random results with the gdb testsuite

random results with the gdb testsuite

[Denis PILAT]

Hi all

I am having random results using the gdb testsuite (with dejagnu 1.4.4),
especially for some test like gdb.base/display.exp ,
gdb.base/long_long.exp and gdb.cp/ctti.exp.
Sometime they are successful, sometime 50% fails.

Has anybody already experimented such random behavior ?
Thanks for your answer.
Denis

Re: random results with the gdb testsuite

[teawater]

I test them in GDB-cvs version a lot of times but not got any fail.
Which testsuite and wich GDB did use?
And maybe you can post your "gdb.log" if you don't mind.

BTW Maybe you need post this mail to gdb@sourceware.org.

On Thu, Oct 9, 2008 at 17:19, Denis PILAT <denis.pilat@st.com> wrote:
> Hi all
>
> I am having random results using the gdb testsuite (with dejagnu 1.4.4),
> especially for some test like gdb.base/display.exp ,
> gdb.base/long_long.exp and gdb.cp/ctti.exp.
> Sometime they are successful, sometime 50% fails.
>
> Has anybody already experimented such random behavior ?
> Thanks for your answer.
> Denis
>
>

Re: random results with the gdb testsuite

[Michael Snyder]

I haven't seen those failures.

teawater wrote:
> I test them in GDB-cvs version a lot of times but not got any fail.
> Which testsuite and wich GDB did use?
> And maybe you can post your "gdb.log" if you don't mind.
> 
> BTW Maybe you need post this mail to gdb@sourceware.org.
> 
> On Thu, Oct 9, 2008 at 17:19, Denis PILAT <denis.pilat@st.com> wrote:
>> Hi all
>>
>> I am having random results using the gdb testsuite (with dejagnu 1.4.4),
>> especially for some test like gdb.base/display.exp ,
>> gdb.base/long_long.exp and gdb.cp/ctti.exp.
>> Sometime they are successful, sometime 50% fails.
>>
>> Has anybody already experimented such random behavior ?
>> Thanks for your answer.
>> Denis
>>
>>

Re: random results with the gdb testsuite

[Denis PILAT]

Here is part of the log file. At a specific time  (p/f *(long long *)) 
it seems like the debugger hangs up, then all further test are UNRESOLVED.

If I run this test and don't touch my keyboard, don't switch the current 
window, then it's 100% successful.

If I'm using the console window intensively after having run this test, 
switching window, using keyboard,  then I get the following log where 
apparently my debugger crash, or just close.
I can't find a way to debug that problem.
Of course all test passed well if executed by hand !

I'm using gdb 6.8 testsuite, hosted on linux with an internal ST 
debugger based as well on gdb6.8.


$63 = 100100011010001010110011110001001101010111100110111101111
(gdb) PASS: gdb.base/long_long.exp: p/t *(long long *)ll
p/a *(long long *)ll
$64 = 0x89abcdef
(gdb) PASS: gdb.base/long_long.exp: p/a *(long long *)ll
p/f *(long long *)ll
$65 = 3.5127005640885037e-303
(gdb) PASS: gdb.base/long_long.exp: p/f *(long long *)ll
p/c *(long long *)ll
$66 = -17 'ERROR: Process no longer exists
UNRESOLVED: gdb.base/long_long.exp: p/c *(long long *)ll
ERROR: Couldn't send x/w w to GDB.
UNRESOLVED: gdb.base/long_long.exp: set examine size to w
ERROR: Couldn't send x/x w to GDB.
 ...


-- 
Denis

teawater wrote:
> I test them in GDB-cvs version a lot of times but not got any fail.
> Which testsuite and wich GDB did use?
> And maybe you can post your "gdb.log" if you don't mind.
>
> BTW Maybe you need post this mail to gdb@sourceware.org.
>
> On Thu, Oct 9, 2008 at 17:19, Denis PILAT <denis.pilat@st.com> wrote:
>   
>> Hi all
>>
>> I am having random results using the gdb testsuite (with dejagnu 1.4.4),
>> especially for some test like gdb.base/display.exp ,
>> gdb.base/long_long.exp and gdb.cp/ctti.exp.
>> Sometime they are successful, sometime 50% fails.
>>
>> Has anybody already experimented such random behavior ?
>> Thanks for your answer.
>> Denis
>>
>>
>>     
>
>   

Re: random results with the gdb testsuite

[Mark Kettenis]

> Date: Fri, 10 Oct 2008 10:22:50 +0200
> From: Denis PILAT <denis.pilat@st.com>
> 
> Here is part of the log file. At a specific time  (p/f *(long long *)) 
> it seems like the debugger hangs up, then all further test are UNRESOLVED.

That almost certainly means GDB crashed, which means there is a bug in
our code.

