Zag

[fallen bobcat]

u do this and then sysctl can tune dir.some_knob

[heavy sandal]

ah neat

[fallen bobcat]

the way freebsd does it is kinda crazy, historically sysctl didn't have names but rather IDs (so like kern.whatever would be something like [1, 2]), now it uses string names but to handle legacy code it still uses numerical IDs, and to find names it searches into a special ID

so you have to do 2 syscalls for sysctl, one to get the IDs from the string name and one to set the value

and it's a linear search of names in that entry 💀

what I do is have an AVL tree ordered by lexicographical ordering of the names in each directory, and the linux subsystem will simply convert file paths to the standard . syntax I guess

[heavy sandal]

consider: std.StaticStringMap
if everything is defined statically anyway

[fallen bobcat]

idk how that would work

[heavy sandal]

oh right tree not strict map

[fallen bobcat]

no it would work

but idk how to write it

because the way I do things rn parent is a type so I can't insert to it directly, I put nodes into an executable section and traverse it and insert them:

ah I don't think StaticStringMap would work here because I need to add the KVs one by one basically

[heavy sandal]

you could somehow build it all at comptime? idk

[fallen bobcat]

there is probably a way to build the tree at comptime though

yeah, I just don't know how

[heavy sandal]

if you declared it all in one place itd be easy

but idk how youd do it here

[fallen bobcat]

do you think I can declare the type with const first then refer to the storage and add it to a comptime hashmap or something

like

[heavy sandal]

comptime global vars dont exist

is the issue

[fallen bobcat]

const Holder = struct { var storage = ...};
comptime_hashmap.insert(&name, &holder.storage);

ah

im sure I could build it at compile time

it's just very annoying

[heavy sandal]

yeah

[fallen bobcat]

I wish zig had some type of compile-time section or something, where you could iterate over all variables in that section and do stuff on them

Kind of a crazy concept but I think it'd be useful

[fallen bobcat]

@languid canyon what is that linux thing where you have variables that are read-only after boot

[languid canyon]

What are you asking exactly

Just a section thats changed to ro after init

__ro_after_boot I think

[fallen bobcat]

what is it called

ah ok thanks

[languid canyon]

Or after init

Just grep

[fallen bobcat]

I've added tunables, which are cmdline parameters but better:

for dynamic parameters you might wanna set on boot

i saw it in xnu code and thought it was kinda like sysctl but it wasn't, though I will continue my sysctl-like thing too

[vital summit]

this looks like linux' sysctls

[fallen bobcat]

it's not

also tf you mean "linux"'s sysctls 😭

[vital summit]

like the way they're defined

[fallen bobcat]

linux sysctl is cancer

BSD-style sysctl tho is pretty good

[fallen bobcat]

I mean the syntax is similar but it's not really related, sysctl will be another layer on top of that for stuff that can be changed at runtime

[vital summit]

i only meant syntax lol idk if they're related

[fallen bobcat]

this is my sysctl-like thing

I wont name it tunable tho I'll have to find another name

[fallen bobcat]

the larp never stops

i added basic hardening features to the allocator

[fallen bobcat]

@heavy sandal LOL

[heavy sandal]

lmao what did you do

[fallen bobcat]

