Date: Thu, 6 Jan 2011 08:18:11 -0500 (EST) From: Rick Macklem <rmacklem@uoguelph.ca> To: perryh@pluto.rain.com Cc: marek sal <marek_sal@wp.pl>, freebsd-stable@freebsd.org, milu@dat.pl, jyavenard@gmail.com, jhb@freebsd.org Subject: Re: NFSv4 - how to set up at FreeBSD 8.1 ? Message-ID: <1634633777.173061.1294319891074.JavaMail.root@erie.cs.uoguelph.ca> In-Reply-To: <4d257864.YjgoS235V8eKvUX2%perryh@pluto.rain.com>
next in thread | previous in thread | raw e-mail | index | archive | help
> John Baldwin <jhb@freebsd.org> wrote: > > > ... even NFS UDP mounts maintain their own set of "socket" state > > to manage retries and retransmits for UDP RPCs. > > Not according to what I remember of the SunOS NFS documentation, > which indicated that the driving force behind using UDP instead of > TCP was to have the server be _completely_ stateless. (Of course > locking is inherently stateful; they made it very clear that the > locking protocol was considered to be an adjunct rather than part > of the NFS protocol itself.) > For UDP, in the server all requests show up at socket/port 2049. They pretty quickly discovered that retries of non-idempotent RPCs trashed things, so the Duplicate Request Cache was invented, which is really state that doesn't have to be recovered after a server crash. (By Chet Jacuzak at DEC, if I recall correctly, who is living on a little island on a lake up in Maine, last I heard.) My recollection of why Sun didn't use TCP was that "they knew that the overhead would be excessive", which wasn't completely untrue, given the speed of an MC68020. > It's been quite a few years since I read that, and I didn't get > into the details, but I suppose the handle returned to a client (in > response to a mount or open request) must have contained both a > representation of the inode number and a unique identification of > the filesystem (so that, in the case where server crash recovery > included a newfs and reload from backup, the FS ID would not match > and the client would get a "stale handle" response). All of the > retry and retransmit burden had to have been managed by the client, > for both reading and writing. Yea, it depended on how the backup was done. To avoid "stale handle" the backup/reload had to retain the same i-nodes, including the generation number in them. (But, then, those 1980s SMD disks never trashed the file systems, or did they?:-) You shouldn't get me reminising on the good ole days, rick
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?1634633777.173061.1294319891074.JavaMail.root>