Zinnia

[autumn jasper]

which is kind of like turnstiles but much more coarse grained

and removing the need for the turnstile block

[storm bobcat]

It's a ticket lock but more scalable because no cache line bounce

or less

[autumn jasper]

like if two threads are contending for two different PTE locks but the addresses of those PTEs happen to collide in this hash table, then when one of them is unlocked itll signal that event flag and both threads will be awoken and one of them will just hav eto go back to sleep after finding the PTE still contended

is a risk

[storm bobcat]

The same idea can be used for mutices using mcs nodes

[near tartan]

i never implemented pte locks tbh

[autumn jasper]

they needed this bc they had like 1 available PTE bit to use lol

they could have used a spinlock but wanted to be fancy i suppose

[near tartan]

i think the ISAs i want to target have maybe 2 bits in common?

i want x86_64, aarch64, riscv64rva, and loongarch64

maybe powerpc

[autumn jasper]

powerpc doesnt even have multi level page tables

it has a hash table that people normally put a multi level table on top of and treat the hash table like an in-memory software refilled TLB

NT and Linux both do that on PPC

[near tartan]

i wanted to write generic paging logic

[autumn jasper]

the ppc hash table is particularly problematic though

bc you can get page faults due to a previous hash table collision that overwrote another PTE

and this overwriting loses information

[near tartan]

ew

[autumn jasper]

which is bad for certain vmm designs

NT and Linux both sometimes want to store info in PTEs that is not reflected anywhere else, for example the location in swap space of a private anonymous page that has been removed from memory

[autumn jasper]

so they do this on ppc

contrast this however with mach derivative vmms like AIX's and XNU's

mach treats the page tables as like a cache for information in the memory objects

so theyre able to service these hash collision faults just by looking up a memory object and filling the PTE in with the info there

and its fine if they overwrite some other PTE in the process bc the info it stored is fully reflected in the memory object

so AIX and XNU (Mac OS X) do not simulate a multi level page table on PPC

weird instance of a thing where NT and Linux align with one another but not with like, AIX and XNU lol

is how page tables are treated by their vmms

[near tartan]

I'm not sure if i want to couple my VMM to the page tables

i mean i could

[autumn jasper]

its probably a bad idea to

current meta that the vmm heads in here have moved to is that the mach idea is the more correct one

disposable page tables are good

not just for portability but also for memory usage

basically you can throw away page tables for processes willy nilly at any time if you do page tracking like mach does, where the page tables hold no unique info

so that becomes a source of memory reclamation

the memory objects (or other container such as uvm amaps) are probably/usually more compact than the page tables so this is a win

under this scheme throwing away page tables is basically invalidating a cache, so it has a performance effect (causes page faults to repopulate said cache later), so youd only want to start doing that when memory is really low

[near tartan]

that's so dumb then

if it's less portable and uses more memory

[autumn jasper]

NT at least has a solution to the memory thing which is that it can swap out the page tables

linux has no solution the page tables are just wired

a page table holding the swap location of swapped out anonymous pages is just eternal on linux, you dont get to reclaim it after all the contained pages are gone, it only goes away when the process dies or the range is munmap()d which sucks

on NT the page table becomes swappable itself when all contained anon pages are out of memory

and then the page table's own swap location is stored in the upper level page table entry and so on

under NT the whole page table hierarchy for a process can get swapped out in this manner, from the bottom up

i believe

which is obviously more complicated than a scheme where you can just throw them away at any time without doing any writes to disk

at least it does SOMETHING about page table memory though unlike linux

linux is the only major kernel without a solution to this specific thing i think

i wonder if you could DoS linux by like allocating single pages of private anon memory spaced 2MB apart to allocate 1 page table each

and just keep doing this until the OOM killer has to run because it cant swap out or free those page tables

assuming u tiled all of userspace with pages spaced 2mb apart

with 4 level paging

u could cause it to allocate up to apparently 256GB of wired memory in the form of page tables

im sure youd hit quota limits long before that

but itd be fun to test

[marble salmon]

a scheme that makes pts disposable still has to store swap locations somewhere though

is the DS that stores swap locations swappable in NT?

