Hi Thomas, Zhangjin,
On Mon, Jul 10, 2023 at 05:51:39PM +0200, Thomas Weißschuh wrote:
On 2023-07-10 15:23:40+0800, Zhangjin Wu wrote:
On Sat, Jul 08, 2023 at 11:26:42PM +0800, Zhangjin Wu wrote:
[..]
As a preparation, this creates the architecture specific directory and moves tools/include/nolibc/arch-<ARCH>.h to tools/include/nolibc/<ARCH>/arch.h.
I'm sorry but I still don't understand what it *provides*. I'm reading it as "we *can* do this so let's do it". But what is the specific purpose of adding this extra directory structure ? It's really unclear to me and worries me that it'll only result in complicating maintenance by adding even more files, thus even more "include" lines and cross dependencies.
Willy, I was assuming you had a look at the discussion between Thomas and me, so, I didn't add the link to our discussion, it is more about the 'clarity' of code "include" [1].
I have proposed the idea in the discussion but got no response yet, so, sent this revision for more discussion, obviously, it is better to discuss more there and get more feedback from Thomas and you.
To be honest I got overwhelmed at some point and instead of figuring out to which series' I already responded and which not I only responded to those where I had time to do so immediately.
Keeping the amount of in-flight serieses lower would help this.
+1 on this. First it's difficult for me to assign contiguous time on the subject so I can't grasp all series at once, and I'm terribly bad at context-switching, which takes even more time and induces confusion. Less topics at once, more focused with less reviews of reorganizations will definitely help.
The v0 included "crt.h" before "arch.h", Thomas suggested me include "crt.h" in arch_<ARCH>.h, just like the "compiler.h" did. His suggestion did inspire me to think about how to treat the relationship among crt.h, sys.h and arch.h.
The idea behind is we have many directions to divide nolibc to different parts/modules:
Again above I'm seeing an opportunity but no explanation of why this is needed. Thomas already mentioned that you're speaking about just trying to factor out a few tens of lines. I'm not seeing *why* we need to re-split everything yet again.
one is arch specific (arch.h) and non-arch specific (the others)
This method is used by us currently, It is very good to put all of the arch specific parts together to simplify (in the files to be added/maintained) the porting of a new architecture.
But to be honest, It also confuse the modularity a little, for example, like sys.h, crt.h should be a core function/feature of nolibc, arch.h is not so. arch.h only provides the necessary minimal assembly "pieces".
But that's precisely the principle: keep arch-specific stuff as minimal as possible, keep most of the rest generic but easily overloadable if needed as we know that archs are not all 1:1 equivalent.
both sys.h and crt.h are not a sub modules of arch.h (although they have minimal arch specific code), so, like sys.h, crt.h should be included in the top-level headers, not in arch.h,
Why ? Keep in mind that these are only include files, to in the end, *all* of them are included. The ordering is the only thing that really matters.
reversely, the minimal arch specific should be included in crt.h. To do so and to avoid include the non-crt part, the split of arch.h is required, and therefore, the <ARCH>/ is created to put the divided <ARCH>/sys.h and <ARCH>/crt.h, otherwise, there will be many sys-<ARCH>.h and crt-<ARCH>.h in the top-level directory of nolibc.
Then doesn't it prove that you don't need that crt-<ARCH>.h and that instead it should just be in arch-<ARCH> like the rest of the same arch ?
another is the parallel functions/features (like crt.h, sys.h, stack protector ...)
This is used by musl and glibc, before sending this proposal, I have taken a look at both of them, musl is simpler and clearer, we apply the similar method:
musl: crt/crt1.c #include "crt_arch.h" /* arch/<ARCH>/crt_arch.h */
In musl crt_arch.h seems to be used in different ways. So it makes sense to split it from syscall_arch.h. In nolibc there is no such distinction. And everything will end up in a global namespace anyways.
Exactly. Musl is musl and nolibc is nolibc. Musl is a regular libc in that it provides a .so that is built from many .c files. As such it's desirable to split along certain edges. nolibc contains no single C file. It's only meant to be included as-is in the user's C file. This changes a lot of things, even in terms of splitting. Also keep in mind that musl is a general-purpose libc, and that some distros are entirely built on it. nolibc doesn't have such goal nor expectation, the first user was a preinit code I wrote long ago, and the second one is rcutorture which contains a while() loop around gettimeofday() IIRC. We must not just blindly imitate other components' choices because they work, when we're dealing with different constraints. If ours are acceptable, no need to complicate everything.
With this method, the "crt_arch.h + crt.h" together provide the C RunTime (startup code, stack protector, environ and _auxv currently) function, the "sys_arch.h + sys.h" together provide the syscall definitions. The arch specific parts are hidden behind, and only require to include the crt_arch.h in crt.h and sys_arch.h in sys.h, no need to include the whole arch.h for all.
Everything is included all the time. *everything*. The more files we create, the more "#ifdef FOO_H" gets evaluated, and the more maintenance burden it adds.
As a summary, the core obvious reason here is, to this crt.h itself, it is ok for us to include crt.h in arch.h in code side, but reversely, I do prefer to include arch.h (and therefore the crt_arch.h) in crt.h, crt.h is the core function should be exported, arch.h is not, it only provide some low-level helpers for crt.h. If we treat sys.h as a core function and crt.h as a arch specific thing, it does confuse a little. This reorg may also help the similar future functions who require arch specific support, but of course, it does require to add/maintain more files for a new architecture, but it also allow to develop/debug at a smaller fineness.
In current stage, include crt.h in arch.h is not that unacceptable, but
Why would it be more unacceptable in the future?
if reorg is a better direction, why not do it currently, because we do have two functions (crt.h and sys.h) in <ARCH>/, if only one, it is not urgent ;-)
Is this explanation better than before? welcome to discuss more ;-)
Personally I'm not convinced :-)
Neither am I.
The arch-specific code in nolibc in mainline is currentl ~200 lines per arch. With this series in general it will be even less. If at some point there are many more architectures it may make sense to introduce an arch/ directory then.
And even then I'm not convinced because the number of archs will remain low anyway.
Like musl, if required, another top-level arch/ may be required to put all of the <ARCH>/ directories together to clean up the top-level nolibc directory.
At the moment in mainline there are 26 files in nolibc. That does not seem excessive, in fact it looks to be less than most other kernel directories.
Indeed :-) Note that it started with a single one!
As explained in another reply, it is really hard to write a just ok commit message for every change, sometimes, the justification is 'obvious' to some develoers who have the background information or who have dicussed together, sometimes,
Sometimes yes, but most of the series come with propositions to improve something. The commits making the major changes (and the cover letter) should justify why this is a desirable change, what it implies not to have it and what it may imply to have it, what possible alternatives were considered and dropped sometimes (e.g. when hesitating between two approaches), etc.
Thanks, Willy