500 cleaned update

[palainp]

This PR is a rebase against current main + updates some constants to be the same as in linux.

The bad news is that I'm currently in a dead end with mmap/munmap:
I first wanted to use malloc/posix_mem_align as it's available with dlmalloc as a mmap replacement, but going on that path led me to search how Ocaml uses mmap/munmap. If I understand correctly, Ocaml requests memory with mmap (more than necessary, maybe to ensure some alignments constraints?) and then uses munmap to "free" pages it won't use:

• So we can't use free here as some pages haven't been allocated with malloc
• If I noop munmap it works, but it leads to more memory usage/memory wasted than necessary (which isn't great as Ocaml forces an additional 2MB stack for each domains)

[palainp]

The last commit re-uses the bmap allocator from mirage-xen (https://github.com/mirage/mirage-xen/blob/main/lib/bindings/bmap.c) and permits to both mmap & munmap. So far it works with qubes mirage firewall (transmit traffic) and I have the same bandwith on an ADSL link (which isn't an indicator per se as the bandwidth is limited).

I haven't done any other performance tests yet (the bmap allocation is O(n) with n=2048 pages, and munmap is O(1)) because I'm more worried about having to set a magic constant ("mmap size", here 8MB) in the libc startup function.

[palainp]

So having an allocation for mmap leads to bad things, the Ocaml runtime runs out of memory at some point. So I'm switching to another memory allocation system:

• the bitmap system for managing free/used pages
• a binary buddy system for memory request smaller than a page

This currently works with tutorial/hello unikernel, but segfault with application/https and leave the CPU spinning with qubes-mirage-firewall. I've left debug trace to continue some work on that early 2024 :)

[palainp]

It seems to work now, qubes-mirage-firewall works and routes traffic :)

So far:

• I have a very similar bandwidth with FTTH (I expected the simplest allocation algorithm to be slowest and maybe be a bottleneck, maybe with more heap memory (currently using 32MB) the O(n) complexity will shows some issues),
• the Ocaml runtime uses a few extra MB at startup (13MB 5.1 vs 8MB with 4.14.1, 2MB is due to domain stack allocation, I haven't checked for the other 3) but memory usage is similar after a while (due to uncollected Cstruct bigarrays?),
• I still need the manual GC calls with qubes-mirage-firewall

[palainp]

Closing this as @shym has been working on 5.2 integration and the allocator isn't well adapted.