[musl] [PATCH] Zero the leading stack canary byte

[musl] [PATCH] Zero the leading stack canary byte

[jvoisin]

This reduces entropy of the canary from 64-bit to 56-bit in exchange for
mitigating non-terminated C string overflows.

Checking the byte order should be "good enough", since I think that the stacks
on all architectures supported by musl are growing downwards. Worse case, this
can always be improved later if needed.

This is taken from https://github.com/GrapheneOS/platform_bionic/commit/7024d880b51f03a796ff8832f1298f2f1531fd7b
---
 src/env/__stack_chk_fail.c | 7 ++++++-
 1 file changed, 6 insertions(+), 1 deletion(-)

diff --git a/src/env/__stack_chk_fail.c b/src/env/__stack_chk_fail.c
index bf5a280a..45f948fe 100644
--- a/src/env/__stack_chk_fail.c
+++ b/src/env/__stack_chk_fail.c
@@ -2,6 +2,11 @@
 #include <stdint.h>
 #include "pthread_impl.h"
 
+// Sacrifice 8 bits of entropy to mitigate non-terminated C string overflows
+static const uintptr_t canary_mask = __BYTE_ORDER == __BIG_ENDIAN ?
+  0x00ffffffffffffffUL :
+  0xffffffffffffff00UL ;
+
 uintptr_t __stack_chk_guard;
 
 void __init_ssp(void *entropy)
@@ -9,7 +14,7 @@ void __init_ssp(void *entropy)
 	if (entropy) memcpy(&__stack_chk_guard, entropy, sizeof(uintptr_t));
 	else __stack_chk_guard = (uintptr_t)&__stack_chk_guard * 1103515245;
 
-	__pthread_self()->canary = __stack_chk_guard;
+	__pthread_self()->canary = __stack_chk_guard & canary_mask;
 }
 
 void __stack_chk_fail(void)
-- 
2.30.2

Re: [musl] [PATCH] Zero the leading stack canary byte

[Rich Felker]

On Sun, Dec 12, 2021 at 07:34:40PM +0100, jvoisin wrote:
> This reduces entropy of the canary from 64-bit to 56-bit in exchange for
> mitigating non-terminated C string overflows.
> 
> Checking the byte order should be "good enough", since I think that the stacks
> on all architectures supported by musl are growing downwards. Worse case, this
> can always be improved later if needed.
> 
> This is taken from https://github.com/GrapheneOS/platform_bionic/commit/7024d880b51f03a796ff8832f1298f2f1531fd7b
> ---
>  src/env/__stack_chk_fail.c | 7 ++++++-
>  1 file changed, 6 insertions(+), 1 deletion(-)
> 
> diff --git a/src/env/__stack_chk_fail.c b/src/env/__stack_chk_fail.c
> index bf5a280a..45f948fe 100644
> --- a/src/env/__stack_chk_fail.c
> +++ b/src/env/__stack_chk_fail.c
> @@ -2,6 +2,11 @@
>  #include <stdint.h>
>  #include "pthread_impl.h"
>  
> +// Sacrifice 8 bits of entropy to mitigate non-terminated C string overflows
> +static const uintptr_t canary_mask = __BYTE_ORDER == __BIG_ENDIAN ?
> +  0x00ffffffffffffffUL :
> +  0xffffffffffffff00UL ;
> +
>  uintptr_t __stack_chk_guard;
>  
>  void __init_ssp(void *entropy)
> @@ -9,7 +14,7 @@ void __init_ssp(void *entropy)
>  	if (entropy) memcpy(&__stack_chk_guard, entropy, sizeof(uintptr_t));
>  	else __stack_chk_guard = (uintptr_t)&__stack_chk_guard * 1103515245;
>  
> -	__pthread_self()->canary = __stack_chk_guard;
> +	__pthread_self()->canary = __stack_chk_guard & canary_mask;
>  }
>  
>  void __stack_chk_fail(void)
> -- 
> 2.30.2

As written this patch assumes a 64-bit uintptr_t, which isn't ok.
Indeed 56 bits should be fine on a 64-bit arch, but dropping from 32
to 24 on a 32-bit arch severely weakens the protection. So it probably
needs to be conditional on 64-bit.

