffies; 1530ccb26b5aSJan Kara 15314f8ad655SJan Kara if (wbc->pages_skipped) { 15324f8ad655SJan Kara /* 15334f8ad655SJan Kara * writeback is not making progress due to locked 15344f8ad655SJan Kara * buffers. Skip this inode for now. 15354f8ad655SJan Kara */ 1536b35250c0SJan Kara redirty_tail_locked(inode, wb); 15374f8ad655SJan Kara return; 15384f8ad655SJan Kara } 15394f8ad655SJan Kara 1540ccb26b5aSJan Kara if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { 1541ccb26b5aSJan Kara /* 1542ccb26b5aSJan Kara * We didn't write back all the pages. nfs_writepages() 1543ccb26b5aSJan Kara * sometimes bales out without doing anything. 1544ccb26b5aSJan Kara */ 1545ccb26b5aSJan Kara if (wbc->nr_to_write <= 0) { 1546ccb26b5aSJan Kara /* Slice used up. Queue for next turn. */ 1547ccb26b5aSJan Kara requeue_io(inode, wb); 1548ccb26b5aSJan Kara } else { 1549ccb26b5aSJan Kara /* 1550ccb26b5aSJan Kara * Writeback blocked by something other than 1551ccb26b5aSJan Kara * congestion. Delay the inode for some time to 1552ccb26b5aSJan Kara * avoid spinning on the CPU (100% iowait) 1553ccb26b5aSJan Kara * retrying writeback of the dirty page/inode 1554ccb26b5aSJan Kara * that cannot be performed immediately. 1555ccb26b5aSJan Kara */ 1556b35250c0SJan Kara redirty_tail_locked(inode, wb); 1557ccb26b5aSJan Kara } 1558ccb26b5aSJan Kara } else if (inode->i_state & I_DIRTY) { 1559ccb26b5aSJan Kara /* 1560ccb26b5aSJan Kara * Filesystems can dirty the inode during writeback operations, 1561ccb26b5aSJan Kara * such as delayed allocation during submission or metadata 1562ccb26b5aSJan Kara * updates after data IO completion. 1563ccb26b5aSJan Kara */ 1564b35250c0SJan Kara redirty_tail_locked(inode, wb); 15650ae45f63STheodore Ts'o } else if (inode->i_state & I_DIRTY_TIME) { 1566a2f48706STheodore Ts'o inode->dirtied_when = jiffies; 1567c7f54084SDave Chinner inode_io_list_move_locked(inode, wb, &wb->b_dirty_time); 15685afced3bSJan Kara inode->i_state &= ~I_SYNC_QUEUED; 1569ccb26b5aSJan Kara } else { 1570ccb26b5aSJan Kara /* The inode is clean. Remove from writeback lists. */ 1571f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, wb); 1572ccb26b5aSJan Kara } 1573ccb26b5aSJan Kara } 1574ccb26b5aSJan Kara 1575ccb26b5aSJan Kara /* 1576da0c4c60SEric Biggers * Write out an inode and its dirty pages (or some of its dirty pages, depending 1577da0c4c60SEric Biggers * on @wbc->nr_to_write), and clear the relevant dirty flags from i_state. 1578da0c4c60SEric Biggers * 1579da0c4c60SEric Biggers * This doesn't remove the inode from the writeback list it is on, except 1580da0c4c60SEric Biggers * potentially to move it from b_dirty_time to b_dirty due to timestamp 1581da0c4c60SEric Biggers * expiration. The caller is otherwise responsible for writeback list handling. 1582da0c4c60SEric Biggers * 1583da0c4c60SEric Biggers * The caller is also responsible for setting the I_SYNC flag beforehand and 1584da0c4c60SEric Biggers * calling inode_sync_complete() to clear it afterwards. 15851da177e4SLinus Torvalds */ 15861da177e4SLinus Torvalds static int 1587cd8ed2a4SYan Hong __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 15881da177e4SLinus Torvalds { 15891da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 1590251d6a47SWu Fengguang long nr_to_write = wbc->nr_to_write; 159101c03194SChristoph Hellwig unsigned dirty; 15921da177e4SLinus Torvalds int ret; 15931da177e4SLinus Torvalds 15944f8ad655SJan Kara WARN_ON(!(inode->i_state & I_SYNC)); 15951da177e4SLinus Torvalds 15969fb0a7daSTejun Heo trace_writeback_single_inode_start(inode, wbc, nr_to_write); 15979fb0a7daSTejun Heo 15981da177e4SLinus Torvalds ret = do_writepages(mapping, wbc); 15991da177e4SLinus Torvalds 160026821ed4SChristoph Hellwig /* 160126821ed4SChristoph Hellwig * Make sure to wait on the data before writing out the metadata. 160226821ed4SChristoph Hellwig * This is important for filesystems that modify metadata on data 16037747bd4bSDave Chinner * I/O completion. We don't do it for sync(2) writeback because it has a 16047747bd4bSDave Chinner * separate, external IO completion path and ->sync_fs for guaranteeing 16057747bd4bSDave Chinner * inode metadata is written back correctly. 160626821ed4SChristoph Hellwig */ 16077747bd4bSDave Chinner if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { 160826821ed4SChristoph Hellwig int err = filemap_fdatawait(mapping); 16091da177e4SLinus Torvalds if (ret == 0) 16101da177e4SLinus Torvalds ret = err; 16111da177e4SLinus Torvalds } 16121da177e4SLinus Torvalds 16135547e8aaSDmitry Monakhov /* 16141e249cb5SEric Biggers * If the inode has dirty timestamps and we need to write them, call 16151e249cb5SEric Biggers * mark_inode_dirty_sync() to notify the filesystem about it and to 16161e249cb5SEric Biggers * change I_DIRTY_TIME into I_DIRTY_SYNC. 16171e249cb5SEric Biggers */ 16181e249cb5SEric Biggers if ((inode->i_state & I_DIRTY_TIME) && 161983dc881dSEric Biggers (wbc->sync_mode == WB_SYNC_ALL || 16201e249cb5SEric Biggers time_after(jiffies, inode->dirtied_time_when + 16211e249cb5SEric Biggers dirtytime_expire_interval * HZ))) { 16221e249cb5SEric Biggers trace_writeback_lazytime(inode); 16231e249cb5SEric Biggers mark_inode_dirty_sync(inode); 16241e249cb5SEric Biggers } 16251e249cb5SEric Biggers 16261e249cb5SEric Biggers /* 1627da0c4c60SEric Biggers * Get and clear the dirty flags from i_state. This needs to be done 1628da0c4c60SEric Biggers * after calling writepages because some filesystems may redirty the 1629da0c4c60SEric Biggers * inode during writepages due to delalloc. It also needs to be done 1630da0c4c60SEric Biggers * after handling timestamp expiration, as that may dirty the inode too. 16315547e8aaSDmitry Monakhov */ 1632250df6edSDave Chinner spin_lock(&inode->i_lock); 16335547e8aaSDmitry Monakhov dirty = inode->i_state & I_DIRTY; 16340ae45f63STheodore Ts'o inode->i_state &= ~dirty; 16359c6ac78eSTejun Heo 16369c6ac78eSTejun Heo /* 16379c6ac78eSTejun Heo * Paired with smp_mb() in __mark_inode_dirty(). This allows 16389c6ac78eSTejun Heo * __mark_inode_dirty() to test i_state without grabbing i_lock - 16399c6ac78eSTejun Heo * either they see the I_DIRTY bits cleared or we see the dirtied 16409c6ac78eSTejun Heo * inode. 16419c6ac78eSTejun Heo * 16429c6ac78eSTejun Heo * I_DIRTY_PAGES is always cleared together above even if @mapping 16439c6ac78eSTejun Heo * still has dirty pages. The flag is reinstated after smp_mb() if 16449c6ac78eSTejun Heo * necessary. This guarantees that either __mark_inode_dirty() 16459c6ac78eSTejun Heo * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. 16469c6ac78eSTejun Heo */ 16479c6ac78eSTejun Heo smp_mb(); 16489c6ac78eSTejun Heo 16499c6ac78eSTejun Heo if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 16509c6ac78eSTejun Heo inode->i_state |= I_DIRTY_PAGES; 165108276bdaSDavid Howells else if (unlikely(inode->i_state & I_PINNING_FSCACHE_WB)) { 165208276bdaSDavid Howells if (!(inode->i_state & I_DIRTY_PAGES)) { 165308276bdaSDavid Howells inode->i_state &= ~I_PINNING_FSCACHE_WB; 165408276bdaSDavid Howells wbc->unpinned_fscache_wb = true; 165508276bdaSDavid Howells dirty |= I_PINNING_FSCACHE_WB; /* Cause write_inode */ 165608276bdaSDavid Howells } 165708276bdaSDavid Howells } 16589c6ac78eSTejun Heo 1659250df6edSDave Chinner spin_unlock(&inode->i_lock); 16609c6ac78eSTejun Heo 166126821ed4SChristoph Hellwig /* Don't write the inode if only I_DIRTY_PAGES was set */ 16620ae45f63STheodore Ts'o if (dirty & ~I_DIRTY_PAGES) { 1663a9185b41SChristoph Hellwig int err = write_inode(inode, wbc); 16641da177e4SLinus Torvalds if (ret == 0) 16651da177e4SLinus Torvalds ret = err; 16661da177e4SLinus Torvalds } 166708276bdaSDavid Howells wbc->unpinned_fscache_wb = false; 16684f8ad655SJan Kara trace_writeback_single_inode(inode, wbc, nr_to_write); 16694f8ad655SJan Kara return ret; 16704f8ad655SJan Kara } 16714f8ad655SJan Kara 16724f8ad655SJan Kara /* 1673da0c4c60SEric Biggers * Write out an inode's dirty data and metadata on-demand, i.e. separately from 1674da0c4c60SEric Biggers * the regular batched writeback done by the flusher threads in 1675da0c4c60SEric Biggers * writeback_sb_inodes(). @wbc controls various aspects of the write, such as 1676da0c4c60SEric Biggers * whether it is a data-integrity sync (%WB_SYNC_ALL) or not (%WB_SYNC_NONE). 