Floating point problems on UML - help needed
From: Richard Weinberger
Date: Wed Aug 17 2011 - 16:58:30 EST
Hi,
Gunnar reported that some Java program does not work proper within
UserModeLinux.
After looking closer at the problem I was able to reduce it to a small C
program. (Program is attached.)
It looks like FPU registers get sometimes lost after switching between
two or more threads.
It also happens not always, that's why my test program contains a
infinite loop. After a few million iterations the program abort()s.
I can reproduce the issue on both x86 and x86_64, the host's or UML's
kernel version does not matter.
I've tested 2.6.18 to 3.1-rc2.
Interestingly the problem occurs not on my old Pentium 4 machines.
One P4 has HT the other not.
Only "newer" CPUs are affected.
First I thought it's a race in _switch_to(), but adding
unblock/block_signals() did not help.
Currently I'm running out of ideas.
I'm not an expert in this area of UML. :-(
Any idea what goes wrong here?
Thanks,
//richard
#include <stdio.h>
#include <stdlib.h>
#include <pthread.h>
static int xxx(float f)
{
if(f <= 0.0){
printf("wrong f!: %f\n", f);
return -1;
}
return 0;
}
static void *fun(void *arg)
{
float f = 5.0;
for(;;){
if(xxx(f) < 0)
abort();
}
return NULL;
}
int main()
{
pthread_t t1, t2;
pthread_create(&t1, NULL, fun, NULL);
pthread_create(&t2, NULL, fun, NULL);
pthread_join(t1, NULL);
pthread_join(t2, NULL);
return 0;
}