On 11/18/2020 5:17 PM, David Ahern wrote:
On 11/18/20 5:41 PM, Jacob Keller wrote:
On 11/18/2020 11:22 AM, Parav Pandit wrote:
From: David Ahern <dsahern@xxxxxxxxx>
Sent: Wednesday, November 18, 2020 11:33 PM
With Connectx-4 Lx for example the netdev can have at most 63 queues
leaving 96 cpu servers a bit short - as an example of the limited number of
queues that a nic can handle (or currently exposes to the OS not sure which).
If I create a subfunction for ethernet traffic, how many queues are allocated
to it by default, is it managed via ethtool like the pf and is there an impact to
the resources used by / available to the primary device?
Jason already answered it with details.
Thanks a lot Jason.
Short answer to ethtool question, yes, ethtool can change num queues for subfunction like PF.
Default is same number of queues for subfunction as that of PF in this patchset.
But what is the mechanism for partitioning the global resources of the
device into each sub function?
Is it just evenly divided into the subfunctions? is there some maximum
limit per sub function?
I hope it is not just evenly divided; it should be user controllable. If
I create a subfunction for say a container's networking, I may want to
only assign 1 Rx and 1 Tx queue pair (or 1 channel depending on
terminology where channel includes Rx, Tx and CQ).
I think we need a way to expose and configure policy for resources
associated with each type of auxiliary device.
For ex: default, min and max queues and interrupt vectors.
Once an auxiliary device is created, the user should be able to
configure the resources within the allowed min-max values.
