Re: regression from your recent change to x86's copy_user_handle_tail()

["Jan Beulich" <JBeulich@suse.com>]

>>> On 23.04.15 at 17:33, <torvalds@linux-foundation.org> wrote:
> On Tue, Apr 21, 2015 at 11:33 PM, Jan Beulich <JBeulich@suse.com> wrote:
>>
>> while the description of commit cae2a173fe certainly makes sense, the
>> change itself ignores the __probe_kernel_write() code path, for which
>> the destination address is expected to be in kernel space but accesses
>> may still fault. I.e. the use of plain memset() causes
>> __probe_kernel_write() to oops rather than return an error. Shouldn't
>> the "(unsigned long)to >= TASK_SIZE_MAX" be relaxed to take the
>> effect of set_fs() into account?
> 
> Hmm. I think you're right.  So something like
> 
>     --- a/arch/x86/lib/usercopy_64.c
>     +++ b/arch/x86/lib/usercopy_64.c
>     @@ -82,7 +82,7 @@ copy_user_handle_tail(char *to, char *from, unsigned len)
>             clac();
> 
>             /* If the destination is a kernel buffer, we always clear the end */
>     -       if ((unsigned long)to >= TASK_SIZE_MAX)
>     +       if (!__addr_ok(to))
>                     memset(to, 0, len);
>             return len;
>      }
> 
> which will effectively say "only if we copy from user mode to kernel
> mode" because if we use "set_fs(KERNEL_DS)" then kernel addresses will
> also be __addr_ok..
> 
> Did you have a test-case for this? I guess we're talking odd ftrace
> uses or kgdb?

So I thought you meant something you could try. The above fixes
the issue for me, i.e.

Tested-by: Jan Beulich <jbeulich@suse.com>

Thanks, Jan