uACPI - a portable and easy-to-integrate ACPI implementation | Operating System Development | Page 37

calm latch Jan 31, 2025, 10:18 PM

#

Most of my servers are single threaded*, but I had plans to do this eventually (if it gets big I guess)

jaunty fox Jan 31, 2025, 10:19 PM

#

i love this trick

calm latch Jan 31, 2025, 10:20 PM

#

also, you don't have to deal with concurrency as much if you don't want to with microkernels

gentle peak Jan 31, 2025, 10:20 PM

#

that's interesting

calm latch Jan 31, 2025, 10:20 PM

#

now that I think of it

gentle peak Jan 31, 2025, 10:20 PM

#

I wonder how the performance of the two approaches compare on modern hardware

jaunty fox Jan 31, 2025, 10:21 PM

#

nowadays with ASIDs and assuming that most interdomain calls will be to do a small amount of work, i think more preferable is to try to call into a cpu local thread

#

i expect the overhead of migrating between processors will dominate the gain from not losing the l1 cache, which you may lose anyway if the system is loaded enough that the calling thread doesn't simply leave the core unburdened but is rather scheduled out

left orbit Jan 31, 2025, 10:35 PM

#

would it be stupid to just use a stripped down linux as a base for a "microkernel" based operating system lol

#

i had this idea just now, just disable everything that can be disabled and implement everything in userspace meme

loud ice Jan 31, 2025, 10:37 PM

#

you can't disable all that much on linux i think

mortal yoke Jan 31, 2025, 10:37 PM

#

it would probably be a decent amount of work to expose the stuff you need to userspace and even if you strip away a lot of the config options its going to be monolithic having almost all of the syscalls

gentle peak Jan 31, 2025, 10:40 PM

#

a lot of the abstractions that linux uses to expose info that drivers need to userspace rely on having fully functional drivers for those devices in the first place

left orbit Jan 31, 2025, 10:41 PM

#

i dont care

gentle peak Jan 31, 2025, 10:41 PM

#

plus it has no way to process IRQs fully in userspace, you need a hardirq handler for every device in kernel

left orbit Jan 31, 2025, 10:41 PM

#

i just want pci in userspace

dense steppe Jan 31, 2025, 10:41 PM

#

well Linux is open source. so you can fork it if you want to rewrite it into a microkernel

left orbit Jan 31, 2025, 10:41 PM

#

exactly

#

i can just add my own stuff to it

dense steppe Jan 31, 2025, 10:41 PM

#

but you would have to rewrite a lot

gentle peak Jan 31, 2025, 10:41 PM

#

it'd probably be more effort than just writing a microkernel

mortal yoke Jan 31, 2025, 10:42 PM

#

yeah its likely easier to design a kernel with it in mind from the start than to try converting linux

left orbit Jan 31, 2025, 10:42 PM

#

:(

calm latch Jan 31, 2025, 10:43 PM

#

small functional microkernel is like 10k LoC

leaden fox Jan 31, 2025, 10:52 PM

#

left orbit i just want pci in userspace

Real af

#

uacpi in userspace

calm latch Jan 31, 2025, 10:56 PM

#

leaden fox uacpi in userspace

i have uacpi in userspace

#

and pci

left orbit Jan 31, 2025, 10:57 PM

#

yeah i mean, you can do port io in userspace + you can map mmio so it's not that crazy

#

i like the idea of just using userspace apis for everything in uacpi tho lol

north hull Jan 31, 2025, 11:59 PM

#

ok I mean

#

at least I beat Nyaux

gentle peak Feb 1, 2025, 12:05 AM

#

I think that's slightly lower than what nyaux gets without ubsan actually

#

Idk how much has changed since then but when I fixed its allocator bug it got 1M

#

Might be due to cpu model though

north hull Feb 1, 2025, 12:06 AM

#

oh

#

this is with -O0 and ubsan

#

lemme try optimized build

#

nvm im trolling

#

its O1

left orbit Feb 1, 2025, 12:08 AM

#

ubsan had a decent hit in my kernel actually

north hull Feb 1, 2025, 12:09 AM

#

so

#

not sure what to make of that

#

so it like

#

swings by 400k randomly

left orbit Feb 1, 2025, 12:10 AM

#

bad timekeeping?

#

or are you running under wsl2?

north hull Feb 1, 2025, 12:10 AM

#

I just changed to using hpet

gentle peak Feb 1, 2025, 12:10 AM

#

Probably host system scheduling

north hull Feb 1, 2025, 12:10 AM

#

left orbit or are you running under wsl2?

qemu on native linux

slim panther Feb 1, 2025, 12:10 AM

#

north hull swings by 400k randomly

tends to happen

#

my kernel swings around 200k points from time to time

north hull Feb 1, 2025, 12:11 AM

#

gentle peak Feb 1, 2025, 12:11 AM

#

old proxima on my cpu has multiple performance ranges: 4.0-4.4M, 5.0-5.3M, 6.5-6.7M, 7.5-7.7M

north hull Feb 1, 2025, 12:12 AM

#

gentle peak Feb 1, 2025, 12:12 AM

#

I believe each range corresponds to a different host cpu speed setting

north hull Feb 1, 2025, 12:12 AM

#

so now it swung by 700k+

gentle peak Feb 1, 2025, 12:12 AM

#

north hull so now it swung by 700k+

you probably have a range at ~500k and at ~1.1M

north hull Feb 1, 2025, 12:13 AM

#

I suppose yeah

#

probably host system stuff

slim panther Feb 1, 2025, 12:13 AM

#

just take whatever is the highest meme

north hull Feb 1, 2025, 12:13 AM

#

lol

#

I dont want to submit this as an actual score

#

im using a shitty heap I cooked up to see how much my heap is impacting perf

#

before using this heap I got 34k

gentle peak Feb 1, 2025, 12:13 AM

#

Proxima's scores are the ~median of the highest performance range

north hull Feb 1, 2025, 12:13 AM

#

so time to make a slab allocator ig

#

lol

calm latch Feb 1, 2025, 12:24 AM

#

pmos has gotten 3.3M with dlmalloc, which is best-fit trl

#

on top of first fit vmm and segregated fit pmm in kernel

frank canopy Feb 1, 2025, 12:27 AM

#

sadly i still dont have a new number with my new mm/alloc, thanks to my page allocator is panicking like halfway through init

north holly Feb 1, 2025, 4:36 AM

#

#

was scrolling through my thread's old messages

#

#1141057599584878645 message

fiery turtle Feb 1, 2025, 9:41 AM

#

Lol

fiery turtle Feb 1, 2025, 9:41 AM

#

north hull

Letsgo

vale isle Feb 1, 2025, 9:42 AM

#

i don't remember which microkernel came up with it but there's microkernels where tasks can choose which vmm they want to use (which are also userspace

#

iirc fuchsia has that

fiery turtle Feb 1, 2025, 10:06 AM

#

@vagrant hull btw did u figure out the ironclad bug on your laptop?

urban galleon Feb 1, 2025, 10:34 AM

#

Can uACPI Generate aml

zenith crescent Feb 1, 2025, 10:36 AM

#

urban galleon Can uACPI Generate aml

What do you mean by generate aml?

urban galleon Feb 1, 2025, 10:37 AM

#

Something like iASL

fiery turtle Feb 1, 2025, 10:38 AM

#

Its not a compiler

urban galleon Feb 1, 2025, 10:38 AM

#

ok

leaden fox Feb 1, 2025, 11:48 AM

#

vale isle i don't remember which microkernel came up with it but there's microkernels wher...

L4 probably as well

calm latch Feb 1, 2025, 11:51 AM

#

vale isle i don't remember which microkernel came up with it but there's microkernels wher...

Do you mean pagers?

vale isle Feb 1, 2025, 11:51 AM

#

yep

calm latch Feb 1, 2025, 11:51 AM

#

It's very common for microkernels

#

Mach had it

#

I have it

vale isle Feb 1, 2025, 11:52 AM

#

some designs have pagers per process, some have pagers shared across processes with a default pager

calm latch Feb 1, 2025, 11:52 AM

#

But that's more of a pmm though

vale isle Feb 1, 2025, 11:52 AM

#

i thought it handled paging in and out too?

calm latch Feb 1, 2025, 11:52 AM

#

vale isle some designs have pagers per process, some have pagers shared across processes w...

I have pagers per memory objects

vale isle Feb 1, 2025, 11:53 AM

#

ah

#

how does that work

calm latch Feb 1, 2025, 11:53 AM

#

vale isle i thought it handled paging in and out too?

But it's up to the kernel to decide (I think)

#

That's what I'll do

calm latch Feb 1, 2025, 11:53 AM

#

vale isle how does that work

When a thread accesses a page that is not present, it is stopped and kernel asks the pager to provide it

vale isle Feb 1, 2025, 11:53 AM

#

i don't really want to overengineer my pmm/vmm in my first attempt at a microkernel so I'll keep them in kernel and might move them out at some point

calm latch Feb 1, 2025, 11:54 AM

#

Even Linux has it

calm latch Feb 1, 2025, 11:54 AM

#

vale isle i don't really want to overengineer my pmm/vmm in my first attempt at a microker...

How do you do mmap though

vale isle Feb 1, 2025, 11:54 AM

#

idk i never got to that point

#

I'll need to read that some papers

torpid root Feb 1, 2025, 12:16 PM

#

hmmmmmmmmmmmmmmmmmmmmm

#

sus

#

both values

#

I don't return higher half address from any of my api functions

gentle peak Feb 1, 2025, 12:23 PM

#

that's not higher half, that's kernel image

#

which is even more fucked

#

actually addr is supposed to be a physical address there so it's worse than i thought

torpid root Feb 1, 2025, 12:26 PM

#

oof

#

yeah missed a few F s there lol

#

smh how can this even happen

rustic compass Feb 1, 2025, 12:34 PM

#

whats the advantage of a per process vmm?

mortal yoke Feb 1, 2025, 12:36 PM

#

?

rustic compass Feb 1, 2025, 12:37 PM

#

was a question to a discussion further up the chat

north holly Feb 1, 2025, 12:38 PM

#

torpid root smh how can this even happen

Your kernel is cursee

#

*curses

#

**cursed

rustic compass Feb 1, 2025, 12:41 PM

#

its the next generation os, a os for 128bit cpu's

torpid root Feb 1, 2025, 12:45 PM

#

as far as I can tell, it happens after this line: https://github.com/uACPI/uACPI/blob/80b4e1ec7f1774b445a4f3011c49251c2396e3e4/source/registers.c#L552

#

before unlock

#

so in do_write_register

#

hmm but I don't think that calls uacpi_kernel_map

#

definitely after line 533

#

@fiery turtle any tips? 🥺

north holly Feb 1, 2025, 12:52 PM

#

Uncurse your kernel

fiery turtle Feb 1, 2025, 12:54 PM

#

torpid root <@353593948523986945> any tips? 🥺

No idea sorry lol

#

Printf debugging

#

Do write most likely dispatches into your api

calm latch Feb 1, 2025, 2:08 PM

#

rustic compass whats the advantage of a per process vmm?

It's not per process, but more of per mmap region, with several regions being able to have the same pager

#

So that filesystems in userspace can read data on demand from disk and whatnot

torpid root Feb 1, 2025, 2:24 PM

#

is it trying to execute the stack

#

smh

slim panther Feb 1, 2025, 2:34 PM

#

torpid root is it trying to execute the stack

smh

torpid root Feb 1, 2025, 2:34 PM

#

truer words have never been spoken

#

how the hell do I trace it if stack is trashed

calm latch Feb 1, 2025, 2:37 PM

#

gdb?

