11da177e4SLinus Torvalds /* 21da177e4SLinus Torvalds * fs/fs-writeback.c 31da177e4SLinus Torvalds * 41da177e4SLinus Torvalds * Copyright (C) 2002, Linus Torvalds. 51da177e4SLinus Torvalds * 61da177e4SLinus Torvalds * Contains all the functions related to writing back and waiting 71da177e4SLinus Torvalds * upon dirty inodes against superblocks, and writing back dirty 81da177e4SLinus Torvalds * pages against inodes. ie: data writeback. Writeout of the 91da177e4SLinus Torvalds * inode itself is not handled here. 101da177e4SLinus Torvalds * 11e1f8e874SFrancois Cami * 10Apr2002 Andrew Morton 121da177e4SLinus Torvalds * Split out of fs/inode.c 131da177e4SLinus Torvalds * Additions for address_space-based writeback 141da177e4SLinus Torvalds */ 151da177e4SLinus Torvalds 161da177e4SLinus Torvalds #include <linux/kernel.h> 17630d9c47SPaul Gortmaker #include <linux/export.h> 181da177e4SLinus Torvalds #include <linux/spinlock.h> 195a0e3ad6STejun Heo #include <linux/slab.h> 201da177e4SLinus Torvalds #include <linux/sched.h> 211da177e4SLinus Torvalds #include <linux/fs.h> 221da177e4SLinus Torvalds #include <linux/mm.h> 23bc31b86aSWu Fengguang #include <linux/pagemap.h> 2403ba3782SJens Axboe #include <linux/kthread.h> 251da177e4SLinus Torvalds #include <linux/writeback.h> 261da177e4SLinus Torvalds #include <linux/blkdev.h> 271da177e4SLinus Torvalds #include <linux/backing-dev.h> 28455b2864SDave Chinner #include <linux/tracepoint.h> 29719ea2fbSAl Viro #include <linux/device.h> 3021c6321fSTejun Heo #include <linux/memcontrol.h> 3107f3f05cSDavid Howells #include "internal.h" 321da177e4SLinus Torvalds 33d0bceac7SJens Axboe /* 34bc31b86aSWu Fengguang * 4MB minimal write chunk size 35bc31b86aSWu Fengguang */ 36bc31b86aSWu Fengguang #define MIN_WRITEBACK_PAGES (4096UL >> (PAGE_CACHE_SHIFT - 10)) 37bc31b86aSWu Fengguang 38cc395d7fSTejun Heo struct wb_completion { 39cc395d7fSTejun Heo atomic_t cnt; 40cc395d7fSTejun Heo }; 41cc395d7fSTejun Heo 42bc31b86aSWu Fengguang /* 43c4a77a6cSJens Axboe * Passed into wb_writeback(), essentially a subset of writeback_control 44c4a77a6cSJens Axboe */ 4583ba7b07SChristoph Hellwig struct wb_writeback_work { 46c4a77a6cSJens Axboe long nr_pages; 47c4a77a6cSJens Axboe struct super_block *sb; 480dc83bd3SJan Kara unsigned long *older_than_this; 49c4a77a6cSJens Axboe enum writeback_sync_modes sync_mode; 506e6938b6SWu Fengguang unsigned int tagged_writepages:1; 5152957fe1SH Hartley Sweeten unsigned int for_kupdate:1; 5252957fe1SH Hartley Sweeten unsigned int range_cyclic:1; 5352957fe1SH Hartley Sweeten unsigned int for_background:1; 547747bd4bSDave Chinner unsigned int for_sync:1; /* sync(2) WB_SYNC_ALL writeback */ 55ac7b19a3STejun Heo unsigned int auto_free:1; /* free on completion */ 5698754bf7STejun Heo unsigned int single_wait:1; 5798754bf7STejun Heo unsigned int single_done:1; 580e175a18SCurt Wohlgemuth enum wb_reason reason; /* why was writeback initiated? */ 59c4a77a6cSJens Axboe 608010c3b6SJens Axboe struct list_head list; /* pending work list */ 61cc395d7fSTejun Heo struct wb_completion *done; /* set if the caller waits */ 6203ba3782SJens Axboe }; 6303ba3782SJens Axboe 64a2f48706STheodore Ts'o /* 65cc395d7fSTejun Heo * If one wants to wait for one or more wb_writeback_works, each work's 66cc395d7fSTejun Heo * ->done should be set to a wb_completion defined using the following 67cc395d7fSTejun Heo * macro. Once all work items are issued with wb_queue_work(), the caller 68cc395d7fSTejun Heo * can wait for the completion of all using wb_wait_for_completion(). Work 69cc395d7fSTejun Heo * items which are waited upon aren't freed automatically on completion. 70cc395d7fSTejun Heo */ 71cc395d7fSTejun Heo #define DEFINE_WB_COMPLETION_ONSTACK(cmpl) \ 72cc395d7fSTejun Heo struct wb_completion cmpl = { \ 73cc395d7fSTejun Heo .cnt = ATOMIC_INIT(1), \ 74cc395d7fSTejun Heo } 75cc395d7fSTejun Heo 76cc395d7fSTejun Heo 77cc395d7fSTejun Heo /* 78a2f48706STheodore Ts'o * If an inode is constantly having its pages dirtied, but then the 79a2f48706STheodore Ts'o * updates stop dirtytime_expire_interval seconds in the past, it's 80a2f48706STheodore Ts'o * possible for the worst case time between when an inode has its 81a2f48706STheodore Ts'o * timestamps updated and when they finally get written out to be two 82a2f48706STheodore Ts'o * dirtytime_expire_intervals. We set the default to 12 hours (in 83a2f48706STheodore Ts'o * seconds), which means most of the time inodes will have their 84a2f48706STheodore Ts'o * timestamps written to disk after 12 hours, but in the worst case a 85a2f48706STheodore Ts'o * few inodes might not their timestamps updated for 24 hours. 86a2f48706STheodore Ts'o */ 87a2f48706STheodore Ts'o unsigned int dirtytime_expire_interval = 12 * 60 * 60; 88a2f48706STheodore Ts'o 897ccf19a8SNick Piggin static inline struct inode *wb_inode(struct list_head *head) 907ccf19a8SNick Piggin { 917ccf19a8SNick Piggin return list_entry(head, struct inode, i_wb_list); 927ccf19a8SNick Piggin } 937ccf19a8SNick Piggin 9415eb77a0SWu Fengguang /* 9515eb77a0SWu Fengguang * Include the creation of the trace points after defining the 9615eb77a0SWu Fengguang * wb_writeback_work structure and inline functions so that the definition 9715eb77a0SWu Fengguang * remains local to this file. 9815eb77a0SWu Fengguang */ 9915eb77a0SWu Fengguang #define CREATE_TRACE_POINTS 10015eb77a0SWu Fengguang #include <trace/events/writeback.h> 10115eb77a0SWu Fengguang 102774016b2SSteven Whitehouse EXPORT_TRACEPOINT_SYMBOL_GPL(wbc_writepage); 103774016b2SSteven Whitehouse 104d6c10f1fSTejun Heo static bool wb_io_lists_populated(struct bdi_writeback *wb) 105d6c10f1fSTejun Heo { 106d6c10f1fSTejun Heo if (wb_has_dirty_io(wb)) { 107d6c10f1fSTejun Heo return false; 108d6c10f1fSTejun Heo } else { 109d6c10f1fSTejun Heo set_bit(WB_has_dirty_io, &wb->state); 11095a46c65STejun Heo WARN_ON_ONCE(!wb->avg_write_bandwidth); 111766a9d6eSTejun Heo atomic_long_add(wb->avg_write_bandwidth, 112766a9d6eSTejun Heo &wb->bdi->tot_write_bandwidth); 113d6c10f1fSTejun Heo return true; 114d6c10f1fSTejun Heo } 115d6c10f1fSTejun Heo } 116d6c10f1fSTejun Heo 117d6c10f1fSTejun Heo static void wb_io_lists_depopulated(struct bdi_writeback *wb) 118d6c10f1fSTejun Heo { 119d6c10f1fSTejun Heo if (wb_has_dirty_io(wb) && list_empty(&wb->b_dirty) && 120766a9d6eSTejun Heo list_empty(&wb->b_io) && list_empty(&wb->b_more_io)) { 121d6c10f1fSTejun Heo clear_bit(WB_has_dirty_io, &wb->state); 12295a46c65STejun Heo WARN_ON_ONCE(atomic_long_sub_return(wb->avg_write_bandwidth, 12395a46c65STejun Heo &wb->bdi->tot_write_bandwidth) < 0); 124766a9d6eSTejun Heo } 125d6c10f1fSTejun Heo } 126d6c10f1fSTejun Heo 127d6c10f1fSTejun Heo /** 128d6c10f1fSTejun Heo * inode_wb_list_move_locked - move an inode onto a bdi_writeback IO list 129d6c10f1fSTejun Heo * @inode: inode to be moved 130d6c10f1fSTejun Heo * @wb: target bdi_writeback 131d6c10f1fSTejun Heo * @head: one of @wb->b_{dirty|io|more_io} 132d6c10f1fSTejun Heo * 133d6c10f1fSTejun Heo * Move @inode->i_wb_list to @list of @wb and set %WB_has_dirty_io. 134d6c10f1fSTejun Heo * Returns %true if @inode is the first occupant of the !dirty_time IO 135d6c10f1fSTejun Heo * lists; otherwise, %false. 136d6c10f1fSTejun Heo */ 137d6c10f1fSTejun Heo static bool inode_wb_list_move_locked(struct inode *inode, 138d6c10f1fSTejun Heo struct bdi_writeback *wb, 139d6c10f1fSTejun Heo struct list_head *head) 140d6c10f1fSTejun Heo { 141d6c10f1fSTejun Heo assert_spin_locked(&wb->list_lock); 142d6c10f1fSTejun Heo 143d6c10f1fSTejun Heo list_move(&inode->i_wb_list, head); 144d6c10f1fSTejun Heo 145d6c10f1fSTejun Heo /* dirty_time doesn't count as dirty_io until expiration */ 146d6c10f1fSTejun Heo if (head != &wb->b_dirty_time) 147d6c10f1fSTejun Heo return wb_io_lists_populated(wb); 148d6c10f1fSTejun Heo 149d6c10f1fSTejun Heo wb_io_lists_depopulated(wb); 150d6c10f1fSTejun Heo return false; 151d6c10f1fSTejun Heo } 152d6c10f1fSTejun Heo 153d6c10f1fSTejun Heo /** 154d6c10f1fSTejun Heo * inode_wb_list_del_locked - remove an inode from its bdi_writeback IO list 155d6c10f1fSTejun Heo * @inode: inode to be removed 156d6c10f1fSTejun Heo * @wb: bdi_writeback @inode is being removed from 157d6c10f1fSTejun Heo * 158d6c10f1fSTejun Heo * Remove @inode which may be on one of @wb->b_{dirty|io|more_io} lists and 159d6c10f1fSTejun Heo * clear %WB_has_dirty_io if all are empty afterwards. 160d6c10f1fSTejun Heo */ 161d6c10f1fSTejun Heo static void inode_wb_list_del_locked(struct inode *inode, 162d6c10f1fSTejun Heo struct bdi_writeback *wb) 163d6c10f1fSTejun Heo { 164d6c10f1fSTejun Heo assert_spin_locked(&wb->list_lock); 165d6c10f1fSTejun Heo 166d6c10f1fSTejun Heo list_del_init(&inode->i_wb_list); 167d6c10f1fSTejun Heo wb_io_lists_depopulated(wb); 168d6c10f1fSTejun Heo } 169d6c10f1fSTejun Heo 170f0054bb1STejun Heo static void wb_wakeup(struct bdi_writeback *wb) 1715acda9d1SJan Kara { 172f0054bb1STejun Heo spin_lock_bh(&wb->work_lock); 173f0054bb1STejun Heo if (test_bit(WB_registered, &wb->state)) 174f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 175f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 1765acda9d1SJan Kara } 1775acda9d1SJan Kara 178f0054bb1STejun Heo static void wb_queue_work(struct bdi_writeback *wb, 1796585027aSJan Kara struct wb_writeback_work *work) 1806585027aSJan Kara { 181f0054bb1STejun Heo trace_writeback_queue(wb->bdi, work); 1826585027aSJan Kara 183f0054bb1STejun Heo spin_lock_bh(&wb->work_lock); 18498754bf7STejun Heo if (!test_bit(WB_registered, &wb->state)) { 18598754bf7STejun Heo if (work->single_wait) 18698754bf7STejun Heo work->single_done = 1; 1875acda9d1SJan Kara goto out_unlock; 18898754bf7STejun Heo } 189cc395d7fSTejun Heo if (work->done) 190cc395d7fSTejun Heo atomic_inc(&work->done->cnt); 191f0054bb1STejun Heo list_add_tail(&work->list, &wb->work_list); 192f0054bb1STejun Heo mod_delayed_work(bdi_wq, &wb->dwork, 0); 1935acda9d1SJan Kara out_unlock: 194f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 19503ba3782SJens Axboe } 1961da177e4SLinus Torvalds 197cc395d7fSTejun Heo /** 198cc395d7fSTejun Heo * wb_wait_for_completion - wait for completion of bdi_writeback_works 199cc395d7fSTejun Heo * @bdi: bdi work items were issued to 200cc395d7fSTejun Heo * @done: target wb_completion 201cc395d7fSTejun Heo * 202cc395d7fSTejun Heo * Wait for one or more work items issued to @bdi with their ->done field 203cc395d7fSTejun Heo * set to @done, which should have been defined with 204cc395d7fSTejun Heo * DEFINE_WB_COMPLETION_ONSTACK(). This function returns after all such 205cc395d7fSTejun Heo * work items are completed. Work items which are waited upon aren't freed 206cc395d7fSTejun Heo * automatically on completion. 207cc395d7fSTejun Heo */ 208cc395d7fSTejun Heo static void wb_wait_for_completion(struct backing_dev_info *bdi, 209cc395d7fSTejun Heo struct wb_completion *done) 210cc395d7fSTejun Heo { 211cc395d7fSTejun Heo atomic_dec(&done->cnt); /* put down the initial count */ 212cc395d7fSTejun Heo wait_event(bdi->wb_waitq, !atomic_read(&done->cnt)); 213cc395d7fSTejun Heo } 214cc395d7fSTejun Heo 215703c2708STejun Heo #ifdef CONFIG_CGROUP_WRITEBACK 216703c2708STejun Heo 2172a814908STejun Heo /* parameters for foreign inode detection, see wb_detach_inode() */ 2182a814908STejun Heo #define WB_FRN_TIME_SHIFT 13 /* 1s = 2^13, upto 8 secs w/ 16bit */ 2192a814908STejun Heo #define WB_FRN_TIME_AVG_SHIFT 3 /* avg = avg * 7/8 + new * 1/8 */ 2202a814908STejun Heo #define WB_FRN_TIME_CUT_DIV 2 /* ignore rounds < avg / 2 */ 2212a814908STejun Heo #define WB_FRN_TIME_PERIOD (2 * (1 << WB_FRN_TIME_SHIFT)) /* 2s */ 2222a814908STejun Heo 2232a814908STejun Heo #define WB_FRN_HIST_SLOTS 16 /* inode->i_wb_frn_history is 16bit */ 2242a814908STejun Heo #define WB_FRN_HIST_UNIT (WB_FRN_TIME_PERIOD / WB_FRN_HIST_SLOTS) 2252a814908STejun Heo /* each slot's duration is 2s / 16 */ 2262a814908STejun Heo #define WB_FRN_HIST_THR_SLOTS (WB_FRN_HIST_SLOTS / 2) 2272a814908STejun Heo /* if foreign slots >= 8, switch */ 2282a814908STejun Heo #define WB_FRN_HIST_MAX_SLOTS (WB_FRN_HIST_THR_SLOTS / 2 + 1) 2292a814908STejun Heo /* one round can affect upto 5 slots */ 2302a814908STejun Heo 23121c6321fSTejun Heo void __inode_attach_wb(struct inode *inode, struct page *page) 23221c6321fSTejun Heo { 23321c6321fSTejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 23421c6321fSTejun Heo struct bdi_writeback *wb = NULL; 23521c6321fSTejun Heo 23621c6321fSTejun Heo if (inode_cgwb_enabled(inode)) { 23721c6321fSTejun Heo struct cgroup_subsys_state *memcg_css; 23821c6321fSTejun Heo 23921c6321fSTejun Heo if (page) { 24021c6321fSTejun Heo memcg_css = mem_cgroup_css_from_page(page); 24121c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 24221c6321fSTejun Heo } else { 24321c6321fSTejun Heo /* must pin memcg_css, see wb_get_create() */ 24421c6321fSTejun Heo memcg_css = task_get_css(current, memory_cgrp_id); 24521c6321fSTejun Heo wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 24621c6321fSTejun Heo css_put(memcg_css); 24721c6321fSTejun Heo } 24821c6321fSTejun Heo } 24921c6321fSTejun Heo 25021c6321fSTejun Heo if (!wb) 25121c6321fSTejun Heo wb = &bdi->wb; 25221c6321fSTejun Heo 25321c6321fSTejun Heo /* 25421c6321fSTejun Heo * There may be multiple instances of this function racing to 25521c6321fSTejun Heo * update the same inode. Use cmpxchg() to tell the winner. 25621c6321fSTejun Heo */ 25721c6321fSTejun Heo if (unlikely(cmpxchg(&inode->i_wb, NULL, wb))) 25821c6321fSTejun Heo wb_put(wb); 25921c6321fSTejun Heo } 26021c6321fSTejun Heo 261703c2708STejun Heo /** 26287e1d789STejun Heo * locked_inode_to_wb_and_lock_list - determine a locked inode's wb and lock it 26387e1d789STejun Heo * @inode: inode of interest with i_lock held 26487e1d789STejun Heo * 26587e1d789STejun Heo * Returns @inode's wb with its list_lock held. @inode->i_lock must be 26687e1d789STejun Heo * held on entry and is released on return. The returned wb is guaranteed 26787e1d789STejun Heo * to stay @inode's associated wb until its list_lock is released. 