Skip to contentSkip to conversation

fs: Fix missed inode write during fsync

PATCH Superseded Jan Kara 4 participants 20 messages May 11 – May 13

Subsystems (6)

v1 of 5 Superseded Latest is v5 →

lore
Share & tools
lore

Jan Kara

author reviewer
#0 lore raw

cover fs: Fix missed inode write during fsync

Hello,

this patch series fixes the possibly missing inode write during fsync(2) for filesystems using generic metadata bh tracking. The inherent problem is that .write_inode methods clear inode dirty bit but they only copy inode contents into to the buffer cache. Because buffer carrying the inode is shared among multiple inodes, it cannot be tracked by the generic metadata bh tracking infrastructure and thus nothing is tracking that buffer needs to be written out to maintain fsync(2) guarantees. Normally, this gets taken care of by .write_inode checking for WB_SYNC_ALL writeback and submitting & waiting for the buffer in that case however if flush worker ends up writing the inode before data integrity writeback, this mechanism is broken.

This patch series adds a way for filesystems to track metadata block number which contains the inode metadata and then uses this information to writeout the buffer on fsync.

Honza

author reviewer
#1 lore raw

PATCH 1/9 affs: Drop support for metadata bh tracking

AFFS did all the hard work of tracking metadata bhs dirtied for an inode but it actually never used this information as affs_file_fsync() just calls sync_blockdev() to writeback all filesystem metadata bhs. After a discussion with AFFS maintainer nobody cares about AFFS performance so let's keep this affs_file_fsync() behavior and just drop all the pointless tracking from AFFS.

CC David Sterba <[email protected]>
---
 fs/affs/affs.h     |  1 -
 fs/affs/amigaffs.c | 12 ++++++------
 fs/affs/file.c     | 25 +++++++++++--------------
 fs/affs/inode.c    | 13 +++++--------
 fs/affs/namei.c    |  9 ++++-----
 fs/affs/super.c    |  1 -
 6 files changed, 26 insertions(+), 35 deletions(-)
fs/affs/affs.h +0 -1
index a0caf6ace860..406a0ef63e7b 100644
@@ -44,7 +44,6 @@ struct affs_inode_info {
 	struct mutex i_link_lock;		/* Protects internal inode access. */
 	struct mutex i_ext_lock;		/* Protects internal inode access. */
 #define i_hash_lock i_ext_lock
-	struct mapping_metadata_bhs i_metadata_bhs;
 	u32	 i_blkcnt;			/* block count */
 	u32	 i_extcnt;			/* extended block count */
 	u32	*i_lc;				/* linear cache of extended blocks */
fs/affs/amigaffs.c +6 -6
index bed4fc805e8e..6cc0fc9a4cbf 100644
@@ -57,7 +57,7 @@ affs_insert_hash(struct inode *dir, struct buffer_head *bh)
 		AFFS_TAIL(sb, dir_bh)->hash_chain = cpu_to_be32(ino);
 
 	affs_adjust_checksum(dir_bh, ino);
-	mmb_mark_buffer_dirty(dir_bh, &AFFS_I(dir)->i_metadata_bhs);
+	mark_buffer_dirty(dir_bh);
 	affs_brelse(dir_bh);
 
 	inode_set_mtime_to_ts(dir, inode_set_ctime_current(dir));
@@ -100,7 +100,7 @@ affs_remove_hash(struct inode *dir, struct buffer_head *rem_bh)
 			else
 				AFFS_TAIL(sb, bh)->hash_chain = ino;
 			affs_adjust_checksum(bh, be32_to_cpu(ino) - hash_ino);
-			mmb_mark_buffer_dirty(bh, &AFFS_I(dir)->i_metadata_bhs);
+			mark_buffer_dirty(bh);
 			AFFS_TAIL(sb, rem_bh)->parent = 0;
 			retval = 0;
 			break;
@@ -180,7 +180,7 @@ affs_remove_link(struct dentry *dentry)
 			affs_unlock_dir(dir);
 			goto done;
 		}
-		mmb_mark_buffer_dirty(link_bh, &AFFS_I(inode)->i_metadata_bhs);
+		mark_buffer_dirty(link_bh);
 
 		memcpy(AFFS_TAIL(sb, bh)->name, AFFS_TAIL(sb, link_bh)->name, 32);
 		retval = affs_insert_hash(dir, bh);
@@ -188,7 +188,7 @@ affs_remove_link(struct dentry *dentry)
 			affs_unlock_dir(dir);
 			goto done;
 		}
-		mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+		mark_buffer_dirty(bh);
 
 		affs_unlock_dir(dir);
 		iput(dir);
@@ -203,7 +203,7 @@ affs_remove_link(struct dentry *dentry)
 			__be32 ino2 = AFFS_TAIL(sb, link_bh)->link_chain;
 			AFFS_TAIL(sb, bh)->link_chain = ino2;
 			affs_adjust_checksum(bh, be32_to_cpu(ino2) - link_ino);
-			mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+			mark_buffer_dirty(bh);
 			retval = 0;
 			/* Fix the link count, if bh is a normal header block without links */
 			switch (be32_to_cpu(AFFS_TAIL(sb, bh)->stype)) {
@@ -306,7 +306,7 @@ affs_remove_header(struct dentry *dentry)
 	retval = affs_remove_hash(dir, bh);
 	if (retval)
 		goto done_unlock;
-	mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(bh);
 
 	affs_unlock_dir(dir);
 
fs/affs/file.c +11 -14
index 144b17482d12..23e088a7ed4f 100644
@@ -140,14 +140,14 @@ affs_alloc_extblock(struct inode *inode, struct buffer_head *bh, u32 ext)
 	AFFS_TAIL(sb, new_bh)->parent = cpu_to_be32(inode->i_ino);
 	affs_fix_checksum(sb, new_bh);
 
-	mmb_mark_buffer_dirty(new_bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(new_bh);
 
 	tmp = be32_to_cpu(AFFS_TAIL(sb, bh)->extension);
 	if (tmp)
 		affs_warning(sb, "alloc_ext", "previous extension set (%x)", tmp);
 	AFFS_TAIL(sb, bh)->extension = cpu_to_be32(blocknr);
 	affs_adjust_checksum(bh, blocknr - tmp);
-	mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(bh);
 
 	AFFS_I(inode)->i_extcnt++;
 	mark_inode_dirty(inode);
@@ -581,7 +581,7 @@ affs_extent_file_ofs(struct inode *inode, u32 newsize)
 		memset(AFFS_DATA(bh) + boff, 0, tmp);
 		be32_add_cpu(&AFFS_DATA_HEAD(bh)->size, tmp);
 		affs_fix_checksum(sb, bh);
-		mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+		mark_buffer_dirty(bh);
 		size += tmp;
 		bidx++;
 	} else if (bidx) {
@@ -603,7 +603,7 @@ affs_extent_file_ofs(struct inode *inode, u32 newsize)
 		AFFS_DATA_HEAD(bh)->size = cpu_to_be32(tmp);
 		affs_fix_checksum(sb, bh);
 		bh->b_state &= ~(1UL << BH_New);
-		mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+		mark_buffer_dirty(bh);
 		if (prev_bh) {
 			u32 tmp_next = be32_to_cpu(AFFS_DATA_HEAD(prev_bh)->next);
 
