gaming extensions to the kernel
Hi all, I have managed to create some exciting, gaming-specific extensions to the OpenBSD kernel, specifically for an arm64 raspberry pi 4. I would like to turn this into a product that people enjoy if possible and I would be happy to make something that benefits the OpenBSD community as well somehow. I am enjoying working on OpenBSD and am genuinely happy to give something back if I can. I started a discussion on other channels about this and got quite a bit of resistance, mainly because I wasn't planning to send diffs for what I am doing. My reasoning for not sending them is that the changes I made could create security issues for ordinary users, and I think that it would be a nightmare to maintain only to be able to play smoother games on a single platform, which in the grand scheme of things is quite small. To give you an idea, I am giving exclusive access to 3 out of 4 cpu cores to a game and I give the game quite a few pages of contiguous memory for the framebuffer. I give all that back to openbsd when the game ends. OpenBSD cannot interrupt the game on those 3 cores, it can only kill the game if needed. That's not stuff that should go into the official kernel, right? What I was thinking was more like "I go on and try to make and sell my product and when I make money I donate a percentage of the profits to the OpenBSD Foundation". Is that acceptable? Or alternatively, what is the "right" way of doing something like that? Thanks :-) Alessandro
allocating contiguous memory in user space
Hi all, is there any way in openbsd to allocate contiguous memory pages in user space? Thanks, A
Fwd: rethinking terminal login with security in mind
Sorry, my keyboard went crazy and the message was sent incomplete. Continuing: normally the entry of username is immediately followed by the password entry. However, if the OS is busy for any reason between the two entries, character echo is still on. If I don't notice that, I may start typing the password before the OS stops echoing and so I show it to anybody around who cares to look. Wouldn't it be better to have a way to turn off echoing of characters as soon as I entered my username, regardless of whether the OS is busy or not? Sorry again for the double message. Best, A -- Forwarded message - From: Alessandro Pistocchi Date: Wed, May 5, 2021 at 1:39 AM Subject: rethinking terminal login with security in mind To: OpenBSD misc Hi all, I am a new user. I have been using openbsd for the last few weeks on a raspberry pi 4. I have used other unix flavours in the past. I was wondering, what about changing how echoing of characters work when logging in from the terminal? Every unix I tried, including openbsd, asks for the username and then for the password. There is an opportunity for password sniffing there. What happens is that if I entern
rethinking terminal login with security in mind
Hi all, I am a new user. I have been using openbsd for the last few weeks on a raspberry pi 4. I have used other unix flavours in the past. I was wondering, what about changing how echoing of characters work when logging in from the terminal? Every unix I tried, including openbsd, asks for the username and then for the password. There is an opportunity for password sniffing there. What happens is that if I entern
Re: Fwd: umm_map returns unaligned address?
Hi,
apologies again.
I am not familiar with cvs so I have used the diff command between an
original sys folder and mine.
I am attaching the diff file and two files to be put in sys/kern. Sorry but
the diff
command did not include the content of the two files in the diff itself.
Please notice that this is not code that is intended to be put into the
openbsd kernel,
as it would probably not be interesting to the general users.
What I am doing here is to have a syscall that stops the scheduler on 3 of
the 4 cores,
runs my own code on those while the calling process runs on the other core
and restarts
scheduling on the 3 cores when the process terminates.
The shared memory is used for the 3 cores to communicate with the process.
I know that it might sound crazy to you but it does make a lot of sense for
what
I am trying to achieve.
Regarding the mapping, I hope that it works well enough even with
reordering issues and other stuff.
It does for the example I am sending and I don't need much more than that.
What I was flagging is just that sometimes uvm_map returns an address that
is not
aligned to PAGE_SIZE ( I printed it out and it has 0x004 in the lower 12
bits). On the
other hand uvm_unmap has an assertion that panics if the address passed to
it is not
page aligned. I believe that there could be a bug somewhere.
In previous runs I was getting constantly an address which is not aligned
but now
I get an address that is aligned.
Best :-)
Alessandro
On Sat, Apr 24, 2021 at 2:32 AM Theo de Raadt wrote:
> Alessandro Pistocchi wrote:
>
> > During the syscall I allocate some memory that I want to share between
> the
> > kernel and the calling process.
>
> When you get the mapping working, it will not work as well as you like.
>
> Compiler re-ordering of writes & reads, caches, write buffers, and other
> details such as non-TSO will creat problems, meaning you'll need very
> careful
> data-structure, and great care with co-dependent fields, otherwise one side
> can fool the other.
