especially since most stuff is stubbed to a no-op
#uACPI - a portable and easy-to-integrate ACPI implementation
fiery turtle
fiery turtle
lol
slim panther
what do i pass to table_find
fiery turtle
what do i pass to table_find
"HPET"?
slim panther
doesn't work
fiery turtle
show code
vagrant hull
i should maybe just return rdtsc
dusky glade
@slim panther
fiery turtle
i should maybe just return rdtsc
kvm_clock to convert tsc to ns
fiery turtle
_by_signature
dusky glade
uacpi_table_find_by_signature
slim panther
ah
dusky glade
https://github.com/rayanmargham/NyauxKC/blob/master/kernel/src/timers/hpet.c some reference code from nyaux if u need it
slim panther
wait and how do i convert this to the actual table
vagrant hull
panic("The Nyaux Kernel does not support Devices without a HPET\nThe "
"Reason for this is becausei hate tsc with a burning desire. pr in a "
"different calibration timer if u want one so bad lol");```loli have the best spelling
slim panther
is there a hpet_table struct?
dusky glade
😡
fiery turtle
is there a hpet_table struct?
acpi.h
north holly
is there a hpet_table struct?
acpi_hpet
dusky glade
```c
panic("The Nyaux Kernel does not support Devices without a HPET\nThe "
...
wait why did i say devices
was i on a sleep tablet when i wrote that
💀
fiery turtle
were you high when writing that
dusky glade
no
but the sleep med i take before i go to sleep
gives me a eurphoic response as a side effect
before it makes me sleepy
fiery turtle
hpet works but tsc + kvm_clock is lightning fast and accurate
dusky glade
im cold
vagrant hull
im hot
vagrant hull
i am actually hot tho
vagrant hull
yeah
i started wearing a heavier tshirt than usual
and it doesn't take much for me to start feeling hot
vagrant hull
i think so
dusky glade
play co-op with me
vagrant hull
no i need to wire this hpet thing in
north holly
no i need to wire this hpet thing in
Look at obos code, if you dare https://github.com/OBOS-dev/obos/blob/userspace-work/src%2Foboskrnl%2Farch%2Fx86_64%2Ftimer.c
rustic compass
does uacpi have a stripped down version that only contains the table api? i want to put uacpi in userspace but need some tables at boot and dont want to have uacpi two times
fiery turtle
well early_table_access mode exists, which only needs map and unmap, but there isn't a mode which removes references to other apis
i had that feature idea for later
rustic compass
its still a bit until i need uacpi, maybe by the time thats supported
fiery turtle
right now u can stub out all other api to just be unimplemented and it will work fine technically
if you use LTO it will probably remove all other code
rustic compass
does uacpi move the tables or just copies them to an internal buffer?
fiery turtle
neither
the internal buffer stores state needed to access those tables
like the virtual address where they're mapped, reference count etc
rustic compass
so it preserves the tables in memory?
fiery turtle
yeah
rustic compass
so that two instances could access it
fiery turtle
yeah ofc
vagrant hull
is 100MHz the right hpet freq for qemu?
dusky glade
idk about mhz but for nanoseconds for me its 1 hpet tick per 10ns
slim panther
@fiery turtle success! need to impl the rest
fiery turtle
nice nice
mortal yoke
is 100MHz the right hpet freq for qemu?
yes
vagrant hull
ty
fiery turtle
bro good thing i decided to integrate uacpi into linux
it found actually important warnings
like functions that should be static not being static
loud ice
but you do have spinlocks right?
vagrant hull
yeah
fiery turtle
lmao how
vagrant hull
wdym how?
loud ice
what if you take a spinlock instead of a real mutex acquire
median crest
Mint secretelt has a phd
vagrant hull
what if you take a spinlock instead of a real mutex acquire
as i mentioned above, my kernel panics if you try to take a spinlock with ints enabled
vagrant hull
whats your allocator
it's a really simple bitmap allocator, actually
vagrant hull
it's the most simple C memcpy ever lol
loud ice
something seems not quite right
fiery turtle
^
vagrant hull
i am returning time like this, idk, maybe it's an issue? c uacpi_u64 uacpi_kernel_get_nanoseconds_since_boot(void) { return hpet_read_counter() / (1000000000 / hpet_frequency); }
hpet_frequency is in Hz
read_counter() reads the main hpet counter straight from mmio
dusky glade
...
loud ice
i am returning time like this, idk, maybe it's an issue? ```c
uacpi_u64 uacpi_ke...
it is
i think
vagrant hull
what's the issue?
loud ice
your calculation seems to be a little bit off
dusky glade
i am returning time like this, idk, maybe it's an issue? ```c
uacpi_u64 uacpi_ke...
yuh
vagrant hull
yeah
loud ice
your clock runs 100 times fast
vagrant hull
how come
vagrant hull
uint64_t counter_clk_period = tmp >> 32;
hpet_frequency = 1000000000000000 / counter_clk_period;```i use this division
dusky glade
```c
uint64_t counter_clk_period = tmp >> 32;
hpet_frequency = 100000000...
thats wrong
dusky glade
yea
vagrant hull
then it's right?
dusky glade
huh
loud ice
i am returning time like this, idk, maybe it's an issue? ```c
uacpi_u64 uacpi_ke...
i think this timekeeping logic is off
vagrant hull
it's volatile
dusky glade
okay
vagrant hull
i think this timekeeping logic is off
yeah but how?