@@ -613,8 +613,7 @@ affs_extent_file_ofs(struct inode *inode, u32 newsize)
 					     bidx, tmp_next);
 			AFFS_DATA_HEAD(prev_bh)->next = cpu_to_be32(bh->b_blocknr);
 			affs_adjust_checksum(prev_bh, bh->b_blocknr - tmp_next);
-			mmb_mark_buffer_dirty(prev_bh,
-					      &AFFS_I(inode)->i_metadata_bhs);
+			mark_buffer_dirty(prev_bh);
 			affs_brelse(prev_bh);
 		}
 		size += bsize;
@@ -733,7 +732,7 @@ static int affs_write_end_ofs(const struct kiocb *iocb,
 		AFFS_DATA_HEAD(bh)->size = cpu_to_be32(
 			max(boff + tmp, be32_to_cpu(AFFS_DATA_HEAD(bh)->size)));
 		affs_fix_checksum(sb, bh);
-		mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+		mark_buffer_dirty(bh);
 		written += tmp;
 		from += tmp;
 		bidx++;
@@ -766,13 +765,12 @@ static int affs_write_end_ofs(const struct kiocb *iocb,
 						     bidx, tmp_next);
 				AFFS_DATA_HEAD(prev_bh)->next = cpu_to_be32(bh->b_blocknr);
 				affs_adjust_checksum(prev_bh, bh->b_blocknr - tmp_next);
-				mmb_mark_buffer_dirty(prev_bh,
-					&AFFS_I(inode)->i_metadata_bhs);
+				mark_buffer_dirty(prev_bh);
 			}
 		}
 		affs_brelse(prev_bh);
 		affs_fix_checksum(sb, bh);
-		mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+		mark_buffer_dirty(bh);
 		written += bsize;
 		from += bsize;
 		bidx++;
@@ -801,14 +799,13 @@ static int affs_write_end_ofs(const struct kiocb *iocb,
 						     bidx, tmp_next);
 				AFFS_DATA_HEAD(prev_bh)->next = cpu_to_be32(bh->b_blocknr);
 				affs_adjust_checksum(prev_bh, bh->b_blocknr - tmp_next);
-				mmb_mark_buffer_dirty(prev_bh,
-						&AFFS_I(inode)->i_metadata_bhs);
+				mark_buffer_dirty(prev_bh);
 			}
 		} else if (be32_to_cpu(AFFS_DATA_HEAD(bh)->size) < tmp)
 			AFFS_DATA_HEAD(bh)->size = cpu_to_be32(tmp);
 		affs_brelse(prev_bh);
 		affs_fix_checksum(sb, bh);
-		mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+		mark_buffer_dirty(bh);
 		written += tmp;
 		from += tmp;
 		bidx++;
@@ -945,7 +942,7 @@ affs_truncate(struct inode *inode)
 	}
 	AFFS_TAIL(sb, ext_bh)->extension = 0;
 	affs_fix_checksum(sb, ext_bh);
-	mmb_mark_buffer_dirty(ext_bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(ext_bh);
 	affs_brelse(ext_bh);
 
 	if (inode->i_size) {
fs/affs/inode.c +5 -8
index 5dd1b016bcb0..d4a3f381c4bc 100644
@@ -206,7 +206,7 @@ affs_write_inode(struct inode *inode, struct writeback_control *wbc)
 		}
 	}
 	affs_fix_checksum(sb, bh);
-	mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(bh);
 	affs_brelse(bh);
 	affs_free_prealloc(inode);
 	return 0;
@@ -266,11 +266,8 @@ affs_evict_inode(struct inode *inode)
 	if (!inode->i_nlink) {
 		inode->i_size = 0;
 		affs_truncate(inode);
-	} else {
-		mmb_sync(&AFFS_I(inode)->i_metadata_bhs);
 	}
 
-	mmb_invalidate(&AFFS_I(inode)->i_metadata_bhs);
 	clear_inode(inode);
 	affs_free_prealloc(inode);
 	cache_page = (unsigned long)AFFS_I(inode)->i_lc;
@@ -305,7 +302,7 @@ affs_new_inode(struct inode *dir)
 	bh = affs_getzeroblk(sb, block);
 	if (!bh)
 		goto err_bh;
-	mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(bh);
 	affs_brelse(bh);
 
 	inode->i_uid     = current_fsuid();
@@ -393,17 +390,17 @@ affs_add_entry(struct inode *dir, struct inode *inode, struct dentry *dentry, s3
 		AFFS_TAIL(sb, bh)->link_chain = chain;
 		AFFS_TAIL(sb, inode_bh)->link_chain = cpu_to_be32(block);
 		affs_adjust_checksum(inode_bh, block - be32_to_cpu(chain));
-		mmb_mark_buffer_dirty(inode_bh, &AFFS_I(inode)->i_metadata_bhs);
+		mark_buffer_dirty(inode_bh);
 		set_nlink(inode, 2);
 		ihold(inode);
 	}
 	affs_fix_checksum(sb, bh);
-	mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(bh);
 	dentry->d_fsdata = (void *)(long)bh->b_blocknr;
 
 	affs_lock_dir(dir);
 	retval = affs_insert_hash(dir, bh);
-	mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(bh);
 	affs_unlock_dir(dir);
 	affs_unlock_link(inode);
 
fs/affs/namei.c +4 -5
index c3c6532da4b0..57d8d755aada 100644
@@ -373,7 +373,7 @@ affs_symlink(struct mnt_idmap *idmap, struct inode *dir,
 	}
 	*p = 0;
 	inode->i_size = i + 1;
-	mmb_mark_buffer_dirty(bh, &AFFS_I(inode)->i_metadata_bhs);
+	mark_buffer_dirty(bh);
 	affs_brelse(bh);
 	mark_inode_dirty(inode);
 
@@ -443,8 +443,7 @@ affs_rename(struct inode *old_dir, struct dentry *old_dentry,
 	/* TODO: move it back to old_dir, if error? */
 
 done:
-	mmb_mark_buffer_dirty(bh,
-			&AFFS_I(retval ? old_dir : new_dir)->i_metadata_bhs);
+	mark_buffer_dirty(bh);
 	affs_brelse(bh);
 	return retval;
 }
@@ -497,8 +496,8 @@ affs_xrename(struct inode *old_dir, struct dentry *old_dentry,
 	retval = affs_insert_hash(old_dir, bh_new);
 	affs_unlock_dir(old_dir);
 done:
-	mmb_mark_buffer_dirty(bh_old, &AFFS_I(new_dir)->i_metadata_bhs);
-	mmb_mark_buffer_dirty(bh_new, &AFFS_I(old_dir)->i_metadata_bhs);
+	mark_buffer_dirty(bh_old);
+	mark_buffer_dirty(bh_new);
 	affs_brelse(bh_old);
 	affs_brelse(bh_new);
 	return retval;
fs/affs/super.c +0 -1
index 079f36e1ddec..8451647f3fea 100644
@@ -108,7 +108,6 @@ static struct inode *affs_alloc_inode(struct super_block *sb)
 	i->i_lc = NULL;
 	i->i_ext_bh = NULL;
 	i->i_pa_cnt = 0;
-	mmb_init(&i->i_metadata_bhs, &i->vfs_inode.i_data);
 
 	return &i->vfs_inode;
 }
Signature · 1 line
2.51.0
author reviewer
#2 lore raw

PATCH 2/9 ext4: Allocate mapping_metadata_bhs struct on demand

In reply to Jan Kara

Currently every ext4 inode gets mapping_metadata_bhs struct although it is only needed when running without a journal and only for inodes where any metadata was dirtied. Allocate mapping_metadata_bhs struct on demand when dirtying the first metadata buffer for the inode.

