nah i noticed this yesterday as well and that was before the builtin was replaced
#uACPI - a portable and easy-to-integrate ACPI implementation
fiery turtle
popcount was only used in one place in code
and its not triggered by proxima
the toshiba laptop right?
gentle peak
yes
gentle peak
can't test, threw away the laptop battery a while ago due to expansion
fiery turtle
ah right
calm latch
You can try testing pmOS 
gentle peak
searched in my message history a bit and it seems like on dec 26 i did still get lid events
so i'm downloading an image from ci from that day and testing with that
left orbit
god bless continuous integration
fiery turtle
searched in my message history a bit and it seems like on dec 26 i did still get...
yeah there should be a lid notify
for GPE 02
the one with this
left orbit
holy hell
fiery turtle
searched in my message history a bit and it seems like on dec 26 i did still get...
whats interesting is i didnt really change much since dec 26
gentle peak
maybe i downloaded the wrong build but i'm not getting lid events on this either
oh wait nevermind i am
i apparently just wasn't patient enough?
fiery turtle
uhh
gentle peak
maybe that was the issue with the newer build too lol
fiery turtle
can you retry the newer one lmao
gentle peak
i apparently just wasn't patient enough?
like i didn't keep the lid closed long enough
fiery turtle
this while is probably so slow it takes a while
nopunintended
fiery turtle
like i didn't keep the lid closed long enough
i guess it makes sense for them to have a slight timeout
or maybe a polling interval of some sort
gentle peak
yeah i'm getting LID_(0x80), VALZ(0x8F), LID_(0x80) on latest proxima (with uacpi master)
testing 1.0-rc now
gentle peak
getting them on 1.0-rc too
gentle peak
yeah
fiery turtle
ok nice
left orbit
least broken laptop firmware
gentle peak
yes
fiery turtle
does it work?
gentle peak
it does
fiery turtle
and shuts down fine and stuff
gentle peak
yeah perfectly fine
gentle peak
no problem
fiery turtle
alright, I guess thats it, im adding davix to the leaderboard and merging this PR
then pushing the 1.0 tag
gentle peak
testing pmos too for the fun of it
because this laptop had broken fb before on there
still broken unfortunately
power button works fine regardless though
fiery turtle
most stable microkernel
calm latch
The framebuffer driver is in userspace, so technically the kernel is indeed stable 
fiery turtle
it did shutdown so i guess
fiery turtle
holy hell
yeah this is basically textbook example of why u dont make a recursive aml interpreter
calm latch
Still, needing to press enter in VMs is weird
Does QEMU use keyboard controller for shutting down?
fiery turtle
no
calm latch
The new 32 bit build...
It works in all vms (including bochs) and tripple faults on my laptop
fiery turtle
uACPI 1.0 is finally LIVE 
kind mantle
fiery turtle
the 1.0 leaderboard
jaunty fox
Congrats on the big 1
fiery turtle
Congrats on the big 1
thanks for believing in the project!
kind mantle
the 1.0 leaderboard
Shouldn't you, to fairly compare kernels, run them all on the same type of CPU?
So e.g. they'd send you ISOs and you run it.
torpid root
that needs a volunteer with a good cpu
fiery turtle
i have a shitty cpu that doesnt perform well, plus people want to track progress relative to their cpu and not the cpu that i have
kind mantle
that needs a volunteer with a good cpu
The CPU doesn't need to be "good", it needs to not be abhorrently slow.
fiery turtle
the top1 score is definitely not the best cpu on the list either lol
loud ice
i'm really curious what the difference between proxima and the rest is
torpid root
The CPU doesn't need to be "good", it needs to not be abhorrently slow.
I'd say that the volunteer part is the most important
jaunty fox
thanks for believing in the project!
No worries, you have certainly established yourself as an authority on acpi
fiery turtle
i'm really curious what the difference between proxima and the rest is
insane allocator
loud ice
insane allocator
i dunno it feels super weird
kind mantle
I would be willing to run benchmarks, my CPU isn't the most recent but it's decent. It's a Ryzen 5 5600.
loud ice
its. a very large margin
vast kestrel
A fast path allocator does wonders
fiery turtle
what can i say monkuous is really good at what they do
jaunty fox
i dunno it feels super weird
No smp
fiery turtle
this score is with smp support
left orbit
the allocator actually is really that fast
fiery turtle
ikr
vast kestrel
Like, I think we saw that with a lock in the alloc path it dropped dramatically
fiery turtle
monkuous profiled every function call and micro optimized it
kind mantle
monkuous profiled every function call and micro optimized it
lmao nice
hasty plinth
Congrats on the 1.0 🎉
fiery turtle
thanks a lot!
fiery turtle
I'll have to add magazines to keyronex
keyronex is still not on the list which makes me sad
kind mantle
I would be willing to run benchmarks, my CPU isn't the most recent but it's dece...
But either way, let me know if you guys want to take me up on my offer. I think it'd be cool if the leaderboard is all run on the same CPU.
jaunty fox
keyronex is still not on the list which makes me sad
I have figures in this thread somewhere
sterile egret
May you never go crazy and smash up an expensive sports car 
fiery turtle
I have figures in this thread somewhere
those are probably severely outdated
jaunty fox
But either way, let me know if you guys want to take me up on my offer. I think ...
I agree that there should be a clearinghouse for scores so we can compare them fairly
flat badge
infy, are you also updating the Managarm bindings?
fiery turtle
But either way, let me know if you guys want to take me up on my offer. I think ...
i agree
jaunty fox
It should be someone with decent statistical skill
fiery turtle
infy, are you also updating the Managarm bindings?
no disk space to build managarm 
flat badge
wtf lol
fiery turtle
its like 30gb
gentle peak
i'm really curious what the difference between proxima and the rest is
Hydrogen (Proxima's kernel) is always built with -O3 LTO (it relies on it to keep performance that would otherwise be sacrificed for code cleanliness)
flat badge
no, it's 3 gb
fiery turtle
is it
gentle peak
So there's a lot of inlining
flat badge
you don't need to build every package ;D
fiery turtle
maybe ill give it a try
jaunty fox
Hydrogen (Proxima's kernel) is always built with -O3 LTO (it relies on it to kee...
Lto, that'll be helping no doubt
flat badge
by now we also have automatic pulling of deps, so you only need to do xbstrap install --rebuild managarm-kernel in an empty build dir (after setting up xbstrap correctly ofc)
jaunty fox
I don't use 03 they say it makes hash out of good code
fiery turtle
by now we also have automatic pulling of deps, so you only need to do `xbstrap i...
oh cool
left orbit
yeah inlining the entirety of the kernel into kmain certainly helps 
fiery turtle
yeah lto is insane
gentle peak
With LTO the allocator's fast path is a few instructions that are inlined basically everywhere
jaunty fox
With LTO the allocator's fast path is a few instructions that are inlined basica...
