Linux Virtualization - Nov 2014 - [PATCH v4 10/10] x86: Support compiling out userspace IO (iopl and ioperm)

On Sun, 2 Nov 2014 09:33:01 -0800
Josh Triplett <josh at joshtriplett.org> wrote:
> On the vast majority of modern systems, no processes will use the
> userspsace IO syscalls, iopl and ioperm.  Add a new config option,
> CONFIG_X86_IOPORT, to support configuring them out of the kernel
> entirely.  Most current systems do not run programs using these
> syscalls, so X86_IOPORT does not depend on EXPERT, though it does still
> default to y.
This isn't unreasonable but there are drivers with userspace helpers that
use iopl/ioperm type functionality where you should be doing a SELECT of
X86_IOPORT. The one that comes to mind is the uvesa driver. From a quick
scan it may these days be the only mainstream one that needs the select
adding.

Some X servers for legacy cards still use io port access. There are also
a couple of other highly non-obvious userspace users that hang on for
some systems - eg some older servers DMI and error records can only by
read via a real mode BIOS call so management tools have no choice but to
go the lrmi/io path.

Still makes sense IMHO.

From a code perspective however you could define IO_BITMAP_LONGS to 0,
add an IO_BITMAP_SIZE (defined as LONGS + 1 or 0) and as far as I can see
gcc would then optimise out a lot of the code you are ifdeffing

Alan

> > This isn't unreasonable but there are drivers with userspace
helpers that
> > use iopl/ioperm type functionality where you should be doing a SELECT
of
> > X86_IOPORT. The one that comes to mind is the uvesa driver. From a
quick
> > scan it may these days be the only mainstream one that needs the
select
> > adding.
> 
> Should kernel drivers really express dependencies that only their
> (current instances of) corresponding userspace components need?
> Something seems wrong about that.
uvesafb will always need X86_IOPORT. It's kind of implicit in the design.
I'm not suggesting that fbdev should select X86_IOPORT but in the uvesafb
case at least it's completely useless to have one and not the other.
> IO_BITMAP_LONGS already gets defined to (0/sizeof(long)).  And as far as
> I can tell, that would only work for init_tss_io, not anything else.
> Even then, that would only work with a zero-size array left around in
> tss_struct, which doesn't seem appropriate.  The remaining ifdefs wrap
> code that GCC could not constant-fold away, and making that code
> constant-foldable seems significantly more invasive than the ifdefs.
OK