Features/MiniDebugInfo

MiniDebugInfo

Summary

Install minimal debuginfo by default to increase quality of bug reports and allow easier support for profiling and user space tracing.

Owner

• Name: Alexander Larsson
• Email: alexl@redhat.com

Current status

• Targeted release: Fedora 18
• Last updated: 2012-05-07
• Percentage of completion: 0%

Detailed Description

The debuginfo packages in Fedora are very useful if you're a developer and you are debugging a bug on your machine. However, they are very large, so they cannot be deployed in a wider scale. In many usecases you really only need enough information to be able to create backtraces, which is far smaller and could be installed by default.

The main advantage of this is a general improvement of bug report quality (as reported via ABRT or manually), as the reports will have complete backtraces from all users. Additionally things like system wide profiling and userspace tracing would also be possible without large debuginfo installation.

Debug info for backtraces relies on two types of information, the function names in the symbol tables, and (optionally) the linenumber debug information. I wrote a script that takes our current debuginfo packages and generates minimal debuginfo files containing only the minimal function name information and optionally line number information. Here are the sizes it generates for some F17 packages:

Package	Size of executables	symbols	xz symbols	symbols+lines	xz symbols+lines
coreutils	4851 Kb	15.46%	3.92%	38.82%	14.15%
gcc	13003 Kb	3.54%	.87%	23.68%	10.13%
gdb	5062 Kb	3.98%	.85%	18.67%	8.25%
libreoffice-core	142897 Kb	17.96%	2.38%	61.22%	11.76%
libreoffice-writer	11501 Kb	18.52%	2.73%	71.27%	12.42%
cairo	808 Kb	11.95%	2.46%	33.56%	13.08%
gimp	14828 Kb	8.89%	1.89%	29.43%	7.90%
gtk2	6007 Kb	11.46%	2.22%	28.20%	8.72%
gtk3	5302 Kb	8.73%	1.64%	26.08%	8.37%
webkitgtk	31147 Kb	18.51%	2.68%	98.03%	13.75%
webkitgtk3	31143 Kb	18.50%	2.68%	97.97%	13.76%
nautilus	2039 Kb	14.59%	2.89%	33.09%	10.20%
evolution	7492 Kb	8.16%	1.82%	25.95%	7.96%
xulrunner	25533 Kb	43.78%	6.97%	108.84%	22.11%
qt3	10458 Kb	.89%	.21%	21.58%	7.18%
qt	12377 Kb	14.65%	2.17%	58.66%	12.34%
qtwebkit	25989 Kb	19.03%	2.83%	92.24%	13.29%
kdelibs	36231 Kb	10.20%	1.84%	36.91%	9.35%
kde-baseapps	8030 Kb	11.48%	2.46%	32.44%	9.22%
samba	17508 Kb	3.25%	.63%	16.90%	5.31%
httpd	2496 Kb	11.83%	2.97%	25.53%	9.94%
llvm-libs	23176 Kb	2.78%	.42%	33.83%	9.41%
emacs	12445 Kb	1.17%	.30%	4.98%	2.29%
Total	450334 Kb	15.21%	2.32%	58.24%	11.42%

On my F17 installation /usr is 8771 Mb, and 1861 Mb of that is made up of ELF binaries. Leading to the following estimations for full-distro sizes:

	symbols	xz symbols	symbols+lines	xz symbols+lines
Estimated minidebuginfo size	283Mb	43Mb	1084Mb	213Mb
Estimated percentage of /usr	3.2%	0.5%	12.3%	2.4%

Compared to the current size of a Fedora install and the large churn of package update an on-disk increase of 43Mb is quite modest, so we should be able to at least ship the compressed minimal symbol table information by default.

There are several choices to be made here:

• Compressed or uncompressed

Storing the data compresses will use less diskspace, but may cause
an additional performance and memory hit when using the data.

This also affects how the "normal" debuginfo is handled. If
accessing the minidebuginfo isn't too slow, then that information
could be left out of the .debuginfo packages, making them smaller.

• Keep the data in the original or in a separate package.

If the data is kept in the originating binaries (I have patches that
allows this while using compression) then supporting minidebuginfo is
transparent to users, but non-optional.

If we create new minidebuginfo packages for the minidebuginfo then
the data can be made optional, but will require modifications
to the package handling in order to automatically install and update
the new packages. It also makes the change quite visible to users, as
the new packages will be visible in the UI in a lot of places.
 

• Do we add line number information

The line number information is an additional piece of information
that can be quite useful, but it is also quite a lot larger than
just the symbols.

These affect the implementation chosen and must be decided initially.

Benefit to Fedora

We get better bug reports by default, even when offline or on a bad link, without users having to wait while downloading large amounts of data, or having to send possibly privacy sensitive core information to anyone. Additionally this will work for all packages built on the system (not only the official packages, but also site-specific packages) and will work with or without ABRT (some sites may not want to send bug reports to ABRT but collect them locally).

It will also make it easier to do things like system wide profiling, userspace dynamic probes and causual debugging.

Scope

Generation of the minidebuginfo just requires a few lines in /usr/lib/rpm/find-debuginfo.sh, and possibly some additional rpm macros to create new subpackages.

Consumers of debuginfo (mainly gdb, but also profilers, etc) must be changed to look in additional places for extra debug info.

If we chose to use separate rpms for the minidebuginfo we also needs modifications to the build system, the package mirroring system, the installer, the package installation tools, etc.

This feature requires a mass rebuild to generate the new minidebuginfo for all packages.

How To Test

Running any application under gdb you should be able to get a backtrace with function names for all frames by doing:

1. Run "gdb <program name>" 2. Type "run" 3. Do something 4. Press ctrl-c to get back to gdb 5. Type "bt" and watch the full backtrace

User Experience

The end-user experience should be minimally changed. The new debug information is mainly targeted towards developers.

Dependencies

• gdb
• rpm
• libunwind

If using separate minidebug packages

• yum
• anaconda
• pkg-kit
• Possibly other package handling apps

Contingency Plan

None necessary, revert to previous release behaviour.

Documentation

• No relevant documentation

Release Notes

• Fedora 18 ships with a small subset of debug information by default. Full debug information is still available in the debuginfo packages.

Comments and Discussion

• See Talk:Features/MiniDebugInfo