Bug #2098: xfs/ext4 non-idempotent transaction - Ceph - Ceph

Actions

Copy link

Bug #2098

closed

xfs/ext4 non-idempotent transaction

Added by Samuel Just about 12 years ago. Updated about 12 years ago.

Status:

Resolved

Priority:

Urgent

Assignee:

Sage Weil

Category:

OSD

Target version:

% Done:

Source:

Development

Tags:

Backport:

Regression:

Severity:

Reviewed:

Affected Versions:

ceph-qa-suite:

Pull request ID:

Crash signature (v1):

Crash signature (v2):

Description

Forcing a sync after a non-idempotent transaction is not adequate to ensure correctness during journal replay.

Consider the following transaction:

clone(a,b)
write(a)
[sync]

A crash between the write and the sync will cause the b to have the wrong contents since this transaction will be replayed. collection_add presents a particular problem during recovery:

T1:
touch(a)
write(a)

T2:
write(a)
unlink(b)
link(a,b)
unlink(a)
[sync]

A crash between the unlink and the sync will cause T2 (but not necessarily T1) to be replayed resulting b failing to contain the data written during T1.

Related issues 1 (0 open — 1 closed)