Kernel issues on R4000/R4000 SC and MC

Kernel issues on R4000/R4000 SC and MC

[Ralf Baechle]

Since the introduction of rmap into the kernel memory managment has been
broken for the SC and MC versions of R4000 and R4400 versions.

Interestingly enough this was only ever reported once in May 2006 and at
that time the report just didn't make sense, so it ended up being ignored,
see http://www.linux-mips.org/cgi-bin/mesg.cgi?a=linux-mips&i=1146433653.18721.10.camel%40pkunk.wgtn.cat-it.co.nz

Anyway, the issue boiled up again last week and was supposedly fixed for
linux-2.6.17-rc7 which I've just merged.   I'd like to ask somebody with
one of the affected CPUs to test this.  Below Nick Piggin's test program.

Thanks,

  Ralf

#define _GNU_SOURCE
#include <stdlib.h>
#include <stdio.h>
#include <unistd.h>
#include <sys/mman.h>
#include <assert.h>

int main(void)
{
	char *mem, *mem2, *frag;
	size_t length = getpagesize();
	unsigned int i;

	mem = mmap(NULL, length, PROT_READ, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
	if (mem == MAP_FAILED)
		perror("mmap"), exit(1);
	/* fault in a zero page */
	for (i = 0; i < length; i++)
		assert(mem[i] == 0);

	/* create an obstacle */
	frag = mmap(NULL, length, PROT_READ, MAP_ANONYMOUS|MAP_PRIVATE, -1, 0);
	if (frag == MAP_FAILED)
		perror("mmap"), exit(1);
	for (i = 0; i < length; i++)
		assert(frag[i] == 0);
	
	mem2 = mremap(mem, length, length*2, MREMAP_MAYMOVE);
	if (mem2 == MAP_FAILED)
		perror("mremap"), exit(1);
	
	if (mem == mem2)
		printf("Did not move old mappings!\n");
	else {
		unsigned int order = 3;
		unsigned long size = length << order;
		unsigned long mask = (size - 1) & ~(length-1);

		printf("Moved ZERO_PAGE vaddr from %p to %p\n", mem, mem2);
		if (((unsigned long)mem&mask) == ((unsigned long)mem2&mask))
			printf("This will not change physical ZERO_PAGEs\n");
		else
			printf("This will change physical ZERO_PAGEs\n");
	}

	if (munmap(frag, length) == -1)
		perror("munmap"), exit(1);

	if (munmap(mem2, length*2) == -1)
		perror("munmap"), exit(1);

	exit(0);
}

Re: Kernel issues on R4000/R4000 SC and MC

[Atsushi Nemoto]

On Wed, 31 Jan 2007 13:09:26 +0000, Ralf Baechle <ralf@linux-mips.org> wrote:
> Anyway, the issue boiled up again last week and was supposedly fixed for
> linux-2.6.17-rc7 which I've just merged.   I'd like to ask somebody with
> one of the affected CPUs to test this.  Below Nick Piggin's test program.

What is the expected output of the test program?

---
Atsushi Nemoto

Re: Kernel issues on R4000/R4000 SC and MC

[Thomas Bogendoerfer]

On Thu, Feb 01, 2007 at 07:07:17PM +0900, Atsushi Nemoto wrote:
> On Wed, 31 Jan 2007 13:09:26 +0000, Ralf Baechle <ralf@linux-mips.org> wrote:
> > Anyway, the issue boiled up again last week and was supposedly fixed for
> > linux-2.6.17-rc7 which I've just merged.   I'd like to ask somebody with
> > one of the affected CPUs to test this.  Below Nick Piggin's test program.
> 
> What is the expected output of the test program?

I can confirm, that it kills R4400SC/MC machines. I saw some
"detected softlock" messages. With just one zero page (CPU without VCE)
everything is fine

Thomas.

-- 
Crap can work. Given enough thrust pigs will fly, but it's not necessary a
good idea.                                                [ RFC1925, 2.3 ]