From @steve-m-hay

Testing modperl (svn trunk, not the current CPAN release) with httpd-2.4.10 and recent 5.21.x perls I find that things look ok up to and including 5.21.6, but the mod_perl-enabled server fails to start with 5.21.7 and 5.21.8.

This function in modperl_util.c is where things go wrong​:

U16 *modperl_code_attrs(pTHX_ CV *cv) {
  MAGIC *mg;

  if (!SvMAGICAL(cv)) {
  sv_magic((SV*)cv, (SV *)NULL, PERL_MAGIC_ext, NULL, -1);
  }

  mg = mg_find((SV*)cv, PERL_MAGIC_ext);
  return &(mg->mg_private);
}

The cv passed in comes from cv = get_cv(name, FALSE) where name is "main​::add_my_version".

SvMAGICAL(cv) is true so sv_magic() is not called, and then mg_find() returns NULL and things all go wrong from there.

The "main​::add_my_version" in question is this in modperl_extra.pl​:

sub test_add_version_component {
  Apache2​::ServerUtil->server->push_handlers(
  PerlPostConfigHandler => \&add_my_version);

  sub add_my_version {
  my ($conf_pool, $log_pool, $temp_pool, $s) = @​_;
  $s->add_version_component("world domination series/2.0");
  return Apache2​::Const​::OK;
  }
}

If I move sub add_my_version {} outside of sub test_add_version_component {} like this​:

sub test_add_version_component {
  Apache2​::ServerUtil->server->push_handlers(
  PerlPostConfigHandler => \&add_my_version);
}

sub add_my_version {
  my ($conf_pool, $log_pool, $temp_pool, $s) = @​_;
  $s->add_version_component("world domination series/2.0");
  return Apache2​::Const​::OK;
}

then SvMAGICAL(cv) is now false so sv_magic() is called, and then mg_find() returns non-NULL and all is well.

Is this an expected change in behaviour between 5.21.6 and 5.21.7 that mod_perl needs to keep up with, or a bug in 5.21.7?

(I haven't tried the latest bleadperl yet, but I don't see anything since 5.21.8 that looks like it would affect this.)

From @steve-m-hay

The change in behaviour can be seen with the attached XS module.
Under 5.21.6 "make test" outputs "ok".
Under 5.21.7 it outputs "not ok - SvMAGICAL returned TRUE".

From @steve-m-hay

Foo-1.00.tar.gz

From [Unknown Contact. See original ticket]

The change in behaviour can be seen with the attached XS module.
Under 5.21.6 "make test" outputs "ok".
Under 5.21.7 it outputs "not ok - SvMAGICAL returned TRUE".

From @Leont

On Tue, Jan 27, 2015 at 3​:15 PM, Steve Hay <perlbug-followup@​perl.org>
wrote​:

Testing modperl (svn trunk, not the current CPAN release) with
httpd-2.4.10 and recent 5.21.x perls I find that things look ok up to and
including 5.21.6, but the mod_perl-enabled server fails to start with
5.21.7 and 5.21.8.

This function in modperl_util.c is where things go wrong​:

U16 *modperl_code_attrs(pTHX_ CV *cv) {
MAGIC *mg;

if \(\!SvMAGICAL\(cv\)\) \{
   sv\_magic\(\(SV\*\)cv\, \(SV \*\)NULL\, PERL\_MAGIC\_ext\, NULL\, \-1\);
\}

mg = mg\_find\(\(SV\*\)cv\, PERL\_MAGIC\_ext\);
return &\(mg\->mg\_private\);

}

The cv passed in comes from cv = get_cv(name, FALSE) where name is
"main​::add_my_version".

SvMAGICAL(cv) is true so sv_magic() is not called, and then mg_find()
returns NULL and things all go wrong from there.

The code is naive at best. It is assuming there is no magic on the cv from
any other type. If there is non-ext magic SvMAGICAL will be true but
mg_find(PERL_MAGIC_ext) will return NULL. I would suggest something like
this (untested)​:

U16 *modperl_code_attrs(pTHX_ CV *cv) {
  MAGIC *mg;

  if (!(SvMAGICAL(cv) && mg = mg_find((SV*)cv, PERL_MAGIC_ext))) {
  mg = sv_magicext((SV*)cv, (SV *)NULL, PERL_MAGIC_ext, NULL, NULL,
-1);
  }

  return &(mg->mg_private);
}

Actually, you may prefer to use mg_findext and an explicit vtable (even if
it's empty), but that's another discussion.

Leon

The RT System itself - Status changed from 'new' to 'open'

From @tonycoz

On Tue Jan 27 14​:52​:38 2015, LeonT wrote​:

The code is naive at best. It is assuming there is no magic on the cv from
any other type. If there is non-ext magic SvMAGICAL will be true but
mg_find(PERL_MAGIC_ext) will return NULL. I would suggest something like
this (untested)​:

I had something pretty similar to this written up when I saw your post, but I do wonder...

PERL_DL_NONLAZY=1 "/home/tony/perl/blead/bin/perl5.21.8" "-Iblib/lib" "-Iblib/arch" test.pl
SV = PVCV(0x1d027d0) at 0x1cfa658
  REFCNT = 1
  FLAGS = (RMG,DYNFILE)
  MAGIC = 0x1cd9400
  MG_VIRTUAL = &PL_vtbl_backref
  MG_TYPE = PERL_MAGIC_backref(<)
  MG_OBJ = 0x1cd1640
  COMP_STASH = 0x1cd14f0 "Foo"
  START = 0x1cd9320 ===> 1
  ROOT = 0x1cd92e0
  GVGV​::GV = 0x1d410b8 "Foo" :​: "sub2"
  FILE = "blib/lib/Foo.pm"
  DEPTH = 0
  FLAGS = 0x1000
  OUTSIDE_SEQ = 454
  PADLIST = 0x1cfd8f0
  OUTSIDE = 0x1cd1568 (sub1)
not ok - SvMAGICAL returned TRUE

why the reference started getting backref magic.

Tony

From @tonycoz

On Tue Jan 27 15​:14​:41 2015, tonyc wrote​:

why the reference started getting backref magic.

a70f21d is why.

Tony

From @steve-m-hay

Thank you both for your helpful replies.I should have realized what the cause of the problem there was, but I wouldn't have known whether that was an intentional change in perl or not.

The suggested fix (plus some extra parens) did the trick​:
https://svn.apache.org/viewvc?view=revision&revision=1655200

@steve-m-hay - Status changed from 'open' to 'resolved'