[autumn jasper]

NT doesnt have disposable PTs so i dont get the relevance of the question

the data structure that stores swap locations on NT (and Linux) is the hardware PTE of the private anon page

which is eventually swappable on NT (after all contained pages have been fully removed from memory) and wired on Linux

[marble salmon]

the relevance is that disposable PTs does not necessarily mean that more memory is swapable

[marble salmon]

but this answers my question ^

[autumn jasper]

planned scheme for mintia2 is that the data structures for this (memory objects for shared memory (like files and shared anon pages) and amaps (for private anon pages)) will tend to be much more compact than page tables

and will go in paged pool

and be swappable

a page of paged pool will have 1 pin for each active page tracking entry it contains though

the pin is removed when the tracked page is fully removed from memory and only exists in swap

this is to avoid deadlocks from having to allocate pages in order to free pages

(ie having to allocate a page to read a swapped-out paged pool page into, in order to read a tracking PTE to find a page to swap out)

(this is the same reason NT page tables are pinned until the last contained page is fully gone)

[uncut jackal]

it's iBoot2

it's been known before apple silicon macs became a thing

there's been Project Sandcastle and checkm8 and stuff u know

[hard lark]

sorry thats unrelated, but we can see nt's pte struct is a union of like:

ULONGLONG Long;
MMPTE_HARDWARE Hard;
MMPTE_PROTOTYPE Proto; 
MMPTE_SOFTWARE Soft;
MMPTE_TIMESTAMP TimeStamp;
MMPTE_TRANSITION Trans;
MMPTE_SUBSECTION Subsect;
MMPTE_LIST List;
MMPTE_ACTIVE_PAGE_TABLE_LINKS PageTableLinks;

maybe you know what type of pte is a TIMESTAMP mmpte

[autumn jasper]

probably some 500 IQ synchronization thing

[near tartan]

the BSD larp is advancing

[sick tangle]

TIMESTAMP is a format that records the TB flush generation at which a mapping was removed, allowing to determine later whether a TB flush entry must be issued

[hard lark]

got new info

[hard lark]

interesting

[hard lark]

whats meant by buddy in _MI_ACTIVE_PFN?

https://www.vergiliusproject.com/kernels/x64/windows-11/25h2/_MI_ACTIVE_PFN

[near tartan]

i had a stupid idea

for a microkernel:

• servers can register syscalls
• during registration, it passes a set of descriptors for each argument that the handler expects
• these descriptors can then specify things like "this argument is a pointer and its length is described by argument n" or "this argument is a handle"

or maybe some kind of bytecode can be used for this

ebpf

[marsh holly]

lua

[near tartan]

omfg i have an idea

printf string

eh nah

that's cancer

[storm bobcat]

yea but how do you enforce this

[near tartan]

wdym

[storm bobcat]

i mean the idea behind this is security/validation, right?

[near tartan]

no

[storm bobcat]

what is it then

[near tartan]

this would be a form of IPC

the kernel would catch the syscall, check how it needs to transform the arguments, and call the handler in the other process

[storm bobcat]

oh yeah ok

i mean this is fairly common

[near tartan]

idk it's something dumb i thought of

is it?

[storm bobcat]

well u send a bunch of descriptors or something for handles and pointers

[near tartan]

i don't think i'm talking about the same thing here

i mean that when a server registers the syscall handler, it tells the kernel what to do with each argument

[storm bobcat]

same concept as doing it client-side idk

ig that makes sense

[near tartan]

not really

because the kernel needs to translate object handles between processes, right?

[storm bobcat]

yeah but i mean oftentimes it's just included in the request the client sends

[near tartan]

yeah i guess

[storm bobcat]

instead of done in the server

[near tartan]

but then the kernel knows about the format before hand

my idea was to do this dynamically

[storm bobcat]

so you want to set up such schema for every kind of request or something?