---
 fs/ext4/ext4.h      |  2 +-
 fs/ext4/ext4_jbd2.c | 24 +++++++++++++++++++++---
 fs/ext4/fsync.c     | 12 ++++++++----
 fs/ext4/inode.c     |  9 +++++----
 fs/ext4/super.c     |  8 +++++---
 5 files changed, 40 insertions(+), 15 deletions(-)
fs/ext4/ext4.h +1 -1
index 94283a991e5c..6bb29a20420f 100644
@@ -1117,7 +1117,7 @@ struct ext4_inode_info {
 	struct rw_semaphore i_data_sem;
 	struct inode vfs_inode;
 	struct jbd2_inode *jinode;
-	struct mapping_metadata_bhs i_metadata_bhs;
+	struct mapping_metadata_bhs *i_metadata_bhs;
 
 	/*
 	 * File creation time. Its function is same as that of
fs/ext4/ext4_jbd2.c +21 -3
index 9a8c225f2753..74f05bd0cdde 100644
@@ -350,6 +350,21 @@ int __ext4_journal_get_create_access(const char *where, unsigned int line,
 	return 0;
 }
 
+static void ext4_inode_attach_mmb(struct inode *inode)
+{
+	struct mapping_metadata_bhs *mmb;
+
+	/*
+	 * It's difficult to handle failure when marking buffer dirty without
+	 * leaving filesystem corrupyted
+	 */
+	mmb = kmalloc_obj(*mmb, GFP_KERNEL | __GFP_NOFAIL);
+	mmb_init(mmb, inode->i_mapping);
+	/* Someone swapped another mmb before us? */
+	if (cmpxchg(&EXT4_I(inode)->i_metadata_bhs, NULL, mmb))
+		kfree(mmb);
+}
+
 int __ext4_handle_dirty_metadata(const char *where, unsigned int line,
 				 handle_t *handle, struct inode *inode,
 				 struct buffer_head *bh)
@@ -389,11 +404,14 @@ int __ext4_handle_dirty_metadata(const char *where, unsigned int line,
 					 err);
 		}
 	} else {
-		if (inode)
+		if (inode) {
+			if (!EXT4_I(inode)->i_metadata_bhs)
+				ext4_inode_attach_mmb(inode);
 			mmb_mark_buffer_dirty(bh,
-					      &EXT4_I(inode)->i_metadata_bhs);
-		else
+					      EXT4_I(inode)->i_metadata_bhs);
+		} else {
 			mark_buffer_dirty(bh);
+		}
 		if (inode && inode_needs_sync(inode)) {
 			sync_dirty_buffer(bh);
 			if (buffer_req(bh) && !buffer_uptodate(bh)) {
fs/ext4/fsync.c +8 -4
index 924726dcc85f..e25d365e1179 100644
@@ -46,6 +46,7 @@
 static int ext4_sync_parent(struct inode *inode)
 {
 	struct dentry *dentry, *next;
+	struct mapping_metadata_bhs *mmb;
 	int ret = 0;
 
 	if (!ext4_test_inode_state(inode, EXT4_STATE_NEWENTRY))
@@ -68,9 +69,12 @@ static int ext4_sync_parent(struct inode *inode)
 		 * through ext4_evict_inode()) and so we are safe to flush
 		 * metadata blocks and the inode.
 		 */
-		ret = mmb_sync(&EXT4_I(inode)->i_metadata_bhs);
-		if (ret)
-			break;
+		mmb = READ_ONCE(EXT4_I(inode)->i_metadata_bhs);
+		if (mmb) {
+			ret = mmb_sync(mmb);
+			if (ret)
+				break;
+		}
 		ret = sync_inode_metadata(inode, 1);
 		if (ret)
 			break;
@@ -89,7 +93,7 @@ static int ext4_fsync_nojournal(struct file *file, loff_t start, loff_t end,
 	};
 	int ret;
 
-	ret = mmb_fsync_noflush(file, &EXT4_I(inode)->i_metadata_bhs,
+	ret = mmb_fsync_noflush(file, EXT4_I(inode)->i_metadata_bhs,
 				start, end, datasync);
 	if (ret)
 		return ret;
fs/ext4/inode.c +5 -4
index c2c2d6ac7f3d..3e66e9510909 100644
@@ -195,9 +195,8 @@ void ext4_evict_inode(struct inode *inode)
 			ext4_warning_inode(inode, "data will be lost");
 
 		truncate_inode_pages_final(&inode->i_data);
-		/* Avoid mballoc special inode which has no proper iops */
-		if (!EXT4_SB(inode->i_sb)->s_journal)
-			mmb_sync(&EXT4_I(inode)->i_metadata_bhs);
+		if (EXT4_I(inode)->i_metadata_bhs)
+			mmb_sync(EXT4_I(inode)->i_metadata_bhs);
 		goto no_delete;
 	}
 
@@ -3451,6 +3450,7 @@ static bool ext4_release_folio(struct folio *folio, gfp_t wait)
 static bool ext4_inode_datasync_dirty(struct inode *inode)
 {
 	journal_t *journal = EXT4_SB(inode->i_sb)->s_journal;
+	struct mapping_metadata_bhs *mmb;
 
 	if (journal) {
 		if (jbd2_transaction_committed(journal,
@@ -3461,8 +3461,9 @@ static bool ext4_inode_datasync_dirty(struct inode *inode)
 		return true;
 	}
 
+	mmb = READ_ONCE(EXT4_I(inode)->i_metadata_bhs);
 	/* Any metadata buffers to write? */
-	if (mmb_has_buffers(&EXT4_I(inode)->i_metadata_bhs))
+	if (mmb && mmb_has_buffers(mmb))
 		return true;
 	return inode_state_read_once(inode) & I_DIRTY_DATASYNC;
 }
fs/ext4/super.c +5 -3
index 6a77db4d3124..92134ea4620c 100644
@@ -1430,7 +1430,7 @@ static struct inode *ext4_alloc_inode(struct super_block *sb)
 	INIT_WORK(&ei->i_rsv_conversion_work, ext4_end_io_rsv_work);
 	ext4_fc_init_inode(&ei->vfs_inode);
 	spin_lock_init(&ei->i_fc_lock);
-	mmb_init(&ei->i_metadata_bhs, &ei->vfs_inode.i_data);
+	ei->i_metadata_bhs = NULL;
 	return &ei->vfs_inode;
 }
 
@@ -1527,8 +1527,10 @@ static void destroy_inodecache(void)
 void ext4_clear_inode(struct inode *inode)
 {
 	ext4_fc_del(inode);
-	if (!EXT4_SB(inode->i_sb)->s_journal)
-		mmb_invalidate(&EXT4_I(inode)->i_metadata_bhs);
+	if (EXT4_I(inode)->i_metadata_bhs) {
+		mmb_invalidate(EXT4_I(inode)->i_metadata_bhs);
+		kfree(EXT4_I(inode)->i_metadata_bhs);
+	}
 	clear_inode(inode);
 	ext4_discard_preallocations(inode);
 	/*
Signature · 1 line
2.51.0
author reviewer
#3 lore raw

PATCH 3/9 fs: Writeout inode buffer from mmb_sync()

In reply to Jan Kara

Currently metadata bh tracking does not track inode buffers because they are usually shared by several inodes and so our linked list tracking cannot be used. On fsync we call sync_inode_metadata() to write inode instead where filesystems' .write_inode methods detect data integrity writeback and take care to submit inode buffer to disk and wait for it in that case. This is however racy as for example flush worker can submit normal (WB_SYNC_NONE) inode writeback first, which makes the inode clean and copies the inode to the buffer but doesn't submit the buffer for IO. Thus sync_inode_metadata() call does nothing and we fail to persist inode buffer to disk on fsync(2).

Fix the problem by allowing filesystem to set the number of block backing the inode in mmb structure and mmb_sync() then takes care to writeout corresponding buffer and wait for it.

