Zinnia

[marsh holly]

https://github.com/Mathewnd/Astral/blob/rewrite/kernel-src%2Fio%2Fkeyboard.c#L109

It literally just sends the packets from all keyboards to the ringbuffer of this specific keyboard

[near tartan]

[late bramble]

is it really that easy?

[near tartan]

yeah

as long as you're doing fbdev

[marsh holly]

Xorg is piss easy

[late bramble]

🎉

[near tartan]

i just need to implement fbdev and i guess i also need sockets

[late bramble]

MeniXorg

i have sockets, but no fbdev rn

[rugged shard]

Menix Org

[near tartan]

idk how sockets work even xd

[rugged shard]

black magic

[late bramble]

these are sockets in french:

[marsh holly]

Socket Listen socket connect accept write read wirte read close

[near tartan]

no on the kernel

is it just pipes with extra steps

[marsh holly]

Pretty much

At least stream ones

[late bramble]

I only have socketpair and unix ones (on filesystem) iirc

[grave peak]

Well bidirectional

SCM RIGHTS is the hard part

[marsh holly]

Some bsd even implements pipes using sockets

[near tartan]

struct socket {
    struct pipe a, b;
};

[marsh holly]

https://tenor.com/view/взгляд-2000-ярдов-война-war-soldier-gif-3632617944134077161

[late bramble]

what's that?

[marsh holly]

We do not speak of the wicked ancillary data

[marsh holly]

Fd passing

[grave peak]

Giving other processes your file descriptors

[late bramble]

sounds nice

[grave peak]

Including mmaped buffers, or literally anything

[near tartan]

lace

[grave peak]

Its the backbone of everything

[near tartan]

sounds like microkernel capability passing

[grave peak]

Thats why its called rights

Its basically capabilities

Another word for fd

[marsh holly]

Openssh uses it for privilege separation

Wine uses it for something

Not sure what

[grave peak]

Window managers use it for shared buffers

Etc

[near tartan]

i should probably first do fbdev before going straight into drm

even if it's just simpledrm

[grave peak]

If you're doing xorg then u can do that

[near tartan]

for the moment i'll be doing xorg

just makes everything easier

[grave peak]

Yeah that requires very little

They literally used to ship GPU drivers as part of xorg

In pre DRM days

[near tartan]

yeah

[rugged shard]

what the

[grave peak]

Yeah

[marsh holly]

Yeah lmao

[grave peak]

Thats why xorg supports dynamic modules

[storm bobcat]

dont they still? with the stuff like xf86-xorg-amd or whatever

[near tartan]

nobody's using it at least

[near tartan]

i think i fixed this

but bash doesn't like me anymore

[cyan nexus]

make your own bash and call it mash

or try fish

[near tartan]

💀

fish is rust

that needs a port of that first

first time that menix has survived running for longer than 1 minute

I'm trying to see how long it can stay alive

[grave peak]

Run fastfetch in a loop

[near tartan]

how

12mins in and so far no crash

[cyan nexus]

spam to console

using a for loop

[near tartan]

it might be killing itself at some point when it runs out of ram

[cyan nexus]

it shouldn't tho

[near tartan]

because i don't actually free any memory after process destruction

[cyan nexus]

ah

[near tartan]

at least for the page tables

[cyan nexus]

why not, lol

[near tartan]

effort

[cyan nexus]

...

[near tartan]

i will eventually lol

[cyan nexus]

it's literaly a few lines of code

[near tartan]

i mean i'd have to traverse the page tables

[cyan nexus]

i have separate page tables per process

[near tartan]

so do i

[cyan nexus]

so i just have to loop into the process table to free them

tho i might need to find a faster way

[near tartan]

the problem is

you need to either parse the tables

or record the list of physical pages in use for it

and you need to consider CoW

[cyan nexus]

huh?

oh

CoW

[grave peak]

Bash has loops

[cyan nexus]

i didn't do it yet

[near tartan]

well i do CoW

so it's not as easy

wooo 15 minutes in

note that this is running with smp=4

so that also works pretty well

[cyan nexus]

why errors on top?

[near tartan]

didnt implement sysconf lol

oh noo

it's over

i did cat

and now i'm stuck forever

[hoary cave]

fym

the CoW object should know how many references to each page there are

[near tartan]

yes bro

[hoary cave]

ok

[near tartan]

that's what i mean bruh

[hoary cave]

so then when it's dropped

you can just handle it there

[cyan nexus]

why XD

[hoary cave]

memory objects are behind Arc

no ^C

[near tartan]

what is syscall cancellation anyways

btw the fix for this is incredibly stupid.

[hoary cave]

when menix cancellable syscalls

[near tartan]

after nvme

once that works i'm speedrunning x11

so some hardening + sockets + fbdev + evdev

[rugged shard]

do a fork bomb

[near tartan]

lol

okay without any sort of process freeing i can spawn 305 instances of fastfetch with 8 gigs of ram

[vast lotus]

well that would explain why none of my kernel changes seem to be taking effect

[near tartan]

what's the context lol

[vast lotus]

i'm only changing the actual disk image

but it's not booting off of that

[near tartan]

did you mix up threads?

[vast lotus]

oh

yes

[near tartan]

lmfao

[grave peak]

lol

[modern hamlet]

not a worry

i think the best way to do it is 1 page table per process

period

now it might seem stupid to suggest a naive idea like that, but let's be honest, are you really going to clone page tables anyway

no, just fill those in at page fault time

"but where am i going to get my private anonymous mappings" from amaps that you create for those mappings

so instead of copying-on-write the page tables, create-them-on-write

now i realize im preaching something i dont practice, but thats only because i havent implemented fork yet

[near tartan]

fair enough

i just need to figure out the best way of recording which pages are used

either i just traverse all levels manually

or save the pages in a vec

[storm bobcat]

New fish is rust, real fish is C++

[near tartan]

real

[storm bobcat]

No like

For real

[near tartan]

i know

fish 4.0 is rust

[storm bobcat]

They're rewriting it in rust for no reason

[near tartan]

yeah

but whatever

not like rust is unportable

[storm bobcat]

Run the fireworks test

[near tartan]

source?

[storm bobcat]

it's in boron

But like unironically do try running it in the kernel

it will stress test your scheduler

[near tartan]

in the kernel kernel or in userspace?

[storm bobcat]

kernel

[near tartan]

ah

yeah i'll try

[storm bobcat]

I guess you can run it in userspace if you have a way of creating threads

And mapping the framebuffer

[near tartan]

not yet

[storm bobcat]

But the problem is it will most likely find bugs

Which is annoying

Like I've been debugging it for around a week but I fixed like 3 major bugs and now it runs fine without smp, just gotta figure out why it's dead locking for no reason on smp (it seems a thread at IPL DPC is being rescheduled???)

[near tartan]

i'll see

my scheduler seems more robust recently (speculation)

ever since i fixed the irqlock shit

especially with smp

[storm bobcat]

It's cool because it creates a shitton of threads and you can visually see them running

[near tartan]

yeah

[cyan nexus]

might do that too

[near tartan]

good to see that my xbps repo is working

it's just not signed but those are details

[spark surge]

neat!

[marsh holly]

going to replace your packages with slightly broken versions and you will have no way of knowing

😈

[near tartan]

i wanna port xorg

[marble salmon]

If it's not signed you can't xbps-install from it

[near tartan]

yeah you need to rindex it

but i was just trying to see if it can fetch the packages from the repodata

[marble salmon]

unsigned rindex is not enough

you need a signed rindex

for http/https repositories

for xbps repos on the local fs it doesn't matter

[late bramble]

do it

now

[near tartan]

i did

it just doesn't run yet because i want drm

[late bramble]

implement drm

[near tartan]

and I don't have sockets

i am going to

[late bramble]

implement sockets

i wanna see menix running MINECRAFT

[near tartan]

ive actually never used sockets in C directly

aren't they just bidirectional pipes

[late bramble]

what a great way to learn

uh kinda

as long as your are using the unix ones I think that's basically it?

if you want to do networking though it becomes more complicated

[near tartan]

i probably want to do that

[late bramble]

yeah but networking isn't required for xorg as far as i am aware

only unix ones

[near tartan]

networking is something for later

[late bramble]

gotta send that ping and receive a pong

[near tartan]

one of the goals is hosting its own website

[late bramble]

yeah that would be neat

currently maestro doesn't have a website and that's an issue I'll have to fix

[near tartan]

a friend of mine made one for me

it's a bit too flashy imo but I'm appreciative

https://menix-os.org

nice embed

the website is total overkill

I'd rather have something like my own website

https://mrvn.sh

[late bramble]

looks nice to me

[near tartan]

yeah because your blog also looks like that

[late bramble]

i started making a website myself for maestro (repo private for now) where I was using https://github.com/copy/v86 to make a demo

[near tartan]

im a minimalist when it comes to design

[late bramble]

well I like minimalism too, but I also like pretty stuff

[near tartan]

i like xfce, that should tell you how much i like pretty things

[late bramble]

right

I use KDE plasma and my design model for Maestro is macOS (the glassmorphism version, not the new liquid stuff)

[near tartan]

i think the two biggest issues right now are input and output

[late bramble]

on what?

[near tartan]

menix

i dont know how to best handle input

sure i could use the evdev interface

if i don't then I'm likely losing libinput

and in turn most wayland stuff

[late bramble]

I think I'll do that

[near tartan]

or I'll bite the bullet and resort to xorg

maybe someone ends up making a modern x11 server

[near tartan]

but you're linux compatible

I don't really want to do that unless i absolutely have to

that's also why I'm using drm over linux fbdev

[late bramble]

to be more precise, I implement a subset of Linux.
Because I don't want (or can) support every single feature (some of which are legacy stuff anyways).
And if you want all of the Linux features, just use Linux directly

[near tartan]

yes but you're source compatible

i don't want to have to use linux headers for example

[late bramble]

tbh I don't even know what I am doing at this point

[near tartan]

just use mlibc

[late bramble]

oh yeah I do use those

[near tartan]

like

[late bramble]

I surely will try it at some point

[near tartan]

"i want my own kernel but it depends on another Kernel's interface"

???

[late bramble]

yeah

[near tartan]

the more people use mlibc the better

because patches will help all sysdeps

[marsh holly]

Or a wayland xf86-input-keyboard/mouse equivalent

Idk how wayland works

[near tartan]

wayland compositors depend on libinput

which is only really designed for evdev

[marble salmon]

you could just add your own input interface to an existing compositor

[marsh holly]

Me when I write nonportable software

[late bramble]

the opposite kinda feels like:

[near tartan]

but that's kinda the point

if i made a linux compatible kernel I'd rather use linux

[marble salmon]

i wonder how hard it is to make a simple wayland compositor

[near tartan]

with wlroots it should be piss easy

[late bramble]

I started writing a xorg server (in BLAZING FAST RUST 🚀🚀🚀) long ago but I think I'll turn it into a wayland compositor instead

[near tartan]

you can hack together something in less than 1000 lines

[marsh holly]

Yeah someone should make a wayland compositor thats as portable as xorg

[near tartan]

[near tartan]

write it in C

so it's actually portable

[late bramble]

no u

[near tartan]

tbh just write a portable libinput

just one more layer of indirection

[marsh holly]

