sv_upgrade() loses 64-bit alignment, causing SIGBUS on sparc #9567
Comments
From @ntyniThis is a bug report for perl from Niko Tyni <ntyni@debian.org> As seen in <http://bugs.debian.org/505415>, calling sv_chop() or otherwise This was uncovered by the JSON-XS-2.23 test suite failing for us on On 5.10.0, calling sv_chop() will upgrade the SV into an SVt_PVIV due Calling the SvGROW() macro afterwards will cast sv->sv_any to an XPV I'm attaching a test program that mimics what JSON-XS does when compiled This particular failure mode has been fixed in bleadperl with change However, it seems to me that this is a more general problem that should The same test program, when compiled with -O2 but without -DUSE_CHOP Note that using a debugging perl build (-DDEBUGGING) seems to defeat the crash. Details for the bleadperl SIGBUS: Program received signal SIGBUS, Bus error. The preprocessor output for the above SvGROW() invocation is (((XPV*) (sv)->sv_any)->xpv_len < (((XPV*) (sv)->sv_any)->xpv_cur + 1) ? Perl_sv_grow(my_perl, sv,((XPV*) (sv)->sv_any)->xpv_cur + 1) : ((sv)->sv_u.svu_pv)); The instruction that causes the bus error is a double-word load (ldd) 0x00010708 <doit+56>: ld [ %i1 ], %g1 Quoting Julien Cristau in the Debian bug:
Flags: Site configuration information for perl 5.11.0: Configured by niko at Sat Nov 15 18:40:57 UTC 2008. Summary of my perl5 (revision 5 version 11 subversion 0) configuration: Locally applied patches: @INC for perl 5.11.0: Environment for perl 5.11.0: |
From @ntyni#include <EXTERN.h>
#include <perl.h>
static PerlInterpreter *my_perl;
SV *sv;
void doit(void) {
sv = newSVpv("foo", 3);
#ifdef USE_CHOP
sv_chop(sv, SvPV_nolen(sv) + 1);
#else
SvUPGRADE(sv, SVt_PVIV);
#endif
SvGROW (sv, SvCUR (sv) + 1);
}
int main(int argc, char **argv, char **env)
{
int i;
my_perl = perl_alloc();
for (i=0; i<10; i++) {
printf("try %d\n", i);
doit();
}
}
|
From @nwc10On Sun, Nov 16, 2008 at 04:16:14AM -0800, Niko Tyni wrote:
Yes, it's a more general problem.
What is __alignof__(XPVIV) and sizeof(XPVIV) on the same system?
I see that -fno-strict-aliasing is on. That might prove relevant in my reply Nicholas Clark |
|
The RT System itself - Status changed from 'new' to 'open' |
From @ntyniOn Sun, Nov 16, 2008 at 05:15:36PM +0000, Nicholas Clark wrote:
8 and 24.
I didn't explicitly ask for it, so it would seem to be on by default. |
From @nwc10On Sun, Nov 16, 2008 at 08:17:17PM +0200, Niko Tyni wrote:
whereas __alignof__(xpv_allocated) and sizeof(xpv_allocated) are 8 and 4, If so, that would explain a lot, and it's my fault.
It is Perl's default for gcc. I think it may not be relevant now. Nicholas Clark |
From @ntyniOn Mon, Nov 17, 2008 at 03:44:18PM +0000, Nicholas Clark wrote:
The other way around, otherwise right. XPV: alignof: 8 sizeof: 16
Looking forward to the explanation :) Thanks for your work, |
From au@hcsd.deHi, I just ran into the same bug with Perl 5.8.8 compiled with Sun's Regards, Stephan Summary of my perl5 (revision 5 version 8 subversion 8) configuration: Characteristics of this binary (from libperl): |
From [Unknown Contact. See original ticket]Hi, I just ran into the same bug with Perl 5.8.8 compiled with Sun's Regards, Stephan Summary of my perl5 (revision 5 version 8 subversion 8) configuration: Characteristics of this binary (from libperl): |
From au@hcsd.deJust to provide the core file details... # dbx -f /opt/mssperl5/bin/perl core |
From [Unknown Contact. See original ticket]Just to provide the core file details... # dbx -f /opt/mssperl5/bin/perl core |
From au@hcsd.deQuoting myself...
I have fixed the SIGBUS by upgrading to Perl 5.8.9 and using a less Regards, Stephan |
From @jimcXPV and XPVIV are both special bodies, they use the ghost-fields trick, I dont have a 64 bit machine, so I can only speculate, /* 8 bytes on most ILP32 with IEEE doubles */ /* 12 */ |
From @nwc10On Mon, Nov 17, 2008 at 08:52:28PM +0200, Niko Tyni wrote:
Sorry for the delay. Does change 69ba284 I don't have access to sparc Debian, but I do have limited acces to sparc bash-2.05$ ./perl -Ilib align.pl >align.c && gcc -Wall -o align align.c && ./align XPVIV 24 8 XPVAV 32 8 XPVHV 32 8 XPVCV 64 8 XPVFM 64 8 XPVIO 88 8 regexp 96 8 bash-2.05$ sizeof() ^ ^ __alignof__() So I think the new way preserves the alignment constraint. If I've got this right, I'll explain further if necessary. I'm not committing Nicholas Clark #!perl -w print <<'EOH'; int main() { my %first = ( my @types = qw(XPV XPVIV XPVAV XPVHV XPVCV XPVFM); foreach my $raw (@types) { print <<'EOH'; __END__ |
From @ntyniOn Fri, Jul 17, 2009 at 08:44:40PM +0100, Nicholas Clark wrote:
Yes, I can confirm it fixes it all for me on both blead and maint-5.10,
This is the output on maint-5.10 here: XPV 16 8 XPVIV 24 8 XPVAV 32 8 XPVHV 32 8 XPVCV 64 8 XPVFM 72 8 Many thanks for your work! Much appreciated. |
From @nwc10Resolved and released. I want to close the RT ticket now, but I still Nicholas Clark |
|
@nwc10 - Status changed from 'open' to 'resolved' |
Migrated from rt.perl.org#60574 (status was 'resolved')
Searchable as RT60574$
The text was updated successfully, but these errors were encountered: