Bug#177136: File browsers browsing SGI served NFS directories don't display all directories.
Package: general
Version: all kernels from 2.4.18 to 2.4.20, unstable distribution for
last 6 months.
Severity: important
(Well, important for sites with SGI machines).
This one is difficult to explain and track. It effects many programs,
such as mozilla, evolution and java apps., suggesting a library.
The actual bug is that file browsers on many application, most notably
mozilla, but also some Java applications (like MagicDraw-5.5i) loose
directories on an NFS service in an unpredictable way. That is, the
directories are not displayed in the file browser. Which directories go
missing is not predictable, but the lost directories are persistently
lost until there is a change in the directory containing the "missing"
entries. Then the missing directories may, or may not appear, depending
on the phase of the moon :)
The NFS file service is being served by an SGI running IRIX 6.5. Output
of uname on the NFS server is
IRIX <hostname> 6.5 10181059 IP32
I have only been able to reproduce this for SGI NFS servers, using both
Version 2 and 3 protocols.
On the linux side, my kernel is 2.4.19-rc2, but the problem presists
across many kernels.
The problem is not present in the current stable branch running kernels
in the range that do display the problem on the unstable branch. This
makes it unlikely to be a linux kernel thing. That it occurs in a
variety of apps suggests a library somewhere that is not handling
something stupid that the SGI is serving up as a directory listing.
The problem occurs on a variety of different linux machines. It is not
hardware dependent.
To add to the mystery, ls works fine. Solid as a rock, as is file IO etc
to/from the server. In all other respects except for these strange
effects in some file browsers, the NFS is solid. (BTW take it for
granted that all permissions and ownership issues were considered and
examined till the cows came home). Examining the way ls opperates, it
seems to use open() with O_DIRECTORY flag rather than opendir() and
friends. Maybe this is a start?????
Konqueror works fine. It does not have any problem scanning the
directories on the server.
There seemed to be some coincidence that many directories that go
missing have the same name as a username that is served over YP (or NIS
these days). But that well may be a false lead, as we have encountered
exceptions.
All in all, a really strange one.
libc6 is 2.3.1-9, but the problem has persisted through many versions of
libc6. The problem pre-dates the move to gcc-3.2.2-0 on the current
unstable branch.
Cheers,
_________________________________________________________
P. Hornby
CSIRO Exploration & Mining ,
AARC, 26 Dick Perry Ave,
Kensington, WA 6151 Australia
Phone +61 8 6436 8500 Fax +61 8 6436 8555
Reply to: