#Crescent

1 messages · Page 2 of 1

cyan slate
#

Those can get complex

#

U can try uacpi In your kernel with the same ec driver meme

molten wyvern
#

its possible, I have done some manual testing with the fields but that very likely didn't cover a lot of cases

cyan slate
#

Yeah I barely have test cases for those

#

And frankly I dont remember how edge cases are supposed to work there

molten wyvern
#

the current field read/write logic that I have isn't even that complex, I don't see anything immediately wrong with it (though ofc it doesn't mean that its correct either)

#

ig if I would add uacpi maybe slightly more properly than last time I could try comparing the writes/reads again

cyan slate
#

That would be the easiest thing to do

#

Diffing the io

#

Or u can look at e.g. what _REG does

#

Maybe notice something suspicious that can misbehave in qacpi

cyan slate
#

Btw did you figure it out?

molten wyvern
cyan slate
#

Thats probably more useful than figuring out firmware issues

molten wyvern
#

I found out at least one possibly related minor difference and its that I register the ec handler after initializing the namespace

#

I didn't try if that makes any difference tho

cyan slate
#

And yes it makes a huge diff

#

Because it does OSYS checks in REG potentially

#

Basically hw with ecdt have almost no op _REG

#

Hw without ecdt has huge REG usually

molten wyvern
#

yeah I already do it before even loading the namespace if there is an ecdt

cyan slate
#

How do u run REG before loading the namespace?

molten wyvern
#

wait no yeah I misremembered lol

cyan slate
#

Lmfao

molten wyvern
#

before initializing the namespace

cyan slate
#

Yeah

#

So your behavior seems to match managarm then

#

Or?

molten wyvern
#

ah right, it only discovers the ec before initializing but doesn't install the addr space handler

#

nvm then

molten wyvern
#

also other thing that I found now by comparing the test runner output when feeding my dsdt/ssdts is that some fields are read in different order (maybe its related to the fact that I don't immediately evaluate fields to their values)

cyan slate
#

Wdym by immediately?

molten wyvern
#

or actually idk, in ```
OperationRegion (ICLK, SystemMemory, (SBRG + ((ICKP << 0x10) + 0x8000)), 0x40)

cyan slate
#

Ah I see

#

I dont think its specified by the spec

#

But I think nt does left to right too

#

Do u think it matters in this case?

molten wyvern
#

I'd hope it doesn't, that was just one difference that I found

cyan slate
#

I should test nt to see what order it does it in

#

But iirc left to right

karmic coral
#

are SBRG and ICKP registers?

cyan slate
#

As its the simplest way

cyan slate
karmic coral
#

lmao wtf

cyan slate
#

Ikr

molten wyvern
#

though to be fair I don't think the difference between uacpi comes from that because the other operand of each of the exprs in there is an integer

# 
uacpi
1 0X44AC4EAD = 0X9B
1 0X44AB9830 = 0X9B
1 0X44AC4334 = 0X9B
1 0X44AC4335 = 0X9B
1 0X44AC4336 = 0X9B
1 0X44AC4337 = 0X9B
1 0X44AC4567 = 0X9B
1 0X44AC4568 = 0X9B
qacpi
1 0X44AC4EAD = 0X9B
1 0X44AB9830 = 0X9B
1 0X44AC4567 = 0X9B
1 0X44AC4568 = 0X9B
1 0X44AC4334 = 0X9B
1 0X44AC4335 = 0X9B
1 0X44AC4336 = 0X9B
1 0X44AC4337 = 0X9B
#

or actually it probably is thonk

#

but like idk how exactly

#

actually wait, so uacpi evaluates the field right when it sees it and not when its used in an expr?

cyan slate
#

Well depends on what u mean

#

But I think so

#

If there's an AddOp(lhs, ...)

