On Tue, May 13, 2025 at 12:55:29PM +0200, Michal Kubiak wrote:
> The current implementation of the Tx scheduler allows the tree to be
> rebuilt as the user adds more Tx queues to the VSI. In such a case,
> additional child nodes are added to the tree to support the new number
> of queues.
> Unfortunately, this algorithm does not take into account that the limit
> of the VSI support node may be exceeded, so an additional node in the
> VSI layer may be required to handle all the requested queues.
>
> Such a scenario occurs when adding XDP Tx queues on machines with many
> CPUs. Although the driver still respects the queue limit returned by
> the FW, the Tx scheduler was unable to add those queues to its tree
> and returned one of the errors below.
>
> Such a scenario occurs when adding XDP Tx queues on machines with many
> CPUs (e.g. at least 321 CPUs, if there is already 128 Tx/Rx queue pairs).
> Although the driver still respects the queue limit returned by the FW,
> the Tx scheduler was unable to add those queues to its tree and returned
> the following errors:
>
> Failed VSI LAN queue config for XDP, error: -5
> or:
> Failed to set LAN Tx queue context, error: -22
>
> Fix this problem by extending the tree rebuild algorithm to check if the
> current VSI node can support the requested number of queues. If it
> cannot, create as many additional VSI support nodes as necessary to
> handle all the required Tx queues. Symmetrically, adjust the VSI node
> removal algorithm to remove all nodes associated with the given VSI.
> Also, make the search for the next free VSI node more restrictive. That is,
> add queue group nodes only to the VSI support nodes that have a matching
> VSI handle.
> Finally, fix the comment describing the tree update algorithm to better
> reflect the current scenario.
>
> Fixes: b0153fdd7e8a ("ice: update VSI config dynamically")
> Reviewed-by: Dawid Osuchowski <[email protected]>
> Reviewed-by: Przemek Kitszel <[email protected]>
> Signed-off-by: Michal Kubiak <[email protected]>
Reviewed-by: Simon Horman <[email protected]>
