Recently I have been preparing a reviewer response for my paper about the Generalized Functional Pruning Optimal Partitioning (GFPOP) algorithm. Pre-print is arXiv:1810.00117 and code is in R package PeakSegDisk.

I have addressed all the reviewer comments, but CRAN will not accept a new submission unless I fix all the problems shown on the check page. That is a sensible policy, but it has been extremely difficult for me to figure out how to fix the problem, since I am unable to reproduce the error message on my own computer. Here is what it looked like on CRAN:

> names.list <- PeakSegFPOP_disk(tmp, "10.5")
==3342== Syscall param write(buf) points to uninitialised byte(s)
==3342==    at 0x53FFB15: write (in /usr/lib64/libpthread-2.29.so)
==3342==    by 0x72451E5: ??? (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x7284CC4: std::basic_filebuf<char, std::char_traits<char> >::_M_convert_to_external(char*, long) (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x728514B: std::basic_filebuf<char, std::char_traits<char> >::overflow(int) (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x7284E55: std::basic_filebuf<char, std::char_traits<char> >::_M_terminate_output() (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x7284F9E: std::basic_filebuf<char, std::char_traits<char> >::_M_seek(long, std::_Ios_Seekdir, __mbstate_t) (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x7285455: std::basic_filebuf<char, std::char_traits<char> >::seekoff(long, std::_Ios_Seekdir, std::_Ios_Openmode) (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x72AD1FE: std::ostream::seekp(long, std::_Ios_Seekdir) (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x48C38C8: seek_element (packages/tests-vg/PeakSegDisk/src/PeakSegFPOPLog.cpp:91)
==3342==    by 0x48C38C8: write (packages/tests-vg/PeakSegDisk/src/PeakSegFPOPLog.cpp:131)
==3342==    by 0x48C38C8: PeakSegFPOP_disk(char*, char*) (packages/tests-vg/PeakSegDisk/src/PeakSegFPOPLog.cpp:405)
==3342==    by 0x48C8507: PeakSegFPOP_interface(char**, char**) (packages/tests-vg/PeakSegDisk/src/interface.cpp:11)
==3342==    by 0x498FFD: do_dotCode (svn/R-devel/src/main/dotcode.c:1746)
==3342==    by 0x4CCC8F: bcEval (svn/R-devel/src/main/eval.c:6775)
==3342==  Address 0x100919a8 is 56 bytes inside a block of size 8,192 alloc'd
==3342==    at 0x4839593: operator new[](unsigned long) (/builddir/build/BUILD/valgrind-3.15.0/coregrind/m_replacemalloc/vg_replace_malloc.c:433)
==3342==    by 0x7284B23: std::basic_filebuf<char, std::char_traits<char> >::_M_allocate_internal_buffer() (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x7288E36: std::basic_filebuf<char, std::char_traits<char> >::open(char const*, std::_Ios_Openmode) (in /usr/lib64/libstdc++.so.6.0.26)
==3342==    by 0x48C31CC: open (/usr/include/c++/9/fstream:1180)
==3342==    by 0x48C31CC: init (packages/tests-vg/PeakSegDisk/src/PeakSegFPOPLog.cpp:73)
==3342==    by 0x48C31CC: PeakSegFPOP_disk(char*, char*) (packages/tests-vg/PeakSegDisk/src/PeakSegFPOPLog.cpp:263)
==3342==    by 0x48C8507: PeakSegFPOP_interface(char**, char**) (packages/tests-vg/PeakSegDisk/src/interface.cpp:11)
==3342==    by 0x498FFD: do_dotCode (svn/R-devel/src/main/dotcode.c:1746)
==3342==    by 0x4CCC8F: bcEval (svn/R-devel/src/main/eval.c:6775)
==3342==    by 0x4E4DFF: Rf_eval (svn/R-devel/src/main/eval.c:620)
==3342==    by 0x4E694E: R_execClosure (svn/R-devel/src/main/eval.c:1780)
==3342==    by 0x4E7694: Rf_applyClosure (svn/R-devel/src/main/eval.c:1706)
==3342==    by 0x4E4EF4: Rf_eval (svn/R-devel/src/main/eval.c:743)
==3342==    by 0x4E908E: do_set (svn/R-devel/src/main/eval.c:2808)
==3342==  Uninitialised value was created by a heap allocation
==3342==    at 0x4838E86: operator new(unsigned long) (/builddir/build/BUILD/valgrind-3.15.0/coregrind/m_replacemalloc/vg_replace_malloc.c:344)
==3342==    by 0x48C3431: allocate (/usr/include/c++/9/ext/new_allocator.h:114)
==3342==    by 0x48C3431: allocate (/usr/include/c++/9/bits/alloc_traits.h:444)
==3342==    by 0x48C3431: _M_get_node (/usr/include/c++/9/bits/stl_list.h:438)
==3342==    by 0x48C3431: _M_create_node<double, int, double, double&, double&, int, bool> (/usr/include/c++/9/bits/stl_list.h:630)
==3342==    by 0x48C3431: _M_insert<double, int, double, double&, double&, int, bool> (/usr/include/c++/9/bits/stl_list.h:1907)
==3342==    by 0x48C3431: emplace_back<double, int, double, double&, double&, int, bool> (/usr/include/c++/9/bits/stl_list.h:1223)
==3342==    by 0x48C3431: PeakSegFPOP_disk(char*, char*) (packages/tests-vg/PeakSegDisk/src/PeakSegFPOPLog.cpp:284)
==3342==    by 0x48C8507: PeakSegFPOP_interface(char**, char**) (packages/tests-vg/PeakSegDisk/src/interface.cpp:11)
==3342==    by 0x498FFD: do_dotCode (svn/R-devel/src/main/dotcode.c:1746)
==3342==    by 0x4CCC8F: bcEval (svn/R-devel/src/main/eval.c:6775)
==3342==    by 0x4E4DFF: Rf_eval (svn/R-devel/src/main/eval.c:620)
==3342==    by 0x4E694E: R_execClosure (svn/R-devel/src/main/eval.c:1780)
==3342==    by 0x4E7694: Rf_applyClosure (svn/R-devel/src/main/eval.c:1706)
==3342==    by 0x4E4EF4: Rf_eval (svn/R-devel/src/main/eval.c:743)
==3342==    by 0x4E908E: do_set (svn/R-devel/src/main/eval.c:2808)
==3342==    by 0x4E5162: Rf_eval (svn/R-devel/src/main/eval.c:695)
==3342==    by 0x5137FC: Rf_ReplIteration (svn/R-devel/src/main/main.c:260)
==3342==    by 0x5137FC: Rf_ReplIteration (svn/R-devel/src/main/main.c:200)
==3342== 
> 

