Lustre discuss - Jan 2007 - 1.4.8, shared root clusters and flock...

Hi,

Just wondering is anyone else hitting this:

1.4.8 (correctly I guess!) started rejecting locks unless the "flock"
mount
option was used, as per changelog note for bug #10743.  I guess in the
past it faked success but the file wasn''t really locked?

Among other things, on RHEL4:
modprobing locks...
rpm locks...
sqlite (and therefore yum) locks...

This has blocked migration to 1.4.8 in our shared root setup. While the < 
1.4.8 behaviour worked okay for us (in so far as most locks don''t
matter
all that much anyway and so booting proceeds just fine if they''re
fake), I
certainly don''t consider it valid for us to have relied on it, so I
don''t
consider this bug-fix behaviour change a bug: Obviously, we were 
skating on rather thin ice, but we kind of knew that already...

=> So, what''s the latest on file lock support in general in lustre?

Actually using the "flock" mount option doesn''t seem to be a
solution, as
AFAIK it''s known incomplete in 1.4.x (?): Unless flock is supposed to
be
working fully in 1.4.8, in which case, well, maybe there is a bug to be 
found...

Hello!

On Fri, Jan 05, 2007 at 06:02:02PM +0000, David Golden wrote:
> Just wondering is anyone else hitting this:
> 1.4.8 (correctly I guess!) started rejecting locks unless the
"flock" mount
> option was used, as per changelog note for bug #10743.  I guess in the
> past it faked success but the file wasn''t really locked?
File was locked, but only on a that node, other nodes won''t see it.
> => So, what''s the latest on file lock support in general in
lustre?
We are trying to fix when we learn about problems with it.
(see bug 11415 if you plan to use fcntl locks with anything above 2.6.5 kernel)
> Actually using the "flock" mount option doesn''t seem to
be a solution, as
> AFAIK it''s known incomplete in 1.4.x (?): Unless flock is supposed
to be
> working fully in 1.4.8, in which case, well, maybe there is a bug to be 
> found...
Well, we found one already.
I wonder if there are any comprehensive flock/fcntl locking testsuites
we can try to use?

Bye,
    Oleg