#

(If it pagefaults in the same place, just put a watchpoint at that address)

torpid root Feb 1, 2025, 2:38 PM

#

invalid opcode

calm latch Feb 1, 2025, 2:39 PM

#

-fno-omit-framd-pointer and %rbp?

torpid root Feb 1, 2025, 2:41 PM

#

just the last value of rip and then nothing

torpid root Feb 1, 2025, 2:58 PM

#

watch is so slow

#

g_register_mappings+88 is being set to an address in stack

north holly Feb 1, 2025, 3:01 PM

#

Is this the qemu blob?

torpid root Feb 1, 2025, 3:01 PM

#

?

north holly Feb 1, 2025, 3:02 PM

#

the one that breaks in your kernel

torpid root Feb 1, 2025, 3:02 PM

#

yes

#

maybe, I haven't tested anything else

#

it broke after 1.0

north holly Feb 1, 2025, 3:02 PM

#

idk maybe your kernel is cursed or smth

torpid root Feb 1, 2025, 3:02 PM

#

nooo

north holly Feb 1, 2025, 3:03 PM

#

or you have elves playing around with your computer

torpid root Feb 1, 2025, 3:09 PM

#

this kernel can go and shit itself
I'm done smh

leaden fox Feb 1, 2025, 3:17 PM

#

osdev gang just shilled uacpi to FreeBSD at fossdem

#

Upstream uacpi coming soon to the BSDs

fiery turtle Feb 1, 2025, 3:26 PM

#

Lol for real?

slim panther Feb 1, 2025, 3:26 PM

#

leaden fox osdev gang just shilled uacpi to FreeBSD at fossdem

lmao

leaden fox Feb 1, 2025, 3:29 PM

#

fiery turtle Lol for real?

to obiwan himself

slim panther Feb 1, 2025, 3:29 PM

#

based

fiery turtle Feb 1, 2025, 3:34 PM

#

Based af

vale isle Feb 1, 2025, 3:50 PM

#

based af

#

imagine uacpi in bsds before uacpi in linux

#

i wonder what the BSDs would get in the leaderboards

#

ig what's left is uacpi in NT

#

imagine all major operating systems using uacpi

#

major operating systems include reactos and bsds and haiku

gentle peak Feb 1, 2025, 3:57 PM

#

vale isle ig what's left is uacpi in NT

may i introduce you to https://www.youtube.com/watch?v=gUEzkglsXn4

YouTube

The_DarkFire_

ReactOS ACPI / uACPI windows driver wakes up windows 10 for the fir...

▶ Play video

vale isle Feb 1, 2025, 3:58 PM

#

I'm aware of that yes

#

darkfire should upstream it to microsoft

gentle peak Feb 1, 2025, 3:59 PM

#

true

vale isle Feb 1, 2025, 3:59 PM

#

:^3

#

uacpi will become the reference implementation

slim panther Feb 1, 2025, 3:59 PM

#

gentle peak may i introduce you to https://www.youtube.com/watch?v=gUEzkglsXn4

pure ecstasy

vale isle Feb 1, 2025, 3:59 PM

#

yea that's 😎

#

i should play around with reactos one day

#

or potentially make my own NT clone

#

or hear me out: darwin clone?

fiery turtle Feb 1, 2025, 4:03 PM

#

vale isle or potentially make my own NT clone

Based

vale isle Feb 1, 2025, 4:04 PM

#

cloning a proprietary os bug-for-bug (including driver interfaces and stuff) sounds fun

#

NT is the top candidate there sadly

left orbit Feb 1, 2025, 4:05 PM

#

i am working on my own, sort of, nt clone

vale isle Feb 1, 2025, 4:05 PM

#

vale isle NT is the top candidate there sadly

wish there was an alternative

#

because NT is already done kinda

#

with reactos

left orbit Feb 1, 2025, 4:05 PM

#

i want an nt like architecture but native posix support

vale isle Feb 1, 2025, 4:06 PM

#

take NT 4 and fix its posix subsystem

left orbit Feb 1, 2025, 4:06 PM

#

idk how that last part is gonna work yet, but we'll see

vale isle Feb 1, 2025, 4:06 PM

#

i wonder how unhappy windows drivers would get if they were ran in userspace

mortal yoke Feb 1, 2025, 4:07 PM

#

at least amd gpu driver would not be happy lol

vale isle Feb 1, 2025, 4:07 PM

#

vale isle i wonder how unhappy windows drivers would get if they were ran in userspace

surely microsoft abstracts away quite a lot of stuff in the driver interface

left orbit Feb 1, 2025, 4:07 PM

#

considering you provide enough apis it should just work

#

yes they do

#

well, i mean

vale isle Feb 1, 2025, 4:08 PM

#

then why don't drivers just work in WoA?

left orbit Feb 1, 2025, 4:08 PM

#

most stuff is abstracted away by the kernel itself already

#

probably because x86->aarch64 jit in kernel isn't that great of an idea

mortal yoke Feb 1, 2025, 4:08 PM

#

mortal yoke at least amd gpu driver would not be happy lol

it uses cr registers for something (I think virtualization detection), though ofc you could emulate that with sigsegv handler (and same for the fixed-address shared data page that contains stuff like the system time)

vale isle Feb 1, 2025, 4:09 PM

#

left orbit probably because x86->aarch64 jit in kernel isn't that great of an idea

hmm

#

don't give me ideas

#

i had a x86 emulator and was thinking about adding dynarec to it

left orbit Feb 1, 2025, 4:10 PM

#

you mean bad ideas?

vale isle Feb 1, 2025, 4:10 PM

#

ye

left orbit Feb 1, 2025, 4:10 PM

#

no problem

mortal yoke Feb 1, 2025, 4:10 PM

#

mortal yoke it uses cr registers for something (I think virtualization detection), though of...

this stuff

left orbit Feb 1, 2025, 4:11 PM

#

right

#

you can probably patch these ahead of time

#

or when transpiling code

vale isle Feb 1, 2025, 4:12 PM

#

mortal yoke this stuff

how do you know this

left orbit Feb 1, 2025, 4:12 PM

#

it's in the public kernel headers

vale isle Feb 1, 2025, 4:12 PM

#

ah

left orbit Feb 1, 2025, 4:12 PM

#

drivers use this

mortal yoke Feb 1, 2025, 4:12 PM

#

yeah the driver kit headers contain this stuff

vale isle Feb 1, 2025, 4:12 PM

#

I thought qwinci just broke an NDA

left orbit Feb 1, 2025, 4:13 PM

#

the same shit is mapped in user mode at 0x7ffe0000

#

though as read only memory

#

in kernel it's mapped read write

#

for obvious reasons

vale isle Feb 1, 2025, 4:13 PM

#

a clone could map it read write for drivers

#

NT clone but microkernel :3

left orbit Feb 1, 2025, 4:13 PM

#

drivers can already write to it

#

in kernel space it's mapped read write

vale isle Feb 1, 2025, 4:14 PM

#

no for running it in userspace

left orbit Feb 1, 2025, 4:14 PM

#

oh

#

yeah good luck

#

you'll probably have to emulate it

vale isle Feb 1, 2025, 4:14 PM

#

probably

#

sadge

#

I'm not doing this yet

left orbit Feb 1, 2025, 4:14 PM

#

luckily you can catch exceptions when accessing addresses > user space end iirc

#

i mean i guess that is kinda obvious

vale isle Feb 1, 2025, 4:15 PM

#

I know how hard would it be from reactos lol

jaunty fox Feb 1, 2025, 4:15 PM

#

leaden fox osdev gang just shilled uacpi to FreeBSD at fossdem

kinda based

#

and under BSD tutelage uacpi remains an independent project

fiery turtle Feb 1, 2025, 4:15 PM

#

left orbit i am working on my own, sort of, nt clone

Damn, which language and do u have a src

left orbit Feb 1, 2025, 4:15 PM

#

C

#

not yet

jaunty fox Feb 1, 2025, 4:15 PM

#

rather than getting sucked up into the linux code and the beloved kernel api getting replaced with direct calls

fiery turtle Feb 1, 2025, 4:15 PM

#

LETSFUCKINGGOOOOOO

left orbit Feb 1, 2025, 4:15 PM

#

i will probably put it on github the moment i have a working loader

frank canopy Feb 1, 2025, 4:15 PM

#

oh hey same, though mine has already diverged a lot and is zig

#

more like nt-inspired than an nt clone

fiery turtle Feb 1, 2025, 4:16 PM

#

left orbit i will probably put it on github the moment i have a working loader

Loader?

left orbit Feb 1, 2025, 4:16 PM

#

yeah im just inspired by the nt architecture, i dont really aspire to be an actual nt clone

left orbit Feb 1, 2025, 4:16 PM

#

fiery turtle Loader?

prepares some structures for the kernel, loads it and the drivers

#

not sure what it'll do just yet

fiery turtle Feb 1, 2025, 4:16 PM

#

And thats loaded by limine?

left orbit Feb 1, 2025, 4:16 PM

#

but i know i'll do the initial loading and mapping of kernel and some drivers

#

yeah, for now

fiery turtle Feb 1, 2025, 4:17 PM

#

Like eir

left orbit Feb 1, 2025, 4:17 PM

#

basically

fiery turtle Feb 1, 2025, 4:17 PM

#

I see

left orbit Feb 1, 2025, 4:17 PM

#

i want to have a pure efi loader at some point too

fiery turtle Feb 1, 2025, 4:17 PM

#

Is your kernel a pe

left orbit Feb 1, 2025, 4:17 PM

#

but not yet, that's a bit more work than limine :^)

#

nope

#

i use elf

fiery turtle Feb 1, 2025, 4:17 PM

#

Ah

left orbit Feb 1, 2025, 4:18 PM

#

while pe would be much easier to work with, and im more famiilar with it, i think i want to use elf

mortal yoke Feb 1, 2025, 4:18 PM

#

one nice thing I discovered is that with clang you can use x86_64-unknown-windows-elf target to get the ms calling conv (as it doesn't support gcc's -mabi=ms) in an elf file

frank canopy Feb 1, 2025, 4:18 PM

#

ive considered pe now that limine can boot it but zig cant get stack traces on pe without windows being there but it can iterate on the base pointer easily on elf

jaunty fox Feb 1, 2025, 4:19 PM

#

gcc and clang are generating PEs with DWARF sections for debug info iirc

#

not the true and authentic format normally used for them

mortal yoke Feb 1, 2025, 4:20 PM

#

stack unwinding is broken tho as gdb doesn't support seh and afaik they still generate seh unwind info

#

even with frame pointers enabled I couldn't get it to work nicely, stuff like finish or stepping out of a function just continued the execution and backtrace only showed the current frame (or the one directly before the current one if lucky)

frank canopy Feb 1, 2025, 4:22 PM

#

yeah its a bit of a mess, thats why i dont try to use pe for my zig stuff too

fiery turtle Feb 1, 2025, 4:25 PM

#

left orbit while pe would be much easier to work with, and im more famiilar with it, i thin...

Why is that?

