[Date Prev][Date Next] [Thread Prev][Thread Next] [Date Index] [Thread Index]

Re: Debian GNU/Linux on tablet hardware

W dniu 28.10.2011 17:46, Jeremiah Foster pisze:

Android's Linux kernels are supported (maintained?) by Linaro.

With my Linaro hat on I must object. Depending on what you meant the statement above is either highly inaccurate or simply untrue.

Hence the question mark. :)

I think what I originally meant is that we don't focus solely on Android.

"Improved" is the word I would use, that implies neither support nor maintenance as we are between the vendors (that are also part of Linaro) and upstream kernel community (that we are a part of) and we have no control of either side and their actions. As to what we do check our FAQ (http://www.linaro.org/faqs/) and read on.

Android kernel situation is complicated and varies per board/SoC. What Linaro does is try to upstream and unify the kernel for Linaro member companies SoCs.

What does that mean in practice?

Disclaimer: I'm not a kernel developer. I have experience in the non-Intel part of the world but I'm not the sort of person with up-to-date hands-on experience. For those folks please look at traffic in linaro-dev@lists.linaro.org and at our patchwork instance at patches.linaro.org. You can see what kind of patches we push upstream and if they have landed yet.

As for your question, read on.

A lot of ARM devices have a BSP kernel (android or not) that is prepared by some 3rd party (sometimes also by the vendor themselves if the device is a clone of the reference design) and that kernel is generally not pushed upstream.

This affects, by far, the vast majority of devices out there (I'd say that nothing gets pushed by those companies simply because their work mode does not require such a step - we are working on educating them in the benefits of working both towards products and common code base).

The ARM tree in the upstream kernel is, again, by far, the largest of all the other architectures. If I remember correctly it is in fact larger than *all* the other trees combined. The reasons for this are complicated but can be generally simplified to code duplication between the different devices and greater diversity in the actual hardware as compared to other platforms.

To get ARM Linux healthy we need to reduce that clutter and make sure that support for the latest and greatest hardware is upstream quickly and the code is being actively maintained by everyone.

This is far from finished and uniform. The "BSP" kernel that hardware vendors provide is not supported by Linaro and in fact often contains code that cannot go upstream.

What does it use this proprietary code for? To know the APIs or to get other hardware interface info? Isn't that a little risky? Won't proprietary, and potentially patented IP leak into the Linaro work? (Not that I believe in IP.)

The term proprietary is a bit misleading, the code IS released as GPL. It is simply there to support some parts (often userspace or "firmware") that is not open sourced and cannot be for all practical considerations.

As for the patches in general, there are different reasons why they are not suitable for being proposed and included upstream:

1) Shabby code, against old trees, copy-pasted from another similar device, maintenance hell. This is, by my unqualified judgment, the vast majority of the problem.

2) Code that has no good place in the kernel just yet because the kernel interfaces are insufficient for the extremely complicated world of ARM SoCs. Off the top of my, unqualified, head: power management, memory management, everything related to graphics and probably many more. Here the reason for not being upstream is that there is no consensus on how to do something (how to support a certain class of SoC components) that could be applied to many vendors and non-arm world as well. Here the people that write the BSP cannot solve the problem and just implement their own solution to meet the deadline. Such code is often very good but there are many similar solutions that are quite nontrivial to merge into one sensible codebase. One such example is memory management where we have no less than 3 or 4 competing interfaces from various companies and there is a working group inside Linaro and the greater Linux community that tries to solve this problem.

3) Bits that enable proprietary userspace drivers. The reasons are obvious. This could be related to lots of different things, not only graphics as people often think. IP and software edge (optimizations that make otherwise identical hardware perform better than competition) is probably a big motivation here. The IP protection is not only used as in "don't steal our stuff" but rather "hey, with this being binary it is harder to prove that we violate a specific patent you have". In retrospective this is a thing those companies obviously need. Just look at how many Android handset vendors pay to, for example, Microsoft, for patents that allegedly apply there. The world of graphics is riddled with patents and I'm sure that a big money-laden hardware vendor is a good target for whoever owns the patents.

Linaro has several trees, including a grand unification tree that tries to support all the member companies chips in one tree (and one binary, thanks to device trees) but this effort is years away (my personal estimate, I don't speak for the organization). In addition we have several trees for normal/androidized kernel for each board. In the latest 2011.10 release hardware was not supported in 100% on any board that I'm aware of.

Having said that the term "supported" seems inappropriate to me. We do work on those boards though.

How would you define it?

By "work" I meant "we are *working* on making the kernel and userspace on those boards better in each release". Better is shared amongst:

1) More patches landed upstream, thus less delta.
2) Less duplication within the kernel (better code), more device tree usage, closer to having one kernel binary that supports several different boards. 3) Better power management, stability, performance, more features, bells and whistles. 4) Less delta from the android variant to the normal variant. More discussion and more consensus on how to join the two worlds.

And let's not forget, my own personal favorite, more validation. The code is tested both manually and automatically and the scope, coverage and quality is pushed forward each time.

Anything that runs Android can run GNU/Linux.

This is a gross oversimplification IMHO. You usually get androidized BSP kernel from a few months/years ago with binary parts that have no corresponding source code. Good luck booting vanilla kernel there.

But it appears to me that all the official boards that are targets for Linaro can run a vanilla kernel, is that not the case? If not, what BSP stuff are you referring to - graphics acceleration?

No I don't think that is the case. A quick glance at http://git.linaro.org/gitweb will show you how many trees we have. Except for the explicit upstream trees that Nico maintains none of the important changes in other trees are upstream today (at least not yet).

Even the Linaro various kernels (which are _not_ the BSP kernels) fail to work sensibly on all of the boards today. Next-gen boards are usually the ones with weakest support (although that is rapidly changing, thanks to what we are doing). Often most primitive board features work (like, it kind of boots with a specific boot loader and the CPU runs) but anything you definitely want those boards to do: power management, stability, sound, graphics, 3D & multimedia, wifi/bluetooth, FM radio, GPS(?), DSP suppoet, ARMv7 optimizations, is simply not there.

Best regards

PS: I think that you should ask those questions on @linaro-dev. I could be talking nonsense here and the people that really know simply did not see this message. Therefore I'm cross-posting to linaro-dev.

Reply to: