[LU-5262] setfacl on striped directory will cause the striped directory crash

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.6.0
• Labels: HB
• Severity: 3
• Assignee: nasf (Inactive)
• Reporter: nasf (Inactive)

Description

The mdd_acl_set() sets size unexpectedly on the master object and cause the master object's directory content is truncated.

This corruption can be easily reproduced by the sanity test_125, unfortunately, there is no way to read the master object directory content directly until we work on the patch for LU-5223.

Comments

[nasf (Inactive)]

The patch has been landed to master

[Di Wang]

Ah it seems the problem is that in lod_attr_get, it tries to merge the size from all sub-stripes, then in lod_attr_set it set the merged size to master object.
probably we can just simply drop the merging in lod_attr_get, which is actually unnecessary, see http://review.whamcloud.com/#/c/10841/ And this will also fix LU-5130

[Alex Zhuravlev]

MDD/LOD/OFD should not try to do this of course. I'm saying about something more like LASSERT() in OSD.

[nasf (Inactive)]

Then all OSDs need to check LA_SIZE. Why not control that in MDD/LOD, then no need to worry about new OSD in future.

[Alex Zhuravlev]

probably it makes sense to check in OSD for LA_SIZE in case of non-regular file.

[nasf (Inactive)]

The basic idea of the patch is that: "NOT set size on directory".

The LU-5223 depends on this patch, because if without the patch, the master object may be truncated unexpectedly when setfacl, then the iteration the master object to build LMV EA will get failure because the master directory (for shards FID and name) will be regarded empty although it is not empty.

Sanity test_125 test such case, that is why LU-5223 failed at sanity.

[nasf (Inactive)]

Here is the patch:
http://review.whamcloud.com/10870

It is an important patch, should be landed to Lustre-2.6

[Andreas Dilger]

Isn't the fix here to clear LA_SIZE (and most others except LA_MODE) from attr->la_valid bits in mdd_xattr_set() before mdo_attr_set() is called so that these values are not all changed? Otherwise, there is a chance that this will set stale values into the inode if it is changed concurrently (e.g. pdirops). It isn't clear why LU-5223 is needed to fix this?