> Date: Mon, 22 Feb 2010 13:41:07 -0800 > From: "H.J. Lu" > > On Mon, Feb 22, 2010 at 1:30 PM, Mark Kettenis wrote: > >> Date: Mon, 22 Feb 2010 13:06:31 -0800 > >> From: "H.J. Lu" > >> > >> On Mon, Feb 22, 2010 at 11:50 AM, Mark Kettenis wrote: > >> >> Date: Mon, 22 Feb 2010 12:03:03 -0500 > >> >> From: Daniel Jacobowitz > >> >> > >> >> On Mon, Feb 22, 2010 at 05:56:58PM +0100, Mark Kettenis wrote: > >> >> > I've looked at the Linux kernel sources for the kernel on my > >> >> > workstation (2.6.27 in its OpenSUSE incarnation), and the only way to > >> >> > distinguish between a 32-bit and a 64-bit process seems to be to > >> >> > attempt to write one of the debug address registers with a value > >> >> > that's larger than 0xffffffff.  If that fails, you have a 32-bit > >> >> > process, otherwise it's a 64-bit process. > >> >> > >> >> Yuck :-(  But I didn't see anything else either. > >> > > >> > Indeed. > >> > > >> >> Is there an eflags bit for this?  Even if so, IIRC, we may not want to > >> >> use it; it's possible to run 32-bit code in a 64-bit process and some > >> >> overly clever programs may do so. > >> > > >> > Nope, there is no %eflags/%rflags bit for this.  Not quite sure what > >> > running 32-bit code in a 64-bit process actually means.  But I'd guess > >> > you want the 64-bit view on the registers in that case. > >> > > >> > Anyway, I think it's probably best if HJ leaves this bit out of this > >> > diff for now.  We can revisit the issue when AVX support is > >> > introduced. > >> > > >> > >> Please see if my latest patch is OK: > >> > >> --- > >>  /* Get CS register.  */ > >>   errno = 0; > >>   cs = ptrace (PTRACE_PEEKUSER, tid, > >>                offsetof (struct user_regs_struct, cs), 0); > >>   if (errno != 0) > >>     perror_with_name (_("Couldn't get CS register")); > >> > >>   /* Value of CS register: > >>      1. 64bit: 0x33. > >>      2. 32bit: 0x23. > >>    */ > >>   if (cs == 0x33) > >>     return tdesc_amd64_linux; > >>   else > >>     return tdesc_i386_linux; > >> --- > >> > >> In kernel, there is > >> > >>       regs->cs = test_thread_flag(TIF_64BIT_ILP32) ? __USER_CS : __USER32_CS; > > > > I fear that's rather fragile.  I mean, the actual value of > > __USER_CS/__USER32_CS is just an implementation detail isn't it? > > > > That is how strace checks 32bit process on Linux/x86-64 I have > discussed it with Peter and Suresh. It is very unlikely Linux kernel will > break strace. In any case, we will add a new ptrace option to Linux > 2.6.35 to get TIF_64BIT_ILP32 among other things. I will update gdb > to try the new ptrace option first and then fail back to CS register. OK, that makes it less fragile than I thought. But please use #defines with meaningful names instead of magic constants.