RcppCore/RcppParallel

ODR issue for gcc san with CRAN version

Opened this issue · 5 comments

I am getting the following ODR violation; CRAN asked me to fix ODR violations in my package, and this one came up:

==136164==ERROR: AddressSanitizer: odr-violation (0x7efd2f2f1240):
  [1] size=168 '__itt__ittapi_global' ../../src/tbb/tools_api/ittnotify_static.c:231:14
  [2] size=168 '__itt__ittapi_global' ../../src/tbb/tools_api/ittnotify_static.c:231:14
These globals were registered at these points:
  [1]:
    #0 0x7efd3c896b78  (/usr/lib/x86_64-linux-gnu/libasan.so.6+0x37b78)
    #1 0x7efd2f24ab91 in _sub_I_00099_1 (/usr/local/RDsan/lib/R/site-library/RcppParallel/lib/libtbbmalloc.so.2+0xd6b91)
    #2 0x7efd3d244dbd  (/lib64/ld-linux-x86-64.so.2+0x11dbd)

  [2]:
    #0 0x7efd3c896b78  (/usr/lib/x86_64-linux-gnu/libasan.so.6+0x37b78)
    #1 0x7efd2f5dec70 in _sub_I_00099_1 (/usr/local/RDsan/lib/R/site-library/RcppParallel/lib/libtbb.so.2+0x240c70)
    #2 0x7efd3d244dbd  (/lib64/ld-linux-x86-64.so.2+0x11dbd)

This is because both tbb and tbbmalloc define the (identical) object here:

/* static part descriptor which handles. all notification api attributes. */
__itt_global _N_(_ittapi_global) = {
ITT_MAGIC, /* identification info */
ITT_MAJOR, ITT_MINOR, API_VERSION_BUILD, /* version info */
0, /* api_initialized */
0, /* mutex_initialized */
0, /* atomic_counter */
MUTEX_INITIALIZER, /* mutex */
NULL, /* dynamic library handle */
NULL, /* error_handler */
(const char**)&dll_path, /* dll_path_ptr */
(__itt_api_info*)&api_list, /* api_list_ptr */
NULL, /* next __itt_global */
NULL, /* thread_list */
NULL, /* domain_list */
NULL, /* string_list */
__itt_collection_normal /* collection state */
};

I tried playing around to see if there was a simple way to avoid that collision, but I couldn't find anything on first glance.

For some reason, some compiled optimized versions do not see this ODR violation; I think it doesn't (yet) occur with CRAN.

It does; the CRAN team has notified me already. And it's apparent because that same symbol is defined in both libraries:

$ nm libtbb.dylib | grep ittapi_global
00000000000268e0 d ___itt__ittapi_global
$ nm libtbbmalloc.dylib | grep ittapi_global
0000000000015c30 d ___itt__ittapi_global

I'm not sure what the right fix is on TBB's side, though. Should they be sharing a single definition? Should the name of the version in tbbmalloc be different?

I am unsure too. We simply redefined for each binary that we produced. If you follow that approach you would pick one and change any name that refers to it. This could probably by some build script.

Note to self, the solution here is probably to try defining INTEL_ITTNOTIFY_PREFIX differently for tbb + tbbmalloc:

#ifndef INTEL_ITTNOTIFY_PREFIX
# define INTEL_ITTNOTIFY_PREFIX __itt_
#endif /* INTEL_ITTNOTIFY_PREFIX */
#ifndef INTEL_ITTNOTIFY_POSTFIX
# define INTEL_ITTNOTIFY_POSTFIX _ptr_
#endif /* INTEL_ITTNOTIFY_POSTFIX */