[LU-14726] quota accounting invalid on OST

• Type: Bug
• Resolution: Unresolved
• Priority: Minor
• Environment: Lustre 2.12.6 + patches

Epic/Theme

quota

• Severity: 3
• Assignee: Etienne Aujames
• Reporter: Etienne Aujames

Description

We observed invalid quota accounting for some groups on several OST:

"lctl get_param osd-ldiskfs.<OST>.quota_slave.acct_group" returned 2-7 greater number of inodes  that with "find -gid <gid>" on the ldiskfs OST target.

After dropping dentry caches on all OSS, the accounting values came back to normal.

echo 2 > /proc/sys/vm/drop_caches

For example, for the gid 34132 on OST001e:

• Before dropping cache: accounting space used = 38705302288* kbytes ~ 36T
• After dropping cache: accounting space used = 13905337816 kbytes ~ 12T
• diff = 24855082348 kbytes ~ 23T (-64%)

This filesystem is used for small files operations (SSD disks for targets).

We think that issue occurred after repeatedly:

• untar 30 000 000 files (total ~ 100T)
• stats this files
• delete this files

The tool used to tar/untar is mpifileutils (https://github.com/hpc/mpifileutils) used with 360 process (dispatched on 10 lustre clients).