Lets crompromise, all of you write in v

[near tartan]

[marble salmon]

uWayland

[near tartan]

i mean

udrm failed because I'm lazy and it's much smarter to just support lil instead

[late bramble]

|| je méchant ||

[near tartan]

but something like libinput-ng isn't a terribly stupid idea

libinput just has a shitton of code

[late bramble]

I wanna do network and graphics

[near tartan]

oh god

i found a stupid pmm bug

first time in a while that this happened

apparently a zero sized allocation is trying to free pages on an already freed allocation

because you need to free at least one page

so i had two consecutive allocations, one size 0, one size 4096. freeing just one of them works, but coincidentally i wanted to free both right after each other

[hybrid island]

Why do you have a 0 sized alloc

[near tartan]

and since the first free already removes the USED flag from the pfndb entry it ended up double freeing

[near tartan]

I'm pretty sure i have to

[hybrid island]

proceeds not to tell me why

[near tartan]

wtf do you mean why

[hybrid island]

I'm curious

[near tartan]

i mean now i just panic on zero sized allocation

it's an oversight

rust doesn't allocate ZSTs i think

but this was a manual phys allocation

[near tartan]

:D

[hybrid island]

nice

Does that mean you'll have actual storage drivers soon then?

[near tartan]

yes

[hybrid island]

yay

[near tartan]

and probably an ext2

[hybrid island]

Do you have that one yet?

[near tartan]

no lol i don't have block drivers

when that works i can focus on running x11

after nvme first thing i'll do is drm and kms

[hybrid island]

You could probaby nab large parts of ext2 from BadgerOS since that's got a working Rust ext2 impl

[near tartan]

license?

[hybrid island]

MIT

[near tartan]

but idk if the interfaces are too similar

[hybrid island]

They're probably similar, though I'd expect the names for certain things to be slightly different

From what I know about your VFS it should™ fit together more or less, since you also do a VNodeOps-like thing

[near tartan]

i kinda follow linux naming

i have a superblock trait

which is an fs instance

[hybrid island]

Well if anything you could always use it for reference

[near tartan]

sure

do you also implement writing

[hybrid island]

Full base ext2 read/write

[near tartan]

[hybrid island]

Even checks compat flags

[near tartan]

i'll 99% need vfs adjustments anyways

since tmpfs has this big exception on page caching

with no backing device

[hybrid island]

I should note that page caching happens on the backing device for my VFS

[near tartan]

huh

[hybrid island]

I know, not modern or whatever, but it works

[hybrid island]

blkdevs have a page cache before them, and that's it.

[near tartan]

oh well i have a per vnode page cache

[hybrid island]

vfs caches some superblock stuff too

[near tartan]

and when you map a node it uses freebsd like pagers

[hybrid island]

I would suggest that if you do intend to take from my VFS, that you make similar block device APIs to mine

[near tartan]

cc @fair lintel

[near tartan]

okay nice my nvme driver can now also write into the image again

[near tartan]

seems to work

[hybrid island]

Nice

[near tartan]

alright nvme driver is merged

[late bramble]

The neverminde driver

[knotty umbra]

I saw this once, and after that couldn't really read it of any other way. NVM Express -> NeVer Mind Express

[near tartan]

i'm still working on some speedup improvements

my block io interface is extremely primitive

current read speed is 300MB/s with a block size of 0x2000

[near tartan]

today I'm finishing the rest of the broken ports

once those work i can try the drm modesetting example

[near tartan]

ok nice, modesetting example built successfully

[near tartan]

more mlibc fixes....

[near tartan]

marvin@fw13 ~/r/m/s/menix (main)> tokei kernel/ drivers/
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Language              Files        Lines         Code     Comments       Blanks
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 LD Script                 2          219          161           15           43
 TOML                      6           91           74            2           15
─────────────────────────────────────────────────────────────────────────────────
 Rust                    115        16470        13769          507         2194
 |- Markdown              71          788            0          778           10
 (Total)                            17258        13769         1285         2204
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
 Total                   123        17568        14004         1302         2262
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

damn

[somber solar]

is that cloc but rewritten in BLAZINGLY FAST RUST 🚀 🚀 🚀

[near tartan]

yes

[near tartan]

yay

i'm hacking together a few openrc scripts to replace whatever upstream provides

because those scripts kinda suck

[near tartan]

fastfetch working again

(ignore the IP that's just a placeholder until mlibc makes getifaddrs work on !linux)

[marsh holly]

67

[somber solar]

256 in u8 is impressive

:D

[marsh holly]

Lmaooo just noticed

[near tartan]

67 subnet in ipv4

but yeah it's just a string in fastfetch

[idle flower]

the forbidden ipv4 address

[full topaz]

I thought /23 was wierd, but /67? that feels more cursed than pandoras box

[near tartan]

why is 23 weird

[storm bobcat]

https://tenor.com/view/67-67-kid-edit-analog-horror-phonk-gif-3349401281762803381

[full topaz]

odd number

and it just feels wierd

[near tartan]

it's fairly common

[near tartan]

i still kinda want to port an existing init system

openrc is weird, there's a shit ton of shell scripts

not sure if i like that

ideally i'd like to provide my own files

looking at this table right now

[marsh holly]

/etc/rc or sysvinit

[near tartan]

what's /etc/rc

[marsh holly]

Super ancient way of doing init stuff

You just have init run the shell file /etc/rc

Ita what astral does

[near tartan]

do you have like a astral-utils thing?

or wait you have the linux option you can just use that

[marsh holly]

Neither

But I do have some astral specific topls

Tools

I have considered proting sysvinit to astral

Especially as it outgrows the etc rc thing

[near tartan]

i was considering dinit

just something more modern

[near tartan]

we got partitions

hell yeah

[hoary cave]

why is it registering the namespace device after it discovers partitions

💀

but that's still really cool progress

finally menix !dead project

[near tartan]

it's registering the ns block device and that function scans for partitions before registering

[ionic lava]

Is this qemu run? Show me nvme options, please. I can't make nvme storage to appear (ufs too). Only host controller does.