[near tartan]

pretty much

(note that i won't probably do this, i'm just brainstorming new methods)

[near tartan]

i'll make an example

[idle flower]

how about you push all io through shared memory ring buffers? thats what i plan to do, that way you completely bypass the kernel (after ipc setup) and get async io basically for free on top

[near tartan]

that doesn't work

you still need to translate handles

well, in my proof of concept i already have all data-only IPC via shared memory

the problem is passing object handles along

i have a channel for this, similar to what Zircon does

where you have a handle buffer and a data buffer

[autumn jasper]

I appreciate that the DEC "TB" terminology has persisted in NT all these years

[near tartan]

okay i did some digging

kqueue is pretty cool

i will still expose epoll via mlibc, but internally route it as kqueue

[near tartan]

not so sure about signalfd

i don't want to lose my wayland ports

[marsh holly]

"i don't want to lose my slop"

🥀

[near tartan]

wayland good tho

[marsh holly]

"being kicked in the balls good tho" bro stop

[near tartan]

being kicked in the balls by xorg

[near tartan]

maybe i should drop the linux FDs and become a true BSD-like

XFCE is good enough

https://www.youtube.com/watch?v=_RO9beApnec i got this in my yt recommendations

i MUST larp

[near tartan]

nahhhh this guy actually installed CDE

crazy

[rugged shard]

https://tenor.com/view/openbsd-puffer-fish-baiacu-gif-25742155

[teal hawk]

Extremely cool

[jaunty bay]

well it is if you’re into that

[near tartan]

okay it seems eventfd is actually a thing on BSDs too

https://man.netbsd.org/eventfd.2

timerfd as well

although no signalfd

which makes sense, since kqueue exists

[near tartan]

@hybrid island are you still working on lily

i have more macro soup soon

[near tartan]

i'm surprised with how simple the NetBSD schedulers are

both sched_4bsd and sched_m2

they're less than 500 eloc

[storm bobcat]

the logic is in tdq

or runq

https://github.com/NetBSD/src/blob/trunk/sys/kern/kern_runq.c

this

tho the netbsd schedulers kinda suck

[near tartan]

any recommendations

[storm bobcat]

freebsd is better

[near tartan]

ULE?

[storm bobcat]

yes

[near tartan]

will look into it tomorrow

[storm bobcat]

didnt you already have it

[near tartan]

in rust

[storm bobcat]

yea

[near tartan]

and also a bastardized version

[storm bobcat]

you should be able to translate

https://github.com/rdmsr/zag/blob/master/src/kern/sched.zig

this is better than ULE in some ways

[near tartan]

from what i've seen UVM is also quite nice

but there's a lot of code

[storm bobcat]

not really on the same level tho

a scheduler is easy to reimplement

uvm is a whole ass subsystem

[near tartan]

yeah i saw

[storm bobcat]

there's a phd thesis on it

well it was made for a phd thesis

i should make netbsd's slab allocator better

[marsh holly]

(gpt had it)

[near tartan]

[storm bobcat]

fr? 😭

[near tartan]

only the parts that didn't work

i was too lazy to debug

but that's more or less why i even want to rewrite

if i slop things there's no point in doing it

no fun

[storm bobcat]

i used an llm to check over it and it did find a few errors i made

but it was mostly just me copy pasting code around and forgetting to adapt it

[autumn jasper]

In what ways

[storm bobcat]

the load balancing stuff in particular

that's the only way

[autumn jasper]

ask the llm about minting's

[autumn jasper]

weren't you gonna send me a PR with these improvements for mintia2 ...

[storm bobcat]

eventually

but now i have j*b so I have less time

[near tartan]

i am free from my job for the rest of the week

[storm bobcat]

and it's draining blue collar work so I dont have much time at night to do a lot of stuff

[near tartan]

so i want to make it count

is nbsds turnstile impl useful

[storm bobcat]

afaik it's pretty similar to solaris's original code

so yea

they do the solaris lock pointer thing tho

[near tartan]

 * Turnstiles are described in detail in:
 *
 *    Solaris Internals: Core Kernel Architecture, Jim Mauro and
 *        Richard McDougall.

