Re: uapi headers userspace build results

From: Randy Dunlap
Date: Tue Jun 19 2018 - 14:17:28 EST


On 06/18/2018 06:47 PM, Jayant Chowdhary wrote:
> Hi Randy,
>
> On 06/12/2018 05:07 PM, Randy Dunlap wrote:
>> On 06/12/2018 01:39 PM, Jayant Chowdhary wrote:
>>> Hi Randy,
>>>
>>> On 06/11/2018 10:49 PM, Randy Dunlap wrote:
>>>> Hi,
>>>>
>>>> Here is what I have so far. It begins with a makefile and some
>>>> template files that are added to. There's a good bit of Perl also.
>>>>
>>>> I put all of these files in tools/uapi/ and run them from there.
>>>>
>>>> There is one .c file generated for each .h file in builddir/usr/include
>>>> (O=builddir).
>>>>
>>>
>>> Thanks for this! I wrote a small Makefile (uapi-compile.mk) which I'd put in
>>> tools/build (I can change this to tools/uapi, if that is more apt).
>>
>> Your makefile foo is much better than mine is.
>> Yes, I think that it deserves to be in its own sub-directory.
>>
>>> uapi-compile.mk straight-away compiles the uapi headers, without pulling them
>>> into any generated c source files. It may also be invoked with an environment
>>
>> Hm, I didn't even know that is possible.
>>
>>> variable 'UAPI_DIR' specifying the directory, for which the user would like to
>>> compile headers. This way we can test a directory at a time as well. In your
>>
>> Yes, good, I was planning to make a way to restrict the build to certain sub-dirs.
>>
>>> opinion, would this be simpler to have rather than having to auto-generate c
>>> source files including each uapi header and also autog-enerating the make
>>> targets? I feel like this approach would make maintaining these makefiles/
>>> scripts easier as well.
>>
>> Sure, this is much better than my scripts.
>>
>>>> Out of 889 header files, I see 45 errors. That is better than I expected.
>>>>
>>>> The makefiles and scripts are attached (tar), as well as the output (I used
>>>> 'make -ik' so that make would keep going after errors and attempt to build
>>>> all target files).
>>>>
>>>> have fun!
>>>>
>>>
>>> I did a 'make ARCH=arm64 headers_install' from the kernel source's root, and
>>> then a 'make -kf uapi-compile.mk all > build.log 2>&1' to compile all the
>>> headers. Out of 864 headers, I see 20 compilation failures.
>>>
>>> I'm attaching uapi-compile.mk and the build.log file along.
>>
>> I have some usage comments.
>>
>> Since I ran 'make ARCH=x86_64 O=xx64 headers_install', I had to modify
>> uapi-compile.mk to use that SRC_DIR:
>>
>> SRC_DIR :=../../xx64
>>
>> Also, I first tried to make BDIR as a sub-directory of tools/uapi/ and
>> uapi-compile.mk did not work (when using BDIR=BDIR).
>> Then I did 'mkdir ../../xx64/BDIR' and specified BDIR=../../xx64/BDIR and
>> that worked. But: that sub-dir is not used:
>>
>> gcc -I../../xx64/usr/include/ --include=../../xx64/usr/include/linux/posix_types.h --include=../../xx64/usr/include/asm-generic/ipcbuf.h --include=stdarg.h --include=stdint.h --include=stddef.h -c ../../xx64/usr/include//linux/caif/caif_socket.h -o ../../xx64/BDIR/../../xx64/usr/include//linux/caif/caif_socket.o
>> [see the next comment]
>>
>> Oh, this makefile builds the .o files in the same sub-dirs as their
>> respective .h files. I don't especially like that, but as long as
>> make clean works, it will do. [and make clean does work]
>>
>
> Thanks for these comments. I'll take care of them in my patch-set. I've got a
> couple of questions for you. Since most of the errors were found in the
> include/uapi/linux directory, I tried investigating why.

Please also repost your latest patch-set.

> 1) I found that multiple headers depend on the definition of types such as
> pid_t, which have no definition in the set of uapi headers. There is a
> definition (of pid_t) in include/linux/types.h, and I thought we could try
> exposing that in the set of uapi headers. One problem I can see with that is
> that the header has some definitions which depend on kernel configs: eg:
> CONFIG_ARCH_DMA_ADDR_T_64BIT. Since user-land programs shouldn't really assume
> kernel configs, I was thinking we should re-factor this header so that
> appropriate parts can be exposed to user-land.

Sure, that's worth a try. Mostly on a case-by-case basis.

I see that (at least in the distro that I am using)
/usr/include/asm-generic/posix_types.h has a typedef for __kernel_pid_t.
I wonder if that could be co-opted, but I expect that this would have
type/size issues.

OTOH, 'man getpid' uses pid_t and refers to <sys/types.h> and <unistd.h>,
so there should already be a pid_t for userspace. Just #include more
headers files. :)

> 2) Some headers try to expose information which should probably not be exposed
> to user-land. eg: wait_queue_head in linux/coda_psdev.h (this header should
> probably be removed altogether ?)

I suppose that header file describes a kernel-to-userspace ("Venus") interface,
so maybe not removed altogether. But the wait_queue_head_t part of it should
just be some padding/reserved field (of what size/type?).


> Do you have better ideas ?

slow and steady.

--
~Randy