Re: random results with the gdb testsuite

[Denis PILAT]

Mark Kettenis wrote:
>> Date: Fri, 10 Oct 2008 10:22:50 +0200
>> From: Denis PILAT <denis.pilat@st.com>
>>
>> Here is part of the log file. At a specific time  (p/f *(long long *)) 
>> it seems like the debugger hangs up, then all further test are UNRESOLVED.
>>     
>
> That almost certainly means GDB crashed, which means there is a bug in
> our code.
>
>
>   
I would say, in *my* code since I'm the only one who got this problem ...
My problem is how to debug that.

Re: random results with the gdb testsuite

[Daniel Jacobowitz]

On Fri, Oct 10, 2008 at 10:22:50AM +0200, Denis PILAT wrote:
> Here is part of the log file. At a specific time  (p/f *(long long *)) it 
> seems like the debugger hangs up, then all further test are UNRESOLVED.
>
> If I run this test and don't touch my keyboard, don't switch the current  
> window, then it's 100% successful.
>
> If I'm using the console window intensively after having run this test,  
> switching window, using keyboard,  then I get the following log where  
> apparently my debugger crash, or just close.
> I can't find a way to debug that problem.

Normally Windows (you did say this was Windows, right?) can be
configured to pop up a dialog when a program crashes.  You can't
necessarily get into GDB from there, but you may be able to get some
kind of dump.  I think the Cygwin folks have instructions on getting
stack dumps or even core dumps from that point.

-- 
Daniel Jacobowitz
CodeSourcery

Re: random results with the gdb testsuite

[Denis PILAT]

Daniel Jacobowitz wrote:
> On Fri, Oct 10, 2008 at 10:22:50AM +0200, Denis PILAT wrote:
>   
>> Here is part of the log file. At a specific time  (p/f *(long long *)) it 
>> seems like the debugger hangs up, then all further test are UNRESOLVED.
>>
>> If I run this test and don't touch my keyboard, don't switch the current  
>> window, then it's 100% successful.
>>
>> If I'm using the console window intensively after having run this test,  
>> switching window, using keyboard,  then I get the following log where  
>> apparently my debugger crash, or just close.
>> I can't find a way to debug that problem.
>>     
>
> Normally Windows (you did say this was Windows, right?) can be
> configured to pop up a dialog when a program crashes.  You can't
> necessarily get into GDB from there, but you may be able to get some
> kind of dump.  I think the Cygwin folks have instructions on getting
> stack dumps or even core dumps from that point.
>   
Sorry, I said window but not in the Microsoft sense. I'm using kde in a 
linux RHEL3.
My debugger terminates, but I can't determine how, valgrind detects 
nothing, it's like if the process is "killed -9".

Re: random results with the gdb testsuite

[Daniel Jacobowitz]

On Fri, Oct 10, 2008 at 02:53:23PM +0200, Denis PILAT wrote:
> Sorry, I said window but not in the Microsoft sense. I'm using kde in a  
> linux RHEL3.
> My debugger terminates, but I can't determine how, valgrind detects  
> nothing, it's like if the process is "killed -9".

Unless your ulimit is messed up, you should get a core dump after this
happens.

-- 
Daniel Jacobowitz
CodeSourcery

Re: random results with the gdb testsuite

[Denis PILAT]

Daniel Jacobowitz wrote:
> On Fri, Oct 10, 2008 at 02:53:23PM +0200, Denis PILAT wrote:
>   
>> Sorry, I said window but not in the Microsoft sense. I'm using kde in a  
>> linux RHEL3.
>> My debugger terminates, but I can't determine how, valgrind detects  
>> nothing, it's like if the process is "killed -9".
>>     
>
> Unless your ulimit is messed up, you should get a core dump after this
> happens.
>   
No, I already checked, it's "coredumpsize  unlimited".

Re: random results with the gdb testsuite

[Michael Snyder]

Denis PILAT wrote:
> 
> Daniel Jacobowitz wrote:
>> On Fri, Oct 10, 2008 at 10:22:50AM +0200, Denis PILAT wrote:
>>
>>> Here is part of the log file. At a specific time  (p/f *(long long *)) it
>>> seems like the debugger hangs up, then all further test are UNRESOLVED.
>>>
>>> If I run this test and don't touch my keyboard, don't switch the current
>>> window, then it's 100% successful.
>>>
>>> If I'm using the console window intensively after having run this test,
>>> switching window, using keyboard,  then I get the following log where
>>> apparently my debugger crash, or just close.
>>> I can't find a way to debug that problem.
>>>
>> Normally Windows (you did say this was Windows, right?) can be
>> configured to pop up a dialog when a program crashes.  You can't
>> necessarily get into GDB from there, but you may be able to get some
>> kind of dump.  I think the Cygwin folks have instructions on getting
>> stack dumps or even core dumps from that point.
>>
> Sorry, I said window but not in the Microsoft sense. I'm using kde in a
> linux RHEL3.
> My debugger terminates, but I can't determine how, valgrind detects
> nothing, it's like if the process is "killed -9".