[near tartan]

yes it's qemu run

-device nvme,serial=FAKE_SERIAL_ID,drive=disk

[ionic lava]

Hm. That's all? Weird. I have the very same thing and doesn't show up. I guess it shouldn't matter what type of virtual disk one uses, right?

[near tartan]

qemu: ovmf/ovmf-code-$(ARCH).fd build-$(ARCH)/menix.img
        qemu-system-$(ARCH) $(QEMUFLAGS) \
                -drive format=raw,file=build-$(ARCH)/menix.img,if=none,id=disk \
                -device nvme,serial=FAKE_SERIAL_ID,drive=disk

this is the full one

[fiery marlin]

cool

[spark surge]

awesome

[somber solar]

nice

[hybrid island]

niiiice

[potent hamlet]

niceee

[somber solar]

niiiiceee

[near tartan]

next week i will add the drm subsystem (for real this time)

together with plainfb

[grave peak]

based

[near tartan]

when the build system is so shit that you need to patch it yourself

[potent hamlet]

https://articles.manugarg.com/aboutelfauxiliaryvectors

which way is the auxv oriented

is the first thing at the top of the stack?

[near tartan]

no

the first thing you write is the terminator

this is what i do

[potent hamlet]

is that the entire thing?

can i just steal it

[near tartan]

for auxvals? yes

[potent hamlet]

elf auxv for mlibc yeah

wait wtf why am i on a thread on not on the misc channel

[near tartan]

bruh

[potent hamlet]

mb bro 💀

[near tartan]

i mean my entire ELF implementation is only 580 lines (including the declarations)

[potent hamlet]

damn does it support dynamically linked stuff as well

[near tartan]

ofc

[potent hamlet]

580 lines is pretty concise tbh

[near tartan]

https://github.com/menix-os/menix/blob/main/kernel/src/vfs/exec/elf.rs

[potent hamlet]

ty

[near tartan]

okay i ported dinit now

was less annoying than i thought

but now it wants unix sockets 😭

[storm bobcat]

not really but tbf he has constants in the file

[near tartan]

the actual logic is just 200 lines

including stupid rust formatting

[somber solar]

does it need /proc and /sys

[near tartan]

no

it didn't even need any patches

[somber solar]

nice

[near tartan]

the build system is shit though

[near tartan]

it also wants to use sockets very early on

[vast lotus]

can't you just do make PREFIX=/usr? iirc the = operator only sets if the variable is unset

[vast lotus]

elf dynamic linking is incredibly easy on the loader side, all you do is load a second elf file if pt_interp specifies one and call that instead of the main binary (passing info about the main binary through auxv)

the actual library loading and relocation etc is the interpreter's job, the loader doesn't care about it

[somber solar]

what's ?= for then?

[vast lotus]

I think that's the same for :=

[near tartan]

no :)

[vast lotus]

:= evaluates and sets the var to the result whereas = sets the var to the expression and evaluates when it's used

[near tartan]

it ignores it for some fucking reason

[vast lotus]

that's fucked

[near tartan]

maybe because sub makes?

nah that doesn't make sense

dinit does that

i hate hand rolled makefiles

[modern hamlet]

no thats ?=

= just sets the value regardless

[near tartan]

trying to run os-test is going great

[fair lintel]

lmao

[near tartan]

whatever the fuck that means

[fair lintel]

oh bruh without signals and futexes it is not going to work, like at all

[near tartan]

funni

also

[near tartan]

you will cringe at this

i didn't actually implement getdirents

so to even run make i need to specify the file xdddd

[late bramble]

I have signals but not futexes, rip

[cyan nexus]

XD

[near tartan]

idk how to even implement signals

i haven't found any good resources

besides "just look at the code bro"

[late bramble]

man 7 signal good

[near tartan]

that's just the user api tho

idk how the kernel side is supposed to work

[late bramble]

look at the section Execution of signal handlers

it has the steps the kernel does to execute a signal handler

This paragraph in particular is important (to me):

From the kernel's point of view, execution of the signal handler code is exactly the same as the execution of any other user-space code.  That is to
say, the kernel does not record any special state information indicating that the thread is currently executing inside a signal handler.  All neces‐
sary  state information is maintained in user-space registers and the user-space stack.  The depth to which nested signal handlers may be invoked is
thus limited only by the user-space stack (and sensible software design!).

[fair lintel]

yeah, for signals you decide what to do based on the signal and the handling state

so for signals that terminate the process, terminate the process

and make that info available via waitpid

for signal handlers, push a signal frame to the stack and set up the rest of the state, set PC to the handler, go

etc.

[late bramble]

sigreturn just takes the info on the user stack and restores the state according to it

[fair lintel]

futexes should be quite easy to get going, I can only recommend doing it

[late bramble]

I have to understand how they work first

[fair lintel]

man 2 futex

[late bramble]

yeah

[fair lintel]

although linux obv has a quite large API surface for it

for mlibc, wait and wake are enough

[somber solar]

@late bramble have you switched to using mlibc?

or planning to

[late bramble]

I have not switched but plan trying it at least

I like musl, but I haven't tested mlibc

[marble salmon]

wtf, you have bash but no readdir?

