[PATCH v4] Documentation/gpu: VM_BIND locking document

Thomas Hellström thomas.hellstrom at linux.intel.com
Tue Nov 21 08:41:55 UTC 2023


Hi, Boris

On 11/16/23 15:47, Boris Brezillon wrote:
> On Thu, 16 Nov 2023 14:53:50 +0100
> Thomas Hellström <thomas.hellstrom at linux.intel.com> wrote:
>
>> On Thu, 2023-11-16 at 14:27 +0100, Boris Brezillon wrote:
>>> On Thu, 16 Nov 2023 12:48:45 +0100
>>> Thomas Hellström <thomas.hellstrom at linux.intel.com> wrote:
>>>    
>>>> Hi, Boris,
>>>>
>>>> Thanks for reviewing. Some comments below:
>>>>
I'm going to send out an updated version today with I think all of 
Danilo's comments and must of yours addressed. While I added more 
references to GPUVM, mainly as code examples and explanations, I've 
intentionally left out the "This is a driver lock and this is a gpvum 
lock distinction", The reason is twofold. First I think that when we get 
userptr into gpvum, gpuvm needs to be aware of most if not all locks. 
Second, since this document is an implementation guideline and gpuvm is 
an implementation it makes more sense to me to add pointers from the 
GPUVM documentation to the VM_BIND locking guideline, and that could be 
a task to be looked at after merging this together with implementing the 
userptr stuff. The most important thing at this point is that the 
document doesn't conflict in any way with the gpuvm implementation, and 
I've fixed those parts where I missed the separate lock protecting the 
GEM object's vm_bo list that you pointed out.

I strongly think this is the right way to go but if you disagree to the 
point where you're not willing to provide an ack or R-B, let me know and 
we can look at adding what's missing.

Thanks,

Thomas




More information about the dri-devel mailing list