[LU-2634] short symlinks on MDT with "extents" have EXT4_EXTENTS_FL set

• Type: Technical task
• Resolution: Fixed
• Priority: Blocker
• Fix Version/s: Lustre 2.4.0, Lustre 2.1.5
• Affects Version/s: Lustre 2.4.0, Lustre 2.1.2, Lustre 2.1.5
• Labels: LB
• Environment: MDT needs to be formatted with {{--mkfsoptions="-O extents"}}. This is not a normal configuration, but has been seen in the field
• Assignee: Emoly Liu
• Reporter: Andreas Dilger

Description

Short symlinks on MDT filesystems formatted with the "extents" feature appear to be created with the EXT4_EXTENTS_FL in osd-ldiskfs, but that shouldn't be happening. e2fsck considers this a corruption and deletes the symlink.

While we have never formatted MDT filesystems with "extents" enabled, some users have done this, or enabled it after formatting, and the MDS should not corrupt such filesystems.

Issue Links

• is related to: LU-2627 /bin/ls gets Input/output error

Comments

[Peter Jones]

Landed for 2.4

[Emoly Liu]

b2_1 port is at http://review.whamcloud.com/5458

[Emoly Liu]

Thanks for Andreas' comment! I saw the code in ext4/ldiskfs_symlink(). Now I know my previous change to ext4/ldiskfs is not a right place. It should be fixed in lustre osd-ldiskfs level not ext4/ldiskfs level.

BTW, I test the change above in osd_ldiskfs_writelink() and it works.

[Andreas Dilger]

Looking at the ext4 code in the kernel, it shows something like:

static int ext4_symlink(struct inode *dir,
                        struct dentry *dentry, const char *symname)
{
        if (l < EXT4_N_BLOCKS * 4) {
                /* clear the extent format for fast symlink */
                ext4_clear_inode_flag(inode, EXT4_INODE_EXTENTS);
                inode->i_op = &ext4_fast_symlink_inode_operations;
        }

It appears we should be doing the same thing in our own code:

 static int osd_ldiskfs_writelink(struct inode *inode, char *buffer, int buflen)
 {
+        /* clear the extent format for fast symlink */
+        ldiskfs_clear_inode_flag(inode, LDISKFS_INODE_EXTENTS);

         memcpy((char *)&LDISKFS_I(inode)->i_data, (char *)buffer, buflen);
         LDISKFS_I(inode)->i_disksize = buflen;
         i_size_write(inode, buflen);
         inode->i_sb->s_op->dirty_inode(inode);

         return 0;
 }

[Emoly Liu]

patch tracking at http://review.whamcloud.com/5154

[Emoly Liu]

Since the feature extents is disabled by default in current mkfs.lustre code, the patch only needs to fix the problem in the following two situations:
1. if the users still format MDT with "-O extents", we will prevent setting EXT4_INODE_EXTENTS flag to the new fast symlink inodes;
2. if the users disable extents feature but the old data already have extents flag, we will clear EXT4_INODE_EXTENTS for the old fast symlink inodes and write back to disk to avoid e2fsck error.

Now the patch can pass my local test for the above two situations. BTW, because the part of patch for situation 1 has avoided to set that extent flag for new inodes, it's hard to reproduce situation 2 in conf-sanity test. I simulate it manually and it works.

[Andreas Dilger]

Sorry, it seems mkfs.lustre works too hard to clear the "extents" option from the MDT feature list. It could also be enabled with tune2fs after formatting.

[Emoly Liu]

When I use --mkfsoptions=\\\"-O ^extents,extents\\\", the error "Fast symlink xxx has EXTENT_FL set. Clear? no" can be reproduced.

[Emoly Liu]

Probably because option "extents" is in front of "^extents"?

-O extents,dirdata,uninit_bg,^extents,dir_nlink,quota,huge_file,flex_bg