On Thu, Feb 16, 2023 at 12:40 PM David Woodhouse <dw...@infradead.org> wrote:
> On Thu, 2023-02-16 at 09:29 -1000, Richard Henderson wrote: > > On 2/16/23 09:02, David Woodhouse wrote: > > > It wouldn't be beyond the wit of man to extend qemu-user to support the > > > similar personality variations for SCO/Solaris/etc. using that as a > > > guide. > > > > Not beyond wit but perhaps beyond patience. > > > > It would certainly be possible to emulate the "easy middle" of one POSIX > guest on a > > different POSIX host. But the dusty corners are going to get in the > way, where we > > currently rely on guest and host having identical semantics, and pass > the system call > > through to the host. > > > > It's a big job. > > True, but the existing iBCS / linux-abi kernel patches should highlight > a lot of those dusty corners. > So one thing to understand, the iBCS is a separate ABI. This means that you'd have to rewrite everything from the syscall dispatch on down. Even if this were relevant, it would be a huge job. A lot would depend on how much of Solaris is used. Solaris is ELF, which is good (it would be even worse if it were SunOS 4, then it's a.out and a whole lot of other complication that's more of a bsd-user thing). However, as others have pointed out, linux-user assumes a linux kernel. While one can run linux-user on FreeBSD with its Linux ABI implementation, even that's quite limited in what it can do (I needed to do this for some kexec support I was adding to FreeBSD boot loader that ran as a Linux binary). I had to make tweaks to FreeBSD's emulation to make it work, and that was for a binary that used only 10 system calls, no threads, no signals, nothing "messy" and apart from some extensions to 64-bits, nothing that wasn't in 7th Edition Unix. And there's also a number of special filesystems, IIRC, on Solaris that are used like linux's /sys and /proc filesystems, but with different details. And a million other details. Knowing the details isn't enough, assuming you could know them from cribbing from existing code. You have to actually go implement the details and that would be a very tedious job. Even if you kept it to a subset that your program uses... I started on a Venix emulator (ancient Unix V7 port to 8088/8086 micros I cut my teeth on), and even that was daunting. Now, with 3 years of bsd-user hacking and upstreaming under my belt, it would be easier, but there's a *HUGE* learning curve to understand the CPU, its exception model, how system calls are handled, how memory is mapped, etc. And the 'assume we're on linux' is definitely leveraged for memory mapping in the existing linux-user code if you were to copy it all as a starting point. When people say it's a big job, they are underselling it somewhat. It would be a big job for the maintainers of linux-user who have all the context and know where the gotchas are. For anybody else, learning everything you need to know itself is a big job. Your best bet is qemu-sparc-system + Solaris install. Warner P.S. Sorry to go into partial rant mode, but 4 years ago when I started helping the folks working with bsd-user, I thought how hard could it be... now I know...
