A little background - Since Redhat stopped updating gcc versions for RHEL6, we had to implement our own gcc5 deployment.
We can't upgrade to RHEL7 at this time, nor will it solve this problem as Redhat hadn't deployed gcc5 for rhel7 yet.
Our current issue is with /usr/src/kernels/$(uname -r)/include/linux/compiler-gcc5.h
That path and the compiler-gcc*.h files are owned by kernel-devel, and a gcc5 one exists only for fedora (only for distros where gcc5 has been introduced officialy).
I can easily create an rpm that will own the file (the reason is mainly so we will receive a conflict if RH ever deploys kerenl-devel with that file, so we can stop deploying our own) for the current kernel, but I'd like to use %triggerin to copy the file to any new kernel-devel installed on the system.
using uname -r wouldn't work for 2 reasons:
it's evaluated at rpmbuild time, not install time.
it evaluates the current running kernel, not the future one just installed.
I can overcome 1 by using a %post script, I think, but can't figure out how to overcome 2, nor how to own the file by my rpm in it's final location (can I %ghost a path I don't know in advance, or use something like dkms?)
Is there a way to know the exact version of the rpm that triggers the %triggerin and incorporate it into the path?
Is there an alternate solution?
Related
Currently I am trying to install a rpm secured_soft_2.0.0.rpm and i am
unable to install it as we already have secured_soft_1.3.0 installed.
Requirement is that we need to have both the versions installed.
Complexities :
These package inturn have dependent rpm's (lot of them ) and all these
interdependent rpm's also have versions
ex: secured_soft_1.3.0 works only with packages which are of version 1.3,
and secured_soft_2.0.0.rpm work only with dependecies of version 2.0 only.
So all these dependencies also need to be reinstalled and even these
dependenies should be parallely installed, without deleting old.
Finally , both these versions contain shared libraries and these shared
lib's do not have version numbers in their name.
#rpm -ivh secured_soft_2.0.0.rpm
error: Failed dependencies:
init-class >= 1.4.17.1-1 is needed by secured_soft_2.0.0.rpm
init-connection-interface >= 2.0.11.0 is needed by secured_soft_2.0.0.rpm
init-logger >= 2.0.11.0 is needed by secured_soft_2.0.0.rpm
init-security >= 2.0.11.0 is needed by secured_soft_2.0.0.rpm
As i have specified we already have secured_soft_1.3.0.rpm installed and
above dependencies are also available but of different version.
So we need the to install above dependencies and also need the old version's
of dependencies for the old rpm's to work
ex : secured_soft_2.0.0.rpm has libArt.so libSec.so and so on
which are copied to /usr/lib
Similarly secured_soft_1.3.0.rpm also has libArt.so libSec.so and so on
which are already available in /usr/lib
I tried to rename the so's but still iam not able to install.
Is it possible to change the location for these so's and get the things done
Is there any way we can do it.
At the moment, iam stuck here and would need advice on this
Appreciate any help on this.
Since the programs use the same filenames, and you need to put them on the same machine, you might be able to move the older version to another directory tree and make it work there.
You can do this with many applications which do not have compiled-in pathnames.
For instance,
install the older version (this sounds like where you are starting from)
use rpm -ql for each of the packages containing unversioned executables, libraries and associated files.
use tar to capture an archive of those files, relative to /usr (but omitting directories not owned by the packages).
create a new directory, e.g., /usr/local/myapp and untar the older version there.
update configuration files in the new location as needed
For applications such as this, I would run in a script that updates PATH (and perhaps sets LD_LIBRARY_PATH) to force the program to run from the new location. You can verify if this works using tools such as strace and lsof, i.e., by looking for the files that the program opens.
Once you have the older version working properly in the new location, you can uninstall its rpms and install the new version of the application.
Caveat: If the newer package is copied from a newer version of the operating system, however, the task is likely to be beyond your ability, whether or not you choose the alternative approach of recompiling the newer packages to fit on the existing system.