loud ice
the correct way to calculate it is hpet_ticks() * ticks_per_ns
dusky glade
^
loud ice
ticks_per_ns = 1e9 / freq_in_hz
dusky glade
the correct way to calculate it is `hpet_ticks() * ticks_per_ns`
this is what i also do
loud ice
so the correct value is hpet_ticks() * 1e9 / freq_in_hz
fiery turtle
ld: (.text+0x6331f3): undefined reference to `uacpi_kernel_free'
ld: (.text+0x633205): undefined reference to `uacpi_kernel_free'
ld: (.text+0x633216): undefined reference to `uacpi_kernel_free'
ld: (.text+0x633223): undefined reference to `uacpi_kernel_free'
ld: (.text+0x63337a): undefined reference to `uacpi_kernel_alloc'
ld: vmlinux.o: in function `uacpi_free_namespace_node_info':
(.text+0x6334aa): undefined reference to `uacpi_kernel_free'
ld: vmlinux.o: in function `uacpi_device_matches_pnp_id':
(.text+0x633550): undefined reference to `uacpi_kernel_free'
ld: (.text+0x633578): undefined reference to `uacpi_kernel_free'
ld: (.text+0x633585): undefined reference to `uacpi_kernel_free'
ld: vmlinux.o: in function `uacpi_get_pci_routing_table':
(.text+0x633a75): undefined reference to `uacpi_kernel_alloc'
ld: (.text+0x633bd1): undefined reference to `uacpi_kernel_free'
ld: vmlinux.o: in function `uacpi_free_pci_routing_table':
(.text+0x633dea): undefined reference to `uacpi_kernel_free'
ld: vmlinux.o: in function `uacpi_free_dynamic_string':
(.text+0x633e1a): undefined reference to `uacpi_kernel_free'
oh my god
fiery turtle
now to implement these
dusky glade
this is just kernel api not implemented
i know
loud ice
uACPICA when
slim panther
i have this lol
return hpet.regs->main_counter * (hpet.period / 1000000);
im not saying oh my god to that
loud ice
i have this lol
```c
return hpet.regs->main_counter * (hpet.period / 100...
yeah that also works
loud ice
though i'd calculate hpet.period / 1... statically
the best way to do it is to do what i do iirc
fiery turtle
drivers/acpi/uACPI/source/utilities.c:931:26: error: no previous prototype for ‘find_one_device’ [-Werror=missing-prototypes]
931 | uacpi_iteration_decision find_one_device(
| ^~~~~~~~~~~~~~~
This warning should really be enabled everywhere
loud ice
or actually idk if i do it
median crest
Just use kvmclock and get a free nanosecond clock 
dusky glade
FINALLY BACK TO EARTH
loud ice
that seems a lot more accurate
slim panther
dusky glade
NYAUX IS CLOSE TO YOU
vagrant hull
not sure
dusky glade
HEAR ME ROAR MINT NYAUX WILL DESTROY CYNIX INTO THE GROUND
vagrant hull
seems a bit low
slim panther
i will submit a benchmark later
loud ice
theres no way you are doing 5 million though
fiery turtle
^
vagrant hull
i mean i am also surprised if i do 5 mil
but 50k seems a bit low too lol
both values seem off
dusky glade
but 50k seems a bit low too lol
no u deserve that spot
fiery turtle
with a bitmap allocator its imaginable
loud ice
yeah i'd assume thats why its slow
slim panther
menix has phd memcpy
dusky glade
menix has phd memcpy
show
mortal yoke
btw nyaux has a optimized memcpy and memset
its also likely not optimized for small copies
loud ice
its a phd memcpy so it must be good
also lets go pitust shitos 10000 uacpi race edition
how do i make an iso again
mortal yoke
dw im gonna steal memix's
just add an if and do a simple byte loop for small ones
loud ice
bye
median crest
Cya
dusky glade
the timer seems 10 times off
no its completely accurate this is cope
dusky glade
just add an if and do a simple byte loop for small ones
if statement adds more cycles 😡
mortal yoke
it doesn't really matter
dusky glade
what if statement should i check for how big of a count
mortal yoke
though tbf the rep overhead for small copies doesn't likely matter that much anyway
dusky glade
though tbf the rep overhead for small copies doesn't likely matter that much any...
no i want phd memcpy in nyaux
so
dusky glade
what if statement should i check for how big of a count
anwer
anwer
bro i cannot spell
help
median crest
Just rely on memory being already initalized with the values you want and panic otherwise
mortal yoke
anwer
idk I don't remember, you could test it yourself with gbenchmark
probably something around 128 or whatever
dusky glade
idk I don't remember, you could test it yourself with gbenchmark
ill js wait for mr marvin to come home and to steal bros memcpy 
easy
torpid root
the timer seems 10 times off
rep movsb and a faster allocator fixes that
loud ice
i know your step 1.5 of adding 10 memory corruption bugs
dusky glade
dusky glade
actually no u need to do step 0.5
well i was gonna say step 0.5 is to make a 16 bit bootloader and calling it your operating system then ask in #osdev-misc-0 how to make windows
everyone had that phase ngl
or ask chatgpt to generate ur entire kernel
dusky glade
my step 0.5 was learning c
fair
north holly
obos could never
Obos does though?
dusky glade
😡
median crest
Stable
north holly
Stable enough
Rather
Doesn't crash when you blow on it once
Only when you mess with code that you shouldn't mess with
Or mess with code with // Abandon all hope, ye who enter here
dusky glade
😡
fiery turtle
i think most submissions are now done so I'll probably close the initial list in like 3-4 hours
loud ice
lmfao rust's shitty compiler-builtins doesnt support __popcountdi2 because its not called by llvm apparantly?
lmao
mortal yoke
lmfao rust's shitty compiler-builtins doesnt support `__popcountdi2` because its...
seems to line up with my findings on clang, it generates inline code for the popcnt builtin
(when its not supported natively, if there is the popcnt instruction then it will use that)
loud ice
oh huh
interesting
i just called rust count_ones
which doesnt generate the call
i think
hopefully lol
okay
i now have uacpistub os
fiery turtle
that could actually be very useful
loud ice
lol torture test
dusky glade
why rust @loud ice
fiery turtle
he wants his os optimized out
torpid root
when the compiler inlines everything into a single function
loud ice
why rust <@696410219663851552>
because
loud ice
when the compiler inlines everything into a single function
i mean that is the definition of performance :^)
least schizo rust code
torpid root
i think most submissions are now done so I'll probably close the initial list in...
but mah new allocator that I'm definitely not procrastinating writing!
