Re: Fw: Re: 2.5 kernel regression in alarm() syscall behaviour?

Re: Fw: Re: 2.5 kernel regression in alarm() syscall behaviour?

[george anzinger]

Andrew Morton wrote:
> 
> Begin forwarded message:
> 
> Date: Mon, 14 Jul 2003 09:39:33 -0500
> From: Amos Waterland <apw@us.ibm.com>
> To: linux-kernel@vger.kernel.org
> Subject: Re: 2.5 kernel regression in alarm() syscall behaviour?
> 
> 
> I think Wes' mail client mangled his testcase a bit.  Here is a cleaned
> up version.

I suppose we are going to have a lot of these.  The test calls alarm 
which sets up an itimer for the specified number of seconds and 
returns the number of seconds remaining on the old itimer.  If any 
useconds remain, seconds is boosted by 1.  The test expects the number 
returned to be the same as what was sent, i.e. 1 second wait is 
expected to return 1 second if it is immeadiatly queried.

The problem with this test is that it assumes that seconds can be 
translated into jiffies with out any error.  Jiffies, however, is now 
defined to be close but not equal to 1/HZ.  In fact, on the x86 
jiffies is 999848 nano seconds.  The conversion of a second with the 
proper round up gives 1001 jiffies and converting this back to seconds 
gives 1.000847848 seconds.  It is this 0.000847848 that is forcing the 
subject test to report a number higher than expected.

IMHO it is the test that is wrong, not the kernel.

-g
> 
> Compile with:
> 
>   % gcc -Wall -Werror alarm.c -o alarm
> 
> Output on 2.4 kernel is:
> 
>   % ./alarm
>   [1058193354] alarm(0), want retval:0; got retval:0 (PASS)
>   ...
>   [1058193354] alarm(9), want retval:8; got retval:8 (PASS)
>   0/10 tests failed
> 
> Output on 2.5 kernel is: many failures.  The number of failures go down
> when the system is heavily stressed.
> 
> ---- Begin alarm.c ----
> 
> #include <unistd.h>
> #include <stdio.h>
> #include <sys/time.h>
> 
> #define MINVAL 0
> #define MAXVAL 10
> #define NOALARM 0
> 
> int main(int argc, char **argv)
> {
>     int retval = 0, failed = 0, tests = MAXVAL, prev = 0, curr = 0;
>     struct timeval time;
> 
>     if (argc > 1)
>         if (sscanf(argv[1], "%d", &tests) != 1)
>             return 1;
> 
>     for (curr = MINVAL; curr < tests; curr++) {
>         retval = alarm(curr);
>         gettimeofday(&time, NULL);
>         printf("[%li] alarm(%d), want retval:%d; ",
>                time.tv_sec, curr, prev);
>         /* was there a previous alarm? */
>         if (retval == NOALARM && prev == NOALARM) {
>             printf("got retval:0 (PASS)");
>         } else if (retval == NOALARM && prev > NOALARM) {
>             printf("got retval:0 (FAIL)");
>             failed++;
>         } else if (retval != prev) {
>             printf("got retval:%d (FAIL)", retval);
>             failed++;
>         } else {
>             printf("got retval:%d (PASS)", retval);
>         }
>         printf("\n");
>         prev = curr;
>     }
>     printf("%d/%d tests failed\n", failed, tests);
>     return failed;
> }
> 
> ---- End alarm.c ----
> -
> To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
> the body of a message to majordomo@vger.kernel.org
> More majordomo info at  http://vger.kernel.org/majordomo-info.html
> Please read the FAQ at  http://www.tux.org/lkml/
> 
> 

-- 
George Anzinger   george@mvista.com
High-res-timers:  http://sourceforge.net/projects/high-res-timers/
Preemption patch: http://www.kernel.org/pub/linux/kernel/people/rml

Re: Fw: Re: 2.5 kernel regression in alarm() syscall behaviour?