---
 fs/buffer.c        | 34 +++++++++++++++++++++++-----------
 include/linux/fs.h |  1 +
 2 files changed, 24 insertions(+), 11 deletions(-)
fs/buffer.c +23 -11
index b0b3792b1496..dba29a45346b 100644
@@ -477,12 +477,14 @@ EXPORT_SYMBOL(mark_buffer_async_write);
  * using RCU, grab the lock, verify we didn't race with somebody detaching the
  * bh / moving it to different inode and only then proceeding.
  */
+#define INVALID_BLK (~0ULL)
 
 void mmb_init(struct mapping_metadata_bhs *mmb, struct address_space *mapping)
 {
 	spin_lock_init(&mmb->lock);
 	INIT_LIST_HEAD(&mmb->list);
 	mmb->mapping = mapping;
+	mmb->inode_blk = INVALID_BLK;
 }
 EXPORT_SYMBOL(mmb_init);
 
@@ -593,8 +595,18 @@ int mmb_sync(struct mapping_metadata_bhs *mmb)
 			}
 		}
 	}
-
 	spin_unlock(&mmb->lock);
+
+	/* Writeout inode buffer head */
+	if (mmb->inode_blk != INVALID_BLK) {
+		bh = sb_find_get_block(mmb->mapping->host->i_sb, mmb->inode_blk);
+		write_dirty_buffer(bh, REQ_SYNC);
+		wait_on_buffer(bh);
+		if (!buffer_uptodate(bh))
+			err = -EIO;
+		brelse(bh);
+	}
+
 	blk_finish_plug(&plug);
 	spin_lock(&mmb->lock);
 
@@ -646,18 +658,18 @@ int mmb_fsync_noflush(struct file *file, struct mapping_metadata_bhs *mmb,
 	if (err)
 		return err;
 
-	if (mmb)
-		ret = mmb_sync(mmb);
 	if (!(inode_state_read_once(inode) & I_DIRTY_ALL))
-		goto out;
+		goto sync_buffers;
 	if (datasync && !(inode_state_read_once(inode) & I_DIRTY_DATASYNC))
-		goto out;
-
-	err = sync_inode_metadata(inode, 1);
-	if (ret == 0)
-		ret = err;
-
-out:
+		goto sync_buffers;
+
+	ret = sync_inode_metadata(inode, 1);
+sync_buffers:
+	if (mmb) {
+		err = mmb_sync(mmb);
+		if (ret == 0)
+			ret = err;
+	}
 	/* check and advance again to catch errors after syncing out buffers */
 	err = file_check_and_advance_wb_err(file);
 	if (ret == 0)
include/linux/fs.h +1 -0
index 11559c513dfb..435a41e4c90f 100644
@@ -446,6 +446,7 @@ extern const struct address_space_operations empty_aops;
 /* Structure for tracking metadata buffer heads associated with the mapping */
 struct mapping_metadata_bhs {
 	struct address_space *mapping;	/* Mapping bhs are associated with */
+	sector_t inode_blk;	/* Number of block containing the inode */
 	spinlock_t lock;	/* Lock protecting bh list */
 	struct list_head list;	/* The list of bhs (b_assoc_buffers) */
 };
Signature · 1 line
2.51.0
#4 lore raw

Re: [PATCH 3/9] fs: Writeout inode buffer from mmb_sync()

Quoted 94 lines · Jan Kara

Currently metadata bh tracking does not track inode buffers because they are usually shared by several inodes and so our linked list tracking cannot be used. On fsync we call sync_inode_metadata() to write inode instead where filesystems' .write_inode methods detect data integrity writeback and take care to submit inode buffer to disk and wait for it in that case. This is however racy as for example flush worker can submit normal (WB_SYNC_NONE) inode writeback first, which makes the inode clean and copies the inode to the buffer but doesn't submit the buffer for IO. Thus sync_inode_metadata() call does nothing and we fail to persist inode buffer to disk on fsync(2).

Fix the problem by allowing filesystem to set the number of block backing the inode in mmb structure and mmb_sync() then takes care to writeout corresponding buffer and wait for it.

---
 fs/buffer.c        | 34 +++++++++++++++++++++++-----------
 include/linux/fs.h |  1 +
 2 files changed, 24 insertions(+), 11 deletions(-)
fs/buffer.c +23 -11
index b0b3792b1496..dba29a45346b 100644
@@ -477,12 +477,14 @@ EXPORT_SYMBOL(mark_buffer_async_write);
  * using RCU, grab the lock, verify we didn't race with somebody detaching the
  * bh / moving it to different inode and only then proceeding.
  */
+#define INVALID_BLK (~0ULL)
 
 void mmb_init(struct mapping_metadata_bhs *mmb, struct address_space *mapping)
 {
 	spin_lock_init(&mmb->lock);
 	INIT_LIST_HEAD(&mmb->list);
 	mmb->mapping = mapping;
+	mmb->inode_blk = INVALID_BLK;
 }
 EXPORT_SYMBOL(mmb_init);
 
@@ -593,8 +595,18 @@ int mmb_sync(struct mapping_metadata_bhs *mmb)
 			}
 		}
 	}
-
 	spin_unlock(&mmb->lock);
+
+	/* Writeout inode buffer head */
+	if (mmb->inode_blk != INVALID_BLK) {
+		bh = sb_find_get_block(mmb->mapping->host->i_sb, mmb->inode_blk);
+		write_dirty_buffer(bh, REQ_SYNC);
+		wait_on_buffer(bh);
+		if (!buffer_uptodate(bh))
+			err = -EIO;
+		brelse(bh);
+	}
+
 	blk_finish_plug(&plug);
 	spin_lock(&mmb->lock);
 
@@ -646,18 +658,18 @@ int mmb_fsync_noflush(struct file *file, struct mapping_metadata_bhs *mmb,
 	if (err)
 		return err;
 
-	if (mmb)
-		ret = mmb_sync(mmb);
 	if (!(inode_state_read_once(inode) & I_DIRTY_ALL))
-		goto out;
+		goto sync_buffers;
 	if (datasync && !(inode_state_read_once(inode) & I_DIRTY_DATASYNC))
-		goto out;
-
-	err = sync_inode_metadata(inode, 1);
-	if (ret == 0)
-		ret = err;
-
-out:
+		goto sync_buffers;
+
+	ret = sync_inode_metadata(inode, 1);
+sync_buffers:
+	if (mmb) {
+		err = mmb_sync(mmb);
+		if (ret == 0)
+			ret = err;
+	}
 	/* check and advance again to catch errors after syncing out buffers */
 	err = file_check_and_advance_wb_err(file);
 	if (ret == 0)
include/linux/fs.h +0 -0
index 11559c513dfb..435a41e4c90f 100644

include/linux/fs.h extern const struct address_space_operations empty_aops; · line 449

Quoted 4 lines
include/linux/fs.h +1 -0
@@ -446,6 +446,7 @@ extern const struct address_space_operations empty_aops;
 /* Structure for tracking metadata buffer heads associated with the mapping */
 struct mapping_metadata_bhs {
 	struct address_space *mapping;	/* Mapping bhs are associated with */
+	sector_t inode_blk;	/* Number of block containing the inode */

This is great, thanks!

author reviewer
#5 lore raw

PATCH 4/9 ext2: Fix possibly missing inode write on fsync(2)

In reply to Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's inode table block on fsync(2).

---
 fs/ext2/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/ext2/inode.c +2 -1
index 74aca5eb572d..6ce832da944f 100644
@@ -1612,6 +1612,7 @@ static int __ext2_write_inode(struct inode *inode, int do_sync)
 	} else for (n = 0; n < EXT2_N_BLOCKS; n++)
 		raw_inode->i_block[n] = ei->i_data[n];
 	mark_buffer_dirty(bh);
+	ei->i_metadata_bhs.inode_blk = bh->b_blocknr;
 	if (do_sync) {
 		sync_dirty_buffer(bh);
 		if (buffer_req(bh) && !buffer_uptodate(bh)) {
@@ -1627,7 +1628,7 @@ static int __ext2_write_inode(struct inode *inode, int do_sync)
 
 int ext2_write_inode(struct inode *inode, struct writeback_control *wbc)
 {
-	return __ext2_write_inode(inode, wbc->sync_mode == WB_SYNC_ALL);
+	return __ext2_write_inode(inode, 0);
 }
 
 int ext2_getattr(struct mnt_idmap *idmap, const struct path *path,
Signature · 1 line
2.51.0
author reviewer
#6 lore raw

PATCH 5/9 udf: Fix possibly missing inode write on fsync(2)

In reply to Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's block on fsync(2).