What happens if you run gdb by hand, and repeat each command
from the log?

Re: random results with the gdb testsuite

[Denis PILAT]

Michael Snyder wrote:
> Denis PILAT wrote:
>>
>> Daniel Jacobowitz wrote:
>>> On Fri, Oct 10, 2008 at 10:22:50AM +0200, Denis PILAT wrote:
>>>
>>>> Here is part of the log file. At a specific time  (p/f *(long long 
>>>> *)) it
>>>> seems like the debugger hangs up, then all further test are 
>>>> UNRESOLVED.
>>>>
>>>> If I run this test and don't touch my keyboard, don't switch the 
>>>> current
>>>> window, then it's 100% successful.
>>>>
>>>> If I'm using the console window intensively after having run this 
>>>> test,
>>>> switching window, using keyboard,  then I get the following log where
>>>> apparently my debugger crash, or just close.
>>>> I can't find a way to debug that problem.
>>>>
>>> Normally Windows (you did say this was Windows, right?) can be
>>> configured to pop up a dialog when a program crashes.  You can't
>>> necessarily get into GDB from there, but you may be able to get some
>>> kind of dump.  I think the Cygwin folks have instructions on getting
>>> stack dumps or even core dumps from that point.
>>>
>> Sorry, I said window but not in the Microsoft sense. I'm using kde in a
>> linux RHEL3.
>> My debugger terminates, but I can't determine how, valgrind detects
>> nothing, it's like if the process is "killed -9".
>
> What happens if you run gdb by hand, and repeat each command
> from the log?
Everything's fine if done by hand. I get the testsuite expected results. 
Like if I run the test without doind anything else on my linux.
>

Re: random results with the gdb testsuite

[Michael Snyder]

Denis PILAT wrote:
> 
> Michael Snyder wrote:
>> Denis PILAT wrote:
>>> Daniel Jacobowitz wrote:
>>>> On Fri, Oct 10, 2008 at 10:22:50AM +0200, Denis PILAT wrote:
>>>>
>>>>> Here is part of the log file. At a specific time  (p/f *(long long
>>>>> *)) it
>>>>> seems like the debugger hangs up, then all further test are
>>>>> UNRESOLVED.
>>>>>
>>>>> If I run this test and don't touch my keyboard, don't switch the
>>>>> current
>>>>> window, then it's 100% successful.
>>>>>
>>>>> If I'm using the console window intensively after having run this
>>>>> test,
>>>>> switching window, using keyboard,  then I get the following log where
>>>>> apparently my debugger crash, or just close.
>>>>> I can't find a way to debug that problem.
>>>>>
>>>> Normally Windows (you did say this was Windows, right?) can be
>>>> configured to pop up a dialog when a program crashes.  You can't
>>>> necessarily get into GDB from there, but you may be able to get some
>>>> kind of dump.  I think the Cygwin folks have instructions on getting
>>>> stack dumps or even core dumps from that point.
>>>>
>>> Sorry, I said window but not in the Microsoft sense. I'm using kde in a
>>> linux RHEL3.
>>> My debugger terminates, but I can't determine how, valgrind detects
>>> nothing, it's like if the process is "killed -9".
>> What happens if you run gdb by hand, and repeat each command
>> from the log?
> Everything's fine if done by hand. I get the testsuite expected results.
> Like if I run the test without doind anything else on my linux.

Ah, how frustrating.  I have occasionally run into that type
of behavior.  It can be very challenging to debug.  It implies
that the testing framework itself is perturbing the system.

Could be timing issues, could be memory locations being shifted around.

There's a load address randomization feature in modern linuxen,
have you tried disabling that?

Other than that, sticking "printfs" into the expect scripts...
;-(

Re: random results with the gdb testsuite

[Thiago Jung Bauermann]

El jue, 09-10-2008 a las 11:19 +0200, Denis PILAT escribió:
> I am having random results using the gdb testsuite (with dejagnu 1.4.4),
> especially for some test like gdb.base/display.exp ,
> gdb.base/long_long.exp and gdb.cp/ctti.exp.
> Sometime they are successful, sometime 50% fails.

Yes for display.exp, no for the other two.

> Has anybody already experimented such random behavior ?

I definitely see random results in testsuite runs on ppc-linux and
ppc64-linux. This is an outdated and incomplete list of testcases which
randomly fail for me:

gdb.base/display.exp
gdb.base/checkpoint.exp
gdb.base/multi-forks.exp
gdb.threads/schedlock.exp
gdb.threads/watchthreads2.exp

There was some work recently to fix some random failures in the
testsuite, so maybe a few of the above are fixed by now, haven't
checked.
-- 
[]'s
Thiago Jung Bauermann
IBM Linux Technology Center