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 * 111da177e4SLinus Torvalds * 10Apr2002 akpm@zip.com.au 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> 17f5ff8422SJens Axboe #include <linux/module.h> 181da177e4SLinus Torvalds #include <linux/spinlock.h> 191da177e4SLinus Torvalds #include <linux/sched.h> 201da177e4SLinus Torvalds #include <linux/fs.h> 211da177e4SLinus Torvalds #include <linux/mm.h> 221da177e4SLinus Torvalds #include <linux/writeback.h> 231da177e4SLinus Torvalds #include <linux/blkdev.h> 241da177e4SLinus Torvalds #include <linux/backing-dev.h> 251da177e4SLinus Torvalds #include <linux/buffer_head.h> 2607f3f05cSDavid Howells #include "internal.h" 271da177e4SLinus Torvalds 281da177e4SLinus Torvalds /** 291da177e4SLinus Torvalds * __mark_inode_dirty - internal function 301da177e4SLinus Torvalds * @inode: inode to mark 311da177e4SLinus Torvalds * @flags: what kind of dirty (i.e. I_DIRTY_SYNC) 321da177e4SLinus Torvalds * Mark an inode as dirty. Callers should use mark_inode_dirty or 331da177e4SLinus Torvalds * mark_inode_dirty_sync. 341da177e4SLinus Torvalds * 351da177e4SLinus Torvalds * Put the inode on the super block's dirty list. 361da177e4SLinus Torvalds * 371da177e4SLinus Torvalds * CAREFUL! We mark it dirty unconditionally, but move it onto the 381da177e4SLinus Torvalds * dirty list only if it is hashed or if it refers to a blockdev. 391da177e4SLinus Torvalds * If it was not hashed, it will never be added to the dirty list 401da177e4SLinus Torvalds * even if it is later hashed, as it will have been marked dirty already. 411da177e4SLinus Torvalds * 421da177e4SLinus Torvalds * In short, make sure you hash any inodes _before_ you start marking 431da177e4SLinus Torvalds * them dirty. 441da177e4SLinus Torvalds * 451da177e4SLinus Torvalds * This function *must* be atomic for the I_DIRTY_PAGES case - 461da177e4SLinus Torvalds * set_page_dirty() is called under spinlock in several places. 471da177e4SLinus Torvalds * 481da177e4SLinus Torvalds * Note that for blockdevs, inode->dirtied_when represents the dirtying time of 491da177e4SLinus Torvalds * the block-special inode (/dev/hda1) itself. And the ->dirtied_when field of 501da177e4SLinus Torvalds * the kernel-internal blockdev inode represents the dirtying time of the 511da177e4SLinus Torvalds * blockdev's pages. This is why for I_DIRTY_PAGES we always use 521da177e4SLinus Torvalds * page->mapping->host, so the page-dirtying time is recorded in the internal 531da177e4SLinus Torvalds * blockdev inode. 541da177e4SLinus Torvalds */ 551da177e4SLinus Torvalds void __mark_inode_dirty(struct inode *inode, int flags) 561da177e4SLinus Torvalds { 571da177e4SLinus Torvalds struct super_block *sb = inode->i_sb; 581da177e4SLinus Torvalds 591da177e4SLinus Torvalds /* 601da177e4SLinus Torvalds * Don't do this for I_DIRTY_PAGES - that doesn't actually 611da177e4SLinus Torvalds * dirty the inode itself 621da177e4SLinus Torvalds */ 631da177e4SLinus Torvalds if (flags & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) { 641da177e4SLinus Torvalds if (sb->s_op->dirty_inode) 651da177e4SLinus Torvalds sb->s_op->dirty_inode(inode); 661da177e4SLinus Torvalds } 671da177e4SLinus Torvalds 681da177e4SLinus Torvalds /* 691da177e4SLinus Torvalds * make sure that changes are seen by all cpus before we test i_state 701da177e4SLinus Torvalds * -- mikulas 711da177e4SLinus Torvalds */ 721da177e4SLinus Torvalds smp_mb(); 731da177e4SLinus Torvalds 741da177e4SLinus Torvalds /* avoid the locking if we can */ 751da177e4SLinus Torvalds if ((inode->i_state & flags) == flags) 761da177e4SLinus Torvalds return; 771da177e4SLinus Torvalds 781da177e4SLinus Torvalds if (unlikely(block_dump)) { 791da177e4SLinus Torvalds struct dentry *dentry = NULL; 801da177e4SLinus Torvalds const char *name = "?"; 811da177e4SLinus Torvalds 821da177e4SLinus Torvalds if (!list_empty(&inode->i_dentry)) { 831da177e4SLinus Torvalds dentry = list_entry(inode->i_dentry.next, 841da177e4SLinus Torvalds struct dentry, d_alias); 851da177e4SLinus Torvalds if (dentry && dentry->d_name.name) 861da177e4SLinus Torvalds name = (const char *) dentry->d_name.name; 871da177e4SLinus Torvalds } 881da177e4SLinus Torvalds 891da177e4SLinus Torvalds if (inode->i_ino || strcmp(inode->i_sb->s_id, "bdev")) 901da177e4SLinus Torvalds printk(KERN_DEBUG 911da177e4SLinus Torvalds "%s(%d): dirtied inode %lu (%s) on %s\n", 921da177e4SLinus Torvalds current->comm, current->pid, inode->i_ino, 931da177e4SLinus Torvalds name, inode->i_sb->s_id); 941da177e4SLinus Torvalds } 951da177e4SLinus Torvalds 961da177e4SLinus Torvalds spin_lock(&inode_lock); 971da177e4SLinus Torvalds if ((inode->i_state & flags) != flags) { 981da177e4SLinus Torvalds const int was_dirty = inode->i_state & I_DIRTY; 991da177e4SLinus Torvalds 1001da177e4SLinus Torvalds inode->i_state |= flags; 1011da177e4SLinus Torvalds 1021da177e4SLinus Torvalds /* 1031da177e4SLinus Torvalds * If the inode is locked, just update its dirty state. 