navtree js output varies from ASLR

[bmwiedemann]

Describe the bug
doxygen produces nondeterministic navtreedata/index.js output from certain input (see attached libzypp-min.tar.gz) unless ASLR is disabled.

Expected behavior
Output shoud be deterministic using the same doxygen binary, input and environment.

To Reproduce
in Linux with doxygen-1.9.1, do

tar xf libzypp-min.tar.gz
cd libzypp-min
for i in $(seq 1 10) ; do ./build.sh ; done|sort|uniq -c
      4 2b83f77db2e4c27380ed01f2dafd8295  -
      6 cefbad1d7e1d909b1002eea8870544b2  -

Version
doxygen-1.9.1 in openSUSE-Tumbleweed 20210525 x86_64

Additional context
similar to #8158 and #7474

[albert-github]

The reproducibility issues are very inconvenient. I haven't looked at this problem specific, but recently we fixed an issue in respect so problem in the sorting in the master version.

• could you try the master version to see whether the master version still exhibits the problem?

[albert-github]

I just tried the 1.9.1 on Cygwin and here the problem does not occur. I get the always ca101aafea79a3953c3da5fa911f335f (this is different from yours but can be due to paths).

• The example is nicely small, so can you save the different html directories as well and publish them here as well?

I see that ASLR means: Address space layout randomization. I don't know whether or not Cygwin is using this (I don't think so), is there a way that I can enable it on (any) system?
Problem is probably a non initialized variable or pointer, that sometimes hit a "strange" value.

[bmwiedemann]

https://cygwin.com/pipermail/cygwin/2020-March/244107.html suggests to compile with g++ -Xlinker --dynamicbase options. Also it needs Vista or newer.

To experiment, you can also try the samples at https://github.com/bmwiedemann/theunreproduciblepackage/tree/master/aslr

[bmwiedemann]

With master I only got order variations in the .doxytag file. I can try to trace down that one, too.
Can you link me the sorting issue fix?

[albert-github]

I think the fix is in doxygen.cpp (bottom of function readDir):

  if (resultList)
  {
    // sort the resulting list to make the order platform independent.
    std::sort(resultList->begin(),
              resultList->end(),
              [](const auto &f1,const auto &f2) { return qstricmp(f1.c_str(),f2.c_str())<0; });
  }

as part of commit e965e5d for the issue #8560 (see comments #8560 (comment) and #8560 (comment)).

[albert-github]

Could you publish the difference in the .doxytag files / the complete files that are different so that we can have a quick look and maybe give some pointers?

[albert-github]

@bmwiedemann

• any progress?
• what are the differences in the .doxytag files / where can we find them?

[bmwiedemann]

https://rb.zq1.de/temp/libzypp-17.25.10-build.tar.xz (21MB) has the two versions as
a/libzypp-17.25.10/build/doc/autodoc/libzypp.doxytag
b/libzypp-17.25.10/build/doc/autodoc/libzypp.doxytag

along with the sources.
disorderfs might help to debug such issue from filesystem-readdir-order.

[albert-github]

Thanks for the information!
How long lasting is the xz file as it is on a temp directory? I copied the the file already, but not sure whether other developers / maintainers will still have access later on.

Looking at the both directories I see differences in the following files:

• libzypp-17.25.10/build/CMakeCache.txt
  just a change in the name of the used node / machine
• libzypp-17.25.10/build/CMakeFiles/3.20.2/CMakeSystem.cmake:
  signalling a.o other differences are similar

  < set(CMAKE_HOST_SYSTEM "Linux-5.3.18-lp152.50-default")
  ---
  > set(CMAKE_HOST_SYSTEM "Linux-5.3.18-lp152.66-default")
  

• libzypp-17.25.10/build/CMakeFiles/CMakeError.log
  looks like differences is some CMake directory names like CMakeFiles/cmTC_4e123.dir versus CMakeFiles/cmTC_d3530.dir
• libzypp-17.25.10/build/CMakeFiles/CMakeOutput.log
  see CMakeError.log above
• libzypp-17.25.10/build/DartConfiguration.tcl
  just a change in the name of the used node / machine
• libzypp-17.25.10/build/install_manifest.txt
  file has the same information but in a different order!
• libzypp-17.25.10/build/doc/autodoc/libzypp.doxytag: the important files with the differences. the files contain the same information but in a different order (seen the remark at the top, so added here the both tag files again: Uploading doxytag.tar.gz…).
  It looks like to have to do with with the compound kind="file" in case the files names are the same but the paths are different we see these double files names as (from the tag file with the number of occurrences):

      2     <name>PoolImpl.cc</name>
      2     <name>PoolImpl.h</name>
      2     <name>Resolver.cc</name>
      2     <name>Resolver.h</name>
      2     <name>Types.h</name>
      3     <name>Downloader.cc</name>
      3     <name>Downloader.h</name>
  

  I saw the problem occur a.o. with Downloader.h

@doxygen might this be a sorting problem where only the file name is taken into consideration and not the full path?

[bmwiedemann]

Using the full path can have its own downsides when people want to be able to get identical output from building in /home/foo and /home/bar (Debian does ; openSUSE mostly uses constant build path)
Relative paths can be a way out or content-based disambiguation.

[albert-github]

I didn't mean using full paths but using internally for the sorting not only the name but as a second level the path.

[doxygen]

Please verify if the commit I just pushed fixes the unpredictable ordering issue.

[bmwiedemann]

My tests still show differences in the .doxytag file using master (commit 251e18b7). Also that issue is no more about ASLR, but related to filesystem-order. Maybe we should open another issue to make that clearer?

[albert-github]

Commit 251e18b7? I don't see this in the doxygen tree (I searched the complete git log for 251e18b7), so it is unclear, to me, which version you are using.
The doxygen commit to be used is 6fc45ba (or newer)

[bmwiedemann]

Ah, indeed it was 6fc45ba ; was just looking at a file hash in the diff.

[albert-github]

@bmwiedemann

• probably best again the doxytag files ...
• when you have it, might be useful, can you also provide the console output of the erroneous runs?, i.e. the QUIET=NO output)

@doxygen
at first glance maybe in index.cpp, function writeFileIndex here something is done depending on FULL_PATH_NAMES (including sorting)

[bmwiedemann]

doxytags.tar.gz

[albert-github]

Looks like it has to do with:

.../libzypp-17.25.10/zypp/Resolver.h
.../libzypp-17.25.10/zypp/solver/detail/Resolver.h

and

....libzypp-17.25.10/zypp/Resolver.cc
.../libzypp-17.25.10/zypp/solver/detail/Resolver.cc

@bmwiedemann any chance for the console output?