[storm bobcat]

yea i told you that

solaris book and freebsd book both go over it

tbh I'm not really sure what the XNU and freebsd impls do much more than solaris

XNU's is like 3k lines or something

uses pairing heaps

i decided against that because it wasnt really worth it i thought

[marsh holly]

ewww I hope they find a cure

NEETs of the world, unite! we have nothing to lose but our chains
🚬
🪰

[near tartan]

i will larp more and call my include dir sys/

[storm bobcat]

you should larp NT instead

much cooler

[near tartan]

nah

[storm bobcat]

NT directory structure i mean

/ layout

[near tartan]

KiLlingMySelf

idk how it looks

[marsh holly]

KeAcquireBitches

[storm bobcat]

i like ke_acquire_lock

dont have to be PascalCase

[near tartan]

i don't like the prefixes everywhere

[storm bobcat]

cool strict-ish layering though

it just gives order to the thing and forces you to separate subsystems which i like

[near tartan]

wtf is wrong with clang

[autumn jasper]

Then it's DG/UX larp

[storm bobcat]

did DG/UX do layering

[autumn jasper]

Yes

Independently

[near tartan]

@brisk totem help what is clang smoking

[storm bobcat]

i remember i have the source code but never ended up really digging through it

[autumn jasper]

It also did prefixes on everything with 2-3 letter abbreviations but as snake case instead of pascal case

[storm bobcat]

I dont like the NT inconsistencies tho

like, Ki vs Exp

[autumn jasper]

then don't imitate them

i do Kep

and Exp

and -u for stuff exported to other subsystems but not to kernel modules

[storm bobcat]

lol i forgot it did this

[storm bobcat]

how can you larp then

[autumn jasper]

isn't it awesome ..

[storm bobcat]

oh clanker finally responded

to the mintia thing

idk the mintia code enough to say

[near tartan]

this is my kernel

[autumn jasper]

I will review these

I'm p sure the last one is wrong it just doesn't understand the reaping logic

[storm bobcat]

you should do this for every function instead

[near tartan]

[storm bobcat]

yea i see the prefix things in dgux but it's unclear what those prefixes all mean

[autumn jasper]

NT has a piss prefix

And a PlayStation Portable prefix

are those clear?

[storm bobcat]

clearer

I found the meaning of cm at least, it's "channel manager"

there was something in the late 80s/early 90s that lead people to name everything "X manager"

[autumn jasper]

No it's not

Its configuration manager

[marsh holly]

claude please generate me huge comments make no mistakes

[storm bobcat]

in dgux

[autumn jasper]

oh

thought u meant NT

[storm bobcat]

yea ik NT is cache manager, they just happen to share that prefix ig

[marsh holly]

this guy is larping so hard he even copied the copyright header

[storm bobcat]

they also share ps but idk what that's about yet in dgux

[autumn jasper]

we figured out why NT and DG/UX both do these ridiculous gigantic comments on everything

[storm bobcat]

i think its just an assembly thing

[autumn jasper]

it's bc they were both the product of mainframe OS engineers who had been doing asm for like a decade, where that was basically mandatory to make giant kernel codebases in asm maintainable, doing C for the first time

and carrying over the habit

yeah

[marsh holly]

oooooooooooooooooooooor

they had secret cia llm tech before it was public

[autumn jasper]

don't care about whatever joke ur about to make

blocked u preemptively

[marsh holly]

😭😭

[storm bobcat]

they always call it the "PS subsystem" and never say what it actually is 😭

[near tartan]

bro this IS a file from NetBSD

[autumn jasper]

Process Structure

[storm bobcat]

it seems like ipc stuff

no in dgux again

[autumn jasper]

oh dgux

wow

[storm bobcat]

I kinda assumed PS meant process subsystem in NT, ig structure makes more sense though otherwise all other prefixes would have S

[autumn jasper]

I realized it meant process structure when i saw that phrase being used everywhere in like VMS and Mica writings

[marsh holly]

same

or just

ProcesS

[storm bobcat]

I just call it proc because I'm a loser

the best i can figure out on PS is it's something like "process synchronization" since it seems to contain a bunch of IPC primitives

[near tartan]

struct lwp

[storm bobcat]

yeah lwp is weird terminology

i think only netbsd uses that

though dflybsd has lwkt

@autumn jasper there's also this

[autumn jasper]

I'll report back how many of these r substantial

@storm bobcat give it my turnstile impl...

