__xmknod, __sysv_signal

__xmknod, __sysv_signal

[M Farkas-Dyck]

A dynamic-linked binary blob I have wants these. Are these appropriate
to include in musl? I have a patch ready.

Re: __xmknod, __sysv_signal

[Rich Felker]

On Fri, Apr 18, 2014 at 09:18:33PM -0500, M Farkas-Dyck wrote:
> A dynamic-linked binary blob I have wants these. Are these appropriate
> to include in musl? I have a patch ready.

Send the patch and I'll review it.

For __xmknod, I think this is just the ABI symbol glibc uses for
mknod, and we could certainly add it (though I question whether there
are any useful programs using it that couldn't just be rebuilt against
musl).

For __sysv_signal, I'm not sure what we should do. Its semantics are
different from signal(), but it's doubtful that any program actually
intended to request the sysv semantics (they're fundamentally broken
and have race conditions that renders any program using them buggy) so
it might just make more sense to have it be an alias for signal than
to actually implement the broken sysv behavior. (Most likely is that
glibc's weird behavior with certain feature test macros caused a
program to inadvertently pull in the broken sysv version of signal
rather than the default one.)

Rich

Re: __xmknod, __sysv_signal

[M Farkas-Dyck]

On 18/04/2014, Rich Felker <dalias@libc.org> wrote:
> For __xmknod, I think this is just the ABI symbol glibc uses for
> mknod, and we could certainly add it (though I question whether there
> are any useful programs using it that couldn't just be rebuilt against
> musl).

I wish to build ghc against musl, but it's like making yogurt: I need
ghc to build ghc, and the ghc folks in their great wisdom distribute
binaries... dynamic-linked against GNU turds.

I myself doubt whether we ought to include them and pollute the ABI,
but I thought that others might be in a like situation, which is why I
ask. I am now using ghc with a locally-patched musl, so I hope to not
need these symbols further myself, but the ultimate choice is yours.

---
 src/stat/__xmknod.c | 6 ++++++
 1 file changed, 6 insertions(+)
 create mode 100644 src/stat/__xmknod.c

diff --git a/src/stat/__xmknod.c b/src/stat/__xmknod.c
new file mode 100644
index 0000000..62499de
--- /dev/null
+++ b/src/stat/__xmknod.c
@@ -0,0 +1,6 @@
+#include <sys/stat.h>
+
+int __xmknod(int ver, const char *path, mode_t mode, dev_t *dev)
+{
+	return mknod (path, mode, dev);
+}
-- 
1.8.5.2

---
 src/signal/signal.c | 1 +
 1 file changed, 1 insertion(+)

diff --git a/src/signal/signal.c b/src/signal/signal.c
index c0f063e..29e03c8 100644
--- a/src/signal/signal.c
+++ b/src/signal/signal.c
@@ -13,3 +13,4 @@ void (*signal(int sig, void (*func)(int)))(int)
 }

 weak_alias(signal, bsd_signal);
+weak_alias(signal, __sysv_signal);
-- 
1.8.5.2

Re: __xmknod, __sysv_signal

[Rich Felker]

Sorry for taking a while to review this. I'd like to get it committed,
but a few questions...

On Fri, Apr 18, 2014 at 10:11:02PM -0500, M Farkas-Dyck wrote:
> ---
>  src/stat/__xmknod.c | 6 ++++++
>  1 file changed, 6 insertions(+)
>  create mode 100644 src/stat/__xmknod.c
> 
> diff --git a/src/stat/__xmknod.c b/src/stat/__xmknod.c
> new file mode 100644
> index 0000000..62499de
> --- /dev/null
> +++ b/src/stat/__xmknod.c
> @@ -0,0 +1,6 @@
> +#include <sys/stat.h>
> +
> +int __xmknod(int ver, const char *path, mode_t mode, dev_t *dev)
> +{
> +	return mknod (path, mode, dev);
> +}

I think this should be *dev or something; mknod takes dev_t, not
dev_t*.

Also I don't really like having this in src/stat, but we don't really
have a dedicated place for ABI-compat junk yet...

> ---
>  src/signal/signal.c | 1 +
>  1 file changed, 1 insertion(+)
> 
> diff --git a/src/signal/signal.c b/src/signal/signal.c
> index c0f063e..29e03c8 100644
> --- a/src/signal/signal.c
> +++ b/src/signal/signal.c
> @@ -13,3 +13,4 @@ void (*signal(int sig, void (*func)(int)))(int)
>  }
> 
>  weak_alias(signal, bsd_signal);
> +weak_alias(signal, __sysv_signal);
> -- 
> 1.8.5.2

Probably ok.

Rich

Re: __xmknod, __sysv_signal

[M Farkas-Dyck]

On 30/04/2014, Rich Felker <dalias@libc.org> wrote:
> Sorry for taking a while to review this. I'd like to get it committed,
> but a few questions...

Ah, wasn't sure whether you meant to do this.

>> +int __xmknod(int ver, const char *path, mode_t mode, dev_t *dev)
>> +{
>> +	return mknod (path, mode, dev);
>> +}
>
> I think this should be *dev or something; mknod takes dev_t, not
> dev_t*.

Yes, that seems a typo in [1].

> Also I don't really like having this in src/stat, but we don't really
> have a dedicated place for ABI-compat junk yet...

I don't like it either.

What is musl's general policy on ABI compat? The FAQ says solely that
"musl aims for a degree of feature-compatibility", not what degree. Is
full binary compatibility with glibc the goal?

If we mean to include such, we ought to choose where to keep the code first.

