[LU-5863] Can't mount MGS due to label being less than 8 characters long.

• Type: Bug
• Resolution: Fixed
• Priority: Critical
• Fix Version/s: Lustre 2.7.0
• Affects Version/s: Lustre 2.7.0
• Labels: HB patch
• Severity: 3
• Assignee: James Nunez (Inactive)
• Reporter: James A Simmons

Description

The patch for LU-5383 fixed a out bound for ldd_svname. It test to see if ldd_svname is at least 8 characters in length but the function in lustre_disk.h server_make_name() sets ldd_svname for the mgs to MGS which is to short. This can prevent the MGS from mounting.

Issue Links

• is duplicated by: LU-5876 MGS volumes with short names do not mount since LU-5383 utils: fix array index out of bounds

Comments

[Patrick Farrell (Inactive)]

In case it influences the decision to know what people are doing in the real world, Cray is increasingly using separate MGS and MDS volumes for ease of backup and recovery.
Also for DNE - it's convenient to have MDS volumes all identical.

[Jian Yu]

Hi Andreas,

Clearly, we need a conf-sanity test that formats a separate MGS and mounts it, since we can't possibly have had such a test if this bug slipped through.

I found that conf-sanity test 21d is a basic test case that starting a separate MGS. However, autotest system always uses the configuration of combined MGT and MDT, and then test 21d has been always skipped. Should we enhance autotest system to add a configuration of separated MGT and MDT, and run a test session with this configuration?
For now and this ticket, I'll just add a new test case that formats $fs2mds_DEV as a separate MGT and mounts it.

[Jian Yu]

Sure. I'll do.

[James A Simmons]

Could you finish it off. I'm working on a few other tickets.

[Jian Yu]

Hi James Simmons,
Are you going to add a regression test case into conf-sanity.sh or would you like me to do this?

[Andreas Dilger]

The patch fixes the core of the problem, but it doesn't add a test. Clearly, we need a conf-sanity test that formats a separate MGS and mounts it, since we can't possibly have had such a test if this bug slipped through.

[James Nunez (Inactive)]

Landed to master (pre-2.7)

[Cliff White (Inactive)]

And i have confirmed that we do NOT test internally with a separate MGS - so this was a complete escape.

[Cliff White (Inactive)]

And, I can't make a new clean filesystem, either. Hyperion is down
MGSNID="192.168.120.5@o2ib"
MGSDEV="/dev/mapper/iws10_4"
MDSDEV="/dev/mapper/iws10_1"
mkfs.lustre --reformat --mgs --device-size=$((512 * 1048576)) --fsname lustre $MGSDEV
mkfs.lustre --reformat --mdt --index=0 --mkfsoptions='-i 4096' --mgsnid=$MGSNID --device-size=$((2048 * 1048576)) --fsname lustre $MDSDEV

mkdir /mnt/mgs
mkdir /mnt/mds

mount -t lustre $MGSDEV /mnt/mgs
mount -t lustre $MDSDEV /mnt/mds
~
Fails with same error

[Cliff White (Inactive)]

This bug has stopped all testing on the stable filesystem, as the stable filesystem (formatted with older Lustre ) can no longer be mounted.
arg[4] = /dev/mapper/iws11_4
arg[5] = /mnt/mgs
source = /dev/mapper/iws11_4 (/dev/mapper/iws11_4), target = /mnt/mgs
options = rw
checking for existing Lustre data: found
Reading CONFIGS/mountdata
mount.lustre: invalid name 'MGS'

I really don't want to nuke the stable side, so a fix would be appreciated

[James Nunez (Inactive)]

Thanks for the patch, James.