16774f8ad655SJan Kara * 1678da0c4c60SEric Biggers * To prevent the inode from going away, either the caller must have a reference 1679da0c4c60SEric Biggers * to the inode, or the inode must have I_WILL_FREE or I_FREEING set. 16804f8ad655SJan Kara */ 1681aaf25593STejun Heo static int writeback_single_inode(struct inode *inode, 16824f8ad655SJan Kara struct writeback_control *wbc) 16834f8ad655SJan Kara { 1684aaf25593STejun Heo struct bdi_writeback *wb; 16854f8ad655SJan Kara int ret = 0; 16864f8ad655SJan Kara 16874f8ad655SJan Kara spin_lock(&inode->i_lock); 16884f8ad655SJan Kara if (!atomic_read(&inode->i_count)) 16894f8ad655SJan Kara WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 16904f8ad655SJan Kara else 16914f8ad655SJan Kara WARN_ON(inode->i_state & I_WILL_FREE); 16924f8ad655SJan Kara 16934f8ad655SJan Kara if (inode->i_state & I_SYNC) { 1694da0c4c60SEric Biggers /* 1695da0c4c60SEric Biggers * Writeback is already running on the inode. For WB_SYNC_NONE, 1696da0c4c60SEric Biggers * that's enough and we can just return. For WB_SYNC_ALL, we 1697da0c4c60SEric Biggers * must wait for the existing writeback to complete, then do 1698da0c4c60SEric Biggers * writeback again if there's anything left. 1699da0c4c60SEric Biggers */ 17004f8ad655SJan Kara if (wbc->sync_mode != WB_SYNC_ALL) 17014f8ad655SJan Kara goto out; 1702169ebd90SJan Kara __inode_wait_for_writeback(inode); 17034f8ad655SJan Kara } 17044f8ad655SJan Kara WARN_ON(inode->i_state & I_SYNC); 17054f8ad655SJan Kara /* 1706da0c4c60SEric Biggers * If the inode is already fully clean, then there's nothing to do. 1707da0c4c60SEric Biggers * 1708da0c4c60SEric Biggers * For data-integrity syncs we also need to check whether any pages are 1709da0c4c60SEric Biggers * still under writeback, e.g. due to prior WB_SYNC_NONE writeback. If 1710da0c4c60SEric Biggers * there are any such pages, we'll need to wait for them. 17114f8ad655SJan Kara */ 17120ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL) && 1713f9b0e058SJan Kara (wbc->sync_mode != WB_SYNC_ALL || 1714f9b0e058SJan Kara !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) 17154f8ad655SJan Kara goto out; 17164f8ad655SJan Kara inode->i_state |= I_SYNC; 1717b16b1debSTejun Heo wbc_attach_and_unlock_inode(wbc, inode); 17184f8ad655SJan Kara 1719cd8ed2a4SYan Hong ret = __writeback_single_inode(inode, wbc); 17201da177e4SLinus Torvalds 1721b16b1debSTejun Heo wbc_detach_inode(wbc); 1722aaf25593STejun Heo 1723aaf25593STejun Heo wb = inode_to_wb_and_lock_list(inode); 1724250df6edSDave Chinner spin_lock(&inode->i_lock); 17254f8ad655SJan Kara /* 17264e3c51f4SSvyatoslav Feldsherov * If the inode is freeing, its i_io_list shoudn't be updated 17274e3c51f4SSvyatoslav Feldsherov * as it can be finally deleted at this moment. 17284e3c51f4SSvyatoslav Feldsherov */ 17294e3c51f4SSvyatoslav Feldsherov if (!(inode->i_state & I_FREEING)) { 17304e3c51f4SSvyatoslav Feldsherov /* 17314e3c51f4SSvyatoslav Feldsherov * If the inode is now fully clean, then it can be safely 17324e3c51f4SSvyatoslav Feldsherov * removed from its writeback list (if any). Otherwise the 17334e3c51f4SSvyatoslav Feldsherov * flusher threads are responsible for the writeback lists. 17344f8ad655SJan Kara */ 17350ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1736f3b6a6dfSRoman Gushchin inode_cgwb_move_to_attached(inode, wb); 1737cbfecb92SLukas Czerner else if (!(inode->i_state & I_SYNC_QUEUED)) { 1738cbfecb92SLukas Czerner if ((inode->i_state & I_DIRTY)) 1739846a3351SJing Xia redirty_tail_locked(inode, wb); 1740cbfecb92SLukas Czerner else if (inode->i_state & I_DIRTY_TIME) { 1741cbfecb92SLukas Czerner inode->dirtied_when = jiffies; 17424e3c51f4SSvyatoslav Feldsherov inode_io_list_move_locked(inode, 17434e3c51f4SSvyatoslav Feldsherov wb, 17444e3c51f4SSvyatoslav Feldsherov &wb->b_dirty_time); 17454e3c51f4SSvyatoslav Feldsherov } 1746cbfecb92SLukas Czerner } 1747cbfecb92SLukas Czerner } 1748846a3351SJing Xia 17494f8ad655SJan Kara spin_unlock(&wb->list_lock); 17501c0eeaf5SJoern Engel inode_sync_complete(inode); 17514f8ad655SJan Kara out: 17524f8ad655SJan Kara spin_unlock(&inode->i_lock); 17531da177e4SLinus Torvalds return ret; 17541da177e4SLinus Torvalds } 17551da177e4SLinus Torvalds 1756a88a341aSTejun Heo static long writeback_chunk_size(struct bdi_writeback *wb, 17571a12d8bdSWu Fengguang struct wb_writeback_work *work) 1758d46db3d5SWu Fengguang { 1759d46db3d5SWu Fengguang long pages; 1760d46db3d5SWu Fengguang 1761d46db3d5SWu Fengguang /* 1762d46db3d5SWu Fengguang * WB_SYNC_ALL mode does livelock avoidance by syncing dirty 1763d46db3d5SWu Fengguang * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX 1764d46db3d5SWu Fengguang * here avoids calling into writeback_inodes_wb() more than once. 1765d46db3d5SWu Fengguang * 1766d46db3d5SWu Fengguang * The intended call sequence for WB_SYNC_ALL writeback is: 1767d46db3d5SWu Fengguang * 1768d46db3d5SWu Fengguang * wb_writeback() 1769d46db3d5SWu Fengguang * writeback_sb_inodes() <== called only once 1770d46db3d5SWu Fengguang * write_cache_pages() <== called once for each inode 1771d46db3d5SWu Fengguang * (quickly) tag currently dirty pages 1772d46db3d5SWu Fengguang * (maybe slowly) sync all tagged pages 1773d46db3d5SWu Fengguang */ 1774d46db3d5SWu Fengguang if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) 1775d46db3d5SWu Fengguang pages = LONG_MAX; 17761a12d8bdSWu Fengguang else { 1777a88a341aSTejun Heo pages = min(wb->avg_write_bandwidth / 2, 1778dcc25ae7STejun Heo global_wb_domain.dirty_limit / DIRTY_SCOPE); 17791a12d8bdSWu Fengguang pages = min(pages, work->nr_pages); 17801a12d8bdSWu Fengguang pages = round_down(pages + MIN_WRITEBACK_PAGES, 17811a12d8bdSWu Fengguang MIN_WRITEBACK_PAGES); 17821a12d8bdSWu Fengguang } 1783d46db3d5SWu Fengguang 1784d46db3d5SWu Fengguang return pages; 1785d46db3d5SWu Fengguang } 1786d46db3d5SWu Fengguang 178703ba3782SJens Axboe /* 1788f11c9c5cSEdward Shishkin * Write a portion of b_io inodes which belong to @sb. 1789edadfb10SChristoph Hellwig * 1790d46db3d5SWu Fengguang * Return the number of pages and/or inodes written. 17910ba13fd1SLinus Torvalds * 17920ba13fd1SLinus Torvalds * NOTE! This is called with wb->list_lock held, and will 17930ba13fd1SLinus Torvalds * unlock and relock that for each inode it ends up doing 17940ba13fd1SLinus Torvalds * IO for. 1795f11c9c5cSEdward Shishkin */ 1796d46db3d5SWu Fengguang static long writeback_sb_inodes(struct super_block *sb, 1797d46db3d5SWu Fengguang struct bdi_writeback *wb, 1798d46db3d5SWu Fengguang struct wb_writeback_work *work) 179903ba3782SJens Axboe { 1800d46db3d5SWu Fengguang struct writeback_control wbc = { 1801d46db3d5SWu Fengguang .sync_mode = work->sync_mode, 1802d46db3d5SWu Fengguang .tagged_writepages = work->tagged_writepages, 1803d46db3d5SWu Fengguang .for_kupdate = work->for_kupdate, 1804d46db3d5SWu Fengguang .for_background = work->for_background, 18057747bd4bSDave Chinner .for_sync = work->for_sync, 1806d46db3d5SWu Fengguang .range_cyclic = work->range_cyclic, 1807d46db3d5SWu Fengguang .range_start = 0, 1808d46db3d5SWu Fengguang .range_end = LLONG_MAX, 1809d46db3d5SWu Fengguang }; 1810d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1811d46db3d5SWu Fengguang long write_chunk; 181268f4c6ebSZhihao Cheng long total_wrote = 0; /* count both pages and inodes */ 1813d46db3d5SWu Fengguang 181403ba3782SJens Axboe while (!list_empty(&wb->b_io)) { 18157ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1816aaf25593STejun Heo struct bdi_writeback *tmp_wb; 181768f4c6ebSZhihao Cheng long wrote; 1818edadfb10SChristoph Hellwig 1819edadfb10SChristoph Hellwig if (inode->i_sb != sb) { 1820d46db3d5SWu Fengguang if (work->sb) { 1821edadfb10SChristoph Hellwig /* 1822edadfb10SChristoph Hellwig * We only want to write back data for this 1823edadfb10SChristoph Hellwig * superblock, move all inodes not belonging 1824edadfb10SChristoph Hellwig * to it back onto the dirty list. 1825edadfb10SChristoph Hellwig */ 1826f758eeabSChristoph Hellwig redirty_tail(inode, wb); 182766f3b8e2SJens Axboe continue; 182866f3b8e2SJens Axboe } 1829edadfb10SChristoph Hellwig 1830edadfb10SChristoph Hellwig /* 1831edadfb10SChristoph Hellwig * The inode belongs to a different superblock. 