[LU-2561] Newly created file is same size as directory

• Status: Resolved
• Project: Lustre
• Fix Version/s: Lustre 2.4.0
• Type: Bug
• Priority: Blocker
• Reporter: Christopher Morrone
• Assignee: Lai Siyao
• Resolution: Fixed
• Labels: MB, sequoia

Environment

Lustre 2.3.54-2chaos, 2.3.57-2chaos. (github.com/chaos/lustre) ZFS OSTs.

Sub-Tasks

Key	Summary	Type	Status	Assignee
LU-2583	Review and test patch for LU-2561	Technical task	Closed	Prakash Surya

• Key: Summary
• Summary: Type
• Type: Status
• Status: Assignee
• Severity: 3

Description

We have a subdirectory where any newly created files, e.g. "touch newfile", is created with a file size that matches the size of the directory, as listed in "ls -la .".

I'm not sure yet what is specially about this directory. I can't reproduce the problem in directories that I just "mkdir" myself.

However, if "cp -a" the problem subdirectory somewhere else in the filesystem, I AM able to then recreate the problem there. In the original problem directory, the directory size is 10720, and any newly created file is 10720 in size. In one copy that I made, the directory size is 9728 bytes, and any newly created files are 9728 bytes.

Very strange. I'm not yet sure what is special about the directory contents that makes it able to trigger this bug.

I am able to read the files, and they just contain zeroes.

Issue Links

• is related to: LU-9663 sanity test_103a: LU-2561 test failed

Comments

[Christopher Morrone]

Ned Bass came through with a reproducer!

Found a reproducer.  Seems to be related to having a system.posix_acl_default xattr.

# grove530 /p/lstest/bass6 > mkdir test
# grove530 /p/lstest/bass6 > cd test
# grove530 /p/lstest/bass6/test > getfacl --access . | setfacl -d -M- .
# grove530 /p/lstest/bass6/test > touch x
# grove530 /p/lstest/bass6/test > ls -al
total 12
drwx------+ 2 root  root  5632 Jan  2 17:32 .
drwx------+ 3 bass6 bass6 5632 Jan  2 17:32 ..
-rw-------  1 root  root  5632 Jan  2 17:32 x
# grove530 /p/lstest/bass6/test > getfattr -d -m. .
# file: .
system.posix_acl_default=0sAgAAAAEABwD/////BAAAAP////8gAAAA/////w==
[cut]

[Peter Jones]

Lai

Could you please look into this one?

Thanks

Peter

[Ned Bass]

I verified that this occurs with both ldiskfs and ZFS backends.

[Lai Siyao]

The cause is that mdd_declare_create() mistakenly declare attr set on child when default acl exists for parent, instead this should be done for parent and when child is directory only.

I'll commit a fix later.

[Lai Siyao]

Patch is on http://review.whamcloud.com/#change,4956

[Christopher Morrone]

Patch works for us. Just needs landing at this point. Removing "topsequoia" label.

[Christopher Morrone]

Patch 4956 landed on master. Ticket can be closed.