What sort of allocator is it?
gentle peak
Slab
flat badge
building Managarm with LTO doesn't make a huge difference
jaunty fox
With magazines?
flat badge
but Managarm doesn't have a lock free allocator fast path
gentle peak
No magazines, but it doesn't matter because the benchmark is done with -smp 1
jaunty fox
They do at least let locking be elided
gentle peak
With magazines you can replace 2*cmpxchg with pushf+cli+sti
It'd probably improve perf a bit but I doubt it's that much
mortal yoke
pushf/cli/sti also affects the score decently if you do it in alloc fast path
flat badge
yeah
calm latch
pmOS is -O3 lto
gentle peak
cli/sti isn't that expensive right? Iirc they don't cause vmexits at least (unlike cr8)
flat badge
putting another pushf cli sti in the managarm syscall path gives a measurable performance drop in glxgears
gentle peak
i think their mutexes are 8ns to acquire
Mutex acquire is actually the major performance bottleneck of the allocator currently
Not because of the acquire itself but because of the memory ordering requirements
fiery turtle
ah
mortal yoke
cli/sti isn't that expensive right? Iirc they don't cause vmexits at least (unli...
my bet is that it would be the pushf/popf but I haven't verified whether that's the case
fiery turtle
Mutex acquire is actually the major performance bottleneck of the allocator curr...
how does your schedule_work work if it needs to allocate a task that is called from an irq handler?
flat badge
no, the cli sti are expensive
fiery turtle
that might sleep on the mutex
gentle peak
how does your schedule_work work if it needs to allocate a task that is called f...
IRQ handlers are not allowed to allocate
loud ice
yeah cli/sti is pretty pricey
fiery turtle
IRQ handlers are not allowed to allocate
yeah so how does it work currently
gentle peak
An ACPI IRQ simply signals a semaphore which a softirq task is waiting on
fiery turtle
ah ok
vast kestrel
If you don't use the allocator from interrupt handler you can replace cli/sti with a preempt count
gentle peak
Good point
mortal yoke
no, the cli sti are expensive
interesting, is it related to the fact how they have the delay and whatever? because aren't they fundamentally just setting a bit in a register?
fiery turtle
If you don't use the allocator from interrupt handler you can replace cli/sti wi...
what is that? like block preemption until the counter drops?
vast kestrel
(while still keeping the per core semantics obviously)
gentle peak
ah ok
Before I switched to a softirq task I just statically allocated enough to have 256 things in flight at a time
vast kestrel
Yeah basically, if you get a preemption with the counter being non zero toy just return
And preempt manually at zero
fiery turtle
i see
vast kestrel
The cost is basically a per cpu count and an if at the decrement
fiery turtle
And preempt manually at zero
like call yield after it drops? wouldnt that steal your entire slice potentially
vast kestrel
Your slice is already up if it preempted in the middle
fiery turtle
but u dont know that
vast kestrel
You keep a mark
flat badge
the problem with preemption locks is that you probably don't want to acquire a spinlock while they are active though
vast kestrel
You increase the count, if a preemption interrupt happens toy mark it, and when decrementing and it's marked for preempt you preempt manually
Yeah
Preemption locks are obviously only for the per-cpu path
gentle peak
What I do for {enable,disable}_preempt:
- Separate TASK_RUNNING and TASK_READY
- On yield always set state to TASK_READY, even if preemption is disabled
- When counter reaches 0 and state isn't TASK_RUNNING, yield
fiery turtle
You increase the count, if a preemption interrupt happens toy mark it, and when ...
ah
mortal yoke
interesting, is it related to the fact how they have the delay and whatever? bec...
according to uica it looks like its specifically the sti that's expensive, 4x more than cli so maybe its indeed the effects and whatever (and popf is almost 5x more than cli)
gentle peak
Popf isn't used though
Pushf is, but it's popped into a regular register
At least usually
jaunty fox
In keyronex can do an ipl raise instead
gentle peak
Unless you do lazy IPL that has to write to cr8 which causes a vmexit
fiery turtle
In keyronex can do an ipl raise instead
thats a cr8 write tho right
jaunty fox
It won't be reflected with a cr8 write unless an interrupt arrives while ipl raised
gentle peak
That's even worse
vast kestrel
Cr8 is a bit more expensive than cli/sti iirc
fiery turtle
not a bit
gentle peak
Oh ok lazy ipl helps a lot with that yeah
fiery turtle
like a ton
vast kestrel
Ignoring vmexit
fiery turtle
in vms
gentle peak
Cr8 is a bit more expensive than cli/sti iirc
On real hardware probably a bit, but in VMs it's incredibly expensive
mortal yoke
Popf isn't used though
ah yeah
jaunty fox
If one does, cr8 is written and interrupt left unhandled for later
fiery turtle
If one does, cr8 is written and interrupt left unhandled for later
why write it in the irq handler? cant u just keep it at what it was
mortal yoke
then you are going to get the irq again
fiery turtle
hmm
gentle peak
Level triggered IRQs + IRQs are already so expensive you might as well write cr8 and avoid further IRQs
fiery turtle
who would unset it tho
gentle peak
Then you can't receive higher priority interrupts
fiery turtle
and that yeah
mortal yoke
I wonder why wouldn't they use the vmx tpr virtualization if its supported tho
flat badge
who would unset it tho
you'd unset it when exiting the critical section
mortal yoke
that would avoid vm exits on cr8 writes/reads afaik
fiery turtle
i wonder why that needs a vmexit in the first place
why cant that be emulated by the virtual irq chip
jaunty fox
It shouldn't with vTPR or whatever Iots called now but unfortunately it seems that's not exploited
fiery turtle
i guess because its not a priority for kvm
since linux or bsds dont use cr8
and nt has lazy cr8 probably
flat badge
it's not clear that these strategies are worth it though
fiery turtle
which
gentle peak
I doubt lazy IF update is worth it
flat badge
sure, they will improve the uacpi benchmark but they pessimize workloads that actually use interrupts
fiery turtle
its just a different system
gentle peak
CR8 probably because the only reason to use cr8 is because you want interrupt priority for other reasons
flat badge
lazy IF update is strictly superior to lazy CR8 update in terms of throughput
gentle peak
Yeah but lazy IF update doesn't preserve the priority stuff
gentle peak
From what I've seen the interrupt priority stuff is rarely done for performance reasons
fiery turtle
and masking of some interrupts at some ipls
yup
and cli/sti is better for uacpi benchmark too
gentle peak
Honestly the best part of uacpi 1.0 is that i can finally remove __popcountdi2.S
fiery turtle
ikr
flat badge
I'm aware that there are other uses of CR8
i'm only talking about the optimization here
fiery turtle
nah its not a superior approach for perf
gentle peak
Oh yeah for optimization it's definitely not worth it over just plain cli/sti
fiery turtle
i dont think there was an argument for that
flat badge
well, it is definitely faster in the uacpi benchmark
flat badge
because cli / sti have a real (and non-negligible) cost while lazy IF=0 has no extra cost if there is no interrupt
and there are no (or very little) interrupts during namespace initialization
gentle peak
I just stubbed out save_disable_irq and restore_irq (so no cli/sti at all during uACPI benchmark), performance is still ~within the same range
vale isle
why does cli / sti have a cost? is it that "sti doesn't apply until the next instruction is done executing" thing?
flat badge
I just stubbed out `save_disable_irq` and `restore_irq` (so no cli/sti *at all* ...
that's surprising
gentle peak
that's surprising
Eh, not that surprising, almost none of my kernel APIs disable IRQs
Except the ones that are absolutely required to
So it's not called particularly often
flat badge
for example, just inserting an irq lock + unlock into Managarm's syscall path is way more impactful than disabling LTO
not for the uacpi benchmark ofc but for performance in userspace benchmarks
gentle peak
I haven't used the profiler in a while but the last time I did, during the entire uacpi benchmark, save_disable_irq and restore_irq were called a grand total of 4 times each
Compared to for example mutex_lock (17k) and kalloc (10k)
flat badge
how do you protect against IRQs that interrupt a spinlock?
gentle peak
Spinlocks always disable IRQs, I just don't use spinlocks if I can avoid them
fiery turtle
how do you protect against IRQs that interrupt a spinlock?
