[LU-6528] permission denied over NFS (intermittent

• Type: Bug
• Resolution: Fixed
• Priority: Major
• Fix Version/s: Lustre 2.8.0
• Affects Version/s: Lustre 2.5.3
• Severity: 3
• Assignee: Lai Siyao
• Reporter: Frederik Ferner (Inactive)

Description

(This might be related to LU-5730 but as the symptoms are different, I'm opening a new ticket.)

Our users are

Issue Links

• is related to: LU-7630 permission denied over NFS
• is related to: LU-5730 intermittent I/O errors for some directories
• is related to: LU-13791 Capabilities are not effective

Comments

[Frederik Ferner (Inactive)]

could you verify that the test user has EXEC access right to all directories in "/dls/p45/data/2015/cm12185-2/tmp/frederik-test/top_D19Rnh"?

Yes, I've have checked and this is the case.

Thanks,
Frederik

[Lai Siyao]

The log shows lustre client .lookup_nd failed with -13, and it looks quite possible to be inside kernel function reconnect_path() when it finds a disconnected dentry:
1. ->get_parent() found parent.
2. ->get_name() found child dentry name.
3. ->lookup_one_len() on parent for this name failed with -EACCES, MDS log shows it doesn't have EXEC right to parent.

I'm not clear of the cause yet, but first could you verify that the test user has EXEC access right to all directories in "/dls/p45/data/2015/cm12185-2/tmp/frederik-test/top_D19Rnh"?

[Frederik Ferner (Inactive)]

Also, we are receiving an increasing number of reports from users who are trying to access files and directories written days ago and are receiving intermittent permission denied (for example when running ls). This is obviously much harder to debug as we are not (yet) able to reproduce at will...

[Frederik Ferner (Inactive)]

This is on our test file system, I believe it should be fairly quite at most times, so I don't think there are other clients creating files but I'm not absolutely sure.

I've noticed (now) that on the lustre client it seems the lnet/debug_mb was quite small (only 8) during the previous test, not sure if this meant that some of the interesting information has been overwritten, so I've increased this now (to 32) and repeated the test. I've verified that the string "setattr" now exists in the client dump file as well.

I'll upload the new dump files and /var/log/message from the lustre client/nfs server after reproducing it again: The error was

Traceback (most recent call last):
  File "/home/bnh65367/bin/LU-6528-reproducer1.sh", line 15, in <module>
    sub_dir = tempfile.mkdtemp(prefix='sub_', dir=top_dir)
  File "/usr/lib64/python2.6/tempfile.py", line 318, in mkdtemp
    _os.mkdir(file, 0700)
OSError: [Errno 13] Permission denied: '/dls/p45/data/2015/cm12185-2/tmp/frederik-test/top_D19Rnh/sub_4GFmt6'

[Lai Siyao]

I don't see any failure in both logs, but I see there are create/setattr in MDS logs, while not found in client logs. Actually client log shows it's idle (only ping is sent to MDS).

Is there other lustre clients which is creating files simultaneously? And when does client log starts? If it does contain all the logs of this creation, I'm afraid _os.mkdir() failed before it's really done on lustre client, that is, NFS server returned -EACCES before calling into lustre code. If this is true, could you `echo 32767 > /proc/sys/sunrpc/nfsd_debug`, and tried to reproduce, and then upload messages which contains NFSD logs?

[Frederik Ferner (Inactive)]

I have just managed to reproduce it, triggering 'lctl dk' on both lustre client/aka NFS server and MDS server directly after the permission denied message, both files are attached, so I'm hoping there might be some information in the dumps.

This was the error:

File "/home/bnh65367/bin/LU-6528-reproducer1.sh", line 15, in <module>
sub_dir = tempfile.mkdtemp(prefix='sub_', dir=top_dir)
File "/usr/lib64/python2.6/tempfile.py", line 318, in mkdtemp
_os.mkdir(file, 0700)
OSError: [Errno 13] Permission denied: '/dls/p45/data/2015/cm12185-2/tmp/frederik-test/top_39JlTW/sub_ju6qhK'

[Frederik Ferner (Inactive)]

Lai,

I'm struggling slightly to gather the logs.

Initially I was going to use one of our (more powerful) test clients as NFS server, but failed to reproduce the issue with that machine. On the much less powerful machine I used earlier to reproduce the issue on our test system, I failed to trigger it after enabling the debug daemon. I'm currently trying without debug daemon again, hoping to catch it in time to use lctl dk to gather debug logs which contain the interesting information.

One difference between the NFS servers where I was able to reproduce the issue and the one where I couldn't is that the later use o2ib for lnet and the others ethernet only. I don't see how this would matter but thought I'd mention it.

Cheers,
Frederik

[Frederik Ferner (Inactive)]

Malcolm,

yes we are using LDAP, SSSD is not installed on our MDS servers (or the NFS servers), it might be installed on the clients but as you mention secondary groups upcall, I assume the MDS is what you were after.

Cheers,
Frederik

[Malcolm Haak (Inactive)]

This is an odd question but are you using LDAP?
If so are you using SSSD?
If so do you have enumerate enabled?
We ran into this exact issue when enumerate was not enabled as it breaks the secondary groups upcall.

[Lai Siyao]

I tested locally with nfsv3 and nfsv4, but couldn't reproduce. Once you have debug logs, please put it somewhere, and then I'll look into them.

[Peter Jones]

Lai

Could you please assist with this issue?

Thanks

Peter