[late bramble]

I am actually refactoring my getdents right now because I am simplifying filesystem implementations

[near tartan]

yeah

the problem is that i'm stuck on the interface design

do i just return a Vec<uapi::dirent> from the VFS?

since i can't just let the kernel write to the buffer in unmapped memory

cc @late bramble

[late bramble]

I am passing a closure to my filesystem implementation, which calls the closure for each entry.
The closure writes the entry in userspace memory and returns a boolean saying if we can continue or must stop

[somber solar]

hmm this made me think

I could use std::generator

it looks nice :P

iirc the behaviour is undefined if the contents of the directory change between getdents calls

[near tartan]

what worries me is the offset parameter

[somber solar]

d_off?

[near tartan]

yeah

[somber solar]

is it used by anything?

[near tartan]

no idea

[late bramble]

it is whatever you want

you can use it as a cookie that you interpret the way you want in your filesystem implementation

iirc getdents uses the offset in the open file description to know where the last getdents call finished, so that it can resume at the right location. The offset can contain any value just like d_off

[near tartan]

how stupid would it be to implement read() for dirfd

instead of having a separate syscall

[late bramble]

that does not sound practical

but doable I guess?

[near tartan]

i mean readdir already exists

and opendir

[late bramble]

yeah but they use getdents underneath

[marsh holly]

the olden unices could do that

but iirc they would literally shit out the actual directory file data which blegh

[late bramble]

yeah

[near tartan]

okay scrapping that

then i don't need to care about the offset argument

that's not a thing on mlibc sys_ level

[near tartan]

[near tartan]

i have decided on a goal before 2026

i want to have signals, (unix) sockets and Weston working by the end of the year

[marsh holly]

bro do you have 48 hours in a day

[near tartan]

yes

[marsh holly]

send method I need that too

[near tartan]

weston can't be that hard

it's just drm + evdev

signals don't have to be complete

I'll just do the bare minimum for those

and unix sockets should be relatively okay

unless I'm underestimating things

i will have the PTO until next year so i have enough time to get shit done

[marble salmon]

I don't think you need signal support to make weston work

You need signal support for other things though (such as ctrl+q working)

[idle flower]

https://tenor.com/view/teach-me-doctor-strange-benedict-cumberbatch-gif-8215241

[near tartan]

oh nice then

i already started a half baked plainfb drm impl

so ig the biggest pain point will be sockets

[near tartan]

because i only want weston to draw something this year

it doesn't have to work properly yet

[late bramble]

how

[past leaf]

sockets are easier than signals

and Xorg used signals

(well you can disable them)

[near tartan]

wayland

[near tartan]

i've been working to split out some wayland linux APIs in mlibc so i can use libinput and libwayland too

namely epoll, signalfd and timerfd

@grave peak DRM commences today

[grave peak]

letsgo

[near tartan]

drm update

today i did CRTC stuff

maybe we'll see some colors today???

[near tartan]

note to self for today:

move uapi to rust

write drm uapi bits

use the dumb drm impl instead of trying to make something fancy

[marble salmon]

uapi?

[grave peak]

maybe linux userspace headers

[marble salmon]

also what do you mean by "dumb drm impl"?

[grave peak]

dumbfb only or smth

[near tartan]

user api

technically it's user abi

[marble salmon]

yeah but what do you mean by that?

[near tartan]

currently i generate the bits from C headers

if i wrote them in rust instead, i could e.g. derive extra macros

and it also seems like bindgen doesn't understand size_t, so those types are all wrong

[marble salmon]

i'd probably re-use @fair lintel's rust abi gen script

seems significant easier than maintaining out of band code

[near tartan]

yeah that was my thought

I'd still have to maintain the bits manually

for annotations and integration with e.g. usercpy

[near tartan]

i thought of copying linux/managarm drm stuff, but after seeing that Aero does exactly what i had thought of before, I'll just do that instead

i had thought to myself "no way that will work"

probably another certified overthinker moment ™

[marble salmon]

you can just impl traits as long as you're in the same module or crate

[near tartan]

yeah but with bindgen that doesn't really work iirc

[marble salmon]

what does Aero do?

[near tartan]

it just stores a bunch of fields in every mode objecr and then implements a ModeObject trait over it

[marble salmon]

also, you can use newtype patterns

if you want to impl extra stuff on your abi types

[marble salmon]

why is that "dumb"?

[near tartan]

dumb as in simple

like dumbfb

ig it's a bad choice of words

[grave peak]

thats literally what its called in DRM

dumb framebuffer

[marble salmon]

no, dumb buffers in DRM are a specific thing

what Marvin is talking about is an implementation strategy

[near tartan]

yes

[marble salmon]

unrelated to dumb buffers

[grave peak]

ah

[near tartan]

i mean i also have to impl dumb buffers

anyways

time to start working

[near tartan]

okay we're getting somewhere

the modsetting example that I'm using was able to detect a crtc

if i hook up the plane logic we can probably see some results very soon

[hoary cave]

cool stuff

[near tartan]

bruh

wayland feels like a fucking waste of time

weston has os-compatibility.h, but puts linux specific stuff everywhere unconditionally

[hoary cave]

ur compiling linux specific source files and wonder why it dies?

[near tartan]

bro

i should stay with x11

tbh

[hoary cave]

lol

weak

[modern hamlet]

linux-file-sync.c

dont compile that file

[hoary cave]

that's what i said

[modern hamlet]

yeah

[near tartan]

is this rage bait

[hoary cave]

no?

