[RFC][PATCH 0/5 v2] DMA-BUF Heaps (destaging ION)

[Benjamin Gaignard]

Le mer. 13 mars 2019 à 23:31, John Stultz <john.stultz at linaro.org> a écrit :
>
> On Wed, Mar 13, 2019 at 1:11 PM Liam Mark <lmark at codeaurora.org> wrote:
> > On Tue, 5 Mar 2019, John Stultz wrote:
> > >
> > > Eventual TODOS:
> > > * Reimplement page-pool for system heap (working on this)
> > > * Add stats accounting to system/cma heaps
> > > * Make the kselftest actually useful
> > > * Add other heaps folks see as useful (would love to get
> > >   some help from actual carveout/chunk users)!
> >
> > We use a modified carveout heap for certain secure use cases.
>
> Cool! It would be great to see if you have any concerns about adding
> such a secure-carveout heap to this framework. I suspect it would be
> fairly similar to how its integrated into ION, but particularly I'd be
> interested in issues around the lack of private flags and other
> allocation arguments like alignment.
>
> > Although there would probably be some benefit in discssing how the dma-buf
> > heap framework may want to support
> > secure heaps in the future it is a large topic which I assume you don't
> > want to tackle now.
>
> So I suspect others (Benjamin?) would have a more informed opinion on
> the details, but the intent is to allow secure heap implementations.
> I'm not sure what areas of concern you have for this allocation
> framework in particular?

yes I would be great to understand how you provide the information to
tell that a dmabuf
is secure (or not) since we can't add flag in dmabuf structure itself.
An option is manage
the access rights when a device attach itself to the dmabuf but in
this case you need define
a list of allowed devices per heap...
If you have a good solution for secure heaps you are welcome :-)

Benjamin
>
> > We don't have any non-secure carveout heap use cases but the client use
> > case I have seen usually revolve around
> > wanting large allocations to succeed very quickly.
> > For example I have seen camera use cases which do very large allocations
> > on camera bootup from the carveout heap, these allocations would come from
> > the carveout heap and fallback to the system heap when the carveout heap
> > was full.
> > Actual non-secure carveout heap can perhaps provide more detail.
>
> Yea, I'm aware that folks still see carveout as preferable to CMA due
> to more consistent/predictable allocation latency.  I think we still
> have the issue that we don't have bindings to establish/configure
> carveout regions w/ dts, and I'm not really wanting to hold up the
> allocation API on that issue.
>
>
> > Since we are making some fundamental changes to how ION worked and since
> > Android is likely also be the largest user of the dma-buf heaps framework
> > I think it would be good
> > to have a path to resolve the issues which are currently preventing
> > commercial Android releases from moving to the upstream version of ION.
>
> Yea, I do see solving the cache management efficiency issues as
> critical for the dmabuf heaps to be actually usable (my previous
> version of this patchset accidentally had my hacks to improve
> performance rolled in!).  And there are discussions going on in
> various channels to try to figure out how to either change Android to
> use dma-bufs more in line with how upstream expects, or what more
> generic dma-buf changes we may need to allow Android to use dmabufs
> with the expected performance they need.
>
> > I can understand if you don't necessarily want to put all/any of these
> > changes into the dma-buf heaps framework as part of this series, but my
> > hope is we can get
> > the upstream community and the Android framework team to agree on what
> > upstreamable changes to dma-buf heaps framework, and/or the Android
> > framework, would be required in order for Android to move to the upstream
> > dma-buf heaps framework for commercial devices.
>
> Yes. Though I also don't want to get the bigger dma-buf usage
> discussion (which really affects all dmabuf exporters) too tied up
> with this patch sets attempt to provide a usable allocation interface.
> Part of the problem that I think we've seen with ION is that there is
> a nest of of related issues, and the entire thing is just too big to
> address at once, which I think is part of why ION has sat in staging
> for so long. This patchset just tries to provide an dmabuf allocation
> interface, and a few example exporter heap types.
>
> > I don't mean to make this specific to Android, but my assumption is that
> > many of the ION/dma-buf heaps issues which affect Android would likely
> > affect other new large users of the dma-buf heaps framework, so if we
> > resolve it for Android we would be helping these future users as well.
> > And I do understand that some the issues facing Android may need to be
> > resolved by making changes to Android framework.
>
> While true, I also think some of the assumptions in how the dma-bufs
> are used (pre-attachment of all devices, etc) are maybe not so
> realistic given how Android is using them.  I do want to explore if
> Android can change how they use dma-bufs, but I also worry that we
> need to think about how we could loosen the expectations for dma-bufs,
> as well as trying to figure out how to support things folks have
> brought up like partial cache maintenance.
>
> > I think it would be helpful to try and get as much of this agreed upon as
> > possible before the dma-buf heaps framework moves out of staging.
> >
> > As part of my review I will highlight some of the issues which would
> > affect Android.
> > In my comments I will apply them to the system heap since that is what
> > Android currently uses for a lot of its use cases.
> > I realize that this new framework provides more flexibility to heaps, so
> > perhaps some of these issues can be solved by creating a new type of
> > system heap which Android can use, but even if the solution involves
> > creating a new system heap I would like to make sure that this "new"
> > system heap is upstreamable.
>
> So yea, I do realize I'm dodging the hard problem here, but I think
> the cache-management/usage issue is far more generic.
>
> You're right that this implementation give a lot of flexibility to the
> exporter heaps in how they implement the dmabuf ops (just like how
> other device drivers that are dmabuf exporters have the same
> flexibility), but I very much agree we don't want to add a system and
> then later a "system-android" heap. So yea, a reasonable amount of
> caution is warranted here.
>
> Thanks so much for the review and feedback! I'll try to address things
> as I can as I'm traveling this week (so I may be a bit spotty).
>
> thanks
> -john