ipc/shm: Fix shmat mmap nil-page protection
authorDavidlohr Bueso <dave@stgolabs.net>
Mon, 27 Feb 2017 22:28:24 +0000 (14:28 -0800)
committerWilly Tarreau <w@1wt.eu>
Tue, 20 Jun 2017 12:04:23 +0000 (14:04 +0200)
commit 95e91b831f87ac8e1f8ed50c14d709089b4e01b8 upstream.

The issue is described here, with a nice testcase:

    https://bugzilla.kernel.org/show_bug.cgi?id=192931

The problem is that shmat() calls do_mmap_pgoff() with MAP_FIXED, and
the address rounded down to 0.  For the regular mmap case, the
protection mentioned above is that the kernel gets to generate the
address -- arch_get_unmapped_area() will always check for MAP_FIXED and
return that address.  So by the time we do security_mmap_addr(0) things
get funky for shmat().

The testcase itself shows that while a regular user crashes, root will
not have a problem attaching a nil-page.  There are two possible fixes
to this.  The first, and which this patch does, is to simply allow root
to crash as well -- this is also regular mmap behavior, ie when hacking
up the testcase and adding mmap(...  |MAP_FIXED).  While this approach
is the safer option, the second alternative is to ignore SHM_RND if the
rounded address is 0, thus only having MAP_SHARED flags.  This makes the
behavior of shmat() identical to the mmap() case.  The downside of this
is obviously user visible, but does make sense in that it maintains
semantics after the round-down wrt 0 address and mmap.

Passes shm related ltp tests.

Link: http://lkml.kernel.org/r/1486050195-18629-1-git-send-email-dave@stgolabs.net
Signed-off-by: Davidlohr Bueso <dbueso@suse.de>
Reported-by: Gareth Evans <gareth.evans@contextis.co.uk>
Cc: Manfred Spraul <manfred@colorfullife.com>
Cc: Michael Kerrisk <mtk.manpages@googlemail.com>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
Signed-off-by: Willy Tarreau <w@1wt.eu>
ipc/shm.c

index 08b14f69d6cfdb2176c56c431a2ba3e5d25d0fd0..ddfad445242cd1d07e9de25a91121c6c92fa5fe8 100644 (file)
--- a/ipc/shm.c
+++ b/ipc/shm.c
@@ -1041,8 +1041,8 @@ out_unlock1:
  * "raddr" thing points to kernel space, and there has to be a wrapper around
  * this.
  */
-long do_shmat(int shmid, char __user *shmaddr, int shmflg, ulong *raddr,
-             unsigned long shmlba)
+long do_shmat(int shmid, char __user *shmaddr, int shmflg,
+             ulong *raddr, unsigned long shmlba)
 {
        struct shmid_kernel *shp;
        unsigned long addr;
@@ -1063,8 +1063,13 @@ long do_shmat(int shmid, char __user *shmaddr, int shmflg, ulong *raddr,
                goto out;
        else if ((addr = (ulong)shmaddr)) {
                if (addr & (shmlba - 1)) {
-                       if (shmflg & SHM_RND)
-                               addr &= ~(shmlba - 1);     /* round down */
+                       /*
+                        * Round down to the nearest multiple of shmlba.
+                        * For sane do_mmap_pgoff() parameters, avoid
+                        * round downs that trigger nil-page and MAP_FIXED.
+                        */
+                       if ((shmflg & SHM_RND) && addr >= shmlba)
+                               addr &= ~(shmlba - 1);
                        else
 #ifndef __ARCH_FORCE_SHMLBA
                                if (addr & ~PAGE_MASK)