Re: [PATCH] Fix OOPS in mmap_region() when merging adjacent VM_LOCKED file segments

[Maksim Yevmenkin <maksim.yevmenkin@gmail.com>]

On Thu, Jan 29, 2009 at 2:48 PM, Randy Dunlap <randy.dunlap@oracle.com> wrote:
> Maksim Yevmenkin wrote:
>> On Thu, Jan 29, 2009 at 12:48 PM, Linus Torvalds
>> <torvalds@linux-foundation.org> wrote:
>>> On Thu, 29 Jan 2009, Linus Torvalds wrote:
>>>> THIS PATCH IS TOTALLY UNTESTED!
>>> Well, it boots. FWIW. I've not really tested anything interesting with it,
>>> but any potential breakage is at least not catastrophic and immediate.
>>>
>>>> diff --git a/mm/mmap.c b/mm/mmap.c
>>>> index 8d95902..3f78ead 100644
>>>> --- a/mm/mmap.c
>>>> +++ b/mm/mmap.c
>>>> @@ -769,6 +769,10 @@ struct vm_area_struct *vma_merge(struct mm_struct *mm,
>>>>       if (vm_flags & VM_SPECIAL)
>>>>               return NULL;
>>>>
>>>> +     /* Anonymous shared mappings are unsharable */
>>>> +     if ((vm_flags & VM_SHARED) && !file)
>>>> +             return NULL;
>>>> +
>>> .. and I think this part of it is actually unnecessary, because what
>>> happens is that a shared anon mapping is turned into a shmem mapping when
>>> it is inserted, and that actually ends up allocating a file for it. So the
>>> vma->vm_file for anon mappings will not match a NULL file pointer
>>> _anyway_, so there's no way it would end up merging.
>>>
>>> So my patch can be further simplified, I think, to just the following.
>>> Even more total lines removed.
>>>
>>> I still want somebody else to look at and think about it, though.
>>
>> Just to confirm. This patch also appear to fix the immediate issue for us.
>
> Is there a (small) test program available?

Yes, it was in the original (first) email. Here it is again

/*
 * Program to provoke kernel NULL pointer de-reference during
 * mmap(...MAP_LOCKED...) in Linux 2.6.28.
 *
 * 1. Create a 32KB test file in /tmp (avoids mlock limit on all recent
 *    Linuxes).
 * 2. mmap it with MAP_LOCKED from top to bottom.  (Provokes the oops,
 *    since vmas can be merged in this case.)
 * 3. Clean up.
 *
 * Compile:
 *
 *	gcc maplock-bug.c -o maplog-bug
 */

#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <stdint.h>
#include <sys/types.h>
#include <sys/stat.h>
#include <fcntl.h>
#include <sys/mman.h>

#define	SIZE	(32*1024)	/* Will get rounded down to page size if nec. */

static char tmp[] = "./maplock-bug.XXXXXX";
static char junkbuf[SIZE];

int
main(void)
{
    int fd;
    int ps = getpagesize();
    size_t sz = (SIZE / ps) * ps;
    void **addrs;
    off_t off;
    int i;

    if ((addrs = malloc((sz / ps) * sizeof (*addrs))) == 0) {
	perror("malloc");
	exit(1);
    }

    if ((fd = mkstemp(tmp)) < 0) {
	perror("mkstemp");
	exit(1);
    }

    if (write(fd, junkbuf, sz) != sz) {
	perror("write");
	exit(1);
    }

    if (close(fd) < 0) {
	perror("close");
	exit(1);
    }

    if ((fd = open(tmp, O_RDONLY)) < 0) {
	perror("open");
	exit(1);
    }

    for (off = sz - ps, i = 0; off >= 0; off -= ps, i++) {
	if ((addrs[i] =
	     mmap(0, ps, PROT_READ, MAP_SHARED|MAP_LOCKED,
		  fd, off)) == MAP_FAILED) {
	    perror("mmap");
	    exit(1);
	}

	printf("Mapped offset 0x%jx at %p\n",
	       (uintmax_t)off, addrs[i]);
    }

    if (close(fd) < 0) {
	perror("close");
	exit(1);
    }

    for (i = 0; i < sz / ps; i++) {
	if (munmap(addrs[i], ps) < 0) {
	    perror("munmap");
	    exit(1);
	}
	printf("Unmapped %p\n", addrs[i]);
    }

    if (unlink(tmp) < 0) {
	perror("unlink");
	exit(1);
    }

    printf("Done\n");
}

Thanks,
max