no code paths that need spinlocks are touched by the benchmark i guess
flat badge
how does the 17k mutex_lock operations match with 4 cli then?
gentle peak
Mutexes don't disable IRQs unless they're contended
In the uncontended path lock and unlock are both just one cmpxchg
And given how early uacpi is initialized, if they're contended at all it's a bug and deadlock
flat badge
so what happens if CPU1 does mutex_lock (uncontended), then CPU2 locks (contended) and CPU1 receives an IRQ?
vagrant hull
congrats on the 1.0 lol
fiery turtle
congrats on the 1.0 lol
thx!
gentle peak
Eventually CPU1 will get back into the critical section and unlock it
Same way userspace mutexes work at all
north holly
Ryurjr7dkryd
flat badge
so the task on CPU2 will sleep in this case?
north holly
gentle peak
Yeah
north holly
1.0 released
flat badge
ah
gentle peak
If mutex_lock is called while contended, the mutex state is updated to reflect that, and the thread sleeps on the mutex
Then in mutex_unlock it tries to cmpxchg the locked-uncontended state to unlocked, and if that fails it must be contended and it wakes up one of the sleeping threads
flat badge
yeah i get how the mechanism works i didn't think of this though
fiery turtle
yeah pretty smart
flat badge
also because we do quite a bit of work in contexts where we can't sleep
fiery turtle
monkuous is just huge brained
flat badge
you can only do this operation if all work happens on threads though
i.e., you need one stack per concurrent unit of work
which doesn't apply to us
fiery turtle
yeah with fibers this wont work
gentle peak
Ah you have stackless coroutines in kernel?
flat badge
yes
gentle peak
That'd make this more complicated yeah
fiery turtle
yeah managarm does all uacpi work in fibers too
get_thread_id is implemented as currentFiber() iirc
flat badge
fibers in the kernel do have their own stack though
but some stuff is done on work queues that don't have their own stack
but even on your kernel, this approach has the same performance characteristics than locking a spinlock with preempt disabled (but IRQs on), right?
in particular, CPU2 in my example will pay at least the latency of the IRQ on CPU1
in reality, it will pay more since a task switch is more expensive than spinning
gentle peak
It does
But the same performance characteristics apply to mutexes and spinlocks in userspace
Or well I suppose the spinlocks are a bit better in kernel because you can avoid the IRQ latency
Hm
I should test this once I've got a more active userspace
Anyway the general consensus for userspace mutex vs spinlocks is that it's not worth optimizing for the heavily contended case
flat badge
your strategy is probably good for workloads where you always have enough tasks to run on all cpus
gentle peak
Yeah I tend not to consider multi-CPU workloads beyond "does it work"
Hydrogen is intended to be a single-CPU kernel with multi-CPU functional but not super great
fiery turtle
Hydrogen is intended to be a single-CPU kernel with multi-CPU functional but not...
why is that
gentle peak
If you want to have a good multi-CPU kernel your scheduler will get very complicated very quickly due to hyperthreading and heterogenous CPUs
I didn't want to deal with that, and I decided that if I was gonna half-ass multi-CPU support in the scheduler I wasn't gonna focus on good multi-CPU support in other areas either
fiery turtle
hm
leaden fox
If you want to have a good multi-CPU kernel your scheduler will get very complic...
how does hyperthreading make it more complicated
gentle peak
If you have two CPU-intensive threads you have to take care not to schedule them on the same physical CPU
Otherwise you will have terrible performance
frank canopy
and nt has lazy cr8 probably
it used to afaik but iirc nt doesn't have lazy cr8
or at least the wdk headers I have irql as macros that set cr8 directly
calm latch
in contrast to just cli / sti
cli/sti doesn't work for ring 3
(which is how I use cr8)
fiery turtle
breaks how?
torpid root
gpf
fiery turtle
how 💀
torpid root
idk
the current one just switches between pf and inv opcode
fiery turtle
which opcode even
fiery turtle
fiery turtle
idk neither addr2line nor objdump shows anything at rip
x/10i $rip
torpid root
yeah I'm not in the mood for debugging
why do I even do this
I'll go play factorio
fiery turtle
lol
torpid root
congrats on v1 btw
uacpi in linux 2025
fiery turtle
thanks!
slim panther
is v1 out yet?
slim panther
LETS GO
slim panther
we shall be implementing
strong heath
@fiery turtle I have upstreamed support if you want to add Ironclad now
strong heath
actually dont add it, the ops/s are too slow for now and its going to look like shit
ill optimize it and give them to you
strong heath
I have self respect
fiery turtle
Its an Ada kernel so I think its fair
Reactos is also in that neighbourhood I think
strong heath
I can speed it up a lot ill just do so
fiery turtle
Sure
fiery turtle
context?
north holly
did ironclad have acpi before this
strong heath
context?
(0000000007) unable to attach (PCI_Config) operation region \_SB_.PC00.THC0.TH
(0000000007) CR: not found
(0000000007) aborting table load due to previous error: not found
(0000000007) #0 in \()
oh no the new line is expected
sorry I forgot that I am splitting them
yeah its THCR not found
fiery turtle
this means you returned NOT_FOUND from your pci_device_open
fiery turtle
is this qemu?
strong heath
nah this is real hardware
fiery turtle
ah ok
strong heath
its a decently new intel board
strong heath
thats the only error though so thats encouraging
fiery turtle
well considering it aborted the entire table load due to it
or maybe it was one of the SSDTs, since u didnt paste the entire log idk
strong heath
nah its the DSDT
fiery turtle
i mean
+ function PCI_Device_Open
+ (Address : PCI_Address;
+ Handle : out System.Address) return Status
+ is
+ Ret : constant uACPI_PCI_Acc := new uACPI_PCI;
+ Success : Boolean;
+ begin
+ Arch.PCI.Search_Device
+ (Bus => Address.Bus,
+ Slot => Address.Device,
+ Func => Address.Func,
+ Result => Ret.Dev,
+ Success => Success);
+ if Success then
+ Handle := C3.To_Address (C3.Object_Pointer (Ret));
+ return Status_OK;
+ else
+ Handle := System.Null_Address;
+ return Status_Not_Found;
+ end if;
+ end PCI_Device_Open;
u do return not found here
strong heath
yeah I do, it is not coming from nowhere
fiery turtle
can u send the dump btw?
strong heath
ill do when I can
fiery turtle
sure
strong heath
I gotta sort some other stuff right now that should let me debug it easier
fiery turtle
btw your github mirror is severely outdated
strong heath
yeah it is, thank you
fiery turtle
yeah I do, it is not coming from nowhere
Did you figure it out?
strong heath
I was fixing another thing, ill let you know when I get around debugging that
rustic compass
does uacpi support acpi 1.1? asking for a friend
rustic compass
i'm thinking about buying a am2 board with an athlon 64X2
and i would like to test uacpi sometime in the future on that
fiery turtle
uacpi should work on any piece of hw
and id be curious to see some of the old more fucked up aml
slow drift
and id be curious to see some of the old more fucked up aml
I do have an HP Mini 100e (made in 2009-2010) if you think it's old enough to be worth testing uACPI on it. I'm not sure what you'd define as "old AML"
fiery turtle
Damn
fiery turtle
I do have an HP Mini 100e (made in 2009-2010) if you think it's old enough to be...
Old is pre vista
But yeah 2009 is pretty old
leaden fox
that has an ACPI EC
I have a feeling that's gonna be saner than modern aml
slow drift
is there any image (OS) that I can test uACPI with on that notebook? I haven't ported uACPI to my kernel and even if I did, I'd have to write some proper tests...
