Infinite loop in gcov_exit

[Reetesh Ranjan]

Release version: 5-2016-q3-update
Using binary package
Host: Mac OSX, Windows

All other mandatory details are available in the description below:

I was trying to do code coverage for my embedded project using 5-2016-q3-update and GNU ARM Eclipse (http://gnuarmeclipse.github.io/). I broadly followed the concept of manually breaking into 'gcov_exit' and using 'gdb dump binary' to create the gcda files, as described in https://mcuoneclipse.com/2014/12/26/code-coverage-for-embedded-target-with-eclipse-gcc-and-gcov/. However; gcov_exit hanged infinitely. For a simple project of not more than 25 source files, it would take minutes and won't return. To debug this, I made a copy of respective gcc and libgcc code and tailored it as needed and plugged it into my project so that I could see what the hang is about. It turned out that the 1st instance of these loops "for (gi_ptr = list; gi_ptr; gi_ptr = gi_ptr->next)" just won't finish. On debugging it appeared that the list was circular and not a null-terminated one. I cannot figure it out why that is as I studied gcc code, and __gcov_root.list construction through __gcov_init does not seem to create any circular lists. But it appeared to be so, because when I changed the loops to expect a circular list rather than a null-terminated singly linked list, things worked. Also, I manually debugged the functions called by gcov_exit to see that there are exactly as many gcov_info nodes in the __gcov_root.list as the number of source files covered, and the change to expect a circular list made sense. I had to end up changing 3 instances of such loops and then I was able to generated the gcda files.

The ported/tailored version of the minimum gcc/libgcc code has been made available at: https://github.com/reeteshranjan/libgcov-embedded

I saw that the toolchain is based on svn://gcc.gnu.org/svn/gcc/branches/ARM/; and I picked the code from gcc trunk. However; it seems at least from embedded-6-branch onwards that the gcov source is the same as I could see in gcc trunk that time.

I have also attached a patch I made using the embedded-6-branch which clearly shows the for loop changes made.

I also did a sample GNU ARM Eclipse project to use the above, which is available at:
https://github.com/reeteshranjan/libgcov-embedded-example

Steps to reproduce and analyze:
- Setup any embedded project using GNU ARM Eclipse. Follow the steps in https://mcuoneclipse.com/2014/12/26/code-coverage-for-embedded-target-with-eclipse-gcc-and-gcov/ to generate coverage data by breaking into gcov_exit or __gcov_flush (whichever is available). The hang in gcov_exit should be reproducible.
- Now create another project from scratch without using the stub code as described in https://mcuoneclipse.com/2014/12/26/code-coverage-for-embedded-target-with-eclipse-gcc-and-gcov/
- Plugin https://github.com/reeteshranjan/libgcov-embedded using the instructions in README.md provided in this repository into the new project.
- In the port/libgcc/libgcov-driver.c file in the above, search for the loops "for (gi_ptr = list; gi_ptr && !(gi_ptr == list && break_loop_counter); gi_ptr = gi_ptr->next, break_loop_counter++)". These are the changes to expect a circular list and not a null-terminated singly linked list. Change these lines to how they are in original gcc code which is "for (gi_ptr = list; gi_ptr; gi_ptr = gi_ptr->next)"
- Follow the README instructions on how to break in gcov-io.c to use gdb instructions to manually create gcda files. You won't hit this breakpoint because of the hang in first of these for loops in 'compute_summary' function (now reproducible as we changed the code to use the original libgcc version)
- Now revert the for loop changes back
- Now the breakpoints should hit once for each source file

[Tejas Belagod (belagod-tejas)]

Hi Reetesh,

Thanks for suggesting the patch.

I suggest you post a patch against trunk to <email address hidden> and get it reviewed by GCOV maintainers upstream.

Thanks,
Tejas.

[Reetesh Ranjan (reeteshranjan)]

Hi Tejas,

This is my first ever gcc contrib. I am not sure what you expect me to send to the patches email? Should I send the whole description to them again?

