Re: [PATCH] locking/atomic/x86: Introduce try_cmpxchg64

[Maxim Levitsky]

On Mon, 2022-05-16 at 15:14 +0000, Sean Christopherson wrote:
> On Mon, May 16, 2022, Maxim Levitsky wrote:
> > On Mon, 2022-05-16 at 14:08 +0000, Sean Christopherson wrote:
> > > On Mon, May 16, 2022, Maxim Levitsky wrote:
> > > > On Wed, 2022-05-11 at 21:54 +0200, Uros Bizjak wrote:
> > > > > On Wed, May 11, 2022 at 6:04 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> > > > > > On Wed, May 11, 2022, Uros Bizjak wrote:
> > > > > > > On Wed, May 11, 2022 at 9:54 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> > > > > > > > Still, does 32bit actually support that stuff?
> > > > > > > 
> > > > > > > Unfortunately, it does:
> > > > > > > 
> > > > > > > kvm-intel-y        += vmx/vmx.o vmx/vmenter.o vmx/pmu_intel.o vmx/vmcs12.o \
> > > > > > >                vmx/evmcs.o vmx/nested.o vmx/posted_intr.o
> > > > > > > 
> > > > > > > And when existing cmpxchg64 is substituted with cmpxchg, the
> > > > > > > compilation dies for 32bits with:
> > > > > > 
> > > > > > ...
> > > > > > 
> > > > > > > > Anyway, your patch looks about right, but I find it *really* hard to
> > > > > > > > care about 32bit code these days.
> > > > > > > 
> > > > > > > Thanks, this is also my sentiment, but I hope the patch will enable
> > > > > > > better code and perhaps ease similar situation I have had elsewhere.
> > > > > > 
> > > > > > IMO, if we merge this it should be solely on the benefits to 64-bit code.  Yes,
> > > > > > KVM still supports 32-bit kernels, but I'm fairly certain the only people that
> > > > > > run 32-bit KVM are KVM developers.  32-bit KVM has been completely broken for
> > > > > > multiple releases at least once, maybe twice, and no one ever complained.
> > > > > 
> > > > > Yes, the idea was to improve cmpxchg64 with the implementation of
> > > > > try_cmpxchg64 for 64bit targets. However, the issue with 32bit targets
> > > > > stood in the way, so the effort with 32-bit implementation was mainly
> > > > > to unblock progression for 64-bit targets.
> > > > 
> > > > Would that allow tdp mmu to work on 32 bit?
> > > 
> > > From a purely technical perspective, there's nothing that prevents enabling the
> > > TDP MMU on 32-bit kernels.  The TDP MMU is 64-bit only to simplify the implementation
> > > and to reduce the maintenance and validation costs.
> > 
> > I understand exactly that, so the question, will this patch help make the tdp
> > mmu work transparently on 32 bit kernels? I  heard that 64 bit cmpxchg was
> > one of the main reasons that it is 64 bit only.
> 
> I don't think it moves the needled much, e.g. non-atomic 64-bit accesses are still
> problematic, and we'd have to update the TDP MMU to deal with PAE paging (thanks
> NPT).  All those problems are solvable, it's purely a matter of the ongoing costs
> to solve them.
> 
> > I am asking because there was some talk to eliminate the direct mode from the
> > legacy non tdp mmu, which would simplify its code by a lot, but then it will
> > make 32 bit kernel fail back to shadowing mmu.
> 
> Simplify which code?  Between the nonpaging code and direct shadow pages in
> indirect MMUs, the vast majority of the "direct" support in the legacy MMU needs
> to be kept even if TDP support is dropped.  And the really nasty stuff, e.g. PAE
> roots, would need to be carried over to the TDP MMU.
> 

I guess this makes sense. I haven't researched the code well enough to know the exact answer.
I was just curious if this patch makes any difference :)

Thanks!

Best regards,
	Maxim Levitsky