1041da177e4SLinus Torvalds * The unlocker will place the inode on the appropriate 1051da177e4SLinus Torvalds * superblock list, based upon its state. 1061da177e4SLinus Torvalds */ 1071da177e4SLinus Torvalds if (inode->i_state & I_LOCK) 1081da177e4SLinus Torvalds goto out; 1091da177e4SLinus Torvalds 1101da177e4SLinus Torvalds /* 1111da177e4SLinus Torvalds * Only add valid (hashed) inodes to the superblock's 1121da177e4SLinus Torvalds * dirty list. Add blockdev inodes as well. 1131da177e4SLinus Torvalds */ 1141da177e4SLinus Torvalds if (!S_ISBLK(inode->i_mode)) { 1151da177e4SLinus Torvalds if (hlist_unhashed(&inode->i_hash)) 1161da177e4SLinus Torvalds goto out; 1171da177e4SLinus Torvalds } 1181da177e4SLinus Torvalds if (inode->i_state & (I_FREEING|I_CLEAR)) 1191da177e4SLinus Torvalds goto out; 1201da177e4SLinus Torvalds 1211da177e4SLinus Torvalds /* 1221da177e4SLinus Torvalds * If the inode was already on s_dirty or s_io, don't 1231da177e4SLinus Torvalds * reposition it (that would break s_dirty time-ordering). 1241da177e4SLinus Torvalds */ 1251da177e4SLinus Torvalds if (!was_dirty) { 1261da177e4SLinus Torvalds inode->dirtied_when = jiffies; 1271da177e4SLinus Torvalds list_move(&inode->i_list, &sb->s_dirty); 1281da177e4SLinus Torvalds } 1291da177e4SLinus Torvalds } 1301da177e4SLinus Torvalds out: 1311da177e4SLinus Torvalds spin_unlock(&inode_lock); 1321da177e4SLinus Torvalds } 1331da177e4SLinus Torvalds 1341da177e4SLinus Torvalds EXPORT_SYMBOL(__mark_inode_dirty); 1351da177e4SLinus Torvalds 1361da177e4SLinus Torvalds static int write_inode(struct inode *inode, int sync) 1371da177e4SLinus Torvalds { 1381da177e4SLinus Torvalds if (inode->i_sb->s_op->write_inode && !is_bad_inode(inode)) 1391da177e4SLinus Torvalds return inode->i_sb->s_op->write_inode(inode, sync); 1401da177e4SLinus Torvalds return 0; 1411da177e4SLinus Torvalds } 1421da177e4SLinus Torvalds 1431da177e4SLinus Torvalds /* 1446610a0bcSAndrew Morton * Redirty an inode: set its when-it-was dirtied timestamp and move it to the 1456610a0bcSAndrew Morton * furthest end of its superblock's dirty-inode list. 1466610a0bcSAndrew Morton * 1476610a0bcSAndrew Morton * Before stamping the inode's ->dirtied_when, we check to see whether it is 1486610a0bcSAndrew Morton * already the most-recently-dirtied inode on the s_dirty list. If that is 1496610a0bcSAndrew Morton * the case then the inode must have been redirtied while it was being written 1506610a0bcSAndrew Morton * out and we don't reset its dirtied_when. 1516610a0bcSAndrew Morton */ 1526610a0bcSAndrew Morton static void redirty_tail(struct inode *inode) 1536610a0bcSAndrew Morton { 1546610a0bcSAndrew Morton struct super_block *sb = inode->i_sb; 1556610a0bcSAndrew Morton 1566610a0bcSAndrew Morton if (!list_empty(&sb->s_dirty)) { 1576610a0bcSAndrew Morton struct inode *tail_inode; 1586610a0bcSAndrew Morton 1596610a0bcSAndrew Morton tail_inode = list_entry(sb->s_dirty.next, struct inode, i_list); 1606610a0bcSAndrew Morton if (!time_after_eq(inode->dirtied_when, 1616610a0bcSAndrew Morton tail_inode->dirtied_when)) 1626610a0bcSAndrew Morton inode->dirtied_when = jiffies; 1636610a0bcSAndrew Morton } 1646610a0bcSAndrew Morton list_move(&inode->i_list, &sb->s_dirty); 1656610a0bcSAndrew Morton } 1666610a0bcSAndrew Morton 1676610a0bcSAndrew Morton /* 1681da177e4SLinus Torvalds * Write a single inode's dirty pages and inode data out to disk. 1691da177e4SLinus Torvalds * If `wait' is set, wait on the writeout. 1701da177e4SLinus Torvalds * 1711da177e4SLinus Torvalds * The whole writeout design is quite complex and fragile. We want to avoid 1721da177e4SLinus Torvalds * starvation of particular inodes when others are being redirtied, prevent 1731da177e4SLinus Torvalds * livelocks, etc. 1741da177e4SLinus Torvalds * 1751da177e4SLinus Torvalds * Called under inode_lock. 1761da177e4SLinus Torvalds */ 1771da177e4SLinus Torvalds static int 1781da177e4SLinus Torvalds __sync_single_inode(struct inode *inode, struct writeback_control *wbc) 1791da177e4SLinus Torvalds { 1801da177e4SLinus Torvalds unsigned dirty; 1811da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 1821da177e4SLinus Torvalds struct super_block *sb = inode->i_sb; 1831da177e4SLinus Torvalds int wait = wbc->sync_mode == WB_SYNC_ALL; 1841da177e4SLinus Torvalds int ret; 1851da177e4SLinus Torvalds 1861da177e4SLinus Torvalds BUG_ON(inode->i_state & I_LOCK); 1871da177e4SLinus Torvalds 1881da177e4SLinus Torvalds /* Set I_LOCK, reset I_DIRTY */ 1891da177e4SLinus Torvalds dirty = inode->i_state & I_DIRTY; 1901da177e4SLinus Torvalds inode->i_state |= I_LOCK; 1911da177e4SLinus Torvalds inode->i_state &= ~I_DIRTY; 1921da177e4SLinus Torvalds 1931da177e4SLinus Torvalds spin_unlock(&inode_lock); 1941da177e4SLinus Torvalds 1951da177e4SLinus Torvalds ret = do_writepages(mapping, wbc); 1961da177e4SLinus Torvalds 1971da177e4SLinus Torvalds /* Don't write the inode if only I_DIRTY_PAGES was set */ 1981da177e4SLinus Torvalds if (dirty & (I_DIRTY_SYNC | I_DIRTY_DATASYNC)) { 1991da177e4SLinus Torvalds int err = write_inode(inode, wait); 2001da177e4SLinus Torvalds if (ret == 0) 2011da177e4SLinus Torvalds ret = err; 2021da177e4SLinus Torvalds } 2031da177e4SLinus Torvalds 2041da177e4SLinus Torvalds if (wait) { 2051da177e4SLinus Torvalds int err = filemap_fdatawait(mapping); 2061da177e4SLinus Torvalds if (ret == 0) 2071da177e4SLinus Torvalds ret = err; 2081da177e4SLinus Torvalds } 2091da177e4SLinus Torvalds 2101da177e4SLinus Torvalds spin_lock(&inode_lock); 2111da177e4SLinus Torvalds inode->i_state &= ~I_LOCK; 2121da177e4SLinus Torvalds if (!(inode->i_state & I_FREEING)) { 2131da177e4SLinus Torvalds if (!(inode->i_state & I_DIRTY) && 2141da177e4SLinus Torvalds mapping_tagged(mapping, PAGECACHE_TAG_DIRTY)) { 2151da177e4SLinus Torvalds /* 2161da177e4SLinus Torvalds * We didn't write back all the pages. nfs_writepages() 2171da177e4SLinus Torvalds * sometimes bales out without doing anything. Redirty 2181da177e4SLinus Torvalds * the inode. It is still on sb->s_io. 2191da177e4SLinus Torvalds */ 2201da177e4SLinus Torvalds if (wbc->for_kupdate) { 2211da177e4SLinus Torvalds /* 2221da177e4SLinus Torvalds * For the kupdate function we leave the inode 2231da177e4SLinus Torvalds * at the head of sb_dirty so it will get more 2241da177e4SLinus Torvalds * writeout as soon as the queue becomes 2251da177e4SLinus Torvalds * uncongested. 2261da177e4SLinus Torvalds */ 2271da177e4SLinus Torvalds inode->i_state |= I_DIRTY_PAGES; 2281da177e4SLinus Torvalds list_move_tail(&inode->i_list, &sb->s_dirty); 2291da177e4SLinus Torvalds } else { 2301da177e4SLinus Torvalds /* 2311da177e4SLinus Torvalds * Otherwise fully redirty the inode so that 2321da177e4SLinus Torvalds * other inodes on this superblock will get some 2331da177e4SLinus Torvalds * writeout. Otherwise heavy writing to one 2341da177e4SLinus Torvalds * file would indefinitely suspend writeout of 2351da177e4SLinus Torvalds * all the other files. 2361da177e4SLinus Torvalds */ 2371da177e4SLinus Torvalds inode->i_state |= I_DIRTY_PAGES; 2381da177e4SLinus Torvalds inode->dirtied_when = jiffies; 2391da177e4SLinus Torvalds list_move(&inode->i_list, &sb->s_dirty); 2401da177e4SLinus Torvalds } 2411da177e4SLinus Torvalds } else if (inode->i_state & I_DIRTY) { 2421da177e4SLinus Torvalds /* 2431da177e4SLinus Torvalds * Someone redirtied the inode while were writing back 2441da177e4SLinus Torvalds * the pages. 2451da177e4SLinus Torvalds */ 2466610a0bcSAndrew Morton redirty_tail(inode); 2471da177e4SLinus Torvalds } else if (atomic_read(&inode->i_count)) { 2481da177e4SLinus Torvalds /* 2491da177e4SLinus Torvalds * The inode is clean, inuse 2501da177e4SLinus Torvalds */ 2511da177e4SLinus Torvalds list_move(&inode->i_list, &inode_in_use); 2521da177e4SLinus Torvalds } else { 2531da177e4SLinus Torvalds /* 2541da177e4SLinus Torvalds * The inode is clean, unused 2551da177e4SLinus Torvalds */ 2561da177e4SLinus Torvalds list_move(&inode->i_list, &inode_unused); 2571da177e4SLinus Torvalds } 2581da177e4SLinus Torvalds } 2591da177e4SLinus Torvalds wake_up_inode(inode); 2601da177e4SLinus Torvalds return ret; 2611da177e4SLinus Torvalds } 2621da177e4SLinus Torvalds 2631da177e4SLinus Torvalds /* 2647f04c26dSAndrea Arcangeli * Write out an inode's dirty pages. Called under inode_lock. Either the 2657f04c26dSAndrea Arcangeli * caller has ref on the inode (either via __iget or via syscall against an fd) 2667f04c26dSAndrea Arcangeli * or the inode has I_WILL_FREE set (via generic_forget_inode) 2671da177e4SLinus Torvalds */ 2681da177e4SLinus Torvalds static int 2697f04c26dSAndrea Arcangeli __writeback_single_inode(struct inode *inode, struct writeback_control *wbc) 2701da177e4SLinus Torvalds { 2711da177e4SLinus Torvalds wait_queue_head_t *wqh; 2721da177e4SLinus Torvalds 2737f04c26dSAndrea Arcangeli if (!atomic_read(&inode->i_count)) 274659603efSAndrea Arcangeli WARN_ON(!(inode->i_state & (I_WILL_FREE|I_FREEING))); 2757f04c26dSAndrea Arcangeli else 2767f04c26dSAndrea Arcangeli WARN_ON(inode->i_state & I_WILL_FREE); 2777f04c26dSAndrea Arcangeli 2781da177e4SLinus Torvalds if ((wbc->sync_mode != WB_SYNC_ALL) && (inode->i_state & I_LOCK)) { 2794b89eed9SLinus Torvalds struct address_space *mapping = inode->i_mapping; 2804b89eed9SLinus Torvalds int ret; 2814b89eed9SLinus Torvalds 2821da177e4SLinus Torvalds list_move(&inode->i_list, &inode->i_sb->s_dirty); 2834b89eed9SLinus Torvalds 2844b89eed9SLinus Torvalds /* 2854b89eed9SLinus Torvalds * Even if we don't actually write the inode itself here, 2864b89eed9SLinus Torvalds * we can at least start some of the data writeout.. 