ext4: fast symlink test should not rely on i_blocks
PATCH 1/1 ext4: fast symlink test should not rely on i_blocks
ext4_inode_info->i_data is the storage area for 4 types of data:
a) Extents data
b) Inline data
c) Block map
d) Fast symlink data (symlink length < 60)
Extents data case is positively identified by EXT4_INODE_EXTENTS flag. Inline data case is also obvious because of EXT4_INODE_INLINE_DATA flag.
Distinguishing c) and d) however requires additional logic. This currently relies on i_blocks count. After subtracting external xattr block from i_blocks, if it is greater than 0 then we know that some data blocks exist, so there must be a block map.
This logic got broken after ea_inode feature was added. That feature charges the data blocks of external xattr inodes to the referencing inode and so adds them to the i_blocks. To fix this, we could subtract ea_inode blocks by iterating through all xattr entries and then check whether remaining i_blocks count is zero. Besides being complicated, this won't change the fact that the current way of distinguishing between c) and d) is fragile.
The alternative solution is to test whether i_size is less than 60 to determine fast symlink case. ext4_symlink() uses the same test to decide whether to store the symlink in i_data. There is one caveat to address before this can work though.
If an inode's i_nlink is zero during eviction, its i_size is set to zero and its data is truncated. If system crashes before inode is removed from the orphan list, next boot orphan cleanup may find the inode with zero i_size. So, a symlink that had its data stored in a block may now appear to be a fast symlink. The solution used in this patch is to treat i_size = 0 as a non-fast symlink case. A zero sized symlink is not legal so the only time this can happen is the mentioned scenario. This is also logically correct because a i_size = 0 symlink has no data stored in i_data.
--- fs/ext4/inode.c | 20 +++++++++++++------- 1 file changed, 13 insertions(+), 7 deletions(-)
fs/ext4/inode.c +1 -5
@@ -144,16 +144,12 @@ static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,/* * Test whether an inode is a fast symlink. +* A fast symlink has its symlink data stored in ext4_inode_info->i_data. */ int ext4_inode_is_fast_symlink (struct inode *inode) { -int ea_blocks =EXT4_I (inode)->i_file_acl ? -EXT4_CLUSTER_SIZE (inode->i_sb) >>9 :0 ; - -if (ext4_has_inline_data (inode)) -return 0 ;
Quoted 42 lines · Tahsin Erdogan
ext4_inode_info->i_datais the storage area for 4 types of data:a) Extents data
b) Inline data
c) Block map
d) Fast symlink data (symlink length < 60)Extents data case is positively identified by
EXT4_INODE_EXTENTSflag. Inline data case is also obvious because ofEXT4_INODE_INLINE_DATAflag.Distinguishing c) and d) however requires additional logic. This currently relies on
i_blockscount. After subtracting external xattr block fromi_blocks, if it is greater than 0 then we know that some data blocks exist, so there must be a block map.This logic got broken after
ea_inodefeature was added. That feature charges the data blocks of external xattr inodes to the referencing inode and so adds them to thei_blocks. To fix this, we could subtractea_inodeblocks by iterating through all xattr entries and then check whether remainingi_blockscount is zero. Besides being complicated, this won't change the fact that the current way of distinguishing between c) and d) is fragile.The alternative solution is to test whether
i_sizeis less than 60 to determine fast symlink case.ext4_symlink()uses the same test to decide whether to store the symlink ini_data. There is one caveat to address before this can work though.If an inode's
i_nlinkis zero during eviction, itsi_sizeis set to zero and its data is truncated. If system crashes before inode is removed from the orphan list, next boot orphan cleanup may find the inode with zeroi_size. So, a symlink that had its data stored in a block may now appear to be a fast symlink. The solution used in this patch is to treati_size= 0 as a non-fast symlink case. A zero sized symlink is not legal so the only time this can happen is the mentioned scenario. This is also logically correct because ai_size= 0 symlink has no data stored ini_data.
