Re: [PATCH 2/2] x86/ldt: allow to disable modify_ldt at runtime

[Willy Tarreau <w@1wt.eu>]

On Mon, Aug 03, 2015 at 11:45:24AM -0700, Andy Lutomirski wrote:
> I'm not entirely convinced that the lock bit should work this way.  At
> some point, we might want a setting for "32-bit only" or even "32-bit,
> present, not non-conforming only" (like we do unconditionally for
> set_thread_area).  When we do that, having -1 act like 0 might be
> confusing.
> 
> I'd actually favor rigging it up to support enumerated values and/or
> the word "locked" somewhere in the text.  So we could have "0", "1
> locked", "1" or even "enabled" "enabled locked", "disabled", "disabled
> locked", "safe 32-bit", "safe 32-bit locked", etc.

Got it, that makes sense indeed. I asked myself whether we'd use more
than these 3 values, and estimated that "locked on" didn't make much
sense here, and I thought that nobody would like to manipulate such
things using bitmaps. But with words like this it can indeed make
sense.

I feel like it's probably part of a larger project then. Do you think
we should step back and only support 0/1 for now ? I also have the
patch available.

> I'll add an explicit 16-bit check to my infinite todo list for the asm
> part.  Now that the synchronous modify_ldt code is merged, it won't be
> racy, and it would make a 32-bit only mode actually be useful (except
> maybe on AMD -- someone needs to test just how badly broken IRET is on
> AMD systems -- I know that AMD has IRET-to-16-bit differently broken
> from Intel, and that causes test-cast failures.  Grump.)
> 
> P.S. Hey CPU vendors: please consider stopping your utter suckage when
> it comes to critical system instructions.  Intel and AMD both
> terminally screwed up IRET in multiple ways that clearly took actual
> effort.  Intel screwed up SYSRET pretty badly (AFAIK every single
> 64-bit OS has had at least one root hole as a result), and AMD screwed
> SYSRET up differently (userspace crash bug that requires a performance
> hit to mitigate because no one at AMD realized that one might preempt
> a process during a syscall).

Well the good thing is that SYSRET reused the LOADALL opcode so at
least this one cannot be screwed on 64-bit :-) It would have helped us
to emulate IRET though.

> P.P.S. You know what would be *way* better than allowing IRET to
> fault?  Just allow IRET to continue executing the next instruction on
> failure (I'm talking about #GP, #NP, and #SS here, not page faults).
> 
> P.P.P.S.  Who thought that IRET faults unmasking NMIs made any sense
> whatsoever when NMIs run on an IST stack?  Seriously, people?

A dedicated flag "don't clear NMI yet" would have been nice in EFLAGS
so that the software stack could set it in fault handlers. It would be
one-shot and always cleared by IRET. That would have been very handy.

Willy