I talked to R-core member Tomas Kalibera at useR 2019 in Toulouse, and he recommended I try reproducing the error using R-hub. Luckily I was able to reproduce the valgrind error using rhub::check_with_valgrind, which uses the rhub/debian-gcc-release docker image.

My old mac does not support the new Docker Desktop, so I had to install Docker Toolbox.

At first I had some problems because Docker was unable to access files/folders on my Mac. The key thing to know about this setup is that Docker will actually be run inside a Linux Virtualbox on the mac. So we have to make sure the Virtualbox can access the files/folders.

The Virtualbox Guest additions docs says that shared folders can be mounted on Linux guests using

mount -t vboxsf [-o OPTIONS] sharename mountpoint

so I opened up the Virtualbox linux command line and ran

sudo mount -t vboxsf Users /Users

which worked. Note that Users is a keyword/name for the shared folder, that can be set by going into Virtualbox settings. I start a new docker container using

docker run \
 --name test \
 -d \
 -it \
 -v /Users/maudelaperriere/toby:/home/toby \
 rhub/debian-gcc-release

Note that -d is for detached, -it is for interactive, and -v is for volume mounting. Since it starts detached, you can get a commandline inside using

docker attach test

Inside the docker image you can suspend it using C-P C-Q.

Then I install the CRAN version of my package and its deps using

docker exec test R -e 'install.packages("PeakSegDisk")'

Then I test to make sure I get the same error as on CRAN:

docker exec test R -d 'valgrind --track-origins=yes' \\
 -e 'library(PeakSegDisk);example(PeakSegFPOP_disk)'

I did. I then created this valgrind.sh script in my R package directory:

#!/bin/bash
set -o errexit
R CMD INSTALL .
R -d 'valgrind --track-origins=yes' -e 'library(PeakSegDisk);example(PeakSegFPOP_file)'

Then I run it from the mac command line via

docker exec -w /home/toby/R/PeakSegDisk test bash valgrind.sh

It reproduces the valgrind output I saw on CRAN:

