trolled
#pmOS - microkernel OS for RISC-V, x86, and LoongArch
toxic torrent
toxic torrent
fair point
carmine nacelle
(because of bad design)
carmine nacelle
The same way that my pthread_mutex_unlock can also ENOMEM
carmine nacelle
from libc to posix server?
Yes
toxic torrent
ahh
carmine nacelle
Though idk I don't even have signals fully implemented
toxic torrent
signals are :(
carmine nacelle
i have processes and threads in kernel, but posix layer in user
I only have threads in kernel
And for signals I have userspace thread per process
But idk
I'll see
"rustc has very good error messages"
nvm, I'm dumb
toxic torrent
i have a microkernel question
how does one start a posix server without the posix system
like
wouldn't you need a lot of ipc calls to itself?
because i noticed you're not compiling with -ffreestanding
carmine nacelle
all my ipc is async
and the library thread is what makes the calls
lib/libc/src/generic/worker_thread/worker_main.c
Idk if it's a good design...
toxic torrent
i mean, how are you able to use a hosted environment for e.g. bootstrapd
carmine nacelle
It doesn't use POSIX API
And everything is done lazilly
I mean it does
Nvm
It calls malloc and friends
toxic torrent
same thing with managarm
carmine nacelle
It doesn't use the things provided by posix server
toxic torrent
i don't see how i'm supposed to bootstrap e.g. the posix server without a vfs
if that needs an interpreter
lol
agile patrol
Come on, why you do that
toxic torrent
?
carmine nacelle
It doesn't use the things provided by posix server
It just loads stuff directly without the posix server?
carmine nacelle
Also, I don't have exec* functions yet
Because I have no filesystems
(For which I'm porting Rust)
mellow sleet
i don't see how i'm supposed to bootstrap e.g. the posix server without a vfs
by not writing a microkernel 
I would just not use posix functions in the posix server
toxic torrent
maybe im misunderstanding things
because e.g. posix-subsystem on managarm has a PT_INTERP entry
carmine nacelle
I would just not use posix functions in the posix server
That's boring
agile patrol
I just dont know what I would typing
carmine nacelle
because e.g. posix-subsystem on managarm has a PT_INTERP entry
Just link it statically?
hexed acorn
same thing with managarm
in mammogram the kernel handles the small set of posix requests needed to dynamically link and run posix-subsystem and all the drivers etc
hexed acorn
for the vfs it serves stuff from the initrd and whatever you upload at runtime
mellow sleet
fake microkernel
hexed acorn
it's called being pragmatic
toxic torrent
i kinda want to keep the posix server self contained
agile patrol
I am offering $12
carmine nacelle
I also have a very small ELF loader in kernel to load bootstrapd
toxic torrent
same
carmine nacelle
You can also just use a flat binary
mellow sleet
you should do scheduling in userspace or yotrure not a real microkernel
carmine nacelle
I'm planning to do that 
Idr why
But I've been thinking about it a lot and it made sense
vernal bay
I think Tanenbaum put the line at the VFS
toxic torrent
performance is cringe anyways
carmine nacelle
My kernel has no knowledge of vfs
carmine nacelle
But I've been thinking about it a lot and it made sense
As in the kernel maintains the priority queues, and some userspace server decides on priorities, migrates the tasks and stuff like that
I remembered why
Linux added userspace schedulers in 6.12
carmine nacelle
As in the kernel maintains the priority queues, and some userspace server decide...
It made a lot of sense in my mind
agile patrol
Where is everyone,,, now is morning
carmine nacelle
hexed acorn
My kernel has no knowledge of vfs
in managarm all the posix glue for servers is contained in one file so the rest of the kernel doesn't know about it
technically we could get rid of it
by having a second set of mlibc sysdeps or not using mlibc at all
toxic torrent
@agile patrol can you stop posting random things to derail the discussion
hexed acorn
by having a second set of mlibc sysdeps or not using mlibc at all
but that'd be a major downgrade since we make heavy use of the standard c++ library in drivers and posix-subsystem
carmine nacelle
I also use C++
agile patrol
<@450856877472088066> can you stop posting random things to derail the discussio...
Sorry
agile patrol
I also use C++
Wow interisting person
hexed acorn
And lib(std)c++
yeah but i mean stuff like std::map and std::shared_ptr etc
not just freestanding stuff
carmine nacelle
toxic torrent
banger
carmine nacelle
the only freestanding program is kernel, everything else is fully hosted
hexed acorn
yeah and for that you need at least a skeleton of a libc
hexed acorn
and for managarm instead of having a separate libc or whatever we just have the minimum needed to run mlibc programs in the kernel instead
carmine nacelle
hexed acorn
although it's still all done via ipc, the programs just talk over ipc to the kernel
toxic torrent
how dumb would it be to just have the posix system be a freestanding binary 
carmine nacelle
and for managarm instead of having a separate libc or whatever we just have the ...
where do you use that?
I just wrote my own freestanding containers and stuff for the kernel
mellow sleet
although it's still all done via ipc, the programs just talk over ipc to the ker...
this is the way
the only real syscall imple,mented should be write()
hexed acorn
I just wrote my own freestanding containers and stuff for the kernel
nono the kernel doesn't use mlibc
toxic torrent
the only real syscall imple,mented should be write()
ioctl*
hexed acorn
nono the kernel doesn't use mlibc
i mean, posix-subsystem and drivers use regular managarm mlibc, and the kernel handles more or less the minimum ipc requests to facilitate that
hexed acorn
i mean, posix-subsystem and drivers use regular managarm mlibc, and the kernel h...
more or less you get stdin/stdout/stderr tied to the kernel log, the initrd in place of the vfs (read only too iirc), and mmap
carmine nacelle
My stdin/stdout/stderr just waits for the log server to appear (the first time they are used), and libc also only talks to vfs when you call open() or whatnot
It's done without vfs
hexed acorn
well we need the vfs since posix-subsystem et al are dynamically linked
and we use dynamic libraries for more than just libc
carmine nacelle
more or less you get stdin/stdout/stderr tied to the kernel log, the initrd in p...
doing that in init server seems reasonable
hexed acorn
i mean we could, but mlibc static linking support came after all this was in place
carmine nacelle
I don't have dynamic linking yet
mellow sleet
more or less you get stdin/stdout/stderr tied to the kernel log, the initrd in p...
is the initrd mapped in the pid 1 address space?
or is it ipc
hexed acorn
is the initrd mapped in the pid 1 address space?
no, you do the same requests as you'd do to posix-subsystem to access files off of it
just to an ipc stream where the other lane is listened to by the kernel instead of posix-subsystem
carmine nacelle
my bootstrap server can (could in the past, it's probably broken now) also act as a filesystem
mellow sleet
no, you do the same requests as you'd do to posix-subsystem to access files off ...
oh damn
carmine nacelle
so tmpfs driver could just open() the module passed by the bootloader
hexed acorn
it is a small impurity in the microkernelness but it's self-contained in a 700 line file (+ associated header file)
the other code just does runService(name, controlHandle, myNewThread); and that spins up the coroutine that does the request handling
hexed acorn
it is a small impurity in the microkernelness but it's self-contained in a 700 l...
looking at the file, we handle: sigaction (noop since no signals), openat, isatty, close, mmap, read, write (only on stdin/out/err), seek, readdir
toxic torrent
free
inland temple
I am offering $12
please just wire
agile patrol
what are you offering it for?