---
 fs/udf/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/udf/inode.c +2 -1
index 67bcf83758c8..4102d3482319 100644
@@ -1707,7 +1707,7 @@ void udf_update_extra_perms(struct inode *inode, umode_t mode)
 
 int udf_write_inode(struct inode *inode, struct writeback_control *wbc)
 {
-	return udf_update_inode(inode, wbc->sync_mode == WB_SYNC_ALL);
+	return udf_update_inode(inode, 0);
 }
 
 static int udf_sync_inode(struct inode *inode)
@@ -1937,6 +1937,7 @@ static int udf_update_inode(struct inode *inode, int do_sync)
 
 	/* write the data blocks */
 	mark_buffer_dirty(bh);
+	iinfo->i_metadata_bhs.inode_blk = bh->b_blocknr;
 	if (do_sync) {
 		sync_dirty_buffer(bh);
 		if (buffer_write_io_error(bh)) {
Signature · 1 line
2.51.0
author reviewer
#7 lore raw

PATCH 6/9 fat: Fix possibly missing inode write on fsync(2)

In reply to Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's buffer on fsync(2).

---
 fs/fat/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/fat/inode.c +2 -1
index 28f78df086ef..4ca00b7a618b 100644
@@ -907,6 +907,7 @@ static int __fat_write_inode(struct inode *inode, int wait)
 	}
 	spin_unlock(&sbi->inode_hash_lock);
 	mark_buffer_dirty(bh);
+	MSDOS_I(inode)->i_metadata_bhs.inode_blk = bh->b_blocknr;
 	err = 0;
 	if (wait)
 		err = sync_dirty_buffer(bh);
@@ -925,7 +926,7 @@ static int fat_write_inode(struct inode *inode, struct writeback_control *wbc)
 		err = fat_clusters_flush(sb);
 		mutex_unlock(&MSDOS_SB(sb)->s_lock);
 	} else
-		err = __fat_write_inode(inode, wbc->sync_mode == WB_SYNC_ALL);
+		err = __fat_write_inode(inode, 0);
 
 	return err;
 }
Signature · 1 line
2.51.0
maintainer
#8 lore raw

Re: [PATCH 6/9] fat: Fix possibly missing inode write on fsync(2)

Quoted 13 lines · Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's buffer on fsync(2).

---
 fs/fat/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/fat/inode.c +0 -0
index 28f78df086ef..4ca00b7a618b 100644

fs/fat/inode.c static int __fat_write_inode(struct inode *inode, int wait) · line 910

Quoted 4 lines
fs/fat/inode.c +1 -0
@@ -907,6 +907,7 @@ static int __fat_write_inode(struct inode *inode, int wait)
 	}
 	spin_unlock(&sbi->inode_hash_lock);
 	mark_buffer_dirty(bh);
+	MSDOS_I(inode)->i_metadata_bhs.inode_blk = bh->b_blocknr;

When inode position was changed/removed, this will point the wrong block. And maybe sync a unrelated block and wait.

Quoted 12 lines
fs/fat/inode.c +1 -1
 	err = 0;
 	if (wait)
 		err = sync_dirty_buffer(bh);
@@ -925,7 +926,7 @@ static int fat_write_inode(struct inode *inode, struct writeback_control *wbc)
 		err = fat_clusters_flush(sb);
 		mutex_unlock(&MSDOS_SB(sb)->s_lock);
 	} else
-		err = __fat_write_inode(inode, wbc->sync_mode == WB_SYNC_ALL);
+		err = __fat_write_inode(inode, 0);
 
 	return err;
 }
Signature · 1 line
OGAWA Hirofumi <[email protected]>
author reviewer
#9 lore raw

Re: [PATCH 6/9] fat: Fix possibly missing inode write on fsync(2)

Quoted 10 lines · OGAWA Hirofumi
Quoted 8 lines · Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's buffer on fsync(2).

---
 fs/fat/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
Quoted 12 lines

fs/fat/inode.c static int __fat_write_inode(struct inode *inode, int wait) · line 910

Quoted 9 lines
fs/fat/inode.c +1 -0
index 28f78df086ef..4ca00b7a618b 100644
@@ -907,6 +907,7 @@ static int __fat_write_inode(struct inode *inode, int wait)
 	}
 	spin_unlock(&sbi->inode_hash_lock);
 	mark_buffer_dirty(bh);
+	MSDOS_I(inode)->i_metadata_bhs.inode_blk = bh->b_blocknr;

When inode position was changed/removed, this will point the wrong block. And maybe sync a unrelated block and wait.

So I didn't realize that e.g. rename does change the backing inode block. But given we set i_metadata_bhs.inode_blk on each inode write, inode_blk should always contain the current position where the inode was written so fsync should be syncing the right block. Or am I still missing something?

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
maintainer
#10 lore raw

Re: [PATCH 6/9] fat: Fix possibly missing inode write on fsync(2)

Quoted 24 lines · Jan Kara
Quoted 10 lines · OGAWA Hirofumi
Quoted 8 lines · Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's buffer on fsync(2).

---
 fs/fat/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
Quoted 12 lines

fs/fat/inode.c static int __fat_write_inode(struct inode *inode, int wait) · line 910

Quoted 9 lines
fs/fat/inode.c +1 -0
index 28f78df086ef..4ca00b7a618b 100644
@@ -907,6 +907,7 @@ static int __fat_write_inode(struct inode *inode, int wait)
 	}
 	spin_unlock(&sbi->inode_hash_lock);
 	mark_buffer_dirty(bh);
+	MSDOS_I(inode)->i_metadata_bhs.inode_blk = bh->b_blocknr;

When inode position was changed/removed, this will point the wrong block. And maybe sync a unrelated block and wait.

Quoted 4 lines

So I didn't realize that e.g. rename does change the backing inode block. But given we set i_metadata_bhs.inode_blk on each inode write, inode_blk should always contain the current position where the inode was written so fsync should be syncing the right block. Or am I still missing something?

I didn't check the case of rename completely, just recalled it when I saw this code, need confirm/check. But at least, the case of remove will leave it even after the block is reused.

Signature · 1 line
OGAWA Hirofumi <[email protected]>
author reviewer
#11 lore raw

Re: [PATCH 6/9] fat: Fix possibly missing inode write on fsync(2)

Quoted 26 lines · OGAWA Hirofumi
Quoted 24 lines · Jan Kara
Quoted 22 lines · OGAWA Hirofumi
Quoted 13 lines · Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's buffer on fsync(2).

---
 fs/fat/inode.c | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)
fs/fat/inode.c +0 -0
index 28f78df086ef..4ca00b7a618b 100644

fs/fat/inode.c static int __fat_write_inode(struct inode *inode, int wait) · line 910

Quoted 4 lines
fs/fat/inode.c +1 -0
@@ -907,6 +907,7 @@ static int __fat_write_inode(struct inode *inode, int wait)
 	}
 	spin_unlock(&sbi->inode_hash_lock);
 	mark_buffer_dirty(bh);
+	MSDOS_I(inode)->i_metadata_bhs.inode_blk = bh->b_blocknr;

When inode position was changed/removed, this will point the wrong block. And maybe sync a unrelated block and wait.

Quoted 8 lines
Quoted 4 lines

So I didn't realize that e.g. rename does change the backing inode block. But given we set i_metadata_bhs.inode_blk on each inode write, inode_blk should always contain the current position where the inode was written so fsync should be syncing the right block. Or am I still missing something?

I didn't check the case of rename completely, just recalled it when I saw this code, need confirm/check. But at least, the case of remove will leave it even after the block is reused.

Right. fat_detach() should set i_metadata_bhs.inode_blk to INVALID_BLK, thanks for catching that. I was thinking whether we should set i_metadata_bhs.inode_blk in fat_attach() instead of during inode dirtying. It would be somewhat more obviously correct but it could lead to unnecessary flushing in case the directory block gets dirtied by some other entry in it while the inode we are fsyncing got never dirtied. IMHO that's a sensible tradeoff so I'd do that but what is your opinion?

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
maintainer
#12 lore raw

Re: [PATCH 6/9] fat: Fix possibly missing inode write on fsync(2)

Quoted 11 lines · Jan Kara
Quoted 3 lines

I didn't check the case of rename completely, just recalled it when I saw this code, need confirm/check. But at least, the case of remove will leave it even after the block is reused.

