Re: linux-yocto: ktypes/tiny and some questions along the way


Bruce Ashfield <bruce.ashfield@...>
 

On 11-12-14 01:07 PM, Darren Hart wrote:


On 12/14/2011 08:59 AM, Bruce Ashfield wrote:
On 11-12-12 06:17 PM, Darren Hart wrote:
On 12/11/2011 09:14 PM, Bruce Ashfield wrote:
On 11-12-09 5:52 PM, Darren Hart wrote:
Bruce,

I'm looking at introducing a new kernel type, ktypes/tiny.

Tiny will define a core set of kernel policy options, such as proc, sys,
devtmpfs, futex, epoll, elf bin format, etc. It will not enable any
drivers, filesystems, debug options, or networking options by default.
This would be the responsibility of the BSP to add a named feature
implementing this.

Taking a look at the ktypes we have now resulted in some questions:

dvhart@envy:~/source/linux/linux-yocto-3.0/meta/cfg/kernel-cache/ktypes
$ tree
.
├── base
│ ├── base.cfg
│ ├── base.scc
│ ├── hardware.cfg
│ ├── hardware.kcf
│ ├── non-hardware.cfg
│ └── non-hardware.kcf
├── preempt-rt
│ ├── preempt-rt.cfg
│ └── preempt-rt.scc
├── standard
│ ├── perf-force-include-of-stdbool.h.patch
│ ├── perf-hard-code-NO_LIBPERL-NO_LIBPYTHON.patch
│ ├── standard-patches.scc
│ ├── standard.cfg
│ ├── standard.scc
│ └── x86-add-TIF_32BIT-compatibility-define.patch

These form a hiearchy, each inheriting from the layer beneath like so:

base/standard/preempt-rt

As I dig into this I see that some policy is infact laid down by base,
including things like:

CONFIG_MODULES=y
CONFIG_INET=y
CONFIG_PREEMPT=y
CONFIG_NFS_FS=y
CONFIG_MSDOS_PARTITION=y

These pull in the IP stack, the block layer, etc. Because of this, I
can't really inherit from base for ktypes/tiny. I would like to inherit
the hardware and non-hardware kcf files though, as well as any patches
that might make their way into base. Which leads me to standard. I would
like to see a much smaller set of config policy options set in base.
Most likely these should be exactly what we agree on for tiny, and tiny
wouldn't add any additional policy as it implements only what is
required for a Yocto Project built kernel.

NOTE: kernel version is 3.0+
$ cat base/base.scc | head -n 1
set_kernel_version 2.6.37
FYI: nothing uses this at the moment. But I fixed that here.


There are three patches in standard, only two of which (the perf ones)
are listed in the standard.scc. As I believe any kernel we build should
have these, I would like to see any global patches applied to base,
leaving standard to define policy, and include named features.
standard only includes 3 patches because they were a bit hard to
classify elsewhere. In a more fully populated kernel, it does include
quite a bit more directly.
Sorry for the slow reply, I got bogged down trying to write some
python code yesterday.


directly = git repository commits?

Sort of. Or what I meant by directly was the "patch<foo>" right
in standard.scc. As you well know, for the most part, standard pulls in
patches by including other features, and those are just as direct
as a patch/commit listed right in the standard kernel .scc file.

The yocto kernels have thus far had relatively few additional features,
but if that changes, the standard kernel will have more direct changes
being applied to the branches.

I'm seeing a problem with the coupling of patches and configs. I
understand why it made sense to do at the time, but if we want to base
all branches off the same sources (and it's clear we do) then our
infrastructure should allow us to prepare that source base separately
from the config. Applying the config and then clearing it out is a
rather ugly solution IMO.
There are configs that absolutely should be coupled with patches,
and configs that don't need to be. We are obviously talking about
configs that don't need to be, and we definitely have the flexibility
to keep them apart.






With these changes, I could add ktypes/tiny as follows:
$ cat tiny/tiny.scc
include ktypes/base
branch tiny
include features/xyz/xyz.scc
include cfg/abc.scc

Another point of interest is preempt-rt, as I can see people wanting to
build tiny preempt-rt. I think the best approach here is to create
ktypes/tiny/preempt-rt-tiny/preempt-rt-tiny.scc.

