unaligned exception handling

unaligned exception handling

[Kip Walker]

After inspecting a strange case in the mips64 kernel with address
errors, I'm starting to think there's a problem in the do_ade()
implementation.  I think the 32-bit kernel may have a similar problem,
but I haven't really inspected it.  The issue is where the kernel's
emulation of an address error causes another address error (NOT a page
fault).

Basically, I don't see how the exception table stuff in
emulate_load_store_insn is going to work.  Consider this scenario:

- user process does a 'sw' (for example) to an illegal address
  above xuseg but below xsseg
- do_ade calls emulate_load_store_insn, which tries swl/swr
- the swl again hits an illegal address, this time in the
  kernel's context
- do_ade does NOT check the exception table for the swl
- emulate_load_store_insn goes to the 'swl' part of the switch
- die_if_kernel DOES __die before the SIGBUS is delivered.

So I don't see how the ex_table stuff is useful at all.  Shouldn't
do_ade() do the exception table grovelling before calling
emulate_load_store_insn?

Kip

Re: unaligned exception handling

[Carsten Langgaard]

You are right there is a potential problem in the 64-bit kernel, when
emulating load/store instructions.
I have previously addressed this problem (and also send a patch).

The problem is the check_axs macro, that checks the access before
actually doing the emulation (any load/stores).
It's a simple copy from the 32-bit kernel, where you do a check for user
space by simple check the most significant bit.
That's fine in the 32-bit world, but it's not sufficient in the 64-bit
case.

Locally I have a check_axs that look like this:

/*
 * User code may only access USEG;
 * Kernel code may access the entire address space, except the area
between
 * USEG (XUSEG) and KSEG0.
 */
#define check_axs(pc,a,s)      \
        if (((pc < KUSIZE) && (((a) | ((a)+(s))) >= KUSIZE)) ||  \
     ((((a) | ((a)+(s))) < K0BASE) &&    \
      (((a) | ((a)+(s))) >= KUSIZE)))    \
  goto sigbus;



Hope that helps you.
/Carsten



Kip Walker wrote:

> After inspecting a strange case in the mips64 kernel with address
> errors, I'm starting to think there's a problem in the do_ade()
> implementation.  I think the 32-bit kernel may have a similar problem,
> but I haven't really inspected it.  The issue is where the kernel's
> emulation of an address error causes another address error (NOT a page
> fault).
>
> Basically, I don't see how the exception table stuff in
> emulate_load_store_insn is going to work.  Consider this scenario:
>
> - user process does a 'sw' (for example) to an illegal address
>   above xuseg but below xsseg
> - do_ade calls emulate_load_store_insn, which tries swl/swr
> - the swl again hits an illegal address, this time in the
>   kernel's context
> - do_ade does NOT check the exception table for the swl
> - emulate_load_store_insn goes to the 'swl' part of the switch
> - die_if_kernel DOES __die before the SIGBUS is delivered.
>
> So I don't see how the ex_table stuff is useful at all.  Shouldn't
> do_ade() do the exception table grovelling before calling
> emulate_load_store_insn?
>
> Kip

--
_    _ ____  ___   Carsten Langgaard   Mailto:carstenl@mips.com
|\  /|||___)(___   MIPS Denmark        Direct: +45 4486 5527
| \/ |||    ____)  Lautrupvang 4B      Switch: +45 4486 5555
  TECHNOLOGIES     2750 Ballerup       Fax...: +45 4486 5556
                   Denmark             http://www.mips.com