Thanks,
Reetesh

> On 31-May-2017, at 6:29 PM, Tejas Belagod <email address hidden> wrote:
>
> Hi Reetesh,
>
> Thanks for suggesting the patch.
>
> I suggest you post a patch against trunk to <email address hidden> and
> get it reviewed by GCOV maintainers upstream.
>
> Thanks,
> Tejas.
>
> --
> You received this bug notification because you are subscribed to the bug
> report.
> https://bugs.launchpad.net/bugs/1694644
>
> Title:
> Infinite loop in gcov_exit
>
> Status in GNU ARM Embedded Toolchain:
> New
>
> Bug description:
> Release version: 5-2016-q3-update
> Using binary package
> Host: Mac OSX, Windows
>
> All other mandatory details are available in the description below:
>
> I was trying to do code coverage for my embedded project using
> 5-2016-q3-update and GNU ARM Eclipse
> (http://gnuarmeclipse.github.io/). I broadly followed the concept of
> manually breaking into 'gcov_exit' and using 'gdb dump binary' to
> create the gcda files, as described in
> https://mcuoneclipse.com/2014/12/26/code-coverage-for-embedded-target-
> with-eclipse-gcc-and-gcov/. However; gcov_exit hanged infinitely. For
> a simple project of not more than 25 source files, it would take
> minutes and won't return. To debug this, I made a copy of respective
> gcc and libgcc code and tailored it as needed and plugged it into my
> project so that I could see what the hang is about. It turned out that
> the 1st instance of these loops "for (gi_ptr = list; gi_ptr; gi_ptr =
> gi_ptr->next)" just won't finish. On debugging it appeared that the
> list was circular and not a null-terminated one. I cannot figure it
> out why that is as I studied gcc code, and __gcov_root.list
> construction through __gcov_init does not seem to create any circular
> lists. But it appeared to be so, because when I changed the loops to
> expect a circular list rather than a null-terminated singly linked
> list, things worked. Also, I manually debugged the functions called by
> gcov_exit to see that there are exactly as many gcov_info nodes in the
> __gcov_root.list as the number of source files covered, and the change
> to expect a circular list made sense. I had to end up changing 3
> instances of such loops and then I was able to generated the gcda
> files.
>
> The ported/tailored version of the minimum gcc/libgcc code has been
> made available at: https://github.com/reeteshranjan/libgcov-embedded
>
> I saw that the toolchain is based on
> svn://gcc.gnu.org/svn/gcc/branches/ARM/; and I picked the code from
> gcc trunk. However; it seems at least from embedded-6-branch onwards
> that the gcov source is the same as I could see in gcc trunk that
> time.
>
> I have also attached a patch I made using the embedded-6-branch which
> clearly shows the for loop changes made.
>
> I also did a sample GNU ARM Eclipse project to use the above, which is available at:
> https://github.com/reeteshranjan/libgcov-embedded-example
>
> Steps to reproduce and analyze:
> - Setup any embedded p...

[Tejas Belagod (belagod-tejas)]

Hi Reetesh,

Really appreciate your keennes to contribute.

Here is a detailed guide on how to contribute to gcc - https://gcc.gnu.org/contribute.html.

Thanks,
Tejas.

[Petr Belskiy (petr.belskiy)]

I have the same issue in my internal project.

[eDeviser (wolle3)]

Hi Tejas,

I read your comment at https://mcuoneclipse.com/2014/12/26/code-coverage-for-embedded-target-with-eclipse-gcc-and-gcov/ and your description at https://github.com/reeteshranjan/libgcov-embedded .

It seems, that it is the same problem for me. Did you had any success in posting a patch against trunk?

[Tejas Belagod (belagod-tejas)]

Hi eDeviser,

I think it was Reetesh Ranjan's patch (see above) - I don't know if he managed to get it submitted upstream to the gcc lists.

Thanks,
Tejas.

[Reetesh Ranjan (reeteshranjan)]

Hi,

I talked to the gcc team. The summary of the discussion is:

My patch works around the core issue of circular lists being created. It's not a fix in the sense it does not prevent the circular lists from getting formed. I was told to find the root cause of what created the circular lists and provide a patch for that.

Sadly I have not had time to spend on this new investigation.

Meanwhile, using the whole mechanism I have provided in GitHub is something we can use, if that's any help.

Thanks
Reetesh

> On 11-Sep-2017, at 6:54 PM, Tejas Belagod <email address hidden> wrote:
>
> Hi eDeviser,
>
> I think it was Reetesh Ranjan's patch (see above) - I don't know if he
> managed to get it submitted upstream to the gcc lists.
>
> Thanks,
> Tejas.
>
> --
> You received this bug notification because you are subscribed to the bug
> report.
> https://bugs.launchpad.net/bugs/1694644
>
> Title:
> Infinite loop in gcov_exit
>
> Status in GNU ARM Embedded Toolchain:
> New
>
> Bug description:
> Release version: 5-2016-q3-update
> Using binary package
> Host: Mac OSX, Windows
>
> All other mandatory details are available in the description below:
>
> I was trying to do code coverage for my embedded project using
> 5-2016-q3-update and GNU ARM Eclipse
> (http://gnuarmeclipse.github.io/). I broadly followed the concept of
> manually breaking into 'gcov_exit' and using 'gdb dump binary' to
> create the gcda files, as described in
> https://mcuoneclipse.com/2014/12/26/code-coverage-for-embedded-target-
> with-eclipse-gcc-and-gcov/. However; gcov_exit hanged infinitely. For
> a simple project of not more than 25 source files, it would take
> minutes and won't return. To debug this, I made a copy of respective
> gcc and libgcc code and tailored it as needed and plugged it into my
> project so that I could see what the hang is about. It turned out that
> the 1st instance of these loops "for (gi_ptr = list; gi_ptr; gi_ptr =
> gi_ptr->next)" just won't finish. On debugging it appeared that the
> list was circular and not a null-terminated one. I cannot figure it
> out why that is as I studied gcc code, and __gcov_root.list
> construction through __gcov_init does not seem to create any circular
> lists. But it appeared to be so, because when I changed the loops to
> expect a circular list rather than a null-terminated singly linked
> list, things worked. Also, I manually debugged the functions called by
> gcov_exit to see that there are exactly as many gcov_info nodes in the
> __gcov_root.list as the number of source files covered, and the change
> to expect a circular list made sense. I had to end up changing 3
> instances of such loops and then I was able to generated the gcda
> files.
>
> The ported/tailored version of the minimum gcc/libgcc code has been
> made available at: https://github.com/reeteshranjan/libgcov-embedded
>
> I saw that the toolchain is based on
> svn://gcc.gnu.org/svn/gcc/branches/ARM/; and I picked the code from
> gcc trunk. However; it seems at least from embedded-6-branch onwards
> that the gcov source is the same as I could see in gcc trunk that
> time.
>
> I have also attached a p...