Re: [PATCH v3 4/8] x86: speed up int3-based patching using direct write

[Petr Mladek <pmladek@suse.cz>]

Masami Hiramatsu píše v Čt 14. 11. 2013 v 20:36 +0900:
> (2013/11/14 19:41), Petr Mladek wrote:
> > @@ -837,7 +882,7 @@ static void recover_iter(struct text_poke_bp_iter *iterator,
> >  	}
> >  
> >  	/* Finally, put back the first byte from the old code */
> > -	err = text_poke(addr, old_opcode, sizeof(bp_int3));
> > +	err = text_poke_part(addr, old_opcode, sizeof(bp_int3));
> >  	/* we can not continue if the interrupt is still there */
> >  	BUG_ON(err);
> >  }
> > @@ -865,7 +910,9 @@ static void recover_iter(struct text_poke_bp_iter *iterator,
> >   * It is a bit more paranoid than text_poke_bp because it checks the actual
> >   * code before patching. This is a good practice proposed by the ftrace code.
> >   *
> > - * Note: This function must be called under text_mutex.
> > + * Note: This function must be called under text_mutex. Also the caller is
> > + * responsible for making the patched code read-write, for example using
> > + * set_kernel_text_rw() and set_all_modules_text_rw()
> 
> OK, I got the reason why we can do it.
> But if so, we should change all existing caller of text_poke_bp() to
> call it in this patch, or they must fail at this point.

Ah, I used "text_poke_part" only in "text_poke_bp_iter" in the end. This
function is currently used only in ftrace that sets the code read-write
in "ftrace_arch_code_modify_prepare", see arch/x86/kernel/ftrace.c.

Note that "text_poke_bp" still uses "text_poke". It seems that the
remapping is faster if you patch only one address.

So, we should be on the safe side.

Thanks a lot for review.

Best Regards,
Petr