metadata: Force tree re-read after successful flush
Once we flush the journal and write new tree file we need to re-read it to refresh internal data structures (and mmap data from the right file). We originally left this work on meta_tree_refresh_locked() and meta_tree_needs_rereading() respectively where we checked the rotated bit. In detail, metabuilder wrote a new temp tree file, then explicitly opened the current (old) one, wrote the rotated bit and atomically replaced the temp file. Then the metadata daemon having mmapped the old file detected the rotated bit and scheduled journal and tree file reopen+reread. However in concurrent environment like NFS homedir where multiple metadata daemons are handling the same database we may run in a race and not getting the rotated bit detected properly. This led to an infinite loop between meta_journal_add_entry() - meta_tree_flush_locked() - meta_tree_refresh_locked() - meta_journal_add_entry() since we had full journal, didn't detect the rotation and since the files were already unlinked, there was no force to break that loop. This patch forces tree file re-read after successful flush to prevent this issue. https://bugzilla.gnome.org/show_bug.cgi?id=637095
parent
051556ee
Please register or sign in to comment