Zinnia

[near tartan]

because just casting a slice from user memory is a bad idea obviously

[brisk totem]

if you agree that you didnt hear it from me i do know how to make a DST with whatever pair pointer-sized pair of numbers you want stored in a reference to it

[near tartan]

nah i think then it's just easier to write functions in IovecIterator

[brisk totem]

yeah

i think linux calls this an iov_iter

[near tartan]

it's very practical because you don't need to care about whether or not a buffer is in userspace or not

[brisk totem]

just make it so that the buffer is always in userspace /hj

[storm bobcat]

That's just operator overloading isn't it

[brisk totem]

no

you can overload all the operators to make it kinda work

[marble salmon]

what's even the issue

[brisk totem]

although i think until we get place projections it cant quite work properly

[near tartan]

nothing

[marsh holly]

You also kind of need it for proper writev/readv/readmsg/sendmsg support

[marble salmon]

imo overloading stuff based on this property is a bad idea

[near tartan]

hm?

[brisk totem]

why?

sometimes you want to read from files in kernelmode

i mean i wouldnt but i can see why some people would think this is a good idea

its probably handy for like sockets or something?

[marble salmon]

then just read into the page cache

[brisk totem]

except there you could just make a kernel api...

[brisk totem]

well no you want to move from the page cache to a normalish buffer

[marble salmon]

pointers to userspace also need to remain in the same address space

there are just too many differences

you can't take the physical address of them

without locking memory

in particular, differences that imply that you need to handle them differently

at the site where they are used

[near tartan]

are you saying i should have different APIs for kernel vs user accesses in e.g. the VFS code?

[marble salmon]

yes

[brisk totem]

i would just not have kernel accesses

[near tartan]

i think i use it to bootstrap

i.e. File::read and ::mmap for init

[brisk totem]

to bootstrap what?

[somber solar]

init

[near tartan]

ok so wait

[brisk totem]

btw

you know that rust has a never type right

[near tartan]

yeah

[brisk totem]

you dont have to use unreachable!()

[near tartan]

where

[brisk totem]

kerne/src/syscall/process.rs

[somber solar]

is that noreturn

[brisk totem]

no

its a bottom type

you can have it in a union

and then that case of the union is unconstructible

and you can match on it to construct whatever value you want

[somber solar]

monostate but better?

[brisk totem]

monostate but it behaves properly in every case 🙃

wait no its not monostate

its an unconstructible type

you cannot make a value of type !

[near tartan]

ok but what's the change here?

-> Result<!, Errno>

?

[brisk totem]

ye

[near tartan]

👍

btw thanks for the reviews

i could use all of them

even if they're nitpicks

[somber solar]

pitust could you take a look at my shitty vmm when I finish it?
you seem to know stuff about it

[near tartan]

i would like to get interrupts working but idfk what I'm doing

especially msis with nvme

[marble salmon]

iirc that's not stablized, is it?

i.e., in arbitrary places instead of just function returns

[brisk totem]

yeah but nightly is better anyway

[near tartan]

i have to use nightly regardless

[marble salmon]

you don't have to use nightly

[near tartan]

you do

if you want to write a kernel

[marble salmon]

why?

[brisk totem]

you need a custom target spec

[near tartan]

and build-std is unstable

[brisk totem]

yeah

[marble salmon]

just patch rustc

[brisk totem]

you need nightly to build the patched rustc

[near tartan]

[marble salmon]

amazing

why can't you just use no_std?

and a freestanding target?

surely there must be a way to do bare metal apps on stable?

[near tartan]

different abi

[brisk totem]

different code model

and abi

and no prebuilt binaries

[near tartan]

and in my case i also have dynamic linking enabled

[marble salmon]

i know ppl who work with rust on bare metal in regulated industries and i'm 90% sure that they cannot use nightly

[near tartan]

maybe

ask them

[marble salmon]

even we use stable only at work

(but we don't do bare metal rust)

[near tartan]

yeah that makes sense

maybe they use nightly cargo and stable rust

afaik build-std is done on the cargo level

[brisk totem]

ok then just RUSTC_BOOTSTRAP=1

thats TOTALLY different right

[brisk totem]

well the stdlib is built with effectively nightly

[marble salmon]

well the difference is that the stable compiler is actually, you know, stable

[brisk totem]

not with RUSTC_BOOTSTRAP=1 it isnt

[marble salmon]

well, it's built from a stable branch etc

it may have nightly features enabled during bootstrap but the point is that it went through way more QA than a random nightly build

with random nightly features enabled

[magic charm]

i increasingly feel freebsd is doing the right thing

by autogenerating the de/marshalling code from SAL annotations

[hybrid island]

Autogenerated marshalling is 100% the way

Now you can't forget to enforce certain preconditions, especially if you have a type system that can more rigidly enforce it like C++ and Rust

[near tartan]

@brisk totem i added smap

everything still works so far

[brisk totem]

nice!

[near tartan]

i cherry picked stuff over from the socket branch

[grave peak]

damn vgpu

[near tartan]

yeah

it works

with drm

[grave peak]

even 3d?

[near tartan]

lol no

dumb buffers

[grave peak]

ah

[near tartan]

you forget that i'm not goated

[grave peak]

tfym

3d soon

[near tartan]

UserSlice

[grave peak]

u can always verify user crap later

[near tartan]

i really want to get sockets working with a cool rust abstraction

[grave peak]

3d is where its at

[near tartan]

but i can't think of anything

i have this shit rn

console=/dev/fbcon seems to work

[grave peak]

is this a bash port or just fastfetch

[near tartan]

just fastfetch

bash works but there's no input yet

[grave peak]

ah

[near tartan]

lol#

i totally forgot zinnia has a working nvme driver

i did console=/dev/nvme0n1

[grave peak]

what does console= do?

