Re: Compat 32-bit syscall entry from 64-bit task!?

[Will Drewry <wad@chromium.org>]

On Thu, Jan 19, 2012 at 1:21 PM, Linus Torvalds
<torvalds@linux-foundation.org> wrote:
> On Thu, Jan 19, 2012 at 8:01 AM, Jamie Lokier <jamie@shareable.org> wrote:
>> Andrew Lutomirski wrote:
>>> It's reasonable, obvious, and even more wrong than it appears.  On
>>> Xen, there's an extra 64-bit GDT entry, and it gets used by default.
>>> (I got bitten by this in some iteration of the vsyscall emulation
>>> patches -- see user_64bit_mode for the correct and
>>> unusable-from-user-mode way to do this.)
>>
>> Here it is:
>>
>>        static inline bool user_64bit_mode(struct pt_regs *regs)
>
> This is pointless, even if it worked, which it clearly doesn't on Xen
> (or other random situations).
>
> Why would you care?
>
> The issue is *not* whether somebody is running in 32-bit mode or 64-bit mode.
>
> The problem is the system call itself, and that can be 32-bit or
> 64-bit independently of the execution mode. So knowing the user-mode
> mode is simply not relevant.
>
> In the kernel, we know this with the TS_COMPAT flag - exactly because
> it's impossible to tell from any actual CPU state. So *that* is the
> flag you need to figure out, and currently the kernel doesn't export
> it any way (but my suggested patch would export it in the high bits of
> rflags).

Would it be worth considering changing the return from
task_user_regset_view, like:

--- a/arch/x86/kernel/ptrace.c
+++ b/arch/x86/kernel/ptrace.c
@@ -1311,7 +1311,11 @@ void update_regset_xstate_info(unsigned int
size, u64 xstate_mask)
 const struct user_regset_view *task_user_regset_view(struct task_struct *task)
 {
 #ifdef CONFIG_IA32_EMULATION
-       if (test_tsk_thread_flag(task, TIF_IA32))
+       /* If the task is in a syscall, then the TS_COMPAT status
+        * is more accurate than the personality.
+        */
+       if (test_tsk_thread_flag(task, TIF_IA32) ||
+           task_thread_info(task)->status & TS_COMPAT)
 #endif
 #if defined CONFIG_X86_32 || defined CONFIG_IA32_EMULATION
                return &user_x86_32_view;


This would make TS_COMPAT behave like a personality change.
PTRACE_POKEUSR and PEEKUSR would still access the 64-bit view with no
compat info (just like with TIF_IA32 tasks), but PTRACE_[GS]ETREGS
would return/expect 32-bit struct user_struct_regs.  This would result
in the tracer needing to check the returned regs to see if it was
fully populated (which seems heinous), but it would export the
TS_COMPAT state.

Right now, if a 64-bit tracer changes the regs for a TS_COMPAT call,
the args will be 32-bit truncated (for better or worse). Of course, on
trace_syscall_leave, 64-bit registers won't be truncated so it maybe
makes less sense.

Perhaps this was considered and discarded as being obviously broken,
but it wasn't clear cut to me.

Thanks!
will