Re: The end of embedded Linux?

From: Nicolas Pitre (nico@cam.org)
Date: Mon Oct 07 2002 - 12:45:35 EST


On Mon, 7 Oct 2002, Mark Mielke wrote:

> On Mon, Oct 07, 2002 at 09:02:33AM -0700, David S. Miller wrote:
> > From: Nicolas Pitre <nico@cam.org>
> > Date: Mon, 7 Oct 2002 12:05:16 -0400 (EDT)
> > 2) Not inlining inb() and friend reduce the bloat but then you further
> > impact performances on CPUs which are generally many order of
> > magnitude slower than current desktop machines.
> > I don't buy this one. You are saying that the overhead of a procedure
> > call is larger than the overhead of going out over the I/O bus to
> > touch a device?
>
> I think the key phrase is 'further impact'.
>
> If anything, the procedure call increases latency.
>
> Although... I don't see why CONFIG_TINY wouldn't be able to decide whether
> inb() should be inlined or not...

Please don't mix up the issues.

The problems with inb() and friends as it stands in the embedded world right
now as to do with code cleanness not kernel image bloat. Nothing to be
solved with CONFIG_TINY. Please let's keep those issues separate.

Here's the IO macro issue: On some embedded platforms the IO bus is only 8
bit wide or only 16 bit wide, or address lines are shifted so registers
offsets are not the same, etc. All this because embedded platforms are
often using standard ISA peripheral chipsets since they can be easily glued
to any kind of bare buses or static memory banks.

The nice thing here is the fact that only by modifying inb() and friends you
can reuse most current kernel drivers without further modifications.
However the modifs to inb() are often different whether the peripheral in
question is wired to a static memory bank, to the PCMCIA space or onto some
expansion board via a CPLD or other weirdness some hardware designers are
pleased to come with. Hence no global inb() and friend tweaking is possible
without some performance hit by using a runtime fixup based on the address
passed to them.

We therefore end up with something that looks like this in each drivers for
which a fixup is needed:

#ifdef CONFIG_ASSABET_NEPONSET

/*
 * These functions allow us to handle IO addressing as we wish - this
 * ethernet controller can be connected to a variety of busses. Some
 * busses do not support 16 bit or 32 bit transfers. --rmk
 */
static inline u8 smc_inb(u_int base, u_int reg)
{
        u_int port = base + reg * 4;

        return readb(port);
}

static inline u16 smc_inw(u_int base, u_int reg)
{
        u_int port = base + reg * 4;

        return readb(port) | readb(port + 4) << 8;
}

static inline void smc_outb(u8 val, u_int base, u_int reg)
{
        u_int port = base + reg * 4;

        writeb(val, port);
}

static inline void smc_outw(u16 val, u_int base, u_int reg)
{
        u_int port = base + reg * 4;

        writeb(val, port);
        writeb(val >> 8, port + 4);
}

#endif

As you can see such code duplicated multiple times for all bus arrangements
in existence out there is just not pretty and was refused by Alan. We lack
a global lightweight IO abstraction to nicely override the default IO macros
for individual drivers at compile time to fix that problem optimally and
keep the driver proper clean.

Nicolas

-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/



This archive was generated by hypermail 2b29 : Mon Oct 07 2002 - 22:00:59 EST