1832edadfb10SChristoph Hellwig * Bounce back to the caller to unpin this and 1833edadfb10SChristoph Hellwig * pin the next superblock. 1834edadfb10SChristoph Hellwig */ 1835d46db3d5SWu Fengguang break; 1836edadfb10SChristoph Hellwig } 1837edadfb10SChristoph Hellwig 18389843b76aSChristoph Hellwig /* 1839331cbdeeSWanpeng Li * Don't bother with new inodes or inodes being freed, first 1840331cbdeeSWanpeng Li * kind does not need periodic writeout yet, and for the latter 18419843b76aSChristoph Hellwig * kind writeout is handled by the freer. 18429843b76aSChristoph Hellwig */ 1843250df6edSDave Chinner spin_lock(&inode->i_lock); 18449843b76aSChristoph Hellwig if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { 1845b35250c0SJan Kara redirty_tail_locked(inode, wb); 1846250df6edSDave Chinner spin_unlock(&inode->i_lock); 18477ef0d737SNick Piggin continue; 18487ef0d737SNick Piggin } 1849cc1676d9SJan Kara if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { 1850cc1676d9SJan Kara /* 1851cc1676d9SJan Kara * If this inode is locked for writeback and we are not 1852cc1676d9SJan Kara * doing writeback-for-data-integrity, move it to 1853cc1676d9SJan Kara * b_more_io so that writeback can proceed with the 1854cc1676d9SJan Kara * other inodes on s_io. 1855cc1676d9SJan Kara * 1856cc1676d9SJan Kara * We'll have another go at writing back this inode 1857cc1676d9SJan Kara * when we completed a full scan of b_io. 1858cc1676d9SJan Kara */ 1859cc1676d9SJan Kara requeue_io(inode, wb); 186010e14073SJchao Sun spin_unlock(&inode->i_lock); 1861cc1676d9SJan Kara trace_writeback_sb_inodes_requeue(inode); 1862cc1676d9SJan Kara continue; 1863cc1676d9SJan Kara } 1864f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 1865f0d07b7fSJan Kara 18664f8ad655SJan Kara /* 18674f8ad655SJan Kara * We already requeued the inode if it had I_SYNC set and we 18684f8ad655SJan Kara * are doing WB_SYNC_NONE writeback. So this catches only the 18694f8ad655SJan Kara * WB_SYNC_ALL case. 18704f8ad655SJan Kara */ 1871169ebd90SJan Kara if (inode->i_state & I_SYNC) { 1872169ebd90SJan Kara /* Wait for I_SYNC. This function drops i_lock... */ 1873169ebd90SJan Kara inode_sleep_on_writeback(inode); 1874169ebd90SJan Kara /* Inode may be gone, start again */ 1875ead188f9SJan Kara spin_lock(&wb->list_lock); 1876169ebd90SJan Kara continue; 1877169ebd90SJan Kara } 18784f8ad655SJan Kara inode->i_state |= I_SYNC; 1879b16b1debSTejun Heo wbc_attach_and_unlock_inode(&wbc, inode); 1880169ebd90SJan Kara 1881a88a341aSTejun Heo write_chunk = writeback_chunk_size(wb, work); 1882d46db3d5SWu Fengguang wbc.nr_to_write = write_chunk; 1883d46db3d5SWu Fengguang wbc.pages_skipped = 0; 1884250df6edSDave Chinner 1885169ebd90SJan Kara /* 1886169ebd90SJan Kara * We use I_SYNC to pin the inode in memory. While it is set 1887169ebd90SJan Kara * evict_inode() will wait so the inode cannot be freed. 1888169ebd90SJan Kara */ 1889cd8ed2a4SYan Hong __writeback_single_inode(inode, &wbc); 1890d46db3d5SWu Fengguang 1891b16b1debSTejun Heo wbc_detach_inode(&wbc); 1892d46db3d5SWu Fengguang work->nr_pages -= write_chunk - wbc.nr_to_write; 189368f4c6ebSZhihao Cheng wrote = write_chunk - wbc.nr_to_write - wbc.pages_skipped; 189468f4c6ebSZhihao Cheng wrote = wrote < 0 ? 0 : wrote; 189568f4c6ebSZhihao Cheng total_wrote += wrote; 1896590dca3aSChris Mason 1897590dca3aSChris Mason if (need_resched()) { 1898590dca3aSChris Mason /* 1899590dca3aSChris Mason * We're trying to balance between building up a nice 1900590dca3aSChris Mason * long list of IOs to improve our merge rate, and 1901590dca3aSChris Mason * getting those IOs out quickly for anyone throttling 1902590dca3aSChris Mason * in balance_dirty_pages(). cond_resched() doesn't 1903590dca3aSChris Mason * unplug, so get our IOs out the door before we 1904590dca3aSChris Mason * give up the CPU. 1905590dca3aSChris Mason */ 1906008f75a2SChristoph Hellwig blk_flush_plug(current->plug, false); 1907590dca3aSChris Mason cond_resched(); 1908590dca3aSChris Mason } 1909590dca3aSChris Mason 1910aaf25593STejun Heo /* 1911aaf25593STejun Heo * Requeue @inode if still dirty. Be careful as @inode may 1912aaf25593STejun Heo * have been switched to another wb in the meantime. 1913aaf25593STejun Heo */ 1914aaf25593STejun Heo tmp_wb = inode_to_wb_and_lock_list(inode); 19154f8ad655SJan Kara spin_lock(&inode->i_lock); 19160ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 191768f4c6ebSZhihao Cheng total_wrote++; 1918aaf25593STejun Heo requeue_inode(inode, tmp_wb, &wbc); 19194f8ad655SJan Kara inode_sync_complete(inode); 19200f1b1fd8SDave Chinner spin_unlock(&inode->i_lock); 1921590dca3aSChris Mason 1922aaf25593STejun Heo if (unlikely(tmp_wb != wb)) { 1923aaf25593STejun Heo spin_unlock(&tmp_wb->list_lock); 1924aaf25593STejun Heo spin_lock(&wb->list_lock); 1925aaf25593STejun Heo } 1926aaf25593STejun Heo 1927d46db3d5SWu Fengguang /* 1928d46db3d5SWu Fengguang * bail out to wb_writeback() often enough to check 1929d46db3d5SWu Fengguang * background threshold and other termination conditions. 1930d46db3d5SWu Fengguang */ 193168f4c6ebSZhihao Cheng if (total_wrote) { 1932d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1933d46db3d5SWu Fengguang break; 1934d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1935d46db3d5SWu Fengguang break; 19361da177e4SLinus Torvalds } 19378bc3be27SFengguang Wu } 193868f4c6ebSZhihao Cheng return total_wrote; 1939f11c9c5cSEdward Shishkin } 194038f21977SNick Piggin 1941d46db3d5SWu Fengguang static long __writeback_inodes_wb(struct bdi_writeback *wb, 1942d46db3d5SWu Fengguang struct wb_writeback_work *work) 1943f11c9c5cSEdward Shishkin { 1944d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1945d46db3d5SWu Fengguang long wrote = 0; 1946f11c9c5cSEdward Shishkin 1947f11c9c5cSEdward Shishkin while (!list_empty(&wb->b_io)) { 19487ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1949f11c9c5cSEdward Shishkin struct super_block *sb = inode->i_sb; 1950f11c9c5cSEdward Shishkin 1951eb6ef3dfSKonstantin Khlebnikov if (!trylock_super(sb)) { 19520e995816SWu Fengguang /* 1953eb6ef3dfSKonstantin Khlebnikov * trylock_super() may fail consistently due to 19540e995816SWu Fengguang * s_umount being grabbed by someone else. Don't use 19550e995816SWu Fengguang * requeue_io() to avoid busy retrying the inode/sb. 19560e995816SWu Fengguang */ 19570e995816SWu Fengguang redirty_tail(inode, wb); 1958d19de7edSChristoph Hellwig continue; 1959334132aeSChristoph Hellwig } 1960d46db3d5SWu Fengguang wrote += writeback_sb_inodes(sb, wb, work); 1961eb6ef3dfSKonstantin Khlebnikov up_read(&sb->s_umount); 1962f11c9c5cSEdward Shishkin 1963d46db3d5SWu Fengguang /* refer to the same tests at the end of writeback_sb_inodes */ 1964d46db3d5SWu Fengguang if (wrote) { 1965d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1966d46db3d5SWu Fengguang break; 1967d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1968f11c9c5cSEdward Shishkin break; 1969f11c9c5cSEdward Shishkin } 1970d46db3d5SWu Fengguang } 197166f3b8e2SJens Axboe /* Leave any unwritten inodes on b_io */ 1972d46db3d5SWu Fengguang return wrote; 197366f3b8e2SJens Axboe } 197466f3b8e2SJens Axboe 19757d9f073bSWanpeng Li static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, 19760e175a18SCurt Wohlgemuth enum wb_reason reason) 1977edadfb10SChristoph Hellwig { 1978d46db3d5SWu Fengguang struct wb_writeback_work work = { 1979d46db3d5SWu Fengguang .nr_pages = nr_pages, 1980d46db3d5SWu Fengguang .sync_mode = WB_SYNC_NONE, 1981d46db3d5SWu Fengguang .range_cyclic = 1, 19820e175a18SCurt Wohlgemuth .reason = reason, 1983d46db3d5SWu Fengguang }; 1984505a666eSLinus Torvalds struct blk_plug plug; 1985edadfb10SChristoph Hellwig 1986505a666eSLinus Torvalds blk_start_plug(&plug); 1987f758eeabSChristoph Hellwig spin_lock(&wb->list_lock); 1988424b351fSWu Fengguang if (list_empty(&wb->b_io)) 1989f9cae926SJan Kara queue_io(wb, &work, jiffies); 1990d46db3d5SWu Fengguang __writeback_inodes_wb(wb, &work); 