Re: [PATCH v2 2/5] mm: avoid unnecessary flush on change_huge_pmd()

[Dave Hansen]

On 10/26/21 1:07 PM, Nadav Amit wrote:
> I just wonder how come the R/W-clearing and the P-clearing cause concurrent
> dirty bit setting to behave differently. I am not a hardware guy, but I would
> imagine they would be the same...

First of all, I think the non-atomic properties where a PTE can go:

	W=1,D=0 // original
	W=0,D=0 // software clears W
	W=0,D=1 // hardware sets D

were a total implementation accident.  It wasn't someone being clever
and since the behavior was architecturally allowed and well-tolerated by
software it was around for a while.  I think I was the one that asked
that it get fixed for shadow stacks, and nobody pushed back on it too
hard as far as I remember.  I don't think it was super hard to fix.

Why do the Present/Accessed and Write/Dirty pairs act differently?  I
think it's a total implementation accident and wasn't by design.

The KNL erratum was an erratum and wasn't codified in the architecture
because it actually broke things.  The pre-CET Write/Dirty behavior
didn't break software  to a level it was considered an erratum.  It gets
to live on as allowed in the architecture.