SMP users needed to test patch

[Stathis Kamperis-2]

Hello everyone!

I'd like to mark mq*() syscall as MPSAFE, but before that I need
someone to test them in an SMP capable machine running SMP kernel. I
only have UP machines around.

So, if anyone is able and kind enough, here are some directions on how
to do it. I assume s\he is running HEAD.

cd /usr/src
fetch http://leaf.dragonflybsd.org/~beket/mq-mpsafe.diff
git apply mq-mpsafe.diff
make buildkernel
make installkernel
reboot
git clone git://gitweb.dragonflybsd.org/~beket/pcca-tests.git
cd pcca-tests/mqueue.h
make && make -k run
cd etc
make
sysctl -w kern.mqueue.mq_prio_max=200
./t_mq_parallel_threads
./t_mq_parallel_fork

If you encounter problems you can mail me off-list to sort them out. I
know that it's messy, so no hard feelings if no one volunteers :)
Thank you for considering.

Best regards,
Stathis Kamperis

[YONETANI Tomokazu-4]

Hi,

On Wed, Nov 04, 2009 at 02:30:38PM +0200, Stathis Kamperis wrote:
> git clone git://gitweb.dragonflybsd.org/~beket/pcca-tests.git
> cd pcca-tests/mqueue.h
> make && make -k run
> cd etc
> make
> sysctl -w kern.mqueue.mq_prio_max=200
> ./t_mq_parallel_threads
> ./t_mq_parallel_fork

Are we suppose to perform the tests (make -k run and the last two lines)
as root or a non-root user?  I just tried that on vkernel, but if I run
them as root, t_mq_open, t_mq_open_umask dump core, and any subsequent
runs dump core too, no matter run as root or not.  The version of
pcca-tests is 0c458bb3cf5d4943093c9b05ca70cc298f09bb46.

Cheers.

[Stathis Kamperis-2]

2009/11/5 YONETANI Tomokazu <qhwt+dfly@...>:
> Hi,
Hi Yonetani & thank you for bothering!

> On Wed, Nov 04, 2009 at 02:30:38PM +0200, Stathis Kamperis wrote:
>> git clone git://gitweb.dragonflybsd.org/~beket/pcca-tests.git
>> cd pcca-tests/mqueue.h
>> make && make -k run
>> cd etc
>> make
>> sysctl -w kern.mqueue.mq_prio_max=200
>> ./t_mq_parallel_threads
>> ./t_mq_parallel_fork
>
> Are we suppose to perform the tests (make -k run and the last two lines)
> as root or a non-root user?  I just tried that on vkernel, but if I run

As non-root.
(All the tests in the repository assume non-root runtime context. I
should have mentioned that, sorry.)

> them as root, t_mq_open, t_mq_open_umask dump core, and any subsequent
> runs dump core too, no matter run as root or not.  The version of
That's "expected".

When the tests failed for the first time (due to being ran as root) ,
the message queues remained in the system. So every subsequent attempt
to run the tests again, failed because those queues were already
present (opened) in the system. I will fix it by installing a signal
handler that cleans up the mess upon SIGABRT. (Actually, IIRC, some
tests may already have such a thing.)

Most important is that vkernel didn't crash or hang :)
I assume that you didn't pass -n1 to the vkernel, right ?

> pcca-tests is 0c458bb3cf5d4943093c9b05ca70cc298f09bb46.
>
> Cheers.
>

Thanks again.

Best regards,
Stathis Kamperis

[YONETANI Tomokazu-4]

On Thu, Nov 05, 2009 at 01:55:33AM +0200, Stathis Kamperis wrote:

> > Are we suppose to perform the tests (make -k run and the last two lines)
> > as root or a non-root user? 〓I just tried that on vkernel, but if I run
> As non-root.
> (All the tests in the repository assume non-root runtime context. I
> should have mentioned that, sorry.)
>
> > them as root, t_mq_open, t_mq_open_umask dump core, and any subsequent
> > runs dump core too, no matter run as root or not. 〓The version of
> That's "expected".
>
> When the tests failed for the first time (due to being ran as root) ,
> the message queues remained in the system. So every subsequent attempt
> to run the tests again, failed because those queues were already
> present (opened) in the system. I will fix it by installing a signal
> handler that cleans up the mess upon SIGABRT. (Actually, IIRC, some
> tests may already have such a thing.)

OK, thanks for the explanation.

> Most important is that vkernel didn't crash or hang :)
> I assume that you didn't pass -n1 to the vkernel, right ?

I used -n3 and -n16 but vkernel didn't crash or hang in either cases,
so I tried it on the real kernel too, but so far all tests ran fine and
I haven't managed to crash :).
Oh, t_mq_send failed at line 56 after raising kern.mqueue.mq_prio_max to 200,
but if I put it back to default (32), the test passes. Is it expected too?

Cheers.

[Stathis Kamperis-2]

> Oh, t_mq_send failed at line 56 after raising kern.mqueue.mq_prio_max to 200,
> but if I put it back to default (32), the test passes. Is it expected too?
Thanks for discovering a double defect :)

First, my test case uses the MQ_PRIO_MAX constant (default 32) as
defined in sys/mqueue.h to figure out the largest possible value for
message queue priorities.

This is wrong.
It should get the value via sysconf() and _SC_MQ_PRIO_MAX. That said,
when you set max priorities to 200, the test assumed that 32 was the
largest possible value and that 2*32 would certainly be an invalid
priority value. And it expected mq_send(..., 64); to fail, which it
didn't since the real upper value for priorities was 200. When you
restored the limit to its default value, 64 was a "legitimate invalid"
value and the test passed.

Second, we don't expose the real values for MQ_OPEN_MAX, MQ_PRIO_MAX
and _POSIX_MESSAGE_PASSING.

I'll fix both of them tomorrow.

>
> Cheers.
>

Thanks, for once more.


Best regards,
Stathis Kamperis

[Stathis Kamperis-2]

2009/11/5 Stathis Kamperis <ekamperi@...>:
> Second, we don't expose the real values for MQ_OPEN_MAX, MQ_PRIO_MAX
> and _POSIX_MESSAGE_PASSING.
In the sysconf() level that is.

[Stathis Kamperis-2]

2009/11/4 Stathis Kamperis <ekamperi@...>:

> Hello everyone!
>
> I'd like to mark mq*() syscall as MPSAFE, but before that I need
> someone to test them in an SMP capable machine running SMP kernel. I
> only have UP machines around.
>
> So, if anyone is able and kind enough, here are some directions on how
> to do it. I assume s\he is running HEAD.
>
> cd /usr/src
> fetch http://leaf.dragonflybsd.org/~beket/mq-mpsafe.diff
> git apply mq-mpsafe.diff
> make buildkernel
> make installkernel
> reboot
> git clone git://gitweb.dragonflybsd.org/~beket/pcca-tests.git
> cd pcca-tests/mqueue.h
> make && make -k run
> cd etc
> make
> sysctl -w kern.mqueue.mq_prio_max=200
> ./t_mq_parallel_threads
> ./t_mq_parallel_fork

Unfortunately, I forgot to mention one crucial step. After the patch
application one should type 'make sysent' inside sys/kern, so that the
syscall-related files be regenerated, namely the init_sysent.c. Or
even better, *I* should have added the init_sysent.c changes in my
patch. Since buildkernel doesn't regenerate these files automatically,
I'm afraid that the tests ran with giant lock held upon syscall
invocation.

I apologize for wasting people`s time :(  I'll just try to get
physical access on an SMP cabable machine & run them on my own.

Thank you both.

Best regardis,
Stathis Kamperis