Re: modules: per-segment policy

[Maxime Villard <max%m00nbsd.net@localhost>]

Le 08/06/2016 à 08:57, Martin Husemann a écrit :
> On Wed, Jun 08, 2016 at 05:57:47AM +0800, Paul Goyette wrote:
> > All of this sounds like a good idea.  however, there are at least one or
> > two PRs out there related to limits on the amount of address space that
> > modules can occupy, on i386.

Where?

> > It would seem to me that taking each
> > module and mapping it in multipple page-sized pieces would end up taking
> > more total space (due to page-rounding) and possibly exacerbate the
> > current situation.

With three different chunks, we lose at most 12285 bytes in padding. No one
cares about losing such a ridiculously small amount of memory. There are
many places in the kernel that could be optimized to save up much more
space.

> How about creating a vmmap for module_text and another for module_data,
>  guessing the size split (upfront, statically) from statisitics in
> /stand/modules ?

You are not going to guess it properly. A driver could have a lot of firmware
code in data and only few instructions; a syscall could have a lot of
instructions and no data at all. I don't think this even reduces memory usage,
since you also need to allocate memory for the map headers and other things.