You are viewing a plain text version of this content. The canonical link for it is here.
Posted to modperl@perl.apache.org by Alex Hunsaker <ba...@gmail.com> on 2010/06/28 20:15:39 UTC
[PATCH] END blocks may segfault
Hi!
perl: 5.10.1, 5.12.1 (i686-linux, x86_64-linux)
mod_perl: 2.0.4
Basically if you have end blocks that modify/add END blocks things
might go crazy.
A simple test case for this is:
package test:
END { eval "END { }" for 1..10 }
gives:
Not a CODE reference.
END failed--call queue aborted.
if I add another END {}:
package test:
END { eval "END { }" for 1..10 }
END { }
I get:
Undefined subroutine &main:: called.
END failed--call queue aborted.
This was found as I was getting strange segfaults when trying to use
NYTProf. Nicholas Clark was able to pinpoint the problem.
See:
http://groups.google.com/group/develnytprof-dev/msg/4ff26e6b3d95861f
thread: http://groups.google.com/group/develnytprof-dev/browse_thread/thread/20d2ea800e8936f
Ive included some relevant bits here (again from Nicholas):
> Of course, I looked at the core perl source to see how END blocks were run,
> and the behaviour was safe with perl. perl unshifts each CV from the list
> before it calls it. I should have remembered that inevitably mod_perl is
> "special".
>
> On reflection, I think that this *is* a bug in mod_perl, as they're going
> to get tripped up in similar fashion if any code run as part of an END block
> contains a string eval containing an END block.
Now the simple fix would be to just make modperl shift things off as
well (see attached patch to make it use Perl_call_list). Course if
they use ModPerl::Global::special_list_(call|clear) like
t/modperl/endav.t things break. As when we go to run END blocks or
call special_list_call() they will have been shifted off. This might
be acceptable behavior for END blocks, but I can see not wanting to
break other things that use modperl_perl_call_use. Maybe, maybe not.
Thoughts?
Another option would be to copy/local() the array in
modpler_perl_call_list. Of course that still wont get it "right"
because we wont run any newly defined subs... But at least we wont
segfault/crap out.
Thoughts?
Re: [PATCH] END blocks may segfault
Posted by Alex Hunsaker <ba...@gmail.com>.
>>> On Wed, Jun 30, 2010 at 02:49, Tim Bunce <Ti...@pobox.com> wrote:
>>> > I suggest the code shift items off the main array, like perl does,
>>> > but also push those items onto a new temp array.
Find attached a patch that does the above (perhaps naively), passes a
make test and fixes the reported problems for me.
Re: [PATCH] END blocks may segfault
Posted by Alex Hunsaker <ba...@gmail.com>.
On Wed, Jun 30, 2010 at 16:37, Tim Bunce <Ti...@pobox.com> wrote:
> On Wed, Jun 30, 2010 at 09:24:42AM -0600, Alex Hunsaker wrote:
>> On Wed, Jun 30, 2010 at 02:49, Tim Bunce <Ti...@pobox.com> wrote:
>> > I suggest the code shift items off the main array, like perl does,
>> > but also push those items onto a new temp array.
> I presume mod_perl doesn't shift the array because it wants the array to
> persist to be used again later.
Yeah... Basically when running under ModPerl::Registry it tries to
treat your script exactly like it is running under cgi. So when the
script finish it runs end blocks. When the next request comes in you
still want those end blocks to run.
>From the ModPerl::Registry documentation:
| END blocks encountered during compilation of a script, are called
| after the script has completed its run, including subsequent
| invocations when the script is cached in memory. This is
| assuming that the script itself doesn't define a package on its
| own. If the script defines its own package, the END blocks in the
| scope of that package will be executed at the end of the interpretor's life.
> Otherwise running 11 the first and 21 the second is the right behavior
> for this (very unusual) test case.
Yeah... I think you are right, I was completely over thinking things I think.
Re: [PATCH] END blocks may segfault
Posted by Tim Bunce <Ti...@pobox.com>.
On Wed, Jun 30, 2010 at 09:24:42AM -0600, Alex Hunsaker wrote:
> On Wed, Jun 30, 2010 at 02:49, Tim Bunce <Ti...@pobox.com> wrote:
> > On Tue, Jun 29, 2010 at 09:50:00PM -0700, Fred Moyer wrote:
> >> I think getting rid of the segfault is a good thing. But if the main
> >> problem is issues with NYTProf, then it seems like this change won't
> >> solve the core problem of autogenerated null end blocks faulting. I'm
> >> don't fully understand the scope of these issues yet, so my answers
> >> (and questions) may not be completely helpful.
>
> Well, the point of the test case is it does not use NYTProf. Sure It
> was found while playing with NYTProf. :)
>
> > I suggest the code shift items off the main array, like perl does,
> > but also push those items onto a new temp array.
> > When there are no more items in the main array it would copy
> > the items back again (av_make() is handy for that).
>
> Wouldn't the list keep getting longer each time? For example take
> your test case:
> END { eval "END { }" for 1..10 }
>
> The first time we run them we should shift of 11 items. The next time
> 21, etc. Am I missing something? Perhaps some cool trick? Thats why
> I was thinking local() or equiv would make the most sense.
I presume mod_perl doesn't shift the array because it wants the array to
persist to be used again later. If not then it should shift.
Otherwise running 11 the first and 21 the second is the right behavior
for this (very unusual) test case.
