Re: [PATCH v1 12/14] x86/boot: tweak a20.c for better code generation

[David Laight <david.laight.linux@gmail.com>]

On Fri, 23 Jan 2026 20:24:55 -0800
"H. Peter Anvin" <hpa@zytor.com> wrote:

> On 2026-01-23 19:00, Maciej W. Rozycki wrote:
> > On Wed, 21 Jan 2026, David Laight wrote:
> >   
> >> No loops needed.  
> > 
> >  A loop is needed because there can be a considerable delay from issuing 
> > the I/O request to flip the A20 gate till the circuitry responding.  This 
> > is particularly true with the command issued to the 8042 device, which is 
> > a microcontroller running its own firmware that needs it time to process 
> > an incoming request to drive one of the microcontroller's GPIOs.  There 
> > was a reason for port 0x92 circuitry later added to the PC architecture 
> > with the IBM PS/2 being called the "fast A20 gate".
> >   
> 
> Indeed. I thought I had responded to this already but I hadn't, apparently.
> 
> Note that the "long" delay is 2^21 loops! That number wasn't taken out of the
> air, either; we found machines that actually needed that many iterations.

Ok, so you need a loop because it might take ages for the value read from
0x1000200 to change.
But there is no need to keep changing the value.
The comments in the code don't really stress that.

> In the case where A20 is enabled already, the loop terminates on either the
> first or second iteration (the second iteration is when the value at 0x1000200
> is exactly 1 higher than the value at 0x200.
> 
> Modern machines (Nehalem+) already have A20 enabled, and most machines of the
> i686+ generation implement int 0x15 function 0x2401.

I know some of the history.
And just read some more of the gory details...

A20 being disabled is there to make a 286 compatible with the older 8086 PCs
and any software that relied on address wrapping (rather than using it to get
an extra ~64kB in real mode).
That would be for dos and win 3.11...

The only 8088 and 286 cpu I used were on IO cards.

> 
> 	-hpa
>