        self.cpus = @ptrCast(@alignCast(cpu_zone.alloc(.{}) catch @panic("Failed to allocate per-CPU magazine state")));
``` I think it doesnt like this

[heavy sandal]

hm

idk why that wouldnt work tho

is the thing

[fallen bobcat]

well I do pretty weird stuff

this is a []rtl.CachePadded(Cpu) type so it may not like this

since it's like an inline type or whatever

[heavy sandal]

idk i do shit like that fairly often

only way to really be sure is to run a releasesafe or debug build of the compiler tho

and that takes like five years to build

[fallen bobcat]

what's funny is that this is 0.16 too

[fallen bobcat]

@crimson sand the NT functions for the memory thresholds are MiInitializeWorkingSetManagerParameters and MiUpdatePageThresholdsDpc

[fallen bobcat]

fkgjfkgj

perkenrle time

prekernel will do a bunch of stuff: detect cpuid, start APs, map kernel and pfndb

infy about to roast me

[languid canyon]

Why start aps in the prekernel? You will have to restart them for hotplug or resume from suspend anyway in the kernel

[fallen bobcat]

yeah ik, but I want the kernel to boot in a nice cozy environment

where all APs are spinning in their idle thread already

[languid canyon]

Ig if you decide to never support those features then its fine

[fallen bobcat]

i have to have duplicated code for pagemaps anyway so 🤷‍♂️

[fallen bobcat]

nah I do want to eventually

but like

it's duplicated code

not that bad

[fallen bobcat]

btw how does hotplug work, is it through ACPI?

[languid canyon]

Yeah, firmware notification on the cpu object in aml

[fallen bobcat]

i could always somehow have the same assembly file for the prekernel and kernel

like I don't really think it's a big issue

same thing for pagemaps, i have to have duplicated code kinda

[languid canyon]

I mean sure

[fallen bobcat]

see, in C I would share the headers

but I can't here

idk maybe I can but too much work

btw @solid marsh for hot(un)plug how do you protect for_each_online_cpu?

do you have a lock around the cpu count or something?

or is it atomic and it'll get picked up next time

[solid marsh]

i haven't implemented that yet

i thought of avoiding the lock by allocating all the cpu structures at startup

[fallen bobcat]

ig a lock would work but it'd be shitty

cuz when a cpu goes offline you wanna wait till everyone is out of these loops

[solid marsh]

and then those structures have a flag indicating whether they're online

so for each online cpu you're interested in, you can check the bitmap and the structure

[fallen bobcat]

you could probably do it through some kind of RCU thing

[solid marsh]

you avoid having a global lock, and if you set up transitivity well you only need the lock in the structure

yeah, I thought about that too

but I'll have to see how well it plays with hotplug

[heavy sandal]

i do think acpi is guaranteed to have an object for each possible cpu/slot iirc so at least you dont have to allocate new for a hotplugged cpu

[fallen bobcat]

you wait till all cpus are in a quiescent state and decrease the CPU count only then

it'd be annoying because you'd have to park the cpus for a while until everyone is quiescent

but I don't think you really have a choice with unplug anyway

[fallen bobcat]

@heavy sandal how would you do this

[fallen bobcat]

or some kind of IPI actually

no that wouldn't work nvm

[heavy sandal]

so i can think of two ways and one is way more fucked up than the other but both kinda suck

[solid marsh]

for hotunplugging cpus I'll have a percpu task which manages exactly this

[heavy sandal]

option a is to use externs/exports and have the kernel link to the prekernel for shared stuff. downside is you lose access to zig slices, error unions, and generally anything that isnt extern
option b is you compile the prekernel and kernel as a single executable and then split the prekernel out using objcopy somehow
option c is you just @import the same file from both and live with the code being duplicated in memory - this is what i do for my bootloader protocol file by having the root files for the kernel and loader be one directory level lower than they arguably should

[solid marsh]

why would you need a cpu count? except for informative purposes

[fallen bobcat]

Idk how do you do for each cpu

What I do is do a for loop from 0 to ncpus

[solid marsh]

yeah, but ncpus stays constant

[fallen bobcat]

no

I have a compile time ncpus like Linux

And a runtime ncpus

Compile time is for max CPUs

[solid marsh]

what is the datastructure you traverse with the loop?

[heavy sandal]

just loop every possible cpu and skip the thing if the cpu isnt present no?

[fallen bobcat]

This is convenient because zig has no macros and it would suck to do it through an iterator everytime

[fallen bobcat]

True but boring

[fallen bobcat]

Depends, but this is a fairly common thing

[solid marsh]

the compile time constant is the maximum amount of cpus the kernel can handle

[fallen bobcat]

And annoying

[heavy sandal]

acpi will tell you what possible cpus there are so its not like youre looping to max cpus

[solid marsh]

then depending on the system the maximum amount of cpus is <= that

[fallen bobcat]

Yeah but then I have to not forget to do that

Unless I make it an iterator

[fallen bobcat]

Yeah that's what I do

But I have a runtime ncpus

[solid marsh]

and you say that it can vary during runtime?

uh

i don't understand why you would want that

[heavy sandal]

yeah i dont understand either

[solid marsh]

like, let's go back to the bitmap example

what if you have cpus 0..49 and remove cpu number 21

ncpus=49, okay

how does this help you

[fallen bobcat]

Well if CPUs were unplugged yea

[fallen bobcat]

If the machine has 4 CPUs I can just do for (0..4) to go through every cpu

compile time ncpus is for cpumask

[solid marsh]

yes

[heavy sandal]

the zig std bit set has an iterator already btw

just gives indices of next set bit

[fallen bobcat]

Ah

I have my own

[heavy sandal]

just put all the possible cpus in a slice and skip not present ones imo, i dont see whats the problem about it especially if you either make an iterator or a helper for it

[fallen bobcat]

because iterators suck in zig

and i dont have macros

[heavy sandal]

what sucks about them?

[fallen bobcat]

also this would require to allocate it no?

[fallen bobcat]

clunky to use

[heavy sandal]

yes but only once because you get the slots ahead of time

[fallen bobcat]

idk why runtime ncpus doesnt work then

it's literally just the same thing

but you add if (!online(i)) continue

[heavy sandal]

how so? ive never found

while(it.next()) |item| {}

to be clunky

[heavy sandal]

yeah this is literally what im suggesting

[solid marsh]

what we are trying to say is that the number is useless, unless in your datastructure you remove offline cpus from being able to be iterated over

then that number is obviously needed

[fallen bobcat]

dont you want to iterate over all cpus at some point

I do that a bunch of times

[solid marsh]

yes

[fallen bobcat]

I also use it for my counters that must go through every cpu

[heavy sandal]

you dont need the number of active ones though because of the basic if(!online) continue

unless we've been spending who knows how long using different mental definitions of what ncpus refers to and not realised

[solid marsh]

i understand his definition of ncpus as "amount of cpus online"

[heavy sandal]

yeah same

[fallen bobcat]

yeah

[heavy sandal]

and thats the one i dont see a use for

[solid marsh]

so if you have 4 possible cores and 2 are online, ncpus=2

[fallen bobcat]

which is why it can vary

[solid marsh]

^

[heavy sandal]

i dont see how that helps if its easier and way less sync required to just iterate possible cpus and skip offline ones

[fallen bobcat]

how do you iterate over online CPUs then? you iterate over the bitmap?

[solid marsh]

yes

and skip offline ones

[fallen bobcat]

well the bitmap has the same issue then

[heavy sandal]

or over a generic list of possible ones and skip offline

[solid marsh]

