[PATCH] fixing userspace memory dereference in security.c

[Serge E. Hallyn]

On Wed, Oct 06, 2021 at 07:03:56PM -0400, T. Williams wrote:
>  security/security.c | 2 ++
>  1 file changed, 2 insertions(+)
> 
> diff --git a/security/security.c b/security/security.c
> index 9ffa9e9c5c55..7c41b5d732ab 100644
> --- a/security/security.c
> +++ b/security/security.c
> @@ -1737,6 +1737,8 @@ int security_kernel_read_file(struct file *file, enum
> kernel_read_file_id id,
>         int ret;
> 
>         ret = call_int_hook(kernel_read_file, 0, file, id, contents);
> +       if (ret > 0)
> +               return -EINVAL;
>         if (ret)
>                 return ret;
>         return ima_read_file(file, id, contents);
> -- 
> 2.25.1
> 
> This commit is to fix a userspace address dereference found by
> syzkaller.
> The crash is triggered by passing a file descriptor to an incorrectly
> formed kernel module to finit_module.
> 
> Kernel/module.c:4175 : Within the finit_module, info.len is set to the
> return value from kernel_read_file_from_fd. This value is then
> dereferenced by memcmp within module_sig_check from inside load_module.
> The value is 0xb000000 so the kernel dereferences user memory and kernel
> panics.

Hi,

thanks for sending this.  For some reason, I can't seem to find this
message-id in lore.kernel.org to see if there were ever any replies.

There is indeed a problem, although I think a more concise explanation
is:

1. security_kernel_read_file() returns any non-zero return value to mean
permission denied
2. kernel_read_file() returns > 0 meaning number of bytes read
3. hen kernel_read_file() gets any non-zero rv from security_kernel_read_file(),
it returns that value unchanged.

Since kernel_read_file() is the only caller of security_kernel_read_file(),
I think your patch is good, except you should also change the comment above
it to read

 * Return: Returns 0 if permission is granted, < 0 on error.

Reviewed-by: Serge Hallyn <serge at hallyn.com>

I think the reason it's not been a practical problem is because while
security_kernel_read_file() will honor a >0 error from an LSM, no
LSM implementation of that hook does that.  (Only loadpin and selinux
implement it)

-serge