#

Where lhs is a field

#

Its evaluated right away yes

molten wyvern
#

as in (SBRG + ((ICKP << 0x10) + 0x8000)) when it first sees SBRG it evaluates it before pushing the value to the stack and continuing evaluating the other half of it

cyan slate
#

Yeah

#

I think so?

molten wyvern
#

yeah that probably explains it, ig I should do that too instead of pushing field object to the stack

cyan slate
#

I think its supposed to be kinda transparent in that way

cyan slate
molten wyvern
#

yeah Ill do, there are still a lot more in the test runner log that Ill have to manually compare because there's other trace info in between KEKW

cyan slate
#

I think uacpi would automatically convert that field name into a fake FieldReadOp

#

So the addop won't even know there was a field

cyan slate
clever shell
#

qwinci are u a crossont

molten wyvern
clever shell
molten wyvern
#

after that change at least the reads/writes done during the namespace load match, for the init I am not sure as the order is different (though I did spot one difference regarding the ec's reg, it sets OSTY to 7 whereas with uacpi it sets it to 1)

cyan slate
#

Why is your init order different

#

Oh also uacpi test runner disables some osi strings I think for testing

molten wyvern
#

doesn't uacpi do it depth first?

cyan slate
#

By top down u mean bfs?

#

Can u try commenting out the osi interace disabling code btw

#

(In uacpi test runner)

molten wyvern
#

yeah Ill try, that might be it as it does a check on the disabled osi string

cyan slate
#

Ye that would make sense

molten wyvern
cyan slate
#

Same order as uacpi

#

This is a depth first lol

molten wyvern
#

I wonder why is it different then KEKW

cyan slate
#

💀

molten wyvern
#

actually wait, I need to push the children in the opposite order that's probably why (so basically rn it inits the last child first)

#

which might be bad if it depends on the childs before it being initialized already

#

back to comparing the reads/writes ig

cyan slate
#

And it did break some hw iirc

molten wyvern
#

now ig its time to finally retest it, I didn't check all of the init addresses but I think they match now

#

I also got rid of reads from fields marked as preserved if the bit offset within an access size is zero and the bit width == access size in bits

cyan slate
molten wyvern
#

yes

#

I hope that it works now because if it doesn't I have no idea what to do about it

cyan slate
#

Btw a funny thing about fields

#

Ipmi, smbus and others utilize a read-then-write approach

#

What this means is

#

Local0 = FILD = Local0

#

Writes the request to FILD

#

The buffer that gets written is modified with the response right away

karmic coral
#

is that evaluated as (Local0 = FILD) = Local0 in aml?

cyan slate
#

And that rvalue gets stored to local0

karmic coral
#

ah no, i misunderstood

cyan slate
karmic coral
#

so Local0 = (FILD = Local0)

cyan slate
#

Yeah

#

Basically as if the operator= also mutated the inputs

molten wyvern
#

does the aml compiler handle that?

cyan slate
#

In c++

cyan slate
#

Im working on that in uacpi right now

molten wyvern
#

how do you even handle that lol

#

do you just read the field before assigning to it and then store the value read at the start to the target/push to the stack?

cyan slate
#

Well ill have the address space write handler also modify the input I guess

cyan slate
#

The contents of local0 are written to the field

#

And local0 is magically modified with the response

karmic coral
#

now write a constant instead meme

cyan slate
#

Or rather, a write to that field magically returns a response

cyan slate
karmic coral
#

FLD = Zero?

cyan slate
#

Well thats a buffer of size 8

#

For all intents and purposes

molten wyvern
cyan slate
karmic coral
#

temporary one i assume?

cyan slate
#

Yeah

#

It exists for the duration of the op

#

All those protocols have different buffer sizes they expect on input so u have to dynamically allocate and memcpy anyway

#

So it doesnt matter if aml writes a small value

molten wyvern
#

I don't think the ec thing happens now btw, at least so far it hasn't occurred

cyan slate
#

Good find

#

Uapci had this bug as well early on

#

Its only intuitive to push at the front

molten wyvern
#

I just hope that the field related change didn't break anything KEKW basically now it immediately reads the field when it encounters a name that resolves to a field

cyan slate
#

Thats the expected behavior I think

#

Well make sure u didnt break refof or derefof

molten wyvern
#

yeah I don't think I did, I specifically made it push the field object itself if a supername is expected (basically the same way it works for methods)

cyan slate
molten wyvern
#

yes

cyan slate
#

Ah

molten wyvern
#

anyway now ig its time to get back to the fun stuff after spending 5h on this annoying thing mostly consisting of comparing logs KEKW

