On 11/5/21 9:57 AM, Michael Olbrich wrote:
Hi,

On Thu, Nov 04, 2021 at 11:06:13AM +0100, Christian Melki wrote:
Bump.

Right, this got lost on my side, sorry.


Np. :)

No more comments? Action?

On 10/15/21 3:34 PM, Christian Melki wrote:
On 10/15/21 3:12 PM, Michael Olbrich wrote:
Hi,

On Wed, Oct 13, 2021 at 09:29:24AM +0200, Christian Melki wrote:
I've been correcting various hardcoded paths in ptxdist packages
and whatnot
lately. This started with ptxdist not coping well with toolchains that
adhere to the ABI path for various architectures.

Yesterday, I dug into the problem that on x86_64 (/lib64), all
.pc.in -> .pc
file transformations looked broken with hardcoded paths. At
first I thought
it had something to do with autoconf or pkg-config but after a
while I found
this.

scripts/lib/ptxd_make_world_install_mangle_pc.awk

Which after a while made me realize that there is still a lot of code in
ptxdist core stuff that assumes that lib-paths are only /lib and nothing
else.

So. I'm presenting a two options here.

1. Fix all ptxdist core stuff, because really, ptxdist should be more
flexible than hardcoded paths. Esp. for libs.

2. Split ptxd_get_lib_dir, because, ld.so path should not be
assumed to be
the same as main library install path. So ptxd_get_ld_lib_dir which does
what it does today and install ld there and ptxd_get_lib_dir =
/lib and be
done with all the userspace library transformations.

So, put everything in lib/ and only the ld.so in lib64/, right?


Yes. And depending on toolchain configuration.. + glibc, but not the
rest of the userspace libs as you say. Since the copy_toolchain routines
query the linker by itself afaiu.

Good point about copy_toolchain. I don't mind if glibc (etc.) is installed
in /lib64 as well. Whatever is easier to start with. We can change that
later if necessary.


Ok. I'll prepare a patch for hardcoding ptxdist to /lib for non toolchain related things. I will be naive/stupid/simple, but the discussion about configurability etc can start there.

Either way, the important bit is the hardcoded path to ld.so in the ABI
declaration. That's the one you have to hit, otherwise you'll have
bricked userspace. :) The rest can be solved with ld.so.conf.

So /usr/lib is not in the default search path and /etc/ld.so.conf is
necessary in this case?


Hmm. If we do nothing then I guess that depends on the toolchain. If the toolchain is configured with an extended "user-defined-trusted-dirs" option you probably don't need an ld.so.conf for the default search paths. But otherwise, I'd say yes? Can't remember. I use an ld.so.conf that always includes the usual suspects and prepare a cache when I boot (after ld.so has been hit).

To me it looks like glibc expects the default usr search path to be the usr/ld.so path (/lib64, /usr/lib64)
This is what it looks like during glibc generation:
---
(glibc) Makefile:
# Build a file mentioning all trustworthy directories to look for shared



# libraries when using LD_LIBRARY_PATH in a setuid program. The user can


# add directories to the list by defining $(user-defined-trusted-dirs)



# before starting make.



$(objpfx)trusted-dirs.h: $(objpfx)trusted-dirs.st; @:
$(objpfx)trusted-dirs.st: Makefile $(..)Makeconfig
        $(make-target-directory)
echo "$(subst :, ,$(default-rpath) $(user-defined-trusted-dirs))" \
        | $(AWK) -f gen-trusted-dirs.awk > ${@:st=T};
        echo '#define DL_DST_LIB "$(notdir $(slibdir))"' >> ${@:st=T}
        $(move-if-change) ${@:st=T} ${@:st=h}
        touch $@

---
(non-extended, generated glibc header, on x86_64) trusted-dirs.h:
#define SYSTEM_DIRS \
  "/lib64/" "\0" "/usr/lib64/"

#define SYSTEM_DIRS_LEN \
  7, 11

#define SYSTEM_DIRS_MAX_LEN     11
#define DL_DST_LIB "lib64"


So. Number one probably requires a lot of more work and a lot of
headache.

And we'll probably break it more often.

Pretty much guaranteed, yes.

Number two should be rather straightforward, atleast in theory.

There will be some packages that will use lib64/ anyways, because they
query the toolchain directly. We have a platform-foo/sysroot-host/lib64
symlink because of that.

Any thoughts?

I actually really like option 2. It sounds like a lot less work and
problems in the long run.

While I'd like a more generic solution, I can absolutely value things
that works and are simple. So, less problems it is. :)

:-)

If necessary, we could also make lib64/ a symlink to lib/ (in
ptxd_make_world_install_prepare() and in the final rootfs). That way, it
wouldn't matter which one the packages are using.
But I'd like to avoid that if possible.

Mmm. I botched my rules to always include a separate lib64 now, but
that's just a hack.

Michael


_______________________________________________
ptxdist mailing list
[email protected]
To unsubscribe, send a mail with subject "unsubscribe" to 
[email protected]

Reply via email to