26887e1d789STejun Heo */ 26987e1d789STejun Heo static struct bdi_writeback * 27087e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 27187e1d789STejun Heo __releases(&inode->i_lock) 27287e1d789STejun Heo __acquires(&wb->list_lock) 27387e1d789STejun Heo { 27487e1d789STejun Heo while (true) { 27587e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 27687e1d789STejun Heo 27787e1d789STejun Heo /* 27887e1d789STejun Heo * inode_to_wb() association is protected by both 27987e1d789STejun Heo * @inode->i_lock and @wb->list_lock but list_lock nests 28087e1d789STejun Heo * outside i_lock. Drop i_lock and verify that the 28187e1d789STejun Heo * association hasn't changed after acquiring list_lock. 28287e1d789STejun Heo */ 28387e1d789STejun Heo wb_get(wb); 28487e1d789STejun Heo spin_unlock(&inode->i_lock); 28587e1d789STejun Heo spin_lock(&wb->list_lock); 28687e1d789STejun Heo wb_put(wb); /* not gonna deref it anymore */ 28787e1d789STejun Heo 288aaa2cacfSTejun Heo /* i_wb may have changed inbetween, can't use inode_to_wb() */ 289aaa2cacfSTejun Heo if (likely(wb == inode->i_wb)) 29087e1d789STejun Heo return wb; /* @inode already has ref */ 29187e1d789STejun Heo 29287e1d789STejun Heo spin_unlock(&wb->list_lock); 29387e1d789STejun Heo cpu_relax(); 29487e1d789STejun Heo spin_lock(&inode->i_lock); 29587e1d789STejun Heo } 29687e1d789STejun Heo } 29787e1d789STejun Heo 29887e1d789STejun Heo /** 29987e1d789STejun Heo * inode_to_wb_and_lock_list - determine an inode's wb and lock it 30087e1d789STejun Heo * @inode: inode of interest 30187e1d789STejun Heo * 30287e1d789STejun Heo * Same as locked_inode_to_wb_and_lock_list() but @inode->i_lock isn't held 30387e1d789STejun Heo * on entry. 30487e1d789STejun Heo */ 30587e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 30687e1d789STejun Heo __acquires(&wb->list_lock) 30787e1d789STejun Heo { 30887e1d789STejun Heo spin_lock(&inode->i_lock); 30987e1d789STejun Heo return locked_inode_to_wb_and_lock_list(inode); 31087e1d789STejun Heo } 31187e1d789STejun Heo 312682aa8e1STejun Heo struct inode_switch_wbs_context { 313682aa8e1STejun Heo struct inode *inode; 314682aa8e1STejun Heo struct bdi_writeback *new_wb; 315682aa8e1STejun Heo 316682aa8e1STejun Heo struct rcu_head rcu_head; 317682aa8e1STejun Heo struct work_struct work; 318682aa8e1STejun Heo }; 319682aa8e1STejun Heo 320682aa8e1STejun Heo static void inode_switch_wbs_work_fn(struct work_struct *work) 321682aa8e1STejun Heo { 322682aa8e1STejun Heo struct inode_switch_wbs_context *isw = 323682aa8e1STejun Heo container_of(work, struct inode_switch_wbs_context, work); 324682aa8e1STejun Heo struct inode *inode = isw->inode; 325682aa8e1STejun Heo struct bdi_writeback *new_wb = isw->new_wb; 326682aa8e1STejun Heo 327682aa8e1STejun Heo /* 328682aa8e1STejun Heo * By the time control reaches here, RCU grace period has passed 329682aa8e1STejun Heo * since I_WB_SWITCH assertion and all wb stat update transactions 330682aa8e1STejun Heo * between unlocked_inode_to_wb_begin/end() are guaranteed to be 331682aa8e1STejun Heo * synchronizing against mapping->tree_lock. 332682aa8e1STejun Heo */ 333682aa8e1STejun Heo spin_lock(&inode->i_lock); 334682aa8e1STejun Heo 335682aa8e1STejun Heo inode->i_wb_frn_winner = 0; 336682aa8e1STejun Heo inode->i_wb_frn_avg_time = 0; 337682aa8e1STejun Heo inode->i_wb_frn_history = 0; 338682aa8e1STejun Heo 339682aa8e1STejun Heo /* 340682aa8e1STejun Heo * Paired with load_acquire in unlocked_inode_to_wb_begin() and 341682aa8e1STejun Heo * ensures that the new wb is visible if they see !I_WB_SWITCH. 342682aa8e1STejun Heo */ 343682aa8e1STejun Heo smp_store_release(&inode->i_state, inode->i_state & ~I_WB_SWITCH); 344682aa8e1STejun Heo 345682aa8e1STejun Heo spin_unlock(&inode->i_lock); 346682aa8e1STejun Heo 347682aa8e1STejun Heo iput(inode); 348682aa8e1STejun Heo wb_put(new_wb); 349682aa8e1STejun Heo kfree(isw); 350682aa8e1STejun Heo } 351682aa8e1STejun Heo 352682aa8e1STejun Heo static void inode_switch_wbs_rcu_fn(struct rcu_head *rcu_head) 353682aa8e1STejun Heo { 354682aa8e1STejun Heo struct inode_switch_wbs_context *isw = container_of(rcu_head, 355682aa8e1STejun Heo struct inode_switch_wbs_context, rcu_head); 356682aa8e1STejun Heo 357682aa8e1STejun Heo /* needs to grab bh-unsafe locks, bounce to work item */ 358682aa8e1STejun Heo INIT_WORK(&isw->work, inode_switch_wbs_work_fn); 359682aa8e1STejun Heo schedule_work(&isw->work); 360682aa8e1STejun Heo } 361682aa8e1STejun Heo 362682aa8e1STejun Heo /** 363682aa8e1STejun Heo * inode_switch_wbs - change the wb association of an inode 364682aa8e1STejun Heo * @inode: target inode 365682aa8e1STejun Heo * @new_wb_id: ID of the new wb 366682aa8e1STejun Heo * 367682aa8e1STejun Heo * Switch @inode's wb association to the wb identified by @new_wb_id. The 368682aa8e1STejun Heo * switching is performed asynchronously and may fail silently. 369682aa8e1STejun Heo */ 370682aa8e1STejun Heo static void inode_switch_wbs(struct inode *inode, int new_wb_id) 371682aa8e1STejun Heo { 372682aa8e1STejun Heo struct backing_dev_info *bdi = inode_to_bdi(inode); 373682aa8e1STejun Heo struct cgroup_subsys_state *memcg_css; 374682aa8e1STejun Heo struct inode_switch_wbs_context *isw; 375682aa8e1STejun Heo 376682aa8e1STejun Heo /* noop if seems to be already in progress */ 377682aa8e1STejun Heo if (inode->i_state & I_WB_SWITCH) 378682aa8e1STejun Heo return; 379682aa8e1STejun Heo 380682aa8e1STejun Heo isw = kzalloc(sizeof(*isw), GFP_ATOMIC); 381682aa8e1STejun Heo if (!isw) 382682aa8e1STejun Heo return; 383682aa8e1STejun Heo 384682aa8e1STejun Heo /* find and pin the new wb */ 385682aa8e1STejun Heo rcu_read_lock(); 386682aa8e1STejun Heo memcg_css = css_from_id(new_wb_id, &memory_cgrp_subsys); 387682aa8e1STejun Heo if (memcg_css) 388682aa8e1STejun Heo isw->new_wb = wb_get_create(bdi, memcg_css, GFP_ATOMIC); 389682aa8e1STejun Heo rcu_read_unlock(); 390682aa8e1STejun Heo if (!isw->new_wb) 391682aa8e1STejun Heo goto out_free; 392682aa8e1STejun Heo 393682aa8e1STejun Heo /* while holding I_WB_SWITCH, no one else can update the association */ 394682aa8e1STejun Heo spin_lock(&inode->i_lock); 395682aa8e1STejun Heo if (inode->i_state & (I_WB_SWITCH | I_FREEING) || 396682aa8e1STejun Heo inode_to_wb(inode) == isw->new_wb) { 397682aa8e1STejun Heo spin_unlock(&inode->i_lock); 398682aa8e1STejun Heo goto out_free; 399682aa8e1STejun Heo } 400682aa8e1STejun Heo inode->i_state |= I_WB_SWITCH; 401682aa8e1STejun Heo spin_unlock(&inode->i_lock); 402682aa8e1STejun Heo 403682aa8e1STejun Heo ihold(inode); 404682aa8e1STejun Heo isw->inode = inode; 405682aa8e1STejun Heo 406682aa8e1STejun Heo /* 407682aa8e1STejun Heo * In addition to synchronizing among switchers, I_WB_SWITCH tells 408682aa8e1STejun Heo * the RCU protected stat update paths to grab the mapping's 409682aa8e1STejun Heo * tree_lock so that stat transfer can synchronize against them. 410682aa8e1STejun Heo * Let's continue after I_WB_SWITCH is guaranteed to be visible. 411682aa8e1STejun Heo */ 412682aa8e1STejun Heo call_rcu(&isw->rcu_head, inode_switch_wbs_rcu_fn); 413682aa8e1STejun Heo return; 414682aa8e1STejun Heo 415682aa8e1STejun Heo out_free: 416682aa8e1STejun Heo if (isw->new_wb) 417682aa8e1STejun Heo wb_put(isw->new_wb); 418682aa8e1STejun Heo kfree(isw); 419682aa8e1STejun Heo } 420682aa8e1STejun Heo 42187e1d789STejun Heo /** 422b16b1debSTejun Heo * wbc_attach_and_unlock_inode - associate wbc with target inode and unlock it 423b16b1debSTejun Heo * @wbc: writeback_control of interest 424b16b1debSTejun Heo * @inode: target inode 425b16b1debSTejun Heo * 426b16b1debSTejun Heo * @inode is locked and about to be written back under the control of @wbc. 427b16b1debSTejun Heo * Record @inode's writeback context into @wbc and unlock the i_lock. On 428b16b1debSTejun Heo * writeback completion, wbc_detach_inode() should be called. This is used 429b16b1debSTejun Heo * to track the cgroup writeback context. 430b16b1debSTejun Heo */ 431b16b1debSTejun Heo void wbc_attach_and_unlock_inode(struct writeback_control *wbc, 432b16b1debSTejun Heo struct inode *inode) 433b16b1debSTejun Heo { 434b16b1debSTejun Heo wbc->wb = inode_to_wb(inode); 4352a814908STejun Heo wbc->inode = inode; 4362a814908STejun Heo 4372a814908STejun Heo wbc->wb_id = wbc->wb->memcg_css->id; 4382a814908STejun Heo wbc->wb_lcand_id = inode->i_wb_frn_winner; 4392a814908STejun Heo wbc->wb_tcand_id = 0; 4402a814908STejun Heo wbc->wb_bytes = 0; 4412a814908STejun Heo wbc->wb_lcand_bytes = 0; 4422a814908STejun Heo wbc->wb_tcand_bytes = 0; 4432a814908STejun Heo 444b16b1debSTejun Heo wb_get(wbc->wb); 445b16b1debSTejun Heo spin_unlock(&inode->i_lock); 446b16b1debSTejun Heo } 447b16b1debSTejun Heo 448b16b1debSTejun Heo /** 4492a814908STejun Heo * wbc_detach_inode - disassociate wbc from inode and perform foreign detection 4502a814908STejun Heo * @wbc: writeback_control of the just finished writeback 451b16b1debSTejun Heo * 452b16b1debSTejun Heo * To be called after a writeback attempt of an inode finishes and undoes 453b16b1debSTejun Heo * wbc_attach_and_unlock_inode(). Can be called under any context. 4542a814908STejun Heo * 4552a814908STejun Heo * As concurrent write sharing of an inode is expected to be very rare and 4562a814908STejun Heo * memcg only tracks page ownership on first-use basis severely confining 4572a814908STejun Heo * the usefulness of such sharing, cgroup writeback tracks ownership 4582a814908STejun Heo * per-inode. While the support for concurrent write sharing of an inode 4592a814908STejun Heo * is deemed unnecessary, an inode being written to by different cgroups at 4602a814908STejun Heo * different points in time is a lot more common, and, more importantly, 4612a814908STejun Heo * charging only by first-use can too readily lead to grossly incorrect 4622a814908STejun Heo * behaviors (single foreign page can lead to gigabytes of writeback to be 4632a814908STejun Heo * incorrectly attributed). 4642a814908STejun Heo * 4652a814908STejun Heo * To resolve this issue, cgroup writeback detects the majority dirtier of 4662a814908STejun Heo * an inode and transfers the ownership to it. To avoid unnnecessary 4672a814908STejun Heo * oscillation, the detection mechanism keeps track of history and gives 4682a814908STejun Heo * out the switch verdict only if the foreign usage pattern is stable over 4692a814908STejun Heo * a certain amount of time and/or writeback attempts. 4702a814908STejun Heo * 4712a814908STejun Heo * On each writeback attempt, @wbc tries to detect the majority writer 4722a814908STejun Heo * using Boyer-Moore majority vote algorithm. In addition to the byte 4732a814908STejun Heo * count from the majority voting, it also counts the bytes written for the 4742a814908STejun Heo * current wb and the last round's winner wb (max of last round's current 4752a814908STejun Heo * wb, the winner from two rounds ago, and the last round's majority 4762a814908STejun Heo * candidate). Keeping track of the historical winner helps the algorithm 4772a814908STejun Heo * to semi-reliably detect the most active writer even when it's not the 4782a814908STejun Heo * absolute majority. 4792a814908STejun Heo * 4802a814908STejun Heo * Once the winner of the round is determined, whether the winner is 4812a814908STejun Heo * foreign or not and how much IO time the round consumed is recorded in 4822a814908STejun Heo * inode->i_wb_frn_history. If the amount of recorded foreign IO time is 4832a814908STejun Heo * over a certain threshold, the switch verdict is given. 484b16b1debSTejun Heo */ 485b16b1debSTejun Heo void wbc_detach_inode(struct writeback_control *wbc) 486b16b1debSTejun Heo { 4872a814908STejun Heo struct bdi_writeback *wb = wbc->wb; 4882a814908STejun Heo struct inode *inode = wbc->inode; 4892a814908STejun Heo u16 history = inode->i_wb_frn_history; 4902a814908STejun Heo unsigned long avg_time = inode->i_wb_frn_avg_time; 4912a814908STejun Heo unsigned long max_bytes, max_time; 4922a814908STejun Heo int max_id; 4932a814908STejun Heo 4942a814908STejun Heo /* pick the winner of this round */ 4952a814908STejun Heo if (wbc->wb_bytes >= wbc->wb_lcand_bytes && 4962a814908STejun Heo wbc->wb_bytes >= wbc->wb_tcand_bytes) { 4972a814908STejun Heo max_id = wbc->wb_id; 4982a814908STejun Heo max_bytes = wbc->wb_bytes; 4992a814908STejun Heo } else if (wbc->wb_lcand_bytes >= wbc->wb_tcand_bytes) { 5002a814908STejun Heo max_id = wbc->wb_lcand_id; 5012a814908STejun Heo max_bytes = wbc->wb_lcand_bytes; 5022a814908STejun Heo } else { 5032a814908STejun Heo max_id = wbc->wb_tcand_id; 5042a814908STejun Heo max_bytes = wbc->wb_tcand_bytes; 5052a814908STejun Heo } 5062a814908STejun Heo 5072a814908STejun Heo /* 5082a814908STejun Heo * Calculate the amount of IO time the winner consumed and fold it 5092a814908STejun Heo * into the running average kept per inode. If the consumed IO 5102a814908STejun Heo * time is lower than avag / WB_FRN_TIME_CUT_DIV, ignore it for 5112a814908STejun Heo * deciding whether to switch or not. This is to prevent one-off 5122a814908STejun Heo * small dirtiers from skewing the verdict. 5132a814908STejun Heo */ 5142a814908STejun Heo max_time = DIV_ROUND_UP((max_bytes >> PAGE_SHIFT) << WB_FRN_TIME_SHIFT, 5152a814908STejun Heo wb->avg_write_bandwidth); 5162a814908STejun Heo if (avg_time) 5172a814908STejun Heo avg_time += (max_time >> WB_FRN_TIME_AVG_SHIFT) - 5182a814908STejun Heo (avg_time >> WB_FRN_TIME_AVG_SHIFT); 5192a814908STejun Heo else 5202a814908STejun Heo avg_time = max_time; /* immediate catch up on first run */ 5212a814908STejun Heo 5222a814908STejun Heo if (max_time >= avg_time / WB_FRN_TIME_CUT_DIV) { 5232a814908STejun Heo int slots; 5242a814908STejun Heo 5252a814908STejun Heo /* 5262a814908STejun Heo * The switch verdict is reached if foreign wb's consume 5272a814908STejun Heo * more than a certain proportion of IO time in a 5282a814908STejun Heo * WB_FRN_TIME_PERIOD. This is loosely tracked by 16 slot 5292a814908STejun Heo * history mask where each bit represents one sixteenth of 5302a814908STejun Heo * the period. Determine the number of slots to shift into 5312a814908STejun Heo * history from @max_time. 5322a814908STejun Heo */ 5332a814908STejun Heo slots = min(DIV_ROUND_UP(max_time, WB_FRN_HIST_UNIT), 5342a814908STejun Heo (unsigned long)WB_FRN_HIST_MAX_SLOTS); 5352a814908STejun Heo history <<= slots; 5362a814908STejun Heo if (wbc->wb_id != max_id) 5372a814908STejun Heo history |= (1U << slots) - 1; 5382a814908STejun Heo 5392a814908STejun Heo /* 5402a814908STejun Heo * Switch if the current wb isn't the consistent winner. 5412a814908STejun Heo * If there are multiple closely competing dirtiers, the 5422a814908STejun Heo * inode may switch across them repeatedly over time, which 5432a814908STejun Heo * is okay. The main goal is avoiding keeping an inode on 5442a814908STejun Heo * the wrong wb for an extended period of time. 5452a814908STejun Heo */ 546682aa8e1STejun Heo if (hweight32(history) > WB_FRN_HIST_THR_SLOTS) 547682aa8e1STejun Heo inode_switch_wbs(inode, max_id); 5482a814908STejun Heo } 5492a814908STejun Heo 5502a814908STejun Heo /* 5512a814908STejun Heo * Multiple instances of this function may race to update the 5522a814908STejun Heo * following fields but we don't mind occassional inaccuracies. 5532a814908STejun Heo */ 5542a814908STejun Heo inode->i_wb_frn_winner = max_id; 5552a814908STejun Heo inode->i_wb_frn_avg_time = min(avg_time, (unsigned long)U16_MAX); 5562a814908STejun Heo inode->i_wb_frn_history = history; 5572a814908STejun Heo 558b16b1debSTejun Heo wb_put(wbc->wb); 559b16b1debSTejun Heo wbc->wb = NULL; 560b16b1debSTejun Heo } 561b16b1debSTejun Heo 562b16b1debSTejun Heo /** 5632a814908STejun Heo * wbc_account_io - account IO issued during writeback 5642a814908STejun Heo * @wbc: writeback_control of the writeback in progress 5652a814908STejun Heo * @page: page being written out 5662a814908STejun Heo * @bytes: number of bytes being written out 5672a814908STejun Heo * 5682a814908STejun Heo * @bytes from @page are about to written out during the writeback 5692a814908STejun Heo * controlled by @wbc. Keep the book for foreign inode detection. See 5702a814908STejun Heo * wbc_detach_inode(). 5712a814908STejun Heo */ 5722a814908STejun Heo void wbc_account_io(struct writeback_control *wbc, struct page *page, 5732a814908STejun Heo size_t bytes) 5742a814908STejun Heo { 5752a814908STejun Heo int id; 5762a814908STejun Heo 5772a814908STejun Heo /* 5782a814908STejun Heo * pageout() path doesn't attach @wbc to the inode being written 5792a814908STejun Heo * out. This is intentional as we don't want the function to block 5802a814908STejun Heo * behind a slow cgroup. Ultimately, we want pageout() to kick off 5812a814908STejun Heo * regular writeback instead of writing things out itself. 5822a814908STejun Heo */ 5832a814908STejun Heo if (!wbc->wb) 5842a814908STejun Heo return; 5852a814908STejun Heo 5862a814908STejun Heo rcu_read_lock(); 5872a814908STejun Heo id = mem_cgroup_css_from_page(page)->id; 5882a814908STejun Heo rcu_read_unlock(); 5892a814908STejun Heo 5902a814908STejun Heo if (id == wbc->wb_id) { 5912a814908STejun Heo wbc->wb_bytes += bytes; 5922a814908STejun Heo return; 5932a814908STejun Heo } 5942a814908STejun Heo 5952a814908STejun Heo if (id == wbc->wb_lcand_id) 5962a814908STejun Heo wbc->wb_lcand_bytes += bytes; 5972a814908STejun Heo 5982a814908STejun Heo /* Boyer-Moore majority vote algorithm */ 5992a814908STejun Heo if (!wbc->wb_tcand_bytes) 6002a814908STejun Heo wbc->wb_tcand_id = id; 6012a814908STejun Heo if (id == wbc->wb_tcand_id) 6022a814908STejun Heo wbc->wb_tcand_bytes += bytes; 6032a814908STejun Heo else 6042a814908STejun Heo wbc->wb_tcand_bytes -= min(bytes, wbc->wb_tcand_bytes); 6052a814908STejun Heo } 6062a814908STejun Heo 6072a814908STejun Heo /** 608703c2708STejun Heo * inode_congested - test whether an inode is congested 609703c2708STejun Heo * @inode: inode to test for congestion 610703c2708STejun Heo * @cong_bits: mask of WB_[a]sync_congested bits to test 611703c2708STejun Heo * 612703c2708STejun Heo * Tests whether @inode is congested. @cong_bits is the mask of congestion 613703c2708STejun Heo * bits to test and the return value is the mask of set bits. 614703c2708STejun Heo * 615703c2708STejun Heo * If cgroup writeback is enabled for @inode, the congestion state is 616703c2708STejun Heo * determined by whether the cgwb (cgroup bdi_writeback) for the blkcg 617703c2708STejun Heo * associated with @inode is congested; otherwise, the root wb's congestion 618703c2708STejun Heo * state is used. 619703c2708STejun Heo */ 620703c2708STejun Heo int inode_congested(struct inode *inode, int cong_bits) 621703c2708STejun Heo { 6225cb8b824STejun Heo /* 6235cb8b824STejun Heo * Once set, ->i_wb never becomes NULL while the inode is alive. 6245cb8b824STejun Heo * Start transaction iff ->i_wb is visible. 6255cb8b824STejun Heo */ 626aaa2cacfSTejun Heo if (inode && inode_to_wb_is_valid(inode)) { 6275cb8b824STejun Heo struct bdi_writeback *wb; 6285cb8b824STejun Heo bool locked, congested; 6295cb8b824STejun Heo 6305cb8b824STejun Heo wb = unlocked_inode_to_wb_begin(inode, &locked); 6315cb8b824STejun Heo congested = wb_congested(wb, cong_bits); 6325cb8b824STejun Heo unlocked_inode_to_wb_end(inode, locked); 6335cb8b824STejun Heo return congested; 634703c2708STejun Heo } 635703c2708STejun Heo 636703c2708STejun Heo return wb_congested(&inode_to_bdi(inode)->wb, cong_bits); 637703c2708STejun Heo } 638703c2708STejun Heo EXPORT_SYMBOL_GPL(inode_congested); 639703c2708STejun Heo 640f2b65121STejun Heo /** 64198754bf7STejun Heo * wb_wait_for_single_work - wait for completion of a single bdi_writeback_work 64298754bf7STejun Heo * @bdi: bdi the work item was issued to 64398754bf7STejun Heo * @work: work item to wait for 64498754bf7STejun Heo * 64598754bf7STejun Heo * Wait for the completion of @work which was issued to one of @bdi's 64698754bf7STejun Heo * bdi_writeback's. The caller must have set @work->single_wait before 64798754bf7STejun Heo * issuing it. This wait operates independently fo 64898754bf7STejun Heo * wb_wait_for_completion() and also disables automatic freeing of @work. 64998754bf7STejun Heo */ 65098754bf7STejun Heo static void wb_wait_for_single_work(struct backing_dev_info *bdi, 65198754bf7STejun Heo struct wb_writeback_work *work) 65298754bf7STejun Heo { 65398754bf7STejun Heo if (WARN_ON_ONCE(!work->single_wait)) 65498754bf7STejun Heo return; 65598754bf7STejun Heo 65698754bf7STejun Heo wait_event(bdi->wb_waitq, work->single_done); 65798754bf7STejun Heo 65898754bf7STejun Heo /* 65998754bf7STejun Heo * Paired with smp_wmb() in wb_do_writeback() and ensures that all 66098754bf7STejun Heo * modifications to @work prior to assertion of ->single_done is 66198754bf7STejun Heo * visible to the caller once this function returns. 66298754bf7STejun Heo */ 66398754bf7STejun Heo smp_rmb(); 66498754bf7STejun Heo } 66598754bf7STejun Heo 66698754bf7STejun Heo /** 667f2b65121STejun Heo * wb_split_bdi_pages - split nr_pages to write according to bandwidth 668f2b65121STejun Heo * @wb: target bdi_writeback to split @nr_pages to 669f2b65121STejun Heo * @nr_pages: number of pages to write for the whole bdi 670f2b65121STejun Heo * 671f2b65121STejun Heo * Split @wb's portion of @nr_pages according to @wb's write bandwidth in 672f2b65121STejun Heo * relation to the total write bandwidth of all wb's w/ dirty inodes on 673f2b65121STejun Heo * @wb->bdi. 674f2b65121STejun Heo */ 675f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 676f2b65121STejun Heo { 677f2b65121STejun Heo unsigned long this_bw = wb->avg_write_bandwidth; 678f2b65121STejun Heo unsigned long tot_bw = atomic_long_read(&wb->bdi->tot_write_bandwidth); 679f2b65121STejun Heo 680f2b65121STejun Heo if (nr_pages == LONG_MAX) 681f2b65121STejun Heo return LONG_MAX; 682f2b65121STejun Heo 683f2b65121STejun Heo /* 684f2b65121STejun Heo * This may be called on clean wb's and proportional distribution 685f2b65121STejun Heo * may not make sense, just use the original @nr_pages in those 686f2b65121STejun Heo * cases. In general, we wanna err on the side of writing more. 687f2b65121STejun Heo */ 688f2b65121STejun Heo if (!tot_bw || this_bw >= tot_bw) 689f2b65121STejun Heo return nr_pages; 690f2b65121STejun Heo else 691f2b65121STejun Heo return DIV_ROUND_UP_ULL((u64)nr_pages * this_bw, tot_bw); 692f2b65121STejun Heo } 693f2b65121STejun Heo 694db125360STejun Heo /** 695db125360STejun Heo * wb_clone_and_queue_work - clone a wb_writeback_work and issue it to a wb 696db125360STejun Heo * @wb: target bdi_writeback 697db125360STejun Heo * @base_work: source wb_writeback_work 698db125360STejun Heo * 699db125360STejun Heo * Try to make a clone of @base_work and issue it to @wb. If cloning 700db125360STejun Heo * succeeds, %true is returned; otherwise, @base_work is issued directly 701db125360STejun Heo * and %false is returned. In the latter case, the caller is required to 702db125360STejun Heo * wait for @base_work's completion using wb_wait_for_single_work(). 703db125360STejun Heo * 704db125360STejun Heo * A clone is auto-freed on completion. @base_work never is. 705db125360STejun Heo */ 706db125360STejun Heo static bool wb_clone_and_queue_work(struct bdi_writeback *wb, 707db125360STejun Heo struct wb_writeback_work *base_work) 708db125360STejun Heo { 709db125360STejun Heo struct wb_writeback_work *work; 710db125360STejun Heo 711db125360STejun Heo work = kmalloc(sizeof(*work), GFP_ATOMIC); 712db125360STejun Heo if (work) { 713db125360STejun Heo *work = *base_work; 714db125360STejun Heo work->auto_free = 1; 715db125360STejun Heo work->single_wait = 0; 716db125360STejun Heo } else { 717db125360STejun Heo work = base_work; 718db125360STejun Heo work->auto_free = 0; 719db125360STejun Heo work->single_wait = 1; 720db125360STejun Heo } 721db125360STejun Heo work->single_done = 0; 722db125360STejun Heo wb_queue_work(wb, work); 723db125360STejun Heo return work != base_work; 724db125360STejun Heo } 725db125360STejun Heo 726db125360STejun Heo /** 727db125360STejun Heo * bdi_split_work_to_wbs - split a wb_writeback_work to all wb's of a bdi 728db125360STejun Heo * @bdi: target backing_dev_info 729db125360STejun Heo * @base_work: wb_writeback_work to issue 730db125360STejun Heo * @skip_if_busy: skip wb's which already have writeback in progress 731db125360STejun Heo * 732db125360STejun Heo * Split and issue @base_work to all wb's (bdi_writeback's) of @bdi which 733db125360STejun Heo * have dirty inodes. If @base_work->nr_page isn't %LONG_MAX, it's 734db125360STejun Heo * distributed to the busy wbs according to each wb's proportion in the 735db125360STejun Heo * total active write bandwidth of @bdi. 736db125360STejun Heo */ 737db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 738db125360STejun Heo struct wb_writeback_work *base_work, 739db125360STejun Heo bool skip_if_busy) 740db125360STejun Heo { 741db125360STejun Heo long nr_pages = base_work->nr_pages; 742db125360STejun Heo int next_blkcg_id = 0; 743db125360STejun Heo struct bdi_writeback *wb; 744db125360STejun Heo struct wb_iter iter; 745db125360STejun Heo 746db125360STejun Heo might_sleep(); 747db125360STejun Heo 748db125360STejun Heo if (!bdi_has_dirty_io(bdi)) 749db125360STejun Heo return; 750db125360STejun Heo restart: 751db125360STejun Heo rcu_read_lock(); 752db125360STejun Heo bdi_for_each_wb(wb, bdi, &iter, next_blkcg_id) { 753db125360STejun Heo if (!wb_has_dirty_io(wb) || 754db125360STejun Heo (skip_if_busy && writeback_in_progress(wb))) 755db125360STejun Heo continue; 756db125360STejun Heo 757db125360STejun Heo base_work->nr_pages = wb_split_bdi_pages(wb, nr_pages); 758db125360STejun Heo if (!wb_clone_and_queue_work(wb, base_work)) { 759db125360STejun Heo next_blkcg_id = wb->blkcg_css->id + 1; 760db125360STejun Heo rcu_read_unlock(); 761db125360STejun Heo wb_wait_for_single_work(bdi, base_work); 762db125360STejun Heo goto restart; 763db125360STejun Heo } 764db125360STejun Heo } 765db125360STejun Heo rcu_read_unlock(); 766db125360STejun Heo } 767db125360STejun Heo 768f2b65121STejun Heo #else /* CONFIG_CGROUP_WRITEBACK */ 769f2b65121STejun Heo 77087e1d789STejun Heo static struct bdi_writeback * 77187e1d789STejun Heo locked_inode_to_wb_and_lock_list(struct inode *inode) 77287e1d789STejun Heo __releases(&inode->i_lock) 77387e1d789STejun Heo __acquires(&wb->list_lock) 77487e1d789STejun Heo { 77587e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 77687e1d789STejun Heo 77787e1d789STejun Heo spin_unlock(&inode->i_lock); 77887e1d789STejun Heo spin_lock(&wb->list_lock); 77987e1d789STejun Heo return wb; 78087e1d789STejun Heo } 78187e1d789STejun Heo 78287e1d789STejun Heo static struct bdi_writeback *inode_to_wb_and_lock_list(struct inode *inode) 78387e1d789STejun Heo __acquires(&wb->list_lock) 78487e1d789STejun Heo { 78587e1d789STejun Heo struct bdi_writeback *wb = inode_to_wb(inode); 78687e1d789STejun Heo 78787e1d789STejun Heo spin_lock(&wb->list_lock); 78887e1d789STejun Heo return wb; 78987e1d789STejun Heo } 79087e1d789STejun Heo 791f2b65121STejun Heo static long wb_split_bdi_pages(struct bdi_writeback *wb, long nr_pages) 792f2b65121STejun Heo { 793f2b65121STejun Heo return nr_pages; 794f2b65121STejun Heo } 795f2b65121STejun Heo 796db125360STejun Heo static void bdi_split_work_to_wbs(struct backing_dev_info *bdi, 797db125360STejun Heo struct wb_writeback_work *base_work, 798db125360STejun Heo bool skip_if_busy) 799db125360STejun Heo { 800db125360STejun Heo might_sleep(); 801db125360STejun Heo 802db125360STejun Heo if (bdi_has_dirty_io(bdi) && 803db125360STejun Heo (!skip_if_busy || !writeback_in_progress(&bdi->wb))) { 804db125360STejun Heo base_work->auto_free = 0; 805db125360STejun Heo base_work->single_wait = 0; 806db125360STejun Heo base_work->single_done = 0; 807db125360STejun Heo wb_queue_work(&bdi->wb, base_work); 808db125360STejun Heo } 809db125360STejun Heo } 810db125360STejun Heo 811703c2708STejun Heo #endif /* CONFIG_CGROUP_WRITEBACK */ 812703c2708STejun Heo 813c00ddad3STejun Heo void wb_start_writeback(struct bdi_writeback *wb, long nr_pages, 814c00ddad3STejun Heo bool range_cyclic, enum wb_reason reason) 815b6e51316SJens Axboe { 816c00ddad3STejun Heo struct wb_writeback_work *work; 817c00ddad3STejun Heo 818c00ddad3STejun Heo if (!wb_has_dirty_io(wb)) 819c00ddad3STejun Heo return; 820c00ddad3STejun Heo 821c00ddad3STejun Heo /* 822c00ddad3STejun Heo * This is WB_SYNC_NONE writeback, so if allocation fails just 823c00ddad3STejun Heo * wakeup the thread for old dirty data writeback 824c00ddad3STejun Heo */ 825c00ddad3STejun Heo work = kzalloc(sizeof(*work), GFP_ATOMIC); 826c00ddad3STejun Heo if (!work) { 827c00ddad3STejun Heo trace_writeback_nowork(wb->bdi); 828c00ddad3STejun Heo wb_wakeup(wb); 829c00ddad3STejun Heo return; 830c00ddad3STejun Heo } 831c00ddad3STejun Heo 832c00ddad3STejun Heo work->sync_mode = WB_SYNC_NONE; 833c00ddad3STejun Heo work->nr_pages = nr_pages; 834c00ddad3STejun Heo work->range_cyclic = range_cyclic; 835c00ddad3STejun Heo work->reason = reason; 836ac7b19a3STejun Heo work->auto_free = 1; 837c00ddad3STejun Heo 838c00ddad3STejun Heo wb_queue_work(wb, work); 839d3ddec76SWu Fengguang } 840d3ddec76SWu Fengguang 841c5444198SChristoph Hellwig /** 8429ecf4866STejun Heo * wb_start_background_writeback - start background writeback 8439ecf4866STejun Heo * @wb: bdi_writback to write from 844c5444198SChristoph Hellwig * 845c5444198SChristoph Hellwig * Description: 8466585027aSJan Kara * This makes sure WB_SYNC_NONE background writeback happens. When 8479ecf4866STejun Heo * this function returns, it is only guaranteed that for given wb 8486585027aSJan Kara * some IO is happening if we are over background dirty threshold. 