#

I were starting to implement ppoll

#

though maybe I should start with signals first because they kinda interact with everything else

cyan slate
molten wyvern
#

signals KEKW

image.png
# 
struct sigaction act {
    .sa_handler = [](int num) {
        printf("SIGSEGV handler AAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAAA %d\n", num);
        while (true) {
            asm volatile("");
        }
    },
    .sa_mask {},
    .sa_flags {},
    .sa_restorer {}
};
sigaction(SIGSEGV, &act, nullptr);
*reinterpret_cast<volatile int*>(0) = 0;
#

now I just have to make tgkill ig, I already made a basic skeleton for it but it needs a pid to process table and I didn't add that yet

#

and sigrestore but that shouldn't be hard

scenic rampart
#

Signals my behated

molten wyvern
#

lol

#

I mean yeah they are kinda weird but at the same time I think they are kinda cool

molten wyvern
#

now ig its time for the poll, I have the basic signal stuff including tgkill/kill implemented (apart from SIGSTOP/SIGCONT but I don't want to handle that rn)

#

I just hope that I didn't introduce any deadlocks by using mutexes to protect the signal stuff instead of irq disabling spinlocks KEKW

#

I think its fine as the mutex is only taken inside sigaction and whatever else like that and inside the pre-emption timer but only if it came from userspace and a thread can't be both in sigaction and in userspace at the same time

#

idk if its really necessary to even protect the signal context from concurrent access by multiple threads but I couldn't find any info about the thread safety of sigaction

#

or signal safety too ig as there could be a signal that it installs handler for being triggered at the same time

#

well its specified to be signal safe so ig I need the mutex anyway

molten wyvern
#

I think I fixed the scheduler + event stuff now hopefully (previously it had weird workarounds), I don't want to touch it anymore KEKW

#

at least it works with 18 cores + 80 spawned windows in different processes

ivory cloak
#

"abandon all hope ye who enter here"

molten wyvern
#

lol

#

there are many different locks that interact with each other and are acquired in different order in different parts of the code

#

one per-thread lock that protects it from being moved to an other cpu, one per-thread lock to protect its run status and some other locks like per-scheduler sleeping threads list lock

#

now I think I can get back to the poll again 3rd time

#

as I implemented a wait multiple events function too, I hope that it works

molten wyvern
#

also one thing which I have no idea how to implement is syscalls that are supposed to return EINTR if interrupted by a signal

hot cobalt
molten wyvern
#

ah, yeah I guess that isn't too bad then as I'd just have to keep track whether a thread is inside an interruptible syscall and then if a signal is queued check check that and unblock the thread if its blocked (and set some marker inside it to tell it the wake reason ig)

molten wyvern
#

bash kinda works, though if I type anything nothing appears KEKW it does display the prompt again when I press enter tho which is weird

image.png
scenic rampart
#

did you implement tty echoing if its in raw mode (readline)

molten wyvern
#

I didn't implement tty at all KEKW I just did this, then in poll if there is input in input_events I return POLLIN and in read I just get inputs from the queue and convert them to characters

image.png
#

and in stdout vnode's write I just print whatever is put there using the kernel print

molten wyvern
cyan slate
molten wyvern
#

I wonder what is it then

#

at least it seems to be working otherwise, I typed echo hello from crescent and pressed enter

image.png
cyan slate
#

I could be retarded but I distinctly remember bash doing write(2) with the char and enabling echoing causing the char to appear twice

scenic rampart
#

you can probably look at termios to see what it wants

ivory cloak
molten wyvern
#

apparently it just wanted a somewhat sane termios out of TCGETS and now it mostly displays it correctly (except sometimes there are random spaces between the characters)

#

actually nvm that was just because I accidentally gave it a winsize with 18 cols lol

molten wyvern
dire galleon
#

thanks for the ping! I'm keen to have a read through this

molten wyvern
#

I should add some comments to it though, not everything of it is self-explanatory (though the sdm does a decent job at explaining most of the things)