1991f758eeabSChristoph Hellwig spin_unlock(&wb->list_lock); 1992505a666eSLinus Torvalds blk_finish_plug(&plug); 1993edadfb10SChristoph Hellwig 1994d46db3d5SWu Fengguang return nr_pages - work.nr_pages; 199566f3b8e2SJens Axboe } 199666f3b8e2SJens Axboe 199703ba3782SJens Axboe /* 199803ba3782SJens Axboe * Explicit flushing or periodic writeback of "old" data. 199903ba3782SJens Axboe * 200003ba3782SJens Axboe * Define "old": the first time one of an inode's pages is dirtied, we mark the 200103ba3782SJens Axboe * dirtying-time in the inode's address_space. So this periodic writeback code 200203ba3782SJens Axboe * just walks the superblock inode list, writing back any inodes which are 200303ba3782SJens Axboe * older than a specific point in time. 200403ba3782SJens Axboe * 200503ba3782SJens Axboe * Try to run once per dirty_writeback_interval. But if a writeback event 200603ba3782SJens Axboe * takes longer than a dirty_writeback_interval interval, then leave a 200703ba3782SJens Axboe * one-second gap. 200803ba3782SJens Axboe * 2009f9cae926SJan Kara * dirtied_before takes precedence over nr_to_write. So we'll only write back 201003ba3782SJens Axboe * all dirty pages if they are all attached to "old" mappings. 201103ba3782SJens Axboe */ 2012c4a77a6cSJens Axboe static long wb_writeback(struct bdi_writeback *wb, 201383ba7b07SChristoph Hellwig struct wb_writeback_work *work) 201403ba3782SJens Axboe { 2015d46db3d5SWu Fengguang long nr_pages = work->nr_pages; 2016f9cae926SJan Kara unsigned long dirtied_before = jiffies; 2017a5989bdcSJan Kara struct inode *inode; 2018d46db3d5SWu Fengguang long progress; 2019505a666eSLinus Torvalds struct blk_plug plug; 202003ba3782SJens Axboe 2021505a666eSLinus Torvalds blk_start_plug(&plug); 2022e8dfc305SWu Fengguang spin_lock(&wb->list_lock); 202303ba3782SJens Axboe for (;;) { 202403ba3782SJens Axboe /* 2025d3ddec76SWu Fengguang * Stop writeback when nr_pages has been consumed 202603ba3782SJens Axboe */ 202783ba7b07SChristoph Hellwig if (work->nr_pages <= 0) 202803ba3782SJens Axboe break; 202903ba3782SJens Axboe 203003ba3782SJens Axboe /* 2031aa373cf5SJan Kara * Background writeout and kupdate-style writeback may 2032aa373cf5SJan Kara * run forever. Stop them if there is other work to do 2033aa373cf5SJan Kara * so that e.g. sync can proceed. They'll be restarted 2034aa373cf5SJan Kara * after the other works are all done. 2035aa373cf5SJan Kara */ 2036aa373cf5SJan Kara if ((work->for_background || work->for_kupdate) && 2037f0054bb1STejun Heo !list_empty(&wb->work_list)) 2038aa373cf5SJan Kara break; 2039aa373cf5SJan Kara 2040aa373cf5SJan Kara /* 2041d3ddec76SWu Fengguang * For background writeout, stop when we are below the 2042d3ddec76SWu Fengguang * background dirty threshold 204303ba3782SJens Axboe */ 2044aa661bbeSTejun Heo if (work->for_background && !wb_over_bg_thresh(wb)) 204503ba3782SJens Axboe break; 204603ba3782SJens Axboe 20471bc36b64SJan Kara /* 20481bc36b64SJan Kara * Kupdate and background works are special and we want to 20491bc36b64SJan Kara * include all inodes that need writing. Livelock avoidance is 20501bc36b64SJan Kara * handled by these works yielding to any other work so we are 20511bc36b64SJan Kara * safe. 20521bc36b64SJan Kara */ 2053ba9aa839SWu Fengguang if (work->for_kupdate) { 2054f9cae926SJan Kara dirtied_before = jiffies - 2055ba9aa839SWu Fengguang msecs_to_jiffies(dirty_expire_interval * 10); 20561bc36b64SJan Kara } else if (work->for_background) 2057f9cae926SJan Kara dirtied_before = jiffies; 2058028c2dd1SDave Chinner 20595634cc2aSTejun Heo trace_writeback_start(wb, work); 2060e8dfc305SWu Fengguang if (list_empty(&wb->b_io)) 2061f9cae926SJan Kara queue_io(wb, work, dirtied_before); 206283ba7b07SChristoph Hellwig if (work->sb) 2063d46db3d5SWu Fengguang progress = writeback_sb_inodes(work->sb, wb, work); 2064edadfb10SChristoph Hellwig else 2065d46db3d5SWu Fengguang progress = __writeback_inodes_wb(wb, work); 20665634cc2aSTejun Heo trace_writeback_written(wb, work); 2067028c2dd1SDave Chinner 206803ba3782SJens Axboe /* 206971fd05a8SJens Axboe * Did we write something? Try for more 2070e6fb6da2SWu Fengguang * 2071e6fb6da2SWu Fengguang * Dirty inodes are moved to b_io for writeback in batches. 2072e6fb6da2SWu Fengguang * The completion of the current batch does not necessarily 2073e6fb6da2SWu Fengguang * mean the overall work is done. So we keep looping as long 2074e6fb6da2SWu Fengguang * as made some progress on cleaning pages or inodes. 207571fd05a8SJens Axboe */ 2076d46db3d5SWu Fengguang if (progress) 207703ba3782SJens Axboe continue; 2078a5989bdcSJan Kara /* 2079e6fb6da2SWu Fengguang * No more inodes for IO, bail 2080a5989bdcSJan Kara */ 2081b7a2441fSWu Fengguang if (list_empty(&wb->b_more_io)) 208203ba3782SJens Axboe break; 208303ba3782SJens Axboe /* 20848010c3b6SJens Axboe * Nothing written. Wait for some inode to 20858010c3b6SJens Axboe * become available for writeback. Otherwise 20868010c3b6SJens Axboe * we'll just busyloop. 208703ba3782SJens Axboe */ 20885634cc2aSTejun Heo trace_writeback_wait(wb, work); 208903ba3782SJens Axboe inode = wb_inode(wb->b_more_io.prev); 2090250df6edSDave Chinner spin_lock(&inode->i_lock); 2091f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 2092169ebd90SJan Kara /* This function drops i_lock... */ 2093169ebd90SJan Kara inode_sleep_on_writeback(inode); 2094f0d07b7fSJan Kara spin_lock(&wb->list_lock); 209503ba3782SJens Axboe } 2096e8dfc305SWu Fengguang spin_unlock(&wb->list_lock); 2097505a666eSLinus Torvalds blk_finish_plug(&plug); 209803ba3782SJens Axboe 2099d46db3d5SWu Fengguang return nr_pages - work->nr_pages; 210003ba3782SJens Axboe } 210103ba3782SJens Axboe 210203ba3782SJens Axboe /* 210383ba7b07SChristoph Hellwig * Return the next wb_writeback_work struct that hasn't been processed yet. 210403ba3782SJens Axboe */ 2105f0054bb1STejun Heo static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) 210603ba3782SJens Axboe { 210783ba7b07SChristoph Hellwig struct wb_writeback_work *work = NULL; 210803ba3782SJens Axboe 2109f87904c0SKhazhismel Kumykov spin_lock_irq(&wb->work_lock); 2110f0054bb1STejun Heo if (!list_empty(&wb->work_list)) { 2111f0054bb1STejun Heo work = list_entry(wb->work_list.next, 211283ba7b07SChristoph Hellwig struct wb_writeback_work, list); 211383ba7b07SChristoph Hellwig list_del_init(&work->list); 211403ba3782SJens Axboe } 2115f87904c0SKhazhismel Kumykov spin_unlock_irq(&wb->work_lock); 211683ba7b07SChristoph Hellwig return work; 211703ba3782SJens Axboe } 211803ba3782SJens Axboe 21196585027aSJan Kara static long wb_check_background_flush(struct bdi_writeback *wb) 21206585027aSJan Kara { 2121aa661bbeSTejun Heo if (wb_over_bg_thresh(wb)) { 21226585027aSJan Kara 21236585027aSJan Kara struct wb_writeback_work work = { 21246585027aSJan Kara .nr_pages = LONG_MAX, 21256585027aSJan Kara .sync_mode = WB_SYNC_NONE, 21266585027aSJan Kara .for_background = 1, 21276585027aSJan Kara .range_cyclic = 1, 21280e175a18SCurt Wohlgemuth .reason = WB_REASON_BACKGROUND, 21296585027aSJan Kara }; 21306585027aSJan Kara 21316585027aSJan Kara return wb_writeback(wb, &work); 21326585027aSJan Kara } 21336585027aSJan Kara 21346585027aSJan Kara return 0; 21356585027aSJan Kara } 21366585027aSJan Kara 213703ba3782SJens Axboe static long wb_check_old_data_flush(struct bdi_writeback *wb) 213803ba3782SJens Axboe { 213903ba3782SJens Axboe unsigned long expired; 214003ba3782SJens Axboe long nr_pages; 214103ba3782SJens Axboe 214269b62d01SJens Axboe /* 214369b62d01SJens Axboe * When set to zero, disable periodic writeback 214469b62d01SJens Axboe */ 214569b62d01SJens Axboe if (!dirty_writeback_interval) 214669b62d01SJens Axboe return 0; 214769b62d01SJens Axboe 214803ba3782SJens Axboe expired = wb->last_old_flush + 214903ba3782SJens Axboe msecs_to_jiffies(dirty_writeback_interval * 10); 215003ba3782SJens Axboe if (time_before(jiffies, expired)) 215103ba3782SJens Axboe return 0; 215203ba3782SJens Axboe 215303ba3782SJens Axboe wb->last_old_flush = jiffies; 2154cdf01dd5SLinus Torvalds nr_pages = get_nr_dirty_pages(); 215503ba3782SJens Axboe 2156c4a77a6cSJens Axboe if (nr_pages) { 215783ba7b07SChristoph Hellwig struct