RE: [RFC please help] membarrier: Rewrite sync_core_before_usermode()

From: David Laight
Date: Fri Jan 01 2021 - 13:38:41 EST


From: Andy Lutomirski
> Sent: 29 December 2020 00:36
...
> I mean that the mapping from the name "sync_core" to its semantics is
> x86 only. The string "sync_core" appears in the kernel only in
> arch/x86, membarrier code, membarrier docs, and a single SGI driver
> that is x86-only. Sure, the idea of serializing things is fairly
> generic, but exactly what operations serialize what, when things need
> serialization, etc is quite architecture specific.
>
> Heck, on 486 you serialize the instruction stream with JMP.

Did the 486 even have a memory cache?
Never mind separate I&D caches.
Without branch prediction or an I$ a jmp is enough.
No idea how the dual 486 box we had actually behaved.

For non-SMP the x86 cpus tend to still be compatible with
the original 8086 - so are pretty much fully coherent.
ISTR the memory writes will invalidate I$ lines.

But there was some hardware compatibility that meant a load
of Pentium-75 systems were 'scavenged' from development for
a customer - we got faster P-266 boxes as replacements.

OTOH we never did work out how to do the required 'barrier'
when switching a Via C3 to and from 16-bit mode.
Sometimes it worked, other times the cpu went AWOL.
Best guess was that it sometimes executed pre-decoded
instructions for the wrong mode when returning from the
function call that flipped modes.

Then there is the P-Pro era Intel doc that says that IOR/IOW
aren't sequenced wrt memory accesses.
Fortunately all x86 processors have sequenced them.
Which is what the current docs say.

David

-
Registered Address Lakeside, Bramley Road, Mount Farm, Milton Keynes, MK1 1PT, UK
Registration No: 1397386 (Wales)