#

and there is also generic x86 instruction emulation code in that file that I want to move out from there because it can be shared between different virtualization impls

ivory cloak
#

whoa crescent has a hypervisor

molten wyvern
#

I mean it can display the first line that seabios prints to vga but that's as far as it gets

ivory cloak
#

still pretty damn cool

dire galleon
#

yeah agreed, thats pretty cool

molten wyvern
# 
[desktop]: connection accepted
[kernel][x86]: pagefault caused by userspace read from FFFFFFFFE5A4B050 (instruction fetch)
    at FFFFFFFFE5A4B050
[kernel][x86]: sending SIGSEGV to evm (thread evm)
[kernel][signal]: killing thread evm because of unhandled signal 11
[kernel][sched]: destroying exited thread evm
[kernel][sched]: destroying empty process evm
[kernel][x86]: pagefault caused by kernel write to FFFF80010046E338 (non present page)
    at FFFFFFFF80076E8D
[kernel][x86]: sending SIGSEGV to user process (thread user thread)
[kernel][signal]: killing thread user thread because of unhandled signal 11
[kernel][sched]: destroying exited thread user thread

map a0000 (size 20000) -> 0x7b5000
run vm
[kernel][x86]: EXCEPTION: general protection fault at FFFFFFFF800CD593
``` I discovered more randomly happening faults, the gp fault happens only sometimes and its I am pretty sure its because `*(u64*) (xstate + 512 + 16)` is not zero like its supposed to be in a host xstate that I save before vmenter and load after vmexit (and the load specifically is where it faults)
#

idk how does it get to a non-zero state tho, if I add asserts it goes away too (or well at least it occurs less often so I haven't managed to get it to reproduce)

molten wyvern
#

apparently its 0x400 when entering the run function before running xsave too, I have no idea where does it get changed (its zeroed initially)

#

and breaking after its allocated and putting a watchpoint didn't help either, it doesn't show any accesses to that location at all (and in info tlb that's the only address that I can see referring to the same physical page alongside the hhdm but the hhdm address doesn't get hit if watchpointed either)

molten wyvern
#

apparently it gets modified by the vmptrld nooo that's certainly not the address that I give to it

#

actually I might know why, maybe its because I leave a temporary test vmcs current and free the backing page and then when I execute vmptrld again it writes the state of that vmcs that was stored on the cpu to memory before making the new vmcs current

#

I think that was it, doesn't look like the simd state gets modified anymore

#

and that other issue might also have been related, though some different issues remain if you launch multiple instances of the vm at the same time (likely because I don't make the vmcs current in any of the helper functions so things like writing to the vm's control registers might affect some random other active vm)

#

but that shouldn't be hard to fix, Ill just have to keep track of the currently active vmcs and if its different than the one the program wants to modify change it using vmptrld

molten wyvern
#

I also found other safety issues with it like the owning vm wasn't really being referenced by the vcpus so if you'd have closed the vm handle and tried to run a vcpu funny things could happen (in the best case nothing and in worst the kernel using random freed memory as the ept along with potentially faulting if it can't traverse the map properly)

#

should be fixed now tho

cyan slate
#

btw whats the current situation with crescent suspend support

molten wyvern
#

well I haven't really touched it in some time, last time it worked in qemu and its implemented for some devices like the ps2 and bochs fb

#

but on the laptop it didn't really properly work (or well it did likely enter some kind of state as the power light went off for like 2s but after that it lit back up with screen off)

cyan slate
#

interesting

#

should try obos on it

molten wyvern
#

has the fb been properly restored on some real hw using it

#

though well ig that doesn't really matter for the purposes of seeing if it spuriously wakes up

cyan slate
molten wyvern
#

so either it suspends and resumes very short time afterwards or it triple faulted or did something else that caused a reboot

cyan slate
#

@ivory cloak

#

do u have amd btw?

molten wyvern
#

no

ivory cloak
#

I don't have amd

cyan slate
#

yet another obos triple fualt

ivory cloak
#

obos doesn't run on amd, and I am yet to debug it

cyan slate
#

qwinci has an intel

ivory cloak
#

bruh

#

how new

#

it works on my haswell cpu

molten wyvern
#

11gen

#

i5 1135g7

#

I can also try it on my desktop rq now that I have it on an usb, brb

cyan slate
#

maybe that last iso is just bad

#

do u have a new one

ivory cloak
#

I'll post the latest obos iso

#

just let me remove the userspace stuff from the initrd to make it smaller

#

wait nvm dont use that

#

uacpi causes it to hang or smth

#

somewhere during driver loading