or also that

[fallen bobcat]

that would be slow if your CONFIG_NCPUS is high

[solid marsh]

i think that you rarely have to do such iteration

[fallen bobcat]

how do you protect races in the bitmap?

[fallen bobcat]

i have a few

[heavy sandal]

no because you get a system-local max during init from acpi

[solid marsh]

one of the few cases I can think of right now is for picking on which cpu a task should be assigned on creation or when its affinity is changed

[heavy sandal]

separate from your os's supported max

[fallen bobcat]

even in virtual machines?

also where is that

[solid marsh]

yes

just iterate over the madt

[fallen bobcat]

the madt will have entries for all possible cpus?

[solid marsh]

yes

[heavy sandal]

yeah the madt cpu entries (also the cpu device/processor objects) will have status flags for not-present cpus

[solid marsh]

the only weird thing about the madt in x86 is that, well

there's 2 bits: enabled, and online capable

the spec says that entries where enabled and online capable are 0 must be ignored

[fallen bobcat]

ah it's the first bit

[solid marsh]

so enabled means "i am running from boot", and online capable means "i am a hotplug CPU"

only problem is, no one fucking SETS the second bit

[fallen bobcat]

also I use it e.g in the allocator to allocate per-cpu state

[solid marsh]

not even when there are real hotplug cpus

so you will have to find the processor objects

[fallen bobcat]

so i have to do AML in the prekernel?

[heavy sandal]

madt is fixed, you use the namespace objects for changeable ones

[solid marsh]

so in most computers, if you have both flags 0, the madt entry is just for padding

and in virtual machines (one of the very few places you'll find hotplug cpus), it may be hotpluggable

[fallen bobcat]

see i have a bunch

[heavy sandal]

if the madt has both bits 0 then the cpu isnt plugged in or its padding

[fallen bobcat]

very helpful

[solid marsh]

the spec says that if both are 0 it should be ignored

well, must I think

so i have no idea why no one sets the online capable bit lmao

[fallen bobcat]

so this?

that is fucked

[solid marsh]

yes

a little bit

if you see the dmesg of a machine with padding entries, it will report as present even the padding cpus

and then trim them off

[fallen bobcat]

ig what i could do is make the for_each_online_cpu use the cpumask with only the online cpu bits set

so that way I dont have to figure out the number of possible cpus

yeah maybe I just dont support cpu hotplug

only hotunplug ig

[heavy sandal]

i think this is what linux does btw but idk the details offhand so i may be completely wrong

[fallen bobcat]

what is the std bit thing you said btw

i have that in my bitmap

[heavy sandal]

https://ziglang.org/documentation/master/std/#std.bit_set.Dynamic

also static version for comptime length

[fallen bobcat]

ah lol I literally reinvented ArrayBitSet

i think

[solid marsh]

https://elixir.bootlin.com/linux/v7.1.2/source/arch/x86/kernel/acpi/boot.c#L164

[fallen bobcat]

https://elixir.bootlin.com/linux/v7.1.2/source/include/linux/cpumask.h#L1140

that is what linux does

makes sense

[solid marsh]

so, apparently, from ACPI 6.3, the online capable bit is implemented, so if your tables are at that revision or greater, you can use that bit to know whether your cpu is usable (hotpluggable)
otherwise, if the enable bit is not set, for real hw is padding, and if in qemu, it's a hotplug cpu

@languid canyon insanity

[fallen bobcat]

idk how they protect against racy mask updates tho

[solid marsh]

there's a cpu write lock

[fallen bobcat]

they don't take it in the iterator tho

[solid marsh]

yeah, because you dont update in the iterator

[fallen bobcat]

no but

in parallel there's a cpu starting

like, one CPU is doing a for_each_online_cpu and another is starting a new cpu

or worse, one CPU is doing a for_each_online_cpu and another is removing a CPU

[solid marsh]

you then have a lock outside that protecting

[fallen bobcat]

so you do have a lock in the iterator then

or they let it race for most stuff

[languid canyon]

Yeah thats how thats set up, I have the 6.3 check in my kernel

[solid marsh]

nice

[fallen bobcat]

can i hire one of you two to do hardware/firmware bullshit

so i can do actually interesting stuff

[solid marsh]

how much $$$ are we talking about

[fallen bobcat]

you will get 10 million$ in zagcoin

[solid marsh]

welp

there is also a cpus read lock

[fallen bobcat]

ah yeah

ok that makes sense

thank you clanker

[solid marsh]

well, as i say

[fallen bobcat]

to me that doesnt sound really good, unless it's RW lock or something?

[solid marsh]

it is not that common to have to iterate over all online cpus

[fallen bobcat]

true

but still nasty

[languid canyon]

Well it says read lock right there so probably

[fallen bobcat]

ig they could use a seqlock actually

[solid marsh]

and if you still really really want to remove the lock

[fallen bobcat]

i wonder

[solid marsh]

you could have an enable bit in the cpu structure too

so that enabled bit in cpu -> enabled bit in bitmap

[languid canyon]

Imo its not needed, you hotplug extremely rarely, so 99.9% of the time u grab the fast path read lock

[fallen bobcat]

yeah if it's rwlock sure

yeah it is

[solid marsh]

well actually, if you decide to duplicate the bit then you 100% also want a lock in the cpu structure

so at this point better be linux and use a rwlock

[fallen bobcat]

yeah rwlock works here

seqlock could work too

[languid canyon]

Could it?

[fallen bobcat]

why not

[languid canyon]

Seqlock allows intentional races, here you would be accessing freed data or trying to talk to a cpu thats been offlined before the second checkpoint

You really wanna block offlining while iterating this stuff

[fallen bobcat]

youd do it on the cpu mask or something

yeah no ok I dont think it'd work well

[languid canyon]

Like the nice thing is the way offlining is designed is the kernel only gets a request, but it decides exactly when ejection happens so you can postpone it until the lock is free

[fallen bobcat]

sorry i hadn't read this yet, I was planning on doing c) but I wonder if I import a kernel module would it also import all the stuff it imports?

or since zig is lazy it'll see those symbols are unused

[fallen bobcat]

yeah that's what I said earlier

you have to have a lock somehow to avoid the cpu state being iterated on

fucking hell i have so much stuff to do and I am so slow

if i worked full time on it im sure it'd still take like 1 year to get where i want

[heavy sandal]

yeah i think so lol

[fallen bobcat]

I'd rather not rely on std anyway

ugly function names

(/s but mostly because it's not stable and there are no docs)

[heavy sandal]

ime it will just be lazy

[fallen bobcat]

and I don't trust the std implementations

SmpAllocator ruined my trust of the std

[heavy sandal]

thats your decision lol

[fallen bobcat]

Nah I use it here and there it's just that particular thing I didn't know it had

I guess its useful but I reimplemented it anyway

[heavy sandal]

yeah

all depends on context

[fallen bobcat]

How do you know the std so much

[heavy sandal]

when i need a data structure i go looking in std first

this one in particular i use a lot though so i know it from that

but also just idk this is the sort of thing im good at learning and absorbing

like i still know half of the nonsense obscure things from the dotnet standard libraries despite not having used most of them

like my beloved conditionalweaktable

ive still never used it but i sure do know it

[fallen bobcat]

I'm thinking I can move all large pages to the loader

i mean handling of them

that'll simplify the kernel pmap code

[fallen bobcat]

@heavy sandal is there some kind of zig allocator i can just give a page alloc function to and it'll work

ideally id use arena allocator but i need another allocator for that

ah i think i can just use arena allocator with a custom page allocator my bad

i think

idk

[heavy sandal]

debugallocator

which I think is getting renamed to safeallocator? idk

it also needs a mutex type iirc and is a bit slow tho

because it traces everything and does leak checks and all sorts of stuff like that

[fallen bobcat]

Hey @vital summit is the initgraph stuff fully compile time or are there steps that are added dynamically? Like if you detect some device of some sort do you add that step to the initgraph or do you run the init manually in one of the steps

If so how does that work? Initgraph steps can add other steps?

https://github.com/zinnia-os/zinnia/blob/main/kernel/src/device/dt/mod.rs#L457

ah ig you register the tasks but u just dont do anything

[vital summit]

all steps are static, but i have considered something like "don't init anything acpi if no rsdp"

[fallen bobcat]

cuz I'm sure I can cook something up with zig comptime and do it all at compile time

but if steps are added dynamically then not really

mhm ig u can have some kind of "features" attribute to init steps

and like if features has acpi and no acpi: dont run it

[vital summit]

i wanted to do subtrees

so you'd register a stage with a tree

[fallen bobcat]

ooh yeah

that makes sense

sounds good

[vital summit]

something like "if the root is disabled, I don't run any leaves"

[fallen bobcat]

that way you can still build all trees statically but you can choose to not run a bunch of steps at once

[vital summit]

yeah

what's your opinion on kernel modules

i considered to just remove them to cut down on boot complexity

also might give boosts with lto

[fallen bobcat]

I guess it depends, I do want to have them but it'll be annoying because zig and C aren't natively compatible (so id have to have a bunch of extern wrappers)

ig the best option is to choose whether or not they're compiled in or linked

[vital summit]

in rust it's been easy so far

[fallen bobcat]

do you expose a C api? or just rust?

[vital summit]

just rust

[fallen bobcat]

yeah ideally I'd expose a C api

[vital summit]

it does mangling with a buildid

so kernel abi is always the same

but i feel like it causes more trouble than it solves

[heavy sandal]

fwiw zig doesnt have a defined extern abi as of yet so linked-in modules written in zig would have to work over the C abi despite all being written in zig

[fallen bobcat]

I think you have pretty much everything as modules?

[vital summit]

only irqchips and serial is in kernel

[fallen bobcat]

ig that makes sense

i was about to say you should have the core drivers in the kernel at least

[vital summit]

now that i started to write aarch64 drivers i realize that this might be too coarse for all the little devices

there are so many little drivers i need

[fallen bobcat]

I'd make modules for any "big" driver then

stuff like nvme vs ahci

ig it's hard to draw the line

[vital summit]

exactly

in linux you can make this configurable

but cargo is really inflexible here

[fallen bobcat]

wait, so I can't even link zig code with zig code right?

i need a C api

[heavy sandal]

no you can

its just that any things that cross the boundary have to be extern-compatible

so no error unions, slices, tagged unions, or auto structs

[fallen bobcat]

makes sense

but then it's annoying af

[heavy sandal]

yeah

[fallen bobcat]

like realyl annoying

because I have to write wrappers over every kernel API

[heavy sandal]

if you compile them in as part of the same compilation you can use zig apis tho

[vital summit]

they should make c abi ffi items

rust has the same issue

well in rust it throws a warning about ffi safety

[fallen bobcat]

i should've just written in C smh

[vital summit]

fr

[heavy sandal]

its because these modern langs tend to lean towards single compilation rather than linking in stuff ig

and static everything

so ABI isnt a priority

[fallen bobcat]

maybe

but for a language like zig, which wants to replace C

[heavy sandal]

yeah

[fallen bobcat]

that's just unacceptable lol

[heavy sandal]

if it wasnt still generally early in dev i would be more upset about it lol

[fallen bobcat]

i can't wait till i have so much aura i can just nudge the language development towards my direction à la microsoft engineers with intel engineers

[vital summit]

i wish c would have some macro shit to do name mangling

then you could do macro generics

[fallen bobcat]

wdym

[heavy sandal]

theres a nonzero chance my kernel will be among if not the first zig project to do full dynamic linking and loading where everything is written in zig actually, once ive got that to the proof of concept level ill start poking the zig people about ABI stuff

[vital summit]

like vec(unsigned int)

as a macro

and that would define a struct

[fallen bobcat]

can't you do this now

[vital summit]

no

how

[fallen bobcat]

like anonymous structs are now equal

[vital summit]

no

[fallen bobcat]

if they're declared the same

[vital summit]

false

[fallen bobcat]

I can swear I saw something like this as either a gnu extension or as a c26 thing

[vital summit]

only named structs are allowed to be redeclared if they contain the same fields

in c23 at least

idk if gnu has an extension

[fallen bobcat]

my main reason for using zig was generics kinda but now I realize that I don't really use them

[vital summit]

i only use them for containers

[fallen bobcat]

I don't use any arraylist/vec

[vital summit]

i write pretty C like rust

[heavy sandal]

...yknow kinda unrelated but i wonder how badly zig/llvm would implode if i mix higher and lower half stuff in the same binary somehow

[vital summit]

how does zig cope with writing to rbx in inline asm

[fallen bobcat]

yeah but stop procrastinating or I will be the first

[heavy sandal]

i use zig because i am chronically unable to read macros to save my life

i look at C macros and my eyes glaze over and my brain shuts down

[fallen bobcat]

no ABI is also annoying for thread switching

but u can just make the entry points use C abi

I think

[vital summit]

that's what i do

but it's annoying

[fallen bobcat]

i make the C ABI entry point call the zig function

[heavy sandal]

i shared a macro that linux uses for declaring elf notes from C with a zig dev while asking for help on something and the person is like yeah this shouldnt be too hard to do in zig and im over here and cant mentally parse the macro at all even though its simple

[heavy sandal]

you dont need full abi for entry just calling convention

[fallen bobcat]

yeah callign convention i meant

[heavy sandal]

yeah

[vital summit]

as much as i like rust, in a kernel it has helped me very little, but in turn caused me a ton of toolchain issues and implicit UB that's harder to spot

[fallen bobcat]

honestly I don't even understand how that works, like wont you do context switches in the middle of zig functions?

like i know it works but idk why

[heavy sandal]

the interrupt handling of the cpu handles the in the middle of functions part

is the secret

[fallen bobcat]

i wrote this recently

[heavy sandal]

and so the actual context switch is always explicit

[fallen bobcat]

no because context switches occur not just in interrupts

[heavy sandal]

yes but its always explicit is my point. you always explicitly call a function to perform the yield/switch

thats why it works

it took me like a week to work that out lmao

[fallen bobcat]

ah and that function is C callconv

so it works

yea ok

[heavy sandal]

it doesnt even have to be C callconv actually, you can put the swap in inline asm and have the inline asm clobber feature do the register saves for you, but thats advanced nonsense lmao

but yeah thats the idea

[fallen bobcat]

this is peak zig code lol

[heavy sandal]

took me a week to wrap my head around it but thats because i went straight for the advanced bullshit version with inline asm instead of a C callconv

but thats me lol

[fallen bobcat]

the storage trick is my favorite

i use it everywhere

it's crazy it works

[heavy sandal]

yeah

[heavy sandal]

storage trick?

[fallen bobcat]

you declare a type so that you can declare a variable inside it

like in insert() here

