TBB UBSAN error reported in client package

[kevinushey]

The gaston CRAN package, which uses RcppParallel, has a number of UBSAN check warnings that seem to be coming from TBB. Copied from https://www.stats.ox.ac.uk/pub/bdr/memtests/clang-UBSAN/gaston-Ex.Rout:

/data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/task.h:721:30: runtime error: member call on address 0x7f4f2ee8bd00 which does not point to an object of type 'tbb::internal::scheduler'
0x7f4f2ee8bd00: note: object is of type 'tbb::internal::custom_scheduler<tbb::internal::IntelSchedulerTraits>'
 00 00 00 00  c8 5b 50 2f 4f 7f 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  80 6f c9 2e
              ^~~~~~~~~~~~~~~~~~~~~~~
              vptr for 'tbb::internal::custom_scheduler<tbb::internal::IntelSchedulerTraits>'
SUMMARY: AddressSanitizer: undefined-behavior /data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/task.h:721:30 in 
/data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/parallel_reduce.h:177:19: runtime error: member call on address 0x7f4f2ee6bb40 which does not point to an object of type 'tbb::task'
0x7f4f2ee6bb40: note: object has invalid vptr
 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00
              ^~~~~~~~~~~~~~~~~~~~~~~
              invalid vptr
SUMMARY: AddressSanitizer: undefined-behavior /data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/parallel_reduce.h:177:19 in 
/data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/task.h:688:9: runtime error: member call on address 0x7f4f2ee6bb40 which does not point to an object of type 'tbb::task'
0x7f4f2ee6bb40: note: object has invalid vptr
 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00
              ^~~~~~~~~~~~~~~~~~~~~~~
              invalid vptr
SUMMARY: AddressSanitizer: undefined-behavior /data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/task.h:688:9 in 
/data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/parallel_reduce.h:178:31: runtime error: member call on address 0x7f4f2ee6bb40 which does not point to an object of type 'tbb::task'
0x7f4f2ee6bb40: note: object has invalid vptr
 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  00 00 00 00
              ^~~~~~~~~~~~~~~~~~~~~~~
              invalid vptr
SUMMARY: AddressSanitizer: undefined-behavior /data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/parallel_reduce.h:178:31 in 
/data/gannet/ripley/R/test-clang/RcppParallel/include/tbb/task.h:946:23: runtime error: member call on address 0x7f4f2ee8bd00 which does not point to an object of type 'tbb::internal::scheduler'
0x7f4f2ee8bd00: note: object is of type 'tbb::internal::custom_scheduler<tbb::internal::IntelSchedulerTraits>'
 00 00 00 00  c8 5b 50 2f 4f 7f 00 00  00 00 00 00 00 00 00 00  00 00 00 00 00 00 00 00  80 6f c9 2e
              ^~~~~~~~~~~~~~~~~~~~~~~
              vptr for 'tbb::internal::custom_scheduler<tbb::internal::IntelSchedulerTraits>'

I'm guessing these issues may be resolved when we update TBB sometime in the future.

[HervePerdry]

Many thanks Kevin for reporting the issue. Note that there are similar messages in the UBSAN test of RcppParallel, when executing test.distance, test.innerproduct, test.sum and test.transform.

Another remark: the same commands with the same arguments are executed several times in gaston examples; the UBSAN warnings are issued only the first time, so there might be side effects/unitialized objects involved.

[jolars]

Is there any update on this?

[eddelbuettel]

You could try using one of the UBSAN Docker containers and see if you can still reproduce it. Some parts of the software stack have changed...

[jolars]

Thank you, Dirk. Unfortunately, I couldn't manage to figure out how to compile with clang using the docker image. The latest check on CRAN is, however, only a few days old and still shows the error. Moreover, almost all of the packages on CRAN that import RcppParallel exhibit the same UBSAN-Clang warnings.

[eddelbuettel]

Maybe someone sufficiently motivated (hey, I am looking your way ;-) ) should file a bug with the TBB software then. IIRC it is from code we just bring in ...

[jolars]

Haha, well... before that I think that it might be wise to update RcppParallel to a newer incarnation of TBB. It is now far behind and if these bugs are related to TBB they may well have been fixed by now (even if I cannot find an appropriate item in the relase notes). You have pointed this out yourself.

[eddelbuettel]

Could you maybe fork RcppParallel and see if you could attempt such an update as a pull request?

Everybody over here is busy too but we'll help... promised!

[jolars]

I am probably much to much of a novice for this, but I'll see what I can do!

[kevinushey]

Looks like this is resolved in the development version of RcppParallel (courtesy of the upgraded version of TBB). We should consider pushing out a new version of RcppParallel in the new year.

[HughParsonage]

We should consider pushing out a new version of RcppParallel in the new year.

Happy New Year!

I import RcppParallel in a package of mine (thank you!) and am planning a package update in the next week or so. However, I believe my package is showing UBSAN warnings due to this issue on CRAN. I'm more than happy to modify my package or my submission date to fit in with RcppParallel, so what would you prefer I do?

1. Wait until you submit your package before updating mine (though I must update this month),
2. Update my package before you submit your update, or
3. Take out the parts of the package that use RcppParallel and submit independently.

Thanks again.

[kevinushey]

We actually have submitted a new version of RcppParallel to CRAN, but have not yet received a response from the CRAN team.

[kbenoit]

Just to say that we (quanteda package developers) are delighted to hear of this, since our recently submitted package is being held in CRAN purgatory for similar UBSAN issues. Thanks!

clang-UBSAN https://www.stats.ox.ac.uk/pub/bdr/memtests/clang-UBSAN/quanteda
gcc-UBSAN https://www.stats.ox.ac.uk/pub/bdr/memtests/gcc-UBSAN/quanteda

[kevinushey]

I'm still waiting for a response from the CRAN team.