[LU-4272] lu_device_fini()) ASSERTION( cfs_atomic_read(&d->ld_ref) == 0 ) failed from lovsub_device_free

• Type: Bug
• Resolution: Duplicate
• Priority: Critical
• Fix Version/s: Lustre 2.8.0
• Affects Version/s: Lustre 2.6.0
• Labels: nfs
• Severity: 3
• Assignee: Niu Yawei (Inactive)
• Reporter: Oleg Drokin

Description

Apparently we have a problem with inode cleaning on unmount exposed by at least nfs.

The easiest reproducer I have right now is this:

sh llmount.sh
echo 0 >/proc/sys/lnet/panic_on_lbug
service nfs start
mount localhost:/mnt/lustre /mnt/nfs -t nfs
touch /mnt/nfs/file
ls -l /mnt/lustre
cp -f /etc/passwd /mnt/nfs/file

umount /mnt/nfs
service nfs stop
sh llmountcleanup.sh

Reading the logs, It appears that after the write file dentry is deleted, but because we have locks on inode - it stays in place (unlike when we do not use nfs where it also goes away after sync).
Then on unmount file inode is found, but is considered dirty and not cleaned, meaning the layout lock and objects remain in place.
Now we get to kill_super and that blows up trying to clean up lov.

Issue Links

• is related to: LU-6794 memory leak in Lustre NFS support code, LASSERT() at unmount
• is related to: LU-2613 opening and closing file can generate 'unreclaimable slab' space