8496585027aSJan Kara * Caller need not hold sb s_umount semaphore. 850c5444198SChristoph Hellwig */ 8519ecf4866STejun Heo void wb_start_background_writeback(struct bdi_writeback *wb) 852c5444198SChristoph Hellwig { 8536585027aSJan Kara /* 8546585027aSJan Kara * We just wake up the flusher thread. It will perform background 8556585027aSJan Kara * writeback as soon as there is no other work to do. 8566585027aSJan Kara */ 8579ecf4866STejun Heo trace_writeback_wake_background(wb->bdi); 8589ecf4866STejun Heo wb_wakeup(wb); 8591da177e4SLinus Torvalds } 8601da177e4SLinus Torvalds 8611da177e4SLinus Torvalds /* 862a66979abSDave Chinner * Remove the inode from the writeback list it is on. 863a66979abSDave Chinner */ 864a66979abSDave Chinner void inode_wb_list_del(struct inode *inode) 865a66979abSDave Chinner { 86687e1d789STejun Heo struct bdi_writeback *wb; 867a66979abSDave Chinner 86887e1d789STejun Heo wb = inode_to_wb_and_lock_list(inode); 869d6c10f1fSTejun Heo inode_wb_list_del_locked(inode, wb); 87052ebea74STejun Heo spin_unlock(&wb->list_lock); 871f758eeabSChristoph Hellwig } 872a66979abSDave Chinner 873a66979abSDave Chinner /* 8746610a0bcSAndrew Morton * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 8756610a0bcSAndrew Morton * furthest end of its superblock's dirty-inode list. 8766610a0bcSAndrew Morton * 8776610a0bcSAndrew Morton * Before stamping the inode's ->dirtied_when, we check to see whether it is 87866f3b8e2SJens Axboe * already the most-recently-dirtied inode on the b_dirty list. If that is 8796610a0bcSAndrew Morton * the case then the inode must have been redirtied while it was being written 8806610a0bcSAndrew Morton * out and we don't reset its dirtied_when. 8816610a0bcSAndrew Morton */ 882f758eeabSChristoph Hellwig static void redirty_tail(struct inode *inode, struct bdi_writeback *wb) 8836610a0bcSAndrew Morton { 88403ba3782SJens Axboe if (!list_empty(&wb->b_dirty)) { 88566f3b8e2SJens Axboe struct inode *tail; 8866610a0bcSAndrew Morton 8877ccf19a8SNick Piggin tail = wb_inode(wb->b_dirty.next); 88866f3b8e2SJens Axboe if (time_before(inode->dirtied_when, tail->dirtied_when)) 8896610a0bcSAndrew Morton inode->dirtied_when = jiffies; 8906610a0bcSAndrew Morton } 891d6c10f1fSTejun Heo inode_wb_list_move_locked(inode, wb, &wb->b_dirty); 8926610a0bcSAndrew Morton } 8936610a0bcSAndrew Morton 8946610a0bcSAndrew Morton /* 89566f3b8e2SJens Axboe * requeue inode for re-scanning after bdi->b_io list is exhausted. 896c986d1e2SAndrew Morton */ 897f758eeabSChristoph Hellwig static void requeue_io(struct inode *inode, struct bdi_writeback *wb) 898c986d1e2SAndrew Morton { 899d6c10f1fSTejun Heo inode_wb_list_move_locked(inode, wb, &wb->b_more_io); 900c986d1e2SAndrew Morton } 901c986d1e2SAndrew Morton 9021c0eeaf5SJoern Engel static void inode_sync_complete(struct inode *inode) 9031c0eeaf5SJoern Engel { 904365b94aeSJan Kara inode->i_state &= ~I_SYNC; 9054eff96ddSJan Kara /* If inode is clean an unused, put it into LRU now... */ 9064eff96ddSJan Kara inode_add_lru(inode); 907365b94aeSJan Kara /* Waiters must see I_SYNC cleared before being woken up */ 9081c0eeaf5SJoern Engel smp_mb(); 9091c0eeaf5SJoern Engel wake_up_bit(&inode->i_state, __I_SYNC); 9101c0eeaf5SJoern Engel } 9111c0eeaf5SJoern Engel 912d2caa3c5SJeff Layton static bool inode_dirtied_after(struct inode *inode, unsigned long t) 913d2caa3c5SJeff Layton { 914d2caa3c5SJeff Layton bool ret = time_after(inode->dirtied_when, t); 915d2caa3c5SJeff Layton #ifndef CONFIG_64BIT 916d2caa3c5SJeff Layton /* 917d2caa3c5SJeff Layton * For inodes being constantly redirtied, dirtied_when can get stuck. 918d2caa3c5SJeff Layton * It _appears_ to be in the future, but is actually in distant past. 919d2caa3c5SJeff Layton * This test is necessary to prevent such wrapped-around relative times 9205b0830cbSJens Axboe * from permanently stopping the whole bdi writeback. 921d2caa3c5SJeff Layton */ 922d2caa3c5SJeff Layton ret = ret && time_before_eq(inode->dirtied_when, jiffies); 923d2caa3c5SJeff Layton #endif 924d2caa3c5SJeff Layton return ret; 925d2caa3c5SJeff Layton } 926d2caa3c5SJeff Layton 9270ae45f63STheodore Ts'o #define EXPIRE_DIRTY_ATIME 0x0001 9280ae45f63STheodore Ts'o 929c986d1e2SAndrew Morton /* 9300e2f2b23SWang Sheng-Hui * Move expired (dirtied before work->older_than_this) dirty inodes from 931697e6fedSJan Kara * @delaying_queue to @dispatch_queue. 9322c136579SFengguang Wu */ 933e84d0a4fSWu Fengguang static int move_expired_inodes(struct list_head *delaying_queue, 9342c136579SFengguang Wu struct list_head *dispatch_queue, 9350ae45f63STheodore Ts'o int flags, 936ad4e38ddSCurt Wohlgemuth struct wb_writeback_work *work) 9372c136579SFengguang Wu { 9380ae45f63STheodore Ts'o unsigned long *older_than_this = NULL; 9390ae45f63STheodore Ts'o unsigned long expire_time; 9405c03449dSShaohua Li LIST_HEAD(tmp); 9415c03449dSShaohua Li struct list_head *pos, *node; 942cf137307SJens Axboe struct super_block *sb = NULL; 9435c03449dSShaohua Li struct inode *inode; 944cf137307SJens Axboe int do_sb_sort = 0; 945e84d0a4fSWu Fengguang int moved = 0; 9465c03449dSShaohua Li 9470ae45f63STheodore Ts'o if ((flags & EXPIRE_DIRTY_ATIME) == 0) 9480ae45f63STheodore Ts'o older_than_this = work->older_than_this; 949a2f48706STheodore Ts'o else if (!work->for_sync) { 950a2f48706STheodore Ts'o expire_time = jiffies - (dirtytime_expire_interval * HZ); 9510ae45f63STheodore Ts'o older_than_this = &expire_time; 9520ae45f63STheodore Ts'o } 9532c136579SFengguang Wu while (!list_empty(delaying_queue)) { 9547ccf19a8SNick Piggin inode = wb_inode(delaying_queue->prev); 9550ae45f63STheodore Ts'o if (older_than_this && 9560ae45f63STheodore Ts'o inode_dirtied_after(inode, *older_than_this)) 9572c136579SFengguang Wu break; 958a8855990SJan Kara list_move(&inode->i_wb_list, &tmp); 959a8855990SJan Kara moved++; 9600ae45f63STheodore Ts'o if (flags & EXPIRE_DIRTY_ATIME) 9610ae45f63STheodore Ts'o set_bit(__I_DIRTY_TIME_EXPIRED, &inode->i_state); 962a8855990SJan Kara if (sb_is_blkdev_sb(inode->i_sb)) 963a8855990SJan Kara continue; 964cf137307SJens Axboe if (sb && sb != inode->i_sb) 965cf137307SJens Axboe do_sb_sort = 1; 966cf137307SJens Axboe sb = inode->i_sb; 9675c03449dSShaohua Li } 9685c03449dSShaohua Li 969cf137307SJens Axboe /* just one sb in list, splice to dispatch_queue and we're done */ 970cf137307SJens Axboe if (!do_sb_sort) { 971cf137307SJens Axboe list_splice(&tmp, dispatch_queue); 972e84d0a4fSWu Fengguang goto out; 973cf137307SJens Axboe } 974cf137307SJens Axboe 9755c03449dSShaohua Li /* Move inodes from one superblock together */ 9765c03449dSShaohua Li while (!list_empty(&tmp)) { 9777ccf19a8SNick Piggin sb = wb_inode(tmp.prev)->i_sb; 9785c03449dSShaohua Li list_for_each_prev_safe(pos, node, &tmp) { 9797ccf19a8SNick Piggin inode = wb_inode(pos); 9805c03449dSShaohua Li if (inode->i_sb == sb) 9817ccf19a8SNick Piggin list_move(&inode->i_wb_list, dispatch_queue); 9822c136579SFengguang Wu } 9832c136579SFengguang Wu } 984e84d0a4fSWu Fengguang out: 985e84d0a4fSWu Fengguang return moved; 9865c03449dSShaohua Li } 9872c136579SFengguang Wu 9882c136579SFengguang Wu /* 9892c136579SFengguang Wu * Queue all expired dirty inodes for io, eldest first. 9904ea879b9SWu Fengguang * Before 9914ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 9924ea879b9SWu Fengguang * =============> gf edc BA 9934ea879b9SWu Fengguang * After 9944ea879b9SWu Fengguang * newly dirtied b_dirty b_io b_more_io 9954ea879b9SWu Fengguang * =============> g fBAedc 9964ea879b9SWu Fengguang * | 9974ea879b9SWu Fengguang * +--> dequeue for IO 9982c136579SFengguang Wu */ 999ad4e38ddSCurt Wohlgemuth static void queue_io(struct bdi_writeback *wb, struct wb_writeback_work *work) 10002c136579SFengguang Wu { 1001e84d0a4fSWu Fengguang int moved; 10020ae45f63STheodore Ts'o 1003f758eeabSChristoph Hellwig assert_spin_locked(&wb->list_lock); 10044ea879b9SWu Fengguang list_splice_init(&wb->b_more_io, &wb->b_io); 10050ae45f63STheodore Ts'o moved = move_expired_inodes(&wb->b_dirty, &wb->b_io, 0, work); 10060ae45f63STheodore Ts'o moved += move_expired_inodes(&wb->b_dirty_time, &wb->b_io, 10070ae45f63STheodore Ts'o EXPIRE_DIRTY_ATIME, work); 1008d6c10f1fSTejun Heo if (moved) 1009d6c10f1fSTejun Heo wb_io_lists_populated(wb); 1010ad4e38ddSCurt Wohlgemuth trace_writeback_queue_io(wb, work, moved); 101166f3b8e2SJens Axboe } 101266f3b8e2SJens Axboe 1013a9185b41SChristoph Hellwig static int write_inode(struct inode *inode, struct writeback_control *wbc) 101466f3b8e2SJens Axboe { 10159fb0a7daSTejun Heo int ret; 10169fb0a7daSTejun Heo 10179fb0a7daSTejun Heo if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) { 10189fb0a7daSTejun Heo trace_writeback_write_inode_start(inode, wbc); 10199fb0a7daSTejun Heo ret = inode->i_sb->s_op->write_inode(inode, wbc); 10209fb0a7daSTejun Heo trace_writeback_write_inode(inode, wbc); 10219fb0a7daSTejun Heo return ret; 10229fb0a7daSTejun Heo } 102303ba3782SJens Axboe return 0; 102466f3b8e2SJens Axboe } 102508d8e974SFengguang Wu 10262c136579SFengguang Wu /* 1027169ebd90SJan Kara * Wait for writeback on an inode to complete. Called with i_lock held. 1028169ebd90SJan Kara * Caller must make sure inode cannot go away when we drop i_lock. 102901c03194SChristoph Hellwig */ 1030169ebd90SJan Kara static void __inode_wait_for_writeback(struct inode *inode) 1031169ebd90SJan Kara __releases(inode->i_lock) 1032169ebd90SJan Kara __acquires(inode->i_lock) 103301c03194SChristoph Hellwig { 103401c03194SChristoph Hellwig DEFINE_WAIT_BIT(wq, &inode->i_state, __I_SYNC); 103501c03194SChristoph Hellwig wait_queue_head_t *wqh; 103601c03194SChristoph Hellwig 103701c03194SChristoph Hellwig wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 103858a9d3d8SRichard Kennedy while (inode->i_state & I_SYNC) { 1039250df6edSDave Chinner spin_unlock(&inode->i_lock); 104074316201SNeilBrown __wait_on_bit(wqh, &wq, bit_wait, 104174316201SNeilBrown TASK_UNINTERRUPTIBLE); 1042250df6edSDave Chinner spin_lock(&inode->i_lock); 104358a9d3d8SRichard Kennedy } 104401c03194SChristoph Hellwig } 104501c03194SChristoph Hellwig 104601c03194SChristoph Hellwig /* 1047169ebd90SJan Kara * Wait for writeback on an inode to complete. Caller must have inode pinned. 1048169ebd90SJan Kara */ 1049169ebd90SJan Kara void inode_wait_for_writeback(struct inode *inode) 1050169ebd90SJan Kara { 1051169ebd90SJan Kara spin_lock(&inode->i_lock); 1052169ebd90SJan Kara __inode_wait_for_writeback(inode); 1053169ebd90SJan Kara spin_unlock(&inode->i_lock); 1054169ebd90SJan Kara } 1055169ebd90SJan Kara 1056169ebd90SJan Kara /* 1057169ebd90SJan Kara * Sleep until I_SYNC is cleared. This function must be called with i_lock 1058169ebd90SJan Kara * held and drops it. It is aimed for callers not holding any inode reference 1059169ebd90SJan Kara * so once i_lock is dropped, inode can go away. 1060169ebd90SJan Kara */ 1061169ebd90SJan Kara static void inode_sleep_on_writeback(struct inode *inode) 1062169ebd90SJan Kara __releases(inode->i_lock) 1063169ebd90SJan Kara { 1064169ebd90SJan Kara DEFINE_WAIT(wait); 1065169ebd90SJan Kara wait_queue_head_t *wqh = bit_waitqueue(&inode->i_state, __I_SYNC); 1066169ebd90SJan Kara int sleep; 1067169ebd90SJan Kara 1068169ebd90SJan Kara prepare_to_wait(wqh, &wait, TASK_UNINTERRUPTIBLE); 1069169ebd90SJan Kara sleep = inode->i_state & I_SYNC; 1070169ebd90SJan Kara spin_unlock(&inode->i_lock); 1071169ebd90SJan Kara if (sleep) 1072169ebd90SJan Kara schedule(); 1073169ebd90SJan Kara finish_wait(wqh, &wait); 1074169ebd90SJan Kara } 1075169ebd90SJan Kara 1076169ebd90SJan Kara /* 1077ccb26b5aSJan Kara * Find proper writeback list for the inode depending on its current state and 1078ccb26b5aSJan Kara * possibly also change of its state while we were doing writeback. Here we 1079ccb26b5aSJan Kara * handle things such as livelock prevention or fairness of writeback among 1080ccb26b5aSJan Kara * inodes. This function can be called only by flusher thread - noone else 1081ccb26b5aSJan Kara * processes all inodes in writeback lists and requeueing inodes behind flusher 1082ccb26b5aSJan Kara * thread's back can have unexpected consequences. 1083ccb26b5aSJan Kara */ 1084ccb26b5aSJan Kara static void requeue_inode(struct inode *inode, struct bdi_writeback *wb, 1085ccb26b5aSJan Kara struct writeback_control *wbc) 1086ccb26b5aSJan Kara { 1087ccb26b5aSJan Kara if (inode->i_state & I_FREEING) 1088ccb26b5aSJan Kara return; 1089ccb26b5aSJan Kara 1090ccb26b5aSJan Kara /* 1091ccb26b5aSJan Kara * Sync livelock prevention. Each inode is tagged and synced in one 1092ccb26b5aSJan Kara * shot. If still dirty, it will be redirty_tail()'ed below. Update 1093ccb26b5aSJan Kara * the dirty time to prevent enqueue and sync it again. 1094ccb26b5aSJan Kara */ 1095ccb26b5aSJan Kara if ((inode->i_state & I_DIRTY) && 1096ccb26b5aSJan Kara (wbc->sync_mode == WB_SYNC_ALL || wbc->tagged_writepages)) 1097ccb26b5aSJan Kara inode->dirtied_when = jiffies; 1098ccb26b5aSJan Kara 10994f8ad655SJan Kara if (wbc->pages_skipped) { 11004f8ad655SJan Kara /* 11014f8ad655SJan Kara * writeback is not making progress due to locked 11024f8ad655SJan Kara * buffers. Skip this inode for now. 11034f8ad655SJan Kara */ 11044f8ad655SJan Kara redirty_tail(inode, wb); 11054f8ad655SJan Kara return; 11064f8ad655SJan Kara } 11074f8ad655SJan Kara 1108ccb26b5aSJan Kara if (mapping_tagged(inode->i_mapping, PAGECACHE_TAG_DIRTY)) { 1109ccb26b5aSJan Kara /* 1110ccb26b5aSJan Kara * We didn't write back all the pages. nfs_writepages() 1111ccb26b5aSJan Kara * sometimes bales out without doing anything. 1112ccb26b5aSJan Kara */ 1113ccb26b5aSJan Kara if (wbc->nr_to_write <= 0) { 1114ccb26b5aSJan Kara /* Slice used up. Queue for next turn. */ 1115ccb26b5aSJan Kara requeue_io(inode, wb); 1116ccb26b5aSJan Kara } else { 1117ccb26b5aSJan Kara /* 1118ccb26b5aSJan Kara * Writeback blocked by something other than 1119ccb26b5aSJan Kara * congestion. Delay the inode for some time to 1120ccb26b5aSJan Kara * avoid spinning on the CPU (100% iowait) 1121ccb26b5aSJan Kara * retrying writeback of the dirty page/inode 1122ccb26b5aSJan Kara * that cannot be performed immediately. 1123ccb26b5aSJan Kara */ 1124ccb26b5aSJan Kara redirty_tail(inode, wb); 1125ccb26b5aSJan Kara } 1126ccb26b5aSJan Kara } else if (inode->i_state & I_DIRTY) { 1127ccb26b5aSJan Kara /* 1128ccb26b5aSJan Kara * Filesystems can dirty the inode during writeback operations, 1129ccb26b5aSJan Kara * such as delayed allocation during submission or metadata 1130ccb26b5aSJan Kara * updates after data IO completion. 