[near tartan]

opens a file for stdin/out/err for init

[grave peak]

oh

[brisk totem]

you should def make sure that your API for making them is secure by construction, i.e. that you dont have to call a validation function in every usage site to make sure you dont accidentally copyout to userspace

or well i guess its not as big of an issue with smap properly enforced

[near tartan]

wdym

currently the only "safe" way of accessing user memory is through UserPtr

which handles SMAP and page faults

[brisk totem]

yeah good

[near tartan]

are there any important cases where having the same API for kernel and user memory is useful? (cc @marsh holly since you have code for iovec_iterator_t in the VFS)

[storm bobcat]

what is iovec

[near tartan]

#include <sys/uio.h>

[hoary cave]

scatter gather

[brisk totem]

only vaguely

like in socket code MAYBE

and in elf loading code

[near tartan]

yeah currently i only have it in elf loading code really

[storm bobcat]

just realized

there are a bunch of names here that start with Z

nvm no there arent

we're the only two

[near tartan]

[brisk totem]

but for the elf code you can just load a flat binary

and then have it be the real elf loader

aka the vdso loader route

[storm bobcat]

do your elf loading in userspace

like a man

[near tartan]

no

i am a pussy

[hoary cave]

kityt

[hoary cave]

[brisk totem]

get some help then

[near tartan]

i have rebased most of the patches from the socket branch onto main

(minus the big vfs changes)

this afternoon i want to actually get some shit done

let's see...

• switch to iovecs

• rebase the vfs changes

• port rust std

std will be the most annoying part i think

[storm bobcat]

now that you switched back to rust I dont fw this project anymore

😢

[near tartan]

it's over

[lofty copper]

Nah

[near tartan]

how so

[lofty copper]

Or like it's annoying, but not too difficult

[near tartan]

sure

i need to do the mlibc definitions too

sadly I can't use a generator so ig i do it by hand

i could make an abi checker maybe

[lofty copper]

Managarm is mainline

Just steal from there

[near tartan]

but not mlibc

managarm isn't unix

managarm is linux_like

[lofty copper]

Huh

[near tartan]

yeah

[lofty copper]

Then idk

There's a bunch of other unix-like operating systems

[near tartan]

maybe haiku

[lofty copper]

Also, most of it is abstracted away anyway?

[lofty copper]

Hurd

Redox

[near tartan]

the problem for me is libc

i have the base std done

i just need libc

but all structs and functions in there are random

[somber solar]

I have lib::maybe_uspan
incredible name, I know

it can either hold kernel or user pointer

I use it for my vfs methods

of course every read and write is checked

[near tartan]

i wonder if it's enough to make the iovec iterator thingy in a way that can't be Copy, Clone, Sync, Send

that would mean the only way that struct can be safely constructed is if it's pointing to an address in the active address space

[hoary cave]

Alright

[near tartan]

because you can't send it to another thread, everything stays contained

ofc you can construct raw iovecs from thin air at all times

they just don't carry any meaning since you can't do I/O with them safely

[hoary cave]

so what is the use for an iovec if you can't do io on it

that's the only use for it

[near tartan]

you use it with the iovec iterator

you pass a list of iovecs to this iterator struct, which allows you to do reads and writes

the raw iovec is meant as something like an FFI

and the iterator is a safe wrapper around it

[hoary cave]

okay

[near tartan]

@hoary cave i got further with iovec

read and write now uses it

[late bramble]

Womenix

[near tartan]

[near tartan]

i should figure out serial interrupts

this is like the most basic ass form of input

i need to see if the coreutils do their shit right

[somber solar]

Womenix

[neon flicker]

Womenix

[near tartan]

[modern hamlet]

pretty sure they're using the same "legacy IRQ" system as ps/2

meaning you need to setup IOAPIC and stuff

unless you have a PCI serial device

[near tartan]

i dont have that either

do i need to parse acpi tables for ioapic configuration

[modern hamlet]

yes im pretty sure you do

luckily you dont need to know how to do aml

[near tartan]

like what for example?

do i need to configure the pin polarities?

[modern hamlet]

4 different types of headers

[near tartan]

wut

[somber solar]

it's not that hard

[modern hamlet]

https://github.com/iProgramMC/Boron/blob/master/drivers/halx86/source/ioapic.c heres my implementation that i stole from somebody else

[near tartan]

yeah but idk where to look

[modern hamlet]

i believe it was astral but not sure

[near tartan]

i have discovered the ioapic and i have code to configure it

ig im just missing the setup

[somber solar]

https://github.com/ilobilix/kernel/blob/master/kernel/source/arch/x86_64/system/ioapic.cpp here's my implementation that I probably stole from somebody else but I don't remember

[near tartan]

I've got an idea

I'm going to make Zinnia BSD-like

it already has a BSD license and some BSD ioctls

most of the work will be getting BSD functions into mlibc

[marsh holly]

Freebsd abi compat

[near tartan]

nah

i could, but i explicitly want to help with mlibc

because i believe it's a project that benefits way more people than a shitty hobby kernel

[median forge]

port bsd libc

[near tartan]

no

[near tartan]

@somber solar

what is this

[somber solar]

where

oh

pic thingy

that connects the two pics

I don't remember the name

[near tartan]

ohhhh

yea i masked the pic completely

is that bad?

[somber solar]

nope

you should do that if you use ioapic

[near tartan]

k

[grave peak]

linux only unconditionally masks if doing x2apic + iommu

[near tartan]

wait so

[grave peak]

otherwise it masks per pin

when enabling a specific ioapic interrupt

[near tartan]

what am i even supposed to do here

your code reads like there are no GSI overrides if PIC is disabled

[somber solar]

I think you need to unmask irq 2 in pic even with ioapic

[grave peak]

if u want to be able to use pic as a fallback

[near tartan]

i dont

[somber solar]

I think isos are for just that

[near tartan]

oh

hm

well like i said i just want to get com1 interrupts

idfk how

[somber solar]

I have code for that too

wait a sec

https://github.com/ilobilix/kernel/blob/master/kernel/source/arch/x86_64/drivers/output/uart8250.cpp

[spark surge]

what base revision do you use?

