I herewith report that the portable branch as distributed by Void Linux
works on Void Linux. (I could not get the 6.6.1p1 that I myself
downloaded to work, though it did compile and install. I have not
looked at what, if any, difference(s) there may be between the two.) My
use of Opensmtpd is simply for myself on my home machine, with a couple
different email accounts on a smarthost.
RPTN
On Tue, Nov 19, 2019 at 08:07:10AM +, gil...@poolp.org wrote:
> Hello,
>
> OpenSMTPD portable was first released many many years ago thanks to the work
> of Charles Longeau.
> IIRC this was his first OpenBSD portable project and he bootstrapped it from
> OpenSSH.
> It served us well and opened us to the Linux world but it also came with its
> share of errors inherent to a new project.
>
> Unless some of you are interested, I won't enter the details, but the basic
> idea is that the portable version has a library which contains the bits that
> were detected as missing/broken on a target system and, because of it's done,
> it may contain a bit more than we think it does. In most cases, this isn't a
> problem because if OpenSMTPD uses the portable version of strwhatever()
> instead of libc's, it isn't really a big deal.
>
>
> There are two cases however that bit us:
>
> The first one was the "OpenSMTPD portable on OpenBSD" case. For years, I kept
> repeating that running OpenSMTPD portable on OpenBSD was not supported and
> that this would lead to unexpected issues. This became a reality at some
> point when building portable on OpenBSD resulted in an executable that would
> crash at startup despite being fundamentally the same code as the master
> branch which would work just fine. It turned out that this was because,
> despite being on OpenBSD, some bits from the compat layer were still built
> and these bits not being the system ones, they didn't comply with our
> pledge().
>
> The second one came shortly before 6.6.0p1 as people started reporting
> crashes on some systems and the issue was traced back to an incorrect
> inclusion of arc4random() related functions. If a target system was using
> LibreSSL, which brings arc4random(), we would somehow end up using the
> LibreSSL one but use our own entropy stirring function and blow up in pieces.
>
>
> There are other downsides to the way the portable layer works and while we
> had work-arounds so people wouldn't notice on systems we were actively
> supporting (Linux, FreeBSD, ...), it is less than perfect for developers
> willing to bring new targets. Freddy Dissaux who worked on bringing OpenSMTPD
> to Solaris hit several of these downsides.
>
>
> Anyways, I have started reworking the layer so that it conditionally builds
> what it needs.
> The compat layer on OpenBSD is now empty as it should and the resulting
> executable runs reliably as it should too.
> There's still work required because code set aside, we also have headers
> issues (for which we have a workaround).
>
>
> I REALLY need feedback from portable users running the development branch.
>
> Ihor Antonov has setup a CI that lets us spot failures to build on Linux
> glibc/musl, however:
>
> - we don't have CI for FreeBSD, NetBSD, DFlyBSD, OSX, Solaris
> - CI won't spot either one of the two issues I solved above as they occur at
> startup / runtime
>
>
> You can help us make the portable branch really good in a very short time and
> on the long term by testing and reporting regularly that the portable branch
> still works for you.
>
>
> Gilles
>
>
--