>
> That is why the copyin/copyout approach is used as a high-level
> serializing-forcing
> primitive in most operating systems.
>
> But since you don't show your whole diff.. I suspect this is a waste of
> time.
>
struct uvm_object *game_uvm_object;
void game_mode_create_shared_memory_region(struct proc *p) {
game_uvm_object = uao_create(game_memory_size, 0);
if(!game_uvm_object) return;
// TODO(ale): make sure that this memory cannot be swapped out
uao_reference(game_uvm_object);
if(uvm_map(kernel_map, (vaddr_t *)_memory, round_page(game_memory_size), game_uvm_object,
0, 0, UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE,
MAP_INHERIT_SHARE, MADV_NORMAL, 0))) {
uao_detach(game_uvm_object);
game_uvm_object = 0;
return;
}
uao_reference(game_uvm_object);
if(uvm_map(>p_vmspace->vm_map, (vaddr_t *)_memory_in_proc_space, round_page(game_memory_size), game_uvm_object,
0, 0, UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE,
MAP_INHERIT_NONE, MADV_NORMAL, 0))) {
// TODO(ale): the kernel mapping returns an address for game_memory that is 4 bytes more than page aligned. This panics.
uvm_unmap(kernel_map, trunc_page((vaddr_t)game_memory), round_page((vaddr_t)game_memory + game_memory_size));
game_memory = 0;
uao_detach(game_uvm_object);
uao_detach(game_uvm_object);
game_uvm_object = 0;
return;
}
game_mode_proc = p;
}
void game_mode_free_shared_memory_region(void) {
if(game_memory_in_proc_space)uvm_unmap(_mode_proc->p_vmspace->vm_map, trunc_page((vaddr_t)game_memory_in_proc_space),
round_page((vaddr_t)game_memory_in_proc_space + game_memory_size));
// TODO(ale): the kernel mapping returns an address for game_memory that is 4 bytes more than page aligned. This panics.
if(game_memory)uvm_unmap(kernel_map, trunc_page((vaddr_t)game_memory), round_page((vaddr_t)game_memory + game_memory_size));
if(game_uvm_object){
uao_detach(game_uvm_object);
uao_detach(game_uvm_object);
}
game_memory_in_proc_space = game_memory = 0;
game_uvm_object = 0;
}
// static inline void *phisical_from_virtual(struct proc *p, void *va) {
// paddr_t pa;
// vaddr_t faddr = trunc_page((vaddr_t)va), off = ((vaddr_t)va) - faddr;
// if (pmap_extract(vm_map_pmap(>p_vmspace->vm_map), faddr, ) == FALSE) panic("vmapbuf: null page frame");
// return (void *)(pa+(paddr_t)off);
// }
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#include
#define VBLANK_THREAD 1
#define AUDIO_THREAD 2
#define INPUT_THREAD 3
void *game_memory;
void *game_memory_in_proc_space;
vsize_t game_memory_size;
void sched_start_secondary_cpus(void);
void sched_stop_secondary_cpus(void);
void game_mode_create_shared_memory_region(struct proc *p);
void game_mode_free_shared_memory_region(void);
v
Re: umm_map returns unaligned address?
Sorry, I didn't see previous replies because the emails went to the spam
folder
for some reason.
I didn't mean to waste anybody's time. Apologies. I replied to your
subsequent points below
and I'll send the diff in a new email.
On Sat, Apr 24, 2021 at 2:24 AM Philip Guenther wrote:
> On Fri, Apr 23, 2021 at 3:13 PM Alessandro Pistocchi <
> apukbusin...@gmail.com> wrote:
>
>> -- Forwarded message -
>> From: Alessandro Pistocchi
>> Date: Fri, Apr 23, 2021 at 1:55 PM
>> Subject: umm_map returns unaligned address?
>> To:
>>
>>
>> Hi all,
>>
>> I am fairly new to openbsd so if this is something obvious that I missed
>> please be understanding.
>>
>> I am adding a syscall to openbsd 6.8. I am working on a raspberry pi.
>>
>> During the syscall I allocate some memory that I want to share between the
>> kernel
>> and the calling process.
>>
>> When it's time to wrap up and unmap the memory, I unmap it both from the
>> kernel
>> map and from the process map.
>>
>> The unmapping from the process map goes fine, the unmapping from the
>> kernel
>> map
>> fails by saying that the virtual address in kernel map is not aligned to
>> the page size
>> ( it's actually 4 bytes off ).
>>
>> What have I missed? I assumed that umm_map would return a page aligned
>> virtual
>> address for the kernel mapping as well.