#

Also why not zig btw

left orbit Feb 1, 2025, 4:29 PM

#

PE is just

here's a list of sections, load them into memory
here's a list of relocations, u have to implement like 3 types of relocations - done
here's a list of external functions i want and where to find them, pls put the addresses there

ELF is similar in some regards, it's not hard by any means, but i am not familiar with ELF relocations at all, and external symbols in ELF executables work different and i find it weird how they work, also there is two different types of relocations? rel and rela, from what i gathered it's just implicit addend vs explicit addend, but i prefer the way PE does thing

#

also i thought about zig for a moment but i don't want to fight the compiler, i just want to write working code meme

frank canopy Feb 1, 2025, 4:30 PM

#

side note on the stack trace stuff, im really looking forward to https://github.com/ziglang/zig/pull/22077 so i can use the std stack trace printing and also get more info than i already have from the trace print

gentle peak Feb 1, 2025, 4:30 PM

#

rel vs rela is easy, but each architecture has like 10 ways to calculate the relocated address

frank canopy Feb 1, 2025, 4:31 PM

#

left orbit also i thought about zig for a moment but i don't want to fight the compiler, i ...

ive not had to fight the zig compiler much at all for my zig kernel so its def not bad compared to what it used to be

north holly Feb 1, 2025, 4:31 PM

#

leaden fox osdev gang just shilled uacpi to FreeBSD at fossdem

based

gentle peak Feb 1, 2025, 4:31 PM

#

on one hand it allows for some very nice features (like ifunc) but on the other it is much more annoying to implement than pe relocation

left orbit Feb 1, 2025, 4:31 PM

#

frank canopy ive not had to fight the zig compiler much at all for my zig kernel so its def n...

yeah no, i'm not saying that, i just don't feel like trying to do dynamic linking in zig

#

i haven't done that and i don't want to find out what it takes lol

mortal yoke Feb 1, 2025, 4:32 PM

#

mortal yoke stack unwinding is broken tho as gdb doesn't support seh and afaik they still ge...

actually looks like it does support it thonk idk why it didn't really work even with dwarf info then

frank canopy Feb 1, 2025, 4:32 PM

#

left orbit yeah no, i'm not saying that, i just don't feel like trying to do dynamic linkin...

oh god yeah if youre doing dynamic linking then i understand. zig is very much not built to do that at the osdev level in my experience

vagrant hull Feb 1, 2025, 6:53 PM

#

fiery turtle <@379163868364275712> btw did u figure out the ironclad bug on your laptop?

we're @ FOSDEM now we did not have time

fiery turtle Feb 1, 2025, 6:54 PM

#

ah

#

have fun

vagrant hull Feb 1, 2025, 6:55 PM

#

we shilled uACPI to obiwan after their FreeBSD modern standby talk

#

i like making fun of the ACPICA maintainer smashing the corvette

fiery turtle Feb 1, 2025, 6:57 PM

#

leaden fox osdev gang just shilled uacpi to FreeBSD at fossdem

we already know meme

fiery turtle Feb 1, 2025, 6:57 PM

#

vagrant hull i like making fun of the ACPICA maintainer smashing the corvette

lmfao

north holly Feb 1, 2025, 7:05 PM

#

vagrant hull we shilled uACPI to obiwan after their FreeBSD modern standby talk

vagrant hull Feb 1, 2025, 7:06 PM

#

lol

dense steppe Feb 1, 2025, 7:07 PM

#

I just imagine like Obi-Wan sitting in front of a computer and being like... I feel a disturbance in the VMM...

torpid root Feb 2, 2025, 11:39 AM

#

I have uncursed my kernel and I hate myself even more now

#

when I added v1.0 api sized io functions, I kept my old one that had a switch with sizeof(type) cases. I passed 8, 16 and 32 as width instead of 1, 2 and 4 from the respective uacpi functions. I was calling fucking std::unreachable in default: case

fiery turtle Feb 2, 2025, 11:44 AM

#

bruh

#

good job uncursing it!

torpid root Feb 2, 2025, 11:47 AM

#

with v1 I get more points on my shitty cpu than on i7-13700k previously

#

but could a kind good cpu haver test it?

#

qemu-system-x86_64 -serial stdio -cpu host,migratable=off -M q35 -accel kvm -cdrom image-x86_64.iso

📎 image-x86_64.iso

#

run a few times as difference can be a hundred thousand points

gentle peak Feb 2, 2025, 11:54 AM

#

~4.1M/s on ryzen 5800x

torpid root Feb 2, 2025, 11:58 AM

#

thank you

mortal yoke Feb 2, 2025, 11:59 AM

#

I should also update to uacpi 1.0 and see what score I get

#

(on the nt thing, I implemented some of the stubbed uacpi apis like events/mutexes yesterday as I got around to implementing KeReleaseSemaphore/KeWaitForSingleObject)

torpid root Feb 2, 2025, 12:09 PM

#

gentle peak ~4.1M/s on ryzen 5800x

could you try this as well? I get 900k more points

📎 image-x86_64.iso

gentle peak Feb 2, 2025, 12:09 PM

#

~4.9M

calm latch Feb 2, 2025, 12:09 PM

#

I wonder how IA32 pmOS will perform (since AMD64 is just an insane mess) thonk

mortal yoke Feb 2, 2025, 12:10 PM

#

mortal yoke I should also update to uacpi 1.0 and see what score I get

apparently its around the same

torpid root Feb 2, 2025, 12:10 PM

#

gentle peak ~4.9M

@fiery turtle bigplead

gentle peak Feb 2, 2025, 12:11 PM

#

i can get an exact number to put on the leaderboard if you want

#

[00:00:00.002049] [info] uacpi: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 4906870/s)

fiery turtle Feb 2, 2025, 12:12 PM

#

mortal yoke apparently its around the same

wait how tf do u get more than proxima

gentle peak Feb 2, 2025, 12:12 PM

#

that's slightly less than proxima

#

by like 400k

fiery turtle Feb 2, 2025, 12:12 PM

#

torpid root <@353593948523986945> <:bigplead:1299010072411639909>

queued LETSFUCKINGGOOOOOO

fiery turtle Feb 2, 2025, 12:12 PM

#

gentle peak that's slightly less than proxima

ah

#

still pretty insane

#

didnt know your os was that fast

torpid root Feb 2, 2025, 12:13 PM

#

fiery turtle queued <:LETSFUCKINGGOOOOOO:1003833035688513606>

thanks!
anything else queued? because that would put me in the second place on the current leaderboard, or third with qwinci kernel included

#

monkuous is an osdev magician

fiery turtle Feb 2, 2025, 12:14 PM

#

i think davix improvements also but they havent submitted them for the leaderboard yet

mortal yoke Feb 2, 2025, 12:15 PM

#

maybe I should also get this thing on the leaderboard after I implement the rest of the functions (well its mostly just the pci stuff that's missing), though Ill have to also come up with a name for it and put it to git KEKW

#

I think a big factor to the score when compared to crescent is that I didn't spam cli;pushf;sti everywhere lol

dense steppe Feb 2, 2025, 12:16 PM

#

fiery turtle i think davix improvements also but they havent submitted them for the leaderboa...

wrong... I have plans on improving my allocator but haven't actually done anything yet

#

also, it fluctuates between like 600K and 2M

#

might be host power management stuff

fiery turtle Feb 2, 2025, 1:22 PM

#

mortal yoke maybe I should also get this thing on the leaderboard after I implement the rest...

what about qacpi?

jaunty fox Feb 2, 2025, 1:23 PM

#

i should submit a score to the leaderboard some day

fiery turtle Feb 2, 2025, 1:24 PM

#

that would be a welcome addition

jaunty fox Feb 2, 2025, 1:25 PM

#

unfortunately i haven't one yet as keyronex itself is missing high resolution time of day (that was only in the testbed kernel) and that's a task to be dealt with after i've done the driver framework work and some usb stuff

fiery turtle Feb 2, 2025, 1:25 PM

#

makes sense

left orbit Feb 2, 2025, 1:28 PM

#

torpid root could you try this as well? I get 900k more points

[00:00:00.004587] [info] uacpi: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 5531368/s)

#

intel W once again

#

i wonder if i could get a better score more reliably by running linux on bare metal

#

a tiny upgrade
[00:00:00.003999] [info] uacpi: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 5662213/s)

mortal yoke Feb 2, 2025, 1:42 PM

#

fiery turtle what about qacpi?

Ill have to test after I integrate it + come up with some kind of abstraction

fiery turtle Feb 2, 2025, 1:43 PM

#

no like

#

are u gonna have both integrated or

gentle peak Feb 2, 2025, 1:53 PM

#

@fiery turtle i think i found a uacpi bug, if in create_gpe_block any of the uacpi_map_gas_noalloc calls fail after gas_validate, the cleanup uninstall_gpe_block calls uacpi_unmap_gas_nofree on the register whose map call failed, which tries to call gas->unmap, which is null because the map call didn't fully complete

#

to reproduce this just make uacpi_kernel_io_map unconditionally return an error, calling uacpi_namespace_load will trigger the bug (with qemu's blob at least)

fiery turtle Feb 2, 2025, 1:55 PM

#

doesnt it check for reg width before unmapping which would be zero there?

gentle peak Feb 2, 2025, 1:56 PM

#

it's not zero because gas_validate succeeded

north holly Feb 2, 2025, 1:56 PM

#

can reprod on obos

gentle peak Feb 2, 2025, 1:56 PM

#

it happens when gas_validate succeeds and either of the map calls fail

fiery turtle Feb 2, 2025, 1:56 PM

#

ah yeah ur right

#

these fields should not be filled on failure

#

ill fix it thx

north holly Feb 2, 2025, 1:57 PM

#

fiery turtle Feb 2, 2025, 1:58 PM

#

?

gentle peak Feb 2, 2025, 1:58 PM

#

stack trace i assume

north holly Feb 2, 2025, 1:58 PM

#

just a stack trace

fiery turtle Feb 2, 2025, 1:58 PM

#

ah

mortal yoke Feb 2, 2025, 2:00 PM

#

fiery turtle are u gonna have both integrated or

ah, yeah I think it would be nice if I can just come up with good enough abstraction so that I don't have to duplicate code or make it full of ifdefs lol

fiery turtle Feb 2, 2025, 2:01 PM

#

are you using c++ btw?

torpid root Feb 2, 2025, 2:03 PM

#

left orbit a tiny upgrade `[00:00:00.003999] [info] uacpi: successfully loaded 1 AML blob, ...

thank you very much
infy I hope you'll use this one :p

fiery turtle Feb 2, 2025, 2:05 PM

#

which cpu

left orbit Feb 2, 2025, 2:05 PM

#

i7-13700k

fiery turtle Feb 2, 2025, 2:05 PM

#

sure

fiery turtle Feb 2, 2025, 2:05 PM

#

left orbit a tiny upgrade `[00:00:00.003999] [info] uacpi: successfully loaded 1 AML blob, ...

slim panther Feb 2, 2025, 2:05 PM

#

left orbit a tiny upgrade `[00:00:00.003999] [info] uacpi: successfully loaded 1 AML blob, ...

holy crap

left orbit Feb 2, 2025, 2:06 PM

#

i am slowly starting to dominate the uacpi benchmarking scene

#

with 4 scores being tested by me meme

slim panther Feb 2, 2025, 2:06 PM

#

noooo im moving even lower

#

kms

left orbit Feb 2, 2025, 2:06 PM

#

step it up smh

mortal yoke Feb 2, 2025, 2:06 PM

#

fiery turtle are you using c++ btw?

yes

#

its literally crescent but nt'fied troll

torpid root Feb 2, 2025, 2:08 PM

#

slim panther holy crap

I was compiling uacpi before with optimisations disabled for some reason

slim panther Feb 2, 2025, 2:08 PM

#

stare

#

meanwhile i'm on -O3

torpid root Feb 2, 2025, 2:08 PM

#

me too

left orbit Feb 2, 2025, 2:08 PM

#

i am planning on having a decent IO + PnP subsystem so hopefully i can just have a pluggable acpi.sys

slim panther Feb 2, 2025, 2:08 PM

#

for some reason i am faulting in uacpi now

calm latch Feb 2, 2025, 2:08 PM

#

I'm not at home to submit the new build nooo

calm latch Feb 2, 2025, 2:08 PM

#

slim panther meanwhile i'm on -O3

pmOS is -O3 + lto + AVX512 trl

slim panther Feb 2, 2025, 2:09 PM

#

fym avx

left orbit Feb 2, 2025, 2:09 PM

#

it's userspace

slim panther Feb 2, 2025, 2:09 PM

#

oh

#

fair

gentle peak Feb 2, 2025, 2:12 PM

#

that makes me wonder, would it be cheating to run uacpi as a single-threaded userspace server and take advantage of that by stubbing out mutexes

