Re: [patch] x86: improved memory barrier implementation

[Alan Cox <alan@lxorguk.ukuu.org.uk>]

> The only alternative is to assume a weak memory model, and add the required
> barriers to spin_unlock -- something that has been explicitly avoided, but

We have the barriers in spin_unlock already for Pentium Pro and IDT
Winchip systems. The Winchip explicitly supports out of order store (and
was about 10-20% faster with it set this way), the PPro has some annoying
little store ordering bugs which is why we xchgb out of spin_unlock in
those cases.

> The complaint that broken old ppros get more broken seems to be bogus as well:
> wmb on i386 never catered to the old ppro to start with. The errata seem to

There are only a couple of awkward cases with the PPro and we carefully
work around them. The other half of the work is the flush_write_buffers.

> Buggy ppros: I have no interest in, given the vague errata and apparently
> unfixable problems with lockless code; anybody is welcome to try adding
> whatever they like to help them work, but again please don't confuse that
> with having anything to do with this patch...

PPro was fixed, made to work and brutally tested on quad PPro. There are
cases that wouldn't work (PPro + AGP which never happens) and there are
cases that need specific handling (PPro + Voodoo card and some other 3D
bits) which are handled currently in the X server. The rest is fine.

> Winchip: can any of these CPUs with ooostores do SMP? If not, then smp_wmb
> can also be a simple barrier on i386 too.

The IDT Winchip can do SMP apparently. Nobody has ever seen an SMP
winchip board so I don't think it matters and even if one existed we
wouldn't support it anyway

However 
- You've not shown the patch has any performance gain
- You've probably broken Pentium Pro
- and for modern processors its still not remotely clear your patch is
correct because of NT stores.

So NAK

Alan