Re: [patch 1/4] Kprobes support for IA64

[Keshavamurthy Anil S <anil.s.keshavamurthy@intel.com>]

On Thu, May 26, 2005 at 10:49:02AM +1000, Keith Owens wrote:
> On Wed, 25 May 2005 10:00:18 -0400,
> "Alan D. Brunelle" <Alan.Brunelle@hp.com> wrote:
> >Isn't the real issue here that if kprobes attempts to put in a 'break
> >0x80200' into a B-slot that it instead becomes a 'break.b 0' -- as the
> >break.b does not accept an immediate value?
> 
> break.b is a B9 type instruction, which does take an imm21 value.  It
> is the hardware that does not store imm21 in CR.IIM when break.b is
> issued.
> 
> >Kprobes does have the two cases covered in traps.c (case 0 - when a
> >B-slot break is used, and case 0x80200 for a non-B-slot break). But this
> >doesn't seem very clean. (If it was decided that one should not overload
> >the break 0 case, and instead use a uniquely defined break number, then
> >it fails on a B-slot probe. If it is OK to overload the break 0 case,
> >why have another break number at all?)
> 
> Mainly for documentation when looking at the assembler code.  break 0
> is used for BUG(), coding a different value in the break instruction
> for the debugger helps the person debugging the debugger :(.  I have no
> problem with coding two cases in ia64_bad_break() in order to work
> around the hardware "feature".

I agree with Keith, when a person taking a instructin dump, a 
different value will help uniquely identify that this
is a kprobe break instruction which is a replaced instrucion
of the original instruction. So we will leave with what we have
now, i.e handle the same with two cases.


> 
> Also consider the case where your debugger allows users to code a
> deliberate entry to the debugger, like KDB_ENTER().  That case always
> requires a separate break imm21 value, because the break point is not
> known to the debugger until the code is executed.
>