[heavy sandal]

oh yeah that

its so slick

it is (or can be) equivalent to a static local in C afaik

[fallen bobcat]

yeah but ig here I'd do it with a macro in C

[heavy sandal]

mhm

the first time i saw that storage trick was in the csprng in the zig std

[fallen bobcat]

i just bullshitted the compiler until i made something that worked

[heavy sandal]

which uses that trick to keep the csprng state as a threadlocal static local variable

[polar crystal]

yo im curous is this the spleen or terminus font or some else? 🤔

[fallen bobcat]

it's berkeley mono

you're right it kinda looks like spleen a bit ig

[heavy sandal]

tbh the more the merrier, the more people are suffering trying to do dynamic linking/loading of pure zig code the more likely the zig people are to consider actually making a defined ABI for stuff

[polar crystal]

i use too much OpenBSD 😭

appreciate the reply

[fallen bobcat]

do you know ProFont

it looks great too

[heavy sandal]

fwiw it wasnt until last month that llvm even had a defined ABI for bitints

[polar crystal]

oh damn never heard of it

indeed does look great

[fallen bobcat]

i saw a guy post it on twitter the other day on netbsd and thought it looked awesome

[heavy sandal]

oh thats a pretty good font yeah

[fallen bobcat]

nothing can beat the sun firmware font though

[heavy sandal]

i forget what font i use, im not on my dev machine to check and need to leave in like 5 minutes anyway

[polar crystal]

nah nah

no one can beat openindianas tty font

[fallen bobcat]

isn't that just spleen

[polar crystal]

thought it was spleen at fiirst as well

[heavy sandal]

i do know what my favorite pixel font is, its https://media.charlesleifer.com/playground/notes/img/z-peep-12pt.png

[polar crystal]

but it has some differences

[fallen bobcat]

yea but it's pretty hard to get your point across to devs

[heavy sandal]

yeah

[fallen bobcat]

no one listened to me last time i tried improving zig 😭

[heavy sandal]

with what, the allocator improvements? i think that was more the team not having enough people to review shit lmao

[polar crystal]

it uses its own bitmap font

[fallen bobcat]

yeah

[polar crystal]

if i think about it tho and look at it, its actually kinda fine shyt

[fallen bobcat]

at least someone agrees 😭

.

[polar crystal]

skyrim fonts better imo

/jk

[vital summit]

what does this do

[fallen bobcat]

Read the comment 😭

[vital summit]

i see

[polar crystal]

i also have a question

what does it do

[fallen bobcat]

like genuinely or not

cuz the comment seems self-explanatory

[polar crystal]

nah im kidding

[polar crystal]

oh yea about that

what does the comment do

[fallen bobcat]

it's a pretty useful thing actually, I stole it from freebsd

i didnt know __start was a thing

[polar crystal]

finally someone who steals from the BSDs 🙏

[vital summit]

huh

[fallen bobcat]

the linker will generate symbols for sections that dont start with .

[vital summit]

TIL

[fallen bobcat]

also jesus how do apple devs manage to implement freebsd stuff but somehow with 10x more lines

there's a freebsd file that's 600 lines, xnu implements it in 3k

[fallen bobcat]

I cooked up something very cool

[heavy sandal]

oh?

[polar crystal]

ho?

[languid canyon]

?ho

[drifting blade]

?

[solid marsh]

hohoho

[fallen bobcat]

https://github.com/freebsd/freebsd-src/pull/2316

this

[crimson sand]

linux pelt?

[fallen bobcat]

similar yea

[crimson sand]

iirc it was criticized in android world

https://static.linaro.org/connect/sfo17/Presentations/SFO17-307 WALT vs PELT.pdf

[fallen bobcat]

yeah

i know

they have WALT

iirc PELT kinda fixes it by having a util_est too now

but anyway freebsd's target is not like android

[crimson sand]

wdym

[fallen bobcat]

freebsd runs on desktops

and servers

[crimson sand]

gaming?

idk how good pelt is for that

[fallen bobcat]

you're aware that PELT is the linux default right?

the android people didnt like it because power savings or something

[crimson sand]

heard different stories about linux performance in workloads of this kind

[fallen bobcat]

this isnt even full-blown PELT, it's just an additional heuristic to do load balancing

[crimson sand]

like inversions and stuff ruining perf

[fallen bobcat]

basically instead of only checking the number of threads, you also check the actual CPU load of these threads

so in a case where CPU0 has 10 evil threads and CPU1 has 10 mostly idle threads it'll chose CPU1

lol lots of PRs are just straight up slop 💀

[languid canyon]

Why did you decide to implement it for freebsd instead of your own kernel

[fallen bobcat]

Because I will implement it in my own

But like might as well improve an actually useful kernel

[languid canyon]

I see

[solid marsh]

based tbh

i hope it doesn't drown in the sea of slop and it is noticed

[fallen bobcat]

it's weird because they have like 10 ways to send patches

also it's funny because you can't compile freebsd from linux, so I had to boot a VM, modify the kernel and install it, then it crashes -> boot into the old kernel and try again

do this like 5 times until it works, then copy your code back from ssh over to your local machine 😭

[languid canyon]

Why cant you compile it?

[fallen bobcat]

you just can't

it expects a freebsd host

[rocky yoke]

Can't you just run it in nested bhyve?

[fallen bobcat]

maybe, but atp it's easier to just reboot and check if it works

[rocky yoke]

But yeah very cool work

[fallen bobcat]

• i learned how to use ddb and sysctl

so

[rocky yoke]

Did freebsd not have any per thread load tracking?

[fallen bobcat]

no, for load balancing they only kept track of the number of threads

[rocky yoke]

And for top etc?

[fallen bobcat]

@inner python brought this up to me and told me this was shitty so i started looking into PELT by his suggestion

[fallen bobcat]

that's a different mechanism, but not used for load balancing

[rocky yoke]

Isn't it also just an exponentially moving average, at least on linux? What's the difference

[fallen bobcat]

yeah

it's a set of ewmas on linux

this is just one ewma

[rocky yoke]

Ah

[fallen bobcat]

wait idk what you mean

do you mean the % load as reported?

I meant PELT is basically a set of ewmas

[rocky yoke]

I mean both

[fallen bobcat]

well I'm not sure what Linux does, but FreeBSD's % estimation relies on a WMA too but it decays much slower, so it can't really be used for making balancing decisions like that, I assume Linux also does something similar (slower-moving wma for % estimation)

[languid canyon]

But yeah thats pretty cool work, i dont understand any of it (probably because i know nothing about ule)

[rocky yoke]

it's not really ULE specific, right? It's a load estimation that can be used for load balancing decisions

[fallen bobcat]

yeah

it's basically just an ewma of the time spent runnable (running or in the runqueue) so that you can make load balancing decisions more accurately based on cpu usage vs thread count

it's simpler math than linux's because linux doesn't have ticks and does it based on periods of 1024us instead, so you end up with fractional periods etc which is trickier to handle

[fallen bobcat]

the next file i will autistically focus on is the scheduler

i have a bunch of ideas to improve it

[fallen bobcat]

holy shit i had an epiphany

I wasn't sure how to embed the slab metadata in the struct page (also hadn't really thought about it) because it's basically like so:

struct slab {
  struct list_entry link;
  uintptr_t base;
  size_t capacity;
  uint64_t bitmap[];
};

This sounded annoying cuz the bitmap is sized differently directly in the struct.
I figured out you can basically shorten it to:

struct slab {
    struct list_entry link;
    uint8_t kind : 1;
    union {
        pfn_t head_pfn; // points to slab with bitmap in it
        uintptr_t bitmap; // Either an inline bitmap if capacity < 64 or a pointer to a bitmap
    };
};

Capacity is redundant if we remove slab coloring (which wasnt really that useful anyway), we can store it in the zone itself.