#

like you're not sacrificing capability

fiery turtle Feb 2, 2025, 2:13 PM

#

sounds fair

calm latch Feb 2, 2025, 2:14 PM

#

I can't avoid mutexes in malloc because I have a libc thread for every process koreablob

gentle peak Feb 2, 2025, 2:14 PM

#

just make a uacpi-specific malloc implementation

fiery turtle Feb 2, 2025, 2:14 PM

#

malloc server 💀

gentle peak Feb 2, 2025, 2:14 PM

#

malloc implemented with ipc

#

actually that's just mmap with a lower granularity

north holly Feb 2, 2025, 2:15 PM

#

bruh why is lto compiling out all uacpi's symbols

left orbit Feb 2, 2025, 2:15 PM

#

why wouldn't it

north holly Feb 2, 2025, 2:15 PM

#

nothing in my static library persists after enabling lto in uacpi

left orbit Feb 2, 2025, 2:15 PM

#

they aren't needed

#

like most of them

#

oh trl

#

you dont enable lto in uacpi

#

you enable lto in your kernel

north holly Feb 2, 2025, 2:16 PM

#

sounds fair

calm latch Feb 2, 2025, 2:16 PM

#

left orbit you dont enable lto in uacpi

?

left orbit Feb 2, 2025, 2:16 PM

#

from what i understand he compiles uacpi and makes a static library out of it

#

which then he links into his kernel

north holly Feb 2, 2025, 2:16 PM

#

yes

gentle peak Feb 2, 2025, 2:17 PM

#

you do still need to enable lto for the static library

#

but you need to make sure to use gcc-ar instead of normal ar

left orbit Feb 2, 2025, 2:17 PM

#

ah there u go

#

i thought lto is a linker arg

#

wasnt sure how it affects the compilation itself

gentle peak Feb 2, 2025, 2:17 PM

#

it's both compiler and linker

north holly Feb 2, 2025, 2:17 PM

#

lto is linker+compiler iirc

gentle peak Feb 2, 2025, 2:18 PM

#

lto isn't possible unless the compiler generates IR instead of machine code

left orbit Feb 2, 2025, 2:18 PM

#

ohh right

north holly Feb 2, 2025, 2:19 PM

#

anyway I'll deal with that later

#

(never, I don't give a damn about lto atm)

#

I want to debug /bin/sleep

#

(it just hangs)

leaden fox Feb 2, 2025, 2:26 PM

#

uacpi shilling to haiku done

fiery turtle Feb 2, 2025, 2:27 PM

#

lmfao basesd

left orbit Feb 2, 2025, 2:28 PM

#

holy based

slim panther Feb 2, 2025, 2:28 PM

#

zased

#

what does haiku use atm

fiery turtle Feb 2, 2025, 2:29 PM

#

haiku is nice although i know 0 about beos or whatever its basd on

fiery turtle Feb 2, 2025, 2:29 PM

#

slim panther what does haiku use atm

acpica

left orbit Feb 2, 2025, 2:29 PM

#

🤢

#

acpica compat layer when

slim panther Feb 2, 2025, 2:29 PM

#

haiku pr when

north holly Feb 2, 2025, 2:29 PM

#

acpica

fiery turtle Feb 2, 2025, 2:29 PM

#

when haiku maintainers get around to it meme

fiery turtle Feb 2, 2025, 3:39 PM

#

gentle peak <@353593948523986945> i think i found a uacpi bug, if in `create_gpe_block` any ...

can u let me know if this fixes it:

--- a/source/io.c
+++ b/source/io.c
@@ -838,10 +838,6 @@ uacpi_status uacpi_map_gas_noalloc(
     if (ret != UACPI_STATUS_OK)
         return ret;

-    out_mapped->access_bit_width = access_bit_width;
-    out_mapped->total_bit_width = total_width;
-    out_mapped->bit_offset = gas->register_bit_offset;
-
     if (gas->address_space_id == UACPI_ADDRESS_SPACE_SYSTEM_MEMORY) {
         out_mapped->mapping = uacpi_kernel_map(gas->address, total_width / 8);
         if (uacpi_unlikely(out_mapped->mapping == UACPI_NULL))
@@ -860,6 +856,10 @@ uacpi_status uacpi_map_gas_noalloc(
         out_mapped->unmap = unmap_gas_io;
     }

+    out_mapped->access_bit_width = access_bit_width;
+    out_mapped->total_bit_width = total_width;
+    out_mapped->bit_offset = gas->register_bit_offset;
+
     return UACPI_STATUS_OK;
 }

gentle peak Feb 2, 2025, 3:44 PM

#

yeah, works now

fiery turtle Feb 2, 2025, 3:44 PM

#

cool thanks for being vigilant lol

gentle peak Feb 2, 2025, 3:45 PM

#

wasn't really a matter of being vigilant, i made a very small uacpi impl to test the performance of my rewritten kernel and ran into it

fiery turtle Feb 2, 2025, 3:45 PM

#

ah

fiery turtle Feb 2, 2025, 3:46 PM

#

gentle peak wasn't really a matter of being vigilant, i made a very small uacpi impl to test...

how did that go btw

gentle peak Feb 2, 2025, 3:46 PM

#

pretty much the same performance, which is what i expected and wanted, just wanted to make sure i didn't accidentally regress anything

fiery turtle Feb 2, 2025, 3:46 PM

#

is that in kernel land still?

gentle peak Feb 2, 2025, 3:47 PM

#

yes

fiery turtle Feb 2, 2025, 3:47 PM

#

ah

#

are you keeping it this way?

gentle peak Feb 2, 2025, 3:47 PM

#

no - this is temporary and small, just to test performance of basic kernel stuff. the only reason i did it was to check i didn't introduce any regressions compared to the prev rewrite

fiery turtle Feb 2, 2025, 3:47 PM

#

makes sense

dense steppe Feb 2, 2025, 4:27 PM

#

@fiery turtle is there a way to replace e.g. uacpi_kernel_alloc with my own definition, so that I can e.g. inline some stuff if __builtin_constant_p (size)?

gentle peak Feb 2, 2025, 4:28 PM

#

lto

dense steppe Feb 2, 2025, 4:28 PM

#

lto is kinda "meh" in this situation

mortal yoke Feb 2, 2025, 4:29 PM

#

put the definition in uacpi_libc.h

#

its included in like all uacpi sources

gentle peak Feb 2, 2025, 4:30 PM

#

that doesnt really work because kernel_api.h still defines the prototype

dense steppe Feb 2, 2025, 4:30 PM

#

yep.

mortal yoke Feb 2, 2025, 4:30 PM

#

how does it not work? the compiler will then see the definition

dense steppe Feb 2, 2025, 4:30 PM

#

because multiple definitions of a global symbol

mortal yoke Feb 2, 2025, 4:31 PM

#

you make it inline?

dense steppe Feb 2, 2025, 4:31 PM

#

cannot change the definition in uacpi/kernel_api.h though

mortal yoke Feb 2, 2025, 4:31 PM

#

I'd assume its enough if the definition is marked inline

dense steppe Feb 2, 2025, 4:32 PM

#

basically this is my kmalloc:

mortal yoke Feb 2, 2025, 4:35 PM

#

I think it would work if you'd just make it normal inline inside uacpi_libc.h (though then you need to also instantiate it once in some source file because the way inline works in c is stupid)

gentle peak Feb 2, 2025, 4:35 PM

#

if you make it normal inline you'll get multiple definitions

hollow elm Feb 2, 2025, 4:36 PM

#

in c++ you'd get sane inline behavior :^)

dense steppe Feb 2, 2025, 4:37 PM

#

mortal yoke I think it would work if you'd just make it normal inline inside uacpi_libc.h (t...

nope... does not work

#

gentle peak Feb 2, 2025, 4:38 PM

#

dense steppe basically this is my kmalloc:

if you can get this definition to be included before kernel_api.h it'll work

#

you do have to use static in that case

#

if it's included afterwards, you get a compiler error

mortal yoke Feb 2, 2025, 4:40 PM

#

ah yeah

gentle peak Feb 2, 2025, 4:41 PM

#

but like

dense steppe Feb 2, 2025, 4:41 PM

#

it's the world biggest hack but it works

gentle peak Feb 2, 2025, 4:41 PM

#

why isn't lto sufficient for this

dense steppe Feb 2, 2025, 4:41 PM

#

but I won't use it, because it's too hacky

mortal yoke Feb 2, 2025, 4:41 PM

#

how is it hacky

dense steppe Feb 2, 2025, 4:41 PM

#

gentle peak why isn't lto sufficient for this

my build system is... weird

gentle peak Feb 2, 2025, 4:41 PM

#

ah

dense steppe Feb 2, 2025, 4:42 PM

#

I guess I can just

void *
uacpi_kernel_alloc (uacpi_size size)
{
        if (size == 32)
                return kmalloc (32);
        else
                return kmalloc (size);
}

so it will slightly optimise the common-case of allocating 32-byte objects

#

but there are bigger, more important performance improvements to do

hollow elm Feb 2, 2025, 4:46 PM

#

maybe also add likely there?

dense steppe Feb 2, 2025, 4:48 PM

#

nahh

#

the branch predictor is sufficient

#

also I don't want to move that other kmalloc to a cold section

gentle peak Feb 2, 2025, 4:49 PM

#

__builtin_expect actually makes a pretty huge difference in performance

#

i forgot the exact numbers but it's very much not insignificant

dense steppe Feb 2, 2025, 4:49 PM

#

dense steppe I guess I can just ```c void * uacpi_kernel_alloc (uacpi_size size) { if...

but anyways, I won't do this. Instead I will work on CPU-local object bins for slab

#

which requires runtime allocation of CPU-local storage, the way I plan on doing it.