I mean there are a bunch but like a non-cursed one
left orbit
managarm, pmOS, proxima, astral
north holly
obos
gentle peak
Is it 64 bit?
north holly
obos <:trl:829316641191886888>
only crashes on about 82.74% of real hardware that I have tested
slow drift
obos <:trl:829316641191886888>
non-cursed
slow drift
Is it 64 bit?
wait a sec I'm not sure
calm latch
this is a semi-recent build
slow drift
yeah it's 64-bit
slow drift
this is a semi-recent build
give me a few minutes and I'll test
for whatever reason it just shows a black screen - no Limine menu
I've previously booted Limine with no issues on this laptop so idk what's up
also on my PC it boots fine so it isn't an issue with the bootable media
the ISO supports legacy BIOS, right?
slow drift
weird
calm latch
From USB drive
I've just changed my build system to generate a hard disk image instead
gentle peak
this is a semi-recent build
for what it's worth this iso didn't do anything on my laptop while it was in bios mode
i had to put it in uefi mode to even get to limine
calm latch
I can upload it when I come home
vagrant hull
check how you build it and see if it is like the template
worth noting that a bunch of legacy BIOS systems won't boot GPT partitioned devices in legacy CSM mode but that's usually only if they know what a GPT is
slow drift
it doesn't have UEFI or CSM support, it's a BIOS system, so most likely it's an issue with the ISO
managarm does a thor panic related to AHCI or sth, so I couldn't test uACPI 
but Limine does boot, no issues on that
fiery turtle
Gn
slow drift
ty
calm latch
worth noting that a bunch of legacy BIOS systems won't boot GPT partitioned devi...
It's xorriso'es image
I'm too dumb for this, how do I install limine onto the image?
calm latch
where do i get limine?
left orbit
either v8.x-branch or compile it yourself
binary branch i think
sorry
v8.x-binary is the branch name
if you get that you can run make to build the host limine tool
calm latch
did you run `limine bios-install` on it?
I was running xorriso -as mkisofs -b boot/limine/limine-bios-cd.bin -no-emul-boot -boot-load-size 4 -boot-info-table --efi-boot boot/limine/limine-uefi-cd.bin -efi-boot-part --efi-boot-image --protective-msdos-label isodir -o $(ISO)
gentle peak
if you only do that, the iso's only bootable on bios if you're actually booting from a cd (or if you have a weird bios)
limine bios-install is necessary to make it bootable from a usb as well
calm latch
Does this look right?
limine/limine bios-install pmos.img
Physical block size of 512 bytes.
Installing to GPT. Logical block size of 512 bytes.
Secondary header at LBA 0x3ffff.
Secondary header valid.
GPT partition NOT specified. Attempting GPT embedding.
New maximum count of partition entries: 44.
Stage 2 to be located at 0x1a00 and 0x7ffd600.
Reminder: Remember to copy the limine-bios.sys file in either
the root, /boot, /limine, or /boot/limine directories of
one of the partitions on the device, or boot will fail!
Limine BIOS stages installed successfully!
(I mean I can just try it...)
gentle peak
seems fine yeah
vagrant hull
i mean assuming pmos.img is the previously created iso, yes
calm latch
It seems to have created a bunch of GPT entries...
left orbit
GPT partition NOT specified. Attempting GPT embedding.
limine embeds itself in the GPT
vagrant hull
you need to respect the size of the partition table
you're likely reading past the end of it
calm latch
It seems to have created a bunch of GPT entries...
I've just wrote this code so idk how it works
vagrant hull
lol
calm latch
it's probably wrong*
I am supposedly respecting the size
Anyway, this is the new hdd image
Which booted in QEMU with bios
bruh my wifi is so bad it can't push the commit over ssh
I'm doing this ```C++
struct GPTHeader {
uint8_t signature[8];
uint32_t revision;
uint32_t header_size;
uint32_t header_crc32;
uint32_t reserved;
uint64_t current_lba;
uint64_t backup_lba;
uint64_t first_usable_lba;
uint64_t last_usable_lba;
uint8_t disk_guid[16];
uint64_t partition_entry_lba;
uint32_t num_partition_entries;
uint32_t partition_entry_size;
uint32_t partition_entry_array_crc32;
};
// ... read gpt
auto gpt_entry_size = gpt->partition_entry_size;
auto gpt_entry_count = gpt->num_partition_entries;
auto gpt_partition_array_size = gpt_entry_size * gpt_entry_count;
// ... read gpt array
std::vectorDisk::Partition partitions;
for (size_t offset = 0; offset < gpt_partition_array_size; offset += gpt_entry_size) {
auto *entry = reinterpret_cast<GPTPartitionEntry *>(reinterpret_cast<char *>(gpt_ptr) + offset);
if (guid_zero(entry->type_guid))
continue;
std::string guid = guid_to_string(entry->type_guid);
printf("GPT partition: type %s, start %" PRIu64 ", end %" PRIu64 "\n", guid.c_str(),
entry->first_lba, entry->last_lba);
partitions.push_back({entry->first_lba, entry->last_lba});
}
@slow drift
./partitions.hh:42:15: error: static assertion failed due to requirement 'sizeof(GPTHeader) == 92'
42 | static_assert(sizeof(GPTHeader) == 92);
| ^~~~~~~~~~~~~~~~~~~~~~~
num_partition_entries
calm latch
num_partition_entries
It's 44
vagrant hull
yeah
vagrant hull
why should it be 2?
vagrant hull
it's the number of partition entries the table holds
free or non-free
44 means there are 44 available slots
calm latch
why am I reading past the table though
vagrant hull
idk, i am just assuming you are if you're reading garbage data
GPT isn't hard
you can easily figure it out with wikipedia alone, really
it's way more than enough
calm latch
I'm reading that stuff from entries 0, 1, 2, 3, 4, 8, 9
Maybe my AHCI driver is reading bogus data
And entries 8 and 9 contain the right data....
Which are lba_size * 2 
(my theory is that my driver is ignoring the start sector and reading from 0)
frank canopy
uacpi feature request (not a huge deal but might be nice): let me have (by a define probably) a separate alloc function for smallish object allocations compared to larger or variable size allocations
frank canopy
ig
i was thinking more for possibly pooling some of the structs vs variable sized stuff
but that would probably take a lot more work to add
maybe im just used to zig and passing around allocators lol
like it also might be cool to have support for an arena for if theres a lot of short-lived allocations during init or whatever
left orbit
just do it in ur alloc function
frank canopy
yeah ig
calm latch
Even EC works, nice
slow drift
yep
vast kestrel
Old is pre vista
I might have a 1999 PC with acpi now that I'm thinking about it
fiery turtle
daamn
fiery turtle
I might have a 1999 PC with acpi now that I'm thinking about it
maybe at least try to get a dump from it
calm latch
I can try porting pmOS real quick
vale isle
fastest port in the west??
strong heath
Did you figure it out?
