[linuxsh-dev] Re: bad alignment fixup

[Philipp R. <pr...@pa...>]

On Mon, Nov 20, 2000 at 03:27:01PM +0100, Jesper Skov wrote:
> >>>>> "David" == David Woodhouse <dw...@in...> writes:
> David> You mean to search the exception table before attempting the
> David> fixup, and to refuse to do the fixup if there's an entry for
> David> $pc? I don't like that very much. I think we _should_ be fixing
> David> up unaligned userspace accesses. Not only when they're accessed
> David> from the kernel, but when they happen entirely in userspace.
> 
> I don't agree. Nasty DOS you could get out of that - simply make an
> infinite loop read from odd addresses and the kernel will be wasting
> tons of time fixing up those accesses. User space apps should die if
> they do unaligned access.

AIUI, the main cost of emulating unaligned accesses is the user->kernel->
user context switch.  user->kernel->SIGBUS handler->user shouldn't be
significantly more expensive but avoid all security holes I know about.

> David> At the moment, I'm more concerned about what happens when the
> David> offending insn is in a branch delay slot. It looks like
> David> regs->pc will be pointing to the branch insn iff it was taken,
> David> so in that case, we have to fix up the insn at $pc+2 and then
> David> also emulate the branch insn follow the branch.
> 
> Urgh. Messy.

delay slots are :/