[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]

Re: [Linux-cluster] SCSI-3 reservation on RHEL 5.0 native cluster



Sadek, Abdel napisaƂ(a):
> Hello;
> we need to understand how it works in order for our developers to make it work with our own multipath driver.

Hello,

scsi_reserve and fence_scsi do not work with multipath environments.
scsi_reserve do registration and reservation at startup, and fence_scsi
do unregister.

Here is some information form conversation with Ryan O'Hara - person who
wrote fence_scsi and scsi_reserve scripts.

--cut---
There are a few reasons for this. First, it is possible the creating
registrations/reservations at startup and unregistering at fence time
could ping-pong the I/O paths. That is assuming that I/O down either
path will automatically switch the active I/O path. So at best it would
work with only active/active controllers. Second, discovering the I/O
paths in a multipath environment does not work. All devices (I/O paths
in this case) must be registered, so each node must create a
registration for each device (I/O path) using its key. This also does
not work. Finally, fencing is unreliable with scsi reservations in a
multipath environment. For example, suppose an I/O path fails for one
node. Assume that node is selected to fence another node. It will need
to unregister the key for the faulty node for all I/O paths. If the node
responsible for performing fencing does not have all I/O paths
available, it will not be able to unregister the key for the faulty
node. This is very problematic.

These are a few reasons why you cannot use scsi reservations as a fence
method in a multipath environment.

---cut---

For my opinion You have two choices:
1. Change multipath driver to work correctly with scsi fencing
2. Change firmware in You storage where there it only check of host ID.
In this situation WWN of HBA isn't checked and You can connect all of
You pathes.

Please give me also information which multipath driver do You use.

Hope this makes sense.

Best Regards
Maciej Bogucki


[Date Prev][Date Next]   [Thread Prev][Thread Next]   [Thread Index] [Date Index] [Author Index]