1131ccb26b5aSJan Kara */ 1132ccb26b5aSJan Kara redirty_tail(inode, wb); 11330ae45f63STheodore Ts'o } else if (inode->i_state & I_DIRTY_TIME) { 1134a2f48706STheodore Ts'o inode->dirtied_when = jiffies; 1135d6c10f1fSTejun Heo inode_wb_list_move_locked(inode, wb, &wb->b_dirty_time); 1136ccb26b5aSJan Kara } else { 1137ccb26b5aSJan Kara /* The inode is clean. Remove from writeback lists. */ 1138d6c10f1fSTejun Heo inode_wb_list_del_locked(inode, wb); 1139ccb26b5aSJan Kara } 1140ccb26b5aSJan Kara } 1141ccb26b5aSJan Kara 1142ccb26b5aSJan Kara /* 11434f8ad655SJan Kara * Write out an inode and its dirty pages. Do not update the writeback list 11444f8ad655SJan Kara * linkage. That is left to the caller. The caller is also responsible for 11454f8ad655SJan Kara * setting I_SYNC flag and calling inode_sync_complete() to clear it. 11461da177e4SLinus Torvalds */ 11471da177e4SLinus Torvalds static int 1148cd8ed2a4SYan Hong __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 11491da177e4SLinus Torvalds { 11501da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 1151251d6a47SWu Fengguang long nr_to_write = wbc->nr_to_write; 115201c03194SChristoph Hellwig unsigned dirty; 11531da177e4SLinus Torvalds int ret; 11541da177e4SLinus Torvalds 11554f8ad655SJan Kara WARN_ON(!(inode->i_state & I_SYNC)); 11561da177e4SLinus Torvalds 11579fb0a7daSTejun Heo trace_writeback_single_inode_start(inode, wbc, nr_to_write); 11589fb0a7daSTejun Heo 11591da177e4SLinus Torvalds ret = do_writepages(mapping, wbc); 11601da177e4SLinus Torvalds 116126821ed4SChristoph Hellwig /* 116226821ed4SChristoph Hellwig * Make sure to wait on the data before writing out the metadata. 116326821ed4SChristoph Hellwig * This is important for filesystems that modify metadata on data 11647747bd4bSDave Chinner * I/O completion. We don't do it for sync(2) writeback because it has a 11657747bd4bSDave Chinner * separate, external IO completion path and ->sync_fs for guaranteeing 11667747bd4bSDave Chinner * inode metadata is written back correctly. 116726821ed4SChristoph Hellwig */ 11687747bd4bSDave Chinner if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync) { 116926821ed4SChristoph Hellwig int err = filemap_fdatawait(mapping); 11701da177e4SLinus Torvalds if (ret == 0) 11711da177e4SLinus Torvalds ret = err; 11721da177e4SLinus Torvalds } 11731da177e4SLinus Torvalds 11745547e8aaSDmitry Monakhov /* 11755547e8aaSDmitry Monakhov * Some filesystems may redirty the inode during the writeback 11765547e8aaSDmitry Monakhov * due to delalloc, clear dirty metadata flags right before 11775547e8aaSDmitry Monakhov * write_inode() 11785547e8aaSDmitry Monakhov */ 1179250df6edSDave Chinner spin_lock(&inode->i_lock); 11809c6ac78eSTejun Heo 11815547e8aaSDmitry Monakhov dirty = inode->i_state & I_DIRTY; 1182a2f48706STheodore Ts'o if (inode->i_state & I_DIRTY_TIME) { 1183a2f48706STheodore Ts'o if ((dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) || 1184a2f48706STheodore Ts'o unlikely(inode->i_state & I_DIRTY_TIME_EXPIRED) || 1185a2f48706STheodore Ts'o unlikely(time_after(jiffies, 1186a2f48706STheodore Ts'o (inode->dirtied_time_when + 1187a2f48706STheodore Ts'o dirtytime_expire_interval * HZ)))) { 11880ae45f63STheodore Ts'o dirty |= I_DIRTY_TIME | I_DIRTY_TIME_EXPIRED; 11890ae45f63STheodore Ts'o trace_writeback_lazytime(inode); 11900ae45f63STheodore Ts'o } 1191a2f48706STheodore Ts'o } else 1192a2f48706STheodore Ts'o inode->i_state &= ~I_DIRTY_TIME_EXPIRED; 11930ae45f63STheodore Ts'o inode->i_state &= ~dirty; 11949c6ac78eSTejun Heo 11959c6ac78eSTejun Heo /* 11969c6ac78eSTejun Heo * Paired with smp_mb() in __mark_inode_dirty(). This allows 11979c6ac78eSTejun Heo * __mark_inode_dirty() to test i_state without grabbing i_lock - 11989c6ac78eSTejun Heo * either they see the I_DIRTY bits cleared or we see the dirtied 11999c6ac78eSTejun Heo * inode. 12009c6ac78eSTejun Heo * 12019c6ac78eSTejun Heo * I_DIRTY_PAGES is always cleared together above even if @mapping 12029c6ac78eSTejun Heo * still has dirty pages. The flag is reinstated after smp_mb() if 12039c6ac78eSTejun Heo * necessary. This guarantees that either __mark_inode_dirty() 12049c6ac78eSTejun Heo * sees clear I_DIRTY_PAGES or we see PAGECACHE_TAG_DIRTY. 12059c6ac78eSTejun Heo */ 12069c6ac78eSTejun Heo smp_mb(); 12079c6ac78eSTejun Heo 12089c6ac78eSTejun Heo if (mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) 12099c6ac78eSTejun Heo inode->i_state |= I_DIRTY_PAGES; 12109c6ac78eSTejun Heo 1211250df6edSDave Chinner spin_unlock(&inode->i_lock); 12129c6ac78eSTejun Heo 12130ae45f63STheodore Ts'o if (dirty & I_DIRTY_TIME) 12140ae45f63STheodore Ts'o mark_inode_dirty_sync(inode); 121526821ed4SChristoph Hellwig /* Don't write the inode if only I_DIRTY_PAGES was set */ 12160ae45f63STheodore Ts'o if (dirty & ~I_DIRTY_PAGES) { 1217a9185b41SChristoph Hellwig int err = write_inode(inode, wbc); 12181da177e4SLinus Torvalds if (ret == 0) 12191da177e4SLinus Torvalds ret = err; 12201da177e4SLinus Torvalds } 12214f8ad655SJan Kara trace_writeback_single_inode(inode, wbc, nr_to_write); 12224f8ad655SJan Kara return ret; 12234f8ad655SJan Kara } 12244f8ad655SJan Kara 12254f8ad655SJan Kara /* 12264f8ad655SJan Kara * Write out an inode's dirty pages. Either the caller has an active reference 12274f8ad655SJan Kara * on the inode or the inode has I_WILL_FREE set. 12284f8ad655SJan Kara * 12294f8ad655SJan Kara * This function is designed to be called for writing back one inode which 12304f8ad655SJan Kara * we go e.g. from filesystem. Flusher thread uses __writeback_single_inode() 12314f8ad655SJan Kara * and does more profound writeback list handling in writeback_sb_inodes(). 12324f8ad655SJan Kara */ 12334f8ad655SJan Kara static int 12344f8ad655SJan Kara writeback_single_inode(struct inode *inode, struct bdi_writeback *wb, 12354f8ad655SJan Kara struct writeback_control *wbc) 12364f8ad655SJan Kara { 12374f8ad655SJan Kara int ret = 0; 12384f8ad655SJan Kara 12394f8ad655SJan Kara spin_lock(&inode->i_lock); 12404f8ad655SJan Kara if (!atomic_read(&inode->i_count)) 12414f8ad655SJan Kara WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 12424f8ad655SJan Kara else 12434f8ad655SJan Kara WARN_ON(inode->i_state & I_WILL_FREE); 12444f8ad655SJan Kara 12454f8ad655SJan Kara if (inode->i_state & I_SYNC) { 12464f8ad655SJan Kara if (wbc->sync_mode != WB_SYNC_ALL) 12474f8ad655SJan Kara goto out; 12484f8ad655SJan Kara /* 1249169ebd90SJan Kara * It's a data-integrity sync. We must wait. Since callers hold 1250169ebd90SJan Kara * inode reference or inode has I_WILL_FREE set, it cannot go 1251169ebd90SJan Kara * away under us. 12524f8ad655SJan Kara */ 1253169ebd90SJan Kara __inode_wait_for_writeback(inode); 12544f8ad655SJan Kara } 12554f8ad655SJan Kara WARN_ON(inode->i_state & I_SYNC); 12564f8ad655SJan Kara /* 1257f9b0e058SJan Kara * Skip inode if it is clean and we have no outstanding writeback in 1258f9b0e058SJan Kara * WB_SYNC_ALL mode. We don't want to mess with writeback lists in this 1259f9b0e058SJan Kara * function since flusher thread may be doing for example sync in 1260f9b0e058SJan Kara * parallel and if we move the inode, it could get skipped. So here we 1261f9b0e058SJan Kara * make sure inode is on some writeback list and leave it there unless 1262f9b0e058SJan Kara * we have completely cleaned the inode. 12634f8ad655SJan Kara */ 12640ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL) && 1265f9b0e058SJan Kara (wbc->sync_mode != WB_SYNC_ALL || 1266f9b0e058SJan Kara !mapping_tagged(inode->i_mapping, PAGECACHE_TAG_WRITEBACK))) 12674f8ad655SJan Kara goto out; 12684f8ad655SJan Kara inode->i_state |= I_SYNC; 1269b16b1debSTejun Heo wbc_attach_and_unlock_inode(wbc, inode); 12704f8ad655SJan Kara 1271cd8ed2a4SYan Hong ret = __writeback_single_inode(inode, wbc); 12721da177e4SLinus Torvalds 1273b16b1debSTejun Heo wbc_detach_inode(wbc); 1274f758eeabSChristoph Hellwig spin_lock(&wb->list_lock); 1275250df6edSDave Chinner spin_lock(&inode->i_lock); 12764f8ad655SJan Kara /* 12774f8ad655SJan Kara * If inode is clean, remove it from writeback lists. Otherwise don't 12784f8ad655SJan Kara * touch it. See comment above for explanation. 12794f8ad655SJan Kara */ 12800ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1281d6c10f1fSTejun Heo inode_wb_list_del_locked(inode, wb); 12824f8ad655SJan Kara spin_unlock(&wb->list_lock); 12831c0eeaf5SJoern Engel inode_sync_complete(inode); 12844f8ad655SJan Kara out: 12854f8ad655SJan Kara spin_unlock(&inode->i_lock); 12861da177e4SLinus Torvalds return ret; 12871da177e4SLinus Torvalds } 12881da177e4SLinus Torvalds 1289a88a341aSTejun Heo static long writeback_chunk_size(struct bdi_writeback *wb, 12901a12d8bdSWu Fengguang struct wb_writeback_work *work) 1291d46db3d5SWu Fengguang { 1292d46db3d5SWu Fengguang long pages; 1293d46db3d5SWu Fengguang 1294d46db3d5SWu Fengguang /* 1295d46db3d5SWu Fengguang * WB_SYNC_ALL mode does livelock avoidance by syncing dirty 1296d46db3d5SWu Fengguang * inodes/pages in one big loop. Setting wbc.nr_to_write=LONG_MAX 1297d46db3d5SWu Fengguang * here avoids calling into writeback_inodes_wb() more than once. 1298d46db3d5SWu Fengguang * 1299d46db3d5SWu Fengguang * The intended call sequence for WB_SYNC_ALL writeback is: 1300d46db3d5SWu Fengguang * 1301d46db3d5SWu Fengguang * wb_writeback() 1302d46db3d5SWu Fengguang * writeback_sb_inodes() <== called only once 1303d46db3d5SWu Fengguang * write_cache_pages() <== called once for each inode 1304d46db3d5SWu Fengguang * (quickly) tag currently dirty pages 1305d46db3d5SWu Fengguang * (maybe slowly) sync all tagged pages 1306d46db3d5SWu Fengguang */ 1307d46db3d5SWu Fengguang if (work->sync_mode == WB_SYNC_ALL || work->tagged_writepages) 1308d46db3d5SWu Fengguang pages = LONG_MAX; 13091a12d8bdSWu Fengguang else { 1310a88a341aSTejun Heo pages = min(wb->avg_write_bandwidth / 2, 1311dcc25ae7STejun Heo global_wb_domain.dirty_limit / DIRTY_SCOPE); 13121a12d8bdSWu Fengguang pages = min(pages, work->nr_pages); 13131a12d8bdSWu Fengguang pages = round_down(pages + MIN_WRITEBACK_PAGES, 13141a12d8bdSWu Fengguang MIN_WRITEBACK_PAGES); 13151a12d8bdSWu Fengguang } 1316d46db3d5SWu Fengguang 1317d46db3d5SWu Fengguang return pages; 1318d46db3d5SWu Fengguang } 1319d46db3d5SWu Fengguang 132003ba3782SJens Axboe /* 1321f11c9c5cSEdward Shishkin * Write a portion of b_io inodes which belong to @sb. 1322edadfb10SChristoph Hellwig * 1323d46db3d5SWu Fengguang * Return the number of pages and/or inodes written. 1324f11c9c5cSEdward Shishkin */ 1325d46db3d5SWu Fengguang static long writeback_sb_inodes(struct super_block *sb, 1326d46db3d5SWu Fengguang struct bdi_writeback *wb, 1327d46db3d5SWu Fengguang struct wb_writeback_work *work) 132803ba3782SJens Axboe { 1329d46db3d5SWu Fengguang struct writeback_control wbc = { 1330d46db3d5SWu Fengguang .sync_mode = work->sync_mode, 1331d46db3d5SWu Fengguang .tagged_writepages = work->tagged_writepages, 1332d46db3d5SWu Fengguang .for_kupdate = work->for_kupdate, 1333d46db3d5SWu Fengguang .for_background = work->for_background, 13347747bd4bSDave Chinner .for_sync = work->for_sync, 1335d46db3d5SWu Fengguang .range_cyclic = work->range_cyclic, 1336d46db3d5SWu Fengguang .range_start = 0, 1337d46db3d5SWu Fengguang .range_end = LLONG_MAX, 1338d46db3d5SWu Fengguang }; 1339d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1340d46db3d5SWu Fengguang long write_chunk; 1341d46db3d5SWu Fengguang long wrote = 0; /* count both pages and inodes */ 1342d46db3d5SWu Fengguang 134303ba3782SJens Axboe while (!list_empty(&wb->b_io)) { 13447ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1345edadfb10SChristoph Hellwig 1346edadfb10SChristoph Hellwig if (inode->i_sb != sb) { 1347d46db3d5SWu Fengguang if (work->sb) { 1348edadfb10SChristoph Hellwig /* 1349edadfb10SChristoph Hellwig * We only want to write back data for this 1350edadfb10SChristoph Hellwig * superblock, move all inodes not belonging 1351edadfb10SChristoph Hellwig * to it back onto the dirty list. 1352edadfb10SChristoph Hellwig */ 1353f758eeabSChristoph Hellwig redirty_tail(inode, wb); 135466f3b8e2SJens Axboe continue; 135566f3b8e2SJens Axboe } 1356edadfb10SChristoph Hellwig 1357edadfb10SChristoph Hellwig /* 1358edadfb10SChristoph Hellwig * The inode belongs to a different superblock. 1359edadfb10SChristoph Hellwig * Bounce back to the caller to unpin this and 1360edadfb10SChristoph Hellwig * pin the next superblock. 1361edadfb10SChristoph Hellwig */ 1362d46db3d5SWu Fengguang break; 1363edadfb10SChristoph Hellwig } 1364edadfb10SChristoph Hellwig 13659843b76aSChristoph Hellwig /* 1366331cbdeeSWanpeng Li * Don't bother with new inodes or inodes being freed, first 1367331cbdeeSWanpeng Li * kind does not need periodic writeout yet, and for the latter 13689843b76aSChristoph Hellwig * kind writeout is handled by the freer. 13699843b76aSChristoph Hellwig */ 1370250df6edSDave Chinner spin_lock(&inode->i_lock); 13719843b76aSChristoph Hellwig if (inode->i_state & (I_NEW | I_FREEING | I_WILL_FREE)) { 1372250df6edSDave Chinner spin_unlock(&inode->i_lock); 1373fcc5c222SWu Fengguang redirty_tail(inode, wb); 13747ef0d737SNick Piggin continue; 13757ef0d737SNick Piggin } 1376cc1676d9SJan Kara if ((inode->i_state & I_SYNC) && wbc.sync_mode != WB_SYNC_ALL) { 1377cc1676d9SJan Kara /* 1378cc1676d9SJan Kara * If this inode is locked for writeback and we are not 1379cc1676d9SJan Kara * doing writeback-for-data-integrity, move it to 1380cc1676d9SJan Kara * b_more_io so that writeback can proceed with the 1381cc1676d9SJan Kara * other inodes on s_io. 1382cc1676d9SJan Kara * 1383cc1676d9SJan Kara * We'll have another go at writing back this inode 1384cc1676d9SJan Kara * when we completed a full scan of b_io. 1385cc1676d9SJan Kara */ 1386cc1676d9SJan Kara spin_unlock(&inode->i_lock); 1387cc1676d9SJan Kara requeue_io(inode, wb); 1388cc1676d9SJan Kara trace_writeback_sb_inodes_requeue(inode); 1389cc1676d9SJan Kara continue; 1390cc1676d9SJan Kara } 1391f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 1392f0d07b7fSJan Kara 13934f8ad655SJan Kara /* 13944f8ad655SJan Kara * We already requeued the inode if it had I_SYNC set and we 13954f8ad655SJan Kara * are doing WB_SYNC_NONE writeback. So this catches only the 13964f8ad655SJan Kara * WB_SYNC_ALL case. 13974f8ad655SJan Kara */ 1398169ebd90SJan Kara if (inode->i_state & I_SYNC) { 1399169ebd90SJan Kara /* Wait for I_SYNC. This function drops i_lock... */ 1400169ebd90SJan Kara inode_sleep_on_writeback(inode); 1401169ebd90SJan Kara /* Inode may be gone, start again */ 1402ead188f9SJan Kara spin_lock(&wb->list_lock); 1403169ebd90SJan Kara continue; 1404169ebd90SJan Kara } 14054f8ad655SJan Kara inode->i_state |= I_SYNC; 1406b16b1debSTejun Heo wbc_attach_and_unlock_inode(&wbc, inode); 1407169ebd90SJan Kara 1408a88a341aSTejun Heo write_chunk = writeback_chunk_size(wb, work); 1409d46db3d5SWu Fengguang wbc.nr_to_write = write_chunk; 1410d46db3d5SWu Fengguang wbc.pages_skipped = 0; 1411250df6edSDave Chinner 1412169ebd90SJan Kara /* 1413169ebd90SJan Kara * We use I_SYNC to pin the inode in memory. While it is set 1414169ebd90SJan Kara * evict_inode() will wait so the inode cannot be freed. 