[near tartan]

why

3

[spark surge]

because on base revision 5+ Limine will mask the ExtINT LVT in the local APIC(s) which means that masked or not masked, you're not gonna get PIC interrupts

in any case it's fully masked, in any base revision

[spark surge]

like if you wanted to get PIC interrupts with base rev 5+ you'd have to unmask the ExtINT

masking ExtINT is the "proper" way to disable the PIC

per se

[grave peak]

damn not using aml to detect serial

😭

[somber solar]

:P

how do I do that?

[spark surge]

imagine not assuming that ISA peripherals just exist

[near tartan]

i need to write a uacpi driver wrapper

i dont like the way uacpi does discovery

but that's bound to C ig

@somber solar irq_swtich

[spark surge]

the wiki says that masking all PIC IRQs = "disabled" but masking ExtINT is more of a real disabling than that

[somber solar]

yeah I'm good at naming how could you tell

[spark surge]

you should do both anyways

[near tartan]

and good at spelling

[somber solar]

indeed

[near tartan]

@grave peak have you considered removing anon structs and unions from uacpi headers

:3

[grave peak]

Which

[spark surge]

this

[near tartan]

it's not an issue at all but it makes codegen a bit ugly on langs that don't have it

[spark surge]

using enums in C is stupid

[near tartan]

example

[spark surge]

even more so on public interfaces

[grave peak]

Who was talking about enums lol

[spark surge]

oh fuck

i hallucinated, sorry

still, it applies i think :p

[near tartan]

[spark surge]

real

[grave peak]

Nah enums aren't going away

[spark surge]

:(

[grave peak]

For anon structs ill think about it

[near tartan]

i'm only talking about anonymous types

[spark surge]

why do you like enums

in C

real q

[near tartan]

what's the problem with enums?

[grave peak]

Just a clearer intent when reading code, instead of int thing

[near tartan]

especially in c23 they're good

[grave peak]

Yep

[spark surge]

yeah in C23

but like

[near tartan]

nobody here isn't using that

[spark surge]

what person making libraries is targeting C23

especially source libraries

[near tartan]

i've never had issues with enums?

[somber solar]

@grave peak I couldn't find any info about it iirc

[near tartan]

even before c23

[grave peak]

Cant you like make a custom helper or just tell zig the struct contains one item that you want instead of using raw bindings?

[spark surge]

you can use enums, just not in public interfaces, that's my take, but whatever, irrelevant here

[near tartan]

no

[grave peak]

Why not?

[spark surge]

wait, zig?

[near tartan]

rust

nobody writing c89

[grave peak]

Oh thats rust, even worse

[spark surge]

not C89 but usually something like C99 or C11

[near tartan]

both have enums?

what's the problem with it lol

[grave peak]

@near tartan I could add like a no_anon_struct define, would that help?

[near tartan]

how would that look

[grave peak]

Ifdef no_anon_struct { field } else union { field1, field2, field3 }

Something like that

[spark surge]

it doesn't matter lol, it's a matter of compatibility and type

[near tartan]

just give all unnamed unions a name?

that would cover all cases

[spark surge]

you can give types to defines but not to enums

enums are always int

[grave peak]

That breaks all C code tho

[spark surge]

(not in C23 but whatever)

[grave peak]

Does it not?

[near tartan]

no?

[grave peak]

Cuz then u have to have a name for the member, which means u have to add a tbl.thing.ptr

[near tartan]

wut

no?

[spark surge]

i think infy is right

[grave peak]

Yep

Adding a name means u can no longer access members directly in C

[spark surge]

yeah

[grave peak]

Which makes all C code as ugly as your rust code

[spark surge]

yeah, i'd be against making unions/structs named

[grave peak]

Barring that, it just breaks literally all users

So that's a bit too much

[near tartan]

oh

god fucking damn it

i hate C

[spark surge]

removing enums won't :p

[grave peak]

But putting that behind an ifdef is fine perhaps

[spark surge]

i hate C too <3

[near tartan]

no that's useless

[grave peak]

Why lol, u can just pass it to rust

And you will have no ugly names

[near tartan]

no?

[grave peak]

No what

[near tartan]

you have the same issue then lol

[grave peak]

?

[spark surge]

@grave peak don't tell me there isn't some instance where you wish you could have typed enum members

[somber solar]

sleep marvin

[grave peak]

You will have tbl.ptr

In rust

[spark surge]

barring C23 of course

[grave peak]

I cant in uacpi even if I wanted

It supports openshitcom etc

[somber solar]

lol

[spark surge]

tf

oh

watcom

[near tartan]

oh wait lol

tfw

[grave peak]

marving trolling today

[near tartan]

im mad okay

[spark surge]

what does that have to do? what i mean is that using #defines over enums means you can give them types beyond int

[near tartan]

i'm pissed off that i can't figure out how to properly setup the ioapic

[spark surge]

not to mention that having functions take enum parameters or return enums is kind of a meme in C as well

[grave peak]

whats the point exactly?