It is NVidia hardware... Difficult to understanding. Move to private chat if you want money
vernal bay
.!naenae @agile patrol no soliciting in DMs!
long grailBOT
hexed acorn
i mean we could, but mlibc static linking support came after all this was in pla...
also on the note of static linking, im not a fan of it, esp. in this case since i'd have to relink everything after fixing mlibc bugs or whatever, and the dependency on libc is not tracked by meson (since it's brought in by the compiler itself), so that'd be a pain
vernal bay
.!naenae @Dd-Aing no soliciting in other project's threads
long grailBOT
Command executed by lukflug
vernal bay
.!naenae 450856877472088066 no soliciting in other project's threads
long grailBOT
long grailBOT
lunar island
you should do scheduling in userspace or yotrure not a real microkernel
ethan kernel has it
carmine nacelle
bruh wsl crashed
left dew
lol wtf did u do
eager rose
maybe it got OOM'd
carmine nacelle
my laptop has 32gb of ram
Anyway, the only things missing are this rust dumbness (or more likely my skill issue), ioctl and a couple of broken things
carmine nacelle
I think this is the only thing that I need to add to my libc/Rust?
inland temple
if you would use mlibc you could just autogenerate the bindings for the libc crate
we have a script that uses libclang to parse mlibc headers and output rust libc bindings that you just copy+paste into the crate
and then you need a 2-line patch to the crade to pub mod managarm; or equivalent
pearl narwhal
we have a script that uses libclang to parse mlibc headers and output rust libc ...
does managarm have a patchset for rust on linux-mlibc?
if no i think that should be a thing
carmine nacelle
too late
I did it manually
(though I might switch to mlibc later once I stabilize my API)
pearl narwhal
if no i think that should be a thing
if yes I should use it in nixpkgs
inland temple
does managarm have a patchset for rust on linux-mlibc?
I don't think the question makes sense
like
linux-mlibc has no relation to managarm
inland temple
yeah still what does that refer to
pearl narwhal
e.g. for gcc there's a fork i can use
but if there's no such thing for rust I'll have to do it myself (either using the rust binding generator in nixpkgs or uploading the fork somewhere
inland temple
so for providing the (patched) libc crate, we pull the libc crate source, copy/paste our autogenerated bindings, and apply a small patch for it to include our bindings
pearl narwhal
ah so there's no downstream fork for it or anything and is just patched by the build system?
inland temple
I'm not sure how relevant that would be on linux with mlibc tho
carmine nacelle
Someone can also just upstream mlibc to rust
pearl narwhal
I'm not sure how relevant that would be on linux with mlibc tho
I would assume it needs something to build against mlibc at least
inland temple
ah so there's no downstream fork for it or anything and is just patched by the b...
yeah, we just have a small patch file we apply on top of the crate source
generally we do not maintain forked repos of anything except gcc and binutils
we always just ship patch files that we apply on top of the source
toxic torrent
wait do i understand it correctly, if mlibc was upstream, would that enable it for all projects using it?
carmine nacelle
I have Rust??
carmine nacelle
wait do i understand it correctly, if mlibc was upstream, would that enable it f...
I think it just makes it easier to port
You still have to add OS-specific stuff?
toxic torrent
isn't that then handled by libc?
left dew
is rust std entirely implemented in terms of libc calls?
toxic torrent
would be omega based
toxic torrent
duality of man
left dew
nah i fully stand behind implementing std in terms of syscalls
linking against libc in non-C code is cringe
after all it's the C library...
not the Rust library
toxic torrent
for a big os, maybe
but its tedious if you have to do it on your own
libc my beloved
dense carbon
after all it's the C library...
it's the library to the OS
OSes and languages always went together historically
inland temple
wait do i understand it correctly, if mlibc was upstream, would that enable it f...
no
it needs to be per-os
obv as the ABI and options selection changes
what I will attempt to do is to upstream the (auto-generated) managarm bindings
pearl narwhal
is rust std entirely implemented in terms of libc calls?
depends on the target
it needs to be per-os yes but quite a lot can be libc calls i believe
pearl narwhal
<:bruh:761542241381908491>I have Rust??
does it hello the world
carmine nacelle
does it hello the world
Idk I had lectures
I'll test now
It probably won't link
How do I give cargo sysroot?
Do I just use normal lld?
hexed acorn
are you rewriting pmos in rust?
hexed acorn
ah
carmine nacelle
Like I normally link C/C++ with clang --target=x86_64-pmos --sysroot=/home/misha/pmos/i8042/../sysroot ./ports.o ./main.o ./timers.o -o i8042.elf or whatever
Is there no way to see why it's failing?
I'm too dumb for this...
hexed acorn
no such file or directory: crt0.o?
carmine nacelle
bruh
I didn't see that
I'm blind
Do I even need crt0?
If I'm linking with libc anyway
pearl narwhal
probably?
-nodefaultlibs so you need it yes
also test it with rust-lld too when you have crt0
carmine nacelle
Though I should have it
carmine nacelle
Does it not search in usr/lib?
carmine nacelle
-nodefaultlibs so you need it yes
(time to recompile the compiler)
(I wonder how long my laptop will survive compiling rustc on battery)
carmine nacelle
Ok, I suppose I only need compiler_rt?
Bruh it also wants this for some reason
Which I don't have
And compiler rt?
bruh my libc was compiled for i686 for some reason??
Wtf
(like this stuff builds fine on i686 and riscv)
Idk how, but I compiled it
nevermind
it just casually failed
carmine nacelle
carmine nacelle
Wtf
Bruh I have no idea
(and how it just magically works on i686)
I think copying random things from chatgpt solved it...
(totally perfectly working build system)
wtf wdym
The functions are for Real Time Solaris (Real Time means guaranteed response in a time boundary), which has since been integrated into mainline Solaris and is part of libc. Librt is present only for historical, backward compatibility reasons and you do not need to reference it.
twilit talon
bruh wsl crashed
what's interesting is i've literally never had my wsl crash, but it crashed today out of nowhere
twilit talon
had my first time in like 5 years today
carmine nacelle
This is not even that bad!
What if I just link it with clang++?
This is like an extreme jank
(it didn't work)
carmine nacelle
Missing functions to hello world
(I guess I can just stub them)
(except writev)
(it should be easy)
(famous last words)
export RUSTFLAGS="-C link-arg=--sysroot=$HOME/pmos/sysroot/ -C linker=clang++ -C link-arg=--target=x86_64-unknown-pmos -C link-arg=-lunwind -C link-arg=-static"
inland temple
you will get absolutely shafted by libc ngl
carmine nacelle
?
carmine nacelle
Ok, I have no idea how to implement poll
Since everything is IPC...
ok, I think I have an idea
(though my libc is kinda casually becoming a huge mess)
carmine nacelle
Time to test
Ok, Rust really wants poll
(But it is working I'd imagine)
(enough for libc to printf the error)
carmine nacelle
does it hello the world
!!!
carmine nacelle
Idk, now
(I need to unstub poll though)
(and a couple of other places)
(but it works for println)
I wonder if panic works
pearl narwhal
imagine if it fails to fail
pearl narwhal
does unwind work??
pearl narwhal
ah it probably does
carmine nacelle
Where does it get variables?
carmine nacelle
It's broken
It's probably trying to open /proc/self or whatever, which I don't have
but still
carmine nacelle
you will get absolutely shafted by libc ngl
I didn't
inland temple
surprising
inland temple
It's probably trying to open /proc/self or whatever, which I don't have
isn't that linux specific though?