[storm bobcat]

which file is it

[near tartan]

anyways

[storm bobcat]

ah KeTurnstile

[near tartan]

good night, please dont spam so much so i don't have to scroll far

[storm bobcat]

@autumn jasper

the only useful one is #1

[sick tangle]

Is its source code open?

[sick tangle]

The cache manager in NT is Cc

[magic charm]

it's an old solarisism (at least i think solaris was the first to use the term, but others might have preceded it)

primarily existing because of a distinction drawn between userland threads and the substrate that they run on (M:N threading models)

[autumn jasper]

these were unfortunately all legit

i fixed them

[autumn jasper]

not legally but its extremely unlikely anybody would come after you for studying a tape dump of an early version of its sources of DG/UX for the Nova architecture from 1989

considering its been gone since the 90s

i dont even know who owns it

[autumn jasper]

#1 is wrong, it would make no sense to use the priority of the highest priority timeshared thread for this

its intentional that it gets smothered underneath all real time / interactive threads and that only the priority of the former is used if any are in the waiter heap

[hybrid island]

Yeah

Bring me the soupy macros

[hard lark]

can you share where it is

[autumn jasper]

this did inspire a change: timeshared threads no longer bestow their priority and do not reside in the max heap

the code was also generally cleaned up around there

[near tartan]

today I'm going to try to get an ELF loaded from memory

signalfd is stupid

the bsd epoll shim redefines read as a macro

[nocturne rampart]

Wait didn't you do it before

[near tartan]

yes

but i want to move to kqueue

and kqueue has a signal filter

but signalfd is accessed with a read()

[nocturne rampart]

Ohh

So now bsd larp-

[near tartan]

yes

literally

[leaden stirrup]

are there any ideas on supporting shared locks?

maybe it's already there, but I didn't get it

[autumn jasper]

it is not there

its quite problematic to try to do

as a result ive just not been using shared locks

except for locks that are like at the very tippy top of the kernel's locking hierarchy AND which are only taken by threads that are probably not doing anything latency sensitive

such as a shared lock for the filesystem name cache

[hard lark]

iirc even xnu doesnt support

and they have huge turnstile implementation

[autumn jasper]

solaris just keeps track of the first shared acquirer and does PI through him only

any additional ones are forgotten

i dont even bother with that lmfao

[hard lark]

one thing xnu does and oyu dont is WaiterPriorityChanged handling

they have turnstile_update_thread_priority_chain

[dreamy scaffold]

I don't think anyone does

[autumn jasper]

I have heard from reliable sources that autoboost does

I don't know if it's airtight though or if it can fail to track all shared holders sometimes

[dreamy scaffold]

maybe they're implementing it in a more selective and more expensive way

or maybe their design is just smarter to begin with... I don't know

[dreamy scaffold]

is it kernel mode only?

[uncut jackal]

netbsd's linux syscall emulation layer probably implements that too lol

so like it may not be idiomatic but there is an implementation in the source

[near tartan]

wait so nbsd's timerfd is fake?

[uncut jackal]

no idea

[near tartan]

i might leech off of the netbsd mlibc port

[uncut jackal]

which is in really early stages

it could be useful in the future

[near tartan]

ye

[uncut jackal]

the netbsd port will have the linux option unsupported (could change in the future)

[near tartan]

i don't support the linux option anyways

you could enable the epoll option tho

then copy the epoll shim to mlibc

[uncut jackal]

it will have the glibc option supported but I also want it to build without the glibc option

pure bsd+posix+ansi

[uncut jackal]

@near tartan

netbsd-mlibc is back to where it was

(it segfaults on flush i think)

[near tartan]

nice

hm

[uncut jackal]

yea i couldn't debug it last time

[near tartan]

because of time or you didn't know what it was?

[uncut jackal]

seemingly it flushes after the file descriptor is closed?

[uncut jackal]

latter

[near tartan]

i mean if you build with debug symbols you should see where it dies

you have the coredump

[uncut jackal]

yea that's what I did

it dies somewhere in exit

[near tartan]

can you show the disasm?