It will make my score surpass managarm!
fiery turtle
i can wait for a bit if you're actually serious
fiery turtle
u can always add in the updated score later
but if u actually finish it today ill add it
north holly
u can always add in the updated score later
If I switch to a bump allocator for uacpi and I get better perf can I use that as my score /hj
vast kestrel
I wonder if its possible to somehow hack it so uacpi will have an arena
fiery turtle
i mean if u make a private heap for uacpi probably
vast kestrel
well, some things change global context
north holly
Wassan arena
vast kestrel
I guess not really, since anything changing the namespace will have to be kept
fiery turtle
yea
dusky glade
he wants his os optimized out
lmao
torpid root
osdev teams up to get obos to be faster than managarm
fiery turtle
microsoft with 1B dollar sponsorship to get obos to 5M/s
mortal yoke
just run it on a good enough computer
north holly
osdev teams up to get obos to be faster than managarm
I think most people have a stroke reading obos code, unfortunately
And so do I
I avoid reading most obos code
mortal yoke
lol
loud ice
fiery turtle
Big
mortal yoke
lol
I think my "worst" code is my page allocator mostly because originally it had fine-grained locking but then I just slapped a giant lock over the whole thing after I realized there were issues with the locking
but otherwise its fine I think
north holly
My worst looking code is probably everything
median crest
My biggest issue is vmm and mutex probably
loud ice
hits a todo on mutexes
north holly
My worst looking code is probably everything
If I had to choose one thing I'd look into the pf handler, which works, but the code is messy
north holly
If I had to choose one thing I'd look into the pf handler, which works, but the ...
Actually, it's the x86_64 entry code
In the part that loads kernel symbols into a tree
mortal yoke
I like this part of my kernel 
north holly
I like this part of my kernel <:troll:796858892985434152>
Uh oh
mortal yoke
jump to higher half + stack pointer adjust for hhdm
loud ice
lmfao
wait
what
oh
ohhhh
but why
this is maximum ub
unless arch_start is an assembly self-relocator
gentle peak
I'm pretty sure there's a double-release bug here
don't see how it could be an issue in my kernel, but can anyone confirm? seems like it happens if uacpi_kernel_install_interrupt_handler unconditionally returns an error
mortal yoke
unless arch_start is an assembly self-relocator
the way it works is that I rely on the compiler generating relocation-less pie code with -fno-PIE for the entry code (before that asm) and the base of the kernel is set to higher half in the linker script so the functions/globals after that are at the correct addresses
north holly
don't see how it could be an issue in my kernel, but can anyone confirm? seems l...
If this is the qemu blob, then I am pretty sure I assert stuff like that in my mutez
*mutex
Wait I don't
loud ice
the way it works is that I rely on the compiler generating relocation-less pie c...
uuuuuuu
thats
bad
vagrant hull
whatever i'm back
north holly
I just return success if the lock is already released
gentle peak
If this is the qemu blob, then I am pretty sure I assert stuff like that in my m...
No the blob loads fine, the timing info is printed earlier, this seems to happen because my uacpi_kernel_install_interrupt_handler returns UACPI_STATUS_UNIMPLEMENTED
fiery turtle
I'm pretty sure there's a double-release bug here
Can u get a backtrace of where its called from?
mortal yoke
bad
it works tho, idk if there would have been any other way to do it without needing a prekernel
gentle peak
Can u get a backtrace of where its called from?
The panicking release?
This one has all the line numbers
vagrant hull
whatever i'm back
the hpet counter increments by 1 billion a second, so it's not 100MHz so it's off by 10
mortal yoke
it works tho, idk if there would have been any other way to do it without needin...
anyway this is unrelated to uacpi, maybe go to #1097930216032178276 for the reply if you want lol
vagrant hull
the hpet counter increments by 1 billion a second, so it's not 100MHz so it's of...
so it's not cope
vagrant hull
i wouldn't say that yet
but i believe so
i need to make sure i am doing time properly
gentle peak
anyway, i'm pretty satisfied with this performance considering i started this kernel rewrite literally today
loud ice
500k sounds fine i think
fiery turtle
anyway, i'm pretty satisfied with this performance considering i started this ke...
That's quite decent
Wanna get added to notable projects?
loud ice
@fiery turtle am i allowed to call the handler immediatly in uacpi_kernel_schedule_work?
slim panther
he wants his os optimized out
poweron and poweroff state are the same so optimize everything out
fiery turtle
<@353593948523986945> am i allowed to call the handler immediatly in `uacpi_kern...
Ofc
loud ice
assuming im not in an irq context
vagrant hull
the hpet counter increments by 1 billion a second, so it's not 100MHz so it's of...
ok i don't think it's off, i think i got it wrong, it's actually 50k/s lol
loud ice
did you do the "spin in a loop for 1 second worth of hpet ticks" thing
vagrant hull
i am just printing it over and over and it does 1 billion ticks per second
fiery turtle
I think I designed it specifically so lazy impl like that should work
gentle peak
Wanna get added to notable projects?
Nah, the source isn't public yet and won't be for a while
fiery turtle
Ah ok
vagrant hull
i am just printing it over and over and it does 1 billion ticks per second
it actually makes me wonder if @median crest could do the same and verify if it's the case for him as well
that it does 1 billion ticks a second
fiery turtle
So yeah its probably the bitmap holding you back
fiery turtle
if i do nothing uacpi uses the memcpy and friends i provide right?
Nope the other way around
vagrant hull
so how do i make it use my memcpy
fiery turtle
Take a look at libc.h
vagrant hull
tedium
slim panther
why not ship a hyper optimized memcpy 
fiery turtle
Indeed
vagrant hull
do i really have to do -Duacpi_memcpy=memcpy???
mortal yoke
so how do i make it use my memcpy
you add -DUACPI_OVERRIDE_LIBC and create an uacpi_libc.h somewhere in include path where you do (for an example) ```c
#define uacpi_memcpy __builtin_memcpy
#define uacpi_memset __builtin_memset
^
vagrant hull
💀
fiery turtle
What lol
fiery turtle
Taking suggestions
vagrant hull
like i wouldn't provide the impls at all
everyone should realistically have those functions anyways
loud ice
yeah
its also forbidden to not have those functions afaik
because gcc/clang will call them anyway
vagrant hull
the mem* ones yes
fiery turtle
Maybe but whenever I want to port uacpi its annoying af to implement all missing ones
loud ice
for memcpy on gcc i believe theres some funny meme you can do
vagrant hull
then at least do that for the mem* ones
fiery turtle
Id rather have my own trusted ones by default than have some beginner complain about uacpi bugs
slim panther
aren't you forced to have memcpy for struct derefs?
slim panther
because
my_struct = *struct_ptr;
invokes memcpy
loud ice
imo you should default to calling memcpy anyway
fiery turtle
Yup
vagrant hull
you know what may slow down my impl
a lot
the fact that i memset allocations to 0 always
and that i poison frees
fiery turtle
Definitely
loud ice
probably lol
vagrant hull
and that the memset is naive
loud ice
do you have profiles?
vagrant hull
no
i mean i can use the qemu profiler actually
i have it here in the repo handy
let's see
fiery turtle
True
median crest
I wanted to use that shell script profiler but uacpi was too fast
loud ice
do it in a loop
loud ice
tbh yeah
i think llenotre's qemu instrumentation thing might work?
idk
or alternatively you could try copious amounts of dtrace?
vagrant hull
or try without kvm actually lol
fiery turtle
Yeah actually
vagrant hull
it seems like a lot of time is spent in vmm.c
fiery turtle
This one has all the line numbers
now that i think about it, can u give me the name of the node that is being uninstalled here?
vagrant hull
it seems like a lot of time is spent in vmm.c
some of it in spinlocks
fiery turtle
i dont understand why it enters this branch
fiery turtle
that means u cause a lot of vmexits probably
median crest
Check for lots of cr8 writes or something
slim panther
@fiery turtle why is thread_id a void*?
fiery turtle
recursive mutexes
why do u ask btw, are your thread-ids non representable as void*?
gentle peak
now that i think about it, can u give me the name of the node that is being unin...
this one? or am i looking in the wrong func/var?
slim panther
no they are, they're just usizes and i thought maybe you want a thread pointer
fiery turtle
this one? or am i looking in the wrong func/var?
correct yeah
vagrant hull
replacing the mem functions with rep movsb and friends barely made any difference
fiery turtle
thanks
fiery turtle
this one? or am i looking in the wrong func/var?
which unimplemented caused this?
slim panther
gentle peak
which unimplemented caused this?
the only unimplemented thing that's being called is uacpi_kernel_install_interrupt_handler, but i haven't checked whether that's the actual cause of the issue, it could be unrelated
fiery turtle
the only unimplemented thing that's being called is `uacpi_kernel_install_interr...
install_interrupt_handler is called way before namespace_load
gentle peak
i'm logging all calls to unimplemented functions, and that's the only one that gets logged, and it gets logged right before the panic
gentle peak
set a breakpoint, this is the very first call to the function
fiery turtle
nvm i dont know how uacpi works apparently lmao
fair enough i must've moved it at some point
will fix it shortly
vagrant hull
ok i made it not poison on frees and not zero on allocs when not needed, and now it's 64k
idk, i have a feeling the allocator is not the issue
median crest
Stub out some uneeded you figure out whats the issue functions one by one until
slim panther
@fiery turtle how can i wire up an interrupt handler? the function signature doesn't match the one from uacpi
vagrant hull
realistically, how allocation heavy is this?
fiery turtle
<@353593948523986945> how can i wire up an interrupt handler? the function signa...
Wrap in one that does?
slim panther
but that only works for one handler?
fiery turtle
realistically, how allocation heavy is this?
Very
vagrant hull
is this allocation heavy to the point where things like cache locality and whatnot start to matter?
loud ice
holy shit my kernel is so fast!1!!!!11
fiery turtle
but that only works for one handler?
Dont u have a way to provide a context pointer to an interrupt
fiery turtle
holy shit my kernel is so fast!1!!!!11
Based
loud ice
(i always return 0 in my timer shit)
slim panther
Dont u have a way to provide a context pointer to an interrupt
not sure
loud ice
lol
slim panther
probably not
loud ice
ok whats the default tsc rate
gentle peak
whatever your host is
slim panther
i do i think
vagrant hull
i had a slab allocator but i removed it because i figured i did not need it
fiery turtle
Whats data?
slim panther
priv basically
fiery turtle
Thats fine
slim panther
okay
fiery turtle
is this allocation heavy to the point where things like cache locality and whatn...
Maybe
loud ice
doesnt help
fiery turtle
okay i have a real timer
Good starting point
fiery turtle
doesnt help
Like at all?
median crest
I had seen no diff between O2 and O3 in astral too
loud ice
Like at all?
nope
i was at -Copt-level=1 before tbh
so its not super shocking
also i have stubbed mutexes still
fiery turtle
Ah
loud ice
and i think i have some UB somewhere
okay i turned off spammy e9 logging and now its at 83
this is all tcg btw
fiery turtle
Try kvm?
loud ice
gotta move it to my other laptop first
fiery turtle
Should be about 800k
fiery turtle
Not bad for a 3 hour old kernel
loud ice
lmfao
vagrant hull
loud ice
its full of stubs
vagrant hull
i backported the Lyre slab allocator
loud ice
im on 909
fiery turtle
i backported the Lyre slab allocator
Nice
loud ice
i wonder how much of that is just my shit ass malloc
which isnt optimized for performance by far
vagrant hull
run-to-run variance
loud ice
holy shit the variance of this is bad yeah
okay then lets kill factorio
i think this is the most i see
probably just the fact that my cpu is worse than mint's at least somewhat
and my probably worse allocator
vagrant hull
the slab allocator i backported still zeroes out on all allocs so
loud ice
oh and i fill allocations
ah yeah same
actually i double clear on zeroed allocs
but yeah this is kinda a bad test just because its way too fast
loud ice
do they have more?
ugh i dont want to deal with ovmf
and i dont want to deal with graphics rendering either
median crest
Mint's cpu and qwinci's cpu are 3x faster than mine on the benchmark lol
loud ice
how much do i get? qemu-system-x86_64 -hda bs/reblast.iso -M q35,smm=off -cpu host,tsc-frequency=1000000000 -d int -accel kvm -debugcon stdio
vagrant hull
on mine?
loud ice
yeah
vagrant hull
aight one sec
loud ice
thanks
vagrant hull
uacpi: successfully loaded 1 AML blob, 1702 ops in 0ms (avg 1854648/s)
calm latch
Btw, linux (https://www.kernel.org/doc/Documentation/virt/kvm/timekeeping.txt) says this
The TSC is a CPU-local clock in most implementations. This means, on SMP
platforms, the TSCs of different CPUs may start at different times depending
on when the CPUs are powered on. Generally, CPUs on the same die will share
the same clock, however, this is not always the case.