[Amos Waterland]

On Mon, Jul 14, 2003 at 05:16:06PM -0700, george anzinger wrote:
> I suppose we are going to have a lot of these.  The test calls alarm
> which sets up an itimer for the specified number of seconds and
> returns the number of seconds remaining on the old itimer.  If any
> useconds remain, seconds is boosted by 1.  The test expects the number
> returned to be the same as what was sent, i.e. 1 second wait is
> expected to return 1 second if it is immeadiatly queried.
> 
> The problem with this test is that it assumes that seconds can be
> translated into jiffies with out any error.  Jiffies, however, is now
> defined to be close but not equal to 1/HZ.  In fact, on the x86
> jiffies is 999848 nano seconds.  The conversion of a second with the
> proper round up gives 1001 jiffies and converting this back to seconds
> gives 1.000847848 seconds.  It is this 0.000847848 that is forcing the
> subject test to report a number higher than expected.
> 
> IMHO it is the test that is wrong, not the kernel.

Thanks for your explanation.

The language of the SuSv3[1] seems a little vague:

  If there is a previous alarm() request with time remaining, alarm()
  shall return a non-zero value that is the number of seconds until the
  previous request would have generated a SIGALRM signal.

It leaves open to interpretation the question of what the return value
should be if the delta between calling alarm(1) and alarm(n) is near 0s,
about .5s, and near 1s.  I think the issue can be boiled down to an
implementation choice of floor(delta), round(delta), ceil(delta).

Linux 2.4, Solaris, OpenBSD take the floor() approach.  Running
alarm2.c[2] on them gives results similar to the following; where
'delta' is the difference in microseconds between calling alarm(1) and
alarm(2), and 'ret' is the return value of the alarm(2) call:

t1      t2      delta   ret
------  ------  ------  ------
940284  947913  7629    1
948336  57230   108894  1
57697   266358  208661  1
266826  576979  310153  1
577455  986953  409498  1
987447  497109  509662  1
497576  106690  609114  1
107173  817304  710131  1
817767  626956  809189  1
Alarm clock

That is, even when 809ms have passed between calling alarm(1) and
alarm(2), Linux 2.4/Solaris/OpenBSD report that there is 1s "until the
previous request would have generated a SIGALRM signal."

Since 2.5 is incrementing tv_sec if any tv_usec remain, it is much
closer to taking a ceil() approach: 1.000847848 gets converted to 2,
rather than 1.

I guess the question boils down to: does Linux 2.6 want to make a
different implementation choice than 2.4 and other Unix variants?

----

[1] http://www.opengroup.org/onlinepubs/007904975/functions/alarm.html

[2]

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

/* Returns the difference in usecs between now and before.
 */
long delta_t(struct timeval now, struct timeval before)
{
    if (now.tv_sec == before.tv_sec && now.tv_usec >= before.tv_usec) {
        return now.tv_usec - before.tv_usec;
    } else if (now.tv_sec > before.tv_sec) {
        return (now.tv_usec + 1000000) - before.tv_usec;
    }

    return -1;
}

int main(int argc, char **argv)
{
    int i, ret, fails = 0;
    struct timeval t1, t2;

    printf("t1\tt2\tdelta\tret\n------\t------\t------\t------\n");

    for (i = 0; ; i += 100000) {
        gettimeofday(&t1, NULL);

        alarm(1);
        usleep(i);
        ret = alarm(2);

        gettimeofday(&t2, NULL);

        printf("%li\t%li\t%li\t%i\n", t1.tv_usec,
               t2.tv_usec, delta_t(t2, t1), ret);
    }

    return fails;
}

Re: Fw: Re: 2.5 kernel regression in alarm() syscall behaviour?