[uncut jackal]

it seemingly calls flush once then closes then calls flush again

[uncut jackal]

not rn

I'm going off memory

[near tartan]

(probably move this to the mlibc channel too)

but thanks

[storm bobcat]

why? what I did is make interactivity its own priority range, easier to manage wrt PI

i love how there's a discussion about mintia and zag in a thread named zinnia, it was meant to be...

[autumn jasper]

Because no priority inversion can result by having a timeshared thread block on a lower priority other timeshared thread

By nature they will all get a turn Eventually

And are assumed not to require responsiveness

[storm bobcat]

yeah but the priority inheritance is so that they get the same turn

[autumn jasper]

So may as well save the cycles

[storm bobcat]

maybe idk

[autumn jasper]

actually there is a problem which is idle priority threads

you could get priority inversion if a timeshared thread blocks on an idle priority thread

bc idle priority threads never run until all timeshared threads have blocked

oopz

[leaden stirrup]

one case I know of is when the per thread slots run out

I'm also not sure what it does if the allocation of a head from the lookaside fails

[near tartan]

okay i need to find a way to deal with login1 bullshit so i can upstream plasma

i think i'm going to make a portable logind that poeple can use

my solution for now is quite simple, it just stubs the hell out of everything and sends static data over

i know i love changing my mind regarding rewrites

but for now it's still the most effective solution to just fix what i have

this means i have to redo:

• the scheduler (i will adapt the design from Zag)
• signal handling
• VFS operations

the problem with VFS is that currently i separate between an open file and a vnode

this leaves me with two layers of file operations

i think for me it makes a bit more sense to just have FileOps be part of <NodeType>Ops

for the scheduler, i need to start respecting CPU affinity and i want it to be aware of CPU topology

i am also refactoring the arch:: abstraction so it's less confusing

maybe i can make it a public trait and pass it as a generic parameter to functions that call it 🤔

i saw that

but yeah, regarding , i'm not using it anymore to do kernel code

while it did fix some issues, the code is a liability and i don't want it around anymore

lastly, i want to add support for direct EFI boot

this isn't some vendetta against the limine protocol, but it's something i've wanted to do for a long time

i'm still going to use limine as the bootloader

[uncut jackal]

the birth of ulogind

[near tartan]

it's very simple

it's like 3 files

like, i already made it

i just need to wire it up to meson

[silver birch]

why do you think this is wrong

there are operations which are only appropriate for an inode and others for an open file

[near tartan]

well i found out that every open file needs a backing inode/vnode record anyways

[silver birch]

yeah

[near tartan]

so now I'm not sure if i want to keep that design, or just route everything through the inode

[storm bobcat]

werent you rewriting

[near tartan]

read above

[marsh holly]

I love making choices but I would never ever ever ever never actually do what I chose to do

[storm bobcat]

Zag mentioned 🗣️ 🗣️ WTF is a useful kernel???

[near tartan]

translation: I'm a chud with too little time on my hands to go through with a full rewrite

[near tartan]

i spent my vacation working on managarm and now i have like no time left to work on my shit 😔

i just come home exhausted af

[marsh holly]

I have spent the past 5 days vibecoding late into the night/not sleeping

All I get is more chud rage

[storm bobcat]

y

[near tartan]

soyboy website

[marsh holly]

Some stuff that is not really in the scope of this server

For an event

[marsh holly]

do nusoi really

[marsh holly]

Which had to be working last week so we just had to lock the fuck in to get it working now

Its been fun I could larp with 3 other people in a hotel lobby as SF founders

[near tartan]

btw @storm bobcat i ended up implementing the initgraph subtrees

these depend on the dt.parse-blob stage at the top of the screenshot

so they get all skipped

[nocturne rampart]

I'm assuming this is similar to the Linux efi stub?