So far all I have found out is that uacpi is asking to find a device going by (segment, bus, device, func) (0. 0. 0x10, 0x6), both Ironclad and Linux do not find such device
you will have my acpi dump on your dms shortly
fiery turtle
Some hw checks non existent devices and then checks if VDID == FFFF
but ill take a look
So tldr you must support that
strong heath
so what should I return in my bindings for those cases, should I check that, and if it is true, report a fake device?
fiery turtle
Make a fake device object that just reads at the correct place where that device would've been
strong heath
thats very stupid
north holly
Stupid hardware
fiery turtle
That's aml for you
fiery turtle
thats very stupid
Or you could unconditionally return FF on reads from it
But Linux just does what I suggested first
Aka it uses the raw pci helpers
fiery turtle
@strong heath lol i was right
Device (THC0)
{
Name (_ADR, 0x00100006) // _ADR: Address
Name (RSTL, Zero)
OperationRegion (THCR, PCI_Config, Zero, 0x0100)
Field (THCR, ByteAcc, NoLock, Preserve)
{
VDID, 32
}
If ((VDID != 0xFFFFFFFF))
{
it basically creates the body of the device only if its actually present
otherwise the body is skipped
hollow elm
amazing
fiery turtle
nice
how old are these
oh wait damn this is itanium
daamn GPE blocks live in system memory on it
its not hardware reduced acpi
[0ACh 0172 00Ch] PM1A Control Block : [Generic Address Structure]
[0ACh 0172 001h] Space ID : 00 [SystemMemory]
[0ADh 0173 001h] Bit Width : 10
[0AEh 0174 001h] Bit Offset : 00
[0AFh 0175 001h] Encoded Access Width : 00 [Undefined/Legacy]
[0B0h 0176 008h] Address : 00000000FF5C100C
Method (_PTS, 1, NotSerialized) // _PTS: Prepare To Sleep
{
Debug = Arg0
}
lol
fiery turtle
Huh
[094h 0148 00Ch] PM1A Event Block : [Generic Address Structure]
[094h 0148 001h] Space ID : 01 [SystemIO]
[095h 0149 001h] Bit Width : 10
[096h 0150 001h] Bit Offset : 00
[097h 0151 001h] Encoded Access Width : 00 [Undefined/Legacy]
[098h 0152 008h] Address : 0000000000000400
SystemIO on itanium?
thats the only mention of SystemIO
❯ grep -rni 'SystemIO'
rx2800-i2/dbgp.dsl:27:[028h 0040 001h] Space ID : 01 [SystemIO]
rx2800-i2/facp.dsl:106:[094h 0148 001h] Space ID : 01 [SystemIO]
rx2800-i2/facp.dsl:120:[0ACh 0172 001h] Space ID : 01 [SystemIO]
rx2800-i2/facp.dsl:141:[0D0h 0208 001h] Space ID : 01 [SystemIO]
rx2800-i2/facp.dsl:148:[0DCh 0220 001h] Space ID : 01 [SystemIO]
rx2800-i2/spcr.dsl:27:[028h 0040 001h] Space ID : 01 [SystemIO]
Debug = "Dumping OpReg Info..."
Debug = " DSIX is: "
Debug = DSIX /* \DRBL.SET_.DSIX */
Debug = " DCIX is: "
Debug = DCIX /* \DRBL.SET_.DCIX */
Debug = " SVNP is: "
Debug = SVNP /* \DRBL.SET_.SVNP */
Debug = " DBLL is: "
Debug = DBLL /* \DRBL.SET_.DBLL */
Debug = "Storing DSIX"
DCIX = DSIX /* \DRBL.SET_.DSIX */
Debug = "Storing SVNP"
SVNP = One
Debug = "Storing DBLL"
DBLL = One
most production ready aml
lol take a look at this
its hilarious
[uACPI][TRACE] [AML DEBUG] String => "Init of scratch RAM complete!"
[uACPI][TRACE] [AML DEBUG] String => "Fatal Error: "
[uACPI][TRACE] [AML DEBUG] String => "_SB_._INI"
[uACPI][TRACE] [AML DEBUG] String => "SCRAM Data struct access init failed."
[uACPI][TRACE] [AML DEBUG] Integer => 0x00000003
Fatal firmware error: type: 1 code: 1 arg: 1
[uACPI][ERROR] loop time out after running for 3 seconds
[uACPI][ERROR] aborting method invocation due to previous error: hanging AML while loop
[uACPI][ERROR] #0 in \CLIB.ERRR()
[uACPI][ERROR] #1 in \_SB_.SBA0._INI()
it literally
calls Fatal
then hangs
While (Ones)
{
Sleep (0x03E8)
}
are u insane
very cool to see anyway
server hw is built differently
north holly
I can tell 
north holly
0x3e8 = 1000
vast kestrel
maybe at least try to get a dump from it
Wasn't able to export the dump, but it does have a bit of funny stuff
It has a method named strc that compares two strings
It has a mutex object it uses for protecting some communication buffer
It has a power button device and an acpi power source device (whatever that is)
I might be able to export it, not sure
fiery turtle
It has a method named strc that compares two strings
Thats a classic
fiery turtle
I might be able to export it, not sure
USB?
vast kestrel
It's complex cause its a work machine
I can def test something on it, not sure I can bring stuff out of it
dense steppe
so I was looking through uACPI sources, and I found this:
Some kernels may want to use an in-kernel implementation of the rwmutex.
so... feature request (can make an issue on GitHub if you want me to): the ability to define a macro somewhere that tells uACPI to use an additional kernel API for rwmutexes.
fiery turtle
i mean sure that could be added ig, not sure how common that is tho
frank canopy
I will be having an rwlock in kernel idk about others
calm latch
I have full posix api in userspace so...
fiery turtle
barely anyone bothers reading about other existing uacpi options so im not sure its worth it
but ill think about it
deft canopy
https://pastebin.com/gUmDAVjN
truly the best AML ever, isnt it? :P
fiery turtle
not many people even replace libc stuff
fiery turtle
truly the best AML ever, isnt it? :P
yeah pretty funny, thanks for the dumps!
vast kestrel
I mean, is the rwlock used that much and is implemented that differently
fiery turtle
basically yeah, its used once in uacpi for ns walks
and if your kernel has it, is it like much faster or
vast kestrel
Big brain, instead of mutex and event just take a futex like api :^)
dense steppe
and if your kernel has it, is it like much faster or
It would really only replace (up to) two mutex acquisitions with one rwmutex acquisition (which can potentially be faster on high reader contention). It may in fact pessimise the writer (I find it hard to imagine an rwmutex writer acquisition being faster than a normal mutex acquisition).
dense steppe
not many people even replace libc stuff
That's because we're lazy (and for me, because I speedran uACPI --- currently working on IOAPIC stuff so I can actually handle events)
fiery turtle
true
dense steppe
It would really only replace (up to) two mutex acquisitions with one rwmutex acq...
However an in-kernel rwmutex would allow e.g. preventing writer starvation by making readers block even if the rwmutex is held by readers, if there is a writer waiting.
fiery turtle
truly the best AML ever, isnt it? :P
btw does itanium actually have io ports?
deft canopy
btw does itanium actually have io ports?
sorta, only for systems with x86 compat I believe
fiery turtle
interesting
deft canopy
if you need to access the IO ports from outside of x86 mode then you need to map them into a 64KiB region of the address space or smth
but idk if you can map them if the x86 compat mode isn't implemented on the processor
fiery turtle
another reason why uacpi should have map api for ports
deft canopy
but idk if you can map them if the x86 compat mode isn't implemented on the proc...
I suspect you probably cant, but presumably they'd be wired to nothing even if you could
vast kestrel
However an in-kernel rwmutex would allow e.g. preventing writer starvation by ma...