Building new/custom packages is one route to recompiling the newer version. If you have the source-RPMs for each part, that is a starting point:
extract the files from the source-RPM, e.g., using a script such as unrpm (see for example HowTo: Extract an RPM Package Files Without Installing It), and
copy those extracted files to their as-expected locations in your build-tree, e.g., $HOME/rpmbuild/SOURCES and $HOME/rpmbuild/SPECS
modify the spec-file to use the alternative location
build the new/modified package using the modified spec-file.
No, out of the box, you cannot.
I'd highly recommend looking into Docker, where you can throw each one into their own container and let them take care of all their dependency problems.
Consider these 4 pieces of software:
COBRA 2.05
LibSBML 5.10
MATLAB R2013a (Also known as 8.1, 64-bit; MATLAB no longer supports 32-bit Linux anyway)
A 64-bit Linux OS (such as Ubuntu 14.04 or the latest Mint but not restricted to them)
Intro
The COBRA toolbox is an optimization suite that runs on top of MATLAB aimed at the development of MATLAB code for metabolic network modelling. Such a "network" is a system of equations that can have a very large number of equations and variables (such as thousands). Therefore, routines to read and write those large models according to some format specification are a must-have, and COBRA uses the standard SBML for that.
Problem
Unlike the Windows versions, the Linux binary packages do not integrate well out-of-the-box: to begin with, the pre-compiled Linux binary of libSBML (open-source) available for download does not come with MATLAB support. If one tries to use the pre-compiled libSBML, COBRA won't find the "MATLAB bindings" and therefore won't be able to, for example, read and write SBML XML files from the disk in a m-script.
The question
What needs to be done to make COBRA 2.05 running on top of MATLAB R2013a under Linux (Ubuntu 14.04 or the latest Mint, but this is not likely distro-specific) able to read and write SBML XML files? In other words, what needs to be done system-wide to make COBRA pass its own testSBML test?
This is how I got it working and what I could learn from all the hassle regarding how my Linux box works. I hope I am not forgetting/missing/mistaking anything.
1. MATLAB
1.1. Install MATLAB
Install it in its default location (you will need root access for this), don't be stubborn like I tried to be. Why:
1.1.1. Integration
It is very likely you will want to install some other software that uses the MATLAB framework at some point, and in the real world software doesn't always find other software even if you know how to tell it where to look for.
1.1.2. Make your life easier
Even though it seems like a good idea to install a big software in a place where you have lots of available space and that you can use in multiple machines (specially in Linux, which doesn't have that abomination called Registry, and has symbolic links), that would only perhaps be a good idea - apart from item 1.1.1 - if that place is a partition with a Linux filesystem, since at some point, some executable/script will need execution permission, and mounting the entire partition with execution permission for all its files is rather unsafe. Therefore, do not put MATLAB in an NTFS partition of an external HD; perhaps creating a Linux partition in the external HD just for Linux-specific stuff could work for this matter, but how much hassle is that?
1.2. Setup MATLAB so people and other software can launch it
Even though I have seen somewhere that the MATLAB installer eventually shows an option to create symbolic links in the system path for convenience, it didn't in my case. But that is OK, since I would have to replace the symbolic link /usr/local/bin/matlab by the following shell-script (same path, same filename) anyway:
#!/bin/sh
export LD_LIBRARY_PATH=/usr/local/lib:/usr/lib/x86_64-linux-gnu:$LD_LIBRARY_PATH
exec /usr/local/MATLAB/R2013a/bin/matlab $*
OBS: That LD_LIBRARY_PATH is needed for MATLAB to find SBML bindings later and to be able to use them. Also, if you install some third-party solver such as TOMLAB, you will most likely need to add some more paths in this little launcher script.