#

the thing is, since slab uses this dynamic runtime allocation of CPU-local storage, the implementation of dynamic runtime CPU-local storage allocation cannot use slab

#

though that's not a problem (just don't support free and it's fine)

hallow marten Feb 2, 2025, 4:52 PM

#

solaris fixes this by allocating cpu-local storage using vmem

#

I'll probably do something similar

dense steppe Feb 2, 2025, 4:52 PM

#

rustc (vmem shill)

hallow marten Feb 2, 2025, 4:54 PM

#

the solaris allocation stack is very well made

#

there's a reason why everyone stole it meme

frank canopy Feb 2, 2025, 4:55 PM

#

dense steppe also I don't want to move that other `kmalloc` to a cold section

theres different branch preditiction stuff for cold vs just unlikely

dense steppe Feb 2, 2025, 4:56 PM

#

for now I'm just going to bump-allocate from a 16KiB or so statically allocated region. later I might reserve like 1MiB virtual address space and only actually populate it with physical pages when needed

#

(16KiB/CPU)

dense steppe Feb 2, 2025, 5:25 PM

#

this is what the fastpath of my new slab allocator implementation will look like

#

sadly, it loses the ability to allocate memory in IRQ context

#

I can have a separate IRQ-safe allocator

vast kestrel Feb 2, 2025, 5:26 PM

#

Don't you want to preempt_on even if num is zero

dense steppe Feb 2, 2025, 5:26 PM

#

yeah... but that's done later

#

I should probably builtin_expect num to not be zero

dense steppe Feb 2, 2025, 5:27 PM

#

vast kestrel Don't you want to preempt_on even if num is zero

but yeah... good point

#

I moved the initial preempt_off () to outside of the first branch

#

won't affect perf

#

will make code simpler

#

👍

#

ok... the fastpath for allocation should be like 20 instructions or so

#

but I gotta write the slowpath too

leaden fox Feb 2, 2025, 7:28 PM

#

slim panther haiku pr when

Might try porting uacpi

#

They have their own fork of ACPICA basically

#

No clue how much they changed

north holly Feb 2, 2025, 7:29 PM

#

port uacpi to the apollo 10

#

space craft

fiery turtle Feb 2, 2025, 7:39 PM

#

leaden fox They have their own fork of ACPICA basically

thats every os

#

because u cant integrate it without forking

#

u have to change headers and stuff i think

leaden fox Feb 2, 2025, 7:40 PM

#

Didn't know ACPICA was that bad

fiery turtle Feb 2, 2025, 7:40 PM

#

lol

#

its just a different design decision

#

and it was made for linux ig

leaden fox Feb 2, 2025, 7:41 PM

#

fiery turtle and it was made for linux ig

Which Linux forked for themselves?

fiery turtle Feb 2, 2025, 7:41 PM

#

sorta

#

acpica has native linux support scripts inside the repo

#

and headers as well

#

you can kind of upstream support with your headers if your os is popular

slim panther Feb 2, 2025, 7:50 PM

#

fiery turtle because u cant integrate it without forking

lol

calm latch Feb 2, 2025, 8:14 PM

#

👀

📎 pmos-hyper.img

#

(could someone with a fast CPU test it?)

gentle peak Feb 2, 2025, 8:20 PM

#

1.35M on my 5800x

#

using qemu-system-x86_64 -accel kvm -cpu host,migratable=off -M q35,smm=off -drive format=raw,file=pmos-hyper.img (yes i know it's i386)

#

do you have an x86_64 image for comparison?

calm latch Feb 2, 2025, 8:24 PM

#

no

#

i mean yes

#

i need to fix it

left orbit Feb 2, 2025, 8:25 PM

#

on my system i get this

#

i think this one is a hard outlier meme

gentle peak Feb 2, 2025, 8:26 PM

#

what the fuck that's faster than proxima and it's running in userspace

left orbit Feb 2, 2025, 8:26 PM

#

hold on

fiery turtle Feb 2, 2025, 8:26 PM

#

pmos i386 is now the top of the leaderboard lmfao

gentle peak Feb 2, 2025, 8:26 PM

#

actually it probably wouldn't be faster than proxima if proxima didn't suffer from the intel lock cmpxchg shit

calm latch Feb 2, 2025, 8:27 PM

#

i need to add fast (or rather not insanely slow) ipc stuff to x86_64 and check if it gets the same performance...

left orbit Feb 2, 2025, 8:28 PM

#

it varies wildly but i think it's due to wsl, mostly consistent around 5.5-6M

calm latch Feb 2, 2025, 8:28 PM

#

like to see if this performance is credible

gentle peak Feb 2, 2025, 8:28 PM

#

i need to resume work on sirius (my i386 os) damn

fiery turtle Feb 2, 2025, 8:29 PM

#

is this speedup really due to this being i386

#

like cache locality and stuff

gentle peak Feb 2, 2025, 8:29 PM

#

probably not

calm latch Feb 2, 2025, 8:29 PM

#

calm latch i need to add fast (or rather not insanely slow) ipc stuff to x86_64 and check i...

possibly due to this

fiery turtle Feb 2, 2025, 8:30 PM

#

proxima doesnt use ipc at all and it gets less

calm latch Feb 2, 2025, 8:30 PM

#

and also temp mapping are 2 tlb misses instead of 4

#

hold on let me fix x86_64 build

#

(I don't even know if it compiles)

#

Also I have some "temporary" chatgpt generated code in timers, so they could be broken

left orbit Feb 2, 2025, 8:31 PM

#

☠️

fiery turtle Feb 2, 2025, 8:32 PM

#

inb4 this is actually 800K

calm latch Feb 2, 2025, 8:33 PM

#

It was giving about the same score as 64 bit build on my PC though

#

(about 1.3M on 5900X)

gentle peak Feb 2, 2025, 8:34 PM

#

i'm looking at your malloc source code and it looks like you're using spinlocks for malloc

#

which could be where the speedup comes from, since spinlocks generally don't suffer from the intel cmpxchg shit

calm latch Feb 2, 2025, 8:34 PM

#

i'm using spinlocks everywhere

fiery turtle Feb 2, 2025, 8:35 PM

#

gentle peak which could be where the speedup comes from, since spinlocks generally don't suf...

wdym?

calm latch Feb 2, 2025, 8:35 PM

#

left orbit hold on

But this is Intel

fiery turtle Feb 2, 2025, 8:35 PM

#

dont spinlocks need cmpxchg as well

calm latch Feb 2, 2025, 8:35 PM

#

It's slow on AMD

gentle peak Feb 2, 2025, 8:35 PM

#

fiery turtle wdym?

spinlocks can use xchg instead of cmpxchg

calm latch Feb 2, 2025, 8:35 PM

#

oh

fiery turtle Feb 2, 2025, 8:36 PM

#

gentle peak spinlocks can use xchg instead of cmpxchg

hm how

gentle peak Feb 2, 2025, 8:36 PM

#

while (!xchg(&lock->state, true)) pause();

fiery turtle Feb 2, 2025, 8:36 PM

#

oh

gentle peak Feb 2, 2025, 8:37 PM

#

also i have no idea why the difference between intel and amd is so huge there

#

like sure iretq's cpu is quite a bit better than ryzen 5xxx but it's not 7 times better

calm latch Feb 2, 2025, 8:37 PM

#

some memory memes?

#

what if I use hhdm? thonk

#

(anyway, I'm trying to build the 64 bit image)

left orbit Feb 2, 2025, 8:40 PM

#

how can you not use cmpxchg for spinlocks?

#

what difference does that make

fiery turtle Feb 2, 2025, 8:40 PM

#

gentle peak ```c while (!xchg(&lock->state, true)) pause(); ```

.

left orbit Feb 2, 2025, 8:40 PM

#

you still need to know if it was previously locked or not

calm latch Feb 2, 2025, 8:40 PM

#

cache invalidation?

gentle peak Feb 2, 2025, 8:40 PM

#

whatever intel's doing with cache coherence for lock cmpxchg absolutely destroys performance

left orbit Feb 2, 2025, 8:40 PM

#

cmpxchg does the compare for you

fiery turtle Feb 2, 2025, 8:41 PM

#

left orbit you still need to know if it was previously locked or not

xchg returns the old value

gentle peak Feb 2, 2025, 8:41 PM

#

and it doesn't do the same for xchg

left orbit Feb 2, 2025, 8:41 PM

#

cmpxchg compares and then exchanges them, so you dont get the old value but you get the flags

#

also i dont see why you couldn't use xchg for whatever else you want to implement

#

and i think that the slow down might come from the lock prefix, but i am unsure lol

gentle peak Feb 2, 2025, 8:42 PM

#

xchg has implied lock

#

you're right that it comes from the prefix (normal cmpxchg doesn't have the problem) but xchg with a memory operand implies lock so

frank canopy Feb 2, 2025, 8:43 PM

#

yep

#

xchg always has a lock per the docs

dense steppe Feb 2, 2025, 8:44 PM

#

This is my memory allocator fastpath now. I haven't looked at the assembly for the old thing, but I'd assume it is a fair bit worse than this.

Anyway, this makes basically no difference in the benchmark 😭

frank canopy Feb 2, 2025, 8:44 PM

#

and yeah xchg true works for a spinlock because if its already true then the swap has no semantic effect

left orbit Feb 2, 2025, 8:44 PM

#

gentle peak xchg has implied lock

ah interesting

frank canopy Feb 2, 2025, 8:45 PM

#

i use bit test things for mine because i want my spinlocks to be true 1-bit so i can have them ORd with an align > 1 pointer

calm latch Feb 2, 2025, 8:58 PM

#

64 bit image

📎 pmos.img

dense steppe Feb 2, 2025, 8:59 PM

#

:p

calm latch Feb 2, 2025, 9:00 PM

#

calm latch 64 bit image

I can also make one with 32 bit uacpi server and 64 bit kernel

gentle peak Feb 2, 2025, 9:00 PM

#

that'd be interesting to compare

#

that one gives me the same score as the 32 bit image so i doubt it has much to do with bitness

dense steppe Feb 2, 2025, 9:01 PM

#

calm latch 64 bit image

#

world record

calm latch Feb 2, 2025, 9:01 PM

#

do you have tsc on?

dense steppe Feb 2, 2025, 9:01 PM

#

oh. forgot -cpu host,migratable=off

#

now it is more reasonable

#

sometimes getting speeds around 3.6M, sometimes 1.4M

#

host CPU scheduling skill issue

calm latch Feb 2, 2025, 9:04 PM

#

Which CPU is it?

calm latch Feb 2, 2025, 9:05 PM

#

calm latch I can also make one with 32 bit uacpi server and 64 bit kernel

(I need to figure out how to extract the server out of the prevous image meme )

hallow marten Feb 2, 2025, 9:05 PM

#

mount the iso and copy the files

calm latch Feb 2, 2025, 9:06 PM

#

windows is refusing to mount it

#

it's just a fat32 partition

gentle peak Feb 2, 2025, 9:08 PM

#

which one is it?

calm latch Feb 2, 2025, 9:08 PM

#

devicesd.elf

gentle peak Feb 2, 2025, 9:08 PM

#

📎 devicesd.elf

calm latch Feb 2, 2025, 9:08 PM

#

thx

gentle peak Feb 2, 2025, 9:08 PM

#

also idk if you're already aware but the 2nd partition on that image (the ext2 one) is malformed

#

[277050.930059] EXT4-fs (loop0p2): mounting ext2 file system using the ext4 subsystem
[277050.930062] EXT4-fs (loop0p2): bad geometry: block count 16384 exceeds size of device (16256 blocks)```

calm latch Feb 2, 2025, 9:09 PM

#

i was not, but I didn't get to write the filesystems stuff yet

dense steppe Feb 2, 2025, 9:10 PM

#

calm latch Which CPU is it?

ryzen 9 7950x3d

#

apparently half the cores (the ones with extra L3) run at a lower clock speed

#

but also my host runs some CPU cores at lower clock speeds when they are not under load, I think

#

This is a major source of Davix uACPI benchmark skill issues

left orbit Feb 2, 2025, 9:12 PM

#

you can use taskset to pin a task to a certain cpu

#

BUT idk if that works with qemu kvm

dense steppe Feb 2, 2025, 9:12 PM

#

yeah

calm latch Feb 2, 2025, 9:12 PM

#

gentle peak

Somehow the kernel is having a skill issue loading it

fiery turtle Feb 2, 2025, 9:12 PM

#

left orbit BUT idk if that works with qemu kvm

it does

dense steppe Feb 2, 2025, 9:12 PM

#

I should probably setup libvirt

fiery turtle Feb 2, 2025, 9:12 PM

#

a vpcu is just a thread

#

*vcpu

dense steppe Feb 2, 2025, 9:12 PM

#

yeah

calm latch Feb 2, 2025, 9:12 PM

#

calm latch Somehow the kernel is having a skill issue loading it

(let me fix it)

left orbit Feb 2, 2025, 9:12 PM

#

fiery turtle a vpcu is just a thread

yeah that's what i thought but i wasn't sure

dense steppe Feb 2, 2025, 9:12 PM

#

a vcpu is just a host thread that calls ioctl (kvmfd, KVM_RUN) or something

fiery turtle Feb 2, 2025, 9:12 PM

#

calm latch Somehow the kernel is having a skill issue loading it

your 64-bit kernel has compatibility mode support?

