On 07/13/2018 03:55 PM, Severin Gehwolf wrote:

On Fri, 2018-07-13 at 15:30 +0200, Mikolaj Izdebski wrote:

...

On 07/13/2018 11:00 AM, Severin Gehwolf wrote:

...

See this bug for some background: https://bugzilla.redhat.com/show_bug.cgi?id=1600426

Any Java package gets a "Requires: javapackages-tools" whether they want it or not. It's part of the Java Packaging Guidelines[1]. Does anybody remember the rationale as to why that exists? If the answer is directory ownership for dirs like /usr/lib/jvm/ et.al. I'd consider changing the guidelines to auto-requires javapackages-filesystem in F29+.

Typically packages require javapackages-tools during runtime for one or both of two following major reasons:

1. Directory ownership. javapackages-tools requires

javapackages-filesystem, which owns many directories into which Java packages install their files.

2. Java functions library. javapackages-tools contains library of shell

functions at /usr/share/java-utils/java-functions which is imported by application launcher scripts written in shell. These functions may be used to help loading system-wide or application-specific configuration, constructing application classpath, determining proper JVM/JDK to use with the application, and finally invoking JVM with appropriate flags, including ABRT agent flags, and finally invoking JVM.

(There are also other minor, but valid reasons to require javapckages-tools, such as for configuration files it contains.)

Changing auto-requires from javapackages-tools to javapackages-filesystem and rebuilding all packages would break most of launcher scripts. Explicit requires on javapackages-tools would have to be added to packages that have launchers in order to fix them, reverting some of benefits of the change.

Question is how many of the Java packages use javapackages-tools for 2). Any ideas how we'd figure this out in some automated way?

Should be easy. Check packages that require javapackages-tools and check if their executable files import /usr/share/java-utils/java-functions.

...

Lets look at case of byteman, mentioned in the bug. Switching auto-requires to javapackages-filesystem would not achieve much, as byteman ships launcher scripts in /usr/bin/ that require javapackages-tools to function properly, so explicit requires on javapackages-tools would probably need to be added. Moreover, byteman requires objectweb-asm, which also contains launcher script and therefore would still need to require javapackages-tools for reason nr 2. above.

Not really. byteman does not use what you describe as 2) from javapackages-tools. That's just upstream functionality. Byteman also doesn't require objectweb-asm (at runtime; its BR only), so that's not good of an argument either. It would run fine with just java-openjdk (JDK 10).

Right, byteman does not use java-functions library. After I checked its file lists wrongly assumed it followed packaging guidelines for launcher scripts, but it does not.

Also from what I can see byteman does require objectweb-asm during runtime, but this is probably a bug as byteman bundles ASM, IIRC.

$ dnf -q --repo rawhide repoquery --requires byteman | grep asm mvn(org.ow2.asm:asm) mvn(org.ow2.asm:asm-analysis) mvn(org.ow2.asm:asm-commons) mvn(org.ow2.asm:asm-tree)

$ dnf -q --repo rawhide repoquery --whatprovides 'mvn(org.ow2.asm:asm)' objectweb-asm-0:6.2-1.fc29.noarch

...

For the auto-requries to have desired effect many packages would need to be split into library and application parts. Like for example, maven binary package contains only launcher script (plus manpage and bash completion associated with the script), while maven-lib contains library code.

Perhaps that would be a change for the better?

Perhaps.

Hi Mikolaj,

On Fri, 2018-07-13 at 15:30 +0200, Mikolaj Izdebski wrote:

On 07/13/2018 11:00 AM, Severin Gehwolf wrote:

...

See this bug for some background: https://bugzilla.redhat.com/show_bug.cgi?id=1600426

Any Java package gets a "Requires: javapackages-tools" whether they want it or not. It's part of the Java Packaging Guidelines[1]. Does anybody remember the rationale as to why that exists? If the answer is directory ownership for dirs like /usr/lib/jvm/ et.al. I'd consider changing the guidelines to auto-requires javapackages-filesystem in F29+.

Typically packages require javapackages-tools during runtime for one or both of two following major reasons:

1. Directory ownership. javapackages-tools requires

javapackages-filesystem, which owns many directories into which Java packages install their files.

2. Java functions library. javapackages-tools contains library of shell

functions at /usr/share/java-utils/java-functions which is imported by application launcher scripts written in shell. These functions may be used to help loading system-wide or application-specific configuration, constructing application classpath, determining proper JVM/JDK to use with the application, and finally invoking JVM with appropriate flags, including ABRT agent flags, and finally invoking JVM.

(There are also other minor, but valid reasons to require javapckages-tools, such as for configuration files it contains.)

Changing auto-requires from javapackages-tools to javapackages-filesystem and rebuilding all packages would break most of launcher scripts. Explicit requires on javapackages-tools would have to be added to packages that have launchers in order to fix them, reverting some of benefits of the change.

[...]

Some data points from a rawhide query from today.

A) Total packages with "Requires: javapackages-tools": 4204 B) Total packages with "Requires: javapackages-tools" that are *not* -javadoc sub-packages: 2991 C) Total packages with "Requires: javapackages-tools", not a -javadoc sub-package and installing something in {,/usr}/bin: 145

Implementing this change would require modifying many Java packages to add explicit requires to spec files. It would not affect users of Java applications, which would still need to require javapackages-tools. It would only possibly affect installations with Java libraries, but no applications, provided that these libraries don't have dependencies on applications.

Given from the data above it would entail to change on the order of 145 (binary) packages[1].

I've also run specific analysis on those 145 packages in order to figure out whether or not they source /usr/share/java-utils/java- functions from javapackages-tools. It appears there are (roughly) 82 SRPM packages which may need changing[2]. 5 of those already use explicit requires for javapackages-tools[3].

Note that this doesn't seem to capture packages like ant. ant itself doesn't depend on javapackages-tools, but it depends on ant-lib which does. Yet, the ant package installs binaries which use functions exported by javapackages-tools, not ant-lib. There is also a case of installed symlinks like for package eclipse-platform which symlinks to ant. While eclipse-platform is a false positive, ant is a false negative.

So to me it seems like a lot of effort for not much benefit. But I'm not opposed to the change, as long as people implementing it will take responsibility for fixing all packages broken by it. Due to nature of the change, affecting many packages across the distribution, I think it should be a system-wide change approved by FESCo.

If you think a system-wide change is needed for this, then so be it. I'll help get packages fixed and get explicit javapackages-tools requirement added for the packages that need it.

Given that there are currently about 80 SRPM to potentially change, it seems a reasonable effort. After all, all other packages will benefit from the change.

Thanks, Severin

[1] See attachment "bin_containing_packages_javapackages_tools.txt" for a list of specific binary RPM packages. [2] See attachment "pkgs_needing_javapackages-tools_srpms_name.txt" [3] console-image-viewer eclipse gradle liquibase will-crash