loud ice
okay so now question
how do people make their shit faster
i guess faster malloc?
hmm lets try that
vagrant hull
i actually don't even remember why i removed the slab allocator
damn i get 1.6 mil consistently now
all i did compared to before is allow 2048 byte allocs to also go via slabs
gentle peak
qemu kvm: 1705 ops, 1088140 ops/s
real hw (haswell mobile cpu): 4962 ops, 122947 ops/s
i wonder why the difference is so big?
gentle peak
you dont do as many expensive ops on real hw
real hw is like 10x slower though
vagrant hull
does uACPI map and unmap the same regions often?
gentle peak
does that really just come down to how much better my host system's cpu is?
loud ice
does uACPI map and unmap the same regions often?
idk i just return hhdm
fiery turtle
qemu kvm: 1705 ops, 1088140 ops/s
real hw (haswell mobile cpu): 4962 ops, 122947...
Maybe aml does a sleep() inside for example
loud ice
any reason you cant do that?
fiery turtle
A lot of factors
fiery turtle
does uACPI map and unmap the same regions often?
Probably not
loud ice
oh i guess you need to actually make sure its mapped
gentle peak
Maybe aml does a sleep() inside for example
sleep and stall are both stubs rn
loud ice
just map all of physmem
gentle peak
but yeah probably just firmware devs being as good at writing AML as they are at writing firmware (not very)
vagrant hull
any reason you cant do that?
well, as per base revision 3 of the Limine protocol, and because of the fact that you should always map stuff that isn't usable memory and friends, i do map it, in hhdm, but i do map it
fiery turtle
Yeah thats why we test on qemu so we have the same aml for all
loud ice
well, as per base revision 3 of the Limine protocol, and because of the fact tha...
yeah true
slim panther
@fiery turtle in kernel_install_interrupt_handler, is ctx optional?
vagrant hull
so yeah i was wondering that because checking if it was already mapped and not mapping it in that case may speed things up
if the same regions are mapped over and over
fiery turtle
<@353593948523986945> in kernel_install_interrupt_handler, is ctx optional?
Wdym optional
Its uacpics context for that irq
slim panther
i probably have to modify my irq implementation, right? i can only pass one thing as a context
i'm confused
fiery turtle
That it wants to be passed to its handler
fiery turtle
i probably have to modify my irq implementation, right? i can only pass one thin...
U can make a global variable if u want
gentle peak
qemu kvm: 1705 ops, 1088140 ops/s
real hw (haswell mobile cpu): 4962 ops, 122947...
tested it on my actual host without QEMU in the way: 10097 ops, 976768 ops/s
fiery turtle
alright i get 2.5 mil now
What did u do
vagrant hull
slab allocator + allow 2k slabs + don't zero out unless calloc
fiery turtle
tested it on my actual host without QEMU in the way: 10097 ops, 976768 ops/s
So real hw close to qemu then?
fiery turtle
slab allocator + allow 2k slabs + don't zero out unless calloc
Nice
gentle peak
yeah pretty close
fiery turtle
Depends on the blob I guess
fiery turtle
3M+ is PhD territory
gentle peak
I do think QEMU uses faster ops because it's faster even though anything that would be a vmexit on there is much cheaper on actual hw
fiery turtle
no mutexes though
Yeah thats not very fair
vagrant hull
idk what my mutexes would even be, other than spinlocks without cli/sti
fiery turtle
Once you implement them
fiery turtle
Yeah
vagrant hull
slab allocators are pretty cool
fiery turtle
Indeed
slim panther
what exactly does the interrupt handler return
fiery turtle
Handled/not handled
loud ice
slab allocator + allow 2k slabs + don't zero out unless calloc
damn thats significantly faster than mine
how do you select the slab bucket?
loud ice
im gonna assume its something smarter than just a loop?
loud ice
ah huh
slim panther
Ever heard of a thing called struct lol
i didn't want to make one 😢
tbh i could also do a [static 2]
loud ice
wtf why is my slab so much worse then
or what else is the problem?
maybe my malloc is crap?
fiery turtle
Pog
vagrant hull
it's not much but it's honest work
fiery turtle
Mathwend recently implemented it you might want to ask him
slim panther
womp womp
fiery turtle
Where do I look for documentation on kvm-clock?
Its just one msr and a struct in memory so pretty simple stuff
median crest
Where do I look for documentation on kvm-clock?
Theres a linux doc page talking about it
slim panther
[ 0.080972] [ 0] acpi: successfully loaded 1 AML blob, 1816 ops in 50ms (avg 36304/s)
woe
calm latch
vagrant hull
fiery turtle
https://github.com/mintsuki/cynix
Niice
vagrant hull
no license, all rights reserved :^)
fiery turtle
I like how it just has uacpi libc in root
vagrant hull
yeah
gentle peak
Where do I look for documentation on kvm-clock?
https://docs.kernel.org/virt/kvm/x86/cpuid.html shows how to detect it
https://docs.kernel.org/virt/kvm/x86/msr.html shows how to use it
gentle peak
https://github.com/mintsuki/cynix
aren't the uacpi_kernel_io_{read,write} supposed to be PIO on x86, not MMIO? or am i implementing them wrong?
vagrant hull
i thought it was mmio
fiery turtle
pio ofc
vagrant hull
damn
so like
how
then what the hell is it the same for kernel_io_map?
that makes no sense
gentle peak
this is how i implement them https://hst.sh/utixukoket.csharp
kernel_io_map is for probably for platforms where "PIO" is just MMIO with an offset
fiery turtle
kernel_io_map is for probably for platforms where "PIO" is just MMIO with an off...