[spark surge]

if you need them to be typed

[grave peak]

or u could just do enum : type

[spark surge]

that is C23 only

or C++

so that's a no-go

if you care about compat

[near tartan]

i should make myself a uacpi wrapper

[storm bobcat]

write your own aml interpreter

[somber solar]

uacpi C++ modules support when

[grave peak]

Device (UAR1)
{
    Name (_HID, EisaId ("PNP0500") /* Standard PC COM Serial Port */)  // _HID: Hardware ID
    Name (_UID, One)  // _UID: Unique ID
    Name (_DDN, "COM1")
    Method (_STA, 0, NotSerialized)  // _STA: Status
    {
        If ((FUIO (Zero) != 0x0F))
        {
            Return (0x0F)
        }

        Return (Zero)
    }

    Method (_CRS, 0, Serialized)  // _CRS: Current Resource Settings
    {
        Name (BUF0, ResourceTemplate ()
        {
            IO (Decode16,
                0x02E8, 
                0x02E8,
                0x01,
                0x08,
                _Y2C)
            IRQNoFlags (_Y2D)
                {3}
        })
        CreateByteField (BUF0, \_SB.PCI0.UAR1._CRS._Y2C._MIN, IOLO)  
        CreateByteField (BUF0, 0x03, IOHI)
        CreateByteField (BUF0, \_SB.PCI0.UAR1._CRS._Y2C._MAX, IORL) 
        CreateByteField (BUF0, 0x05, IORH)
        CreateWordField (BUF0, \_SB.PCI0.UAR1._CRS._Y2D._INT, IRQL)
        Local0 = FUIO (Zero)
        Switch (ToInteger (Local0))
        {
            Case (Zero)
            {
                IOLO = 0xE8
                IOHI = 0x02
                IORL = 0xE8
                IORH = 0x02
            }
            Case (One)
            {
                IOLO = 0xF8
                IOHI = 0x02
                        IORL = 0xF8
                        IORH = 0x02
                    }
                    Case (0x02)
                    {
                        IOLO = 0xE8
                        IOHI = 0x03
                        IORL = 0xE8
                        IORH = 0x03
                    }
                    Case (0x03)
                    {
                        IOLO = 0xF8
                        IOHI = 0x03
                        IORL = 0xF8
                        IORH = 0x03
                    }

                }

                IRQL = (One << (FRUI (Zero) & 0x0F))
                Return (BUF0)
            }
        }

Some examples from real hw

[somber solar]

I don't speak aml lol

PNP0500

[grave peak]

PNP0500 + uacpi_get_current_resources

[storm bobcat]

huh kinda funny they have an IRQL field

[grave peak]

so does that help strip a prefix completely?

[near tartan]

yeah

oh wait

no

[storm bobcat]

what does the switch case do? is FUIO the thing you pass in to find a port

i mean

is Zero the first parameter or something

[grave peak]

it is yes

Method (FUIO, 1, Serialized)
{
    If ((IER0 == One))
    {
        If ((WUR0 == Arg0))
        {
            Return (Zero)
        }
    }

    If ((IER1 == One))
    {
        If ((WUR1 == Arg0))
        {
            Return (One)
        }
    }

    If ((IER2 == One))
    {
        If ((WUR2 == Arg0))
        {
            Return (0x02)
        }
    }

    If ((IER3 == One))
    {
        If ((WUR3 == Arg0))
        {
            Return (0x03)
        }
    }

    Return (0x0F)
}

its some internal firmware code

who knows

[storm bobcat]

ah ok so you call that function to get the port

[grave peak]

only stuff prefixed with _ is public API

[storm bobcat]

bruh

[near tartan]

duh

[somber solar]

I found something about spcr and dbg2 tables as well

[grave peak]

@somber solar serial devices in aml also often times have possible_resources() and set_resources() so u can choose any IRQ u want and any IO port you want

out of the supported values

[somber solar]

nice

well I know what I'm doing next

[grave peak]

yeah thats for early serial before aml

not all hw has it but yeah

[somber solar]

tlb shootdown can wait

[storm bobcat]

so what device can you actually assume is present for early logging lol

[somber solar]

what if I want to log before early table access :P

[storm bobcat]

I chose serial because at least on real hardware it might actually exist vs debugcon

[grave peak]

dbg2/spcr tells you

in hypervisors u can test for 0xE9

[grave peak]

well

[somber solar]

can you just write to com1 and hope it exists?

[grave peak]

u can try that

[storm bobcat]

implement a log buffer

and log when u find a port

[grave peak]

^

[storm bobcat]

in fact infy is implementing uLog as we speak

[somber solar]

hmmm

[spark surge]

yes

[somber solar]

really?

[grave peak]

straight up lies

[storm bobcat]

[somber solar]

lol

[spark surge]

well, sort of, rather, you can check if it exists

[grave peak]

if anything he should make one instead

[spark surge]

without AML

[storm bobcat]

who, me?

[grave peak]

ofc

[storm bobcat]

just convert my zig back to C++

[grave peak]

your ring supports 32 bit

[storm bobcat]

[somber solar]

okie

gonna get inspired

[storm bobcat]

that relies on zig shenanigans though

[grave peak]

zig to C bindings

[spark surge]

ring?

[storm bobcat]

since I can do u<arbitrary> so I do u<usize bits - 2>

[somber solar]

even more pnps PNP0500 PNP0501 PNP0510 PNP0511

[grave peak]

yeah theres a few

[somber solar]

if only you could have that many ways of detecting pit

[spark surge]

here's the thing: for ISA devices you can assume that they are either where you'd expect them to be, or they are not, but there's not going to be something else in their place

that's just not a thing that happens

[storm bobcat]

yeah thats why I opted for com over debugcon

[spark surge]

so you can test if they are there with readbacks/init and whatnot

[grave peak]

yeah they're scared af to break nt boot

[somber solar]

I do both com and e9

[storm bobcat]

well I suppose I could setup detection

afaik e9 is easy to do

[somber solar]

you just read from it

[storm bobcat]

yea

[grave peak]

e9 detection is in(e9) == e9 lol

[storm bobcat]

well

in_hypervisor && in(e9) == e9

[grave peak]

ye

[spark surge]

tbf i don't think there is much of an advantage to using e9 once you have com anyways

