[LU-91] Impossible to use quotas on RHEL6.0

• Type: Bug
• Resolution: Fixed
• Priority: Blocker
• Fix Version/s: Lustre 2.1.0
• Affects Version/s: Lustre 2.0.0
• Environment: RHEL6.0 GA with kernel 2.6.32-71
• Severity: 3
• Bugzilla ID: 23,707

Epic

RHEL6 ext4 ldiskfs quotacheck quotas

• Assignee: Johann Lombardi (Inactive)
• Reporter: Diego Moreno (Inactive)

Description

It's not possible for us to use quotas on our Lustre 2.0 over RHEL6.0.

When we try to do "lfs quotacheck -ug /fs1/" from a client, and being fs1 a Lustre fs, command hangs without replying (you can see the logs on bugzilla#23707).

After investigation, we found that Lustre have some conflicts with new code introduced in kernel's quotas code (fs/quota/). The Lustre stack would be like this:

ll_quota_on (> return sb>s_qcop->quota_on(sb, off, ver, name... (with ver=QFMT_VFS_V0), returning code =-1)
fsfilt_ext3_quotactl
quota_onoff
fsfilt_ext3_quotacheck

Then, in the ldiskfs (ext4) stack we have the next entries:

v2_read_file_info (from line "dqopt->ops[type]->read_file_info(sb, type) on vfs_load_quota_inode) rc=-1
vfs_load_quota_inode
vfs_quota_on_path
ldiskfs_quota_on

The -1 value comes from the beginning of v2_read_file_info function:

static int v2_read_file_info(struct super_block *sb, int type)
{
struct v2_disk_dqinfo dinfo;
struct v2_disk_dqheader dqhead;
struct mem_dqinfo *info = sb_dqinfo(sb, type);
struct qtree_mem_dqinfo *qinfo;
ssize_t size;
unsigned int version;

if (!v2_read_header(sb, type, &dqhead))
return -1;
version = le32_to_cpu(dqhead.dqh_version);
if ((info->dqi_fmt_id == QFMT_VFS_V0 && version != 0) ||
(info->dqi_fmt_id == QFMT_VFS_V1 && version != 1))
return -1;

The first condition statement (!v2_read_header(sb, type, &dqhead)) is false, the second is true so v2_read_file_info returns '1'. Values are "info>dqi_fmt_id = QFMT_VFS_V0" and "version = 1", so there is a version mismatch between the version in the header of quotas file and the version stored on the ldiskfs super block (info->dqi_fmt_id).

This new condition statement was introduced in 2.6.33-rc (commit 869835dfad3eb6f7d90c3255a24b084fea82f30d "quota: Improve checking of quota file header") and then it was accepted for RHEL6.0 GA so Lustre would hit this problem in any new kernel with this commit on it.

I was looking where do we initialize the dqhead.dqh_version (this is the 'bad' value as we use QFMT_VFS_V0 for quotas on Lustre, isn't it?) but I didn't find it. I also looked how ext4 initialize this value but I didn't find it. This is the first time I'm looking at the quotas code so I ask for help to somebody knowing more than me on quotas:

• Do you know where do Lustre or ldiskfs initialize the headers value (&sb->s_dquot->files[0].dqh_version)? Adding traces in Lustre code didn't help me.

• Do you think is possible to have a work-around? Using QFMT_VFS_V1 instead of QFMT_VFS_V0 doesn't seem a proper WA...

Comments

[Build Master (Inactive)]

Integrated in lustre-master » x86_64,el6 #13
LU-91 Fix quota format problem with RHEL6 and kernels >= 2.6.33

Brian J. Murrell : cc6dc918f19cbabdcb7333d7dccd48fd6e3d72cf
Files :

• lustre/lvfs/fsfilt_ext3.c

[Build Master (Inactive)]

Integrated in reviews-rhel6 #55
LU-91 Fix quota format problem with RHEL6 and kernels >= 2.6.33

Brian J. Murrell : cc6dc918f19cbabdcb7333d7dccd48fd6e3d72cf
Files :

• lustre/lvfs/fsfilt_ext3.c

[Build Master (Inactive)]

Integrated in reviews-centos5 #571
LU-91 Fix quota format problem with RHEL6 and kernels >= 2.6.33

Brian J. Murrell : cc6dc918f19cbabdcb7333d7dccd48fd6e3d72cf
Files :

• lustre/lvfs/fsfilt_ext3.c

[Build Master (Inactive)]

Integrated in lustre-master » x86_64,ubuntu #13
LU-91 Fix quota format problem with RHEL6 and kernels >= 2.6.33

Brian J. Murrell : cc6dc918f19cbabdcb7333d7dccd48fd6e3d72cf
Files :

• lustre/lvfs/fsfilt_ext3.c

[Peter Jones]

Fix is now landed to master

[Build Master (Inactive)]

Integrated in lustre-master-centos5 #162
LU-91 Fix quota format problem with RHEL6 and kernels >= 2.6.33

Oleg Drokin : b25eb219a157dda49a57e14f7bbc400a52a10a9d
Files :

• lustre/lvfs/fsfilt_ext3.c

[Build Master (Inactive)]

Integrated in reviews-centos5 #552
LU-91 Fix quota format problem with RHEL6 and kernels >= 2.6.33

Johann Lombardi : 93171cb31cacbd12d976a71ae056775c8e4583b9
Files :

• lustre/lvfs/fsfilt_ext3.c

[Johann Lombardi (Inactive)]

Updated patch pushed to gerrit:
http://review.whamcloud.com/#change,268

[Johann Lombardi (Inactive)]

Thanks for the quick feedback. If we don't want to break older kernels (like RHEL5), we need to add something like:
#ifdef QFMT_VFS_V1
... use QFMT_VFS_V1
#else
... use QFMT_VFS_V0
#endif

Actually, we might just want to use something like QFMT_LUSTRE in the lustre code and define it as appropriate.
I will push such a patch through gerrit.

[Diego Moreno (Inactive)]

Patch changing QFMT_VFS_V0 to QFMT_VFS_V1 for quotas initialization on recent kernels.