1415169ebd90SJan Kara */ 1416cd8ed2a4SYan Hong __writeback_single_inode(inode, &wbc); 1417d46db3d5SWu Fengguang 1418b16b1debSTejun Heo wbc_detach_inode(&wbc); 1419d46db3d5SWu Fengguang work->nr_pages -= write_chunk - wbc.nr_to_write; 1420d46db3d5SWu Fengguang wrote += write_chunk - wbc.nr_to_write; 14214f8ad655SJan Kara spin_lock(&wb->list_lock); 14224f8ad655SJan Kara spin_lock(&inode->i_lock); 14230ae45f63STheodore Ts'o if (!(inode->i_state & I_DIRTY_ALL)) 1424d46db3d5SWu Fengguang wrote++; 14254f8ad655SJan Kara requeue_inode(inode, wb, &wbc); 14264f8ad655SJan Kara inode_sync_complete(inode); 14270f1b1fd8SDave Chinner spin_unlock(&inode->i_lock); 1428169ebd90SJan Kara cond_resched_lock(&wb->list_lock); 1429d46db3d5SWu Fengguang /* 1430d46db3d5SWu Fengguang * bail out to wb_writeback() often enough to check 1431d46db3d5SWu Fengguang * background threshold and other termination conditions. 1432d46db3d5SWu Fengguang */ 1433d46db3d5SWu Fengguang if (wrote) { 1434d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1435d46db3d5SWu Fengguang break; 1436d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1437d46db3d5SWu Fengguang break; 14381da177e4SLinus Torvalds } 14398bc3be27SFengguang Wu } 1440d46db3d5SWu Fengguang return wrote; 1441f11c9c5cSEdward Shishkin } 144238f21977SNick Piggin 1443d46db3d5SWu Fengguang static long __writeback_inodes_wb(struct bdi_writeback *wb, 1444d46db3d5SWu Fengguang struct wb_writeback_work *work) 1445f11c9c5cSEdward Shishkin { 1446d46db3d5SWu Fengguang unsigned long start_time = jiffies; 1447d46db3d5SWu Fengguang long wrote = 0; 1448f11c9c5cSEdward Shishkin 1449f11c9c5cSEdward Shishkin while (!list_empty(&wb->b_io)) { 14507ccf19a8SNick Piggin struct inode *inode = wb_inode(wb->b_io.prev); 1451f11c9c5cSEdward Shishkin struct super_block *sb = inode->i_sb; 1452f11c9c5cSEdward Shishkin 1453eb6ef3dfSKonstantin Khlebnikov if (!trylock_super(sb)) { 14540e995816SWu Fengguang /* 1455eb6ef3dfSKonstantin Khlebnikov * trylock_super() may fail consistently due to 14560e995816SWu Fengguang * s_umount being grabbed by someone else. Don't use 14570e995816SWu Fengguang * requeue_io() to avoid busy retrying the inode/sb. 14580e995816SWu Fengguang */ 14590e995816SWu Fengguang redirty_tail(inode, wb); 1460d19de7edSChristoph Hellwig continue; 1461334132aeSChristoph Hellwig } 1462d46db3d5SWu Fengguang wrote += writeback_sb_inodes(sb, wb, work); 1463eb6ef3dfSKonstantin Khlebnikov up_read(&sb->s_umount); 1464f11c9c5cSEdward Shishkin 1465d46db3d5SWu Fengguang /* refer to the same tests at the end of writeback_sb_inodes */ 1466d46db3d5SWu Fengguang if (wrote) { 1467d46db3d5SWu Fengguang if (time_is_before_jiffies(start_time + HZ / 10UL)) 1468d46db3d5SWu Fengguang break; 1469d46db3d5SWu Fengguang if (work->nr_pages <= 0) 1470f11c9c5cSEdward Shishkin break; 1471f11c9c5cSEdward Shishkin } 1472d46db3d5SWu Fengguang } 147366f3b8e2SJens Axboe /* Leave any unwritten inodes on b_io */ 1474d46db3d5SWu Fengguang return wrote; 147566f3b8e2SJens Axboe } 147666f3b8e2SJens Axboe 14777d9f073bSWanpeng Li static long writeback_inodes_wb(struct bdi_writeback *wb, long nr_pages, 14780e175a18SCurt Wohlgemuth enum wb_reason reason) 1479edadfb10SChristoph Hellwig { 1480d46db3d5SWu Fengguang struct wb_writeback_work work = { 1481d46db3d5SWu Fengguang .nr_pages = nr_pages, 1482d46db3d5SWu Fengguang .sync_mode = WB_SYNC_NONE, 1483d46db3d5SWu Fengguang .range_cyclic = 1, 14840e175a18SCurt Wohlgemuth .reason = reason, 1485d46db3d5SWu Fengguang }; 1486edadfb10SChristoph Hellwig 1487f758eeabSChristoph Hellwig spin_lock(&wb->list_lock); 1488424b351fSWu Fengguang if (list_empty(&wb->b_io)) 1489ad4e38ddSCurt Wohlgemuth queue_io(wb, &work); 1490d46db3d5SWu Fengguang __writeback_inodes_wb(wb, &work); 1491f758eeabSChristoph Hellwig spin_unlock(&wb->list_lock); 1492edadfb10SChristoph Hellwig 1493d46db3d5SWu Fengguang return nr_pages - work.nr_pages; 149466f3b8e2SJens Axboe } 149566f3b8e2SJens Axboe 149603ba3782SJens Axboe /* 149703ba3782SJens Axboe * Explicit flushing or periodic writeback of "old" data. 149803ba3782SJens Axboe * 149903ba3782SJens Axboe * Define "old": the first time one of an inode's pages is dirtied, we mark the 150003ba3782SJens Axboe * dirtying-time in the inode's address_space. So this periodic writeback code 150103ba3782SJens Axboe * just walks the superblock inode list, writing back any inodes which are 150203ba3782SJens Axboe * older than a specific point in time. 150303ba3782SJens Axboe * 150403ba3782SJens Axboe * Try to run once per dirty_writeback_interval. But if a writeback event 150503ba3782SJens Axboe * takes longer than a dirty_writeback_interval interval, then leave a 150603ba3782SJens Axboe * one-second gap. 150703ba3782SJens Axboe * 150803ba3782SJens Axboe * older_than_this takes precedence over nr_to_write. So we'll only write back 150903ba3782SJens Axboe * all dirty pages if they are all attached to "old" mappings. 151003ba3782SJens Axboe */ 1511c4a77a6cSJens Axboe static long wb_writeback(struct bdi_writeback *wb, 151283ba7b07SChristoph Hellwig struct wb_writeback_work *work) 151303ba3782SJens Axboe { 1514e98be2d5SWu Fengguang unsigned long wb_start = jiffies; 1515d46db3d5SWu Fengguang long nr_pages = work->nr_pages; 15160dc83bd3SJan Kara unsigned long oldest_jif; 1517a5989bdcSJan Kara struct inode *inode; 1518d46db3d5SWu Fengguang long progress; 151903ba3782SJens Axboe 15200dc83bd3SJan Kara oldest_jif = jiffies; 15210dc83bd3SJan Kara work->older_than_this = &oldest_jif; 152203ba3782SJens Axboe 1523e8dfc305SWu Fengguang spin_lock(&wb->list_lock); 152403ba3782SJens Axboe for (;;) { 152503ba3782SJens Axboe /* 1526d3ddec76SWu Fengguang * Stop writeback when nr_pages has been consumed 152703ba3782SJens Axboe */ 152883ba7b07SChristoph Hellwig if (work->nr_pages <= 0) 152903ba3782SJens Axboe break; 153003ba3782SJens Axboe 153103ba3782SJens Axboe /* 1532aa373cf5SJan Kara * Background writeout and kupdate-style writeback may 1533aa373cf5SJan Kara * run forever. Stop them if there is other work to do 1534aa373cf5SJan Kara * so that e.g. sync can proceed. They'll be restarted 1535aa373cf5SJan Kara * after the other works are all done. 1536aa373cf5SJan Kara */ 1537aa373cf5SJan Kara if ((work->for_background || work->for_kupdate) && 1538f0054bb1STejun Heo !list_empty(&wb->work_list)) 1539aa373cf5SJan Kara break; 1540aa373cf5SJan Kara 1541aa373cf5SJan Kara /* 1542d3ddec76SWu Fengguang * For background writeout, stop when we are below the 1543d3ddec76SWu Fengguang * background dirty threshold 154403ba3782SJens Axboe */ 1545aa661bbeSTejun Heo if (work->for_background && !wb_over_bg_thresh(wb)) 154603ba3782SJens Axboe break; 154703ba3782SJens Axboe 15481bc36b64SJan Kara /* 15491bc36b64SJan Kara * Kupdate and background works are special and we want to 15501bc36b64SJan Kara * include all inodes that need writing. Livelock avoidance is 15511bc36b64SJan Kara * handled by these works yielding to any other work so we are 15521bc36b64SJan Kara * safe. 15531bc36b64SJan Kara */ 1554ba9aa839SWu Fengguang if (work->for_kupdate) { 15550dc83bd3SJan Kara oldest_jif = jiffies - 1556ba9aa839SWu Fengguang msecs_to_jiffies(dirty_expire_interval * 10); 15571bc36b64SJan Kara } else if (work->for_background) 15580dc83bd3SJan Kara oldest_jif = jiffies; 1559028c2dd1SDave Chinner 1560d46db3d5SWu Fengguang trace_writeback_start(wb->bdi, work); 1561e8dfc305SWu Fengguang if (list_empty(&wb->b_io)) 1562ad4e38ddSCurt Wohlgemuth queue_io(wb, work); 156383ba7b07SChristoph Hellwig if (work->sb) 1564d46db3d5SWu Fengguang progress = writeback_sb_inodes(work->sb, wb, work); 1565edadfb10SChristoph Hellwig else 1566d46db3d5SWu Fengguang progress = __writeback_inodes_wb(wb, work); 1567d46db3d5SWu Fengguang trace_writeback_written(wb->bdi, work); 1568028c2dd1SDave Chinner 1569e98be2d5SWu Fengguang wb_update_bandwidth(wb, wb_start); 157003ba3782SJens Axboe 157103ba3782SJens Axboe /* 157271fd05a8SJens Axboe * Did we write something? Try for more 1573e6fb6da2SWu Fengguang * 1574e6fb6da2SWu Fengguang * Dirty inodes are moved to b_io for writeback in batches. 1575e6fb6da2SWu Fengguang * The completion of the current batch does not necessarily 1576e6fb6da2SWu Fengguang * mean the overall work is done. So we keep looping as long 1577e6fb6da2SWu Fengguang * as made some progress on cleaning pages or inodes. 157871fd05a8SJens Axboe */ 1579d46db3d5SWu Fengguang if (progress) 158003ba3782SJens Axboe continue; 1581a5989bdcSJan Kara /* 1582e6fb6da2SWu Fengguang * No more inodes for IO, bail 1583a5989bdcSJan Kara */ 1584b7a2441fSWu Fengguang if (list_empty(&wb->b_more_io)) 158503ba3782SJens Axboe break; 158603ba3782SJens Axboe /* 15878010c3b6SJens Axboe * Nothing written. Wait for some inode to 15888010c3b6SJens Axboe * become available for writeback. Otherwise 15898010c3b6SJens Axboe * we'll just busyloop. 159003ba3782SJens Axboe */ 159103ba3782SJens Axboe if (!list_empty(&wb->b_more_io)) { 1592d46db3d5SWu Fengguang trace_writeback_wait(wb->bdi, work); 159303ba3782SJens Axboe inode = wb_inode(wb->b_more_io.prev); 1594250df6edSDave Chinner spin_lock(&inode->i_lock); 1595f0d07b7fSJan Kara spin_unlock(&wb->list_lock); 1596169ebd90SJan Kara /* This function drops i_lock... */ 1597169ebd90SJan Kara inode_sleep_on_writeback(inode); 1598f0d07b7fSJan Kara spin_lock(&wb->list_lock); 159903ba3782SJens Axboe } 160003ba3782SJens Axboe } 1601e8dfc305SWu Fengguang spin_unlock(&wb->list_lock); 160203ba3782SJens Axboe 1603d46db3d5SWu Fengguang return nr_pages - work->nr_pages; 160403ba3782SJens Axboe } 160503ba3782SJens Axboe 160603ba3782SJens Axboe /* 160783ba7b07SChristoph Hellwig * Return the next wb_writeback_work struct that hasn't been processed yet. 160803ba3782SJens Axboe */ 1609f0054bb1STejun Heo static struct wb_writeback_work *get_next_work_item(struct bdi_writeback *wb) 161003ba3782SJens Axboe { 161183ba7b07SChristoph Hellwig struct wb_writeback_work *work = NULL; 161203ba3782SJens Axboe 1613f0054bb1STejun Heo spin_lock_bh(&wb->work_lock); 1614f0054bb1STejun Heo if (!list_empty(&wb->work_list)) { 1615f0054bb1STejun Heo work = list_entry(wb->work_list.next, 161683ba7b07SChristoph Hellwig struct wb_writeback_work, list); 161783ba7b07SChristoph Hellwig list_del_init(&work->list); 161803ba3782SJens Axboe } 1619f0054bb1STejun Heo spin_unlock_bh(&wb->work_lock); 162083ba7b07SChristoph Hellwig return work; 162103ba3782SJens Axboe } 162203ba3782SJens Axboe 1623cdf01dd5SLinus Torvalds /* 1624cdf01dd5SLinus Torvalds * Add in the number of potentially dirty inodes, because each inode 1625cdf01dd5SLinus Torvalds * write can dirty pagecache in the underlying blockdev. 1626cdf01dd5SLinus Torvalds */ 1627cdf01dd5SLinus Torvalds static unsigned long get_nr_dirty_pages(void) 1628cdf01dd5SLinus Torvalds { 1629cdf01dd5SLinus Torvalds return global_page_state(NR_FILE_DIRTY) + 1630cdf01dd5SLinus Torvalds global_page_state(NR_UNSTABLE_NFS) + 1631cdf01dd5SLinus Torvalds get_nr_dirty_inodes(); 1632cdf01dd5SLinus Torvalds } 1633cdf01dd5SLinus Torvalds 16346585027aSJan Kara static long wb_check_background_flush(struct bdi_writeback *wb) 16356585027aSJan Kara { 1636aa661bbeSTejun Heo if (wb_over_bg_thresh(wb)) { 16376585027aSJan Kara 16386585027aSJan Kara struct wb_writeback_work work = { 16396585027aSJan Kara .nr_pages = LONG_MAX, 16406585027aSJan Kara .sync_mode = WB_SYNC_NONE, 16416585027aSJan Kara .for_background = 1, 16426585027aSJan Kara .range_cyclic = 1, 16430e175a18SCurt Wohlgemuth .reason = WB_REASON_BACKGROUND, 16446585027aSJan Kara }; 16456585027aSJan Kara 16466585027aSJan Kara return wb_writeback(wb, &work); 16476585027aSJan Kara } 16486585027aSJan Kara 16496585027aSJan Kara return 0; 16506585027aSJan Kara } 16516585027aSJan Kara 165203ba3782SJens Axboe static long wb_check_old_data_flush(struct bdi_writeback *wb) 165303ba3782SJens Axboe { 165403ba3782SJens Axboe unsigned long expired; 165503ba3782SJens Axboe long nr_pages; 165603ba3782SJens Axboe 165769b62d01SJens Axboe /* 165869b62d01SJens Axboe * When set to zero, disable periodic writeback 165969b62d01SJens Axboe */ 166069b62d01SJens Axboe if (!dirty_writeback_interval) 166169b62d01SJens Axboe return 0; 166269b62d01SJens Axboe 166303ba3782SJens Axboe expired = wb->last_old_flush + 166403ba3782SJens Axboe msecs_to_jiffies(dirty_writeback_interval * 10); 166503ba3782SJens Axboe if (time_before(jiffies, expired)) 166603ba3782SJens Axboe return 0; 166703ba3782SJens Axboe 166803ba3782SJens Axboe wb->last_old_flush = jiffies; 1669cdf01dd5SLinus Torvalds nr_pages = get_nr_dirty_pages(); 167003ba3782SJens Axboe 1671c4a77a6cSJens Axboe if (nr_pages) { 167283ba7b07SChristoph Hellwig struct wb_writeback_work work = { 1673c4a77a6cSJens Axboe .nr_pages = nr_pages, 1674c4a77a6cSJens Axboe .sync_mode = WB_SYNC_NONE, 1675c4a77a6cSJens Axboe .for_kupdate = 1, 1676c4a77a6cSJens Axboe .range_cyclic = 1, 16770e175a18SCurt Wohlgemuth .reason = WB_REASON_PERIODIC, 1678c4a77a6cSJens Axboe }; 1679c4a77a6cSJens Axboe 168083ba7b07SChristoph Hellwig return wb_writeback(wb, &work); 1681c4a77a6cSJens Axboe } 168203ba3782SJens Axboe 168303ba3782SJens Axboe return 0; 168403ba3782SJens Axboe } 168503ba3782SJens Axboe 168603ba3782SJens Axboe /* 168703ba3782SJens Axboe * Retrieve work items and do the writeback they describe 168803ba3782SJens Axboe */ 168925d130baSWanpeng Li static long wb_do_writeback(struct bdi_writeback *wb) 169003ba3782SJens Axboe { 169183ba7b07SChristoph Hellwig struct wb_writeback_work *work; 1692c4a77a6cSJens Axboe long wrote = 0; 169303ba3782SJens Axboe 16944452226eSTejun Heo set_bit(WB_writeback_running, &wb->state); 1695f0054bb1STejun Heo while ((work = get_next_work_item(wb)) != NULL) { 1696cc395d7fSTejun Heo struct wb_completion *done = work->done; 169798754bf7STejun Heo bool need_wake_up = false; 169883ba7b07SChristoph Hellwig 1699f0054bb1STejun Heo trace_writeback_exec(wb->bdi, work); 1700455b2864SDave Chinner 170183ba7b07SChristoph Hellwig wrote += wb_writeback(wb, work); 170203ba3782SJens Axboe 170398754bf7STejun Heo if (work->single_wait) { 170498754bf7STejun Heo WARN_ON_ONCE(work->auto_free); 170598754bf7STejun Heo /* paired w/ rmb in wb_wait_for_single_work() */ 170698754bf7STejun Heo smp_wmb(); 170798754bf7STejun Heo work->single_done = 1; 170898754bf7STejun Heo need_wake_up = true; 170998754bf7STejun Heo } else if (work->auto_free) { 171083ba7b07SChristoph Hellwig kfree(work); 171198754bf7STejun Heo } 171298754bf7STejun Heo 1713cc395d7fSTejun Heo if (done && atomic_dec_and_test(&done->cnt)) 171498754bf7STejun Heo need_wake_up = true; 171598754bf7STejun Heo 171698754bf7STejun Heo if (need_wake_up) 1717cc395d7fSTejun Heo wake_up_all(&wb->bdi->wb_waitq); 171803ba3782SJens Axboe } 171903ba3782SJens Axboe 172003ba3782SJens Axboe /* 172103ba3782SJens Axboe * Check for periodic writeback, kupdated() style 172203ba3782SJens Axboe */ 172303ba3782SJens Axboe wrote += wb_check_old_data_flush(wb); 17246585027aSJan Kara wrote += wb_check_background_flush(wb); 17254452226eSTejun Heo clear_bit(WB_writeback_running, &wb->state); 172603ba3782SJens Axboe 172703ba3782SJens Axboe return wrote; 172803ba3782SJens Axboe } 172903ba3782SJens Axboe 173003ba3782SJens Axboe /* 173103ba3782SJens Axboe * Handle writeback of dirty data for the device backed by this bdi. Also 1732839a8e86STejun Heo * reschedules periodically and does kupdated style flushing. 