Right. fat_detach() should set i_metadata_bhs.inode_blk to INVALID_BLK, thanks for catching that. I was thinking whether we should set i_metadata_bhs.inode_blk in fat_attach() instead of during inode dirtying. It would be somewhat more obviously correct but it could lead to unnecessary flushing in case the directory block gets dirtied by some other entry in it while the inode we are fsyncing got never dirtied. IMHO that's a sensible tradeoff so I'd do that but what is your opinion?

IMO, the marker should be cleared like b_assoc_buffers or I_DIRTY_* flags after each sync. Otherwise, because the block is shared with other inodes, it would sync/wait the unrelated dirty easily.

[And more serious implementation, looks like it should be cleared at similar points or such with b_assoc_buffers is cleared to minimize unrelated sync/wait.]

Signature · 1 line
OGAWA Hirofumi <[email protected]>
author reviewer
#13 lore raw

Re: [PATCH 6/9] fat: Fix possibly missing inode write on fsync(2)

Quoted 14 lines · OGAWA Hirofumi
Quoted 11 lines · Jan Kara
Quoted 3 lines

I didn't check the case of rename completely, just recalled it when I saw this code, need confirm/check. But at least, the case of remove will leave it even after the block is reused.

Right. fat_detach() should set i_metadata_bhs.inode_blk to INVALID_BLK, thanks for catching that. I was thinking whether we should set i_metadata_bhs.inode_blk in fat_attach() instead of during inode dirtying. It would be somewhat more obviously correct but it could lead to unnecessary flushing in case the directory block gets dirtied by some other entry in it while the inode we are fsyncing got never dirtied. IMHO that's a sensible tradeoff so I'd do that but what is your opinion?

Quoted 3 lines

IMO, the marker should be cleared like b_assoc_buffers or I_DIRTY_* flags after each sync. Otherwise, because the block is shared with other inodes, it would sync/wait the unrelated dirty easily.

Well, even if we do that, we should still clear inode_blk in fat_detach() AFAICT as nobody has to sync the inode before unlinking it.

Regarding clearing of inode_blk in mmb_sync() - yes, that makes sense could be done although then we have to be careful about races of mmb_sync() with .write_inode so that always guarantee mmb_sync() after .write_inode will persist the buffer. I'll see how complex that will get.

Quoted 3 lines

[And more serious implementation, looks like it should be cleared at similar points or such with b_assoc_buffers is cleared to minimize unrelated sync/wait.]

OTOH this doesn't really make much sense. We need to handle b_assoc_buffers when bh is getting reclaimed so that we can free the bh. We deliberately track inode block number and not bh pointer in mapping_metadata_bhs so that bhs backing inodes can be freed independently as it's infeasible to track down all mapping_metadata_bhs structs that might be referencing this bh.

And yes, I'm aware that in some corner cases the simple tracking can result in mmb_sync() writing out inode buffer although the inode itself was already persisted but exact tracking is way too expensive and not worth it for simple filesystems using this infrastructure. If we implement clearing of inode_blk in mmb_sync(), then we'll write out inode buffer unnecessarily at most once after the inode gets dirty which is IMO a reasonable middle ground between complexity of the tracking and unnecessary writeback.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
author reviewer
#14 lore raw

PATCH 7/9 minix: Fix possibly missing inode write on fsync(2)

In reply to Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's buffer on fsync(2).

---
 fs/minix/inode.c | 9 +--------
 1 file changed, 1 insertion(+), 8 deletions(-)
fs/minix/inode.c +1 -8
index 9c6bac248907..e3e05c9308bd 100644
@@ -693,14 +693,7 @@ static int minix_write_inode(struct inode *inode, struct writeback_control *wbc)
 		bh = V2_minix_update_inode(inode);
 	if (!bh)
 		return -EIO;
-	if (wbc->sync_mode == WB_SYNC_ALL && buffer_dirty(bh)) {
-		sync_dirty_buffer(bh);
-		if (buffer_req(bh) && !buffer_uptodate(bh)) {
-			printk("IO error syncing minix inode [%s:%08llx]\n",
-				inode->i_sb->s_id, inode->i_ino);
-			err = -EIO;
-		}
-	}
+	minix_i(inode)->i_metadata_bhs.inode_blk = bh->b_blocknr;
 	brelse (bh);
 	return err;
 }
Signature · 1 line
2.51.0
author reviewer
#15 lore raw

PATCH 8/9 bfs: Fix possibly missing inode write on fsync(2)

In reply to Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's buffer on fsync(2).

---
 fs/bfs/inode.c | 6 +-----
 1 file changed, 1 insertion(+), 5 deletions(-)
fs/bfs/inode.c +1 -5
index 19e49c8cf750..16d351b2f122 100644
@@ -165,11 +165,7 @@ static int bfs_write_inode(struct inode *inode, struct writeback_control *wbc)
 	di->i_eoffset = cpu_to_le32(i_sblock * BFS_BSIZE + inode->i_size - 1);
 
 	mark_buffer_dirty(bh);
-	if (wbc->sync_mode == WB_SYNC_ALL) {
-		sync_dirty_buffer(bh);
-		if (buffer_req(bh) && !buffer_uptodate(bh))
-			err = -EIO;
-	}
+	BFS_I(inode)->i_metadata_bhs.inode_blk = bh->b_blocknr;
 	brelse(bh);
 	mutex_unlock(&info->bfs_lock);
 	return err;
Signature · 1 line
2.51.0
author reviewer
#16 lore raw

PATCH 9/9 ext4: Use mmb infrastructure for inode buffer writeout

In reply to Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's inode table block on fsync(2) in nojournal mode (from ext4_sync_parent() and ext4_fsync_nojournal()). This significantly simplifies the code as we don't have to explicitely handle inode buffer writeback in ext4_write_inode() and thus we can also remove sync_inode_metadata() calls from ext4_sync_parent() and ext4_write_inode() call from ext4_fsync_nojournal().

---
 fs/ext4/ext4_jbd2.c |  2 +-
 fs/ext4/ext4_jbd2.h |  2 ++
 fs/ext4/fsync.c     | 12 ------------
 fs/ext4/inode.c     | 24 +++++-------------------
 4 files changed, 8 insertions(+), 32 deletions(-)
fs/ext4/ext4_jbd2.c +1 -1
index 74f05bd0cdde..6bbaf72108fd 100644
@@ -350,7 +350,7 @@ int __ext4_journal_get_create_access(const char *where, unsigned int line,
 	return 0;
 }
 
-static void ext4_inode_attach_mmb(struct inode *inode)
+void ext4_inode_attach_mmb(struct inode *inode)
 {
 	struct mapping_metadata_bhs *mmb;
 
fs/ext4/ext4_jbd2.h +2 -0
index 63d17c5201b5..2a01b8279c88 100644
@@ -122,6 +122,8 @@
 #define EXT4_HT_EXT_CONVERT     11
 #define EXT4_HT_MAX             12
 
+void ext4_inode_attach_mmb(struct inode *inode);
+
 int
 ext4_mark_iloc_dirty(handle_t *handle,
 		     struct inode *inode,
fs/ext4/fsync.c +0 -12
index e25d365e1179..af84489e57c6 100644
@@ -75,9 +75,6 @@ static int ext4_sync_parent(struct inode *inode)
 			if (ret)
 				break;
 		}
-		ret = sync_inode_metadata(inode, 1);
-		if (ret)
-			break;
 	}
 	dput(dentry);
 	return ret;
@@ -87,10 +84,6 @@ static int ext4_fsync_nojournal(struct file *file, loff_t start, loff_t end,
 				int datasync, bool *needs_barrier)
 {
 	struct inode *inode = file->f_inode;
-	struct writeback_control wbc = {
-		.sync_mode = WB_SYNC_ALL,
-		.nr_to_write = 0,
-	};
 	int ret;
 
 	ret = mmb_fsync_noflush(file, EXT4_I(inode)->i_metadata_bhs,
@@ -98,11 +91,6 @@ static int ext4_fsync_nojournal(struct file *file, loff_t start, loff_t end,
 	if (ret)
 		return ret;
 
