Re: [PATCH 09/10] mm/hmm: allow to mirror vma of a file on a DAX backed filesystem

[Dan Williams]

On Tue, Jan 29, 2019 at 7:03 PM Jerome Glisse <jglisse@xxxxxxxxxx> wrote:
[..]
> > >     1) Convert ODP to use HMM underneath so that we share code between
> > >     infiniband ODP and GPU drivers. ODP do support DAX today so i can
> > >     not convert ODP to HMM without also supporting DAX in HMM otherwise
> > >     i would regress the ODP features.
> > >
> > >     2) I expect people will be running GPGPU on computer with file that
> > >     use DAX and they will want to use HMM there too, in fact from user-
> > >     space point of view wether the file is DAX or not should only change
> > >     one thing ie for DAX file you will never be able to use GPU memory.
> > >
> > >     3) I want to convert as many user of GUP to HMM (already posted
> > >     several patchset to GPU mailing list for that and i intend to post
> > >     a v2 of those latter on). Using HMM avoids GUP and it will avoid
> > >     the GUP pin as here we abide by mmu notifier hence we do not want to
> > >     inhibit any of the filesystem regular operation. Some of those GPU
> > >     driver do allow GUP on DAX file. So again i can not regress them.
> >
> > Is this really a GUP to HMM conversion, or a GUP to mmu_notifier
> > solution? It would be good to boil this conversion down to the base
> > building blocks. It seems "HMM" can mean several distinct pieces of
> > infrastructure. Is it possible to replace some GUP usage with an
> > mmu_notifier based solution without pulling in all of HMM?
>
> Kind of both, some of the GPU driver i am converting will use HMM for
> more than just this GUP reason. But when and for what hardware they
> will use HMM for is not something i can share (it is up to each vendor
> to announce their hardware and feature on their own timeline).

Typically a spec document precedes specific hardware announcement and
Linux enabling is gated on public spec availability.

> So yes you could do the mmu notifier solution without pulling HMM
> mirror (note that you do not need to pull all of HMM, HMM as many
> kernel config option and for this you only need to use HMM mirror).
> But if you are not using HMM then you will just be duplicating the
> same code as HMM mirror. So i believe it is better to share this
> code and if we want to change core mm then we only have to update
> HMM while keeping the API/contract with device driver intact.

No. Linux should not end up with the HMM-mm as distinct from the
Core-mm. For long term maintainability of Linux, the target should be
one mm.

> This
> is one of the motivation behind HMM ie have it as an impedence layer
> between mm and device drivers so that mm folks do not have to under-
> stand every single device driver but only have to understand the
> contract HMM has with all device driver that uses it.

This gets to heart of my critique of the approach taken with HMM. The
above statement is antithetical to
Documentation/process/stable-api-nonsense.rst. If HMM is trying to set
expectations that device-driver projects can write to a "stable" HMM
api then HMM is setting those device-drivers up for failure.

The possibility of refactoring driver code *across* vendors is a core
tenet of Linux maintainability. If the refactoring requires the API
exported to drivers to change then so be it. The expectation that all
out-of-tree device-drivers should have is that the API they are using
in kernel version X may not be there in version X+1. Having the driver
upstream is the only surefire insurance against that thrash.

HMM seems a bold experiment in trying to violate Linux development norms.

> Also having each driver duplicating this code increase the risk of
> one getting a little detail wrong. The hope is that sharing same
> HMM code with all the driver then everyone benefit from debugging
> the same code (i am hopping i do not have many bugs left :))

"each driver duplicating code" begs for refactoring driver code to
common code and this refactoring is hindered if it must adhere to an
"HMM" api.