carmine nacelle
idk
inland temple
it it is that should be guarded behind a OS == Linux check
carmine nacelle
I don't really care I guess
Go needed to open that to print backtraces
bruh I'm having a skill issue with git
hexed knot
Go needed to open that to print backtraces
gc shouldn't need it...
i guess the shitty go compiler might though
although
its not a thing on macos, or the BSDs
so
hexed knot
no
carmine nacelle
So I'm just gonna use it
hexed knot
go green threading asm is like
not hard
lol
also if you are cool with a mildly scuffed version of go there's also tinygo
pearl narwhal
but rust doesn't need weird assembly shenanigans for userspace threads
rust doesn't have a thread runtime
in std
hexed knot
also if you are cool with a mildly scuffed version of go there's also tinygo
which has LLVM do the magic for you
hexed knot
yeah rust also takes the "let's have a stupid% performance degradation in pathological cases" approach
instead of "let's realloc() the thread stack when it gets too full"
carmine nacelle
idk
I just need to figure out how to implement my kernel's interface in Rust now...
carmine nacelle
bruh I'm having a skill issue with git
Ok, I managed to push it to gitlab
I guess I'll just delete the local repo and clone it again 
carmine nacelle
I think I need to finally add handles to IPC
The interface gets clunkier every time...
(I'm currently writing mbus-inspired mechanism for device discovery)
hexed knot
cool!
carmine nacelle
Like they could both be useful to distinguish various async IPC message replies, create permission memes, and also have some system to track the servers programs dying and not replying to IPC
The only issue is that I need to have some mechanism for handles to be shared between several threads
(for which I could use task groups which I already have, now that I think of it)
The only thing I'm worried about is performance, since it would be allocating more stuff for every IPC message...
carmine nacelle
(I'm currently writing mbus-inspired mechanism for device discovery)
Also, I want to try and write that server in Rust
hexed knot
Also, I want to try and write that server in Rust
do you have an async runtime or something like that?
carmine nacelle
(I'm currently writing mbus-inspired mechanism for device discovery)
I think I can also tie it to the init server (which I don't have at the moment), so that it can directly start necessary servers for published devices, without doing what I do now, which is running drivers, and then having them fork
carmine nacelle
do you have an async runtime or something like that?
No, but that can be synchronous
carmine nacelle
My C++ stuff is async
hexed knot
at least for parity
carmine nacelle
would still be cool to have async support
I need to investigate how it's done
In Rust
hexed knot
I need to investigate how it's done
basically: like C++, but dumber
there is a "Future" trait
which has one function named poll
which takes a context and returns a future status
the context contains a waker
when you want to do io, you return a future which, when polled, tries to read/write/whatever
and if it blocks, it arranges for the waker to be woken up when it is ready
so yeah like c++ but dumber
pearl narwhal
basically: like C++, but dumber
its not that dumb tbh
hexed knot
yes it is tbh
the C++ approach produces faster code
because their wakeup logic wakes up a coroutine instead of queuing a coroutine to be scheduled
hexed knot
in rust, you need a scheduler
yeah
like the original top-level waker is provided by the scheduler
carmine nacelle
hexed knot
yes.
hexed knot
"the" scheduler = the userspace future scheduler that is
yeah C++ coros are a lot easier to work with
the main disadvantage is that they allocate implicitly
carmine nacelle
Like I have an event loops, which block on reading from IPC queues
hexed knot
and aiui they allocate one allocation more
carmine nacelle
Like I have an event loops, which block on reading from IPC queues
And then when a message arrives, I just run the blocked coroutine which that message was for
Or create new coroutines
(idk what's the term for awaking)
hexed knot
And then when a message arrives, I just run the blocked coroutine which that mes...
yeah and in rust you ALSO need a scheduling queue to go along with it
so a message arriving would unblock a future
and then you run the unblocked futures until all of them are blocked
(yes this is stupid)
carmine nacelle
So the future would have just 1 coroutine in its queue?
Ok I will not use coroutines for now then
(Also, I need to implement handles first anyway)
hexed knot
So the future would have just 1 coroutine in its queue?
a rust "future" is like a c++ "coroutine"
carmine nacelle
this sounds weird
hexed knot
except its implemented differently
since in rust futures only run when the future scheduler decides they should
and not immediatly when work completes
pearl narwhal
in rust, you need a scheduler
it's hard to implement but it's basically what you need to do if you want to not provide a thread runtime in the std lib
hexed knot
it's hard to implement but it's basically what you need to do if you want to not...
no its not lol
C++ doesnt need it
pearl narwhal
libstdcxx provides a thread runtime tho right?
hexed knot
no
hexed knot
libstd provides the same amount of thread stuff as libc++ and libstdc++
pearl narwhal
tbh I'm not well versed into this so i can't really say anything
but my experience with rust async (including implementing a simple thread runtime) was nice
the biggest issue is that the ecosystem is split between async and sync
sans-io ftw
carmine nacelle
C++ allows you to do wtf you want with coroutines
hexed knot
tbh I'm not well versed into this so i can't really say anything
the main difference is that C++ coroutine wakeup isnt handled by the scheduler, but by the individual coroutine being woken up
which makes impl significantly easier
its also faster in certain semi-pathological cases
toxic torrent
the main disadvantage is that they allocate implicitly
that's only a minor inconvenience
toxic torrent
it's negligible imo
hexed knot
probably
the important factors are that rust wakes up the root coroutine
and not the leaf coroutine
which is not negligible
carmine nacelle
Bruh my Rust abstractions to kernel API just work..?
(I think I'll write more tomorrow, if I have time)
inland temple
based
carmine nacelle
Now I just need to implement the parsing of the stuff that I send over IPC...
(It will be C, C++ and Rust programs talking together)
(also I maybe really need to look at using xbstrap or something, since Makefiles are really rough for building 2 compilers and 12 executables...)
hexed knot
Bruh my Rust abstractions to kernel API just work..?
now implement pop_front_async
also damn you have std, very nice
carmine nacelle
now implement pop_front_async
I need to implement handles for IPC first
hexed knot
handles for ipc?
carmine nacelle
Like for IPC ports/chanels/whatever
carmine nacelle
Yes
hexed knot
ah
carmine nacelle
Like I send a message and give provide a handle to reply
(that's what I want to do)
hexed knot
ahhh
carmine nacelle
And then I can distinguish where messages come from
carmine nacelle
And then I can distinguish where messages come from
And it would also be a great permission system
hexed knot
which allows the sender to, well, send exactly one message in return
carmine nacelle
they have a concept of a "send once" port right
But I want to have one port, and give out different "handles" to it
Idk
I don't have a mechanism to block waiting for more than 1 port
And I don't think I need to
hexed knot
But I want to have one port, and give out different "handles" to it
yeah thats the point
carmine nacelle
And then I can distinguish where messages come from
What I do at the moment is I usually include "user_arg" field inside the messages, which can be used to know which message was replied to
hexed knot
when you do an ipc call, you send (call arguments, send once port right)
carmine nacelle
But it's obnoxious and not universal
hexed knot
and then when they return, they send a message to that send once port right
carmine nacelle
when you do an ipc call, you send (call arguments, send once port right)
I just send
(at the moment)
But I want it to send a handle, along with message
Which would resolve to a port
But the sender would know the id of the handle which was assigned when a message was sent
So even though the messages would all still come back to the same 1 port
(which I can read from in an event loop)
I can universally have sync IPC-lookalike behaviour
To know which coroutine to unblock/schedule/resume/whatever by the handle
Idk
(I think I'm repeating myself)
carmine nacelle
But I want it to send a handle, along with message
Like I've already added a mechanism to send memory objects
Which in reality are handles for memory objects
It shouldn't be too difficult
And since my ports are bound to threads, a lot of the stuff could be CPU-local
And I have RCU in kernel...