[near tartan]

yes

it's going to be a bit tricky with cargo but i'll figure something out

[nocturne rampart]

Zinnia UKI images confirmed

[near tartan]

💀

maybe i can also add Hyper support

can't hurt

[uncut jackal]

I'm going to be really honest

multiple boot protocol support is only really useful if the protocols are like genuinely limiting what platforms a kernel can get booted on

[shy turret]

Multiboot2 aswell

[uncut jackal]

like multiboot2+limine+hyper wouldn't really be useful if you're only targeting x86

[shy turret]

I think the point is just that you can and to show it's properly abstracted

[uncut jackal]

abstract it between more useful stuff then idk

[near tartan]

i am targeting every arch that is supported by rust at some point

right now it's only x86_64, aarch64, riscv64

[uncut jackal]

cool

[near tartan]

but I'll probably be limited by what mlibc supports anyways

[uncut jackal]

if you want to support direct EFI boot then most of the bringup (e.g. SMP) will need to be done by your kernel right?

[shy turret]

You want to do that yourself anyways

[uncut jackal]

at that point I'd just do efi+linux boot protocol

[near tartan]

i already do that

i think the linux boot protocol is legacy and quite shit

[uncut jackal]

linux boot protocol for risc isas*

[near tartan]

i use 0 features from limine that are limine specific

[uncut jackal]

that's literally just a jump and usually 1-2 registers

[near tartan]

yeah the x86 one not so much

[near tartan]

because i knew from the beginning that this day is going to come

[uncut jackal]

I limit myself to architectures supported by limine rn

so I go all out on limine features

[uncut jackal]

yea because that is legacy

EFI probably covers enough

[near tartan]

ye

[near tartan]

i don't see the point quite honestly

doing smp startup is not really hard

[uncut jackal]

yea it's not

but supporting a much more barebones protocol and a much more fleshed out protocol together doesn't really make sense to me

the boundary of where the kernel's responsibilities start is vague then

like if limine will always be booted by EFI and you support EFI why add that extra jump in between

if limine is to be used as a boot manager then do efi chainload ig

[near tartan]

all i need from a boot protocol is:

• RSDP or DTB address
• memory map
• hhdm mapping for bootstrap
• initramfs

[uncut jackal]

now what I would get is something like

linux-risc+efi+iboot2 or other vendor specific

[near tartan]

even on apple i can use efi

the issue with pure efi boot will probably be loading an initramfs

[uncut jackal]

UKI is not actually that bad of an idea

[near tartan]

i don't like ukis even on Linux

[uncut jackal]

or you could take some ideas from the android world

dtb partition initrd partition etc

on disk

you'd need a disk driver but filesystem implementations (other than whatever you're using for your initramfs) can reside in modules or userspace or wherever

you can even a/b the initrd partition

for system upgrades

[near tartan]

this sounds funny

[uncut jackal]

it is

[near tartan]

having a partition with just a big cpio archive xd

[uncut jackal]

it's not a partition with just a big cpio archive

it is a partition with no fs that is just a cpio archive

[near tartan]

yes?

that's the same sentence

[uncut jackal]

i thought you meant like

a partition with an fs with just 1 file

[near tartan]

nah

[uncut jackal]

that is the initrd

i think the reason android kinda does that is because the "disk" is emmc/ufs and the boot chain is not the only thing that reads those partitions

[near tartan]

my initramfs parsing is in kernel anyways, would be hard to load modules otherwise

[uncut jackal]

like they have firmware blob partitions

that are not read by the kernel usually

the other chips themselves read them

[near tartan]

yeah i have seen something like this on a few risc-v sbcs

[uncut jackal]

i found this online

real list of partitions from an android phone

there's arm trusted firmware stuff

there's modem stuff

[near tartan]

nice date

[uncut jackal]

there's dsp stuff

there's a logo partition ffs

I've modified that before btw

changing boot logos is fun

[near tartan]

honestly, i wonder what it would take to boot zinnia on a phone

[uncut jackal]

i will do osdev for a phone if I get a cheap second hand easily moddable phone

[near tartan]

an older one ofc

maybe a pixel?

