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

Re: Kconfig vs tool chain capabilities



On 25.08.20 14:20, Bertrand Marquis wrote:


On 25 Aug 2020, at 12:22, Jürgen Groß <jgross@xxxxxxxx> wrote:

On 25.08.20 13:16, Bertrand Marquis wrote:
On 25 Aug 2020, at 12:06, Jürgen Groß <jgross@xxxxxxxx> wrote:

On 25.08.20 12:17, Bertrand Marquis wrote:
On 25 Aug 2020, at 10:49, Jürgen Groß <jgross@xxxxxxxx> wrote:

On 25.08.20 10:48, Jan Beulich wrote:
On 25.08.2020 10:08, Jürgen Groß wrote:
On 25.08.20 09:48, Jan Beulich wrote:
On 25.08.2020 09:43, Jürgen Groß wrote:
On 25.08.20 09:34, Jan Beulich wrote:
On 25.08.2020 09:12, Jürgen Groß wrote:
I think both problems can be solved at the same time via the following
approach:

- collect the data which is reflected in today's CONFIG_ variables in a
      single script and store it in a file, e.g in a format like:

      CC_IS_GCC y
      GCC_VERSION 70500
      CLANG_VERSION 0
      CC_HAS_VISIBILITY_ATTRIBUTE y

- check the tool data at each build to match the contents of that file
      and either fail the build or update the file and rerun kconfig if they
      don't match (I think failing the build and requiring to run a
      "make config" would be the better approach)

- fill the CONFIG_ variable contents from that file in kconfig instead
      of issuing the single shell commands

While I agree this is a possible model to use (but still not the
one we've inherited from Linux), I fail to see how this addresses
my "developers should be aware of what they do (not) build and
test" concern: There'd still be dependencies of Kconfig options
on the tool chain capabilities, and their prompts therefore would
still be invisible without the tool chain having the needed
capabilities. IOW I only see this to address 2), but not 1).

Sorry, I fail to see a problem here.

What sense does it make to be able to configure an option which the
tools don't support?

Take CET as an example (chosen because that's the one which
already uses the Kconfig approach, despite my disagreement): It's
quite relevant to know whether you're testing Xen with it enabled,
or with it disabled (and hence you potentially missing changes you
need to make to relevant code portions).

Correct me if I'm wrong, but assuming my suggested changes being made,
wouldn't a .config file setup once with CET enabled (and I assume you'd
try to enable CET on purpose when trying to test CET and you'd realize
not being able to do so in case your tools don't support CET) ensure
you'd never been hit by surprise when some tool updates would remove
CET support?
Probably, but that's not my point. With a CET-incapable tool chain
you're not prompted whether to enable CET in the first place, when
creating the initial .config. It is this unawareness of a crucial
part of code not getting built and tested (and likely unknowingly)
that I dislike. In fact, after Andrew's patches had gone in, it
had taken me a while to realize that in certain of my builds I don't
have CET enabled (despite me having done nothing to disable it), and
hence those builds working fine are meaningless for any changes
affecting CET code in any way.

Yes, this is the result of letting some options depend on others.

This is what I meant regarding the architecture: there are e.g. multiple
source files in drivers/char/ being built only for ARM or X86, in spite
of being located outside arch/. And yet you don't see this as a problem,
even if you are not able to select those drivers to be built when using
"the other" arch. They are silently disabled. Just like CET in case of
an incapable tool chain.

So IMO either we ban "depends on" from our Kconfig files (no, I don't
want to do that), or we use it as designed and make it as user friendly
as possible. In case we as developers have a special test case then we
need to check the .config whether the desired settings are really
present. Having those settings depending on tool capabilities in a
specific file will make this check much easier.

And BTW, I can't see how setting the tolls' capabilities from e.g.
arch/x86/Rules.mk is better in any way (see how CONFIG_INDIRECT_THUNK
got its value in older Xen versions like 4.12).

We can't have everything and I believe automatically disabling features
which can't work with the current tools is a sane decision. Doing this
via Kconfig is the better approach compared to Makefile sniplets IMO.
That sounds like a nice feature to have some compiler or tools options that
can be selected or activated in Kconfig. There are some compiler options
mandatory to handle some erratas or XSA that one might want to explicitely
select.
I am bit unsure about the part where some kconfig options would only
be available or not depending on some tests with the compiler being doing
prior to opening the editor. I would guess the menuconfig process would
have to first run some tests and then generated some HAS_ configuration
options depending on the result of the tests.
Did i got the idea right here ?
Is this something somebody tried to do ?
As a user I would more expect that the build process would tell me that my
configuration is invalid because i selected something that is not supported
by my compiler. I might have the chance to just fix my build to use the right
compiler (like by mistake using x86 toolchain to compile for arm).
We should also be careful not to silently ignore some configuration option if
one is changing the compiler and the new one does not support an option.
A user would have his configuration and compile using it without
passing through the editor interface and might need to be aware that a part
of his configuration is not valid anymore because the tools he is using changed.
This is something that could occur a lot when using a distribution toolchain.
Also there are some compiler option changing so i would more think that
there should be generic configuration options so that in the makefiles we
could have the opportunity to add different compiler options for different
toolchains depending on the version or the type of the toolchain.
To be clear i would see something like:
in kconfig:
COMPILER_OPTION_XXX
        bool “activate XXX compiler flag
in Makefile:
ifeq ($(CONFIG_COMPILER_OPTION_XXX), true)
test_compiler_cxx:
        $(CC) -xxx dummy.c -o dummy || $(error Your compiler does not support 
-xxx)
cc-flags += -xxx
endif
The syntax is wrong here but you get the idea :-)

Ah, okay, this is another approach, which might be even more flexible.
It would allow to control compiler flags instead of more high level
features.
We might have both, this would also allow to have more high level features 
which are
doing both adding compiler flags and other stuff,

In case we want to go that route we should default COMPILER_OPTION_XXX
to the current tool capabilities in order to avoid longer try-and-error
loops.
I am not quite sure how you want to achieve this cleanly.

Something like (picked an actual example from x86):

config HAS_COMPILER_OPTION_IBR
        bool "Select compiler option -mindirect-branch-register"
        default $(cc-option,-mindirect-branch-register)
          blah blah blah


Nice :-)
Definitely i would add a “default y if EXPERT” or something equivalent.

Uh, rather not. I as a developer don't want to have change the config
manually just because a new HAS_COMPILER_OPTION_ has been added my tools
don't understand (yet). The default action should require no user
intervention, even as expert.


Juergen



 


Rackspace

Lists.xenproject.org is hosted with RackSpace, monitoring our
servers 24x7x365 and backed by RackSpace's Fanatical Support®.