Tim.
p.s. I'm not familar enough with mod_perl not know the all issues
surrounding how it uses PL_endav. So if this thread continues I'd be
grateful if someone could give some background/context. Thanks.
Re: [PATCH] END blocks may segfault
Posted by Alex Hunsaker <ba...@gmail.com>.
On Wed, Jun 30, 2010 at 02:49, Tim Bunce <Ti...@pobox.com> wrote:
> On Tue, Jun 29, 2010 at 09:50:00PM -0700, Fred Moyer wrote:
>> I think getting rid of the segfault is a good thing. But if the main
>> problem is issues with NYTProf, then it seems like this change won't
>> solve the core problem of autogenerated null end blocks faulting. I'm
>> don't fully understand the scope of these issues yet, so my answers
>> (and questions) may not be completely helpful.
Well, the point of the test case is it does not use NYTProf. Sure It
was found while playing with NYTProf. :)
> I suggest the code shift items off the main array, like perl does,
> but also push those items onto a new temp array.
> When there are no more items in the main array it would copy
> the items back again (av_make() is handy for that).
Wouldn't the list keep getting longer each time? For example take
your test case:
END { eval "END { }" for 1..10 }
The first time we run them we should shift of 11 items. The next time
21, etc. Am I missing something? Perhaps some cool trick? Thats why
I was thinking local() or equiv would make the most sense.
Hrm?
Re: [PATCH] END blocks may segfault
Posted by Tim Bunce <Ti...@pobox.com>.
On Tue, Jun 29, 2010 at 09:50:00PM -0700, Fred Moyer wrote:
> On Mon, Jun 28, 2010 at 11:15 AM, Alex Hunsaker <ba...@gmail.com> wrote:
>
> > Another option would be to copy/local() the array in
> > modpler_perl_call_list. Of course that still wont get it "right"
> > because we wont run any newly defined subs... But at least we wont
> > segfault/crap out.
> >
> > Thoughts?
>
> I think getting rid of the segfault is a good thing. But if the main
> problem is issues with NYTProf, then it seems like this change won't
> solve the core problem of autogenerated null end blocks faulting. I'm
> don't fully understand the scope of these issues yet, so my answers
> (and questions) may not be completely helpful.
I suggest the code shift items off the main array, like perl does,
but also push those items onto a new temp array.
When there are no more items in the main array it would copy
the items back again (av_make() is handy for that).
Tim.
Re: [PATCH] END blocks may segfault
Posted by Fred Moyer <fr...@redhotpenguin.com>.
On Mon, Jun 28, 2010 at 11:15 AM, Alex Hunsaker <ba...@gmail.com> wrote:
> Hi!
>
> perl: 5.10.1, 5.12.1 (i686-linux, x86_64-linux)
> mod_perl: 2.0.4
>
> Basically if you have end blocks that modify/add END blocks things
> might go crazy.
This sounds like it is similar to something Tim B. brought up a month
or so ago dealing with the new NYTProf release, but I wasn't able to
find the thread. Is this issue related?
> A simple test case for this is:
> package test:
> END { eval "END { }" for 1..10 }
>
> gives:
> Not a CODE reference.
> END failed--call queue aborted.
>
> if I add another END {}:
> package test:
> END { eval "END { }" for 1..10 }
> END { }
>
> I get:
> Undefined subroutine &main:: called.
> END failed--call queue aborted.
>
> This was found as I was getting strange segfaults when trying to use
> NYTProf. Nicholas Clark was able to pinpoint the problem.
>
> See:
> http://groups.google.com/group/develnytprof-dev/msg/4ff26e6b3d95861f
> thread: http://groups.google.com/group/develnytprof-dev/browse_thread/thread/20d2ea800e8936f
>
> Ive included some relevant bits here (again from Nicholas):
>
>> Of course, I looked at the core perl source to see how END blocks were run,
>> and the behaviour was safe with perl. perl unshifts each CV from the list
>> before it calls it. I should have remembered that inevitably mod_perl is
>> "special".
>>
>> On reflection, I think that this *is* a bug in mod_perl, as they're going
>> to get tripped up in similar fashion if any code run as part of an END block
>> contains a string eval containing an END block.
That doesn't seem like a widespread use case, but I can see how it
could occur unexpectedly and generate a fair amount of pain. But from
the times that I have used string eval's, it was always accompanied by
some pain.
> Now the simple fix would be to just make modperl shift things off as
> well (see attached patch to make it use Perl_call_list). Course if
> they use ModPerl::Global::special_list_(call|clear) like
> t/modperl/endav.t things break. As when we go to run END blocks or
> call special_list_call() they will have been shifted off. This might
> be acceptable behavior for END blocks, but I can see not wanting to
> break other things that use modperl_perl_call_use. Maybe, maybe not.
> Thoughts?
ModPerl::RegistryCooker appears to use the special_list calls.
http://perl.apache.org/docs/2.0/api/ModPerl/Global.html
So the fix could introduce some breakage into deployments which use MP::RC
> Another option would be to copy/local() the array in
> modpler_perl_call_list. Of course that still wont get it "right"
> because we wont run any newly defined subs... But at least we wont
> segfault/crap out.
>
> Thoughts?
I think getting rid of the segfault is a good thing. But if the main
problem is issues with NYTProf, then it seems like this change won't
solve the core problem of autogenerated null end blocks faulting. I'm
don't fully understand the scope of these issues yet, so my answers
(and questions) may not be completely helpful.