Quoted 4 lines
Fixes 74c5bfa651af ("ext4: xattr inode deduplication")Suggested-by Andreas Dilger <[email protected]>Signed-off-by Tahsin Erdogan <[email protected]>
The unfortunate bit is that this makes the inode impossible to undelete, but I don't think that is a huge concern for symlinks.
Quoted 46 lines
---
fs/ext4/inode.c| 20 +++++++++++++-------1 file changed, 13 insertions(+), 7 deletions(-)fs/ext4/inode.c +13
-8@@ -144,16 +144,12 @@ static int ext4_meta_trans_blocks(struct inode *inode, int lblocks,/* * Test whether an inode is a fast symlink. +* A fast symlink has its symlink data stored in ext4_inode_info->i_data. */ int ext4_inode_is_fast_symlink (struct inode *inode) { -int ea_blocks =EXT4_I (inode)->i_file_acl ? -EXT4_CLUSTER_SIZE (inode->i_sb) >>9 :0 ; - -if (ext4_has_inline_data (inode)) -return 0 ; - -return (S_ISLNK (inode->i_mode) && inode->i_blocks - ea_blocks ==0 ); +return S_ISLNK (inode->i_mode) && inode->i_size && + (inode->i_size <EXT4_N_BLOCKS *4 ); }/* @@ -261,6 +257,16 @@ void ext4_evict_inode(struct inode *inode)if (IS_SYNC (inode))ext4_handle_sync (handle); + +/* +* Set inode->i_size to 0 before calling ext4_truncate(). We need +* special handling of symlinks here because i_size is used to +* determine whether ext4_inode_info->i_data contains symlink data or +* block mappings. Setting i_size to 0 will remove its fast symlink +* status. Erase i_data so that it becomes a valid empty block map. +*/ +if (ext4_inode_is_fast_symlink (inode)) +memset (EXT4_I (inode)->i_data,0 ,sizeof (EXT4_I (inode)->i_data)); inode->i_size =0 ; err =ext4_mark_inode_dirty (handle, inode);if (err) { --
Quoted 2 lines
2.13.2.725.g09c95d1e9-goog
Cheers, Andreas
Quoted 44 lines · Andreas Dilger
Quoted 44 lines · Tahsin Erdogan
ext4_inode_info->i_datais the storage area for 4 types of data:a) Extents data
b) Inline data
c) Block map
d) Fast symlink data (symlink length < 60)Extents data case is positively identified by
EXT4_INODE_EXTENTSflag. Inline data case is also obvious because ofEXT4_INODE_INLINE_DATAflag.Distinguishing c) and d) however requires additional logic. This currently relies on
i_blockscount. After subtracting external xattr block fromi_blocks, if it is greater than 0 then we know that some data blocks exist, so there must be a block map.This logic got broken after
ea_inodefeature was added. That feature charges the data blocks of external xattr inodes to the referencing inode and so adds them to thei_blocks. To fix this, we could subtractea_inodeblocks by iterating through all xattr entries and then check whether remainingi_blockscount is zero. Besides being complicated, this won't change the fact that the current way of distinguishing between c) and d) is fragile.The alternative solution is to test whether
i_sizeis less than 60 to determine fast symlink case.ext4_symlink()uses the same test to decide whether to store the symlink ini_data. There is one caveat to address before this can work though.If an inode's
i_nlinkis zero during eviction, itsi_sizeis set to zero and its data is truncated. If system crashes before inode is removed from the orphan list, next boot orphan cleanup may find the inode with zeroi_size. So, a symlink that had its data stored in a block may now appear to be a fast symlink. The solution used in this patch is to treati_size= 0 as a non-fast symlink case. A zero sized symlink is not legal so the only time this can happen is the mentioned scenario. This is also logically correct because ai_size= 0 symlink has no data stored ini_data.Fixes 74c5bfa651af ("ext4: xattr inode deduplication")
Quoted 7 lines
Quoted 2 lines
Suggested-by Andreas Dilger <[email protected]>Signed-off-by Tahsin Erdogan <[email protected]>The unfortunate bit is that this makes the inode impossible to undelete, but I don't think that is a huge concern for symlinks.
Reviewed-by Andreas Dilger <[email protected]>
Thanks, applied.
- Ted