[george anzinger]

Amos Waterland wrote:
> On Mon, Jul 14, 2003 at 05:16:06PM -0700, george anzinger wrote:
> 
>>I suppose we are going to have a lot of these.  The test calls alarm
>>which sets up an itimer for the specified number of seconds and
>>returns the number of seconds remaining on the old itimer.  If any
>>useconds remain, seconds is boosted by 1.  The test expects the number
>>returned to be the same as what was sent, i.e. 1 second wait is
>>expected to return 1 second if it is immeadiatly queried.
>>
>>The problem with this test is that it assumes that seconds can be
>>translated into jiffies with out any error.  Jiffies, however, is now
>>defined to be close but not equal to 1/HZ.  In fact, on the x86
>>jiffies is 999848 nano seconds.  The conversion of a second with the
>>proper round up gives 1001 jiffies and converting this back to seconds
>>gives 1.000847848 seconds.  It is this 0.000847848 that is forcing the
>>subject test to report a number higher than expected.
>>
>>IMHO it is the test that is wrong, not the kernel.
> 
> 
> Thanks for your explanation.
> 
> The language of the SuSv3[1] seems a little vague:
> 
>   If there is a previous alarm() request with time remaining, alarm()
>   shall return a non-zero value that is the number of seconds until the
>   previous request would have generated a SIGALRM signal.
> 
> It leaves open to interpretation the question of what the return value
> should be if the delta between calling alarm(1) and alarm(n) is near 0s,
> about .5s, and near 1s.  I think the issue can be boiled down to an
> implementation choice of floor(delta), round(delta), ceil(delta).
> 
> Linux 2.4, Solaris, OpenBSD take the floor() approach.  Running
> alarm2.c[2] on them gives results similar to the following; where
> 'delta' is the difference in microseconds between calling alarm(1) and
> alarm(2), and 'ret' is the return value of the alarm(2) call:
> 
> t1      t2      delta   ret
> ------  ------  ------  ------
> 940284  947913  7629    1
> 948336  57230   108894  1
> 57697   266358  208661  1
> 266826  576979  310153  1
> 577455  986953  409498  1
> 987447  497109  509662  1
> 497576  106690  609114  1
> 107173  817304  710131  1
> 817767  626956  809189  1
> Alarm clock
> 
> That is, even when 809ms have passed between calling alarm(1) and
> alarm(2), Linux 2.4/Solaris/OpenBSD report that there is 1s "until the
> previous request would have generated a SIGALRM signal."
> 
> Since 2.5 is incrementing tv_sec if any tv_usec remain, it is much
> closer to taking a ceil() approach: 1.000847848 gets converted to 2,
> rather than 1.
> 
> I guess the question boils down to: does Linux 2.6 want to make a
> different implementation choice than 2.4 and other Unix variants?

It is the glibc "alarm()" code that is doing this, not the kernel.

AND, gosh, I was hoping the standard would give some guidance here, 
but it completely ignores the issue.

-g
> 
> ----
> 
> [1] http://www.opengroup.org/onlinepubs/007904975/functions/alarm.html
> 
> [2]
> 
> #include <unistd.h>
> #include <stdio.h>
> #include <sys/time.h>
> 
> /* Returns the difference in usecs between now and before.
>  */
> long delta_t(struct timeval now, struct timeval before)
> {
>     if (now.tv_sec == before.tv_sec && now.tv_usec >= before.tv_usec) {
>         return now.tv_usec - before.tv_usec;
>     } else if (now.tv_sec > before.tv_sec) {
>         return (now.tv_usec + 1000000) - before.tv_usec;
>     }
> 
>     return -1;
> }
> 
> int main(int argc, char **argv)
> {
>     int i, ret, fails = 0;
>     struct timeval t1, t2;
> 
>     printf("t1\tt2\tdelta\tret\n------\t------\t------\t------\n");
> 
>     for (i = 0; ; i += 100000) {
>         gettimeofday(&t1, NULL);
> 
>         alarm(1);
>         usleep(i);
>         ret = alarm(2);
> 
>         gettimeofday(&t2, NULL);
> 
>         printf("%li\t%li\t%li\t%i\n", t1.tv_usec,
>                t2.tv_usec, delta_t(t2, t1), ret);
>     }
> 
>     return fails;
> }
> 
> 

-- 
George Anzinger   george@mvista.com
High-res-timers:  http://sourceforge.net/projects/high-res-timers/
Preemption patch: http://www.kernel.org/pub/linux/kernel/people/rml