Date:   Tue, 14 Feb 2017 12:36:48 -0800
From:   Kees Cook <keescook@...omium.org>
To:     Hoeun Ryu <hoeun.ryu@...il.com>,
        Mark Rutland <mark.rutland@....com>
Cc:     LKML <linux-kernel@...r.kernel.org>,
        "kernel-hardening@...ts.openwall.com" 
        <kernel-hardening@...ts.openwall.com>
Subject: Re: [PATCH] usercopy: add testcases to check zeroing on failure of usercopy

On Mon, Feb 13, 2017 at 5:44 PM, Hoeun Ryu <hoeun.ryu@...il.com> wrote:
>
>
>> On Feb 14, 2017, at 4:24 AM, Kees Cook <keescook@...omium.org> wrote:
>>
>>> On Mon, Feb 13, 2017 at 10:33 AM, Kees Cook <keescook@...omium.org> wrote:
>>>> On Sat, Feb 11, 2017 at 10:13 PM, Hoeun Ryu <hoeun.ryu@...il.com> wrote:
>>>> In the hardend usercopy, the destination buffer will be zeroed if
>>>> copy_from_user/get_user fails. This patch adds testcases for it.
>>>> The destination buffer is set with non-zero value before illegal
>>>> copy_from_user/get_user is executed and the buffer is compared to
>>>> zero after usercopy is done.
>>>>
>>>> Signed-off-by: Hoeun Ryu <hoeun.ryu@...il.com>
>>>> ---
>>>> lib/test_user_copy.c | 17 +++++++++++++++++
>>>> 1 file changed, 17 insertions(+)
>>>>
>>>> diff --git a/lib/test_user_copy.c b/lib/test_user_copy.c
>>>> index 0ecef3e..54bd898 100644
>>>> --- a/lib/test_user_copy.c
>>>> +++ b/lib/test_user_copy.c
>>>> @@ -41,11 +41,18 @@ static int __init test_user_copy_init(void)
>>>>        char *bad_usermem;
>>>>        unsigned long user_addr;
>>>>        unsigned long value = 0x5A;
>>>> +       char *zerokmem;
>>>>
>>>>        kmem = kmalloc(PAGE_SIZE * 2, GFP_KERNEL);
>>>>        if (!kmem)
>>>>                return -ENOMEM;
>>>>
>>>> +       zerokmem = kzalloc(PAGE_SIZE * 2, GFP_KERNEL);
>>>> +       if (!zerokmem) {
>>>> +               kfree(kmem);
>>>> +               return -ENOMEM;
>>>> +       }
>>>> +
>>>>        user_addr = vm_mmap(NULL, 0, PAGE_SIZE * 2,
>>>>                            PROT_READ | PROT_WRITE | PROT_EXEC,
>>>>                            MAP_ANONYMOUS | MAP_PRIVATE, 0);
>>>> @@ -69,25 +76,35 @@ static int __init test_user_copy_init(void)
>>>>                    "legitimate put_user failed");
>>>>
>>>>        /* Invalid usage: none of these should succeed. */
>>>> +       memset(kmem, 0x5A, PAGE_SIZE);
>>>>        ret |= test(!copy_from_user(kmem, (char __user *)(kmem + PAGE_SIZE),
>>>>                                    PAGE_SIZE),
>>>>                    "illegal all-kernel copy_from_user passed");
>>>> +       ret |= test(memcmp(zerokmem, kmem, PAGE_SIZE),
>>>> +                   "zeroing failure for illegal all-kernel copy_from_user");
>>>> +       memset(bad_usermem, 0x5A, PAGE_SIZE);
>>>
>>> Oh, actually, ha-ha: this isn't legal: it's a direct copy from kernel
>>> to userspace. :) This needs a copy_to_user()... (and same for the
>>> memcmp...)
>
> I just came up with that usercopy doesn't check the buffer is valid when zeroing happens. I mean if the buffer is wrong address pointing other kernel objects or user space address, is it possible for zeroing to overwrite the address ?

The overwrite happening even when the address is "wrong" seems like a
bug to me, but it's sort of already too late (a bad kernel address
would have already been a target for a userspace copy), but if
something has gone really wrong (i.e. attacker doesn't have control
over the source buffer) this does give a "write 0" primitive.

Mark Rutland noticed some order-of-operations issues here too, and his
solution is pretty straight forward: move the checks outside the
failure path. If the kernel target is demonstrably bad, then the
process will be killed before the write 0 happens. (In the non-const
case at least...)

(Oh, btw, I just noticed that x86's copy_from_user() already does the
check before _copy_from_user() can do the memset, so x86 is already
"ok" in this regard.)

>>>> +       ret |= test(memcmp(zerokmem, kmem, sizeof(value)),
>>>> +                   "zeroing failure for illegal get_user");
>
> Actually on my x86_64 (qemu), this testcase fails.
> The generic get_user has zeroing but the one of arch x86 does not.
> Do we need to propagate zeroing to the other arch specific get_user code ?

Hm, this didn't fail for me on x86 nor arm. Or, at least, my updated
test doesn't fail:

       value = 0x5A;
       ret |= test(!get_user(value, (unsigned long __user *)kmem),
                   "illegal get_user passed");
       ret |= test(value != 0, "zeroing failure for illegal get_user");

I see the zeroing in the x86 uaccess.h, though it's a bit obfuscated:

#define get_user(x, ptr)                                                \
({                                                                      \
        int __ret_gu;                                                   \
        register __inttype(*(ptr)) __val_gu asm("%"_ASM_DX);            \
        register void *__sp asm(_ASM_SP);                               \
        __chk_user_ptr(ptr);                                            \
        might_fault();                                                  \
        asm volatile("call __get_user_%P4"                              \
                     : "=a" (__ret_gu), "=r" (__val_gu), "+r" (__sp)    \
                     : "0" (ptr), "i" (sizeof(*(ptr))));                \
        (x) = (__force __typeof__(*(ptr))) __val_gu;                    \
        __builtin_expect(__ret_gu, 0);                                  \
})

__ret_gu is the 0 or -EFAULT (from the __get_user_* assembly), and x
is set unconditionally to __val_gu, which gets zero-set by the same
assembly.

Regardless, I'll expand the tests to check all get_user() sizes...

-Kees

-- 
Kees Cook
Pixel Security

Powered by blists - more mailing lists