Re: [PATCH] arm: Remove IO memcpy for Big-Endian

[Linus Walleij]

On Tue, Dec 3, 2024 at 11:08 AM Arnd Bergmann <arnd@xxxxxxxx> wrote:

> - There is one ARMv5 "BE32" based platform, the ixp4xx, which
>   works differently, and this in turn allows multiple configurations
>   for its buses where a byte-swap is performed in the PCI
>   controller.

Actually it affects all memory-mapped I/O. E.g.
drivers/irqchip/irq-ixp4xx.c
for this reason IXP4xx drivers that are not PCI usually
use __raw_readl() and __raw_writel() to let the bus
do its thing.

IXP4xx PCI drivers are however *not* doing this, because
PCI is (IIUC) specified to be little-endian, and the hardware
deals with this. (Except for the PCI host driver itself, using
__raw accesses.)

> Before we apply your patch, I think the minimum would be to
> have Linus Walleij try it out on an an ixp4xx with a driver
> that uses these functions.

I applied the patch, compiled v6.13-rc1 HEAD for the
Linksys NSLU2 (oldest system there is! also a classic!)

+ Boots cleanly
+ Mounts root - this is on a USB drive which is on a
   PCI-attached USB controller, so PCI is tested
+ Reaches userspace and console
+ Gets network and IP address
+ LuCI web user inteface works (browsed to web server on device)
+ Dropbear SSH login works fine

Tested-by: Linus Walleij <linus.walleij@xxxxxxxxxx>

I doubt any IXP4xx driver uses ioread/iowrite. The
MTD driver does not for example, because of other weirdness
with address swizzling (drivers/mtd/maps/physmap-ixp4xx.c)
For similar reasons drivers/ata/pata_ixp4xx_cf.c is
using readw/writew exclusively.

It also looks like a good idea:
Reviewed-by: Linus Walleij <linus.walleij@xxxxxxxxxx>

Your,
Linus Walleij