#

thats pretty cool

fiery turtle Feb 2, 2025, 9:13 PM

#

dense steppe a vcpu is just a host thread that calls `ioctl (kvmfd, KVM_RUN)` or something

IOCTL_KVM_VCPU_RUN or something

left orbit Feb 2, 2025, 9:13 PM

#

close enough

calm latch Feb 2, 2025, 9:13 PM

#

fiery turtle your 64-bit kernel has compatibility mode support?

I mean I've ported userspace to 32 bits a month or so ago

calm latch Feb 2, 2025, 9:14 PM

#

calm latch 64 bit image

@left orbit can you try it?

left orbit Feb 2, 2025, 9:14 PM

#

sure 1s

#

that an hdd image too?

calm latch Feb 2, 2025, 9:14 PM

#

yes

fiery turtle Feb 2, 2025, 9:15 PM

#

calm latch I mean I've ported userspace to 32 bits a month or so ago

do u have like a compat syscall layer as well

calm latch Feb 2, 2025, 9:15 PM

#

No

left orbit Feb 2, 2025, 9:15 PM

#

i get this pretty consistently

calm latch Feb 2, 2025, 9:15 PM

#

calm latch No

Or yes

#

Idk what does that mean

fiery turtle Feb 2, 2025, 9:15 PM

#

how do u make sure u dont truncate anything when returning stuff to a 32-bit program

calm latch Feb 2, 2025, 9:15 PM

#

The kernel knows that the process is 32 bits and expects arguments to be in different registers

calm latch Feb 2, 2025, 9:16 PM

#

fiery turtle how do u make sure u dont truncate anything when returning stuff to a 32-bit pro...

My syscalls return 2 64 bits integers, one for result and one for whatever was requested

#

through registers

#

that fits into 32 bits registers as well

fiery turtle Feb 2, 2025, 9:17 PM

#

yes but lets say it does an mmap

gentle peak Feb 2, 2025, 9:17 PM

#

do you also do stuff like making sure pointers stay within 4G?

fiery turtle Feb 2, 2025, 9:17 PM

#

u have to make sure to have an upper bound at 4g

calm latch Feb 2, 2025, 9:17 PM

#

gentle peak do you also do stuff like making sure pointers stay within 4G?

yes

gentle peak Feb 2, 2025, 9:17 PM

#

damn nice

fiery turtle Feb 2, 2025, 9:17 PM

#

yeah

calm latch Feb 2, 2025, 9:17 PM

#

I have a function which checks the limit

#

(Which also changes for 3/4/5 paging levels with 64 bit stuff)

#

(on risc-v, since the x86_64 paging code is a mess and doesn't support that)

#

Like I do 3/1 split on 32 bit kernel, and full 4GB is available with 64 bit one

dense steppe Feb 2, 2025, 9:19 PM

#

anyway I wrote some code to collect slab_alloc statistics, and here are the results. notably nr_expensive_free is very high.

#

I should totally write some code to only actually free objects in a background worker thread or something

#

I seem to be getting relatively consistent results in the 2.2M-2.3M range (so long as the guest is scheduled on the correct host CPU cores)

calm latch Feb 2, 2025, 9:20 PM

#

(also this is the full build, which doesn't exit(1) after loading acpi namespace if someone's interested)

#

this one

📎 pmos-hyper_0.img

gentle peak Feb 2, 2025, 9:23 PM

#

the PS2 driver seems nonoptimal

calm latch Feb 2, 2025, 9:23 PM

#

idk interrupts are half-broken

gentle peak Feb 2, 2025, 9:24 PM

#

i get 1.8M on that build

fiery turtle Feb 2, 2025, 9:25 PM

#

maybe if u use spinlocks in your allocator youd speed proxima up on iretq's cpu as well

left orbit Feb 2, 2025, 9:26 PM

#

you must optimize for intel

gentle peak Feb 2, 2025, 9:26 PM

#

oh definitely

#

here's a main branch build with spinlock heap, and its results on my cpu:

uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 7933442/s)

📎 proxima.iso

#

oh that's with cmpxchg spinlocks so not even the best it can get

left orbit Feb 2, 2025, 9:28 PM

#

7.9M is crazy

gentle peak Feb 2, 2025, 9:29 PM

#

interestingly using just plain old xchg for spinlocks destroys perf on AMD

#

i wonder how it'd do on intel

📎 proxima.iso

left orbit Feb 2, 2025, 9:30 PM

#

lol classic

#

proxima malware

gentle peak Feb 2, 2025, 9:30 PM

#

drat

dense steppe Feb 2, 2025, 9:30 PM

#

jesus fkn christ I'm trying to upload an image of my kernel for you guys to test but am having severe network skill issue or something

#

it takes ages to upload a 10.27MB file

left orbit Feb 2, 2025, 9:31 PM

#

WOW

#

WINDOWS SECURITY

#

AMAZING

#

BLACK BOX

#

I LOVE WINDOWS

dense steppe Feb 2, 2025, 9:31 PM

#

can someone with less host cpu skill issues than me run this? qemu-system-x86_64 -cdrom bootable.iso -enable-kvm -cpu host,migratable=off -M q35 -debugcon stdio

📎 davix.iso.gz

#

finally it fkin uploaded

left orbit Feb 2, 2025, 9:32 PM

#

gentle peak i wonder how it'd do on intel

averages around 2.6M

#

weird

gentle peak Feb 2, 2025, 9:32 PM

#

ah

#

how about the cmpxchg one? (the first one)

left orbit Feb 2, 2025, 9:32 PM

#

WHA

#

WHY IS EVERYTHING A VIRUS TODAY

calm latch Feb 2, 2025, 9:32 PM

#

trl

gentle peak Feb 2, 2025, 9:32 PM

#

dense steppe can someone with less host cpu skill issues than me run this? `qemu-system-x86_6...

dense steppe Feb 2, 2025, 9:33 PM

#

gentle peak interestingly using just plain old `xchg` for spinlocks *destroys* perf on AMD

so this would be bad then?

/**
 * Optimistically try to acquire the lock with an atomic exchange. Returns true
 * for success.
 */
static inline bool
spin_lock_fastpath (spinlock_t *lock)
{
        return atomic_exchange (&lock->value, 1, memory_order_acquire) == 0;
}

left orbit Feb 2, 2025, 9:33 PM

#

gentle peak how about the cmpxchg one? (the first one)

7.1M

gentle peak Feb 2, 2025, 9:33 PM

#

apparently yeah

dense steppe Feb 2, 2025, 9:33 PM

#

gentle peak

wtf how do you have more host cpu skill issues than me?

gentle peak Feb 2, 2025, 9:33 PM

#

left orbit 7.1M

ok what the fuck was i wrong about the fucking cmpxchg shit

dense steppe Feb 2, 2025, 9:33 PM

#

dense steppe so this would be bad then? ```c /** * Optimistically try to acquire the lock wi...

so what's preferable over this then?

gentle peak Feb 2, 2025, 9:34 PM

#

normal cmpxchg, apparently?

dense steppe Feb 2, 2025, 9:34 PM

#

slowpath for completeness:

/**
 * Slowpath for spinlock acquisition. Implements TTAS semantics.
 */
static inline void
spin_lock_slowpath (spinlock_t *lock)
{
        for (;;) {
                do
                        arch_relax ();
                while (atomic_load_relaxed (&lock->value));

                if (spin_lock_fastpath (lock))
                        break;
        }
}

left orbit Feb 2, 2025, 9:35 PM

#

dense steppe can someone with less host cpu skill issues than me run this? `qemu-system-x86_6...

around 6.3M

dense steppe Feb 2, 2025, 9:35 PM

#

oh btw my allocator is lock-free in the fast-case (nr_quickbin_hit and nr_quickbin_free)

dense steppe Feb 2, 2025, 9:35 PM

#

left orbit around 6.3M

what CPU?

left orbit Feb 2, 2025, 9:35 PM

#

[    0.128253] uACPI: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6364623/s)
[    0.127614] uACPI: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6361916/s)```

#

13700k

gentle peak Feb 2, 2025, 9:35 PM

#

dense steppe wtf how do you have more host cpu skill issues than me?

so the pci errors are intended?

dense steppe Feb 2, 2025, 9:35 PM

#

gentle peak so the pci errors are intended?

yes. the PCI funcs aren't implemented yet lol

gentle peak Feb 2, 2025, 9:35 PM

#

i thought that was an indicator that something was wrong which is why i included them

#

ah

dense steppe Feb 2, 2025, 9:35 PM

#

but it doesn't matter since it doesn't execute any PCI stuff for the benchmark

dense steppe Feb 2, 2025, 9:36 PM

#

left orbit ``` [ 0.128253] uACPI: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6...

cc @fiery turtle

left orbit Feb 2, 2025, 9:36 PM

#

[    0.128499] uACPI: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6210795/s)
[    0.129090] uACPI: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6347728/s)
[    0.128887] uACPI: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6377479/s)
[    0.127661] uACPI: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6365027/s)```

dense steppe Feb 2, 2025, 9:36 PM

#

troll2

left orbit Feb 2, 2025, 9:36 PM

#

pretty much consistent except for wsl mishaps as usual

#

but i filter those out

fiery turtle Feb 2, 2025, 9:36 PM

#

left orbit ``` [ 0.128253] uACPI: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6...

#

bruh everyone is getting 7mil these days

dense steppe Feb 2, 2025, 9:36 PM

#

yeah lol

fiery turtle Feb 2, 2025, 9:37 PM

#

i remember when 400k was a lot

dense steppe Feb 2, 2025, 9:37 PM

#

except on my CPU

calm latch Feb 2, 2025, 9:37 PM

#

left orbit hold on

So this is like normal now?

left orbit Feb 2, 2025, 9:37 PM

#

oh i don't think so ☠️

#

i haven't managed to replicate that

#

tbh i haven't seen wsl issues affect the score positively

#

so that must have been a fluke on your side

#

idk how that could have happened though

gentle peak Feb 2, 2025, 9:38 PM

#

@left orbit as a sanity check, could you test these two? the first is the latest non-rewrite CI build, and the second is the exact same except built on my machine

📎 proxima-ci.iso 📎 proxima.iso

left orbit Feb 2, 2025, 9:39 PM

#

man i just wish it wasn't a frickin virus 😢

#

let me just wget in wsl

gentle peak Feb 2, 2025, 9:39 PM

#

they both get ~7.5M/s on my machine

gentle peak Feb 2, 2025, 9:39 PM

#

left orbit man i just wish it wasn't a frickin virus 😢

jokes on you they are actually payloads that escape the hypervisor and run a crypto miner in the nt kernel

#

anyway yeah i wish windows's "this might be a virus" notifications were based on more than just "oh this hash hasn't been downloaded a lot"

left orbit Feb 2, 2025, 9:41 PM

#

they are though

dense steppe Feb 2, 2025, 9:41 PM

#

this doesn't make much of a difference for me

#

(Ryzen 9 7950X3D)

#

(I also added builtin_expect to __spin_lock)

left orbit Feb 2, 2025, 9:42 PM

#

anyway, both isos get me around 5.3-5.5M

gentle peak Feb 2, 2025, 9:42 PM

#

what the fuck

#

does intel have some sort of special optimization for cmpxchg-store

#

none of these mutexes are contested, the only difference should be 1 cmpxchg 1 store vs 2 cmpxchg

#

last one before i give up looking for this issue: this one uses acquire for lock and release for unlock instead of acq-rel for both

📎 proxima.iso

#

i seriously doubt that'll do anything but it's the only thing i can think of

left orbit Feb 2, 2025, 9:46 PM

#

this is like +50-100K

#

maybe less not very consistent

gentle peak Feb 2, 2025, 9:46 PM

#

as i thought

#

i have one more thing to try but it's gonna take a bit to implement

#

let's see how this worked again

left orbit Feb 2, 2025, 9:46 PM

#

holy

dense steppe Feb 2, 2025, 9:47 PM

#

I haven't looked that much at the paper

#

but what I implemented with quickbins today is similar to how I imagine magazines to be

fiery turtle Feb 2, 2025, 9:47 PM

#

gentle peak let's see how this worked again

LETSFUCKINGGOOOOOO

#

10M here we come

dense steppe Feb 2, 2025, 9:47 PM

#

1020M when

#

trl

gentle peak Feb 2, 2025, 9:48 PM

#

i think i'm gonna do per-thread magazines because my {enable,disable}_preempt is slow and it's not like i'm gonna keep this around when i'm already in the middle of a rewrite

left orbit Feb 2, 2025, 9:48 PM

#

per thread?

