Remaining agenda for 0.9.8

Remaining agenda for 0.9.8

[Rich Felker]

PowerPC port issues:
- adding dynamic linking support
- resolving which definition of long double we'll use/support
- checking for remaining omissions/bugs/etc.

Improving app compat:
- integrating sys/mtio.h
- integrating sys/io.h port io stuff
- exposing sigreturn stuff needed by libunwind (?)

In addition, there are a few things still pending that probably won't
make it into this release cycle, but I'd like to keep them in mind
anyway:

- priority inheritance mutexes
- named subarch support (armeb, arm-hardfloat, mips-softfloat, etc.)
- math_errhandling for archs without fenv
- major documentation improvement

Anything I missed, or additional requests? I'm hoping to get a release
out pretty quickly since the last release cycle was rather long, and
this one doesn't have any invasive changes in it that would need heavy
testing before release.

Rich

Re: Remaining agenda for 0.9.8

[Szabolcs Nagy]

* Rich Felker <dalias@aerifal.cx> [2012-11-14 20:55:00 -0500]:
> Anything I missed, or additional requests? I'm hoping to get a release
> out pretty quickly since the last release cycle was rather long, and

i fixed a couple of math bugs

eg. exp10l had constants trucated to double

there were several functions (scalbn, modf, nextafter)
which should be correctly rounded but the return value
was not stored, so they returned a value with excess
precision
(this can cause bad breakage unless the caller saves
the result)

some __invtrig globals are removed

these should be fixed

there are various other improvements

a major change is that i changed STRICT_ASSIGN to
no longer do the volatile thing
(it's not needed on gcc if either -ffloat-store or
c99 standard excess precision handling is present,
hopefully other compilers are not as broken as gcc)

and i added the c11 CMPLX into complex.h
(ifdefed for c11)

Re: Remaining agenda for 0.9.8

[Isaac Dunham]

On Wed, 14 Nov 2012 20:55:00 -0500
Rich Felker <dalias@aerifal.cx> wrote:

> PowerPC port issues:
> - adding dynamic linking support
> - resolving which definition of long double we'll use/support
> - checking for remaining omissions/bugs/etc.
> 
> Improving app compat:
> - integrating sys/mtio.h
Everything it provides is implemented at the kernel level, so you can just drop it into include/sys/

> - integrating sys/io.h port io stuff
> - exposing sigreturn stuff needed by libunwind (?)
> 
> In addition, there are a few things still pending that probably won't
> make it into this release cycle, but I'd like to keep them in mind
> anyway:
> 
> - priority inheritance mutexes
> - named subarch support (armeb, arm-hardfloat, mips-softfloat, etc.)
It may be best to do this via -D/#ifdef, for situations where you need multiple sets of oddities.
I'm thinking of (mipsel-softfloat):
SUBARCH_CFLAGS=-D__SOFTFLOAT=1
SUBARCH=el-softfloat #ld.so = ld-musl-${ARCH}${SUBARCH}.so.1
> - math_errhandling for archs without fenv
> - major documentation improvement
In tree or out-of-tree?
> 
> Anything I missed, or additional requests? I'm hoping to get a release
> out pretty quickly since the last release cycle was rather long, and
> this one doesn't have any invasive changes in it that would need heavy
> testing before release.
For releases that do need more testing, what would you think of rc tarballs? While git can be built with only libc as a dependency* (NO_PYTHON=1 NO_PERL=1 MSGFMT=true... if I remember right), "pull from git" does still limit your audience.

*Either I missed some variable for the installation, or you will also need to install gettext(-tiny) if you don't install with make -i install, because it counts on *.po existing.
-- 
Isaac Dunham <idunham@lavabit.com>

Re: Remaining agenda for 0.9.8

[Rich Felker]

On Thu, Nov 15, 2012 at 10:59:12PM -0800, Isaac Dunham wrote:
> On Wed, 14 Nov 2012 20:55:00 -0500
> Rich Felker <dalias@aerifal.cx> wrote:
> 
> > PowerPC port issues:
> > - adding dynamic linking support
> > - resolving which definition of long double we'll use/support
> > - checking for remaining omissions/bugs/etc.
> > 
> > Improving app compat:
> > - integrating sys/mtio.h
> Everything it provides is implemented at the kernel level, so you
> can just drop it into include/sys/

Yes, I just wanted to clean up the inconsistent spacing and stuff..