Also, zeroing the first byte means we can no longer catch buffer
overflows of the form "off-by-one string length". This seems
unfortunate. Putting the 0 byte at the end would solve that at the
expense of allowing the canary value to be leaked via missing
termination bugs, and overall I would lean towards catching actual
buffer overflow bugs vs stopping canary leaks.

Rich

Re: [musl] [PATCH] Zero the leading stack canary byte

[jvoisin]

> As written this patch assumes a 64-bit uintptr_t, which isn't ok.
> Indeed 56 bits should be fine on a 64-bit arch, but dropping from 32
> to 24 on a 32-bit arch severely weakens the protection. So it probably
> needs to be conditional on 64-bit.
Will do.

> Also, zeroing the first byte means we can no longer catch buffer
> overflows of the form "off-by-one string length". This seems
> unfortunate. Putting the 0 byte at the end would solve that at the
> expense of allowing the canary value to be leaked via missing
> termination bugs, and overall I would lean towards catching actual
> buffer overflow bugs vs stopping canary leaks.
As discussed on IRC, what about zero'ing the second byte instead? This
would allow to catch overflows, as well as preventing canary
leaks/overwrite via string-manipulating functions.

Re: [musl] [PATCH] Zero the leading stack canary byte

[Rich Felker]

On Mon, Dec 13, 2021 at 01:24:38PM +0100, jvoisin wrote:
> > As written this patch assumes a 64-bit uintptr_t, which isn't ok.
> > Indeed 56 bits should be fine on a 64-bit arch, but dropping from 32
> > to 24 on a 32-bit arch severely weakens the protection. So it probably
> > needs to be conditional on 64-bit.
> Will do.
> 
> > Also, zeroing the first byte means we can no longer catch buffer
> > overflows of the form "off-by-one string length". This seems
> > unfortunate. Putting the 0 byte at the end would solve that at the
> > expense of allowing the canary value to be leaked via missing
> > termination bugs, and overall I would lean towards catching actual
> > buffer overflow bugs vs stopping canary leaks.
> As discussed on IRC, what about zero'ing the second byte instead? This
> would allow to catch overflows, as well as preventing canary
> leaks/overwrite via string-manipulating functions.

That seems like a good idea too. Note that it can be done without
masking logic and large constants just with:

((char *)&__stack_chk_guard)[1] = 0;

And beginning or end could have been done just by reducing the memcpy
length by 1.

Rich

[musl] [PATCH] Zero the leading stack canary byte

[jvoisin]

This reduces entropy of the canary from 64-bit to 56-bit in exchange for
mitigating non-terminated C string overflows by setting the second byte
of the canary to NULL, so that off-by-one can still be detected.

This is taken from https://github.com/GrapheneOS/platform_bionic/commit/7024d880b51f03a796ff8832f1298f2f1531fd7b
---
 src/env/__stack_chk_fail.c | 8 ++++++++
 1 file changed, 8 insertions(+)

diff --git a/src/env/__stack_chk_fail.c b/src/env/__stack_chk_fail.c
index bf5a280a..be31a88a 100644
--- a/src/env/__stack_chk_fail.c
+++ b/src/env/__stack_chk_fail.c
@@ -9,6 +9,14 @@ void __init_ssp(void *entropy)
 	if (entropy) memcpy(&__stack_chk_guard, entropy, sizeof(uintptr_t));
 	else __stack_chk_guard = (uintptr_t)&__stack_chk_guard * 1103515245;
 
+#if UINTPTR_MAX >= 0xffffffffffffffff
+	/* Sacrifice 8 bits of entropy on 64bit to prevent leaking/overwriting the
+	 * canary via string-manipulation functions. The NULL byte is on the second
+	 * byte so that off-by-ones can still be detected. Endianness is taken care
+	 * of automatically. */
+	((char *)&__stack_chk_guard)[1] = 0;
+#endif
+
 	__pthread_self()->canary = __stack_chk_guard;
 }
 
-- 
2.30.2