[LU-5949] Verify e2fsck inode badness is working for duplicate blocks

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Labels: e2fsck
• Severity: 3
• Assignee: Andreas Dilger
• Reporter: Andreas Dilger

Description

A number of cases of corrupt inodes have lead to large numbers of duplicate blocks during e2fsck runs. Corrupt inodes should be detected as such during e2fsck scanning and be cleared instead of invoking duplicate block resolution in pass1b.

This may be complicated by running "e2fsck -fn" that does not clear these corrupt inodes during pass1. Need to verify that inode badness is working properly, and possibly figure a way that "e2fsck -fn"

Issue Links

• is related to: LU-13650 ldiskfs: enable metadata_csum on MDT/OST
• is related to: LU-12913 fsck found > 1M multilply-claimed blocks
• is related to: LU-8465 parallel e2fsck performance at scale

Comments

[Gerrit Updater]

Andreas Dilger ([e-mail]) merged in patch https://review.whamcloud.com/41450/
Subject: LU-5949 e2fsck: call delete_inode() properly
Project: tools/e2fsprogs
Branch: master-lustre
Current Patch Set:
Commit: 4aea203f2d490f53596eb21233bf0c186a5b3679

[Gerrit Updater]

Andreas Dilger ([e-mail]) uploaded a new patch: https://review.whamcloud.com/41450
Subject: LU-5949 e2fsck: call delete_inode() properly
Project: tools/e2fsprogs
Branch: master-lustre
Current Patch Set: 1
Commit: 7257c1a81bcf13ad7a88da7bb54625a61aba8163

[Gerrit Updater]

Andreas Dilger ([e-mail]) merged in patch https://review.whamcloud.com/41328/
Subject: LU-5949 e2fsck: simplify inode badness handling
Project: tools/e2fsprogs
Branch: master-lustre
Current Patch Set:
Commit: 8725134d27574605c3da74b46df909e419d02b63

[Andreas Dilger]

One difficulty with checking the badness before doing the duplicate scanning is that the outcome depends on the order the inodes were processed in. The e2fsck block bitmap only stores if a block is used for pass1, assuming that no duplicate blocks exist in the filesystem. If e2fsck finds any duplicate blocks it has to re-scan the whole filesystem to find the inodes that are sharing blocks.

It is possible that we could avoid pass1 in some cases by clearing the bad inode rather than adding the blocks into the bitmap, assuming that the good inode is processed first and adds its blocks to the bitmap, and then the bad inode is found later and verified and does not add which would avoid pass1b sometimes.

The other proposal that Ted Ts'o had was to check all inodes in a block at the same time. That makes it easier to detect if a whole block is corrupted.

[Dongyang Li]

I believe inode_badness patch just checks the inode fields without looking at the dup blocks at the moment, and offers to delete the inode in pass2 and pass4 if the badness is over the threshold.

dup blocks are handled independently by e2fsck: allow deleting or zeroing shared blocks.

I think we could check if the badness is over before pass1b, and skip/delete the inode before scanning for dup blocks.