The /proc/sys/vm/pagecache tunable and associated kernel code will not prevent
the system form hanging up and eventually panic()'ng running the oom-kill.c test
program.  This test program forks off several processes/threads that allocate
all of the RAM in anonymous regions on a system without any swap space.  That
patch was intended to prevent real-world mixed workloads from causing NUMA
systems to hang and eventually panic() like this.  We are working on upstream
fixes such as re-structureing the VM code so it can get into this type of
lock-up and queued spinlocks so that it wont stay in this state if it does. 
Neither of these can be considered for RHEL4 because it will break the kABI and
will be too unstable for a maintance release.

You can probably prevent the panic and hang form happening by making the OOM
killer more aggressive but at the cost of killing many more processes than it
should:

------------------------------------------------------------------------------
--- linux-2.6.9/mm/oom_kill.c.orig      2008-03-05 14:31:16.000000000 -0500
+++ linux-2.6.9/mm/oom_kill.c   2008-03-05 14:32:39.000000000 -0500
@@ -252,6 +252,9 @@ void out_of_memory(int gfp_mask)
        unsigned long now, since;
  
        spin_lock(&oom_lock);
+
+       goto justkill;
+
        now = jiffies;
        since = now - last;
        last = now;
@@ -292,6 +295,7 @@ void out_of_memory(int gfp_mask)
         */
        lastkill = now;
  
+justkill:
        printk("oom-killer: gfp_mask=0x%x\n", gfp_mask);
  
        /* oom_kill() sleeps */
------------------------------------------------------------------------------

Larry Woodman

Cai, just to be clear here, are you saying that the z-stream RHEL4U6 kernel with
the pagecache patch is worse than the standard RHEL4U6 kernel as far as hanging
and panic()'ng is concerned when running that repro-oom program or ar you saying
that it still acts th same ???

Larry

I'd say it is not a regression from standard RHEL4U6 kernel to Z-stream RHEL4U6.
As far as I remember, in standard RHEL4U6 kernel, the test case did hang the
machine, but did not panic, because of NMI watchdog is not working there. In
Z-stream RHEL4U6 kernel, as the NMI watchdog bug has been fixed, panic has been
seen.

Based on the comment #5, I am going to close this bug as DEFERRED.

Use this program:

#include <stdlib.h>
#include <unistd.h>
#include <sys/mman.h>
#include <errno.h>
#include <stdio.h>
main(int argc,char *argv[])
{
        unsigned long   siz;
        char    *ptr1;
        char    *i;
        int     option;

        if ((argc <= 1)||(argc >3)) {
                printf("bad args, usage: memory <Log2 of memsize> <loop otion: 0=exit,1=loop,2=spin>\n");
                exit(-1);
        }
        siz = ((long)1<<atol(argv[1]));
        ptr1 = (char *)malloc(siz);
        if (!ptr1){
                printf("Unable to malloc\n");
                exit(-1);
        }
        option = atoi(argv[2]);
        printf("option=%d\n", option);
        printf("mmaping %ld anonymous bytes\n", siz);
        ptr1 = (char *)mmap((void *)0,siz,PROT_READ|PROT_WRITE,MAP_ANONYMOUS|MAP_PRIVATE,-1,0);
        if (ptr1 == (char *)-1) {
                printf("ptr1 = %lx\n", ptr1);
                perror("");
        }
        printf("touching %d pages\n", siz/4096);
loop:
        for (i=ptr1; i<ptr1+siz-1; i+=4096) {
                *i=(char)'i';
        }
        if (option==1)
                goto loop;
        else if (option==2)
                while(1);
        else
                printf("exiting\n");
}