2874b89eed9SLinus Torvalds */ 2884b89eed9SLinus Torvalds spin_unlock(&inode_lock); 2894b89eed9SLinus Torvalds ret = do_writepages(mapping, wbc); 2904b89eed9SLinus Torvalds spin_lock(&inode_lock); 2914b89eed9SLinus Torvalds return ret; 2921da177e4SLinus Torvalds } 2931da177e4SLinus Torvalds 2941da177e4SLinus Torvalds /* 2951da177e4SLinus Torvalds * It's a data-integrity sync. We must wait. 2961da177e4SLinus Torvalds */ 2971da177e4SLinus Torvalds if (inode->i_state & I_LOCK) { 2981da177e4SLinus Torvalds DEFINE_WAIT_BIT(wq, &inode->i_state, __I_LOCK); 2991da177e4SLinus Torvalds 3001da177e4SLinus Torvalds wqh = bit_waitqueue(&inode->i_state, __I_LOCK); 3011da177e4SLinus Torvalds do { 3021da177e4SLinus Torvalds spin_unlock(&inode_lock); 3031da177e4SLinus Torvalds __wait_on_bit(wqh, &wq, inode_wait, 3041da177e4SLinus Torvalds TASK_UNINTERRUPTIBLE); 3051da177e4SLinus Torvalds spin_lock(&inode_lock); 3061da177e4SLinus Torvalds } while (inode->i_state & I_LOCK); 3071da177e4SLinus Torvalds } 3081da177e4SLinus Torvalds return __sync_single_inode(inode, wbc); 3091da177e4SLinus Torvalds } 3101da177e4SLinus Torvalds 3111da177e4SLinus Torvalds /* 3121da177e4SLinus Torvalds * Write out a superblock's list of dirty inodes. A wait will be performed 3131da177e4SLinus Torvalds * upon no inodes, all inodes or the final one, depending upon sync_mode. 3141da177e4SLinus Torvalds * 3151da177e4SLinus Torvalds * If older_than_this is non-NULL, then only write out inodes which 3161da177e4SLinus Torvalds * had their first dirtying at a time earlier than *older_than_this. 3171da177e4SLinus Torvalds * 3181da177e4SLinus Torvalds * If we're a pdlfush thread, then implement pdflush collision avoidance 3191da177e4SLinus Torvalds * against the entire list. 3201da177e4SLinus Torvalds * 3211da177e4SLinus Torvalds * WB_SYNC_HOLD is a hack for sys_sync(): reattach the inode to sb->s_dirty so 3221da177e4SLinus Torvalds * that it can be located for waiting on in __writeback_single_inode(). 3231da177e4SLinus Torvalds * 3241da177e4SLinus Torvalds * Called under inode_lock. 3251da177e4SLinus Torvalds * 3261da177e4SLinus Torvalds * If `bdi' is non-zero then we're being asked to writeback a specific queue. 3271da177e4SLinus Torvalds * This function assumes that the blockdev superblock's inodes are backed by 3281da177e4SLinus Torvalds * a variety of queues, so all inodes are searched. For other superblocks, 3291da177e4SLinus Torvalds * assume that all inodes are backed by the same queue. 3301da177e4SLinus Torvalds * 3311da177e4SLinus Torvalds * FIXME: this linear search could get expensive with many fileystems. But 3321da177e4SLinus Torvalds * how to fix? We need to go from an address_space to all inodes which share 3331da177e4SLinus Torvalds * a queue with that address_space. (Easy: have a global "dirty superblocks" 3341da177e4SLinus Torvalds * list). 3351da177e4SLinus Torvalds * 3361da177e4SLinus Torvalds * The inodes to be written are parked on sb->s_io. They are moved back onto 3371da177e4SLinus Torvalds * sb->s_dirty as they are selected for writing. This way, none can be missed 3381da177e4SLinus Torvalds * on the writer throttling path, and we get decent balancing between many 3391da177e4SLinus Torvalds * throttled threads: we don't want them all piling up on __wait_on_inode. 3401da177e4SLinus Torvalds */ 3411da177e4SLinus Torvalds static void 3421da177e4SLinus Torvalds sync_sb_inodes(struct super_block *sb, struct writeback_control *wbc) 3431da177e4SLinus Torvalds { 3441da177e4SLinus Torvalds const unsigned long start = jiffies; /* livelock avoidance */ 3451da177e4SLinus Torvalds 3461da177e4SLinus Torvalds if (!wbc->for_kupdate || list_empty(&sb->s_io)) 3471da177e4SLinus Torvalds list_splice_init(&sb->s_dirty, &sb->s_io); 3481da177e4SLinus Torvalds 3491da177e4SLinus Torvalds while (!list_empty(&sb->s_io)) { 3501da177e4SLinus Torvalds struct inode *inode = list_entry(sb->s_io.prev, 3511da177e4SLinus Torvalds struct inode, i_list); 3521da177e4SLinus Torvalds struct address_space *mapping = inode->i_mapping; 3531da177e4SLinus Torvalds struct backing_dev_info *bdi = mapping->backing_dev_info; 3541da177e4SLinus Torvalds long pages_skipped; 3551da177e4SLinus Torvalds 3561da177e4SLinus Torvalds if (!bdi_cap_writeback_dirty(bdi)) { 3579852a0e7SAndrew Morton redirty_tail(inode); 3587b0de42dSDavid Howells if (sb_is_blkdev_sb(sb)) { 3591da177e4SLinus Torvalds /* 3601da177e4SLinus Torvalds * Dirty memory-backed blockdev: the ramdisk 3611da177e4SLinus Torvalds * driver does this. Skip just this inode 3621da177e4SLinus Torvalds */ 3631da177e4SLinus Torvalds continue; 3641da177e4SLinus Torvalds } 3651da177e4SLinus Torvalds /* 3661da177e4SLinus Torvalds * Dirty memory-backed inode against a filesystem other 3671da177e4SLinus Torvalds * than the kernel-internal bdev filesystem. Skip the 3681da177e4SLinus Torvalds * entire superblock. 