Incoming changes for GVNIC

Incoming changes for GVNIC

[Garrett D'Amore]

[-- Attachment #1: Type: text/plain, Size: 1325 bytes --]

This is mostly a heads up....

We (WEKA) use various drivers with specific alignment requirements, which causes us to need to need to use multibuffer (scatter/gather) functionality in various drivers.

Unfortunately, the GVNIC driver was ... very... buggy in this regard (in fact it doesn't work at all in this case!), especially for the DQO mode used on gen 3 (C3) instances.  It turns out it was also buggy in that it was not properly resetting the device on teardown, creating a situation where DMA could be occurring to memory regions after process exit (and thus to invalid memory!)

I've fixed this in our code (its still under review and testing internally) -- but I'd like to upstream these fixes too.  (For benefit of anyone who may have concerns about my "credentials" -- I'm well known for my work in NIC drivers (and many others) in Solaris/illumos, and to a lesser extent (and longer ago) NetBSD.

Anyway, I think I have the detailed instructions for submitting changes to DPDK, but as this is code that is associated with a vendor (Google), I thought I'd reach out first -- if there is a specific code owner here I'd be happy to work with them.

My changes are based on a cherry pick of this driver's code from the upstream 24.07, but mostly we (WEKA) are using DPDK 24.03.

Thanks in advance.

[-- Attachment #2: Type: text/html, Size: 1587 bytes --]

Re: Incoming changes for GVNIC

[Thomas Monjalon]

Hello,

Thanks for the heads up.

As you can see the GVE driver was written by Intel & Google.
I Cc the maintainers from Google.

Please make sure they are part of any future communication.
You can use this option when sending patches:
	--cc-cmd devtools/get-maintainer.sh


14/09/2024 22:10, Garrett D'Amore:
> This is mostly a heads up....
> 
> We (WEKA) use various drivers with specific alignment requirements, which causes us to need to need to use multibuffer (scatter/gather) functionality in various drivers.
> 
> Unfortunately, the GVNIC driver was ... very... buggy in this regard (in fact it doesn't work at all in this case!), especially for the DQO mode used on gen 3 (C3) instances.  It turns out it was also buggy in that it was not properly resetting the device on teardown, creating a situation where DMA could be occurring to memory regions after process exit (and thus to invalid memory!)
> 
> I've fixed this in our code (its still under review and testing internally) -- but I'd like to upstream these fixes too.  (For benefit of anyone who may have concerns about my "credentials" -- I'm well known for my work in NIC drivers (and many others) in Solaris/illumos, and to a lesser extent (and longer ago) NetBSD.
> 
> Anyway, I think I have the detailed instructions for submitting changes to DPDK, but as this is code that is associated with a vendor (Google), I thought I'd reach out first -- if there is a specific code owner here I'd be happy to work with them.
> 
> My changes are based on a cherry pick of this driver's code from the upstream 24.07, but mostly we (WEKA) are using DPDK 24.03.
> 
> Thanks in advance.