I don't think fork() would work with userspace where all buffers are shared. It certainly doesn't work now. The driver needs to be notified that a buffer or texture is shared to ensure data coherency between processes, and the driver must execute decompression and other render passes when a buffer or texture is being shared for the first time. Those aren't called when fork() is called.
Yeah, that's why you can install handlers which run before/after fork() is executed. But to summarize it is illegal for OpenGL, so we don't really need to worry about it.
For compute there are a couple of use cases though, but even those are not real world ones as far as I know.
But see below.
Marek
On Mon, Jan 17, 2022 at 9:34 AM Felix Kuehling <felix.kuehling@xxxxxxx> wrote:
Am 2022-01-17 um 9:21 a.m. schrieb Christian König:
> Am 17.01.22 um 15:17 schrieb Felix Kuehling:
>> Am 2022-01-17 um 6:44 a.m. schrieb Christian König:
>>> Am 14.01.22 um 18:40 schrieb Felix Kuehling:
>>>> Am 2022-01-14 um 12:26 p.m. schrieb Christian König:
>>>>> Am 14.01.22 um 17:44 schrieb Daniel Vetter:
>>>>>> Top post because I tried to catch up on the entire discussion here.
>>>>>>
>>>>>> So fundamentally I'm not opposed to just close this fork() hole
>>>>>> once and
>>>>>> for all. The thing that worries me from a upstream/platform pov is
>>>>>> really
>>>>>> only if we don't do it consistently across all drivers.
>>>>>>
>>>>>> So maybe as an idea:
>>>>>> - Do the original patch, but not just for ttm but all gem rendernode
>>>>>> drivers at least (or maybe even all gem drivers, no idea), with
>>>>>> the
>>>>>> below discussion cleaned up as justification.
>>>>> I know of at least one use case which this will break.
>>>>>
>>>>> A couple of years back we had a discussion on the Mesa mailing list
>>>>> because (IIRC) Marek introduced a background thread to push command
>>>>> submissions to the kernel.
>>>>>
>>>>> That broke because some compositor used to initialize OpenGL and then
>>>>> do a fork(). This indeed worked previously (no GPUVM at that time),
>>>>> but with the addition of the backround thread obviously broke.
>>>>>
>>>>> The conclusion back then was that the compositor is broken and needs
>>>>> fixing, but it still essentially means that there could be people out
>>>>> there with really old userspace where this setting would just break
>>>>> the desktop.
>>>>>
>>>>> I'm not really against that change either, but at least in theory we
>>>>> could make fork() work perfectly fine even with VMs and background
>>>>> threads.
>>>> You may regret this if you ever try to build a shared virtual address
>>>> space between GPU and CPU. Then you have two processes (parent and
>>>> child) sharing the same render context and GPU VM address space.
>>>> But the
>>>> CPU address spaces are different. You can't maintain consistent shared
>>>> virtual address spaces for both processes when the GPU address
>>>> space is
>>>> shared between them.
>>> That's actually not much of a problem.
>>>
>>> All you need to do is to use pthread_atfork() and do the appropriate
>>> action in parent/child to clean up your context:
>>> https://man7.org/linux/man-pages/man3/pthread_atfork.3.html
>> Thunk already does that. However, it's not foolproof. pthread_atfork
>> hanlders aren't called when the process is forked with a clone call.
>
> Yeah, but that's perfectly intentional. clone() is usually used to
> create threads.
Clone can be used to create new processes. Maybe not the common use today.
>
>>> The rest is just to make sure that all shared and all private data are
>>> kept separate all the time. Sharing virtual memory is already done for
>>> decades this way, it's just that nobody ever did it with a statefull
>>> device like GPUs.
>> My concern is not with sharing or not sharing data. It's with sharing
>> the address space itself. If you share the render node, you share GPU
>> virtual address space. However CPU address space is not shared between
>> parent and child. That's a fundamental mismatch between the CPU world
>> and current GPU driver implementation.
>
> Correct, but even that is easily solvable. As I said before you can
> hang this state on a VMA and let it be cloned together with the CPU
> address space.
I'm not following. The address space I'm talking about is struct
amdgpu_vm. It's associated with the render node file descriptor.
Inheriting and using that file descriptor in the child inherits the
amdgpu_vm. I don't see how you can hang that state on any one VMA.
But you don't really need that. You can bind the VM to your VMA mapping and clone that as necessary.
I'm not sure how else I should describe that, as far as I know the kernel that would be rather trivial to do.
Cloning all the userspace state like Marek described above is the much harder part.
Regards,
Christian.
To be consistent with the CPU, you'd need to clone the GPU address space
(struct amdgpu_vm) in the child process. That means you need a new
render node file descriptor that imports all the BOs from the parent
address space. It's a bunch of extra work to fork a process, that you're
proposing to immediately undo with an atfork handler. So I really don't
see the point.
Regards,
Felix
>
> Since VMAs are informed about their cloning (in opposite to file
> descriptors) it's trivial to even just clone kernel data on first access.
>
> Regards,
> Christian.
>
>>
>> Regards,
>> Felix
>>
>>
>>> Regards,
>>> Christian.
>>>
>>>> Regards,
>>>> Felix
>>>>
>