3691da177e4SLinus Torvalds */ 3701da177e4SLinus Torvalds break; 3711da177e4SLinus Torvalds } 3721da177e4SLinus Torvalds 3731da177e4SLinus Torvalds if (wbc->nonblocking && bdi_write_congested(bdi)) { 3741da177e4SLinus Torvalds wbc->encountered_congestion = 1; 3757b0de42dSDavid Howells if (!sb_is_blkdev_sb(sb)) 3761da177e4SLinus Torvalds break; /* Skip a congested fs */ 3771da177e4SLinus Torvalds list_move(&inode->i_list, &sb->s_dirty); 3781da177e4SLinus Torvalds continue; /* Skip a congested blockdev */ 3791da177e4SLinus Torvalds } 3801da177e4SLinus Torvalds 3811da177e4SLinus Torvalds if (wbc->bdi && bdi != wbc->bdi) { 3827b0de42dSDavid Howells if (!sb_is_blkdev_sb(sb)) 3831da177e4SLinus Torvalds break; /* fs has the wrong queue */ 3841da177e4SLinus Torvalds list_move(&inode->i_list, &sb->s_dirty); 3851da177e4SLinus Torvalds continue; /* blockdev has wrong queue */ 3861da177e4SLinus Torvalds } 3871da177e4SLinus Torvalds 3881da177e4SLinus Torvalds /* Was this inode dirtied after sync_sb_inodes was called? */ 3891da177e4SLinus Torvalds if (time_after(inode->dirtied_when, start)) 3901da177e4SLinus Torvalds break; 3911da177e4SLinus Torvalds 3921da177e4SLinus Torvalds /* Was this inode dirtied too recently? */ 3931da177e4SLinus Torvalds if (wbc->older_than_this && time_after(inode->dirtied_when, 3941da177e4SLinus Torvalds *wbc->older_than_this)) 3951da177e4SLinus Torvalds break; 3961da177e4SLinus Torvalds 3971da177e4SLinus Torvalds /* Is another pdflush already flushing this queue? */ 3981da177e4SLinus Torvalds if (current_is_pdflush() && !writeback_acquire(bdi)) 3991da177e4SLinus Torvalds break; 4001da177e4SLinus Torvalds 4011da177e4SLinus Torvalds BUG_ON(inode->i_state & I_FREEING); 4021da177e4SLinus Torvalds __iget(inode); 4031da177e4SLinus Torvalds pages_skipped = wbc->pages_skipped; 4041da177e4SLinus Torvalds __writeback_single_inode(inode, wbc); 4051da177e4SLinus Torvalds if (wbc->sync_mode == WB_SYNC_HOLD) { 4061da177e4SLinus Torvalds inode->dirtied_when = jiffies; 4071da177e4SLinus Torvalds list_move(&inode->i_list, &sb->s_dirty); 4081da177e4SLinus Torvalds } 4091da177e4SLinus Torvalds if (current_is_pdflush()) 4101da177e4SLinus Torvalds writeback_release(bdi); 4111da177e4SLinus Torvalds if (wbc->pages_skipped != pages_skipped) { 4121da177e4SLinus Torvalds /* 4131da177e4SLinus Torvalds * writeback is not making progress due to locked 4141da177e4SLinus Torvalds * buffers. Skip this inode for now. 4151da177e4SLinus Torvalds */ 416f57b9b7bSAndrew Morton redirty_tail(inode); 4171da177e4SLinus Torvalds } 4181da177e4SLinus Torvalds spin_unlock(&inode_lock); 4191da177e4SLinus Torvalds iput(inode); 4204ffc8444SOGAWA Hirofumi cond_resched(); 4211da177e4SLinus Torvalds spin_lock(&inode_lock); 4221da177e4SLinus Torvalds if (wbc->nr_to_write <= 0) 4231da177e4SLinus Torvalds break; 4241da177e4SLinus Torvalds } 4251da177e4SLinus Torvalds return; /* Leave any unwritten inodes on s_io */ 4261da177e4SLinus Torvalds } 4271da177e4SLinus Torvalds 4281da177e4SLinus Torvalds /* 4291da177e4SLinus Torvalds * Start writeback of dirty pagecache data against all unlocked inodes. 4301da177e4SLinus Torvalds * 4311da177e4SLinus Torvalds * Note: 4321da177e4SLinus Torvalds * We don't need to grab a reference to superblock here. If it has non-empty 4331da177e4SLinus Torvalds * ->s_dirty it's hadn't been killed yet and kill_super() won't proceed 4341da177e4SLinus Torvalds * past sync_inodes_sb() until both the ->s_dirty and ->s_io lists are 4351da177e4SLinus Torvalds * empty. Since __sync_single_inode() regains inode_lock before it finally moves 4361da177e4SLinus Torvalds * inode from superblock lists we are OK. 4371da177e4SLinus Torvalds * 4381da177e4SLinus Torvalds * If `older_than_this' is non-zero then only flush inodes which have a 4391da177e4SLinus Torvalds * flushtime older than *older_than_this. 4401da177e4SLinus Torvalds * 4411da177e4SLinus Torvalds * If `bdi' is non-zero then we will scan the first inode against each 4421da177e4SLinus Torvalds * superblock until we find the matching ones. One group will be the dirty 4431da177e4SLinus Torvalds * inodes against a filesystem. Then when we hit the dummy blockdev superblock, 4441da177e4SLinus Torvalds * sync_sb_inodes will seekout the blockdev which matches `bdi'. Maybe not 4451da177e4SLinus Torvalds * super-efficient but we're about to do a ton of I/O... 4461da177e4SLinus Torvalds */ 4471da177e4SLinus Torvalds void 4481da177e4SLinus Torvalds writeback_inodes(struct writeback_control *wbc) 4491da177e4SLinus Torvalds { 4501da177e4SLinus Torvalds struct super_block *sb; 4511da177e4SLinus Torvalds 4521da177e4SLinus Torvalds might_sleep(); 4531da177e4SLinus Torvalds spin_lock(&sb_lock); 4541da177e4SLinus Torvalds restart: 4551da177e4SLinus Torvalds sb = sb_entry(super_blocks.prev); 4561da177e4SLinus Torvalds for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) { 4571da177e4SLinus Torvalds if (!list_empty(&sb->s_dirty) || !list_empty(&sb->s_io)) { 4581da177e4SLinus Torvalds /* we're making our own get_super here */ 4591da177e4SLinus Torvalds sb->s_count++; 4601da177e4SLinus Torvalds spin_unlock(&sb_lock); 4611da177e4SLinus Torvalds /* 4621da177e4SLinus Torvalds * If we can't get the readlock, there's no sense in 4631da177e4SLinus Torvalds * waiting around, most of the time the FS is going to 4641da177e4SLinus Torvalds * be unmounted by the time it is released. 