This or microkernels
vagrant hull
so is io_map supposed to be a no-op in my case?
fiery turtle
For kernel level x86 drivers yeah
vagrant hull
the thing is that io is quite confusing, if it was pio i would've understood it better, but io could be mmio
fiery turtle
Well there's already a memory read
slim panther
what does uacpi use the pci_{read,write} fns for?
slim panther
can i guarantee that these functions will be called after e.g. mcfg enumeration
fiery turtle
They won't be called before namespace_load
slim panther
okay good
slim panther
get owned
torpid root
that shrimply doesn't work on my kernel
some addresses are not mapped, or I forgot to map them somewhere
median crest
I play safe and barely rely on the hhdm outside of physical memory
slim panther
i forgot to have a mapping api for the kernel
calm latch
aren't the `uacpi_kernel_io_{read,write}` supposed to be PIO on x86, not MMIO? o...
Btw, does uACPI only call this only during the initialization?
gentle peak
from the docs it seems like it's something that can be triggered by AML
(the raw_io family explicitly says it can't be triggered by AML, and the normal io family does not)
calm latch
I play safe and barely rely on the hhdm outside of physical memory
I don't use hhdm at all and don't even map it when stiching to my own page tables from limine 
torpid root
I wish I was
alright I'm too tired for this allocator thingie
maybe in a few days
slim panther
@fiery turtle what page protection does kernel_map need? RW?
calm latch
RO crashes it
fiery turtle
<@353593948523986945> what page protection does kernel_map need? RW?
Rw ye
fiery turtle
we're up to 63k
!!
slim panther
94k
loud ice
tcg?
slim panther
what's that
loud ice
are you using qemu with TCG on
slim panther
i'm getting a pretty consistent average of 90k
loud ice
or KVM
slim panther
kvm
loud ice
holy fuck what are you doing
im getting 90k on TCG
admittedly i cant call my thing a kernel
slim panther
how do i do tcg
slim panther
holy fuck what are you doing
phd memcpy
loud ice
lol
loud ice
phd memcpy
that should mean its not 90k lol
more like 900k
are you sure your timer logic is correct?
loud ice
does it match rdtsc with -cpu host,tsc-frequency=1000000000 (which makes it a nanosecond-precision timer)
calm latch
uACPI: cutting edge timer stress test
slim panther
actually, maybe the timer is a bit slow
it's like 80% real time
it's not very precise lol
i told it to wait 1 second
maybe it's the printing
loud ice
it's not very precise lol
this looks like expected drift
your drift is 1ms per second?
so margin of error
slim panther
hpet
loud ice
ah
slim panther
my timer is accurate so far though
idk why it's so slow :(
@loud ice what's your qemu command
loud ice
<@696410219663851552> what's your qemu command
qemu-system-x86_64 -hda bs/reblast.iso -M q35,smm=off -cpu host,tsc-frequency=1000000000 -d int -accel kvm -debugcon stdio
slim panther
nope, still 90k
funnily enough with fb enabled it's faster somehow
over serial it's only 36k
yea no, i can't get past 90k
without kvm i get 30k
gentle peak
is the irq passed to uacpi_kernel_install_interrupt_handler a GSI or an ISA IRQ? if it's GSI, how do you determine whether it's active low/high and level/edge triggered?
north holly
is the `irq` passed to `uacpi_kernel_install_interrupt_handler` a GSI or an ISA ...
Gso
*gsi
And the ioapic redirection table entry things tell you
The trigger mode and polarity
gentle peak
like, the ones already in the ioapic registers, which were written by firmware?
i'd prefer not to rely on those if possible
north holly
like, the ones already in the ioapic registers, which were written by firmware?
Acpi ones
From the madt
gentle peak
interrupt source overrides? those are only for isa irqs
mortal yoke
usually the irq passed to that function is an isa irq
I think its also allowed to be a gsi tho (likely only on platforms where there is no legacy pic at all)
north holly
interrupt source overrides? those are only for isa irqs
Whatever type two
Eas
*was
mortal yoke
I think its also allowed to be a gsi tho (likely only on platforms where there i...
System vector the SCI interrupt is wired to in 8259 mode. On systems that do not contain the 8259, this field contains the Global System interrupt number of the SCI interrupt. OSPM is required to treat the ACPI SCI interrupt as a sharable, level, active low interrupt.
fiery turtle
is the `irq` passed to `uacpi_kernel_install_interrupt_handler` a GSI or an ISA ...
ISA
Its usually overridden in madt so be careful
gentle peak
alright thanks
yeah that's why i asked, i wanted to know if i needed to check the isos
fiery turtle
sounds correct
dusky glade
@calm latch pmOS powerbutton works just fine
fiery turtle
set a breakpoint, this is the very first call to the function
--- a/source/namespace.c
+++ b/source/namespace.c
@@ -194,11 +194,14 @@ uacpi_status uacpi_initialize_namespace(void)
void uacpi_deinitialize_namespace(void)
{
+ uacpi_status ret;
uacpi_namespace_node *current, *next = UACPI_NULL;
uacpi_u32 depth = 1;
current = uacpi_namespace_root();
+ ret = uacpi_namespace_write_lock();
+
while (depth) {
next = next == UACPI_NULL ? current->child : next->next;
@@ -233,6 +236,9 @@ void uacpi_deinitialize_namespace(void)
// This node has no children, move on to its peer
}
+ if (ret == UACPI_STATUS_OK)
+ uacpi_namespace_write_unlock();
+
uacpi_object_unref(g_uacpi_rt_ctx.root_object);
can you please test if this fix works
gentle peak
yep, that fixed it
fiery turtle
nice
gentle peak
one interesting thing i noticed: when booting with ovmf instead of seabios, uacpi is consistently faster by ~100k-200k ops/s (consistently 1.2-1.3M/s on ovmf, 900k-1.1M/s on seabios)
fiery turtle
different aml perhaps?
gentle peak
doubt it, they both have exactly the same number of ops (1705)
fiery turtle
This is the final call for anyone who wants to make it to the initial uacpi notable project & leaderboard list
ill be finalizing that pr shortly
@jaunty fox would you like to be mentioned? you have a pretty good score
jaunty fox
<@391791686881247234> would you like to be mentioned? you have a pretty good sco...
i have no objection to being mentioned
though i am wondering what the test methodology is
who carries them out?
slim panther
<@391791686881247234> would you like to be mentioned? you have a pretty good sco...
what's considered decent/good
fiery turtle
1M+ i guess
slim panther
what are the bottlenecks again
mortal yoke
also ubsan
slim panther
wait memset too?