[1] http://refspecs.linuxbase.org/LSB_3.1.1/LSB-Core-generic/LSB-Core-generic/baselib---xmknod-1.html

Re: __xmknod, __sysv_signal

[Rich Felker]

On Tue, May 06, 2014 at 11:55:48AM -0500, M Farkas-Dyck wrote:
> What is musl's general policy on ABI compat? The FAQ says solely that
> "musl aims for a degree of feature-compatibility", not what degree. Is
> full binary compatibility with glibc the goal?

While I don't think it's spelled out anywhere, the hope is to make it
so any strictly conforming POSIX program build against glibc also
works with musl dropped in. Programs using extensions that musl also
provides should work too. Programs using glibc features that musl does
not provide, or depending on glibc bugs, are not intended to be
supported.

> If we mean to include such, we ought to choose where to keep the code first.

Similar things are scattered here and there; see the junk in
src/ctype.

Rich

Re: __xmknod, __sysv_signal

[Rob Landley]

On 05/06/14 22:17, Rich Felker wrote:
> On Tue, May 06, 2014 at 11:55:48AM -0500, M Farkas-Dyck wrote:
>> What is musl's general policy on ABI compat? The FAQ says solely that
>> "musl aims for a degree of feature-compatibility", not what degree. Is
>> full binary compatibility with glibc the goal?
> 
> While I don't think it's spelled out anywhere, the hope is to make it
> so any strictly conforming POSIX program build against glibc also
> works with musl dropped in. Programs using extensions that musl also
> provides should work too. Programs using glibc features that musl does
> not provide, or depending on glibc bugs, are not intended to be
> supported.
> 
>> If we mean to include such, we ought to choose where to keep the code first.
> 
> Similar things are scattered here and there; see the junk in
> src/ctype.


Given that:

#ifdef __MUSL__
#include <gnuisms.h>
#endif

is off the table because musl is the platonic ideal C library, anonymous
and infinite, and the idea of having musl-specific idiosynrasies that
aren't necessarily considered the default (let alone building software
for other libraries with 90% market share and wildly different behavior)
will never come up in practice...

There's a gnu extension called #include_next that lets you have a second
set of headers inserted before the first set. It is, of course, a gnu
extension.

The library side seems easy enough to deal with via some sort of -lcrap
added to the build, if you don't want it in libc itself.

Or you could have a "./configure --legacy" to selectively enable this
sort of stuff in libc.so and the headers, and then require people to use
something like autoconf to determine whether or not this instance of
libc was built with that because a #define would be coddling them.

Rob

Re: __xmknod, __sysv_signal

[Rich Felker]

On Wed, May 07, 2014 at 05:18:30AM -0500, Rob Landley wrote:
> On 05/06/14 22:17, Rich Felker wrote:
> > On Tue, May 06, 2014 at 11:55:48AM -0500, M Farkas-Dyck wrote:
> >> What is musl's general policy on ABI compat? The FAQ says solely that
> >> "musl aims for a degree of feature-compatibility", not what degree. Is
> >> full binary compatibility with glibc the goal?
> > 
> > While I don't think it's spelled out anywhere, the hope is to make it
> > so any strictly conforming POSIX program build against glibc also
> > works with musl dropped in. Programs using extensions that musl also
> > provides should work too. Programs using glibc features that musl does
> > not provide, or depending on glibc bugs, are not intended to be
> > supported.
> > 
> >> If we mean to include such, we ought to choose where to keep the code first.
> > 
> > Similar things are scattered here and there; see the junk in
> > src/ctype.
> 
> 
> Given that:
> 
> #ifdef __MUSL__
> #include <gnuisms.h>
> #endif
> 
> is off the table because musl is the platonic ideal C library, anonymous

There is no such ideological reason. As explained in the FAQ and many
times before, the reasons it's not provided are purly practical:

1. Every single time somebody has requested it, they were trying to do
something wrong, and in most case, what they were doing would have
broken support for subsequent versions of musl by hard-coding an
assumption that was about to become incorrect. (So, providing such a
macro would have left us getting the package fixed twice or more times
rather than just once, and would have made us look bad for "breaking
things" despite it being the package maintainer's fault.)

2. Hard-coding idiosynchracies of a particular platform is wrong
unless they are bugs that you know the platform has and which the
vendor has refused to fix (see: Windows), and which cannot be detected
at build time. In all other cases, the correct behavior is to detect,
either with configure-type checks, or with macros indicating the
presence of particular features/conformance profiles (like the
_POSIX_* macros from unistd.h).

> and infinite, and the idea of having musl-specific idiosynrasies that
> aren't necessarily considered the default (let alone building software
> for other libraries with 90% market share and wildly different behavior)
> will never come up in practice...
> 
> There's a gnu extension called #include_next that lets you have a second
> set of headers inserted before the first set. It is, of course, a gnu
> extension.
> 
> The library side seems easy enough to deal with via some sort of -lcrap
> added to the build, if you don't want it in libc itself.
> 
> Or you could have a "./configure --legacy" to selectively enable this
> sort of stuff in libc.so and the headers, and then require people to use
> something like autoconf to determine whether or not this instance of
> libc was built with that because a #define would be coddling them.

This is useless. The symbols in question are ABI-only, not API. You
never use them when compiling new binaries. The only way the reference
to these symbols ever comes to exist is by virtue of glibc mess in
their headers and libc_nonshared.a, which redirects certain standard
functions to use these glibc-specific interfaces to libc.so rather
than just using the public names.

Rich