Strange results with changing HZ

Strange results with changing HZ

[Wojciech Kromer]

My system is:
 - 2.4.25 kernel from denx.de
 - uclibc (gcc 3.4.2) + busybox
 - changed HZ to 1000
 - rebuild kernel only

AFAIR there souldn't be any need to recompile user apps.

I have strange results.
Sleep and timeout selects works fine, but alarm functions seems to be 
too fast.
For example included program or ping waits too short.



it is OK:

# time alarm_test
alarm fired after 0.099166s
real    0m 0.11s
user    0m 0.00s
sys     0m 0.00s
#
# time sleep 1
real    0m 1.01s
user    0m 0.00s
sys     0m 0.00s
# time sleep 10
real    0m 10.01s
user    0m 0.00s
sys     0m 0.00s


but:

# alarm_test
alarm fired after 0.099709s
# alarm_test
alarm fired after 0.099710s


------------------------------------ alarm_test.c :

#include <stdio.h>
#include <signal.h>
#include <unistd.h>
#include <sys/time.h>

static void sighand(int sig)
{
}

int main()
{
 struct timeval tv1, tv2;
 sigset_t set;

 sigfillset(&set);
 sigdelset(&set, SIGALRM);
 sigdelset(&set, SIGINT);

 signal(SIGALRM, sighand);
 gettimeofday(&tv1, NULL);

 alarm(1);
 sigsuspend(&set);
 gettimeofday(&tv2, NULL);

 tv1.tv_sec = tv2.tv_sec - tv1.tv_sec - (tv2.tv_usec < tv1.tv_usec);
 tv1.tv_usec = tv2.tv_usec - tv1.tv_usec + (tv2.tv_usec < tv1.tv_usec) * 
1000000;

 printf("alarm fired after %ld.%06lds\n",
 tv1.tv_sec, tv1.tv_usec);

 return 0;
}

Re: Strange results with changing HZ

[David Jander]

On Monday 21 November 2005 08:44, Wojciech Kromer wrote:
> My system is:
>  - 2.4.25 kernel from denx.de
>  - uclibc (gcc 3.4.2) + busybox
>  - changed HZ to 1000
>  - rebuild kernel only
>
> AFAIR there souldn't be any need to recompile user apps.

Just an uneducated guess, but are you sure uclibc doesn't depend on HZ at all?
Some things might be just straight syscalls, others might be implemented in 
userspace (inside uclibc)... for example alarm(1) programming an alarm on 
jiffies+1/HZ, and HZ being compiled in from kernel headers ?
I may just be talking rubbish though.

Greetings,

-- 
David Jander

Re: Strange results with changing HZ

[Wojciech Kromer]

Dnia 2005-11-21 12:43, Użytkownik David Jander napisał:

>On Monday 21 November 2005 08:44, Wojciech Kromer wrote:
>  
>
>>My system is:
>> - 2.4.25 kernel from denx.de
>> - uclibc (gcc 3.4.2) + busybox
>> - changed HZ to 1000
>> - rebuild kernel only
>>
>>AFAIR there souldn't be any need to recompile user apps.
>>    
>>
>
>Just an uneducated guess, but are you sure uclibc doesn't depend on HZ at all?
>Some things might be just straight syscalls, others might be implemented in 
>userspace (inside uclibc)... for example alarm(1) programming an alarm on 
>jiffies+1/HZ, and HZ being compiled in from kernel headers ?
>I may just be talking rubbish though.
>  
>

Hmm, after recompiling whole uclibc stuff it seem to work same.
There should be kernel problem. Sleep & select works fine, but alarm 
works 10 times faster.

Anyone tries to change HZ to 1000?