omw to phd memset
but tbh my allocator might be my bottleneck
shitty little slab
north holly
Is mimalloc any good
fiery turtle
u mean namespace_load?
slim panther
during the benchmark test or w/e
fiery turtle
like every
slim panther
because if so i know what's bottlenecking it lmao
im printing a todo message on every call
rofl
fiery turtle
that costs a lot yeah
slim panther
let me remove it
loud ice
Is mimalloc any good
mimalloc is really good
slim panther
could that because i have some qemu devices?
torpid root
if they interface with acpi then probably yes
slim panther
qemu-system-x86_64 -bios /usr/share/ovmf/x64/OVMF.4m.fd -machine q35 -cpu max -name "menix x86_64" -serial stdio -no-reboot -no-shutdown -m 2G -smp 16 -drive format=raw,file=fat:rw:build/install,if=none,id=fatfs -device virtio-gpu -netdev user,id=net0 -device virtio-net,disable-modern=on,netdev=net0 -device nvme,serial=deadbeef,drive=fatfs -enable-kvm
nice, but u have quite a lot of stubbed out right
slim panther
generally it's 1704/1705
1708 with only an nvme
slim panther
nice, but u have quite a lot of stubbed out right
yea, is there anything that's considered "cheating"?
fiery turtle
kinda, yeah
not actually implementing an api is sort of cheating
because u avoid quite a lot of work
fiery turtle
even io read and stuff?
torpid root
is it cheating if I do something like not zero out allocated memory when trying to get the highest score possible?
but keep that memset in the kernel afterwards
gentle peak
like, not zero out memory allocated by calloc?
slim panther
i have allocations, raw mem rw, raw io rw, pci rw, io map, timer, spin, interrupts
torpid root
like, not zero out memory allocated by calloc?
malloc, beacause I zero that too
gentle peak
i don't think that'd be cheating, normal malloc isn't required to be zeroed
torpid root
i have allocations, raw mem rw, raw io rw, pci rw, io map, timer, spin, interrup...
my mutex is basically a spinlock for now
slim panther
lol
idk how to design a nice mutex interface
though a mutex makes little sense at the time where i'm calling uacpi_init
it's the only process that's running then
1.9 mil
north holly
damn me
fiery turtle
is it cheating if I do something like not zero out allocated memory when trying ...
no, ofc not
fiery turtle
1.9 mil <:meme:575445094589661215>
how
slim panther
idk
maybe the timer is being bad again
oh yea no it tried to configure lapic without hpet 
okay 1.9 might have been bogus, but i can reliably get 1.1-1.4M
calm latch
Why CPU do you have?
mortal yoke
if you have a rep movsb/stosb for them then doing anything else is unlikely to make them a lot faster
slim panther
i dont
i had a for loop
i set bytes so the address is qword aligned, then i do qword copies for memcpy
slim panther
does that work on arm
loud ice
yea
mortal yoke
with mmu + cache enabled yes
loud ice
if you have the mmu or caches off what the fuck are you doing
i believe windows does this btw
mortal yoke
if you have the mmu or caches off what the fuck are you doing
linux boot protocol
loud ice
linux boot protocol <:meme:575445094589661215>
the first thing you do when you take over is you turn on caches and mmu
mortal yoke
yes but if you want to do that in eg. c/c++ or whatever
loud ice
you cant
mortal yoke
you can
loud ice
cringe
and a perf penalty for the whole kernel
for the 1 line you save for caches
i believe its only caches too
slim panther
okay the new memset got me +100k
gentle peak
for anyone trying to maximize their score, here's some very rudimentary profiling data
slim panther
howd you do that
loud ice
for anyone trying to maximize their score, here's some very rudimentary profilin...
neat
so alloc, calloc, free
plus memcpy and memset
gentle peak
howd you do that
recorded starting time at the start of each function, end time at the end, and some glue code to list it all
loud ice
now record the call stack and see what allocates the most
gentle peak
worth doing tbh but i feel like at that point it'd be easier to hook up a proper profiler
mortal yoke
i believe its only caches too
maybe I should move the cache code to the startup asm in my kernel I could also implement the entire thing in asm (so I wouldn't have to rely on the compiler generating relocation-less pie code) but I don't like writing paging code in asm
gentle peak
just because your sample rate is way too low
aren't there profiling solutions that instrument the code to automatically do what i've done here but for every function?
loud ice
maybe I should move the cache code to the startup asm in my kernel <:KEKW:913708...
you really should do that
loud ice
aren't there profiling solutions that instrument the code to automatically do wh...
-minstrument-mcount kinda?
but its entry only i believe
i use it btw
gentle peak
damn
gentle peak
you need to remove -
the quotes take care of that
mortal yoke
they don't
gentle peak
ah it's -finstrument-functions
loud ice
ah
slim panther
okay i believe now it's just my allocator slowing things down :(
north holly
ah it's `-finstrument-functions`
yeah I found that
and I don't really know how to get it to work
in a kernel
properly
fiery turtle
for anyone trying to maximize their score, here's some very rudimentary profilin...
oh nice
jaunty fox
for anyone trying to maximize their score, here's some very rudimentary profilin...
log callers are killers under kvm what with the slow portio
if i log in uacpi_kernel_map it cuts me from
[uacpi]: successfully loaded 1 AML blob, 2356 ops in 1ms (avg 1857465/s)
immediately to
[uacpi]: successfully loaded 1 AML blob, 2356 ops in 2ms (avg 836618/s)
slim panther
yea i had the same issue
i was dangling around 60k because i logged "todo" on every unimplemented fn
without i'm at 1.3mil
gentle peak
log callers are killers under kvm what with the slow portio
in that profiling data the only function that calls another profiled function is uacpi_kernel_calloc (which calls uacpi_kernel_alloc), and none of the calls log anything (except uacpi_kernel_log of course)
so i believe i avoided that in the data
but yeah it's a huge perf drain if you're logging a lot
for anyone that wants to make their own profiling data, here's the code i used: https://hst.sh/uqegidozov.cpp
it's not particularly optimized because i just wanted to get it implemented quickly, but none of the expensive parts are measured in the actual data (unless your profiled functions call other profiled functions)
gentle peak
fixed a bug that caused the heap to expand on each allocation, and uh...