wb_writeback_work work = { 2158c4a77a6cSJens Axboe .nr_pages = nr_pages, 2159c4a77a6cSJens Axboe .sync_mode = WB_SYNC_NONE, 2160c4a77a6cSJens Axboe .for_kupdate = 1, 2161c4a77a6cSJens Axboe .range_cyclic = 1, 21620e175a18SCurt Wohlgemuth .reason = WB_REASON_PERIODIC, 2163c4a77a6cSJens Axboe }; 2164c4a77a6cSJens Axboe 216583ba7b07SChristoph Hellwig return wb_writeback(wb, &work); 2166c4a77a6cSJens Axboe } 216703ba3782SJens Axboe 216803ba3782SJens Axboe return 0; 216903ba3782SJens Axboe } 217003ba3782SJens Axboe 217185009b4fSJens Axboe static long wb_check_start_all(struct bdi_writeback *wb) 217285009b4fSJens Axboe { 217385009b4fSJens Axboe long nr_pages; 217485009b4fSJens Axboe 217585009b4fSJens Axboe if (!test_bit(WB_start_all, &wb->state)) 217685009b4fSJens Axboe return 0; 217785009b4fSJens Axboe 217885009b4fSJens Axboe nr_pages = get_nr_dirty_pages(); 217985009b4fSJens Axboe if (nr_pages) { 218085009b4fSJens Axboe struct wb_writeback_work work = { 218185009b4fSJens Axboe .nr_pages = wb_split_bdi_pages(wb, nr_pages), 218285009b4fSJens Axboe .sync_mode = WB_SYNC_NONE, 218385009b4fSJens Axboe .range_cyclic = 1, 218485009b4fSJens Axboe .reason = wb->start_all_reason, 218585009b4fSJens Axboe }; 218685009b4fSJens Axboe 218785009b4fSJens Axboe nr_pages = wb_writeback(wb, &work); 218885009b4fSJens Axboe } 218985009b4fSJens Axboe 219085009b4fSJens Axboe clear_bit(WB_start_all, &wb->state); 219185009b4fSJens Axboe return nr_pages; 219285009b4fSJens Axboe } 219385009b4fSJens Axboe 219485009b4fSJens Axboe 219503ba3782SJens Axboe /* 219603ba3782SJens Axboe * Retrieve work items and do the writeback they describe 219703ba3782SJens Axboe */ 219825d130baSWanpeng Li static long wb_do_writeback(struct bdi_writeback *wb) 219903ba3782SJens Axboe { 220083ba7b07SChristoph Hellwig struct wb_writeback_work *work; 2201c4a77a6cSJens Axboe long wrote = 0; 220203ba3782SJens Axboe 22034452226eSTejun Heo set_bit(WB_writeback_running, &wb->state); 2204f0054bb1STejun Heo while ((work = get_next_work_item(wb)) != NULL) { 22055634cc2aSTejun Heo trace_writeback_exec(wb, work); 220683ba7b07SChristoph Hellwig wrote += wb_writeback(wb, work); 22074a3a485bSTahsin Erdogan finish_writeback_work(wb, work); 220803ba3782SJens Axboe } 220903ba3782SJens Axboe 221003ba3782SJens Axboe /* 221185009b4fSJens Axboe * Check for a flush-everything request 221285009b4fSJens Axboe */ 221385009b4fSJens Axboe wrote += wb_check_start_all(wb); 221485009b4fSJens Axboe 221585009b4fSJens Axboe /* 221603ba3782SJens Axboe * Check for periodic writeback, kupdated() style 221703ba3782SJens Axboe */ 221803ba3782SJens Axboe wrote += wb_check_old_data_flush(wb); 22196585027aSJan Kara wrote += wb_check_background_flush(wb); 22204452226eSTejun Heo clear_bit(WB_writeback_running, &wb->state); 222103ba3782SJens Axboe 222203ba3782SJens Axboe return wrote; 222303ba3782SJens Axboe } 222403ba3782SJens Axboe 222503ba3782SJens Axboe /* 222603ba3782SJens Axboe * Handle writeback of dirty data for the device backed by this bdi. Also 2227839a8e86STejun Heo * reschedules periodically and does kupdated style flushing. 222803ba3782SJens Axboe */ 2229f0054bb1STejun Heo void wb_workfn(struct work_struct *work) 223003ba3782SJens Axboe { 2231839a8e86STejun Heo struct bdi_writeback *wb = container_of(to_delayed_work(work), 2232839a8e86STejun Heo struct bdi_writeback, dwork); 223303ba3782SJens Axboe long pages_written; 223403ba3782SJens Axboe 223568f23b89STheodore Ts'o set_worker_desc("flush-%s", bdi_dev_name(wb->bdi)); 223603ba3782SJens Axboe 2237839a8e86STejun Heo if (likely(!current_is_workqueue_rescuer() || 22384452226eSTejun Heo !test_bit(WB_registered, &wb->state))) { 223903ba3782SJens Axboe /* 2240f0054bb1STejun Heo * The normal path. Keep writing back @wb until its 2241839a8e86STejun Heo * work_list is empty. Note that this path is also taken 2242f0054bb1STejun Heo * if @wb is shutting down even when we're running off the 2243839a8e86STejun Heo * rescuer as work_list needs to be drained. 224403ba3782SJens Axboe */ 2245839a8e86STejun Heo do { 224625d130baSWanpeng Li pages_written = wb_do_writeback(wb); 2247455b2864SDave Chinner trace_writeback_pages_written(pages_written); 2248f0054bb1STejun Heo } while (!list_empty(&wb->work_list)); 2249839a8e86STejun Heo } else { 2250253c34e9SArtem Bityutskiy /* 2251839a8e86STejun Heo * bdi_wq can't get enough workers and we're running off 2252839a8e86STejun Heo * the emergency worker. Don't hog it. Hopefully, 1024 is 2253839a8e86STejun Heo * enough for efficient IO. 2254253c34e9SArtem Bityutskiy */ 2255f0054bb1STejun Heo pages_written = writeback_inodes_wb(wb, 1024, 2256839a8e86STejun Heo WB_REASON_FORKER_THREAD); 2257839a8e86STejun Heo trace_writeback_pages_written(pages_written); 225803ba3782SJens Axboe } 225903ba3782SJens Axboe 2260f0054bb1STejun Heo if (!list_empty(&wb->work_list)) 2261b8b78495SJan Kara wb_wakeup(wb); 22626ca738d6SDerek Basehore else if (wb_has_dirty_io(wb) && dirty_writeback_interval) 2263f0054bb1STejun Heo wb_wakeup_delayed(wb); 226403ba3782SJens Axboe } 226503ba3782SJens Axboe 226603ba3782SJens Axboe /* 2267595043e5SJens Axboe * Start writeback of `nr_pages' pages on this bdi. If `nr_pages' is zero, 2268595043e5SJens Axboe * write back the whole world. 2269595043e5SJens Axboe */ 2270595043e5SJens Axboe static void __wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2271e8e8a0c6SJens Axboe enum wb_reason reason) 2272595043e5SJens Axboe { 2273595043e5SJens Axboe struct bdi_writeback *wb; 2274595043e5SJens Axboe 2275595043e5SJens Axboe if (!bdi_has_dirty_io(bdi)) 2276595043e5SJens Axboe return; 2277595043e5SJens Axboe 2278595043e5SJens Axboe list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 2279e8e8a0c6SJens Axboe wb_start_writeback(wb, reason); 2280595043e5SJens Axboe } 2281595043e5SJens Axboe 2282595043e5SJens Axboe void wakeup_flusher_threads_bdi(struct backing_dev_info *bdi, 2283595043e5SJens Axboe enum wb_reason reason) 2284595043e5SJens Axboe { 2285595043e5SJens Axboe rcu_read_lock(); 2286e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2287595043e5SJens Axboe rcu_read_unlock(); 2288595043e5SJens Axboe } 2289595043e5SJens Axboe 2290595043e5SJens Axboe /* 22919ba4b2dfSJens Axboe * Wakeup the flusher threads to start writeback of all currently dirty pages 229203ba3782SJens Axboe */ 22939ba4b2dfSJens Axboe void wakeup_flusher_threads(enum wb_reason reason) 229403ba3782SJens Axboe { 2295b8c2f347SChristoph Hellwig struct backing_dev_info *bdi; 2296b8c2f347SChristoph Hellwig 229751350ea0SKonstantin Khlebnikov /* 229851350ea0SKonstantin Khlebnikov * If we are expecting writeback progress we must submit plugged IO. 229951350ea0SKonstantin Khlebnikov */ 2300008f75a2SChristoph Hellwig blk_flush_plug(current->plug, true); 230151350ea0SKonstantin Khlebnikov 2302b8c2f347SChristoph Hellwig rcu_read_lock(); 2303595043e5SJens Axboe list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) 2304e8e8a0c6SJens Axboe __wakeup_flusher_threads_bdi(bdi, reason); 2305b8c2f347SChristoph Hellwig rcu_read_unlock(); 230603ba3782SJens Axboe } 230703ba3782SJens Axboe 2308a2f48706STheodore Ts'o /* 2309a2f48706STheodore Ts'o * Wake up bdi's periodically to make sure dirtytime inodes gets 2310a2f48706STheodore Ts'o * written back periodically. We deliberately do *not* check the 2311a2f48706STheodore Ts'o * b_dirtytime list in wb_has_dirty_io(), since this would cause the 2312a2f48706STheodore Ts'o * kernel to be constantly waking up once there are any dirtytime 2313a2f48706STheodore Ts'o * inodes on the system. So instead we define a separate delayed work 2314a2f48706STheodore Ts'o * function which gets called much more rarely. (By default, only 2315a2f48706STheodore Ts'o * once every 12 hours.) 