173303ba3782SJens Axboe */ 1734f0054bb1STejun Heo void wb_workfn(struct work_struct *work) 173503ba3782SJens Axboe { 1736839a8e86STejun Heo struct bdi_writeback *wb = container_of(to_delayed_work(work), 1737839a8e86STejun Heo struct bdi_writeback, dwork); 173803ba3782SJens Axboe long pages_written; 173903ba3782SJens Axboe 1740f0054bb1STejun Heo set_worker_desc("flush-%s", dev_name(wb->bdi->dev)); 1741766f9164SPeter Zijlstra current->flags |= PF_SWAPWRITE; 174203ba3782SJens Axboe 1743839a8e86STejun Heo if (likely(!current_is_workqueue_rescuer() || 17444452226eSTejun Heo !test_bit(WB_registered, &wb->state))) { 174503ba3782SJens Axboe /* 1746f0054bb1STejun Heo * The normal path. Keep writing back @wb until its 1747839a8e86STejun Heo * work_list is empty. Note that this path is also taken 1748f0054bb1STejun Heo * if @wb is shutting down even when we're running off the 1749839a8e86STejun Heo * rescuer as work_list needs to be drained. 175003ba3782SJens Axboe */ 1751839a8e86STejun Heo do { 175225d130baSWanpeng Li pages_written = wb_do_writeback(wb); 1753455b2864SDave Chinner trace_writeback_pages_written(pages_written); 1754f0054bb1STejun Heo } while (!list_empty(&wb->work_list)); 1755839a8e86STejun Heo } else { 1756253c34e9SArtem Bityutskiy /* 1757839a8e86STejun Heo * bdi_wq can't get enough workers and we're running off 1758839a8e86STejun Heo * the emergency worker. Don't hog it. Hopefully, 1024 is 1759839a8e86STejun Heo * enough for efficient IO. 1760253c34e9SArtem Bityutskiy */ 1761f0054bb1STejun Heo pages_written = writeback_inodes_wb(wb, 1024, 1762839a8e86STejun Heo WB_REASON_FORKER_THREAD); 1763839a8e86STejun Heo trace_writeback_pages_written(pages_written); 176403ba3782SJens Axboe } 176503ba3782SJens Axboe 1766f0054bb1STejun Heo if (!list_empty(&wb->work_list)) 17676ca738d6SDerek Basehore mod_delayed_work(bdi_wq, &wb->dwork, 0); 17686ca738d6SDerek Basehore else if (wb_has_dirty_io(wb) && dirty_writeback_interval) 1769f0054bb1STejun Heo wb_wakeup_delayed(wb); 1770455b2864SDave Chinner 1771839a8e86STejun Heo current->flags &= ~PF_SWAPWRITE; 177203ba3782SJens Axboe } 177303ba3782SJens Axboe 177403ba3782SJens Axboe /* 177503ba3782SJens Axboe * Start writeback of `nr_pages' pages. If `nr_pages' is zero, write back 177603ba3782SJens Axboe * the whole world. 177703ba3782SJens Axboe */ 17780e175a18SCurt Wohlgemuth void wakeup_flusher_threads(long nr_pages, enum wb_reason reason) 177903ba3782SJens Axboe { 1780b8c2f347SChristoph Hellwig struct backing_dev_info *bdi; 1781b8c2f347SChristoph Hellwig 178247df3ddeSJan Kara if (!nr_pages) 178347df3ddeSJan Kara nr_pages = get_nr_dirty_pages(); 1784b8c2f347SChristoph Hellwig 1785b8c2f347SChristoph Hellwig rcu_read_lock(); 1786f2b65121STejun Heo list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 1787f2b65121STejun Heo struct bdi_writeback *wb; 1788f2b65121STejun Heo struct wb_iter iter; 1789f2b65121STejun Heo 1790f2b65121STejun Heo if (!bdi_has_dirty_io(bdi)) 1791f2b65121STejun Heo continue; 1792f2b65121STejun Heo 1793f2b65121STejun Heo bdi_for_each_wb(wb, bdi, &iter, 0) 1794f2b65121STejun Heo wb_start_writeback(wb, wb_split_bdi_pages(wb, nr_pages), 1795f2b65121STejun Heo false, reason); 1796f2b65121STejun Heo } 1797b8c2f347SChristoph Hellwig rcu_read_unlock(); 179803ba3782SJens Axboe } 179903ba3782SJens Axboe 1800a2f48706STheodore Ts'o /* 1801a2f48706STheodore Ts'o * Wake up bdi's periodically to make sure dirtytime inodes gets 1802a2f48706STheodore Ts'o * written back periodically. We deliberately do *not* check the 1803a2f48706STheodore Ts'o * b_dirtytime list in wb_has_dirty_io(), since this would cause the 1804a2f48706STheodore Ts'o * kernel to be constantly waking up once there are any dirtytime 1805a2f48706STheodore Ts'o * inodes on the system. So instead we define a separate delayed work 1806a2f48706STheodore Ts'o * function which gets called much more rarely. (By default, only 1807a2f48706STheodore Ts'o * once every 12 hours.) 1808a2f48706STheodore Ts'o * 1809a2f48706STheodore Ts'o * If there is any other write activity going on in the file system, 1810a2f48706STheodore Ts'o * this function won't be necessary. But if the only thing that has 1811a2f48706STheodore Ts'o * happened on the file system is a dirtytime inode caused by an atime 1812a2f48706STheodore Ts'o * update, we need this infrastructure below to make sure that inode 1813a2f48706STheodore Ts'o * eventually gets pushed out to disk. 1814a2f48706STheodore Ts'o */ 1815a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w); 1816a2f48706STheodore Ts'o static DECLARE_DELAYED_WORK(dirtytime_work, wakeup_dirtytime_writeback); 1817a2f48706STheodore Ts'o 1818a2f48706STheodore Ts'o static void wakeup_dirtytime_writeback(struct work_struct *w) 1819a2f48706STheodore Ts'o { 1820a2f48706STheodore Ts'o struct backing_dev_info *bdi; 1821a2f48706STheodore Ts'o 1822a2f48706STheodore Ts'o rcu_read_lock(); 1823a2f48706STheodore Ts'o list_for_each_entry_rcu(bdi, &bdi_list, bdi_list) { 1824001fe6f6STejun Heo struct bdi_writeback *wb; 1825001fe6f6STejun Heo struct wb_iter iter; 1826001fe6f6STejun Heo 1827001fe6f6STejun Heo bdi_for_each_wb(wb, bdi, &iter, 0) 1828001fe6f6STejun Heo if (!list_empty(&bdi->wb.b_dirty_time)) 1829f0054bb1STejun Heo wb_wakeup(&bdi->wb); 1830a2f48706STheodore Ts'o } 1831a2f48706STheodore Ts'o rcu_read_unlock(); 1832a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 1833a2f48706STheodore Ts'o } 1834a2f48706STheodore Ts'o 1835a2f48706STheodore Ts'o static int __init start_dirtytime_writeback(void) 1836a2f48706STheodore Ts'o { 1837a2f48706STheodore Ts'o schedule_delayed_work(&dirtytime_work, dirtytime_expire_interval * HZ); 1838a2f48706STheodore Ts'o return 0; 1839a2f48706STheodore Ts'o } 1840a2f48706STheodore Ts'o __initcall(start_dirtytime_writeback); 1841a2f48706STheodore Ts'o 18421efff914STheodore Ts'o int dirtytime_interval_handler(struct ctl_table *table, int write, 18431efff914STheodore Ts'o void __user *buffer, size_t *lenp, loff_t *ppos) 18441efff914STheodore Ts'o { 18451efff914STheodore Ts'o int ret; 18461efff914STheodore Ts'o 18471efff914STheodore Ts'o ret = proc_dointvec_minmax(table, write, buffer, lenp, ppos); 18481efff914STheodore Ts'o if (ret == 0 && write) 18491efff914STheodore Ts'o mod_delayed_work(system_wq, &dirtytime_work, 0); 18501efff914STheodore Ts'o return ret; 18511efff914STheodore Ts'o } 18521efff914STheodore Ts'o 185303ba3782SJens Axboe static noinline void block_dump___mark_inode_dirty(struct inode *inode) 185403ba3782SJens Axboe { 185503ba3782SJens Axboe if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) { 185603ba3782SJens Axboe struct dentry *dentry; 185703ba3782SJens Axboe const char *name = "?"; 185803ba3782SJens Axboe 185903ba3782SJens Axboe dentry = d_find_alias(inode); 186003ba3782SJens Axboe if (dentry) { 186103ba3782SJens Axboe spin_lock(&dentry->d_lock); 186203ba3782SJens Axboe name = (const char *) dentry->d_name.name; 186303ba3782SJens Axboe } 186403ba3782SJens Axboe printk(KERN_DEBUG 186503ba3782SJens Axboe "%s(%d): dirtied inode %lu (%s) on %s\n", 186603ba3782SJens Axboe current->comm, task_pid_nr(current), inode->i_ino, 186703ba3782SJens Axboe name, inode->i_sb->s_id); 186803ba3782SJens Axboe if (dentry) { 186903ba3782SJens Axboe spin_unlock(&dentry->d_lock); 187003ba3782SJens Axboe dput(dentry); 187103ba3782SJens Axboe } 187203ba3782SJens Axboe } 187303ba3782SJens Axboe } 187403ba3782SJens Axboe 187503ba3782SJens Axboe /** 187603ba3782SJens Axboe * __mark_inode_dirty - internal function 187703ba3782SJens Axboe * @inode: inode to mark 187803ba3782SJens Axboe * @flags: what kind of dirty (i.e. I_DIRTY_SYNC) 187903ba3782SJens Axboe * Mark an inode as dirty. Callers should use mark_inode_dirty or 188003ba3782SJens Axboe * mark_inode_dirty_sync. 188103ba3782SJens Axboe * 188203ba3782SJens Axboe * Put the inode on the super block's dirty list. 188303ba3782SJens Axboe * 188403ba3782SJens Axboe * CAREFUL! We mark it dirty unconditionally, but move it onto the 188503ba3782SJens Axboe * dirty list only if it is hashed or if it refers to a blockdev. 188603ba3782SJens Axboe * If it was not hashed, it will never be added to the dirty list 188703ba3782SJens Axboe * even if it is later hashed, as it will have been marked dirty already. 188803ba3782SJens Axboe * 188903ba3782SJens Axboe * In short, make sure you hash any inodes _before_ you start marking 189003ba3782SJens Axboe * them dirty. 189103ba3782SJens Axboe * 189203ba3782SJens Axboe * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 189303ba3782SJens Axboe * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 189403ba3782SJens Axboe * the kernel-internal blockdev inode represents the dirtying time of the 189503ba3782SJens Axboe * blockdev's pages. This is why for I_DIRTY_PAGES we always use 189603ba3782SJens Axboe * page->mapping->host, so the page-dirtying time is recorded in the internal 189703ba3782SJens Axboe * blockdev inode. 189803ba3782SJens Axboe */ 18990ae45f63STheodore Ts'o #define I_DIRTY_INODE (I_DIRTY_SYNC | I_DIRTY_DATASYNC) 190003ba3782SJens Axboe void __mark_inode_dirty(struct inode *inode, int flags) 190103ba3782SJens Axboe { 190203ba3782SJens Axboe struct super_block *sb = inode->i_sb; 19030ae45f63STheodore Ts'o int dirtytime; 19040ae45f63STheodore Ts'o 19050ae45f63STheodore Ts'o trace_writeback_mark_inode_dirty(inode, flags); 190603ba3782SJens Axboe 190703ba3782SJens Axboe /* 190803ba3782SJens Axboe * Don't do this for I_DIRTY_PAGES - that doesn't actually 190903ba3782SJens Axboe * dirty the inode itself 191003ba3782SJens Axboe */ 19110ae45f63STheodore Ts'o if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC | I_DIRTY_TIME)) { 19129fb0a7daSTejun Heo trace_writeback_dirty_inode_start(inode, flags); 19139fb0a7daSTejun Heo 191403ba3782SJens Axboe if (sb->s_op->dirty_inode) 1915aa385729SChristoph Hellwig sb->s_op->dirty_inode(inode, flags); 19169fb0a7daSTejun Heo 19179fb0a7daSTejun Heo trace_writeback_dirty_inode(inode, flags); 191803ba3782SJens Axboe } 19190ae45f63STheodore Ts'o if (flags & I_DIRTY_INODE) 19200ae45f63STheodore Ts'o flags &= ~I_DIRTY_TIME; 19210ae45f63STheodore Ts'o dirtytime = flags & I_DIRTY_TIME; 192203ba3782SJens Axboe 192303ba3782SJens Axboe /* 19249c6ac78eSTejun Heo * Paired with smp_mb() in __writeback_single_inode() for the 19259c6ac78eSTejun Heo * following lockless i_state test. See there for details. 192603ba3782SJens Axboe */ 192703ba3782SJens Axboe smp_mb(); 192803ba3782SJens Axboe 19290ae45f63STheodore Ts'o if (((inode->i_state & flags) == flags) || 19300ae45f63STheodore Ts'o (dirtytime && (inode->i_state & I_DIRTY_INODE))) 193103ba3782SJens Axboe return; 193203ba3782SJens Axboe 193303ba3782SJens Axboe if (unlikely(block_dump)) 193403ba3782SJens Axboe block_dump___mark_inode_dirty(inode); 193503ba3782SJens Axboe 1936250df6edSDave Chinner spin_lock(&inode->i_lock); 19370ae45f63STheodore Ts'o if (dirtytime && (inode->i_state & I_DIRTY_INODE)) 19380ae45f63STheodore Ts'o goto out_unlock_inode; 193903ba3782SJens Axboe if ((inode->i_state & flags) != flags) { 194003ba3782SJens Axboe const int was_dirty = inode->i_state & I_DIRTY; 194103ba3782SJens Axboe 194252ebea74STejun Heo inode_attach_wb(inode, NULL); 194352ebea74STejun Heo 19440ae45f63STheodore Ts'o if (flags & I_DIRTY_INODE) 19450ae45f63STheodore Ts'o inode->i_state &= ~I_DIRTY_TIME; 194603ba3782SJens Axboe inode->i_state |= flags; 194703ba3782SJens Axboe 194803ba3782SJens Axboe /* 194903ba3782SJens Axboe * If the inode is being synced, just update its dirty state. 195003ba3782SJens Axboe * The unlocker will place the inode on the appropriate 195103ba3782SJens Axboe * superblock list, based upon its state. 195203ba3782SJens Axboe */ 195303ba3782SJens Axboe if (inode->i_state & I_SYNC) 1954250df6edSDave Chinner goto out_unlock_inode; 195503ba3782SJens Axboe 195603ba3782SJens Axboe /* 195703ba3782SJens Axboe * Only add valid (hashed) inodes to the superblock's 195803ba3782SJens Axboe * dirty list. Add blockdev inodes as well. 195903ba3782SJens Axboe */ 196003ba3782SJens Axboe if (!S_ISBLK(inode->i_mode)) { 19611d3382cbSAl Viro if (inode_unhashed(inode)) 1962250df6edSDave Chinner goto out_unlock_inode; 196303ba3782SJens Axboe } 1964a4ffdde6SAl Viro if (inode->i_state & I_FREEING) 1965250df6edSDave Chinner goto out_unlock_inode; 196603ba3782SJens Axboe 196703ba3782SJens Axboe /* 196803ba3782SJens Axboe * If the inode was already on b_dirty/b_io/b_more_io, don't 196903ba3782SJens Axboe * reposition it (that would break b_dirty time-ordering). 