[PATCH] x86: fix possible spectre-v1 in do_get_thread_area()

[PATCH] x86: fix possible spectre-v1 in do_get_thread_area()

[Dianzhang Chen]

The idx in do_get_thread_area() is controlled by userspace via syscall: ptrace(defined in kernel/ptrace.c), hence leading to a potential exploitation of the Spectre variant 1 vulnerability.
The idx can be controlled from: ptrace -> arch_ptrace -> do_get_thread_area.

Fix this by sanitizing idx before using it to index p->thread.tls_array.

Signed-off-by: Dianzhang Chen <dianzhangchen0@gmail.com>
---
 arch/x86/kernel/tls.c | 9 +++++++--
 1 file changed, 7 insertions(+), 2 deletions(-)

diff --git a/arch/x86/kernel/tls.c b/arch/x86/kernel/tls.c
index a5b802a..e3dc05b 100644
--- a/arch/x86/kernel/tls.c
+++ b/arch/x86/kernel/tls.c
@@ -5,6 +5,7 @@
 #include <linux/user.h>
 #include <linux/regset.h>
 #include <linux/syscalls.h>
+#include <linux/nospec.h>
 
 #include <linux/uaccess.h>
 #include <asm/desc.h>
@@ -220,15 +221,19 @@ int do_get_thread_area(struct task_struct *p, int idx,
 		       struct user_desc __user *u_info)
 {
 	struct user_desc info;
+	int index = idx - GDT_ENTRY_TLS_MIN;
 
 	if (idx == -1 && get_user(idx, &u_info->entry_number))
 		return -EFAULT;
 
-	if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
+	if (index < 0 || index > GDT_ENTRY_TLS_MAX - GDT_ENTRY_TLS_MIN)
 		return -EINVAL;
 
+	index = array_index_nospec(index,
+				GDT_ENTRY_TLS_MAX - GDT_ENTRY_TLS_MIN + 1);
+
 	fill_user_desc(&info, idx,
-		       &p->thread.tls_array[idx - GDT_ENTRY_TLS_MIN]);
+		       &p->thread.tls_array[index]);
 
 	if (copy_to_user(u_info, &info, sizeof(info)))
 		return -EFAULT;
-- 
2.7.4

Re: [PATCH] x86: fix possible spectre-v1 in do_get_thread_area()

[Thomas Gleixner]

On Fri, 24 May 2019, Dianzhang Chen wrote:

> The idx in do_get_thread_area() is controlled by userspace via syscall: ptrace(defined in kernel/ptrace.c), hence leading to a potential exploitation of the Spectre variant 1 vulnerability.
> The idx can be controlled from: ptrace -> arch_ptrace -> do_get_thread_area.
> 
> Fix this by sanitizing idx before using it to index p->thread.tls_array.
> 
> Signed-off-by: Dianzhang Chen <dianzhangchen0@gmail.com>
> ---
>  arch/x86/kernel/tls.c | 9 +++++++--
>  1 file changed, 7 insertions(+), 2 deletions(-)
> 
> diff --git a/arch/x86/kernel/tls.c b/arch/x86/kernel/tls.c
> index a5b802a..e3dc05b 100644
> --- a/arch/x86/kernel/tls.c
> +++ b/arch/x86/kernel/tls.c
> @@ -5,6 +5,7 @@
>  #include <linux/user.h>
>  #include <linux/regset.h>
>  #include <linux/syscalls.h>
> +#include <linux/nospec.h>
>  
>  #include <linux/uaccess.h>
>  #include <asm/desc.h>
> @@ -220,15 +221,19 @@ int do_get_thread_area(struct task_struct *p, int idx,
>  		       struct user_desc __user *u_info)
>  {
>  	struct user_desc info;
> +	int index = idx - GDT_ENTRY_TLS_MIN;
>  
>  	if (idx == -1 && get_user(idx, &u_info->entry_number))
>  		return -EFAULT;

This is broken in case of idx == -1 because index is not reevaluated after
idx is copied from u_info. You have to calculate index _AFTER_ that.
  
> -	if (idx < GDT_ENTRY_TLS_MIN || idx > GDT_ENTRY_TLS_MAX)
> +	if (index < 0 || index > GDT_ENTRY_TLS_MAX - GDT_ENTRY_TLS_MIN)
>  		return -EINVAL;
>  
> +	index = array_index_nospec(index,
> +				GDT_ENTRY_TLS_MAX - GDT_ENTRY_TLS_MIN + 1);

What about defining the array size and using it here and in the sanity
check above?

> +
>  	fill_user_desc(&info, idx,
> -		       &p->thread.tls_array[idx - GDT_ENTRY_TLS_MIN]);
> +		       &p->thread.tls_array[index]);

Please get rid of the line break. The line now fits into 80 char.

Thanks,

	tglx