Re: [PATCH v6 4/4] arm64/syscalls: Specific usage of verify_pre_usermode_state

[Catalin Marinas]

On Wed, Apr 05, 2017 at 11:14:34AM -0700, Thomas Garnier wrote:
> On Wed, Apr 5, 2017 at 10:49 AM, Catalin Marinas
> <catalin.marinas@xxxxxxx> wrote:
> > On Wed, Apr 05, 2017 at 07:36:17AM -0700, Thomas Garnier wrote:
> >> On Wed, Apr 5, 2017 at 7:22 AM, Catalin Marinas <catalin.marinas@xxxxxxx> wrote:
> >> > On Tue, Apr 04, 2017 at 10:47:27AM -0700, Thomas Garnier wrote:
> >> >> +
> >> >>       ldr     x1, [tsk, #TSK_TI_FLAGS]
> >> >>       and     x2, x1, #_TIF_WORK_MASK
> >> >>       cbnz    x2, work_pending
> >> >> @@ -779,6 +788,12 @@ finish_ret_to_user:
> >> >>       kernel_exit 0
> >> >>  ENDPROC(ret_to_user)
> >> >>
> >> >> +addr_limit_fail:
> >> >> +     stp     x0, lr, [sp,#-16]!
> >> >> +     bl      asm_verify_pre_usermode_state
> >> >> +     ldp     x0, lr, [sp],#16
> >> >> +     ret     lr
> >> >
> >> > Where is this supposed to return? What is the value of lr when branching
> >> > to addr_limit_fail?
> >>
> >> It is not supposed to return. Do you think I should remove stp, ldp,
> >> ret and jut add a brk 0x100 or jmp/call a break/bug function?
> >
> > Can you not just make addr_limit_fail a C function which never returns
> > (similar to what we to with bad_mode() on arm64)? Since addr_limit_fail
> > is only called when the segment is not the right one, I don't really see
> > why you need another call to asm_verify_pre_usermode_state() to do a
> > similar check again. Just panic in addr_limit_fail (unless I
> > misunderstood what you are trying to achieve).
> 
> Calling asm_verify_pre_usermode_state has the advantage of having a
> clear BUG_ON for the error (versus a panic description).
> 
> What do you think about replacing asm_verify_pre_usermode_state by a
> "address_limit_fail" function that still calls
> verify_pre_usermode_state but panic afterwards (because it should
> never return)?
> 
> The assembly code would be easier to understand and in case of error
> the BUG_ON is clear for the user.

It looks fine to me, though I'd have to see the patch.

-- 
Catalin