--- Additional comment from Raghavendra G on 2016-07-08 01:58:08 EDT --- Seems like there is a leak in inodes: # grep fuse.itable.active glusterdump.2460.dump.1467123831 | tail -1 [xlator.mount.fuse.itable.active.2904698] As can be seen above there are 2904698 "active" inodes, which strongly points towards a leak. regards, Raghavendra --- Additional comment from Raghavendra G on 2016-07-08 02:30:54 EDT --- [raghu@unused 1353774]$ grep ia_type=1 ./glusterdump.2460.dump.1467123831 | wc -l 3457676 [raghu@unused 1353774]$ grep ia_type=2 ./glusterdump.2460.dump.1467123831 | wc -l 637 Seems like most of the inodes that leaked are regular files. --- Additional comment from Raghavendra G on 2016-07-08 04:44:31 EDT --- I think leak was introduced by http://review.gluster.org/#/c/11892/ for inodes on which "need-lookup" is set (typically the ones returned by readdirp), we call entry resolution even though loc->inode is set (as inode was already present in itable). However, fuse_resolve_entry_cbk assumes loc->inode is unset always, thereby overriding loc->inode with a new value without doing an unref. regards, Raghavendra
Clone of bug 1353774 which is duplicate of bug 1294759 which is CLOSED CURRENTRELEASE. Closing this one too.