On Wed, Jun 24, 2015 at 07:10:33AM +0200, Patrik Lundin wrote:
>
> The summary for now looks like this:
> Working: amd64, sparc64
> Broken: i386, macppc
>
> Is there some relation between threading and 32/64 bit? It is the main
> thing that sticks out currently, since sparc64 rules out an endian issue
> as far as i can tell.
>
So there might be some light at the end of the opendnssec tunnel.
I have noticed that the thread stack size is smaller on i386 than it is
on amd64 using this code snippet:
---
#include <stdio.h>
#include <pthread.h>
pthread_attr_t attr;
size_t stacksize;
int main (){
pthread_attr_init(&attr);
pthread_attr_getstacksize(&attr, &stacksize);
printf("stacksize: %d\n", stacksize);
return 0;
}
---
While amd64 and sparc64 reports 524288, i386 says 262144. It looks like I have
been able to stop opendnssec from crashing by modifying a wrapper function to
pass an attribute variable to pthread_create(), increasing the stack size:
---
pthread_attr_t attr;
size_t stacksize;
pthread_attr_init(&attr);
pthread_attr_getstacksize(&attr, &stacksize);
pthread_attr_setstacksize(&attr, stacksize * 2);
---
The above was inspired from this thread:
http://openbsd-archive.7691.n7.nabble.com/Default-Posix-thread-stack-size-td47465.html
Anyway, I have sent a message regarding this to one of the nlnet people. I will
report back with how things goes. I am not sure if this is the correct fix, or
if one should instead store less data on the stack or something.
--
Patrik Lundin
