Bug #38395: luminous: write following remove might access previous onode - bluestore - Ceph

Actions

Copy link

Bug #38395

closed

luminous: write following remove might access previous onode

Added by Igor Fedotov about 5 years ago. Updated about 5 years ago.

Status:

Resolved

Priority:

High

Assignee:

Igor Fedotov

Target version:

% Done:

Source:

Tags:

Backport:

Regression:

Severity:

2 - major

Reviewed:

Affected Versions:

Ceph - v12.2.10

ceph-qa-suite:

Pull request ID:

26540

Crash signature (v1):

Crash signature (v2):

Description

So the sequence is as follows:
T1:
remove A

T2:
touch A
write A

In Luminous there is a chance that A is removed from the cache (as no ref is taken) before T1 lands to DB.
Subsequent B might access outdated onode A from DB and proceed with invalid A state.
Which finally results in, e.g. duplicate removal from the allocator.
Final symptoms remind #38049 and #36541 but the root cause is different.

This has been fixed in Mimic+ by
https://github.com/ceph/ceph/pull/18196/commits/0347518d02eda4e0e2da5241f1d77bc7304d59fb
and follow-ups.

Files

debug-osd.10.log.gz (240 KB) debug-osd.10.log.gz

Igor Fedotov, 02/20/2019 10:25 AM