4651da177e4SLinus Torvalds */ 4661da177e4SLinus Torvalds if (down_read_trylock(&sb->s_umount)) { 4671da177e4SLinus Torvalds if (sb->s_root) { 4681da177e4SLinus Torvalds spin_lock(&inode_lock); 4691da177e4SLinus Torvalds sync_sb_inodes(sb, wbc); 4701da177e4SLinus Torvalds spin_unlock(&inode_lock); 4711da177e4SLinus Torvalds } 4721da177e4SLinus Torvalds up_read(&sb->s_umount); 4731da177e4SLinus Torvalds } 4741da177e4SLinus Torvalds spin_lock(&sb_lock); 4751da177e4SLinus Torvalds if (__put_super_and_need_restart(sb)) 4761da177e4SLinus Torvalds goto restart; 4771da177e4SLinus Torvalds } 4781da177e4SLinus Torvalds if (wbc->nr_to_write <= 0) 4791da177e4SLinus Torvalds break; 4801da177e4SLinus Torvalds } 4811da177e4SLinus Torvalds spin_unlock(&sb_lock); 4821da177e4SLinus Torvalds } 4831da177e4SLinus Torvalds 4841da177e4SLinus Torvalds /* 4851da177e4SLinus Torvalds * writeback and wait upon the filesystem's dirty inodes. The caller will 4861da177e4SLinus Torvalds * do this in two passes - one to write, and one to wait. WB_SYNC_HOLD is 4871da177e4SLinus Torvalds * used to park the written inodes on sb->s_dirty for the wait pass. 4881da177e4SLinus Torvalds * 4891da177e4SLinus Torvalds * A finite limit is set on the number of pages which will be written. 4901da177e4SLinus Torvalds * To prevent infinite livelock of sys_sync(). 4911da177e4SLinus Torvalds * 4921da177e4SLinus Torvalds * We add in the number of potentially dirty inodes, because each inode write 4931da177e4SLinus Torvalds * can dirty pagecache in the underlying blockdev. 4941da177e4SLinus Torvalds */ 4951da177e4SLinus Torvalds void sync_inodes_sb(struct super_block *sb, int wait) 4961da177e4SLinus Torvalds { 4971da177e4SLinus Torvalds struct writeback_control wbc = { 4981da177e4SLinus Torvalds .sync_mode = wait ? WB_SYNC_ALL : WB_SYNC_HOLD, 499111ebb6eSOGAWA Hirofumi .range_start = 0, 500111ebb6eSOGAWA Hirofumi .range_end = LLONG_MAX, 5011da177e4SLinus Torvalds }; 502b1e7a8fdSChristoph Lameter unsigned long nr_dirty = global_page_state(NR_FILE_DIRTY); 503fd39fc85SChristoph Lameter unsigned long nr_unstable = global_page_state(NR_UNSTABLE_NFS); 5041da177e4SLinus Torvalds 5051da177e4SLinus Torvalds wbc.nr_to_write = nr_dirty + nr_unstable + 5061da177e4SLinus Torvalds (inodes_stat.nr_inodes - inodes_stat.nr_unused) + 5071da177e4SLinus Torvalds nr_dirty + nr_unstable; 5081da177e4SLinus Torvalds wbc.nr_to_write += wbc.nr_to_write / 2; /* Bit more for luck */ 5091da177e4SLinus Torvalds spin_lock(&inode_lock); 5101da177e4SLinus Torvalds sync_sb_inodes(sb, &wbc); 5111da177e4SLinus Torvalds spin_unlock(&inode_lock); 5121da177e4SLinus Torvalds } 5131da177e4SLinus Torvalds 5141da177e4SLinus Torvalds /* 5151da177e4SLinus Torvalds * Rather lame livelock avoidance. 5161da177e4SLinus Torvalds */ 5171da177e4SLinus Torvalds static void set_sb_syncing(int val) 5181da177e4SLinus Torvalds { 5191da177e4SLinus Torvalds struct super_block *sb; 5201da177e4SLinus Torvalds spin_lock(&sb_lock); 5211da177e4SLinus Torvalds sb = sb_entry(super_blocks.prev); 5221da177e4SLinus Torvalds for (; sb != sb_entry(&super_blocks); sb = sb_entry(sb->s_list.prev)) { 5231da177e4SLinus Torvalds sb->s_syncing = val; 5241da177e4SLinus Torvalds } 5251da177e4SLinus Torvalds spin_unlock(&sb_lock); 5261da177e4SLinus Torvalds } 5271da177e4SLinus Torvalds 5281da177e4SLinus Torvalds /** 52967be2dd1SMartin Waitz * sync_inodes - writes all inodes to disk 53067be2dd1SMartin Waitz * @wait: wait for completion 5311da177e4SLinus Torvalds * 5321da177e4SLinus Torvalds * sync_inodes() goes through each super block's dirty inode list, writes the 5331da177e4SLinus Torvalds * inodes out, waits on the writeout and puts the inodes back on the normal 5341da177e4SLinus Torvalds * list. 5351da177e4SLinus Torvalds * 5361da177e4SLinus Torvalds * This is for sys_sync(). fsync_dev() uses the same algorithm. The subtle 5371da177e4SLinus Torvalds * part of the sync functions is that the blockdev "superblock" is processed 5381da177e4SLinus Torvalds * last. This is because the write_inode() function of a typical fs will 5391da177e4SLinus Torvalds * perform no I/O, but will mark buffers in the blockdev mapping as dirty. 