-	/* Force writeout of inode table buffer to disk */
-	ret = ext4_write_inode(inode, &wbc);
-	if (ret)
-		return ret;
-
 	ret = ext4_sync_parent(inode);
 
 	if (test_opt(inode->i_sb, BARRIER))
fs/ext4/inode.c +5 -19
index 3e66e9510909..09506b4de1b2 100644
@@ -5786,24 +5786,6 @@ int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
 
 		err = ext4_fc_commit(EXT4_SB(inode->i_sb)->s_journal,
 						EXT4_I(inode)->i_sync_tid);
-	} else {
-		struct ext4_iloc iloc;
-
-		err = __ext4_get_inode_loc_noinmem(inode, &iloc);
-		if (err)
-			return err;
-		/*
-		 * sync(2) will flush the whole buffer cache. No need to do
-		 * it here separately for each inode.
-		 */
-		if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
-			sync_dirty_buffer(iloc.bh);
-		if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
-			ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
-					       "IO error syncing inode");
-			err = -EIO;
-		}
-		brelse(iloc.bh);
 	}
 	return err;
 }
@@ -6348,7 +6330,11 @@ int ext4_mark_iloc_dirty(handle_t *handle,
 
 	/* the do_update_inode consumes one bh->b_count */
 	get_bh(iloc->bh);
-
+	if (!ext4_handle_valid(handle)) {
+		if (!EXT4_I(inode)->i_metadata_bhs)
+			ext4_inode_attach_mmb(inode);
+		EXT4_I(inode)->i_metadata_bhs->inode_blk = iloc->bh->b_blocknr;
+	}
 	/* ext4_do_update_inode() does jbd2_journal_dirty_metadata */
 	err = ext4_do_update_inode(handle, inode, iloc);
 	put_bh(iloc->bh);
Signature · 1 line
2.51.0
#17 lore raw

Re: [PATCH 9/9] ext4: Use mmb infrastructure for inode buffer writeout

Quoted 109 lines · Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's inode table block on fsync(2) in nojournal mode (from ext4_sync_parent() and ext4_fsync_nojournal()). This significantly simplifies the code as we don't have to explicitely handle inode buffer writeback in ext4_write_inode() and thus we can also remove sync_inode_metadata() calls from ext4_sync_parent() and ext4_write_inode() call from ext4_fsync_nojournal().

---
 fs/ext4/ext4_jbd2.c |  2 +-
 fs/ext4/ext4_jbd2.h |  2 ++
 fs/ext4/fsync.c     | 12 ------------
 fs/ext4/inode.c     | 24 +++++-------------------
 4 files changed, 8 insertions(+), 32 deletions(-)
fs/ext4/ext4_jbd2.c +1 -1
index 74f05bd0cdde..6bbaf72108fd 100644
@@ -350,7 +350,7 @@ int __ext4_journal_get_create_access(const char *where, unsigned int line,
 	return 0;
 }
 
-static void ext4_inode_attach_mmb(struct inode *inode)
+void ext4_inode_attach_mmb(struct inode *inode)
 {
 	struct mapping_metadata_bhs *mmb;
 
fs/ext4/ext4_jbd2.h +2 -0
index 63d17c5201b5..2a01b8279c88 100644
@@ -122,6 +122,8 @@
 #define EXT4_HT_EXT_CONVERT     11
 #define EXT4_HT_MAX             12
 
+void ext4_inode_attach_mmb(struct inode *inode);
+
 int
 ext4_mark_iloc_dirty(handle_t *handle,
 		     struct inode *inode,
fs/ext4/fsync.c +0 -12
index e25d365e1179..af84489e57c6 100644
@@ -75,9 +75,6 @@ static int ext4_sync_parent(struct inode *inode)
 			if (ret)
 				break;
 		}
-		ret = sync_inode_metadata(inode, 1);
-		if (ret)
-			break;
 	}
 	dput(dentry);
 	return ret;
@@ -87,10 +84,6 @@ static int ext4_fsync_nojournal(struct file *file, loff_t start, loff_t end,
 				int datasync, bool *needs_barrier)
 {
 	struct inode *inode = file->f_inode;
-	struct writeback_control wbc = {
-		.sync_mode = WB_SYNC_ALL,
-		.nr_to_write = 0,
-	};
 	int ret;
 
 	ret = mmb_fsync_noflush(file, EXT4_I(inode)->i_metadata_bhs,
@@ -98,11 +91,6 @@ static int ext4_fsync_nojournal(struct file *file, loff_t start, loff_t end,
 	if (ret)
 		return ret;
 
-	/* Force writeout of inode table buffer to disk */
-	ret = ext4_write_inode(inode, &wbc);
-	if (ret)
-		return ret;
-
 	ret = ext4_sync_parent(inode);
 
 	if (test_opt(inode->i_sb, BARRIER))
fs/ext4/inode.c +0 -18
index 3e66e9510909..09506b4de1b2 100644
@@ -5786,24 +5786,6 @@ int ext4_write_inode(struct inode *inode, struct writeback_control *wbc)
 
 		err = ext4_fc_commit(EXT4_SB(inode->i_sb)->s_journal,
 						EXT4_I(inode)->i_sync_tid);
-	} else {
-		struct ext4_iloc iloc;
-
-		err = __ext4_get_inode_loc_noinmem(inode, &iloc);
-		if (err)
-			return err;
-		/*
-		 * sync(2) will flush the whole buffer cache. No need to do
-		 * it here separately for each inode.
-		 */
-		if (wbc->sync_mode == WB_SYNC_ALL && !wbc->for_sync)
-			sync_dirty_buffer(iloc.bh);
-		if (buffer_req(iloc.bh) && !buffer_uptodate(iloc.bh)) {
-			ext4_error_inode_block(inode, iloc.bh->b_blocknr, EIO,
-					       "IO error syncing inode");
-			err = -EIO;
-		}
-		brelse(iloc.bh);
 	}
 	return err;
 }

fs/ext4/inode.c int ext4_mark_iloc_dirty(handle_t *handle, · line 6336

Quoted 8 lines
fs/ext4/inode.c +4 -1
@@ -6348,7 +6330,11 @@ int ext4_mark_iloc_dirty(handle_t *handle,
 
 	/* the do_update_inode consumes one bh->b_count */
 	get_bh(iloc->bh);
-
+	if (!ext4_handle_valid(handle)) {
+		if (!EXT4_I(inode)->i_metadata_bhs)
+			ext4_inode_attach_mmb(inode);
+		EXT4_I(inode)->i_metadata_bhs->inode_blk = iloc->bh->b_blocknr;

The series is great overall. The only thing I think we should change is that we should hide this

EXT4_I(inode)->i_metadata_bhs->inode_blk = iloc->bh->b_blocknr;

behind a dedicated static inline/regular function call instead of open-coding it everywhere. Can then also be paired with some VFS_WARN_ON_ONCE() to detect garbage bh->b_blocknr.

author reviewer
#18 lore raw

Re: [PATCH 9/9] ext4: Use mmb infrastructure for inode buffer writeout

Quoted 10 lines · Christian Brauner
Quoted 9 lines · Jan Kara

Use mmb inode buffer writeout infrastructure to reliably write out inode's inode table block on fsync(2) in nojournal mode (from ext4_sync_parent() and ext4_fsync_nojournal()). This significantly simplifies the code as we don't have to explicitely handle inode buffer writeback in ext4_write_inode() and thus we can also remove sync_inode_metadata() calls from ext4_sync_parent() and ext4_write_inode() call from ext4_fsync_nojournal().

...