[near tartan]

THEY UNCONDITIONALLY DEPEND ON THIS

[hoary cave]

how does freebsd "work" around this, for example?

i'm sure weston runs on non linux platforms

[modern hamlet]

that is true

it does include the files unconditionally

(unless i am looking at the wrong thing)

[near tartan]

it also uses the functions in the main code

[modern hamlet]

i wasnt aware of that before you mentioned it though

[hoary cave]

me neither

but it's a weston issue

[near tartan]

it doesn't apparently

[hoary cave]

not a wayland issue

[near tartan]

there is no weston in freshports

[near tartan]

yes bro you're not the first to say this

but if EVERY single compositor uses linux specific shit

it might as well be

[marble salmon]

FreeBSD probably just has these headers

[near tartan]

eh with a bunch of fixes i got it to build

probably won't work correctly though

ehhhh fuck it

x11 is going to be easier for now

i'll put this on a different branch

[near tartan]

[radiant veldt]

Oh my God Minix

[near tartan]

rust my beloved

i have added more DRM functionality

the test app is trying to create a dumb buffer now

i also finally fixed getdents

[near tartan]

@hoary cave can u believe it
i finally locked tf in

sockets are half working

xorg is trying to do stuff

i need to read up on how x11 architecture actually works

[grave peak]

letsgo

[near tartan]

i also tried doing ps2 with evdev

im writing a test application to do input

idk if i want to have kernel vt support

delegating everything to userspace seems easier

[near tartan]

thank you rust-analyzer

[near tartan]

more drm work happened today

if i get the time I'm sure i can get something drawn on the screen today

i need to mess a bit with mapping buffers and so on, but it shouldn't be a problem

[near tartan]

we have a mode?????

huge???

[fair lintel]

yooooooooo

pic

[near tartan]

nothing drawn yet

💀

rn all it does is take over control over fbcon

will send pics when it does shit

but drm at least says it has a mode with the correct crtc

[marsh holly]

the clock is ticking lil vro

🕰️

[near tartan]

💀 💀 💀 💀

not happening

[hoary cave]

chop chop

[fair lintel]

no pressure

[hoary cave]

totally no pressure

[marsh holly]

no rush but u have 6 hours

[near tartan]

im cooked wallahi

[near tartan]

copying this table was more aids than it should have been

[magic charm]

that's got to be worth something anyway

the first run of xorg only reveals lots more stuff to work on and fix, in my experience anyway, so i expect the same for wayland

[near tartan]

true

[near tartan]

2026 year of the menix desktop

[near tartan]

we're so close

[marsh holly]

Yoooo

[hoary cave]

based

[hybrid island]

Do you not support signalling usermode yet?

[near tartan]

no

too annoying

[marsh holly]

if (!page_fault()) {
raise_signal_sync(SIGSEGV);
}

[near tartan]

[near tartan]

IT WORKS

@hoary cave

[hoary cave]

finally

now xeyes time

🔥

[near tartan]

the rendering is a bit messed up because sleep is a noop

[marsh holly]

yoooooo menix is real

[near tartan]

now it only has to be cleaned up and work with virtio gpu and so on

[marsh holly]

u should add virito gpu accel

[near tartan]

💀

not even leo managed to get it working

i'll be happy if it can modeset

[near tartan]

ok virtio gpu is surprisingly not that bad

it doesn't work tho

[marble salmon]

we have virtio gpu as well, the problem was only accel

[near tartan]

figured, i'm not doing accel for a long time

[marble salmon]

i think it should be solvable though

[near tartan]

maybe i can transfer some knowledge to managarm

[marble salmon]

might have been some ABI meme, not sure

[near tartan]

just for fun i tried just vibe coding a virtio-gpu driver and it turned out utterly useless as always

[fair lintel]

expected

[near tartan]

i'm giving up on the virtio gpu shit

for now

it does do some interesting things

but no image

[marble salmon]

You need to break that down to smaller tasks if you want to do it using AI

[hoary cave]

i already told him that at some point

[near tartan]

no

i actually don't want to do it with ai

i just did it for the lols

it did give me a good virtio impl though

[near tartan]

[grave peak]

damn this makes me wanna work on ultra again

gttnh consumed me entirely

[near tartan]

drm without actual gem is much simpler than i thought

[hoary cave]

go go go

[near tartan]

lock in

[hoary cave]

this guy gets it

[near tartan]

the ai did get modesetting working tho

i could change the resolution to 4k

shame there's no drm test application that uses PAGE_FLIP

[grave peak]

damn menix is both vibecoded and is rust

full hype mode

[near tartan]

eh

iirc there isn't anything vibe coded commited yet

[grave peak]

i might use cursor for ultra as well

its not super useful but for some stuff ig

[near tartan]

nah

i just use vscode with the agent shit

[hoary cave]

copilot is goat

it's cheap and you probably have it for free :^)

[near tartan]

i might start using ai for churn work tho

cursor is big overkill imo

Here's an idea

make the clanker write a drm example that uses ATOMIC and PAGE_FLIP

[grave peak]

yeah ai shit is great at tiny test applications or scripts etc

[fair lintel]

yeah LLMs can be helpful for unit tests

[near tartan]

personally I'd like to get os-test working at some point

it's a nice metric of accuracy

[marble salmon]

LLMs can also be good for actual features but you have to approach it differently than just "implement virtio gpu"

[near tartan]

ofc

(pretends that that's not literally what i did)

i just kept feeding it logs

[marble salmon]

You have to write something like:

• We will now implement a virtio-gpu driver. The pci bindings that we'll need are in foo.rs. Look at the virtio spec at some.url to find the virtq layout. Implement a virtq data structure for non-unified layout. Ignore MSIs and unified layout. Ignore configuration change IRQs.
• Implement the 2d resource creation request for virtio-gpu. Set up the driver to such a point that this request can be tested in qemu

etc

[near tartan]

that seems logical

i might do that tbh

[marble salmon]