If it's needed then uacpi can implement itself
sterile egret
NICE
fiery turtle
median crest
i mean sure that could be added ig, not sure how common that is tho
I have funni rwlock
dusky glade
then hangs
```
While (Ones)
{
Sleep (0x03E8)
}
```
bru 😭
left orbit
does anyone have any insight into how uacpi uses memcpy/memset/memmove? i wonder if small copies are more dominant than bigger copies (>64 bytes)
i "wrote" (not exactly, with a little help of llvm ) some very optimized memory operation routines and i wonder if uacpi will make good use of that
not sure how fast they are in practice, but on paper (according to IACA 3.0) they should be able to process roughly 5-7x more data than a naive memcpy/memmove imlpementation, so i hope my work is not gonna go to waste
loud ice
i think most of the impact is from clang recognizing memcpy/memset as an intrinsic
and not from it being extra fasts
fiery turtle
does anyone have any insight into how uacpi uses memcpy/memset/memmove? i wonder...
most should be tiny
left orbit
i hope this will at least speed up my allocator
when i have to copy/zero out big blocks of memory
but honestly this doesn't happen very often, so idk
maybe i am just overengineering it 😭
rustic compass
could be usefull on process creation when copying elf segments around
left orbit
yeah i am quite far from that
but it will maybe slightly speed up my prekernel, since i am loading the kernel and modules manually :^)
rustic compass
could you share the algorithm?
left orbit
i am basically doing 64, 32, then single byte copies
unrolling those loops apparently produces very fast code
it's like 200 lines of assembly compared to 10-20 but yeah
i also addiitonally prefetch the next 64 byte blocks using prefetcht1, idk if that has any significance but it's there for both 64- and 32-byte cases
and the bytewise copy is done using a jump table where each case falls through, so basically a duff's device
rustic compass
could be significant as the cpu can copy stuff around in cache and at the same time fetch from ram
left orbit
hopefully, the IACA analyzer shows significant throughput increase for my optimized version
also my naive versions were backend bound where this one is frontend bound
though tbf it was tested against skylake uarch so idk how well it compares to modern uarches
uICA (https://uica.uops.info/) doesn't seem to support anything newer and i cant be bothered to set up the tools myself
my memmove (as i said, i helped myself a little bit with LLVM) was actually unrolled into a 128 byte loop so there's a little discrepancy between that and the memcpy, and i was too lazy to go back and unroll my memcpy too lol
rustic compass
tbh its somewhat annoying that cpus dont feature buildin memcpy/set engines like programming a vector register and the cpu does it in the background
like a simd extension just for memcpy/set
torpid ferry
urban galleon
fresh laurel
https://cdn.discordapp.com/attachments/1244456349341974679/1255204505792806984/t...
just stop making up words
fiery turtle
last commit yesterday
north holly
Uacpi thread dea
fiery turtle
and more like complete
calm latch
I'll try and test uACPI on my friend's Asus EEE 700 in a few days
fiery turtle
nice
north hull
I think the AML on my laptop is cursed
I get a good few errors with uacpi and acpica on linux too
urban galleon
where is the post for uDRM or the gpu thing
calm latch
I'll try and test uACPI on my friend's Asus EEE 700 in a few days
It's a 2007 Intel Atom netbook
fiery turtle
I think the AML on my laptop is cursed
what are they
north hull
what are they
I'll take a picture in a little bit
pine leaf
where is the post for uDRM or the gpu thing
urban galleon
thanks
fiery turtle
5 years
calm latch
After UltraOS gets Mesa?
fiery turtle
nah ill use one of the osses from here as the test bed
north holly
nah ill use one of the osses from here as the test bed
obos
By then I would definitely have a mesa port
100%
fiery turtle
nah by 5 years i mean a usable library
left orbit
uacpi 1.0 out, can't wait for the ukernel :^)
kind mantle
uLibc when? 
left orbit
uLibc when? <:troll:796858892985434152>
petition to rename mlibc to ulibc
slim panther
already exists
hollow elm
there's uclibc but i'm not aware of ulibc
slim panther
oh i probably mean that then
slim panther
\\\\\
hollow elm
\\<
vale isle
miaow
gentle peak
unsuccessful = timed out
oh and if the timeout is 0 you're supposed to do it without blocking at all
north hull
thats fine
wait so am I waiting for the counter to go to zero
and if it goes to zero before timeout its successfull
gentle peak
you're waiting for the counter to be non-zero
north hull
ah
gentle peak
if that happens before timeout, decrement it and return success
it's just a semaphore
left orbit
least fucked firmware
north hull
there's even some more that wraps the fb
north hull
least fucked firmware
lol
frank canopy
ran into an issue finalizing my new MM stuff and hooking it into uacpi
that being that uacpi doesnt support sized unmap
fresh laurel
hey im thinking on using uACPI on my project
does it need an allocator or specific stuff?
frank canopy
theres a header, kernel_api.h. you need to define the things in there
fresh laurel
oh ok
i already have an allocator so i guess im done
it uses a page allocator right?
or its byte based
hollow elm
it needs more or less malloc/free
fresh laurel
well shi, my allocator is page based
i guess its fixing time
i could make a suballocator
that allocates more pages if there are not enough bytes
frank canopy
i could make a suballocator
thats the normal way to do that yeah
fresh laurel
yeah
i still need to make a new allocator
the bitmap allocator is kinda bad for efi memory maps
frank canopy
does uacpi need to do mallocs bigger than a 2M large page?
if so then fuck that but if not then my new mm can at least handle every allocation it needs
fiery turtle
that being that uacpi doesnt support sized unmap
Wdym? It does tho
fiery turtle
<@353593948523986945> a bit late but here you go
What the fuck
Does your linux dmesg look similar?
Although acpica would abort the entire table on first duplicate method
fiery turtle
does uacpi need to do mallocs bigger than a 2M large page?
Its not that simple, for example it uses dynamic arrays to store a lot of things, so in theory it depends on the aml being executed, if it has an insane deeply nested expression for example
fiery turtle
that being that uacpi doesnt support sized unmap
Not sure what you mean but unmap takes a length parameter
north hull
Does your linux dmesg look similar?
fairly
it throws less errors
but yes
its absolutely cooked
fiery turtle
yeah thats because acpica cant handle graceful skipping of some ops and aborts right away
idk what your firmware developers were smoking
do they not test at all
fiery turtle
did u update your bios?
fiery turtle
then again ive seen a person here who has multiple copies of the same SSDT in their pc
north hull
lmao'
fiery turtle
I probably should do that huh
thats how they fix aml yeah
north hull
ill give that a shot
considering this is the main thing I use for testing
I'd like not to get blasted with errors every boot
lol
ah of course they ship bios updates as fucking windows executables
fuck
here is a snippet from dmesg
fiery turtle
💀
vale isle
I was browsing around in the coreboot docs as a girl does on a random friday
it looks like coreboot people have their own DSL for device trees that compiles to AML?
probably also compiles to FDT?