OBS 2: In my case, the installation script didn't automatically create any launchers or shortcuts, but I have found an iconless and extension-less Matlab 8.01 file and a matlab icon as a png file, and that first file was a template .desktop file that I could edit to fit my needs and put in /usr/share/applications so the Unity Dash would find it. The contents of this Matlab.desktop file are:
[Desktop Entry]
Type=Application
Icon=/usr/local/MATLAB/R2013a/Matlab.png
Name=Matlab 8.01
Comment="Start Matlab 8.01"
Exec=/usr/local/MATLAB/R2013a/bin/matlab -desktop
Categories=Development;
Name=Matlab 8.01
GenericName=Matlab 8.01
Comment="Start Matlab 8.01"
2. libSBML
2.1. Install libSBML
libSBML is provided by a deb package specific for Ubuntu (and for CentOS), and also versions for several flavours of Windows and MacOSX (their home page: http://sbml.org/Software/libSBML). Guess which is the only platform whose binaries weren't compiled with MATLAB support? Linux, of course. That means you will need to compile from source (and that the deb package is therefore useless to you). To compile:
2.1.1. Install dependencies
The dependency libxml2-dev (from software manager or from a terminal):
sudo apt-get install libxml2-dev
2.1.2. Save yourself some time in the future
Usually, one would do configure, make and then make install. But this is not recommended for the same reason as installing anything that doesn't come in a pretty little package: you will loose control of which files went where, and will need to keep the source-code to be able to run make uninstall if you need to uninstall it later. So, install checkinstall and use it to replace the step make install, since checkinstall creates a package for your system that can be later uninstalled or reinstalled just as any regular packaged software (from software manager or from a terminal):
sudo apt-get install checkinstall
2.1.3. Configure the compiling-process
Get LibSBML source code and extract it to some folder. From a terminal, navigate to that folder and configure the compilation:
./configure --with-matlab
OBS: with the with-matlab flag, the configure script will fail it it cannot find an executable whose filename is matlab. If it fails, it outputs that the matlab file could not be found, but the test it performs is actually both for the existence of the file and whether it is executable. So, if the file is in an NTFS partition, configure will fail even if it finds the file, but will tell you the file couldn't be found. You can enforce it to look for the executable in /path/to/matlab/root by passing (it will look for a bin folder inside that path, and for the executable inside that bin folder):
./configure --with-matlab=/path/to/matlab/root
OBS: This will install libSBML in the default location: /usr/local/lib. Again, it is a good idea to just let it install in its default location, but if you need to change it, you can pass the path with the flag: --prefix=/your/installation/path
OBS 2: You might ask why libSBML needs to find and execute matlab to be compiled with support for it: it needs to fire up MATLAB later to build MEX-files (compiled MATLAB code), so I would speculate you wouldn't be able to install libSBML after all if your MATLAB has no compiler to generate MEX-files.
2.1.4. Build and install libSBML
make
checkinstall
VERY IMPORTANT OBS:
I) checkinstall asks for confirmation of the metadata of the package it is about to create. In my case, the string for the version field came by default as "Source" (without the quotes), which caused checkinstall to fail because dpkg (the system tool that actually builds the deb file) failed complaining the version number must start with, well, a number. So, save yourself some time and make sure the string in the version field starts with a number (i.e. "5.10", without the quotes obviously)
II) checkinstall asks if you want to exclude from the future package files that the make install command would put in your home folder and says it is a good idea to exclude them. LibSBML has a test.xml file that it needs to be in the $HOME folder later, or else it won't let you integrate with MATLAB. And even though it tells you a test.xml is missing, it doesn't tell you where that file should be or if that file was something that came with the library. I only noticed it because checkinstall had found a $HOME/test.xml reference earlier (that I excluded from the package, of course) and I had found that odd. So, save yourself some time and exclude $HOME/test.xml from the package generated by checkinstall, and then search for test.xml inside the source-code folder and copy it to $HOME as soon as libSBML finishes being installed by checkinstall.
2.2. Integrate libSBML to MATLAB
Fire up MATLAB, navigate to where the SBML MATLAB-bindings were installed in step 2.1.5 (in my case: /usr/local/lib) and run the file installSBML.m that should be there.
