10:52 a.m.
Thanks Frank. If Ganesha is the only server, do you think additions can be
missed - I see that we do mdcache_dirent_add()
from mdcache_alloc_and_check_handle(). Wouldn't that be good enough to
cover additions or are there cases we might miss it? place_new_dirent()
also invalidates if it can't place new dirent.
Thanks,
Pradeep
On Mon, Mar 1, 2021 at 8:02 AM Frank Filz <ffilzlnx(a)mindspring.com> wrote:
The danger in not invalidating is that something that adds an entry
races
with the deletion such that we miss the addition. Or even a deletion
external to Ganesha.
Frank
*From:* Pradeep [mailto:pradeepthomas@gmail.com]
*Sent:* Sunday, February 28, 2021 11:19 PM
*To:* Frank Filz <ffilzlnx(a)mindspring.com>; Daniel Gryniewicz <
dang(a)redhat.com>; Ganesha-devel <devel(a)lists.nfs-ganesha.org>
*Subject:* [NFS-Ganesha-Devel] unlink of one dirent could invalidate all.
Hello,
In mdcache_unlink, ganesha removes the dirent from its parent by
calling mdcache_dirent_remove(). It also clears MDCACHE_TRUST_ATTRS. So, if
a subsequent GETATTR comes for the parent (which can happen in the same
compound from nfs4_op_remove), it frees all entries at the end of
mdcache_refresh_attrs(). In this scenario, isn't it possible to keep the
rest of dirents intact?
I'm thinking if we update parents attrs in unlink itself, would it prevent
any further refreshes? The use case here is large directories with frequent
additions/deletions/enumerations.
Thanks,
Pradeep