Re: [RFC][PATCH] kprobes: kprobe-booster for ia64

[Masami Hiramatsu <mhiramat@redhat.com>]

Hello Tony,

Thank you for review.
Luck, Tony wrote:
> +/* Insert a long branch code */
> +static void __kprobes set_brl_inst(void *from, void *to)
> +{
> +	s64 rel = ((s64) to - (s64) from) >> 4;
> +	bundle_t *brl;
> +	brl = (bundle_t *) ((u64) from & ~0xf);
> +	brl->quad0.template = 0x05;	/* [MLX](stop) */
> +	brl->quad0.slot0 = NOP_M_INST;	/* nop.m 0x0 */
> +	brl->quad0.slot1_p0 = ((rel >> 20) & 0x7fffffffff) << 2;
> +	brl->quad1.slot1_p1 = (((rel >> 20) & 0x7fffffffff) << 2) >> (64 - 46);
> +	/* brl.cond.sptk.many.clr rel<<4 (qp=0) */
> +	brl->quad1.slot2 = BRL_INST(rel >> 59, rel & 0xfffff);
> +}
> 
> Is this function operating directly on the instruction bundle in the
> kernel code?  If so, the bundle will contain some "interesting"
> intermediate values as each of the "brl->xxx = yyy;" are processed.
> If another cpu should try to execute this bundle while the modifications
> are in progress, then the results will be not what you want.

No, this function writes brl on kprobe's instruction buffer
(kprobe.ainsn.insn->bundle[1]) when preparing the kprobe.
So, other cpu never execute it, and we don't need to write it atomically.

> On cpu models that support st16 you could construct the brl instruction
> and use a single st16 to update the whole bundle atomically.
> 
> -Tony

Thank you again,

-- 
Masami Hiramatsu

Software Engineer
Hitachi Computer Products (America) Inc.
Software Solutions Division

e-mail: mhiramat@redhat.com