very interesting
device pci 15.0 on
chip drivers/i2c/generic
register "hid" = ""ELAN0000""
register "desc" = ""ELAN Touchpad""
register "irq" = "ACPI_IRQ_LEVEL_LOW(GPP_A21_IRQ)"
register "detect" = "1"
register "wake" = "GPE0_DW0_21"
device i2c 15 on end
end
end # I2C #0
Scope (\_SB.PCI0.I2C0)
{
Device (D015)
{
Name (_HID, "ELAN0000") // _HID: Hardware ID
Name (_UID, Zero) // _UID: Unique ID
Name (_DDN, "ELAN Touchpad") // _DDN: DOS Device Name
Method (_STA, 0, NotSerialized) // _STA: Status
{
Return (0x0F)
}
Name (_CRS, ResourceTemplate () // _CRS: Current Resource Settings
{
I2cSerialBusV2 (0x0015, ControllerInitiated, 400000,
AddressingMode7Bit, "\\_SB.PCI0.I2C0",
0x00, ResourceConsumer, , Exclusive, )
Interrupt (ResourceConsumer, Level, ActiveLow, Exclusive, ,, )
{
0x0000002D,
}
})
Name (_S0W, ACPI_DEVICE_SLEEP_D3_HOT) // _S0W: S0 Device Wake State
Name (_PRW, Package (0x02) // _PRW: Power Resources for Wake
{
0x15, // GPE #21
0x03 // Sleep state S3
})
}
}
for acpidumps we could probably compile all mainline device trees in coreboot to AML
that would be lots of chromebooks
so this is a great source of production arm AML
probably pretty much all chromebooks out there
fiery turtle
damn i thought this was gonna be a unique feature for shkwve
sterile egret
frank canopy
Not sure what you mean but unmap takes a length parameter
wait what? I'll have to double check then because my impl never has lol
ok yeah im just stupid apparently
fiery turtle
Lol
gentle peak
non-sized unmap would be horrible to deal with
frank canopy
yeah for whatever reason the zig export for it that ive got has never had a length param and ig last night i wasnt thinking enough to double check the header
i still gotta figure out why my new map thing does a safety-checked index out of bounds (which panics because im on a safe build mode) but thats a thing for later today
vagrant hull
i tested Ironclad on my laptop and trying to shutdown made it turn off the fans only but the laptop stayed on
and i had to force turn it off
i was scared it would overheat lol
i wonder if it's an Ironclad or uACPI issue
maybe missing EC handling or some crap?
fiery turtle
i wonder if it's an Ironclad or uACPI issue
Test a different OS to find out
Pmos/obos/managarm
fiery turtle
maybe missing EC handling or some crap?
Could be yeah
fiery turtle
Pmos/obos/managarm
Or proxima as well
Anything in logs BTW
The fact that it stayed on tells me prepare to sleep probably failed
vagrant hull
i see
vagrant hull
no
vagrant hull
alright
gentle peak
don't use the latest one from github actions, that one's using a just-started kernel rewrite
fiery turtle
Wait
gentle peak
doesn't have uacpi yet
vagrant hull
this?
fiery turtle
A rewrite AGAIN?
gentle peak
yeah well the rewrite was kinda necessary considering i'm completely changing what i want the kernel/userspace boundary to look like
fiery turtle
it was going so well 
fiery turtle
yeah well the rewrite was kinda necessary considering i'm completely changing wh...
Whats the new idea?
gentle peak
microkernel
vagrant hull
i love how it's always like "yeah just go dig in logs and determine which of the many arbitrarily thrown around ISOs from days ago is safe to test with"
gentle peak
wanted to be able to write drivers in funny languages like java and python
that's really the main reason
fiery turtle
Interesting
gentle peak
this is the latest pre-rewrite image
fiery turtle
Although proxima might not shutdown because your laptop probably routes the power button via ec so it just won't get the notif
fiery turtle
Is it a pc or a laptop
vagrant hull
no
fiery turtle
Then its not routed via gpio at least
gentle peak
you never know
fiery turtle
True
gentle peak
wanted to be able to write drivers in funny languages like java and python
speaking of which, i wonder how uacpi would perform if the kernel api was written in javascript
fiery turtle
💀
gentle peak
yeah probably
but then what if the mmap regions for libuacpi.so are backed by an ahci driver written in python
this is the kinda stuff i wanna experiment with
gentle peak
i thought of that but the old plan for proxima was Yet Another Monolithic Unix Clone and i realized that i would lose motivation to work on that extremely quickly
vagrant hull
i booted it, it says uncompressing init, then it doesn't say anything after that, then i press the power button, it says shutting down in 3/2/1, and it shuts down (i assume successfully)
fiery turtle
Does the laptop shutdown?
vagrant hull
yes
fiery turtle
Ironclad bug ig
gentle peak
i booted it, it says uncompressing init, then it doesn't say anything after that...
you mean executing init followed by hello world? because if nothing's printed after uncompressing init it shouldn't have even init'd uacpi yet
gentle peak
yeah
fiery turtle
Lol
vagrant hull
because the hello world is at the top
gentle peak
i guess it could wrap around if your laptop has a very large number of acpi tables
fiery turtle
I mean if it says shutdown in 3 it did init successfully
vagrant hull
it has a bunch of ACPI tables
gentle peak
this is the qemu output for that iso fwiw
fiery turtle
this is the qemu output for that iso fwiw
I'm gonna miss this kernel
vagrant hull
btw for the curious it does 1.2M
fiery turtle
this is the latest pre-rewrite image
This is 1.0 correct?
gentle peak
also side note i have no idea what happened because i changed literally nothing in the kernel (i've mostly been working on userspace) but a few days ago the uacpi performance spiked up from 7.0-7.2M to 7.5-7.6M
gentle peak
This is 1.0 correct?
let me check
fiery turtle
also side note i have no idea what happened because i changed literally nothing ...
Damn
gentle peak
it's uacpi 1bd2930030ef819519de32dd1f7a8f999440308f
which is the first commit of 1.0
fiery turtle
Ok nice
frank canopy
ok time to start adding printfs to uacpi i guess, this is looking like a use-after-free of map/unmap stuff in early table access rn and i cant get my interrupt handler to give me the last few stack frames correctly sadly
frank canopy
...it also might help if i set the log level so my printfs actually get executed
fiery turtle
Your issue confused me since I read it right as I woke up and I was like wait it doesnt have a length?
gentle peak
I'm gonna miss this kernel
next time proxima's benchmark score gets updated it probably won't be at the top anymore
fiery turtle
Indeed
gentle peak
gonna miss that
fiery turtle
That's why I want this ukernel to be a separate project
gentle peak
well it's not like i'm deleting the repo or something, right now the rewrite's a separate git branch and once it's mature enough i'll just rename main to pre-rewrite
fiery turtle
Is it gonna have full smp support btw
fiery turtle
well it's not like i'm deleting the repo or something, right now the rewrite's a...
Oh well
gentle peak
Is it gonna have full smp support btw
sort of? i'm planning to fully take smp into account where i can but i'm probably never gonna get around to making a proper smp scheduler
frank canopy
Your issue confused me since I read it right as I woke up and I was like wait it...
yeah that was just me being tired and missing it in the header lol
vale isle
i thought of that but the old plan for proxima was Yet Another Monolithic Unix C...
yet another microkernel unix clone is more interesting and you can do your funny stuff later
fiery turtle
Lol
fiery turtle
sort of? i'm planning to fully take smp into account where i can but i'm probabl...
