1. Name

2. Motivation

3. inotify vs fanotify

4. Installing

5. How to use

6. Example of usage

7. Other uses

8. Clustering

9. Known building issues

10. FreeBSD support

11. Support

12. Developing

13. Articles

14. See also

15. Name

Why "clsync"? The first name of the utility was "insync" (due to inotify) but then I suggested to use "fanotify" instead of "inotify" and utility has been renamed to "fasync". After that I started to intensively write the program and I faced with some problems in "fanotify". So I was have to temporary fallback to "inotify" then I decided that the best name is "Runtime Sync" or "Live Sync" but "rtsync" is a name of some corporation and "lsync" is busy by "lsyncd". So I called it "clsync" that should be interpreted as "lsync but on c" due to "lsyncd" that written on "LUA" and may be used for the same purposes.

UPD: Also I was have to add somekind of clustering support. It's a multicast notifing subsystem to prevent loops on bidirection syncing. So "clsync" also can be interpreted as "cluster live sync". ;)

2. Motivation

This utility has been written for two purposes:

• for making high availability clusters
• for making backups of them

To do a HA cluster I've tried a lot of different solutions, like "simple rsync by cron", "glusterfs", "ocfs2 over drbd", "common mirrorable external storage", "incron + perl + rsync", "inosync", "lsyncd" and so on. When I started to write the utility we were using "lsyncd", "ceph" and "ocfs2 over drbd". However all of this solutions doesn't arrange me, so I was have to write own utility for this purpose.

To do backups we also tried a lot of different solution, and again I was have to write own utility for this purpose.

The best known (for me) replacement for this utility is "lsyncd", however:

• It's code is >½ on LUA. There a lot of problems connected with it, for example:
  • It's more difficult to maintain the code with ordinary sysadmin.
  • It really eats 100% CPU sometimes.
  • It requires LUA libs, that cannot be easily installed to few of our systems.
• It's a little buggy. That may be easily fixed for our cases, but LUA. :(
• It doesn't support pthread or something like that. It's necessary to serve huge directories with a lot of containers right.
• It cannot run rsync for a pack of files. It runs rsync for every event. :(
• Sometimes, it's too complex in configuration for our situation.
• It can't set another event-collecting delay for big files. We don't want to sync big files (>1GiB) so often as ordinary files.
• Shared object (.so file) cannot be used as rsync-wrapper.
• It doesn't support kqueue/bsm

Sorry, if I'm wrong. Let me know if it is, please :). "lsyncd" - is really interesting and useful utility, just it's not appropriate for us.

UPD.: Also clsync had been used to replace incron/csync2/etc in HPC-clusters for syncing /etc/{passwd,shadow,group,shells} files.

3. inotify vs fanotify:

It's said that fanotify is much better than inotify. So I started to write this program with using of fanotify. However I encountered the problem, that fanotify was unable to catch some important events at the moment of writing the program, like "directory creation" or "file deletion". So I switched to "inotify", leaving the code for "fanotify" in the safety... So, don't use "fanotify" in this utility ;).

4. Installing

Debian/ubuntu-users can try to install it directly with apt-get:

apt-get install clsync

If it's required to install clsync from the source, first of all, you should install dependencies to compile it. On debian-like systems you should execute something like:

apt-get install libglib2.0-dev autoreconf gcc

Next step is generating Makefile. To do that usually it's enought to execute:

autoreconf -i && ./configure

Next step is compiling. To compile usually it's enough to execute:

make

Next step is installing. To install usually it's enough to execute:

su -c 'make install'

5. How to use

How to use is described in "man" ;). What is not described, you can ask me personally (see "Support").