[uncut jackal]

pixels are expensive and have never entered the market here

[near tartan]

pixels aren't expensive

[uncut jackal]

they are here

because they are officially not in this market

[uncut jackal]

ok now find one in turkey

or better yet gift me one

[near tartan]

come to fosdem i might

[uncut jackal]

i actually might

i currently have a turkish special passport

gotta make use of it before it expires

i wonder if I can get my university to fund a 10-15 person trip to fosdem

they're already funding my trip to poland in september

lol

partially*

[near tartan]

lol

[uncut jackal]

I'll be in poland in september

for ERC

european rover competition

[uncut jackal]

turkey has a 2-tiered regular passport

special and personal i think?

we don't call them that so I might have remembered the english names wrong

we call them green and red

they're both regular people passports but red is for everyone and green is for state workers and family

green has visa-free entrance to pretty much everywhere

except the us and uk

i am not a state worker I am the child of one so mine is going to expire when I'm 25

i have not made use of it yet

i think android has some superpartition magic btw

"partition"s are kinda more like volumes

i think

i haven't looked too much into this

[near tartan]

ew

[uncut jackal]

i don't think they have a set size

[near tartan]

i should add apfs support

wonder how bad it is to impl

[uncut jackal]

but they're definitely not just one fs

so idk what they're doing to make that work

[uncut jackal]

does linux have apfs?

[near tartan]

think so

[uncut jackal]

i don't think I succeeded in mounting macos in my friends asahi mac

[near tartan]

there seems to be a fuse driver

[uncut jackal]

fun fact

one of the SBCs in our rover

runs ubuntu 20.04 userspace with linux 7.1.4

the SBC got mainlined in like 6.18

[near tartan]

cursed

[uncut jackal]

but we needed 20.04

which is on like

5.8 or smth

[near tartan]

but i mean the kernel shouldn't break userspace anyways

[uncut jackal]

ye

[uncut jackal]

I have created nix tooling that builds this

with debootstrap

the kernel is the latest nixpkgs aarch64 kernel

lol

[near tartan]

i should port nix to zinnia

for

oh man so many cool things i could be doing

i don't know where to start

well, i guess i should sort out my outstanding 80 file diff

[uncut jackal]

i have decided i want to port nix to managarm

[near tartan]

port it to zinnia too

[uncut jackal]

i mean

[near tartan]

idk it should compile, it's just c++ right?

[uncut jackal]

nix assumes a lot of unix stuff

windows port is damn hard (idk if it has happened yet)

[near tartan]

zinnia is posix

[uncut jackal]

but

ports to bsds etc exist

it's doing great

will probably have some bugs in mlibc

then a-ok

[near tartan]

whatever fixes you get in mlibc will probably also help me

[uncut jackal]

I have lots of bugs to fix in mlibc fwiw

mlibc in nixpkgs can't get binutils to build

because I build with tests enabled

[near tartan]

huh

[uncut jackal]

even a single test in some random dependency failing makes the build not go thru

which i want

it finds a lot of bugs

nix runs the test suite of anything i attempt to build for mlibc

[marsh holly]

Mlibc can build binutils without the tests (done it for astral) but Ive never tried it with the tests

[uncut jackal]

the issue is not Binutils yet

the dependencies don't build with tests

[tough thistle]

You should take a look at xens credit1

They do some cool stuff in relation to prioritising cores close to the task during migration, aswell as cores without noisy neighbours and weighting SMT cores accordingly

Although I’d say it’s a lot more affinity / topology aware than the credit2, it doesn’t really handle load as well as the credit2, which replenishes credits differently (which was a source of performance degradation in credit1 due to them replenishing at fixed intervals of time) , doesn’t fall victim to the multiboost problem, and puts an upper bound on cpu utilisation using a deferrable server

credit2 is a lot smarter in how it decides what cores to push/pull from based on what migrations will do to the load of the runqueues involved

[silver birch]

https://klipy.com/gifs/waow-based-1

[near tartan]

actually, good news about running zinnia on a phone

my old phone is a oneplus 9

which has an open bootloader

[nova tiger]

random side quest

[near tartan]

.

only limitation is that it's going to be EL1 only

but that's fine

[near tartan]

apparently it's a really good target because i can sideload limine

they use a stripped down edk2 it seems

[shy turret]

https://www.goretroid.com/products/retroid-pocket-6-handheld
This thing can boot limine and other shit too, IIRC it has UEFI too??

Could be another good target

And it's funny

[near tartan]

cursed

i have other aarch sbcs

that i want to support as well

[near tartan]

https://github.com/zinnia-os/minilogind

if anyone cares

[near tartan]

i LOVE xonotics build system