06 April 2013

NetBSD support upcoming

My first real UNIX ever was a NetBSD. That was now about 22 years ago. I am still sentimental about it. I had installed it last about 8 years ago. And I still like it. Back in the days, it was the first UNIX to encounter for me, running on Amiga hardware, first of a friend, then on my own.

Recently, there had been support for Nuitka on FreeBSD added. A lot of people use it on the web, and some want to use Nuitka to improve their Python performance, so this is kind of relevant.

There were issues resolved, but in the end, something was with Clang on FreeBSD 8, that I could for the life of it, not resolve remotely. So I attempted to install it myself. Using “virt-install”, these things are a breeze now. I had already done it with CentOS6 before to test the RPM repositories of Nuitka. That “virt-install” is a wonderful thing by itself, making virtualisation somewhat useful. It’s only a pity, that I can’t just install other qemu support architectures. I would love to checkout Nuitka on PowerPC.

Note

If you could check out Nuitka on other Linux architectures than x86_64, x86, or arm, that would be great.

This is the report of getting NetBSD supported. It was a quite an interesting story that I would like to share with you.

Naivly I was assuming, that it would be just for fun, and that Nuitka will work right away. Little did I know.

On FreeBSD 9 the minimal install medium was chosen, and entered its ports collection, installed git, cloned Nuitka, and ran the tests, successfully right away. Now that is unfair, in the Nuitka there were tons of “Linuxism” already removed. In fact, it had to work, and on the newest FreeBSD (version 9.1) and then it did. Great!

Note

If you would like to add Nuitka to FreeBSD’s ports, please do so. It should be really easy now.

On NetBSD, things were unfortunately a little different. I also chose minimal system. After going through “pkg source” boot strap and git install, I cloned Nuitka, and then tried to start it. First off, it couldn’t locate “python” at all. I am using /usr/bin/env python already. But Python2 was on the system. I ended up creating the “python” link myself. What I should have done according to “#netbsd” is to install the software, and indeed, python2.7 setup.py install gives an installation of Nuitka that is executable.

Next up, you need to know about “Fibers” in Nuitka. These are used for C co-routines, used to implement Python generators. They have an interface, that is very close to makecontext/swapcontext routines in C.

For ARM and x86_64 Linux we have optimized code, that switches faster, but other platforms, including x86 Linux, use the generic implementation, also because it normally is very fast already.

Now you have to know that since 2001 the interface is deprecated and shall not be used. And next up, is that on NetBSD, makecontext gave a segfault only. So I ran to “#netbsd” and asked.

Now that was a very friendly experience. Of course, I had to give a rationale for using an obsolete interface. It’s not quite obvious, why threads wouldn’t be a better choice. And maybe they are, but they definitely have more overhead associated, and if they never run at the same time, why use them.

Ultimately it helped to point out, that for a user of 22 years, an interface that is only obsolete for 11 years, is not quite as horrifying as for others.

And they helped me through it. And it turns out, funny thing. For the context to setup, you are allocating a stack to use for the C routine, and you “get” the current context, then you make a new one. All the examples have a certain order of doing it. And my code did it the other way around. No system but NetBSD noticed.

On FreeBSD and Linux, it didn’t matter. But it seems, that the needed getcontext call was overwriting my stack pointer now with the current stack. And makecontext deeply hated that, a lot. It was preparing that stack to be used, while it was in usage. Doesn’t sound like a good task to give to it, right? My fault truly, because every example on every man page, on all systems, was doing it differently. But then they were also all using arrays from the local stack, so quite obviously that was not real code.

So that was fixed, and all good? No! Next thing was it crashed when free happened in Python on a compiled frame object, in a later part of a test that heavily uses generators. Turns out, malloc information was corrupted. I had to suspect the generic “Fiber” code, but that took me a while to figure out.

And how could my simple malloc and free do that, and make it happen. When I knew that a context would not longer be used (the generator has finished, the generator object deleted, etc), I would look at the context handle stack pointer and free it.

But that pointer changed. Something totally unexpected (by me obviously), but it also explains the earlier problem. For all systems, I had used so far, this pointer was not being changed, and remained the same. So I could free it from there. It worked fine, but not on NetBSD. And it wasn’t correct anywhere.

It seems NetBSD is doing something clever, since instead of saving the stack pointer register in a separate area, it saves it to that place originally specified. It’s quite obviously an improvement, in that you save the pointer.

It’s only bad, that now to make up for this savings, I have added the pointer in a separate field, which won’t be changed, so I can free it again. If one needs it again, and that’s not unlikely, you have to remember it elsewhere. So maybe that idea is not that clever. But it surely was wrong by me to assume that the provided value would not be touched.

So, these are 2 bugs it found. The wrong order of calls. And the usage of a pointer, that may have been changed. This can only help with other systems, or possibly architectures under Linux.

While this is all description of nasty problems, it’s also the report of the solution, and it was big fun. I would also like to compliment “#netbsd” for being very helpful and friendly with my porting of Nuitka. I highly enjoyed doing so. It was a lot of fun. I know that it’s probably on a very tiny amount of people that uses both NetBSD and Nuitka, but still.

If this Nuitka project were about market share, it wouldn’t exist. And I can work for market share on another day.