6. Example of usage

Example of usage, that works on my PC is in directory "examples". Just run "clsync-start-rsyncdirect.sh" and try to create/modify/delete files/dirs in "example/testdir/from". All modifications should appear (with some delay) in directory "example/testdir/to" ;)

For dummies:

pushd /tmp
git clone https://github.com/xaionaro/clsync
cd clsync
autoreconf -fi
./configure
make
export PATH_OLD="$PATH"
export PATH="$(pwd):$PATH"
cd examples
./clsync-start-rsyncdirect.sh
export PATH="$PATH_OLD"

Now you can try to make changes in directory "/tmp/clsync/examples/testdir/from" (in another terminal). Wait about 7 seconds after the changes and check directory "/tmp/clsync/examples/testdir/to". To finish the experiment press ^C (control+c) in clsync's terminal.

cd ../..
rm -rf clsync
popd

Note: There's no need to change PATH's value if clsync is installed system-wide, e.g. with

make install

For dummies, again (with "make install"):

pushd /tmp
git clone https://github.com/xaionaro/clsync
cd clsync
autoreconf -fi
./configure
make
sudo make install
cd examples
./clsync-start-rsyncdirect.sh

Directory "/tmp/clsync/examples/testdir/from" is now synced to "/tmp/clsync/examples/testdir/to" with 7 seconds delay. To terminate the clsync press ^C (control+c) in clsync's terminal.

cd ..
sudo make uninstall
cd ..
rm -rf clsync
popd

For really dummies or/and lazy users, there's a video demonstration: http://ut.mephi.ru/oss/clsync

7. Other uses

For example, command

ionice -c 3 clsync -L /dev/shm/clsync --exit-on-no-events -x 23 -x 24 -M rsyncdirect -S $(which rsync) -W /path/from -D /path/to -d1

may be used to copy "/path/from" into "/path/to" with sync up of changes made (in "/path/from") while the copying. It will copy new changes over and over until there will be no changes, and then clsync will exit. It may be used as atomicity-like recursive copy.

Or command

clsync -w5 -t5 -T5 -x1 -W /var/www/site.example.org/root -Mdirect -Schown --uid 0 --gid 0 -Ysyslog -b1 -- --from=root www-data:www-data %INCLUDE-LIST%

may be used to fix files owner in runtime. This may be used as a temporary solution for fixing file privileges of misconfigured web-servers (it's well-known problem of apache users).

8. Clustering

I've started to implement support of bi-directional syncing with using multicast notifing of other nodes. However it became a long task, so it was suspended for next releases.

However let's solve next hypothetical problem. For example, you're using LXC and trying to replicate containers between two servers (to make failover and load balancing).

In this case you have to sync containers in both directions. However, if you just run clsync to sync containers to neighboring node on both of them, you'll get sync-loop [file-update on A causes file-update on B causes file-update on A causes ...].

Well, in this case I with my colleagues were using separate directories for every node of cluster (e.g. "/srv/nodes/<NODE NAME>/containers/<CONTAINERS>") and syncing every directory only in one direction. That was failover with load-balancing, but very unconvenient. So I've started to write code for bi-directional syncing, however it's no time to complete it :(. So Andrew Savchenko proposed to run one clsync-instance per container. And this's really good solution. It's just need to start clsync-process when container starts and stop the process when containers stops. The only problem is split-brain, that can be solved two ways:

• by human every time;
• by scripts that chooses which variant of container to save.

Example of the script is just a script that calls "find" on both sides to determine which side has the latest changes :)

UPD: I've added option "--modification-signature" that helps to prevent syncing file, that is not changed. You can easily use it to prevent sync-loops for bi-directional syncing.

9. Known building issues

May be problems with "configuring" or compilation. In this case just try next command: echo '#define REVISION "-custom"' > revision.h; gcc -std=gnu99 -D_FORTIFY_SOURCE=2 -DPARANOID -pipe -Wall -ggdb3 --param ssp-buffer-size=4 -fstack-check -fstack-protector-all -Xlinker -zrelro -pthread $(pkg-config --cflags glib-2.0) $(pkg-config --libs glib-2.0) -ldl *.c -o /tmp/clsync

10. FreeBSD support

clsync has been ported to FreeBSD.

FreeBSD doesn't support inotify, so there're 3.5 ways to use clsync on it:

• using libinotify;
• using BSM API (with or without a prefetcher thread);
• using kqueue/kevent directly.

Here's an excerpt from the manpage:

 Possible values:
        inotify
               inotify(7) [Linux, (FreeBSD via libinotify)]

               Native, fast, reliable and well tested Linux FS monitor subsystem.

