Re: [RFC PATCH net-next 02/10] ipv4: Add a sysctl to control multipath hash fields

[David Ahern <dsahern@gmail.com>]

On 5/2/21 10:22 AM, Ido Schimmel wrote:
> From: Ido Schimmel <idosch@nvidia.com>
> 
> A subsequent patch will add a new multipath hash policy where the packet
> fields used for multipath hash calculation are determined by user space.
> This patch adds a sysctl that allows user space to set these fields.
> 
> The packet fields are represented using a bitmap and are common between
> IPv4 and IPv6 to allow user space to use the same numbering across both
> protocols. For example, to hash based on standard 5-tuple:
> 
>  # sysctl -w net.ipv4.fib_multipath_hash_fields=0-2,4-5
>  net.ipv4.fib_multipath_hash_fields = 0-2,4-5
> 
> More fields can be added in the future, if needed.
> 
> The 'need_outer' and 'need_inner' variables are set in the control path
> to indicate whether dissection of the outer or inner flow is needed.
> They will be used by a subsequent patch to allow the data path to avoid
> dissection of the outer or inner flow when not needed.
> 
> Signed-off-by: Ido Schimmel <idosch@nvidia.com>
> ---
>  Documentation/networking/ip-sysctl.rst | 29 ++++++++++++++++
>  include/net/ip_fib.h                   | 46 ++++++++++++++++++++++++++
>  include/net/netns/ipv4.h               |  4 +++
>  net/ipv4/fib_frontend.c                | 24 ++++++++++++++
>  net/ipv4/sysctl_net_ipv4.c             | 32 ++++++++++++++++++
>  5 files changed, 135 insertions(+)
> 
> diff --git a/Documentation/networking/ip-sysctl.rst b/Documentation/networking/ip-sysctl.rst
> index c2ecc9894fd0..8ab61f4edf02 100644
> --- a/Documentation/networking/ip-sysctl.rst
> +++ b/Documentation/networking/ip-sysctl.rst
> @@ -100,6 +100,35 @@ fib_multipath_hash_policy - INTEGER
>  	- 1 - Layer 4
>  	- 2 - Layer 3 or inner Layer 3 if present
>  
> +fib_multipath_hash_fields - list of comma separated ranges
> +	When fib_multipath_hash_policy is set to 3 (custom multipath hash), the
> +	fields used for multipath hash calculation are determined by this
> +	sysctl.
> +
> +	The format used for both input and output is a comma separated list of
> +	ranges (e.g., "0-2" for source IP, destination IP and IP protocol).
> +	Writing to the file will clear all previous ranges and update the
> +	current list with the input.
> +
> +	Possible fields are:
> +
> +	== ============================
> +	 0 Source IP address
> +	 1 Destination IP address
> +	 2 IP protocol
> +	 3 Unused
> +	 4 Source port
> +	 5 Destination port
> +	 6 Inner source IP address
> +	 7 Inner destination IP address
> +	 8 Inner IP protocol
> +	 9 Inner Flow Label
> +	10 Inner source port
> +	11 Inner destination port
> +	== ============================
> +
> +	Default: 0-2 (source IP, destination IP and IP protocol)

since you are already requiring a name to id conversion, why not just
use a bitmask here as the input? if the value is a 32-bit bitmask do you
need bitmap_zalloc and its overhead?

Also, you could implement the current default using this scheme since
you have the default fields as the current L3 policy.