[somber solar]

omw to implement e9 on real hardware

[grave peak]

tbh its actually harder
if ((bochs_bios_in_smbios || in_hypervisor) && in(e9) == e9)

[spark surge]

yeah

just checking the e9 readback is enough

imo

[somber solar]

I mean when would you ever read back 0xe9 from port 0xe9 if it wasn't present

[spark surge]

never

[grave peak]

this is for detection

[spark surge]

if a port isn't connected to anything you usually read back ff

on real hw (and probably emulators/qemu)

[grave peak]

yeah but if there's something else at that port you may cause side effects

[spark surge]

unlikely

[grave peak]

still possible

but yes

its not like a standard isa port or anything

[spark surge]

that's why you don't use e9 if you have com :p

[somber solar]

can one detect absence of pit?

[grave peak]

no

well only by probing it

[spark surge]

yeah by probing, but like

that's not enough

for example ARL has a PIT but it's clock gated by default

most probing mechanisms would tell you there is a PIT there

[grave peak]

what did u mean btw

[spark surge]

but if you wait for it to tick, it never will

because it's clock gated

[near tartan]

it fails to generate if i leave it blank

[grave peak]

i mean who needs it for anything other than calibration

[spark surge]

so yeah, no, heuristics are a better approach

[grave peak]

oh ok

so would ifdef help you then?

[spark surge]

e.g. you can assume there is a PIT if a machine is old enough to not have e.g. invariant TSC

[somber solar]

I should mask pit if something else is used

[modern hamlet]

does that matter

[spark surge]

like the IRQ? yes. ideally you should have all IRQs/GSIs masked unless you need them lol

[grave peak]

bragging rights

[somber solar]

lol

[near tartan]

no

just ignore it

[grave peak]

bruh

cant u pass c flags from rust?

also theres proper bindings being worked on anyway so

[spark surge]

this is the case btw if you use Limine

[near tartan]

yes but it doesn't help me bro

[spark surge]

the only thing that is not masked as per spec are MSIs which are left as firmware set them up, but that is usually all masked as well

[grave peak]

how lol

[near tartan]

i still have to type the whole path

it's not like C where it skips anon structs

or not

[grave peak]

not if i get rid of the union entirely

if that define is passed

[somber solar]

better a proper name instead of anon_1

[near tartan]

but i want the union?

i dont get it

[grave peak]

oh ok if u want it then yes

[spark surge]

it would waste memory

but it would work

[grave peak]

i was thiking u have to cast it from rust anyway

could just be a void pointer

[near tartan]

not talking about the cast

the field name

[grave peak]

if u have to do the cast anyway u can just have one single field and no need for a union

the entire idea of a union is to minimize the number of casts required

[near tartan]

it's not so deep

[grave peak]

😭

[spark surge]

ah i thought you wanted to just not use a union if you passed that define

[somber solar]

okay so I think I'll:

1. dump everything to com1 until table access
2. check for spcr/dbg2 when tables are available
3. evaluate pnps once acpi is initalised

[spark surge]

like, and have multiple fields instead

[near tartan]

LIMINE_PTR(x)

[grave peak]

thats what i thought

it can have multiple fields if marvin really wanted to

[near tartan]

i dont care anymore :3

[spark surge]

like as i said, it's a bit more wasteful, but it would also be similar enough™

but a void ptr or something would also work ig

[grave peak]

marvin probably regrets bringing it up

[near tartan]

yeah

[somber solar]

lol

[near tartan]

this isn't useful

[grave peak]

at least with that option u can have a readable prefix

in your builder thing i mean

rust doesnt have anon unions at all right

[somber solar]

c just can't stop winning

[near tartan]

es

[grave peak]

real

[storm bobcat]

why is smbios even useful

[grave peak]

why is it useful to know what system you're running on?

[near tartan]

i wrote a tool once that can register itself in a netbox setup by fetching all dmi info and system info and uploading it

[storm bobcat]

yes unironically

why does it matter what the bios vendor is

[storm bobcat]

the actual proper solution is to not do that btw

like just have a log buffer

[grave peak]

It contains a ton more than just the bios vendor

[storm bobcat]

do you find it via uefi

is that code upstream

[grave peak]

Same as rsdp

[grave peak]

Which

[storm bobcat]

your smbios code

[grave peak]

In the pr that's open

[storm bobcat]

why do you do prs

on your own

[grave peak]

CI

Limine has a way to request smbios tables I think

[storm bobcat]

do you do it with hyper?

[grave peak]

Yes

[storm bobcat]

im trying to depend as less as possible on the bootloader but that kinda stuff is optional anyway so I think I will depend on it for that

[near tartan]

for what?

[grave peak]

Well in bios mode u can find smbios yourself

[near tartan]

smbios?

[grave peak]

In UEFI mode its in the configuration table same as rsdp

So it requires boot services

[storm bobcat]

yea

[storm bobcat]

yes

what is smbios useful for though

what kinda stuff can I do

[grave peak]

https://elixir.bootlin.com/linux/v6.19.10/A/ident/dmi_system_id

this is the main use case

[near tartan]

it can tell you about the chassis type, installed memory modules

stuff like that

[grave peak]

just run sudo dmidecode to get an idea

[near tartan]

it can tell you stuff about a specific dimm

[storm bobcat]

i will crash if the manufacturer is not crucial

[somber solar]

why not both

[near tartan]

should the kernel have a generic way of allocating msis?

considering drivers using msis are also generic

[modern hamlet]

Yes

[near tartan]

But not for platform IRQs right?

or that too

[modern hamlet]

On x86_64 I think you have all the flexibility you can get

You can even remap legacy IRQs onto any vector

So IRQ 0 you could map to vector 0xAF or whatever for example

[near tartan]

and then what

[modern hamlet]

?