#

wait that should actually work, it's using uacpi master

#

@molten wyvern use that iso

#

pls

#

if all goes well it hangs (intentionally) in the r8169 driver

#

after printing the mac address

molten wyvern
#

Ill try, with the old iso on this desktop it just hanged with black screen and I had to force shutdown by long power press

ivory cloak
#

thanks

cyan slate
molten wyvern
#

no it was on my laptop

#

the new iso also has the same behaviour on the laptop, it just straight up reboots right after boot to it

ivory cloak
#

bruh

#

can you help me debug?

#

if you have time rn

cyan slate
#

ill test obos suspend on my 5 laptops before 1.0 release meme

#

later this week

molten wyvern
#

looks like it has the same hang with no screen output on the desktop as with the old iso too

ivory cloak
molten wyvern
#

on the desktop I assume its hang because it doesn't reboot

ivory cloak
#

ok

molten wyvern
#

and on the laptop its likely triple fault

ivory cloak
#

without output?

ivory cloak
#

a triple fault on the laptop

#

debug the laptop first

#

I'm going to add a hang somewhere

#

and send you an iso

# 
Arch_KernelEntryBootstrap:
        cmp rsi, 0x554c5442
        je .ok ; should triple fault on failure
        xor rax,rax
        jmp rax ; All hope is lost.```
that's the only code that intentionally triple faults
#

it checks that the hyper magic passed in rsi is what it should be

#

one of the first things that should happen is you getting logs if all goes well, so it's in early early boot

#

I'm tryna send the iso but discord is being annoying

#

fucking hell

molten wyvern
#

Ill not be able to test it right away anyway, Ill have to first play the daily overwatch match and eat something troll

ivory cloak
#

discord is being restarted

#

yeah anyway I gtg now too

#

I'll send the iso later

molten wyvern
#

I am trying to think what exactly should I do, if I want to port any kind of desktop apps then I basically need xorg or I have to rewrite parts of the apps to use my custom window protocol (which also means that I can't run any binary apps like linux factorio)

#

unless I'd do something similar to xwayland but that sounds painful

winged parrot
#

how about wlcrescent :^)

#

i would not bother messing with xorg

molten wyvern
winged parrot
molten wyvern
#

ah, ig that could work (though first of all I'd have to make the crescent windowing thing not be shit and see what exactly would it need to be able to provide such a layer)

winged parrot
#

i would definitely not bother doing xorg

#

wayland is way nicer to work with

errant echo
#

just layer xwayland on wlcrescent ultrameme

unreal flame
#

Xwayland isn’t bad tbh

clever shell
molten wyvern
#

I took a closer look at the phone touchscreen that I wanted to write a driver for and apparently the i2c serial engine its connected with is literally the only one that has disabled fifo support which means that I need a driver for some annoying serial dma thing called gpi (all the other i2c serial engines that are used to connect audio amp, fm radio, nfc and vibrator support fifo)

winged parrot
#

lol

#

ofc

molten wyvern
#

ig it kinda makes sense considering that you get a lot of data from it but still idk why does it have to have it disabled

molten wyvern
#

at least it has some things with proper public specs like the vibrator (ti drv2624), the amplifier (ti tas2563) and the fm radio (rtc6226) looks to have but its made by some chinese company called richwave that requires some kind of signup to get it and I couldn't get past that (it just generates some kind of undescriptive error)

#

about the nfc thing I am not sure, its compatible within the dtb is just tms,nfc and nothing comes up with that except some belgium based software company (might be that they made the driver but the chip is made by some other company)

molten wyvern
#

also apparently my ps2 driver is broken somehow again at least on the other laptop that I have (I haven't really even touched it after I improved it other than changing timeouts to be in ns) nooo ps2 really sucks

#

it just hangs somewhere (I think its within the controller init or port detection stuff but Ill have to add more prints to there)

clever shell
scenic rampart
#

Qwinci pushed code to astral, now he is forever cursed

molten wyvern
#

lol

clever shell
ivory cloak
#

Just how nyaux pushed code to obos

#

And his kernel got cursed

molten wyvern
#

its also annoying to debug because its the lenovo laptop that I have from work that doesn't have a proper ethernet port so no pxe

#

I wonder if I have tested obos on it

clever shell
molten wyvern
#

idk if this laptop even has a ps2 connected trackpad, it might not

cyan slate
#

aml will have answers

molten wyvern
#

according to it it looks like it would have

image.png image.png
cyan slate
#

yup

#

probably a synaptics one too

molten wyvern
#

probably, all the other lenovo laptops that I have seen have had synaptics too

#

about the other my own laptop I don't remember, it might have elantech

