4.29. cluster and gfs2-utils

Bug Fix

BZ#849048

Previously, it was not possible to specify start-up options to the dlm_controld daemon. As a consequence, certain features were not working as expected. With this update, it is possible to use the /etc/sysconfig/cman configuration file to specify dlm_controld start-up options, thus fixing this bug.

Bug Fixes

BZ#707115

The cluster and gfs2-utils packages have been upgraded to upstream version 3.0.12.1, which provides a number of bug fixes over the previous version.

BZ#713977

Previously, when a custom multicast address was configured, the configuration parser incorrectly set the default value of the time-to-live (TTL) variable for multicast packet to 0. As a consequence, cluster nodes were not able to communicate with each other. With this update, the default TTL value is set to 1, which fixes the problem.

BZ#726777

A section describing the "suborg" option for the fence_cisco_usc agent was not present in the RELAX NG schema which is used to validate the cluster.conf file. As a consequence, validation of cluster.conf failed even if the file was valid. The suborg section has been added to the RELAX NG schema and cluster.conf is now validated correctly.

BZ#707091

Building the resource group index for a new GFS2 file system using the mkfs.gfs2 utility used all the space allocated. If the file system filled up completely, no room was left to write a new rindex entry. As a consequence, the gfs2_grow utility was unable to expand the file system. The mkfs.gfs2 utility has been modified so that enough space is now allocated for the entire rindex file, and one extra rindex entry. The gfs2_grow source code has been modified to utilize the unused rindex space. As a result, gfs2_grow is now able to expand a completely full GFS2 file system.

BZ#678585

GFS2 POSIX (Portable Operating System Interface) lock operations (implemented in Distributed Lock Manager, also known as DLM) are not interruptible when they wait for another POSIX lock. Previously, processes that created a deadlock with POSIX locks could not be killed to resolve the problem, and one node had to be reset. DLM now uses a new kernel feature that allows the waiting process to be killed, and information about the killed process is now passed to the dlm_controld daemon to be cleaned up. Processes deadlocked on GFS2 POSIX locks can now be recovered by killing one or more of them.

BZ#719135

Prior to this update, boundaries for the locktable and label fields in the GFS2 superblock were not properly checked by the tunegfs2 tool. As a consequence, running the "gfs2_tool sb" command could terminate unexpectedly with buffer overflow. In addition, invalid characters could be printed when using tunegfs2 to change locktable or label to a minimum or maximum length (63 characters). The tunegfs tool has been modified to check the correct boundaries of the locktable and label fields. As a result, tunegfs2 no longer creates invalid locktables or labels, and therefore gfs2_tool prints the superblock values properly.

BZ#740385

When executing the cman utility by using the init script with enabled debugging, a file descriptor leaked. The file pointed to the file descriptor would continue to grow endlessly, filling up the /tmp file system. This update ensures that the file descriptor is closed after a successful cman startup. Space in /tmp is now released correctly.

BZ#695795

The cman utility implements a complex set of checks to configure the Totem protocol. One of the checks that copies the configuration data was incorrect and the transport protocol option was not handled correctly as a consequence. A patch has been applied to address this issue and cman now handles the transport option properly.

BZ#679566

When the user executed the "gfs2_edit savemeta" command to save the metadata for a target GFS2 file system, not all of the directory information was saved for large directories. If the metadata was restored to another device, the fsck.gfs2 tool found directory corruption because of a missing leaf block. This was due to gfs2_edit treating the directory leaf index (also known as the directory hash table) like a normal data file. With this update, gfs2_edit's savemeta function is modified to actually read all the data (the directory hash table) for large directories and traverse the hash table, saving all the leaf blocks. Now, all leaf blocks are saved properly.

BZ#679080

When the fsck.gfs2 tool was resolving block references and no valid reference was found, the reference list became empty. As a consequence, fsck.gfs2 check in pass1b terminated unexpectedly with a segmentation fault. With this update, pass1b is modified to check that the list is empty. The segmentation fault no longer occurs and fsck.gfs2 proceeds as expected.

BZ#731775

The dlm_controld daemon passed error results back to the kernel for POSIX unlock operations flagged with CLOSE. As a consequence, the kernel displayed the "dlm: dev_write no op" error messages, most of them when using non-POSIX locks, flocks. The dlm_controld daemon has been fixed to not pass error results to the kernel for POSIX unlock operations flagged with CLOSE. As a result, error messages no longer appear.

BZ#729071

Previously, the mount.gfs2 utility passed the "loop" option to the GFS2 kernel module which treated it as an invalid option. Mounting a GFS2 file system on loopback devices failed with an "Invalid argument" error message. With this update, mount.gfs2 is modified to avoid passing the "loop" option to the kernel. Mounting GFS2 systems on loopback devices now works as expected.

BZ#728230

Missing sanity checks related to the length of a cluster name caused the cman utility to fail to start. The correct sanity checks have been implemented with this update. The cman utility starts successfully and informs the user of the incorrect value of the cluster name, if necessary.

