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

[Denys Vlasenko <vda.linux@googlemail.com>]

On Saturday 21 January 2012 00:49, Indan Zupancic wrote:
> On Fri, January 20, 2012 23:41, H. Peter Anvin wrote:
> > On 01/20/2012 02:40 PM, Roland McGrath wrote:
> >> If you change the size of a regset, then the new full size will be the size
> >> of the core file notes.  Existing userland tools will not be expecting
> >> this, they expect a known exact size.  If you need to add new stuff, it
> >> really is easier all around to add a new regset flavor.  When adding a new
> >> one, you can make it variable-sized from the start so as to be extensible
> >> in the future.  We did this for NT_X86_XSTATE, for example.
> >>
> >
> > Yes, that definitely seems cleaner.
> 
> I would prefer Linus' way of just stuffing it into cs. Jamie also wanted
> a bit telling in what mode the userspace is running. That's 3 bits in total,
> with one bit telling whether the other bits are valid or not. Anything else?

There is actually a bunch of ptrace-specific stuff we want to return.

For example, Oleg wants to be able to print *which syscall*,
(along with its arguments if possible) is restarted when
we restart the ERESTART_RESTARTBLOCK-returning syscall.
Which happens every time strace attaches to a process sleeping
in nanosleep or poll, for example. We get just

$ strace -p 1234
Process 1234 attached - interrupt to quit
restart_syscall(<... resuming interrupted call ...>_

and that's it.

Returning syscall and its parameters require several words,
not a few bits.

> Maybe a bit telling whether it is syscall entry or exit?

Yes, this one too. This is one of longstanding annoyances
that this information is not exposed.

> As all this is very x86_64 specific and cs is already used to figure out
> the mode, it seems overkill to add a new regset just for this.
> 
> It's a lot easier for existing code to add an extra cs check than to use
> different register sets and different ptrace commands.

You don't understand. Returning new bits in cs will break *existing*
programs. This is generally a bad thing. For example, old strace binaries
on new kernel will complain:

        switch (x86_64_regs.cs) {
                case 0x23: currpers = 1; break;
                case 0x33: currpers = 0; break;
                default:
                        fprintf(stderr, "Unknown value CS=0x%08X while "
                                 "detecting personality of process "
                                 "PID=%d\n", (int)x86_64_regs.cs, tcp->pid);
                        currpers = current_personality;
                        break;
        }

when they'll see unfamiliar x86_64_regs.cs value.

-- 
vda