5401da177e4SLinus Torvalds * What we want to do is to perform all that dirtying first, and then write 5411da177e4SLinus Torvalds * back all those inode blocks via the blockdev mapping in one sweep. So the 5421da177e4SLinus Torvalds * additional (somewhat redundant) sync_blockdev() calls here are to make 5431da177e4SLinus Torvalds * sure that really happens. Because if we call sync_inodes_sb(wait=1) with 5441da177e4SLinus Torvalds * outstanding dirty inodes, the writeback goes block-at-a-time within the 5451da177e4SLinus Torvalds * filesystem's write_inode(). This is extremely slow. 5461da177e4SLinus Torvalds */ 547618f0636SKirill Korotaev static void __sync_inodes(int wait) 5481da177e4SLinus Torvalds { 5491da177e4SLinus Torvalds struct super_block *sb; 5501da177e4SLinus Torvalds 551618f0636SKirill Korotaev spin_lock(&sb_lock); 552618f0636SKirill Korotaev restart: 553618f0636SKirill Korotaev list_for_each_entry(sb, &super_blocks, s_list) { 554618f0636SKirill Korotaev if (sb->s_syncing) 555618f0636SKirill Korotaev continue; 556618f0636SKirill Korotaev sb->s_syncing = 1; 557618f0636SKirill Korotaev sb->s_count++; 558618f0636SKirill Korotaev spin_unlock(&sb_lock); 559618f0636SKirill Korotaev down_read(&sb->s_umount); 560618f0636SKirill Korotaev if (sb->s_root) { 561618f0636SKirill Korotaev sync_inodes_sb(sb, wait); 5621da177e4SLinus Torvalds sync_blockdev(sb->s_bdev); 5631da177e4SLinus Torvalds } 564618f0636SKirill Korotaev up_read(&sb->s_umount); 565618f0636SKirill Korotaev spin_lock(&sb_lock); 566618f0636SKirill Korotaev if (__put_super_and_need_restart(sb)) 567618f0636SKirill Korotaev goto restart; 568618f0636SKirill Korotaev } 569618f0636SKirill Korotaev spin_unlock(&sb_lock); 570618f0636SKirill Korotaev } 571618f0636SKirill Korotaev 572618f0636SKirill Korotaev void sync_inodes(int wait) 573618f0636SKirill Korotaev { 574618f0636SKirill Korotaev set_sb_syncing(0); 575618f0636SKirill Korotaev __sync_inodes(0); 576618f0636SKirill Korotaev 5771da177e4SLinus Torvalds if (wait) { 5781da177e4SLinus Torvalds set_sb_syncing(0); 579618f0636SKirill Korotaev __sync_inodes(1); 5801da177e4SLinus Torvalds } 5811da177e4SLinus Torvalds } 5821da177e4SLinus Torvalds 5831da177e4SLinus Torvalds /** 5841da177e4SLinus Torvalds * write_inode_now - write an inode to disk 5851da177e4SLinus Torvalds * @inode: inode to write to disk 5861da177e4SLinus Torvalds * @sync: whether the write should be synchronous or not 5871da177e4SLinus Torvalds * 5887f04c26dSAndrea Arcangeli * This function commits an inode to disk immediately if it is dirty. This is 5897f04c26dSAndrea Arcangeli * primarily needed by knfsd. 5907f04c26dSAndrea Arcangeli * 5917f04c26dSAndrea Arcangeli * The caller must either have a ref on the inode or must have set I_WILL_FREE. 5921da177e4SLinus Torvalds */ 5931da177e4SLinus Torvalds int write_inode_now(struct inode *inode, int sync) 5941da177e4SLinus Torvalds { 5951da177e4SLinus Torvalds int ret; 5961da177e4SLinus Torvalds struct writeback_control wbc = { 5971da177e4SLinus Torvalds .nr_to_write = LONG_MAX, 5981da177e4SLinus Torvalds .sync_mode = WB_SYNC_ALL, 599111ebb6eSOGAWA Hirofumi .range_start = 0, 600111ebb6eSOGAWA Hirofumi .range_end = LLONG_MAX, 6011da177e4SLinus Torvalds }; 6021da177e4SLinus Torvalds 6031da177e4SLinus Torvalds if (!mapping_cap_writeback_dirty(inode->i_mapping)) 60449364ce2SAndrew Morton wbc.nr_to_write = 0; 6051da177e4SLinus Torvalds 6061da177e4SLinus Torvalds might_sleep(); 6071da177e4SLinus Torvalds spin_lock(&inode_lock); 6081da177e4SLinus Torvalds ret = __writeback_single_inode(inode, &wbc); 6091da177e4SLinus Torvalds spin_unlock(&inode_lock); 6101da177e4SLinus Torvalds if (sync) 6111da177e4SLinus Torvalds wait_on_inode(inode); 6121da177e4SLinus Torvalds return ret; 6131da177e4SLinus Torvalds } 6141da177e4SLinus Torvalds EXPORT_SYMBOL(write_inode_now); 6151da177e4SLinus Torvalds 6161da177e4SLinus Torvalds /** 6171da177e4SLinus Torvalds * sync_inode - write an inode and its pages to disk. 6181da177e4SLinus Torvalds * @inode: the inode to sync 6191da177e4SLinus Torvalds * @wbc: controls the writeback mode 6201da177e4SLinus Torvalds * 6211da177e4SLinus Torvalds * sync_inode() will write an inode and its pages to disk. It will also 6221da177e4SLinus Torvalds * correctly update the inode on its superblock's dirty inode lists and will 6231da177e4SLinus Torvalds * update inode->i_state. 6241da177e4SLinus Torvalds * 6251da177e4SLinus Torvalds * The caller must have a ref on the inode. 