fs/exec.c: account for argv/envp pointers
authorKees Cook <keescook@chromium.org>
Fri, 23 Jun 2017 22:08:57 +0000 (15:08 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Fri, 23 Jun 2017 23:15:56 +0000 (16:15 -0700)
When limiting the argv/envp strings during exec to 1/4 of the stack limit,
the storage of the pointers to the strings was not included.  This means
that an exec with huge numbers of tiny strings could eat 1/4 of the stack
limit in strings and then additional space would be later used by the
pointers to the strings.

For example, on 32-bit with a 8MB stack rlimit, an exec with 1677721
single-byte strings would consume less than 2MB of stack, the max (8MB /
4) amount allowed, but the pointers to the strings would consume the
remaining additional stack space (1677721 * 4 == 6710884).

The result (1677721 + 6710884 == 8388605) would exhaust stack space
entirely.  Controlling this stack exhaustion could result in
pathological behavior in setuid binaries (CVE-2017-1000365).

[akpm@linux-foundation.org: additional commenting from Kees]
Fixes: b6a2fea39318 ("mm: variable length argument support")
Link: http://lkml.kernel.org/r/20170622001720.GA32173@beast
Signed-off-by: Kees Cook <keescook@chromium.org>
Acked-by: Rik van Riel <riel@redhat.com>
Acked-by: Michal Hocko <mhocko@suse.com>
Cc: Alexander Viro <viro@zeniv.linux.org.uk>
Cc: Qualys Security Advisory <qsa@qualys.com>
Cc: <stable@vger.kernel.org>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
fs/exec.c

index 72934df..9041990 100644 (file)
--- a/fs/exec.c
+++ b/fs/exec.c
@@ -220,8 +220,26 @@ static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
 
        if (write) {
                unsigned long size = bprm->vma->vm_end - bprm->vma->vm_start;
+               unsigned long ptr_size;
                struct rlimit *rlim;
 
+               /*
+                * Since the stack will hold pointers to the strings, we
+                * must account for them as well.
+                *
+                * The size calculation is the entire vma while each arg page is
+                * built, so each time we get here it's calculating how far it
+                * is currently (rather than each call being just the newly
+                * added size from the arg page).  As a result, we need to
+                * always add the entire size of the pointers, so that on the
+                * last call to get_arg_page() we'll actually have the entire
+                * correct size.
+                */
+               ptr_size = (bprm->argc + bprm->envc) * sizeof(void *);
+               if (ptr_size > ULONG_MAX - size)
+                       goto fail;
+               size += ptr_size;
+
                acct_arg_size(bprm, size / PAGE_SIZE);
 
                /*
@@ -239,13 +257,15 @@ static struct page *get_arg_page(struct linux_binprm *bprm, unsigned long pos,
                 *    to work from.
                 */
                rlim = current->signal->rlim;
-               if (size > ACCESS_ONCE(rlim[RLIMIT_STACK].rlim_cur) / 4) {
-                       put_page(page);
-                       return NULL;
-               }
+               if (size > READ_ONCE(rlim[RLIMIT_STACK].rlim_cur) / 4)
+                       goto fail;
        }
 
        return page;
+
+fail:
+       put_page(page);
+       return NULL;
 }
 
 static void put_arg_page(struct page *page)