So like I'm hoping the performance shouldn't be bad
carmine nacelle
Like I've already added a mechanism to send memory objects
For zero-copy IO
eager rose
But the sender would know the id of the handle which was assigned when a message...
you could make it possible for the sender to attach a pointer-sized field to the send-once port, that is then included in the response.
carmine nacelle
you could make it possible for the sender to attach a pointer-sized field to the...
The thing is, how do you handle the responses?
Like take a disk driver, it pops a bunch of requests from the message queue, sends commands to disk, blocks for however long the operation takes, and replies
There needs to be some mechanism to explicitly reply to a message
eager rose
yup
carmine nacelle
Another thing I've though about it holding onto messages
eager rose
it responds by writing to the send-once port included with the message it is replying to
carmine nacelle
I don't have a mechanism to block waiting for more than 1 port
But then this is problematic
eager rose
"everything is SCM_RIGHTS" kernel 
eager rose
But then this is problematic
not really. a possible solution: the send-once handle is a 'thin wrapper' around a 'real' port handle, that simply sends a message to the real port along with the auxiliary dara attached to the send-once port
carmine nacelle
not really. a possible solution: the send-once handle is a 'thin wrapper' aroun...
And that's what I'm calling a "handle"
eager rose
and multiple send-once handles can point to the same port
eager rose
"everything is SCM_RIGHTS" kernel <:KEKW:913708372329644042>
literally this
carmine nacelle
(maybe the name is bad)
carmine nacelle
"everything is SCM_RIGHTS" kernel <:KEKW:913708372329644042>
But it's not a bad API
But I really need to think about it
Since I want to have a mechanism to send multiple handles at a time
carmine nacelle
also damn you have std, very nice
Using my libc 
eager rose
send-once port is bad too. because they might not actually be send-once, e.g. if you decide to expand the functionality to be able to reply more than once. it might be worth designing it such that a process can instantiate a special kind of handle to a port that associates auxiliary data with all messages sent to it, and then making the 'send-once' part a usage detail
carmine nacelle
send-once port is bad too. because they might not actually be send-once, e.g. ...
Yeah, so I'm thinking about reply once and reply many, with optional transferable rights
carmine nacelle
And also ports requiring handles by default, but some ports could be public
eager rose
like a general facility for creating these, and it can specify a usage limit
carmine nacelle
And also ports requiring handles by default, but some ports could be public
Although don't even know if that's necessary
I can move my ports discovery mechanism to userspace, and just have that server hand out handles
eager rose
as in open_port("/path/to/port")
you can make ports implicitly private by not giving them a name
hexed knot
send-once port is bad too. because they might not actually be send-once, e.g. ...
mach has a special "send once" port right which explicitly requires one message
carmine nacelle
which are assigned sequentially
And anyone can send to any port
hexed knot
and a "send many" port right
eager rose
mach has a special "send once" port right which explicitly requires one message
idk. I'm not a microkernel expert lol
carmine nacelle
And anyone can send to any port
And you can 1. guess them, 2. send to them once you know them
So like there's no protection
eager rose
yeah, that feels like a fundamental flaw to me
carmine nacelle
And then I have a global namespace kind of thing, which allows to get port id by its name, if the port owner decides to publish it
That's how I discover things atm
carmine nacelle
yeah, that feels like a fundamental flaw to me
It's simple and apparently somewhat effective, the only problem is that there is no protections whatsoever
carmine nacelle
And that's what I'm calling a "handle"
But this will solve everything at once
Another problem is that the kernel doesn't know about processes
So the issue is in making a handle available to different threads which form part of the same process
eager rose
'handles' are not part of this global port namespace, i.e. a malicious or faulty process cannot guess and write to a 'handle'?
carmine nacelle
I don't have them at the moment
carmine nacelle
'handles' are not part of this global port namespace, i.e. a malicious or faulty...
I'm thinking of making them global
Although potentially not
eager rose
So the issue is in making a handle available to different threads which form par...
you can have a refcounted fd table (or, well, fd table equivalent), and tasks can choose to unshare them at fork
carmine nacelle
The issue is not with fork
The only thing that's common in a process inside the kernel is a page table/address space
So there is nothing to latch to
Idk how to explain it
Like a filesystem would give a handle for a file
But then, how do I share that with threads inside a process
eager rose
you can have a refcounted fd table (or, well, fd table equivalent), and tasks ca...
to me, this feels like the 'obvious' solution. keep a pointer to the fdtable in struct task or whatever.
carmine nacelle
to me, this feels like the 'obvious' solution. keep a pointer to the fdtable in...
My fdtable is in libc in userspace
eager rose
a task doesn't have a private handle namespace. rather, it is shared among tasks.
eager rose
My fdtable is in libc in userspace
yeah, I should've called it something else (like handle namespace)
carmine nacelle
task groups (I'm bad at naming things) 
But there can be many
So I will have to pass task_group + handle to ipc_send
Which will be 2 uint64_t's
Which don't fit into i686 registers for syscalls
eager rose
So I will have to pass task_group + handle to ipc_send
why? the kernel should know the current task anyways, right?
carmine nacelle
Yes, but any given task can be part of as many task groups as it likes
eager rose
the pointer to the handle table is not kept in a 'task group', rather, it is refcounted and pointed to by each individual task that makes up the task group
carmine nacelle
I could make handles global, and check that the task is part of task group that the handle is attached to
eager rose
Yes, but any given task can be part of as many task groups as it likes
ok this, this is an interesting design decision.
eager rose
Yes, but any given task can be part of as many task groups as it likes
what is the usecase for this?
carmine nacelle
I could make handles global, and check that the task is part of task group that ...
It is done like that so that task groups can be used to check the object lifetimes (?)
Like a filesystem consumer opens a file, and gives a task group. That task (thread) group can represent a single thread, a process, a set of tasks (threads), maybe two processes
(task is thread)
(I need to work on naming)
carmine nacelle
Like a filesystem consumer opens a file, and gives a task group. That task (thre...
This also applies to disks, devices, and so on
eager rose
Like a filesystem consumer opens a file, and gives a task group. That task (thre...
what is the purpose of doing this? like are there any end users that actually do this?
carmine nacelle
No
eager rose
well then
carmine nacelle
I just create a task group for the process and use that everywhere
But it seems like a nice abstraction
Instead of doing that for the whole process, it could be done for select threads
eager rose
does the kernel know about task groups?
carmine nacelle
It's provided by kernel
and it has a mechanism to notify userspace when there are no threads left in a group
So if the process dies, the files are closed in servers, and so on
(and it's also used for permission checking)
eager rose
Which will be 2 uint64_t's
you could also, like, not use 64-bit integers for both of these.
like what are the odds that someone will exhaust 32 bits of task group id space?
carmine nacelle
idk
carmine nacelle
I could make handles global, and check that the task is part of task group that ...