197003ba3782SJens Axboe */ 197103ba3782SJens Axboe if (!was_dirty) { 197287e1d789STejun Heo struct bdi_writeback *wb; 1973d6c10f1fSTejun Heo struct list_head *dirty_list; 1974a66979abSDave Chinner bool wakeup_bdi = false; 1975500b067cSJens Axboe 197687e1d789STejun Heo wb = locked_inode_to_wb_and_lock_list(inode); 1977253c34e9SArtem Bityutskiy 19780747259dSTejun Heo WARN(bdi_cap_writeback_dirty(wb->bdi) && 19790747259dSTejun Heo !test_bit(WB_registered, &wb->state), 19800747259dSTejun Heo "bdi-%s not registered\n", wb->bdi->name); 198103ba3782SJens Axboe 198203ba3782SJens Axboe inode->dirtied_when = jiffies; 1983a2f48706STheodore Ts'o if (dirtytime) 1984a2f48706STheodore Ts'o inode->dirtied_time_when = jiffies; 1985d6c10f1fSTejun Heo 1986a2f48706STheodore Ts'o if (inode->i_state & (I_DIRTY_INODE | I_DIRTY_PAGES)) 19870747259dSTejun Heo dirty_list = &wb->b_dirty; 1988a2f48706STheodore Ts'o else 19890747259dSTejun Heo dirty_list = &wb->b_dirty_time; 1990d6c10f1fSTejun Heo 19910747259dSTejun Heo wakeup_bdi = inode_wb_list_move_locked(inode, wb, 1992d6c10f1fSTejun Heo dirty_list); 1993d6c10f1fSTejun Heo 19940747259dSTejun Heo spin_unlock(&wb->list_lock); 19950ae45f63STheodore Ts'o trace_writeback_dirty_inode_enqueue(inode); 1996253c34e9SArtem Bityutskiy 1997d6c10f1fSTejun Heo /* 1998d6c10f1fSTejun Heo * If this is the first dirty inode for this bdi, 1999d6c10f1fSTejun Heo * we have to wake-up the corresponding bdi thread 2000d6c10f1fSTejun Heo * to make sure background write-back happens 2001d6c10f1fSTejun Heo * later. 2002d6c10f1fSTejun Heo */ 20030747259dSTejun Heo if (bdi_cap_writeback_dirty(wb->bdi) && wakeup_bdi) 20040747259dSTejun Heo wb_wakeup_delayed(wb); 2005a66979abSDave Chinner return; 2006a66979abSDave Chinner } 2007a66979abSDave Chinner } 2008a66979abSDave Chinner out_unlock_inode: 2009a66979abSDave Chinner spin_unlock(&inode->i_lock); 2010a66979abSDave Chinner 201103ba3782SJens Axboe } 201203ba3782SJens Axboe EXPORT_SYMBOL(__mark_inode_dirty); 201303ba3782SJens Axboe 2014b6e51316SJens Axboe static void wait_sb_inodes(struct super_block *sb) 201566f3b8e2SJens Axboe { 201638f21977SNick Piggin struct inode *inode, *old_inode = NULL; 201738f21977SNick Piggin 201803ba3782SJens Axboe /* 201903ba3782SJens Axboe * We need to be protected against the filesystem going from 202003ba3782SJens Axboe * r/o to r/w or vice versa. 202103ba3782SJens Axboe */ 2022b6e51316SJens Axboe WARN_ON(!rwsem_is_locked(&sb->s_umount)); 202303ba3782SJens Axboe 202455fa6091SDave Chinner spin_lock(&inode_sb_list_lock); 202566f3b8e2SJens Axboe 202638f21977SNick Piggin /* 202738f21977SNick Piggin * Data integrity sync. Must wait for all pages under writeback, 202838f21977SNick Piggin * because there may have been pages dirtied before our sync 202938f21977SNick Piggin * call, but which had writeout started before we write it out. 203038f21977SNick Piggin * In which case, the inode may not be on the dirty list, but 203138f21977SNick Piggin * we still have to wait for that writeout. 203238f21977SNick Piggin */ 2033b6e51316SJens Axboe list_for_each_entry(inode, &sb->s_inodes, i_sb_list) { 2034250df6edSDave Chinner struct address_space *mapping = inode->i_mapping; 203538f21977SNick Piggin 2036250df6edSDave Chinner spin_lock(&inode->i_lock); 2037250df6edSDave Chinner if ((inode->i_state & (I_FREEING|I_WILL_FREE|I_NEW)) || 2038250df6edSDave Chinner (mapping->nrpages == 0)) { 2039250df6edSDave Chinner spin_unlock(&inode->i_lock); 204038f21977SNick Piggin continue; 2041250df6edSDave Chinner } 204238f21977SNick Piggin __iget(inode); 2043250df6edSDave Chinner spin_unlock(&inode->i_lock); 204455fa6091SDave Chinner spin_unlock(&inode_sb_list_lock); 204555fa6091SDave Chinner 204638f21977SNick Piggin /* 204755fa6091SDave Chinner * We hold a reference to 'inode' so it couldn't have been 204855fa6091SDave Chinner * removed from s_inodes list while we dropped the 204955fa6091SDave Chinner * inode_sb_list_lock. We cannot iput the inode now as we can 205055fa6091SDave Chinner * be holding the last reference and we cannot iput it under 205155fa6091SDave Chinner * inode_sb_list_lock. So we keep the reference and iput it 205255fa6091SDave Chinner * later. 205338f21977SNick Piggin */ 205438f21977SNick Piggin iput(old_inode); 205538f21977SNick Piggin old_inode = inode; 205638f21977SNick Piggin 205738f21977SNick Piggin filemap_fdatawait(mapping); 205838f21977SNick Piggin 205938f21977SNick Piggin cond_resched(); 206038f21977SNick Piggin 206155fa6091SDave Chinner spin_lock(&inode_sb_list_lock); 206238f21977SNick Piggin } 206355fa6091SDave Chinner spin_unlock(&inode_sb_list_lock); 206438f21977SNick Piggin iput(old_inode); 206566f3b8e2SJens Axboe } 20661da177e4SLinus Torvalds 2067f30a7d0cSTejun Heo static void __writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 2068f30a7d0cSTejun Heo enum wb_reason reason, bool skip_if_busy) 20691da177e4SLinus Torvalds { 2070cc395d7fSTejun Heo DEFINE_WB_COMPLETION_ONSTACK(done); 207183ba7b07SChristoph Hellwig struct wb_writeback_work work = { 20723c4d7165SChristoph Hellwig .sb = sb, 20733c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_NONE, 20746e6938b6SWu Fengguang .tagged_writepages = 1, 207583ba7b07SChristoph Hellwig .done = &done, 20763259f8beSChris Mason .nr_pages = nr, 20770e175a18SCurt Wohlgemuth .reason = reason, 20783c4d7165SChristoph Hellwig }; 2079e7972912STejun Heo struct backing_dev_info *bdi = sb->s_bdi; 20800e3c9a22SJens Axboe 2081e7972912STejun Heo if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 20826eedc701SJan Kara return; 2083cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2084f30a7d0cSTejun Heo 2085db125360STejun Heo bdi_split_work_to_wbs(sb->s_bdi, &work, skip_if_busy); 2086cc395d7fSTejun Heo wb_wait_for_completion(bdi, &done); 20871da177e4SLinus Torvalds } 2088f30a7d0cSTejun Heo 2089f30a7d0cSTejun Heo /** 2090f30a7d0cSTejun Heo * writeback_inodes_sb_nr - writeback dirty inodes from given super_block 2091f30a7d0cSTejun Heo * @sb: the superblock 2092f30a7d0cSTejun Heo * @nr: the number of pages to write 2093f30a7d0cSTejun Heo * @reason: reason why some writeback work initiated 2094f30a7d0cSTejun Heo * 2095f30a7d0cSTejun Heo * Start writeback on some inodes on this super_block. No guarantees are made 2096f30a7d0cSTejun Heo * on how many (if any) will be written, and this function does not wait 2097f30a7d0cSTejun Heo * for IO completion of submitted IO. 2098f30a7d0cSTejun Heo */ 2099f30a7d0cSTejun Heo void writeback_inodes_sb_nr(struct super_block *sb, 2100f30a7d0cSTejun Heo unsigned long nr, 2101f30a7d0cSTejun Heo enum wb_reason reason) 2102f30a7d0cSTejun Heo { 2103f30a7d0cSTejun Heo __writeback_inodes_sb_nr(sb, nr, reason, false); 2104f30a7d0cSTejun Heo } 21053259f8beSChris Mason EXPORT_SYMBOL(writeback_inodes_sb_nr); 21063259f8beSChris Mason 21073259f8beSChris Mason /** 21083259f8beSChris Mason * writeback_inodes_sb - writeback dirty inodes from given super_block 21093259f8beSChris Mason * @sb: the superblock 2110786228abSMarcos Paulo de Souza * @reason: reason why some writeback work was initiated 21113259f8beSChris Mason * 21123259f8beSChris Mason * Start writeback on some inodes on this super_block. No guarantees are made 21133259f8beSChris Mason * on how many (if any) will be written, and this function does not wait 21143259f8beSChris Mason * for IO completion of submitted IO. 21153259f8beSChris Mason */ 21160e175a18SCurt Wohlgemuth void writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 21173259f8beSChris Mason { 21180e175a18SCurt Wohlgemuth return writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 21193259f8beSChris Mason } 2120d8a8559cSJens Axboe EXPORT_SYMBOL(writeback_inodes_sb); 2121d8a8559cSJens Axboe 2122d8a8559cSJens Axboe /** 212310ee27a0SMiao Xie * try_to_writeback_inodes_sb_nr - try to start writeback if none underway 21243259f8beSChris Mason * @sb: the superblock 21253259f8beSChris Mason * @nr: the number of pages to write 212610ee27a0SMiao Xie * @reason: the reason of writeback 21273259f8beSChris Mason * 212810ee27a0SMiao Xie * Invoke writeback_inodes_sb_nr if no writeback is currently underway. 21293259f8beSChris Mason * Returns 1 if writeback was started, 0 if not. 21303259f8beSChris Mason */ 2131f30a7d0cSTejun Heo bool try_to_writeback_inodes_sb_nr(struct super_block *sb, unsigned long nr, 21320e175a18SCurt Wohlgemuth enum wb_reason reason) 21333259f8beSChris Mason { 213410ee27a0SMiao Xie if (!down_read_trylock(&sb->s_umount)) 2135f30a7d0cSTejun Heo return false; 213610ee27a0SMiao Xie 2137f30a7d0cSTejun Heo __writeback_inodes_sb_nr(sb, nr, reason, true); 21383259f8beSChris Mason up_read(&sb->s_umount); 2139f30a7d0cSTejun Heo return true; 21403259f8beSChris Mason } 214110ee27a0SMiao Xie EXPORT_SYMBOL(try_to_writeback_inodes_sb_nr); 214210ee27a0SMiao Xie 214310ee27a0SMiao Xie /** 214410ee27a0SMiao Xie * try_to_writeback_inodes_sb - try to start writeback if none underway 214510ee27a0SMiao Xie * @sb: the superblock 214610ee27a0SMiao Xie * @reason: reason why some writeback work was initiated 214710ee27a0SMiao Xie * 214810ee27a0SMiao Xie * Implement by try_to_writeback_inodes_sb_nr() 214910ee27a0SMiao Xie * Returns 1 if writeback was started, 0 if not. 215010ee27a0SMiao Xie */ 2151f30a7d0cSTejun Heo bool try_to_writeback_inodes_sb(struct super_block *sb, enum wb_reason reason) 215210ee27a0SMiao Xie { 215310ee27a0SMiao Xie return try_to_writeback_inodes_sb_nr(sb, get_nr_dirty_pages(), reason); 215410ee27a0SMiao Xie } 215510ee27a0SMiao Xie EXPORT_SYMBOL(try_to_writeback_inodes_sb); 21563259f8beSChris Mason 21573259f8beSChris Mason /** 2158d8a8559cSJens Axboe * sync_inodes_sb - sync sb inode pages 2159d8a8559cSJens Axboe * @sb: the superblock 2160d8a8559cSJens Axboe * 2161d8a8559cSJens Axboe * This function writes and waits on any dirty inode belonging to this 21620dc83bd3SJan Kara * super_block. 2163d8a8559cSJens Axboe */ 21640dc83bd3SJan Kara void sync_inodes_sb(struct super_block *sb) 2165d8a8559cSJens Axboe { 2166cc395d7fSTejun Heo DEFINE_WB_COMPLETION_ONSTACK(done); 216783ba7b07SChristoph Hellwig struct wb_writeback_work work = { 21683c4d7165SChristoph Hellwig .sb = sb, 21693c4d7165SChristoph Hellwig .sync_mode = WB_SYNC_ALL, 21703c4d7165SChristoph Hellwig .nr_pages = LONG_MAX, 21713c4d7165SChristoph Hellwig .range_cyclic = 0, 217283ba7b07SChristoph Hellwig .done = &done, 21730e175a18SCurt Wohlgemuth .reason = WB_REASON_SYNC, 21747747bd4bSDave Chinner .for_sync = 1, 21753c4d7165SChristoph Hellwig }; 2176e7972912STejun Heo struct backing_dev_info *bdi = sb->s_bdi; 21773c4d7165SChristoph Hellwig 21786eedc701SJan Kara /* Nothing to do? */ 2179e7972912STejun Heo if (!bdi_has_dirty_io(bdi) || bdi == &noop_backing_dev_info) 21806eedc701SJan Kara return; 2181cf37e972SChristoph Hellwig WARN_ON(!rwsem_is_locked(&sb->s_umount)); 2182cf37e972SChristoph Hellwig 2183db125360STejun Heo bdi_split_work_to_wbs(bdi, &work, false); 2184cc395d7fSTejun Heo wb_wait_for_completion(bdi, &done); 218583ba7b07SChristoph Hellwig 2186b6e51316SJens Axboe wait_sb_inodes(sb); 2187d8a8559cSJens Axboe } 2188d8a8559cSJens Axboe EXPORT_SYMBOL(sync_inodes_sb); 21891da177e4SLinus Torvalds 21901da177e4SLinus Torvalds /** 21911da177e4SLinus Torvalds * write_inode_now - write an inode to disk 21921da177e4SLinus Torvalds * @inode: inode to write to disk 21931da177e4SLinus Torvalds * @sync: whether the write should be synchronous or not 21941da177e4SLinus Torvalds * 21957f04c26dSAndrea Arcangeli * This function commits an inode to disk immediately if it is dirty. This is 21967f04c26dSAndrea Arcangeli * primarily needed by knfsd. 21977f04c26dSAndrea Arcangeli * 21987f04c26dSAndrea Arcangeli * The caller must either have a ref on the inode or must have set I_WILL_FREE. 21991da177e4SLinus Torvalds */ 22001da177e4SLinus Torvalds int write_inode_now(struct inode *inode, int sync) 22011da177e4SLinus Torvalds { 2202f758eeabSChristoph Hellwig struct bdi_writeback *wb = &inode_to_bdi(inode)->wb; 22031da177e4SLinus Torvalds struct writeback_control wbc = { 22041da177e4SLinus Torvalds .nr_to_write = LONG_MAX, 220518914b18SMike Galbraith .sync_mode = sync ? WB_SYNC_ALL : WB_SYNC_NONE, 2206111ebb6eSOGAWA Hirofumi .range_start = 0, 2207111ebb6eSOGAWA Hirofumi .range_end = LLONG_MAX, 22081da177e4SLinus Torvalds }; 22091da177e4SLinus Torvalds 22101da177e4SLinus Torvalds if (!mapping_cap_writeback_dirty(inode->i_mapping)) 221149364ce2SAndrew Morton wbc.nr_to_write = 0; 22121da177e4SLinus Torvalds 22131da177e4SLinus Torvalds might_sleep(); 22144f8ad655SJan Kara return writeback_single_inode(inode, wb, &wbc); 22151da177e4SLinus Torvalds } 22161da177e4SLinus Torvalds EXPORT_SYMBOL(write_inode_now); 22171da177e4SLinus Torvalds 22181da177e4SLinus Torvalds /** 22191da177e4SLinus Torvalds * sync_inode - write an inode and its pages to disk. 22201da177e4SLinus Torvalds * @inode: the inode to sync 22211da177e4SLinus Torvalds * @wbc: controls the writeback mode 22221da177e4SLinus Torvalds * 22231da177e4SLinus Torvalds * sync_inode() will write an inode and its pages to disk. It will also 22241da177e4SLinus Torvalds * correctly update the inode on its superblock's dirty inode lists and will 22251da177e4SLinus Torvalds * update inode->i_state. 22261da177e4SLinus Torvalds * 22271da177e4SLinus Torvalds * The caller must have a ref on the inode. 22281da177e4SLinus Torvalds */ 22291da177e4SLinus Torvalds int sync_inode(struct inode *inode, struct writeback_control *wbc) 22301da177e4SLinus Torvalds { 22314f8ad655SJan Kara return writeback_single_inode(inode, &inode_to_bdi(inode)->wb, wbc); 22321da177e4SLinus Torvalds } 22331da177e4SLinus Torvalds EXPORT_SYMBOL(sync_inode); 2234c3765016SChristoph Hellwig 2235c3765016SChristoph Hellwig /** 2236c691b9d9SAndrew Morton * sync_inode_metadata - write an inode to disk 2237c3765016SChristoph Hellwig * @inode: the inode to sync 2238c3765016SChristoph Hellwig * @wait: wait for I/O to complete. 2239c3765016SChristoph Hellwig * 2240c691b9d9SAndrew Morton * Write an inode to disk and adjust its dirty state after completion. 2241c3765016SChristoph Hellwig * 2242c3765016SChristoph Hellwig * Note: only writes the actual inode, no associated data or other metadata. 2243c3765016SChristoph Hellwig */ 2244c3765016SChristoph Hellwig int sync_inode_metadata(struct inode *inode, int wait) 2245c3765016SChristoph Hellwig { 2246c3765016SChristoph Hellwig struct writeback_control wbc = { 2247c3765016SChristoph Hellwig .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_NONE, 2248c3765016SChristoph Hellwig .nr_to_write = 0, /* metadata-only */ 2249c3765016SChristoph Hellwig }; 2250c3765016SChristoph Hellwig 2251c3765016SChristoph Hellwig return sync_inode(inode, &wbc); 2252c3765016SChristoph Hellwig } 2253c3765016SChristoph Hellwig EXPORT_SYMBOL(sync_inode_metadata); 2254