On Fri, 8 Jan 2021 at 19:43, Warner Losh <imp@bsdimp.com> wrote:
>
> The FreeBSD project has rewritten bsd-user. We've been working on this for quite some time (the earliest commits date from 2013). Maybe a dozen people have worked on this over time, and there's 3 or 4 active developers focused on FreeBSD changes at the moment.
>
> For a while, we'd merge in upstream changes from qemu. This worked great for us, but left us with a big backlog that was hard to upstream. Each of the updates took some time, so we got a little behind.
>
> So, a few years ago, I spent several weeks converting the tangled merge mess into a set of linear patches and started moving that forward. This was around the time 4.0 was released. I only managed to get the rebase forward to 3.1 release at the time before I hit problems related to poor testing environment making it hard to verify newer versions were still working. Plus, we found a few bugs that took a while to resolve for a number of reasons. Now that they are resolved, we're able to use qemu-bsd-user to build ~30k packages for arm, and ~20k for different types of mips in FreeBSD "ports" system. We now have great confidence that it's working well again.
>
> Now that those bugs are resolved, I started trying to forward-port the two-year-old base and immediately found myself hitting a number of problems. A big problem was that I was re-doing a lot of work that was due to innoculous changes upstream that I wouldn't have to do if the bsd-user changes were upstream. These changes get in the way of dealing with the more substantial structural changes in qemu that have happened.
>
> There had been talk of doing a remove and replace update of bsd-user. This talk was before I managed to rebase things as far forward as 3.1 even. This appealed to me because we've accumulated about 150 patches to date, many quite large, and curating them into a set of maybe 400 or 500 changes to match the size and scope of most patches I've seen posted to qemu-devel seemed overwhelming.
>
> However, it's been another year since that plan was hatched, and it's become clear to me that plan won't end in success. The closest I've been able to get is 3.1 when 4.1 was current (about 6 months behind). It's time for a new plan.
>
As one of the developers on the FreeBSD side, I agree with this. In
addition to the reasons cited, reintroducing it would really take a
lot more time and effort and I'm not convinced it would ever be
completed because, IMO, going that route should really entail
redesigning it from the ground-up based on an abstraction of
linux-user. I find that right now I'm playing a lot of catch-up
because we seem to be largely copied from linux-user without later
improvement; there is a really healthy amount of platform-independent
stuff that really should end up ultimately shared in a qemu-user or
something to that effect to mitigate duplication of effort.
Yes. Once we have things upstream the normal 'refactoring' work can be done there, rather than in a side repo where it will quickly go stale. Also, once it's in upstream, then the NetBSD and OpenBSD folks can come back to the table with a much smaller delta to get things working again than they have today.
> So, my new plan is to rebase what changes I can to the tip of master and submit those for review. I'll work with the developers on the FreeBSD side to ensure they are included in reviews in addition to the normal qemu-devel list. This will allow us to pare down the deltas between our code and upstream to allow us to make progress. The changes will be held to the standard 'makes things better'. Given how broken bsd-user is today in qemu upstream, at first that will a very easy standard to make.
>
> The first patch I'll submit will be changing MAINTAINERS to point to me, since I'm acting as the point person in this effort. I'll then re-submit some other changes that I've submitted in the past, but CC the FreeBSD folks that are currently active (they were only CC'd to former developers who lack the time to review).
>
Thanks for taking this on!
You bet.
Warner