The cool trick is that the bitmap can be allocated via a bitmap zone, which itself only embeds its own slab metadata inline (thereby removing any possible circular dependency).
I looked at XNU and they seemed to put everything in the struct page but used a crazy buddy allocator to allocate the bitmaps or something, I think this is much cleanly avoided using a zone for the bitmaps themselves and make them use inline data (although there may be another reason why they do this)

lol i might send another patch to freebsd because they don't do this

dunno why i thought about this, I was doing scheduler stuff

[rocky yoke]

why do you want to include the slab meta data in the struct page?

[fallen bobcat]

Overall less memory usage because you mostly don't waste space on it

[heavy sandal]

yeah

i do it too, its actually quite nice to have

though my slabs use a freelist instead of a bitmap so i dont need to worry about if the bitmap can be inline or not

[fallen bobcat]

There are reasons for doing a bitmap

[heavy sandal]

yeah i know

i didnt say my freelist version was good lol

but storing in the struct page is still good either way

[fallen bobcat]

i wanted to do microsecond accurate PELT in my kernel instead of ticks like in freebsd

the math is mathing

[fallen bobcat]

i may have done something wrong, one CPU has like 300k load the other has 0 wtf

[heavy sandal]

yeah that sounds a bit off

[fallen bobcat]

i mean the numbers should be huge

but like idk why theres such a high imbalance

[heavy sandal]

yeah the concerning one isnt the 300k its the 0

[fallen bobcat]

ok great it seems to work

[fallen bobcat]

perhaps my patch will get reviewed soon 😔

[solid marsh]

based based based

[fallen bobcat]

anyway I added a bunch of stuff to the scheduler which is pretty nice:

• PELT with 1024us period accuracy, which is used as:

1. a primary metric for the long term load balancer
2. a tiebreaker for work-stealing
3. a tiebreaker for pick_cpu() when a thread wakes back up
   The logic behind it being only a tiebreaker in pick_cpu() is that you wanna run ASAP even if the CPU's load avg might be high, as enqueuing on a CPU with potentially many light threads would make you wait for a while

• Each thread keeps a WMA (75% old, 25% new) of its load when it blocks so that the estimated CPU's load is also taken into account when selecting CPUs, so that a case like a few worker threads waking up periodically it doesnt show as a load=0 on that CPU

wow discord is stupid, you can't nest lists

[fallen bobcat]

ah also I'll need to add staleness on the affinity, if the thread hasnt ran for more than 1 second then the affinity is lost

[novel sparrow]

now replace the word sum with the sigma character

[fallen bobcat]

that's a thing?

huh yea

[novel sparrow]

i mean its just a greek letter no?

[fallen bobcat]

nah it's way too ugly

[fallen bobcat]

yeah but typically it's pretty enlarged in math notation

[fallen bobcat]

@heavy sandal do you think you could help me with something rq

[heavy sandal]

maybe, whats the thing?

[fallen bobcat]

for fuck's sake why is zig removing my symbols

when in release build, it strips away the storage:

comptime {
    _ = r.percpu_init_set.insert(&pcpu_init);
}

this is how I call it

like wtf this is just insanely annoying

and I cant use @export sanely because I need a unique name

this is the kinda stuff that is just straight up fucking annoying and stupid

[heavy sandal]

if possible id say use an empty asm block with an input or output set to the thing, i think i ran into something like this with limine protocols

[fallen bobcat]

I did try std.mem.doNotOptimizeAway

on storage

and it didnt work

[heavy sandal]

hm

[drifting blade]

Make an issue

[fallen bobcat]

i tried asking on zulip and no one answered yet

[heavy sandal]

yeah i dont have any other ideas other than to start digging with readelf/objdump

[fallen bobcat]

the zig discord is full of idiots

and issues will get closed

👍

[drifting blade]

Why

[heavy sandal]

yeah the discord is not great

idk that an issue would be closed tho

or why

[fallen bobcat]

i think a reasonable fix is that if linksection is explicitly set then dont optimize it away

[drifting blade]

PR time?

[heavy sandal]

def sounds more like issue time to me

personally

[fallen bobcat]

yeah

im on 0.16 tho

ugh

see this is the bullshit C doesn't have

[drifting blade]

Does Zag not compile with master Zigc?

[fallen bobcat]

it used to but I depend on an external slop project that only works on 0.16

sadly

[drifting blade]

Which one?

[fallen bobcat]

tui library

for my config tool

[drifting blade]

Ah

[fallen bobcat]

it's all because of khitiara that wanted me to remove ncurses

[heavy sandal]

because windows doesnt have ncurses

[drifting blade]

So just make it not run on binbows

[fallen bobcat]

might go back to ncurses honestly idk

the lazy stuff in general is annoying

wish there was a way to disable it

[heavy sandal]

on a per variable basis yeah

[fallen bobcat]

i think this is reasonable tho

[heavy sandal]

yeah

what id do here tbh is make a minimal repro and open an issue

[fallen bobcat]

but I've never managed to actually speak to a zig guy

mhm i will start a thread on zulip perhaps, is it compiler or ecosystem

[heavy sandal]

prooobably compiler tbh

[drifting blade]

Why not make an issue directly, do they really not get reviewed?

[heavy sandal]

they do its just slow

[fallen bobcat]

zulip is like discord kinda

[drifting blade]

Ah k

[heavy sandal]

yeah thats where the devs talk on things that move a bit faster than issues would be good for

[heavy sandal]

tbh that might be timezone stuff too but idk

[fallen bobcat]

andrew is in oregon

[drifting blade]

Meanwhile C3: the dev is in the languages discord server available 24/7

[fallen bobcat]

ig a lot are europeans

true

[heavy sandal]

mhm

[fallen bobcat]

ok zulip post made

time to ping every maintainer

no @heavy sandal ping every maintainer that way i can't be held accountable

[heavy sandal]

lmao

[fallen bobcat]

if zig would add something like __COUNTER__ I could generate unique names

[heavy sandal]

btw can you not just use the export keyword on the storage variable? idk if thatd work for the same reason theres been limine protocol issues before but it might

[fallen bobcat]

if this project wasnt like 18k lines I would probably rewrite in C, it's just annoying that 1. you cant really rely on the compiler that much 2. shit breaks so dependencies are annoying 3. types are cumbersome and verbose to use the @intFromN thing

[fallen bobcat]

but then I cant have any type be in there

[heavy sandal]

right

[fallen bobcat]

like I wanna export functions thru that

[fallen bobcat]

actually no that doesnt work because it's still not a unique name

[vital summit]

18k is not that much tbh

[fallen bobcat]

I already rewrote from C to zig I don't wanna spend time rewriting zig to C

And 18k is quite a lot for what it does lol

[vital summit]

ehj

i would probably spend a day on it

[fallen bobcat]

Are you trolling

[vital summit]

i'm currently rewriting evyrthing in C just to see how different it is

i'm like halfway done

0 clanker

[fallen bobcat]

you can't write 18k lines in a day lol

even 1:1

[languid canyon]

tbh clanker can do rewrites really quickly

especially from something like zig to C

[vital summit]

the point is use 0 clanker anymore

[fallen bobcat]

I'd probably do it using clanker first then clean it up

[vital summit]

i want to rid myself of that

[heavy sandal]

good

[fallen bobcat]

fwiw my OS probably has 0 lines of useful AI code

Only tests

Idk I like zig but sometimes there's just so much bullshit for no reason

[vital summit]

same with rust

[fallen bobcat]

Besides I'd have to find a new name so not doing that

[vital summit]

rust has many QoL things but the toolchain is sooo dogshit

[heavy sandal]

tbh a lot of langs have that tbf, just in different places

[vital summit]

also basically no ENOMEM handling if you use ::alloc

[fallen bobcat]