2316a2f48706STheodore Ts'o * 2317a2f48706STheodore Ts'o * If there is any other write activity going on in the file system, 2318a2f48706STheodore Ts'o * this function won't be necessary. But if the only thing that has 2319a2f48706STheodore Ts'o * happened on the file system is a dirtytime inode caused by an atime 2320a2f48706STheodore Ts'o * update, we need this infrastructure below to make sure that inode 2321a2f48706STheodore Ts'o * eventually gets pushed out to disk. 2322a2f48706STheodore Ts'o */ 2323a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w); 2324a2f48706STheodore Ts'o static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); 2325a2f48706STheodore Ts'o 2326a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w) 2327a2f48706STheodore Ts'o { 2328a2f48706STheodore Ts'o struct backing_dev_info *bdi; 2329a2f48706STheodore Ts'o 2330a2f48706STheodore Ts'o rcu_read_lock(); 2331a2f48706STheodore Ts'o list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 2332001fe6f6STejun Heo struct bdi_writeback *wb; 2333001fe6f6STejun Heo 2334b817525aSTejun Heo list_for_each_entry_rcu(wb, &bdi->wb_list, bdi_node) 23356fdf860fSTejun Heo if (!list_empty(&wb->b_dirty_time)) 23366fdf860fSTejun Heo wb_wakeup(wb); 2337a2f48706STheodore Ts'o } 2338a2f48706STheodore Ts'o rcu_read_unlock(); 2339a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2340a2f48706STheodore Ts'o } 2341a2f48706STheodore Ts'o 2342a2f48706STheodore Ts'o static int __init start_dirtytime_writeback(void) 2343a2f48706STheodore Ts'o { 2344a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 2345a2f48706STheodore Ts'o return 0; 2346a2f48706STheodore Ts'o } 2347a2f48706STheodore Ts'o __initcall(start_dirtytime_writeback); 2348a2f48706STheodore Ts'o 23491efff914STheodore Ts'o int dirtytime_interval_handler(struct ctl_table *table, int write, 23509ca48e20STobias Klauser void *buffer, size_t *lenp, loff_t *ppos) 23511efff914STheodore Ts'o { 23521efff914STheodore Ts'o int ret; 23531efff914STheodore Ts'o 23541efff914STheodore Ts'o ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 23551efff914STheodore Ts'o if (ret == 0 && write) 23561efff914STheodore Ts'o mod_delayed_work(system_wq, &dirtytime_work, 0); 23571efff914STheodore Ts'o return ret; 23581efff914STheodore Ts'o } 23591efff914STheodore Ts'o 236003ba3782SJens Axboe /** 236135d14f27SEric Biggers * __mark_inode_dirty - internal function to mark an inode dirty 23620117d427SMauro Carvalho Chehab * 236303ba3782SJens Axboe * @inode: inode to mark 236435d14f27SEric Biggers * @flags: what kind of dirty, e.g. I_DIRTY_SYNC. This can be a combination of 236535d14f27SEric Biggers * multiple I_DIRTY_* flags, except that I_DIRTY_TIME can't be combined 236635d14f27SEric Biggers * with I_DIRTY_PAGES. 23670117d427SMauro Carvalho Chehab * 236835d14f27SEric Biggers * Mark an inode as dirty. We notify the filesystem, then update the inode's 236935d14f27SEric Biggers * dirty flags. Then, if needed we add the inode to the appropriate dirty list. 237003ba3782SJens Axboe * 237135d14f27SEric Biggers * Most callers should use mark_inode_dirty() or mark_inode_dirty_sync() 237235d14f27SEric Biggers * instead of calling this directly. 237303ba3782SJens Axboe * 237435d14f27SEric Biggers * CAREFUL! We only add the inode to the dirty list if it is hashed or if it 237535d14f27SEric Biggers * refers to a blockdev. Unhashed inodes will never be added to the dirty list 237635d14f27SEric Biggers * even if they are later hashed, as they will have been marked dirty already. 237703ba3782SJens Axboe * 237835d14f27SEric Biggers * In short, ensure you hash any inodes _before_ you start marking them dirty. 237903ba3782SJens Axboe * 238003ba3782SJens Axboe * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 238103ba3782SJens Axboe * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 238203ba3782SJens Axboe * the kernel-internal blockdev inode represents the dirtying time of the 238303ba3782SJens Axboe * blockdev's pages. This is why for I_DIRTY_PAGES we always use 238403ba3782SJens Axboe * page->mapping->host, so the page-dirtying time is recorded in the internal 238503ba3782SJens Axboe * blockdev inode. 238603ba3782SJens Axboe */ 238703ba3782SJens Axboe void __mark_inode_dirty(struct inode *inode, int flags) 238803ba3782SJens Axboe { 238903ba3782SJens Axboe struct super_block *sb = inode->i_sb; 239035d14f27SEric Biggers int dirtytime = 0; 239110e14073SJchao Sun struct bdi_writeback *wb = NULL; 23920ae45f63STheodore Ts'o 23930ae45f63STheodore Ts'o trace_writeback_mark_inode_dirty(inode, flags); 239403ba3782SJens Axboe 2395e2728c56SEric Biggers if (flags & I_DIRTY_INODE) { 239635d14f27SEric Biggers /* 2397cbfecb92SLukas Czerner * Inode timestamp update will piggback on this dirtying. 2398cbfecb92SLukas Czerner * We tell ->dirty_inode callback that timestamps need to 2399cbfecb92SLukas Czerner * be updated by setting I_DIRTY_TIME in flags. 2400cbfecb92SLukas Czerner */ 2401cbfecb92SLukas Czerner if (inode->i_state & I_DIRTY_TIME) { 2402cbfecb92SLukas Czerner spin_lock(&inode->i_lock); 2403cbfecb92SLukas Czerner if (inode->i_state & I_DIRTY_TIME) { 2404cbfecb92SLukas Czerner inode->i_state &= ~I_DIRTY_TIME; 2405cbfecb92SLukas Czerner flags |= I_DIRTY_TIME; 2406cbfecb92SLukas Czerner } 2407cbfecb92SLukas Czerner spin_unlock(&inode->i_lock); 2408cbfecb92SLukas Czerner } 2409cbfecb92SLukas Czerner 2410cbfecb92SLukas Czerner /* 241135d14f27SEric Biggers * Notify the filesystem about the inode being dirtied, so that 241235d14f27SEric Biggers * (if needed) it can update on-disk fields and journal the 241335d14f27SEric Biggers * inode. This is only needed when the inode itself is being 241435d14f27SEric Biggers * dirtied now. I.e. it's only needed for I_DIRTY_INODE, not 241535d14f27SEric Biggers * for just I_DIRTY_PAGES or I_DIRTY_TIME. 241635d14f27SEric Biggers */ 24179fb0a7daSTejun Heo trace_writeback_dirty_inode_start(inode, flags); 241803ba3782SJens Axboe if (sb->s_op->dirty_inode) 2419cbfecb92SLukas Czerner sb->s_op->dirty_inode(inode, 2420cbfecb92SLukas Czerner flags & (I_DIRTY_INODE | I_DIRTY_TIME)); 24219fb0a7daSTejun Heo trace_writeback_dirty_inode(inode, flags); 2422e2728c56SEric Biggers 242335d14f27SEric Biggers /* I_DIRTY_INODE supersedes I_DIRTY_TIME. */ 24240ae45f63STheodore Ts'o flags &= ~I_DIRTY_TIME; 242535d14f27SEric Biggers } else { 242635d14f27SEric Biggers /* 242735d14f27SEric Biggers * Else it's either I_DIRTY_PAGES, I_DIRTY_TIME, or nothing. 242835d14f27SEric Biggers * (We don't support setting both I_DIRTY_PAGES and I_DIRTY_TIME 242935d14f27SEric Biggers * in one call to __mark_inode_dirty().) 243035d14f27SEric Biggers */ 24310ae45f63STheodore Ts'o dirtytime = flags & I_DIRTY_TIME; 243235d14f27SEric Biggers WARN_ON_ONCE(dirtytime && flags != I_DIRTY_TIME); 243335d14f27SEric Biggers } 243403ba3782SJens Axboe 243503ba3782SJens Axboe /* 24369c6ac78eSTejun Heo * Paired with smp_mb() in __writeback_single_inode() for the 24379c6ac78eSTejun Heo * following lockless i_state test. See there for details. 243803ba3782SJens Axboe */ 243903ba3782SJens Axboe smp_mb(); 244003ba3782SJens Axboe 2441cbfecb92SLukas Czerner if ((inode->i_state & flags) == flags) 244203ba3782SJens Axboe return; 244303ba3782SJens Axboe 2444250df6edSDave Chinner spin_lock(&inode->i_lock); 244503ba3782SJens Axboe if ((inode->i_state & flags) != flags) { 244603ba3782SJens Axboe const int was_dirty = inode->i_state & I_DIRTY; 244703ba3782SJens Axboe 244852ebea74STejun Heo inode_attach_wb(inode, NULL); 244952ebea74STejun Heo 245003ba3782SJens Axboe inode->i_state |= flags; 245103ba3782SJens Axboe 245203ba3782SJens Axboe /* 245310e14073SJchao Sun * Grab inode's wb early because it requires dropping i_lock and we 245410e14073SJchao Sun * need to make sure following checks happen atomically with dirty 245510e14073SJchao Sun * list handling so that we don't move inodes under flush worker's 245610e14073SJchao Sun * hands. 245710e14073SJchao Sun */ 245810e14073SJchao Sun if (!was_dirty) { 245910e14073SJchao Sun wb = locked_inode_to_wb_and_lock_list(inode); 246010e14073SJchao Sun spin_lock(&inode->i_lock); 246110e14073SJchao Sun } 246210e14073SJchao Sun 246310e14073SJchao Sun /* 24645afced3bSJan Kara * If the inode is queued for writeback by flush worker, just 24655afced3bSJan Kara * update its dirty state. Once the flush worker is done with 24665afced3bSJan Kara * the inode it will place it on the appropriate superblock 24675afced3bSJan Kara * list, based upon its state. 