SIGSEGV/SIGBUS when accessing after end of mmapped file; why it differs with GCC?

[Konstantin Belousov]

On Wed, Feb 13, 2013 at 12:13:58PM -0500, Ryan Stone wrote:
> On Wed, Feb 13, 2013 at 11:18 AM, <natris at centrum.cz> wrote:
> 
> > Hello,
> > I am porting an application which maps files into the memory and works
> > directly with the memory. When doing this, it can happen that when someone
> > resizes the file so that part of the previously mapped region is no longer
> > backed by the file, synchronous signal is sent to the process which needs
> > to be handled.
> >
> > On all other platforms than FreeBSD I have tested (Solaris, Linux, Darwin,
> > HP-UX) the signal in question is SIGBUS. However on FreeBSD, depending on
> > the >>compiler<< used, it is either SIGBUS or SIGSEGV. When I compile the
> > binary with "gcc version 4.2.1 20070831 patched [FreeBSD], amd64", the
> > signal is SIGBUS, when i use "gcc version 4.7.3 20121103 (prerelease)
> > (FreeBSD Ports Collection)", the signal is SIGSEGV. Please note that one of
> > the versions of gcc between 4.2 and 4.7 also caused SIGSEFV to be sent but
> > this did not matter for me as I did not need to use that version; I however
> > need gcc 4.7 to work because of c++11 stuff that my project has recently
> > started to use.
> >
> > Unfortunately registering signal handler on SIGSEGV is very inconvenient
> > for me; I would prefer to somehow switch the behavior to be sane.
> >
> > Please anyone has an idea whether or how this could be achieved? I have
> > tried to find out why, on single machine, just because of different gcc
> > version, kernel sends different signal but I have never worked with fbsd
> > kernel before and so my search did not succeed so far.
> >
> > Machine in question runs amd64 FreeBSD 9.1-RC2, but this has also happened
> > to me with older version of FreeBSD.
> >
> >
> > gcc 4.2 (same happens also with libstdc++ and co. from gcc 4.2):
> > Program received signal SIGBUS, Bus error.
> > (gdb) i shared
> > From                To                  Syms Read   Shared Object Library
> > 0x0000000800f2ef70  0x0000000800f3ee68  Yes (*)     /libexec/ld-elf.so.1
> > 0x000000080114d710  0x0000000801159748  Yes (*)     /lib/libthr.so.3
> > 0x00000008013c2d30  0x000000080142c656  Yes
> > /usr/local/lib/gcc47/libstdc++.so.6
> > 0x00000008016737c0  0x00000008016891b8  Yes (*)     /lib/libm.so.5
> > 0x0000000801893930  0x00000008018a3088  Yes
> > /usr/local/lib/gcc47/libgcc_s.so.1
> > 0x0000000801ad71d0  0x0000000801ba9358  Yes (*)     /lib/libc.so.7
> >
> > gcc 4.7:
> > Program received signal SIGSEGV, Segmentation fault.
> > (gdb) i shared
> > From                To                  Syms Read   Shared Object Library
> > 0x0000000800f5ef70  0x0000000800f6ee68  Yes (*)     /libexec/ld-elf.so.1
> > 0x000000080117d710  0x0000000801189748  Yes (*)     /lib/libthr.so.3
> > 0x00000008013f2d30  0x000000080145c656  Yes
> > /usr/local/lib/gcc47/libstdc++.so.6
> > 0x00000008016a37c0  0x00000008016b91b8  Yes (*)     /lib/libm.so.5
> > 0x00000008018c3930  0x00000008018d3088  Yes
> > /usr/local/lib/gcc47/libgcc_s.so.1
> > 0x0000000801b071d0  0x0000000801bd9358  Yes (*)     /lib/libc.so.7
> >
> >
> > I would be glad for any hint or information.
> > Kind Regards,
> > Ondrej Kolacek
> > _______________________________________________
> > freebsd-hackers at freebsd.org mailing list
> > http://lists.freebsd.org/mailman/listinfo/freebsd-hackers
> > To unsubscribe, send any mail to "freebsd-hackers-unsubscribe at freebsd.org"
> 
> 
> 
> I think that setting sysctl machdep.prot_fault_translation=1 would do what
> you want.

This might be an indication of the issue with the toolchains you use,
in particular, with the linker or csu.  The default selection for the signal
is based on the note osver, linked into the binary from crt1.o. Either
linker omits the note, or wrong crt1 is used.

You did not specified anything about version of the FreeBSD used, nor
the exact compiler invocations. Using the crystal ball, I see the
r244600 for HEAD and r244904 for stable/9, if you use --gc-sections
flags. This is more or less consistent with what you reported, since
gcc from ports uses binutils from ports, which have newer ld with
bugfix already applied.
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 834 bytes
Desc: not available
URL: <http://lists.freebsd.org/pipermail/freebsd-hackers/attachments/20130213/725e7abb/attachment.sig>