On Tue, Jul 14, 2020 at 12:42:11PM -0700, Dave Hansen wrote:
> On 7/14/20 12:29 PM, Peter Zijlstra wrote:
> > On Tue, Jul 14, 2020 at 12:06:16PM -0700, Ira Weiny wrote:
> >> On Tue, Jul 14, 2020 at 10:44:51AM +0200, Peter Zijlstra wrote:
> >>> So, if I followed along correctly, you're proposing to do a WRMSR per
> >>> k{,un}map{_atomic}(), sounds like excellent performance all-round :-(
> >> Only to pages which have this additional protection, ie not DRAM.
> >>
> >> User mappings of this memory is not affected (would be covered by User
> >> PKeys if
> >> desired). User mappings to persistent memory are the primary use case and
> >> the
> >> performant path.
> > Because performance to non-volatile memory doesn't matter? I think Dave
> > has a better answer here ...
>
> So, these WRMSRs are less evil than normal. They're architecturally
> non-serializing instructions, just like the others in the SDM WRMSR
> documentation:
>
> Note that WRMSR to the IA32_TSC_DEADLINE MSR (MSR index 6E0H)
> and the X2APIC MSRs (MSR indices 802H to 83FH) are not
> serializing.
>
> This section of the SDM needs to be updated for the PKRS. Also note
> that the PKRS WRMSR is similar in its ordering properties to WRPKRU:
>
> WRPKRU will never execute speculatively. Memory accesses
> affected by PKRU register will not execute (even speculatively)
> until all prior executions of WRPKRU have completed execution
> and updated the PKRU register.
>
> Which means we don't have to do silliness like LFENCE before WRMSR to
> get ordering *back*. This is another tidbit that needs to get added to
> the SDM. It should probably also get captured in the changelog.
>
> But, either way, this *will* make accessing PMEM more expensive from the
> kernel. No escaping that. But, we've also got customers saying they
> won't deploy PMEM until we mitigate this stray write issue. Those folks
> are quite willing to pay the increased in-kernel cost for increased
> protection from stray kernel writes. Intel is also quite motivated
> because we really like increasing the number of PMEM deployments. :)
>
> Ira, can you make sure this all gets pulled into the changelogs somewhere?
Yes of course. Thanks for writing that up.
Ira