Quoted 13 lines

fs/ext4/inode.c int ext4_mark_iloc_dirty(handle_t *handle, · line 6336

Quoted 9 lines
fs/ext4/inode.c +4 -1
@@ -6348,7 +6330,11 @@ int ext4_mark_iloc_dirty(handle_t *handle,
 
 	/* the do_update_inode consumes one bh->b_count */
 	get_bh(iloc->bh);
-
+	if (!ext4_handle_valid(handle)) {
+		if (!EXT4_I(inode)->i_metadata_bhs)
+			ext4_inode_attach_mmb(inode);
+		EXT4_I(inode)->i_metadata_bhs->inode_blk = iloc->bh->b_blocknr;

The series is great overall. The only thing I think we should change is that we should hide this

Quoted 5 lines

EXT4_I(inode)->i_metadata_bhs->inode_blk = iloc->bh->b_blocknr;

behind a dedicated static inline/regular function call instead of open-coding it everywhere. Can then also be paired with some VFS_WARN_ON_ONCE() to detect garbage bh->b_blocknr.

Good point. I've created mmb_mark_inode_buffer_dirty() helper for this matching mmb_mark_buffer_dirty() we use for standard metadata buffers. It now also handles dirtying the buffer and synchronization with mmb_sync() clearing the inode_blk.

Honza

Signature · 2 lines
Jan Kara <[email protected]>
SUSE Labs, CR
syzbot Tested-by [email protected] on the v1 series
#19 lore raw
In reply to Jan Kara
syzbot ci has tested the following series

syzbot ci has tested the following series

[v1] fs: Fix missed inode write during fsync
https://lore.kernel.org/all/[email protected]
* [PATCH 1/9] affs: Drop support for metadata bh tracking
* [PATCH 2/9] ext4: Allocate mapping_metadata_bhs struct on demand
* [PATCH 3/9] fs: Writeout inode buffer from mmb_sync()
* [PATCH 4/9] ext2: Fix possibly missing inode write on fsync(2)
* [PATCH 5/9] udf: Fix possibly missing inode write on fsync(2)
* [PATCH 6/9] fat: Fix possibly missing inode write on fsync(2)
* [PATCH 7/9] minix: Fix possibly missing inode write on fsync(2)
* [PATCH 8/9] bfs: Fix possibly missing inode write on fsync(2)
* [PATCH 9/9] ext4: Use mmb infrastructure for inode buffer writeout

and found the following issue:

KASAN: null-ptr-deref Write in write_dirty_buffer

Full report is available here:
https://ci.syzbot.org/series/d987d2d8-3775-4aa9-959f-8a045778888c

***

KASAN: null-ptr-deref Write in write_dirty_buffer
tree:      torvalds
URL:       https://kernel.googlesource.com/pub/scm/linux/kernel/git/torvalds/linux
base:      5d6919055dec134de3c40167a490f33c74c12581
arch:      amd64
compiler:  Debian clang version 21.1.8 (++20251221033036+2078da43e25a-1~exp1~20251221153213.50), Debian LLD 21.1.8
config:    https://ci.syzbot.org/builds/567d596c-ca65-43c9-bd7d-1e60cfe9da2a/config
syz repro: https://ci.syzbot.org/findings/1bc13af8-2d91-4fbd-b43e-fbe72f29ca41/syz_repro

EXT4-fs (loop2): unmounting filesystem 00000000-0000-0000-0000-000000000000.
==================================================================

BUG: KASAN: null-ptr-deref in instrument_atomic_read_write include/linux/instrumented.h:112 [inline]
BUG: KASAN: null-ptr-deref in test_and_set_bit_lock include/asm-generic/bitops/instrumented-lock.h:57 [inline]
BUG: KASAN: null-ptr-deref in trylock_buffer include/linux/buffer_head.h:425 [inline]
BUG: KASAN: null-ptr-deref in lock_buffer include/linux/buffer_head.h:431 [inline]
BUG: KASAN: null-ptr-deref in write_dirty_buffer+0x37/0x190 fs/buffer.c:2760
Write of size 8 at addr 0000000000000000 by task syz-executor/5742

CPU: 1 UID: 0 PID: 5742 Comm: syz-executor Not tainted syzkaller #0 PREEMPT(full) 
Hardware name: QEMU Standard PC (Q35 + ICH9, 2009), BIOS 1.16.2-debian-1.16.2-1 04/01/2014
Call Trace:
 <TASK>
 dump_stack_lvl+0xe8/0x150 lib/dump_stack.c:120
 kasan_report+0x117/0x150 mm/kasan/report.c:595
 check_region_inline mm/kasan/generic.c:-1 [inline]
 kasan_check_range+0x264/0x2c0 mm/kasan/generic.c:200
 instrument_atomic_read_write include/linux/instrumented.h:112 [inline]
 test_and_set_bit_lock include/asm-generic/bitops/instrumented-lock.h:57 [inline]
 trylock_buffer include/linux/buffer_head.h:425 [inline]
 lock_buffer include/linux/buffer_head.h:431 [inline]
 write_dirty_buffer+0x37/0x190 fs/buffer.c:2760
 mmb_sync+0x74c/0xed0 fs/buffer.c:603
 ext4_evict_inode+0x2fa/0x1040 fs/ext4/inode.c:199
 evict+0x61e/0xb10 fs/inode.c:841
 ext4_quota_off+0x470/0x580 fs/ext4/super.c:7326
 ext4_quotas_off fs/ext4/super.c:1195 [inline]
 ext4_put_super+0xdf/0xd80 fs/ext4/super.c:1306
 generic_shutdown_super+0x13d/0x2d0 fs/super.c:646
 kill_block_super+0x44/0x90 fs/super.c:1725
 ext4_kill_sb+0x68/0xb0 fs/ext4/super.c:7494
 deactivate_locked_super+0xbc/0x130 fs/super.c:476
 cleanup_mnt+0x437/0x4d0 fs/namespace.c:1312
 task_work_run+0x1d9/0x270 kernel/task_work.c:233
 resume_user_mode_work include/linux/resume_user_mode.h:50 [inline]
 __exit_to_user_mode_loop kernel/entry/common.c:67 [inline]
 exit_to_user_mode_loop+0xf3/0x4d0 kernel/entry/common.c:98
 __exit_to_user_mode_prepare include/linux/irq-entry-common.h:207 [inline]
 syscall_exit_to_user_mode_prepare include/linux/irq-entry-common.h:238 [inline]
 syscall_exit_to_user_mode include/linux/entry-common.h:318 [inline]
 do_syscall_64+0x33e/0xf80 arch/x86/entry/syscall_64.c:100
 entry_SYSCALL_64_after_hwframe+0x77/0x7f
RIP: 0033:0x7fd8d1b9e017
Code: a2 c7 05 dc 06 25 00 00 00 00 00 eb 96 e8 e1 12 00 00 90 31 f6 e9 09 00 00 00 66 0f 1f 84 00 00 00 00 00 b8 a6 00 00 00 0f 05 <48> 3d 00 f0 ff ff 77 01 c3 48 c7 c2 e8 ff ff ff f7 d8 64 89 02 b8
RSP: 002b:00007ffef04ebf88 EFLAGS: 00000246 ORIG_RAX: 00000000000000a6
RAX: 0000000000000000 RBX: 00007fd8d1c32120 RCX: 00007fd8d1b9e017
RDX: 0000000000000000 RSI: 0000000000000009 RDI: 00007ffef04ec040
RBP: 00007ffef04ec040 R08: 00007ffef04ed040 R09: 00000000ffffffff
R10: 0000000000000000 R11: 0000000000000246 R12: 00007ffef04ed0d0
R13: 00007fd8d1c32120 R14: 0000000000014595 R15: 00007ffef04ed110
 </TASK>
==================================================================

***

If these findings have caused you to resend the series or submit a separate fix, please add the following tag to your commit message:

---
This report is generated by a bot. It may contain errors. syzbot ci engineers can be reached at [email protected].

To test a patch for this bug, please reply with #syz test (should be on a separate line).

The patch should be attached to the email.
Note: arguments like custom git repos and branches are not supported.