Date: Tue, 7 Oct 2014 17:06:02 -0400 (EDT) From: Rick Macklem <rmacklem@uoguelph.ca> To: Garrett Wollman <wollman@csail.mit.edu> Cc: freebsd-fs@freebsd.org Subject: Re: 9.3 NFS client bug? Message-ID: <603326314.59995969.1412715962746.JavaMail.root@uoguelph.ca> In-Reply-To: <1109402573.59384867.1412683413852.JavaMail.root@uoguelph.ca>
next in thread | previous in thread | raw e-mail | index | archive | help
I wrote: > Garrett Wollman wrote: > > <<On Fri, 3 Oct 2014 18:10:16 -0400 (EDT), Rick Macklem > > <rmacklem@uoguelph.ca> said: > > > > > 2 - Try an "oldnfs" mount and see if the old client exhibits the > > > same behaviour. > > > > Same behavior both old and new. I'm doing binary search now to try > > to > > minimize the size of the workload that exhibits the issue. > > > > -GAWollman > > > By ay chance is your ZFS server allocating a ZFS inode (whatever they > call it) with a fileno/inode# that doesn't fit in 32 bits? > (There is a diagnostic printf w.r.t. this for NFSv4, but not NFSv3, > so you might try an NFSv4 mount. The printf starts with "NFSv4 > fileid > 32..".) > Oops, I've realized that this diagnostic won't help if the server is a FreeBSD one, since the FreeBSD server just sets the high order 32bits to 0s and not what the file system returns. (va_filerev is a "long", so it is only 64bits on some arches, yikes.) You could try what jhb@ suggested (try a UFS file system) or just use a different ZFS volume that isn't very full, if you think the 32bit restriction could have been exceeded. Btw, I tried your bonnie++ command on my test machines, but they failed part way through because they couldn't allocate boundary tags. (At least now I can reproduce that easily, although it is the M_NOWAIT case where the thread just loops in the kernel.) rick > Since d_fileno in "struct dirent" is 32bits, NFS just truncates the > 64bit fileid to 32bits (or fills the high order 32bits with 0s for > the server). > ZFS is known to generate fileids with non-zero high order > 32bits->busted. > (And not at all easy to fix. I actually have a somewhat hackish idea > for > a way to make 64bit d_fileno values work without busting backwards > compatibility. > I think I'll post that to freebsd-fs@ and see what others think?) > > rick > _______________________________________________ > freebsd-fs@freebsd.org mailing list > http://lists.freebsd.org/mailman/listinfo/freebsd-fs > To unsubscribe, send any mail to "freebsd-fs-unsubscribe@freebsd.org" >
Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?603326314.59995969.1412715962746.JavaMail.root>