PSFPOP> PeakSegFPOP_file(tmp, pstr)
==279== Syscall param write(buf) points to uninitialised byte(s)
==279==    at 0x4D13471: write (write.c:26)
==279==    by 0x7849D65: ??? (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0x7882AC1: std::basic_filebuf<char, std::char_traits<char> >::_M_convert_to_external(char*, long) (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0x7882E83: std::basic_filebuf<char, std::char_traits<char> >::overflow(int) (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0x7882C0D: std::basic_filebuf<char, std::char_traits<char> >::_M_terminate_output() (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0x7882D0A: std::basic_filebuf<char, std::char_traits<char> >::_M_seek(long, std::_Ios_Seekdir, __mbstate_t) (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0x7883164: std::basic_filebuf<char, std::char_traits<char> >::seekoff(long, std::_Ios_Seekdir, std::_Ios_Openmode) (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0x78A3F55: std::ostream::seekp(long, std::_Ios_Seekdir) (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0xCC649DF: seek_element (PeakSegFPOPLog.cpp:86)
==279==    by 0xCC649DF: write (PeakSegFPOPLog.cpp:126)
==279==    by 0xCC649DF: PeakSegFPOP_disk(char*, char*) (PeakSegFPOPLog.cpp:379)
==279==    by 0xCC69586: PeakSegFPOP_interface(char**, char**) (interface.cpp:14)
==279==    by 0x4937C9B: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x496C0C5: ??? (in /usr/lib/R/lib/libR.so)
==279==  Address 0x9248f58 is 56 bytes inside a block of size 8,192 alloc'd
==279==    at 0x483650F: operator new[](unsigned long) (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==279==    by 0x7882987: std::basic_filebuf<char, std::char_traits<char> >::_M_allocate_internal_buffer() (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0x7886821: std::basic_filebuf<char, std::char_traits<char> >::open(char const*, std::_Ios_Openmode) (in /usr/lib/x86_64-linux-gnu/libstdc++.so.6.0.25)
==279==    by 0xCC644AA: open (fstream:1076)
==279==    by 0xCC644AA: init (PeakSegFPOPLog.cpp:71)
==279==    by 0xCC644AA: PeakSegFPOP_disk(char*, char*) (PeakSegFPOPLog.cpp:236)
==279==    by 0xCC69586: PeakSegFPOP_interface(char**, char**) (interface.cpp:14)
==279==    by 0x4937C9B: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x496C0C5: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x497C16F: Rf_eval (in /usr/lib/R/lib/libR.so)
==279==    by 0x497DD9E: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x497C326: Rf_eval (in /usr/lib/R/lib/libR.so)
==279==    by 0x4981F69: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x496C0C5: ??? (in /usr/lib/R/lib/libR.so)
==279==  Uninitialised value was created by a heap allocation
==279==    at 0x4835DEF: operator new(unsigned long) (in /usr/lib/x86_64-linux-gnu/valgrind/vgpreload_memcheck-amd64-linux.so)
==279==    by 0xCC64718: allocate (new_allocator.h:111)
==279==    by 0xCC64718: allocate (alloc_traits.h:436)
==279==    by 0xCC64718: _M_get_node (stl_list.h:450)
==279==    by 0xCC64718: _M_create_node<double, double, double, double&, double&, int, double> (stl_list.h:642)
==279==    by 0xCC64718: _M_insert<double, double, double, double&, double&, int, double> (stl_list.h:1903)
==279==    by 0xCC64718: emplace_back<double, double, double, double&, double&, int, double> (stl_list.h:1235)
==279==    by 0xCC64718: PeakSegFPOP_disk(char*, char*) (PeakSegFPOPLog.cpp:256)
==279==    by 0xCC69586: PeakSegFPOP_interface(char**, char**) (interface.cpp:14)
==279==    by 0x4937C9B: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x496C0C5: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x497C16F: Rf_eval (in /usr/lib/R/lib/libR.so)
==279==    by 0x497DD9E: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x497C326: Rf_eval (in /usr/lib/R/lib/libR.so)
==279==    by 0x4981F69: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x496C0C5: ??? (in /usr/lib/R/lib/libR.so)
==279==    by 0x497C16F: Rf_eval (in /usr/lib/R/lib/libR.so)
==279==    by 0x497DD9E: ??? (in /usr/lib/R/lib/libR.so)
==279== 

So after a bit of interactive debugging I found the fix, which is explained on this SO post. Essentially when we are dealing with structs in C or classes in C++, we need to initialize the whole memory block in order to avoid errors in valgrind, i.e. need to use memset in the constructor below

PoissonLossPieceLog::PoissonLossPieceLog
(double li, double lo, double co, double m, double M, int i, double prev){
  memset(this, 0, sizeof(PoissonLossPieceLog));
  Rprintf("double=%d int=%d this=%d\n", 
    sizeof(double), sizeof(int), sizeof(PoissonLossPieceLog));
  Linear = li;
  Log = lo;
  Constant = co;
  min_log_mean = m;
  max_log_mean = M;
  data_i = i;
  prev_log_mean = prev;
}

This is pretty strange but the idea is that sizeof(x) is not always equal to sizeof(x.y) + sizeof(x.z), e.g. the code above produces the following output:

double=8 int=4 this=56

which means that there are 52 bytes which are used to store data and 4 bytes which are not (padding). If we omit the call to memset above, then the constructor leaves 4 bytes un-initialized. Normally that is not an issue with valgrind because we refer to all members by name. But since in this code I am serializing instances of this class to disk, the error comes from

memcpy(p, &(*it), sizeof(PoissonLossPieceLog));

which copies all bytes associated with the struct (including the four un-initialized bytes).

Therefore another fix would be to edit the serialization code, to explicitly mention all members of the class that should be serialized. On the one hand that would make the code more complicated (the fix above is only one line) but on the other hand it would make it more efficient.

UPDATE: I ran the sizeof code on my Ubuntu laptop, on which I could not reproduce the error, and I got the following output:

double=8 int=4 this=52

So on that system the compiler does not allocate any padding bytes in the struct, which means that all 52 bytes are initialized, and so it never writes any un-initialized data, so it makes sense that there was no valgrind error on that system.