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 4841920 ops/s)
that's a bigger improvement than i thought it'd be
slim panther
wtf
gentle peak
this is with 0 stubs btw, each kernel api func is fully implemented
slim panther
what allocator do you use?
fiery turtle
u sure your time keeping is accurate?
gentle peak
u sure your time keeping is accurate?
yes, sleeping in a loop for 1s/iter and printing the current time works correctly
fiery turtle
how are you so fast
fiery turtle
fixed a bug that caused the heap to expand on each allocation, and uh...
> `uacp...
@kindred beacon im afraid managarm has been demolished
kindred beacon
fiery turtle
whats your cpu?
gentle peak
what allocator do you use?
pmm: simple free list, not even buddy (all physical allocations that are more than 1 page use a separate api that only works during early boot)
heap (objects <=PAGE_SIZE): allocations are rounded to the next highest power of two (using __builtin_clzl). a separate free list is maintained for each power of two, which is expanded directly from the pmm
heap (objects >PAGE_SIZE): allocations are rounded to a multiple of PAGE_SIZE, vmm does the rest (irrelevant for this benchmark since there's only 1 "large" allocation that uses this)
slim panther
oh okay
gentle peak
whats your cpu?
AMD Ryzen 7 5800X 8-Core Processor
slim panther
i'm using a (less opimal) slab allocator
gentle peak
how are you so fast
i think the speed is a combination of four factors:
- good cpu
- my allocators are designed to only do the bare minimum required to be usable and reliable, which makes them simple and fast
- memcpy/memmove/memset are rep movsb and rep stosb
- my kernel does a lot of allocating and freeing before uacpi init, so my small object allocator's free lists are already populated. i suspect it can always or almost always go through its fast path
vagrant hull
for anyone trying to maximize their score, here's some very rudimentary profilin...
how come map is only called once
that's weird
gentle peak
"modern" cpus (this was introduced over a decade ago) have a feature that optimizes rep stosb and rep movsb
they're pretty much the fastest memcpy/memset you can get (excluding SIMD which isn't usable in kernels)
gentle peak
how come map is only called once
this data is only for uacpi_namespace_load, not any of the other calls. i suspect most of the mapping is done by uacpi_initialize
vagrant hull
ah
gentle peak
i think the speed is a combination of four factors:
- good cpu
- my allocators a...
i moved uacpi init earlier into the boot sequence to see if the last factor made any significant difference:
uacpi: info: successfully loaded 1 AML blob, 1705 ops in 0ms (avg 4605780/s)
so it does have an impact, but not super significant
fiery turtle
i think the speed is a combination of four factors:
- good cpu
- my allocators a...
so which cpu
gentle peak
AMD Ryzen 7 5800X 8-Core Processor
.
dusky glade
bro beats phd code
gentle peak
bro beats phd code
i just make use of phd microcode (my string functions are https://hst.sh/afelovokab.pl)
frank canopy
incredible
gentle peak
still, managarm is a microkernel, isn't it? if that 3.2M figure is uacpi running in userspace, that's more impressive than my monolithic stuff imo
vagrant hull
i don't think it's in userspace
jaunty fox
i moved uacpi init earlier into the boot sequence to see if the last factor made...
remarkable
vagrant hull
not sure, but i clearly remember lai being in thor
fiery turtle
still, managarm is a microkernel, isn't it? if that 3.2M figure is uacpi running...
its in kernel space
@gentle peak would you post your iso so people can test on different cpus?
jaunty fox
my best result yet is [uacpi]: successfully loaded 1 AML blob, 2356 ops in 0ms (avg 2586007/s) but usually it's more 1.8 to 1.9mil
slim panther
i'm getting a very reliable 1.3mil
gentle peak
some real hw benchmarks:
- Haswell laptop (Intel i5-4200M): 579037/s
- Same system as before but on bare metal (Ryzen 5800X): 3880858/s (I suspect the lower result is because linux puts the CPU onto a higher frequency)
gentle peak
<@938223886439436320> would you post your iso so people can test on different cp...
sure, give me a minute
jaunty fox
<@938223886439436320> would you post your iso so people can test on different cp...
i think we need someone to be a clearinghouse for benchmarking
slim panther
yea
vagrant hull
some real hw benchmarks:
- Haswell laptop (Intel i5-4200M): 579037/s
- Same syst...
nah it's very likely not the frequency thing
it's what was mentioned before
like different firmware AML doing different things (and definitely more than QEMU's AML)
jaunty fox
he should be someone that runs probably gnu/linux so KVM can be used and he should have an invariant TSC
vagrant hull
he?
median crest
astral?
fiery turtle
was it not astral?
jaunty fox
leo is also acceptable
median crest
leo had worse perf
fiery turtle
oh damn
median crest
did you mean qwinci
fiery turtle
qwinci had the highest?
median crest
qwinci and mint left me at 2.4mil
fiery turtle
i mean the leaderboard specifies which cpu was used so its not like super unfair
jaunty fox
there might also be other factors in play
yes, the other criterion should be some statistical knowledge (both lukflug and leo will fit that bill)
gentle peak
The command line I used for measuring was qemu-system-x86_64 -accel kvm -cpu host,migratable=off -M q35,smm=off -cdrom proxima.iso
fiery turtle
but if qwinci or mint are down to measure everyone's iso i dont mind it
vagrant hull
i can
jaunty fox
with keyronex tests i see a lot of intertest variability (probably the same will be true of other kernels, i will check this proxima one next) so whoever is testing should be mindful of it
slim panther
i don't have an iso 
gentle peak
The command line I used for measuring was `qemu-system-x86_64 -accel kvm -cpu ho...
oh btw this is built with -O3 and LTO enabled
vagrant hull
gentle peak
in case it makes a difference for comparison
fiery turtle
this is probably important as well
jaunty fox
The command line I used for measuring was `qemu-system-x86_64 -accel kvm -cpu ho...
i observe consistent 3.6mil, congratulations
fiery turtle
now im sad its proprietary
gentle peak
its like, extremely simple
fiery turtle
but i guess can still be on the leaderboard