4.11. cman

BZ#765665

The fence_rhevm fencing agent uses the Red Hat Enterprise Virtualization API to check the power status ("on" or "off") of a virtual machine. In addition to the states of "up" and "down", the API includes other states like "unassigned", "powering_up", "paused", "migrating", "unknown", "not_responding", "wait_for_launch", "reboot_in_progress", "saving_state", "restoring_state", "suspended", "image_illegal", "image_locked" or "powering_down". Previously, only if the machine was in the "up" state, the "on" power status was returned. The "off" status was returned for all other states although the machine was actually running. This allowed for successful fencing even before the machine was really powered off. With this update, the fence_rhevm agent detects power status of a cluster node more conservatively, and the "off" status is returned only if the machine is really powered off, it means in the "off" state.

BZ#811939

Previously, cman did not handle idle connection timeout correctly when fencing a cluster node. Consequently, a connection to a fence device timed out and fencing failed if the "delay" option was set to more than 5 seconds. With this update, the "delay" option is applied before the connection is opened and the fencing thus no longer fails in this scenario.

BZ#861392

The speed of fencing is critical because otherwise, broken nodes have more time to corrupt data. Previously, the operation of the fence_vmware_soap fencing agent was slow when used on the VMWare vSphere platform with hundreds of virtual machines. This update fixes a problem with virtual machines that do not have a valid UUID, which can be created during failed P2V (Physical-to-Virtual) processes. Now, the fencing process is also much faster and it does not terminate if a virtual machines without an UUID is encountered.

Bug Fixes

BZ#674497

With this update, several typographical errors in the fence/agents/scsi/fence_scsi.pl file have been fixed. As a result, the debug messages of this file are now typographically correct.

BZ#745985

Prior to this update, the VMware vSphere 5.0 SOAP API was not listed as a version supported by the fence_vmware_soap fence agent. Consequently, the fence agent did not function properly with virtual machines managed by VMware vSphere 5.0. With this update, VMware vSphere 5.0 has been added to the list of supported interfaces, and is fully compatible with fence_vmware_soap.

BZ#753839

Previously, the fence_scsi man page contained incorrect information about the limitations of the fence_scsi agent. The correct description of these limitations can be found in the in the Cluster Administration guide. With this update, the misleading section has been removed from the manual page in order to avoid the potential confusion.

BZ#782919

Prior to this update, when attempting to create registrations on a multipath device with the fence_scsi_test -o command, some registrations failed without signalization. Consequently, there was a need to verify that all registrations had been successfully created. With this update, a --strict option has been added into the fence_scsi_test program. This option forces the verification step to compare the number of paths to the number of times the registration key appears on the device. Without this option, the verification step only inspects the existence of a registration key on the device. Although it is usually safe to omit the --strict option, it is strongly recommended to use --strict on multipath devices.

BZ#786485

The fence_rhevm fencing agent uses the Red Hat Enterprise Virtualization API to check the power status ("on" or "off") of a virtual machine. In addition to the states of "up" and "down", the API includes other states like "unassigned", "paused", etc. (14 in sum). Previously, only when the machine was in the "up" state, the "on" power status was returned. The "off" status was returned for all other states even though the machine was actually running. This behavior allowed for successful fencing even before the machine was powered off. The bug has been fixed and the "off" status is now returned only in a case the machine is in the "off" state.

BZ#804170

Previously, the cman utility did not apply the --delay option correctly when fencing a cluster node. Consequently, a connection to a fence device timed out and fencing failed when --delay was set to more than 5 seconds. With this update, --delay is applied before the connection is opened and the fencing no longer fails in the described scenario.

BZ#809390

Prior to this update, when the method name setting was left empty in the /etc/cluster/cluster.conf configuration file, an unintended core file was created. Consequently, a fenced cloud terminated with a segmentation fault. This bug has been fixed, and the termination no longer occurs in the aforementioned scenario.

BZ#809481

Due to a bug in the fence_scsi_test function, failing to create a reservation led to an incorrect reset of the error count to zero. The bug has been fixed, and the error count is now incremented properly when the error occurs.

BZ#836654

Previously, the fence_vmware_soap fencing agent operated slowly on the VMware vSphere platform with hundreds of virtual machines (VM). This behavior was caused by requesting for needless attributes. With this update, the unnecessary requests have been removed. This update also handles the VMs with invalid UUIDs, which can occur as a result of the failed P2V (physical to virtual machine) process. As a result, fence_vmware_soap performs fencing with increased speed and the process is no longer terminated when a VM without an UUID is encountered.

BZ#843083

In certain cases, the fence_xvm fencing agent incorrectly reported successful fencing, and ignored a communication issue between the agent and the fence_xvmd fencing host. This bug has been fixed and the communication errors are now reported correctly.

BZ#863567

The new href attribute on the /vms/vm element in the Red Hat Enterprise Virtualization Manager 3.1 API caused the get_id regular expression of the fence_rhevm fencing agent to fail. Consequently, the plug status was not available. With this update, get_id has been modified to allow arbitrary attributes to be added to the element. As a result, the plug status is now correctly shown.

Enhancements

BZ#738705

RHEL5 Cluster Suite now supports using RHEV shared storage disks as the shared storage between cluster nodes. Highly Available Logical Volume Manager (HA-LVM), Clustered Logical Volume Manager (CLVM), and the qDisk daemon are all supported on this storage. Fencing via fence_scsi will not work as shared disks are only exposed as VirtIO or IDE devices.

BZ#741985

A new fence agent, fence_ipdu, that handles the IBM iPDU fence device over the Simple Network Management Protocol (SNMP) has been added. As a result, the cman package now provides compatibility with IBM iPDU.

BZ#782900

Previously, using the qdiskd daemon for multipath devices required tuning with the device-mapper-multipath tool, which was a complex and error-prone process. With this update, the qdiskd input and output operations have been improved to automatically detect the multipath-related timeouts without requiring a manual configuration. As a result, qdiskd can now be easily deployed with device-mapper-multipath.

BZ#810949

Various cman fence agents differ in handling of the end-of-line (EOL) markers. Previously, changing the universal \r\n EOL could make the login process impossible for the fence agent. With this update, an automatic detection of EOL has been added to a fencing library and the described error no longer occurs.

BZ#821857

Prior to this update, it was not possible to specify more than four fencing devices per method with the cman utility. With this update, the maximum number of devices per method has increased to eight.

BZ#836963

The Distributed Lock Manager (DLM) now allows tuning of DLM hash table sizes from the /etc/sysconfig/cman file. The following parameters can be set in the /etc/sysconfig/cman file:

DLM_LKBTBL_SIZE=<size_of_table>
DLM_RSBTBL_SIZE=<size_of_table>
DLM_DIRTBL_SIZE=<size_of_table>

which, in turn, modifies the values in the following files respectively:

/sys/kernel/config/dlm/cluster/lkbtbl_size
/sys/kernel/config/dlm/cluster/rsbtbl_size
/sys/kernel/config/dlm/cluster/dirtbl_size

BZ#856954

Previously, it was not possible to modify the default TCP port (21064) of the Distributed Lock Manager (DLM). With this update, the DLM_TCP_PORT configuration parameter has been added into the /etc/sysconfig/cman file. As a result, the DLM TCP port can be manually configured.

BZ#878998

Support for clusters utilizing VMware's VMDK disk image technology with the multi-writer option is now provided. It is now possible to deploy Global File System 2 (GFS2) on top of VMDK.