[near tartan]

what's the difference to other archs

[modern hamlet]

other archs might not have nearly as much flexibility

e.g. x86 without apic has fixed legacy IRQ numbers which you cant change

[near tartan]

makes sense

[near tartan]

so to answer this question

it's no

right?

[modern hamlet]

so i guess generic to drivers? yes

generic to platforms? not really

[modern hamlet]

platform irqs like what

[near tartan]

idk im stupid

irqs that aren't msis

[modern hamlet]

well yeah

also you'll want such a feature if you implement irq priority levels using the TPR/cr8

[fiery marlin]

are you only planning to support x86?

[modern hamlet]

since those block interrupts depending on the interrupt vector number

[near tartan]

no

[modern hamlet]

e.g. a TPR of 0x8 blocks interrupts with numbers 0x00-0x8F

[fiery marlin]

ok, so far what I'm thinking is that a generic api for allocating interrupt lines (of any kind, msi or something else) is useful for drivers, but the implementation really depends on the platform and interrupt controller configuration.

[near tartan]

yeah figured that much

[fiery marlin]

there's a paper that fadanoid(?) linked recently that I thought was really interesting for this, I dont know if you read it

#1385970208631427173 message this one

[near tartan]

should be doable

i can make allocating int lines a archdep

[marble salmon]

for non MSI, the routing is usually fixed or almost fixed though

I think anybody who wants to design a proper IRQ abstraction needs to look at multiple architectures and IRQ controllers since they are all different

You cannot look at x86 and extrapolate from there

[fiery marlin]

yeah

[marble salmon]

as there is too much x86 specific stuff

[fiery marlin]

I have made that mistake in the past

[fiery marlin]

true, I probably should have said acquiring instead of allocating there.

[marble salmon]

All of these assumptions are wrong:

• IRQ routing is fixed / constrained and not fully dynamic
• IRQ routing is fully dynamic
• The CPU has a concept of an IRQ vector number
• The device has a concept of an IRQ number
• The interrupt number that the device uses matches the one on the CPU side
• There is only one IRQ controller between the device and the CPU
• MSIs are sent directly to the CPU
• You need to send EOI before handling the IRQ
• You need to send EOI after handling the IRQ
• IRQs are either edge or level
• The kernel knows if IRQs are edge or level
• All IRQs have the same priority

[somber solar]

unsigned _BitInt(bits)

[storm bobcat]

You can just use masks

In zig they're useful because you don't have to use masks

But Linux does

[near tartan]

finally have working serial with interrupts

somehow this broke lol

yay

[near tartan]

i'm rewriting zinnia in C++ and i'm doing linux abi compat

[somber solar]

[storm bobcat]

For real?

Ah fuck April fools

[near tartan]

:)

[storm bobcat]

Unironically do that tho

[near tartan]

in another life

that's just fishix tbh

[grave peak]

that would be the good choice tbh

[grave peak]

nah

fishix is a speedrun

[somber solar]

ilobilix :P

[near tartan]

see i want to do bsd source compat

[grave peak]

its not focused on quality etc

[near tartan]

i think it's cool

now that i can actually interact with my kernel i think i can test it much better

i need to add tty discipline

[storm bobcat]

Which

[near tartan]

a bit of everything

maybe netbsd

[storm bobcat]

why

do freebsd

[grave peak]

yeah out of all bsds

[near tartan]

what's the diff

[grave peak]

at least pick the popular one

[storm bobcat]

as much as I love netbsd

just do freebsd

[near tartan]

ok

[storm bobcat]

well nvmm is saner tho

So I'd do freebsd + nvmm

[near tartan]

yeah

[somber solar]

what's the difference between source and abi compat

[grave peak]

what is nvmm

[near tartan]

kvm but good

[grave peak]

source means u have to recompile

abi means u can cp native programs

[storm bobcat]

The good choice imo is doing like I'm doing which is building Linux compat as a separate subsystem that's built on my own design

[somber solar]

ah

[storm bobcat]

virtualization API

[near tartan]

i want to do bsd + mlibc

[storm bobcat]

It's really nice and simple

And qemu supports it

[grave peak]

thats freebsd with extra steps

[storm bobcat]

unlike freebsd jails

[storm bobcat]

Well yes but not cloning freebsd

[near tartan]

the thing is, because i'm not claiming to be literal freebsd, i can do whatever i want and pick my favorite APIs from everywhere

[storm bobcat]

I'm not cloning anything in particular

[grave peak]

buzz lightyear meme

[near tartan]

nah

im gonna remake that gif with linux abi compat

[grave peak]

lol

[storm bobcat]

freebsd has Linux Abi compat

Buzz Lightyear meme

[near tartan]

lol

[grave peak]

linux abi compat used to be a novel concept

[somber solar]

fishix, ultra, mine

[grave peak]

but everyone stole it from me

[somber solar]

who else does it?

[grave peak]

😭

[storm bobcat]

I did it before you did

!

Like two years ago

[grave peak]

damn

[somber solar]

I hadn't even heard of your kernel when I started mine

[grave peak]

probably because it didnt exist

[somber solar]

this rewrite is a bit more than a year old

[grave peak]

has it always been linux compat?

