[ILUG] linux device drivers: port or memory mapped i/o?

[Paul Jakma]

ok the 'whole lot faster' bit is an exaggeration. :)

On Tue, 16 Oct 2001, Padraig Brady wrote:

> Well this depends. For large wads of data the compiler/CPU should
> do a better job @ optimising memory access instructions as
> apposed to inb etc?

i guess so.

but i think the best reason is that processors and chipsets are
geared/optimised towards memory access.

also, just using an IO space IO instruction must surely have some
overhead compared to a programme that just solely reads and writes
away to various memory addresses. (i'm just guessing). else why is it
that some other arches just never bothered with special port IO
instructions?

finally, as most hardware (decent, ie PCI) has been mmap'ed IO, i'll 
bet chipset designers don't do much work on port IO beyond "ok, it 
works".

> Also for hardware compatibility reasons I think, most ports in
> the 0-0x3ff range take almost exactly 1 microsecond to access.

i'll still bet mmap IO is faster. :)

also, might be wrong, but on x86 port IO only gets you to the top of
ISA address space, no? eg, aren't there only 65536 IO ports on x86?
(i'm probably wrong though - dont tell kevin!).

> This has good linux specific info:
> http://www.xml.com/ldd/chapter/book/ch08.html

hmm...

> Padraig.

regards,
-- 
Paul Jakma	paul at clubi.ie	paul at jakma.org	Key ID: 64A2FF6A
Fortune:
If mathematically you end up with the wrong answer, try multiplying by
the page number.