               There's no essential performance profit to use "inotify"  instead  of
               "kevent"  on FreeBSD using "libinotify". It backends to "kevent" any‐
               way.

               FreeBSD users: The libinotify on FreeBSD is still not ready and unus‐
               able for clsync to sync a lot of files and directories.

        kqueue
               kqueue(2) [FreeBSD, (Linux via libkqueue)]

               A  *BSD  kernel  event  notification  mechanism (inc. timer, sockets,
               files etc).

               This monitor subsystem cannot determine file creation event,  but  it
               can determine a directory where something happened. So clsync is have
               to rescan whole dir every  time  on  any  content  change.  Moreover,
               kqueue  requires  an  open()  on  every watched file/dir. But FreeBSD
               doesn't allow to open() symlink itself (without following)  and  it's
               highly  invasively  to open() pipes and devices. So clsync just won't
               call open() on everything except regular files and directories.  Con‐
               sequently,  clsync  cannot  determine  if  something  changed in sym‐
               link/pipe/socket and so on.  However it still  can  determine  if  it
               will  be created or deleted by watching the parent directory and res‐
               caning it on every appropriate event.

               Also this API requires to open every monitored file and directory. So
               it  may  produce  a  huge  amount  of  file descriptors. Be sure that
               kern.maxfiles is big enough (in FreeBSD).

               CPU/HDD expensive way.

               Not well tested. Use with caution!

               Linux users: The libkqueue on Linux is not working. He-he :)

        bsm
               bsm(3) [FreeBSD]

               Basic Security Module (BSM) Audit API.

               This is not a FS monitor subsystem, actually. It's  just  an  API  to
               access  to  audit information (inc. logs).  clsync can setup audit to
               watch FS events and report it into log. After that clsync  will  just
               parse the log via auditpipe(4) [FreeBSD].

               Reliable,  but  hacky  way.  It requires global audit reconfiguration
               that may hopple audit analysis.

               Warning!  FreeBSD has a limit for queued events. In  default  FreeBSD
               kernel it's only 1024 events. So choose one of:
                      - To patch the kernel to increase the limit.
                      - Don't use clsync on systems with too many file events.
                      - Use bsm_prefetch mode (but there's no guarantee in this case
                      anyway).
               See also option --exit-on-sync-skip.

               Not  well  tested.  Use   with   caution!    Also   file   /etc/secu‐
               rity/audit_control will be overwritten with:
                      #clsync

                      dir:/var/audit
                      flags:fc,fd,fw,fm,cl
                      minfree:0
                      naflags:fc,fd,fw,fm,cl
                      policy:cnt
                      filesz:1M
               unless it's already starts with "#clsync\n" ("\n" is a new line char‐
               acter).

        bsm_prefetch
               The same as bsm but all BSM events will be  prefetched  by  an  addi‐
               tional  thread  to prevent BSM queue overflow. This may utilize a lot
               of memory on systems with a high FS events frequency.

               However the thread may be not fast enough to unload  the  kernel  BSM
               queue. So it may overflow anyway.

 The default value on Linux is "inotify". The default value on FreeBSD is "kqueue".

I hope you will send me bugreports to make me able to improve the FreeBSD support :)

11. Support

To get support, you can contact with me this ways:

• Official IRC channel of "clsync": irc.freenode.net#clsync
• Where else can you find me: IRC:SSL+UTF-8 irc.campus.mephi.ru:6695#mephi,xaionaro,xai
• And e-mail: [email protected], [email protected]; PGP pubkey: 0x8E30679C

12. Developing

I started to write "DEVELOPING" and "PROTOCOL" files. You can look there if you wish. ;)

I'll be glad to receive code contribution :)

13. Articles

Russian:

• HA clustering
• syncing to many nodes
• atomic sync

LVEE (Russian):

• clsync - live sync utility (abstract) presentation
• clsync progress: security and porting to freebsd

14. See also

• lrsync

                                           -- Dmitry Yu Okunev <[email protected]> 0x8E30679C