NFS incorrectly update atime

[Bruce Evans]

On Thu, 25 Sep 2003, Rong-en Fan wrote:

>   I'm using FreeBSD 4.8-RELEASE for both NFS server and client,
>   however when i append file from client, it updates both
>   atime and mtime, where atime shouldn't be updated.
>
> $ touch testfile
> $ stat testfile
> Access: Thu Sep 25 01:17:43 2003
> Modify: Thu Sep 25 01:17:43 2003
> Change: Thu Sep 25 01:17:43 2003
> $ echo foobar >> testfile
> $ stat testfile
> Access: Thu Sep 25 01:17:43 2003
> Modify: Thu Sep 25 01:19:15 2003
> Change: Thu Sep 25 01:19:15 2003
> $ echo foobar >> testfile
> $ stat testfile
> Access: Thu Sep 25 01:19:31 2003
> Modify: Thu Sep 25 01:19:31 2003
> Change: Thu Sep 25 01:19:31 2003

This still happens in -current.  I think it is normal nfs brokenness...

>   it only occur when file is not empty, also i've noticed that

Writes to 8K boundaries also don't update the atime.  I tried
"dd if=/dev/zero of=8k count=1 >>foobar".  8K is my server's filesystem
block size.  I'm not sure if this magic number is for the server or
nfs-related.

I think the client just does a read as part of a read-modify-write.
It works with blocks, so it must do a read before write in all cases where
the write is not on a block boundary, and at least FreeBSD ffs servers
have no way to avoid updating the atime on read short of a global mount
-noatime.  Some of the reads, including ones from small test files in your
example, may be from a client cache, so the atime update might be avoided
accidentally, but a read will always be necessay if the cache is cold.

>   http://www.freebsd.org/cgi/query-pr.cgi?pr=44580
>   is the same problem. BTW, i also test it on 5.1-RELEASE,
>   and got the same result.

This was misclassified in the "i386" category where it is hard to see
(unfortunately there is an i386 mailing list which seems to have no
activity except for blackholing i386 PRs).  It has nothing to do with
i386's, so I moved it to the "kern" category.  gnats seems to be broken
-- didn't prompt for a reason.

Bruce