impala-reviews mailing list archives

Site index · List index
Message view « Date » · « Thread »
Top « Date » · « Thread »
From "Philip Zeyliger (Code Review)" <ger...@cloudera.org>
Subject [Impala-ASF-CR] IMPALA-6330, IMPALA-5702: Avoid boost's trim() to workaround crash after dynamic linking.
Date Tue, 09 Jan 2018 23:53:21 GMT
Hello Michael Ho, Impala Public Jenkins, 

I'd like you to reexamine a change. Please visit

    http://gerrit.cloudera.org:8080/8888

to look at the new patch set (#2).

Change subject: IMPALA-6330, IMPALA-5702: Avoid boost's trim() to workaround crash after dynamic
linking.
......................................................................

IMPALA-6330, IMPALA-5702: Avoid boost's trim() to workaround crash after dynamic linking.

Replaces boost::algorithm::trim() with std::string methods when parsing
/proc/self/smaps and adds a trivial unit test for MemInfo::ParseSmaps().

I did *not* replace other uses of trim() with equivalents from
be/src/gutil/strings/strip.h at this moment.

The backstory here is that
TestAdmissionControllerStress::test_admission_controller_with_flags
fails occasionally on dynamically linked builds of Impala. I was able
to reproduce the failure reliably (within 3 tries) with the following:

  $ ./buildall.sh -notests -so -noclean
  $ bin/start-impala-cluster.py  --impalad_args="--memory_maintenance_sleep_time_ms=1"
  $ impala-shell.sh --query 'select max(t.c1), avg(t.c2), min(t.c3), avg(c4), avg(c5), avg(c6)
from (select max(tinyint_col) over (order by int_col) c1, avg(tinyint_col) over (order by
smallint_col) c2, min(tinyint_col) over (order by smallint_col desc) c3, rank() over (order
by int_col desc) c4, dense_rank() over (order by bigint_col) c5, first_value(tinyint_col)
over (order by bigint_col desc) c6 from functional.alltypes) t;'

The stack trace looks like:

  (gdb) bt
  #0  0x00007fe230df2428 in __GI_raise (sig=sig@entry=6) at ../sysdeps/unix/sysv/linux/raise.c:54
  #1  0x00007fe230df402a in __GI_abort () at abort.c:89
  #2  0x00007fe23312026d in __gnu_cxx::__verbose_terminate_handler() () at ../../../../gcc-4.9.2/libstdc++-v3/libsupc++/vterminate.cc:95
  #3  0x00007fe2330d8b66 in __cxxabiv1::__terminate(void (*)()) (handler=<optimized out>)
at ../../../../gcc-4.9.2/libstdc++-v3/libsupc++/eh_terminate.cc:47
  #4  0x00007fe2330d8bb1 in std::terminate() () at ../../../../gcc-4.9.2/libstdc++-v3/libsupc++/eh_terminate.cc:57
  #5  0x00007fe2330d8cb8 in __cxxabiv1::__cxa_throw(void*, std::type_info*, void (*)(void*))
(obj=0x8e54080, tinfo=0x7fe233356210 <typeinfo for std::bad_cast>, dest=0x7fe23311ea70
<std::bad_cast::~bad_cast()>) at ../../../../gcc-4.9.2/libstdc++-v3/libsupc++/eh_throw.cc:87
  #6  0x00007fe233110332 in std::__throw_bad_cast() () at ../../../../../gcc-4.9.2/libstdc++-v3/src/c++11/functexcept.cc:63
  #7  0x00007fe2330e8ad7 in std::use_facet<std::ctype<char> >(std::locale const&)
(__loc=...) at /data/jenkins/workspace/verify-impala-toolchain-package-build/label/ec2-package-ubuntu-16-04/toolchain/source/gcc/build-4.9.2/x86_64-unknown-linux-gnu/libstdc++-v3/include/bits/locale_classes.tcc:137
  #8  0x00000000008d2cdf in void boost::algorithm::trim<std::string>(std::string&,
std::locale const&) ()
  #9  0x00007fe2396d5057 in impala::MemInfo::ParseSmaps() () at /home/philip/src/Impala/be/src/util/mem-info.cc:132
  ...

My best theory is that there's a race/bug, wherein the std::locale* static initialization
work is getting somehow 'reset' by the dynamic linker, when more libraries are linked
in as a result of the query. My evidence to support this theory is scant, but
I do notice that LD_DEBUG=all prints the following when the query is executed
(but not right at startup):

  binding file /home/philip/src/Impala/toolchain/gcc-4.9.2/lib64/libstdc++.so.6 [0] to
  /home/philip/src/Impala/toolchain/gflags-2.2.0-p1/lib/libgflags.so.2.2 [0]:
  normal symbol `std::locale::facet::_S_destroy_c_locale(__locale_struct*&)'

Note that there are BSS segments for some of std::locale::facet::* inside
of libgflags.so.

  $nm toolchain/gflags-2.2.0-p1/lib/libgflags.so | c++filt | grep facet | grep ' B '
  00000000002e2d10 B std::locale::facet::_S_c_locale
  00000000002e2d0c B std::locale::facet::_S_once

I'm not the first to run into variants of these issues, though the results
are fairly unhelpful:

  http://www.boost.org/doc/libs/1_58_0/libs/locale/doc/html/faq.html
  https://stackoverflow.com/questions/26990412/c-boost-crashes-while-using-locale
  https://svn.boost.org/trac10/ticket/4671
  http://clang-developers.42468.n3.nabble.com/std-use-facet-lt-std-ctype-lt-char-gt-gt-crashes-on-linux-td4033967.html
  https://unix.stackexchange.com/questions/719/can-we-get-compiler-information-from-an-elf-binary
  https://stackoverflow.com/questions/42376100/linking-with-library-causes-collate-facet-to-be-missing-from-char
  http://lists.llvm.org/pipermail/cfe-dev/2012-July/023289.html
  https://gcc.gnu.org/ml/libstdc++/2014-11/msg00122.html

Change-Id: I8dd807f869a9359d991ba515177fb2298054520e
---
M be/src/util/mem-info.cc
M be/src/util/proc-info-test.cc
2 files changed, 14 insertions(+), 5 deletions(-)


  git pull ssh://gerrit.cloudera.org:29418/Impala-ASF refs/changes/88/8888/2
-- 
To view, visit http://gerrit.cloudera.org:8080/8888
To unsubscribe, visit http://gerrit.cloudera.org:8080/settings

Gerrit-Project: Impala-ASF
Gerrit-Branch: master
Gerrit-MessageType: newpatchset
Gerrit-Change-Id: I8dd807f869a9359d991ba515177fb2298054520e
Gerrit-Change-Number: 8888
Gerrit-PatchSet: 2
Gerrit-Owner: Philip Zeyliger <philip@cloudera.com>
Gerrit-Reviewer: Impala Public Jenkins
Gerrit-Reviewer: Michael Ho <kwho@cloudera.com>
Gerrit-Reviewer: Philip Zeyliger <philip@cloudera.com>

Mime
  • Unnamed multipart/alternative (inline, 8-Bit, 0 bytes)
View raw message