[somber solar]

yes

but this is the only "successful" one

[grave peak]

ah ok

[somber solar]

I was using mlibc before this rewrite

[near tartan]

i'm doing a mix of everything
drm + evdev + nvmm + posix_devctl + epoll + bsd style device management

[somber solar]

and pe instead of elf

[grave peak]

is nvmm better than kvm?

also is that different from bhyve

[near tartan]

arguably

[storm bobcat]

It's definitely simpler

[near tartan]

streaks implemented it in ironclad and says it's also better

or mint idr

[storm bobcat]

bhyve doesn't support qemu

[storm bobcat]

Most of the logic is in userspace, the actual hypervisor code in netbsd is like 2k lines and really readable

[grave peak]

is bhyve like part of that or

[storm bobcat]

No bhyve is freebsd's own thing

arguably they should've just adopted nvmm

[grave peak]

ohh nvmm is netbsd

[near tartan]

yea

[grave peak]

and bhyve is freebsd i guess

[storm bobcat]

Yes, but now the development is done in dflybsd because the author got mad

[grave peak]

if all the logic is in userspace that probably means a lot of vmexits become userspace vmexits which are incredibly slow btw

[somber solar]

average open source project

[grave peak]

sure its simpler but its not something ud want for production

[near tartan]

another funny thing about being a bsd is devd with a non aids API

(netlink)

[storm bobcat]

Not all of it, I'm not sure how it works

But like it is pretty good from what I've heard

[grave peak]

is it supported by qemu also?

[grave peak]

its kinda funny they keep splitting into more forks

[near tartan]

mitosis

[somber solar]

lol

[near tartan]

opened a pr to mlibc to drop menix

(and rename to zinnia)

[grave peak]

[storm bobcat]

dflybsd is pretty old

[storm bobcat]

nvmm is, bhyve is not

[hoary cave]

i thought you said diddybsd

im cooked

[storm bobcat]

diddybsdm

[drowsy tendon]

diddybdsm

[near tartan]

@marble salmon btw the solution was to turn on interrupts in syscalls

it's the reason the serial interrupt didn't arrive in the kernel

because it did fire in qemu

my ioapic code was mostly fine already

[somber solar]

do you keep track of interrupt state with a counter like preemption?

[near tartan]

yes

i have an IrqLock

but it assumes that a level of 0 == interrupts are enabled

[somber solar]

mine is also broken. it assumes that the last thing unlocked was the first locked

gotta fix that

[near tartan]

lol

well in my case the mask for syscall doesn't have 0x200

so i have to manually enable interrupts in the handler

[somber solar]

doesn't that go in some mask msr

idr

[near tartan]

yes

SFMASK

[near tartan]

anyways, now that i can get interrupts routed and send input to user programs it's time to tackle ttys and ptys

[near tartan]

finally fixed mounting

i still need to do the tty shit

i want to fix some things that i had stubbed out for way too long

• Signals
• Process groups
• devctl instead of ioctl
• ext2
• fix mount and chroot

speaking of, how does one go from initramfs to rootfs? i know linux has pivot_root, but how is this usually implemented?

my idea would've been to just mount the rootfs and chroot into it from init

but idk if dinit can cope with that

maybe I'll just switch back to openrc

[median forge]

just mount some partition to / and exec to /bin/init for example

[near tartan]

oh true

[marble salmon]

chroot is fine for this use case in general

pivot_root is a hack IMO

[near tartan]

yeah it feels like it is

[marble salmon]

it predates bind mounts which can solve this more cleanly

[near tartan]

so i guess my initramfs "init" is just a simple mount + chroot + execve

[marble salmon]

or, at least, the magic chdir change of pivot_root is a hack

[near tartan]

HOLY SHITTTTT

[near tartan]

bro 😭

[somber solar]

[near tartan]

i hate my life

i forgot to swap out the raw CStrs in execve

that's why its dying

i think

[near tartan]

yeah that's it

great

[near tartan]

nice

i'm now loading modules from userspace

[near tartan]

you can now mount partitions by their partuuid

[somber solar]

is that not a symlink?

[near tartan]

it is

just points to the partition device

[somber solar]

why not use /dev/disk/by-partuuid :P

[near tartan]

i dont like subdirs

i also have /dev/drmcard0

[grave peak]

lol

[brisk totem]

mmm, stringly typed kernel APIs

my favorite

[near tartan]

idk how else to do it

[brisk totem]

u8 source[] = {0x53, 0x2c, ..., 0xed};
struct zn_mount_op mount_op = {
  .fstype = ZN_FS_EXT4,
  .target = "/tmp", .target_len = 4,
  .source_type = ZN_MOUNT_BY_PARTUUID, .source = source, .source_len = sizeof(source)
};
if (zn_mount(&mount_op) != 0) { perror("zn_mount"); return 1; }

ur welcome

[near tartan]

that's not very ergonomic

what if you don't want to mount anything

you just want to open the device

[brisk totem]

and then do what with it

[near tartan]

io?

[brisk totem]

yeah i mean like what io

creating and managing partitions?

you can do that with a kernel api 🙃

[near tartan]

💀

that's dumb

[brisk totem]

you kinda need a kernel api for that anyway

you need to sync against the kernel

[near tartan]

not really

in what case

[brisk totem]

well i guess you dont need a lot of sync

you could just tell the kernel to rescan partitions or smth

[near tartan]

for a bit i'll stop working on this to focus on doing mlibc work

[nocturne rampart]

I mean you'll have to end up doing it for ptys

[near tartan]

[nocturne rampart]

/dev/pts/67

I don't like it either

[nocturne rampart]

https://media.discordapp.net/attachments/1349870841965576223/1351919652653240320/attachment.gif

[near tartan]

fixed the ext2 memes

[near tartan]

lol

[somber solar]

fix yo getdents

[near tartan]

yeah

that's tmpfs 😢

yay

[near tartan]

unfortunate

even more unfortunate

i really want sockets

[somber solar]

tty1 is gone

[near tartan]

yeah i renamed it to fbcon

the tty1 was a test

[somber solar]

console is still linked to it :P

[near tartan]

oops

[somber solar]

you probably still want /dev/tty

[near tartan]

yea ofc

but it's fake rn

[hybrid island]

how is the marv

is the zinnia working?