2.2.1. Shared libraries problems
In my case, I had an error due to an old unresolved issue: libstdc++.so.6 not having a reference to GLIBCXX_3.4.15. Turns out that MATLAB was trying to use a libstdc++.so.6.0.13 (libstdc++.so.6 was a symbolic link pointing to this file) that came with it in /usr/local/MATLAB/R2013a/sys/os/glnxa64, which indeed didn't have that reference (one could verify that by issuing:
strings /usr/local/MATLAB/R2013a/sys/os/glnxa64/libstdc++.so.6.0.13 | grep GLIBC
). My system has a libstdc++.so.6.0.19 located in /usr/lib/x86_64-linux-gnu that has that reference, so I enforced MATLAB to use 6.0.19 one by setting the LD_LIBRARY_PATH properly (refer to step 1.2) and also by renaming the libstdc++.so.6 that came with MATLAB to something else so it would not find it and would keep looking until it found my system's. A friend of mine running Linux Mint didn't need to rename anything: for him, setting the LD_LIBRARY_PATH was enough.
2.2.2. Other problems
installSBML.m will fail if it doesn't find that $HOME/test.xml file mentioned in step 2.1.5, regardless of whether the library actually works. It assumes that if it could not test itself using a file that it assumes to be in $HOME, the user shouldn't have the option to install it anyway.
3. COBRA / SBML toolbox
3.1. Setup COBRA
In MATLAB, navigate to <YOUR_COBRA_ROOT_FOLDER_HERE>/external/toolboxes/SBMLToolbox-4.1.0/toolbox and run the file install.m there. You should have all set so it finds the MATLAB-bindings you set up in step 2.2.
3.2. MATLAB setpaths problems
I had to manually edit the file /usr/local/MATLAB/R2013a/toolbox/local/pathdef.m as root to include the folder /usr/local/lib (where libSBML and its MATLAB-bindings are) to make that setting persistent. Every time I restarted MATLAB, its setpath had gone back to the default, no matter if I started MATLAB as root when setting its setpath via the MATLAB GUI.
3.3. Test
Now you have hopefully connected all the dots. Try it: in MATLAB, navigate to <YOUR_COBRA_ROOT_FOLDER_HERE> and issue:
initCobraToolbox
testAll
If you haven't got any third-party solvers installed and configured, it should pass 14 of the 19 tests, including the SBML test (testSBML). Now you can load SBML files into MATLAB and play with them.
I also needed to add a symbolic link from /usr/local/lib/libsbml.so.5 to the MATLAB sys folder by:
sudo ln -s /usr/local/lib/libsbml.so.5 /usr/local/MATLAB/R2014a/sys/os/glnxa64/
This finally made the installation possible.
I installed using Cmake. To do this it is necessary to find the FindMatlab.cmake in the source package and insert the MATLAB path manually!
.............
elseif(EXISTS "/Applications/MATLAB_R2008a.app/")
set(MATLAB_ROOT_PATH "/Applications/MATLAB_R2008a.app/")
endif()
else()
if (EXISTS "/usr/local/MATLAB/R2014a/")
set(MATLAB_ROOT_PATH "/usr/local/MATLAB/R2014a/")
endif()
endif()
..........
FYI, to resolve the shared library issue at point 2.2.1 I needed to install the package matlab-support (in Ubuntu repositories)
Alright so after trying to chase down the dependencies for various pieces of software for the n-th time and replicating work that various people do for all the different linux distributions I would like to know if there is a better way of bundling various pieces of software into one .rpm or .deb file for easier distribution.
My current set up for doing this is a frankenstein monster of various tools but mainly Vagrant and libguestfs (built from source running in Fedora because none of the distributions actually ship it with virt-diff). Here are the steps I currently follow:
Spin up a base OS using either a Vagrant box or by create one from live CDs.
Export the .vmdk and call it base-image.
Spin up an exact replica of the previous image and go to town: use the package manager,
or some other means, to download, compile, and install all the pieces that I need. Once again, export the .vmdk and call it non-base-image.
Make both base images available to the Fedora guest OS that has libguestfs.
Use virt-diff to diff the two images and dump that data to file called diff.
Run several ruby scripts to massage diff into another format that contains the information I need and none of the stuff I don't like things in /var.
Run another script to generate a command script for guestfish with a bunch of copy-out commands.
Run the guestfish script.
Run another script to regenerate the symlinks from diff because guestfish can't do it.
Turn the resulting folder structure into a .deb or .rpm file and ship it.
