I think this is fixed now in trunk and pvfs-2-7-branch. This was indeed
a bug in our readdir as Sam suggested.
The problem occurred if the kernel module did a readdir from PVFS but
wasn't able to fit all of the results into the caller's buffer. At that
point, the directory position has to be set to pick up at the correct
spot on the next iteration. The logic for doing this wasn't setting the
right value.
As it turns out, fixing this bug also corrects this earlier issue with
glibc that I had pointed out without having to implement a directory
seek function:
http://www.beowulf-underground.org/pipermail/pvfs2-developers/2008-January/003813.html
... so I went ahead and pulled that stuff back out.
There are a couple of more general headaches with how the directory
position works in readdir(), but I don't think they would cause anyone
trouble in practice. As far as I know this update covers the oddball
knfsd and glibc cases.
Thanks for the excellent test case description in your bug report!
-Phil
Emmanuel Florac wrote:
OK, I've been testing again and again, and I found a weeeeeeird bug.
When I say weird, I mean really really of the black magic flavour.
So : still testing this NFS stuff, I tried with various configurations
to sort it out, and it really a PVFS2 problem (tried different clients,
and exporting non-pvfs2 FS from the same server, all OK).
The NFS server exports the pvfs2 FS this way :
/mnt/pvfs2 *(rw,async,no_root_squash,no_subtree_check)
On the client, I mount it and run this into the share :
for i in `seq 1 50` ; do touch test.$i ; done ; echo "created $i" ;
echo -n "listing " ; ls | wc -l
So I create 50 files from test.1 to test.50 and so what? I tried again
and again, and always the same : test.26 is missing in the ls output!
Not any other file. When I create test.1 to test.65536, test.26 is
still missing.
BUT if you type "ls test.26", it shows up just OK. Yeah, that's weird.
Then I do
rm * ; ls
and it tells me
test.26
Uh....
If I create 2000 files but starting from 100 (test.100 to test.2100),
when I do
rm *; ls
It gives
test.125
Uh, the 26th file, again... This is a very strange modulo 26 bug
indeed :)
_______________________________________________
Pvfs2-users mailing list
[email protected]
http://www.beowulf-underground.org/mailman/listinfo/pvfs2-users