Note: I believe this fails with .patch.patch
$ cat features/rt/rt.scc | grep patch
patch rt-apply-patch-3.0.10-rt27.patch.patch
Love my "patch.patch". Script went wild on that one, luckily the content
in the branch is king :) I fixed that here when I was auditing 3.0.12+rt
so that looks more sane now.


I'm working on a patch series that implements the suggestions I've made
above. Bruce, do you have any issues with this approach?
Allow me to ramble a bit below ...

No big issues. At some level(s) it is just a shell game. We can move
configs and patches around to the appropriate level to get the building
blocks that we need and get common functionality into a common location.

base was (is) intended to document the branch point from the mainline
kernel, and contain largely configuration and very few patches. Any
important or larger size functional additions go into the standard
kernel.
What is the value of having base and standard as separate branches?
Isn't base easily derivable from the standard git history? Still if it
isn't used anywhere, then the branch is a no-op and adds no complexity,
so that's fine.
base can definitely be derived from git history, or via git merge-base,
but the branch is a very clear delineation of the jumping point
and contains things like the initial commits to .gitignore. It also
provides a common branch point for the meta branch and standard, so
really basic things can be shared. It has also been used in the past
for a simple branch to build the stock korg kernel from where we jump
into added content.

It also reduces complexity in the tools by having a defined/known
branch name that can be used to detect the type of a repo.

So nothing earth shattering, but a series of smaller reasons.
OK... I can certainly work with it.




So branching from base for a new kernel type is something we can do, but
it will risk missing additions (say for example tracing fixes, or the
next super-duper debug via kprobes patch series), since they'll go into
the standard kernel.
OK, that's no good. My moving patches down to base was intended to avoid
just that problem. And given the above, leaving them in standard is
preferred. OK, that poses some issues.... discussed below.

If I just slide all the patches down into base,
why not just call 'base' 'standard' or just make standard have the
configuration you are working on for tiny.

The point I'm attempting to make, is that the base/common point can be
whatever we decide it needs to be, "standard" can be renamed, content
moved up and down, etc. i.e. standard could have it's config changed,
a new branch created off standard ( and options moved there), or
standard could be streamlined and the boards include more common
config options manually vs inheriting them, etc.

I need a bit more time to think about that myself.
The above is fine. Or rather, any of the above can work. Let's discuss
which one :-)
Indeed. It is just a matter of shifting things around to where
we want them.



Another option is to let tiny inherit from standard, but force a
new baseline for configuration options. i.e. your allnoconfig
technique buried in the middle of inheritance tree. With the
optional/required designations I'm finishing up for 1.2 you won't get
hit with reams of redefined configuration warnings. With that set, you
can then go about adding kernel type specific options/features/patches
as we see fit .. and you'll pickup those features that I was talking
about above. If you bear with me by doing some of this manually for
now (setting the baseline), it is easy to add a construct to do that
reset automagically.
I really want to avoid setting a base config, and then backtracking to
reset it to something else in a derived branch. Chasing those overrides,
etc can be a massive pain. Much better to get the main SOURCES and
essential policy into a single branch, then let standard and tiny branch
from there. Intuitively, Standard and Tiny should be siblings, not
ancestors of eachother. Keeping this relationship apparent in the
branching and source and config inheritance should help keep things
clear and maintainable.
Agreed, and that's the trick, we have source changes and configuration
changes. We want to inherit (not cherry pick) both where possible, but
in the end, we sometimes have to chose.

BSPs are an example of this. We try and drive the patches for boards
down as far as possible until they are common. But sometimes, they are
so evil that they have to locked into a room by themselves. To share
those BSP changes between kernel types, they are essentially cherry
picks, as are their config values.

yocto/standard/my-bsp
yocto/standard/preempt-rt/my-bsp

Have the same patches (as can be checked by git patch-id), but they don't
have the same git hash.