>>
>> Here is my code for creating the shared memory chunk:
>>
>
> Stop sending summaries and just send diffs that compile: you don't know
> everything that is relevant and keep leaving out stuff that is. I'm the
> third person to say this.
>
>
>>
>>
>> // memory_size is a multiple of page size
>> uvm_object = uao_create(memory_size, 0);
>> if(!uvm_object) return;
>>
>> // TODO(ale): make sure that this memory cannot be swapped out
>>
>> uao_reference(uvm_object)
>> if(uvm_map(kernel_map, (vaddr_t *), round_page(memory_size),
>> uvm_object,
>>0, 0, UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE,
>>MAP_INHERIT_SHARED, MADV_NORMAL, 0))) {
>>
>
> The cast of is wrong: it's either unnecessary (if memory is of the
> correct type) or totally broken (if it isn't). Why did you think it was
> unnecessary to show how you declared your variables?
>
You're right.
memory is a void * which is zero at call time, vadd_t an unsigned long. I
believed it was ok since I am expecting to receive the address of the start
of the mapping.
>
> You also fail to show your initialization of 'memory'. If you didn't then
> that's ABSOLUTELY wrong and not in line with the existing uses of uvm_map()
> in the kernel. Please consult the uvm_map(9) manpage for what the incoming
> value means.
>
> ...
>
>> uao_reference(uvm_object);
>> if(uvm_map(>p_vmspace->vm_map, _in_proc_space,
>> round_page(memory_size), uvm_object,
>>0, 0, UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE,
>>MAP_INHERIT_NONE, MADV_NORMAL, 0))) {
>> memory = 0;
>>
>
> This error handling is incomplete, lacking an unmap.
>
Yep, I had noticed that and corrected it in my code. However that should
not be related to the problem I get.
>
>
> Philip Guenther
>
Fwd: umm_map returns unaligned address?
-- Forwarded message -
From: Alessandro Pistocchi
Date: Fri, Apr 23, 2021 at 1:55 PM
Subject: umm_map returns unaligned address?
To:
Hi all,
I am fairly new to openbsd so if this is something obvious that I missed
please be understanding.
I am adding a syscall to openbsd 6.8. I am working on a raspberry pi.
During the syscall I allocate some memory that I want to share between the
kernel
and the calling process.
When it's time to wrap up and unmap the memory, I unmap it both from the
kernel
map and from the process map.
The unmapping from the process map goes fine, the unmapping from the kernel
map
fails by saying that the virtual address in kernel map is not aligned to
the page size
( it's actually 4 bytes off ).
What have I missed? I assumed that umm_map would return a page aligned
virtual
address for the kernel mapping as well.
Here is my code for creating the shared memory chunk:
// memory_size is a multiple of page size
uvm_object = uao_create(memory_size, 0);
if(!uvm_object) return;
// TODO(ale): make sure that this memory cannot be swapped out
uao_reference(uvm_object)
if(uvm_map(kernel_map, (vaddr_t *), round_page(memory_size),
uvm_object,
0, 0, UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE,
MAP_INHERIT_SHARED, MADV_NORMAL, 0))) {
uao_detach(uvm_object);
uvm_object = 0;
return;
}
uao_reference(uvm_object);
if(uvm_map(>p_vmspace->vm_map, _in_proc_space,
round_page(memory_size), uvm_object,
0, 0, UVM_MAPFLAG(PROT_READ | PROT_WRITE, PROT_READ | PROT_WRITE,
MAP_INHERIT_NONE, MADV_NORMAL, 0))) {
memory = 0;
uao_detach(uvm_object);
uao_detach(uvm_object);
uvm_object = 0;
return;
}
Thanks,
A
accessing user memory from kernel
Hi all, I am playing around with openbsd kernel source code. Is there any clean way of accessing a process' memory from inside the kernel? Thanks, Alessandro
adding a syscall to openbsd
Hi all, I am trying to figure out how to add a syscall to openbsd 6.8. I am following "OpenBSD Kernel Internals - The Hitchhiker's Guide" by Vladimir Kirillov. (https://atmnis.com/~proger/openkyiv/openkyiv2009_proger_sys.pdf). I could successfully add my syscall and call it by using syscall(CALL_NR) from a basic C program. However, in that document it says that I should also rebuild libc. How do I do that? I tried by doing cd /usr/src/lib/libc and make but when I do make install I get an error: ran lib: /usr/lib/libc.a: no archive map to update. libc.a seems to be updated with the newest version, however libc.so.96.0 is still old. Thanks :-) Alessandro