I feel like C has much less

If any

Bullshit = unneeded verbosity, fighting with the compiler/toolchain, etc

[vital summit]

THIS

they make it so much more annoying than C compilers

i don't get it

[fallen bobcat]

C is so good it's insane

Like when you think about jt

It's like 55 years old

[heavy sandal]

its old enough that people have thrown out the bullshit is i think what c has going tbh

[fallen bobcat]

I don't think there really has been much bullshit

Ever

Probably it's been cleaned up a bit

[vital summit]

especially with C23 there's been many QoL changes

like typeof(x) for example

or constexpr + auto

[fallen bobcat]

but I probably won't rewrite anyway, there's like 1% chance I will. I have to use zig for aura and don't wanna find a new name

[heavy sandal]

i spent a long time fighting build systems and toolchain stuff in 2012 for C stuff so i disagree here, but that shit i was doing is why stuff like meson etc have been invented since

[fallen bobcat]

It's stuff that already existed but standardized so to me it's not life changing

[vital summit]

constexpr has not

bools, nullptr

[heavy sandal]

also C has a ton of random bullshit if you need msvc compat, msvc has so much hot trash and broken nonsense

[vital summit]

i don't support broken implementations

[fallen bobcat]

Based

[heavy sandal]

good

[fallen bobcat]

I think it's fair to assume a GNU C compiler

[heavy sandal]

all the bullshit is contained over in msvc land where we can ignore it because its bad

[vital summit]

i feel like in C it's truly unopinionated, you can do whatever the fuck you want

[fallen bobcat]

Also what I don't really like about not doing C is general compat, modules get annoying to do and you pretty much have to write it all in the same language (like zig)

[vital summit]

wow it's almost like we have the exact same issues

even though we're using different langs

[fallen bobcat]

My dream language would probably be C but with a bit more modern features like stronger types etc without trying to completely reinvent it like a lot of C killer languages

[vital summit]

this is what i wished C++ to be

but it has a bunch of cancer attached to it

[fallen bobcat]

C++ just has too much complexity imo

[heavy sandal]

yeah no

theres a reason i havent touched c++ in years and i want to keep it that way

[fallen bobcat]

I love using C++ and abusing it but it gets tricky to follow

[solid marsh]

c++ is 💀 if you try to use everything

[vital summit]

yeah

you can do the wildest shit in c++

[heavy sandal]

yeah

[vital summit]

see mlibc's sysdep interface

[fallen bobcat]

There's an argument to be made for stackless coroutines tho

They're good

[heavy sandal]

all the C killers seem to try to be as far from C as possible and tbh its not like C is a bad foundation, you can just start closer to it and add stuff instead

[fallen bobcat]

There are probably languages like I'd like, like C3 or odin or whatever I haven't looked at but the ecosystem is non-existent

Everything works with C

[heavy sandal]

yeah

[solid marsh]

c is just that based

sorry

[fallen bobcat]

It is

[heavy sandal]

the lack of stable linkable abi is the only real thing i miss using zig tbh

[vital summit]

in like every lang

idk what's so hard lol

[heavy sandal]

ikr

[solid marsh]

imagine not having a standard

(every compiler makes whatever they want anyways)

[fallen bobcat]

It's our problem, C was made to write OSes, newer languages weren't

[heavy sandal]

i do get why a brand new lang might not commit to a stable abi yet, but it is still annoying not to have

[fallen bobcat]

Maybe we need @long pendant 's new language to save us

[fallen bobcat]

zig just does not commit to anything lol

[heavy sandal]

mhm

[long pendant]

if it ever comes

[heavy sandal]

mood

[long pendant]

i think languages for writing OSes in are critically underexplored

[solid marsh]

problem is no one writes new oses lol

[long pendant]

there are no languages designed for that anymore, only ones from the 60s/70s/80s

[vital summit]

imo a 1.x+ language should absolutely have a standard + abi

[solid marsh]

unless you're (which we are)

[fallen bobcat]

I think jackal is really good for what it is (it's basically C reskinned), I find it really elegant and simple

[heavy sandal]

yeah i agree there, and zig isnt at 1.x yet so im only midlly annoyed at not having an abi yet

[vital summit]

rust is

and that's even worse

[heavy sandal]

yeah

[vital summit]

so GCC-RS is literally just copying rustc

[fallen bobcat]

reminder that zig is like 15 years old lol

[heavy sandal]

crumbles into dust

i really dislike rust so tbh i didnt know that it was 1.x already tbh (as i crumble into dust at how old rust is now)

[long pendant]

it has the major merit that it isn't "actively painful in at least a few ways to write an OS in", which distinguishes it from most others today

[fallen bobcat]

Have you ever tried zig?

[solid marsh]

what special does jackal have

[fallen bobcat]

Curious

[fallen bobcat]

Not much, I just like the syntax

The cool thing are the function templates thing tho

[long pendant]

i'm only loosely familiar with it

[languid canyon]

You can still use the name zag even if the project is in C you know that right

[long pendant]

i followed it a little when it had an IRC channel (2017?) and andrew r kelly was in there

[fallen bobcat]

Basically traits but for functions

[long pendant]

i remember he was a big fan of Cosmic Gate

[fallen bobcat]

But then it doesn't make sense

[fallen bobcat]

I think it still does

And like only Andrew hangs out on there

[long pendant]

latterly it was more interesting to look at again because i found xerox mesa (which is in principle a very remote conceptual ancestor of jackal via pillar via modula-3) surprisingly had its exact same treatment of comptime

[heavy sandal]

wouldnt surprise me tbh

comptime is so good

so much better than macro soup

[fallen bobcat]

Yeah but it's not that useful tbh

[long pendant]

https://xeroxparcarchive.computerhistory.org/cyan/cedar6.1/vm/.VM.mesa!1.html

[fallen bobcat]

Like I had a table to generate using math and I still ended up just precomputing it

using python

because why slow down compile times even more for no reason

[long pendant]

Interval: TYPE = RECORD [page: PageNumber, count: PageCount];
this mesa code is showing exactly the same pattern as zig type declarations

Interval: TYPE = is just a constant declaration (TYPE is just the type of types)

[fallen bobcat]

@heavy sandal we just need to form an osdev cabal with direct contact to Andrew so we can make zig good

[heavy sandal]

lol

not wrong tho tbh

[fallen bobcat]

If zag turns out to be pretty good there's a good chance I'll have influence

[long pendant]

whether or not kelly copied mesa i have no idea, but it's interesting they did the exact same thing

[heavy sandal]

yeah

to both of you

[long pendant]

this of course makes Zag the modern-day Cedar OS

[fallen bobcat]

I'll remember to ask him eventually

Because it is pretty interesting

[heavy sandal]

@fallen bobcat btw for your issue, have you tried doing verbose llvm ir and checking if the thing is in the IR or not?

nvm, someone in zulip checked that, it seems llvm/lld is killing it

@fallen bobcat btw have you tried somehow doing KEEP in the linkerscript for your set sections? since latest testing from others in zulip says its llvm/lld not zig that is removing it

[fallen bobcat]

the idea is to not have to rely on that, unless i can do KEEP(set_*) or something

[heavy sandal]

yeah would def be ideal not to worry about it

[heavy sandal]

that would in theory work yeah

[fallen bobcat]

https://github.com/bagggage/bamos/ i found a zig OS, to me it seems mostly vibecoded 💀

yeah ok def is

[stray kelp]

lol

[drifting blade]

What makes you think that

[fallen bobcat]

one line the comments are perfect english the other the comments are written in very poor english

[inner python]

They could’ve gotten really drunk

[fallen bobcat]

@heavy sandal LOL

this is what I had to do