cyan slate
#

btw if its a thinkpad it has a multiplexed ps/2 as well

molten wyvern
#

yeah it is a t490s

#

is the tracknob (idk what is it called lol) accessed using that mechanism? or does it just share the input with the trackpad?

cyan slate
#

its a separate ps/2 device

#

it has 3

#

the keyboard, the nob, and synaptics

#

multiplexed ps/2 supports up to 4 devices

#

so yeah u need to enable the multiplexed mode for it to work

molten wyvern
#

interesting, I guess I can take a look at that after I have hopefully figured out the other issues

cyan slate
#

its a pretty simple thing honestly

molten wyvern
#

it should still function without the multiplexed mode enabled tho right?

cyan slate
#

yes but in a semi retarded way

#

iirc it tries to combine both

#

into one thing

molten wyvern
#

ah

molten wyvern
#

apparently the reason for the hang was that in my sched unblock function I forgot to re-queue the thread if it was sleeping previously

#

I think that there is some other hang that doesn't always happen while evaluating _BST too

#

the good thing is that sleep apparently at least somewhat works, at least it doesn't randomly wake up on its own (though it also doesn't wake on keyboard/trackpad click) and when waking it with the power button the mute led goes on but the power button keeps pulsing (and display is off) so its not completely working

cyan slate
#

and obos hangs on this laptop because bogus ioapic redirections right?

molten wyvern
#

possibly

cyan slate
#

it calls finalize_gpe, which enables some gpes that have been pending, which makes the cpu hang once it recieves the irq

#

we know that because it doesnt hang if he doesnt install the irq handler

#

annoying bug ngl

molten wyvern
#

yeah

molten wyvern
#

yeah looks like pinning it to the bsp makes that not happen, idk if this is a skill issue in the kernel or some kind of firmware meme nvm it didn't fix it

#

and for the wake at least one possible thing that could cause the waking to fail even on power button could be if the firmware boots the code in protected mode as I rely on cs relative things in it

#

I should probably write a different trampoline for that case because I don't think there is any easy way to make it work on both

#

or well ig I can just have the entry part be set to different parts of it for them as I can reuse the long mode stuff + gdt

molten wyvern
#

I wanted to do kgdb so I could debug the weird stuff like the _BST taking too long to execute but now I realized that its not so easy to implement on top of a normal socket

#

because like I need to have access to some other thread's stuff than the thread where the kgdb socket is served

molten wyvern
#

also I am starting to get braindamage from the intel gpu render prms, they are literally structured the worst way possible and include irrelevant stuff in between (and the fields within the structures aren't documented very good, like wtf does Vertex URB Entry Length in pairs of 128-bit vertex element units mean or how is the urb distributed between different shader threads and are you supposed to restrict the amount of threads based on how much urb space one invocation requires or how does simd8 dispatch work (I can guess that it means that one vertex shader invocation is supposed to process multiple vertices <= 8 but idk if that's what it means) nooo

winged parrot
#

i.e. probably multiples of 32

molten wyvern
#

actually now that you said it I think I saw somewhere mentioned that the vertex attributes always take that amount of space in there

winged parrot
#

lmfao

#

im missing context to understand the rest of them

#

what is a URB?

#

uniform resource buffer?

#

are you doing accel CBT?

molten wyvern
# winged parrot uniform resource buffer?

its either uniform return buffer or universal return buffer, as far as I understand its basically a region of memory used to store input vertex attributes extracted from the vertex buffer to be consumed by the vertex shader (and then the vertex shader puts its outputs to there that are then fed as inputs to the pixel shader)

molten wyvern
cyan slate
molten wyvern
#

trying to make a "driver" to draw an accelerated triangle on an intel gpu

cyan slate
#

oh lol same as me

#

(a few years back)

molten wyvern
#

its where I got the idea from

cyan slate
#

lol

molten wyvern
#

I though about what I want to do with crescent some more and I think I am going to not implement any of the linux/glibc specific things (and I am going to make more options to hzlibc to disable them, shouldn't be hard as they are already in their own folders mostly) so its just going to be an unix-like but not linux-like and then I want to actually try to do the gpu shaders (+ if it goes well enough then make some kind of primitive glsl compiler, I already kinda started one that has some of the parsing done)

#

hopefully its not going to make porting some stuff too hard lol

cyan slate