Drivers in go or python could be fun ig
vale isle
yet another microkernel unix clone is more interesting and you can do your funny...
have your cake and eat it too (have your C/rust/c++ drivers and have your JavaScript ones too
frank canopy
oh ok cool it isnt a bug in uacpi that im getting rn at least
its my own problem that i dont know where its coming from 
but at least i managed to confirm it isnt trying to unmap the address it then page faults on
so back to my own thread for debugging hell ig
frank canopy
if i comment this one line out and it fixes it im gonna be mad ngl
the line in question being the unmap-the-ptes line
it didnt \o/
fiery turtle
frank canopy
oh i bet i know it
frank canopy
ok so dump prints fine and debug prints fine and memcpy page faults accessing the header and i dont know why
tables.c L563
fiery turtle
well the header is the virtual address you have mapped right
whats the value of cr2
frank canopy
fffffc001ffff324 which is the thing that map returned
dense steppe
does your memcpy return dst?
frank canopy
my memcpy is builtin_memcpy
dense steppe
throwback to a bug I had while integrating ACPICA into a previous kernel of mine. it failed because my (then nonconformant) memcpy impl didn't return dst
fiery turtle
that calls into normal mecmpy
frank canopy
just run info tlb ig or whatever
did and it isnt mapped anymore and that has me very confused
dense steppe
and no compile error because the impl was in asm and C land used the builtin
fiery turtle
did and it isnt mapped anymore and that has me very confused
wtf
frank canopy
that calls into normal mecmpy
which is zig's @memcpy in zig compiler_rt here
fiery turtle
what if u disable your libc overrides entirely and let uacpi run with builtins
fiery turtle
thats good ig
left orbit
fwiw my zig kernel used builtins and it worked fine
frank canopy
unless this is some compiler nonsense
left orbit
so that shouldn't be the problem here
frank canopy
i think it cached wrong gimme a sec
idk why it wouldnt have invalidated build cache on changing the headers tho
ok yeah same error using uacpi's version
https://gist.github.com/Khitiara/c7fce93c47de7474563d39c21241dfa2 heres my log btw, ffffffff8013365c is tables.c L184 and ffffffff80133922 is in uacpi_memcpy
in case anyone sees something im missing
ig ill put a debug after the memcpy call just in case its a different one somewhere idk
alright no i was wrong its a different memcpy call
ok it turns out it was this one on line 105
fiery turtle
btw why cant u just use gdb
frank canopy
i cant figure out how to get it to breakpoint in my code booting from limine
fiery turtle
disable kaslr
frank canopy
it isnt on gdb just refuses to let me put in the breakpoint
dense steppe
gdb + qemu is in general just goddamn awful to setup lol. been there, tried that, gave up and did hcf debugging instead
left orbit
you need hbreak with kvm
fiery turtle
with tcg it should work flawlessly
frank canopy
ig i can try again at some point
also with qemu+gdb being annoying i find printf stuff easier
oh GOD if this is the problem im gonna feel so dumb
calm latch
microkernel
2025 microkernels Renaissance
frank canopy
oh GOD if this is the problem im gonna feel so dumb
it was the problem
turns out using max instead of min is a problem sometimes
i needed it to clear up to the first two of a thing with given length. put in max instead of min so it was overrunning when length was 1 and nuking the next pte along which was the rxsdt
fiery turtle
bruh
frank canopy
and it took me a while to find the issue because the table load was recursing to get the dsdt and that made me think the problem was later than it was
anyway now its getting past all the uacpi things and then i think dying in my code later
yep dying in my code now
so all good here
fiery turtle
nice
mortal yoke
i want to make one too...
same tbh, though I already have two osdev projects so I think I will not at least for now lol
mortal yoke
I could theoretically still turn the nt thing to a microkernel because its that early but eh I don't think I want to do that
gentle peak
yeah the reason i didn't make the new proxima iteration a separate project is because it'd have been my 3rd one lmao
dense steppe
I haven't really understood what would make microkernels easier to develop than monolithic kernels.
like, sure, the 'kernel' is much smaller and doesn't have to contain IO or filesystem drivers, but you've still got to do that work somewhere, except now it is in userspace and your kernel needs an appropriate form of IPC.
mortal yoke
also kinda forces you to have a decent driver interface lol
north holly
obos rewrite 6 will be an astral fork
left orbit
yeah splitting everything into smaller pieces doesn't make development easier but it does at the same time
gentle peak
it forces you to have good abstraction layers
gentle peak
with microkernels you can't make a quick hack that bypasses a bunch of abstraction and oops 100k loc later it's very tightly integrated
left orbit
but with a microkernel if you want some piece of information you need an api to get it out of the kernel
gentle peak
except 'this shiy' is in a different userspace process and the kernel doesn't even know it exists or how to find it
dense steppe
currently I am writing a monolithic kernel, and I will continue with that until it is self-hosting and has actually good userspace. but there might come a day when I decide to rewrite (or atleast redesign core subsystems). I believe that I learn things about everything that I implement in every iteration that gives me the ability to make the next iteration better. and I believe this is mostly true regardless of whether you're writing a monolithic or a microkernel.
hallow marten
I haven't really understood what would make microkernels easier to develop than ...
microkernels are harder
frank canopy
oh crap i forgot i set it to debug level to find that bug earlier
calm latch
but with a microkernel if you want some piece of information you need an api to ...
You can just have everything in userspace
left orbit
or that, yeah
you usually need some sort of protocol so processes can talk to each other
so i guess that's the abstraction
dense steppe
the thing with that is, you're still going to have to implement everything that you'd need to implement in a monolithic kernel.
tradeoffs
calm latch
the kernel just needs to have a good IPC interface
and some other things
i think minix does pmm and vmm in userspace
left orbit
yeah, all the kernel really needs is a way to create a thread, start executing user code and manipulate virtual memory
oh
interesting
hallow marten
yeah, all the kernel really needs is a way to create a thread, start executing u...
nah
IPC
left orbit
i prefer memory management in kernel though
calm latch
i think minix does pmm and vmm in userspace
yeah that is extreme
dense steppe
yeah but like... you don't have to write device drivers inside of the kernel... but if you want an OS, you're still going to have to write device drivers
calm latch
but you can write drivers in userspace
left orbit
yeah but the abstractions needed to write them wont write themselves
calm latch
which implies you can do cursed things
left orbit
and that generally determines the quality of everything
and how enjoyable writing new drivers is
hallow marten
a microkernel system is more complex and slower
calm latch
yeah but the abstractions needed to write them wont write themselves
what, you just need to map memory and have a good strategy for interrupts management
hallow marten
you can have a perfectly fine driver interface in the kernel
calm latch
a microkernel system is more complex and slower
not (slower) by much
and doesn't matter probably
dense steppe
with modern CPUs it doesn't matter tbh
gentle peak
a microkernel system is more complex and slower
complexity can be fun if it isn't due to bad design, and if you want maximum performance writing your own os isn't the best option in the first place
jaunty fox
not (slower) by much
from what i have looked at microkernels they are largely behind the times when it comes to modern scalability measures
they don't have to be
kind mantle
a microkernel system is more complex and slower
Is it slow if the vast majority of CPU time is spent in userspace normally anyway?
jaunty fox
as we saw in #ipc message microkernels (or microkernel-like systems) were looking at these issues historically
hallow marten
Is it slow if the vast majority of CPU time is spent in userspace normally anywa...
that is not the problem
the problem is constant context switches
and cache invalidation
jaunty fox
but microkernels have suffered probably from their unfortunate relegation to mostly embedded applications
and embedded people do not look at scalability
gentle peak
I think they're talking about general cache invalidation, afaik the main issue with microkernels for that is that what would be syscalls on monolithic kernels often get processed by daemon threads running on different CPUs
dense steppe
if a microkernel achieves the same end result as a monolithic kernel, are they not one and the same thing? 
kind mantle
if a microkernel achieves the same end result as a monolithic kernel, are they n...
no
gentle peak
so the inputs and outputs of those 'syscalls' aren't in the cache for the server and the client
calm latch
I think they're talking about general cache invalidation, afaik the main issue w...
you can spawn per cpu threads for servers
I don't see it as an issue
gentle peak
yeah but most microkernels don't bother making sure that the server thread that accepts the message is running on the same cpu as the sender
gentle peak
most microkernel problems are implementation issues yeah
jaunty fox
yeah but most microkernels don't bother making sure that the server thread that ...
but is that always best? see #ipc message for a counterpoint