[near tartan]

i'm doing ttys right now

[hybrid island]

cool

[near tartan]

it's kind of working

i have ext2 now

and signals

[hybrid island]

I'm doing UVM-style VMM for BadgerOS ATM

[near tartan]

nice

[hybrid island]

Still some bugs to iron out

[near tartan]

i can start gcc now :3

[hybrid island]

After that I'm going to do IoKit-like device rewrite

[near tartan]

[hybrid island]

lmao

that's not arch

[hollow hedge]

I was thinking about looking into UVM, any resources you'd recommend?

[hybrid island]

The UVM paper combined with NetBSD sauce kodes for it

[near tartan]

very productive day today 😃

i can even start openrc now

somewhat

once i figure out unix sockets we can do xorg

@magic charm iirc you did something with keyronex udp sockets and xorg

[near tartan]

todo for today

• check ppoll impl for bash input
• implement get/setresid, deprecate the individual setid syscalls
• finish ttys
• unix sockets

[magic charm]

I did get xorg happily running remote apps over TCP

[near tartan]

how was that set up? running the server on linux?

[magic charm]

Xorg on keyronex and apps on Linux

[near tartan]

ah

that's cool

i have no idea how to even start with sockets tbh

i have no resources

[brisk totem]

how is it that literally everyone is doing advanced shit

and im still failing at baby allocator implementation

[near tartan]

advanced

bro have you seen ANY of my impls

it's the most cursed dogshit ever

[brisk totem]

yes but you have userspace

[near tartan]

ok and

[brisk totem]

and i have a hello world tier kernel without an allocator

is the solution to just not do good impls then

bc if so then thats lame

[near tartan]

korona said something about that as well

"make it exist first, then make it good"

and i mean my impls aren't the worst

they're above average i guess

[past leaf]

and im stuck in a loop trying to do a page cache before i refactor the kernel 😭

[brisk totem]

lmao

right now i didnt even get to a page cache

[past leaf]

it's literally a year that i say i have to do it

[brisk totem]

im refactoring my allocator rn

[past leaf]

and i never get to do it

[brisk totem]

but my old allocator was garbage

[past leaf]

turns out a 4kloc syscalls.c file is not the best design ever

[near tartan]

my vmem allocator for userspace is a bump allocator

[brisk totem]

fucking what

[past leaf]

that's a good idea tho

for 64bit

[brisk totem]

i mean true

[near tartan]

yeah i have enough virtual space so shrug

[brisk totem]

personally id just have libc choose the address

[past leaf]

like, even if you allocate 1GB/s

[brisk totem]

and then only do MAP_FIXED in kernel mode

[past leaf]

it's literally 2 days of uptime

[near tartan]

ooh good point

[magic charm]

TCP/IP Illustrated vol 2 is excellent, it covers the socket architecture more broadly not just tcp/ip itself. And cross referenced with the actual bsd sockets code from BSD Net 3

[brisk totem]

thats what im gonna do for my kernel bc easier (well im doing MAP_FIXED_NOREPLACE because posix sucks, i want to do strict accounting, making an allocator that refuses to take memory back is fucking insane, and while you can make failure-free munmap/mprotect work its really fucking annoying)

[near tartan]

is it available online?

[past leaf]

yes

[magic charm]

Either vol 2 or vol 3 even specifically has a section on unix domain sockets

[brisk totem]

a unix domain socket is just a duplex pipe + duplex fd passing, right? (with some extra headers to make the two line up right)

so if you dont care about quality of impl, you can just make a pipe of messages where each message owns the data inside

[magic charm]

It supports the full sockets interface including the cmsg api but in principle you might special case and implement it trivially

[brisk totem]

and to optimize this you can store a range over a data ring instead of a separate alloc

[near tartan]

damn

maybe i CAN implement sockets this week

[brisk totem]

yeah thats true, it does kinda suck to have a separate path for uds

hmm. i wanted to make a bitmap for tracking which objects within my slab allocator are in use, but it kinda sucks to reserve 512 bytes from each zone for a bitmap, when almost all objects in practice are big enough that you wont use all of it

[magic charm]

Xorg certainly doesnt demand much and can live without cmsg support and the like

[brisk totem]

i guess its still less than 1% overhead

[past leaf]

i found it much easier to just keep it as a linked list of messages, that can be either control or data

i think i got this from keyronex

[brisk totem]

yeah but linked list slow and ugly

[past leaf]

mitshm-using applications dont really write much data to the socket

xterm is an absolute hog tho

[brisk totem]

xterm is cringe anyway so it doesnt matter

me when unicode

[storm bobcat]

What are you thinking of

rn I have a classic slab thing but I wanna improve it

One thing I can immediately think of is having alloc and free magazines

[brisk totem]

slab with magazines basically

lol

[storm bobcat]

yeah I already do that

[brisk totem]

and i dont have anything 🙃

[storm bobcat]

What's a bit annoying is the hashmap

[brisk totem]

why do you need a hashmap?

[storm bobcat]

For large slabs you need one to convert an object pointer to its bufctl

[brisk totem]

or you can just not have large slabs 🙃

[storm bobcat]

they're useful but not that common

[brisk totem]

idk are they useful

[storm bobcat]

I cant use a tree because that would make the bufctls huge

in vmem I was able to replace the hashmap with a rbtree tho

[brisk totem]

im trying to avoid having a hashmap

by making consumers keep track of the alloc state structures

[storm bobcat]

It's probably not that big of a problem since large slabs aren't that common, but I still need to have some kind of mechanism to resize it

[brisk totem]

just preallocate it

[storm bobcat]

Which would be done in a thread or something

[brisk totem]

a large slab is how big?

[brisk totem]

actually, what do you even use them for?