Bug in VM accounting code, probably exploitable
Subsystems (3)
Hello.
As far as I know, if overcommit is
disabled, the OOM kill should never
happen.
It seems to be the bug in the linux
kernel though (any version I think,
probably also including 2.4.x), which
makes it possible to overcommit almost
arbitrary and provoke an OOM kill
afterwards.
Attached is a program that demonstrates
the bug. Don't forget to "swapoff -a"
before starting it, or touching pages
will take eternity. And the amount of
RAM must be <1Gb, or the prog will not
work:)
On 2.4.25 I get:
---
May 11 22:28:18 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0) May 11 22:28:20 lin syslogd: /var/log/debug: Cannot allocate memory May 11 22:28:18 lin kernel: VM: killing process mozilla-bin May 11 22:28:18 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1f0/0) May 11 22:28:20 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0) May 11 22:28:21 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0) May 11 22:28:21 lin kernel: VM: killing process X May 11 22:28:21 lin gnome-name-server[1254]: input condition is: 0x11, exiting May 11 22:29:00 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0) May 11 22:29:00 lin kernel: VM: killing process overc_test
---
As you can see, the program caused many
other processes to be killed, before it
died itself.
2.6.6 seems to kill only the test
program itself, but there is one
nasty side-effect here, that mprotect()
fails to merge VMAs because one VMA can
end up with VM_ACCOUNT flag set, and
another - without that flag. That makes
several apps of mine to malfuncate.
Also it might be possible to bypass thesecurity_vm_enough_memory() checks, which
may probably be exploitable in some other
ways.
I am also attaching the fix for 2.6
kernels.
Would be nice if someone can confirm
the bug and verify the fix. Perhaps
I am missing something obvious here,
at least the fact that both 2.6 and 2.4
behaves similar, confuses me. And the
fix looks also very strange, but it
seems to work.
#include <stdio.h>
#include <stdlib.h>
#include <unistd.h>
#include <sys/mman.h>
#include <asm/page.h>
#include <errno.h>
#define SIZE (1024 * 1024 * 1024)
int main(int argc, char *argv[])
{
char *addr;
int i;
if ((addr = mmap(0, SIZE, PROT_READ | PROT_WRITE,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0)) != MAP_FAILED) {
printf("You have more than %i memory, increase SIZE\n", SIZE);
return 1;
}
perror("mmap()");
printf("OK, now trying uncommited...\n");
if ((addr = mmap(0, SIZE, PROT_NONE,
MAP_PRIVATE | MAP_ANONYMOUS, -1, 0)) == MAP_FAILED) {
perror("mmap()");
return 1;
}
if (mprotect(addr, SIZE, PROT_READ | PROT_WRITE) == -1) {
perror("mprotect()");
return 1;
}
printf("Memory of size %i successfully committed!\n", SIZE);
printf("overcommit_memory: ");
fflush(stdout);
system("cat /proc/sys/vm/overcommit_memory");
for (i = 0; i < SIZE; i += PAGE_SIZE) {
addr[i] = 1;
} return 0;
}
include/linux/mm.h +1 -1
@@ -126,7 +126,7 @@#define VM_NONLINEAR 0x00800000/* Is non-linear (remap_file_pages) */ /* It makes sense to apply VM_ACCOUNT to this vma. */ -#define VM_MAYACCT (vma) (!!((vma)->vm_flags & VM_HUGETLB)) +#define VM_MAYACCT (vma) (!((vma)->vm_flags & VM_HUGETLB))#ifndef VM_STACK_DEFAULT_FLAGS /* arch can override this */ #define VM_STACK_DEFAULT_FLAGS VM_DATA_DEFAULT_FLAGS
Scanned by evaluation version of Dr.Web antivirus Daemon
http://drweb.ru/unix/
Quoted 4 lines · Stas Sergeev
mprotect()fails to merge VMAs because one VMA can end up withVM_ACCOUNTflag set, and another without that flag. That makes several apps of mine to malfuncate.
And the fix looks also very strange, but it seems to work.