I could just do this
eager rose
yup, but then you lose out on some of the benefits of a 'local' handle table
lunar island
mach has a special "send once" port right which explicitly requires one message
what it does is lock the refcnt to 1
eager rose
and it has a mechanism to notify userspace when there are no threads left in a g...
you could also notify when an end of a "socket" closes. (where a "socket" represents a sender/receiver handle pair, and if the sender dies, yell at the receiver)
carmine nacelle
you could also notify when an end of a "socket" closes. (where a "socket" repre...
do I need it?
for signals
eager rose
you could also notify when an end of a "socket" closes. (where a "socket" repre...
I meant do this instead of notifying when a process dies
lunar island
have an handle namespace per-process
carmine nacelle
I don't have processes
hexed knot
what it does is lock the refcnt to 1
no
eager rose
@carmine nacelle it feels like you are overcomplicating object management by, like, a lot.
hexed knot
but its refcount can be nonzero
(e.g. you can have a different send-once right to the same port)
lunar island
"Send-once rights always have a user reference count of one. Tasks may hold multiple user references for dead names. Each send-once right generated guarantees the receipt of a single message, either a message sent to that send-once right or, if the send-once right is in any way destroyed, a send-once notification."
carmine nacelle
<@350748255929499649> it feels like you are overcomplicating object management b...
If you haven't noticed, everything (not just object management) is more complicated than it should be
lunar island
(e.g. you can have a different send-once right to the same port)
yes
I am talking about the right
not the port
carmine nacelle
"Send-once rights always have a user reference count of one. Tasks may hold mult...
That's what I want to do
eager rose
yup. and iiuc, the (or rather, a) problem is how do we make sure that the sender has the right to respond to the message? and a proposed answer is local handles
carmine nacelle
If you haven't noticed, everything (not just object management) is more complica...
pmOS is about 60k LoC (excluding external libraries), and I don't have a shell, or a filesystem driver
carmine nacelle
yup. and iiuc, the (or rather, a) problem is how do we make sure that the sende...
What's the problem with global though
lunar island
pmOS is about 60k LoC (excluding external libraries), and I don't have a shell, ...
what can it do tho
carmine nacelle
idk
lunar island
yup. and iiuc, the (or rather, a) problem is how do we make sure that the sende...
sender?
lunar island
u just pass an handle along with your message
carmine nacelle
what can it do tho
Like I ported rust std, so I could write ext4 driver with it
eager rose
pmOS is about 60k LoC (excluding external libraries), and I don't have a shell, ...
last time I checked, Davix was 30k LoC, and approx. 20k of those was from a generated test case for a kernel datastructure
but like
my kernel doesn't do much either (hardly anything)
and LoC is a bad measurement unit anyways
lunar island
im currently microkerneling so the kernel isnt doing that much either
eager rose
I should finish writing my in-kernel debugging tools (unwind and symbol lookup), and then implement the page cache so my files can actually have contents lol
carmine nacelle
I had libunwind and libcxxrt with C++ exceptions in kernel, but I've removed that
Now I just kinda keep my fingers crossed that the kernel doesn't crash since it only writes to COM0 on panic
When running it on laptops and whatnot
toxic torrent
Bruh my Rust abstractions to kernel API just work..?
real
carmine nacelle
By default, HashMap uses a hashing algorithm selected to provide resistance against HashDoS attacks. The algorithm is randomly seeded, and a reasonable best-effort is made to generate this seed from a high quality, secure source of randomness provided by the host without blocking the program. Because of this, the randomness of the seed depends on the output quality of the system’s random number coroutine when the seed is created. In particular, seeds generated when the system’s entropy pool is abnormally low such as during system boot may be of a lower quality.
carmine nacelle
Good and bad news
wide spruce
🪨
modest nest
Also second panic while unwinding was in progress
carmine nacelle
I know it's because of syscall that's not implemented by kernel
wide spruce
Imagine syscalls
carmine nacelle
Which function does write call?
carmine nacelle
(ok, I think I'll just not print long strings then and pretend it's working )
carmine nacelle
This is unfun
left dew
you said you'd implement it
carmine nacelle
I'm just gonna use BTreeMap for the time being
carmine nacelle
you said you'd implement it
There are too many things that need to be implemented
left dew
Box<str> pains me for some reason
carmine nacelle
What am I supposed to use instead ?
left dew
i mean, i guess it ends up being functionally equivalent because of Deref<Target=str> (or smth, idk rust lol) but String is a nicer, owned string type
but as i said, Box<str> works as well
carmine nacelle
I've read that it stores 2 usizes instead of 3
Although does Rust do SSO?
SSO has been intentionally left out, because it adds trade-offs (like extra checks on each data access) and you can use a library to implement your own flavor of SSO (there are many crates for this with various levels of compactness and unsafety).
toxic torrent
Nice Rust code
this hurts to read for some reason
restive junco
cause serialization and deserialization hurt to do in higher level languages
speaking from experience
it's much much easier to do it in C
restive junco
cause serialization and deserialization hurt to do in higher level languages
I've done it in JS with a kind of messaging protocol between JS and C (some IoT things) and man it was awful to write that in JS
pearl narwhal
Bruh my Rust abstractions to kernel API just work..?
port tokio and make the api nonblocking
i wonder how hard it is to port tokio
considering it basically reimplements std
something simpler might be prefferred
toxic torrent
i wonder how hard it is to port tokio
pain
pearl narwhal
did u do it?
toxic torrent
no but i looked at tokio internals
pearl narwhal
i wonder if async-std is simpler
or one of the lower level stuff (that don't reimplement std)
toxic torrent
pmos all servers in rust when
pearl narwhal
my opinion is that userspace drivers would be fun in rust
but I won't do them for stuff that needs to be optimised for binary size
toxic torrent
sure
pearl narwhal
e.g. i don't need huge rust binaries for a uart driver
ns16550 and friends would be so simple to implement, doing them in C or C++ wouldn't make you lose anything really
carmine nacelle
pmos all servers in rust when
I'm not rewriting everything in Rust
(yet)
Also I need to implement a bunch of stuff
Namely finish pmbus, fix AHCI, write ext4 driver, implement IPC handled/whatever
At which point I might actually finally implement a slab allocator in kernel
pearl narwhal
name it premarketOS
kindred venture
in rust, x86 means i*86
and with i*86 the only way to do anything on more than 4 bytes is cmpxchg8b
toxic torrent
What the hell
seems reasonable
carmine nacelle
and with i*86 the only way to do anything on more than 4 bytes is cmpxchg8b
so i386 as well?
kindred venture
i386 and i486 don't have cmpxchg8b and thus it is impossible to do atomic ops on 64-bit values
well you can use libatomic but that's implemented with spinlocks
carmine nacelle
ok, that's reasonable then
kindred venture
also note that that's only saying that atomic loads on 8 bytes are only guaranteed to work on rw memory, not that it's not supported
i suspect an atomic load of 8 bytes gets lowered into cmpxchg8b(ptr, x, x)
kindred venture
xmm isn't guaranteed to be available on 32 bit
but if it is yeah that's an option
kindred venture
i686 is a subset of 32 bit x86 that is guaranteed to support sse (like x86_64)
specifically i686 = everything the pentium pro (P6 microarch) supports
kindred venture
i586 def doesn't support xmm, it doesn't even guarantee mmx
but yeah so on i386 and i486 it'd get lowered into libatomic, i586 to cmpxchg8b, i686 to sse (probably)
carmine nacelle
Also maybe floating point instructions?
kindred venture
how does x87 help with atomics
carmine nacelle
is it not atomic as well?
kindred venture
ah you mean like loading something as a double and copying to a cpu reg
idk what the atomicity guarantees were like on pre-i586
on i586+ it'd work probably but then you have cmpxchg8b so why would you
carmine nacelle
isn't cmpxchg8b slow?
like if you just want to read something with relaxed ordering
kindred venture
depends on implementation and whether it's successful
relaxed ordering isn't a thing on x86, everything is acq rel
or well there's some newer simd instructions that might have relaxed ordering idk but in general everything's acq rel
carmine nacelle
but I mean if you just want to read a value from memory, not exchange it
kindred venture
lock cmpxchg8b is def much more expensive than a normal mov yeah
but idk how it compares to fp load + copy to int reg
clang on i586 uses fp ops so i guess that's more performant than cmpxchg8b
carmine nacelle
Bruh, recieving objects with Rust is easy, now the question comes, how do I serialize them...
