Re: "make hierarchy" from main tree breaks c++ headers on older branches with unclear errors if "make install world" is interrupted

From: Enji Cooper <yaneurabeya_at_gmail.com>
Date: Tue, 30 May 2023 23:07:42 UTC
> On May 28, 2023, at 3:53 AM, Dimitry Andric <dim@FreeBSD.org> wrote:
> 
> On 28 May 2023, at 07:18, Enji Cooper <yaneurabeya@gmail.com> wrote:
>> 
>> I just tried to run “make hierarchy” from a main tree on a 13.2-RELEASE system, and doing so completely broke my headers.
>> It took me about 30 minutes to figure out what happened… /usr/include/c++/v1/__string was a header, whereas on :main it’s a directory?!
> 
> Yes, upstream libc++ has split up large headers into multiple components, in particular __string and __type_traits.
> 
> Since we had a file named __string, it had to be somehow replaced by a directory. This is what the distrib-cleanup target in the top-level Makefile does:
> 
> https://github.com/DimitryAndric/freebsd-src/commit/6b13b4a095e3
> 
> Afterwards, the replacement directory named __string is created as part of the regular mtree commands.
> 
> 
> ...
>> I reinstalled the headers by going to lib/libc++ in my releng/13.2 tree, building, and installing all of the 13.2-RELEASE headers.
>> This change has been live for almost a year now on :main — is this a known caveat when doing a source-based upgrade from 13.2-RELEASE* to 14.0-CURRENT, i.e., that the installworld (if interrupted) could break the c++ compiler?
> 
> I think it should already be quite clear that interrupting installworld is risky. If you were just in the middle of replacing libc or rtld, and those were half-written, your system will be completely hosed. (I am speaking from experience. :) The same goes for any system headers or libraries, not only the C++ ones. If you have a half-installed tree, it should not be used for anything except attempting another installworld.

That’s a fair point, however, given that this follows the standard installation workflow…
1. etcupdate pre-run.
2. install kernel
3. reboot
4. make installworld
5. etcupdate post-run
… this could surprise end-users. In particular, the C++ compiler will be broken between step 1 and step 4.

It would be good to bring this fact up in UPDATING or RELNOTES as a warning because it could result in a broken system that’s difficult to recover from.

Cheers,
-Enji