#

wtf

fiery turtle Feb 2, 2025, 9:48 PM

#

gentle peak let's see how this worked again

is this for the new kernel?

dense steppe Feb 2, 2025, 9:48 PM

#

how would you allocate like the lists for storing objects?

gentle peak Feb 2, 2025, 9:48 PM

#

i'll probably implement it in the new kernel as well but rn i'm working on the old one just trying to see how high i could get

dense steppe Feb 2, 2025, 9:48 PM

#

just statically in struct task?

gentle peak Feb 2, 2025, 9:48 PM

#

dense steppe how would you allocate like the lists for storing objects?

extra field of fixed size in struct task

frank canopy Feb 2, 2025, 9:49 PM

#

gentle peak normal cmpxchg, apparently?

reminds me that zig has separate weak and strong cmpxchg builtins, wonder if thats the same difference

dense steppe Feb 2, 2025, 9:49 PM

#

yep. ok

gentle peak Feb 2, 2025, 9:49 PM

#

frank canopy reminds me that zig has separate weak and strong cmpxchg builtins, wonder if tha...

no, weak is allowed to spuriously fail, on x86 they're the exact same

frank canopy Feb 2, 2025, 9:49 PM

#

ah ok

gentle peak Feb 2, 2025, 9:49 PM

#

unlocked cmpxchg is completely broken on any multiprocessor system

#

because it's not atomic

frank canopy Feb 2, 2025, 9:49 PM

#

ah

dense steppe Feb 2, 2025, 9:50 PM

#

gentle peak i think i'm gonna do per-thread magazines because my {enable,disable}_preempt is...

what do you do for {enable,disable}_preempt?

fiery turtle Feb 2, 2025, 9:50 PM

#

dense steppe ~~10~~20M when

i could hit 20M with computed gotos in uacpi interpreter potentially, but thing is those are GCC specific i think and they require everything to be in a switch, which would destroy stack usage in debug mode

#

maybe some smart object caching could help, but i doubt it would win over a ~100ns allocator

dense steppe Feb 2, 2025, 9:50 PM

#

wtf are computed gotos? I've probably heard of them but with a different name

gentle peak Feb 2, 2025, 9:51 PM

#

dense steppe what do you do for `{enable,disable}_preempt`?

the new kernel just increments and decrements, and when it reaches zero does a quick lockless check if it needs to reschedule
the old kernel also just increments or decrements, but when it reaches zero it takes a spinlock

frank canopy Feb 2, 2025, 9:51 PM

#

dense steppe wtf are computed gotos? I've probably heard of them but with a different name

goto a switch with new value to switch on

dense steppe Feb 2, 2025, 9:51 PM

#

fiery turtle maybe some smart object caching could help, but i doubt it would win over a ~100...

I expect the fast-case of my allocator to be like 10-20ns tbh. But I should measure

frank canopy Feb 2, 2025, 9:51 PM

#

called that because its an asm jump to a switch case

#

good for interpreter loops

fiery turtle Feb 2, 2025, 9:51 PM

#

dense steppe I expect the fast-case of my allocator to be like 10-20ns tbh. But I should mea...

damn

gentle peak Feb 2, 2025, 9:51 PM

#

frank canopy ah

cmpxchg itself is atomic as in a single instruction, but not atomic as in other processors and/or devices only see either the full outcome or none of it

dense steppe Feb 2, 2025, 9:51 PM

#

fiery turtle damn

but like

gentle peak Feb 2, 2025, 9:51 PM

#

cmpxchg on multi-processor systems can spuriously succeed

frank canopy Feb 2, 2025, 9:51 PM

#

frank canopy good for interpreter loops

zig's compiler got faster when they added support for them and switched over to them in the compiler

dense steppe Feb 2, 2025, 9:52 PM

#

frank canopy Feb 2, 2025, 9:52 PM

#

iirc

dense steppe Feb 2, 2025, 9:52 PM

#

quite a lot of nr_expensive_free

#

and those take a lock

fiery turtle Feb 2, 2025, 9:52 PM

#

i like these stats u added

#

always appreciate detailed stuff like that

dense steppe Feb 2, 2025, 9:52 PM

#

I just wanted to know how often the fastpath is hit

slim panther Feb 2, 2025, 9:53 PM

#

oh wait

#

@fiery turtle does uacpi_kernel_io_map map unaligned addresses?

fiery turtle Feb 2, 2025, 9:53 PM

#

it's an io address

#

like port io

gentle peak Feb 2, 2025, 9:53 PM

#

trl

dense steppe Feb 2, 2025, 9:54 PM

#

the slowpath steals a page from the slab->pg_full_list and slab->pg_partial_list if there any one. otherwise it allocates one. then it populates the entire magazine and returns one object

dense steppe Feb 2, 2025, 9:54 PM

#

gentle peak <:trl:829316641191886888>

that is just unreasonably large

#

well

#

it depends on how many slabs you have

gentle peak Feb 2, 2025, 9:54 PM

#

i know

#

oh this is the number of objects in the magazine per slab

dense steppe Feb 2, 2025, 9:54 PM

#

yeah

#

I have something similar

#

but I set it to min(31, objs_per_page - 1)

#

for reasons

gentle peak Feb 2, 2025, 9:54 PM

#

so in total it's 3072 rounds

fiery turtle Feb 2, 2025, 9:55 PM

#

thats a minigun magazine

dense steppe Feb 2, 2025, 9:55 PM

#

https://github.com/dbstream/davix/blob/master/mm/pfn_slab.c here is my slab allocator if you want to take a look at it

GitHub

davix/mm/pfn_slab.c at master · dbstream/davix

A hobby operating system kernel targeting x86_64. Contribute to dbstream/davix development by creating an account on GitHub.

#

hopefully it isn't too unreadable

dense steppe Feb 2, 2025, 9:56 PM

#

dense steppe

the slow case (nr_quickbin_miss and nr_expensive_free) is probably a lot slower than 20ns though

fiery turtle Feb 2, 2025, 9:57 PM

#

nooo yet another tabs user

dense steppe Feb 2, 2025, 9:57 PM

#

sorry :p

#

and also yet another tabs=8spaces user

#

but like wtf. why is my kernel getting like 6M on intel and 2M on AMD?

#

gotta do -mtune

calm latch Feb 2, 2025, 10:07 PM

#

calm latch Somehow the kernel is having a skill issue loading it

I've found the skill issue

gentle peak Feb 2, 2025, 10:11 PM

#

oh this is just my luck

#

with magazines it's slower because none of the init that runs before uacpi allocates objects from the same slab as uacpi mainly does

#

i can probably minimize the impact but it's certainly not going to be faster

calm latch Feb 2, 2025, 10:12 PM

#

mixed build

📎 pmos.img

#

(that is 32 bit uACPI/devices server and 64 bit everything else)

hallow marten Feb 2, 2025, 10:14 PM

#

gentle peak with magazines it's *slower* because none of the init that runs before uacpi all...

then your implementation is broken lol

#

for me magazines noticeably improved single-threaded performance

gentle peak Feb 2, 2025, 10:15 PM

#

it probably does once your allocator has been used a bit

#

but uacpi inits too early for that to be the case

hallow marten Feb 2, 2025, 10:16 PM

#

well then it's the same as allocating from a slab

#

then subsequent allocs will be from the cpu cache

gentle peak Feb 2, 2025, 10:17 PM

#

do you allocate an entire magazine if the magazine is empty? i'm going off of the vmem paper which just lets it fill up as objects get allocated

hallow marten Feb 2, 2025, 10:17 PM

#

I do whatever solaris does

dense steppe Feb 2, 2025, 10:18 PM

#

gentle peak with magazines it's *slower* because none of the init that runs before uacpi all...

wtf it took you like 25min to implement magazines

#

well tbh they aren't that hard

gentle peak Feb 2, 2025, 10:19 PM

#

my allocator is tiny it'd only take me like 10m to fully rewrite it

hallow marten Feb 2, 2025, 10:19 PM

#

dense steppe well tbh they aren't that hard

depends on how fancy you want them

gentle peak Feb 2, 2025, 10:19 PM

#

not that hard to add stuff onto

hallow marten Feb 2, 2025, 10:19 PM

#

solaris does fancy dynamic resizing

#

based on lock contention

dense steppe Feb 2, 2025, 10:20 PM

#

hallow marten solaris does fancy dynamic resizing

are there any measurements for if this improves performance?

#

(will not do it regardless, because it seems hard)

hallow marten Feb 2, 2025, 10:21 PM

#

dense steppe are there any measurements for if this improves performance?

the point of doing that is to ensure not too much memory is wasted for rarely used magazines and so that the almost-perfect size is used for regularly used magazines

#

so yea it likely improves performance

#

and memory usage

#

but this is for high scalability stuff, idk how useful it is here

dense steppe Feb 2, 2025, 10:25 PM

#

dense steppe gotta do `-mtune`

ok -mtune=native barely makes a difference

#

I get like 2.4M vs 2.3M when the guest is scheduled properly by the host

gentle peak Feb 2, 2025, 10:48 PM

#

ok i fully copied the strategy documented in the vmem paper and it did indeed improve performance

#

but not by much

#

it went from like 7.5M to 7.7M

#

if i init uacpi 10 times i get 9M on the final one

slim panther Feb 2, 2025, 10:57 PM

#

💀

calm latch Feb 2, 2025, 11:03 PM

#

vmem is cool though

fiery turtle Feb 2, 2025, 11:03 PM

#

gentle peak if i init uacpi 10 times i get 9M on the final one

LETSFUCKINGGOOOOOO

gentle peak Feb 2, 2025, 11:04 PM

#

@left orbit if you want to test proxima+magazines, here it is

📎 proxima.iso

#

if you want to see perf for the first of the eleven runs, -debugcon stdio

#

uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 9112238/s)
this is on my cpu

fiery turtle Feb 2, 2025, 11:05 PM

#

Insane

gentle peak Feb 2, 2025, 11:06 PM

#

the way i've implemented it (thread-local caches) is incredibly stupid if you want actual overall benefits, so i won't be pushing it to git and i won't be submitting it to the leaderboard

#

but it is technically leaderboard-valid so it's fun to see the performance

left orbit Feb 2, 2025, 11:06 PM

#

uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6390171/s)
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6718073/s)
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6768290/s)
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6726713/s)
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6456621/s)
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6670396/s)
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6754508/s)
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 6782695/s)```

north holly Feb 2, 2025, 11:07 PM

#

do you guys have nothing better to do for your OSes?

fiery turtle Feb 2, 2025, 11:07 PM

#

gentle peak the way i've implemented it (thread-local caches) is incredibly stupid if you wa...

