[Bug 251363] use unionfs as a disk-cache for NFS [feature]

From: <bugzilla-noreply_at_freebsd.org>
Date: Wed, 19 May 2021 14:24:13 +0000
https://bugs.freebsd.org/bugzilla/show_bug.cgi?id=251363

Rick Macklem <rmacklem_at_FreeBSD.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |rmacklem_at_FreeBSD.org

--- Comment #21 from Rick Macklem <rmacklem_at_FreeBSD.org> ---
A lack of any cache consistency guarantees is a major
limitation and I tend to agree with Kirk that this
should not be recommended practice.
If it works for you, that's fine, but others would
not understand the limitations of not having cache
consistency and that would probably result in bug
reports related to it.

When I wrote the packrats code, cache consistency was
maintained by the NFSv4 delegation.
Although the code basically worked (I had not written
the recovery code needed after a client reboot), I
left the bits to rot because I did not think it was
widely useful. I may eventually revisit that, since
I can see cases where it could be useful. My intent was
to handle laptops mounting through wifi/wan connections
better, since those are often high latency connections.
I suspect that some of your performance improvement is
a result of the lack of cache consistency. For example,
flushing of all written data to the server upon close(2) can be a big
performance hit, but is required for close->open consistency.
--> I honestly do not know how the Linux fscache code
    works, but I suspect that some variant of cache
    consistency is maintained by it.

Another issue is that the comment in the bugs section
of the mount_unionfs man page still applies (and is in
caps). I don't know the details w.r.t. how unionfs is
broken, but if it were easy to fix, someone would have
done so. I do know that people report issues w.r.t. using
unionfs on an NFS mount and I simply point out to them
that it is broken.
It does not make much sense to implement a new feature on
top of a broken one, imho.

kib_at_ is probably the only person who knows what the breakage
in unionfs is and what needs to be done to fix it.

rick
ps: I don't think comments like "the one and only" are
    appreciated in the context you used it.

-- 
You are receiving this mail because:
You are the assignee for the bug.
Received on Wed May 19 2021 - 14:24:13 UTC

Original text of this message