Re: [RFC PATCH 0/2] x86: fix Xen PV regression caused by NUMA page migration

[David Vrabel <david.vrabel@citrix.com>]

[-- Attachment #1: Type: text/plain, Size: 815 bytes --]

On 21/03/14 18:18, David Vrabel wrote:
> This series should properly fix the Xen PV guest regression introduced
> by 1667918b6483 (mm: numa: clear numa hinting information on
> mprotect).  The previous fix a9c8e4beeeb6 (xen: properly account for
> _PAGE_NUMA during xen pte translations) breaks save/restore
> (migration) and needs to be reverted.
> 
> I've only given this series a minimal amount of testing and would
> appreciate testing by someone who experienced/reproduced the original
> regression.

Attached is a simple test program that triggers the bug.

Patch #2 fixes the regression and I think that it is the correct fix but
I need to confirm it doesn't cause any measurable performance problems
on native x86 (I think it should not).

I don't have any more time to work on this until next week.

David

[-- Attachment #2: mprotect.c --]
[-- Type: text/x-csrc, Size: 1407 bytes --]

#include <sys/mman.h>
#include <stdio.h>
#include <stdlib.h>
#include <string.h>

#define PAGE_SIZE 4096
#define MAGIC 0x79128347ul

int main(void)
{
    void *p;
    volatile unsigned long *m;
    int ret;

    /* Map a frame for the test. */
    m = p = mmap(NULL, PAGE_SIZE, PROT_WRITE|PROT_READ, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
    if (p == MAP_FAILED) {
        perror("mmap");
        exit(1);
    }

    /* Write to the frame. */
    *m = MAGIC;

    /* Set PROT_NONE on the mapping. */
    ret = mprotect(p, PAGE_SIZE, PROT_NONE);
    if (ret < 0) {
        perror("mprotect");
        exit(1);
    }

    /*
     * Set PROT_READ (or any other non-PROT_NONE protection) on the
     * mapping.
     *
     * On a buggy system pte_mknonnuma() will clear _PAGE_PROTNONE (==
     * _PAGE_NUMA) and set _PAGE_PRESENT /without/ doing a PFN to MFN
     * conversion.
     *
     * The resulting mapping will be to the wrong MFN.
     */
    ret = mprotect(p, PAGE_SIZE, PROT_READ);
    if (ret < 0) {
        perror("mprotect");
        exit(1);
    }

    /* Check we're still mapping the original frame. */
    if (*m != MAGIC) {
        printf("FAIL (0x%lx != 0x%lx)\n", *m, MAGIC);
        ret = 1;
    }

    /*
     * On a buggy system, the munmap() will usually trigger a "Bad
     * page" error since the mapping isn't for the original page.
     */
    munmap(p, PAGE_SIZE);

    return ret;
}

[-- Attachment #3: Type: text/plain, Size: 126 bytes --]

_______________________________________________
Xen-devel mailing list
Xen-devel@lists.xen.org
http://lists.xen.org/xen-devel