> > - integrating sys/io.h port io stuff
> > - exposing sigreturn stuff needed by libunwind (?)
> > 
> > In addition, there are a few things still pending that probably won't
> > make it into this release cycle, but I'd like to keep them in mind
> > anyway:
> > 
> > - priority inheritance mutexes
> > - named subarch support (armeb, arm-hardfloat, mips-softfloat, etc.)
> It may be best to do this via -D/#ifdef, for situations where you need multiple sets of oddities.
> I'm thinking of (mipsel-softfloat):
> SUBARCH_CFLAGS=-D__SOFTFLOAT=1
> SUBARCH=el-softfloat #ld.so = ld-musl-${ARCH}${SUBARCH}.so.1

Yeah I was thinking of something like that. Maybe even
ld-musl-$(ARCH)$(ENDIAN)$(ABIVARIANT).so.1 or similar. The things to
think about are how the source tree needs to look to make whatever
option is chosen work. For example, endianness should normally not
affect which files get compiled (endian-agnostic code is pretty much
always the right way to go), but the ABI variant might affect which
asm gets used (e.g. setjmp asm that saves or skips saving floating
point registers, or real vs no-op versions of fenv stuff). Or we could
just say the same asm is going to get used either way, and arrange for
there to exist some global symbols describing the variant, which the
asm could read at runtime.

> > - math_errhandling for archs without fenv
> > - major documentation improvement
> In tree or out-of-tree?

That's a good question. I've gotten some requests for in-tree. The
pros of in-tree are that you can keep it versioned with the code, and
that everything's all in one place, and it looks more "professional"
in releases to have the docs present. The cons are that it makes the
repo and releases a good bit larger, and looks bad in releases if
parts of the docs are outdated at the time of release (if they're
separate, new docs releases can be made separately from new code
releases).

> > Anything I missed, or additional requests? I'm hoping to get a release
> > out pretty quickly since the last release cycle was rather long, and
> > this one doesn't have any invasive changes in it that would need heavy
> > testing before release.
> For releases that do need more testing, what would you think of rc
> tarballs? While git can be built with only libc as a dependency*
> (NO_PYTHON=1 NO_PERL=1 MSGFMT=true... if I remember right), "pull
> from git" does still limit your audience.

The old gitweb had a "download tarball" link. I don't think cgit has
that (for which my server is thankful). RC's are one idea, but I
wonder if daily snapshots might be just as good.

> *Either I missed some variable for the installation, or you will
> also need to install gettext(-tiny) if you don't install with make
> -i install, because it counts on *.po existing.

I think there's some disable option for it.

Rich

Re: Remaining agenda for 0.9.8

[Isaac Dunham]

On Fri, 16 Nov 2012 02:11:35 -0500
Rich Felker <dalias@aerifal.cx> wrote:

> > For releases that do need more testing, what would you think of rc
> > tarballs? While git can be built with only libc as a dependency*
> > (NO_PYTHON=1 NO_PERL=1 MSGFMT=true... if I remember right), "pull
> > from git" does still limit your audience.
> 
> The old gitweb had a "download tarball" link. I don't think cgit has

BTW, I hope you've updated cgit recently: there's a command execution bug that was recently fixed (as well as another security issue).

> that (for which my server is thankful). RC's are one idea, but I
> wonder if daily snapshots might be just as good.

Possibly, especially with "just after release" fixes.


-- 
Isaac Dunham <idunham@lavabit.com>

Re: Remaining agenda for 0.9.8

[Rich Felker]

On Fri, Nov 16, 2012 at 07:02:26AM -0800, Isaac Dunham wrote:
> On Fri, 16 Nov 2012 02:11:35 -0500
> Rich Felker <dalias@aerifal.cx> wrote:
> 
> > > For releases that do need more testing, what would you think of rc
> > > tarballs? While git can be built with only libc as a dependency*
> > > (NO_PYTHON=1 NO_PERL=1 MSGFMT=true... if I remember right), "pull
> > > from git" does still limit your audience.
> > 
> > The old gitweb had a "download tarball" link. I don't think cgit has
> 
> BTW, I hope you've updated cgit recently: there's a command
> execution bug that was recently fixed (as well as another security
> issue).

The attack vector is malicious repositories. If the repository had
malicious commits injected into it, that would be considerably worse
than somebody obtaining limited access on the webserver. :-)

But I will go ahead and upgrade it soon anyway.

Rich

Re: Remaining agenda for 0.9.8

[Rich Felker]

On Wed, Nov 14, 2012 at 08:55:00PM -0500, Rich Felker wrote:
> PowerPC port issues:
> - adding dynamic linking support
> - resolving which definition of long double we'll use/support

Done.

> - checking for remaining omissions/bugs/etc.

Possibly some issues with clone failing to report errors; still
looking into this. Otherwise, I'm not aware of any _known_ problems
left with the powerpc port.

> Improving app compat:
> - integrating sys/mtio.h
> - integrating sys/io.h port io stuff

Done.

> - exposing sigreturn stuff needed by libunwind (?)

I spent some time looking into what's needed for libunwind to work,
and it seems a good bit more ucontext-level stuff is needed before
there's any hope of using it. So the lack of exposed sigreturn is a
relatively small problem.

> In addition, there are a few things still pending that probably won't
> make it into this release cycle, but I'd like to keep them in mind
> anyway:
> 
> - priority inheritance mutexes

I added stubs for most or all of the related interfaces, I think.
Actual implementations will come later.

> - named subarch support (armeb, arm-hardfloat, mips-softfloat, etc.)
> - math_errhandling for archs without fenv
> - major documentation improvement

These will all have to wait, I think.

Rich