Hi folks, I am very excited to announce a real major Keepalived release. This release is 2.0.0 and include a HUGE set of extensions in order to push it to the next level... These extensions are targetting high performances and efficiency. The is the the fruit of 2years of hard work, and I would really like the thanks Quentin Armitage here for having performed such a great job. Quentin, it is a real pleasure to have you on board !
In a near future we are going to target documentation (since it is really out of date) and extensions (such as usability and many others). This code can be found on www.keepalived.org website or on https://github.com/acassen/keepalived. If you choose to build keepalived from master github source tree, you will need to have automake and autoconf installed, and will need to execute ./build_setup in the root of the git tree before running anything else. This is documented in the INSTALL file. ChangeLog for this release is : 2018-05-26 Quentin Armitage <quen...@armitage.org.uk> * keepalived-2.0.0 * Beta branch merge into master branch ! 2years of dev here ! * Transition to master as soon as decision is made to do so Previously keepalived waited one further advert interval before transitioning. This meant that previously if a master went down and sent priority 0 message, there was one extra advert interval before the highest priority backup configured the VIP addresses. Now if vrrp instances have high priorities (i.e. close to 255), then the transition to master and configuration of addresses will now occur in a small multiple of advert_interval/256. * Process interface state changes immediately. Previously keepalived waited for advert timer expiry. The problem was that if an interface went down and came back up before the next timer expiry, and addresses, routes and VMACs that we had configured on that interface would be removed, but we wouldn't know about it. * Add support for hot-swappable NICs This also handles interfaces being deleted and restored. * Add vrrp_track_file option. This allows track_scripts, which are run on a frequent scheduled basis, to be replaced with a vrrp_track_file, which contains a number as a text string which is used in the same way as the exit status from a track script. The track_files are only read if they are changed, so external events can update a track file, rather than their status needing to be detected by polling by track scripts. * Add notify fifos. Rather than sending notifications via notify scripts it is now possible to send notify messages via fifos. Not only does this mean that the overhead of executing script for each notification is removed, but it also guarantees the delivery of notifications in the correct order, whereas if the notification is via scripts, there is no guarantee that the scripts will execute in the desired order if two or more notifications are sent in quick succession. There can be a global fifo to process all notifies, and also separate fifos for vrrp and checkers. It is possible to specify a script for keepalived to execute to process the messages on the fifo(s). * Stop logging address addition/deletions if addresses not ours The -a option can be used to override this behaviour and log all address changes. * Transition to fault state if source address for adverts is deleted from interface * Transition to backup state if a VIP or eVIP is removed When we next transition to master the addresses will be restored. If nopreempt is not set, that will be almost immediately. * Make address owner (priority 255) transition to master immediately * Don't process a received advert if the authentication fails * Ignore invalid received adverts totally Previously the master down timer was being updated, which meant that a backup could be stuck in backup state even if the only received adverts were invalid. * Don't reset timer before sending next advert if receive a lower priority advert. This was stopping a higher priority backup instance to stay in backup state. * Log if receive invalid authentication header * Ignore lower priority adverts when backup (to comply with RFCs) This also means that the master down timer wasn't reset, which was causing a delay to becoming master * Fix first advert interval of vrrp instances in a sync group. * Stop two vrrp instances with preempt delay and equal priorities flip-flopping between master and backup state * Make sync group members transition state at same time When first instance makes transition (i.e. when the trigger event occurs) rather than wait for next timer expiry * Process vrrp track script returning a new status code immediately For all instances (and their sync group members), rather than waiting for the next timer expiry on each instance, the instance will transition update it's state immediately. * On reload, make track scripts inherit the state from before reload. This stops vrrp instances transitioning to down and coming back up once the script has run. * Correct the use of adver_int and master_adver_int * Ensure when leaving fault state that a vrrp instance transitions to backup unless it has priority 