I would like to know if there is a better way to do this. You'd think there would be but I haven't figured it out.
I would definitely consider something along the lines of:
A)
yum list (select your packages/dependencies whatever)
use yumdownloader on the previous list (or use th pkgs you have already downloaded)
createrepo
ship on media with install script that adds the cd repo to repolist, etc.
or B)
first two steps as above, then pack the rpms into an archive build a package that contains all of the above and kicks off the actual install of the rpms (along the lines of rpm -Uvh /tmp/repo/*) as a late script (in the cleanup phase, maybe). Dunno if this can be done avoiding locks on the rpm database.
I think you reached the point of complexity - indeed a frankenstein monster - where you should stop fearing of making proper packages with dependencies. We did this in my previous work - we had a set of fabricated rpm packages - and it was very easy and straightforward, including:
pre/post install scripts
uninstall scripts
dependencies
We never had to do anything you just described. And for the customer, installing even a set of packages was very easy!
You can follow a reference manual of how to build RPM package for more info.
EDIT: If you need a single installation package, then create this master packge, that would contain all the other packages (with dependencies set properly) and installed them in the post-install script (and uninstalled them in the uninstall script).
There are mainly 3 steps to make a package with all dependencies (let it be A, B & C).
A. Gather required files.
There are many way to gather files of the main software and its dependencies. In order to get all the dependices and for error free run you need to use a base OS (i.e live system)
1. Using AppDirAssistant
This app is used by www.portablelinuxapps.org to create portable app directory. They scan and watch for the files accessed by the app to find required.
2. Using chroot & overlayfs
In this method you don't need to boot into live cd instead chroot into it.
a. mount the .iso # /cdrom and
b. mount the filesystem(filesystem.squashfs) # another place, say # /tmp/union/root
c. Bind mount /proc # /tmp/union/root/proc
d. Overlay on it
mount -t overlayfs overlayfs /tmp/union/root -o lowerdir=/tmp/union/root,upperdir=/tmp/union/rw
e. Chroot
chroot /tmp/union/root
Now you can install packages using apt-get or another method (only from the chrooted terminal). All the changed files are stored # /tmp/union/rw. Take files from there.
3. Using manually collected packages
Use package manager to collect dependencies. For example
apt-get install package --print-uris will print download uris for dep packages. Using this uris download packages and extract all (dpkg -x 1.deb ./extracted).
B. Clean garbage files
After gathering files remove unwanted files
C. Pack files
1. Using appImageAssistance
If you manually gathered files then you need to copy appname.desktop file from ./usr/share/applications to root of directory tree. Also copy file named AppRun from another app or extract it from AppDirAssistance.
2. Make a .deb or .rpm using gathered files.
Is the problem primarily that of ensuring that your customers have installed all the standard upstream distro packages necessary for your package to run?
If that's the case, then I believe the most straightforward solution would be to leverage the yum and apt infrastructure to have those tools track down and install the necessary prerequisite packages.
If you supply a native yum/apt repository with complete pre-req specs (the hard work you've apparently already completed). Then the standard system install tool takes care of the rest. See link below for more on creating a personal repository for yum/apt.
For off-line customers, you can supply media with your software, and a mirror - or mirror subset - of the upstream distro, and instructions for adding them to yum config/apt config.
Yum
Creating a Yum Repository in the Fedora Deployment Guide
Apt
How To Setup A Debian Repository on the Debian Wiki
So your customers aren't ever going to install any other software that might specify a different version of those dependencies that you are walking all over, right?
Why not just create your own distro if you're going to go that far?
Or you can just give them a bunch of packages and a single script that does rpm -i dep1 dep2 yourpackage
I have a debian package that I built that contains a tar ball of the files, a control file, and a postinst file. Its built using dpkg-deb and it installs properly using dpkg.
The modification I would like to make is to have the installation directory of the files be determined at runtime based on an environment variable that will be set when dpkg -i is run on the deb file. I echo out the environment variable in the postinst script and I can see that its set properly.
My questions:
1) Is it possible to dynamically determine the installation directory at runtime?
