Bug#554218: perl: FTBFS on hppa: Thread creation failed

[Niko Tyni-3]

Package: perl
Version: 5.10.1-6
Severity: serious

perl/5.10.1-6 failed to build on paer, see

 https://buildd.debian.org/fetch.cgi?pkg=perl&arch=hppa&ver=5.10.1-6&stamp=1256782011&file=log&as=raw

5.10.1-5 built OK, and as there were no related source changes in -6,
I assume the libc NPTL switch is the culprit.

Cc'ing the debian-hppa list. Could somebody please look into this?
I don't know if perl needs to be recompiled against libc6 2.10 first or
if it shows up with the system perl too.

It should be enough to just run 'cd ext/threads-shared && perl t/stress.t'
but I don't quite have the heart to try it out on paer and its chroot
still has libc6 2.9 anyway.

 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 Thread creation failed: pthread_create returned 11 at ../ext/threads-shared/t/stress.t line 77.
 # Warning: 13 threads failed
 # Note: errno 12 = ENOMEM
 # Test failed: 37 threads timed out
 ext/threads-shared/t/stress...................................FAILED at test 1
 
--
Niko Tyni   ntyni@...



--
To UNSUBSCRIBE, email to debian-hppa-REQUEST@...
with a subject of "unsubscribe". Trouble? Contact listmaster@...

[Domenico Andreoli-4]

Hi,

On Tue, Nov 3, 2009 at 8:36 PM, Niko Tyni <ntyni@...> wrote:
> Package: perl
> Version: 5.10.1-6
> Severity: serious
>
> perl/5.10.1-6 failed to build on paer, see

it has just been successfully built on my hppa box. should i upload it?

> 5.10.1-5 built OK, and as there were no related source changes in -6,
> I assume the libc NPTL switch is the culprit.

eglibc is 2.10.1-5, so NPTL switch does not look directly involved here.

today i hit a problem with g++/libstdc++ 4.4.2-2 which i downgraded
to 4.4.2-1 (which is the same version of libstdc++ user on paer). mah..

is perl written in c++?

regards,
Domenico

-----[ Domenico Andreoli, aka cavok
 --[ http://www.dandreoli.com/gpgkey.asc
   ---[ 3A0F 2F80 F79C 678A 8936  4FEE 0677 9033 A20E BC50


--
To UNSUBSCRIBE, email to debian-hppa-REQUEST@...
with a subject of "unsubscribe". Trouble? Contact listmaster@...

[Domenico Andreoli-4]

Hi again,

On Thu, Nov 5, 2009 at 5:33 PM, Domenico Andreoli <cavokz@...> wrote:
> On Tue, Nov 3, 2009 at 8:36 PM, Niko Tyni <ntyni@...> wrote:
>> Package: perl
>> Version: 5.10.1-6
>> Severity: serious
>>
>> perl/5.10.1-6 failed to build on paer, see
>
> it has just been successfully built on my hppa box. should i upload it?

not sure any more. the package is built but the success of t/stress.t is
subject to the number of threads it makes.

trying with 500 threads the virtual memory goes up to 3Gb (i have 4Gb of RAM
and a 32bit kernel, which gives ~3.6Gb of usable memory) then it fails to
create new ones returning ENOMEM, which looks benign. with 350 threads the
test goes nicely sucking almost 3Gb of virtual memory.

while 350 threads performing a simple sum consume al that virtual memory is
not that clear to me as well as how the standard 50 threads of this test left
paer out of memory.

regards,
Domenico

-----[ Domenico Andreoli, aka cavok
 --[ http://www.dandreoli.com/gpgkey.asc
   ---[ 3A0F 2F80 F79C 678A 8936  4FEE 0677 9033 A20E BC50


--
To UNSUBSCRIBE, email to debian-hppa-REQUEST@...
with a subject of "unsubscribe". Trouble? Contact listmaster@...

[Niko Tyni-3]

On Thu, Nov 05, 2009 at 06:01:00PM +0100, Domenico Andreoli wrote:

> On Thu, Nov 5, 2009 at 5:33 PM, Domenico Andreoli <cavokz@...> wrote:
> > On Tue, Nov 3, 2009 at 8:36 PM, Niko Tyni <ntyni@...> wrote:
> >> Package: perl
> >> Version: 5.10.1-6
> >> Severity: serious
> >>
> >> perl/5.10.1-6 failed to build on paer, see
> >
> > it has just been successfully built on my hppa box. should i upload it?
>
> not sure any more. the package is built but the success of t/stress.t is
> subject to the number of threads it makes.
>
> trying with 500 threads the virtual memory goes up to 3Gb (i have 4Gb of RAM
> and a 32bit kernel, which gives ~3.6Gb of usable memory) then it fails to
> create new ones returning ENOMEM, which looks benign. with 350 threads the
> test goes nicely sucking almost 3Gb of virtual memory.
>
> while 350 threads performing a simple sum consume al that virtual memory is
> not that clear to me as well as how the standard 50 threads of this test left
> paer out of memory.

Thanks for looking at this. Carlos already pointed me to the right
direction off-list (see the bug log), and upstream has made the test
more robust in newer versions.

The errno in the build failure is EAGAIN; it seems the NPTL implementation
returns that in case of resource starvation while LinuxThreads didn't.
No idea what resource it's running out of, apparently it's not memory.
In any case it should not make the test fail.

I'll try to get a new version uploaded this weekend with the improved
test, but in the meantime uploading a manual build is fine by me.
Possibly just retrying it on the buildds would do.
--
Niko Tyni   ntyni@...


--
To UNSUBSCRIBE, email to debian-hppa-REQUEST@...
with a subject of "unsubscribe". Trouble? Contact listmaster@...