#

are you going to work on both new and old crescent at the same time?

molten wyvern
#

Ill probably just switch between them, I just want something else custom to work on too lol

cyan slate
#

interesting

#

a glsl compiler sounds painful

#

that shader isa is tough

#

but it would be cool if u figured it out

molten wyvern
#

I ported my gameboy emulator that originally used sdl2 for video + audio to the crescent windowing system + kernel audio interface (I think it would be better with uhda than this old one that I had but I haven't yet made the api fully functional on top of uhda) KEKW

#

it also revealed some issues with window redrawing that Ill have to take a look at

molten wyvern
#

the issue was pretty trivial to fix as I just needed to make some adjustments to the dirty rect calculation to take into account the borders. one other thing that I noticed is that it doesn't properly work on my phone for some reason, it just displays a static snapshot of the emu window

#

looks like it likely has something to do with load balancing (or smp init ig), if I disable the smp bootup it works fine

molten wyvern
#

apparently I forgot to add the scheduler timer callback on the other cpus (and also enable the timer on the other cpus)

#

so now it works, though on real hw the ap bootup sometimes gets stuck

molten wyvern
#

I finally decided that its time to commit all the changes and it literally took like 4 hours to sort them to commits that hopefully compile without any changes 💀 it was like 33 commits lol

#

at least I learned that you shouldn't accumulate multiple months worth of uncommitted changes like this

timber patio
#

lol

molten wyvern
#

today I started trying to implement bluetooth over usb but apparently my xhci driver is somehow broken again so that it doesn't even detect the device in qemu KEKW so I can't really test whether any of it works until I figure out how to fix it and what is the problem with it now

#

after that's figured out (and assuming that the stuff that I wrote actually works) ig I can expose some bt ioctls (or well the equivalent that I have with devlink, its going to need a new btlink or whatever) to userspace so then I can start writing the rest of the billions of layers that there are on top of the core bt inside some kind of bluetooth server lol

cyan slate
#

How does Bluetooth work

molten wyvern
#

idk how does the higher level stuff work yet as I have mostly just looked at the transport layer and that's basically just some command packets to the controller passed as the body of usb setup packets or data packets to other devices yeeted through an usb bulk endpoint (+ an irq in endpoint for events like command completions from the controller)

#

well there are also other transports it can use besides usb like sdio (which needs an sd card driver), that's how it works on eg. qcom phones afaik

cyan slate
#

interestng

molten wyvern
#

it does work poggers I can read the bluetooth clock and it generates a cmd complete event, the problem wasn't really the xhci code but me assuming that qemu would take host usb device vendor/product as hex by default but apparently that's not the case

#

now ig I can try sending some inquiry packets (or well tell the controller to start sending them, you don't have to send the packets themself yourself) and see what devices there are or smth

cyan slate
molten wyvern
#

for which part?

cyan slate
#

in general

molten wyvern
#

writing a driver for an usb bluetooth adapter

cyan slate
#

is there like a spec for it

molten wyvern
#

yes

cyan slate
#

i thought it would require proprietary firmware blobs and stuff

molten wyvern
#

it probably does

cyan slate
#

but works anyway?

molten wyvern
#

well at least some part of it but idk if the actual stuff that sends packets over the air works

cyan slate
#

ah

molten wyvern
#

at least the inquiry works, I get an event if I put my phone into pairing mode

#

but also idk if its possible that the fw is retained in the device from when linux loaded it initially

cyan slate
#

yeah possibly

molten wyvern
#

one problem that Ill have to solve is that there is really no way to respond to hotplugs in userspace and that's also an issue with usb stuff like this even if its not really hotplugged as the usb device driver runs in a different thread that might get scheduled after other stuff

molten wyvern
#

I am starting to not like this bluetooth stuff very much lol, like why does there have to be billion different things that the software must implement nooo first you have the acl packets that you give to the controller, then you have multiple different types of l2cap packets (depending on whether the l2cap channel is reliable or not) that you have to split/combine to/from the acl packet data + re-send if a timeout happens + other features that tcp has too, and to create an l2cap channel to a specific service on the other device you have to use a predefined l2cap channel to first query the device for the services it has using a different protocol on top