The good news is of course that these changes are on the end of branches,
are largely orthogonal to the rest of the kernel (and hence don't
conflict, or need to be shared) and they are automatically done when
updating the kernel. Also to deal with this, the BSPs are split into
kernel type specific options and common options. That way we only
maintain one common set of configuration values for a board and share
it via scc includes.

So I see this as a similar situation/choice, do we go with:

a)
yocto/standard/tiny

b)
yocto/standard
yocto/tiny

c)
yocto/standard/standard-config-items/
yocto/standard/tiny

In all options, we'd drive a decent/common configuration into the base
and both standard and tiny would inherit that, so consider that a
constant.

The rest of the issues/properties of the options are:

a) tiny would inherit standard, patches and config. We'd have to
back off (or exclude (we can do that!)) standard's config values.
But we get the patch inheritance that we want.
Barf.


b) tiny and standard are peers, but tiny needs individual merges
of all features and there aren't any common commits between the
two outside of base (or korg). So this is the 'source complexity'
issue.
Double barf.

They were options .. just not good ones :)



c) tiny inherits standard, but we move out standard's config values
into another branch (they are cheap after all). That way
tiny gets the patches and features, but not the config of standard.
In this example, we might want to kill "base", since we are adding
another layering of branches, although one that would be trivial/small.
This seems the obvious choice to me. Bare with me for some context:
To me as well.


I had a discussion with Richard yesterday regarding distros, images,
policies, etc. The kernel-type selection really belongs to the DISTRO
definition. The preempt-rt is selected currently by manually setting
PREFERRED_PROVIDER in a local config file. This is lame. The right way
to have done this would have been to define a poky-rt distribution which
specified linux-yocto-rt as the PREFERRED_PROVIDER and added rt-tests
(and taskset and chrt to RDEPENDS). Then people set their DISTRO in the
local.conf (this actually makes sense!) and build one of the standard
images like core-image-minimal or core-image-sato, and they get an RT
enabled kernel and the tools commonly needed to work with and analyze
this kernel. It's a much better mechanism.
Yup. I agree with that as well. "distro" seems like a heavy term,
but that's because I'm not a traditional OE kinda-guy. But the
concept is bang on.


So what does that mean here? Well, I suggest we put all the sources in
standard (minus evil BSP patches obviously) and then create a ktype per
DISTRO definition:
And minus -rt at the moment. It still has issues with being enabled
and turned completely off, and it makes merging other out of tree
functionality difficult/manual .. not something I want to take on
everywhere. In the future, we both well know that this will likely be
doable, I'm just saying 'not yet' .. but that doesn't impact what we
are proposing here much (only for the -rt patches themselves).


yocto/base
yocto/standard/base
yocto/standard/poky
yocto/standard/poky-rt
yocto/standard/poky-tiny
I'd want the distro not to be named in the branches, but yes,
that looks ok to me.


Combinatorix become a bit of an issue, as we will likely have need for:

yocto/standard/poky-tiny/base
yocto/standard/poky-tiny/poky-tiny-rt

or something like that.

This clearly delineates the line between common sources (standard/base)
and distro defined policy (poky, poky-rt, and poky-tiny) which includes
the standard kernel config.
I'd still s/poky/something else/g .. but I'd agree with that.




It may not be obvious, but we will need to have BSPs hanging off
the new tiny kernel type as well.

yocto/standard/qemuppc
yocto/tiny/qemppc
yocto/standard/poky/qemuppc
yocto/standard/poky-rt/qemuppc
yocto/standard/poky-tiny/qemuppc

Yes, understood. If we couple this with my earlier email about deleting
unecessary branches, I think this is managable. I believe we should
consider any BSP specific branch to be a problem in need of solving -
e.g. an upstream solution is needed.
+1 .. completely agree. anyone building on the kernel tree as a base
has full flexibility to do what they want for new branches, and they
can be dynamically be created at build time .. so we can keep things
simple and not impact flexibility.



So it comes down to this:

- Is tiny a full kernel type, and takes the "responsibilities" that
this entails, or is it a refinement and streamlining of existing
BSPs and kernel types ... i.e. is it a last minute overlay.
ktype.
that's the right choice :)



I've made both choices in the past and depending on the choice, we
arrange the tree differently.



It's the features and patch balancing that's the hard part, not getting
the configs how we want them. I've lived the madness of the other!
I agree the source-work should not be duplicated.


What we don't want to get into (I've been there, with an old "small"
kernel type) is having a different stack of patches and fixes on
multiple kernel types. The patches start failing, and you end up always
merging a fix in two different ways due to changes in the feature mix
on each kernel type.
Agreed

(This comment more applies to the tiny and -rt
mix). Better mileage might be gained by doing.

yocto/standard/preempt-rt/tiny

versus moving the big patch onto yocto/tiny/preempt-rt, but that implies
that tiny is providing a set of configuration values that clobber what
has been set by both standard and -rt .. something that trades patch
maintenance with config maintenance.
I intend to keep Tiny as config options as much as possible (but I can
see some new Kconfig options from linux-tiny (or tiny-linux, whatever it
is) creeping their way in. Still, I prefer tiny/preempt-rt for the
reasons stated above.
We may need to normalize the names to keep things straight.
There are currently two kernel types in yocto: "standard" and
"preempt-rt". (Other instances of the tooling have more kernel
types).

In this discussion we are talking about adding another kernel
type "tiny". So it would be a peer to preempt-rt, and thus
we wouldn't want to re-use the term "preempt-rt".

Internally in the kernel-cache directories preempt-rt is the "kernel
type" and "rt" is the feature.

So we'd want something more like:

yocto/standard
yocto/standard/tiny
yocto/standard/tiny/rt
yocto/standard/preempt-rt
See above, but I think we are in agreement here.
Yep, I'm not sold on a distro name, but if you change this to:

yocto/base
yocto/standard/cfg (bad name, but I wanted something)
yocto/standard/rt
yocto/standard/tiny

Then the tree is more of a common base ... they are just names after
all! We already have 'yocto' in there, so that's enough specifics for
my taste.

or we flip it around ...

base
standard/yocto
standard/yocto-rt
standard/yocto-tiny

Which looks more like what you proposed, but without the double
specific names.



(That's just a suggestion, and yes, I admit that can also be
confusing to someone simply reading the branches).

No matter how you slice it, there will be some duplication between
tiny/rt and preempt-rt, since you'll want both the kernel configs
and the patches from the rt feature (which is most of the preempt-rt
kernel type). So they will have to be merged twice (automatically
during tree generation of course).

.. and BTW. The names (even "standard", are not hardcoded anywhere),
so if adding new options, kernels, etc, makes some name just seem
wrong, we can change it (in a new release). i.e. standard could become
the (over used) "core" branch .. if it makes things read better.

That begs the question. If a patch is good enough for tiny, and it
is properly #ifdef'd, it can be merged to the standard branch, and
once that is done, we'll have a much easier time maintaining tiny +
stacked features, since those features will already have adapted
to the tiny imported features/source changes.
Agreed.




Working through all that, what I'm saying is that I'd prefer to make
standard do what you want, and keep the base similar to what it is.
We can shuffle undesirable standard configuration items up versus
shuffling all patches and functionality down to yocto/base. But the
approach is fundamentally the same.

i.e. I see no reason why the tiny config wouldn't be fine as
the standard config or as something that makes significant changes
to standard's configuration. But definitely something that wants the
patches and features in standard.
Agreed.
This is the way forward. Some variant of the options that I've listed
above.

I've been through a lot of this in the past, so I can make suggestions
and having a discussion like this is very helpful if anyone peering
in wonders what's going on behind the covers, since this is largely
unseen effort.
Let me know what you think of the distro:branch mapping I have defined
above. I think this is the cleanest way forward. Going to wait to hear
from you before I proceed. Let's make sure we have an agreed upon
approach first.
See above. I'm thinking more about names, but wanted to get
an answer out sooner.

Cheers,

Bruce

Join yocto@lists.yoctoproject.org to automatically receive all group messages.