255 * Remove quick_sync functionality since no longer needed. * Improved code efficiency: * Finding vrrp instance after read timeout * When getting interface information for a new vmac, only request information for that i/f. * Directly update effective priority of vrrp instances when scripts return new status rather than scheduling a thread to do it * Don't run a read timeout on vrrp instance in fault state * Don't run a track script if no vrrp instance is tracking it * Stop checking interface status after every timer expiry since processing interface state changes is now done synchronously * The timeout for the select call had a maximum timeout of 1 second, it now times out only when something needs to happen * The timeout on netlink reads was 500 seconds and this has been extended to 1 day. * Streamline signal handling between main process and child process by using signalfd if available, rather than using a pipe * Minimise searching for an interface struct based on its index by using pointers to the interface structures * Stop opening and closing vrrp scripts before running them. We can detect they are missing from the return of the exec call. * Allow threads that don't need a timeout to never timeout * Calculate the maximum fd number when calling select() rather than specifying the maximum of 1024. * Ignore netlink NEWLINK messages that are only wireless state changes. * Don't check whether timers have expired after select() returns if its timeout didn't expire. * Termination of child processes (scripts) were being handled twice * Don't generate the IP header checksum since the kernel will always generate it. * Maintain pointers to tracking scripts to save seaching a list to find the relevant script. * Vrrp instances to have pointer to interface structure to avoid having to search based in index * Fix the checksum calculation for VRRPv3 unicast peers. * Don't regenerate the full advert packet each time an advert is sent Keepalived now simply updates the necessary fields and calculates the change needed to the checksum. * Detect a vmac interface going down, and make the vrrp instance transition to fault state. Previously the instance would only go down if the underlying interface went down. * Stop weighted track scripts updating priority of sync group members * Make vrrp instances go straight to fault state at startup if a relevant interface is down Previously an instance would start in up state and transition to fault at next timer expiry * Ensure that a sync group starts in backup state unless all members are address owners * Restore master down timer after leaving fault state * Use execve() to execute scripts rather than system(). This saves a fork and an extra process, and also allows the parameters to be parsed once only at startup, rather than each time the script is invoked. * Don't treat a failure to execute a script as a failure of the script * Ensure all scripts receive TERM signal when keepalived terminates * If keepalived is running with an elevated priority, stop running scripts with that elevated priority. * Enable an unweighted tracking script make a vrrp instance which is an address owner transition to fault state * Delay bringing vrrp instances up at startup until after the first completion of the tracking scripts This stops an instance coming up an then being brought back down again after the script completes with a failure. * Reduce number of error messages if a script is not executable * Add linkbeat option per vrrp instance * Fix timer addition on 32-bit systems * Ignore netlink messages for interfaces using linkbeat polling * If priority of vrrp instance changes when in backup due to a vrrp script, reschedule the read timeout * If re-using a VMAC after a reload, ensure it is correctly configured * Don't send priority 0 adverts when transition to fault state unless were in master mode * Identify routes added by keepalived as belonging to keepalived * Enable vrrp instances to be put into fault state if their routes are removed * Add track scripts, track files and track_if to sync groups and deprecate global_tracking (use sync_group_tracking_weight instead, but only if necessary). * Improve AH authentication sequence number handling, and (re)enable sequence number checking for VMACs and sync groups * Remove autoconf/automake generated files from git repo. Script build_setup will create the necessary build environment. * Improve and standardise notifications * Fix not sending RS and VS notifies if omega set * Add no_checker_emails to not send emails every time a checker changes state, but only if a real server changes state * Monitor VIP/eVIP deletion and transition to backup if a VIP/eVIP is removed unloes it is configured with the no-track option. Enjoy, Alexandre _______________________________________________ Please read the documentation before posting - it's available at: http://www.linuxvirtualserver.org/ LinuxVirtualServer.org mailing list - lvs-users@LinuxVirtualServer.org Send requests to lvs-users-requ...@linuxvirtualserver.org or go to http://lists.graemef.net/mailman/listinfo/lvs-users
