On Fri, Oct 12, 2007 at 06:52:10PM +0100, Ralf Baechle wrote:
> On Fri, Oct 12, 2007 at 07:19:38PM +0200, Adrian Bunk wrote:
>
> > I have an objection against this approach:
> >
> > Our __*init*/__*exit* annotations are already a constant source of bugs,
> > and adding more pifalls (e.g. forgotten removal of _i()/_e() when a
> > function is no longer __*init*/__*exit*) doesn't sound like a good plan.
> >
> > Shouldn't it be possible to automatically determine where to put the
> > strings? I don't know enough gcc/ld voodoo for being able to tell
> > whether it is currently possible, and if yes how, but doing it
> > automatically sounds like the only solution that wouldn't result in an
> > unmaintainable mess.
>
> gcc tends to place data such as strings or jump tables generated from
> switches not into a place were it would be easily discardable. The
> latter is the reason that on MIPS we can't discard __exit functions
> at all - a switch table in .rodata might be referencing discarded code
> in .exit.text which makes ld fail. When I discussed this with some gcc
> people a while ago nobody really had a good suggestion to solve this.
- Most of the string annotations are (naturally) dev{init,exit}
annotations, and bugs there are therefore in configurations that have
only extremely low testing coverage during -rc.
- I'm counting 22 annotations in the driver Maciej converted as an
example. When estimating the number of possible annotations based
on the number of C files in the kernel I'm getting a six digit
number.
No matter how hard it would be to teach gcc about it, when thinking of
the amount of __*init*/__*exit* bugs we already have I simply can't
imagine the string annotations as a maintainable solution.
> Ralf
cu
Adrian
--
"Is there not promise of rain?" Ling Tan asked suddenly out
of the darkness. There had been need of rain for many days.
"Only a promise," Lao Er said.
Pearl S. Buck - Dragon Seed
