[linux-lvm] Re: Re: LVM + Multipathing

[Dave Wysochanski]

On Tue, 2007-02-20 at 20:19 +0530, Ritesh Raj Sarraf wrote:
> Luca Berra wrote:
> 
> > i meant that LVM (actually device-mapper) does not do anything special.
> > it just passes the IO requests from the fs layer to the
> > underlying block device and if the underlyng block device returns an IO
> > error then it is passed back to the fs. which will cause ext2 to remount
> > the filesystem readonly.
> 
> Hi,
> 
> Thanks for clarifying.
> 
> In my case, the block device is hidden by the multipathing layer. 
> It is something like:
> (Block Device(Multipathing(LVM(Filesystem) ) ) )
> 
> Now during takeover/giveback, the multipathing layer is intelligent enough to
> wait till <120 seconds before declaring that the path has really gone offline
> and no more paths are available.
> If within the 120 seconds time span, the takeover succeeds, the path is back
> online and everything works fine in a non-LVM setup.
> 
> It is only in an LVM setup that a takeover/giveback ends up with the host OS
> having a filesystem read-only problem.

Interesting.

> Now if I go with your explanation, I shouldn't have had the filesystem read-only
> problem since the I/O is being passed on to the multipathing layer which is
> intelligent enough to wait for N seconds before really sending an I/O Error.
> 
> Are there any timeout options in LVM to allow it to wait for N seconds before
> sending out an error ?
> (I understand that LVM might not be involved, but just wondering).
> 

There are no LVM timeouts like you are suggesting.  LVM just does I/O to
devices like any other application.  The timeout settings you're looking
for should be in the layers below LVM.

What does your /etc/multipath.conf file look like?  Do you have
"no_path_retry" set and/or "queue_if_no_path"?  Both of these settings
will affect how multipath deals with a "no paths available" situation.

There are also settings below multipath, in the low-level driver(s).
Since you are using iscsi, I assume you've got open-iscsi?
You can set node.session.timeo.replacement_timeout to a lower value (it
is 120 by default in /etc/iscsi/iscsid.conf) if you want faster path
failure detection and faster failover.

In a lot of cases, people set the low level driver timeouts smaller (<
30 sec) to allow for quicker failure detection and failover, but a
higher or infinite value for the "no paths available" situation.

You might try open-iscsi.org and/or dm-devel lists.


> 
> Thanks,
> Ritesh