246803ba3782SJens Axboe */ 24695afced3bSJan Kara if (inode->i_state & I_SYNC_QUEUED) 247010e14073SJchao Sun goto out_unlock; 247103ba3782SJens Axboe 247203ba3782SJens Axboe /* 247303ba3782SJens Axboe * Only add valid (hashed) inodes to the superblock's 247403ba3782SJens Axboe * dirty list. Add blockdev inodes as well. 247503ba3782SJens Axboe */ 247603ba3782SJens Axboe if (!S_ISBLK(inode->i_mode)) { 24771d3382cbSAl Viro if (inode_unhashed(inode)) 247810e14073SJchao Sun goto out_unlock; 247903ba3782SJens Axboe } 2480a4ffdde6SAl Viro if (inode->i_state & I_FREEING) 248110e14073SJchao Sun goto out_unlock; 248203ba3782SJens Axboe 248303ba3782SJens Axboe /* 248403ba3782SJens Axboe * If the inode was already on b_dirty/b_io/b_more_io, don't 248503ba3782SJens Axboe * reposition it (that would break b_dirty time-ordering). 248603ba3782SJens Axboe */ 248703ba3782SJens Axboe if (!was_dirty) { 2488d6c10f1fSTejun Heo struct list_head *dirty_list; 2489a66979abSDave Chinner bool wakeup_bdi = false; 2490500b067cSJens Axboe 249103ba3782SJens Axboe inode->dirtied_when = jiffies; 2492a2f48706STheodore Ts'o if (dirtytime) 2493a2f48706STheodore Ts'o inode->dirtied_time_when = jiffies; 2494d6c10f1fSTejun Heo 24950e11f644SChristoph Hellwig if (inode->i_state & I_DIRTY) 24960747259dSTejun Heo dirty_list = &wb->b_dirty; 2497a2f48706STheodore Ts'o else 24980747259dSTejun Heo dirty_list = &wb->b_dirty_time; 2499d6c10f1fSTejun Heo 2500c7f54084SDave Chinner wakeup_bdi = inode_io_list_move_locked(inode, wb, 2501d6c10f1fSTejun Heo dirty_list); 2502d6c10f1fSTejun Heo 25030747259dSTejun Heo spin_unlock(&wb->list_lock); 250410e14073SJchao Sun spin_unlock(&inode->i_lock); 25050ae45f63STheodore Ts'o trace_writeback_dirty_inode_enqueue(inode); 2506253c34e9SArtem Bityutskiy 2507d6c10f1fSTejun Heo /* 2508d6c10f1fSTejun Heo * If this is the first dirty inode for this bdi, 2509d6c10f1fSTejun Heo * we have to wake-up the corresponding bdi thread 2510d6c10f1fSTejun Heo * to make sure background write-back happens 2511d6c10f1fSTejun Heo * later. 2512d6c10f1fSTejun Heo */ 2513f56753acSChristoph Hellwig if (wakeup_bdi && 2514f56753acSChristoph Hellwig (wb->bdi->capabilities & BDI_CAP_WRITEBACK)) 25150747259dSTejun Heo wb_wakeup_delayed(wb); 2516a66979abSDave Chinner return; 2517a66979abSDave Chinner } 2518a66979abSDave Chinner } 251910e14073SJchao Sun out_unlock: 252010e14073SJchao Sun if (wb) 252110e14073SJchao Sun spin_unlock(&wb->list_lock); 2522a66979abSDave Chinner spin_unlock(&inode->i_lock); 252303ba3782SJens Axboe } 252403ba3782SJens Axboe EXPORT_SYMBOL(__mark_inode_dirty); 252503ba3782SJens Axboe 2526e97fedb9SDave Chinner /* 2527e97fedb9SDave Chinner * The @s_sync_lock is used to serialise concurrent sync operations 2528e97fedb9SDave Chinner * to avoid lock contention problems with concurrent wait_sb_inodes() calls. 2529e97fedb9SDave Chinner * Concurrent callers will block on the s_sync_lock rather than doing contending 2530e97fedb9SDave Chinner * walks. The queueing maintains sync(2) required behaviour as all the IO that 2531e97fedb9SDave Chinner * has been issued up to the time this function is enter is guaranteed to be 2532e97fedb9SDave Chinner * completed by the time we have gained the lock and waited for all IO that is 2533e97fedb9SDave Chinner * in progress regardless of the order callers are granted the lock. 2534e97fedb9SDave Chinner */ 2535b6e51316SJens Axboe static void wait_sb_inodes(struct super_block *sb) 253666f3b8e2SJens Axboe { 25376c60d2b5SDave Chinner LIST_HEAD(sync_list); 253838f21977SNick Piggin 253903ba3782SJens Axboe /* 254003ba3782SJens Axboe * We need to be protected against the filesystem going from 254103ba3782SJens Axboe * r/o to r/w or vice versa. 254203ba3782SJens Axboe */ 2543b6e51316SJens Axboe WARN_ON(!rwsem_is_locked(&sb->s_umount)); 254403ba3782SJens Axboe 2545e97fedb9SDave Chinner mutex_lock(&sb->s_sync_lock); 254666f3b8e2SJens Axboe 254738f21977SNick Piggin /* 25486c60d2b5SDave Chinner * Splice the writeback list onto a temporary list to avoid waiting on 25496c60d2b5SDave Chinner * inodes that have started writeback after this point. 25506c60d2b5SDave Chinner * 25516c60d2b5SDave Chinner * Use rcu_read_lock() to keep the inodes around until we have a 25526c60d2b5SDave Chinner * reference. s_inode_wblist_lock protects sb->s_inodes_wb as well as 25536c60d2b5SDave Chinner * the local list because inodes can be dropped from either by writeback 25546c60d2b5SDave Chinner * completion. 255538f21977SNick Piggin */ 25566c60d2b5SDave Chinner rcu_read_lock(); 25576c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 25586c60d2b5SDave Chinner list_splice_init(&sb->s_inodes_wb, &sync_list); 25596c60d2b5SDave Chinner 25606c60d2b5SDave Chinner /* 25616c60d2b5SDave Chinner * Data integrity sync. Must wait for all pages under writeback, because 25626c60d2b5SDave Chinner * there may have been pages dirtied before our sync call, but which had 25636c60d2b5SDave Chinner * writeout started before we write it out. In which case, the inode 25646c60d2b5SDave Chinner * may not be on the dirty list, but we still have to wait for that 25656c60d2b5SDave Chinner * writeout. 25666c60d2b5SDave Chinner */ 25676c60d2b5SDave Chinner while (!list_empty(&sync_list)) { 25686c60d2b5SDave Chinner struct inode *inode = list_first_entry(&sync_list, struct inode, 25696c60d2b5SDave Chinner i_wb_list); 2570250df6edSDave Chinner struct address_space *mapping = inode->i_mapping; 257138f21977SNick Piggin 25726c60d2b5SDave Chinner /* 25736c60d2b5SDave Chinner * Move each inode back to the wb list before we drop the lock 25746c60d2b5SDave Chinner * to preserve consistency between i_wb_list and the mapping 25756c60d2b5SDave Chinner * writeback tag. Writeback completion is responsible to remove 25766c60d2b5SDave Chinner * the inode from either list once the writeback tag is cleared. 25776c60d2b5SDave Chinner */ 25786c60d2b5SDave Chinner list_move_tail(&inode->i_wb_list, &sb->s_inodes_wb); 25796c60d2b5SDave Chinner 25806c60d2b5SDave Chinner /* 25816c60d2b5SDave Chinner * The mapping can appear untagged while still on-list since we 25826c60d2b5SDave Chinner * do not have the mapping lock. Skip it here, wb completion 25836c60d2b5SDave Chinner * will remove it. 25846c60d2b5SDave Chinner */ 25856c60d2b5SDave Chinner if (!mapping_tagged(mapping, PAGECACHE_TAG_WRITEBACK)) 25866c60d2b5SDave Chinner continue; 25876c60d2b5SDave Chinner 25886c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 25896c60d2b5SDave Chinner 2590250df6edSDave Chinner spin_lock(&inode->i_lock); 25916c60d2b5SDave Chinner if (inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) { 2592250df6edSDave Chinner spin_unlock(&inode->i_lock); 25936c60d2b5SDave Chinner 25946c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 259538f21977SNick Piggin continue; 2596250df6edSDave Chinner } 259738f21977SNick Piggin __iget(inode); 2598250df6edSDave Chinner spin_unlock(&inode->i_lock); 25996c60d2b5SDave Chinner rcu_read_unlock(); 260038f21977SNick Piggin 2601aa750fd7SJunichi Nomura /* 2602aa750fd7SJunichi Nomura * We keep the error status of individual mapping so that 2603aa750fd7SJunichi Nomura * applications can catch the writeback error using fsync(2). 2604aa750fd7SJunichi Nomura * See filemap_fdatawait_keep_errors() for details. 2605aa750fd7SJunichi Nomura */ 2606aa750fd7SJunichi Nomura filemap_fdatawait_keep_errors(mapping); 260738f21977SNick Piggin 260838f21977SNick Piggin cond_resched(); 260938f21977SNick Piggin 26106c60d2b5SDave Chinner iput(inode); 26116c60d2b5SDave Chinner 26126c60d2b5SDave Chinner rcu_read_lock(); 26136c60d2b5SDave Chinner spin_lock_irq(&sb->s_inode_wblist_lock); 261438f21977SNick Piggin } 26156c60d2b5SDave Chinner spin_unlock_irq(&sb->s_inode_wblist_lock); 26166c60d2b5SDave Chinner rcu_read_unlock(); 2617e97fedb9SDave Chinner mutex_unlock(&sb->s_sync_lock); 261866f3b8e2SJens Axboe } 26191da177e4SLinus Torvalds 2620f30a7d0cSTejun Heo static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2621f30a7d0cSTejun Heo enum wb_reason reason, bool skip_if_busy) 26221da177e4SLinus Torvalds { 26235b9cce4cSTejun Heo struct backing_dev_info *bdi = sb->s_bdi; 26245b9cce4cSTejun Heo DEFINE_WB_COMPLETION(done, bdi); 262583ba7b07SChristoph Hellwig struct wb_writeback_work work = { 26263c4d7165SChristoph Hellwig .sb = sb, 26273c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_NONE, 26286e6938b6SWu Fengguang .tagged_writepages = 1, 262983ba7b07SChristoph Hellwig .done = &done, 26303259f8beSChris Mason .nr_pages = nr, 26310e175a18SCurt Wohlgemuth .reason = reason, 26323c4d7165SChristoph Hellwig }; 26330e3c9a22SJens Axboe 2634e7972912STejun Heo if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 26356eedc701SJan Kara return; 2636cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2637f30a7d0cSTejun Heo 2638db125360STejun Heo bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); 26395b9cce4cSTejun Heo wb_wait_for_completion(&done); 26401da177e4SLinus Torvalds } 2641f30a7d0cSTejun Heo 2642f30a7d0cSTejun Heo /** 2643f30a7d0cSTejun Heo * writeback_inodes_sb_nr - writeback dirty inodes from given super_block 2644f30a7d0cSTejun Heo * @sb: the superblock 2645f30a7d0cSTejun Heo * @nr: the number of pages to write 2646f30a7d0cSTejun Heo * @reason: reason why some writeback work initiated 2647f30a7d0cSTejun Heo * 2648f30a7d0cSTejun Heo * Start writeback on some inodes on this super_block. No guarantees are made 2649f30a7d0cSTejun Heo * on how many (if any) will be written, and this function does not wait 2650f30a7d0cSTejun Heo * for IO completion of submitted IO. 2651f30a7d0cSTejun Heo */ 2652f30a7d0cSTejun Heo void writeback_inodes_sb_nr(struct super_block *sb, 2653f30a7d0cSTejun Heo unsigned long nr, 2654f30a7d0cSTejun Heo enum wb_reason reason) 2655f30a7d0cSTejun Heo { 2656f30a7d0cSTejun Heo __writeback_inodes_sb_nr(sb, nr, reason, false); 2657f30a7d0cSTejun Heo } 26583259f8beSChris Mason EXPORT_SYMBOL(writeback_inodes_sb_nr); 26593259f8beSChris Mason 26603259f8beSChris Mason /** 26613259f8beSChris Mason * writeback_inodes_sb - writeback dirty inodes from given super_block 26623259f8beSChris Mason * @sb: the superblock 2663786228abSMarcos Paulo de Souza * @reason: reason why some writeback work was initiated 26643259f8beSChris Mason * 26653259f8beSChris Mason * Start writeback on some inodes on this super_block. No guarantees are made 26663259f8beSChris Mason * on how many (if any) will be written, and this function does not wait 26673259f8beSChris Mason * for IO completion of submitted IO. 26683259f8beSChris Mason */ 26690e175a18SCurt Wohlgemuth void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 26703259f8beSChris Mason { 26710e175a18SCurt Wohlgemuth return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 26723259f8beSChris Mason } 2673d8a8559cSJens Axboe EXPORT_SYMBOL(writeback_inodes_sb); 2674d8a8559cSJens Axboe 2675d8a8559cSJens Axboe /** 267610ee27a0SMiao Xie * try_to_writeback_inodes_sb - try to start writeback if none underway 267710ee27a0SMiao Xie * @sb: the superblock 267810ee27a0SMiao Xie * @reason: reason why some writeback work was initiated 267910ee27a0SMiao Xie * 26808264c321SRakesh Pandit * Invoke __writeback_inodes_sb_nr if no writeback is currently underway. 268110ee27a0SMiao Xie */ 26828264c321SRakesh Pandit void try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 268310ee27a0SMiao Xie { 26848264c321SRakesh Pandit if (!down_read_trylock(&sb->s_umount)) 26858264c321SRakesh Pandit return; 26868264c321SRakesh Pandit 26878264c321SRakesh Pandit __writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason, true); 26888264c321SRakesh Pandit up_read(&sb->s_umount); 268910ee27a0SMiao Xie } 269010ee27a0SMiao Xie EXPORT_SYMBOL(try_to_writeback_inodes_sb); 26913259f8beSChris Mason 26923259f8beSChris Mason /** 2693d8a8559cSJens Axboe * sync_inodes_sb - sync sb inode pages 2694d8a8559cSJens Axboe * @sb: the superblock 2695d8a8559cSJens Axboe * 2696d8a8559cSJens Axboe * This function writes and waits on any dirty inode belonging to this 26970dc83bd3SJan Kara * super_block. 2698d8a8559cSJens Axboe */ 26990dc83bd3SJan Kara void sync_inodes_sb(struct super_block *sb) 2700d8a8559cSJens Axboe { 27015b9cce4cSTejun Heo struct backing_dev_info *bdi = sb->s_bdi; 27025b9cce4cSTejun Heo DEFINE_WB_COMPLETION(done, bdi); 270383ba7b07SChristoph Hellwig struct wb_writeback_work work = { 27043c4d7165SChristoph Hellwig .sb = sb, 27053c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_ALL, 27063c4d7165SChristoph Hellwig .nr_pages = LONG_MAX, 27073c4d7165SChristoph Hellwig .range_cyclic = 0, 270883ba7b07SChristoph Hellwig .done = &done, 27090e175a18SCurt Wohlgemuth .reason = WB_REASON_SYNC, 27107747bd4bSDave Chinner .for_sync = 1, 27113c4d7165SChristoph Hellwig }; 27123c4d7165SChristoph Hellwig 2713006a0973STejun Heo /* 2714006a0973STejun Heo * Can't skip on !bdi_has_dirty() because we should wait for !dirty 2715006a0973STejun Heo * inodes under writeback and I_DIRTY_TIME inodes ignored by 2716006a0973STejun Heo * bdi_has_dirty() need to be written out too. 2717006a0973STejun Heo */ 2718006a0973STejun Heo if (bdi == &noop_backing_dev_info) 27196eedc701SJan Kara return; 2720cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2721cf37e972SChristoph Hellwig 27227fc5854fSTejun Heo /* protect against inode wb switch, see inode_switch_wbs_work_fn() */ 27237fc5854fSTejun Heo bdi_down_write_wb_switch_rwsem(bdi); 2724db125360STejun Heo bdi_split_work_to_wbs(bdi, &work, false); 27255b9cce4cSTejun Heo wb_wait_for_completion(&done); 27267fc5854fSTejun Heo bdi_up_write_wb_switch_rwsem(bdi); 272783ba7b07SChristoph Hellwig 2728b6e51316SJens Axboe wait_sb_inodes(sb); 2729d8a8559cSJens Axboe } 2730d8a8559cSJens Axboe EXPORT_SYMBOL(sync_inodes_sb); 27311da177e4SLinus Torvalds 27321da177e4SLinus Torvalds /** 27331da177e4SLinus Torvalds * write_inode_now - write an inode to disk 27341da177e4SLinus Torvalds * @inode: inode to write to disk 27351da177e4SLinus Torvalds * @sync: whether the write should be synchronous or not 27361da177e4SLinus Torvalds * 27377f04c26dSAndrea Arcangeli * This function commits an inode to disk immediately if it is dirty. This is 27387f04c26dSAndrea Arcangeli * primarily needed by knfsd. 27397f04c26dSAndrea Arcangeli * 27407f04c26dSAndrea Arcangeli * The caller must either have a ref on the inode or must have set I_WILL_FREE. 27411da177e4SLinus Torvalds */ 27421da177e4SLinus Torvalds int write_inode_now(struct inode *inode, int sync) 27431da177e4SLinus Torvalds { 27441da177e4SLinus Torvalds struct writeback_control wbc = { 27451da177e4SLinus Torvalds .nr_to_write = LONG_MAX, 274618914b18SMike Galbraith .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, 2747111ebb6eSOGAWA Hirofumi .range_start = 0, 2748111ebb6eSOGAWA Hirofumi .range_end = LLONG_MAX, 27491da177e4SLinus Torvalds }; 27501da177e4SLinus Torvalds 2751f56753acSChristoph Hellwig if (!mapping_can_writeback(inode->i_mapping)) 275249364ce2SAndrew Morton wbc.nr_to_write = 0; 27531da177e4SLinus Torvalds 27541da177e4SLinus Torvalds might_sleep(); 2755aaf25593STejun Heo return writeback_single_inode(inode, &wbc); 27561da177e4SLinus Torvalds } 27571da177e4SLinus Torvalds EXPORT_SYMBOL(write_inode_now); 27581da177e4SLinus Torvalds 27591da177e4SLinus Torvalds /** 2760c691b9d9SAndrew Morton * sync_inode_metadata - write an inode to disk 2761c3765016SChristoph Hellwig * @inode: the inode to sync 2762c3765016SChristoph Hellwig * @wait: wait for I/O to complete. 2763c3765016SChristoph Hellwig * 2764c691b9d9SAndrew Morton * Write an inode to disk and adjust its dirty state after completion. 2765c3765016SChristoph Hellwig * 2766c3765016SChristoph Hellwig * Note: only writes the actual inode, no associated data or other metadata. 2767c3765016SChristoph Hellwig */ 2768c3765016SChristoph Hellwig int sync_inode_metadata(struct inode *inode, int wait) 2769c3765016SChristoph Hellwig { 2770c3765016SChristoph Hellwig struct writeback_control wbc = { 2771c3765016SChristoph Hellwig .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, 2772c3765016SChristoph Hellwig .nr_to_write = 0, /* metadata-only */ 2773c3765016SChristoph Hellwig }; 2774c3765016SChristoph Hellwig 27755662c967SJosef Bacik return writeback_single_inode(inode, &wbc); 2776c3765016SChristoph Hellwig } 2777c3765016SChristoph Hellwig EXPORT_SYMBOL(sync_inode_metadata); 2778