That will (hopefully) cause the llm to produce something reasonable

Then you review what it did, test it, and iterate

[hoary cave]

mfw i already told u that u have to walk the llm through the process

[near tartan]

i know

[hoary cave]

it seems that the llm has a bigger context than u

[near tartan]

vro

[near tartan]

i didn't forget about it

i just yolo'd it

cuz boredom

[marsh holly]

Gen z brain vro

[near tartan]

xooked

[near tartan]

IT WORKS

@grave peak

[median forge]

absolute cinema

[marble salmon]

very nice

[hybrid island]

Very good

[marsh holly]

[grave peak]

LFG

[near tartan]

next i want proper interrupt handling

then i can add evdev support

[near tartan]

i need to go outside

[idle flower]

lgtm

[late bramble]

lmao

&selg

[mental tinsel]

&nut sell

staic true is_busy(cost strut structure selg*);

[near tartan]

ai in 2021

[near tartan]

man :( I've literally been reading managarm IRQ handling until like 6am today

i still don't completely get it 💀

well, i understand the gist of it, but not really the MSI handling

[hoary cave]

what's so different about msi handling

compared to normal irq handling

msis are regular irqs

[near tartan]

the way they're allocated

[hoary cave]

they are allocated on demand

[near tartan]

for example, the part where i got stumped:
nvme requests an MSI, but the kernel handler for that request bails out if there is no msi controller

but the controller is defined only for x86_64

does that mean nvmes don't work on riscv?

[hoary cave]

that means they won't use msis

[hoary cave]

i think platform irqs are generally preferred

but idk

[near tartan]

i would think that msis are preferred

[hoary cave]

¯_(ツ)_/¯

[near tartan]

since they're faster and unique

[hoary cave]

anyway, thats just managarm

[near tartan]

tru

wanna help me sort this out :3

[hoary cave]

it hasn't been tested very well on like

99% of non-x86 hardware

and even 99% of x86 hardware

idk i can try but im retarded

[near tartan]

two retards are better than one

[marsh holly]

interrupt is so simple you just crash on any interrupt and poll instead

hell dont even set up an idt, use triple faults to reschedule

[marble salmon]

MSIs should be preferred even on Managarm

for devices where we support both

[near tartan]

so is my assessment correct that riscv and arm don't support msis yet?

[marble salmon]

iirc i added support for rv but maybe I never pushed that

[near tartan]

okay thanks i was ripping my hair out trying to find the code 😭

i only saw imsic stuff

[marble salmon]

yeah seems that our imsic code isn't integrated into the pcie code yet

it uses MSIs only internally

[near tartan]

okay then i finally completely understand managarms irq handling

this might come in handy for future drivers

[marble salmon]

but yeah, allocation should be simple (at least if you're only allocating one MSI at a time)

[near tartan]

i wonder how i should best handle user memory page faults

do i terminate the process immediately, or do i return an error to the caller?

[marsh holly]

SIGSEGV or EFAULT/EIO/whatever caused the bad page fault in case of a syscall

[near tartan]

so just return an error?

[marsh holly]

yeah

[near tartan]

ight

[near tartan]

working on an abstraction for consoles

I'll change it so instead of every console seeing the same output, everything has its own IO stream, and the stdout for kernel messages can be configured via cmdline

[near tartan]

lol worth a try

[neon flicker]

who tf reacted 🏳️‍🌈 to a message I sent a year ago

[near tartan]

💀

[neon flicker]

smh

[near tartan]

usercopy is going well

(not)

this is super aids

do i really need to construct a stack frame to jump to???

[neon flicker]

r u saying

that before u would blindly write into any pointer

the user gave you

[near tartan]

yes

well, i had the interface for it

just not the impl

so i don't have to update the generic code

just the handling

[neon flicker]

so like, EFAULT just plain

did not exist

[near tartan]

yes

[neon flicker]

krill issue ngl

[near tartan]

idk just never had an app crash 💀

[neon flicker]

(my kernel has a different krill issue, it literally cannot access a user pointer at all, without a view of the memory)

(that is because of kernel page table isolation and also because of SMAP)

[near tartan]

i wonder what's the most efficient checked user memory access method

[neon flicker]

allow any thread not owned by the kernel to PF, and just kill the thread on fault

[near tartan]

hm

look at this 3am autism that i wrote

this is really cursed

(the page fault handler modifies rsi in place and jumps to the function again)

(((what could go wrong possibly)))

but tbh the better solution would be to just accept it and not have fancy closures

just read and write

should i copy user buffers into kernel memory before e.g. doing a write()

what if they're big

[storm bobcat]

do you use asids

if not then kpti is probably a perf nuke

[neon flicker]

no (that is TODO)

(yes my userspace perf is shit)

[storm bobcat]

its probably not that bad but its not very good either

the overhead of kpti i mean

[eternal wharf]

you could just write the actual memory copy in asm

thats what we do in managarm, then in the pf handler we just modify the saved rip to point to the fail label in the function

[rugged shard]

In my pf handler I set rip to this trampoline which is basically what the managarm people do but it lets use my C memcpy https://github.com/9xbt/bentobox/blob/rewrite/kernel%2Farch%2Fx86_64%2Fuser.asm#L54

[hybrid island]

What I did was make a table of "fallible instructions" that the exception handler will just skip over if an exception happens at that address.

So you have no overhead in your code if you don't fault.

