[PATCH v2 3/5] libnvdimm/nvdimm/flush: Allow architecture to override the flush barrier

[Dan Williams]

On Mon, May 18, 2020 at 10:30 PM Aneesh Kumar K.V
<aneesh.kumar at linux.ibm.com> wrote:
>
>
> Hi Dan,
>
> Apologies for the delay in response. I was waiting for feedback from
> hardware team before responding to this email.
>
>
> Dan Williams <dan.j.williams at intel.com> writes:
>
> > On Tue, May 12, 2020 at 8:47 PM Aneesh Kumar K.V
> > <aneesh.kumar at linux.ibm.com> wrote:
> >>
> >> Architectures like ppc64 provide persistent memory specific barriers
> >> that will ensure that all stores for which the modifications are
> >> written to persistent storage by preceding dcbfps and dcbstps
> >> instructions have updated persistent storage before any data
> >> access or data transfer caused by subsequent instructions is initiated.
> >> This is in addition to the ordering done by wmb()
> >>
> >> Update nvdimm core such that architecture can use barriers other than
> >> wmb to ensure all previous writes are architecturally visible for
> >> the platform buffer flush.
> >
> > This seems like an exceedingly bad idea, maybe I'm missing something.
> > This implies that the deployed base of DAX applications using the old
> > instruction sequence are going to regress on new hardware that
> > requires the new instructions to be deployed.
>
>
> pmdk support for ppc64 is still work in progress and there is pull
> request to switch pmdk to use new instruction.

Ok.

>
> https://github.com/tuliom/pmdk/commit/fix-flush
>
> All userspace applications will be switched to use the new
> instructions. The new instructions are designed such that when running on P8
> and P9 they behave as 'dcbf' and 'hwsync'.

Sure, makes sense.

> Applications using new instructions will behave as expected when running
> on P8 and P9. Only future hardware will differentiate between 'dcbf' and
> 'dcbfps'

Right, this is the problem. Applications using new instructions behave
as expected, the kernel has been shipping of_pmem and papr_scm for
several cycles now, you're saying that the DAX applications written
against those platforms are going to be broken on P8 and P9?

> > I'm thinking the kernel
> > should go as far as to disable DAX operation by default on new
> > hardware until userspace asserts that it is prepared to switch to the
> > new implementation. Is there any other way to ensure the forward
> > compatibility of deployed ppc64 DAX applications?
>
> AFAIU there is no released persistent memory hardware on ppc64 platform
> and we need to make sure before applications get enabled to use these
> persistent memory devices, they should switch to use the new
> instruction?

Right, I want the kernel to offer some level of safety here because
everything you are describing sounds like a flag day conversion. Am I
misreading? Is there some other gate that prevents existing users of
of_pmem and papr_scm from having their expectations violated when
running on P8 / P9 hardware? Maybe there's tighter ecosystem control
that I'm just not familiar with, I'm only going off the fact that the
kernel has shipped a non-zero number of NVDIMM drivers that build with
ARCH=ppc64 for several cycles.