Re: [RFC PATCH 00/12] Ion cleanup in preparation for moving out of staging

From: Laura Abbott
Date: Mon Mar 13 2017 - 17:29:22 EST

On 03/13/2017 03:54 AM, Brian Starkey wrote:
> On Sun, Mar 12, 2017 at 02:34:14PM +0100, Benjamin Gaignard wrote:
>> 2017-03-09 18:38 GMT+01:00 Laura Abbott <labbott@xxxxxxxxxx>:
>>> On 03/09/2017 02:00 AM, Benjamin Gaignard wrote:
>>>> 2017-03-06 17:04 GMT+01:00 Daniel Vetter <daniel@xxxxxxxx>:
>>>>> On Mon, Mar 06, 2017 at 11:58:05AM +0100, Mark Brown wrote:
>>>>>> On Mon, Mar 06, 2017 at 11:40:41AM +0100, Daniel Vetter wrote:
>>>>>>> No one gave a thing about android in upstream, so Greg KH just dumped it
>>>>>>> all into staging/android/. We've discussed ION a bunch of times, recorded
>>>>>>> anything we'd like to fix in staging/android/TODO, and Laura's patch
>>>>>>> series here addresses a big chunk of that.
>>>>>>> This is pretty much the same approach we (gpu folks) used to de-stage the
>>>>>>> syncpt stuff.
>>>>>> Well, there's also the fact that quite a few people have issues with the
>>>>>> design (like Laurent). It seems like a lot of them have either got more
>>>>>> comfortable with it over time, or at least not managed to come up with
>>>>>> any better ideas in the meantime.
>>>>> See the TODO, it has everything a really big group (look at the patch for
>>>>> the full Cc: list) figured needs to be improved at LPC 2015. We don't just
>>>>> merge stuff because merging stuff is fun :-)
>>>>> Laurent was even in that group ...
>>>>> -Daniel
>>>> For me those patches are going in the right direction.
>>>> I still have few questions:
>>>> - since alignment management has been remove from ion-core, should it
>>>> be also removed from ioctl structure ?
>>> Yes, I think I'm going to go with the suggestion to fixup the ABI
>>> so we don't need the compat layer and as part of that I'm also
>>> dropping the align argument.
>>>> - can you we ride off ion_handle (at least in userland) and only
>>>> export a dma-buf descriptor ?
>>> Yes, I think this is the right direction given we're breaking
>>> everything anyway. I was debating trying to keep the two but
>>> moving to only dma bufs is probably cleaner. The only reason
>>> I could see for keeping the handles is running out of file
>>> descriptors for dma-bufs but that seems unlikely.
>>>> In the future how can we add new heaps ?
>>>> Some platforms have very specific memory allocation
>>>> requirements (just have a look in the number of gem custom allocator in drm)
>>>> Do you plan to add heap type/mask for each ?
>>> Yes, that was my thinking.
>> My concern is about the policy to adding heaps, will you accept
>> "customs" heap per
>> platforms ? per devices ? or only generic ones ?
>> If you are too strict, we will have lot of out-of-tree heaps and if
>> you accept of of them
>> it will be a nightmare to maintain....
> Are you concerned about actual heaps (e.g. a carveout at 0x80000000 vs
> a carveout at 0x60000000) or heap types?
> For heap types, I think the policy can be strict - if it's generally
> useful then it should live in-tree in ion. Otherwise, it would be
> out-of-tree. I'd expect most "custom" heaps to be parameterisable to
> the point of being generally useful.

I'm willing to be reasonably permissive in what lives in tree. A good
example would be something like a heap for the OMAP tiler which had
weird hardware requirements. The associated devices that go with the
heap should be well supported upstream though.

> For actual heap instances, I would expect them to be communicated via
> reserved-memory regions or something similar, and so the maintenance
> burden is pretty low.

Yes. After the next round of review for this series I'm going to
start thinking about properties for chunk and carveout heaps if nobody
proposes something first.

> The existing query ioctl can allow heap IDs to get assigned
> dynamically at runtime, so there's no need to reserve "bit 6" for
>> Another point is how can we put secure rules (like selinux policy) on
>> heaps since all the allocations
>> go to the same device (/dev/ion) ? For example, until now, in Android
>> we have to give the same
>> access rights to all the process that use ION.
>> It will become problem when we will add secure heaps because we won't
>> be able to distinguish secure
>> processes to standard ones or set specific policy per heaps.
>> Maybe I'm wrong here but I have never see selinux policy checking an
>> ioctl field but if that
>> exist it could be a solution.
> I might be thinking of a different type of "secure", but...
> Should the security of secure heaps be enforced by OS-level
> permissions? I don't know about other architectures, but at least on
> arm/arm64 this is enforced in hardware; it doesn't matter who has
> access to the ion heap, because only secure devices (or the CPU
> running a secure process) is physically able to access the memory
> backing the buffer.
> In fact, in the use-cases I know of, the process asking for the ion
> allocation is not a secure process, and so we wouldn't *want* to
> restrict the secure heap to be allocated from only by secure
> processes.
> -Brian