On Thu 23-03-23 22:18:53, Baokun Li wrote:
On 2023/3/23 19:44, Jan Kara wrote:
fs/ext4/ext4.h | 3 ++- fs/ext4/ext4_jbd2.h | 9 +++++---- fs/ext4/super.c | 14 ++++++++++++++ 3 files changed, 21 insertions(+), 5 deletions(-)
diff --git a/fs/ext4/ext4.h b/fs/ext4/ext4.h index 08b29c289da4..f60967fa648f 100644 --- a/fs/ext4/ext4.h +++ b/fs/ext4/ext4.h @@ -1703,7 +1703,8 @@ struct ext4_sb_info { /* * Barrier between writepages ops and changing any inode's JOURNAL_DATA
* or EXTENTS flag.
* or EXTENTS flag or between changing SHOULD_DIOREAD_NOLOCK flag on* remount and writepages ops.diff --git a/fs/ext4/ext4_jbd2.h b/fs/ext4/ext4_jbd2.h index 0c77697d5e90..d82bfcdd56e5 100644 --- a/fs/ext4/ext4_jbd2.h +++ b/fs/ext4/ext4_jbd2.h @@ -488,6 +488,9 @@ static inline int ext4_free_data_revoke_credits(struct inode *inode, int blocks) return blocks + 2*(EXT4_SB(inode->i_sb)->s_cluster_ratio - 1); } +/* delalloc is a temporary fix to prevent generic/422 test failures*/ +#define EXT4_MOUNT_SHOULD_DIOREAD_NOLOCK (EXT4_MOUNT_DIOREAD_NOLOCK | \
EXT4_MOUNT_DELALLOC)
- This function controls whether or not we should try to go down the
- dioread_nolock code paths, which makes it safe to avoid taking
@@ -499,7 +502,8 @@ static inline int ext4_free_data_revoke_credits(struct inode *inode, int blocks) */ static inline int ext4_should_dioread_nolock(struct inode *inode) {
- if (!test_opt(inode->i_sb, DIOREAD_NOLOCK))
- if (test_opt(inode->i_sb, SHOULD_DIOREAD_NOLOCK) !=
EXT4_MOUNT_SHOULD_DIOREAD_NOLOCK)@@ -507,9 +511,6 @@ static inline int ext4_should_dioread_nolock(struct inode *inode) return 0; if (ext4_should_journal_data(inode)) return 0;
- /* temporary fix to prevent generic/422 test failures */
- if (!test_opt(inode->i_sb, DELALLOC))
return 0;Is there a need for this SHOULD_DIOREAD_NOLOCK? When called from writeback we will be protected by s_writepages_rwsem anyway. When called from other places, we either decide to do dioread_nolock or don't but the situation can change at any instant so I don't see how unifying this check would help. And the new SHOULD_DIOREAD_NOLOCK somewhat obfuscates what's going on.
We're thinking that the mount-related flags in ext4_should_dioread_nolock() might be modified, such as DELALLOC being removed because generic/422 test failures were fixed in some other way, resulting in some unnecessary locking during remount, or for whatever reason a mount-related flag was added to ext4_should_dioread_nolock(), and we didn't make a synchronization change in __ext4_remount() that would cause the problem to recur. So we added this flag to this function (instead of in ext4.h), so that when we change the mount option in ext4_should_dioread_nolock(), we directly change this flag, and we don't have to consider making synchronization changes in __ext4_remount().
We have checked where this function is called and there are two types of calls to this function:
- One category is ext4_do_writepages() and mpage_map_one_extent(), which
are protected by s_writepages_rwsem, the location of the problem; 2. The other type is in ext4_page_mkwrite(), ext4_convert_inline_data_to_extent(), ext4_write_begin() to determine whether to get the block using ext4_get_block_unwritten() or ext4_get_block().
1) If we just started fetching written blocks, it looks like there is no problem; 2) If we start getting unwritten blocks, when DIOREAD_NOLOCK is cleared by remount, we will convert the blocks to written in ext4_map_blocks(). The data=ordered mode ensures that we don't see stale data.
Yes. So do you agree that EXT4_MOUNT_SHOULD_DIOREAD_NOLOCK is not really needed?
Honza