Follow @Openwall on Twitter for new release announcements and other news
[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Sat, 5 Mar 2016 11:30:34 -0500
From: Rich Felker <>
Subject: Re: [PATCH] slim down and avoid undefined behavior in unsetenv

On Sun, Feb 21, 2016 at 12:51:30PM +0300, Alexander Monakov wrote:
> >  		errno = EINVAL;
> >  		return -1;
> >  	}
> Some places in musl tail-call to __syscall_ret(-Exxx) to set errno. I wonder
> if it's accidental or there's a guideline for using one style or the other?

A large portion of the source files in musl are written purely with
public APIs. This is especially desirable for legacy functions, where
there's no reason to optimize them and where being maintenance-free is
the most valuable property, but it's also nice for other files from
the standpoints of maintenance, ease of reading, and ease of reuse in
other software (as drop-ins for systems lacking working versions of
these functions) or libcs.

With this in mind, __syscall_ret is mainly used in places where the
surrounding code already deals directly with making syscalls or does
other things that necessitate use of libc-internal APIs. This might
not be followed everywhere as a strict rule, but it's the intent.

> The only place I imagine the tailcall style might be undesired is sem_trywait,
> where returning failure is not expected to be rare.
> What do you think about a change that introduces __set_errno that accepts
> positive errno and returns -1L? With that change __syscall_ret can become
>     return r < -4095UL ? r : __set_errno(-r);

I don't see much value to it; return __set_errno(e) is functionally
equivalent to return __syscall_ret(-e) and just saves one "neg"
instruction before the tailcall. It might be a nicer abstraction in
places where we don't use syscalls directly, but for the most part
musl avoids internal APIs in those situations anyway, as explained

> > -again:
> [snip]
> > +	for (char **e = __environ; *e; )
> > +		if (!memcmp(name, *e, l) && l[*e] == '=') {
> Here the usage of memcmp requires that it scans buffers left-to-right and
> stops on first mismatch. As I understand the standards do not guarantee that,
> but musl's current implementation does, and is not interposable. Still, a
> gotcha.

I think you're right. Would it be better to switch such usage to
strncmp then?


Powered by blists - more mailing lists

Confused about mailing lists and their use? Read about mailing lists on Wikipedia and check out these guidelines on proper formatting of your messages.