Re: PROBLEM: OOPSes in PREEMPT SMP for AMD Opteron Dual-Core with Memhole Mapping (non tainted kernel)

From: Andi Kleen
Date: Mon Jun 13 2005 - 05:09:43 EST


On Sun, Jun 12, 2005 at 03:29:50PM -0400, Jacob Martin wrote:
> Hardware memhole mapping never seems to work, or causes lockups right away. I
> need to test it further though.
>
> I have discovered that with the following features enabled:
>
> 1. Software memhole mapping
> 2. Continuous,
>
> linux sees the entire 4GB of memory. However, when things start getting
> requested from the upper half, there are Oopses generated. Attached are two
> Oopses that occurred under the test scenario described.

What happens when you boot with numa=off or with numa=noacpi ?

The system seems to believe it has memory in an area not covered
by mem_map.


> launch big memory apps.
>
> I suppose I could write a program to consume/probe the upper memory half.
> Anyone know of a good/quicky way to do that?

You can use the attached program which I often use for similar purposes.
It writes nearly all free memory in a loop and also often triggers memory
problems.

-Andi

#define _GNU_SOURCE 1
#include <stdio.h>
#include <stdlib.h>
#include <string.h>
int main(void)
{
size_t linelen = 0;
char *line = NULL;
unsigned long freemem = 0;
FILE *f = fopen("/proc/meminfo", "r");
while (getdelim(&line, &linelen, '\n', f) > 0) {
if (sscanf(line, "LowFree: %lu", &freemem) == 1)
break;
}

freemem *= 1024;

freemem -= freemem/20;
char *s = malloc(freemem);
if (s) {
long i;
for (;;) {
printf("\nwrite\n");
for (i = 0; i < freemem; i += 10*1024*1024) {
long w = freemem - i;
if (w > 10*1024*1024)
w = 10*1024*1024;
memset(s + i, 0xff, w);
putchar('.');
fflush(stdout);
}
printf("\nread\n");
for (i = 0; i < freemem; i += 10*1024*1024) {
long w = freemem - i;
if (w > 10*1024*1024)
w = 10*1024*1024;
memcpy(s, s + i, w);
putchar('.');
fflush(stdout);
}
}
} else
printf("Cannot allocate memory\n");

return 0;
}