Great find! Someone has got their test the wrong way round. Your patch is good, but since that VM_MAYACCT macro is being used in one place only, and just hiding what it's actually about, I'd prefer the patch below. Against 2.6.6: Andrew, please apply.
include/linux/mm.h +0 -3
@@ -112,9 +112,6 @@ struct vm_area_struct {#define VM_HUGETLB 0x00400000/* Huge TLB Page VM */ #define VM_NONLINEAR 0x00800000/* Is non-linear (remap_file_pages) */ -/* It makes sense to apply VM_ACCOUNT to this vma. */ -#define VM_MAYACCT (vma) (!!((vma)->vm_flags & VM_HUGETLB)) -#ifndef VM_STACK_DEFAULT_FLAGS /* arch can override this */ #define VM_STACK_DEFAULT_FLAGS VM_DATA_DEFAULT_FLAGS#endif
mm/mprotect.c +1 -2
@@ -174,8 +174,7 @@ mprotect_fixup(struct vm_area_struct *vm* a MAP_NORESERVE private mapping to writable will now reserve. */ if (newflags &VM_WRITE ) { -if (!(vma->vm_flags & (VM_ACCOUNT |VM_WRITE |VM_SHARED )) - &&VM_MAYACCT (vma)) { +if (!(vma->vm_flags & (VM_ACCOUNT |VM_WRITE |VM_SHARED |VM_HUGETLB ))) { charged = (end - start) >>PAGE_SHIFT ;if (security_vm_enough_memory (charged))return -ENOMEM ;
Quoted 37 lines · Stas Sergeev
Hello.
As far as I know, if overcommit is
disabled, the OOM kill should never
happen.
It seems to be the bug in the linux
kernel though (any version I think,
probably also including 2.4.x), which
makes it possible to overcommit almost
arbitrary and provoke an OOM kill
afterwards.
Attached is a program that demonstrates
the bug. Don't forget to "swapoff -a"
before starting it, or touching pages
will take eternity. And the amount of
RAM must be <1Gb, or the prog will not
work:)On 2.4.25 I get:
---May 11 22:28:18 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0) May 11 22:28:20 lin syslogd: /var/log/debug: Cannot allocate memory May 11 22:28:18 lin kernel: VM: killing process mozilla-bin May 11 22:28:18 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1f0/0) May 11 22:28:20 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0) May 11 22:28:21 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0) May 11 22:28:21 lin kernel: VM: killing process X May 11 22:28:21 lin gnome-name-server[1254]: input condition is: 0x11, exiting May 11 22:29:00 lin kernel: __alloc_pages: 0-order allocation failed (gfp=0x1d2/0) May 11 22:29:00 lin kernel: VM: killing process overc_test
Quoted 4 lines
---
As you can see, the program caused many
other processes to be killed, before it
died itself.
About v2.4, can you try v2.4.26 with CONFIG_OOM_KILLER=y ?
As for the overcommit, I think it has always been "broken"? (its always possible to overcommit).
Hi Marcelo.
Quoted 4 lines · Marcelo Tosatti
Quoted 3 lines
As you can see, the program caused many
other processes to be killed, before it
died itself.About v2.4, can you try v2.4.26 with
CONFIG_OOM_KILLER=y?
OK, with CONFIG_OOM_KILLER=y (2.4.26) I get this:
---
May 22 16:14:56 lin kernel: Out of Memory: Killed process 1514 (overc_test). May 22 16:14:56 lin kernel: Out of Memory: Killed process 1320 (mozilla-bin). May 22 16:14:56 lin kernel: Out of Memory: Killed process 1406 (mozilla-bin). May 22 16:14:56 lin kernel: Out of Memory: Killed process 1407 (mozilla-bin). May 22 16:14:56 lin kernel: Out of Memory: Killed process 1409 (mozilla-bin).
---
Better than without OOM killer, although
still not perfect - for some reasons it also
decided to kill mozilla *after* killing the
test program. Strange.
Quoted 1 line
As for the overcommit, I think it has always been "broken"? (its always possible to overcommit).
At least I received a couple of messages that
stated that my test program doesn't OOM the
RedHat Fedora 2.4 kernels - mprotect() fails
with ENOMEM. That resembles the correct
behaveour of the fixed 2.6 so I think at least
RedHat got rid of that problem somehow.
Hope that helps.