This here lets me easily use inline assembly snippets of that form: https://github.com/badgeteam/BadgerOS-Kernel/blob/sched3/kernel/cpu/riscv/src/exception.rs#L149

Then my usercopy can do things like this: https://github.com/badgeteam/BadgerOS-Kernel/blob/sched3/kernel/cpu/riscv/src/usercopy.rs#L23

[vast lotus]

I use the facilities designed for C++ exception handling

Specifically, on an unhandled page fault, I unwind the stack until I get to a frame with a personality code that indicates it's a user copy function. This frame has an LSDA pointing to a structure that describes what kind of page faults it's expected to generate and where to return to. If it matches the fault, I restore the unwound registers and set the PC to the value given in the LSDA

[marble salmon]

i doubt that this is generally supported

i read some discussions about that on the gcc mailing list a few years ago and the consensus was that it isn't

[vast lotus]

what aspect of it isn't supported

also remember that i only use the facilities designed for c++ exception handling, i don't actually use c++ exception handling

functions that catch faults in this way are handwritten asm

[marble salmon]

do you compile with -fnon-call-exceptions?

ig if you write the fault handler in asm it's probably fine

• the fault generating code

[vast lotus]

that is only necessary if the fault can be triggered by C code, which it can't

[marble salmon]

ah i see

so you're using eh tables but you're doing all the plumbing yourself?

[vast lotus]

yes

i had the unwind support anyway for backtraces so i decided to just use it for this as well

[grave peak]

do u have that code available anywhere

[vast lotus]

not right now

[grave peak]

new os project?

[vast lotus]

yeah

[grave peak]

damn

whats the idea of it?

[marble salmon]

basically I think for C code, gcc does not promise that the eh info is complete for every possible IP value

even with -fnon-call-exceptions

but ofc that doesn't apply to hand written asm

[vast lotus]

it does if you pass -fexceptions afaict

(which is a valid flag even for C)

[marble salmon]

-fno-call-exceptions also specifically mentions that you cannot just unwind from async signals

even if you pass it

only from synchronous faults

[modern hamlet]

I do this

[modern hamlet]

one thing i borrowed from NT which helps me with this is the concept of an MDL

so basically on every read/write it pins the pages backing the buffer into a structure called MDL which contains the byte offset within a page and references to all of the pages backing that buffer

for really big buffers this is efficient because you dont need to copy from user to kernel memory, just map the same pages into the kernel address space

and the upside is that the drivers dont care about whether the pages are unmapped or not

[vast lotus]

i have extracted the relevant parts of the code https://hst.sh/vehorapate.cpp

[near tartan]

how do you avoid userspace overwriting the data in the buffer while in the kernel

[vast lotus]

why would you want to avoid that

[modern hamlet]

I don't

[hybrid island]

Race conditions

[modern hamlet]

It's undefined behavior

The program is buggy

[vast lotus]

that matters when you're interpreting the data

but read/write usually doesn't do that

[grave peak]

Wait are you making an nt like kernel

[vast lotus]

no i just like the naming scheme and the separation

[hybrid island]

But in general you can't fully avoid this unless you copy all of it upfront so...

[modern hamlet]

exactly

I don't think posix says the data must be written exactly as it was when the write call took place

[hybrid island]

The way I "deal" with it is to simply make an explicit distinction when I read something they could be user memory

[near tartan]

what about ioctls

[hybrid island]

After that you have a copy of (part of) said user memory

[modern hamlet]

In general if you are modifying buffers while you are doing a system call with those, you are asking for trouble

[vast lotus]

in fact the plan is to be ABI compatible with linux (and eventually, only once it works well with precompiled executables, add a separate native abi and start porting stuff)

well, syscall-abi-compatible, i have no plans to replicate procfs and sysfs

[grave peak]

Linux compat is based

Why did you decide to go that route

[vast lotus]

i really don't like having to debug a kernel, a libc, and port quality at the same time

[modern hamlet]

I thought this was mintia386 transpiled code for a sec

[vast lotus]

so i am only doing one at a time

[hybrid island]

I don't get why people like doing Linux compatibility so much

And when I say that I mean ABI

[modern hamlet]

It's a well established interface

[near tartan]

free programs

[hybrid island]

But userspace is still part of the OSDev journey

[vast lotus]

first debugging the kernel with precompiled libc and executables, then once that works debugging ports with glibc and a verified working kernel, then once that works debugging mlibc or a custom libc

[modern hamlet]

People have different goals in osdev

[grave peak]

For me its not a fun part, and I dont enjoy doing it

Porting random crap and building it for 10 hours

[near tartan]

eh

[modern hamlet]

I don't enjoy abiding by specifications (which is why it took a significant amount of pain to write my fork implementation, my partial munmap and overridable mmap, and soon my execve implementation)

[modern hamlet]

You can roll your own userspace

[marsh holly]

Step 1) git clone https://github.com/managarm/bootstrap-managarm

[vast lotus]

if your kernel is already capable of providing the sysdeps needed by mlibc or similar, most of the work necessary to provide linux syscall abi compat is already done

[modern hamlet]

Kind of... but mlibc doesn't strictly require a unix-y file system structure

And by unix-y file system structure I mean devfs and procfs

[vast lotus]

oh yeah full linux abi compat including procfs/sysfs is a very different beast

but most programs either don't need procfs/sysfs or can make do with some manually created files/symlinks in /proc and /sys

[near tartan]

I don't like depending on other projects

[modern hamlet]

When I implement procfs and sysfs I'll do those with a separate driver

I ain't "contaminating" my kernel with that "crap"

[near tartan]

how can you do sysfs as a driver

isn't the point that drivers can expose their stuff via sysfs

unless you have module dependencies or something

[modern hamlet]

The beauty of doing your own OS is that you can ~lie~ and ~pretend~

[marble salmon]

sysfs is separate from drivers on Managarm as well

[near tartan]

yes but on managarm there's a way to communicate between drivers

if you don't have that it's not possible

[vast lotus]

one possible approach is to just only have the generic stuff in sysfs

this would effectively make it appear to programs as if all devices in the system use unknown drivers

[modern hamlet]

Wait what does sysfs have to do with drivers