2) If its possible how would I go about this? I have read about the rules file and the mypackage.install files but I don't know if either of these would allow me to accomplish this.
I could hack it by copying the files to the target location in the posinst script but I would prefer to do it the right way if possible.
Thanks in advance!
So this is what I found out about this problem over the past couple of weeks.
With prepackaged binaries you can't build a debian package with a destination directory dynamicall determined at runtime. I believe that this might be possible if installing a package that is built from source where you can set the install directory using configure. But in this case since these are embedded Ubuntu machines they don't have make so I didn't pursue such an option. I did work out a non traditional method (hack) for installing that did work. Since debian packages simply contain a tar ball relative to / simply build your package relative to a directory under /tmp. In the postinst script you can then determine where to copy the files from the archive into a permanent location.
I expected that after rebooting and the automatic deletion of the subdirectory under /tmp that dpkg might not know that the file package existed. This wasn't a problem. When I ran 'dpkg -l myapp' it showed as still installed. Updating the package using dpkg/apt-get also worked without a hitch.
What I did find is that if you attempted to remove the package using 'dpkg -r myapp' that dpkg would try and remove /tmp which wasn't good. However /tmp isn't easily removed so it never succeeded. Plus in our situation we never remove packages but instead simply upgrade them.
I eventually had to abandon the universal package due to code differences in the sources resulting in having to recompile per platform but I would have left it this way and it did work.
I tried using --instdir to change the install directory of the package and it does relocate the files but dpkg fails since the dpkg file can't be found relative to the new instdir. Using --instdir is sort of like a chroot. I also tried --admindir and --root in various combinations to see if I could use the dpkg system relative to / but install relocate the files but they didn't work. I guess rpm has a relocate option that works but not Ubuntu.
You can also write a script that runs dpkg-deb with a different environment for 6 times, generating 6 different packages. When you make a modification, you simply have to run your script, and all 6 packages gets generated and you can install them on your machines avoiding postinst hacking!
Why not install to a standard location, and simply use a postinst script to create symbolic links to the desired location? This is much cleaner, and shouldn't break anything in dpk -I.
I have a libfoo-devel rpm that I can create, using the trick to override _topdir. Now I want to build a package "bar" which has a BuildRequires 'libfoo-devel". I can't seem to find the Right Way to get access to the contents of libfoo-devel without having to install it on the build host. How should I be doing it?
EDIT:
My build and target distros are both SuSE.
I prefer solutions that don't require mock, since I believe SuSE does not include it in its stock repo.
Subsequent EDIT:
I believe that the answer I seek is in the build package. Perhaps it's SuSE's answer to mock? Or it's the distributed version of the oBS service?
DESCRIPTION
build is a tool to build SuSE Linux
RPMs in a safe and clean way. build
will install a minimal SuSE Linux as
build system into some directory and
will chroot to this system to compile
the package. This way you don't risk
to corrupt your working system (due to
a broken spec file for example), even
if the package does not use BuildRoot.
build searches the spec file for a
BuildRequires: line; if such a line is
found, all the specified rpms are
installed. Otherwise a selection of
default packages are used. Note that
build doesn't automatically resolve
missing dependencies, so the specified
rpms have to be sufficient for the
build.
Note that if you really don't need libfoo-devel installed to build package bar the most sensible alternative would be to remove libfoo-devel from the BuildRequires directive (and maybe put the requirement where it belongs).
However, if you cannot do that for some reason, create a "development" rpm database. Basically it involves using rpm --initdb --root /path/to/fake/root. Then populate it with all of the "target packages" of your standard distro installation.
That's a lot of rpm --install --root /path/to/fake/root --justdb package-name.rpm commands, but maybe you can figure out a way to copy over your /var/lib/rpm/* database files and use those as a starting point. Once you have the alternative rpm database, you can fake the installation of the libfoo-devel package with a --justdb option. Then you'll be home free on the actual rpm build.
If neither mock nor the openSUSE Build Service are a viable choice then you will have to buckle down and install the package, either directly or in a chroot; the package provides files that the SRPM packager has decided are required to build, and hence is in the BuildRequires tag.