BZ#726065

The XML format requires special handling of certain special characters. Handling of these characters was not implemented correctly, which caused the cluster.conf file to not function as expected. Correct handling of the characters has been implemented and cluster.conf now works as expected.

BZ#706141

The exact device/mount paths were not compared due to incorrect logic in mount.gfs2 when trying to find mtab entries for deletion. The original entry was not found during remounts and therefore was not deleted. This resulted in double mtab entries. With this update, the realpath() function is used on the device/mount paths so that they match the content of mtab. As a result, the correct original mtab entry is deleted during a remount, and a replacement entry with the new mount options is inserted in its place.

BZ#720668

Previously, mkfs.gfs2 treated normal files incorrectly as if they were block devices. Attempting to create a GFS2 file system on a normal file caused mkfs.gfs2 to fail with a "not a block device" error message. Additional checks have been added so that mkfs.gfs2 does not call functions specific for block devices on normal files. GFS2 file systems can now be created on normal files. However, use of GFS2 in such cases is not recommended.

BZ#719126

The tunegfs2 command line usage message was not updated to reflect the available arguments which are documented in the man page. As a consequence, tunegfs2 printed an inaccurate usage message. The usage message has been updated and tunegfs2 now prints an accurate message.

BZ#719124

Previously, certain argument validation functions did not return error values, and tunegfs2 therefore printed confusing error messages instead of exiting quietly. Error handling has been improved in these validation functions, and tunegfs2 now exits quietly instead of printing the confusing messages.

BZ#694823

Previously, the gfs2_tool command printed the UUID (Universally Unique Identifier) output in uppercase. Certain applications expecting the output being in lowercase (such as mount) could have malfunctioned as a consequence. With this update, gfs2_tool is modified to print UUIDs in lowercase so that they are in a commonly accepted format.

BZ#735917

The qdisk daemon did not allow cman to upgrade the quorum disk device name. The quorum disk device name was not updated when the device was changed and, in very rare cases, the number of qdiskd votes would therefore not be correct. A new quorum API call has been implemented to update the name and votes of a quorum device. As a result, quorum disk device names and votes are updated consistently and faster than before.

BZ#683104

Prior to this update, the fsck.gfs2 utility used the number of entries in the journal index to look for missing journals. As a consequence, if more than one journal was missing, not all journals were rebuilt and subsequent runs of fsck.gfs2 were needed to recover all the journals. Each node needs its own journal; fsck.gfs2 has therefore been modified to use the "per_node" system directory to determine the correct number of journals to repair. As a result, fsck.gfs2 now repairs all the journals in one run.

BZ#663397

Previously, token timeout intervals of corosync were larger than the time it took a failed node to rejoin the cluster. Consequently, corosync did not detect that a node had failed until it rejoined. The failed node had been added again before the dlm_controld daemon asked corosync for the new member list, but dlm_controld did not notice this change. This eventually caused the DLM (Distributed Lock Manager) lockspace operations to get stuck. With this update, dlm_controld can notice that a node was removed and added between checks by looking for a changed incarnation number. Now, dlm_controld can properly handle nodes that are quickly removed and added again during large token timeouts.

BZ#732991

Previously, if a cluster was configured with a redundant corosync ring, the dlm_controld daemon would log harmless EEXIST errors, "mkdir failed: 17". This update removes these error messages so that they no longer appear.

Enhancements

BZ#733345

The corosync IPC port allows, when configured correctly, non-privileged users to access corosync services. Prior to this update, the cman utility did not handle such connections correctly. As a consequence, users were not able to configure unprivileged access to corosync when it was executed using cman. This update adds support to cman to configure unprivileged access. As a result, configured users and groups can now access corosync services without root privileges.

BZ#680930

This update introduces dynamic schema generation, which provides a lot of flexibility for end users. Users can plug into Red Hat Enterprise Linux High Availability Add-On custom resource and fence agents, and still retain the possibility to validate their cluster.conf file against those agents.

BZ#732635, BZ#735912

This update adds support for Redundant Ring Protocol, which aligns the default configuration of cman with corosync. Note that this enhancement is included as a Technology Preview.

BZ#702313

Previously, gfs2_edit saved GFS2 metadata uncompressed. Saved GFS2 metadata sets could have filled up a lot of storage space, and transferring them (for example, for support and debugging) would be slow. This update adds gzip compression to the metadata saving and restoring functions of gfs2_edit. GFS2 metadata sets are now compressed when saving and decompressed when restoring them. The user can specify the compression level with a command line option.

BZ#704178

With this update, the tunegfs2 utility replaces the superblock manipulating feature of gfs2_tool.

BZ#673575

Previously, the fence_scsi agent did not reboot a node when it was fenced. As a consequence, the node had to be rebooted manually before rejoining the cluster. This update provides a script for detecting loss of SCSI reservations. This can be used in conjunction with the watchdog package in order to reboot a failed host.

BZ#820357

Prior to this update, the cmannotifyd did not correctly generate a cluster status notification message at first cluster startup. This update addresses the problem and now cmannotifyd will correctly trigger the notification hooks when the daemon is started.