Re: [Xenomai] Segfaults and ENOMEM during rt_event_create()

[Philippe Gerum <rpm@xenomai.org>]

On 08/06/2015 09:49 AM, Gilles Chanteperdrix wrote:
> On Thu, Aug 06, 2015 at 09:40:53AM +0200, Philippe Gerum wrote:
>> On 08/05/2015 07:24 PM, Gilles Chanteperdrix wrote:
>>> On Wed, Aug 05, 2015 at 05:42:37PM +0200, Philippe Gerum wrote:
>>>> On 05/08/2015 15:09, Charles Kiorpes wrote:
>>>>> Hello,
>>>>>
>>>>> I am trying to port a fairly large application from Xenomai 2.6.4 to
>>>>> Xenomai 3.0-rc6
>>>>>
>>>>> Most of the port has been decently painless, but there is one issue I
>>>>> can't seem to get past.
>>>>>
>>>>> I am consistently getting ENOMEM errors and segfaults in
>>>>> libcopperplate.so during rt_event_create(), and always at the same
>>>>> spot in my application.
>>>>>
>>>>> If a segfault occurrs, the dmesg output from the segfault looks like this:
>>>>> [   235.422300] PortProcess[1502]: segfault at f1b3ebeb ip b76c14b7 sp
>>>>> bfeae498 error 5 in libcopperplate.so.0.0.0[b76b7000+14000]
>>>>>
>>>>> The application uses a large number of named event flag groups.  At
>>>>> the time that the error occurs, it is looping to create 32 new named
>>>>> event groups.
>>>>>
>>>>> I have increased these values in my kernel config:
>>>>> CONFIG_XENO_OPT_REGISTRY_NRSLOTS=4096
>>>>> CONFIG_XENO_OPT_SYS_HEAPSZ=512
>>>>> CONFIG_XENO_OPT_PRIVATE_HEAPSZ=128
>>>>> CONFIG_XENO_OPT_SHARED_HEAPSZ=128
>>>>>
>>>>> When I dig through the /proc/xenomai filesystem, I see that:
>>>>> - I am only using 383/4096 registry slots
>>>>> - My system heap is about half full (284160KB free of 524288KB)
>>>>> - My shared and private heaps are basically empty
>>>>>
>>>>> My application is launched with --mem-pool-size=400000.  The main heap
>>>>> (in /registry/user/session/system) has about 220KB used at the point
>>>>> that the error occurs.
>>>>> Increasing the mem-pool-size does not change the behavior.  I have
>>>>> also tried further increasing the values of my system (1024), private
>>>>> (256), and shared (256) heapsizes, with no effect.
>>>>>
>>>>> Obviously I've missed something, but I can't figure out where the
>>>>> error is occurring.
>>>>>
>>>>> System details:
>>>>> OS: Debian Jessie 32-bit
>>>>> Kernel: 3.18.12 - patched using Xenomai 3.0-rc6
>>>>> System: Intel Xeon W3503 (x86 architecture)
>>>>> I-Pipe version 1
>>>>>
>>>>
>>>> Any improvement with this patch in?
>>>>
>>>> commit 574c8e2edb99ff523e765cdb8749dd5259784627
>>>> Author: Philippe Gerum <rpm@xenomai.org>
>>>> Date:   Wed Aug 5 16:08:23 2015 +0200
>>>>
>>>>     lib/cobalt: fix private event state accessor
>>>>
>>>> diff --git a/lib/cobalt/internal.c b/lib/cobalt/internal.c
>>>> index 348ca14..62ecb4a 100644
>>>> --- a/lib/cobalt/internal.c
>>>> +++ b/lib/cobalt/internal.c
>>>> @@ -426,7 +426,7 @@ struct cobalt_event_state
>>>> *get_event_state(cobalt_event_t *event)
>>>>  {
>>>>  	return event->flags & COBALT_EVENT_SHARED ?
>>>
>>> I would put parens here.
>>>
>>
>> I'm trying hard, but I don't see why. The standard is pretty clear about
>> the ternary conditional (§6.5.15): its precedence is right above the
>> assignment operators and below anything else, expr0 is evaluated first
>> and followed by a sequence point.
> 
> I had trouble with if (a & b && c), so am now putting parens around
> (a & b) by principle, even if not needed.
> 

I agree in general, but the ternary operator is a bit special since by
design, it won't compete for precedence with the expressions it evaluates.

-- 
Philippe.