Re: Announce: libucontext 0.1.0 - work in progress libc-independent ucontext implementation

[Rich Felker <dalias@libc.org>]

On Tue, Feb 13, 2018 at 09:42:36PM -0600, William Pitcock wrote:
> Hello,
> 
> I am pleased to announce the 0.1.0 release of libucontext, a library
> which implements the ucontext.h functions (getcontext, setcontext,
> makecontext and swapcontext), originally meant for use with gcompat,
> but also useful for applications requiring the functions outside of
> gcompat (such as when building against musl directly).
> 
> Implementation completeness varies based on each architecture, with
> the goal of having complete implementations across all presently
> supported architectures in the next release, but, it should be noted
> that for the most part the implementations provide workable behaviour
> in real-world apps right now.  In other words, it's what you would
> expect for a 0.1.0 release.
> 
> To use these functions, you just link to `-lucontext`, meaning you
> could provide them in $LIBS when running configure scripts and have
> everything most likely work out nicely.
> 
> Download: http://distfiles.dereferenced.org/libucontext/libucontext-0.1.0.tar.xz
> Building should hopefully be straightforward too.
> 
> For Alpine and Adelie distributions, this package is available as
> libucontext in the testing repository.

Looks good. A couple observations:

1. Your implementations don't seem to save or restore signal masks.
This of course makes them much faster and "better" for implementing
coroutines etc., but if applications using them actually expect them
to keep a context-local signal mask, they'll break.

2. Your asm makes effort to save and restore call-clobbered registers.
setcontext does need to restore them if you want to be able to return
to asynchronous contexts (like the ucontext argument to a signal
handler) correctly, but there's no point in saving the call-clobbered
registers in getcontext, since these registers are purely the
getcontext function's "local variables", not a meaningful part of the
context. Likewise the /* we need to restore %ecx because we clobbered
it earlier */ comment in x86 getcontext.S does not make sense.

Rich