Since with C/C++ I just sent structs or vectors directly, but it doesn't seem to be easy with Rust idk
carmine nacelle
I've talked to ChatGPT and it just said to use std::borrow::Cow
My data is represented as C structs
left dew
im afraid that asking chatgpt might not be the best thing you could have done (lol) but oh well, let's see what you come up with
maybe i misunderstood what you're trying to do :^)
carmine nacelle
bytemuck Pod?
left dew
okay, sure
if you just want to bytemush the whole thing over ipc then that would work i guess
strings aren't repr(C) though lol
carmine nacelle
carmine nacelle
if you just want to bytemush the whole thing over ipc then that would work i gue...
I want to bytemuck static stuff, and serialize into Vec or Box the dynamic stuff
Hence Cow
Which is a universal representation of both without writing my own enum for either &[u8] or vec
left dew
yeah i understand, might work
left dew
native rust bindings > generated rust bindings
though if you want you can write rust wrappers around generated bindings
carmine nacelle
I'll just write that stuff myself
Is there a way to implement some sort of universal thing on everything that's bytemuck Pod?
So it translates it to Cow
There's no function overloading in Rust??
Ok, I'm gonna try something...
carmine nacelle
Is it a bad idea to use #![feature(default_field_values)]?
What the fuck
error[E0726]: implicit elided lifetime not allowed here
--> /home/misha/pmos/rust/pmos/src/ipc_msgs.rs:29:18
|
29 | impl Default for IPCBusPublishObject {
| ^^^^^^^^^^^^^^^^^^^ expected lifetime parameter
|
help: indicate the anonymous lifetime
|
29 | impl Default for IPCBusPublishObject<'_> {
| ++++
I'm extremely dumb
carmine nacelle
How am I supposed to do this?
pub const IPC_BUS_PUBLISH_OBJECT_REPLY_NUM: u32 = 0x1b0;
#[repr(C)]
#[derive(Debug, Copy, Clone, Zeroable, Pod)]
pub struct IPCBusPublishObjectReply {
msg_type: u32,
pub flags: u32,
pub result: i32,
reserved: u32,
pub sequence_number: u64,
}
impl IPCBusPublishObjectReply {
pub fn new() -> IPCBusPublishObjectReply {
IPCBusPublishObjectReply {
msg_type: IPC_BUS_PUBLISH_OBJECT_REPLY_NUM,
flags: 0,
result: 0,
reserved: 0,
sequence_number: 0,
}
}
}
Like so that msg_type and reserved are private, but always set
Without writing new for everything
- mut memes
carmine nacelle
efficiently and generically
I don't understand how anything could be more efficient than just casting structs to pointers
pearl narwhal
How am I supposed to do this?
```rust
pub const IPC_BUS_PUBLISH_OBJECT_REPLY_NUM...
make msg_type an enum pls :3
carmine nacelle
And I'm not using json or other stuff for IPC
carmine nacelle
This is my IPC https://gitlab.com/mishakov/pmos/-/blob/main/sysroot/include/pmos...
Like this is simple, effective, fast and easy to use
And works across 3 languages
pearl narwhal
Without writing new for everything
I don't understand what you're looking for
you can implement Default
if that's what you want
or you can pass arguments to new
carmine nacelle
It complained that default fields are private
carmine nacelle
How am I supposed to do this?
```rust
pub const IPC_BUS_PUBLISH_OBJECT_REPLY_NUM...
Like idk so that someone can do ```rust
let msg = IPCBusPublishObjectReply {
flags: 0,
result: whatever,
sequence_number: bruh,
// The rest is default
}
..Default::default()
there's a spread operator
and I would just pass arguments to new
carmine nacelle
Doing ```rust
let msg = IPCBusPublishObjectReply {
flags: 0,
result: whatever,
sequence_number: bruh,
..Default::default()
}
(I've tried it and rustc refused to compile it)
pearl narwhal
yea the issue isn't default
the issue is that you can't construct a struct with private fields this way
also don't put a semicolon there, put a comma
carmine nacelle
I want msg_type to be constant, but it needs to be there
pearl narwhal
I want msg_type to be constant, but it needs to be there
IPCBusPublicObjectsReply::new(IPCBusPublicObjectsReplyFlags::WHATEVER, whatever, bruh)
carmine nacelle
And I don't want to have to copy the struct every time I send a message
pearl narwhal
And I don't want to have to copy the struct every time I send a message
why would it copy this way and not copy your way?
carmine nacelle
`IPCBusPublicObjectsReply::new(IPCBusPublicObjectsReplyFlags::WHATEVER, whatever...
And this initializes the private variables?
pearl narwhal
And this initializes the private variables?
if you do so in new() yes
new isn't special
it's just an associated function (aka static method)
it can take arguments
carmine nacelle
why would it copy this way and not copy your way?
No, I mean I wanted to initially omit the type field and just add it when serializing
But I don't like it
pearl narwhal
yea then keep it in the struct
carmine nacelle
So the struct is just normal rust struct, and the data sent between stuff is different
pearl narwhal
impl IPCBusPublishObjectReply {
pub fn new(flags: u32, result: u32, sequence_number: u64) -> IPCBusPublishObjectReply {
IPCBusPublishObjectReply {
msg_type: IPC_BUS_PUBLISH_OBJECT_REPLY_NUM,
flags,
result,
reserved: 0,
sequence_number,
}
}
}
also please make FLAGS an enum unless it really needs to be bitflags
and if it really needs to be bitflags, use bitflags or whatever the crate was named
only make it an integer if you don't want to depend on that crate
rn you could be passed an invalid flags and would have to check it manually
if you had a IPCBusPublishObjectReplyFlags type where only the valid flags can be constructed, you don't have to check that
carmine nacelle
Flags is basically a reserved field atm
(it doesn't matter for now)
I want to see this working, I will add this stuff later
(later as in before I push the commit)
carmine nacelle
How do I pass int into bitflags! memes?
(bruh, vscode is as lost with Rust as it is with C++)
(idk which is worse)
Creative variable naming (result.result)
What the fuck
Why does it suddenly want linux?
kindred venture
it's being used by your main.rs directly
possibly something lsp did based on your host system?
kindred venture
yeah it autoimported based on your host system
carmine nacelle
Fun, I think it's working
(better screenshot)
The top 4 lines are printed by Rust, the bottom one is printed by AHCId/C++
(Now there is too much stuff to do)
That is (not including all my previous plans):
- Implement requests for stuff from pmbus
- Publish ACPI and PCI devices in pmbus (which needs to be done in C )
- Write an init server, which matches devices in pmbus and starts servers this way as they appear
- Implement handles to radically decrease the obnoxiousness of IPC API (among other things) and provide a good interface for coroutines
- Change blockd to discover disks with pmbus, discover filesystems using it (talking to init server, see step 3), publish partitions onto it and do disk mounting (etc.)
- Finish (though it's almost not even started) the ext4 driver (this implies improving VFS, which I wrote when I was experimenting with ChatGPT in 2023, so the code is very bad)
- Do what @ infy wanted, that is discover PS/2 controllers with ACPI (now if I do 2, everything can be discovered like that...)
- ... improve the build system (it's so bad I'm having to call cargo build and then make (several times) and then qemu manually to run pmOS)
I'm going to sleep, this is too much
(though I guess I can just call cargo with make )
hoary moat
That is (not including all my previous plans):
1. Implement requests for stuff f...
when will pmOS get suspend support????
carmine nacelle
toxic torrent
What
lmao
that warning message is useless
they could've just said "you probably meant to do *obj"
carmine nacelle
carmine nacelle
(I think it's stack overflow? )
How does this even GPF?
Wtf, how does this get to shit in %rcx?
carmine nacelle
Wtf, how does this get to shit in %rcx?
This was a fat UB
carmine nacelle
Ok, I don't know what to do
So I guess I'll finally look at xbstrap
Or something
(improve my build system that is)
carmine nacelle
I don't understand anything
toxic torrent
real
toxic torrent
xbstrap my beloved
pearl narwhal
nixstrap my beloved 😍
carmine nacelle
Idk, I need something which can build (in this order) llvm, my C libraries with CMake, rustc, libc++ and company, kernel with CMake and a bunch of userspace (about 10 servers, with more planned ) with Make and Cargo (basically C, C++ and Rust), and create the root image
Like maybe I'm dumb, but I haven't figured out how to make cargo produce output binary in the same location
That's not in target/{arch}/bin/whatever (idr the exact path)
Since I'm targeting 3 architectures, with more to come
pearl narwhal
I think you can make it "install" to some destination dir
but nix will do it :3 (nixstrap shilling
pearl narwhal
So I can just call cargo install with Make? <:trl:829316641191886888>
look at how gentoo does it
I doubt the gentoo cargo eclass or whatever is complicated
well it also vendors in crates so it likely is but you don't need to do that
but for maximum reproducibility you want to read the cargo.lock file and vendor in the needed crates
then pass --frozen or whatever to cargo so it doesn't fetch from the internet
carmine nacelle
I don't really care about reproducibility, but the way I'm passing files/programs to be loaded is as modules in Limine/Hyper
Like I don't have a proper filesystem
pearl narwhal
but if you don't care about reproducibility, wrapping cargo is fine
carmine nacelle
Like I don't have a proper filesystem
So my makefiles are just kinda hardcoded to assume files in their locations
carmine nacelle
That's not in target/{arch}/bin/whatever (idr the exact path)
And Rust was producing that
Idk my build system is a mess
Maybe I should finally have an init ramdisk or something like that
So I can just dump a bunch of files into a directory and it be dynamic
Without having to change the bootloader configs
pearl narwhal
yea I don't have that problem because I do like '${pkgs.whatever}/bin/whatever'
and the nix recipe will install to the store so where cargo puts it doesn't matter
carmine nacelle
But then I don't want everything in ram (eventually)
Idk I need to think
Maybe I'll consider it
pearl narwhal
you know you could do like string substitution to get the path to the binary in target/
then copy it to your ESP or wherever
for now
pearl narwhal
That's not in target/{arch}/bin/whatever (idr the exact path)
the exact path is target/release/{arch}/whatever
carmine nacelle
Ok I've just managed to do what I want with makefiles
The build system migration effort is on halt until I get S3 sleep working
pearl narwhal
obos but microkernel
hoary moat
I need to implement some cool feature for obos
to get it known for something other than suspend
and crashing all the time
carmine nacelle
bruh I'm having a skill issue with a relative far jump
hoary moat
no
pearl narwhal
no
make it networking™️ cuz you're already working on that
roleplay making an OS in the 90s
hexed acorn
make it AI
oboslop
pearl narwhal
and be like ¡¡ THIS IS THE INTERNET IT IS THE FUTURE !!
pearl narwhal
you mean the implementation that crashes after 5 minutes of a real network?
make it not crash or build an era-appropriate network for it
also 5 minutes is gokd
good
I could never let my stuff run for 5 minutes, not because it wouldn't work but because I couldn't wait that long
carmine nacelle
llvm optimizes my memcpy memes to nopl (%rax)
carmine nacelle
Could it be llvm miscompilation?
I know what's happening...
Fun, my pmm decided that SMP/ACPI bootup vector will be at 0x9E000
(also, maybe hyper port will get smp now that I'm implementing it anyway...)
hexed knot
Could it be llvm miscompilation? <:trl:829316641191886888>
was it?
im gonna guess no
carmine nacelle
I've been copying char * to void **, which llvm was assuming to be aligned to 8
Which was emitting broken assembly
hexed knot
I've been copying char * to void **, which llvm was assuming to be aligned to 8
yeah thats ub
hmm actually
is it
carmine nacelle
(Which was the end of a section in assembly)
twilit talon
i dont think memcpy of anything to anything is ub
carmine nacelle
I declared the variable as char and it worked
Tbh I'm memcpying code to some random place in memory
hexed knot
i dont think memcpy of anything to anything is ub
actually no it can be ub if the two pointers overlap
twilit talon
well ok
hexed knot
or if there is a data race
hexed knot
no its UB i think
kindred venture
the parameters are restrict, passing overlapping pointers is UB
hexed knot
no its clear ub
carmine nacelle
actually no it can be ub if the two pointers overlap
they don't
hexed knot
the parameters are restrict, passing overlapping pointers is UB
yeah ^^
carmine nacelle
It was an issue with alignment
twilit talon
memcpy doesnt really care i think?
hexed knot
i think overlap is just unspecified result
If copying takes place between objects that overlap, the behavior is undefined
N3219 (c23) is pretty explicit about it being UB i think
twilit talon
with restrict it makes sense
carmine nacelle
It's C++23
kindred venture
the alignment UB here isn't in the memcpy, that's just where it manifests itself
it's UB to have a misaligned pointer
hexed knot
yeah id say so
hexed knot
it's UB to *have a misaligned pointer*
actually no
twilit talon
yeah clang's intrinsic rightly assumes that your pointer to a void* is aligned to a void*
kindred venture
actually no
A pointer to an object type may be converted to a pointer to a different object type. If the resulting
pointer is not correctly aligned59) for the referenced type, the behavior is undefined
n3220 6.3.2.4
hexed knot
> A pointer to an object type may be converted to a pointer to a different objec...
ah okay
kindred venture
it's UB to produce a pointer that is unaligned therefore it is UB to have one
hexed knot
n3220 6.3.2.4
6.3.2.4 doesnt seem to say that
kindred venture
6.3.2.3 sorry
hexed knot
it's UB to produce a pointer that is unaligned therefore it is UB to have one
well first of all i don't think thats necessarily true
kindred venture
was looking at the bottom of the page
hexed knot
ah okay
kindred venture
it's UB to produce a pointer that is unaligned therefore it is UB to have one
llvm notices that the pointer you give to memcpy came from a type that has to be aligned, and optimizes memcpy assuming that the pointer is aligned
carmine nacelle
it's UB to produce a pointer that is unaligned therefore it is UB to have one
I had a misaligned global variable
(which wasn't a variable)
hexed knot
I had a misaligned global variable
that is hard UB though anyway
kindred venture
I had a misaligned global variable
how'd you manage that
hexed knot
because llvm can see the alignment
kindred venture
oh wait declaring it in asm?
carmine nacelle
oh wait declaring it in asm?