How did you do it?

gentle peak Feb 2, 2025, 11:07 PM

#

damn, that's worse than i expected

torpid root Feb 2, 2025, 11:07 PM

#

north holly do you guys have nothing better to do for your OSes?

what else is there to osdev?

gentle peak Feb 2, 2025, 11:07 PM

#

fiery turtle How did you do it?

exactly the same as a normal magazines impl except the caches are thread-local instead of cpu-local

fiery turtle Feb 2, 2025, 11:07 PM

#

north holly do you guys have nothing better to do for your OSes?

Than magazines?

left orbit Feb 2, 2025, 11:07 PM

#

hey it's an improvement over what it scored on my pc previously lmao

gentle peak Feb 2, 2025, 11:07 PM

#

yeah but like

fiery turtle Feb 2, 2025, 11:07 PM

#

gentle peak exactly the same as a normal magazines impl except the caches are thread-local i...

Lol

gentle peak Feb 2, 2025, 11:07 PM

#

the later runs should take almost 0 locks in the allocator

left orbit Feb 2, 2025, 11:08 PM

#

oh i just took the ones from debugcon

gentle peak Feb 2, 2025, 11:08 PM

#

they're the right ones

#

dw

left orbit Feb 2, 2025, 11:08 PM

#

ah

slim panther Feb 2, 2025, 11:08 PM

#

okay uh

left orbit Feb 2, 2025, 11:08 PM

#

well i mean i only get one run per boot so

gentle peak Feb 2, 2025, 11:08 PM

#

only reason i said to use debugcon is because the older runs would get wrapped off of the screen

slim panther Feb 2, 2025, 11:08 PM

#

@gentle peak you might want to see this

gentle peak Feb 2, 2025, 11:08 PM

#

oh that shouldn't happen

slim panther Feb 2, 2025, 11:08 PM

#

guess how fast it is on my machine

left orbit Feb 2, 2025, 11:08 PM

#

oh ☠️

#

thats why

gentle peak Feb 2, 2025, 11:09 PM

#

slim panther guess how fast it is on my machine

11M?

slim panther Feb 2, 2025, 11:09 PM

#

13

gentle peak Feb 2, 2025, 11:09 PM

#

DAMN

left orbit Feb 2, 2025, 11:09 PM

#

hooooly

north holly Feb 2, 2025, 11:09 PM

#

fiery turtle Than magazines?

no than maximizing uacpi score

slim panther Feb 2, 2025, 11:09 PM

#

fiery turtle Feb 2, 2025, 11:09 PM

#

What the hell

slim panther Feb 2, 2025, 11:09 PM

#

i thought i couldn't read correctly

gentle peak Feb 2, 2025, 11:09 PM

#

left orbit oh ☠️

fuuuck i should've built this with more debug info than just symbols

#

it doesn't even have line numbers

left orbit Feb 2, 2025, 11:10 PM

#

just look at the disassembly

#

ez

slim panther Feb 2, 2025, 11:10 PM

#

yea okay there is literally no beating that lmfao

fiery turtle Feb 2, 2025, 11:10 PM

#

Monkuous u should upstream this

gentle peak Feb 2, 2025, 11:10 PM

#

now i wonder, what if i combine this with illegal techniques?

slim panther Feb 2, 2025, 11:10 PM

#

meme

fiery turtle Feb 2, 2025, 11:10 PM

#

So no one can ever beat proxima

slim panther Feb 2, 2025, 11:10 PM

#

if i do userspace uacpi do i get to use sse

gentle peak Feb 2, 2025, 11:10 PM

#

yes

fiery turtle Feb 2, 2025, 11:10 PM

#

Yeah

slim panther Feb 2, 2025, 11:10 PM

#

😃

#

i think my current bottleneck is my PMM

#

it's a bitmap atm

fiery turtle Feb 2, 2025, 11:11 PM

#

💀

slim panther Feb 2, 2025, 11:11 PM

#

slim panther

slim panther Feb 2, 2025, 11:11 PM

#

fiery turtle 💀

is that bad

gentle peak Feb 2, 2025, 11:11 PM

#

gentle peak now i wonder, what if i combine this with illegal techniques?

actually this wouldn't do much because the most impactful illegal technique is stubbing mutexes and if you stub mutexes the base allocator is probably faster than the magazine'd one (because the benchmark is single-cpu)

#

i'm gonna try regardless

fiery turtle Feb 2, 2025, 11:12 PM

#

Shouldn't have pinned it since monkuous ain't submitting it

slim panther Feb 2, 2025, 11:12 PM

#

oh

#

removed

fiery turtle Feb 2, 2025, 11:13 PM

#

slim panther is that bad

Well

#

Kinda

gentle peak Feb 2, 2025, 11:13 PM

#

left orbit oh ☠️

i might be completely wrong on this but just looking at the disassembly this might be a uacpi bug actually, because it's lock cmpxchg with dest=0, which shouldn't be possible unless uacpi locks a mutex or spinlock whose kernel api create call returned null

slim panther Feb 2, 2025, 11:13 PM

#

i guess i should do a buddy allocator?

dense steppe Feb 2, 2025, 11:14 PM

#

either buddy

#

or just a linked list

#

both are fine

fiery turtle Feb 2, 2025, 11:14 PM

#

gentle peak i might be completely wrong on this but just looking at the disassembly this *mi...

Why would it randomly repro on qemu

gentle peak Feb 2, 2025, 11:14 PM

#

magazine memory usage idk

fiery turtle Feb 2, 2025, 11:14 PM

#

Hm

#

If it is I wanna know where

gentle peak Feb 2, 2025, 11:14 PM

#

it's in ensure_register_mapped.part.0.lto_priv.0 but i have no idea what that exactly means for the location

#

and as i said this build doesn't have line info

slim panther Feb 2, 2025, 11:15 PM

#

dense steppe or just a linked list

is that slower or faster than buddy?

gentle peak Feb 2, 2025, 11:15 PM

#

it might also be memory corruption, i only realized this just now for some reason

dense steppe Feb 2, 2025, 11:15 PM

#

slim panther is that slower or faster than buddy?

a buddy usually involves multiple linked lists

#

though it is possible to implement bitmap buddies, although those are weird

gentle peak Feb 2, 2025, 11:16 PM

#

@left orbit what qemu options did you use?

#

i wanna try and repro it locally

fiery turtle Feb 2, 2025, 11:16 PM

#

gentle peak it might also be memory corruption, i only realized this just now for some reaso...

Maybe

left orbit Feb 2, 2025, 11:17 PM

#

qemu-system-x86_64 -debugcon stdio -cpu host,migratable=off -M q35 -enable-kvm -cdrom "proxima3.iso" | grep '1705 ops'

fiery turtle Feb 2, 2025, 11:17 PM

#

I mean u did speedrun them basically lmao

left orbit Feb 2, 2025, 11:17 PM

#

gentle peak <@796462670995390524> what qemu options did you use?

^

gentle peak Feb 2, 2025, 11:17 PM

#

ty

#

oh yeah repro

#

that's nice

fiery turtle Feb 2, 2025, 11:17 PM

#

Nice

gentle peak Feb 2, 2025, 11:17 PM

#

i wonder what triggers it since that's basically the same as my command line

#

ok with -M q35,smm=off it works

#

that's very strange

fiery turtle Feb 2, 2025, 11:18 PM

#

💀

left orbit Feb 2, 2025, 11:18 PM

#

good to know we found the cause

dense steppe Feb 2, 2025, 11:18 PM

#

tbh I'm surprised that my allocator rewrite worked like the first time.

#

like

gentle peak Feb 2, 2025, 11:18 PM

#

if you want to test perf that's the flag to use

left orbit Feb 2, 2025, 11:18 PM

#

gaming rn

#

ill test later

gentle peak Feb 2, 2025, 11:18 PM

#

alr

dense steppe Feb 2, 2025, 11:19 PM

#

dense steppe tbh I'm surprised that my allocator rewrite worked like the first time.

"everything seems to be working. what is wrong?"

gentle peak Feb 2, 2025, 11:19 PM

#

now the question: is this reproducible in debug mode

#

it is not

#

fuck

fiery turtle Feb 2, 2025, 11:20 PM

#

Memory corruption nooo

#

Or ub ig

gentle peak Feb 2, 2025, 11:20 PM

#

or (and this is very heavy cope) the slightly inflated kernel binary size triggered the oom

fiery turtle Feb 2, 2025, 11:21 PM

#

Lol

gentle peak Feb 2, 2025, 11:21 PM

#

oh wait no i lied it is reproducible in debug mode

#

i just accidentally passed smm=off

fiery turtle Feb 2, 2025, 11:21 PM

#

LETSFUCKINGGOOOOOO

gentle peak Feb 2, 2025, 11:21 PM

#

not on TCG though unfortunately

#

still i can work with this

dense steppe Feb 2, 2025, 11:22 PM

#

what if you give the VM absurd amounts of memory? like qemu -m 48G?

gentle peak Feb 2, 2025, 11:22 PM

#

still happens so not an oom

fiery turtle Feb 2, 2025, 11:25 PM

#

Just a guess but

#

Smm off might force qemu into acpi mode by default

#

So uacpi won't try to switch back on reset

#

Which would skip that path

gentle peak Feb 2, 2025, 11:27 PM

#

hm

#

don't think that's the case

#

also i found a typo: uacpi_deininitialize_registers

#

deininit

fiery turtle Feb 2, 2025, 11:28 PM

#

Guess not

fiery turtle Feb 2, 2025, 11:28 PM

#

gentle peak also i found a typo: `uacpi_deininitialize_registers`

Lol

gentle peak Feb 2, 2025, 11:29 PM

#

oh wait

#

i put a hlt loop in uacpi_initialize_registers and it doesn't get called in the second iter

#

ininit*

#

oh wait it just doesn't get to the second iter

#

nvm

fiery turtle Feb 2, 2025, 11:30 PM

#

Tier?

#

Nvm

gentle peak Feb 2, 2025, 11:33 PM

#

yeah ok i just confirmed it is indeed a uacpi bug

fiery turtle Feb 2, 2025, 11:33 PM

#

Bruh not again

#

What is it

gentle peak Feb 2, 2025, 11:33 PM

#

fiery turtle Smm off might force qemu into acpi mode by default

you were right btw i just misunderstood at the time

#

uacpi_deininitialize_registers gets called before uacpi_leave_acpi_mode, which uses g_reg_lock but that's already freed by said deinit

dense steppe Feb 2, 2025, 11:34 PM

#

damn UAF

gentle peak Feb 2, 2025, 11:34 PM

#

not really since deinit also sets it to null

dense steppe Feb 2, 2025, 11:34 PM

#

ahh

#

still

fiery turtle Feb 2, 2025, 11:34 PM

#

How does that not crash like always

gentle peak Feb 2, 2025, 11:34 PM

#

but if it weren't for that it'd be uaf yeah

gentle peak Feb 2, 2025, 11:34 PM

#

fiery turtle How does that not crash like always

because it only gets called in uacpi_state_reset and nobody does that ig

fiery turtle Feb 2, 2025, 11:34 PM

#

My tests do

#

Oh wait

#

I pass NO_ACPI_MODE

gentle peak Feb 2, 2025, 11:35 PM

#

do your tests simulate needing to leave acpi mode?

#

ah

fiery turtle Feb 2, 2025, 11:35 PM

#

fuck

#

Good catch

gentle peak Feb 2, 2025, 11:36 PM

#

gentle peak because it only gets called in uacpi_state_reset and nobody does that ig

i'm pretty sure the only times uacpi_state_reset has ever been used directly by any kernel are by me trying to squeeze as much perf as possible out of it

#uACPI - a portable and easy-to-integrate ACPI implementation