On Fri, Mar 30, 2018 at 4:52 PM, Jon Forrest <nobozo@gmail.com> wrote:
[Larry McVoy said ...]
CMU's Mach, mem,...NOTHING like Sun's
VM system, I mean, nothing.
...
Most research projects suffer because the goal of the people who work on it is to hack on it to get their research done,
Two things...
1.) I think we are in complete agreement wrt to code quality - and I think LM is probably correct that Sun had the first real quality implementation.
2.) Paul's question was a little different and I was able to chat with him about it at lunch today. Until a year or so ago, Paul was a compiler/linker guy and was interested less in the OS extension to support shared libraries, but more when did different tool chains start to support it. i.e. How did we get from a.out to ELF.
The key is that make shared libraries real for the tool chain, you need OS support for shared memories. It turns out that Joy's BSD Unix took its shared memory API (mmap) from Tenex's 'PMAP JSYS' (which BTW, the '72 Tenex CACM paper calls out shared libraries modelled after Multics - I was trying to google the API specific and failed - I think its darned near 1x1). I'm also pretty sure that PMAP influenced the VMS shared memory call that Cutler would build in the mid 70s. Not only that, In Rashid's 81 Accent paper, he too referenced Tenex's PMAP too (more in a minute).
Also, please remember, that BSD's mmap was not the first shared memory call for any UNIX. Horton & Johnson can correct me, but I think Dale's team in Columbus had a shared memory in a PDP-11 ??PWB 1.0?? style system - that would eventually begat the System V shmem call (i've got a hard copy of it somewhere in my archives).
What I do not remember, and this is where someone like Steve might know, was could COFF use the Columbus shmem to do shared libs? Certainly by that time, shared libraries for UNIX was being heavily discussed both in and out of AT&T. I remember that either V.2 or V.3 had a shared lib scheme that put the libraries at a static address as the System V kernel support was not very flexible. But to Paul's question, was that AT&T had started to put shared libs into their flavor around the same time? i.e. when did Summit start to support it.
I was trying to get an account of which versions of which OS distributions/releases had some sorts of kernel extensions that allowed shared libraries and that also supported tool chains that could provide them -- i.e. what Paul was asking about. I was not commenting on who's kernel to support same was good or bad. I did comment on the file format. COFF was a dead end because of the way Summit's management introduced it to larger Unix community, even if it was 'better' and solved some of the earlier issues with a.out.
Macho was already there for instance, and it was an 'extensible' version of a.out and had some level of support for architecture of the binary contained within. How v6/v7 a.out got extended to support shared libs, had varied widely by different manufacturers tool chains. Originally, most manufacturers still called it a.out and just hacked on their own version of the toolchain. CMU gave it a different name (macho) and IIRC made the kernel handle both i.e. still supported the original BSD a.out but you were limited in what could be done with it to only the 4.1BSD api. Remember, Mach 1.0, reacting to Accent' failure, was BSD4.1++ so they wanted to be able to run BSD binaries, at least on Vaxen. In fact the way you installed Mach1.0 was build a new kernel boot it on your 4.1BSD system. Binary compatibility was sort of new concept for the University community --- in fact, around that time at UCB I remember Joy once saying he thought it was good to force people to recompile as it ensured the code was not stale.
So moving forward, AT&T's COFF had tried to repair the many different a.outs, but AT&T made other major non-technical missteps so only firms that were late in the Unix world bothered with it. And even after trying to be 'common', it too, got extended in N different incompatible (uncommon) ways.
As I said, I am not sure how ELF really became the universal format, other than timing. AT&T for whatever reason did not have a fit when it was reimplemented by Gnu - but by the Gnu guys doing that, and the FOSS community switching to that version of the compiler, ELF did win in the long run. And fortunately, the Gnu version of ELF and the SVR4 version seems/appears to be were pretty close [as far as I know they matched, but if someone knows otherwise, I defer to them].
As for the kernel stuff, my other point is that >>idea<< of mmap/pmap was kicking around the Unix community by numerous folks hacking on the kernel for a longer time then when it was finally implemented. LM reminds us that you need something for memory sharing to implement shared libraries. But memory sharing added to Unix before shared libraries was added to the tool chains.
I do find it interesting the UCB mmap API call pretty much was all of the Unix kernel implementation agreed to be the memory sharing API and that it happens to be pretty much the same call as Tenex's PMAP. As Paul mentioned at lunch today for him, VMS, Tenex/TOP-20 and eventually Unix it was pretty much the same call for him implementing the user space linker/loader calls, which is part of why he was asking when did the tools start to support it all.
Clem
ᐧ