Yes
carmine nacelle
.equ CODE_SEGMENT, 0x08
.equ DATA_SEGMENT, 0x10
.data
init_vec_stuff:
.code16
.globl acpi_trampoline_begin
acpi_trampoline_begin:
.globl acpi_cpu_startup
acpi_cpu_startup:
cli
movw %cs, %bx
xorl %ecx, %ecx
movw %bx, %cx
addl $(acpi_trampoline_gdt - acpi_cpu_startup), %ecx
movl %ecx, %cs:(acpi_tampoline_gdtr_addr - acpi_cpu_startup)
lgdt %cs:(acpi_trampoline_gdtr - acpi_cpu_startup)
movl %cs:(acpi_trampoline_startup_cr3 - acpi_cpu_startup), %eax
movl $0x01, %eax # Protected mode
movl %eax, %cr0
jmp 1f
1:
.code32
movl %eax, %cr3 # cr3 (Page table)
movl %cr4, %eax # PAE
orl $(1 << 5), %eax
movl %eax, %cr4
movl $(1 << 10), %eax # LME BIT
xorl %edx, %edx
movl $0xC0000080, %ecx # IA32_EFER
wrmsr
# Enable paging
movl $(1 << 0) | (1 << 31), %eax
movl %eax, %cr0
movw $DATA_SEGMENT, %ax
movw %ax, %ds
movw %ax, %ss
movw %ax, %es
movw %ax, %fs
movw %ax, %gs
shr $16, %ebx
leal (acpi_trampoline_kernel_entry - acpi_cpu_startup)(, %ebx, 1), %esp
leal (_64bitst - acpi_cpu_startup)(, %ebx, 1), %eax
pushl %eax
lretl $CODE_SEGMENT
_64bitst:
.code64
jmp *acpi_trampoline_kernel_entry(%rip)
.p2align 4, 0xCC
acpi_trampoline_gdtr:
.space 2 # alignment
acpi_trampoline_gdtr_limit:
.word 23 # 3 entries - 1
acpi_tampoline_gdtr_addr:
.long 0
acpi_trampoline_gdt:
.quad 0
.quad 0x00af9b000000ffff
.quad 0x00af93000000ffff
.globl acpi_trampoline_startup_cr3
acpi_trampoline_startup_cr3:
.space 4
.globl acpi_trampoline_kernel_entry
acpi_trampoline_kernel_entry:
.space 8
scratch_stack:
.space 8
.globl acpi_trampoline_startup_end
acpi_trampoline_startup_end:
Having an unaligned ptr is not UB in C++
does it not include C's language about how converting a pointer requires said pointer to be correctly aligned for the destination type?
carmine nacelle
I memcpy from init_vec_stuff to acpi_trampoline_startup_end
kindred venture
I memcpy from init_vec_stuff to acpi_trampoline_startup_end
ah yeah that'd require them to be declared as char
carmine nacelle
And the acpi_trampoline_startup_end was not aligned to 8 even though I declared it as extern void * in C++
hexed knot
does it not include C's language about how converting a pointer requires said po...
first of all, just because it is C ub to convert it, doesn't mean it is C ub to have a misaligned pointer
carmine nacelle
Since clang didn't let me declare it as void
Memecpy compiled to ```x86asm
ffffffff80013c4d: 31 d2 xorl %edx, %edx
ffffffff80013c4f: 90 nop
; *((char *)dest++) = *((char *)src++);
ffffffff80013c50: 0f b6 34 02 movzbl (%rdx,%rax), %esi
ffffffff80013c54: 41 88 34 16 movb %sil, (%r14,%rdx)
ffffffff80013c58: 0f b6 74 02 01 movzbl 0x1(%rdx,%rax), %esi
ffffffff80013c5d: 41 88 74 16 01 movb %sil, 0x1(%r14,%rdx)
ffffffff80013c62: 0f b6 74 02 02 movzbl 0x2(%rdx,%rax), %esi
ffffffff80013c67: 41 88 74 16 02 movb %sil, 0x2(%r14,%rdx)
ffffffff80013c6c: 0f b6 74 02 03 movzbl 0x3(%rdx,%rax), %esi
ffffffff80013c71: 41 88 74 16 03 movb %sil, 0x3(%r14,%rdx)
ffffffff80013c76: 0f b6 74 02 04 movzbl 0x4(%rdx,%rax), %esi
ffffffff80013c7b: 41 88 74 16 04 movb %sil, 0x4(%r14,%rdx)
ffffffff80013c80: 0f b6 74 02 05 movzbl 0x5(%rdx,%rax), %esi
ffffffff80013c85: 41 88 74 16 05 movb %sil, 0x5(%r14,%rdx)
ffffffff80013c8a: 0f b6 74 02 06 movzbl 0x6(%rdx,%rax), %esi
ffffffff80013c8f: 41 88 74 16 06 movb %sil, 0x6(%r14,%rdx)
ffffffff80013c94: 0f b6 74 02 07 movzbl 0x7(%rdx,%rax), %esi
ffffffff80013c99: 41 88 74 16 07 movb %sil, 0x7(%r14,%rdx)
; while (n--) {
ffffffff80013c9e: 48 83 c2 08 addq $0x8, %rdx
ffffffff80013ca2: 48 39 d1 cmpq %rdx, %rcx
ffffffff80013ca5: 75 a9 jne 0xffffffff80013c50 <limine_main+0x29e0>
Even though the size was 0xb4
hexed knot
what is mozbl wtf
kindred venture
first of all, just because it is C ub to *convert* it, doesn't mean it is C ub t...
in C the only way to get a misaligned pointer is by converting a pointer to a type that has stricter alignment requirements though
hexed knot
ah load
left dew
[although i sure do love a good llvm moment](https://play.rust-lang.org/?version...
is this supposed to crash or something?
carmine nacelle
idk clang memes (I need a better memcpy)
hexed knot
in C the only way to get a misaligned pointer is by converting a pointer to a ty...
actually no
kindred venture
is this supposed to crash or something?
notice how it prints None even though the value being printed is Some(...)
kindred venture
ah fair enough
left dew
notice how it prints None even though the value being printed is Some(...)
oh lol wtf
i did not notice that
for some reaosn
carmine nacelle
btw why does clang unroll memcpy to copies of byte width?
hexed knot
btw why does clang unroll memcpy to copies of byte width?
i have no idea
it shoulkdn't
left dew
unaligned memory?
carmine nacelle
Like without mgeneral-regs-only it unrolls to 4 xmm loads then 4 stores
plain jungle
the wording in the C++ standard is:
Attempting to create an object ([intro.object]) in storage that does not meet the alignment requirements of the object's type is undefined behavior.
hexed knot
unaligned memory?
isel does not have alignment constraints
left dew
fair enough
carmine nacelle
Like without mgeneral-regs-only it unrolls to 4 xmm loads then 4 stores
or avx memes if I enable that