what horror

[heavy sandal]

lol

[fallen bobcat]

i looked a bit at tigerbeetle's code and I want to try refactoring everything a bit so it's cleaner and higher quality (more similar to theirs)

[heavy sandal]

yeah their stuff is very good quality from what ive seen

[fallen bobcat]

one day freebsd PR will get merged...

meanwhile im currently doing prekernel stuff a bit every day

[fallen bobcat]

maybe it's my code, but why does zig use so much stack 😭 16kib is not enough to get through init

ok i pushed the basic prekernel stuff, only maps stuff for the kernel and loads it atm, no pfndb mapping and stuff yet

[heavy sandal]

sometimes copy elision on struct init fails for some reason iirc, mostly because more aggressive copy elision has caused issues with incorrect code in the past and an extra copy is usually preferred to a crash is my only guess other than just saying it's your code

[fallen bobcat]

i think it's because this is a few calls deep and also logging uses a bunch of stack apparently

[heavy sandal]

hm

[fallen bobcat]

or just formatting in general i think

[heavy sandal]

depends ig, for the most part logging should just get inlined to your logfn

oh yeah no formatting is kinda like that sometimes

i wouldnt expect it to be that bad tho

formatting in zig is usually more of a code size issue than a stack one ime

ive wanted for a while to see if theres any good way to do stack size static analysis

[languid canyon]

another C W

i get the exact stack usage i expect with no hidden bs

[fallen bobcat]

I mean you'd get the same in C

I think this is just a standard library thing idk

maybe not

[heavy sandal]

idk

zig isn't doing any hidden stack allocs, it's all about just what functions are used

and how much stack those use

[vital summit]

does it also construct every object on the stack

[fallen bobcat]

ah i figured out why it was using so much stack

[heavy sandal]

oh?

[fallen bobcat]

i did a oopsie and had const features = amd64.cpu_features; in the function

so allocating a new one on the stack

[heavy sandal]

oh yeah no thatll do it

[fallen bobcat]

and that struct is quite big

[heavy sandal]

yeah

[vital summit]

i don't understand

can you translate for a C brain

[fallen bobcat]

big struct on stack go boom

[heavy sandal]

declared a local of that struct type instead of getting a pointer to the global of it

[fallen bobcat]

yea

[heavy sandal]

aka forgot the &

[vital summit]

hm

but that struct doesn't seem too big?

[fallen bobcat]

ig there's other stuff using the stack

[heavy sandal]

19 bools, plus 48 chars, plus however big the CpuVendor type is

[fallen bobcat]

cpuvendor is an enum

[vital summit]

that's nothing tho

[fallen bobcat]

true idk

but it worked

so ig the stack space is limited

[vital summit]

rust uses multiple pages for stack doing nothing xd

[heavy sandal]

so depends on the backing int for the enum it could be anything from 68 to 71 bytes

and then round up for alignment depending on whats after it on the stack

[fallen bobcat]

I've decided that maybe an initgraph thing wouldnt be that useful idk

considering how I do initialization is kinda well defined idk

id have to think about it

[heavy sandal]

yeah i had the same conclusion

my init is sufficiently well defined i dont need to toposort it

[fallen bobcat]

well, I also ended up flattening the whole call stack a bit

now the entry point is in ex and it calls ke.init

[heavy sandal]

mhm

nice

[fallen bobcat]

so the stack is more shallow by like 2-3 stack frames

[heavy sandal]

that might help a bit yeah

[fallen bobcat]

and no layering violation now

is it better to have the whole function exported or should I have a thin callconv C function that just calls ex.init

I dont think it matters

[heavy sandal]

yeah doesnt really matter

[vital summit]

smh the fn decl is ripped from zinnia

[fallen bobcat]

wdym

[vital summit]

i have exactly the same function declaration for my main

[fallen bobcat]

marvin when are you joining the zaggers

[heavy sandal]

id use a specific callconv instead of .c tbh, just to be explicit

[fallen bobcat]

can I do .sysv?

idk i want this to be portable

the loader assumes sysv

[heavy sandal]

yeah thats what i do (though its like .x86_64_sysv or something like that) - i have the arch code specify that arch's calling convention

[vital summit]

my kernel also assumes sysv everywhere

[heavy sandal]

so i use the aarch64 standard one for aarch64 and sysv for x86 everywhere

and that way im never worrying about if some weird target setup i try later accidentally swaps it to a different callconv and breaks things

[fallen bobcat]

@vital summit consider it instead of rewriting zinnia for the 348348th time

[heavy sandal]

(note that this doesnt work for x86_32 if you mix callconvs and use llvm, as llvm for some reason refuses to allow multiple different calling conventions in a 32-bit x86 binary)

[fallen bobcat]

We could be the next managarm 🥺

[vital summit]

yea but I'm retarded

[fallen bobcat]

You can do stuff you're comfortable with

idk what you're good at but I don't like drivers and architecture/hardware code in general

And that's not really tricky

[vital summit]

i will write drivers for you if you can unfuck my rust scheduler

[fallen bobcat]

What's the problem with it

[vital summit]

half of it is vibed because i didn't know how to debug it

and i REALLY don't like it

i tried to understand it, but it kept breaking under SMP so i asked claude to fix it

it works, but i feel like it could be improved a lot

also, i like your design

[fallen bobcat]

what design

[vital summit]

the one you sent me

[fallen bobcat]

ah

yea

[vital summit]

the ule one

[fallen bobcat]

It's basically ule with ntisms stolen from mintia and a bunch of my own stuff

[fallen bobcat]

Why do you want to unfuck the rust code if you're rewriting in C tho

[vital summit]

good point

i'm tired

i will probably lose a lot of github stars if i rewrite

[fallen bobcat]

That's why you need to join the zig hype so you get a bazillion stars

[vital summit]

idk how to feel about zig

it's not quite C, but also not quite rust

[fallen bobcat]

I'm not saying you rewrite in zig Im saying you join a zig project

(cough cough)

[vital summit]

i mean i can do some stuff for it

(but you made it sound like i should abandon zinnia for it)

[fallen bobcat]

You do what you want

[vital summit]

i'm a loser chud

[fallen bobcat]

i did more prekernel stuff, now it maps the pfndb and detects cpu features for the kernel

[heavy sandal]

nice

[fallen bobcat]

I kinda want to do AP startup as well but I'm scared of duplicating ACPI code...

[heavy sandal]

if you want cpu hotplug then youll need ap startup in the main kernel anyway

[solid marsh]

eh

what will you duplicate

getting acpi tables and iterating over the madt

[fallen bobcat]

yeah

that

[solid marsh]

doesnt sound too bad to me lol

[fallen bobcat]

I mean I'll need it eventually anyway if I want NUMA

[heavy sandal]

yeah

[fallen bobcat]

bleh ok fine

[heavy sandal]

youll need at least the acpi table getting for numa

[languid canyon]

Cant you just make an acpi.zig file and share it

[fallen bobcat]

yes but my acpi.zig file rn also has init and stuff

like

sec

https://github.com/rdmsr/zag/blob/master/src/platform/acpi/root.zig

[heavy sandal]

itll get culled anyway

because zig is lazy

[fallen bobcat]

I can reuse the struct definitions I guess

but the functions no

[heavy sandal]

could split the functions maybe

but yeah the structs is already a big help to reuse

[fallen bobcat]

https://github.com/rdmsr/zag/blob/master/src/platform/acpi/root.zig#L263-L287 like this relies on mm.p2v

which is not a loader thing

[heavy sandal]

mm tru

could generic parameterize it

[languid canyon]

does zig support --gc-sections

[heavy sandal]

yes, but also zig is lazy enough that we dont need a preprocessor while still having every single os's support in the std

if you never refer to something it doesnt even get compiled to need to be gc'd