11:39 a.m.
Hi Daniel,
It looks like mdcache_lru_cleanup_try_push() can be improved for the
cases when there are ops executing parallely with an unexport request.
In this case mdcache_lru_cleanup_try_push() is not able to drop the
sentinel ref and mark the entry for cleanup. This is because one extra
ref is taken by the op which is currently executing. The total number
of refs are > 2.
Once the current op execution finishes (after the execution of
mdcache_unexport()) the sentinel ref remains and the entry is also not
marked for cleanup even though no export is referring to it
(entry->first_export_id == -1).
Is it possible to modify mdcache_lru_cleanup_try_push() such that even
if the refcnt is > 2 but if the entry->first_export_id == -1 (the
current export which is going away was the only one referring to this
entry), we should drop the sentinel ref and mark the entry for
cleanup? This will ensure we cleanup the entries when the current op
does put_ref and these are reused faster than other mdcache entries
which are actually pointing to valid exports.
Thanks,
Ashish