6261da177e4SLinus Torvalds */ 6271da177e4SLinus Torvalds int sync_inode(struct inode *inode, struct writeback_control *wbc) 6281da177e4SLinus Torvalds { 6291da177e4SLinus Torvalds int ret; 6301da177e4SLinus Torvalds 6311da177e4SLinus Torvalds spin_lock(&inode_lock); 6321da177e4SLinus Torvalds ret = __writeback_single_inode(inode, wbc); 6331da177e4SLinus Torvalds spin_unlock(&inode_lock); 6341da177e4SLinus Torvalds return ret; 6351da177e4SLinus Torvalds } 6361da177e4SLinus Torvalds EXPORT_SYMBOL(sync_inode); 6371da177e4SLinus Torvalds 6381da177e4SLinus Torvalds /** 6391da177e4SLinus Torvalds * generic_osync_inode - flush all dirty data for a given inode to disk 6401da177e4SLinus Torvalds * @inode: inode to write 64167be2dd1SMartin Waitz * @mapping: the address_space that should be flushed 6421da177e4SLinus Torvalds * @what: what to write and wait upon 6431da177e4SLinus Torvalds * 6441da177e4SLinus Torvalds * This can be called by file_write functions for files which have the 6451da177e4SLinus Torvalds * O_SYNC flag set, to flush dirty writes to disk. 6461da177e4SLinus Torvalds * 6471da177e4SLinus Torvalds * @what is a bitmask, specifying which part of the inode's data should be 648b8887e6eSRandy Dunlap * written and waited upon. 6491da177e4SLinus Torvalds * 6501da177e4SLinus Torvalds * OSYNC_DATA: i_mapping's dirty data 6511da177e4SLinus Torvalds * OSYNC_METADATA: the buffers at i_mapping->private_list 6521da177e4SLinus Torvalds * OSYNC_INODE: the inode itself 6531da177e4SLinus Torvalds */ 6541da177e4SLinus Torvalds 6551da177e4SLinus Torvalds int generic_osync_inode(struct inode *inode, struct address_space *mapping, int what) 6561da177e4SLinus Torvalds { 6571da177e4SLinus Torvalds int err = 0; 6581da177e4SLinus Torvalds int need_write_inode_now = 0; 6591da177e4SLinus Torvalds int err2; 6601da177e4SLinus Torvalds 6611da177e4SLinus Torvalds if (what & OSYNC_DATA) 6621da177e4SLinus Torvalds err = filemap_fdatawrite(mapping); 6631da177e4SLinus Torvalds if (what & (OSYNC_METADATA|OSYNC_DATA)) { 6641da177e4SLinus Torvalds err2 = sync_mapping_buffers(mapping); 6651da177e4SLinus Torvalds if (!err) 6661da177e4SLinus Torvalds err = err2; 6671da177e4SLinus Torvalds } 6681da177e4SLinus Torvalds if (what & OSYNC_DATA) { 6691da177e4SLinus Torvalds err2 = filemap_fdatawait(mapping); 6701da177e4SLinus Torvalds if (!err) 6711da177e4SLinus Torvalds err = err2; 6721da177e4SLinus Torvalds } 6731da177e4SLinus Torvalds 6741da177e4SLinus Torvalds spin_lock(&inode_lock); 6751da177e4SLinus Torvalds if ((inode->i_state & I_DIRTY) && 6761da177e4SLinus Torvalds ((what & OSYNC_INODE) || (inode->i_state & I_DIRTY_DATASYNC))) 6771da177e4SLinus Torvalds need_write_inode_now = 1; 6781da177e4SLinus Torvalds spin_unlock(&inode_lock); 6791da177e4SLinus Torvalds 6801da177e4SLinus Torvalds if (need_write_inode_now) { 6811da177e4SLinus Torvalds err2 = write_inode_now(inode, 1); 6821da177e4SLinus Torvalds if (!err) 6831da177e4SLinus Torvalds err = err2; 6841da177e4SLinus Torvalds } 6851da177e4SLinus Torvalds else 6861da177e4SLinus Torvalds wait_on_inode(inode); 6871da177e4SLinus Torvalds 6881da177e4SLinus Torvalds return err; 6891da177e4SLinus Torvalds } 6901da177e4SLinus Torvalds 6911da177e4SLinus Torvalds EXPORT_SYMBOL(generic_osync_inode); 6921da177e4SLinus Torvalds 6931da177e4SLinus Torvalds /** 6941da177e4SLinus Torvalds * writeback_acquire: attempt to get exclusive writeback access to a device 6951da177e4SLinus Torvalds * @bdi: the device's backing_dev_info structure 6961da177e4SLinus Torvalds * 6971da177e4SLinus Torvalds * It is a waste of resources to have more than one pdflush thread blocked on 6981da177e4SLinus Torvalds * a single request queue. Exclusion at the request_queue level is obtained 6991da177e4SLinus Torvalds * via a flag in the request_queue's backing_dev_info.state. 7001da177e4SLinus Torvalds * 7011da177e4SLinus Torvalds * Non-request_queue-backed address_spaces will share default_backing_dev_info, 7021da177e4SLinus Torvalds * unless they implement their own. Which is somewhat inefficient, as this 7031da177e4SLinus Torvalds * may prevent concurrent writeback against multiple devices. 7041da177e4SLinus Torvalds */ 7051da177e4SLinus Torvalds int writeback_acquire(struct backing_dev_info *bdi) 7061da177e4SLinus Torvalds { 7071da177e4SLinus Torvalds return !test_and_set_bit(BDI_pdflush, &bdi->state); 7081da177e4SLinus Torvalds } 7091da177e4SLinus Torvalds 7101da177e4SLinus Torvalds /** 7111da177e4SLinus Torvalds * writeback_in_progress: determine whether there is writeback in progress 7121da177e4SLinus Torvalds * @bdi: the device's backing_dev_info structure. 713b8887e6eSRandy Dunlap * 714b8887e6eSRandy Dunlap * Determine whether there is writeback in progress against a backing device. 7151da177e4SLinus Torvalds */ 7161da177e4SLinus Torvalds int writeback_in_progress(struct backing_dev_info *bdi) 7171da177e4SLinus Torvalds { 7181da177e4SLinus Torvalds return test_bit(BDI_pdflush, &bdi->state); 7191da177e4SLinus Torvalds } 7201da177e4SLinus Torvalds 7211da177e4SLinus Torvalds /** 7221da177e4SLinus Torvalds * writeback_release: relinquish exclusive writeback access against a device. 7231da177e4SLinus Torvalds * @bdi: the device's backing_dev_info structure 7241da177e4SLinus Torvalds */ 7251da177e4SLinus Torvalds void writeback_release(struct backing_dev_info *bdi) 7261da177e4SLinus Torvalds { 7271da177e4SLinus Torvalds BUG_ON(!writeback_in_progress(bdi)); 7281da177e4SLinus Torvalds clear_bit(BDI_pdflush, &bdi->state); 7291da177e4SLinus Torvalds } 730